diff --git a/CHANGELOG.md b/CHANGELOG.md
index 04ea1c34..5af293ab 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,13 @@ documented here.
 
 This project adheres to [Semantic Versioning](https://semver.org/).
 
+## [0.29.0]
+### Modified
+- The closure given to `apply`, `zip_apply`, `zip_zip_apply` must now modify the
+  first argument inplace, instead of returning a new value. This makes these
+  methods more versatile, and avoid useless clones when using non-Copy scalar
+  types.
+
 ## [0.28.0]
 ### Added
 - Implement `Hash` for `Transform`.
diff --git a/benches/core/matrix.rs b/benches/core/matrix.rs
index d13d54e9..3c483c35 100644
--- a/benches/core/matrix.rs
+++ b/benches/core/matrix.rs
@@ -1,7 +1,4 @@
-use na::{
-    Const, DMatrix, DVector, Dynamic, Matrix2, Matrix3, Matrix4, OMatrix, Vector2, Vector3,
-    Vector4, U10,
-};
+use na::{DMatrix, DVector, Matrix2, Matrix3, Matrix4, OMatrix, Vector2, Vector3, Vector4, U10};
 use rand::Rng;
 use rand_isaac::IsaacRng;
 use std::ops::{Add, Div, Mul, Sub};
@@ -189,7 +186,7 @@ fn axpy(bench: &mut criterion::Criterion) {
 fn tr_mul_to(bench: &mut criterion::Criterion) {
     let a = DMatrix::<f64>::new_random(1000, 1000);
     let b = DVector::<f64>::new_random(1000);
-    let mut c = DVector::new_uninitialized_generic(Dynamic::new(1000), Const::<1>);
+    let mut c = DVector::from_element(1000, 0.0);
 
     bench.bench_function("tr_mul_to", move |bh| bh.iter(|| a.tr_mul_to(&b, &mut c)));
 }
@@ -197,7 +194,7 @@ fn tr_mul_to(bench: &mut criterion::Criterion) {
 fn mat_mul_mat(bench: &mut criterion::Criterion) {
     let a = DMatrix::<f64>::new_random(100, 100);
     let b = DMatrix::<f64>::new_random(100, 100);
-    let mut ab = DMatrix::new_uninitialized_generic(Dynamic::new(100), Dynamic::new(100));
+    let mut ab = DMatrix::<f64>::from_element(100, 100, 0.0);
 
     bench.bench_function("mat_mul_mat", move |bh| {
         bh.iter(|| {
diff --git a/nalgebra-lapack/src/cholesky.rs b/nalgebra-lapack/src/cholesky.rs
index 929f2d40..ea4b1d94 100644
--- a/nalgebra-lapack/src/cholesky.rs
+++ b/nalgebra-lapack/src/cholesky.rs
@@ -6,7 +6,7 @@ use num_complex::Complex;
 
 use na::allocator::Allocator;
 use na::dimension::Dim;
-use na::storage::Storage;
+use na::storage::RawStorage;
 use na::{DefaultAllocator, Matrix, OMatrix, Scalar};
 
 use lapack;
@@ -24,17 +24,17 @@ use lapack;
          OMatrix<T, D, D>: Deserialize<'de>"))
 )]
 #[derive(Clone, Debug)]
-pub struct Cholesky<T, D: Dim>
+pub struct Cholesky<T: Scalar, D: Dim>
 where
     DefaultAllocator: Allocator<T, D, D>,
 {
     l: OMatrix<T, D, D>,
 }
 
-impl<T: Copy, D: Dim> Copy for Cholesky<T, D>
+impl<T: Scalar + Copy, D: Dim> Copy for Cholesky<T, D>
 where
     DefaultAllocator: Allocator<T, D, D>,
-    Owned<T, D, D>: Copy,
+    OMatrix<T, D, D>: Copy,
 {
 }
 
@@ -104,7 +104,7 @@ where
         b: &Matrix<T, R2, C2, S2>,
     ) -> Option<OMatrix<T, R2, C2>>
     where
-        S2: Storage<T, R2, C2>,
+        S2: RawStorage<T, R2, C2>,
         DefaultAllocator: Allocator<T, R2, C2>,
     {
         let mut res = b.clone_owned();
diff --git a/nalgebra-lapack/src/eigen.rs b/nalgebra-lapack/src/eigen.rs
index 49fb72b4..a8f87d85 100644
--- a/nalgebra-lapack/src/eigen.rs
+++ b/nalgebra-lapack/src/eigen.rs
@@ -1,5 +1,3 @@
-use std::fmt;
-
 #[cfg(feature = "serde-serialize")]
 use serde::{Deserialize, Serialize};
 
@@ -11,7 +9,7 @@ use simba::scalar::RealField;
 use crate::ComplexHelper;
 use na::allocator::Allocator;
 use na::dimension::{Const, Dim};
-use na::storage::Storage;
+use na::storage::RawStorage;
 use na::{DefaultAllocator, Matrix, OMatrix, OVector, Scalar};
 
 use lapack;
@@ -34,7 +32,8 @@ use lapack;
          OMatrix<T, D, D>: Deserialize<'de>")
     )
 )]
-pub struct Eigen<T, D: Dim>
+#[derive(Clone, Debug)]
+pub struct Eigen<T: Scalar, D: Dim>
 where
     DefaultAllocator: Allocator<T, D> + Allocator<T, D, D>,
 {
@@ -46,7 +45,7 @@ where
     pub left_eigenvectors: Option<OMatrix<T, D, D>>,
 }
 
-impl<T: Copy, D: Dim> Copy for Eigen<T, D>
+impl<T: Scalar + Copy, D: Dim> Copy for Eigen<T, D>
 where
     DefaultAllocator: Allocator<T, D> + Allocator<T, D, D>,
     OVector<T, D>: Copy,
@@ -54,36 +53,6 @@ where
 {
 }
 
-impl<T: Clone, D: Dim> Clone for Eigen<T, D>
-where
-    DefaultAllocator: Allocator<T, D> + Allocator<T, D, D>,
-    OVector<T, D>: Clone,
-    OMatrix<T, D, D>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self {
-            eigenvalues: self.eigenvalues.clone(),
-            eigenvectors: self.eigenvectors.clone(),
-            left_eigenvectors: self.left_eigenvectors.clone(),
-        }
-    }
-}
-
-impl<T: fmt::Debug, D: Dim> fmt::Debug for Eigen<T, D>
-where
-    DefaultAllocator: Allocator<T, D> + Allocator<T, D, D>,
-    OVector<T, D>: fmt::Debug,
-    OMatrix<T, D, D>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("Eigen")
-            .field("eigenvalues", &self.eigenvalues)
-            .field("eigenvectors", &self.eigenvectors)
-            .field("left_eigenvectors", &self.left_eigenvectors)
-            .finish()
-    }
-}
-
 impl<T: EigenScalar + RealField, D: Dim> Eigen<T, D>
 where
     DefaultAllocator: Allocator<T, D, D> + Allocator<T, D>,
@@ -104,13 +73,11 @@ where
         let ljob = if left_eigenvectors { b'V' } else { b'T' };
         let rjob = if eigenvectors { b'V' } else { b'T' };
 
-        let (nrows, ncols) = m.data.shape();
+        let (nrows, ncols) = m.shape_generic();
         let n = nrows.value();
 
         let lda = n as i32;
 
-        // IMPORTANT TODO: this is still UB.
-
         let mut wr = unsafe { Matrix::new_uninitialized_generic(nrows, Const::<1>).assume_init() };
         // TODO: Tap into the workspace.
         let mut wi = unsafe { Matrix::new_uninitialized_generic(nrows, Const::<1>).assume_init() };
@@ -275,7 +242,7 @@ where
             "Unable to compute the eigenvalue decomposition of a non-square matrix."
         );
 
-        let nrows = m.data.shape().0;
+        let nrows = m.shape_generic().0;
         let n = nrows.value();
 
         let lda = n as i32;
diff --git a/nalgebra-lapack/src/hessenberg.rs b/nalgebra-lapack/src/hessenberg.rs
index dab38c40..7f854cb6 100644
--- a/nalgebra-lapack/src/hessenberg.rs
+++ b/nalgebra-lapack/src/hessenberg.rs
@@ -4,7 +4,7 @@ use num_complex::Complex;
 use crate::ComplexHelper;
 use na::allocator::Allocator;
 use na::dimension::{Const, DimDiff, DimSub, U1};
-use na::storage::Storage;
+use na::storage::RawStorage;
 use na::{DefaultAllocator, Matrix, OMatrix, OVector, Scalar};
 
 use lapack;
@@ -48,7 +48,7 @@ where
 {
     /// Computes the hessenberg decomposition of the matrix `m`.
     pub fn new(mut m: OMatrix<T, D, D>) -> Self {
-        let nrows = m.data.shape().0;
+        let nrows = m.shape_generic().0;
         let n = nrows.value() as i32;
 
         assert!(
@@ -60,7 +60,6 @@ where
             "Unable to compute the hessenberg decomposition of an empty matrix."
         );
 
-        // IMPORTANT TODO: this is still UB.
         let mut tau = unsafe {
             Matrix::new_uninitialized_generic(nrows.sub(Const::<1>), Const::<1>).assume_init()
         };
diff --git a/nalgebra-lapack/src/lib.rs b/nalgebra-lapack/src/lib.rs
index fccf2717..9a027772 100644
--- a/nalgebra-lapack/src/lib.rs
+++ b/nalgebra-lapack/src/lib.rs
@@ -140,7 +140,6 @@ impl ComplexHelper for Complex<f64> {
     }
 }
 
-// This is UB.
 unsafe fn uninitialized_vec<T: Copy>(n: usize) -> Vec<T> {
     let mut res = Vec::new();
     res.reserve_exact(n);
diff --git a/nalgebra-lapack/src/lu.rs b/nalgebra-lapack/src/lu.rs
index 2130fc7e..7d4a5a43 100644
--- a/nalgebra-lapack/src/lu.rs
+++ b/nalgebra-lapack/src/lu.rs
@@ -4,7 +4,7 @@ use num_complex::Complex;
 use crate::ComplexHelper;
 use na::allocator::Allocator;
 use na::dimension::{Const, Dim, DimMin, DimMinimum};
-use na::storage::Storage;
+use na::storage::RawStorage;
 use na::{DefaultAllocator, Matrix, OMatrix, OVector, Scalar};
 
 use lapack;
@@ -61,7 +61,7 @@ where
 {
     /// Computes the LU decomposition with partial (row) pivoting of `matrix`.
     pub fn new(mut m: OMatrix<T, R, C>) -> Self {
-        let (nrows, ncols) = m.data.shape();
+        let (nrows, ncols) = m.shape_generic();
         let min_nrows_ncols = nrows.min(ncols);
         let nrows = nrows.value() as i32;
         let ncols = ncols.value() as i32;
@@ -87,7 +87,7 @@ where
     #[inline]
     #[must_use]
     pub fn l(&self) -> OMatrix<T, R, DimMinimum<R, C>> {
-        let (nrows, ncols) = self.lu.data.shape();
+        let (nrows, ncols) = self.lu.shape_generic();
         let mut res = self.lu.columns_generic(0, nrows.min(ncols)).into_owned();
 
         res.fill_upper_triangle(Zero::zero(), 1);
@@ -100,7 +100,7 @@ where
     #[inline]
     #[must_use]
     pub fn u(&self) -> OMatrix<T, DimMinimum<R, C>, C> {
-        let (nrows, ncols) = self.lu.data.shape();
+        let (nrows, ncols) = self.lu.shape_generic();
         let mut res = self.lu.rows_generic(0, nrows.min(ncols)).into_owned();
 
         res.fill_lower_triangle(Zero::zero(), 1);
@@ -115,7 +115,7 @@ where
     #[inline]
     #[must_use]
     pub fn p(&self) -> OMatrix<T, R, R> {
-        let (dim, _) = self.lu.data.shape();
+        let (dim, _) = self.lu.shape_generic();
         let mut id = Matrix::identity_generic(dim, dim);
         self.permute(&mut id);
 
@@ -191,7 +191,7 @@ where
         b: &Matrix<T, R2, C2, S2>,
     ) -> Option<OMatrix<T, R2, C2>>
     where
-        S2: Storage<T, R2, C2>,
+        S2: RawStorage<T, R2, C2>,
         DefaultAllocator: Allocator<T, R2, C2> + Allocator<i32, R2>,
     {
         let mut res = b.clone_owned();
@@ -209,7 +209,7 @@ where
         b: &Matrix<T, R2, C2, S2>,
     ) -> Option<OMatrix<T, R2, C2>>
     where
-        S2: Storage<T, R2, C2>,
+        S2: RawStorage<T, R2, C2>,
         DefaultAllocator: Allocator<T, R2, C2> + Allocator<i32, R2>,
     {
         let mut res = b.clone_owned();
@@ -227,7 +227,7 @@ where
         b: &Matrix<T, R2, C2, S2>,
     ) -> Option<OMatrix<T, R2, C2>>
     where
-        S2: Storage<T, R2, C2>,
+        S2: RawStorage<T, R2, C2>,
         DefaultAllocator: Allocator<T, R2, C2> + Allocator<i32, R2>,
     {
         let mut res = b.clone_owned();
diff --git a/nalgebra-lapack/src/qr.rs b/nalgebra-lapack/src/qr.rs
index 2209b86e..dc4d81d7 100644
--- a/nalgebra-lapack/src/qr.rs
+++ b/nalgebra-lapack/src/qr.rs
@@ -7,7 +7,7 @@ use num_complex::Complex;
 use crate::ComplexHelper;
 use na::allocator::Allocator;
 use na::dimension::{Const, Dim, DimMin, DimMinimum};
-use na::storage::Storage;
+use na::storage::RawStorage;
 use na::{DefaultAllocator, Matrix, OMatrix, OVector, Scalar};
 
 use lapack;
@@ -54,11 +54,12 @@ where
 {
     /// Computes the QR decomposition of the matrix `m`.
     pub fn new(mut m: OMatrix<T, R, C>) -> Self {
-        let (nrows, ncols) = m.data.shape();
+        let (nrows, ncols) = m.shape_generic();
 
         let mut info = 0;
-        let mut tau =
-            unsafe { Matrix::new_uninitialized_generic(nrows.min(ncols), U1).assume_init() };
+        let mut tau = unsafe {
+            Matrix::new_uninitialized_generic(nrows.min(ncols), Const::<1>).assume_init()
+        };
 
         if nrows.value() == 0 || ncols.value() == 0 {
             return Self { qr: m, tau };
@@ -93,7 +94,7 @@ where
     #[inline]
     #[must_use]
     pub fn r(&self) -> OMatrix<T, DimMinimum<R, C>, C> {
-        let (nrows, ncols) = self.qr.data.shape();
+        let (nrows, ncols) = self.qr.shape_generic();
         self.qr.rows_generic(0, nrows.min(ncols)).upper_triangle()
     }
 }
@@ -119,7 +120,7 @@ where
     #[inline]
     #[must_use]
     pub fn q(&self) -> OMatrix<T, R, DimMinimum<R, C>> {
-        let (nrows, ncols) = self.qr.data.shape();
+        let (nrows, ncols) = self.qr.shape_generic();
         let min_nrows_ncols = nrows.min(ncols);
 
         if min_nrows_ncols.value() == 0 {
diff --git a/nalgebra-lapack/src/schur.rs b/nalgebra-lapack/src/schur.rs
index 7325bb8f..9543fea2 100644
--- a/nalgebra-lapack/src/schur.rs
+++ b/nalgebra-lapack/src/schur.rs
@@ -9,7 +9,7 @@ use simba::scalar::RealField;
 use crate::ComplexHelper;
 use na::allocator::Allocator;
 use na::dimension::{Const, Dim};
-use na::storage::Storage;
+use na::storage::RawStorage;
 use na::{DefaultAllocator, Matrix, OMatrix, OVector, Scalar};
 
 use lapack;
@@ -71,7 +71,7 @@ where
             "Unable to compute the eigenvalue decomposition of a non-square matrix."
         );
 
-        let (nrows, ncols) = m.data.shape();
+        let (nrows, ncols) = m.shape_generic();
         let n = nrows.value();
 
         let lda = n as i32;
@@ -153,15 +153,15 @@ where
     where
         DefaultAllocator: Allocator<Complex<T>, D>,
     {
-        let mut out =
-            unsafe { OVector::new_uninitialized_generic(self.t.data.shape().0, Const::<1>) };
+        let mut out = unsafe {
+            OVector::new_uninitialized_generic(self.t.shape_generic().0, Const::<1>).assume_init()
+        };
 
         for i in 0..out.len() {
-            out[i] = MaybeUninit::new(Complex::new(self.re[i], self.im[i]));
+            out[i] = Complex::new(self.re[i], self.im[i])
         }
 
-        // Safety: all entries have been initialized.
-        unsafe { out.assume_init() }
+        out
     }
 }
 
diff --git a/nalgebra-lapack/src/svd.rs b/nalgebra-lapack/src/svd.rs
index 5bf4758a..872c368d 100644
--- a/nalgebra-lapack/src/svd.rs
+++ b/nalgebra-lapack/src/svd.rs
@@ -6,7 +6,7 @@ use std::cmp;
 
 use na::allocator::Allocator;
 use na::dimension::{Const, Dim, DimMin, DimMinimum, U1};
-use na::storage::Storage;
+use na::storage::RawStorage;
 use na::{DefaultAllocator, Matrix, OMatrix, OVector, Scalar};
 
 use lapack;
@@ -89,7 +89,7 @@ macro_rules! svd_impl(
                                         Allocator<$t, DimMinimum<R, C>> {
 
             fn compute(mut m: OMatrix<$t, R, C>) -> Option<SVD<$t, R, C>> {
-                let (nrows, ncols) = m.data.shape();
+                let (nrows, ncols) = m.shape_generic();
 
                 if nrows.value() == 0 || ncols.value() == 0 {
                     return None;
@@ -99,7 +99,6 @@ macro_rules! svd_impl(
 
                 let lda = nrows.value() as i32;
 
-                // IMPORTANT TODO: this is still UB.
                 let mut u  = unsafe { Matrix::new_uninitialized_generic(nrows, nrows).assume_init() };
                 let mut s  = unsafe { Matrix::new_uninitialized_generic(nrows.min(ncols), Const::<1>).assume_init() };
                 let mut vt = unsafe { Matrix::new_uninitialized_generic(ncols, ncols).assume_init() };
@@ -152,8 +151,8 @@ macro_rules! svd_impl(
             /// been manually changed by the user.
             #[inline]
             pub fn recompose(self) -> OMatrix<$t, R, C> {
-                let nrows           = self.u.data.shape().0;
-                let ncols           = self.vt.data.shape().1;
+                let nrows           = self.u.shape_generic().0;
+                let ncols           = self.vt.shape_generic().1;
                 let min_nrows_ncols = nrows.min(ncols);
 
                 let mut res: OMatrix<_, R, C> = Matrix::zeros_generic(nrows, ncols);
@@ -178,8 +177,8 @@ macro_rules! svd_impl(
             #[inline]
             #[must_use]
             pub fn pseudo_inverse(&self, epsilon: $t) -> OMatrix<$t, C, R> {
-                let nrows           = self.u.data.shape().0;
-                let ncols           = self.vt.data.shape().1;
+                let nrows           = self.u.shape_generic().0;
+                let ncols           = self.vt.shape_generic().1;
                 let min_nrows_ncols = nrows.min(ncols);
 
                 let mut res: OMatrix<_, C, R> = Matrix::zeros_generic(ncols, nrows);
@@ -242,7 +241,7 @@ macro_rules! svd_complex_impl(
                             Allocator<Complex<$t>, R, R>         +
                             Allocator<Complex<$t>, C, C>         +
                             Allocator<$t, DimMinimum<R, C>> {
-            let (nrows, ncols) = m.data.shape();
+            let (nrows, ncols) = m.shape_generic();
 
             if nrows.value() == 0 || ncols.value() == 0 {
                 return None;
diff --git a/nalgebra-lapack/src/symmetric_eigen.rs b/nalgebra-lapack/src/symmetric_eigen.rs
index 7a1f6f2e..f70e9a4d 100644
--- a/nalgebra-lapack/src/symmetric_eigen.rs
+++ b/nalgebra-lapack/src/symmetric_eigen.rs
@@ -9,7 +9,7 @@ use simba::scalar::RealField;
 use crate::ComplexHelper;
 use na::allocator::Allocator;
 use na::dimension::{Const, Dim};
-use na::storage::Storage;
+use na::storage::RawStorage;
 use na::{DefaultAllocator, Matrix, OMatrix, OVector, Scalar};
 
 use lapack;
@@ -89,12 +89,11 @@ where
 
         let jobz = if eigenvectors { b'V' } else { b'T' };
 
-        let nrows = m.data.shape().0;
+        let nrows = m.shape_generic().0;
         let n = nrows.value();
 
         let lda = n as i32;
 
-        // IMPORTANT TODO: this is still UB.
         let mut values =
             unsafe { Matrix::new_uninitialized_generic(nrows, Const::<1>).assume_init() };
         let mut info = 0;
diff --git a/nalgebra-sparse/src/convert/impl_std_ops.rs b/nalgebra-sparse/src/convert/impl_std_ops.rs
index d775fa13..683227e2 100644
--- a/nalgebra-sparse/src/convert/impl_std_ops.rs
+++ b/nalgebra-sparse/src/convert/impl_std_ops.rs
@@ -2,14 +2,16 @@ use crate::convert::serial::*;
 use crate::coo::CooMatrix;
 use crate::csc::CscMatrix;
 use crate::csr::CsrMatrix;
-use nalgebra::storage::Storage;
+use nalgebra::storage::RawStorage;
 use nalgebra::{ClosedAdd, DMatrix, Dim, Matrix, Scalar};
 use num_traits::Zero;
 
-impl<'a, T, R: Dim, C: Dim, S> From<&'a Matrix<T, R, C, S>> for CooMatrix<T>
+impl<'a, T, R, C, S> From<&'a Matrix<T, R, C, S>> for CooMatrix<T>
 where
-    T: Scalar + Zero + PartialEq,
-    S: Storage<T, R, C>,
+    T: Scalar + Zero,
+    R: Dim,
+    C: Dim,
+    S: RawStorage<T, R, C>,
 {
     fn from(matrix: &'a Matrix<T, R, C, S>) -> Self {
         convert_dense_coo(matrix)
@@ -43,10 +45,12 @@ where
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S> From<&'a Matrix<T, R, C, S>> for CsrMatrix<T>
+impl<'a, T, R, C, S> From<&'a Matrix<T, R, C, S>> for CsrMatrix<T>
 where
-    T: Scalar + Zero + PartialEq,
-    S: Storage<T, R, C>,
+    T: Scalar + Zero,
+    R: Dim,
+    C: Dim,
+    S: RawStorage<T, R, C>,
 {
     fn from(matrix: &'a Matrix<T, R, C, S>) -> Self {
         convert_dense_csr(matrix)
@@ -80,10 +84,12 @@ where
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S> From<&'a Matrix<T, R, C, S>> for CscMatrix<T>
+impl<'a, T, R, C, S> From<&'a Matrix<T, R, C, S>> for CscMatrix<T>
 where
-    T: Scalar + Zero + PartialEq,
-    S: Storage<T, R, C>,
+    T: Scalar + Zero,
+    R: Dim,
+    C: Dim,
+    S: RawStorage<T, R, C>,
 {
     fn from(matrix: &'a Matrix<T, R, C, S>) -> Self {
         convert_dense_csc(matrix)
diff --git a/nalgebra-sparse/src/convert/serial.rs b/nalgebra-sparse/src/convert/serial.rs
index ebdf4e65..f84a6583 100644
--- a/nalgebra-sparse/src/convert/serial.rs
+++ b/nalgebra-sparse/src/convert/serial.rs
@@ -7,7 +7,7 @@ use std::ops::Add;
 
 use num_traits::Zero;
 
-use nalgebra::storage::Storage;
+use nalgebra::storage::RawStorage;
 use nalgebra::{ClosedAdd, DMatrix, Dim, Matrix, Scalar};
 
 use crate::coo::CooMatrix;
@@ -16,10 +16,12 @@ use crate::csc::CscMatrix;
 use crate::csr::CsrMatrix;
 
 /// Converts a dense matrix to [`CooMatrix`].
-pub fn convert_dense_coo<T, R: Dim, C: Dim, S>(dense: &Matrix<T, R, C, S>) -> CooMatrix<T>
+pub fn convert_dense_coo<T, R, C, S>(dense: &Matrix<T, R, C, S>) -> CooMatrix<T>
 where
-    T: Scalar + Zero + PartialEq,
-    S: Storage<T, R, C>,
+    T: Scalar + Zero,
+    R: Dim,
+    C: Dim,
+    S: RawStorage<T, R, C>,
 {
     let mut coo = CooMatrix::new(dense.nrows(), dense.ncols());
 
@@ -91,10 +93,10 @@ where
 /// Converts a dense matrix to a [`CsrMatrix`].
 pub fn convert_dense_csr<T, R, C, S>(dense: &Matrix<T, R, C, S>) -> CsrMatrix<T>
 where
-    T: Scalar + Zero + PartialEq,
+    T: Scalar + Zero,
     R: Dim,
     C: Dim,
-    S: Storage<T, R, C>,
+    S: RawStorage<T, R, C>,
 {
     let mut row_offsets = Vec::with_capacity(dense.nrows() + 1);
     let mut col_idx = Vec::new();
@@ -168,10 +170,10 @@ where
 /// Converts a dense matrix to a [`CscMatrix`].
 pub fn convert_dense_csc<T, R, C, S>(dense: &Matrix<T, R, C, S>) -> CscMatrix<T>
 where
-    T: Scalar + Zero + PartialEq,
+    T: Scalar + Zero,
     R: Dim,
     C: Dim,
-    S: Storage<T, R, C>,
+    S: RawStorage<T, R, C>,
 {
     let mut col_offsets = Vec::with_capacity(dense.ncols() + 1);
     let mut row_idx = Vec::new();
diff --git a/nalgebra-sparse/src/coo.rs b/nalgebra-sparse/src/coo.rs
index 679dbdb2..34e5ceec 100644
--- a/nalgebra-sparse/src/coo.rs
+++ b/nalgebra-sparse/src/coo.rs
@@ -57,7 +57,7 @@ impl<T: na::Scalar> CooMatrix<T> {
     /// Panics if any part of the dense matrix is out of bounds of the sparse matrix
     /// when inserted at `(r, c)`.
     #[inline]
-    pub fn push_matrix<R: na::Dim, C: na::Dim, S: nalgebra::storage::Storage<T, R, C>>(
+    pub fn push_matrix<R: na::Dim, C: na::Dim, S: nalgebra::storage::RawStorage<T, R, C>>(
         &mut self,
         r: usize,
         c: usize,
diff --git a/nalgebra-sparse/src/ops/impl_std_ops.rs b/nalgebra-sparse/src/ops/impl_std_ops.rs
index 11d59ded..721023a5 100644
--- a/nalgebra-sparse/src/ops/impl_std_ops.rs
+++ b/nalgebra-sparse/src/ops/impl_std_ops.rs
@@ -6,8 +6,8 @@ use crate::ops::serial::{
     spmm_csc_prealloc, spmm_csr_dense, spmm_csr_pattern, spmm_csr_prealloc,
 };
 use crate::ops::Op;
-use nalgebra::allocator::{Allocator, InnerAllocator};
-use nalgebra::base::storage::Storage;
+use nalgebra::allocator::Allocator;
+use nalgebra::base::storage::RawStorage;
 use nalgebra::constraint::{DimEq, ShapeConstraint};
 use nalgebra::{
     ClosedAdd, ClosedDiv, ClosedMul, ClosedSub, DefaultAllocator, Dim, Dynamic, Matrix, OMatrix,
@@ -28,7 +28,7 @@ macro_rules! impl_bin_op {
             // Note: The Neg bound is currently required because we delegate e.g.
             // Sub to SpAdd with negative coefficients. This is not well-defined for
             // unsigned data types.
-            $($scalar_type: $($bounds + )? Scalar + ClosedAdd + ClosedSub + ClosedMul + Zero + One + Neg<Output=T> + PartialEq)?
+            $($scalar_type: $($bounds + )? Scalar + ClosedAdd + ClosedSub + ClosedMul + Zero + One + Neg<Output=T>)?
         {
             type Output = $ret;
             fn $method(self, $b: $b_type) -> Self::Output {
@@ -272,7 +272,7 @@ macro_rules! impl_spmm_cs_dense {
     ($matrix_type_name:ident, $spmm_fn:ident) => {
         // Implement ref-ref
         impl_spmm_cs_dense!(&'a $matrix_type_name<T>, &'a Matrix<T, R, C, S>, $spmm_fn, |lhs, rhs| {
-            let (_, ncols) = rhs.data.shape();
+            let (_, ncols) = rhs.shape_generic();
             let nrows = Dynamic::new(lhs.nrows());
             let mut result = OMatrix::<T, Dynamic, C>::zeros_generic(nrows, ncols);
             $spmm_fn(T::zero(), &mut result, T::one(), Op::NoOp(lhs), Op::NoOp(rhs));
@@ -301,14 +301,14 @@ macro_rules! impl_spmm_cs_dense {
             T: Scalar + ClosedMul + ClosedAdd + ClosedSub + ClosedDiv + Neg + Zero + One,
             R: Dim,
             C: Dim,
-            S: Storage<T, R, C>,
+            S: RawStorage<T, R, C>,
             DefaultAllocator: Allocator<T, Dynamic, C>,
             // TODO: Is it possible to simplify these bounds?
             ShapeConstraint:
                 // Bounds so that we can turn OMatrix<T, Dynamic, C> into a DMatrixSliceMut
-                  DimEq<U1, <<DefaultAllocator as InnerAllocator<T, Dynamic, C>>::Buffer as Storage<T, Dynamic, C>>::RStride>
+                  DimEq<U1, <<DefaultAllocator as Allocator<T, Dynamic, C>>::Buffer as RawStorage<T, Dynamic, C>>::RStride>
                 + DimEq<C, Dynamic>
-                + DimEq<Dynamic, <<DefaultAllocator as InnerAllocator<T, Dynamic, C>>::Buffer as Storage<T, Dynamic, C>>::CStride>
+                + DimEq<Dynamic, <<DefaultAllocator as Allocator<T, Dynamic, C>>::Buffer as RawStorage<T, Dynamic, C>>::CStride>
                 // Bounds so that we can turn &Matrix<T, R, C, S> into a DMatrixSlice
                 + DimEq<U1, S::RStride>
                 + DimEq<R, Dynamic>
diff --git a/nalgebra-sparse/src/ops/serial/cs.rs b/nalgebra-sparse/src/ops/serial/cs.rs
index 9c799339..db057705 100644
--- a/nalgebra-sparse/src/ops/serial/cs.rs
+++ b/nalgebra-sparse/src/ops/serial/cs.rs
@@ -74,7 +74,7 @@ pub fn spadd_cs_prealloc<T>(
     a: Op<&CsMatrix<T>>,
 ) -> Result<(), OperationError>
 where
-    T: Scalar + ClosedAdd + ClosedMul + Zero + One + PartialEq,
+    T: Scalar + ClosedAdd + ClosedMul + Zero + One,
 {
     match a {
         Op::NoOp(a) => {
diff --git a/nalgebra-sparse/src/ops/serial/csc.rs b/nalgebra-sparse/src/ops/serial/csc.rs
index 5f6868c1..25e59f26 100644
--- a/nalgebra-sparse/src/ops/serial/csc.rs
+++ b/nalgebra-sparse/src/ops/serial/csc.rs
@@ -55,7 +55,7 @@ pub fn spadd_csc_prealloc<T>(
     a: Op<&CscMatrix<T>>,
 ) -> Result<(), OperationError>
 where
-    T: Scalar + ClosedAdd + ClosedMul + Zero + One + PartialEq,
+    T: Scalar + ClosedAdd + ClosedMul + Zero + One,
 {
     assert_compatible_spadd_dims!(c, a);
     spadd_cs_prealloc(beta, &mut c.cs, alpha, a.map_same_op(|a| &a.cs))
diff --git a/nalgebra-sparse/src/ops/serial/csr.rs b/nalgebra-sparse/src/ops/serial/csr.rs
index dc8e937b..fa317bbf 100644
--- a/nalgebra-sparse/src/ops/serial/csr.rs
+++ b/nalgebra-sparse/src/ops/serial/csr.rs
@@ -50,7 +50,7 @@ pub fn spadd_csr_prealloc<T>(
     a: Op<&CsrMatrix<T>>,
 ) -> Result<(), OperationError>
 where
-    T: Scalar + ClosedAdd + ClosedMul + Zero + One + PartialEq,
+    T: Scalar + ClosedAdd + ClosedMul + Zero + One,
 {
     assert_compatible_spadd_dims!(c, a);
     spadd_cs_prealloc(beta, &mut c.cs, alpha, a.map_same_op(|a| &a.cs))
diff --git a/nalgebra-sparse/src/pattern.rs b/nalgebra-sparse/src/pattern.rs
index 8bc71075..85f6bc1a 100644
--- a/nalgebra-sparse/src/pattern.rs
+++ b/nalgebra-sparse/src/pattern.rs
@@ -311,7 +311,7 @@ impl From<SparsityPatternFormatError> for SparseFormatError {
 }
 
 impl fmt::Display for SparsityPatternFormatError {
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         match self {
             SparsityPatternFormatError::InvalidOffsetArrayLength => {
                 write!(f, "Length of offset array is not equal to (major_dim + 1).")
diff --git a/src/base/alias.rs b/src/base/alias.rs
index a866935d..68829d9a 100644
--- a/src/base/alias.rs
+++ b/src/base/alias.rs
@@ -1,9 +1,12 @@
 #[cfg(any(feature = "alloc", feature = "std"))]
 use crate::base::dimension::Dynamic;
 use crate::base::dimension::{U1, U2, U3, U4, U5, U6};
+use crate::base::storage::Owned;
 #[cfg(any(feature = "std", feature = "alloc"))]
 use crate::base::vec_storage::VecStorage;
-use crate::base::{ArrayStorage, Const, Matrix, Owned, Unit};
+use crate::base::{ArrayStorage, Const, Matrix, Unit};
+use crate::storage::OwnedUninit;
+use std::mem::MaybeUninit;
 
 /*
  *
@@ -18,13 +21,16 @@ use crate::base::{ArrayStorage, Const, Matrix, Owned, Unit};
 /// **Because this is an alias, not all its methods are listed here. See the [`Matrix`](crate::base::Matrix) type too.**
 pub type OMatrix<T, R, C> = Matrix<T, R, C, Owned<T, R, C>>;
 
+/// An owned matrix with uninitialized data.
+pub type UninitMatrix<T, R, C> = Matrix<MaybeUninit<T>, R, C, OwnedUninit<T, R, C>>;
+
 /// An owned matrix column-major matrix with `R` rows and `C` columns.
 ///
 /// **Because this is an alias, not all its methods are listed here. See the [`Matrix`](crate::base::Matrix) type too.**
 #[deprecated(
     note = "use SMatrix for a statically-sized matrix using integer dimensions, or OMatrix for an owned matrix using types as dimensions."
 )]
-pub type MatrixMN<T, R, C> = OMatrix<T, R, C>;
+pub type MatrixMN<T, R, C> = Matrix<T, R, C, Owned<T, R, C>>;
 
 /// An owned matrix column-major matrix with `D` columns.
 ///
@@ -277,6 +283,9 @@ pub type OVector<T, D> = Matrix<T, D, U1, Owned<T, D, U1>>;
 /// A statically sized D-dimensional column vector.
 pub type SVector<T, const D: usize> = Matrix<T, Const<D>, U1, ArrayStorage<T, D, 1>>; // Owned<T, Const<D>, U1>>;
 
+/// An owned matrix with uninitialized data.
+pub type UninitVector<T, D> = Matrix<MaybeUninit<T>, D, U1, OwnedUninit<T, D, U1>>;
+
 /// An owned matrix column-major matrix with `R` rows and `C` columns.
 ///
 /// **Because this is an alias, not all its methods are listed here. See the [`Matrix`](crate::base::Matrix) type too.**
diff --git a/src/base/allocator.rs b/src/base/allocator.rs
index 1f639d3d..4d0c27b7 100644
--- a/src/base/allocator.rs
+++ b/src/base/allocator.rs
@@ -1,11 +1,14 @@
 //! Abstract definition of a matrix data storage allocator.
 
-use std::mem::{ManuallyDrop, MaybeUninit};
+use std::any::Any;
 
 use crate::base::constraint::{SameNumberOfColumns, SameNumberOfRows, ShapeConstraint};
 use crate::base::dimension::{Dim, U1};
-use crate::base::storage::ContiguousStorageMut;
-use crate::base::DefaultAllocator;
+use crate::base::{DefaultAllocator, Scalar};
+use crate::storage::{IsContiguous, RawStorageMut};
+use crate::StorageMut;
+use std::fmt::Debug;
+use std::mem::MaybeUninit;
 
 /// A matrix allocator of a memory buffer that may contain `R::to_usize() * C::to_usize()`
 /// elements of type `T`.
@@ -16,12 +19,23 @@ use crate::base::DefaultAllocator;
 ///
 /// Every allocator must be both static and dynamic. Though not all implementations may share the
 /// same `Buffer` type.
-///
-/// If you also want to be able to create uninitizalized or manually dropped memory buffers, see
-/// [`Allocator`].
-pub trait InnerAllocator<T, R: Dim, C: Dim = U1>: 'static + Sized {
+pub trait Allocator<T, R: Dim, C: Dim = U1>: Any + Sized {
     /// The type of buffer this allocator can instanciate.
-    type Buffer: ContiguousStorageMut<T, R, C>;
+    type Buffer: StorageMut<T, R, C> + IsContiguous + Clone + Debug;
+    /// The type of buffer with uninitialized components this allocator can instanciate.
+    type BufferUninit: RawStorageMut<MaybeUninit<T>, R, C> + IsContiguous;
+
+    /// Allocates a buffer with the given number of rows and columns without initializing its content.
+    unsafe fn allocate_uninitialized(nrows: R, ncols: C) -> MaybeUninit<Self::Buffer>;
+    /// Allocates a buffer with the given number of rows and columns without initializing its content.
+    fn allocate_uninit(nrows: R, ncols: C) -> Self::BufferUninit;
+
+    /// Assumes a data buffer to be initialized.
+    ///
+    /// # Safety
+    /// The user must make sure that every single entry of the buffer has been initialized,
+    /// or Undefined Behavior will immediately occur.    
+    unsafe fn assume_init(uninit: Self::BufferUninit) -> Self::Buffer;
 
     /// Allocates a buffer initialized with the content of the given iterator.
     fn allocate_from_iterator<I: IntoIterator<Item = T>>(
@@ -31,45 +45,15 @@ pub trait InnerAllocator<T, R: Dim, C: Dim = U1>: 'static + Sized {
     ) -> Self::Buffer;
 }
 
-/// Same as the [`InnerAllocator`] trait, but also provides methods to build uninitialized buffers,
-/// or buffers whose entries must be manually dropped.
-pub trait Allocator<T, R: Dim, C: Dim = U1>:
-    InnerAllocator<T, R, C>
-    + InnerAllocator<MaybeUninit<T>, R, C>
-    + InnerAllocator<ManuallyDrop<T>, R, C>
-{
-    /// Allocates a buffer with the given number of rows and columns without initializing its content.
-    fn allocate_uninitialized(
-        nrows: R,
-        ncols: C,
-    ) -> <Self as InnerAllocator<MaybeUninit<T>, R, C>>::Buffer;
-
-    /// Assumes a data buffer to be initialized. This operation should be near zero-cost.
-    ///
-    /// # Safety
-    /// The user must make sure that every single entry of the buffer has been initialized,
-    /// or Undefined Behavior will immediately occur.    
-    unsafe fn assume_init(
-        uninit: <Self as InnerAllocator<MaybeUninit<T>, R, C>>::Buffer,
-    ) -> <Self as InnerAllocator<T, R, C>>::Buffer;
-
-    /// Specifies that a given buffer's entries should be manually dropped.
-    fn manually_drop(
-        buf: <Self as InnerAllocator<T, R, C>>::Buffer,
-    ) -> <Self as InnerAllocator<ManuallyDrop<T>, R, C>>::Buffer;
-}
-
-
-/// A matrix reallocator. Changes the size of the memory buffer that initially contains (RFrom ×
-/// CFrom) elements to a smaller or larger size (RTo, CTo).
-pub trait Reallocator<T, RFrom: Dim, CFrom: Dim, RTo: Dim, CTo: Dim>:
+/// A matrix reallocator. Changes the size of the memory buffer that initially contains (`RFrom` ×
+/// `CFrom`) elements to a smaller or larger size (`RTo`, `CTo`).
+pub trait Reallocator<T: Scalar, RFrom: Dim, CFrom: Dim, RTo: Dim, CTo: Dim>:
     Allocator<T, RFrom, CFrom> + Allocator<T, RTo, CTo>
 {
     /// Reallocates a buffer of shape `(RTo, CTo)`, possibly reusing a previously allocated buffer
     /// `buf`. Data stored by `buf` are linearly copied to the output:
     ///
     /// # Safety
-    /// **NO! THIS IS STILL UB!**
     /// * The copy is performed as if both were just arrays (without a matrix structure).
     /// * If `buf` is larger than the output size, then extra elements of `buf` are truncated.
     /// * If `buf` is smaller than the output size, then extra elements of the output are left
@@ -77,8 +61,8 @@ pub trait Reallocator<T, RFrom: Dim, CFrom: Dim, RTo: Dim, CTo: Dim>:
     unsafe fn reallocate_copy(
         nrows: RTo,
         ncols: CTo,
-        buf: <Self as InnerAllocator<T, RFrom, CFrom>>::Buffer,
-    ) -> <Self as InnerAllocator<T, RTo, CTo>>::Buffer;
+        buf: <Self as Allocator<T, RFrom, CFrom>>::Buffer,
+    ) -> <Self as Allocator<T, RTo, CTo>>::Buffer;
 }
 
 /// The number of rows of the result of a componentwise operation on two matrices.
@@ -89,16 +73,23 @@ pub type SameShapeC<C1, C2> = <ShapeConstraint as SameNumberOfColumns<C1, C2>>::
 
 // TODO: Bad name.
 /// Restricts the given number of rows and columns to be respectively the same.
-pub trait SameShapeAllocator<T, R1: Dim, C1: Dim, R2: Dim, C2: Dim>:
+pub trait SameShapeAllocator<T, R1, C1, R2, C2>:
     Allocator<T, R1, C1> + Allocator<T, SameShapeR<R1, R2>, SameShapeC<C1, C2>>
 where
+    R1: Dim,
+    R2: Dim,
+    C1: Dim,
+    C2: Dim,
     ShapeConstraint: SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C2>,
 {
 }
 
-impl<T, R1: Dim, R2: Dim, C1: Dim, C2: Dim> SameShapeAllocator<T, R1, C1, R2, C2>
-    for DefaultAllocator
+impl<T, R1, R2, C1, C2> SameShapeAllocator<T, R1, C1, R2, C2> for DefaultAllocator
 where
+    R1: Dim,
+    R2: Dim,
+    C1: Dim,
+    C2: Dim,
     DefaultAllocator: Allocator<T, R1, C1> + Allocator<T, SameShapeR<R1, R2>, SameShapeC<C1, C2>>,
     ShapeConstraint: SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C2>,
 {
@@ -106,15 +97,19 @@ where
 
 // XXX: Bad name.
 /// Restricts the given number of rows to be equal.
-pub trait SameShapeVectorAllocator<T, R1: Dim, R2: Dim>:
+pub trait SameShapeVectorAllocator<T, R1, R2>:
     Allocator<T, R1> + Allocator<T, SameShapeR<R1, R2>> + SameShapeAllocator<T, R1, U1, R2, U1>
 where
+    R1: Dim,
+    R2: Dim,
     ShapeConstraint: SameNumberOfRows<R1, R2>,
 {
 }
 
-impl<T, R1: Dim, R2: Dim> SameShapeVectorAllocator<T, R1, R2> for DefaultAllocator
+impl<T, R1, R2> SameShapeVectorAllocator<T, R1, R2> for DefaultAllocator
 where
+    R1: Dim,
+    R2: Dim,
     DefaultAllocator: Allocator<T, R1, U1> + Allocator<T, SameShapeR<R1, R2>>,
     ShapeConstraint: SameNumberOfRows<R1, R2>,
 {
diff --git a/src/base/array_storage.rs b/src/base/array_storage.rs
index 0b397c33..65a43c2b 100644
--- a/src/base/array_storage.rs
+++ b/src/base/array_storage.rs
@@ -1,5 +1,4 @@
 use std::fmt::{self, Debug, Formatter};
-use std::mem;
 // use std::hash::{Hash, Hasher};
 #[cfg(feature = "abomonation-serialize")]
 use std::io::{Result as IOResult, Write};
@@ -13,28 +12,43 @@ use serde::ser::SerializeSeq;
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
 #[cfg(feature = "serde-serialize-no-std")]
 use std::marker::PhantomData;
+#[cfg(feature = "serde-serialize-no-std")]
+use std::mem;
 
 #[cfg(feature = "abomonation-serialize")]
 use abomonation::Abomonation;
 
-use crate::allocator::InnerAllocator;
+use crate::base::allocator::Allocator;
 use crate::base::default_allocator::DefaultAllocator;
 use crate::base::dimension::{Const, ToTypenum};
-use crate::base::storage::{
-    ContiguousStorage, ContiguousStorageMut, ReshapableStorage, Storage, StorageMut,
-};
-use crate::base::Owned;
+use crate::base::storage::{IsContiguous, Owned, RawStorage, RawStorageMut, ReshapableStorage};
+use crate::base::Scalar;
+use crate::Storage;
 
 /*
  *
- * Static Storage.
+ * Static RawStorage.
  *
  */
 /// A array-based statically sized matrix data storage.
-#[repr(transparent)]
+#[repr(C)]
 #[derive(Copy, Clone, PartialEq, Eq, Hash)]
 pub struct ArrayStorage<T, const R: usize, const C: usize>(pub [[T; R]; C]);
 
+impl<T, const R: usize, const C: usize> ArrayStorage<T, R, C> {
+    #[inline]
+    pub fn as_slice(&self) -> &[T] {
+        // SAFETY: this is OK because ArrayStorage is contiguous.
+        unsafe { self.as_slice_unchecked() }
+    }
+
+    #[inline]
+    pub fn as_mut_slice(&mut self) -> &mut [T] {
+        // SAFETY: this is OK because ArrayStorage is contiguous.
+        unsafe { self.as_mut_slice_unchecked() }
+    }
+}
+
 // TODO: remove this once the stdlib implements Default for arrays.
 impl<T: Default, const R: usize, const C: usize> Default for ArrayStorage<T, R, C>
 where
@@ -53,10 +67,8 @@ impl<T: Debug, const R: usize, const C: usize> Debug for ArrayStorage<T, R, C> {
     }
 }
 
-unsafe impl<T, const R: usize, const C: usize> Storage<T, Const<R>, Const<C>>
+unsafe impl<T, const R: usize, const C: usize> RawStorage<T, Const<R>, Const<C>>
     for ArrayStorage<T, R, C>
-where
-    DefaultAllocator: InnerAllocator<T, Const<R>, Const<C>, Buffer = Self>,
 {
     type RStride = Const<1>;
     type CStride = Const<R>;
@@ -81,38 +93,36 @@ where
         true
     }
 
-    #[inline]
-    fn into_owned(self) -> Owned<T, Const<R>, Const<C>>
-    where
-        DefaultAllocator: InnerAllocator<T, Const<R>, Const<C>>,
-    {
-        Owned(self)
-    }
-
-    #[inline]
-    fn clone_owned(&self) -> Owned<T, Const<R>, Const<C>>
-    where
-        T: Clone,
-        DefaultAllocator: InnerAllocator<T, Const<R>, Const<C>>,
-    {
-        let it = self.as_slice().iter().cloned();
-        Owned(DefaultAllocator::allocate_from_iterator(
-            self.shape().0,
-            self.shape().1,
-            it,
-        ))
-    }
-
     #[inline]
     unsafe fn as_slice_unchecked(&self) -> &[T] {
         std::slice::from_raw_parts(self.ptr(), R * C)
     }
 }
 
-unsafe impl<T, const R: usize, const C: usize> StorageMut<T, Const<R>, Const<C>>
+unsafe impl<T: Scalar, const R: usize, const C: usize> Storage<T, Const<R>, Const<C>>
     for ArrayStorage<T, R, C>
 where
-    DefaultAllocator: InnerAllocator<T, Const<R>, Const<C>, Buffer = Self>,
+    DefaultAllocator: Allocator<T, Const<R>, Const<C>, Buffer = Self>,
+{
+    #[inline]
+    fn into_owned(self) -> Owned<T, Const<R>, Const<C>>
+    where
+        DefaultAllocator: Allocator<T, Const<R>, Const<C>>,
+    {
+        self
+    }
+
+    #[inline]
+    fn clone_owned(&self) -> Owned<T, Const<R>, Const<C>>
+    where
+        DefaultAllocator: Allocator<T, Const<R>, Const<C>>,
+    {
+        self.clone()
+    }
+}
+
+unsafe impl<T, const R: usize, const C: usize> RawStorageMut<T, Const<R>, Const<C>>
+    for ArrayStorage<T, R, C>
 {
     #[inline]
     fn ptr_mut(&mut self) -> *mut T {
@@ -125,23 +135,12 @@ where
     }
 }
 
-unsafe impl<T, const R: usize, const C: usize> ContiguousStorage<T, Const<R>, Const<C>>
-    for ArrayStorage<T, R, C>
-where
-    DefaultAllocator: InnerAllocator<T, Const<R>, Const<C>, Buffer = Self>,
-{
-}
-
-unsafe impl<T, const R: usize, const C: usize> ContiguousStorageMut<T, Const<R>, Const<C>>
-    for ArrayStorage<T, R, C>
-where
-    DefaultAllocator: InnerAllocator<T, Const<R>, Const<C>, Buffer = Self>,
-{
-}
+unsafe impl<T, const R: usize, const C: usize> IsContiguous for ArrayStorage<T, R, C> {}
 
 impl<T, const R1: usize, const C1: usize, const R2: usize, const C2: usize>
     ReshapableStorage<T, Const<R1>, Const<C1>, Const<R2>, Const<C2>> for ArrayStorage<T, R1, C1>
 where
+    T: Scalar,
     Const<R1>: ToTypenum,
     Const<C1>: ToTypenum,
     Const<R2>: ToTypenum,
@@ -159,8 +158,8 @@ where
 
     fn reshape_generic(self, _: Const<R2>, _: Const<C2>) -> Self::Output {
         unsafe {
-            let data: [[T; R2]; C2] = mem::transmute_copy(&self.0);
-            mem::forget(self.0);
+            let data: [[T; R2]; C2] = std::mem::transmute_copy(&self.0);
+            std::mem::forget(self.0);
             ArrayStorage(data)
         }
     }
@@ -175,7 +174,7 @@ where
 #[cfg(feature = "serde-serialize-no-std")]
 impl<T, const R: usize, const C: usize> Serialize for ArrayStorage<T, R, C>
 where
-    T: Serialize,
+    T: Scalar + Serialize,
 {
     fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
     where
@@ -194,7 +193,7 @@ where
 #[cfg(feature = "serde-serialize-no-std")]
 impl<'a, T, const R: usize, const C: usize> Deserialize<'a> for ArrayStorage<T, R, C>
 where
-    T: Deserialize<'a>,
+    T: Scalar + Deserialize<'a>,
 {
     fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
     where
@@ -211,7 +210,10 @@ struct ArrayStorageVisitor<T, const R: usize, const C: usize> {
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<T, const R: usize, const C: usize> ArrayStorageVisitor<T, R, C> {
+impl<T, const R: usize, const C: usize> ArrayStorageVisitor<T, R, C>
+where
+    T: Scalar,
+{
     /// Construct a new sequence visitor.
     pub fn new() -> Self {
         ArrayStorageVisitor {
@@ -223,7 +225,7 @@ impl<T, const R: usize, const C: usize> ArrayStorageVisitor<T, R, C> {
 #[cfg(feature = "serde-serialize-no-std")]
 impl<'a, T, const R: usize, const C: usize> Visitor<'a> for ArrayStorageVisitor<T, R, C>
 where
-    T: Deserialize<'a>,
+    T: Scalar + Deserialize<'a>,
 {
     type Value = ArrayStorage<T, R, C>;
 
@@ -255,13 +257,13 @@ where
 }
 
 #[cfg(feature = "bytemuck")]
-unsafe impl<T: Copy + bytemuck::Zeroable, const R: usize, const C: usize> bytemuck::Zeroable
-    for ArrayStorage<T, R, C>
+unsafe impl<T: Scalar + Copy + bytemuck::Zeroable, const R: usize, const C: usize>
+    bytemuck::Zeroable for ArrayStorage<T, R, C>
 {
 }
 
 #[cfg(feature = "bytemuck")]
-unsafe impl<T: Copy + bytemuck::Pod, const R: usize, const C: usize> bytemuck::Pod
+unsafe impl<T: Scalar + Copy + bytemuck::Pod, const R: usize, const C: usize> bytemuck::Pod
     for ArrayStorage<T, R, C>
 {
 }
@@ -269,7 +271,7 @@ unsafe impl<T: Copy + bytemuck::Pod, const R: usize, const C: usize> bytemuck::P
 #[cfg(feature = "abomonation-serialize")]
 impl<T, const R: usize, const C: usize> Abomonation for ArrayStorage<T, R, C>
 where
-    T: Abomonation,
+    T: Scalar + Abomonation,
 {
     unsafe fn entomb<W: Write>(&self, writer: &mut W) -> IOResult<()> {
         for element in self.as_slice() {
diff --git a/src/base/blas.rs b/src/base/blas.rs
index 437ce7a7..c19011fd 100644
--- a/src/base/blas.rs
+++ b/src/base/blas.rs
@@ -1,21 +1,9 @@
-//! Implements a subset of the Basic Linear Algebra Subprograms (BLAS), a
-//! standard and highly optimized set of basic vector and matrix operations.
-//!
-//! To avoid unsoundness due to mishandling of uninitialized data, we divide our
-//! methods into two groups: those that take in a `&mut` to a matrix, and those
-//! that return an owned matrix that would otherwise result from setting a
-//! parameter to zero in the other methods.
-
-use crate::{MatrixSliceMut, SimdComplexField, VectorSliceMut};
-#[cfg(feature = "std")]
-use matrixmultiply;
+use crate::{RawStorage, SimdComplexField};
 use num::{One, Zero};
 use simba::scalar::{ClosedAdd, ClosedMul};
-#[cfg(feature = "std")]
-use std::mem;
-use std::mem::MaybeUninit;
 
 use crate::base::allocator::Allocator;
+use crate::base::blas_uninit::{axcpy_uninit, gemm_uninit, gemv_uninit};
 use crate::base::constraint::{
     AreMultipliable, DimEq, SameNumberOfColumns, SameNumberOfRows, ShapeConstraint,
 };
@@ -24,9 +12,10 @@ use crate::base::storage::{Storage, StorageMut};
 use crate::base::{
     DVectorSlice, DefaultAllocator, Matrix, Scalar, SquareMatrix, Vector, VectorSlice,
 };
+use crate::core::uninit::Init;
 
 /// # Dot/scalar product
-impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S>
+impl<T, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S>
 where
     T: Scalar + Zero + ClosedAdd + ClosedMul,
 {
@@ -37,7 +26,7 @@ where
         conjugate: impl Fn(T) -> T,
     ) -> T
     where
-        SB: Storage<T, R2, C2>,
+        SB: RawStorage<T, R2, C2>,
         ShapeConstraint: DimEq<R, R2> + DimEq<C, C2>,
     {
         assert!(
@@ -205,7 +194,7 @@ where
     #[must_use]
     pub fn dot<R2: Dim, C2: Dim, SB>(&self, rhs: &Matrix<T, R2, C2, SB>) -> T
     where
-        SB: Storage<T, R2, C2>,
+        SB: RawStorage<T, R2, C2>,
         ShapeConstraint: DimEq<R, R2> + DimEq<C, C2>,
     {
         self.dotx(rhs, |e| e)
@@ -235,7 +224,7 @@ where
     pub fn dotc<R2: Dim, C2: Dim, SB>(&self, rhs: &Matrix<T, R2, C2, SB>) -> T
     where
         T: SimdComplexField,
-        SB: Storage<T, R2, C2>,
+        SB: RawStorage<T, R2, C2>,
         ShapeConstraint: DimEq<R, R2> + DimEq<C, C2>,
     {
         self.dotx(rhs, T::simd_conjugate)
@@ -262,7 +251,7 @@ where
     #[must_use]
     pub fn tr_dot<R2: Dim, C2: Dim, SB>(&self, rhs: &Matrix<T, R2, C2, SB>) -> T
     where
-        SB: Storage<T, R2, C2>,
+        SB: RawStorage<T, R2, C2>,
         ShapeConstraint: DimEq<C, R2> + DimEq<R, C2>,
     {
         let (nrows, ncols) = self.shape();
@@ -293,10 +282,7 @@ where
     T: Scalar + Zero + ClosedAdd + ClosedMul,
     S: StorageMut<T, D>,
 {
-    /// Computes `self = a * x * c + b * self`, where `a`, `b`, `c` are scalars,
-    /// and `x` is a vector of the same size as `self`.
-    ///
-    /// For commutative scalars, this is equivalent to an [`axpy`] call.
+    /// Computes `self = a * x * c + b * self`.
     ///
     /// If `b` is zero, `self` is never read from.
     ///
@@ -316,34 +302,7 @@ where
         SB: Storage<T, D2>,
         ShapeConstraint: DimEq<D, D2>,
     {
-        assert_eq!(self.nrows(), x.nrows(), "Axcpy: mismatched vector shapes.");
-
-        let rstride1 = self.strides().0;
-        let rstride2 = x.strides().0;
-
-        unsafe {
-            // SAFETY: the conversion to slices is OK because we access the
-            //         elements taking the strides into account.
-            let y = self.data.as_mut_slice_unchecked();
-            let x = x.data.as_slice_unchecked();
-
-            if !b.is_zero() {
-                for i in 0..x.len() {
-                    let y = y.get_unchecked_mut(i * rstride1);
-                    *y = a.inlined_clone()
-                        * x.get_unchecked(i * rstride2).inlined_clone()
-                        * c.inlined_clone()
-                        + b.inlined_clone() * y.inlined_clone();
-                }
-            } else {
-                for i in 0..x.len() {
-                    let y = y.get_unchecked_mut(i * rstride1);
-                    *y = a.inlined_clone()
-                        * x.get_unchecked(i * rstride2).inlined_clone()
-                        * c.inlined_clone();
-                }
-            }
-        }
+        unsafe { axcpy_uninit(Init, self, a, x, c, b) };
     }
 
     /// Computes `self = a * x + b * self`.
@@ -399,38 +358,8 @@ where
         SC: Storage<T, D3>,
         ShapeConstraint: DimEq<D, R2> + AreMultipliable<R2, C2, D3, U1>,
     {
-        let dim1 = self.nrows();
-        let (nrows2, ncols2) = a.shape();
-        let dim3 = x.nrows();
-
-        assert!(
-            ncols2 == dim3 && dim1 == nrows2,
-            "Gemv: dimensions mismatch."
-        );
-
-        if ncols2 == 0 {
-            // NOTE: we can't just always multiply by beta
-            // because we documented the guaranty that `self` is
-            // never read if `beta` is zero.
-            if beta.is_zero() {
-                self.fill(T::zero());
-            } else {
-                *self *= beta;
-            }
-            return;
-        }
-
-        // TODO: avoid bound checks.
-        let col2 = a.column(0);
-        let val = unsafe { x.vget_unchecked(0).inlined_clone() };
-        self.axcpy(alpha.inlined_clone(), &col2, val, beta);
-
-        for j in 1..ncols2 {
-            let col2 = a.column(j);
-            let val = unsafe { x.vget_unchecked(j).inlined_clone() };
-
-            self.axcpy(alpha.inlined_clone(), &col2, val, T::one());
-        }
+        // Safety: this is safe because we are passing Status == Init.
+        unsafe { gemv_uninit(Init, self, alpha, a, x, beta) }
     }
 
     #[inline(always)]
@@ -490,25 +419,6 @@ where
         }
     }
 
-    /// Computes `self = alpha * a * x + beta * self`, where `a` is a **symmetric** matrix, `x` a
-    /// vector, and `alpha, beta` two scalars. DEPRECATED: use `sygemv` instead.
-    #[inline]
-    #[deprecated(note = "This is renamed `sygemv` to match the original BLAS terminology.")]
-    pub fn gemv_symm<D2: Dim, D3: Dim, SB, SC>(
-        &mut self,
-        alpha: T,
-        a: &SquareMatrix<T, D2, SB>,
-        x: &Vector<T, D3, SC>,
-        beta: T,
-    ) where
-        T: One,
-        SB: Storage<T, D2, D2>,
-        SC: Storage<T, D3>,
-        ShapeConstraint: DimEq<D, D2> + AreMultipliable<D2, D2, D3, U1>,
-    {
-        self.sygemv(alpha, a, x, beta)
-    }
-
     /// Computes `self = alpha * a * x + beta * self`, where `a` is a **symmetric** matrix, `x` a
     /// vector, and `alpha, beta` two scalars.
     ///
@@ -709,331 +619,6 @@ where
     }
 }
 
-impl<T, D: Dim, S> Vector<MaybeUninit<T>, D, S>
-where
-    T: Scalar + Zero + ClosedAdd + ClosedMul,
-    S: StorageMut<MaybeUninit<T>, D>,
-{
-    /// Computes `alpha * a * x`, where `a` is a matrix, `x` a vector, and
-    /// `alpha` is a scalar.
-    ///
-    /// `self` must be completely uninitialized, or data leaks will occur. After
-    /// this method is called, all entries in `self` will be initialized.
-    #[inline]
-    pub fn axc<D2: Dim, S2>(
-        &mut self,
-        a: T,
-        x: &Vector<T, D2, S2>,
-        c: T,
-    ) -> VectorSliceMut<T, D, S::RStride, S::CStride>
-    where
-        S2: Storage<T, D2>,
-        ShapeConstraint: DimEq<D, D2>,
-    {
-        let rstride1 = self.strides().0;
-        let rstride2 = x.strides().0;
-
-        // Safety: see each individual remark.
-        unsafe {
-            // We don't mind `x` and `y` not being contiguous, as we'll only
-            // access the elements we're allowed to. (TODO: double check this)
-            let y = self.data.as_mut_slice_unchecked();
-            let x = x.data.as_slice_unchecked();
-
-            // The indices are within range, and only access elements that belong
-            // to `x` and `y` themselves.
-            for i in 0..y.len() {
-                *y.get_unchecked_mut(i * rstride1) = MaybeUninit::new(
-                    a.inlined_clone()
-                        * x.get_unchecked(i * rstride2).inlined_clone()
-                        * c.inlined_clone(),
-                );
-            }
-
-            // We've initialized all elements.
-            self.assume_init_mut()
-        }
-    }
-
-    /// Computes `alpha * a * x`, where `a` is a matrix, `x` a vector, and
-    /// `alpha` is a scalar.
-    ///
-    /// `self` must be completely uninitialized, or data leaks will occur. After
-    /// the method is called, `self` will be completely initialized. We return
-    /// an initialized mutable vector slice to `self` for convenience.
-    #[inline]
-    pub fn gemv_z<R2: Dim, C2: Dim, D3: Dim, SB, SC>(
-        &mut self,
-        alpha: T,
-        a: &Matrix<T, R2, C2, SB>,
-        x: &Vector<T, D3, SC>,
-    ) -> VectorSliceMut<T, D, S::RStride, S::CStride>
-    where
-        T: One,
-        SB: Storage<T, R2, C2>,
-        SC: Storage<T, D3>,
-        ShapeConstraint: DimEq<D, R2> + AreMultipliable<R2, C2, D3, U1>,
-    {
-        let dim1 = self.nrows();
-        let (nrows2, ncols2) = a.shape();
-        let dim3 = x.nrows();
-
-        assert!(
-            ncols2 == dim3 && dim1 == nrows2,
-            "Gemv: dimensions mismatch."
-        );
-
-        if ncols2 == 0 {
-            self.fill_fn(|| MaybeUninit::new(T::zero()));
-
-            // Safety: all entries have just been initialized.
-            unsafe {
-                return self.assume_init_mut();
-            }
-        }
-
-        // TODO: avoid bound checks.
-        let col2 = a.column(0);
-        let val = unsafe { x.vget_unchecked(0).inlined_clone() };
-        let mut init = self.axc(alpha.inlined_clone(), &col2, val);
-
-        // Safety: all indices are within range.
-        unsafe {
-            for j in 1..ncols2 {
-                let col2 = a.column(j);
-                let val = x.vget_unchecked(j).inlined_clone();
-                init.axcpy(alpha.inlined_clone(), &col2, val, T::one());
-            }
-        }
-
-        init
-    }
-
-    #[inline(always)]
-    fn xxgemv_z<D2: Dim, D3: Dim, SB, SC>(
-        &mut self,
-        alpha: T,
-        a: &SquareMatrix<T, D2, SB>,
-        x: &Vector<T, D3, SC>,
-        dot: impl Fn(
-            &DVectorSlice<T, SB::RStride, SB::CStride>,
-            &DVectorSlice<T, SC::RStride, SC::CStride>,
-        ) -> T,
-    ) where
-        T: One,
-        SB: Storage<T, D2, D2>,
-        SC: Storage<T, D3>,
-        ShapeConstraint: DimEq<D, D2> + AreMultipliable<D2, D2, D3, U1>,
-    {
-        let dim1 = self.nrows();
-        let dim2 = a.nrows();
-        let dim3 = x.nrows();
-
-        assert!(
-            a.is_square(),
-            "Symmetric cgemv: the input matrix must be square."
-        );
-        assert!(
-            dim2 == dim3 && dim1 == dim2,
-            "Symmetric cgemv: dimensions mismatch."
-        );
-
-        if dim2 == 0 {
-            return;
-        }
-
-        // TODO: avoid bound checks.
-        let col2 = a.column(0);
-        let val = unsafe { x.vget_unchecked(0).inlined_clone() };
-        let mut res = self.axc(alpha.inlined_clone(), &col2, val);
-
-        res[0] += alpha.inlined_clone() * dot(&a.slice_range(1.., 0), &x.rows_range(1..));
-
-        for j in 1..dim2 {
-            let col2 = a.column(j);
-            let dot = dot(&col2.rows_range(j..), &x.rows_range(j..));
-
-            let val;
-            unsafe {
-                val = x.vget_unchecked(j).inlined_clone();
-                *res.vget_unchecked_mut(j) += alpha.inlined_clone() * dot;
-            }
-            res.rows_range_mut(j + 1..).axpy(
-                alpha.inlined_clone() * val,
-                &col2.rows_range(j + 1..),
-                T::one(),
-            );
-        }
-    }
-
-    /// Computes `self = alpha * a * x`, where `a` is an **hermitian** matrix, `x` a
-    /// vector, and `alpha, beta` two scalars.
-    pub fn hegemv_z<D2: Dim, D3: Dim, SB, SC>(
-        &mut self,
-        alpha: T,
-        a: &SquareMatrix<T, D2, SB>,
-        x: &Vector<T, D3, SC>,
-    ) where
-        T: SimdComplexField,
-        SB: Storage<T, D2, D2>,
-        SC: Storage<T, D3>,
-        ShapeConstraint: DimEq<D, D2> + AreMultipliable<D2, D2, D3, U1>,
-    {
-        self.xxgemv_z(alpha, a, x, |a, b| a.dotc(b))
-    }
-}
-
-impl<T, R1: Dim, C1: Dim, S: StorageMut<MaybeUninit<T>, R1, C1>> Matrix<MaybeUninit<T>, R1, C1, S>
-where
-    T: Scalar + Zero + One + ClosedAdd + ClosedMul,
-    // DefaultAllocator: Allocator<T, R1, C1>,
-{
-    /// Computes `alpha * a * b`, where `a` and `b` are matrices, and `alpha` is
-    /// a scalar.
-    ///
-    /// # Examples:
-    ///
-    /// ```
-    /// # #[macro_use] extern crate approx;
-    /// # use nalgebra::{Matrix2x3, Matrix3x4, Matrix2x4};
-    /// let mut mat1 = Matrix2x4::identity();
-    /// let mat2 = Matrix2x3::new(1.0, 2.0, 3.0,
-    ///                           4.0, 5.0, 6.0);
-    /// let mat3 = Matrix3x4::new(0.1, 0.2, 0.3, 0.4,
-    ///                           0.5, 0.6, 0.7, 0.8,
-    ///                           0.9, 1.0, 1.1, 1.2);
-    /// let expected = mat2 * mat3 * 10.0 + mat1 * 5.0;
-    ///
-    /// mat1.gemm(10.0, &mat2, &mat3, 5.0);
-    /// assert_relative_eq!(mat1, expected);
-    /// ```
-    #[inline]
-    pub fn gemm_z<R2: Dim, C2: Dim, R3: Dim, C3: Dim, SB, SC>(
-        &mut self,
-        alpha: T,
-        a: &Matrix<T, R2, C2, SB>,
-        b: &Matrix<T, R3, C3, SC>,
-    ) -> MatrixSliceMut<T, R1, C1, S::RStride, S::CStride>
-    where
-        SB: Storage<T, R2, C2>,
-        SC: Storage<T, R3, C3>,
-        ShapeConstraint: SameNumberOfRows<R1, R2>
-            + SameNumberOfColumns<C1, C3>
-            + AreMultipliable<R2, C2, R3, C3>,
-    {
-        let ncols1 = self.ncols();
-
-        #[cfg(feature = "std")]
-        {
-            // We assume large matrices will be Dynamic but small matrices static.
-            // We could use matrixmultiply for large statically-sized matrices but the performance
-            // threshold to activate it would be different from SMALL_DIM because our code optimizes
-            // better for statically-sized matrices.
-            if R1::is::<Dynamic>()
-                || C1::is::<Dynamic>()
-                || R2::is::<Dynamic>()
-                || C2::is::<Dynamic>()
-                || R3::is::<Dynamic>()
-                || C3::is::<Dynamic>()
-            {
-                // matrixmultiply can be used only if the std feature is available.
-                let nrows1 = self.nrows();
-                let (nrows2, ncols2) = a.shape();
-                let (nrows3, ncols3) = b.shape();
-
-                // Threshold determined empirically.
-                const SMALL_DIM: usize = 5;
-
-                if nrows1 > SMALL_DIM
-                    && ncols1 > SMALL_DIM
-                    && nrows2 > SMALL_DIM
-                    && ncols2 > SMALL_DIM
-                {
-                    assert_eq!(
-                        ncols2, nrows3,
-                        "gemm: dimensions mismatch for multiplication."
-                    );
-                    assert_eq!(
-                        (nrows1, ncols1),
-                        (nrows2, ncols3),
-                        "gemm: dimensions mismatch for addition."
-                    );
-
-                    // NOTE: this case should never happen because we enter this
-                    // codepath only when ncols2 > SMALL_DIM. Though we keep this
-                    // here just in case if in the future we change the conditions to
-                    // enter this codepath.
-                    if ncols1 == 0 {
-                        self.fill_fn(|| MaybeUninit::new(T::zero()));
-
-                        // Safety: there's no (uninitialized) values.
-                        return unsafe { self.assume_init_mut() };
-                    }
-
-                    let (rsa, csa) = a.strides();
-                    let (rsb, csb) = b.strides();
-                    let (rsc, csc) = self.strides();
-
-                    if T::is::<f32>() {
-                        unsafe {
-                            matrixmultiply::sgemm(
-                                nrows2,
-                                ncols2,
-                                ncols3,
-                                mem::transmute_copy(&alpha),
-                                a.data.ptr() as *const f32,
-                                rsa as isize,
-                                csa as isize,
-                                b.data.ptr() as *const f32,
-                                rsb as isize,
-                                csb as isize,
-                                0.0,
-                                self.data.ptr_mut() as *mut f32,
-                                rsc as isize,
-                                csc as isize,
-                            );
-                        }
-                    } else if T::is::<f64>() {
-                        unsafe {
-                            matrixmultiply::dgemm(
-                                nrows2,
-                                ncols2,
-                                ncols3,
-                                mem::transmute_copy(&alpha),
-                                a.data.ptr() as *const f64,
-                                rsa as isize,
-                                csa as isize,
-                                b.data.ptr() as *const f64,
-                                rsb as isize,
-                                csb as isize,
-                                0.0,
-                                self.data.ptr_mut() as *mut f64,
-                                rsc as isize,
-                                csc as isize,
-                            );
-                        }
-                    }
-
-                    // Safety: all entries have been initialized.
-                    unsafe {
-                        return self.assume_init_mut();
-                    }
-                }
-            }
-        }
-
-        for j1 in 0..ncols1 {
-            // TODO: avoid bound checks.
-            let _ = self
-                .column_mut(j1)
-                .gemv_z(alpha.inlined_clone(), a, &b.column(j1));
-        }
-
-        // Safety: all entries have been initialized.
-        unsafe { self.assume_init_mut() }
-    }
-}
-
 impl<T, R1: Dim, C1: Dim, S: StorageMut<T, R1, C1>> Matrix<T, R1, C1, S>
 where
     T: Scalar + Zero + ClosedAdd + ClosedMul,
@@ -1170,122 +755,9 @@ where
             + SameNumberOfColumns<C1, C3>
             + AreMultipliable<R2, C2, R3, C3>,
     {
-        let ncols1 = self.ncols();
-
-        #[cfg(feature = "std")]
-        {
-            // We assume large matrices will be Dynamic but small matrices static.
-            // We could use matrixmultiply for large statically-sized matrices but the performance
-            // threshold to activate it would be different from SMALL_DIM because our code optimizes
-            // better for statically-sized matrices.
-            if R1::is::<Dynamic>()
-                || C1::is::<Dynamic>()
-                || R2::is::<Dynamic>()
-                || C2::is::<Dynamic>()
-                || R3::is::<Dynamic>()
-                || C3::is::<Dynamic>()
-            {
-                // matrixmultiply can be used only if the std feature is available.
-                let nrows1 = self.nrows();
-                let (nrows2, ncols2) = a.shape();
-                let (nrows3, ncols3) = b.shape();
-
-                // Threshold determined empirically.
-                const SMALL_DIM: usize = 5;
-
-                if nrows1 > SMALL_DIM
-                    && ncols1 > SMALL_DIM
-                    && nrows2 > SMALL_DIM
-                    && ncols2 > SMALL_DIM
-                {
-                    assert_eq!(
-                        ncols2, nrows3,
-                        "gemm: dimensions mismatch for multiplication."
-                    );
-                    assert_eq!(
-                        (nrows1, ncols1),
-                        (nrows2, ncols3),
-                        "gemm: dimensions mismatch for addition."
-                    );
-
-                    // NOTE: this case should never happen because we enter this
-                    // codepath only when ncols2 > SMALL_DIM. Though we keep this
-                    // here just in case if in the future we change the conditions to
-                    // enter this codepath.
-                    if ncols2 == 0 {
-                        // NOTE: we can't just always multiply by beta
-                        // because we documented the guaranty that `self` is
-                        // never read if `beta` is zero.
-                        if beta.is_zero() {
-                            self.fill(T::zero());
-                        } else {
-                            *self *= beta;
-                        }
-                        return;
-                    }
-
-                    if T::is::<f32>() {
-                        let (rsa, csa) = a.strides();
-                        let (rsb, csb) = b.strides();
-                        let (rsc, csc) = self.strides();
-
-                        unsafe {
-                            matrixmultiply::sgemm(
-                                nrows2,
-                                ncols2,
-                                ncols3,
-                                mem::transmute_copy(&alpha),
-                                a.data.ptr() as *const f32,
-                                rsa as isize,
-                                csa as isize,
-                                b.data.ptr() as *const f32,
-                                rsb as isize,
-                                csb as isize,
-                                mem::transmute_copy(&beta),
-                                self.data.ptr_mut() as *mut f32,
-                                rsc as isize,
-                                csc as isize,
-                            );
-                        }
-                        return;
-                    } else if T::is::<f64>() {
-                        let (rsa, csa) = a.strides();
-                        let (rsb, csb) = b.strides();
-                        let (rsc, csc) = self.strides();
-
-                        unsafe {
-                            matrixmultiply::dgemm(
-                                nrows2,
-                                ncols2,
-                                ncols3,
-                                mem::transmute_copy(&alpha),
-                                a.data.ptr() as *const f64,
-                                rsa as isize,
-                                csa as isize,
-                                b.data.ptr() as *const f64,
-                                rsb as isize,
-                                csb as isize,
-                                mem::transmute_copy(&beta),
-                                self.data.ptr_mut() as *mut f64,
-                                rsc as isize,
-                                csc as isize,
-                            );
-                        }
-                        return;
-                    }
-                }
-            }
-        }
-
-        for j1 in 0..ncols1 {
-            // TODO: avoid bound checks.
-            self.column_mut(j1).gemv(
-                alpha.inlined_clone(),
-                a,
-                &b.column(j1),
-                beta.inlined_clone(),
-            );
-        }
+        // SAFETY: this is valid because our matrices are initialized and
+        // we are using status = Init.
+        unsafe { gemm_uninit(Init, self, alpha, a, b, beta) }
     }
 
     /// Computes `self = alpha * a.transpose() * b + beta * self`, where `a, b, self` are matrices.
@@ -1579,33 +1051,78 @@ where
     /// let mid = DMatrix::from_row_slice(3, 3, &[0.1, 0.2, 0.3,
     ///                                           0.5, 0.6, 0.7,
     ///                                           0.9, 1.0, 1.1]);
-    ///
+    /// // The random shows that values on the workspace do not
+    /// // matter as they will be overwritten.
+    /// let mut workspace = DVector::new_random(2);
     /// let expected = &lhs * &mid * lhs.transpose() * 10.0 + &mat * 5.0;
     ///
+    /// mat.quadform_tr_with_workspace(&mut workspace, 10.0, &lhs, &mid, 5.0);
+    /// assert_relative_eq!(mat, expected);
+    pub fn quadform_tr_with_workspace<D2, S2, R3, C3, S3, D4, S4>(
+        &mut self,
+        work: &mut Vector<T, D2, S2>,
+        alpha: T,
+        lhs: &Matrix<T, R3, C3, S3>,
+        mid: &SquareMatrix<T, D4, S4>,
+        beta: T,
+    ) where
+        D2: Dim,
+        R3: Dim,
+        C3: Dim,
+        D4: Dim,
+        S2: StorageMut<T, D2>,
+        S3: Storage<T, R3, C3>,
+        S4: Storage<T, D4, D4>,
+        ShapeConstraint: DimEq<D1, D2> + DimEq<D1, R3> + DimEq<D2, R3> + DimEq<C3, D4>,
+    {
+        work.gemv(T::one(), lhs, &mid.column(0), T::zero());
+        self.ger(alpha.inlined_clone(), work, &lhs.column(0), beta);
+
+        for j in 1..mid.ncols() {
+            work.gemv(T::one(), lhs, &mid.column(j), T::zero());
+            self.ger(alpha.inlined_clone(), work, &lhs.column(j), T::one());
+        }
+    }
+
+    /// Computes the quadratic form `self = alpha * lhs * mid * lhs.transpose() + beta * self`.
+    ///
+    /// This allocates a workspace vector of dimension D1 for intermediate results.
+    /// If `D1` is a type-level integer, then the allocation is performed on the stack.
+    /// Use `.quadform_tr_with_workspace(...)` instead to avoid allocations.
+    ///
+    /// # Examples:
+    ///
+    /// ```
+    /// # #[macro_use] extern crate approx;
+    /// # use nalgebra::{Matrix2, Matrix3, Matrix2x3, Vector2};
+    /// let mut mat = Matrix2::identity();
+    /// let lhs = Matrix2x3::new(1.0, 2.0, 3.0,
+    ///                          4.0, 5.0, 6.0);
+    /// let mid = Matrix3::new(0.1, 0.2, 0.3,
+    ///                        0.5, 0.6, 0.7,
+    ///                        0.9, 1.0, 1.1);
+    /// let expected = lhs * mid * lhs.transpose() * 10.0 + mat * 5.0;
+    ///
     /// mat.quadform_tr(10.0, &lhs, &mid, 5.0);
     /// assert_relative_eq!(mat, expected);
-    pub fn quadform_tr<R3: Dim, C3: Dim, S3, D4: Dim, S4>(
+    pub fn quadform_tr<R3, C3, S3, D4, S4>(
         &mut self,
         alpha: T,
         lhs: &Matrix<T, R3, C3, S3>,
         mid: &SquareMatrix<T, D4, S4>,
         beta: T,
     ) where
+        R3: Dim,
+        C3: Dim,
+        D4: Dim,
         S3: Storage<T, R3, C3>,
         S4: Storage<T, D4, D4>,
-        ShapeConstraint: DimEq<D1, R3> + DimEq<C3, D4>,
-        DefaultAllocator: Allocator<T, R3>,
+        ShapeConstraint: DimEq<D1, D1> + DimEq<D1, R3> + DimEq<C3, D4>,
+        DefaultAllocator: Allocator<T, D1>,
     {
-        let mut work =
-            Matrix::new_uninitialized_generic(R3::from_usize(self.shape().0), Const::<1>);
-        let mut work = work.gemv_z(T::one(), lhs, &mid.column(0));
-
-        self.ger(alpha.inlined_clone(), &work, &lhs.column(0), beta);
-
-        for j in 1..mid.ncols() {
-            work.gemv(T::one(), lhs, &mid.column(j), T::zero());
-            self.ger(alpha.inlined_clone(), &work, &lhs.column(j), T::one());
-        }
+        // TODO: would it be useful to avoid the zero-initialization of the workspace data?
+        let mut work = Matrix::zeros_generic(self.shape_generic().0, Const::<1>);
+        self.quadform_tr_with_workspace(&mut work, alpha, lhs, mid, beta)
     }
 
     /// Computes the quadratic form `self = alpha * rhs.transpose() * mid * rhs + beta * self`.
@@ -1626,34 +1143,79 @@ where
     /// let mid = DMatrix::from_row_slice(3, 3, &[0.1, 0.2, 0.3,
     ///                                           0.5, 0.6, 0.7,
     ///                                           0.9, 1.0, 1.1]);
-    ///
+    /// // The random shows that values on the workspace do not
+    /// // matter as they will be overwritten.
+    /// let mut workspace = DVector::new_random(3);
     /// let expected = rhs.transpose() * &mid * &rhs * 10.0 + &mat * 5.0;
     ///
-    /// mat.quadform(10.0, &mid, &rhs, 5.0);
+    /// mat.quadform_with_workspace(&mut workspace, 10.0, &mid, &rhs, 5.0);
     /// assert_relative_eq!(mat, expected);
-    pub fn quadform<D3: Dim, S3, R4: Dim, C4: Dim, S4>(
+    pub fn quadform_with_workspace<D2, S2, D3, S3, R4, C4, S4>(
         &mut self,
+        work: &mut Vector<T, D2, S2>,
         alpha: T,
         mid: &SquareMatrix<T, D3, S3>,
         rhs: &Matrix<T, R4, C4, S4>,
         beta: T,
     ) where
+        D2: Dim,
+        D3: Dim,
+        R4: Dim,
+        C4: Dim,
+        S2: StorageMut<T, D2>,
         S3: Storage<T, D3, D3>,
         S4: Storage<T, R4, C4>,
-        ShapeConstraint: DimEq<D3, R4> + DimEq<R4, D3> + DimEq<D1, C4>,
-        DefaultAllocator: Allocator<T, D3>,
+        ShapeConstraint:
+            DimEq<D3, R4> + DimEq<D1, C4> + DimEq<D2, D3> + AreMultipliable<C4, R4, D2, U1>,
     {
-        // TODO: figure out why type inference isn't doing its job.
-        let mut work = Matrix::new_uninitialized_generic(D3::from_usize(mid.shape().0), Const::<1>);
-        let mut work = work.gemv_z::<D3, _, _, _, _>(T::one(), mid, &rhs.column(0));
-
+        work.gemv(T::one(), mid, &rhs.column(0), T::zero());
         self.column_mut(0)
-            .gemv_tr(alpha.inlined_clone(), rhs, &work, beta.inlined_clone());
+            .gemv_tr(alpha.inlined_clone(), rhs, work, beta.inlined_clone());
 
         for j in 1..rhs.ncols() {
-            work.gemv::<D3, D3, R4, S3, _>(T::one(), mid, &rhs.column(j), T::zero());
+            work.gemv(T::one(), mid, &rhs.column(j), T::zero());
             self.column_mut(j)
-                .gemv_tr(alpha.inlined_clone(), rhs, &work, beta.inlined_clone());
+                .gemv_tr(alpha.inlined_clone(), rhs, work, beta.inlined_clone());
         }
     }
+
+    /// Computes the quadratic form `self = alpha * rhs.transpose() * mid * rhs + beta * self`.
+    ///
+    /// This allocates a workspace vector of dimension D2 for intermediate results.
+    /// If `D2` is a type-level integer, then the allocation is performed on the stack.
+    /// Use `.quadform_with_workspace(...)` instead to avoid allocations.
+    ///
+    /// ```
+    /// # #[macro_use] extern crate approx;
+    /// # use nalgebra::{Matrix2, Matrix3x2, Matrix3};
+    /// let mut mat = Matrix2::identity();
+    /// let rhs = Matrix3x2::new(1.0, 2.0,
+    ///                          3.0, 4.0,
+    ///                          5.0, 6.0);
+    /// let mid = Matrix3::new(0.1, 0.2, 0.3,
+    ///                        0.5, 0.6, 0.7,
+    ///                        0.9, 1.0, 1.1);
+    /// let expected = rhs.transpose() * mid * rhs * 10.0 + mat * 5.0;
+    ///
+    /// mat.quadform(10.0, &mid, &rhs, 5.0);
+    /// assert_relative_eq!(mat, expected);
+    pub fn quadform<D2, S2, R3, C3, S3>(
+        &mut self,
+        alpha: T,
+        mid: &SquareMatrix<T, D2, S2>,
+        rhs: &Matrix<T, R3, C3, S3>,
+        beta: T,
+    ) where
+        D2: Dim,
+        R3: Dim,
+        C3: Dim,
+        S2: Storage<T, D2, D2>,
+        S3: Storage<T, R3, C3>,
+        ShapeConstraint: DimEq<D2, R3> + DimEq<D1, C3> + AreMultipliable<C3, R3, D2, U1>,
+        DefaultAllocator: Allocator<T, D2>,
+    {
+        // TODO: would it be useful to avoid the zero-initialization of the workspace data?
+        let mut work = Vector::zeros_generic(mid.shape_generic().0, Const::<1>);
+        self.quadform_with_workspace(&mut work, alpha, mid, rhs, beta)
+    }
 }
diff --git a/src/base/blas_uninit.rs b/src/base/blas_uninit.rs
new file mode 100644
index 00000000..2b3c5fc3
--- /dev/null
+++ b/src/base/blas_uninit.rs
@@ -0,0 +1,359 @@
+/*
+ * This file implements some BLAS operations in such a way that they work
+ * even if the first argument (the output parameter) is an uninitialized matrix.
+ *
+ * Because doing this makes the code harder to read, we only implemented the operations that we
+ * know would benefit from this performance-wise, namely, GEMM (which we use for our matrix
+ * multiplication code). If we identify other operations like that in the future, we could add
+ * them here.
+ */
+
+#[cfg(feature = "std")]
+use matrixmultiply;
+use num::{One, Zero};
+use simba::scalar::{ClosedAdd, ClosedMul};
+#[cfg(feature = "std")]
+use std::mem;
+
+use crate::base::constraint::{
+    AreMultipliable, DimEq, SameNumberOfColumns, SameNumberOfRows, ShapeConstraint,
+};
+use crate::base::dimension::{Dim, Dynamic, U1};
+use crate::base::storage::{RawStorage, RawStorageMut};
+use crate::base::uninit::{InitStatus, Initialized};
+use crate::base::{Matrix, Scalar, Vector};
+
+// # Safety
+// The content of `y` must only contain values for which
+// `Status::assume_init_mut` is sound.
+#[allow(clippy::too_many_arguments)]
+unsafe fn array_axcpy<Status, T>(
+    _: Status,
+    y: &mut [Status::Value],
+    a: T,
+    x: &[T],
+    c: T,
+    beta: T,
+    stride1: usize,
+    stride2: usize,
+    len: usize,
+) where
+    Status: InitStatus<T>,
+    T: Scalar + Zero + ClosedAdd + ClosedMul,
+{
+    for i in 0..len {
+        let y = Status::assume_init_mut(y.get_unchecked_mut(i * stride1));
+        *y = a.inlined_clone() * x.get_unchecked(i * stride2).inlined_clone() * c.inlined_clone()
+            + beta.inlined_clone() * y.inlined_clone();
+    }
+}
+
+fn array_axc<Status, T>(
+    _: Status,
+    y: &mut [Status::Value],
+    a: T,
+    x: &[T],
+    c: T,
+    stride1: usize,
+    stride2: usize,
+    len: usize,
+) where
+    Status: InitStatus<T>,
+    T: Scalar + Zero + ClosedAdd + ClosedMul,
+{
+    for i in 0..len {
+        unsafe {
+            Status::init(
+                y.get_unchecked_mut(i * stride1),
+                a.inlined_clone()
+                    * x.get_unchecked(i * stride2).inlined_clone()
+                    * c.inlined_clone(),
+            );
+        }
+    }
+}
+
+/// Computes `self = a * x * c + b * self`.
+///
+/// If `b` is zero, `self` is never read from.
+///
+/// # Examples:
+///
+/// ```
+/// # use nalgebra::Vector3;
+/// let mut vec1 = Vector3::new(1.0, 2.0, 3.0);
+/// let vec2 = Vector3::new(0.1, 0.2, 0.3);
+/// vec1.axcpy(5.0, &vec2, 2.0, 5.0);
+/// assert_eq!(vec1, Vector3::new(6.0, 12.0, 18.0));
+/// ```
+#[inline]
+#[allow(clippy::many_single_char_names)]
+pub unsafe fn axcpy_uninit<Status, T, D1: Dim, D2: Dim, SA, SB>(
+    status: Status,
+    y: &mut Vector<Status::Value, D1, SA>,
+    a: T,
+    x: &Vector<T, D2, SB>,
+    c: T,
+    b: T,
+) where
+    T: Scalar + Zero + ClosedAdd + ClosedMul,
+    SA: RawStorageMut<Status::Value, D1>,
+    SB: RawStorage<T, D2>,
+    ShapeConstraint: DimEq<D1, D2>,
+    Status: InitStatus<T>,
+{
+    assert_eq!(y.nrows(), x.nrows(), "Axcpy: mismatched vector shapes.");
+
+    let rstride1 = y.strides().0;
+    let rstride2 = x.strides().0;
+
+    // SAFETY: the conversion to slices is OK because we access the
+    //         elements taking the strides into account.
+    let y = y.data.as_mut_slice_unchecked();
+    let x = x.data.as_slice_unchecked();
+
+    if !b.is_zero() {
+        array_axcpy(status, y, a, x, c, b, rstride1, rstride2, x.len());
+    } else {
+        array_axc(status, y, a, x, c, rstride1, rstride2, x.len());
+    }
+}
+
+/// Computes `self = alpha * a * x + beta * self`, where `a` is a matrix, `x` a vector, and
+/// `alpha, beta` two scalars.
+///
+/// If `beta` is zero, `self` is never read.
+///
+/// # Examples:
+///
+/// ```
+/// # use nalgebra::{Matrix2, Vector2};
+/// let mut vec1 = Vector2::new(1.0, 2.0);
+/// let vec2 = Vector2::new(0.1, 0.2);
+/// let mat = Matrix2::new(1.0, 2.0,
+///                        3.0, 4.0);
+/// vec1.gemv(10.0, &mat, &vec2, 5.0);
+/// assert_eq!(vec1, Vector2::new(10.0, 21.0));
+/// ```
+#[inline]
+pub unsafe fn gemv_uninit<Status, T, D1: Dim, R2: Dim, C2: Dim, D3: Dim, SA, SB, SC>(
+    status: Status,
+    y: &mut Vector<Status::Value, D1, SA>,
+    alpha: T,
+    a: &Matrix<T, R2, C2, SB>,
+    x: &Vector<T, D3, SC>,
+    beta: T,
+) where
+    Status: InitStatus<T>,
+    T: Scalar + Zero + One + ClosedAdd + ClosedMul,
+    SA: RawStorageMut<Status::Value, D1>,
+    SB: RawStorage<T, R2, C2>,
+    SC: RawStorage<T, D3>,
+    ShapeConstraint: DimEq<D1, R2> + AreMultipliable<R2, C2, D3, U1>,
+{
+    let dim1 = y.nrows();
+    let (nrows2, ncols2) = a.shape();
+    let dim3 = x.nrows();
+
+    assert!(
+        ncols2 == dim3 && dim1 == nrows2,
+        "Gemv: dimensions mismatch."
+    );
+
+    if ncols2 == 0 {
+        if beta.is_zero() {
+            y.apply(|e| Status::init(e, T::zero()));
+        } else {
+            // SAFETY: this is UB if y is uninitialized.
+            y.apply(|e| *Status::assume_init_mut(e) *= beta.inlined_clone());
+        }
+        return;
+    }
+
+    // TODO: avoid bound checks.
+    let col2 = a.column(0);
+    let val = x.vget_unchecked(0).inlined_clone();
+
+    // SAFETY: this is the call that makes this method unsafe: it is UB if Status = Uninit and beta != 0.
+    axcpy_uninit(status, y, alpha.inlined_clone(), &col2, val, beta);
+
+    for j in 1..ncols2 {
+        let col2 = a.column(j);
+        let val = x.vget_unchecked(j).inlined_clone();
+
+        // SAFETY: because y was initialized above, we can use the initialized status.
+        axcpy_uninit(
+            Initialized(status),
+            y,
+            alpha.inlined_clone(),
+            &col2,
+            val,
+            T::one(),
+        );
+    }
+}
+
+/// Computes `self = alpha * a * b + beta * self`, where `a, b, self` are matrices.
+/// `alpha` and `beta` are scalar.
+///
+/// If `beta` is zero, `self` is never read.
+///
+/// # Examples:
+///
+/// ```
+/// # #[macro_use] extern crate approx;
+/// # use nalgebra::{Matrix2x3, Matrix3x4, Matrix2x4};
+/// let mut mat1 = Matrix2x4::identity();
+/// let mat2 = Matrix2x3::new(1.0, 2.0, 3.0,
+///                           4.0, 5.0, 6.0);
+/// let mat3 = Matrix3x4::new(0.1, 0.2, 0.3, 0.4,
+///                           0.5, 0.6, 0.7, 0.8,
+///                           0.9, 1.0, 1.1, 1.2);
+/// let expected = mat2 * mat3 * 10.0 + mat1 * 5.0;
+///
+/// mat1.gemm(10.0, &mat2, &mat3, 5.0);
+/// assert_relative_eq!(mat1, expected);
+/// ```
+#[inline]
+pub unsafe fn gemm_uninit<
+    Status,
+    T,
+    R1: Dim,
+    C1: Dim,
+    R2: Dim,
+    C2: Dim,
+    R3: Dim,
+    C3: Dim,
+    SA,
+    SB,
+    SC,
+>(
+    status: Status,
+    y: &mut Matrix<Status::Value, R1, C1, SA>,
+    alpha: T,
+    a: &Matrix<T, R2, C2, SB>,
+    b: &Matrix<T, R3, C3, SC>,
+    beta: T,
+) where
+    Status: InitStatus<T>,
+    T: Scalar + Zero + One + ClosedAdd + ClosedMul,
+    SA: RawStorageMut<Status::Value, R1, C1>,
+    SB: RawStorage<T, R2, C2>,
+    SC: RawStorage<T, R3, C3>,
+    ShapeConstraint:
+        SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C3> + AreMultipliable<R2, C2, R3, C3>,
+{
+    let ncols1 = y.ncols();
+
+    #[cfg(feature = "std")]
+    {
+        // We assume large matrices will be Dynamic but small matrices static.
+        // We could use matrixmultiply for large statically-sized matrices but the performance
+        // threshold to activate it would be different from SMALL_DIM because our code optimizes
+        // better for statically-sized matrices.
+        if R1::is::<Dynamic>()
+            || C1::is::<Dynamic>()
+            || R2::is::<Dynamic>()
+            || C2::is::<Dynamic>()
+            || R3::is::<Dynamic>()
+            || C3::is::<Dynamic>()
+        {
+            // matrixmultiply can be used only if the std feature is available.
+            let nrows1 = y.nrows();
+            let (nrows2, ncols2) = a.shape();
+            let (nrows3, ncols3) = b.shape();
+
+            // Threshold determined empirically.
+            const SMALL_DIM: usize = 5;
+
+            if nrows1 > SMALL_DIM && ncols1 > SMALL_DIM && nrows2 > SMALL_DIM && ncols2 > SMALL_DIM
+            {
+                assert_eq!(
+                    ncols2, nrows3,
+                    "gemm: dimensions mismatch for multiplication."
+                );
+                assert_eq!(
+                    (nrows1, ncols1),
+                    (nrows2, ncols3),
+                    "gemm: dimensions mismatch for addition."
+                );
+
+                // NOTE: this case should never happen because we enter this
+                // codepath only when ncols2 > SMALL_DIM. Though we keep this
+                // here just in case if in the future we change the conditions to
+                // enter this codepath.
+                if ncols2 == 0 {
+                    // NOTE: we can't just always multiply by beta
+                    // because we documented the guaranty that `self` is
+                    // never read if `beta` is zero.
+                    if beta.is_zero() {
+                        y.apply(|e| Status::init(e, T::zero()));
+                    } else {
+                        // SAFETY: this is UB if Status = Uninit
+                        y.apply(|e| *Status::assume_init_mut(e) *= beta.inlined_clone());
+                    }
+                    return;
+                }
+
+                if T::is::<f32>() {
+                    let (rsa, csa) = a.strides();
+                    let (rsb, csb) = b.strides();
+                    let (rsc, csc) = y.strides();
+
+                    matrixmultiply::sgemm(
+                        nrows2,
+                        ncols2,
+                        ncols3,
+                        mem::transmute_copy(&alpha),
+                        a.data.ptr() as *const f32,
+                        rsa as isize,
+                        csa as isize,
+                        b.data.ptr() as *const f32,
+                        rsb as isize,
+                        csb as isize,
+                        mem::transmute_copy(&beta),
+                        y.data.ptr_mut() as *mut f32,
+                        rsc as isize,
+                        csc as isize,
+                    );
+                    return;
+                } else if T::is::<f64>() {
+                    let (rsa, csa) = a.strides();
+                    let (rsb, csb) = b.strides();
+                    let (rsc, csc) = y.strides();
+
+                    matrixmultiply::dgemm(
+                        nrows2,
+                        ncols2,
+                        ncols3,
+                        mem::transmute_copy(&alpha),
+                        a.data.ptr() as *const f64,
+                        rsa as isize,
+                        csa as isize,
+                        b.data.ptr() as *const f64,
+                        rsb as isize,
+                        csb as isize,
+                        mem::transmute_copy(&beta),
+                        y.data.ptr_mut() as *mut f64,
+                        rsc as isize,
+                        csc as isize,
+                    );
+                    return;
+                }
+            }
+        }
+    }
+
+    for j1 in 0..ncols1 {
+        // TODO: avoid bound checks.
+        // SAFETY: this is UB if Status = Uninit && beta != 0
+        gemv_uninit(
+            status,
+            &mut y.column_mut(j1),
+            alpha.inlined_clone(),
+            a,
+            &b.column(j1),
+            beta.inlined_clone(),
+        );
+    }
+}
diff --git a/src/base/construction.rs b/src/base/construction.rs
index 2a7a80da..ae129f0d 100644
--- a/src/base/construction.rs
+++ b/src/base/construction.rs
@@ -1,6 +1,8 @@
 #[cfg(all(feature = "alloc", not(feature = "std")))]
 use alloc::vec::Vec;
 
+#[cfg(feature = "arbitrary")]
+use crate::base::storage::Owned;
 #[cfg(feature = "arbitrary")]
 use quickcheck::{Arbitrary, Gen};
 
@@ -11,17 +13,49 @@ use rand::{
     Rng,
 };
 
-use std::{iter, mem::MaybeUninit};
+use std::iter;
 use typenum::{self, Cmp, Greater};
 
 use simba::scalar::{ClosedAdd, ClosedMul};
 
-use crate::base::allocator::{Allocator, InnerAllocator};
+use crate::base::allocator::Allocator;
 use crate::base::dimension::{Dim, DimName, Dynamic, ToTypenum};
-use crate::base::storage::Storage;
+use crate::base::storage::RawStorage;
 use crate::base::{
     ArrayStorage, Const, DefaultAllocator, Matrix, OMatrix, OVector, Scalar, Unit, Vector,
 };
+use crate::UninitMatrix;
+use std::mem::MaybeUninit;
+
+/// When "no_unsound_assume_init" is enabled, expands to `unimplemented!()` instead of `new_uninitialized_generic().assume_init()`.
+/// Intended as a placeholder, each callsite should be refactored to use uninitialized memory soundly
+#[macro_export]
+macro_rules! unimplemented_or_uninitialized_generic {
+    ($nrows:expr, $ncols:expr) => {{
+        #[cfg(feature="no_unsound_assume_init")] {
+            // Some of the call sites need the number of rows and columns from this to infer a type, so
+            // uninitialized memory is used to infer the type, as `T: Zero` isn't available at all callsites.
+            // This may technically still be UB even though the assume_init is dead code, but all callsites should be fixed before #556 is closed.
+            let typeinference_helper = crate::base::Matrix::new_uninitialized_generic($nrows, $ncols);
+            unimplemented!();
+            typeinference_helper.assume_init()
+        }
+        #[cfg(not(feature="no_unsound_assume_init"))] { crate::base::Matrix::new_uninitialized_generic($nrows, $ncols).assume_init() }
+    }}
+}
+
+impl<T: Scalar, R: Dim, C: Dim> UninitMatrix<T, R, C>
+where
+    DefaultAllocator: Allocator<T, R, C>,
+{
+    pub fn uninit(nrows: R, ncols: C) -> Self {
+        // SAFETY: this is OK because the dimension automatically match the storage
+        //         because we are building an owned storage.
+        unsafe {
+            Self::from_data_statically_unchecked(DefaultAllocator::allocate_uninit(nrows, ncols))
+        }
+    }
+}
 
 /// # Generic constructors
 /// This set of matrix and vector construction functions are all generic
@@ -29,16 +63,23 @@ use crate::base::{
 /// the dimension as inputs.
 ///
 /// These functions should only be used when working on dimension-generic code.
-impl<T, R: Dim, C: Dim> OMatrix<T, R, C>
+impl<T: Scalar, R: Dim, C: Dim> OMatrix<T, R, C>
 where
     DefaultAllocator: Allocator<T, R, C>,
 {
+    /// Creates a new uninitialized matrix.
+    ///
+    /// # Safety
+    /// If the matrix has a compile-time dimension, this panics
+    /// if `nrows != R::to_usize()` or `ncols != C::to_usize()`.
+    #[inline]
+    pub unsafe fn new_uninitialized_generic(nrows: R, ncols: C) -> MaybeUninit<Self> {
+        Self::from_uninitialized_data(DefaultAllocator::allocate_uninitialized(nrows, ncols))
+    }
+
     /// Creates a matrix with all its elements set to `elem`.
     #[inline]
-    pub fn from_element_generic(nrows: R, ncols: C, elem: T) -> Self
-    where
-        T: Clone,
-    {
+    pub fn from_element_generic(nrows: R, ncols: C, elem: T) -> Self {
         let len = nrows.value() * ncols.value();
         Self::from_iterator_generic(nrows, ncols, iter::repeat(elem).take(len))
     }
@@ -47,10 +88,7 @@ where
     ///
     /// Same as `from_element_generic`.
     #[inline]
-    pub fn repeat_generic(nrows: R, ncols: C, elem: T) -> Self
-    where
-        T: Clone,
-    {
+    pub fn repeat_generic(nrows: R, ncols: C, elem: T) -> Self {
         let len = nrows.value() * ncols.value();
         Self::from_iterator_generic(nrows, ncols, iter::repeat(elem).take(len))
     }
@@ -59,7 +97,7 @@ where
     #[inline]
     pub fn zeros_generic(nrows: R, ncols: C) -> Self
     where
-        T: Zero + Clone,
+        T: Zero,
     {
         Self::from_element_generic(nrows, ncols, T::zero())
     }
@@ -79,37 +117,32 @@ where
     /// The order of elements in the slice must follow the usual mathematic writing, i.e.,
     /// row-by-row.
     #[inline]
-    pub fn from_row_slice_generic(nrows: R, ncols: C, slice: &[T]) -> Self
-    where
-        T: Clone,
-    {
+    pub fn from_row_slice_generic(nrows: R, ncols: C, slice: &[T]) -> Self {
         assert!(
             slice.len() == nrows.value() * ncols.value(),
             "Matrix init. error: the slice did not contain the right number of elements."
         );
 
-        let mut res = Self::new_uninitialized_generic(nrows, ncols);
+        let mut res = Matrix::uninit(nrows, ncols);
         let mut iter = slice.iter();
 
-        for i in 0..nrows.value() {
-            for j in 0..ncols.value() {
-                unsafe {
-                    *res.get_unchecked_mut((i, j)) = MaybeUninit::new(iter.next().unwrap().clone());
+        unsafe {
+            for i in 0..nrows.value() {
+                for j in 0..ncols.value() {
+                    *res.get_unchecked_mut((i, j)) =
+                        MaybeUninit::new(iter.next().unwrap().inlined_clone())
                 }
             }
-        }
 
-        // Safety: all entries have been initialized.
-        unsafe { res.assume_init() }
+            // SAFETY: the result has been fully initialized above.
+            res.assume_init()
+        }
     }
 
     /// Creates a matrix with its elements filled with the components provided by a slice. The
     /// components must have the same layout as the matrix data storage (i.e. column-major).
     #[inline]
-    pub fn from_column_slice_generic(nrows: R, ncols: C, slice: &[T]) -> Self
-    where
-        T: Clone,
-    {
+    pub fn from_column_slice_generic(nrows: R, ncols: C, slice: &[T]) -> Self {
         Self::from_iterator_generic(nrows, ncols, slice.iter().cloned())
     }
 
@@ -120,18 +153,18 @@ where
     where
         F: FnMut(usize, usize) -> T,
     {
-        let mut res = Self::new_uninitialized_generic(nrows, ncols);
+        let mut res = Matrix::uninit(nrows, ncols);
 
-        for j in 0..ncols.value() {
-            for i in 0..nrows.value() {
-                unsafe {
+        unsafe {
+            for j in 0..ncols.value() {
+                for i in 0..nrows.value() {
                     *res.get_unchecked_mut((i, j)) = MaybeUninit::new(f(i, j));
                 }
             }
-        }
 
-        // Safety: all entries have been initialized.
-        unsafe { res.assume_init() }
+            // SAFETY: the result has been fully initialized above.
+            res.assume_init()
+        }
     }
 
     /// Creates a new identity matrix.
@@ -141,7 +174,7 @@ where
     #[inline]
     pub fn identity_generic(nrows: R, ncols: C) -> Self
     where
-        T: Zero + One + Scalar,
+        T: Zero + One,
     {
         Self::from_diagonal_element_generic(nrows, ncols, T::one())
     }
@@ -153,7 +186,7 @@ where
     #[inline]
     pub fn from_diagonal_element_generic(nrows: R, ncols: C, elt: T) -> Self
     where
-        T: Zero + One + Scalar,
+        T: Zero + One,
     {
         let mut res = Self::zeros_generic(nrows, ncols);
 
@@ -171,7 +204,7 @@ where
     #[inline]
     pub fn from_partial_diagonal_generic(nrows: R, ncols: C, elts: &[T]) -> Self
     where
-        T: Zero + Clone,
+        T: Zero,
     {
         let mut res = Self::zeros_generic(nrows, ncols);
         assert!(
@@ -180,7 +213,7 @@ where
         );
 
         for (i, elt) in elts.iter().enumerate() {
-            unsafe { *res.get_unchecked_mut((i, i)) = elt.clone() }
+            unsafe { *res.get_unchecked_mut((i, i)) = elt.inlined_clone() }
         }
 
         res
@@ -205,8 +238,7 @@ where
     #[inline]
     pub fn from_rows<SB>(rows: &[Matrix<T, Const<1>, C, SB>]) -> Self
     where
-        T: Clone,
-        SB: Storage<T, Const<1>, C>,
+        SB: RawStorage<T, Const<1>, C>,
     {
         assert!(!rows.is_empty(), "At least one row must be given.");
         let nrows = R::try_to_usize().unwrap_or_else(|| rows.len());
@@ -225,7 +257,7 @@ where
 
         // TODO: optimize that.
         Self::from_fn_generic(R::from_usize(nrows), C::from_usize(ncols), |i, j| {
-            rows[i][(0, j)].clone()
+            rows[i][(0, j)].inlined_clone()
         })
     }
 
@@ -248,8 +280,7 @@ where
     #[inline]
     pub fn from_columns<SB>(columns: &[Vector<T, R, SB>]) -> Self
     where
-        T: Clone,
-        SB: Storage<T, R>,
+        SB: RawStorage<T, R>,
     {
         assert!(!columns.is_empty(), "At least one column must be given.");
         let ncols = C::try_to_usize().unwrap_or_else(|| columns.len());
@@ -268,7 +299,7 @@ where
 
         // TODO: optimize that.
         Self::from_fn_generic(R::from_usize(nrows), C::from_usize(ncols), |i, j| {
-            columns[j][i].clone()
+            columns[j][i].inlined_clone()
         })
     }
 
@@ -321,6 +352,7 @@ where
 
 impl<T, D: Dim> OMatrix<T, D, D>
 where
+    T: Scalar,
     DefaultAllocator: Allocator<T, D, D>,
 {
     /// Creates a square matrix with its diagonal set to `diag` and all other entries set to 0.
@@ -342,11 +374,11 @@ where
     ///         dm[(2, 0)] == 0.0 && dm[(2, 1)] == 0.0 && dm[(2, 2)] == 3.0);
     /// ```
     #[inline]
-    pub fn from_diagonal<SB: Storage<T, D>>(diag: &Vector<T, D, SB>) -> Self
+    pub fn from_diagonal<SB: RawStorage<T, D>>(diag: &Vector<T, D, SB>) -> Self
     where
-        T: Zero + Scalar,
+        T: Zero,
     {
-        let (dim, _) = diag.data.shape();
+        let (dim, _) = diag.shape_generic();
         let mut res = Self::zeros_generic(dim, dim);
 
         for i in 0..diag.len() {
@@ -366,6 +398,12 @@ where
  */
 macro_rules! impl_constructors(
     ($($Dims: ty),*; $(=> $DimIdent: ident: $DimBound: ident),*; $($gargs: expr),*; $($args: ident),*) => {
+        /// Creates a new uninitialized matrix or vector.
+        #[inline]
+        pub unsafe fn new_uninitialized($($args: usize),*) -> MaybeUninit<Self> {
+            Self::new_uninitialized_generic($($gargs),*)
+        }
+
         /// Creates a matrix or vector with all its elements set to `elem`.
         ///
         /// # Example
@@ -387,10 +425,7 @@ macro_rules! impl_constructors(
         ///         dm[(1, 0)] == 2.0 && dm[(1, 1)] == 2.0 && dm[(1, 2)] == 2.0);
         /// ```
         #[inline]
-        pub fn from_element($($args: usize,)* elem: T) -> Self
-        where
-            T: Clone
-        {
+        pub fn from_element($($args: usize,)* elem: T) -> Self {
             Self::from_element_generic($($gargs, )* elem)
         }
 
@@ -417,10 +452,7 @@ macro_rules! impl_constructors(
         ///         dm[(1, 0)] == 2.0 && dm[(1, 1)] == 2.0 && dm[(1, 2)] == 2.0);
         /// ```
         #[inline]
-        pub fn repeat($($args: usize,)* elem: T) -> Self
-        where
-            T: Clone
-        {
+        pub fn repeat($($args: usize,)* elem: T) -> Self {
             Self::repeat_generic($($gargs, )* elem)
         }
 
@@ -446,9 +478,7 @@ macro_rules! impl_constructors(
         /// ```
         #[inline]
         pub fn zeros($($args: usize),*) -> Self
-        where
-            T: Zero + Clone
-        {
+            where T: Zero {
             Self::zeros_generic($($gargs),*)
         }
 
@@ -504,7 +534,8 @@ macro_rules! impl_constructors(
         ///         dm[(1, 0)] == 3 && dm[(1, 1)] == 4 && dm[(1, 2)] == 5);
         /// ```
         #[inline]
-        pub fn from_fn<F: FnMut(usize, usize) -> T>($($args: usize,)* f: F) -> Self {
+        pub fn from_fn<F>($($args: usize,)* f: F) -> Self
+            where F: FnMut(usize, usize) -> T {
             Self::from_fn_generic($($gargs, )* f)
         }
 
@@ -528,9 +559,7 @@ macro_rules! impl_constructors(
         /// ```
         #[inline]
         pub fn identity($($args: usize,)*) -> Self
-        where
-            T: Zero + One + Scalar
-        {
+            where T: Zero + One {
             Self::identity_generic($($gargs),* )
         }
 
@@ -553,9 +582,7 @@ macro_rules! impl_constructors(
         /// ```
         #[inline]
         pub fn from_diagonal_element($($args: usize,)* elt: T) -> Self
-        where
-            T: Zero + One + Scalar
-        {
+            where T: Zero + One {
             Self::from_diagonal_element_generic($($gargs, )* elt)
         }
 
@@ -582,9 +609,7 @@ macro_rules! impl_constructors(
         /// ```
         #[inline]
         pub fn from_partial_diagonal($($args: usize,)* elts: &[T]) -> Self
-        where
-            T: Zero + Scalar
-        {
+            where T: Zero {
             Self::from_partial_diagonal_generic($($gargs, )* elts)
         }
 
@@ -603,16 +628,14 @@ macro_rules! impl_constructors(
         #[inline]
         #[cfg(feature = "rand")]
         pub fn new_random($($args: usize),*) -> Self
-        where
-            Standard: Distribution<T>
-        {
+            where Standard: Distribution<T> {
             Self::new_random_generic($($gargs),*)
         }
     }
 );
 
 /// # Constructors of statically-sized vectors or statically-sized matrices
-impl<T, R: DimName, C: DimName> OMatrix<T, R, C>
+impl<T: Scalar, R: DimName, C: DimName> OMatrix<T, R, C>
 where
     DefaultAllocator: Allocator<T, R, C>,
 {
@@ -623,19 +646,8 @@ where
     ); // Arguments for non-generic constructors.
 }
 
-impl<T, R: DimName, C: DimName> OMatrix<T, R, C>
-where
-    DefaultAllocator: Allocator<T, R, C>,
-{
-    /// Creates a new uninitialized matrix or vector.
-    #[inline]
-    pub fn new_uninitialized() -> OMatrix<MaybeUninit<T>, R, C> {
-        Self::new_uninitialized_generic(R::name(), C::name())
-    }
-}
-
 /// # Constructors of matrices with a dynamic number of columns
-impl<T, R: DimName> OMatrix<T, R, Dynamic>
+impl<T: Scalar, R: DimName> OMatrix<T, R, Dynamic>
 where
     DefaultAllocator: Allocator<T, R, Dynamic>,
 {
@@ -645,19 +657,8 @@ where
                    ncols);
 }
 
-impl<T, R: DimName> OMatrix<T, R, Dynamic>
-where
-    DefaultAllocator: Allocator<T, R, Dynamic>,
-{
-    /// Creates a new uninitialized matrix or vector.
-    #[inline]
-    pub fn new_uninitialized(ncols: usize) -> OMatrix<MaybeUninit<T>, R, Dynamic> {
-        Self::new_uninitialized_generic(R::name(), Dynamic::new(ncols))
-    }
-}
-
 /// # Constructors of dynamic vectors and matrices with a dynamic number of rows
-impl<T, C: DimName> OMatrix<T, Dynamic, C>
+impl<T: Scalar, C: DimName> OMatrix<T, Dynamic, C>
 where
     DefaultAllocator: Allocator<T, Dynamic, C>,
 {
@@ -667,19 +668,8 @@ where
                    nrows);
 }
 
-impl<T, C: DimName> OMatrix<T, Dynamic, C>
-where
-    DefaultAllocator: Allocator<T, Dynamic, C>,
-{
-    /// Creates a new uninitialized matrix or vector.
-    #[inline]
-    pub fn new_uninitialized(nrows: usize) -> OMatrix<MaybeUninit<T>, Dynamic, C> {
-        Self::new_uninitialized_generic(Dynamic::new(nrows), C::name())
-    }
-}
-
 /// # Constructors of fully dynamic matrices
-impl<T> OMatrix<T, Dynamic, Dynamic>
+impl<T: Scalar> OMatrix<T, Dynamic, Dynamic>
 where
     DefaultAllocator: Allocator<T, Dynamic, Dynamic>,
 {
@@ -689,20 +679,6 @@ where
                    nrows, ncols);
 }
 
-impl<T> OMatrix<T, Dynamic, Dynamic>
-where
-    DefaultAllocator: Allocator<T, Dynamic, Dynamic>,
-{
-    /// Creates a new uninitialized matrix or vector.
-    #[inline]
-    pub fn new_uninitialized(
-        nrows: usize,
-        ncols: usize,
-    ) -> OMatrix<MaybeUninit<T>, Dynamic, Dynamic> {
-        Self::new_uninitialized_generic(Dynamic::new(nrows), Dynamic::new(ncols))
-    }
-}
-
 /*
  *
  * Constructors that don't necessarily require all dimensions
@@ -711,10 +687,8 @@ where
  */
 macro_rules! impl_constructors_from_data(
     ($data: ident; $($Dims: ty),*; $(=> $DimIdent: ident: $DimBound: ident),*; $($gargs: expr),*; $($args: ident),*) => {
-        impl<T, $($DimIdent: $DimBound, )*> OMatrix<T $(, $Dims)*>
-        where
-            DefaultAllocator: Allocator<T $(, $Dims)*>
-        {
+        impl<T: Scalar, $($DimIdent: $DimBound, )*> OMatrix<T $(, $Dims)*>
+        where DefaultAllocator: Allocator<T $(, $Dims)*> {
             /// Creates a matrix with its elements filled with the components provided by a slice
             /// in row-major order.
             ///
@@ -741,10 +715,7 @@ macro_rules! impl_constructors_from_data(
             ///         dm[(1, 0)] == 3 && dm[(1, 1)] == 4 && dm[(1, 2)] == 5);
             /// ```
             #[inline]
-            pub fn from_row_slice($($args: usize,)* $data: &[T]) -> Self
-            where
-                T: Clone
-            {
+            pub fn from_row_slice($($args: usize,)* $data: &[T]) -> Self {
                 Self::from_row_slice_generic($($gargs, )* $data)
             }
 
@@ -771,10 +742,7 @@ macro_rules! impl_constructors_from_data(
             ///         dm[(1, 0)] == 1 && dm[(1, 1)] == 3 && dm[(1, 2)] == 5);
             /// ```
             #[inline]
-            pub fn from_column_slice($($args: usize,)* $data: &[T]) -> Self
-            where
-                T: Clone
-            {
+            pub fn from_column_slice($($args: usize,)* $data: &[T]) -> Self {
                 Self::from_column_slice_generic($($gargs, )* $data)
             }
 
@@ -877,7 +845,7 @@ where
 }
 
 #[cfg(feature = "rand-no-std")]
-impl<T, R: Dim, C: Dim> Distribution<OMatrix<T, R, C>> for Standard
+impl<T: Scalar, R: Dim, C: Dim> Distribution<OMatrix<T, R, C>> for Standard
 where
     DefaultAllocator: Allocator<T, R, C>,
     Standard: Distribution<T>,
@@ -892,10 +860,13 @@ where
 }
 
 #[cfg(feature = "arbitrary")]
-impl<T, R: Dim, C: Dim> Arbitrary for OMatrix<T, R, C>
+impl<T, R, C> Arbitrary for OMatrix<T, R, C>
 where
-    T: Arbitrary + Send,
+    R: Dim,
+    C: Dim,
+    T: Scalar + Arbitrary + Send,
     DefaultAllocator: Allocator<T, R, C>,
+    Owned<T, R, C>: Clone + Send,
 {
     #[inline]
     fn arbitrary(g: &mut Gen) -> Self {
diff --git a/src/base/construction_slice.rs b/src/base/construction_slice.rs
index 650fbfd0..7094bdca 100644
--- a/src/base/construction_slice.rs
+++ b/src/base/construction_slice.rs
@@ -1,11 +1,13 @@
 use crate::base::dimension::{Const, Dim, DimName, Dynamic};
 use crate::base::matrix_slice::{SliceStorage, SliceStorageMut};
-use crate::base::{MatrixSlice, MatrixSliceMutMN};
+use crate::base::{MatrixSlice, MatrixSliceMutMN, Scalar};
 
 use num_rational::Ratio;
 
 /// # Creating matrix slices from `&[T]`
-impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim> MatrixSlice<'a, T, R, C, RStride, CStride> {
+impl<'a, T: Scalar, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
+    MatrixSlice<'a, T, R, C, RStride, CStride>
+{
     /// Creates, without bound-checking, a matrix slice from an array and with dimensions and strides specified by generic types instances.
     ///
     /// # Safety
@@ -55,7 +57,7 @@ impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim> MatrixSlice<'a, T, R, C,
     }
 }
 
-impl<'a, T, R: Dim, C: Dim> MatrixSlice<'a, T, R, C> {
+impl<'a, T: Scalar, R: Dim, C: Dim> MatrixSlice<'a, T, R, C> {
     /// Creates, without bound-checking, a matrix slice from an array and with dimensions specified by generic types instances.
     ///
     /// # Safety
@@ -85,7 +87,7 @@ impl<'a, T, R: Dim, C: Dim> MatrixSlice<'a, T, R, C> {
 
 macro_rules! impl_constructors(
     ($($Dims: ty),*; $(=> $DimIdent: ident: $DimBound: ident),*; $($gargs: expr),*; $($args: ident),*) => {
-        impl<'a, T, $($DimIdent: $DimBound),*> MatrixSlice<'a, T, $($Dims),*> {
+        impl<'a, T: Scalar, $($DimIdent: $DimBound),*> MatrixSlice<'a, T, $($Dims),*> {
             /// Creates a new matrix slice from the given data array.
             ///
             /// Panics if `data` does not contain enough elements.
@@ -101,7 +103,7 @@ macro_rules! impl_constructors(
             }
         }
 
-        impl<'a, T, $($DimIdent: $DimBound, )*> MatrixSlice<'a, T, $($Dims,)* Dynamic, Dynamic> {
+        impl<'a, T: Scalar, $($DimIdent: $DimBound, )*> MatrixSlice<'a, T, $($Dims,)* Dynamic, Dynamic> {
             /// Creates a new matrix slice with the specified strides from the given data array.
             ///
             /// Panics if `data` does not contain enough elements.
@@ -141,7 +143,7 @@ impl_constructors!(Dynamic, Dynamic;
                    nrows, ncols);
 
 /// # Creating mutable matrix slices from `&mut [T]`
-impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
+impl<'a, T: Scalar, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
     MatrixSliceMutMN<'a, T, R, C, RStride, CStride>
 {
     /// Creates, without bound-checking, a mutable matrix slice from an array and with dimensions and strides specified by generic types instances.
@@ -215,7 +217,7 @@ impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
     }
 }
 
-impl<'a, T, R: Dim, C: Dim> MatrixSliceMutMN<'a, T, R, C> {
+impl<'a, T: Scalar, R: Dim, C: Dim> MatrixSliceMutMN<'a, T, R, C> {
     /// Creates, without bound-checking, a mutable matrix slice from an array and with dimensions specified by generic types instances.
     ///
     /// # Safety
@@ -245,7 +247,7 @@ impl<'a, T, R: Dim, C: Dim> MatrixSliceMutMN<'a, T, R, C> {
 
 macro_rules! impl_constructors_mut(
     ($($Dims: ty),*; $(=> $DimIdent: ident: $DimBound: ident),*; $($gargs: expr),*; $($args: ident),*) => {
-        impl<'a, T, $($DimIdent: $DimBound),*> MatrixSliceMutMN<'a, T, $($Dims),*> {
+        impl<'a, T: Scalar, $($DimIdent: $DimBound),*> MatrixSliceMutMN<'a, T, $($Dims),*> {
             /// Creates a new mutable matrix slice from the given data array.
             ///
             /// Panics if `data` does not contain enough elements.
@@ -261,7 +263,7 @@ macro_rules! impl_constructors_mut(
             }
         }
 
-        impl<'a, T, $($DimIdent: $DimBound, )*> MatrixSliceMutMN<'a, T, $($Dims,)* Dynamic, Dynamic> {
+        impl<'a, T: Scalar, $($DimIdent: $DimBound, )*> MatrixSliceMutMN<'a, T, $($Dims,)* Dynamic, Dynamic> {
             /// Creates a new mutable matrix slice with the specified strides from the given data array.
             ///
             /// Panics if `data` does not contain enough elements.
diff --git a/src/base/conversion.rs b/src/base/conversion.rs
index b8a50048..ec7fd936 100644
--- a/src/base/conversion.rs
+++ b/src/base/conversion.rs
@@ -1,10 +1,8 @@
-use std::borrow::{Borrow, BorrowMut};
-use std::convert::{AsMut, AsRef, From, Into};
-use std::mem::{self, ManuallyDrop, MaybeUninit};
-
 #[cfg(all(feature = "alloc", not(feature = "std")))]
 use alloc::vec::Vec;
 use simba::scalar::{SubsetOf, SupersetOf};
+use std::borrow::{Borrow, BorrowMut};
+use std::convert::{AsMut, AsRef, From, Into};
 
 use simba::simd::{PrimitiveSimdValue, SimdValue};
 
@@ -16,7 +14,7 @@ use crate::base::dimension::{
     Const, Dim, DimName, U1, U10, U11, U12, U13, U14, U15, U16, U2, U3, U4, U5, U6, U7, U8, U9,
 };
 use crate::base::iter::{MatrixIter, MatrixIterMut};
-use crate::base::storage::{ContiguousStorage, ContiguousStorageMut, Storage, StorageMut};
+use crate::base::storage::{IsContiguous, RawStorage, RawStorageMut};
 use crate::base::{
     ArrayStorage, DVectorSlice, DVectorSliceMut, DefaultAllocator, Matrix, MatrixSlice,
     MatrixSliceMut, OMatrix, Scalar,
@@ -26,12 +24,17 @@ use crate::base::{DVector, VecStorage};
 use crate::base::{SliceStorage, SliceStorageMut};
 use crate::constraint::DimEq;
 use crate::{IsNotStaticOne, RowSVector, SMatrix, SVector};
+use std::mem::MaybeUninit;
 
 // TODO: too bad this won't work for slice conversions.
-impl<T1, T2, R1: Dim, C1: Dim, R2: Dim, C2: Dim> SubsetOf<OMatrix<T2, R2, C2>>
-    for OMatrix<T1, R1, C1>
+impl<T1, T2, R1, C1, R2, C2> SubsetOf<OMatrix<T2, R2, C2>> for OMatrix<T1, R1, C1>
 where
-    T2: SupersetOf<T1>,
+    R1: Dim,
+    C1: Dim,
+    R2: Dim,
+    C2: Dim,
+    T1: Scalar,
+    T2: Scalar + SupersetOf<T1>,
     DefaultAllocator:
         Allocator<T2, R2, C2> + Allocator<T1, R1, C1> + SameShapeAllocator<T1, R1, C1, R2, C2>,
     ShapeConstraint: SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C2>,
@@ -41,11 +44,11 @@ where
         let (nrows, ncols) = self.shape();
         let nrows2 = R2::from_usize(nrows);
         let ncols2 = C2::from_usize(ncols);
-
-        let mut res = Matrix::new_uninitialized_generic(nrows2, ncols2);
+        let mut res = Matrix::uninit(nrows2, ncols2);
 
         for i in 0..nrows {
             for j in 0..ncols {
+                // Safety: all indices are in range.
                 unsafe {
                     *res.get_unchecked_mut((i, j)) =
                         MaybeUninit::new(T2::from_subset(self.get_unchecked((i, j))));
@@ -53,7 +56,7 @@ where
             }
         }
 
-        // Safety: all entries have been initialized.
+        // Safety: res is now fully initialized.
         unsafe { res.assume_init() }
     }
 
@@ -67,23 +70,25 @@ where
         let (nrows2, ncols2) = m.shape();
         let nrows = R1::from_usize(nrows2);
         let ncols = C1::from_usize(ncols2);
+        let mut res = Matrix::uninit(nrows, ncols);
 
-        let mut res = OMatrix::new_uninitialized_generic(nrows, ncols);
         for i in 0..nrows2 {
             for j in 0..ncols2 {
+                // Safety: all indices are in range.
                 unsafe {
                     *res.get_unchecked_mut((i, j)) =
-                        MaybeUninit::new(m.get_unchecked((i, j)).to_subset_unchecked());
+                        MaybeUninit::new(m.get_unchecked((i, j)).to_subset_unchecked())
                 }
             }
         }
 
-        // Safety: all entries have been initialized.
         unsafe { res.assume_init() }
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S: Storage<T, R, C>> IntoIterator for &'a Matrix<T, R, C, S> {
+impl<'a, T: Scalar, R: Dim, C: Dim, S: RawStorage<T, R, C>> IntoIterator
+    for &'a Matrix<T, R, C, S>
+{
     type Item = &'a T;
     type IntoIter = MatrixIter<'a, T, R, C, S>;
 
@@ -93,7 +98,9 @@ impl<'a, T, R: Dim, C: Dim, S: Storage<T, R, C>> IntoIterator for &'a Matrix<T,
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S: StorageMut<T, R, C>> IntoIterator for &'a mut Matrix<T, R, C, S> {
+impl<'a, T: Scalar, R: Dim, C: Dim, S: RawStorageMut<T, R, C>> IntoIterator
+    for &'a mut Matrix<T, R, C, S>
+{
     type Item = &'a mut T;
     type IntoIter = MatrixIterMut<'a, T, R, C, S>;
 
@@ -103,35 +110,32 @@ impl<'a, T, R: Dim, C: Dim, S: StorageMut<T, R, C>> IntoIterator for &'a mut Mat
     }
 }
 
-impl<T, const D: usize> From<[T; D]> for SVector<T, D> {
+impl<T: Scalar, const D: usize> From<[T; D]> for SVector<T, D> {
     #[inline]
     fn from(arr: [T; D]) -> Self {
-        Self::from_data(ArrayStorage([arr; 1]))
+        unsafe { Self::from_data_statically_unchecked(ArrayStorage([arr; 1])) }
     }
 }
 
-impl<T, const D: usize> From<SVector<T, D>> for [T; D] {
+impl<T: Scalar, const D: usize> From<SVector<T, D>> for [T; D] {
     #[inline]
     fn from(vec: SVector<T, D>) -> Self {
-        let data = ManuallyDrop::new(vec.data.0);
-        // Safety: [[T; D]; 1] always has the same data layout as [T; D].
-        let res = unsafe { (data.as_ptr() as *const [_; D]).read() };
-        mem::forget(data);
-        res
+        // TODO: unfortunately, we must clone because we can move out of an array.
+        vec.data.0[0].clone()
     }
 }
 
-impl<T: Clone, const D: usize> From<[T; D]> for RowSVector<T, D>
+impl<T: Scalar, const D: usize> From<[T; D]> for RowSVector<T, D>
 where
     Const<D>: IsNotStaticOne,
 {
     #[inline]
     fn from(arr: [T; D]) -> Self {
-        SVector::<T, D>::from(arr).transpose_into()
+        SVector::<T, D>::from(arr).transpose()
     }
 }
 
-impl<T: Clone, const D: usize> From<RowSVector<T, D>> for [T; D]
+impl<T: Scalar, const D: usize> From<RowSVector<T, D>> for [T; D]
 where
     Const<D>: IsNotStaticOne,
 {
@@ -144,10 +148,11 @@ where
 macro_rules! impl_from_into_asref_1D(
     ($(($NRows: ident, $NCols: ident) => $SZ: expr);* $(;)*) => {$(
         impl<T, S> AsRef<[T; $SZ]> for Matrix<T, $NRows, $NCols, S>
-        where
-              S: ContiguousStorage<T, $NRows, $NCols> {
+        where T: Scalar,
+              S: RawStorage<T, $NRows, $NCols> + IsContiguous {
             #[inline]
             fn as_ref(&self) -> &[T; $SZ] {
+                // Safety: this is OK thanks to the IsContiguous trait.
                 unsafe {
                     &*(self.data.ptr() as *const [T; $SZ])
                 }
@@ -155,10 +160,11 @@ macro_rules! impl_from_into_asref_1D(
         }
 
         impl<T, S> AsMut<[T; $SZ]> for Matrix<T, $NRows, $NCols, S>
-        where
-              S: ContiguousStorageMut<T, $NRows, $NCols> {
+        where T: Scalar,
+              S: RawStorageMut<T, $NRows, $NCols> + IsContiguous {
             #[inline]
             fn as_mut(&mut self) -> &mut [T; $SZ] {
+                // Safety: this is OK thanks to the IsContiguous trait.
                 unsafe {
                     &mut *(self.data.ptr_mut() as *mut [T; $SZ])
                 }
@@ -182,14 +188,14 @@ impl_from_into_asref_1D!(
     (U13, U1) => 13; (U14, U1) => 14; (U15, U1) => 15; (U16, U1) => 16;
 );
 
-impl<T, const R: usize, const C: usize> From<[[T; R]; C]> for SMatrix<T, R, C> {
+impl<T: Scalar, const R: usize, const C: usize> From<[[T; R]; C]> for SMatrix<T, R, C> {
     #[inline]
     fn from(arr: [[T; R]; C]) -> Self {
-        Self::from_data(ArrayStorage(arr))
+        unsafe { Self::from_data_statically_unchecked(ArrayStorage(arr)) }
     }
 }
 
-impl<T, const R: usize, const C: usize> From<SMatrix<T, R, C>> for [[T; R]; C] {
+impl<T: Scalar, const R: usize, const C: usize> From<SMatrix<T, R, C>> for [[T; R]; C] {
     #[inline]
     fn from(vec: SMatrix<T, R, C>) -> Self {
         vec.data.0
@@ -203,20 +209,22 @@ macro_rules! impl_from_into_asref_borrow_2D(
         ($NRows: ty, $NCols: ty) => ($SZRows: expr, $SZCols: expr);
         $Ref:ident.$ref:ident(), $Mut:ident.$mut:ident()
     ) => {
-        impl<T, S> $Ref<[[T; $SZRows]; $SZCols]> for Matrix<T, $NRows, $NCols, S>
-        where S: ContiguousStorage<T, $NRows, $NCols> {
+        impl<T: Scalar, S> $Ref<[[T; $SZRows]; $SZCols]> for Matrix<T, $NRows, $NCols, S>
+        where S: RawStorage<T, $NRows, $NCols> + IsContiguous {
             #[inline]
             fn $ref(&self) -> &[[T; $SZRows]; $SZCols] {
+                // Safety: OK thanks to the IsContiguous trait.
                 unsafe {
                     &*(self.data.ptr() as *const [[T; $SZRows]; $SZCols])
                 }
             }
         }
 
-        impl<T, S> $Mut<[[T; $SZRows]; $SZCols]> for Matrix<T, $NRows, $NCols, S>
-        where S: ContiguousStorageMut<T, $NRows, $NCols> {
+        impl<T: Scalar, S> $Mut<[[T; $SZRows]; $SZCols]> for Matrix<T, $NRows, $NCols, S>
+        where S: RawStorageMut<T, $NRows, $NCols> + IsContiguous {
             #[inline]
             fn $mut(&mut self) -> &mut [[T; $SZRows]; $SZCols] {
+                // Safety: OK thanks to the IsContiguous trait.
                 unsafe {
                     &mut *(self.data.ptr_mut() as *mut [[T; $SZRows]; $SZCols])
                 }
@@ -244,9 +252,13 @@ impl_from_into_asref_borrow_2D!(
     (U6, U2) => (6, 2); (U6, U3) => (6, 3); (U6, U4) => (6, 4); (U6, U5) => (6, 5); (U6, U6) => (6, 6);
 );
 
-impl<'a, T: Clone, RStride: Dim, CStride: Dim, const R: usize, const C: usize>
+impl<'a, T, RStride, CStride, const R: usize, const C: usize>
     From<MatrixSlice<'a, T, Const<R>, Const<C>, RStride, CStride>>
     for Matrix<T, Const<R>, Const<C>, ArrayStorage<T, R, C>>
+where
+    T: Scalar,
+    RStride: Dim,
+    CStride: Dim,
 {
     fn from(matrix_slice: MatrixSlice<'a, T, Const<R>, Const<C>, RStride, CStride>) -> Self {
         matrix_slice.into_owned()
@@ -254,9 +266,13 @@ impl<'a, T: Clone, RStride: Dim, CStride: Dim, const R: usize, const C: usize>
 }
 
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<'a, T: Clone, C: Dim, RStride: Dim, CStride: Dim>
-    From<MatrixSlice<'a, T, Dynamic, C, RStride, CStride>>
+impl<'a, T, C, RStride, CStride> From<MatrixSlice<'a, T, Dynamic, C, RStride, CStride>>
     for Matrix<T, Dynamic, C, VecStorage<T, Dynamic, C>>
+where
+    T: Scalar,
+    C: Dim,
+    RStride: Dim,
+    CStride: Dim,
 {
     fn from(matrix_slice: MatrixSlice<'a, T, Dynamic, C, RStride, CStride>) -> Self {
         matrix_slice.into_owned()
@@ -264,18 +280,26 @@ impl<'a, T: Clone, C: Dim, RStride: Dim, CStride: Dim>
 }
 
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<'a, T: Clone, R: DimName, RStride: Dim, CStride: Dim>
-    From<MatrixSlice<'a, T, R, Dynamic, RStride, CStride>>
+impl<'a, T, R, RStride, CStride> From<MatrixSlice<'a, T, R, Dynamic, RStride, CStride>>
     for Matrix<T, R, Dynamic, VecStorage<T, R, Dynamic>>
+where
+    T: Scalar,
+    R: DimName,
+    RStride: Dim,
+    CStride: Dim,
 {
     fn from(matrix_slice: MatrixSlice<'a, T, R, Dynamic, RStride, CStride>) -> Self {
         matrix_slice.into_owned()
     }
 }
 
-impl<'a, T: Clone, RStride: Dim, CStride: Dim, const R: usize, const C: usize>
+impl<'a, T, RStride, CStride, const R: usize, const C: usize>
     From<MatrixSliceMut<'a, T, Const<R>, Const<C>, RStride, CStride>>
     for Matrix<T, Const<R>, Const<C>, ArrayStorage<T, R, C>>
+where
+    T: Scalar,
+    RStride: Dim,
+    CStride: Dim,
 {
     fn from(matrix_slice: MatrixSliceMut<'a, T, Const<R>, Const<C>, RStride, CStride>) -> Self {
         matrix_slice.into_owned()
@@ -283,9 +307,13 @@ impl<'a, T: Clone, RStride: Dim, CStride: Dim, const R: usize, const C: usize>
 }
 
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<'a, T: Clone, C: Dim, RStride: Dim, CStride: Dim>
-    From<MatrixSliceMut<'a, T, Dynamic, C, RStride, CStride>>
+impl<'a, T, C, RStride, CStride> From<MatrixSliceMut<'a, T, Dynamic, C, RStride, CStride>>
     for Matrix<T, Dynamic, C, VecStorage<T, Dynamic, C>>
+where
+    T: Scalar,
+    C: Dim,
+    RStride: Dim,
+    CStride: Dim,
 {
     fn from(matrix_slice: MatrixSliceMut<'a, T, Dynamic, C, RStride, CStride>) -> Self {
         matrix_slice.into_owned()
@@ -293,26 +321,37 @@ impl<'a, T: Clone, C: Dim, RStride: Dim, CStride: Dim>
 }
 
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<'a, T: Clone, R: DimName, RStride: Dim, CStride: Dim>
-    From<MatrixSliceMut<'a, T, R, Dynamic, RStride, CStride>>
+impl<'a, T, R, RStride, CStride> From<MatrixSliceMut<'a, T, R, Dynamic, RStride, CStride>>
     for Matrix<T, R, Dynamic, VecStorage<T, R, Dynamic>>
+where
+    T: Scalar,
+    R: DimName,
+    RStride: Dim,
+    CStride: Dim,
 {
     fn from(matrix_slice: MatrixSliceMut<'a, T, R, Dynamic, RStride, CStride>) -> Self {
         matrix_slice.into_owned()
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, RSlice: Dim, CSlice: Dim, RStride: Dim, CStride: Dim, S>
-    From<&'a Matrix<T, R, C, S>> for MatrixSlice<'a, T, RSlice, CSlice, RStride, CStride>
+impl<'a, T, R, C, RSlice, CSlice, RStride, CStride, S> From<&'a Matrix<T, R, C, S>>
+    for MatrixSlice<'a, T, RSlice, CSlice, RStride, CStride>
 where
-    S: Storage<T, R, C>,
+    T: Scalar,
+    R: Dim,
+    C: Dim,
+    RSlice: Dim,
+    CSlice: Dim,
+    RStride: Dim,
+    CStride: Dim,
+    S: RawStorage<T, R, C>,
     ShapeConstraint: DimEq<R, RSlice>
         + DimEq<C, CSlice>
         + DimEq<RStride, S::RStride>
         + DimEq<CStride, S::CStride>,
 {
     fn from(m: &'a Matrix<T, R, C, S>) -> Self {
-        let (row, col) = m.data.shape();
+        let (row, col) = m.shape_generic();
         let row_slice = RSlice::from_usize(row.value());
         let col_slice = CSlice::from_usize(col.value());
 
@@ -327,23 +366,29 @@ where
                 (row_slice, col_slice),
                 (rstride_slice, cstride_slice),
             );
-
-            Self::from_data(data)
+            Matrix::from_data_statically_unchecked(data)
         }
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, RSlice: Dim, CSlice: Dim, RStride: Dim, CStride: Dim, S>
-    From<&'a mut Matrix<T, R, C, S>> for MatrixSlice<'a, T, RSlice, CSlice, RStride, CStride>
+impl<'a, T, R, C, RSlice, CSlice, RStride, CStride, S> From<&'a mut Matrix<T, R, C, S>>
+    for MatrixSlice<'a, T, RSlice, CSlice, RStride, CStride>
 where
-    S: Storage<T, R, C>,
+    T: Scalar,
+    R: Dim,
+    C: Dim,
+    RSlice: Dim,
+    CSlice: Dim,
+    RStride: Dim,
+    CStride: Dim,
+    S: RawStorage<T, R, C>,
     ShapeConstraint: DimEq<R, RSlice>
         + DimEq<C, CSlice>
         + DimEq<RStride, S::RStride>
         + DimEq<CStride, S::CStride>,
 {
     fn from(m: &'a mut Matrix<T, R, C, S>) -> Self {
-        let (row, col) = m.data.shape();
+        let (row, col) = m.shape_generic();
         let row_slice = RSlice::from_usize(row.value());
         let col_slice = CSlice::from_usize(col.value());
 
@@ -358,23 +403,29 @@ where
                 (row_slice, col_slice),
                 (rstride_slice, cstride_slice),
             );
-
-            Matrix::from_data(data)
+            Matrix::from_data_statically_unchecked(data)
         }
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, RSlice: Dim, CSlice: Dim, RStride: Dim, CStride: Dim, S>
-    From<&'a mut Matrix<T, R, C, S>> for MatrixSliceMut<'a, T, RSlice, CSlice, RStride, CStride>
+impl<'a, T, R, C, RSlice, CSlice, RStride, CStride, S> From<&'a mut Matrix<T, R, C, S>>
+    for MatrixSliceMut<'a, T, RSlice, CSlice, RStride, CStride>
 where
-    S: StorageMut<T, R, C>,
+    T: Scalar,
+    R: Dim,
+    C: Dim,
+    RSlice: Dim,
+    CSlice: Dim,
+    RStride: Dim,
+    CStride: Dim,
+    S: RawStorageMut<T, R, C>,
     ShapeConstraint: DimEq<R, RSlice>
         + DimEq<C, CSlice>
         + DimEq<RStride, S::RStride>
         + DimEq<CStride, S::CStride>,
 {
     fn from(m: &'a mut Matrix<T, R, C, S>) -> Self {
-        let (row, col) = m.data.shape();
+        let (row, col) = m.shape_generic();
         let row_slice = RSlice::from_usize(row.value());
         let col_slice = CSlice::from_usize(col.value());
 
@@ -389,22 +440,21 @@ where
                 (row_slice, col_slice),
                 (rstride_slice, cstride_slice),
             );
-
-            Matrix::from_data(data)
+            Matrix::from_data_statically_unchecked(data)
         }
     }
 }
 
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<'a, T> From<Vec<T>> for DVector<T> {
+impl<'a, T: Scalar> From<Vec<T>> for DVector<T> {
     #[inline]
     fn from(vec: Vec<T>) -> Self {
         Self::from_vec(vec)
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S: ContiguousStorage<T, R, C>> From<&'a Matrix<T, R, C, S>>
-    for &'a [T]
+impl<'a, T: Scalar + Copy, R: Dim, C: Dim, S: RawStorage<T, R, C> + IsContiguous>
+    From<&'a Matrix<T, R, C, S>> for &'a [T]
 {
     #[inline]
     fn from(matrix: &'a Matrix<T, R, C, S>) -> Self {
@@ -412,8 +462,8 @@ impl<'a, T, R: Dim, C: Dim, S: ContiguousStorage<T, R, C>> From<&'a Matrix<T, R,
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S: ContiguousStorageMut<T, R, C>> From<&'a mut Matrix<T, R, C, S>>
-    for &'a mut [T]
+impl<'a, T: Scalar + Copy, R: Dim, C: Dim, S: RawStorageMut<T, R, C> + IsContiguous>
+    From<&'a mut Matrix<T, R, C, S>> for &'a mut [T]
 {
     #[inline]
     fn from(matrix: &'a mut Matrix<T, R, C, S>) -> Self {
@@ -421,27 +471,27 @@ impl<'a, T, R: Dim, C: Dim, S: ContiguousStorageMut<T, R, C>> From<&'a mut Matri
     }
 }
 
-impl<'a, T> From<&'a [T]> for DVectorSlice<'a, T> {
+impl<'a, T: Scalar + Copy> From<&'a [T]> for DVectorSlice<'a, T> {
     #[inline]
     fn from(slice: &'a [T]) -> Self {
         Self::from_slice(slice, slice.len())
     }
 }
 
-impl<'a, T> From<DVectorSlice<'a, T>> for &'a [T] {
+impl<'a, T: Scalar> From<DVectorSlice<'a, T>> for &'a [T] {
     fn from(vec: DVectorSlice<'a, T>) -> &'a [T] {
         vec.data.into_slice()
     }
 }
 
-impl<'a, T> From<&'a mut [T]> for DVectorSliceMut<'a, T> {
+impl<'a, T: Scalar + Copy> From<&'a mut [T]> for DVectorSliceMut<'a, T> {
     #[inline]
     fn from(slice: &'a mut [T]) -> Self {
         Self::from_slice(slice, slice.len())
     }
 }
 
-impl<'a, T> From<DVectorSliceMut<'a, T>> for &'a mut [T] {
+impl<'a, T: Scalar> From<DVectorSliceMut<'a, T>> for &'a mut [T] {
     fn from(vec: DVectorSliceMut<'a, T>) -> &'a mut [T] {
         vec.data.into_slice_mut()
     }
@@ -456,7 +506,7 @@ where
 {
     #[inline]
     fn from(arr: [OMatrix<T::Element, R, C>; 2]) -> Self {
-        let (nrows, ncols) = arr[0].data.shape();
+        let (nrows, ncols) = arr[0].shape_generic();
 
         Self::from_fn_generic(nrows, ncols, |i, j| {
             [
@@ -477,7 +527,7 @@ where
 {
     #[inline]
     fn from(arr: [OMatrix<T::Element, R, C>; 4]) -> Self {
-        let (nrows, ncols) = arr[0].data.shape();
+        let (nrows, ncols) = arr[0].shape_generic();
 
         Self::from_fn_generic(nrows, ncols, |i, j| {
             [
@@ -500,7 +550,7 @@ where
 {
     #[inline]
     fn from(arr: [OMatrix<T::Element, R, C>; 8]) -> Self {
-        let (nrows, ncols) = arr[0].data.shape();
+        let (nrows, ncols) = arr[0].shape_generic();
 
         Self::from_fn_generic(nrows, ncols, |i, j| {
             [
@@ -526,7 +576,7 @@ where
     DefaultAllocator: Allocator<T, R, C> + Allocator<T::Element, R, C>,
 {
     fn from(arr: [OMatrix<T::Element, R, C>; 16]) -> Self {
-        let (nrows, ncols) = arr[0].data.shape();
+        let (nrows, ncols) = arr[0].shape_generic();
 
         Self::from_fn_generic(nrows, ncols, |i, j| {
             [
diff --git a/src/base/coordinates.rs b/src/base/coordinates.rs
index 6389ccbe..db66811d 100644
--- a/src/base/coordinates.rs
+++ b/src/base/coordinates.rs
@@ -7,8 +7,8 @@
 use std::ops::{Deref, DerefMut};
 
 use crate::base::dimension::{U1, U2, U3, U4, U5, U6};
-use crate::base::storage::{ContiguousStorage, ContiguousStorageMut};
-use crate::base::Matrix;
+use crate::base::storage::{IsContiguous, RawStorage, RawStorageMut};
+use crate::base::{Matrix, Scalar};
 
 /*
  *
@@ -23,7 +23,7 @@ macro_rules! coords_impl(
         #[repr(C)]
         #[derive(Eq, PartialEq, Clone, Hash, Debug, Copy)]
         #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
-        pub struct $T<T> {
+        pub struct $T<T: Scalar> {
             $(pub $comps: T),*
         }
     }
@@ -31,20 +31,22 @@ macro_rules! coords_impl(
 
 macro_rules! deref_impl(
     ($R: ty, $C: ty; $Target: ident) => {
-        impl<T, S> Deref for Matrix<T, $R, $C, S>
-            where S: ContiguousStorage<T, $R, $C> {
+        impl<T: Scalar, S> Deref for Matrix<T, $R, $C, S>
+            where S: RawStorage<T, $R, $C> + IsContiguous {
             type Target = $Target<T>;
 
             #[inline]
             fn deref(&self) -> &Self::Target {
+                // Safety: this is OK because of the IsContiguous trait.
                 unsafe { &*(self.data.ptr() as *const Self::Target) }
             }
         }
 
-        impl<T, S> DerefMut for Matrix<T, $R, $C, S>
-            where S: ContiguousStorageMut<T, $R, $C> {
+        impl<T: Scalar, S> DerefMut for Matrix<T, $R, $C, S>
+            where S: RawStorageMut<T, $R, $C> + IsContiguous {
             #[inline]
             fn deref_mut(&mut self) -> &mut Self::Target {
+                // Safety: this is OK because of the IsContiguous trait.
                 unsafe { &mut *(self.data.ptr_mut() as *mut Self::Target) }
             }
         }
diff --git a/src/base/default_allocator.rs b/src/base/default_allocator.rs
index 9face98c..2f996008 100644
--- a/src/base/default_allocator.rs
+++ b/src/base/default_allocator.rs
@@ -4,72 +4,50 @@
 //! heap-allocated buffers for matrices with at least one dimension unknown at compile-time.
 
 use std::cmp;
-use std::fmt;
-use std::mem::{self, ManuallyDrop, MaybeUninit};
+use std::mem;
 use std::ptr;
 
 #[cfg(all(feature = "alloc", not(feature = "std")))]
 use alloc::vec::Vec;
 
+use super::Const;
+use crate::base::allocator::{Allocator, Reallocator};
+use crate::base::array_storage::ArrayStorage;
 #[cfg(any(feature = "alloc", feature = "std"))]
 use crate::base::dimension::Dynamic;
-
-use super::Const;
-use crate::base::allocator::{Allocator, InnerAllocator, Reallocator};
-use crate::base::array_storage::ArrayStorage;
 use crate::base::dimension::{Dim, DimName};
-use crate::base::storage::{
-    ContiguousStorage, ContiguousStorageMut, InnerOwned, Storage, StorageMut,
-};
+use crate::base::storage::{RawStorage, RawStorageMut};
+#[cfg(any(feature = "std", feature = "alloc"))]
 use crate::base::vec_storage::VecStorage;
-use crate::U1;
+use crate::base::Scalar;
+use std::mem::{ManuallyDrop, MaybeUninit};
 
 /*
  *
  * Allocator.
  *
  */
-/// A helper struct that controls how the storage for a matrix should be allocated.
-///
-/// This struct is useless on its own. Instead, it's used in trait
 /// An allocator based on `GenericArray` and `VecStorage` for statically-sized and dynamically-sized
 /// matrices respectively.
 #[derive(Copy, Clone, Debug)]
 pub struct DefaultAllocator;
 
 // Static - Static
-impl<T, const R: usize, const C: usize> InnerAllocator<T, Const<R>, Const<C>> for DefaultAllocator {
+impl<T: Scalar, const R: usize, const C: usize> Allocator<T, Const<R>, Const<C>>
+    for DefaultAllocator
+{
     type Buffer = ArrayStorage<T, R, C>;
+    type BufferUninit = ArrayStorage<MaybeUninit<T>, R, C>;
 
     #[inline]
-    fn allocate_from_iterator<I: IntoIterator<Item = T>>(
-        nrows: Const<R>,
-        ncols: Const<C>,
-        iter: I,
-    ) -> Self::Buffer {
-        let mut res = Self::allocate_uninitialized(nrows, ncols);
-        let mut count = 0;
-
-        for (res, e) in res.as_mut_slice().iter_mut().zip(iter.into_iter()) {
-            *res = MaybeUninit::new(e);
-            count += 1;
-        }
-
-        assert!(
-            count == nrows.value() * ncols.value(),
-            "Matrix init. from iterator: iterator not long enough."
-        );
-
-        // Safety: we have initialized all entries.
-        unsafe { <Self as Allocator<T, Const<R>, Const<C>>>::assume_init(res) }
+    unsafe fn allocate_uninitialized(_: Const<R>, _: Const<C>) -> MaybeUninit<Self::Buffer> {
+        mem::MaybeUninit::<Self::Buffer>::uninit()
     }
-}
 
-impl<T, const R: usize, const C: usize> Allocator<T, Const<R>, Const<C>> for DefaultAllocator {
     #[inline]
-    fn allocate_uninitialized(_: Const<R>, _: Const<C>) -> ArrayStorage<MaybeUninit<T>, R, C> {
+    fn allocate_uninit(_: Const<R>, _: Const<C>) -> ArrayStorage<MaybeUninit<T>, R, C> {
         // SAFETY: An uninitialized `[MaybeUninit<_>; _]` is valid.
-        let array = unsafe { MaybeUninit::uninit().assume_init() };
+        let array: [[MaybeUninit<T>; R]; C] = unsafe { MaybeUninit::uninit().assume_init() };
         ArrayStorage(array)
     }
 
@@ -83,41 +61,53 @@ impl<T, const R: usize, const C: usize> Allocator<T, Const<R>, Const<C>> for Def
         ArrayStorage((&uninit as *const _ as *const [_; C]).read())
     }
 
-    /// Specifies that a given buffer's entries should be manually dropped.
     #[inline]
-    fn manually_drop(buf: ArrayStorage<T, R, C>) -> ArrayStorage<ManuallyDrop<T>, R, C> {
-        // SAFETY:
-        // * `ManuallyDrop<T>` and T are guaranteed to have the same layout
-        // * `ManuallyDrop` does not drop, so there are no double-frees
-        // And thus the conversion is safe
-        unsafe { ArrayStorage((&ManuallyDrop::new(buf) as *const _ as *const [_; C]).read()) }
+    fn allocate_from_iterator<I: IntoIterator<Item = T>>(
+        nrows: Const<R>,
+        ncols: Const<C>,
+        iter: I,
+    ) -> Self::Buffer {
+        #[cfg(feature = "no_unsound_assume_init")]
+        let mut res: Self::Buffer = unimplemented!();
+        #[cfg(not(feature = "no_unsound_assume_init"))]
+        let mut res = unsafe { Self::allocate_uninitialized(nrows, ncols).assume_init() };
+        let mut count = 0;
+
+        // Safety: this is OK because the Buffer is known to be contiguous.
+        let res_slice = unsafe { res.as_mut_slice_unchecked() };
+        for (res, e) in res_slice.iter_mut().zip(iter.into_iter()) {
+            *res = e;
+            count += 1;
+        }
+
+        assert!(
+            count == nrows.value() * ncols.value(),
+            "Matrix init. from iterator: iterator not long enough."
+        );
+
+        res
     }
 }
 
 // Dynamic - Static
 // Dynamic - Dynamic
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<T, C: Dim> InnerAllocator<T, Dynamic, C> for DefaultAllocator {
+impl<T: Scalar, C: Dim> Allocator<T, Dynamic, C> for DefaultAllocator {
     type Buffer = VecStorage<T, Dynamic, C>;
+    type BufferUninit = VecStorage<MaybeUninit<T>, Dynamic, C>;
 
     #[inline]
-    fn allocate_from_iterator<I: IntoIterator<Item = T>>(
-        nrows: Dynamic,
-        ncols: C,
-        iter: I,
-    ) -> Self::Buffer {
-        let it = iter.into_iter();
-        let res: Vec<T> = it.collect();
-        assert!(res.len() == nrows.value() * ncols.value(),
-                "Allocation from iterator error: the iterator did not yield the correct number of elements.");
+    unsafe fn allocate_uninitialized(nrows: Dynamic, ncols: C) -> MaybeUninit<Self::Buffer> {
+        let mut res = Vec::new();
+        let length = nrows.value() * ncols.value();
+        res.reserve_exact(length);
+        res.set_len(length);
 
-        VecStorage::new(nrows, ncols, res)
+        mem::MaybeUninit::new(VecStorage::new(nrows, ncols, res))
     }
-}
 
-impl<T, C: Dim> Allocator<T, Dynamic, C> for DefaultAllocator {
     #[inline]
-    fn allocate_uninitialized(nrows: Dynamic, ncols: C) -> VecStorage<MaybeUninit<T>, Dynamic, C> {
+    fn allocate_uninit(nrows: Dynamic, ncols: C) -> VecStorage<MaybeUninit<T>, Dynamic, C> {
         let mut data = Vec::new();
         let length = nrows.value() * ncols.value();
         data.reserve_exact(length);
@@ -143,32 +133,10 @@ impl<T, C: Dim> Allocator<T, Dynamic, C> for DefaultAllocator {
         VecStorage::new(nrows, ncols, new_data)
     }
 
-    #[inline]
-    fn manually_drop(buf: VecStorage<T, Dynamic, C>) -> VecStorage<ManuallyDrop<T>, Dynamic, C> {
-        // Avoids a double-drop.
-        let (nrows, ncols) = buf.shape();
-        let vec: Vec<_> = buf.into();
-        let mut md = ManuallyDrop::new(vec);
-
-        // Safety:
-        // - ManuallyDrop<T> has the same alignment and layout as T.
-        // - The length and capacity come from a valid vector.
-        let new_data =
-            unsafe { Vec::from_raw_parts(md.as_mut_ptr() as *mut _, md.len(), md.capacity()) };
-
-        VecStorage::new(nrows, ncols, new_data)
-    }
-}
-
-// Static - Dynamic
-#[cfg(any(feature = "std", feature = "alloc"))]
-impl<T, R: DimName> InnerAllocator<T, R, Dynamic> for DefaultAllocator {
-    type Buffer = VecStorage<T, R, Dynamic>;
-
     #[inline]
     fn allocate_from_iterator<I: IntoIterator<Item = T>>(
-        nrows: R,
-        ncols: Dynamic,
+        nrows: Dynamic,
+        ncols: C,
         iter: I,
     ) -> Self::Buffer {
         let it = iter.into_iter();
@@ -180,9 +148,24 @@ impl<T, R: DimName> InnerAllocator<T, R, Dynamic> for DefaultAllocator {
     }
 }
 
-impl<T, R: DimName> Allocator<T, R, Dynamic> for DefaultAllocator {
+// Static - Dynamic
+#[cfg(any(feature = "std", feature = "alloc"))]
+impl<T: Scalar, R: DimName> Allocator<T, R, Dynamic> for DefaultAllocator {
+    type Buffer = VecStorage<T, R, Dynamic>;
+    type BufferUninit = VecStorage<MaybeUninit<T>, R, Dynamic>;
+
     #[inline]
-    fn allocate_uninitialized(nrows: R, ncols: Dynamic) -> VecStorage<MaybeUninit<T>, R, Dynamic> {
+    unsafe fn allocate_uninitialized(nrows: R, ncols: Dynamic) -> MaybeUninit<Self::Buffer> {
+        let mut res = Vec::new();
+        let length = nrows.value() * ncols.value();
+        res.reserve_exact(length);
+        res.set_len(length);
+
+        mem::MaybeUninit::new(VecStorage::new(nrows, ncols, res))
+    }
+
+    #[inline]
+    fn allocate_uninit(nrows: R, ncols: Dynamic) -> VecStorage<MaybeUninit<T>, R, Dynamic> {
         let mut data = Vec::new();
         let length = nrows.value() * ncols.value();
         data.reserve_exact(length);
@@ -209,253 +192,59 @@ impl<T, R: DimName> Allocator<T, R, Dynamic> for DefaultAllocator {
     }
 
     #[inline]
-    fn manually_drop(buf: VecStorage<T, R, Dynamic>) -> VecStorage<ManuallyDrop<T>, R, Dynamic> {
-        // Avoids a double-drop.
-        let (nrows, ncols) = buf.shape();
-        let vec: Vec<_> = buf.into();
-        let mut md = ManuallyDrop::new(vec);
+    fn allocate_from_iterator<I: IntoIterator<Item = T>>(
+        nrows: R,
+        ncols: Dynamic,
+        iter: I,
+    ) -> Self::Buffer {
+        let it = iter.into_iter();
+        let res: Vec<T> = it.collect();
+        assert!(res.len() == nrows.value() * ncols.value(),
+                "Allocation from iterator error: the iterator did not yield the correct number of elements.");
 
-        // Safety:
-        // - ManuallyDrop<T> has the same alignment and layout as T.
-        // - The length and capacity come from a valid vector.
-        let new_data =
-            unsafe { Vec::from_raw_parts(md.as_mut_ptr() as *mut _, md.len(), md.capacity()) };
-
-        VecStorage::new(nrows, ncols, new_data)
+        VecStorage::new(nrows, ncols, res)
     }
 }
 
-/// The owned storage type for a matrix.
-#[repr(transparent)]
-pub struct Owned<T, R: Dim, C: Dim>(pub InnerOwned<T, R, C>)
-where
-    DefaultAllocator: InnerAllocator<T, R, C>;
-
-impl<T: Copy, R: Dim, C: Dim> Copy for Owned<T, R, C>
-where
-    DefaultAllocator: InnerAllocator<T, R, C>,
-    InnerOwned<T, R, C>: Copy,
-{
-}
-
-impl<T: Clone, R: Dim, C: Dim> Clone for Owned<T, R, C>
-where
-    DefaultAllocator: InnerAllocator<T, R, C>,
-{
-    fn clone(&self) -> Self {
-        if Self::is_array() {
-            // We first clone the data.
-            let slice = unsafe { self.as_slice_unchecked() };
-            let vec = ManuallyDrop::new(slice.to_owned());
-
-            // We then transmute it back into an array and then an Owned.
-            unsafe { mem::transmute_copy(&*vec.as_ptr()) }
-        } else {
-            // We first clone the data.
-            let clone = ManuallyDrop::new(self.as_vec_storage().clone());
-
-            // We then transmute it back into an Owned.
-            unsafe { mem::transmute_copy(&clone) }
-        }
-
-        // TODO: check that the auxiliary copies are elided.
-    }
-}
-
-impl<T: fmt::Debug, R: Dim, C: Dim> fmt::Debug for Owned<T, R, C>
-where
-    DefaultAllocator: InnerAllocator<T, R, C>,
-{
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        if Self::is_array() {
-            let slice = unsafe { self.as_slice_unchecked() };
-            slice.fmt(f)
-        } else {
-            self.as_vec_storage().fmt(f)
-        }
-    }
-}
-
-impl<T, const R: usize, const C: usize> Owned<T, Const<R>, Const<C>> {
-    fn new(array: [[T; R]; C]) -> Self {
-        Self(ArrayStorage(array))
-    }
-}
-
-impl<T, R: Dim, C: Dim> Owned<T, R, C>
-where
-    DefaultAllocator: InnerAllocator<T, R, C>,
-{
-    /// Returns whether `Self` stores an [`ArrayStorage`]. This is a zero-cost
-    /// operation.
-    const fn is_array() -> bool {
-        R::is_static() && C::is_static()
-    }
-
-    /// Returns whether `Self` stores a [`VecStorage`].
-    const fn is_vec() -> bool {
-        !Self::is_array()
-    }
-
-    /// Returns a reference to the underlying [`VecStorage`].
-    ///
-    /// # Panics
-    /// This method will panic if `Self` does not contain a [`VecStorage`].
-    fn as_vec_storage(&self) -> &VecStorage<T, R, C> {
-        assert!(Self::is_vec());
-
-        // Safety: `self` is transparent and must contain a `VecStorage`.
-        unsafe { &*(self as *const _ as *const _) }
-    }
-
-    /// Returns a mutable reference to the underlying [`VecStorage`].
-    ///
-    /// # Panics
-    /// This method will panic if `Self` does not contain a [`VecStorage`].
-    fn as_vec_storage_mut(&mut self) -> &mut VecStorage<T, R, C> {
-        assert!(Self::is_vec());
-
-        // Safety: `self` is transparent and must contain a `VecStorage`.
-        unsafe { &mut *(self as *mut _ as *mut _) }
-    }
-}
-
-unsafe impl<T, R: Dim, C: Dim> Storage<T, R, C> for Owned<T, R, C>
-where
-    DefaultAllocator: InnerAllocator<T, R, C>,
-{
-    type RStride = U1;
-
-    type CStride = R;
-
-    fn ptr(&self) -> *const T {
-        if Self::is_array() {
-            &self as *const _ as *const T
-        } else {
-            self.as_vec_storage().as_vec().as_ptr()
-        }
-    }
-
-    fn shape(&self) -> (R, C) {
-        if Self::is_array() {
-            (R::default(), C::default())
-        } else {
-            let vec = self.as_vec_storage();
-            (vec.nrows, vec.ncols)
-        }
-    }
-
-    fn strides(&self) -> (Self::RStride, Self::CStride) {
-        if Self::is_array() {
-            (U1::name(), R::default())
-        } else {
-            let vec = self.as_vec_storage();
-            (U1::name(), vec.nrows)
-        }
-    }
-
-    #[inline(always)]
-    fn is_contiguous(&self) -> bool {
-        true
-    }
-
-    unsafe fn as_slice_unchecked(&self) -> &[T] {
-        if Self::is_array() {
-            std::slice::from_raw_parts(
-                self.ptr(),
-                R::try_to_usize().unwrap() * C::try_to_usize().unwrap(),
-            )
-        } else {
-            self.as_vec_storage().as_vec().as_ref()
-        }
-    }
-
-    #[inline(always)]
-    fn into_owned(self) -> Self {
-        self
-    }
-
-    #[inline(always)]
-    fn clone_owned(&self) -> Self
-    where
-        T: Clone,
-    {
-        self.clone()
-    }
-}
-
-unsafe impl<T, R: Dim, C: Dim> StorageMut<T, R, C> for Owned<T, R, C>
-where
-    DefaultAllocator: InnerAllocator<T, R, C>,
-{
-    fn ptr_mut(&mut self) -> *mut T {
-        if Self::is_array() {
-            &mut self as *mut _ as *mut T
-        } else {
-            self.as_vec_storage_mut().as_vec().as_ptr()
-        }
-    }
-
-    unsafe fn as_mut_slice_unchecked(&mut self) -> &mut [T] {
-        if Self::is_array() {
-            std::slice::from_raw_parts(
-                self.ptr_mut(),
-                R::try_to_usize().unwrap() * C::try_to_usize().unwrap(),
-            )
-        } else {
-            self.as_vec_storage_mut().as_vec_mut().as_mut()
-        }
-    }
-}
-
-unsafe impl<T, R: Dim, C: Dim> ContiguousStorage<T, R, C> for Owned<T, R, C> where
-    DefaultAllocator: InnerAllocator<T, R, C>
-{
-}
-
-unsafe impl<T, R: Dim, C: Dim> ContiguousStorageMut<T, R, C> for Owned<T, R, C> where
-    DefaultAllocator: InnerAllocator<T, R, C>
-{
-}
-
 /*
  *
  * Reallocator.
  *
  */
 // Anything -> Static × Static
-impl<T, RFrom: Dim, CFrom: Dim, const RTO: usize, const CTO: usize>
+impl<T: Scalar, RFrom, CFrom, const RTO: usize, const CTO: usize>
     Reallocator<T, RFrom, CFrom, Const<RTO>, Const<CTO>> for DefaultAllocator
 where
+    RFrom: Dim,
+    CFrom: Dim,
     Self: Allocator<T, RFrom, CFrom>,
 {
     #[inline]
     unsafe fn reallocate_copy(
         rto: Const<RTO>,
         cto: Const<CTO>,
-        buf: InnerOwned<T, RFrom, CFrom>,
+        buf: <Self as Allocator<T, RFrom, CFrom>>::Buffer,
     ) -> ArrayStorage<T, RTO, CTO> {
+        #[cfg(feature = "no_unsound_assume_init")]
+        let mut res: ArrayStorage<T, RTO, CTO> = unimplemented!();
+        #[cfg(not(feature = "no_unsound_assume_init"))]
         let mut res =
-            <Self as Allocator<_, Const<RTO>, Const<CTO>>>::allocate_uninitialized(rto, cto);
+            <Self as Allocator<T, Const<RTO>, Const<CTO>>>::allocate_uninitialized(rto, cto)
+                .assume_init();
 
         let (rfrom, cfrom) = buf.shape();
 
         let len_from = rfrom.value() * cfrom.value();
         let len_to = rto.value() * cto.value();
-        ptr::copy_nonoverlapping(
-            buf.ptr(),
-            res.ptr_mut() as *mut T,
-            cmp::min(len_from, len_to),
-        );
+        ptr::copy_nonoverlapping(buf.ptr(), res.ptr_mut(), cmp::min(len_from, len_to));
 
-        // Safety: TODO
-        <Self as Allocator<_, Const<RTO>, Const<CTO>>>::assume_init(res)
+        res
     }
 }
 
 // Static × Static -> Dynamic × Any
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<T, CTo, const RFROM: usize, const CFROM: usize>
+impl<T: Scalar, CTo, const RFROM: usize, const CFROM: usize>
     Reallocator<T, Const<RFROM>, Const<CFROM>, Dynamic, CTo> for DefaultAllocator
 where
     CTo: Dim,
@@ -466,25 +255,25 @@ where
         cto: CTo,
         buf: ArrayStorage<T, RFROM, CFROM>,
     ) -> VecStorage<T, Dynamic, CTo> {
-        let mut res = <Self as Allocator<T, Dynamic, CTo>>::allocate_uninitialized(rto, cto);
+        #[cfg(feature = "no_unsound_assume_init")]
+        let mut res: VecStorage<T, Dynamic, CTo> = unimplemented!();
+        #[cfg(not(feature = "no_unsound_assume_init"))]
+        let mut res =
+            <Self as Allocator<T, Dynamic, CTo>>::allocate_uninitialized(rto, cto).assume_init();
 
         let (rfrom, cfrom) = buf.shape();
 
         let len_from = rfrom.value() * cfrom.value();
         let len_to = rto.value() * cto.value();
-        ptr::copy_nonoverlapping(
-            buf.ptr(),
-            res.ptr_mut() as *mut T,
-            cmp::min(len_from, len_to),
-        );
+        ptr::copy_nonoverlapping(buf.ptr(), res.ptr_mut(), cmp::min(len_from, len_to));
 
-        <Self as Allocator<T, Dynamic, CTo>>::assume_init(res)
+        res
     }
 }
 
 // Static × Static -> Static × Dynamic
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<T, RTo, const RFROM: usize, const CFROM: usize>
+impl<T: Scalar, RTo, const RFROM: usize, const CFROM: usize>
     Reallocator<T, Const<RFROM>, Const<CFROM>, RTo, Dynamic> for DefaultAllocator
 where
     RTo: DimName,
@@ -495,25 +284,27 @@ where
         cto: Dynamic,
         buf: ArrayStorage<T, RFROM, CFROM>,
     ) -> VecStorage<T, RTo, Dynamic> {
-        let mut res = <Self as Allocator<T, RTo, Dynamic>>::allocate_uninitialized(rto, cto);
+        #[cfg(feature = "no_unsound_assume_init")]
+        let mut res: VecStorage<T, RTo, Dynamic> = unimplemented!();
+        #[cfg(not(feature = "no_unsound_assume_init"))]
+        let mut res =
+            <Self as Allocator<T, RTo, Dynamic>>::allocate_uninitialized(rto, cto).assume_init();
 
         let (rfrom, cfrom) = buf.shape();
 
         let len_from = rfrom.value() * cfrom.value();
         let len_to = rto.value() * cto.value();
-        ptr::copy_nonoverlapping(
-            buf.ptr(),
-            res.ptr_mut() as *mut T,
-            cmp::min(len_from, len_to),
-        );
+        ptr::copy_nonoverlapping(buf.ptr(), res.ptr_mut(), cmp::min(len_from, len_to));
 
-        <Self as Allocator<T, RTo, Dynamic>>::assume_init(res)
+        res
     }
 }
 
 // All conversion from a dynamic buffer to a dynamic buffer.
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<T, CFrom: Dim, CTo: Dim> Reallocator<T, Dynamic, CFrom, Dynamic, CTo> for DefaultAllocator {
+impl<T: Scalar, CFrom: Dim, CTo: Dim> Reallocator<T, Dynamic, CFrom, Dynamic, CTo>
+    for DefaultAllocator
+{
     #[inline]
     unsafe fn reallocate_copy(
         rto: Dynamic,
@@ -526,7 +317,7 @@ impl<T, CFrom: Dim, CTo: Dim> Reallocator<T, Dynamic, CFrom, Dynamic, CTo> for D
 }
 
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<T, CFrom: Dim, RTo: DimName> Reallocator<T, Dynamic, CFrom, RTo, Dynamic>
+impl<T: Scalar, CFrom: Dim, RTo: DimName> Reallocator<T, Dynamic, CFrom, RTo, Dynamic>
     for DefaultAllocator
 {
     #[inline]
@@ -541,7 +332,7 @@ impl<T, CFrom: Dim, RTo: DimName> Reallocator<T, Dynamic, CFrom, RTo, Dynamic>
 }
 
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<T, RFrom: DimName, CTo: Dim> Reallocator<T, RFrom, Dynamic, Dynamic, CTo>
+impl<T: Scalar, RFrom: DimName, CTo: Dim> Reallocator<T, RFrom, Dynamic, Dynamic, CTo>
     for DefaultAllocator
 {
     #[inline]
@@ -556,7 +347,7 @@ impl<T, RFrom: DimName, CTo: Dim> Reallocator<T, RFrom, Dynamic, Dynamic, CTo>
 }
 
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<T, RFrom: DimName, RTo: DimName> Reallocator<T, RFrom, Dynamic, RTo, Dynamic>
+impl<T: Scalar, RFrom: DimName, RTo: DimName> Reallocator<T, RFrom, Dynamic, RTo, Dynamic>
     for DefaultAllocator
 {
     #[inline]
diff --git a/src/base/dimension.rs b/src/base/dimension.rs
index cfe66c87..8573dd59 100644
--- a/src/base/dimension.rs
+++ b/src/base/dimension.rs
@@ -2,7 +2,7 @@
 
 //! Traits and tags for identifying the dimension of all algebraic entities.
 
-use std::any::TypeId;
+use std::any::{Any, TypeId};
 use std::cmp;
 use std::fmt::Debug;
 use std::ops::{Add, Div, Mul, Sub};
@@ -11,8 +11,8 @@ use typenum::{self, Diff, Max, Maximum, Min, Minimum, Prod, Quot, Sum, Unsigned}
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
 
-/// Stores the dimension of dynamically-sized algebraic entities.
-#[derive(Clone, Copy, Default, Eq, PartialEq, Debug)]
+/// Dim of dynamically-sized algebraic entities.
+#[derive(Clone, Copy, Eq, PartialEq, Debug)]
 pub struct Dynamic {
     value: usize,
 }
@@ -55,7 +55,7 @@ impl IsNotStaticOne for Dynamic {}
 
 /// Trait implemented by any type that can be used as a dimension. This includes type-level
 /// integers and `Dynamic` (for dimensions not known at compile-time).
-pub trait Dim: 'static + Debug + Copy + Default + PartialEq + Send + Sync {
+pub trait Dim: Any + Debug + Copy + PartialEq + Send + Sync {
     #[inline(always)]
     fn is<D: Dim>() -> bool {
         TypeId::of::<Self>() == TypeId::of::<D>()
@@ -65,16 +65,6 @@ pub trait Dim: 'static + Debug + Copy + Default + PartialEq + Send + Sync {
     /// Dynamic`.
     fn try_to_usize() -> Option<usize>;
 
-    /// Returns whether `Self` has a known compile-time value.
-    fn is_static() -> bool {
-        Self::try_to_usize().is_some()
-    }
-
-    /// Returns whether `Self` does not have a known compile-time value.
-    fn is_dynamic() -> bool {
-        Self::try_to_usize().is_none()
-    }
-
     /// Gets the run-time value of `self`. For type-level integers, this is the same as
     /// `Self::try_to_usize().unwrap()`.
     fn value(&self) -> usize;
@@ -206,10 +196,7 @@ dim_ops!(
     DimMax, DimNameMax, Max, max, cmp::max, DimMaximum, DimNameMaximum, Maximum;
 );
 
-/// A wrapper around const types, which provides the capability of performing
-/// type-level arithmetic. This might get removed if const-generics become
-/// more powerful in the future.
-#[derive(Debug, Copy, Clone, Default, PartialEq, Eq, Hash)]
+#[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
 pub struct Const<const R: usize>;
 
 /// Trait implemented exclusively by type-level integers.
diff --git a/src/base/edition.rs b/src/base/edition.rs
index 94c13b09..0cad0d29 100644
--- a/src/base/edition.rs
+++ b/src/base/edition.rs
@@ -2,9 +2,6 @@ use num::{One, Zero};
 use std::cmp;
 #[cfg(any(feature = "std", feature = "alloc"))]
 use std::iter::ExactSizeIterator;
-#[cfg(any(feature = "std", feature = "alloc"))]
-use std::mem;
-use std::mem::MaybeUninit;
 use std::ptr;
 
 use crate::base::allocator::{Allocator, Reallocator};
@@ -12,8 +9,10 @@ use crate::base::constraint::{DimEq, SameNumberOfColumns, SameNumberOfRows, Shap
 #[cfg(any(feature = "std", feature = "alloc"))]
 use crate::base::dimension::Dynamic;
 use crate::base::dimension::{Const, Dim, DimAdd, DimDiff, DimMin, DimMinimum, DimSub, DimSum, U1};
-use crate::base::storage::{ContiguousStorageMut, ReshapableStorage, Storage, StorageMut};
+use crate::base::storage::{RawStorage, RawStorageMut, ReshapableStorage};
 use crate::base::{DefaultAllocator, Matrix, OMatrix, RowVector, Scalar, Vector};
+use crate::Storage;
+use std::mem::MaybeUninit;
 
 /// # Rows and columns extraction
 impl<T: Scalar + Zero, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
@@ -50,11 +49,11 @@ impl<T: Scalar + Zero, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     where
         I: IntoIterator<Item = &'a usize>,
         I::IntoIter: ExactSizeIterator + Clone,
+        DefaultAllocator: Allocator<T, Dynamic, C>,
     {
         let irows = irows.into_iter();
-        let ncols = self.data.shape().1;
-        let mut res =
-            OMatrix::<T, Dynamic, C>::new_uninitialized_generic(Dynamic::new(irows.len()), ncols);
+        let ncols = self.shape_generic().1;
+        let mut res = Matrix::uninit(Dynamic::new(irows.len()), ncols);
 
         // First, check that all the indices from irows are valid.
         // This will allow us to use unchecked access in the inner loop.
@@ -68,13 +67,15 @@ impl<T: Scalar + Zero, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
             let src = self.column(j);
 
             for (destination, source) in irows.clone().enumerate() {
+                // Safety: all indices are in range.
                 unsafe {
                     *res.vget_unchecked_mut(destination) =
-                        MaybeUninit::new(src.vget_unchecked(*source).clone());
+                        MaybeUninit::new(src.vget_unchecked(*source).inlined_clone());
                 }
             }
         }
 
+        // Safety: res is now fully initialized.
         unsafe { res.assume_init() }
     }
 
@@ -88,32 +89,30 @@ impl<T: Scalar + Zero, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         DefaultAllocator: Allocator<T, R, Dynamic>,
     {
         let icols = icols.into_iter();
-        let nrows = self.data.shape().0;
-        let mut res = Matrix::new_uninitialized_generic(nrows, Dynamic::new(icols.len()));
+        let nrows = self.shape_generic().0;
+        let mut res = Matrix::uninit(nrows, Dynamic::new(icols.len()));
 
         for (destination, source) in icols.enumerate() {
-            for (d, s) in res
-                .column_mut(destination)
-                .iter_mut()
-                .zip(self.column(*source).iter())
-            {
-                *d = MaybeUninit::new(s.clone());
-            }
+            // NOTE: this is basically a copy_frow but wrapping the values insnide of MaybeUninit.
+            res.column_mut(destination)
+                .zip_apply(&self.column(*source), |out, e| {
+                    *out = MaybeUninit::new(e.inlined_clone())
+                });
         }
 
+        // Safety: res is now fully initialized.
         unsafe { res.assume_init() }
     }
 }
 
 /// # Set rows, columns, and diagonal
-impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
+impl<T: Scalar, R: Dim, C: Dim, S: RawStorageMut<T, R, C>> Matrix<T, R, C, S> {
     /// Fills the diagonal of this matrix with the content of the given vector.
     #[inline]
     pub fn set_diagonal<R2: Dim, S2>(&mut self, diag: &Vector<T, R2, S2>)
     where
-        T: Clone,
         R: DimMin<C>,
-        S2: Storage<T, R2>,
+        S2: RawStorage<T, R2>,
         ShapeConstraint: DimEq<DimMinimum<R, C>, R2>,
     {
         let (nrows, ncols) = self.shape();
@@ -121,7 +120,7 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
         assert_eq!(diag.len(), min_nrows_ncols, "Mismatched dimensions.");
 
         for i in 0..min_nrows_ncols {
-            unsafe { *self.get_unchecked_mut((i, i)) = diag.vget_unchecked(i).clone() }
+            unsafe { *self.get_unchecked_mut((i, i)) = diag.vget_unchecked(i).inlined_clone() }
         }
     }
 
@@ -144,8 +143,7 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn set_row<C2: Dim, S2>(&mut self, i: usize, row: &RowVector<T, C2, S2>)
     where
-        T: Clone,
-        S2: Storage<T, U1, C2>,
+        S2: RawStorage<T, U1, C2>,
         ShapeConstraint: SameNumberOfColumns<C, C2>,
     {
         self.row_mut(i).copy_from(row);
@@ -155,8 +153,7 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn set_column<R2: Dim, S2>(&mut self, i: usize, column: &Vector<T, R2, S2>)
     where
-        T: Clone,
-        S2: Storage<T, R2, U1>,
+        S2: RawStorage<T, R2, U1>,
         ShapeConstraint: SameNumberOfRows<R, R2>,
     {
         self.column_mut(i).copy_from(column);
@@ -164,23 +161,23 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
 }
 
 /// # In-place filling
-impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorageMut<T, R, C>> Matrix<T, R, C, S> {
+    /// Sets all the elements of this matrix to the value returned by the closure.
+    #[inline]
+    pub fn fill_with(&mut self, val: impl Fn() -> T) {
+        for e in self.iter_mut() {
+            *e = val()
+        }
+    }
+
     /// Sets all the elements of this matrix to `val`.
     #[inline]
     pub fn fill(&mut self, val: T)
     where
-        T: Clone,
+        T: Scalar,
     {
         for e in self.iter_mut() {
-            *e = val.clone()
-        }
-    }
-
-    /// Sets all the elements of this matrix to `f()`.
-    #[inline]
-    pub fn fill_fn<F: FnMut() -> T>(&mut self, mut f: F) {
-        for e in self.iter_mut() {
-            *e = f();
+            *e = val.inlined_clone()
         }
     }
 
@@ -188,7 +185,7 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn fill_with_identity(&mut self)
     where
-        T: Zero + One + Scalar,
+        T: Scalar + Zero + One,
     {
         self.fill(T::zero());
         self.fill_diagonal(T::one());
@@ -198,13 +195,13 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn fill_diagonal(&mut self, val: T)
     where
-        T: Clone,
+        T: Scalar,
     {
         let (nrows, ncols) = self.shape();
         let n = cmp::min(nrows, ncols);
 
         for i in 0..n {
-            unsafe { *self.get_unchecked_mut((i, i)) = val.clone() }
+            unsafe { *self.get_unchecked_mut((i, i)) = val.inlined_clone() }
         }
     }
 
@@ -212,11 +209,11 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn fill_row(&mut self, i: usize, val: T)
     where
-        T: Clone,
+        T: Scalar,
     {
         assert!(i < self.nrows(), "Row index out of bounds.");
         for j in 0..self.ncols() {
-            unsafe { *self.get_unchecked_mut((i, j)) = val.clone() }
+            unsafe { *self.get_unchecked_mut((i, j)) = val.inlined_clone() }
         }
     }
 
@@ -224,11 +221,11 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn fill_column(&mut self, j: usize, val: T)
     where
-        T: Clone,
+        T: Scalar,
     {
         assert!(j < self.ncols(), "Row index out of bounds.");
         for i in 0..self.nrows() {
-            unsafe { *self.get_unchecked_mut((i, j)) = val.clone() }
+            unsafe { *self.get_unchecked_mut((i, j)) = val.inlined_clone() }
         }
     }
 
@@ -242,11 +239,11 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn fill_lower_triangle(&mut self, val: T, shift: usize)
     where
-        T: Clone,
+        T: Scalar,
     {
         for j in 0..self.ncols() {
             for i in (j + shift)..self.nrows() {
-                unsafe { *self.get_unchecked_mut((i, j)) = val.clone() }
+                unsafe { *self.get_unchecked_mut((i, j)) = val.inlined_clone() }
             }
         }
     }
@@ -261,19 +258,19 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn fill_upper_triangle(&mut self, val: T, shift: usize)
     where
-        T: Clone,
+        T: Scalar,
     {
         for j in shift..self.ncols() {
             // TODO: is there a more efficient way to avoid the min ?
             // (necessary for rectangular matrices)
             for i in 0..cmp::min(j + 1 - shift, self.nrows()) {
-                unsafe { *self.get_unchecked_mut((i, j)) = val.clone() }
+                unsafe { *self.get_unchecked_mut((i, j)) = val.inlined_clone() }
             }
         }
     }
 }
 
-impl<T: Clone, D: Dim, S: StorageMut<T, D, D>> Matrix<T, D, D, S> {
+impl<T: Scalar, D: Dim, S: RawStorageMut<T, D, D>> Matrix<T, D, D, S> {
     /// Copies the upper-triangle of this matrix to its lower-triangular part.
     ///
     /// This makes the matrix symmetric. Panics if the matrix is not square.
@@ -284,7 +281,7 @@ impl<T: Clone, D: Dim, S: StorageMut<T, D, D>> Matrix<T, D, D, S> {
         for j in 0..dim {
             for i in j + 1..dim {
                 unsafe {
-                    *self.get_unchecked_mut((i, j)) = self.get_unchecked((j, i)).clone();
+                    *self.get_unchecked_mut((i, j)) = self.get_unchecked((j, i)).inlined_clone();
                 }
             }
         }
@@ -299,7 +296,7 @@ impl<T: Clone, D: Dim, S: StorageMut<T, D, D>> Matrix<T, D, D, S> {
         for j in 1..self.ncols() {
             for i in 0..j {
                 unsafe {
-                    *self.get_unchecked_mut((i, j)) = self.get_unchecked((j, i)).clone();
+                    *self.get_unchecked_mut((i, j)) = self.get_unchecked((j, i)).inlined_clone();
                 }
             }
         }
@@ -307,7 +304,7 @@ impl<T: Clone, D: Dim, S: StorageMut<T, D, D>> Matrix<T, D, D, S> {
 }
 
 /// # In-place swapping
-impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
+impl<T: Scalar, R: Dim, C: Dim, S: RawStorageMut<T, R, C>> Matrix<T, R, C, S> {
     /// Swaps two rows in-place.
     #[inline]
     pub fn swap_rows(&mut self, irow1: usize, irow2: usize) {
@@ -343,7 +340,7 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
  *
  */
 /// # Rows and columns removal
-impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T: Scalar, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     /*
      *
      * Column removal.
@@ -367,7 +364,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         DefaultAllocator: Reallocator<T, R, C, R, Dynamic>,
     {
         let mut m = self.into_owned();
-        let (nrows, ncols) = m.data.shape();
+        let (nrows, ncols) = m.shape_generic();
         let mut offset: usize = 0;
         let mut target: usize = 0;
         while offset + target < ncols.value() {
@@ -401,7 +398,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         DefaultAllocator: Reallocator<T, R, C, Dynamic, C>,
     {
         let mut m = self.into_owned();
-        let (nrows, ncols) = m.data.shape();
+        let (nrows, ncols) = m.shape_generic();
         let mut offset: usize = 0;
         let mut target: usize = 0;
         while offset + target < nrows.value() * ncols.value() {
@@ -464,7 +461,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         DefaultAllocator: Reallocator<T, R, C, R, DimDiff<C, D>>,
     {
         let mut m = self.into_owned();
-        let (nrows, ncols) = m.data.shape();
+        let (nrows, ncols) = m.shape_generic();
         assert!(
             i + nremove.value() <= ncols.value(),
             "Column index out of range."
@@ -543,7 +540,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         DefaultAllocator: Reallocator<T, R, C, DimDiff<R, D>, C>,
     {
         let mut m = self.into_owned();
-        let (nrows, ncols) = m.data.shape();
+        let (nrows, ncols) = m.shape_generic();
         assert!(
             i + nremove.value() <= nrows.value(),
             "Row index out of range."
@@ -552,7 +549,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         if nremove.value() != 0 {
             unsafe {
                 compress_rows(
-                    &mut m.data.as_mut_slice(),
+                    &mut m.as_mut_slice(),
                     nrows.value(),
                     ncols.value(),
                     i,
@@ -572,7 +569,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
 }
 
 /// # Rows and columns insertion
-impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T: Scalar, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     /*
      *
      * Columns insertion.
@@ -633,7 +630,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         DefaultAllocator: Reallocator<T, R, C, R, DimSum<C, D>>,
     {
         let m = self.into_owned();
-        let (nrows, ncols) = m.data.shape();
+        let (nrows, ncols) = m.shape_generic();
         let mut res = Matrix::from_data(DefaultAllocator::reallocate_copy(
             nrows,
             ncols.add(ninsert),
@@ -717,7 +714,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         DefaultAllocator: Reallocator<T, R, C, DimSum<R, D>, C>,
     {
         let m = self.into_owned();
-        let (nrows, ncols) = m.data.shape();
+        let (nrows, ncols) = m.shape_generic();
         let mut res = Matrix::from_data(DefaultAllocator::reallocate_copy(
             nrows.add(ninsert),
             ncols,
@@ -728,7 +725,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
 
         if ninsert.value() != 0 {
             extend_rows(
-                &mut res.data.as_mut_slice(),
+                &mut res.as_mut_slice(),
                 nrows.value(),
                 ncols.value(),
                 i,
@@ -741,7 +738,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
 }
 
 /// # Resizing and reshaping
-impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T: Scalar, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     /// Resizes this matrix so that it contains `new_nrows` rows and `new_ncols` columns.
     ///
     /// The values are copied such that `self[(i, j)] == result[(i, j)]`. If the result has more
@@ -763,7 +760,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     where
         DefaultAllocator: Reallocator<T, R, C, Dynamic, C>,
     {
-        let ncols = self.data.shape().1;
+        let ncols = self.shape_generic().1;
         self.resize_generic(Dynamic::new(new_nrows), ncols, val)
     }
 
@@ -776,7 +773,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     where
         DefaultAllocator: Reallocator<T, R, C, R, Dynamic>,
     {
-        let nrows = self.data.shape().0;
+        let nrows = self.shape_generic().0;
         self.resize_generic(nrows, Dynamic::new(new_ncols), val)
     }
 
@@ -809,10 +806,10 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         DefaultAllocator: Reallocator<T, R, C, R2, C2>,
     {
         let (nrows, ncols) = self.shape();
-        let mut data = self.data.into_owned();
+        let mut data = self.into_owned();
 
         if new_nrows.value() == nrows {
-            let res = unsafe { DefaultAllocator::reallocate_copy(new_nrows, new_ncols, data.0) };
+            let res = unsafe { DefaultAllocator::reallocate_copy(new_nrows, new_ncols, data.data) };
             let mut res = Matrix::from_data(res);
             if new_ncols.value() > ncols {
                 res.columns_range_mut(ncols..).fill(val);
@@ -832,14 +829,14 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
                         nrows - new_nrows.value(),
                     );
                     res = Matrix::from_data(DefaultAllocator::reallocate_copy(
-                        new_nrows, new_ncols, data.0,
+                        new_nrows, new_ncols, data.data,
                     ));
                 } else {
                     res = Matrix::from_data(DefaultAllocator::reallocate_copy(
-                        new_nrows, new_ncols, data.0,
+                        new_nrows, new_ncols, data.data,
                     ));
                     extend_rows(
-                        &mut res.data.as_mut_slice(),
+                        &mut res.as_mut_slice(),
                         nrows,
                         new_ncols.value(),
                         nrows,
@@ -849,7 +846,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
             }
 
             if new_ncols.value() > ncols {
-                res.columns_range_mut(ncols..).fill(val.clone());
+                res.columns_range_mut(ncols..).fill(val.inlined_clone());
             }
 
             if new_nrows.value() > nrows {
@@ -931,7 +928,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
 
 /// # In-place resizing
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<T: Clone> OMatrix<T, Dynamic, Dynamic> {
+impl<T: Scalar> OMatrix<T, Dynamic, Dynamic> {
     /// Resizes this matrix in-place.
     ///
     /// The values are copied such that `self[(i, j)] == result[(i, j)]`. If the result has more
@@ -942,20 +939,13 @@ impl<T: Clone> OMatrix<T, Dynamic, Dynamic> {
     where
         DefaultAllocator: Reallocator<T, Dynamic, Dynamic, Dynamic, Dynamic>,
     {
-        // IMPORTANT TODO: this method is still UB, and we should decide how to
-        // update the API to take it into account.
-
-        let placeholder = unsafe {
-            Matrix::new_uninitialized_generic(Dynamic::new(0), Dynamic::new(0)).assume_init()
-        };
-        let old = mem::replace(self, placeholder);
-        let new = old.resize(new_nrows, new_ncols, val);
-        let _ = mem::replace(self, new);
+        // TODO: avoid the clone.
+        *self = self.clone().resize(new_nrows, new_ncols, val);
     }
 }
 
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<T: Clone, C: Dim> OMatrix<T, Dynamic, C>
+impl<T: Scalar, C: Dim> OMatrix<T, Dynamic, C>
 where
     DefaultAllocator: Allocator<T, Dynamic, C>,
 {
@@ -970,16 +960,13 @@ where
     where
         DefaultAllocator: Reallocator<T, Dynamic, C, Dynamic, C>,
     {
-        let placeholder =
-            Matrix::from_fn_generic(Dynamic::new(0), self.data.shape().1, |_, _| val.clone());
-        let old = mem::replace(self, placeholder);
-        let new = old.resize_vertically(new_nrows, val);
-        let _ = mem::replace(self, new);
+        // TODO: avoid the clone.
+        *self = self.clone().resize_vertically(new_nrows, val);
     }
 }
 
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<T: Clone, R: Dim> OMatrix<T, R, Dynamic>
+impl<T: Scalar, R: Dim> OMatrix<T, R, Dynamic>
 where
     DefaultAllocator: Allocator<T, R, Dynamic>,
 {
@@ -994,15 +981,18 @@ where
     where
         DefaultAllocator: Reallocator<T, R, Dynamic, R, Dynamic>,
     {
-        let placeholder =
-            Matrix::from_fn_generic(self.data.shape().0, Dynamic::new(0), |_, _| val.clone());
-        let old = mem::replace(self, placeholder);
-        let new = old.resize_horizontally(new_ncols, val);
-        let _ = mem::replace(self, new);
+        // TODO: avoid the clone.
+        *self = self.clone().resize_horizontally(new_ncols, val);
     }
 }
 
-unsafe fn compress_rows<T>(data: &mut [T], nrows: usize, ncols: usize, i: usize, nremove: usize) {
+unsafe fn compress_rows<T: Scalar>(
+    data: &mut [T],
+    nrows: usize,
+    ncols: usize,
+    i: usize,
+    nremove: usize,
+) {
     let new_nrows = nrows - nremove;
 
     if new_nrows == 0 || ncols == 0 {
@@ -1035,7 +1025,13 @@ unsafe fn compress_rows<T>(data: &mut [T], nrows: usize, ncols: usize, i: usize,
 
 // Moves entries of a matrix buffer to make place for `ninsert` emty rows starting at the `i-th` row index.
 // The `data` buffer is assumed to contained at least `(nrows + ninsert) * ncols` elements.
-unsafe fn extend_rows<T>(data: &mut [T], nrows: usize, ncols: usize, i: usize, ninsert: usize) {
+unsafe fn extend_rows<T: Scalar>(
+    data: &mut [T],
+    nrows: usize,
+    ncols: usize,
+    i: usize,
+    ninsert: usize,
+) {
     let new_nrows = nrows + ninsert;
 
     if new_nrows == 0 || ncols == 0 {
@@ -1065,7 +1061,12 @@ unsafe fn extend_rows<T>(data: &mut [T], nrows: usize, ncols: usize, i: usize, n
 /// Extend the number of columns of the `Matrix` with elements from
 /// a given iterator.
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<T, R: Dim, S: Extend<T>> Extend<T> for Matrix<T, R, Dynamic, S> {
+impl<T, R, S> Extend<T> for Matrix<T, R, Dynamic, S>
+where
+    T: Scalar,
+    R: Dim,
+    S: Extend<T>,
+{
     /// Extend the number of columns of the `Matrix` with elements
     /// from the given iterator.
     ///
@@ -1110,6 +1111,7 @@ impl<T, R: Dim, S: Extend<T>> Extend<T> for Matrix<T, R, Dynamic, S> {
 #[cfg(any(feature = "std", feature = "alloc"))]
 impl<T, S> Extend<T> for Matrix<T, Dynamic, U1, S>
 where
+    T: Scalar,
     S: Extend<T>,
 {
     /// Extend the number of rows of a `Vector` with elements
@@ -1128,10 +1130,13 @@ where
 }
 
 #[cfg(any(feature = "std", feature = "alloc"))]
-impl<T, R: Dim, S, RV: Dim, SV> Extend<Vector<T, RV, SV>> for Matrix<T, R, Dynamic, S>
+impl<T, R, S, RV, SV> Extend<Vector<T, RV, SV>> for Matrix<T, R, Dynamic, S>
 where
+    T: Scalar,
+    R: Dim,
     S: Extend<Vector<T, RV, SV>>,
-    SV: Storage<T, RV>,
+    RV: Dim,
+    SV: RawStorage<T, RV>,
     ShapeConstraint: SameNumberOfRows<R, RV>,
 {
     /// Extends the number of columns of a `Matrix` with `Vector`s
diff --git a/src/base/indexing.rs b/src/base/indexing.rs
index bb0adddb..93f41ed3 100644
--- a/src/base/indexing.rs
+++ b/src/base/indexing.rs
@@ -1,8 +1,8 @@
 //! Indexing
 
-use crate::base::storage::{Storage, StorageMut};
+use crate::base::storage::{RawStorage, RawStorageMut};
 use crate::base::{
-    Const, Dim, DimDiff, DimName, DimSub, Dynamic, Matrix, MatrixSlice, MatrixSliceMut, U1,
+    Const, Dim, DimDiff, DimName, DimSub, Dynamic, Matrix, MatrixSlice, MatrixSliceMut, Scalar, U1,
 };
 
 use std::ops;
@@ -310,7 +310,7 @@ fn dimrange_rangetoinclusive_usize() {
 }
 
 /// A helper trait used for indexing operations.
-pub trait MatrixIndex<'a, T, R: Dim, C: Dim, S: Storage<T, R, C>>: Sized {
+pub trait MatrixIndex<'a, T, R: Dim, C: Dim, S: RawStorage<T, R, C>>: Sized {
     /// The output type returned by methods.
     type Output: 'a;
 
@@ -345,7 +345,7 @@ pub trait MatrixIndex<'a, T, R: Dim, C: Dim, S: Storage<T, R, C>>: Sized {
 }
 
 /// A helper trait used for indexing operations.
-pub trait MatrixIndexMut<'a, T, R: Dim, C: Dim, S: StorageMut<T, R, C>>:
+pub trait MatrixIndexMut<'a, T, R: Dim, C: Dim, S: RawStorageMut<T, R, C>>:
     MatrixIndex<'a, T, R, C, S>
 {
     /// The output type returned by methods.
@@ -476,7 +476,7 @@ pub trait MatrixIndexMut<'a, T, R: Dim, C: Dim, S: StorageMut<T, R, C>>:
 ///                         4, 7,
 ///                         5, 8)));
 /// ```
-impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     /// Produces a view of the data at the given index, or
     /// `None` if the index is out of bounds.
     #[inline]
@@ -494,7 +494,7 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     #[must_use]
     pub fn get_mut<'a, I>(&'a mut self, index: I) -> Option<I::OutputMut>
     where
-        S: StorageMut<T, R, C>,
+        S: RawStorageMut<T, R, C>,
         I: MatrixIndexMut<'a, T, R, C, S>,
     {
         index.get_mut(self)
@@ -516,7 +516,7 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn index_mut<'a, I>(&'a mut self, index: I) -> I::OutputMut
     where
-        S: StorageMut<T, R, C>,
+        S: RawStorageMut<T, R, C>,
         I: MatrixIndexMut<'a, T, R, C, S>,
     {
         index.index_mut(self)
@@ -539,7 +539,7 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     #[must_use]
     pub unsafe fn get_unchecked_mut<'a, I>(&'a mut self, index: I) -> I::OutputMut
     where
-        S: StorageMut<T, R, C>,
+        S: RawStorageMut<T, R, C>,
         I: MatrixIndexMut<'a, T, R, C, S>,
     {
         index.get_unchecked_mut(self)
@@ -548,9 +548,12 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
 
 // EXTRACT A SINGLE ELEMENT BY 1D LINEAR ADDRESS
 
-impl<'a, T: 'a, R: Dim, C: Dim, S> MatrixIndex<'a, T, R, C, S> for usize
+impl<'a, T, R, C, S> MatrixIndex<'a, T, R, C, S> for usize
 where
-    S: Storage<T, R, C>,
+    T: Scalar,
+    R: Dim,
+    C: Dim,
+    S: RawStorage<T, R, C>,
 {
     type Output = &'a T;
 
@@ -567,9 +570,12 @@ where
     }
 }
 
-impl<'a, T: 'a, R: Dim, C: Dim, S> MatrixIndexMut<'a, T, R, C, S> for usize
+impl<'a, T, R, C, S> MatrixIndexMut<'a, T, R, C, S> for usize
 where
-    S: StorageMut<T, R, C>,
+    T: Scalar,
+    R: Dim,
+    C: Dim,
+    S: RawStorageMut<T, R, C>,
 {
     type OutputMut = &'a mut T;
 
@@ -577,7 +583,7 @@ where
     #[inline(always)]
     unsafe fn get_unchecked_mut(self, matrix: &'a mut Matrix<T, R, C, S>) -> Self::OutputMut
     where
-        S: StorageMut<T, R, C>,
+        S: RawStorageMut<T, R, C>,
     {
         matrix.data.get_unchecked_linear_mut(self)
     }
@@ -585,9 +591,11 @@ where
 
 // EXTRACT A SINGLE ELEMENT BY 2D COORDINATES
 
-impl<'a, T: 'a, R: Dim, C: Dim, S> MatrixIndex<'a, T, R, C, S> for (usize, usize)
+impl<'a, T: 'a, R, C, S> MatrixIndex<'a, T, R, C, S> for (usize, usize)
 where
-    S: Storage<T, R, C>,
+    R: Dim,
+    C: Dim,
+    S: RawStorage<T, R, C>,
 {
     type Output = &'a T;
 
@@ -595,7 +603,7 @@ where
     #[inline(always)]
     fn contained_by(&self, matrix: &Matrix<T, R, C, S>) -> bool {
         let (rows, cols) = self;
-        let (nrows, ncols) = matrix.data.shape();
+        let (nrows, ncols) = matrix.shape_generic();
         DimRange::contained_by(rows, nrows) && DimRange::contained_by(cols, ncols)
     }
 
@@ -607,9 +615,11 @@ where
     }
 }
 
-impl<'a, T: 'a, R: Dim, C: Dim, S> MatrixIndexMut<'a, T, R, C, S> for (usize, usize)
+impl<'a, T: 'a, R, C, S> MatrixIndexMut<'a, T, R, C, S> for (usize, usize)
 where
-    S: StorageMut<T, R, C>,
+    R: Dim,
+    C: Dim,
+    S: RawStorageMut<T, R, C>,
 {
     type OutputMut = &'a mut T;
 
@@ -617,7 +627,7 @@ where
     #[inline(always)]
     unsafe fn get_unchecked_mut(self, matrix: &'a mut Matrix<T, R, C, S>) -> Self::OutputMut
     where
-        S: StorageMut<T, R, C>,
+        S: RawStorageMut<T, R, C>,
     {
         let (row, col) = self;
         matrix.data.get_unchecked_mut(row, col)
@@ -643,10 +653,12 @@ macro_rules! impl_index_pair {
         $(where $CConstraintType: ty: $CConstraintBound: ident $(<$($CConstraintBoundParams: ty $( = $CEqBound: ty )*),*>)* )*]
     ) =>
     {
-        impl<'a, T: 'a, $R: Dim, $C: Dim, S, $($RTyP : $RTyPB,)* $($CTyP : $CTyPB),*>
-            MatrixIndex<'a, T, $R, $C, S> for ($RIdx, $CIdx)
+        impl<'a, T, $R, $C, S, $($RTyP : $RTyPB,)* $($CTyP : $CTyPB),*> MatrixIndex<'a, T, $R, $C, S> for ($RIdx, $CIdx)
         where
-            S: Storage<T, R, C>,
+            T: Scalar,
+            $R: Dim,
+            $C: Dim,
+            S: RawStorage<T, R, C>,
             $( $RConstraintType: $RConstraintBound $(<$( $RConstraintBoundParams $( = $REqBound )*),*>)* ,)*
             $( $CConstraintType: $CConstraintBound $(<$( $CConstraintBoundParams $( = $CEqBound )*),*>)* ),*
         {
@@ -656,7 +668,7 @@ macro_rules! impl_index_pair {
             #[inline(always)]
             fn contained_by(&self, matrix: &Matrix<T, $R, $C, S>) -> bool {
                 let (rows, cols) = self;
-                let (nrows, ncols) = matrix.data.shape();
+                let (nrows, ncols) = matrix.shape_generic();
                 DimRange::contained_by(rows, nrows) && DimRange::contained_by(cols, ncols)
             }
 
@@ -666,21 +678,23 @@ macro_rules! impl_index_pair {
                 use crate::base::SliceStorage;
 
                 let (rows, cols) = self;
-                let (nrows, ncols) = matrix.data.shape();
+                let (nrows, ncols) = matrix.shape_generic();
 
                 let data =
                     SliceStorage::new_unchecked(&matrix.data,
                         (rows.lower(nrows),  cols.lower(ncols)),
                         (rows.length(nrows), cols.length(ncols)));
 
-                Matrix::from_data(data)
+                Matrix::from_data_statically_unchecked(data)
             }
         }
 
-        impl<'a, T: 'a, $R: Dim, $C: Dim, S, $($RTyP : $RTyPB,)* $($CTyP : $CTyPB),*>
-            MatrixIndexMut<'a, T, $R, $C, S> for ($RIdx, $CIdx)
+        impl<'a, T, $R, $C, S, $($RTyP : $RTyPB,)* $($CTyP : $CTyPB),*> MatrixIndexMut<'a, T, $R, $C, S> for ($RIdx, $CIdx)
         where
-            S: StorageMut<T, R, C>,
+            T: Scalar,
+            $R: Dim,
+            $C: Dim,
+            S: RawStorageMut<T, R, C>,
             $( $RConstraintType: $RConstraintBound $(<$( $RConstraintBoundParams $( = $REqBound )*),*>)* ,)*
             $( $CConstraintType: $CConstraintBound $(<$( $CConstraintBoundParams $( = $CEqBound )*),*>)* ),*
         {
@@ -692,14 +706,14 @@ macro_rules! impl_index_pair {
                 use crate::base::SliceStorageMut;
 
                 let (rows, cols) = self;
-                let (nrows, ncols) = matrix.data.shape();
+                let (nrows, ncols) = matrix.shape_generic();
 
                 let data =
                     SliceStorageMut::new_unchecked(&mut matrix.data,
                         (rows.lower(nrows),  cols.lower(ncols)),
                         (rows.length(nrows), cols.length(ncols)));
 
-                Matrix::from_data(data)
+                Matrix::from_data_statically_unchecked(data)
             }
         }
     }
diff --git a/src/base/iter.rs b/src/base/iter.rs
index b48e8322..b68e1051 100644
--- a/src/base/iter.rs
+++ b/src/base/iter.rs
@@ -5,12 +5,13 @@ use std::marker::PhantomData;
 use std::mem;
 
 use crate::base::dimension::{Dim, U1};
-use crate::base::storage::{Storage, StorageMut};
-use crate::base::{Matrix, MatrixSlice, MatrixSliceMut};
+use crate::base::storage::{RawStorage, RawStorageMut};
+use crate::base::{Matrix, MatrixSlice, MatrixSliceMut, Scalar};
 
 macro_rules! iterator {
     (struct $Name:ident for $Storage:ident.$ptr: ident -> $Ptr:ty, $Ref:ty, $SRef: ty) => {
         /// An iterator through a dense matrix with arbitrary strides matrix.
+        #[derive(Debug)]
         pub struct $Name<'a, T, R: Dim, C: Dim, S: 'a + $Storage<T, R, C>> {
             ptr: $Ptr,
             inner_ptr: $Ptr,
@@ -170,8 +171,8 @@ macro_rules! iterator {
     };
 }
 
-iterator!(struct MatrixIter for Storage.ptr -> *const T, &'a T, &'a S);
-iterator!(struct MatrixIterMut for StorageMut.ptr_mut -> *mut T, &'a mut T, &'a mut S);
+iterator!(struct MatrixIter for RawStorage.ptr -> *const T, &'a T, &'a S);
+iterator!(struct MatrixIterMut for RawStorageMut.ptr_mut -> *mut T, &'a mut T, &'a mut S);
 
 /*
  *
@@ -180,18 +181,18 @@ iterator!(struct MatrixIterMut for StorageMut.ptr_mut -> *mut T, &'a mut T, &'a
  */
 #[derive(Clone, Debug)]
 /// An iterator through the rows of a matrix.
-pub struct RowIter<'a, T, R: Dim, C: Dim, S: Storage<T, R, C>> {
+pub struct RowIter<'a, T, R: Dim, C: Dim, S: RawStorage<T, R, C>> {
     mat: &'a Matrix<T, R, C, S>,
     curr: usize,
 }
 
-impl<'a, T, R: Dim, C: Dim, S: 'a + Storage<T, R, C>> RowIter<'a, T, R, C, S> {
+impl<'a, T, R: Dim, C: Dim, S: 'a + RawStorage<T, R, C>> RowIter<'a, T, R, C, S> {
     pub(crate) fn new(mat: &'a Matrix<T, R, C, S>) -> Self {
         RowIter { mat, curr: 0 }
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S: 'a + Storage<T, R, C>> Iterator for RowIter<'a, T, R, C, S> {
+impl<'a, T, R: Dim, C: Dim, S: 'a + RawStorage<T, R, C>> Iterator for RowIter<'a, T, R, C, S> {
     type Item = MatrixSlice<'a, T, U1, C, S::RStride, S::CStride>;
 
     #[inline]
@@ -219,7 +220,7 @@ impl<'a, T, R: Dim, C: Dim, S: 'a + Storage<T, R, C>> Iterator for RowIter<'a, T
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S: 'a + Storage<T, R, C>> ExactSizeIterator
+impl<'a, T: Scalar, R: Dim, C: Dim, S: 'a + RawStorage<T, R, C>> ExactSizeIterator
     for RowIter<'a, T, R, C, S>
 {
     #[inline]
@@ -229,13 +230,14 @@ impl<'a, T, R: Dim, C: Dim, S: 'a + Storage<T, R, C>> ExactSizeIterator
 }
 
 /// An iterator through the mutable rows of a matrix.
-pub struct RowIterMut<'a, T, R: Dim, C: Dim, S: StorageMut<T, R, C>> {
+#[derive(Debug)]
+pub struct RowIterMut<'a, T, R: Dim, C: Dim, S: RawStorageMut<T, R, C>> {
     mat: *mut Matrix<T, R, C, S>,
     curr: usize,
     phantom: PhantomData<&'a mut Matrix<T, R, C, S>>,
 }
 
-impl<'a, T, R: Dim, C: Dim, S: 'a + StorageMut<T, R, C>> RowIterMut<'a, T, R, C, S> {
+impl<'a, T, R: Dim, C: Dim, S: 'a + RawStorageMut<T, R, C>> RowIterMut<'a, T, R, C, S> {
     pub(crate) fn new(mat: &'a mut Matrix<T, R, C, S>) -> Self {
         RowIterMut {
             mat,
@@ -249,7 +251,9 @@ impl<'a, T, R: Dim, C: Dim, S: 'a + StorageMut<T, R, C>> RowIterMut<'a, T, R, C,
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S: 'a + StorageMut<T, R, C>> Iterator for RowIterMut<'a, T, R, C, S> {
+impl<'a, T, R: Dim, C: Dim, S: 'a + RawStorageMut<T, R, C>> Iterator
+    for RowIterMut<'a, T, R, C, S>
+{
     type Item = MatrixSliceMut<'a, T, U1, C, S::RStride, S::CStride>;
 
     #[inline]
@@ -274,7 +278,7 @@ impl<'a, T, R: Dim, C: Dim, S: 'a + StorageMut<T, R, C>> Iterator for RowIterMut
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S: 'a + StorageMut<T, R, C>> ExactSizeIterator
+impl<'a, T: Scalar, R: Dim, C: Dim, S: 'a + RawStorageMut<T, R, C>> ExactSizeIterator
     for RowIterMut<'a, T, R, C, S>
 {
     #[inline]
@@ -290,18 +294,18 @@ impl<'a, T, R: Dim, C: Dim, S: 'a + StorageMut<T, R, C>> ExactSizeIterator
  */
 #[derive(Clone, Debug)]
 /// An iterator through the columns of a matrix.
-pub struct ColumnIter<'a, T, R: Dim, C: Dim, S: Storage<T, R, C>> {
+pub struct ColumnIter<'a, T, R: Dim, C: Dim, S: RawStorage<T, R, C>> {
     mat: &'a Matrix<T, R, C, S>,
     curr: usize,
 }
 
-impl<'a, T, R: Dim, C: Dim, S: 'a + Storage<T, R, C>> ColumnIter<'a, T, R, C, S> {
+impl<'a, T, R: Dim, C: Dim, S: 'a + RawStorage<T, R, C>> ColumnIter<'a, T, R, C, S> {
     pub(crate) fn new(mat: &'a Matrix<T, R, C, S>) -> Self {
         ColumnIter { mat, curr: 0 }
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S: 'a + Storage<T, R, C>> Iterator for ColumnIter<'a, T, R, C, S> {
+impl<'a, T, R: Dim, C: Dim, S: 'a + RawStorage<T, R, C>> Iterator for ColumnIter<'a, T, R, C, S> {
     type Item = MatrixSlice<'a, T, R, U1, S::RStride, S::CStride>;
 
     #[inline]
@@ -329,7 +333,7 @@ impl<'a, T, R: Dim, C: Dim, S: 'a + Storage<T, R, C>> Iterator for ColumnIter<'a
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S: 'a + Storage<T, R, C>> ExactSizeIterator
+impl<'a, T: Scalar, R: Dim, C: Dim, S: 'a + RawStorage<T, R, C>> ExactSizeIterator
     for ColumnIter<'a, T, R, C, S>
 {
     #[inline]
@@ -339,13 +343,14 @@ impl<'a, T, R: Dim, C: Dim, S: 'a + Storage<T, R, C>> ExactSizeIterator
 }
 
 /// An iterator through the mutable columns of a matrix.
-pub struct ColumnIterMut<'a, T, R: Dim, C: Dim, S: StorageMut<T, R, C>> {
+#[derive(Debug)]
+pub struct ColumnIterMut<'a, T, R: Dim, C: Dim, S: RawStorageMut<T, R, C>> {
     mat: *mut Matrix<T, R, C, S>,
     curr: usize,
     phantom: PhantomData<&'a mut Matrix<T, R, C, S>>,
 }
 
-impl<'a, T, R: Dim, C: Dim, S: 'a + StorageMut<T, R, C>> ColumnIterMut<'a, T, R, C, S> {
+impl<'a, T, R: Dim, C: Dim, S: 'a + RawStorageMut<T, R, C>> ColumnIterMut<'a, T, R, C, S> {
     pub(crate) fn new(mat: &'a mut Matrix<T, R, C, S>) -> Self {
         ColumnIterMut {
             mat,
@@ -359,7 +364,7 @@ impl<'a, T, R: Dim, C: Dim, S: 'a + StorageMut<T, R, C>> ColumnIterMut<'a, T, R,
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S: 'a + StorageMut<T, R, C>> Iterator
+impl<'a, T, R: Dim, C: Dim, S: 'a + RawStorageMut<T, R, C>> Iterator
     for ColumnIterMut<'a, T, R, C, S>
 {
     type Item = MatrixSliceMut<'a, T, R, U1, S::RStride, S::CStride>;
@@ -386,7 +391,7 @@ impl<'a, T, R: Dim, C: Dim, S: 'a + StorageMut<T, R, C>> Iterator
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, S: 'a + StorageMut<T, R, C>> ExactSizeIterator
+impl<'a, T: Scalar, R: Dim, C: Dim, S: 'a + RawStorageMut<T, R, C>> ExactSizeIterator
     for ColumnIterMut<'a, T, R, C, S>
 {
     #[inline]
diff --git a/src/base/matrix.rs b/src/base/matrix.rs
index 8ec78264..6cca767a 100644
--- a/src/base/matrix.rs
+++ b/src/base/matrix.rs
@@ -8,8 +8,7 @@ use std::cmp::Ordering;
 use std::fmt;
 use std::hash::{Hash, Hasher};
 use std::marker::PhantomData;
-use std::mem::{self, ManuallyDrop, MaybeUninit};
-use std::ptr;
+use std::mem;
 
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
@@ -26,14 +25,15 @@ use crate::base::dimension::{Dim, DimAdd, DimSum, IsNotStaticOne, U1, U2, U3};
 use crate::base::iter::{
     ColumnIter, ColumnIterMut, MatrixIter, MatrixIterMut, RowIter, RowIterMut,
 };
-use crate::base::storage::{
-    ContiguousStorage, ContiguousStorageMut, SameShapeStorage, Storage, StorageMut,
-};
+use crate::base::storage::{Owned, RawStorage, RawStorageMut, SameShapeStorage};
 use crate::base::{Const, DefaultAllocator, OMatrix, OVector, Scalar, Unit};
-use crate::{ArrayStorage, MatrixSlice, MatrixSliceMut, SMatrix, SimdComplexField};
+use crate::{ArrayStorage, SMatrix, SimdComplexField, Storage, UninitMatrix};
 
+use crate::storage::IsContiguous;
+use crate::uninit::{Init, InitStatus, Uninit};
 #[cfg(any(feature = "std", feature = "alloc"))]
 use crate::{DMatrix, DVector, Dynamic, VecStorage};
+use std::mem::MaybeUninit;
 
 /// A square matrix.
 pub type SquareMatrix<T, D, S> = Matrix<T, D, D, S>;
@@ -152,8 +152,8 @@ pub type MatrixCross<T, R1, C1, R2, C2> =
 /// Note that mixing `Dynamic` with type-level unsigned integers is allowed. Actually, a
 /// dynamically-sized column vector should be represented as a `Matrix<T, Dynamic, U1, S>` (given
 /// some concrete types for `T` and a compatible data storage type `S`).
-#[repr(transparent)]
-#[derive(Clone, Copy, Debug)]
+#[repr(C)]
+#[derive(Clone, Copy)]
 pub struct Matrix<T, R, C, S> {
     /// The data storage that contains all the matrix components. Disappointed?
     ///
@@ -187,23 +187,44 @@ pub struct Matrix<T, R, C, S> {
     //       from_data_statically_unchecked.
     //       Note that it would probably make sense to just have
     //       the type `Matrix<S>`, and have `T, R, C` be associated-types
-    //       of the `Storage` trait. However, because we don't have
-    //       specialization, this is not possible because these `T, R, C`
-    //       allows us to disambiguate a lot of configurations.
+    //       of the `RawStorage` trait. However, because we don't have
+    //       specialization, this is not bossible because these `T, R, C`
+    //       allows us to desambiguate a lot of configurations.
     _phantoms: PhantomData<(T, R, C)>,
 }
 
-impl<T, R: Dim, C: Dim, S> Default for Matrix<T, R, C, S>
+impl<T, R: Dim, C: Dim, S: fmt::Debug> fmt::Debug for Matrix<T, R, C, S> {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> Result<(), fmt::Error> {
+        formatter
+            .debug_struct("Matrix")
+            .field("data", &self.data)
+            .finish()
+    }
+}
+
+impl<T, R, C, S> Default for Matrix<T, R, C, S>
 where
-    S: Storage<T, R, C> + Default,
+    T: Scalar,
+    R: Dim,
+    C: Dim,
+    S: Default,
 {
     fn default() -> Self {
-        Matrix::from_data(Default::default())
+        Matrix {
+            data: Default::default(),
+            _phantoms: PhantomData,
+        }
     }
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<T, R: Dim, C: Dim, S: Serialize> Serialize for Matrix<T, R, C, S> {
+impl<T, R, C, S> Serialize for Matrix<T, R, C, S>
+where
+    T: Scalar,
+    R: Dim,
+    C: Dim,
+    S: Serialize,
+{
     fn serialize<Ser>(&self, serializer: Ser) -> Result<Ser::Ok, Ser::Error>
     where
         Ser: Serializer,
@@ -213,7 +234,13 @@ impl<T, R: Dim, C: Dim, S: Serialize> Serialize for Matrix<T, R, C, S> {
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<'de, T, R: Dim, C, S: Serialize<'de>> Deserialize<'de> for Matrix<T, R, C, S> {
+impl<'de, T, R, C, S> Deserialize<'de> for Matrix<T, R, C, S>
+where
+    T: Scalar,
+    R: Dim,
+    C: Dim,
+    S: Deserialize<'de>,
+{
     fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
     where
         D: Deserializer<'de>,
@@ -226,7 +253,7 @@ impl<'de, T, R: Dim, C, S: Serialize<'de>> Deserialize<'de> for Matrix<T, R, C,
 }
 
 #[cfg(feature = "abomonation-serialize")]
-impl<T, R: Dim, C: Dim, S: Abomonation> Abomonation for Matrix<T, R, C, S> {
+impl<T: Scalar, R: Dim, C: Dim, S: Abomonation> Abomonation for Matrix<T, R, C, S> {
     unsafe fn entomb<W: Write>(&self, writer: &mut W) -> IOResult<()> {
         self.data.entomb(writer)
     }
@@ -241,7 +268,7 @@ impl<T, R: Dim, C: Dim, S: Abomonation> Abomonation for Matrix<T, R, C, S> {
 }
 
 #[cfg(feature = "compare")]
-impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> matrixcompare_core::Matrix<T>
+impl<T: Scalar, R: Dim, C: Dim, S: RawStorage<T, R, C>> matrixcompare_core::Matrix<T>
     for Matrix<T, R, C, S>
 {
     fn rows(&self) -> usize {
@@ -258,7 +285,7 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> matrixcompare_core::Matrix<T
 }
 
 #[cfg(feature = "compare")]
-impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> matrixcompare_core::DenseAccess<T>
+impl<T: Scalar, R: Dim, C: Dim, S: RawStorage<T, R, C>> matrixcompare_core::DenseAccess<T>
     for Matrix<T, R, C, S>
 {
     fn fetch_single(&self, row: usize, col: usize) -> T {
@@ -267,13 +294,15 @@ impl<T: Clone, R: Dim, C: Dim, S: Storage<T, R, C>> matrixcompare_core::DenseAcc
 }
 
 #[cfg(feature = "bytemuck")]
-unsafe impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> bytemuck::Zeroable for Matrix<T, R, C, S> where
-    S: bytemuck::Zeroable
+unsafe impl<T: Scalar, R: Dim, C: Dim, S: RawStorage<T, R, C>> bytemuck::Zeroable
+    for Matrix<T, R, C, S>
+where
+    S: bytemuck::Zeroable,
 {
 }
 
 #[cfg(feature = "bytemuck")]
-unsafe impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> bytemuck::Pod for Matrix<T, R, C, S>
+unsafe impl<T: Scalar, R: Dim, C: Dim, S: RawStorage<T, R, C>> bytemuck::Pod for Matrix<T, R, C, S>
 where
     S: bytemuck::Pod,
     Self: Copy,
@@ -294,7 +323,7 @@ mod rkyv_impl {
             &self,
             pos: usize,
             resolver: Self::Resolver,
-            out: &mut core::mem::MaybeUninit<Self::Archived>,
+            out: &mut core::meme::MaybeUninit<Self::Archived>,
         ) {
             self.data.resolve(
                 pos + offset_of!(Self::Archived, data),
@@ -328,19 +357,8 @@ mod rkyv_impl {
 }
 
 impl<T, R, C, S> Matrix<T, R, C, S> {
-    /// Creates a new matrix with the given data without statically checking
-    /// that the matrix dimension matches the storage dimension.
-    ///
-    /// There's only two instances in which you should use this method instead
-    /// of the safe counterpart [`from_data`]:
-    /// - You can't get the type checker to validate your matrices, even though
-    ///   you're **certain** that they're of the right dimensions.
-    /// - You want to declare a matrix in a `const` context.
-    ///
-    /// # Safety
-    /// If the storage dimension does not match the matrix dimension, any other
-    /// method called on this matrix may behave erroneously, panic, or cause
-    /// Undefined Behavior.
+    /// Creates a new matrix with the given data without statically checking that the matrix
+    /// dimension matches the storage dimension.
     #[inline(always)]
     pub const unsafe fn from_data_statically_unchecked(data: S) -> Matrix<T, R, C, S> {
         Matrix {
@@ -350,29 +368,50 @@ impl<T, R, C, S> Matrix<T, R, C, S> {
     }
 }
 
-/// # Memory manipulation methods.
-impl<T, R: Dim, C: Dim> OMatrix<T, R, C>
-where
-    DefaultAllocator: Allocator<T, R, C>,
-{
-    /// Allocates a matrix with the given number of rows and columns without initializing its content.
-    pub fn new_uninitialized_generic(nrows: R, ncols: C) -> OMatrix<MaybeUninit<T>, R, C> {
-        OMatrix::from_data(
-            <DefaultAllocator as Allocator<T, R, C>>::allocate_uninitialized(nrows, ncols),
-        )
-    }
-
-    /// Converts this matrix into one whose entries need to be manually dropped. This should be
-    /// near zero-cost.
-    pub fn manually_drop(self) -> OMatrix<ManuallyDrop<T>, R, C> {
-        OMatrix::from_data(<DefaultAllocator as Allocator<T, R, C>>::manually_drop(
-            self.data,
-        ))
+impl<T, const R: usize, const C: usize> SMatrix<T, R, C> {
+    /// Creates a new statically-allocated matrix from the given [`ArrayStorage`].
+    ///
+    /// This method exists primarily as a workaround for the fact that `from_data` can not
+    /// work in `const fn` contexts.
+    #[inline(always)]
+    pub const fn from_array_storage(storage: ArrayStorage<T, R, C>) -> Self {
+        // This is sound because the row and column types are exactly the same as that of the
+        // storage, so there can be no mismatch
+        unsafe { Self::from_data_statically_unchecked(storage) }
     }
 }
 
-/// # More memory manipulation methods.
-impl<T, R: Dim, C: Dim> OMatrix<MaybeUninit<T>, R, C>
+// TODO: Consider removing/deprecating `from_vec_storage` once we are able to make
+// `from_data` const fn compatible
+#[cfg(any(feature = "std", feature = "alloc"))]
+impl<T> DMatrix<T> {
+    /// Creates a new heap-allocated matrix from the given [`VecStorage`].
+    ///
+    /// This method exists primarily as a workaround for the fact that `from_data` can not
+    /// work in `const fn` contexts.
+    pub const fn from_vec_storage(storage: VecStorage<T, Dynamic, Dynamic>) -> Self {
+        // This is sound because the dimensions of the matrix and the storage are guaranteed
+        // to be the same
+        unsafe { Self::from_data_statically_unchecked(storage) }
+    }
+}
+
+// TODO: Consider removing/deprecating `from_vec_storage` once we are able to make
+// `from_data` const fn compatible
+#[cfg(any(feature = "std", feature = "alloc"))]
+impl<T> DVector<T> {
+    /// Creates a new heap-allocated matrix from the given [`VecStorage`].
+    ///
+    /// This method exists primarily as a workaround for the fact that `from_data` can not
+    /// work in `const fn` contexts.
+    pub const fn from_vec_storage(storage: VecStorage<T, Dynamic, U1>) -> Self {
+        // This is sound because the dimensions of the matrix and the storage are guaranteed
+        // to be the same
+        unsafe { Self::from_data_statically_unchecked(storage) }
+    }
+}
+
+impl<T, R: Dim, C: Dim> UninitMatrix<T, R, C>
 where
     DefaultAllocator: Allocator<T, R, C>,
 {
@@ -388,100 +427,29 @@ where
             self.data,
         ))
     }
-
-    /// Assumes a matrix's entries to be initialized, and drops them in place.
-    /// This allows the buffer to be safely reused.
-    ///
-    /// # Safety
-    /// All of the matrix's entries need to be uninitialized. Otherwise,
-    /// Undefined Behavior will be triggered.
-    pub unsafe fn reinitialize(&mut self) {
-        for i in 0..self.nrows() {
-            for j in 0..self.ncols() {
-                ptr::drop_in_place(self.get_unchecked_mut((i, j)));
-            }
-        }
-    }
 }
 
-impl<T, R: Dim, C: Dim, S> Matrix<MaybeUninit<T>, R, C, S> {
-    /// Creates a full slice from `self` and assumes it to be initialized.
-    ///
-    /// # Safety
-    /// The user must make sure that every single entry of the buffer has been initialized,
-    /// or Undefined Behavior will immediately occur.    
-    pub unsafe fn assume_init_ref(&self) -> MatrixSlice<T, R, C, S::RStride, S::CStride>
-    where
-        S: Storage<MaybeUninit<T>, R, C>,
-    {
-        self.full_slice().slice_assume_init()
-    }
-
-    /// Creates a full mutable slice from `self` and assumes it to be initialized.
-    ///
-    /// # Safety
-    /// The user must make sure that every single entry of the buffer has been initialized,
-    /// or Undefined Behavior will immediately occur.    
-    pub unsafe fn assume_init_mut(&mut self) -> MatrixSliceMut<T, R, C, S::RStride, S::CStride>
-    where
-        S: StorageMut<MaybeUninit<T>, R, C>,
-    {
-        self.full_slice_mut().slice_assume_init()
-    }
-}
-
-impl<T, const R: usize, const C: usize> SMatrix<T, R, C> {
-    /// Creates a new statically-allocated matrix from the given [`ArrayStorage`].
-    ///
-    /// This method exists primarily as a workaround for the fact that `from_data` can not
-    /// work in `const fn` contexts.
-    #[inline(always)]
-    pub const fn from_array_storage(storage: ArrayStorage<T, R, C>) -> Self {
-        // Safety: This is sound because the row and column types are exactly
-        // the same as that of the storage, so there can be no mismatch.
-        unsafe { Self::from_data_statically_unchecked(storage) }
-    }
-}
-
-// TODO: Consider removing/deprecating `from_vec_storage` once we are able to make
-// `from_data` const fn compatible
-#[cfg(any(feature = "std", feature = "alloc"))]
-impl<T> DMatrix<T> {
-    /// Creates a new heap-allocated matrix from the given [`VecStorage`].
-    ///
-    /// This method exists primarily as a workaround for the fact that `from_data` can not
-    /// work in `const fn` contexts.
-    pub const fn from_vec_storage(storage: VecStorage<T, Dynamic, Dynamic>) -> Self {
-        // Safety: This is sound because the dimensions of the matrix and the
-        // storage are guaranteed to be the same.
-        unsafe { Self::from_data_statically_unchecked(storage) }
-    }
-}
-
-// TODO: Consider removing/deprecating `from_vec_storage` once we are able to make
-// `from_data` const fn compatible
-#[cfg(any(feature = "std", feature = "alloc"))]
-impl<T> DVector<T> {
-    /// Creates a new heap-allocated matrix from the given [`VecStorage`].
-    ///
-    /// This method exists primarily as a workaround for the fact that `from_data` can not
-    /// work in `const fn` contexts.
-    pub const fn from_vec_storage(storage: VecStorage<T, Dynamic, U1>) -> Self {
-        // Safety: This is sound because the dimensions of the matrix and the
-        // storage are guaranteed to be the same.
-        unsafe { Self::from_data_statically_unchecked(storage) }
-    }
-}
-
-impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     /// Creates a new matrix with the given data.
     #[inline(always)]
     pub fn from_data(data: S) -> Self {
-        // Safety: This is sound because the dimensions of the matrix and the
-        // storage are guaranteed to be the same.
         unsafe { Self::from_data_statically_unchecked(data) }
     }
 
+    /// Creates a new uninitialized matrix with the given uninitialized data
+    pub unsafe fn from_uninitialized_data(data: MaybeUninit<S>) -> MaybeUninit<Self> {
+        let res: Matrix<T, R, C, MaybeUninit<S>> = Matrix {
+            data,
+            _phantoms: PhantomData,
+        };
+        let res: MaybeUninit<Matrix<T, R, C, MaybeUninit<S>>> = MaybeUninit::new(res);
+        // safety: since we wrap the inner MaybeUninit in an outer MaybeUninit above, the fact that the `data` field is partially-uninitialized is still opaque.
+        // with s/transmute_copy/transmute/, rustc claims that `MaybeUninit<Matrix<T, R, C, MaybeUninit<S>>>` may be of a different size from `MaybeUninit<Matrix<T, R, C, S>>`
+        // but MaybeUninit's documentation says "MaybeUninit<T> is guaranteed to have the same size, alignment, and ABI as T", which implies those types should be the same size
+        let res: MaybeUninit<Matrix<T, R, C, S>> = mem::transmute_copy(&res);
+        res
+    }
+
     /// The shape of this matrix returned as the tuple (number of rows, number of columns).
     ///
     /// # Examples:
@@ -493,10 +461,16 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     #[must_use]
     pub fn shape(&self) -> (usize, usize) {
-        let (nrows, ncols) = self.data.shape();
+        let (nrows, ncols) = self.shape_generic();
         (nrows.value(), ncols.value())
     }
 
+    #[inline]
+    #[must_use]
+    pub fn shape_generic(&self) -> (R, C) {
+        self.data.shape()
+    }
+
     /// The number of rows of this matrix.
     ///
     /// # Examples:
@@ -535,7 +509,6 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     /// let slice = mat.slice_with_steps((0, 0), (5, 3), (1, 2));
     /// // The column strides is the number of steps (here 2) multiplied by the corresponding dimension.
     /// assert_eq!(mat.strides(), (1, 10));
-    /// ```
     #[inline]
     #[must_use]
     pub fn strides(&self) -> (usize, usize) {
@@ -595,7 +568,7 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     /// See `relative_eq` from the `RelativeEq` trait for more details.
     #[inline]
     #[must_use]
-    pub fn relative_eq<R2: Dim, C2: Dim, SB>(
+    pub fn relative_eq<R2, C2, SB>(
         &self,
         other: &Matrix<T, R2, C2, SB>,
         eps: T::Epsilon,
@@ -603,6 +576,8 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     ) -> bool
     where
         T: RelativeEq,
+        R2: Dim,
+        C2: Dim,
         SB: Storage<T, R2, C2>,
         T::Epsilon: Copy,
         ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
@@ -617,10 +592,12 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     #[must_use]
     #[allow(clippy::should_implement_trait)]
-    pub fn eq<R2: Dim, C2: Dim, SB>(&self, other: &Matrix<T, R2, C2, SB>) -> bool
+    pub fn eq<R2, C2, SB>(&self, other: &Matrix<T, R2, C2, SB>) -> bool
     where
         T: PartialEq,
-        SB: Storage<T, R2, C2>,
+        R2: Dim,
+        C2: Dim,
+        SB: RawStorage<T, R2, C2>,
         ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
     {
         assert!(self.shape() == other.shape());
@@ -631,10 +608,11 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn into_owned(self) -> OMatrix<T, R, C>
     where
-        T: Clone,
+        T: Scalar,
+        S: Storage<T, R, C>,
         DefaultAllocator: Allocator<T, R, C>,
     {
-        Matrix::from_data(self.data.into_owned().0)
+        Matrix::from_data(self.data.into_owned())
     }
 
     // TODO: this could probably benefit from specialization.
@@ -642,24 +620,24 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     /// Moves this matrix into one that owns its data. The actual type of the result depends on
     /// matrix storage combination rules for addition.
     #[inline]
-    pub fn into_owned_sum<R2: Dim, C2: Dim>(self) -> MatrixSum<T, R, C, R2, C2>
+    pub fn into_owned_sum<R2, C2>(self) -> MatrixSum<T, R, C, R2, C2>
     where
-        T: Clone,
+        T: Scalar,
+        S: Storage<T, R, C>,
+        R2: Dim,
+        C2: Dim,
         DefaultAllocator: SameShapeAllocator<T, R, C, R2, C2>,
         ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
     {
-        // If both storages are the same, we can just return `self.into_owned()`.
-        // Unfortunately, it's not trivial to convince the compiler of this.
-        if TypeId::of::<SameShapeR<R, R2>>() == TypeId::of::<R>()
-            && TypeId::of::<SameShapeC<C, C2>>() == TypeId::of::<C>()
-        {
-            // Safety: we're transmuting from a type into itself, and we make
-            // sure not to leak anything.
+        if TypeId::of::<SameShapeStorage<T, R, C, R2, C2>>() == TypeId::of::<Owned<T, R, C>>() {
+            // We can just return `self.into_owned()`.
+
             unsafe {
-                let mat = self.into_owned();
-                let mat_copy = mem::transmute_copy(&mat);
-                mem::forget(mat);
-                mat_copy
+                // TODO: check that those copies are optimized away by the compiler.
+                let owned = self.into_owned();
+                let res = mem::transmute_copy(&owned);
+                mem::forget(owned);
+                res
             }
         } else {
             self.clone_owned_sum()
@@ -671,19 +649,23 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     #[must_use]
     pub fn clone_owned(&self) -> OMatrix<T, R, C>
     where
-        T: Clone,
+        T: Scalar,
+        S: Storage<T, R, C>,
         DefaultAllocator: Allocator<T, R, C>,
     {
-        Matrix::from_data(self.data.clone_owned().0)
+        Matrix::from_data(self.data.clone_owned())
     }
 
     /// Clones this matrix into one that owns its data. The actual type of the result depends on
     /// matrix storage combination rules for addition.
     #[inline]
     #[must_use]
-    pub fn clone_owned_sum<R2: Dim, C2: Dim>(&self) -> MatrixSum<T, R, C, R2, C2>
+    pub fn clone_owned_sum<R2, C2>(&self) -> MatrixSum<T, R, C, R2, C2>
     where
-        T: Clone,
+        T: Scalar,
+        S: Storage<T, R, C>,
+        R2: Dim,
+        C2: Dim,
         DefaultAllocator: SameShapeAllocator<T, R, C, R2, C2>,
         ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
     {
@@ -691,122 +673,110 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         let nrows: SameShapeR<R, R2> = Dim::from_usize(nrows);
         let ncols: SameShapeC<C, C2> = Dim::from_usize(ncols);
 
-        let mut res = Matrix::new_uninitialized_generic(nrows, ncols);
+        let mut res = Matrix::uninit(nrows, ncols);
 
-        // TODO: use copy_from
-        for j in 0..res.ncols() {
-            for i in 0..res.nrows() {
-                unsafe {
+        unsafe {
+            // TODO: use copy_from?
+            for j in 0..res.ncols() {
+                for i in 0..res.nrows() {
                     *res.get_unchecked_mut((i, j)) =
-                        MaybeUninit::new(self.get_unchecked((i, j)).clone());
+                        MaybeUninit::new(self.get_unchecked((i, j)).inlined_clone());
                 }
             }
-        }
 
-        unsafe { res.assume_init() }
+            // SAFETY: the output has been initialized above.
+            res.assume_init()
+        }
     }
 
-    /// Transposes `self` and store the result into `out`, which will become
-    /// fully initialized.
+    /// Transposes `self` and store the result into `out`.
     #[inline]
-    pub fn transpose_to<R2: Dim, C2: Dim, SB>(&self, out: &mut Matrix<MaybeUninit<T>, R2, C2, SB>)
-    where
-        T: Clone,
-        SB: StorageMut<MaybeUninit<T>, R2, C2>,
+    fn transpose_to_uninit<Status, R2, C2, SB>(
+        &self,
+        status: Status,
+        out: &mut Matrix<Status::Value, R2, C2, SB>,
+    ) where
+        Status: InitStatus<T>,
+        T: Scalar,
+        R2: Dim,
+        C2: Dim,
+        SB: RawStorageMut<Status::Value, R2, C2>,
         ShapeConstraint: SameNumberOfRows<R, C2> + SameNumberOfColumns<C, R2>,
     {
         let (nrows, ncols) = self.shape();
         assert!(
             (ncols, nrows) == out.shape(),
-            "Incompatible shape for transpose-copy."
+            "Incompatible shape for transposition."
         );
 
         // TODO: optimize that.
         for i in 0..nrows {
             for j in 0..ncols {
+                // Safety: the indices are in range.
                 unsafe {
-                    *out.get_unchecked_mut((j, i)) =
-                        MaybeUninit::new(self.get_unchecked((i, j)).clone());
+                    Status::init(
+                        out.get_unchecked_mut((j, i)),
+                        self.get_unchecked((i, j)).inlined_clone(),
+                    );
                 }
             }
         }
     }
 
+    /// Transposes `self` and store the result into `out`.
+    #[inline]
+    pub fn transpose_to<R2, C2, SB>(&self, out: &mut Matrix<T, R2, C2, SB>)
+    where
+        T: Scalar,
+        R2: Dim,
+        C2: Dim,
+        SB: RawStorageMut<T, R2, C2>,
+        ShapeConstraint: SameNumberOfRows<R, C2> + SameNumberOfColumns<C, R2>,
+    {
+        self.transpose_to_uninit(Init, out)
+    }
+
     /// Transposes `self`.
     #[inline]
     #[must_use = "Did you mean to use transpose_mut()?"]
     pub fn transpose(&self) -> OMatrix<T, C, R>
     where
-        T: Clone,
+        T: Scalar,
         DefaultAllocator: Allocator<T, C, R>,
     {
-        let (nrows, ncols) = self.data.shape();
-        let mut res = OMatrix::new_uninitialized_generic(ncols, nrows);
-        self.transpose_to(&mut res);
+        let (nrows, ncols) = self.shape_generic();
 
-        unsafe {
-            // Safety: res is now fully initialized due to the guarantees of transpose_to.
-            res.assume_init()
-        }
-    }
-}
-
-impl<T, R: Dim, C: Dim> OMatrix<T, R, C>
-where
-    DefaultAllocator: Allocator<T, C, R> + Allocator<T, R, C>,
-{
-    /// Transposes `self`. Does not require `T: Clone` like its other counterparts.
-    pub fn transpose_into(self) -> OMatrix<T, C, R> {
-        let (nrows, ncols) = self.data.shape();
-        let mut res = OMatrix::new_uninitialized_generic(ncols, nrows);
-        let mut md = self.manually_drop();
-
-        let (nrows, ncols) = res.shape();
-
-        // TODO: optimize that.
-        for i in 0..nrows {
-            for j in 0..ncols {
-                // Safety: the indices are within range, and since the indices
-                // don't repeat, we don't do any double-drops.
-                unsafe {
-                    *res.get_unchecked_mut((j, i)) =
-                        MaybeUninit::new(ManuallyDrop::take(md.get_unchecked_mut((i, j))));
-                }
-            }
-        }
-
-        unsafe {
-            // Safety: res is now fully initialized, since we've initialized
-            // every single entry.
-            res.assume_init()
-        }
+        let mut res = Matrix::uninit(ncols, nrows);
+        self.transpose_to_uninit(Uninit, &mut res);
+        // Safety: res is now fully initialized.
+        unsafe { res.assume_init() }
     }
 }
 
 /// # Elementwise mapping and folding
-// Todo: maybe make ref versions of these methods that can be used when T is expensive to clone?
-impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     /// Returns a matrix containing the result of `f` applied to each of its entries.
     #[inline]
     #[must_use]
-    pub fn map<T2, F: FnMut(T) -> T2>(&self, mut f: F) -> OMatrix<T2, R, C>
+    pub fn map<T2: Scalar, F: FnMut(T) -> T2>(&self, mut f: F) -> OMatrix<T2, R, C>
     where
-        T: Clone,
+        T: Scalar,
         DefaultAllocator: Allocator<T2, R, C>,
     {
-        let (nrows, ncols) = self.data.shape();
-        let mut res = OMatrix::new_uninitialized_generic(nrows, ncols);
+        let (nrows, ncols) = self.shape_generic();
+        let mut res = Matrix::uninit(nrows, ncols);
 
         for j in 0..ncols.value() {
             for i in 0..nrows.value() {
+                // Safety: all indices are in range.
                 unsafe {
-                    let a = self.data.get_unchecked(i, j).clone();
+                    let a = self.data.get_unchecked(i, j).inlined_clone();
                     *res.data.get_unchecked_mut(i, j) = MaybeUninit::new(f(a));
                 }
             }
         }
 
-        // Safety: all entries have been initialized.
+        // Safety: res is now fully initialized.
         unsafe { res.assume_init() }
     }
 
@@ -819,8 +789,9 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     /// let q2 = q.cast::<f32>();
     /// assert_eq!(q2, Vector3::new(1.0f32, 2.0, 3.0));
     /// ```
-    pub fn cast<T2>(self) -> OMatrix<T2, R, C>
+    pub fn cast<T2: Scalar>(self) -> OMatrix<T2, R, C>
     where
+        T: Scalar,
         OMatrix<T2, R, C>: SupersetOf<Self>,
         DefaultAllocator: Allocator<T2, R, C>,
     {
@@ -840,7 +811,10 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         &self,
         init_f: impl FnOnce(Option<&T>) -> T2,
         f: impl FnMut(T2, &T) -> T2,
-    ) -> T2 {
+    ) -> T2
+    where
+        T: Scalar,
+    {
         let mut it = self.iter();
         let init = init_f(it.next());
         it.fold(init, f)
@@ -850,28 +824,28 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     /// `f` also gets passed the row and column index, i.e. `f(row, col, value)`.
     #[inline]
     #[must_use]
-    pub fn map_with_location<T2: Clone, F: FnMut(usize, usize, T) -> T2>(
+    pub fn map_with_location<T2: Scalar, F: FnMut(usize, usize, T) -> T2>(
         &self,
         mut f: F,
     ) -> OMatrix<T2, R, C>
     where
-        T: Clone,
+        T: Scalar,
         DefaultAllocator: Allocator<T2, R, C>,
     {
-        let (nrows, ncols) = self.data.shape();
-
-        let mut res = OMatrix::new_uninitialized_generic(nrows, ncols);
+        let (nrows, ncols) = self.shape_generic();
+        let mut res = Matrix::uninit(nrows, ncols);
 
         for j in 0..ncols.value() {
             for i in 0..nrows.value() {
+                // Safety: all indices are in range.
                 unsafe {
-                    let a = self.data.get_unchecked(i, j).clone();
+                    let a = self.data.get_unchecked(i, j).inlined_clone();
                     *res.data.get_unchecked_mut(i, j) = MaybeUninit::new(f(i, j, a));
                 }
             }
         }
 
-        // Safety: all entries have been initialized.
+        // Safety: res is now fully initialized.
         unsafe { res.assume_init() }
     }
 
@@ -879,20 +853,17 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     /// `rhs`.
     #[inline]
     #[must_use]
-    pub fn zip_map<T2: Clone, N3, S2, F>(
-        &self,
-        rhs: &Matrix<T2, R, C, S2>,
-        mut f: F,
-    ) -> OMatrix<N3, R, C>
+    pub fn zip_map<T2, N3, S2, F>(&self, rhs: &Matrix<T2, R, C, S2>, mut f: F) -> OMatrix<N3, R, C>
     where
-        T: Clone,
-        S2: Storage<T2, R, C>,
+        T: Scalar,
+        T2: Scalar,
+        N3: Scalar,
+        S2: RawStorage<T2, R, C>,
         F: FnMut(T, T2) -> N3,
         DefaultAllocator: Allocator<N3, R, C>,
     {
-        let (nrows, ncols) = self.data.shape();
-
-        let mut res = OMatrix::new_uninitialized_generic(nrows, ncols);
+        let (nrows, ncols) = self.shape_generic();
+        let mut res = Matrix::uninit(nrows, ncols);
 
         assert_eq!(
             (nrows.value(), ncols.value()),
@@ -902,15 +873,16 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
 
         for j in 0..ncols.value() {
             for i in 0..nrows.value() {
+                // Safety: all indices are in range.
                 unsafe {
-                    let a = self.data.get_unchecked(i, j).clone();
-                    let b = rhs.data.get_unchecked(i, j).clone();
-                    *res.data.get_unchecked_mut(i, j) = MaybeUninit::new(f(a, b));
+                    let a = self.data.get_unchecked(i, j).inlined_clone();
+                    let b = rhs.data.get_unchecked(i, j).inlined_clone();
+                    *res.data.get_unchecked_mut(i, j) = MaybeUninit::new(f(a, b))
                 }
             }
         }
 
-        // Safety: all entries have been initialized.
+        // Safety: res is now fully initialized.
         unsafe { res.assume_init() }
     }
 
@@ -918,22 +890,24 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     /// `b`, and `c`.
     #[inline]
     #[must_use]
-    pub fn zip_zip_map<T2: Clone, N3: Clone, N4, S2, S3, F>(
+    pub fn zip_zip_map<T2, N3, N4, S2, S3, F>(
         &self,
         b: &Matrix<T2, R, C, S2>,
         c: &Matrix<N3, R, C, S3>,
         mut f: F,
     ) -> OMatrix<N4, R, C>
     where
-        T: Clone,
-        S2: Storage<T2, R, C>,
-        S3: Storage<N3, R, C>,
+        T: Scalar,
+        T2: Scalar,
+        N3: Scalar,
+        N4: Scalar,
+        S2: RawStorage<T2, R, C>,
+        S3: RawStorage<N3, R, C>,
         F: FnMut(T, T2, N3) -> N4,
         DefaultAllocator: Allocator<N4, R, C>,
     {
-        let (nrows, ncols) = self.data.shape();
-
-        let mut res = OMatrix::new_uninitialized_generic(nrows, ncols);
+        let (nrows, ncols) = self.shape_generic();
+        let mut res = Matrix::uninit(nrows, ncols);
 
         assert_eq!(
             (nrows.value(), ncols.value()),
@@ -948,55 +922,64 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
 
         for j in 0..ncols.value() {
             for i in 0..nrows.value() {
+                // Safety: all indices are in range.
                 unsafe {
-                    let a = self.data.get_unchecked(i, j).clone();
-                    let b = b.data.get_unchecked(i, j).clone();
-                    let c = c.data.get_unchecked(i, j).clone();
-                    *res.data.get_unchecked_mut(i, j) = MaybeUninit::new(f(a, b, c));
+                    let a = self.data.get_unchecked(i, j).inlined_clone();
+                    let b = b.data.get_unchecked(i, j).inlined_clone();
+                    let c = c.data.get_unchecked(i, j).inlined_clone();
+                    *res.data.get_unchecked_mut(i, j) = MaybeUninit::new(f(a, b, c))
                 }
             }
         }
 
-        // Safety: all entries have been initialized.
+        // Safety: res is now fully initialized.
         unsafe { res.assume_init() }
     }
 
     /// Folds a function `f` on each entry of `self`.
     #[inline]
     #[must_use]
-    pub fn fold<Acc>(&self, mut init: Acc, mut f: impl FnMut(Acc, T) -> Acc) -> Acc
+    pub fn fold<Acc>(&self, init: Acc, mut f: impl FnMut(Acc, T) -> Acc) -> Acc
     where
-        T: Clone,
+        T: Scalar,
     {
-        let (nrows, ncols) = self.data.shape();
+        let (nrows, ncols) = self.shape_generic();
+
+        let mut res = init;
 
         for j in 0..ncols.value() {
             for i in 0..nrows.value() {
+                // Safety: all indices are in range.
                 unsafe {
-                    let a = self.data.get_unchecked(i, j).clone();
-                    init = f(init, a)
+                    let a = self.data.get_unchecked(i, j).inlined_clone();
+                    res = f(res, a)
                 }
             }
         }
 
-        init
+        res
     }
 
     /// Folds a function `f` on each pairs of entries from `self` and `rhs`.
     #[inline]
     #[must_use]
-    pub fn zip_fold<T2: Clone, R2: Dim, C2: Dim, S2, Acc>(
+    pub fn zip_fold<T2, R2, C2, S2, Acc>(
         &self,
         rhs: &Matrix<T2, R2, C2, S2>,
-        mut init: Acc,
+        init: Acc,
         mut f: impl FnMut(Acc, T, T2) -> Acc,
     ) -> Acc
     where
-        T: Clone,
-        S2: Storage<T2, R2, C2>,
+        T: Scalar,
+        T2: Scalar,
+        R2: Dim,
+        C2: Dim,
+        S2: RawStorage<T2, R2, C2>,
         ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
     {
-        let (nrows, ncols) = self.data.shape();
+        let (nrows, ncols) = self.shape_generic();
+
+        let mut res = init;
 
         assert_eq!(
             (nrows.value(), ncols.value()),
@@ -1007,22 +990,21 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         for j in 0..ncols.value() {
             for i in 0..nrows.value() {
                 unsafe {
-                    let a = self.data.get_unchecked(i, j).clone();
-                    let b = rhs.data.get_unchecked(i, j).clone();
-                    init = f(init, a, b)
+                    let a = self.data.get_unchecked(i, j).inlined_clone();
+                    let b = rhs.data.get_unchecked(i, j).inlined_clone();
+                    res = f(res, a, b)
                 }
             }
         }
 
-        init
+        res
     }
 
-    /// Replaces each component of `self` by the result of a closure `f` applied on it.
+    /// Applies a closure `f` to modify each component of `self`.
     #[inline]
-    pub fn apply<F: FnMut(T) -> T>(&mut self, mut f: F)
+    pub fn apply<F: FnMut(&mut T)>(&mut self, mut f: F)
     where
-        T: Clone, // This could be removed by changing the function signature.
-        S: StorageMut<T, R, C>,
+        S: RawStorageMut<T, R, C>,
     {
         let (nrows, ncols) = self.shape();
 
@@ -1030,7 +1012,7 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
             for i in 0..nrows {
                 unsafe {
                     let e = self.data.get_unchecked_mut(i, j);
-                    *e = f(e.clone())
+                    f(e)
                 }
             }
         }
@@ -1039,14 +1021,16 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     /// Replaces each component of `self` by the result of a closure `f` applied on its components
     /// joined with the components from `rhs`.
     #[inline]
-    pub fn zip_apply<T2: Clone, R2: Dim, C2: Dim, S2>(
+    pub fn zip_apply<T2, R2, C2, S2>(
         &mut self,
         rhs: &Matrix<T2, R2, C2, S2>,
-        mut f: impl FnMut(T, T2) -> T,
+        mut f: impl FnMut(&mut T, T2),
     ) where
-        T: Clone, // This could be removed by changing the function signature.
-        S: StorageMut<T, R, C>,
-        S2: Storage<T2, R2, C2>,
+        S: RawStorageMut<T, R, C>,
+        T2: Scalar,
+        R2: Dim,
+        C2: Dim,
+        S2: RawStorage<T2, R2, C2>,
         ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
     {
         let (nrows, ncols) = self.shape();
@@ -1061,8 +1045,8 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
             for i in 0..nrows {
                 unsafe {
                     let e = self.data.get_unchecked_mut(i, j);
-                    let rhs = rhs.get_unchecked((i, j)).clone();
-                    *e = f(e.clone(), rhs)
+                    let rhs = rhs.get_unchecked((i, j)).inlined_clone();
+                    f(e, rhs)
                 }
             }
         }
@@ -1071,16 +1055,21 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     /// Replaces each component of `self` by the result of a closure `f` applied on its components
     /// joined with the components from `b` and `c`.
     #[inline]
-    pub fn zip_zip_apply<T2: Clone, R2: Dim, C2: Dim, S2, N3: Clone, R3: Dim, C3: Dim, S3>(
+    pub fn zip_zip_apply<T2, R2, C2, S2, N3, R3, C3, S3>(
         &mut self,
         b: &Matrix<T2, R2, C2, S2>,
         c: &Matrix<N3, R3, C3, S3>,
-        mut f: impl FnMut(T, T2, N3) -> T,
+        mut f: impl FnMut(&mut T, T2, N3),
     ) where
-        T: Clone, // This could be removed by changing the function signature.
-        S: StorageMut<T, R, C>,
-        S2: Storage<T2, R2, C2>,
-        S3: Storage<N3, R3, C3>,
+        S: RawStorageMut<T, R, C>,
+        T2: Scalar,
+        R2: Dim,
+        C2: Dim,
+        S2: RawStorage<T2, R2, C2>,
+        N3: Scalar,
+        R3: Dim,
+        C3: Dim,
+        S3: RawStorage<N3, R3, C3>,
         ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
         ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
     {
@@ -1101,9 +1090,9 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
             for i in 0..nrows {
                 unsafe {
                     let e = self.data.get_unchecked_mut(i, j);
-                    let b = b.get_unchecked((i, j)).clone();
-                    let c = c.get_unchecked((i, j)).clone();
-                    *e = f(e.clone(), b, c)
+                    let b = b.get_unchecked((i, j)).inlined_clone();
+                    let c = c.get_unchecked((i, j)).inlined_clone();
+                    f(e, b, c)
                 }
             }
         }
@@ -1111,7 +1100,7 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
 }
 
 /// # Iteration on components, rows, and columns
-impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     /// Iterates through this matrix coordinates in column-major order.
     ///
     /// # Examples:
@@ -1168,7 +1157,7 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn iter_mut(&mut self) -> MatrixIterMut<'_, T, R, C, S>
     where
-        S: StorageMut<T, R, C>,
+        S: RawStorageMut<T, R, C>,
     {
         MatrixIterMut::new(&mut self.data)
     }
@@ -1191,7 +1180,7 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn row_iter_mut(&mut self) -> RowIterMut<'_, T, R, C, S>
     where
-        S: StorageMut<T, R, C>,
+        S: RawStorageMut<T, R, C>,
     {
         RowIterMut::new(self)
     }
@@ -1214,13 +1203,13 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     pub fn column_iter_mut(&mut self) -> ColumnIterMut<'_, T, R, C, S>
     where
-        S: StorageMut<T, R, C>,
+        S: RawStorageMut<T, R, C>,
     {
         ColumnIterMut::new(self)
     }
 }
 
-impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
+impl<T: Scalar, R: Dim, C: Dim, S: RawStorageMut<T, R, C>> Matrix<T, R, C, S> {
     /// Returns a mutable pointer to the start of the matrix.
     ///
     /// If the matrix is not empty, this pointer is guaranteed to be aligned
@@ -1257,10 +1246,7 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
     ///
     /// The components of the slice are assumed to be ordered in column-major order.
     #[inline]
-    pub fn copy_from_slice(&mut self, slice: &[T])
-    where
-        T: Clone,
-    {
+    pub fn copy_from_slice(&mut self, slice: &[T]) {
         let (nrows, ncols) = self.shape();
 
         assert!(
@@ -1271,34 +1257,21 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
         for j in 0..ncols {
             for i in 0..nrows {
                 unsafe {
-                    *self.get_unchecked_mut((i, j)) = slice.get_unchecked(i + j * nrows).clone();
+                    *self.get_unchecked_mut((i, j)) =
+                        slice.get_unchecked(i + j * nrows).inlined_clone();
                 }
             }
         }
     }
 
-    /// Fills this matrix with the content of another one via clones. Both must have the same shape.
+    /// Fills this matrix with the content of another one. Both must have the same shape.
     #[inline]
-    pub fn copy_from<R2: Dim, C2: Dim, SB>(&mut self, other: &Matrix<T, R2, C2, SB>)
+    pub fn copy_from<R2, C2, SB>(&mut self, other: &Matrix<T, R2, C2, SB>)
     where
-        T: Clone,
-        SB: Storage<T, R2, C2>,
+        R2: Dim,
+        C2: Dim,
+        SB: RawStorage<T, R2, C2>,
         ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
-    {
-        self.copy_from_fn(other, T::clone)
-    }
-
-    /// Fills this matrix with the content of another one, after applying a function to
-    /// the references of the entries of the other matrix. Both must have the same shape.
-    #[inline]
-    pub fn copy_from_fn<U, R2: Dim, C2: Dim, SB, F>(
-        &mut self,
-        other: &Matrix<U, R2, C2, SB>,
-        mut f: F,
-    ) where
-        SB: Storage<U, R2, C2>,
-        ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
-        F: FnMut(&U) -> T,
     {
         assert!(
             self.shape() == other.shape(),
@@ -1308,71 +1281,20 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
         for j in 0..self.ncols() {
             for i in 0..self.nrows() {
                 unsafe {
-                    *self.get_unchecked_mut((i, j)) = f(other.get_unchecked((i, j)));
+                    *self.get_unchecked_mut((i, j)) = other.get_unchecked((i, j)).inlined_clone();
                 }
             }
         }
     }
 
-    /// Fills this matrix with the content of another one, after applying a function to
-    /// the entries of the other matrix. Both must have the same shape.
+    /// Fills this matrix with the content of the transpose another one.
     #[inline]
-    pub fn move_from<R2: Dim, C2: Dim>(&mut self, other: OMatrix<T, R2, C2>)
+    pub fn tr_copy_from<R2, C2, SB>(&mut self, other: &Matrix<T, R2, C2, SB>)
     where
-        ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
-        DefaultAllocator: Allocator<T, R2, C2>,
-    {
-        self.move_from_fn(other, |e| e)
-    }
-
-    /// Fills this matrix with the content of another one via moves. Both must have the same shape.
-    #[inline]
-    pub fn move_from_fn<U, R2: Dim, C2: Dim, F>(&mut self, other: OMatrix<U, R2, C2>, mut f: F)
-    where
-        ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
-        DefaultAllocator: Allocator<U, R2, C2>,
-        F: FnMut(U) -> T,
-    {
-        assert!(
-            self.shape() == other.shape(),
-            "Unable to move from a matrix with a different shape."
-        );
-
-        let mut md = other.manually_drop();
-
-        for j in 0..self.ncols() {
-            for i in 0..self.nrows() {
-                unsafe {
-                    *self.get_unchecked_mut((i, j)) =
-                        f(ManuallyDrop::take(md.get_unchecked_mut((i, j))));
-                }
-            }
-        }
-    }
-
-    /// Fills this matrix with the content of the transpose another one via clones.
-    #[inline]
-    pub fn tr_copy_from<R2: Dim, C2: Dim, SB>(&mut self, other: &Matrix<T, R2, C2, SB>)
-    where
-        T: Clone,
-        SB: Storage<T, R2, C2>,
+        R2: Dim,
+        C2: Dim,
+        SB: RawStorage<T, R2, C2>,
         ShapeConstraint: DimEq<R, C2> + SameNumberOfColumns<C, R2>,
-    {
-        self.tr_copy_from_fn(other, T::clone)
-    }
-
-    /// Fills this matrix with the content of the transpose of another one, after applying
-    /// a function to the references of the entries of the other matrix. Both must have the
-    /// same shape.
-    #[inline]
-    pub fn tr_copy_from_fn<U, R2: Dim, C2: Dim, SB, F>(
-        &mut self,
-        other: &Matrix<U, R2, C2, SB>,
-        mut f: F,
-    ) where
-        SB: Storage<U, R2, C2>,
-        ShapeConstraint: DimEq<R, C2> + SameNumberOfColumns<C, R2>,
-        F: FnMut(&U) -> T,
     {
         let (nrows, ncols) = self.shape();
         assert!(
@@ -1383,44 +1305,7 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
         for j in 0..ncols {
             for i in 0..nrows {
                 unsafe {
-                    *self.get_unchecked_mut((i, j)) = f(other.get_unchecked((j, i)));
-                }
-            }
-        }
-    }
-
-    /// Fills this matrix with the content of the transpose another one via moves.
-    #[inline]
-    pub fn tr_move_from<R2: Dim, C2: Dim>(&mut self, other: OMatrix<T, R2, C2>)
-    where
-        DefaultAllocator: Allocator<T, R2, C2>,
-        ShapeConstraint: DimEq<R, C2> + SameNumberOfColumns<C, R2>,
-    {
-        self.tr_move_from_fn(other, |e| e)
-    }
-
-    /// Fills this matrix with the content of the transpose of another one, after applying
-    /// a function to the entries of the other matrix. Both must have the same shape.
-    #[inline]
-    pub fn tr_move_from_fn<U, R2: Dim, C2: Dim, F>(&mut self, other: OMatrix<U, R2, C2>, mut f: F)
-    where
-        ShapeConstraint: DimEq<R, C2> + SameNumberOfColumns<C, R2>,
-        DefaultAllocator: Allocator<U, R2, C2>,
-        F: FnMut(U) -> T,
-    {
-        let (nrows, ncols) = self.shape();
-        assert!(
-            (ncols, nrows) == other.shape(),
-            "Unable to move from a matrix with incompatible shape."
-        );
-
-        let mut md = other.manually_drop();
-
-        for j in 0..ncols {
-            for i in 0..nrows {
-                unsafe {
-                    *self.get_unchecked_mut((i, j)) =
-                        f(ManuallyDrop::take(md.get_unchecked_mut((j, i))));
+                    *self.get_unchecked_mut((i, j)) = other.get_unchecked((j, i)).inlined_clone();
                 }
             }
         }
@@ -1429,62 +1314,13 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
     // TODO: rename `apply` to `apply_mut` and `apply_into` to `apply`?
     /// Returns `self` with each of its components replaced by the result of a closure `f` applied on it.
     #[inline]
-    pub fn apply_into<F: FnMut(T) -> T>(mut self, f: F) -> Self
-    where
-        T: Clone,
-    {
+    pub fn apply_into<F: FnMut(&mut T)>(mut self, f: F) -> Self {
         self.apply(f);
         self
     }
 }
 
-impl<T, R: Dim, C: Dim, S: StorageMut<MaybeUninit<T>, R, C>> Matrix<MaybeUninit<T>, R, C, S> {
-    /// Initializes this matrix with the content of another one via clones. Both must have the same shape.
-    #[inline]
-    pub fn copy_init_from<R2: Dim, C2: Dim, SB>(&mut self, other: &Matrix<T, R2, C2, SB>)
-    where
-        T: Clone,
-        SB: Storage<T, R2, C2>,
-        ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
-    {
-        self.copy_from_fn(other, |e| MaybeUninit::new(e.clone()))
-    }
-
-    /// Initializes this matrix with the content of another one, after applying a function to
-    /// the entries of the other matrix. Both must have the same shape.
-    #[inline]
-    pub fn move_init_from<R2: Dim, C2: Dim, SB>(&mut self, other: OMatrix<T, R2, C2>)
-    where
-        SB: Storage<T, R2, C2>,
-        DefaultAllocator: Allocator<T, R2, C2>,
-        ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
-    {
-        self.move_from_fn(other, MaybeUninit::new)
-    }
-
-    /// Initializes this matrix with the content of the transpose another one via clones.
-    #[inline]
-    pub fn tr_copy_init_from<R2: Dim, C2: Dim, SB>(&mut self, other: &Matrix<T, R2, C2, SB>)
-    where
-        T: Clone,
-        SB: Storage<T, R2, C2>,
-        ShapeConstraint: DimEq<R, C2> + SameNumberOfColumns<C, R2>,
-    {
-        self.tr_copy_from_fn(other, |e| MaybeUninit::new(e.clone()))
-    }
-
-    /// Initializes this matrix with the content of the transpose another one via moves.
-    #[inline]
-    pub fn tr_move_init_from<R2: Dim, C2: Dim>(&mut self, other: OMatrix<T, R2, C2>)
-    where
-        DefaultAllocator: Allocator<T, R2, C2>,
-        ShapeConstraint: DimEq<R, C2> + SameNumberOfColumns<C, R2>,
-    {
-        self.tr_move_from_fn(other, MaybeUninit::new)
-    }
-}
-
-impl<T, D: Dim, S: Storage<T, D>> Vector<T, D, S> {
+impl<T, D: Dim, S: RawStorage<T, D>> Vector<T, D, S> {
     /// Gets a reference to the i-th element of this column vector without bound checking.
     #[inline]
     #[must_use]
@@ -1495,7 +1331,7 @@ impl<T, D: Dim, S: Storage<T, D>> Vector<T, D, S> {
     }
 }
 
-impl<T, D: Dim, S: StorageMut<T, D>> Vector<T, D, S> {
+impl<T, D: Dim, S: RawStorageMut<T, D>> Vector<T, D, S> {
     /// Gets a mutable reference to the i-th element of this column vector without bound checking.
     #[inline]
     #[must_use]
@@ -1506,25 +1342,27 @@ impl<T, D: Dim, S: StorageMut<T, D>> Vector<T, D, S> {
     }
 }
 
-impl<T, R: Dim, C: Dim, S: ContiguousStorage<T, R, C>> Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorage<T, R, C> + IsContiguous> Matrix<T, R, C, S> {
     /// Extracts a slice containing the entire matrix entries ordered column-by-columns.
     #[inline]
     #[must_use]
     pub fn as_slice(&self) -> &[T] {
-        self.data.as_slice()
+        // Safety: this is OK thanks to the IsContiguous trait.
+        unsafe { self.data.as_slice_unchecked() }
     }
 }
 
-impl<T, R: Dim, C: Dim, S: ContiguousStorageMut<T, R, C>> Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorageMut<T, R, C> + IsContiguous> Matrix<T, R, C, S> {
     /// Extracts a mutable slice containing the entire matrix entries ordered column-by-columns.
     #[inline]
     #[must_use]
     pub fn as_mut_slice(&mut self) -> &mut [T] {
-        self.data.as_mut_slice()
+        // Safety: this is OK thanks to the IsContiguous trait.
+        unsafe { self.data.as_mut_slice_unchecked() }
     }
 }
 
-impl<T, D: Dim, S: StorageMut<T, D, D>> Matrix<T, D, D, S> {
+impl<T: Scalar, D: Dim, S: RawStorageMut<T, D, D>> Matrix<T, D, D, S> {
     /// Transposes the square matrix `self` in-place.
     pub fn transpose_mut(&mut self) {
         assert!(
@@ -1542,12 +1380,18 @@ impl<T, D: Dim, S: StorageMut<T, D, D>> Matrix<T, D, D, S> {
     }
 }
 
-impl<T: SimdComplexField, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T: SimdComplexField, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     /// Takes the adjoint (aka. conjugate-transpose) of `self` and store the result into `out`.
     #[inline]
-    pub fn adjoint_to<R2: Dim, C2: Dim, SB>(&self, out: &mut Matrix<MaybeUninit<T>, R2, C2, SB>)
-    where
-        SB: StorageMut<MaybeUninit<T>, R2, C2>,
+    fn adjoint_to_uninit<Status, R2, C2, SB>(
+        &self,
+        status: Status,
+        out: &mut Matrix<Status::Value, R2, C2, SB>,
+    ) where
+        Status: InitStatus<T>,
+        R2: Dim,
+        C2: Dim,
+        SB: RawStorageMut<Status::Value, R2, C2>,
         ShapeConstraint: SameNumberOfRows<R, C2> + SameNumberOfColumns<C, R2>,
     {
         let (nrows, ncols) = self.shape();
@@ -1559,14 +1403,29 @@ impl<T: SimdComplexField, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S
         // TODO: optimize that.
         for i in 0..nrows {
             for j in 0..ncols {
+                // Safety: all indices are in range.
                 unsafe {
-                    *out.get_unchecked_mut((j, i)) =
-                        MaybeUninit::new(self.get_unchecked((i, j)).simd_conjugate());
+                    Status::init(
+                        out.get_unchecked_mut((j, i)),
+                        self.get_unchecked((i, j)).simd_conjugate(),
+                    );
                 }
             }
         }
     }
 
+    /// Takes the adjoint (aka. conjugate-transpose) of `self` and store the result into `out`.
+    #[inline]
+    pub fn adjoint_to<R2, C2, SB>(&self, out: &mut Matrix<T, R2, C2, SB>)
+    where
+        R2: Dim,
+        C2: Dim,
+        SB: RawStorageMut<T, R2, C2>,
+        ShapeConstraint: SameNumberOfRows<R, C2> + SameNumberOfColumns<C, R2>,
+    {
+        self.adjoint_to_uninit(Init, out)
+    }
+
     /// The adjoint (aka. conjugate-transpose) of `self`.
     #[inline]
     #[must_use = "Did you mean to use adjoint_mut()?"]
@@ -1574,21 +1433,23 @@ impl<T: SimdComplexField, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S
     where
         DefaultAllocator: Allocator<T, C, R>,
     {
-        let (nrows, ncols) = self.data.shape();
-        let mut res = OMatrix::new_uninitialized_generic(ncols, nrows);
-        self.adjoint_to(&mut res);
+        let (nrows, ncols) = self.shape_generic();
 
+        let mut res = Matrix::uninit(ncols, nrows);
+        self.adjoint_to_uninit(Uninit, &mut res);
+
+        // Safety: res is now fully initialized.
         unsafe { res.assume_init() }
     }
 
     /// Takes the conjugate and transposes `self` and store the result into `out`.
     #[deprecated(note = "Renamed `self.adjoint_to(out)`.")]
     #[inline]
-    pub fn conjugate_transpose_to<R2: Dim, C2: Dim, SB>(
-        &self,
-        out: &mut Matrix<MaybeUninit<T>, R2, C2, SB>,
-    ) where
-        SB: StorageMut<MaybeUninit<T>, R2, C2>,
+    pub fn conjugate_transpose_to<R2, C2, SB>(&self, out: &mut Matrix<T, R2, C2, SB>)
+    where
+        R2: Dim,
+        C2: Dim,
+        SB: RawStorageMut<T, R2, C2>,
         ShapeConstraint: SameNumberOfRows<R, C2> + SameNumberOfColumns<C, R2>,
     {
         self.adjoint_to(out)
@@ -1635,27 +1496,27 @@ impl<T: SimdComplexField, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S
     }
 }
 
-impl<T: SimdComplexField, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
+impl<T: SimdComplexField, R: Dim, C: Dim, S: RawStorageMut<T, R, C>> Matrix<T, R, C, S> {
     /// The conjugate of the complex matrix `self` computed in-place.
     #[inline]
     pub fn conjugate_mut(&mut self) {
-        self.apply(|e| e.simd_conjugate())
+        self.apply(|e| *e = e.simd_conjugate())
     }
 
     /// Divides each component of the complex matrix `self` by the given real.
     #[inline]
     pub fn unscale_mut(&mut self, real: T::SimdRealField) {
-        self.apply(|e| e.simd_unscale(real))
+        self.apply(|e| *e = e.simd_unscale(real))
     }
 
     /// Multiplies each component of the complex matrix `self` by the given real.
     #[inline]
     pub fn scale_mut(&mut self, real: T::SimdRealField) {
-        self.apply(|e| e.simd_scale(real))
+        self.apply(|e| *e = e.simd_scale(real))
     }
 }
 
-impl<T: SimdComplexField, D: Dim, S: StorageMut<T, D, D>> Matrix<T, D, D, S> {
+impl<T: SimdComplexField, D: Dim, S: RawStorageMut<T, D, D>> Matrix<T, D, D, S> {
     /// Sets `self` to its adjoint.
     #[deprecated(note = "Renamed to `self.adjoint_mut()`.")]
     pub fn conjugate_transform_mut(&mut self) {
@@ -1691,13 +1552,12 @@ impl<T: SimdComplexField, D: Dim, S: StorageMut<T, D, D>> Matrix<T, D, D, S> {
     }
 }
 
-impl<T, D: Dim, S: Storage<T, D, D>> SquareMatrix<T, D, S> {
+impl<T: Scalar, D: Dim, S: RawStorage<T, D, D>> SquareMatrix<T, D, S> {
     /// The diagonal of this matrix.
     #[inline]
     #[must_use]
     pub fn diagonal(&self) -> OVector<T, D>
     where
-        T: Clone,
         DefaultAllocator: Allocator<T, D>,
     {
         self.map_diagonal(|e| e)
@@ -1708,9 +1568,8 @@ impl<T, D: Dim, S: Storage<T, D, D>> SquareMatrix<T, D, S> {
     /// This is a more efficient version of `self.diagonal().map(f)` since this
     /// allocates only once.
     #[must_use]
-    pub fn map_diagonal<T2: Clone>(&self, mut f: impl FnMut(T) -> T2) -> OVector<T2, D>
+    pub fn map_diagonal<T2: Scalar>(&self, mut f: impl FnMut(T) -> T2) -> OVector<T2, D>
     where
-        T: Clone,
         DefaultAllocator: Allocator<T2, D>,
     {
         assert!(
@@ -1718,17 +1577,18 @@ impl<T, D: Dim, S: Storage<T, D, D>> SquareMatrix<T, D, S> {
             "Unable to get the diagonal of a non-square matrix."
         );
 
-        let dim = self.data.shape().0;
-        let mut res = OVector::new_uninitialized_generic(dim, Const::<1>);
+        let dim = self.shape_generic().0;
+        let mut res = Matrix::uninit(dim, Const::<1>);
 
         for i in 0..dim.value() {
+            // Safety: all indices are in range.
             unsafe {
                 *res.vget_unchecked_mut(i) =
-                    MaybeUninit::new(f(self.get_unchecked((i, i)).clone()));
+                    MaybeUninit::new(f(self.get_unchecked((i, i)).inlined_clone()));
             }
         }
 
-        // Safety: we have initialized all entries.
+        // Safety: res is now fully initialized.
         unsafe { res.assume_init() }
     }
 
@@ -1744,7 +1604,7 @@ impl<T, D: Dim, S: Storage<T, D, D>> SquareMatrix<T, D, S> {
             "Cannot compute the trace of non-square matrix."
         );
 
-        let dim = self.data.shape().0;
+        let dim = self.shape_generic().0;
         let mut res = T::zero();
 
         for i in 0..dim.value() {
@@ -1792,7 +1652,7 @@ impl<T: SimdComplexField, D: Dim, S: Storage<T, D, D>> SquareMatrix<T, D, S> {
     }
 }
 
-impl<T: Scalar + Zero + One, D: DimAdd<U1> + IsNotStaticOne, S: Storage<T, D, D>>
+impl<T: Scalar + Zero + One, D: DimAdd<U1> + IsNotStaticOne, S: RawStorage<T, D, D>>
     Matrix<T, D, D, S>
 {
     /// Yields the homogeneous matrix for this matrix, i.e., appending an additional dimension and
@@ -1809,13 +1669,13 @@ impl<T: Scalar + Zero + One, D: DimAdd<U1> + IsNotStaticOne, S: Storage<T, D, D>
         );
         let dim = DimSum::<D, U1>::from_usize(self.nrows() + 1);
         let mut res = OMatrix::identity_generic(dim, dim);
-        res.generic_slice_mut::<D, D>((0, 0), self.data.shape())
+        res.generic_slice_mut::<D, D>((0, 0), self.shape_generic())
             .copy_from(self);
         res
     }
 }
 
-impl<T: Scalar + Zero, D: DimAdd<U1>, S: Storage<T, D>> Vector<T, D, S> {
+impl<T: Scalar + Zero, D: DimAdd<U1>, S: RawStorage<T, D>> Vector<T, D, S> {
     /// Computes the coordinates in projective space of this vector, i.e., appends a `0` to its
     /// coordinates.
     #[inline]
@@ -1832,7 +1692,7 @@ impl<T: Scalar + Zero, D: DimAdd<U1>, S: Storage<T, D>> Vector<T, D, S> {
     #[inline]
     pub fn from_homogeneous<SB>(v: Vector<T, DimSum<D, U1>, SB>) -> Option<OVector<T, D>>
     where
-        SB: Storage<T, DimSum<D, U1>>,
+        SB: RawStorage<T, DimSum<D, U1>>,
         DefaultAllocator: Allocator<T, D>,
     {
         if v[v.len() - 1].is_zero() {
@@ -1844,7 +1704,7 @@ impl<T: Scalar + Zero, D: DimAdd<U1>, S: Storage<T, D>> Vector<T, D, S> {
     }
 }
 
-impl<T: Clone + Zero, D: DimAdd<U1>, S: Storage<T, D>> Vector<T, D, S> {
+impl<T: Scalar, D: DimAdd<U1>, S: RawStorage<T, D>> Vector<T, D, S> {
     /// Constructs a new vector of higher dimension by appending `element` to the end of `self`.
     #[inline]
     #[must_use]
@@ -1854,19 +1714,22 @@ impl<T: Clone + Zero, D: DimAdd<U1>, S: Storage<T, D>> Vector<T, D, S> {
     {
         let len = self.len();
         let hnrows = DimSum::<D, U1>::from_usize(len + 1);
-        let mut res = OVector::new_uninitialized_generic(hnrows, Const::<1>);
-        res.generic_slice_mut((0, 0), self.data.shape())
-            .copy_from_fn(self, |e| MaybeUninit::new(e.clone()));
+        let mut res = Matrix::uninit(hnrows, Const::<1>);
+        // This is basically a copy_from except that we warp the copied
+        // values into MaybeUninit.
+        res.generic_slice_mut((0, 0), self.shape_generic())
+            .zip_apply(self, |out, e| *out = MaybeUninit::new(e));
         res[(len, 0)] = MaybeUninit::new(element);
 
+        // Safety: res has been fully initialized.
         unsafe { res.assume_init() }
     }
 }
 
 impl<T, R: Dim, C: Dim, S> AbsDiffEq for Matrix<T, R, C, S>
 where
-    T: AbsDiffEq,
-    S: Storage<T, R, C>,
+    T: Scalar + AbsDiffEq,
+    S: RawStorage<T, R, C>,
     T::Epsilon: Copy,
 {
     type Epsilon = T::Epsilon;
@@ -1886,7 +1749,7 @@ where
 
 impl<T, R: Dim, C: Dim, S> RelativeEq for Matrix<T, R, C, S>
 where
-    T: RelativeEq,
+    T: Scalar + RelativeEq,
     S: Storage<T, R, C>,
     T::Epsilon: Copy,
 {
@@ -1908,8 +1771,8 @@ where
 
 impl<T, R: Dim, C: Dim, S> UlpsEq for Matrix<T, R, C, S>
 where
-    T: UlpsEq,
-    S: Storage<T, R, C>,
+    T: Scalar + UlpsEq,
+    S: RawStorage<T, R, C>,
     T::Epsilon: Copy,
 {
     #[inline]
@@ -1926,9 +1789,10 @@ where
     }
 }
 
-impl<T: PartialOrd, R: Dim, C: Dim, S> PartialOrd for Matrix<T, R, C, S>
+impl<T, R: Dim, C: Dim, S> PartialOrd for Matrix<T, R, C, S>
 where
-    S: Storage<T, R, C>,
+    T: Scalar + PartialOrd,
+    S: RawStorage<T, R, C>,
 {
     #[inline]
     fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
@@ -2017,13 +1881,22 @@ where
     }
 }
 
-impl<T: Eq, R: Dim, C: Dim, S> Eq for Matrix<T, R, C, S> where S: Storage<T, R, C> {}
-
-impl<T: PartialEq, R: Dim, R2: Dim, C: Dim, C2: Dim, S, S2> PartialEq<Matrix<T, R2, C2, S2>>
-    for Matrix<T, R, C, S>
+impl<T, R: Dim, C: Dim, S> Eq for Matrix<T, R, C, S>
 where
-    S: Storage<T, R, C>,
-    S2: Storage<T, R2, C2>,
+    T: Scalar + Eq,
+    S: RawStorage<T, R, C>,
+{
+}
+
+impl<T, R, R2, C, C2, S, S2> PartialEq<Matrix<T, R2, C2, S2>> for Matrix<T, R, C, S>
+where
+    T: Scalar + PartialEq,
+    C: Dim,
+    C2: Dim,
+    R: Dim,
+    R2: Dim,
+    S: RawStorage<T, R, C>,
+    S2: RawStorage<T, R2, C2>,
 {
     #[inline]
     fn eq(&self, right: &Matrix<T, R2, C2, S2>) -> bool {
@@ -2036,7 +1909,7 @@ macro_rules! impl_fmt {
         impl<T, R: Dim, C: Dim, S> $trait for Matrix<T, R, C, S>
         where
             T: Scalar + $trait,
-            S: Storage<T, R, C>,
+            S: RawStorage<T, R, C>,
         {
             fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
                 #[cfg(feature = "std")]
@@ -2140,7 +2013,7 @@ mod tests {
 }
 
 /// # Cross product
-impl<T: Scalar + ClosedAdd + ClosedSub + ClosedMul, R: Dim, C: Dim, S: Storage<T, R, C>>
+impl<T: Scalar + ClosedAdd + ClosedSub + ClosedMul, R: Dim, C: Dim, S: RawStorage<T, R, C>>
     Matrix<T, R, C, S>
 {
     /// The perpendicular product between two 2D column vectors, i.e. `a.x * b.y - a.y * b.x`.
@@ -2150,7 +2023,7 @@ impl<T: Scalar + ClosedAdd + ClosedSub + ClosedMul, R: Dim, C: Dim, S: Storage<T
     where
         R2: Dim,
         C2: Dim,
-        SB: Storage<T, R2, C2>,
+        SB: RawStorage<T, R2, C2>,
         ShapeConstraint: SameNumberOfRows<R, U2>
             + SameNumberOfColumns<C, U1>
             + SameNumberOfRows<R2, U2>
@@ -2176,12 +2049,11 @@ impl<T: Scalar + ClosedAdd + ClosedSub + ClosedMul, R: Dim, C: Dim, S: Storage<T
     /// dynamically-sized matrices and statically-sized 3D matrices.
     #[inline]
     #[must_use]
-    pub fn cross<R2: Dim, C2: Dim, SB>(
-        &self,
-        b: &Matrix<T, R2, C2, SB>,
-    ) -> MatrixCross<T, R, C, R2, C2>
+    pub fn cross<R2, C2, SB>(&self, b: &Matrix<T, R2, C2, SB>) -> MatrixCross<T, R, C, R2, C2>
     where
-        SB: Storage<T, R2, C2>,
+        R2: Dim,
+        C2: Dim,
+        SB: RawStorage<T, R2, C2>,
         DefaultAllocator: SameShapeAllocator<T, R, C, R2, C2>,
         ShapeConstraint: SameNumberOfRows<R, R2> + SameNumberOfColumns<C, C2>,
     {
@@ -2198,7 +2070,7 @@ impl<T: Scalar + ClosedAdd + ClosedSub + ClosedMul, R: Dim, C: Dim, S: Storage<T
                 // TODO: soooo ugly!
                 let nrows = SameShapeR::<R, R2>::from_usize(3);
                 let ncols = SameShapeC::<C, C2>::from_usize(1);
-                let mut res = Matrix::new_uninitialized_generic(nrows, ncols);
+                let mut res = Matrix::uninit(nrows, ncols);
 
                 let ax = self.get_unchecked((0, 0));
                 let ay = self.get_unchecked((1, 0));
@@ -2221,6 +2093,7 @@ impl<T: Scalar + ClosedAdd + ClosedSub + ClosedMul, R: Dim, C: Dim, S: Storage<T
                         - ay.inlined_clone() * bx.inlined_clone(),
                 );
 
+                // Safety: res is now fully initialized.
                 res.assume_init()
             }
         } else {
@@ -2228,7 +2101,7 @@ impl<T: Scalar + ClosedAdd + ClosedSub + ClosedMul, R: Dim, C: Dim, S: Storage<T
                 // TODO: ugly!
                 let nrows = SameShapeR::<R, R2>::from_usize(1);
                 let ncols = SameShapeC::<C, C2>::from_usize(3);
-                let mut res = Matrix::new_uninitialized_generic(nrows, ncols);
+                let mut res = Matrix::uninit(nrows, ncols);
 
                 let ax = self.get_unchecked((0, 0));
                 let ay = self.get_unchecked((0, 1));
@@ -2251,13 +2124,14 @@ impl<T: Scalar + ClosedAdd + ClosedSub + ClosedMul, R: Dim, C: Dim, S: Storage<T
                         - ay.inlined_clone() * bx.inlined_clone(),
                 );
 
+                // Safety: res is now fully initialized.
                 res.assume_init()
             }
         }
     }
 }
 
-impl<T: Scalar + Field, S: Storage<T, U3>> Vector<T, U3, S> {
+impl<T: Scalar + Field, S: RawStorage<T, U3>> Vector<T, U3, S> {
     /// Computes the matrix `M` such that for all vector `v` we have `M * v == self.cross(&v)`.
     #[inline]
     #[must_use]
@@ -2299,9 +2173,10 @@ impl<T: SimdComplexField, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S
     }
 }
 
-impl<T: AbsDiffEq, R: Dim, C: Dim, S> AbsDiffEq for Unit<Matrix<T, R, C, S>>
+impl<T, R: Dim, C: Dim, S> AbsDiffEq for Unit<Matrix<T, R, C, S>>
 where
-    S: Storage<T, R, C>,
+    T: Scalar + AbsDiffEq,
+    S: RawStorage<T, R, C>,
     T::Epsilon: Copy,
 {
     type Epsilon = T::Epsilon;
@@ -2317,8 +2192,9 @@ where
     }
 }
 
-impl<T: RelativeEq, R: Dim, C: Dim, S> RelativeEq for Unit<Matrix<T, R, C, S>>
+impl<T, R: Dim, C: Dim, S> RelativeEq for Unit<Matrix<T, R, C, S>>
 where
+    T: Scalar + RelativeEq,
     S: Storage<T, R, C>,
     T::Epsilon: Copy,
 {
@@ -2339,9 +2215,10 @@ where
     }
 }
 
-impl<T: UlpsEq, R: Dim, C: Dim, S> UlpsEq for Unit<Matrix<T, R, C, S>>
+impl<T, R: Dim, C: Dim, S> UlpsEq for Unit<Matrix<T, R, C, S>>
 where
-    S: Storage<T, R, C>,
+    T: Scalar + UlpsEq,
+    S: RawStorage<T, R, C>,
     T::Epsilon: Copy,
 {
     #[inline]
@@ -2355,9 +2232,12 @@ where
     }
 }
 
-impl<T: Hash, R: Dim, C: Dim, S> Hash for Matrix<T, R, C, S>
+impl<T, R, C, S> Hash for Matrix<T, R, C, S>
 where
-    S: Storage<T, R, C>,
+    T: Scalar + Hash,
+    R: Dim,
+    C: Dim,
+    S: RawStorage<T, R, C>,
 {
     fn hash<H: Hasher>(&self, state: &mut H) {
         let (nrows, ncols) = self.shape();
diff --git a/src/base/matrix_simba.rs b/src/base/matrix_simba.rs
index f3f2d13b..5c259207 100644
--- a/src/base/matrix_simba.rs
+++ b/src/base/matrix_simba.rs
@@ -9,9 +9,11 @@ use crate::base::{DefaultAllocator, OMatrix, Scalar};
  * Simd structures.
  *
  */
-impl<T, R: Dim, C: Dim> SimdValue for OMatrix<T, R, C>
+impl<T, R, C> SimdValue for OMatrix<T, R, C>
 where
     T: Scalar + SimdValue,
+    R: Dim,
+    C: Dim,
     T::Element: Scalar,
     DefaultAllocator: Allocator<T, R, C> + Allocator<T::Element, R, C>,
 {
@@ -42,7 +44,6 @@ where
     fn replace(&mut self, i: usize, val: Self::Element) {
         self.zip_apply(&val, |mut a, b| {
             a.replace(i, b);
-            a
         })
     }
 
@@ -50,7 +51,6 @@ where
     unsafe fn replace_unchecked(&mut self, i: usize, val: Self::Element) {
         self.zip_apply(&val, |mut a, b| {
             a.replace_unchecked(i, b);
-            a
         })
     }
 
diff --git a/src/base/matrix_slice.rs b/src/base/matrix_slice.rs
index 0d65a4fa..261d41e2 100644
--- a/src/base/matrix_slice.rs
+++ b/src/base/matrix_slice.rs
@@ -1,14 +1,13 @@
 use std::marker::PhantomData;
-use std::mem::MaybeUninit;
 use std::ops::{Range, RangeFrom, RangeFull, RangeInclusive, RangeTo};
 use std::slice;
 
-use crate::base::allocator::{Allocator, InnerAllocator};
+use crate::base::allocator::Allocator;
 use crate::base::default_allocator::DefaultAllocator;
 use crate::base::dimension::{Const, Dim, DimName, Dynamic, IsNotStaticOne, U1};
 use crate::base::iter::MatrixIter;
-use crate::base::storage::{ContiguousStorage, ContiguousStorageMut, Storage, StorageMut};
-use crate::base::{Matrix, Owned};
+use crate::base::storage::{IsContiguous, Owned, RawStorage, RawStorageMut, Storage};
+use crate::base::{Matrix, Scalar};
 
 macro_rules! slice_storage_impl(
     ($doc: expr; $Storage: ident as $SRef: ty; $T: ident.$get_addr: ident ($Ptr: ty as $Ref: ty)) => {
@@ -82,7 +81,7 @@ macro_rules! slice_storage_impl(
         impl <'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
             $T<'a, T, R, C, RStride, CStride>
         where
-            Self: ContiguousStorage<T, R, C>
+            Self: RawStorage<T, R, C> + IsContiguous
         {
             /// Extracts the original slice from this storage
             pub fn into_slice(self) -> &'a [T] {
@@ -100,19 +99,19 @@ macro_rules! slice_storage_impl(
 
 slice_storage_impl!("A matrix data storage for a matrix slice. Only contains an internal reference \
                      to another matrix data storage.";
-    Storage as &'a S; SliceStorage.get_address_unchecked(*const T as &'a T));
+    RawStorage as &'a S; SliceStorage.get_address_unchecked(*const T as &'a T));
 
 slice_storage_impl!("A mutable matrix data storage for mutable matrix slice. Only contains an \
                      internal mutable reference to another matrix data storage.";
-    StorageMut as &'a mut S; SliceStorageMut.get_address_unchecked_mut(*mut T as &'a mut T)
+    RawStorageMut as &'a mut S; SliceStorageMut.get_address_unchecked_mut(*mut T as &'a mut T)
 );
 
-impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim> Copy
+impl<'a, T: Scalar, R: Dim, C: Dim, RStride: Dim, CStride: Dim> Copy
     for SliceStorage<'a, T, R, C, RStride, CStride>
 {
 }
 
-impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim> Clone
+impl<'a, T: Scalar, R: Dim, C: Dim, RStride: Dim, CStride: Dim> Clone
     for SliceStorage<'a, T, R, C, RStride, CStride>
 {
     #[inline]
@@ -126,10 +125,10 @@ impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim> Clone
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
+impl<'a, T: Scalar, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
     SliceStorageMut<'a, T, R, C, RStride, CStride>
 where
-    Self: ContiguousStorageMut<T, R, C>,
+    Self: RawStorageMut<T, R, C> + IsContiguous,
 {
     /// Extracts the original slice from this storage
     pub fn into_slice_mut(self) -> &'a mut [T] {
@@ -145,7 +144,7 @@ where
 
 macro_rules! storage_impl(
     ($($T: ident),* $(,)*) => {$(
-        unsafe impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim> Storage<T, R, C>
+        unsafe impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim> RawStorage<T, R, C>
             for $T<'a, T, R, C, RStride, CStride> {
 
             type RStride = RStride;
@@ -182,26 +181,6 @@ macro_rules! storage_impl(
                 }
             }
 
-            #[inline]
-            fn into_owned(self) -> Owned<T, R, C>
-            where
-                T: Clone,
-                DefaultAllocator: Allocator<T, R, C>
-            {
-                self.clone_owned()
-            }
-
-            #[inline]
-            fn clone_owned(&self) -> Owned<T, R, C>
-            where
-                T: Clone,
-                DefaultAllocator: Allocator<T, R, C>
-            {
-                let (nrows, ncols) = self.shape();
-                let it = MatrixIter::new(self).cloned();
-             Owned(   DefaultAllocator::allocate_from_iterator(nrows, ncols, it))
-            }
-
             #[inline]
             unsafe fn as_slice_unchecked(&self) -> &[T] {
                 let (nrows, ncols) = self.shape();
@@ -214,39 +193,29 @@ macro_rules! storage_impl(
                 }
             }
         }
+
+        unsafe impl<'a, T: Scalar, R: Dim, C: Dim, RStride: Dim, CStride: Dim> Storage<T, R, C>
+            for $T<'a, T, R, C, RStride, CStride> {
+            #[inline]
+            fn into_owned(self) -> Owned<T, R, C>
+                where DefaultAllocator: Allocator<T, R, C> {
+                self.clone_owned()
+            }
+
+            #[inline]
+            fn clone_owned(&self) -> Owned<T, R, C>
+                where DefaultAllocator: Allocator<T, R, C> {
+                let (nrows, ncols) = self.shape();
+                let it = MatrixIter::new(self).cloned();
+                DefaultAllocator::allocate_from_iterator(nrows, ncols, it)
+            }
+        }
     )*}
 );
 
 storage_impl!(SliceStorage, SliceStorageMut);
 
-impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
-    SliceStorage<'a, MaybeUninit<T>, R, C, RStride, CStride>
-{
-    /// Assumes a slice storage's entries to be initialized. This operation
-    /// should be near zero-cost.
-    ///
-    /// # Safety
-    /// All of the slice storage's entries must be initialized, otherwise
-    /// Undefined Behavior will be triggered.
-    pub unsafe fn assume_init(self) -> SliceStorage<'a, T, R, C, RStride, CStride> {
-        SliceStorage::from_raw_parts(self.ptr as *const T, self.shape, self.strides)
-    }
-}
-
-impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
-    SliceStorageMut<'a, MaybeUninit<T>, R, C, RStride, CStride>
-{
-    /// Assumes a slice storage's entries to be initialized. This operation should be near zero-cost.
-    ///
-    /// # Safety
-    /// The user must make sure that every single entry of the buffer has been initialized,
-    /// or Undefined Behavior will immediately occur.    
-    pub unsafe fn assume_init(self) -> SliceStorageMut<'a, T, R, C, RStride, CStride> {
-        SliceStorageMut::from_raw_parts(self.ptr as *mut T, self.shape, self.strides)
-    }
-}
-
-unsafe impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim> StorageMut<T, R, C>
+unsafe impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim> RawStorageMut<T, R, C>
     for SliceStorageMut<'a, T, R, C, RStride, CStride>
 {
     #[inline]
@@ -266,37 +235,22 @@ unsafe impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim> StorageMut<T, R,
     }
 }
 
-unsafe impl<'a, T, R: Dim, CStride: Dim> ContiguousStorage<T, R, U1>
-    for SliceStorage<'a, T, R, U1, U1, CStride>
-{
-}
-
-unsafe impl<'a, T, R: Dim, CStride: Dim> ContiguousStorage<T, R, U1>
+unsafe impl<'a, T, R: Dim, CStride: Dim> IsContiguous for SliceStorage<'a, T, R, U1, U1, CStride> {}
+unsafe impl<'a, T, R: Dim, CStride: Dim> IsContiguous
     for SliceStorageMut<'a, T, R, U1, U1, CStride>
 {
 }
 
-unsafe impl<'a, T, R: Dim, CStride: Dim> ContiguousStorageMut<T, R, U1>
-    for SliceStorageMut<'a, T, R, U1, U1, CStride>
-{
-}
-
-unsafe impl<'a, T, R: DimName, C: Dim + IsNotStaticOne> ContiguousStorage<T, R, C>
+unsafe impl<'a, T, R: DimName, C: Dim + IsNotStaticOne> IsContiguous
     for SliceStorage<'a, T, R, C, U1, R>
 {
 }
-
-unsafe impl<'a, T, R: DimName, C: Dim + IsNotStaticOne> ContiguousStorage<T, R, C>
+unsafe impl<'a, T, R: DimName, C: Dim + IsNotStaticOne> IsContiguous
     for SliceStorageMut<'a, T, R, C, U1, R>
 {
 }
 
-unsafe impl<'a, T, R: DimName, C: Dim + IsNotStaticOne> ContiguousStorageMut<T, R, C>
-    for SliceStorageMut<'a, T, R, C, U1, R>
-{
-}
-
-impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     #[inline]
     fn assert_slice_index(
         &self,
@@ -344,7 +298,6 @@ macro_rules! matrix_slice_impl(
      $fixed_slice_with_steps: ident,
      $generic_slice: ident,
      $generic_slice_with_steps: ident,
-     $full_slice: ident,
      $rows_range_pair: ident,
      $columns_range_pair: ident) => {
         /*
@@ -403,14 +356,14 @@ macro_rules! matrix_slice_impl(
         pub fn $rows_generic<RSlice: Dim>($me: $Me, row_start: usize, nrows: RSlice)
             -> $MatrixSlice<'_, T, RSlice, C, S::RStride, S::CStride> {
 
-            let my_shape  = $me.data.shape();
+            let my_shape   = $me.shape_generic();
             $me.assert_slice_index((row_start, 0), (nrows.value(), my_shape.1.value()), (0, 0));
 
             let shape = (nrows, my_shape.1);
 
             unsafe {
                 let data = $SliceStorage::new_unchecked($data, (row_start, 0), shape);
-                Matrix::from_data(data)
+                Matrix::from_data_statically_unchecked(data)
             }
         }
 
@@ -421,16 +374,16 @@ macro_rules! matrix_slice_impl(
             -> $MatrixSlice<'_, T, RSlice, C, Dynamic, S::CStride>
             where RSlice: Dim {
 
-            let my_shape = $me.data.shape();
+            let my_shape   = $me.shape_generic();
             let my_strides = $me.data.strides();
             $me.assert_slice_index((row_start, 0), (nrows.value(), my_shape.1.value()), (step, 0));
 
             let strides = (Dynamic::new((step + 1) * my_strides.0.value()), my_strides.1);
-            let shape = (nrows, my_shape.1);
+            let shape   = (nrows, my_shape.1);
 
             unsafe {
                 let data = $SliceStorage::new_with_strides_unchecked($data, (row_start, 0), shape, strides);
-                Matrix::from_data(data)
+                Matrix::from_data_statically_unchecked(data)
             }
         }
 
@@ -491,33 +444,34 @@ macro_rules! matrix_slice_impl(
         pub fn $columns_generic<CSlice: Dim>($me: $Me, first_col: usize, ncols: CSlice)
             -> $MatrixSlice<'_, T, R, CSlice, S::RStride, S::CStride> {
 
-            let my_shape = $me.data.shape();
+            let my_shape = $me.shape_generic();
             $me.assert_slice_index((0, first_col), (my_shape.0.value(), ncols.value()), (0, 0));
             let shape = (my_shape.0, ncols);
 
             unsafe {
                 let data = $SliceStorage::new_unchecked($data, (0, first_col), shape);
-                Matrix::from_data(data)
+                Matrix::from_data_statically_unchecked(data)
             }
         }
 
+
         /// Extracts from this matrix `ncols` columns skipping `step` columns. Both argument may
         /// or may not be values known at compile-time.
         #[inline]
         pub fn $columns_generic_with_step<CSlice: Dim>($me: $Me, first_col: usize, ncols: CSlice, step: usize)
             -> $MatrixSlice<'_, T, R, CSlice, S::RStride, Dynamic> {
 
-            let my_shape = $me.data.shape();
+            let my_shape   = $me.shape_generic();
             let my_strides = $me.data.strides();
 
             $me.assert_slice_index((0, first_col), (my_shape.0.value(), ncols.value()), (0, step));
 
             let strides = (my_strides.0, Dynamic::new((step + 1) * my_strides.1.value()));
-            let shape = (my_shape.0, ncols);
+            let shape   = (my_shape.0, ncols);
 
             unsafe {
                 let data = $SliceStorage::new_with_strides_unchecked($data, (0, first_col), shape, strides);
-                Matrix::from_data(data)
+                Matrix::from_data_statically_unchecked(data)
             }
         }
 
@@ -537,10 +491,11 @@ macro_rules! matrix_slice_impl(
 
             unsafe {
                 let data = $SliceStorage::new_unchecked($data, start, shape);
-                Matrix::from_data(data)
+                Matrix::from_data_statically_unchecked(data)
             }
         }
 
+
         /// Slices this matrix starting at its component `(start.0, start.1)` and with
         /// `(shape.0, shape.1)` components. Each row (resp. column) of the sliced matrix is
         /// separated by `steps.0` (resp. `steps.1`) ignored rows (resp. columns) of the
@@ -564,7 +519,7 @@ macro_rules! matrix_slice_impl(
 
             unsafe {
                 let data = $SliceStorage::new_unchecked($data, (irow, icol), shape);
-                Matrix::from_data(data)
+                Matrix::from_data_statically_unchecked(data)
             }
         }
 
@@ -581,14 +536,16 @@ macro_rules! matrix_slice_impl(
 
         /// Creates a slice that may or may not have a fixed size and stride.
         #[inline]
-        pub fn $generic_slice<RSlice: Dim, CSlice: Dim>($me: $Me, start: (usize, usize), shape: (RSlice, CSlice))
-            -> $MatrixSlice<T, RSlice, CSlice, S::RStride, S::CStride>
-        {
+        pub fn $generic_slice<RSlice, CSlice>($me: $Me, start: (usize, usize), shape: (RSlice, CSlice))
+            -> $MatrixSlice<'_, T, RSlice, CSlice, S::RStride, S::CStride>
+            where RSlice: Dim,
+                  CSlice: Dim {
+
             $me.assert_slice_index(start, (shape.0.value(), shape.1.value()), (0, 0));
 
             unsafe {
                 let data = $SliceStorage::new_unchecked($data, start, shape);
-                Matrix::from_data(data)
+                Matrix::from_data_statically_unchecked(data)
             }
         }
 
@@ -610,16 +567,10 @@ macro_rules! matrix_slice_impl(
 
             unsafe {
                 let data = $SliceStorage::new_with_strides_unchecked($data, start, shape, strides);
-                Matrix::from_data(data)
+                Matrix::from_data_statically_unchecked(data)
             }
         }
 
-        /// Returns a slice containing the entire matrix.
-        pub fn $full_slice($me: $Me) -> $MatrixSlice<T, R, C, S::RStride, S::CStride> {
-            let (nrows, ncols) = $me.shape();
-            $me.$generic_slice((0, 0), (R::from_usize(nrows), C::from_usize(ncols)))
-        }
-
         /*
          *
          * Splitting.
@@ -633,7 +584,7 @@ macro_rules! matrix_slice_impl(
             -> ($MatrixSlice<'_, T, Range1::Size, C, S::RStride, S::CStride>,
                 $MatrixSlice<'_, T, Range2::Size, C, S::RStride, S::CStride>) {
 
-            let (nrows, ncols) = $me.data.shape();
+            let (nrows, ncols) = $me.shape_generic();
             let strides        = $me.data.strides();
 
             let start1 = r1.begin(nrows);
@@ -654,8 +605,8 @@ macro_rules! matrix_slice_impl(
 
                 let data1  = $SliceStorage::from_raw_parts(ptr1, (nrows1, ncols), strides);
                 let data2  = $SliceStorage::from_raw_parts(ptr2, (nrows2, ncols), strides);
-                let slice1 = Matrix::from_data(data1);
-                let slice2 = Matrix::from_data(data2);
+                let slice1 = Matrix::from_data_statically_unchecked(data1);
+                let slice2 = Matrix::from_data_statically_unchecked(data2);
 
                 (slice1, slice2)
             }
@@ -669,7 +620,7 @@ macro_rules! matrix_slice_impl(
             -> ($MatrixSlice<'_, T, R, Range1::Size, S::RStride, S::CStride>,
                 $MatrixSlice<'_, T, R, Range2::Size, S::RStride, S::CStride>) {
 
-            let (nrows, ncols) = $me.data.shape();
+            let (nrows, ncols) = $me.shape_generic();
             let strides        = $me.data.strides();
 
             let start1 = r1.begin(ncols);
@@ -690,8 +641,8 @@ macro_rules! matrix_slice_impl(
 
                 let data1  = $SliceStorage::from_raw_parts(ptr1, (nrows, ncols1), strides);
                 let data2  = $SliceStorage::from_raw_parts(ptr2, (nrows, ncols2), strides);
-                let slice1 = Matrix::from_data(data1);
-                let slice2 = Matrix::from_data(data2);
+                let slice1 = Matrix::from_data_statically_unchecked(data1);
+                let slice2 = Matrix::from_data_statically_unchecked(data2);
 
                 (slice1, slice2)
             }
@@ -707,9 +658,9 @@ pub type MatrixSliceMut<'a, T, R, C, RStride = U1, CStride = R> =
     Matrix<T, R, C, SliceStorageMut<'a, T, R, C, RStride, CStride>>;
 
 /// # Slicing based on index and length
-impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     matrix_slice_impl!(
-     self: &Self, MatrixSlice, SliceStorage, Storage.get_address_unchecked(), &self.data;
+     self: &Self, MatrixSlice, SliceStorage, RawStorage.get_address_unchecked(), &self.data;
      row,
      row_part,
      rows,
@@ -732,15 +683,14 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
      fixed_slice_with_steps,
      generic_slice,
      generic_slice_with_steps,
-     full_slice,
      rows_range_pair,
      columns_range_pair);
 }
 
 /// # Mutable slicing based on index and length
-impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorageMut<T, R, C>> Matrix<T, R, C, S> {
     matrix_slice_impl!(
-     self: &mut Self, MatrixSliceMut, SliceStorageMut, StorageMut.get_address_unchecked_mut(), &mut self.data;
+     self: &mut Self, MatrixSliceMut, SliceStorageMut, RawStorageMut.get_address_unchecked_mut(), &mut self.data;
      row_mut,
      row_part_mut,
      rows_mut,
@@ -763,29 +713,10 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
      fixed_slice_with_steps_mut,
      generic_slice_mut,
      generic_slice_with_steps_mut,
-     full_slice_mut,
      rows_range_pair_mut,
      columns_range_pair_mut);
 }
 
-impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
-    MatrixSlice<'a, MaybeUninit<T>, R, C, RStride, CStride>
-{
-    /// Assumes a matrix slices's entries to be initialized. This operation should be near zero-cost.
-    pub unsafe fn slice_assume_init(self) -> MatrixSlice<'a, T, R, C, RStride, CStride> {
-        Matrix::from_data(self.data.assume_init())
-    }
-}
-
-impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
-    MatrixSliceMut<'a, MaybeUninit<T>, R, C, RStride, CStride>
-{
-    /// Assumes a matrix slices's entries to be initialized. This operation should be near zero-cost.
-    pub unsafe fn slice_assume_init(self) -> MatrixSliceMut<'a, T, R, C, RStride, CStride> {
-        Matrix::from_data(self.data.assume_init())
-    }
-}
-
 /// A range with a size that may be known at compile-time.
 ///
 /// This may be:
@@ -922,7 +853,7 @@ impl<D: Dim> SliceRange<D> for RangeInclusive<usize> {
 
 // TODO: see how much of this overlaps with the general indexing
 // methods from indexing.rs.
-impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     /// Slices a sub-matrix containing the rows indexed by the range `rows` and the columns indexed
     /// by the range `cols`.
     #[inline]
@@ -936,7 +867,7 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         RowRange: SliceRange<R>,
         ColRange: SliceRange<C>,
     {
-        let (nrows, ncols) = self.data.shape();
+        let (nrows, ncols) = self.shape_generic();
         self.generic_slice(
             (rows.begin(nrows), cols.begin(ncols)),
             (rows.size(nrows), cols.size(ncols)),
@@ -966,7 +897,7 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
 
 // TODO: see how much of this overlaps with the general indexing
 // methods from indexing.rs.
-impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorageMut<T, R, C>> Matrix<T, R, C, S> {
     /// Slices a mutable sub-matrix containing the rows indexed by the range `rows` and the columns
     /// indexed by the range `cols`.
     pub fn slice_range_mut<RowRange, ColRange>(
@@ -978,7 +909,7 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
         RowRange: SliceRange<R>,
         ColRange: SliceRange<C>,
     {
-        let (nrows, ncols) = self.data.shape();
+        let (nrows, ncols) = self.shape_generic();
         self.generic_slice_mut(
             (rows.begin(nrows), cols.begin(ncols)),
             (rows.size(nrows), cols.size(ncols)),
@@ -1004,9 +935,13 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
     }
 }
 
-impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
-    From<MatrixSliceMut<'a, T, R, C, RStride, CStride>>
+impl<'a, T, R, C, RStride, CStride> From<MatrixSliceMut<'a, T, R, C, RStride, CStride>>
     for MatrixSlice<'a, T, R, C, RStride, CStride>
+where
+    R: Dim,
+    C: Dim,
+    RStride: Dim,
+    CStride: Dim,
 {
     fn from(slice_mut: MatrixSliceMut<'a, T, R, C, RStride, CStride>) -> Self {
         let data = SliceStorage {
@@ -1016,6 +951,6 @@ impl<'a, T, R: Dim, C: Dim, RStride: Dim, CStride: Dim>
             _phantoms: PhantomData,
         };
 
-        Matrix::from_data(data)
+        unsafe { Matrix::from_data_statically_unchecked(data) }
     }
 }
diff --git a/src/base/min_max.rs b/src/base/min_max.rs
index 83e62d10..3d390194 100644
--- a/src/base/min_max.rs
+++ b/src/base/min_max.rs
@@ -1,10 +1,10 @@
-use crate::storage::Storage;
+use crate::storage::RawStorage;
 use crate::{ComplexField, Dim, Matrix, Scalar, SimdComplexField, SimdPartialOrd, Vector};
 use num::{Signed, Zero};
 use simba::simd::SimdSigned;
 
 /// # Find the min and max components
-impl<T: Scalar, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T: Scalar, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     /// Returns the absolute value of the component with the largest absolute value.
     /// # Example
     /// ```
@@ -167,7 +167,7 @@ impl<T: Scalar, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     }
 }
 
-impl<T: Scalar + PartialOrd + Signed, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T: Scalar + PartialOrd + Signed, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     /// Computes the index of the matrix component with the largest absolute value.
     ///
     /// # Examples:
@@ -203,7 +203,7 @@ impl<T: Scalar + PartialOrd + Signed, R: Dim, C: Dim, S: Storage<T, R, C>> Matri
 
 // TODO: find a way to avoid code duplication just for complex number support.
 /// # Find the min and max components (vector-specific methods)
-impl<T: Scalar, D: Dim, S: Storage<T, D>> Vector<T, D, S> {
+impl<T: Scalar, D: Dim, S: RawStorage<T, D>> Vector<T, D, S> {
     /// Computes the index of the vector component with the largest complex or real absolute value.
     ///
     /// # Examples:
diff --git a/src/base/mod.rs b/src/base/mod.rs
index fdfbb5c7..88b79dc3 100644
--- a/src/base/mod.rs
+++ b/src/base/mod.rs
@@ -33,10 +33,12 @@ mod unit;
 #[cfg(any(feature = "std", feature = "alloc"))]
 mod vec_storage;
 
+mod blas_uninit;
 #[doc(hidden)]
 pub mod helper;
 mod interpolation;
 mod min_max;
+pub mod uninit;
 
 pub use self::matrix::*;
 pub use self::norm::*;
@@ -50,5 +52,6 @@ pub use self::alias::*;
 pub use self::alias_slice::*;
 pub use self::array_storage::*;
 pub use self::matrix_slice::*;
+pub use self::storage::*;
 #[cfg(any(feature = "std", feature = "alloc"))]
 pub use self::vec_storage::*;
diff --git a/src/base/norm.rs b/src/base/norm.rs
index a8548ddd..c138069d 100644
--- a/src/base/norm.rs
+++ b/src/base/norm.rs
@@ -434,7 +434,7 @@ impl<T: Scalar, R: Dim, C: Dim, S: StorageMut<T, R, C>> Matrix<T, R, C, S> {
     {
         let n = self.norm();
         let le = n.simd_le(min_norm);
-        self.apply(|e| e.simd_unscale(n).select(le, e));
+        self.apply(|e| *e = e.simd_unscale(n).select(le, *e));
         SimdOption::new(n, le)
     }
 
@@ -508,13 +508,8 @@ where
     /// The i-the canonical basis element.
     #[inline]
     fn canonical_basis_element(i: usize) -> Self {
-        assert!(i < D::dim(), "Index out of bound.");
-
         let mut res = Self::zero();
-        unsafe {
-            *res.data.get_unchecked_linear_mut(i) = T::one();
-        }
-
+        res[i] = T::one();
         res
     }
 
diff --git a/src/base/ops.rs b/src/base/ops.rs
index 45a84b35..bbeb6d07 100644
--- a/src/base/ops.rs
+++ b/src/base/ops.rs
@@ -1,29 +1,31 @@
 use num::{One, Zero};
 use std::iter;
-use std::mem::MaybeUninit;
 use std::ops::{
     Add, AddAssign, Div, DivAssign, Index, IndexMut, Mul, MulAssign, Neg, Sub, SubAssign,
 };
 
 use simba::scalar::{ClosedAdd, ClosedDiv, ClosedMul, ClosedNeg, ClosedSub};
 
-use crate::base::allocator::{
-    Allocator, InnerAllocator, SameShapeAllocator, SameShapeC, SameShapeR,
-};
+use crate::base::allocator::{Allocator, SameShapeAllocator, SameShapeC, SameShapeR};
+use crate::base::blas_uninit::gemm_uninit;
 use crate::base::constraint::{
     AreMultipliable, DimEq, SameNumberOfColumns, SameNumberOfRows, ShapeConstraint,
 };
 use crate::base::dimension::{Dim, DimMul, DimName, DimProd, Dynamic};
-use crate::base::storage::{ContiguousStorageMut, Storage, StorageMut};
+use crate::base::storage::{Storage, StorageMut};
+use crate::base::uninit::Uninit;
 use crate::base::{DefaultAllocator, Matrix, MatrixSum, OMatrix, Scalar, VectorSlice};
-use crate::{MatrixSliceMut, SimdComplexField};
+use crate::storage::IsContiguous;
+use crate::uninit::{Init, InitStatus};
+use crate::{RawStorage, RawStorageMut, SimdComplexField};
+use std::mem::MaybeUninit;
 
 /*
  *
  * Indexing.
  *
  */
-impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Index<usize> for Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorage<T, R, C>> Index<usize> for Matrix<T, R, C, S> {
     type Output = T;
 
     #[inline]
@@ -33,10 +35,7 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Index<usize> for Matrix<T, R, C, S>
     }
 }
 
-impl<T, R: Dim, C: Dim, S> Index<(usize, usize)> for Matrix<T, R, C, S>
-where
-    S: Storage<T, R, C>,
-{
+impl<T, R: Dim, C: Dim, S: RawStorage<T, R, C>> Index<(usize, usize)> for Matrix<T, R, C, S> {
     type Output = T;
 
     #[inline]
@@ -52,7 +51,7 @@ where
 }
 
 // Mutable versions.
-impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> IndexMut<usize> for Matrix<T, R, C, S> {
+impl<T, R: Dim, C: Dim, S: RawStorageMut<T, R, C>> IndexMut<usize> for Matrix<T, R, C, S> {
     #[inline]
     fn index_mut(&mut self, i: usize) -> &mut T {
         let ij = self.vector_to_matrix_index(i);
@@ -60,10 +59,7 @@ impl<T, R: Dim, C: Dim, S: StorageMut<T, R, C>> IndexMut<usize> for Matrix<T, R,
     }
 }
 
-impl<T, R: Dim, C: Dim, S> IndexMut<(usize, usize)> for Matrix<T, R, C, S>
-where
-    S: StorageMut<T, R, C>,
-{
+impl<T, R: Dim, C: Dim, S: RawStorageMut<T, R, C>> IndexMut<(usize, usize)> for Matrix<T, R, C, S> {
     #[inline]
     fn index_mut(&mut self, ij: (usize, usize)) -> &mut T {
         let shape = self.shape();
@@ -135,25 +131,27 @@ macro_rules! componentwise_binop_impl(
     ($Trait: ident, $method: ident, $bound: ident;
      $TraitAssign: ident, $method_assign: ident, $method_assign_statically_unchecked: ident,
      $method_assign_statically_unchecked_rhs: ident;
-     $method_to: ident, $method_to_statically_unchecked: ident) => {
+     $method_to: ident, $method_to_statically_unchecked_uninit: ident) => {
+
         impl<T, R1: Dim, C1: Dim, SA: Storage<T, R1, C1>> Matrix<T, R1, C1, SA>
-        where
-            T: Scalar + $bound
-        {
+            where T: Scalar + $bound {
+
             /*
              *
              * Methods without dimension checking at compile-time.
-             * This is useful for code reuse because the sum representative system does not play
-             * nicely with static checks.
+             * This is useful for code reuse because the sum representative system does not plays
+             * easily with static checks.
              *
              */
             #[inline]
-            fn $method_to_statically_unchecked<R2: Dim, C2: Dim, SB, R3: Dim, C3: Dim, SC>(
-                &self, rhs: &Matrix<T, R2, C2, SB>, out: &mut Matrix<MaybeUninit<T>, R3, C3, SC>
-            ) where
-                SB: Storage<T, R2, C2>,
-                SC: StorageMut<MaybeUninit<T>, R3, C3>
-            {
+            fn $method_to_statically_unchecked_uninit<Status, R2: Dim, C2: Dim, SB,
+                                                       R3: Dim, C3: Dim, SC>(&self,
+                                                                     status: Status,
+                                                                     rhs: &Matrix<T, R2, C2, SB>,
+                                                                     out: &mut Matrix<Status::Value, R3, C3, SC>)
+                where Status: InitStatus<T>,
+                      SB: RawStorage<T, R2, C2>,
+                      SC: RawStorageMut<Status::Value, R3, C3> {
                 assert_eq!(self.shape(), rhs.shape(), "Matrix addition/subtraction dimensions mismatch.");
                 assert_eq!(self.shape(), out.shape(), "Matrix addition/subtraction output dimensions mismatch.");
 
@@ -163,31 +161,28 @@ macro_rules! componentwise_binop_impl(
                     if self.data.is_contiguous() && rhs.data.is_contiguous() && out.data.is_contiguous() {
                         let arr1 = self.data.as_slice_unchecked();
                         let arr2 = rhs.data.as_slice_unchecked();
-                        let out = out.data.as_mut_slice_unchecked();
-                        for i in 0..arr1.len() {
-                            *out.get_unchecked_mut(i) = MaybeUninit::new(
-                                arr1.get_unchecked(i).inlined_clone().$method(arr2.get_unchecked(i).inlined_clone()
-                            ));
+                        let out  = out.data.as_mut_slice_unchecked();
+                        for i in 0 .. arr1.len() {
+                            Status::init(out.get_unchecked_mut(i), arr1.get_unchecked(i).inlined_clone().$method(arr2.get_unchecked(i).inlined_clone()));
                         }
                     } else {
-                        for j in 0..self.ncols() {
-                            for i in 0..self.nrows() {
-                                *out.get_unchecked_mut((i, j)) = MaybeUninit::new(
-                                    self.get_unchecked((i, j)).inlined_clone().$method(rhs.get_unchecked((i, j)).inlined_clone())
-                                );
+                        for j in 0 .. self.ncols() {
+                            for i in 0 .. self.nrows() {
+                                let val = self.get_unchecked((i, j)).inlined_clone().$method(rhs.get_unchecked((i, j)).inlined_clone());
+                                Status::init(out.get_unchecked_mut((i, j)), val);
                             }
                         }
                     }
                 }
             }
 
+
             #[inline]
-            fn $method_assign_statically_unchecked<R2: Dim, C2: Dim, SB>(
-                &mut self, rhs: &Matrix<T, R2, C2, SB>
-            ) where
-                SA: StorageMut<T, R1, C1>,
-                SB: Storage<T, R2, C2>
-            {
+            fn $method_assign_statically_unchecked<R2, C2, SB>(&mut self, rhs: &Matrix<T, R2, C2, SB>)
+                where R2: Dim,
+                      C2: Dim,
+                      SA: StorageMut<T, R1, C1>,
+                      SB: Storage<T, R2, C2> {
                 assert_eq!(self.shape(), rhs.shape(), "Matrix addition/subtraction dimensions mismatch.");
 
                 // This is the most common case and should be deduced at compile-time.
@@ -210,12 +205,12 @@ macro_rules! componentwise_binop_impl(
                 }
             }
 
+
             #[inline]
-            fn $method_assign_statically_unchecked_rhs<R2: Dim, C2: Dim, SB>(
-                &self, rhs: &mut Matrix<T, R2, C2, SB>
-            ) where
-                SB: StorageMut<T, R2, C2>
-            {
+            fn $method_assign_statically_unchecked_rhs<R2, C2, SB>(&self, rhs: &mut Matrix<T, R2, C2, SB>)
+                where R2: Dim,
+                      C2: Dim,
+                      SB: StorageMut<T, R2, C2> {
                 assert_eq!(self.shape(), rhs.shape(), "Matrix addition/subtraction dimensions mismatch.");
 
                 // This is the most common case and should be deduced at compile-time.
@@ -250,20 +245,15 @@ macro_rules! componentwise_binop_impl(
              */
             /// Equivalent to `self + rhs` but stores the result into `out` to avoid allocations.
             #[inline]
-            pub fn $method_to<R2: Dim, C2: Dim, SB, R3: Dim, C3: Dim, SC>(
-                &self,
-                rhs: &Matrix<T, R2, C2, SB>,
-                out: &mut Matrix<MaybeUninit<T>, R3, C3, SC>
-            ) where
-                SB: Storage<T, R2, C2>,
-                SC: StorageMut<MaybeUninit<T>, R3, C3>,
-                ShapeConstraint:
-                    SameNumberOfRows<R1, R2> +
-                    SameNumberOfColumns<C1, C2> +
-                    SameNumberOfRows<R1, R3> +
-                    SameNumberOfColumns<C1, C3>
-            {
-                self.$method_to_statically_unchecked(rhs, out)
+            pub fn $method_to<R2: Dim, C2: Dim, SB,
+                              R3: Dim, C3: Dim, SC>(&self,
+                                                    rhs: &Matrix<T, R2, C2, SB>,
+                                                    out: &mut Matrix<T, R3, C3, SC>)
+                where SB: Storage<T, R2, C2>,
+                      SC: StorageMut<T, R3, C3>,
+                      ShapeConstraint: SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C2> +
+                                       SameNumberOfRows<R1, R3> + SameNumberOfColumns<C1, C3> {
+                self.$method_to_statically_unchecked_uninit(Init, rhs, out)
             }
         }
 
@@ -285,14 +275,13 @@ macro_rules! componentwise_binop_impl(
             }
         }
 
-        impl<'a, T, R1: Dim, C1: Dim, R2: Dim, C2: Dim, SA, SB> $Trait<Matrix<T, R2, C2, SB>> for &'a Matrix<T, R1, C1, SA>
-        where
-            T: Scalar + $bound,
-            SA: Storage<T, R1, C1>,
-            SB: Storage<T, R2, C2>,
-            DefaultAllocator: SameShapeAllocator<T, R2, C2, R1, C1>,
-            ShapeConstraint:  SameNumberOfRows<R2, R1> + SameNumberOfColumns<C2, C1>
-        {
+        impl<'a, T, R1, C1, R2, C2, SA, SB> $Trait<Matrix<T, R2, C2, SB>> for &'a Matrix<T, R1, C1, SA>
+            where R1: Dim, C1: Dim, R2: Dim, C2: Dim,
+                  T: Scalar + $bound,
+                  SA: Storage<T, R1, C1>,
+                  SB: Storage<T, R2, C2>,
+                  DefaultAllocator: SameShapeAllocator<T, R2, C2, R1, C1>,
+                  ShapeConstraint:  SameNumberOfRows<R2, R1> + SameNumberOfColumns<C2, C1> {
             type Output = MatrixSum<T, R2, C2, R1, C1>;
 
             #[inline]
@@ -304,14 +293,13 @@ macro_rules! componentwise_binop_impl(
             }
         }
 
-        impl<T, R1: Dim, C1: Dim, R2: Dim, C2: Dim, SA, SB> $Trait<Matrix<T, R2, C2, SB>> for Matrix<T, R1, C1, SA>
-        where
-            T: Scalar + $bound,
-            SA: Storage<T, R1, C1>,
-            SB: Storage<T, R2, C2>,
-            DefaultAllocator: SameShapeAllocator<T, R1, C1, R2, C2>,
-            ShapeConstraint:  SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C2>
-        {
+        impl<T, R1, C1, R2, C2, SA, SB> $Trait<Matrix<T, R2, C2, SB>> for Matrix<T, R1, C1, SA>
+            where R1: Dim, C1: Dim, R2: Dim, C2: Dim,
+                  T: Scalar + $bound,
+                  SA: Storage<T, R1, C1>,
+                  SB: Storage<T, R2, C2>,
+                  DefaultAllocator: SameShapeAllocator<T, R1, C1, R2, C2>,
+                  ShapeConstraint:  SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C2> {
             type Output = MatrixSum<T, R1, C1, R2, C2>;
 
             #[inline]
@@ -320,14 +308,13 @@ macro_rules! componentwise_binop_impl(
             }
         }
 
-        impl<'a, 'b, T, R1: Dim, C1: Dim, R2: Dim, C2: Dim, SA, SB> $Trait<&'b Matrix<T, R2, C2, SB>> for &'a Matrix<T, R1, C1, SA>
-        where
-            T: Scalar + $bound,
-            SA: Storage<T, R1, C1>,
-            SB: Storage<T, R2, C2>,
-            DefaultAllocator: SameShapeAllocator<T, R1, C1, R2, C2>,
-            ShapeConstraint: SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C2>
-        {
+        impl<'a, 'b, T, R1, C1, R2, C2, SA, SB> $Trait<&'b Matrix<T, R2, C2, SB>> for &'a Matrix<T, R1, C1, SA>
+            where R1: Dim, C1: Dim, R2: Dim, C2: Dim,
+                  T: Scalar + $bound,
+                  SA: Storage<T, R1, C1>,
+                  SB: Storage<T, R2, C2>,
+                  DefaultAllocator: SameShapeAllocator<T, R1, C1, R2, C2>,
+                  ShapeConstraint:  SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C2> {
             type Output = MatrixSum<T, R1, C1, R2, C2>;
 
             #[inline]
@@ -335,33 +322,33 @@ macro_rules! componentwise_binop_impl(
                 let (nrows, ncols) = self.shape();
                 let nrows: SameShapeR<R1, R2> = Dim::from_usize(nrows);
                 let ncols: SameShapeC<C1, C2> = Dim::from_usize(ncols);
-                let mut res = Matrix::new_uninitialized_generic(nrows, ncols);
-
-                self.$method_to_statically_unchecked(rhs, &mut res);
+                let mut res = Matrix::uninit(nrows, ncols);
+                self.$method_to_statically_unchecked_uninit(Uninit, rhs, &mut res);
+                // SAFETY: the output has been initialized above.
                 unsafe { res.assume_init() }
             }
         }
 
-        impl<'b, T, R1: Dim, C1: Dim, R2: Dim, C2: Dim, SA, SB> $TraitAssign<&'b Matrix<T, R2, C2, SB>> for Matrix<T, R1, C1, SA>
-        where
-            T: Scalar + $bound,
-            SA: StorageMut<T, R1, C1>,
-            SB: Storage<T, R2, C2>,
-            ShapeConstraint: SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C2>
-        {
+        impl<'b, T, R1, C1, R2, C2, SA, SB> $TraitAssign<&'b Matrix<T, R2, C2, SB>> for Matrix<T, R1, C1, SA>
+            where R1: Dim, C1: Dim, R2: Dim, C2: Dim,
+                  T: Scalar + $bound,
+                  SA: StorageMut<T, R1, C1>,
+                  SB: Storage<T, R2, C2>,
+                  ShapeConstraint: SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C2> {
+
             #[inline]
             fn $method_assign(&mut self, rhs: &'b Matrix<T, R2, C2, SB>) {
                 self.$method_assign_statically_unchecked(rhs)
             }
         }
 
-        impl<T, R1: Dim, C1: Dim, R2: Dim, C2: Dim, SA, SB> $TraitAssign<Matrix<T, R2, C2, SB>> for Matrix<T, R1, C1, SA>
-        where
-            T: Scalar + $bound,
-            SA: StorageMut<T, R1, C1>,
-            SB: Storage<T, R2, C2>,
-            ShapeConstraint: SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C2>
-        {
+        impl<T, R1, C1, R2, C2, SA, SB> $TraitAssign<Matrix<T, R2, C2, SB>> for Matrix<T, R1, C1, SA>
+            where R1: Dim, C1: Dim, R2: Dim, C2: Dim,
+                  T: Scalar + $bound,
+                  SA: StorageMut<T, R1, C1>,
+                  SB: Storage<T, R2, C2>,
+                  ShapeConstraint: SameNumberOfRows<R1, R2> + SameNumberOfColumns<C1, C2> {
+
             #[inline]
             fn $method_assign(&mut self, rhs: Matrix<T, R2, C2, SB>) {
                 self.$method_assign(&rhs)
@@ -372,10 +359,10 @@ macro_rules! componentwise_binop_impl(
 
 componentwise_binop_impl!(Add, add, ClosedAdd;
                           AddAssign, add_assign, add_assign_statically_unchecked, add_assign_statically_unchecked_mut;
-                          add_to, add_to_statically_unchecked);
+                          add_to, add_to_statically_unchecked_uninit);
 componentwise_binop_impl!(Sub, sub, ClosedSub;
                           SubAssign, sub_assign, sub_assign_statically_unchecked, sub_assign_statically_unchecked_mut;
-                          sub_to, sub_to_statically_unchecked);
+                          sub_to, sub_to_statically_unchecked_uninit);
 
 impl<T, R: DimName, C: DimName> iter::Sum for OMatrix<T, R, C>
 where
@@ -574,9 +561,12 @@ where
 
     #[inline]
     fn mul(self, rhs: &'b Matrix<T, R2, C2, SB>) -> Self::Output {
-        let mut res = Matrix::new_uninitialized_generic(self.data.shape().0, rhs.data.shape().1);
-        let _ = self.mul_to(rhs, &mut res);
-        unsafe { res.assume_init() }
+        let mut res = Matrix::uninit(self.shape_generic().0, rhs.shape_generic().1);
+        unsafe {
+            // SAFETY: this is OK because status = Uninit && bevy == 0
+            gemm_uninit(Uninit, &mut res, T::one(), self, rhs, T::zero());
+            res.assume_init()
+        }
     }
 }
 
@@ -634,14 +624,16 @@ where
 // TODO: this is too restrictive:
 //    − we can't use `a *= b` when `a` is a mutable slice.
 //    − we can't use `a *= b` when C2 is not equal to C1.
-impl<T, R1: Dim, C1: Dim, R2: Dim, SA, SB> MulAssign<Matrix<T, R2, C1, SB>>
-    for Matrix<T, R1, C1, SA>
+impl<T, R1, C1, R2, SA, SB> MulAssign<Matrix<T, R2, C1, SB>> for Matrix<T, R1, C1, SA>
 where
+    R1: Dim,
+    C1: Dim,
+    R2: Dim,
     T: Scalar + Zero + One + ClosedAdd + ClosedMul,
     SB: Storage<T, R2, C1>,
-    SA: ContiguousStorageMut<T, R1, C1>,
+    SA: StorageMut<T, R1, C1> + IsContiguous + Clone, // TODO: get rid of the IsContiguous
     ShapeConstraint: AreMultipliable<R1, C1, R2, C1>,
-    DefaultAllocator: Allocator<T, R1, C1> + InnerAllocator<T, R1, C1, Buffer = SA>,
+    DefaultAllocator: Allocator<T, R1, C1, Buffer = SA>,
 {
     #[inline]
     fn mul_assign(&mut self, rhs: Matrix<T, R2, C1, SB>) {
@@ -649,15 +641,17 @@ where
     }
 }
 
-impl<'b, T, R1: Dim, C1: Dim, R2: Dim, SA, SB> MulAssign<&'b Matrix<T, R2, C1, SB>>
-    for Matrix<T, R1, C1, SA>
+impl<'b, T, R1, C1, R2, SA, SB> MulAssign<&'b Matrix<T, R2, C1, SB>> for Matrix<T, R1, C1, SA>
 where
+    R1: Dim,
+    C1: Dim,
+    R2: Dim,
     T: Scalar + Zero + One + ClosedAdd + ClosedMul,
     SB: Storage<T, R2, C1>,
-    SA: ContiguousStorageMut<T, R1, C1>,
+    SA: StorageMut<T, R1, C1> + IsContiguous + Clone, // TODO: get rid of the IsContiguous
     ShapeConstraint: AreMultipliable<R1, C1, R2, C1>,
     // TODO: this is too restrictive. See comments for the non-ref version.
-    DefaultAllocator: Allocator<T, R1, C1> + InnerAllocator<T, R1, C1, Buffer = SA>,
+    DefaultAllocator: Allocator<T, R1, C1, Buffer = SA>,
 {
     #[inline]
     fn mul_assign(&mut self, rhs: &'b Matrix<T, R2, C1, SB>) {
@@ -680,8 +674,9 @@ where
         DefaultAllocator: Allocator<T, C1, C2>,
         ShapeConstraint: SameNumberOfRows<R1, R2>,
     {
-        let mut res = Matrix::new_uninitialized_generic(self.data.shape().1, rhs.data.shape().1);
-        self.tr_mul_to(rhs, &mut res);
+        let mut res = Matrix::uninit(self.shape_generic().1, rhs.shape_generic().1);
+        self.xx_mul_to_uninit(Uninit, rhs, &mut res, |a, b| a.dot(b));
+        // SAFETY: this is OK because the result is now initialized.
         unsafe { res.assume_init() }
     }
 
@@ -695,23 +690,26 @@ where
         DefaultAllocator: Allocator<T, C1, C2>,
         ShapeConstraint: SameNumberOfRows<R1, R2>,
     {
-        let mut res = Matrix::new_uninitialized_generic(self.data.shape().1, rhs.data.shape().1);
-        self.ad_mul_to(rhs, &mut res);
+        let mut res = Matrix::uninit(self.shape_generic().1, rhs.shape_generic().1);
+        self.xx_mul_to_uninit(Uninit, rhs, &mut res, |a, b| a.dotc(b));
+        // SAFETY: this is OK because the result is now initialized.
         unsafe { res.assume_init() }
     }
 
     #[inline(always)]
-    fn xx_mul_to<R2: Dim, C2: Dim, SB, R3: Dim, C3: Dim, SC>(
+    fn xx_mul_to_uninit<Status, R2: Dim, C2: Dim, SB, R3: Dim, C3: Dim, SC>(
         &self,
+        status: Status,
         rhs: &Matrix<T, R2, C2, SB>,
-        out: &mut Matrix<MaybeUninit<T>, R3, C3, SC>,
+        out: &mut Matrix<Status::Value, R3, C3, SC>,
         dot: impl Fn(
             &VectorSlice<'_, T, R1, SA::RStride, SA::CStride>,
             &VectorSlice<'_, T, R2, SB::RStride, SB::CStride>,
         ) -> T,
     ) where
-        SB: Storage<T, R2, C2>,
-        SC: StorageMut<MaybeUninit<T>, R3, C3>,
+        Status: InitStatus<T>,
+        SB: RawStorage<T, R2, C2>,
+        SC: RawStorageMut<Status::Value, R3, C3>,
         ShapeConstraint: SameNumberOfRows<R1, R2> + DimEq<C1, R3> + DimEq<C2, C3>,
     {
         let (nrows1, ncols1) = self.shape();
@@ -740,9 +738,8 @@ where
         for i in 0..ncols1 {
             for j in 0..ncols2 {
                 let dot = dot(&self.column(i), &rhs.column(j));
-                unsafe {
-                    *out.get_unchecked_mut((i, j)) = MaybeUninit::new(dot);
-                }
+                let elt = unsafe { out.get_unchecked_mut((i, j)) };
+                Status::init(elt, dot);
             }
         }
     }
@@ -753,13 +750,13 @@ where
     pub fn tr_mul_to<R2: Dim, C2: Dim, SB, R3: Dim, C3: Dim, SC>(
         &self,
         rhs: &Matrix<T, R2, C2, SB>,
-        out: &mut Matrix<MaybeUninit<T>, R3, C3, SC>,
+        out: &mut Matrix<T, R3, C3, SC>,
     ) where
         SB: Storage<T, R2, C2>,
-        SC: StorageMut<MaybeUninit<T>, R3, C3>,
+        SC: StorageMut<T, R3, C3>,
         ShapeConstraint: SameNumberOfRows<R1, R2> + DimEq<C1, R3> + DimEq<C2, C3>,
     {
-        self.xx_mul_to(rhs, out, |a, b| a.dot(b))
+        self.xx_mul_to_uninit(Init, rhs, out, |a, b| a.dot(b))
     }
 
     /// Equivalent to `self.adjoint() * rhs` but stores the result into `out` to avoid
@@ -768,31 +765,30 @@ where
     pub fn ad_mul_to<R2: Dim, C2: Dim, SB, R3: Dim, C3: Dim, SC>(
         &self,
         rhs: &Matrix<T, R2, C2, SB>,
-        out: &mut Matrix<MaybeUninit<T>, R3, C3, SC>,
+        out: &mut Matrix<T, R3, C3, SC>,
     ) where
         T: SimdComplexField,
         SB: Storage<T, R2, C2>,
-        SC: StorageMut<MaybeUninit<T>, R3, C3>,
+        SC: StorageMut<T, R3, C3>,
         ShapeConstraint: SameNumberOfRows<R1, R2> + DimEq<C1, R3> + DimEq<C2, C3>,
     {
-        self.xx_mul_to(rhs, out, |a, b| a.dotc(b))
+        self.xx_mul_to_uninit(Init, rhs, out, |a, b| a.dotc(b))
     }
 
     /// Equivalent to `self * rhs` but stores the result into `out` to avoid allocations.
     #[inline]
-    pub fn mul_to<'a, R2: Dim, C2: Dim, SB, R3: Dim, C3: Dim, SC>(
+    pub fn mul_to<R2: Dim, C2: Dim, SB, R3: Dim, C3: Dim, SC>(
         &self,
         rhs: &Matrix<T, R2, C2, SB>,
-        out: &'a mut Matrix<MaybeUninit<T>, R3, C3, SC>,
-    ) -> MatrixSliceMut<'a, T, R3, C3, SC::RStride, SC::CStride>
-    where
+        out: &mut Matrix<T, R3, C3, SC>,
+    ) where
         SB: Storage<T, R2, C2>,
-        SC: StorageMut<MaybeUninit<T>, R3, C3>,
+        SC: StorageMut<T, R3, C3>,
         ShapeConstraint: SameNumberOfRows<R3, R1>
             + SameNumberOfColumns<C3, C2>
             + AreMultipliable<R1, C1, R2, C2>,
     {
-        out.gemm_z(T::one(), self, rhs)
+        out.gemm(T::one(), self, rhs, T::zero());
     }
 
     /// The kronecker product of two matrices (aka. tensor product of the corresponding linear
@@ -809,34 +805,31 @@ where
         SB: Storage<T, R2, C2>,
         DefaultAllocator: Allocator<T, DimProd<R1, R2>, DimProd<C1, C2>>,
     {
-        let (nrows1, ncols1) = self.data.shape();
-        let (nrows2, ncols2) = rhs.data.shape();
+        let (nrows1, ncols1) = self.shape_generic();
+        let (nrows2, ncols2) = rhs.shape_generic();
 
-        let mut res = Matrix::new_uninitialized_generic(nrows1.mul(nrows2), ncols1.mul(ncols2));
-
-        {
-            let mut data_res = res.data.ptr_mut();
+        let mut res = Matrix::uninit(nrows1.mul(nrows2), ncols1.mul(ncols2));
+        let mut data_res = res.data.ptr_mut();
 
+        unsafe {
             for j1 in 0..ncols1.value() {
                 for j2 in 0..ncols2.value() {
                     for i1 in 0..nrows1.value() {
-                        unsafe {
-                            let coeff = self.get_unchecked((i1, j1)).inlined_clone();
+                        let coeff = self.get_unchecked((i1, j1)).inlined_clone();
 
-                            for i2 in 0..nrows2.value() {
-                                *data_res = MaybeUninit::new(
-                                    coeff.inlined_clone()
-                                        * rhs.get_unchecked((i2, j2)).inlined_clone(),
-                                );
-                                data_res = data_res.offset(1);
-                            }
+                        for i2 in 0..nrows2.value() {
+                            *data_res = MaybeUninit::new(
+                                coeff.inlined_clone() * rhs.get_unchecked((i2, j2)).inlined_clone(),
+                            );
+                            data_res = data_res.offset(1);
                         }
                     }
                 }
             }
-        }
 
-        unsafe { res.assume_init() }
+            // SAFETY: the result matrix has been initialized by the loop above.
+            res.assume_init()
+        }
     }
 }
 
diff --git a/src/base/properties.rs b/src/base/properties.rs
index 00333708..091d36ef 100644
--- a/src/base/properties.rs
+++ b/src/base/properties.rs
@@ -7,9 +7,10 @@ use simba::scalar::{ClosedAdd, ClosedMul, ComplexField, RealField};
 use crate::base::allocator::Allocator;
 use crate::base::dimension::{Dim, DimMin};
 use crate::base::storage::Storage;
-use crate::base::{DefaultAllocator, Matrix, SquareMatrix};
+use crate::base::{DefaultAllocator, Matrix, Scalar, SquareMatrix};
+use crate::RawStorage;
 
-impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T: Scalar, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     /// The total number of elements of this matrix.
     ///
     /// # Examples:
diff --git a/src/base/scalar.rs b/src/base/scalar.rs
index 80a78594..db9e458d 100644
--- a/src/base/scalar.rs
+++ b/src/base/scalar.rs
@@ -1,38 +1,27 @@
+use std::any::Any;
 use std::any::TypeId;
 use std::fmt::Debug;
 
-/// The basic scalar trait for all structures of `nalgebra`.
+/// The basic scalar type for all structures of `nalgebra`.
 ///
-/// This is by design a very loose trait, and does not make any assumption on
-/// the algebraic properties of `Self`. It has various purposes and objectives:
-/// - Enforces simple and future-proof trait bounds.
-/// - Enables important optimizations for floating point types via specialization.
-/// - Makes debugging generic code possible in most circumstances.
-pub trait Scalar: 'static + Clone + Debug {
+/// This does not make any assumption on the algebraic properties of `Self`.
+pub trait Scalar: Clone + PartialEq + Debug + Any {
     #[inline]
-    /// Tests whether `Self` is the same as the type `T`.
+    /// Tests if `Self` the same as the type `T`
     ///
-    /// Typically used to test of `Self` is an `f32` or an `f64`, which is
-    /// important as it allows for specialization and certain optimizations to
-    /// be made.
-    ///
-    // If the need ever arose to get rid of the `'static` requirement, we could
-    // merely replace this method by two unsafe associated methods `is_f32` and
-    // `is_f64`.
+    /// Typically used to test of `Self` is a f32 or a f64 with `T::is::<f32>()`.
     fn is<T: Scalar>() -> bool {
         TypeId::of::<Self>() == TypeId::of::<T>()
     }
 
-    /// Performance hack: Clone doesn't get inlined for Copy types in debug
-    /// mode, so make it inline anyway.
+    #[inline(always)]
+    /// Performance hack: Clone doesn't get inlined for Copy types in debug mode, so make it inline anyway.
     fn inlined_clone(&self) -> Self {
         self.clone()
     }
 }
 
-// Unfortunately, this blanket impl leads to many misleading compiler messages
-// telling you to implement Copy, even though Scalar is what's really needed.
-impl<T: 'static + Copy + Debug> Scalar for T {
+impl<T: Copy + PartialEq + Debug + Any> Scalar for T {
     #[inline(always)]
     fn inlined_clone(&self) -> T {
         *self
diff --git a/src/base/statistics.rs b/src/base/statistics.rs
index 84a6592a..ebf694a5 100644
--- a/src/base/statistics.rs
+++ b/src/base/statistics.rs
@@ -1,13 +1,12 @@
-use std::mem::MaybeUninit;
-
 use crate::allocator::Allocator;
-use crate::storage::Storage;
+use crate::storage::RawStorage;
 use crate::{Const, DefaultAllocator, Dim, Matrix, OVector, RowOVector, Scalar, VectorSlice, U1};
 use num::Zero;
 use simba::scalar::{ClosedAdd, Field, SupersetOf};
+use std::mem::MaybeUninit;
 
 /// # Folding on columns and rows
-impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T: Scalar, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     /// Returns a row vector where each element is the result of the application of `f` on the
     /// corresponding column of the original matrix.
     #[inline]
@@ -19,16 +18,18 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     where
         DefaultAllocator: Allocator<T, U1, C>,
     {
-        let ncols = self.data.shape().1;
-        let mut res = RowOVector::new_uninitialized_generic(Const::<1>, ncols);
+        let ncols = self.shape_generic().1;
+        let mut res = Matrix::uninit(Const::<1>, ncols);
 
         for i in 0..ncols.value() {
             // TODO: avoid bound checking of column.
+            // Safety: all indices are in range.
             unsafe {
                 *res.get_unchecked_mut((0, i)) = MaybeUninit::new(f(self.column(i)));
             }
         }
 
+        // Safety: res is now fully initialized.
         unsafe { res.assume_init() }
     }
 
@@ -45,16 +46,18 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     where
         DefaultAllocator: Allocator<T, C>,
     {
-        let ncols = self.data.shape().1;
-        let mut res = Matrix::new_uninitialized_generic(ncols, Const::<1>);
+        let ncols = self.shape_generic().1;
+        let mut res = Matrix::uninit(ncols, Const::<1>);
 
         for i in 0..ncols.value() {
             // TODO: avoid bound checking of column.
+            // Safety: all indices are in range.
             unsafe {
                 *res.vget_unchecked_mut(i) = MaybeUninit::new(f(self.column(i)));
             }
         }
 
+        // Safety: res is now fully initialized.
         unsafe { res.assume_init() }
     }
 
@@ -63,22 +66,24 @@ impl<T, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
     #[must_use]
     pub fn compress_columns(
         &self,
-        mut init: OVector<T, R>,
-        f: impl Fn(&mut OVector<T, R>, VectorSlice<T, R, S::RStride, S::CStride>),
+        init: OVector<T, R>,
+        f: impl Fn(&mut OVector<T, R>, VectorSlice<'_, T, R, S::RStride, S::CStride>),
     ) -> OVector<T, R>
     where
         DefaultAllocator: Allocator<T, R>,
     {
+        let mut res = init;
+
         for i in 0..self.ncols() {
-            f(&mut init, self.column(i))
+            f(&mut res, self.column(i))
         }
 
-        init
+        res
     }
 }
 
 /// # Common statistics operations
-impl<T: Scalar, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
+impl<T: Scalar, R: Dim, C: Dim, S: RawStorage<T, R, C>> Matrix<T, R, C, S> {
     /*
      *
      * Sum computation.
@@ -178,7 +183,7 @@ impl<T: Scalar, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         T: ClosedAdd + Zero,
         DefaultAllocator: Allocator<T, R>,
     {
-        let nrows = self.data.shape().0;
+        let nrows = self.shape_generic().0;
         self.compress_columns(OVector::zeros_generic(nrows, Const::<1>), |out, col| {
             *out += col;
         })
@@ -281,10 +286,10 @@ impl<T: Scalar, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         T: Field + SupersetOf<f64>,
         DefaultAllocator: Allocator<T, R>,
     {
-        let (nrows, ncols) = self.data.shape();
+        let (nrows, ncols) = self.shape_generic();
 
         let mut mean = self.column_mean();
-        mean.apply(|e| -(e.inlined_clone() * e));
+        mean.apply(|e| *e = -(e.inlined_clone() * e.inlined_clone()));
 
         let denom = T::one() / crate::convert::<_, T>(ncols.value() as f64);
         self.compress_columns(mean, |out, col| {
@@ -389,7 +394,7 @@ impl<T: Scalar, R: Dim, C: Dim, S: Storage<T, R, C>> Matrix<T, R, C, S> {
         T: Field + SupersetOf<f64>,
         DefaultAllocator: Allocator<T, R>,
     {
-        let (nrows, ncols) = self.data.shape();
+        let (nrows, ncols) = self.shape_generic();
         let denom = T::one() / crate::convert::<_, T>(ncols.value() as f64);
         self.compress_columns(OVector::zeros_generic(nrows, Const::<1>), |out, col| {
             out.axpy(denom.inlined_clone(), &col, T::one())
diff --git a/src/base/storage.rs b/src/base/storage.rs
index 1f06a11e..7ef7e152 100644
--- a/src/base/storage.rs
+++ b/src/base/storage.rs
@@ -2,27 +2,32 @@
 
 use std::ptr;
 
-use crate::base::allocator::{Allocator, InnerAllocator, SameShapeC, SameShapeR};
+use crate::base::allocator::{Allocator, SameShapeC, SameShapeR};
 use crate::base::default_allocator::DefaultAllocator;
 use crate::base::dimension::{Dim, U1};
-use crate::base::Owned;
+use crate::base::Scalar;
 
 /*
  * Aliases for allocation results.
  */
+/// The data storage for the sum of two matrices with dimensions `(R1, C1)` and `(R2, C2)`.
+pub type SameShapeStorage<T, R1, C1, R2, C2> =
+    <DefaultAllocator as Allocator<T, SameShapeR<R1, R2>, SameShapeC<C1, C2>>>::Buffer;
 
 // TODO: better name than Owned ?
 /// The owned data storage that can be allocated from `S`.
-pub type InnerOwned<T, R, C = U1> = <DefaultAllocator as InnerAllocator<T, R, C>>::Buffer;
+pub type Owned<T, R, C = U1> = <DefaultAllocator as Allocator<T, R, C>>::Buffer;
 
-/// The data storage for the sum of two matrices with dimensions `(R1, C1)` and `(R2, C2)`.
-pub type SameShapeStorage<T, R1, C1, R2, C2> = Owned<T, SameShapeR<R1, R2>, SameShapeC<C1, C2>>;
+/// The owned data storage that can be allocated from `S`.
+pub type OwnedUninit<T, R, C = U1> = <DefaultAllocator as Allocator<T, R, C>>::BufferUninit;
 
 /// The row-stride of the owned data storage for a buffer of dimension `(R, C)`.
-pub type RStride<T, R, C = U1> = <InnerOwned<T, R, C> as Storage<T, R, C>>::RStride;
+pub type RStride<T, R, C = U1> =
+    <<DefaultAllocator as Allocator<T, R, C>>::Buffer as RawStorage<T, R, C>>::RStride;
 
 /// The column-stride of the owned data storage for a buffer of dimension `(R, C)`.
-pub type CStride<T, R, C = U1> = <InnerOwned<T, R, C> as Storage<T, R, C>>::CStride;
+pub type CStride<T, R, C = U1> =
+    <<DefaultAllocator as Allocator<T, R, C>>::Buffer as RawStorage<T, R, C>>::CStride;
 
 /// The trait shared by all matrix data storage.
 ///
@@ -33,7 +38,7 @@ pub type CStride<T, R, C = U1> = <InnerOwned<T, R, C> as Storage<T, R, C>>::CStr
 /// should **not** allow the user to modify the size of the underlying buffer with safe methods
 /// (for example the `VecStorage::data_mut` method is unsafe because the user could change the
 /// vector's size so that it no longer contains enough elements: this will lead to UB.
-pub unsafe trait Storage<T, R: Dim, C: Dim = U1>: Sized {
+pub unsafe trait RawStorage<T, R: Dim, C: Dim = U1>: Sized {
     /// The static stride of this storage's rows.
     type RStride: Dim;
 
@@ -118,17 +123,17 @@ pub unsafe trait Storage<T, R: Dim, C: Dim = U1>: Sized {
     ///
     /// Call the safe alternative `matrix.as_slice()` instead.
     unsafe fn as_slice_unchecked(&self) -> &[T];
+}
 
+pub unsafe trait Storage<T, R: Dim, C: Dim = U1>: RawStorage<T, R, C> {
     /// Builds a matrix data storage that does not contain any reference.
     fn into_owned(self) -> Owned<T, R, C>
     where
-        T: Clone,
         DefaultAllocator: Allocator<T, R, C>;
 
     /// Clones this data storage to one that does not contain any reference.
     fn clone_owned(&self) -> Owned<T, R, C>
     where
-        T: Clone,
         DefaultAllocator: Allocator<T, R, C>;
 }
 
@@ -137,7 +142,7 @@ pub unsafe trait Storage<T, R: Dim, C: Dim = U1>: Sized {
 /// Note that a mutable access does not mean that the matrix owns its data. For example, a mutable
 /// matrix slice can provide mutable access to its elements even if it does not own its data (it
 /// contains only an internal reference to them).
-pub unsafe trait StorageMut<T, R: Dim, C: Dim = U1>: Storage<T, R, C> {
+pub unsafe trait RawStorageMut<T, R: Dim, C: Dim = U1>: RawStorage<T, R, C> {
     /// The matrix mutable data pointer.
     fn ptr_mut(&mut self) -> *mut T;
 
@@ -212,40 +217,37 @@ pub unsafe trait StorageMut<T, R: Dim, C: Dim = U1>: Storage<T, R, C> {
     unsafe fn as_mut_slice_unchecked(&mut self) -> &mut [T];
 }
 
-/// A matrix storage that is stored contiguously in memory.
-///
-/// The storage requirement means that for any value of `i` in `[0, nrows * ncols - 1]`, the value
-/// `.get_unchecked_linear` returns one of the matrix component. This trait is unsafe because
-/// failing to comply to this may cause Undefined Behaviors.
-pub unsafe trait ContiguousStorage<T, R: Dim, C: Dim = U1>: Storage<T, R, C> {
-    /// Converts this data storage to a contiguous slice.
-    fn as_slice(&self) -> &[T] {
-        // SAFETY: this is safe because this trait guarantees the fact
-        //         that the data is stored contiguously.
-        unsafe { self.as_slice_unchecked() }
-    }
+pub unsafe trait StorageMut<T, R: Dim, C: Dim = U1>:
+    Storage<T, R, C> + RawStorageMut<T, R, C>
+{
 }
 
-/// A mutable matrix storage that is stored contiguously in memory.
+unsafe impl<S, T, R, C> StorageMut<T, R, C> for S
+where
+    R: Dim,
+    C: Dim,
+    S: Storage<T, R, C> + RawStorageMut<T, R, C>,
+{
+}
+
+/// Marker trait indicating that a storage is stored contiguously in memory.
 ///
 /// The storage requirement means that for any value of `i` in `[0, nrows * ncols - 1]`, the value
 /// `.get_unchecked_linear` returns one of the matrix component. This trait is unsafe because
 /// failing to comply to this may cause Undefined Behaviors.
-pub unsafe trait ContiguousStorageMut<T, R: Dim, C: Dim = U1>:
-    ContiguousStorage<T, R, C> + StorageMut<T, R, C>
-{
-    /// Converts this data storage to a contiguous mutable slice.
-    fn as_mut_slice(&mut self) -> &mut [T] {
-        // SAFETY: this is safe because this trait guarantees the fact
-        //         that the data is stored contiguously.
-        unsafe { self.as_mut_slice_unchecked() }
-    }
-}
+pub unsafe trait IsContiguous {}
 
 /// A matrix storage that can be reshaped in-place.
-pub trait ReshapableStorage<T, R1: Dim, C1: Dim, R2: Dim, C2: Dim>: Storage<T, R1, C1> {
+pub trait ReshapableStorage<T, R1, C1, R2, C2>: RawStorage<T, R1, C1>
+where
+    T: Scalar,
+    R1: Dim,
+    C1: Dim,
+    R2: Dim,
+    C2: Dim,
+{
     /// The reshaped storage type.
-    type Output: Storage<T, R2, C2>;
+    type Output: RawStorage<T, R2, C2>;
 
     /// Reshapes the storage into the output storage type.
     fn reshape_generic(self, nrows: R2, ncols: C2) -> Self::Output;
diff --git a/src/base/swizzle.rs b/src/base/swizzle.rs
index 0c471301..6ed05d81 100644
--- a/src/base/swizzle.rs
+++ b/src/base/swizzle.rs
@@ -1,5 +1,5 @@
-use crate::base::{DimName,  ToTypenum, Vector, Vector2, Vector3};
-use crate::storage::Storage;
+use crate::base::{DimName, Scalar, ToTypenum, Vector, Vector2, Vector3};
+use crate::storage::RawStorage;
 use typenum::{self, Cmp, Greater};
 
 macro_rules! impl_swizzle {
@@ -11,7 +11,7 @@ macro_rules! impl_swizzle {
                 #[must_use]
                 pub fn $name(&self) -> $Result<T>
                 where D::Typenum: Cmp<typenum::$BaseDim, Output=Greater> {
-                    $Result::new($(self[$i].clone()),*)
+                    $Result::new($(self[$i].inlined_clone()),*)
                 }
             )*
         )*
@@ -19,7 +19,7 @@ macro_rules! impl_swizzle {
 }
 
 /// # Swizzling
-impl<T: Clone, D, S: Storage<T, D>> Vector<T, D, S>
+impl<T: Scalar, D, S: RawStorage<T, D>> Vector<T, D, S>
 where
     D: DimName + ToTypenum,
 {
diff --git a/src/base/uninit.rs b/src/base/uninit.rs
new file mode 100644
index 00000000..7fc5f84e
--- /dev/null
+++ b/src/base/uninit.rs
@@ -0,0 +1,76 @@
+use std::mem::MaybeUninit;
+
+// # Safety
+// This trait must not be implemented outside of this crate.
+pub unsafe trait InitStatus<T>: Copy {
+    type Value;
+    fn init(out: &mut Self::Value, t: T);
+    unsafe fn assume_init_ref(t: &Self::Value) -> &T;
+    unsafe fn assume_init_mut(t: &mut Self::Value) -> &mut T;
+}
+
+#[derive(Copy, Clone, Debug, PartialEq, Eq)]
+pub struct Init;
+#[derive(Copy, Clone, Debug, PartialEq, Eq)]
+pub struct Uninit;
+#[derive(Copy, Clone, Debug, PartialEq, Eq)]
+pub struct Initialized<Status>(pub Status);
+
+unsafe impl<T> InitStatus<T> for Init {
+    type Value = T;
+
+    #[inline(always)]
+    fn init(out: &mut T, t: T) {
+        *out = t;
+    }
+
+    #[inline(always)]
+    unsafe fn assume_init_ref(t: &T) -> &T {
+        t
+    }
+
+    #[inline(always)]
+    unsafe fn assume_init_mut(t: &mut T) -> &mut T {
+        t
+    }
+}
+
+unsafe impl<T> InitStatus<T> for Uninit {
+    type Value = MaybeUninit<T>;
+
+    #[inline(always)]
+    fn init(out: &mut MaybeUninit<T>, t: T) {
+        *out = MaybeUninit::new(t);
+    }
+
+    #[inline(always)]
+    unsafe fn assume_init_ref(t: &MaybeUninit<T>) -> &T {
+        std::mem::transmute(t.as_ptr()) // TODO: use t.assume_init_ref()
+    }
+
+    #[inline(always)]
+    unsafe fn assume_init_mut(t: &mut MaybeUninit<T>) -> &mut T {
+        std::mem::transmute(t.as_mut_ptr()) // TODO: use t.assume_init_mut()
+    }
+}
+
+unsafe impl<T, Status: InitStatus<T>> InitStatus<T> for Initialized<Status> {
+    type Value = Status::Value;
+
+    #[inline(always)]
+    fn init(out: &mut Status::Value, t: T) {
+        unsafe {
+            *Status::assume_init_mut(out) = t;
+        }
+    }
+
+    #[inline(always)]
+    unsafe fn assume_init_ref(t: &Status::Value) -> &T {
+        Status::assume_init_ref(t)
+    }
+
+    #[inline(always)]
+    unsafe fn assume_init_mut(t: &mut Status::Value) -> &mut T {
+        Status::assume_init_mut(t)
+    }
+}
diff --git a/src/base/unit.rs b/src/base/unit.rs
index 73fcd6dd..fa869c09 100644
--- a/src/base/unit.rs
+++ b/src/base/unit.rs
@@ -10,7 +10,7 @@ use abomonation::Abomonation;
 
 use crate::allocator::Allocator;
 use crate::base::DefaultAllocator;
-use crate::storage::{InnerOwned, Storage};
+use crate::storage::RawStorage;
 use crate::{Dim, Matrix, OMatrix, RealField, Scalar, SimdComplexField, SimdRealField};
 
 /// A wrapper that ensures the underlying algebraic entity has a unit norm.
@@ -113,10 +113,10 @@ mod rkyv_impl {
 
 impl<T, R, C, S> PartialEq for Unit<Matrix<T, R, C, S>>
 where
-    T: PartialEq,
+    T: Scalar + PartialEq,
     R: Dim,
     C: Dim,
-    S: Storage<T, R, C>,
+    S: RawStorage<T, R, C>,
 {
     #[inline]
     fn eq(&self, rhs: &Self) -> bool {
@@ -126,10 +126,10 @@ where
 
 impl<T, R, C, S> Eq for Unit<Matrix<T, R, C, S>>
 where
-    T: Eq,
+    T: Scalar + Eq,
     R: Dim,
     C: Dim,
-    S: Storage<T, R, C>,
+    S: RawStorage<T, R, C>,
 {
 }
 
@@ -228,7 +228,7 @@ impl<T> Unit<T> {
     /// Wraps the given reference, assuming it is already normalized.
     #[inline]
     pub fn from_ref_unchecked(value: &T) -> &Self {
-        unsafe { &*(value as *const _ as *const _) }
+        unsafe { &*(value as *const T as *const Self) }
     }
 
     /// Retrieves the underlying value.
@@ -331,7 +331,7 @@ impl<T> Deref for Unit<T> {
 
     #[inline]
     fn deref(&self) -> &T {
-        unsafe { &*(self as *const _ as *const T) }
+        unsafe { &*(self as *const Self as *const T) }
     }
 }
 
@@ -344,7 +344,6 @@ where
     T: From<[<T as simba::simd::SimdValue>::Element; 2]>,
     T::Element: Scalar,
     DefaultAllocator: Allocator<T, R, C> + Allocator<T::Element, R, C>,
-    InnerOwned<T::Element, R, C>: Clone,
 {
     #[inline]
     fn from(arr: [Unit<OMatrix<T::Element, R, C>>; 2]) -> Self {
@@ -361,7 +360,6 @@ where
     T: From<[<T as simba::simd::SimdValue>::Element; 4]>,
     T::Element: Scalar,
     DefaultAllocator: Allocator<T, R, C> + Allocator<T::Element, R, C>,
-    InnerOwned<T::Element, R, C>: Clone,
 {
     #[inline]
     fn from(arr: [Unit<OMatrix<T::Element, R, C>>; 4]) -> Self {
@@ -380,7 +378,6 @@ where
     T: From<[<T as simba::simd::SimdValue>::Element; 8]>,
     T::Element: Scalar,
     DefaultAllocator: Allocator<T, R, C> + Allocator<T::Element, R, C>,
-    InnerOwned<T::Element, R, C>: Clone,
 {
     #[inline]
     fn from(arr: [Unit<OMatrix<T::Element, R, C>>; 8]) -> Self {
@@ -403,7 +400,6 @@ where
     T: From<[<T as simba::simd::SimdValue>::Element; 16]>,
     T::Element: Scalar,
     DefaultAllocator: Allocator<T, R, C> + Allocator<T::Element, R, C>,
-    InnerOwned<T::Element, R, C>: Clone,
 {
     #[inline]
     fn from(arr: [Unit<OMatrix<T::Element, R, C>>; 16]) -> Self {
diff --git a/src/base/vec_storage.rs b/src/base/vec_storage.rs
index 494e2090..f5b0b01c 100644
--- a/src/base/vec_storage.rs
+++ b/src/base/vec_storage.rs
@@ -4,14 +4,12 @@ use std::io::{Result as IOResult, Write};
 #[cfg(all(feature = "alloc", not(feature = "std")))]
 use alloc::vec::Vec;
 
-use crate::allocator::InnerAllocator;
+use crate::base::allocator::Allocator;
 use crate::base::constraint::{SameNumberOfRows, ShapeConstraint};
 use crate::base::default_allocator::DefaultAllocator;
 use crate::base::dimension::{Dim, DimName, Dynamic, U1};
-use crate::base::storage::{
-    ContiguousStorage, ContiguousStorageMut, ReshapableStorage, Storage, StorageMut,
-};
-use crate::base::{Owned, Vector};
+use crate::base::storage::{IsContiguous, Owned, RawStorage, RawStorageMut, ReshapableStorage};
+use crate::base::{Scalar, Vector};
 
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{
@@ -19,20 +17,22 @@ use serde::{
     ser::{Serialize, Serializer},
 };
 
+use crate::Storage;
 #[cfg(feature = "abomonation-serialize")]
 use abomonation::Abomonation;
 
 /*
  *
- * Storage.
+ * RawStorage.
  *
  */
 /// A Vec-based matrix data storage. It may be dynamically-sized.
+#[repr(C)]
 #[derive(Eq, Debug, Clone, PartialEq)]
 pub struct VecStorage<T, R: Dim, C: Dim> {
     data: Vec<T>,
-    pub(crate) nrows: R,
-    pub(crate) ncols: C,
+    nrows: R,
+    ncols: C,
 }
 
 #[cfg(feature = "serde-serialize")]
@@ -142,6 +142,18 @@ impl<T, R: Dim, C: Dim> VecStorage<T, R, C> {
     pub fn is_empty(&self) -> bool {
         self.len() == 0
     }
+
+    /// A slice containing all the components stored in this storage in column-major order.
+    #[inline]
+    pub fn as_slice(&self) -> &[T] {
+        &self.data[..]
+    }
+
+    /// A mutable slice containing all the components stored in this storage in column-major order.
+    #[inline]
+    pub fn as_mut_slice(&mut self) -> &mut [T] {
+        &mut self.data[..]
+    }
 }
 
 impl<T, R: Dim, C: Dim> From<VecStorage<T, R, C>> for Vec<T> {
@@ -156,10 +168,7 @@ impl<T, R: Dim, C: Dim> From<VecStorage<T, R, C>> for Vec<T> {
  * Dynamic − Dynamic
  *
  */
-unsafe impl<T, C: Dim> Storage<T, Dynamic, C> for VecStorage<T, Dynamic, C>
-where
-    DefaultAllocator: InnerAllocator<T, Dynamic, C, Buffer = Self>,
-{
+unsafe impl<T, C: Dim> RawStorage<T, Dynamic, C> for VecStorage<T, Dynamic, C> {
     type RStride = U1;
     type CStride = Dynamic;
 
@@ -183,29 +192,34 @@ where
         true
     }
 
-    #[inline]
-    fn into_owned(self) -> Owned<T, Dynamic, C> {
-        Owned(self)
-    }
-
-    #[inline]
-    fn clone_owned(&self) -> Owned<T, Dynamic, C>
-    where
-        T: Clone,
-    {
-        Owned(self.clone())
-    }
-
     #[inline]
     unsafe fn as_slice_unchecked(&self) -> &[T] {
         &self.data
     }
 }
 
-unsafe impl<T, R: DimName> Storage<T, R, Dynamic> for VecStorage<T, R, Dynamic>
+unsafe impl<T: Scalar, C: Dim> Storage<T, Dynamic, C> for VecStorage<T, Dynamic, C>
 where
-    DefaultAllocator: InnerAllocator<T, R, Dynamic, Buffer = Self>,
+    DefaultAllocator: Allocator<T, Dynamic, C, Buffer = Self>,
 {
+    #[inline]
+    fn into_owned(self) -> Owned<T, Dynamic, C>
+    where
+        DefaultAllocator: Allocator<T, Dynamic, C>,
+    {
+        self
+    }
+
+    #[inline]
+    fn clone_owned(&self) -> Owned<T, Dynamic, C>
+    where
+        DefaultAllocator: Allocator<T, Dynamic, C>,
+    {
+        self.clone()
+    }
+}
+
+unsafe impl<T, R: DimName> RawStorage<T, R, Dynamic> for VecStorage<T, R, Dynamic> {
     type RStride = U1;
     type CStride = R;
 
@@ -229,34 +243,39 @@ where
         true
     }
 
-    #[inline]
-    fn into_owned(self) -> Owned<T, R, Dynamic> {
-        Owned(self)
-    }
-
-    #[inline]
-    fn clone_owned(&self) -> Owned<T, R, Dynamic>
-    where
-        T: Clone,
-    {
-        Owned(self.clone())
-    }
-
     #[inline]
     unsafe fn as_slice_unchecked(&self) -> &[T] {
         &self.data
     }
 }
 
+unsafe impl<T: Scalar, R: DimName> Storage<T, R, Dynamic> for VecStorage<T, R, Dynamic>
+where
+    DefaultAllocator: Allocator<T, R, Dynamic, Buffer = Self>,
+{
+    #[inline]
+    fn into_owned(self) -> Owned<T, R, Dynamic>
+    where
+        DefaultAllocator: Allocator<T, R, Dynamic>,
+    {
+        self
+    }
+
+    #[inline]
+    fn clone_owned(&self) -> Owned<T, R, Dynamic>
+    where
+        DefaultAllocator: Allocator<T, R, Dynamic>,
+    {
+        self.clone()
+    }
+}
+
 /*
  *
- * StorageMut, ContiguousStorage.
+ * RawStorageMut, ContiguousStorage.
  *
  */
-unsafe impl<T, C: Dim> StorageMut<T, Dynamic, C> for VecStorage<T, Dynamic, C>
-where
-    DefaultAllocator: InnerAllocator<T, Dynamic, C, Buffer = Self>,
-{
+unsafe impl<T, C: Dim> RawStorageMut<T, Dynamic, C> for VecStorage<T, Dynamic, C> {
     #[inline]
     fn ptr_mut(&mut self) -> *mut T {
         self.data.as_mut_ptr()
@@ -268,18 +287,13 @@ where
     }
 }
 
-unsafe impl<T, C: Dim> ContiguousStorage<T, Dynamic, C> for VecStorage<T, Dynamic, C> where
-    DefaultAllocator: InnerAllocator<T, Dynamic, C, Buffer = Self>
-{
-}
+unsafe impl<T, R: Dim, C: Dim> IsContiguous for VecStorage<T, R, C> {}
 
-unsafe impl<T, C: Dim> ContiguousStorageMut<T, Dynamic, C> for VecStorage<T, Dynamic, C> where
-    DefaultAllocator: InnerAllocator<T, Dynamic, C, Buffer = Self>
-{
-}
-
-impl<T, C1: Dim, C2: Dim> ReshapableStorage<T, Dynamic, C1, Dynamic, C2>
-    for VecStorage<T, Dynamic, C1>
+impl<T, C1, C2> ReshapableStorage<T, Dynamic, C1, Dynamic, C2> for VecStorage<T, Dynamic, C1>
+where
+    T: Scalar,
+    C1: Dim,
+    C2: Dim,
 {
     type Output = VecStorage<T, Dynamic, C2>;
 
@@ -293,8 +307,11 @@ impl<T, C1: Dim, C2: Dim> ReshapableStorage<T, Dynamic, C1, Dynamic, C2>
     }
 }
 
-impl<T, C1: Dim, R2: DimName> ReshapableStorage<T, Dynamic, C1, R2, Dynamic>
-    for VecStorage<T, Dynamic, C1>
+impl<T, C1, R2> ReshapableStorage<T, Dynamic, C1, R2, Dynamic> for VecStorage<T, Dynamic, C1>
+where
+    T: Scalar,
+    C1: Dim,
+    R2: DimName,
 {
     type Output = VecStorage<T, R2, Dynamic>;
 
@@ -308,10 +325,7 @@ impl<T, C1: Dim, R2: DimName> ReshapableStorage<T, Dynamic, C1, R2, Dynamic>
     }
 }
 
-unsafe impl<T, R: DimName> StorageMut<T, R, Dynamic> for VecStorage<T, R, Dynamic>
-where
-    DefaultAllocator: InnerAllocator<T, R, Dynamic, Buffer = Self>,
-{
+unsafe impl<T, R: DimName> RawStorageMut<T, R, Dynamic> for VecStorage<T, R, Dynamic> {
     #[inline]
     fn ptr_mut(&mut self) -> *mut T {
         self.data.as_mut_ptr()
@@ -323,8 +337,11 @@ where
     }
 }
 
-impl<T, R1: DimName, C2: Dim> ReshapableStorage<T, R1, Dynamic, Dynamic, C2>
-    for VecStorage<T, R1, Dynamic>
+impl<T, R1, C2> ReshapableStorage<T, R1, Dynamic, Dynamic, C2> for VecStorage<T, R1, Dynamic>
+where
+    T: Scalar,
+    R1: DimName,
+    C2: Dim,
 {
     type Output = VecStorage<T, Dynamic, C2>;
 
@@ -338,8 +355,11 @@ impl<T, R1: DimName, C2: Dim> ReshapableStorage<T, R1, Dynamic, Dynamic, C2>
     }
 }
 
-impl<T, R1: DimName, R2: DimName> ReshapableStorage<T, R1, Dynamic, R2, Dynamic>
-    for VecStorage<T, R1, Dynamic>
+impl<T, R1, R2> ReshapableStorage<T, R1, Dynamic, R2, Dynamic> for VecStorage<T, R1, Dynamic>
+where
+    T: Scalar,
+    R1: DimName,
+    R2: DimName,
 {
     type Output = VecStorage<T, R2, Dynamic>;
 
@@ -368,16 +388,6 @@ impl<T: Abomonation, R: Dim, C: Dim> Abomonation for VecStorage<T, R, C> {
     }
 }
 
-unsafe impl<T, R: DimName> ContiguousStorage<T, R, Dynamic> for VecStorage<T, R, Dynamic> where
-    DefaultAllocator: InnerAllocator<T, R, Dynamic, Buffer = Self>
-{
-}
-
-unsafe impl<T, R: DimName> ContiguousStorageMut<T, R, Dynamic> for VecStorage<T, R, Dynamic> where
-    DefaultAllocator: InnerAllocator<T, R, Dynamic, Buffer = Self>
-{
-}
-
 impl<T, R: Dim> Extend<T> for VecStorage<T, R, Dynamic> {
     /// Extends the number of columns of the `VecStorage` with elements
     /// from the given iterator.
@@ -407,9 +417,12 @@ impl<'a, T: 'a + Copy, R: Dim> Extend<&'a T> for VecStorage<T, R, Dynamic> {
     }
 }
 
-impl<T: Clone, R: Dim, RV: Dim, SV> Extend<Vector<T, RV, SV>> for VecStorage<T, R, Dynamic>
+impl<T, R, RV, SV> Extend<Vector<T, RV, SV>> for VecStorage<T, R, Dynamic>
 where
-    SV: Storage<T, RV>,
+    T: Scalar,
+    R: Dim,
+    RV: Dim,
+    SV: RawStorage<T, RV>,
     ShapeConstraint: SameNumberOfRows<R, RV>,
 {
     /// Extends the number of columns of the `VecStorage` with vectors
diff --git a/src/debug/random_orthogonal.rs b/src/debug/random_orthogonal.rs
index 2cfbec26..c9684238 100644
--- a/src/debug/random_orthogonal.rs
+++ b/src/debug/random_orthogonal.rs
@@ -1,52 +1,24 @@
-use std::fmt;
-
+#[cfg(feature = "arbitrary")]
+use crate::base::storage::Owned;
 #[cfg(feature = "arbitrary")]
 use quickcheck::{Arbitrary, Gen};
 
 use crate::base::allocator::Allocator;
-use crate::base::dimension::{Dim, DimName, Dynamic};
+use crate::base::dimension::{Dim, Dynamic};
+use crate::base::Scalar;
 use crate::base::{DefaultAllocator, OMatrix};
 use crate::linalg::givens::GivensRotation;
-use crate::storage::Owned;
 use simba::scalar::ComplexField;
 
 /// A random orthogonal matrix.
-pub struct RandomOrthogonal<T, D: Dim = Dynamic>
+#[derive(Clone, Debug)]
+pub struct RandomOrthogonal<T: Scalar, D: Dim = Dynamic>
 where
     DefaultAllocator: Allocator<T, D, D>,
 {
     m: OMatrix<T, D, D>,
 }
 
-impl<T: Copy, D: DimName> Copy for RandomOrthogonal<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D>,
-    Owned<T, D, D>: Copy,
-{
-}
-
-impl<T: Clone, D: Dim> Clone for RandomOrthogonal<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D>,
-    Owned<T, D, D>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self { m: self.m.clone() }
-    }
-}
-
-impl<T: fmt::Debug, D: Dim> fmt::Debug for RandomOrthogonal<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D>,
-    Owned<T, D, D>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("RandomOrthogonal")
-            .field("m", &self.m)
-            .finish()
-    }
-}
-
 impl<T: ComplexField, D: Dim> RandomOrthogonal<T, D>
 where
     DefaultAllocator: Allocator<T, D, D>,
diff --git a/src/debug/random_sdp.rs b/src/debug/random_sdp.rs
index 3e119946..a915f2fc 100644
--- a/src/debug/random_sdp.rs
+++ b/src/debug/random_sdp.rs
@@ -1,50 +1,25 @@
-use std::fmt;
-
+#[cfg(feature = "arbitrary")]
+use crate::base::storage::Owned;
 #[cfg(feature = "arbitrary")]
 use quickcheck::{Arbitrary, Gen};
 
 use crate::base::allocator::Allocator;
 use crate::base::dimension::{Dim, Dynamic};
-use crate::base::{DefaultAllocator, OMatrix, Owned};
+use crate::base::Scalar;
+use crate::base::{DefaultAllocator, OMatrix};
 use simba::scalar::ComplexField;
 
 use crate::debug::RandomOrthogonal;
 
 /// A random, well-conditioned, symmetric definite-positive matrix.
-pub struct RandomSDP<T, D: Dim = Dynamic>
+#[derive(Clone, Debug)]
+pub struct RandomSDP<T: Scalar, D: Dim = Dynamic>
 where
     DefaultAllocator: Allocator<T, D, D>,
 {
     m: OMatrix<T, D, D>,
 }
 
-impl<T: Copy, D: Dim> Copy for RandomSDP<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D>,
-    Owned<T, D, D>: Copy,
-{
-}
-
-impl<T: Clone, D: Dim> Clone for RandomSDP<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D>,
-    Owned<T, D, D>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self { m: self.m.clone() }
-    }
-}
-
-impl<T: fmt::Debug, D: Dim> fmt::Debug for RandomSDP<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D>,
-    Owned<T, D, D>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("RandomSDP").field("m", &self.m).finish()
-    }
-}
-
 impl<T: ComplexField, D: Dim> RandomSDP<T, D>
 where
     DefaultAllocator: Allocator<T, D, D>,
diff --git a/src/geometry/dual_quaternion.rs b/src/geometry/dual_quaternion.rs
index 6ad5bef5..6dd8936d 100644
--- a/src/geometry/dual_quaternion.rs
+++ b/src/geometry/dual_quaternion.rs
@@ -2,15 +2,15 @@
 #![allow(clippy::op_ref)]
 
 use crate::{
-    Isometry3, Matrix4, Normed, OVector, Point3, Quaternion, SimdRealField, Translation3, Unit,
-    UnitQuaternion, Vector3, Zero, U8,
+    Isometry3, Matrix4, Normed, OVector, Point3, Quaternion, Scalar, SimdRealField, Translation3,
+    Unit, UnitQuaternion, Vector3, Zero, U8,
 };
 use approx::{AbsDiffEq, RelativeEq, UlpsEq};
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
 use std::fmt;
 
-use simba::scalar::RealField;
+use simba::scalar::{ClosedNeg, RealField};
 
 /// A dual quaternion.
 ///
@@ -46,16 +46,16 @@ pub struct DualQuaternion<T> {
     pub dual: Quaternion<T>,
 }
 
-impl<T: Eq> Eq for DualQuaternion<T> {}
+impl<T: Scalar + Eq> Eq for DualQuaternion<T> {}
 
-impl<T: PartialEq> PartialEq for DualQuaternion<T> {
+impl<T: Scalar> PartialEq for DualQuaternion<T> {
     #[inline]
     fn eq(&self, right: &Self) -> bool {
         self.real == right.real && self.dual == right.dual
     }
 }
 
-impl<T: Zero + Clone> Default for DualQuaternion<T> {
+impl<T: Scalar + Zero> Default for DualQuaternion<T> {
     fn default() -> Self {
         Self {
             real: Quaternion::default(),
@@ -267,7 +267,10 @@ where
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<T: Serialize> Serialize for DualQuaternion<T> {
+impl<T: SimdRealField> Serialize for DualQuaternion<T>
+where
+    T: Serialize,
+{
     fn serialize<S>(&self, serializer: S) -> Result<<S as Serializer>::Ok, <S as Serializer>::Error>
     where
         S: Serializer,
@@ -277,7 +280,10 @@ impl<T: Serialize> Serialize for DualQuaternion<T> {
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<'a, T: Deserialize<'a>> Deserialize<'a> for DualQuaternion<T> {
+impl<'a, T: SimdRealField> Deserialize<'a> for DualQuaternion<T>
+where
+    T: Deserialize<'a>,
+{
     fn deserialize<Des>(deserializer: Des) -> Result<Self, Des::Error>
     where
         Des: Deserializer<'a>,
@@ -293,14 +299,9 @@ impl<'a, T: Deserialize<'a>> Deserialize<'a> for DualQuaternion<T> {
     }
 }
 
-impl<T> DualQuaternion<T> {
-    // TODO: Cloning shouldn't be necessary.
-    // TODO: rename into `into_vector` to appease clippy.
-    fn to_vector(self) -> OVector<T, U8>
-    where
-        T: Clone,
-    {
-        (self.as_ref().clone()).into()
+impl<T: RealField> DualQuaternion<T> {
+    fn to_vector(self) -> OVector<T, U8> {
+        (*self.as_ref()).into()
     }
 }
 
@@ -356,14 +357,14 @@ impl<T: RealField + UlpsEq<Epsilon = T>> UlpsEq for DualQuaternion<T> {
 /// A unit quaternions. May be used to represent a rotation followed by a translation.
 pub type UnitDualQuaternion<T> = Unit<DualQuaternion<T>>;
 
-impl<T: PartialEq> PartialEq for UnitDualQuaternion<T> {
+impl<T: Scalar + ClosedNeg + PartialEq + SimdRealField> PartialEq for UnitDualQuaternion<T> {
     #[inline]
     fn eq(&self, rhs: &Self) -> bool {
         self.as_ref().eq(rhs.as_ref())
     }
 }
 
-impl<T: Eq> Eq for UnitDualQuaternion<T> {}
+impl<T: Scalar + ClosedNeg + Eq + SimdRealField> Eq for UnitDualQuaternion<T> {}
 
 impl<T: SimdRealField> Normed for DualQuaternion<T> {
     type Norm = T::SimdRealField;
@@ -391,7 +392,10 @@ impl<T: SimdRealField> Normed for DualQuaternion<T> {
     }
 }
 
-impl<T> UnitDualQuaternion<T> {
+impl<T: SimdRealField> UnitDualQuaternion<T>
+where
+    T::Element: SimdRealField,
+{
     /// The underlying dual quaternion.
     ///
     /// Same as `self.as_ref()`.
@@ -410,12 +414,7 @@ impl<T> UnitDualQuaternion<T> {
     pub fn dual_quaternion(&self) -> &DualQuaternion<T> {
         self.as_ref()
     }
-}
 
-impl<T: SimdRealField> UnitDualQuaternion<T>
-where
-    T::Element: SimdRealField,
-{
     /// Compute the conjugate of this unit quaternion.
     ///
     /// # Example
@@ -617,7 +616,7 @@ where
     #[must_use]
     pub fn sclerp(&self, other: &Self, t: T) -> Self
     where
-        T: RealField + RelativeEq<Epsilon = T>,
+        T: RealField,
     {
         self.try_sclerp(other, t, T::default_epsilon())
             .expect("DualQuaternion sclerp: ambiguous configuration.")
@@ -637,7 +636,7 @@ where
     #[must_use]
     pub fn try_sclerp(&self, other: &Self, t: T, epsilon: T) -> Option<Self>
     where
-        T: RealField + RelativeEq<Epsilon = T>,
+        T: RealField,
     {
         let two = T::one() + T::one();
         let half = T::one() / two;
diff --git a/src/geometry/dual_quaternion_construction.rs b/src/geometry/dual_quaternion_construction.rs
index d692d781..ea4c7ee2 100644
--- a/src/geometry/dual_quaternion_construction.rs
+++ b/src/geometry/dual_quaternion_construction.rs
@@ -1,5 +1,5 @@
 use crate::{
-    DualQuaternion, Isometry3, Quaternion, SimdRealField, Translation3, UnitDualQuaternion,
+    DualQuaternion, Isometry3, Quaternion, Scalar, SimdRealField, Translation3, UnitDualQuaternion,
     UnitQuaternion,
 };
 use num::{One, Zero};
@@ -7,7 +7,7 @@ use num::{One, Zero};
 use quickcheck::{Arbitrary, Gen};
 use simba::scalar::SupersetOf;
 
-impl<T> DualQuaternion<T> {
+impl<T: Scalar> DualQuaternion<T> {
     /// Creates a dual quaternion from its rotation and translation components.
     ///
     /// # Example
@@ -60,7 +60,7 @@ impl<T> DualQuaternion<T> {
     /// let q2 = q.cast::<f32>();
     /// assert_eq!(q2, DualQuaternion::from_real(Quaternion::new(1.0f32, 2.0, 3.0, 4.0)));
     /// ```
-    pub fn cast<To>(self) -> DualQuaternion<To>
+    pub fn cast<To: Scalar>(self) -> DualQuaternion<To>
     where
         DualQuaternion<To>: SupersetOf<Self>,
     {
@@ -156,7 +156,7 @@ impl<T: SimdRealField> UnitDualQuaternion<T> {
     /// let q2 = q.cast::<f32>();
     /// assert_eq!(q2, UnitDualQuaternion::<f32>::identity());
     /// ```
-    pub fn cast<To>(self) -> UnitDualQuaternion<To>
+    pub fn cast<To: Scalar>(self) -> UnitDualQuaternion<To>
     where
         UnitDualQuaternion<To>: SupersetOf<Self>,
     {
diff --git a/src/geometry/dual_quaternion_conversion.rs b/src/geometry/dual_quaternion_conversion.rs
index 2afffe26..94ef9e97 100644
--- a/src/geometry/dual_quaternion_conversion.rs
+++ b/src/geometry/dual_quaternion_conversion.rs
@@ -24,7 +24,8 @@ use crate::geometry::{
 
 impl<T1, T2> SubsetOf<DualQuaternion<T2>> for DualQuaternion<T1>
 where
-    T2: SupersetOf<T1>,
+    T1: SimdRealField,
+    T2: SimdRealField + SupersetOf<T1>,
 {
     #[inline]
     fn to_superset(&self) -> DualQuaternion<T2> {
@@ -48,7 +49,8 @@ where
 
 impl<T1, T2> SubsetOf<UnitDualQuaternion<T2>> for UnitDualQuaternion<T1>
 where
-    T2: SupersetOf<T1>,
+    T1: SimdRealField,
+    T2: SimdRealField + SupersetOf<T1>,
 {
     #[inline]
     fn to_superset(&self) -> UnitDualQuaternion<T2> {
diff --git a/src/geometry/dual_quaternion_ops.rs b/src/geometry/dual_quaternion_ops.rs
index 151b2e05..2a1527ec 100644
--- a/src/geometry/dual_quaternion_ops.rs
+++ b/src/geometry/dual_quaternion_ops.rs
@@ -56,21 +56,21 @@ use std::ops::{
     Add, AddAssign, Div, DivAssign, Index, IndexMut, Mul, MulAssign, Neg, Sub, SubAssign,
 };
 
-impl<T> AsRef<[T; 8]> for DualQuaternion<T> {
+impl<T: SimdRealField> AsRef<[T; 8]> for DualQuaternion<T> {
     #[inline]
     fn as_ref(&self) -> &[T; 8] {
-        unsafe { &*(self as *const _ as *const _) }
+        unsafe { &*(self as *const Self as *const [T; 8]) }
     }
 }
 
-impl<T> AsMut<[T; 8]> for DualQuaternion<T> {
+impl<T: SimdRealField> AsMut<[T; 8]> for DualQuaternion<T> {
     #[inline]
     fn as_mut(&mut self) -> &mut [T; 8] {
-        unsafe { &mut *(self as *mut _ as *mut _) }
+        unsafe { &mut *(self as *mut Self as *mut [T; 8]) }
     }
 }
 
-impl<T> Index<usize> for DualQuaternion<T> {
+impl<T: SimdRealField> Index<usize> for DualQuaternion<T> {
     type Output = T;
 
     #[inline]
@@ -79,7 +79,7 @@ impl<T> Index<usize> for DualQuaternion<T> {
     }
 }
 
-impl<T> IndexMut<usize> for DualQuaternion<T> {
+impl<T: SimdRealField> IndexMut<usize> for DualQuaternion<T> {
     #[inline]
     fn index_mut(&mut self, i: usize) -> &mut T {
         &mut self.as_mut()[i]
diff --git a/src/geometry/isometry.rs b/src/geometry/isometry.rs
index 74e2f05d..f8e63d07 100755
--- a/src/geometry/isometry.rs
+++ b/src/geometry/isometry.rs
@@ -15,7 +15,7 @@ use simba::simd::SimdRealField;
 
 use crate::base::allocator::Allocator;
 use crate::base::dimension::{DimNameAdd, DimNameSum, U1};
-use crate::base::storage::InnerOwned;
+use crate::base::storage::Owned;
 use crate::base::{Const, DefaultAllocator, OMatrix, SVector, Scalar, Unit};
 use crate::geometry::{AbstractRotation, Point, Translation};
 
@@ -53,6 +53,7 @@ use crate::geometry::{AbstractRotation, Point, Translation};
 /// # Conversion to a matrix
 /// * [Conversion to a matrix <span style="float:right;">`to_matrix`…</span>](#conversion-to-a-matrix)
 ///
+#[repr(C)]
 #[derive(Debug)]
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
 #[cfg_attr(
@@ -79,6 +80,7 @@ pub struct Isometry<T, R, const D: usize> {
 #[cfg(feature = "abomonation-serialize")]
 impl<T, R, const D: usize> Abomonation for Isometry<T, R, D>
 where
+    T: SimdRealField,
     R: Abomonation,
     Translation<T, D>: Abomonation,
 {
@@ -104,7 +106,10 @@ mod rkyv_impl {
     use crate::{base::Scalar, geometry::Translation};
     use rkyv::{offset_of, project_struct, Archive, Deserialize, Fallible, Serialize};
 
-    impl<T: Archive, R: Archive, const D: usize> Archive for Isometry<T, R, D> {
+    impl<T: Scalar + Archive, R: Archive, const D: usize> Archive for Isometry<T, R, D>
+    where
+        T::Archived: Scalar,
+    {
         type Archived = Isometry<T::Archived, R::Archived, D>;
         type Resolver = (R::Resolver, <Translation<T, D> as Archive>::Resolver);
 
@@ -127,8 +132,8 @@ mod rkyv_impl {
         }
     }
 
-    impl<T: Serialize<S>, R: Serialize<S>, S: Fallible + ?Sized, const D: usize> Serialize<S>
-        for Isometry<T, R, D>
+    impl<T: Scalar + Serialize<S>, R: Serialize<S>, S: Fallible + ?Sized, const D: usize>
+        Serialize<S> for Isometry<T, R, D>
     where
         T::Archived: Scalar,
     {
@@ -140,7 +145,7 @@ mod rkyv_impl {
         }
     }
 
-    impl<T: Archive, R: Archive, _D: Fallible + ?Sized, const D: usize>
+    impl<T: Scalar + Archive, R: Archive, _D: Fallible + ?Sized, const D: usize>
         Deserialize<Isometry<T, R, D>, _D> for Isometry<T::Archived, R::Archived, D>
     where
         T::Archived: Scalar + Deserialize<T, _D>,
@@ -155,9 +160,9 @@ mod rkyv_impl {
     }
 }
 
-impl<T: hash::Hash, R: hash::Hash, const D: usize> hash::Hash for Isometry<T, R, D>
+impl<T: Scalar + hash::Hash, R: hash::Hash, const D: usize> hash::Hash for Isometry<T, R, D>
 where
-    InnerOwned<T, Const<D>>: hash::Hash,
+    Owned<T, Const<D>>: hash::Hash,
 {
     fn hash<H: hash::Hasher>(&self, state: &mut H) {
         self.translation.hash(state);
@@ -165,9 +170,12 @@ where
     }
 }
 
-impl<T: Copy, R: Copy, const D: usize> Copy for Isometry<T, R, D> where InnerOwned<T, Const<D>>: Copy {}
+impl<T: Scalar + Copy, R: Copy, const D: usize> Copy for Isometry<T, R, D> where
+    Owned<T, Const<D>>: Copy
+{
+}
 
-impl<T: Clone, R: Clone, const D: usize> Clone for Isometry<T, R, D> {
+impl<T: Scalar, R: Clone, const D: usize> Clone for Isometry<T, R, D> {
     #[inline]
     fn clone(&self) -> Self {
         Self {
@@ -630,7 +638,7 @@ where
  * Display
  *
  */
-impl<T: Scalar + fmt::Display, R, const D: usize> fmt::Display for Isometry<T, R, D>
+impl<T: RealField + fmt::Display, R, const D: usize> fmt::Display for Isometry<T, R, D>
 where
     R: fmt::Display,
 {
diff --git a/src/geometry/isometry_construction.rs b/src/geometry/isometry_construction.rs
index fe09b5cd..9b855599 100644
--- a/src/geometry/isometry_construction.rs
+++ b/src/geometry/isometry_construction.rs
@@ -1,5 +1,5 @@
 #[cfg(feature = "arbitrary")]
-use crate::base::storage::InnerOwned;
+use crate::base::storage::Owned;
 #[cfg(feature = "arbitrary")]
 use quickcheck::{Arbitrary, Gen};
 
@@ -97,7 +97,7 @@ where
     T: SimdRealField + Arbitrary + Send,
     T::Element: SimdRealField,
     R: AbstractRotation<T, D> + Arbitrary + Send,
-    InnerOwned<T, crate::Const<D>>: Send,
+    Owned<T, crate::Const<D>>: Send,
 {
     #[inline]
     fn arbitrary(rng: &mut Gen) -> Self {
diff --git a/src/geometry/orthographic.rs b/src/geometry/orthographic.rs
index 441ecd2d..b349a621 100644
--- a/src/geometry/orthographic.rs
+++ b/src/geometry/orthographic.rs
@@ -18,29 +18,27 @@ use crate::base::{Matrix4, Vector, Vector3};
 use crate::geometry::{Point3, Projective3};
 
 /// A 3D orthographic projection stored as a homogeneous 4x4 matrix.
-#[repr(transparent)]
+#[repr(C)]
 pub struct Orthographic3<T> {
     matrix: Matrix4<T>,
 }
 
-impl<T: Copy> Copy for Orthographic3<T> {}
+impl<T: RealField> Copy for Orthographic3<T> {}
 
-impl<T: Clone> Clone for Orthographic3<T> {
+impl<T: RealField> Clone for Orthographic3<T> {
     #[inline]
     fn clone(&self) -> Self {
-        Self {
-            matrix: self.matrix.clone(),
-        }
+        Self::from_matrix_unchecked(self.matrix)
     }
 }
 
-impl<T: fmt::Debug> fmt::Debug for Orthographic3<T> {
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+impl<T: RealField> fmt::Debug for Orthographic3<T> {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> Result<(), fmt::Error> {
         self.matrix.fmt(f)
     }
 }
 
-impl<T: PartialEq> PartialEq for Orthographic3<T> {
+impl<T: RealField> PartialEq for Orthographic3<T> {
     #[inline]
     fn eq(&self, right: &Self) -> bool {
         self.matrix == right.matrix
@@ -64,7 +62,7 @@ where
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<T: Serialize> Serialize for Orthographic3<T> {
+impl<T: RealField + Serialize> Serialize for Orthographic3<T> {
     fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
     where
         S: Serializer,
@@ -74,7 +72,7 @@ impl<T: Serialize> Serialize for Orthographic3<T> {
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<'a, T: Deserialize<'a>> Deserialize<'a> for Orthographic3<T> {
+impl<'a, T: RealField + Deserialize<'a>> Deserialize<'a> for Orthographic3<T> {
     fn deserialize<Des>(deserializer: Des) -> Result<Self, Des::Error>
     where
         Des: Deserializer<'a>,
@@ -85,8 +83,31 @@ impl<'a, T: Deserialize<'a>> Deserialize<'a> for Orthographic3<T> {
     }
 }
 
-/// # Basic methods and casts.
 impl<T> Orthographic3<T> {
+    /// Wraps the given matrix to interpret it as a 3D orthographic matrix.
+    ///
+    /// It is not checked whether or not the given matrix actually represents an orthographic
+    /// projection.
+    ///
+    /// # Example
+    /// ```
+    /// # use nalgebra::{Orthographic3, Point3, Matrix4};
+    /// let mat = Matrix4::new(
+    ///     2.0 / 9.0, 0.0,        0.0,         -11.0 / 9.0,
+    ///     0.0,       2.0 / 18.0, 0.0,         -22.0 / 18.0,
+    ///     0.0,       0.0,       -2.0 / 999.9, -1000.1 / 999.9,
+    ///     0.0,       0.0,        0.0,         1.0
+    /// );
+    /// let proj = Orthographic3::from_matrix_unchecked(mat);
+    /// assert_eq!(proj, Orthographic3::new(1.0, 10.0, 2.0, 20.0, 0.1, 1000.0));
+    /// ```
+    #[inline]
+    pub const fn from_matrix_unchecked(matrix: Matrix4<T>) -> Self {
+        Self { matrix }
+    }
+}
+
+impl<T: RealField> Orthographic3<T> {
     /// Creates a new orthographic projection matrix.
     ///
     /// This follows the OpenGL convention, so this will flip the `z` axis.
@@ -130,11 +151,8 @@ impl<T> Orthographic3<T> {
     /// assert_relative_eq!(proj.project_point(&p8), Point3::new(-1.0, -1.0, -1.0));
     /// ```
     #[inline]
-    pub fn new(left: T, right: T, bottom: T, top: T, znear: T, zfar: T) -> Self
-    where
-        T: RealField,
-    {
-        let matrix = Matrix4::identity();
+    pub fn new(left: T, right: T, bottom: T, top: T, znear: T, zfar: T) -> Self {
+        let matrix = Matrix4::<T>::identity();
         let mut res = Self::from_matrix_unchecked(matrix);
 
         res.set_left_and_right(left, right);
@@ -146,10 +164,7 @@ impl<T> Orthographic3<T> {
 
     /// Creates a new orthographic projection matrix from an aspect ratio and the vertical field of view.
     #[inline]
-    pub fn from_fov(aspect: T, vfov: T, znear: T, zfar: T) -> Self
-    where
-        T: RealField,
-    {
+    pub fn from_fov(aspect: T, vfov: T, znear: T, zfar: T) -> Self {
         assert!(
             znear != zfar,
             "The far plane must not be equal to the near plane."
@@ -192,10 +207,7 @@ impl<T> Orthographic3<T> {
     /// ```
     #[inline]
     #[must_use]
-    pub fn inverse(&self) -> Matrix4<T>
-    where
-        T: RealField,
-    {
+    pub fn inverse(&self) -> Matrix4<T> {
         let mut res = self.to_homogeneous();
 
         let inv_m11 = T::one() / self.matrix[(0, 0)];
@@ -229,7 +241,6 @@ impl<T> Orthographic3<T> {
     /// ```
     #[inline]
     #[must_use]
-    // TODO: rename into `into_homogeneous` to appease clippy.
     pub fn to_homogeneous(self) -> Matrix4<T> {
         self.matrix
     }
@@ -265,8 +276,7 @@ impl<T> Orthographic3<T> {
     #[inline]
     #[must_use]
     pub fn as_projective(&self) -> &Projective3<T> {
-        // Safety: Self and Projective3 are both #[repr(transparent)] of a matrix.
-        unsafe { &*(self as *const _ as *const _) }
+        unsafe { &*(self as *const Orthographic3<T> as *const Projective3<T>) }
     }
 
     /// This transformation seen as a `Projective3`.
@@ -279,7 +289,6 @@ impl<T> Orthographic3<T> {
     /// ```
     #[inline]
     #[must_use]
-    // TODO: rename into `into_projective` to appease clippy.
     pub fn to_projective(self) -> Projective3<T> {
         Projective3::from_matrix_unchecked(self.matrix)
     }
@@ -311,10 +320,7 @@ impl<T> Orthographic3<T> {
     pub fn unwrap(self) -> Matrix4<T> {
         self.matrix
     }
-}
 
-/// # Mathematical methods.
-impl<T: RealField> Orthographic3<T> {
     /// The left offset of the view cuboid.
     ///
     /// ```
diff --git a/src/geometry/perspective.rs b/src/geometry/perspective.rs
index 5007b26b..d5a6fe42 100644
--- a/src/geometry/perspective.rs
+++ b/src/geometry/perspective.rs
@@ -34,7 +34,7 @@ impl<T: RealField> Clone for Perspective3<T> {
 }
 
 impl<T: RealField> fmt::Debug for Perspective3<T> {
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> Result<(), fmt::Error> {
         self.matrix.fmt(f)
     }
 }
@@ -158,8 +158,7 @@ impl<T: RealField> Perspective3<T> {
     #[inline]
     #[must_use]
     pub fn as_projective(&self) -> &Projective3<T> {
-        // Safety: Self and Projective3 are both #[repr(transparent)] of a matrix.
-        unsafe { &*(self as *const _ as *const _) }
+        unsafe { &*(self as *const Perspective3<T> as *const Projective3<T>) }
     }
 
     /// This transformation seen as a `Projective3`.
diff --git a/src/geometry/point.rs b/src/geometry/point.rs
index d73c4f22..098b5c2a 100644
--- a/src/geometry/point.rs
+++ b/src/geometry/point.rs
@@ -5,7 +5,6 @@ use std::fmt;
 use std::hash;
 #[cfg(feature = "abomonation-serialize")]
 use std::io::{Result as IOResult, Write};
-use std::mem::{ManuallyDrop, MaybeUninit};
 
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
@@ -15,13 +14,11 @@ use abomonation::Abomonation;
 
 use simba::simd::SimdPartialOrd;
 
-use crate::allocator::InnerAllocator;
 use crate::base::allocator::Allocator;
 use crate::base::dimension::{DimName, DimNameAdd, DimNameSum, U1};
 use crate::base::iter::{MatrixIter, MatrixIterMut};
-use crate::base::{Const, DefaultAllocator, OVector};
-use crate::storage::InnerOwned;
-use crate::Scalar;
+use crate::base::{Const, DefaultAllocator, OVector, Scalar};
+use std::mem::MaybeUninit;
 
 /// A point in an euclidean space.
 ///
@@ -42,16 +39,17 @@ use crate::Scalar;
 /// achieved by multiplication, e.g., `isometry * point` or `rotation * point`. Some of these transformation
 /// may have some other methods, e.g., `isometry.inverse_transform_point(&point)`. See the documentation
 /// of said transformations for details.
-#[repr(transparent)]
-pub struct OPoint<T, D: DimName>
+#[repr(C)]
+#[derive(Debug, Clone)]
+pub struct OPoint<T: Scalar, D: DimName>
 where
-    DefaultAllocator: InnerAllocator<T, D>,
+    DefaultAllocator: Allocator<T, D>,
 {
     /// The coordinates of this point, i.e., the shift from the origin.
     pub coords: OVector<T, D>,
 }
 
-impl<T: hash::Hash, D: DimName> hash::Hash for OPoint<T, D>
+impl<T: Scalar + hash::Hash, D: DimName> hash::Hash for OPoint<T, D>
 where
     DefaultAllocator: Allocator<T, D>,
 {
@@ -60,37 +58,15 @@ where
     }
 }
 
-impl<T: Copy, D: DimName> Copy for OPoint<T, D>
+impl<T: Scalar + Copy, D: DimName> Copy for OPoint<T, D>
 where
     DefaultAllocator: Allocator<T, D>,
     OVector<T, D>: Copy,
 {
 }
 
-impl<T: Clone, D: DimName> Clone for OPoint<T, D>
-where
-    DefaultAllocator: Allocator<T, D>,
-    OVector<T, D>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self::from(self.coords.clone())
-    }
-}
-
-impl<T: fmt::Debug, D: DimName> fmt::Debug for OPoint<T, D>
-where
-    DefaultAllocator: Allocator<T, D>,
-    OVector<T, D>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("OPoint")
-            .field("coords", &self.coords)
-            .finish()
-    }
-}
-
 #[cfg(feature = "bytemuck")]
-unsafe impl<T, D: DimName> bytemuck::Zeroable for OPoint<T, D>
+unsafe impl<T: Scalar, D: DimName> bytemuck::Zeroable for OPoint<T, D>
 where
     OVector<T, D>: bytemuck::Zeroable,
     DefaultAllocator: Allocator<T, D>,
@@ -98,7 +74,7 @@ where
 }
 
 #[cfg(feature = "bytemuck")]
-unsafe impl<T, D: DimName> bytemuck::Pod for OPoint<T, D>
+unsafe impl<T: Scalar, D: DimName> bytemuck::Pod for OPoint<T, D>
 where
     T: Copy,
     OVector<T, D>: bytemuck::Pod,
@@ -107,10 +83,10 @@ where
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<T: Serialize, D: DimName> Serialize for OPoint<T, D>
+impl<T: Scalar, D: DimName> Serialize for OPoint<T, D>
 where
     DefaultAllocator: Allocator<T, D>,
-    <DefaultAllocator as InnerAllocator<T, D>>::Buffer: Serialize,
+    <DefaultAllocator as Allocator<T, D>>::Buffer: Serialize,
 {
     fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
     where
@@ -121,10 +97,10 @@ where
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<'a, T: Deserialize<'a>, D: DimName> Deserialize<'a> for OPoint<T, D>
+impl<'a, T: Scalar, D: DimName> Deserialize<'a> for OPoint<T, D>
 where
     DefaultAllocator: Allocator<T, D>,
-    <DefaultAllocator as InnerAllocator<T, D>>::Buffer: Deserialize<'a>,
+    <DefaultAllocator as Allocator<T, D>>::Buffer: Deserialize<'a>,
 {
     fn deserialize<Des>(deserializer: Des) -> Result<Self, Des::Error>
     where
@@ -139,6 +115,7 @@ where
 #[cfg(feature = "abomonation-serialize")]
 impl<T, D: DimName> Abomonation for OPoint<T, D>
 where
+    T: Scalar,
     OVector<T, D>: Abomonation,
     DefaultAllocator: Allocator<T, D>,
 {
@@ -155,7 +132,7 @@ where
     }
 }
 
-impl<T, D: DimName> OPoint<T, D>
+impl<T: Scalar, D: DimName> OPoint<T, D>
 where
     DefaultAllocator: Allocator<T, D>,
 {
@@ -173,9 +150,8 @@ where
     /// ```
     #[inline]
     #[must_use]
-    pub fn map<T2, F: FnMut(T) -> T2>(&self, f: F) -> OPoint<T2, D>
+    pub fn map<T2: Scalar, F: FnMut(T) -> T2>(&self, f: F) -> OPoint<T2, D>
     where
-        T: Clone,
         DefaultAllocator: Allocator<T2, D>,
     {
         self.coords.map(f).into()
@@ -187,19 +163,16 @@ where
     /// ```
     /// # use nalgebra::{Point2, Point3};
     /// let mut p = Point2::new(1.0, 2.0);
-    /// p.apply(|e| e * 10.0);
+    /// p.apply(|e| *e = *e * 10.0);
     /// assert_eq!(p, Point2::new(10.0, 20.0));
     ///
     /// // This works in any dimension.
     /// let mut p = Point3::new(1.0, 2.0, 3.0);
-    /// p.apply(|e| e * 10.0);
+    /// p.apply(|e| *e = *e * 10.0);
     /// assert_eq!(p, Point3::new(10.0, 20.0, 30.0));
     /// ```
     #[inline]
-    pub fn apply<F: FnMut(T) -> T>(&mut self, f: F)
-    where
-        T: Clone,
-    {
+    pub fn apply<F: FnMut(&mut T)>(&mut self, f: F) {
         self.coords.apply(f)
     }
 
@@ -221,45 +194,25 @@ where
     #[inline]
     #[must_use]
     pub fn to_homogeneous(&self) -> OVector<T, DimNameSum<D, U1>>
-    where
-        T: One + Clone,
-        D: DimNameAdd<U1>,
-        DefaultAllocator: Allocator<T, DimNameSum<D, U1>>,
-    {
-        let mut res = OVector::<_, DimNameSum<D, U1>>::new_uninitialized();
-        for i in 0..D::dim() {
-            unsafe {
-                *res.get_unchecked_mut(i) = MaybeUninit::new(self.coords[i].clone());
-            }
-        }
-
-        res[(D::dim(), 0)] = MaybeUninit::new(T::one());
-
-        unsafe { res.assume_init() }
-    }
-
-    /// Converts this point into a vector in homogeneous coordinates, i.e., appends a `1` at the
-    /// end of it. Unlike [`to_homogeneous`], this method does not require `T: Clone`.
-    pub fn into_homogeneous(self) -> OVector<T, DimNameSum<D, U1>>
     where
         T: One,
         D: DimNameAdd<U1>,
         DefaultAllocator: Allocator<T, DimNameSum<D, U1>>,
     {
-        let mut res = OVector::<_, DimNameSum<D, U1>>::new_uninitialized();
-        let mut md = self.manually_drop();
+        // TODO: this is mostly a copy-past from Vector::push.
+        //       But we can’t use Vector::push because of the DimAdd bound
+        //       (which we don’t use because we use DimNameAdd).
+        //       We should find a way to re-use Vector::push.
+        let len = self.len();
+        let mut res = crate::Matrix::uninit(DimNameSum::<D, U1>::name(), Const::<1>);
+        // This is basically a copy_from except that we warp the copied
+        // values into MaybeUninit.
+        res.generic_slice_mut((0, 0), self.coords.shape_generic())
+            .zip_apply(&self.coords, |out, e| *out = MaybeUninit::new(e));
+        res[(len, 0)] = MaybeUninit::new(T::one());
 
-        for i in 0..D::dim() {
-            unsafe {
-                *res.get_unchecked_mut(i) =
-                    MaybeUninit::new(ManuallyDrop::take(md.coords.get_unchecked_mut(i)));
-            }
-        }
-
-        unsafe {
-            *res.get_unchecked_mut(D::dim()) = MaybeUninit::new(T::one());
-            res.assume_init()
-        }
+        // Safety: res has been fully initialized.
+        unsafe { res.assume_init() }
     }
 
     /// Creates a new point with the given coordinates.
@@ -322,7 +275,9 @@ where
     /// assert_eq!(it.next(), Some(3.0));
     /// assert_eq!(it.next(), None);
     #[inline]
-    pub fn iter(&self) -> MatrixIter<T, D, Const<1>, InnerOwned<T, D>> {
+    pub fn iter(
+        &self,
+    ) -> MatrixIter<'_, T, D, Const<1>, <DefaultAllocator as Allocator<T, D>>::Buffer> {
         self.coords.iter()
     }
 
@@ -346,7 +301,9 @@ where
     ///
     /// assert_eq!(p, Point3::new(10.0, 20.0, 30.0));
     #[inline]
-    pub fn iter_mut(&mut self) -> MatrixIterMut<T, D, Const<1>, InnerOwned<T, D>> {
+    pub fn iter_mut(
+        &mut self,
+    ) -> MatrixIterMut<'_, T, D, Const<1>, <DefaultAllocator as Allocator<T, D>>::Buffer> {
         self.coords.iter_mut()
     }
 
@@ -364,7 +321,7 @@ where
     }
 }
 
-impl<T: AbsDiffEq, D: DimName> AbsDiffEq for OPoint<T, D>
+impl<T: Scalar + AbsDiffEq, D: DimName> AbsDiffEq for OPoint<T, D>
 where
     T::Epsilon: Copy,
     DefaultAllocator: Allocator<T, D>,
@@ -382,7 +339,7 @@ where
     }
 }
 
-impl<T: RelativeEq, D: DimName> RelativeEq for OPoint<T, D>
+impl<T: Scalar + RelativeEq, D: DimName> RelativeEq for OPoint<T, D>
 where
     T::Epsilon: Copy,
     DefaultAllocator: Allocator<T, D>,
@@ -404,7 +361,7 @@ where
     }
 }
 
-impl<T: UlpsEq, D: DimName> UlpsEq for OPoint<T, D>
+impl<T: Scalar + UlpsEq, D: DimName> UlpsEq for OPoint<T, D>
 where
     T::Epsilon: Copy,
     DefaultAllocator: Allocator<T, D>,
@@ -420,9 +377,9 @@ where
     }
 }
 
-impl<T: Eq, D: DimName> Eq for OPoint<T, D> where DefaultAllocator: Allocator<T, D> {}
+impl<T: Scalar + Eq, D: DimName> Eq for OPoint<T, D> where DefaultAllocator: Allocator<T, D> {}
 
-impl<T: PartialEq, D: DimName> PartialEq for OPoint<T, D>
+impl<T: Scalar, D: DimName> PartialEq for OPoint<T, D>
 where
     DefaultAllocator: Allocator<T, D>,
 {
@@ -432,7 +389,7 @@ where
     }
 }
 
-impl<T: PartialOrd, D: DimName> PartialOrd for OPoint<T, D>
+impl<T: Scalar + PartialOrd, D: DimName> PartialOrd for OPoint<T, D>
 where
     DefaultAllocator: Allocator<T, D>,
 {
@@ -497,7 +454,7 @@ where
  * Display
  *
  */
-impl<T: fmt::Display, D: DimName> fmt::Display for OPoint<T, D>
+impl<T: Scalar + fmt::Display, D: DimName> fmt::Display for OPoint<T, D>
 where
     DefaultAllocator: Allocator<T, D>,
 {
diff --git a/src/geometry/point_construction.rs b/src/geometry/point_construction.rs
index 94876c18..d2393146 100644
--- a/src/geometry/point_construction.rs
+++ b/src/geometry/point_construction.rs
@@ -1,5 +1,3 @@
-use std::mem::{ManuallyDrop, MaybeUninit};
-
 #[cfg(feature = "arbitrary")]
 use quickcheck::{Arbitrary, Gen};
 
@@ -22,23 +20,10 @@ use simba::scalar::{ClosedDiv, SupersetOf};
 use crate::geometry::Point;
 
 /// # Other construction methods
-impl<T, D: DimName> OPoint<T, D>
+impl<T: Scalar, D: DimName> OPoint<T, D>
 where
     DefaultAllocator: Allocator<T, D>,
 {
-    /// Creates a new point with uninitialized coordinates.
-    #[inline]
-    pub fn new_uninitialized() -> OPoint<MaybeUninit<T>, D> {
-        OPoint::from(OVector::new_uninitialized_generic(D::name(), Const::<1>))
-    }
-
-    /// Converts `self` into a point whose coordinates must be manually dropped.
-    /// This should be zero-cost.
-    #[inline]
-    pub fn manually_drop(self) -> OPoint<ManuallyDrop<T>, D> {
-        OPoint::from(self.coords.manually_drop())
-    }
-
     /// Creates a new point with all coordinates equal to zero.
     ///
     /// # Example
@@ -57,9 +42,9 @@ where
     #[inline]
     pub fn origin() -> Self
     where
-        T: Zero + Clone,
+        T: Zero,
     {
-        Self::from(OVector::<_, D>::zeros())
+        Self::from(OVector::from_element(T::zero()))
     }
 
     /// Creates a new point from a slice.
@@ -77,11 +62,8 @@ where
     /// assert_eq!(pt, Point3::new(1.0, 2.0, 3.0));
     /// ```
     #[inline]
-    pub fn from_slice(components: &[T]) -> Self
-    where
-        T: Clone,
-    {
-        Self::from(OVector::<_, D>::from_row_slice(components))
+    pub fn from_slice(components: &[T]) -> Self {
+        Self::from(OVector::from_row_slice(components))
     }
 
     /// Creates a new point from its homogeneous vector representation.
@@ -139,7 +121,7 @@ where
     /// let pt2 = pt.cast::<f32>();
     /// assert_eq!(pt2, Point2::new(1.0f32, 2.0));
     /// ```
-    pub fn cast<To>(self) -> OPoint<To, D>
+    pub fn cast<To: Scalar>(self) -> OPoint<To, D>
     where
         OPoint<To, D>: SupersetOf<Self>,
         DefaultAllocator: Allocator<To, D>,
@@ -169,7 +151,7 @@ where
 }
 
 #[cfg(feature = "rand-no-std")]
-impl<T, D: DimName> Distribution<OPoint<T, D>> for Standard
+impl<T: Scalar, D: DimName> Distribution<OPoint<T, D>> for Standard
 where
     Standard: Distribution<T>,
     DefaultAllocator: Allocator<T, D>,
@@ -182,10 +164,10 @@ where
 }
 
 #[cfg(feature = "arbitrary")]
-impl<T: Arbitrary + Send, D: DimName> Arbitrary for OPoint<T, D>
+impl<T: Scalar + Arbitrary + Send, D: DimName> Arbitrary for OPoint<T, D>
 where
+    <DefaultAllocator as Allocator<T, D>>::Buffer: Send,
     DefaultAllocator: Allocator<T, D>,
-    crate::base::storage::InnerOwned<T, D>: Clone + Send,
 {
     #[inline]
     fn arbitrary(g: &mut Gen) -> Self {
@@ -201,7 +183,7 @@ where
 // NOTE: the impl for Point1 is not with the others so that we
 // can add a section with the impl block comment.
 /// # Construction from individual components
-impl<T> Point1<T> {
+impl<T: Scalar> Point1<T> {
     /// Initializes this point from its components.
     ///
     /// # Example
@@ -220,7 +202,7 @@ impl<T> Point1<T> {
 }
 macro_rules! componentwise_constructors_impl(
     ($($doc: expr; $Point: ident, $Vector: ident, $($args: ident:$irow: expr),*);* $(;)*) => {$(
-        impl<T> $Point<T> {
+        impl<T: Scalar> $Point<T> {
             #[doc = "Initializes this point from its components."]
             #[doc = "# Example\n```"]
             #[doc = $doc]
diff --git a/src/geometry/point_conversion.rs b/src/geometry/point_conversion.rs
index b564f0ad..f35a9fc6 100644
--- a/src/geometry/point_conversion.rs
+++ b/src/geometry/point_conversion.rs
@@ -2,7 +2,7 @@ use num::{One, Zero};
 use simba::scalar::{ClosedDiv, SubsetOf, SupersetOf};
 use simba::simd::PrimitiveSimdValue;
 
-use crate::base::allocator::{Allocator, InnerAllocator};
+use crate::base::allocator::Allocator;
 use crate::base::dimension::{DimNameAdd, DimNameSum, U1};
 use crate::base::{Const, DefaultAllocator, Matrix, OVector, Scalar};
 
@@ -19,7 +19,8 @@ use crate::{DimName, OPoint};
 
 impl<T1, T2, D: DimName> SubsetOf<OPoint<T2, D>> for OPoint<T1, D>
 where
-    T2: SupersetOf<T1>,
+    T1: Scalar,
+    T2: Scalar + SupersetOf<T1>,
     DefaultAllocator: Allocator<T1, D> + Allocator<T2, D>,
 {
     #[inline]
@@ -43,6 +44,7 @@ where
 impl<T1, T2, D> SubsetOf<OVector<T2, DimNameSum<D, U1>>> for OPoint<T1, D>
 where
     D: DimNameAdd<U1>,
+    T1: Scalar,
     T2: Scalar + Zero + One + ClosedDiv + SupersetOf<T1>,
     DefaultAllocator: Allocator<T1, D>
         + Allocator<T2, D>
@@ -54,7 +56,7 @@ where
     #[inline]
     fn to_superset(&self) -> OVector<T2, DimNameSum<D, U1>> {
         let p: OPoint<T2, D> = self.to_superset();
-        p.into_homogeneous()
+        p.to_homogeneous()
     }
 
     #[inline]
@@ -64,25 +66,25 @@ where
 
     #[inline]
     fn from_superset_unchecked(v: &OVector<T2, DimNameSum<D, U1>>) -> Self {
-        let coords = v.generic_slice((0, 0), (D::name(), Const::<1>)) / v[D::dim()].clone();
+        let coords = v.generic_slice((0, 0), (D::name(), Const::<1>)) / v[D::dim()].inlined_clone();
         Self {
             coords: crate::convert_unchecked(coords),
         }
     }
 }
 
-impl<T: Zero + One, D: DimName> From<OPoint<T, D>> for OVector<T, DimNameSum<D, U1>>
+impl<T: Scalar + Zero + One, D: DimName> From<OPoint<T, D>> for OVector<T, DimNameSum<D, U1>>
 where
     D: DimNameAdd<U1>,
     DefaultAllocator: Allocator<T, DimNameSum<D, U1>> + Allocator<T, D>,
 {
     #[inline]
     fn from(t: OPoint<T, D>) -> Self {
-        t.into_homogeneous()
+        t.to_homogeneous()
     }
 }
 
-impl<T, const D: usize> From<[T; D]> for Point<T, D> {
+impl<T: Scalar, const D: usize> From<[T; D]> for Point<T, D> {
     #[inline]
     fn from(coords: [T; D]) -> Self {
         Point {
@@ -91,19 +93,16 @@ impl<T, const D: usize> From<[T; D]> for Point<T, D> {
     }
 }
 
-impl<T, const D: usize> From<Point<T, D>> for [T; D]
-where
-    T: Clone,
-{
+impl<T: Scalar, const D: usize> From<Point<T, D>> for [T; D] {
     #[inline]
     fn from(p: Point<T, D>) -> Self {
         p.coords.into()
     }
 }
 
-impl<T, D: DimName> From<OVector<T, D>> for OPoint<T, D>
+impl<T: Scalar, D: DimName> From<OVector<T, D>> for OPoint<T, D>
 where
-    DefaultAllocator: InnerAllocator<T, D>,
+    DefaultAllocator: Allocator<T, D>,
 {
     #[inline]
     fn from(coords: OVector<T, D>) -> Self {
@@ -111,81 +110,85 @@ where
     }
 }
 
-impl<T: Scalar + PrimitiveSimdValue, const D: usize> From<[Point<T::Element, D>; 2]> for Point<T, D>
+impl<T: Scalar + Copy + PrimitiveSimdValue, const D: usize> From<[Point<T::Element, D>; 2]>
+    for Point<T, D>
 where
     T: From<[<T as simba::simd::SimdValue>::Element; 2]>,
-    T::Element: Scalar,
+    T::Element: Scalar + Copy,
+    <DefaultAllocator as Allocator<T::Element, Const<D>>>::Buffer: Copy,
 {
     #[inline]
     fn from(arr: [Point<T::Element, D>; 2]) -> Self {
-        Self::from(OVector::from([
-            arr[0].coords.clone(),
-            arr[1].coords.clone(),
-        ]))
+        Self::from(OVector::from([arr[0].coords, arr[1].coords]))
     }
 }
 
-impl<T: Scalar + PrimitiveSimdValue, const D: usize> From<[Point<T::Element, D>; 4]> for Point<T, D>
+impl<T: Scalar + Copy + PrimitiveSimdValue, const D: usize> From<[Point<T::Element, D>; 4]>
+    for Point<T, D>
 where
     T: From<[<T as simba::simd::SimdValue>::Element; 4]>,
-    T::Element: Scalar,
+    T::Element: Scalar + Copy,
+    <DefaultAllocator as Allocator<T::Element, Const<D>>>::Buffer: Copy,
 {
     #[inline]
     fn from(arr: [Point<T::Element, D>; 4]) -> Self {
         Self::from(OVector::from([
-            arr[0].coords.clone(),
-            arr[1].coords.clone(),
-            arr[2].coords.clone(),
-            arr[3].coords.clone(),
+            arr[0].coords,
+            arr[1].coords,
+            arr[2].coords,
+            arr[3].coords,
         ]))
     }
 }
 
-impl<T: Scalar + PrimitiveSimdValue, const D: usize> From<[Point<T::Element, D>; 8]> for Point<T, D>
+impl<T: Scalar + Copy + PrimitiveSimdValue, const D: usize> From<[Point<T::Element, D>; 8]>
+    for Point<T, D>
 where
     T: From<[<T as simba::simd::SimdValue>::Element; 8]>,
-    T::Element: Scalar,
+    T::Element: Scalar + Copy,
+    <DefaultAllocator as Allocator<T::Element, Const<D>>>::Buffer: Copy,
 {
     #[inline]
     fn from(arr: [Point<T::Element, D>; 8]) -> Self {
         Self::from(OVector::from([
-            arr[0].coords.clone(),
-            arr[1].coords.clone(),
-            arr[2].coords.clone(),
-            arr[3].coords.clone(),
-            arr[4].coords.clone(),
-            arr[5].coords.clone(),
-            arr[6].coords.clone(),
-            arr[7].coords.clone(),
+            arr[0].coords,
+            arr[1].coords,
+            arr[2].coords,
+            arr[3].coords,
+            arr[4].coords,
+            arr[5].coords,
+            arr[6].coords,
+            arr[7].coords,
         ]))
     }
 }
 
-impl<T: Scalar + PrimitiveSimdValue, const D: usize> From<[Point<T::Element, D>; 16]>
+impl<T: Scalar + Copy + PrimitiveSimdValue, const D: usize> From<[Point<T::Element, D>; 16]>
     for Point<T, D>
 where
     T: From<[<T as simba::simd::SimdValue>::Element; 16]>,
-    T::Element: Scalar,
+    T::Element: Scalar + Copy,
+    <DefaultAllocator as Allocator<T::Element, Const<D>>>::Buffer: Copy,
 {
     #[inline]
     fn from(arr: [Point<T::Element, D>; 16]) -> Self {
         Self::from(OVector::from([
-            arr[0].coords.clone(),
-            arr[1].coords.clone(),
-            arr[2].coords.clone(),
-            arr[3].coords.clone(),
-            arr[4].coords.clone(),
-            arr[5].coords.clone(),
-            arr[6].coords.clone(),
-            arr[7].coords.clone(),
-            arr[8].coords.clone(),
-            arr[9].coords.clone(),
-            arr[10].coords.clone(),
-            arr[11].coords.clone(),
-            arr[12].coords.clone(),
-            arr[13].coords.clone(),
-            arr[14].coords.clone(),
-            arr[15].coords.clone(),
+            arr[0].coords,
+            arr[1].coords,
+            arr[2].coords,
+            arr[3].coords,
+            arr[4].coords,
+            arr[5].coords,
+            arr[6].coords,
+            arr[7].coords,
+            arr[8].coords,
+            arr[9].coords,
+            arr[10].coords,
+            arr[11].coords,
+            arr[12].coords,
+            arr[13].coords,
+            arr[14].coords,
+            arr[15].coords,
         ]))
     }
 }
diff --git a/src/geometry/point_coordinates.rs b/src/geometry/point_coordinates.rs
index b9bd69a3..984a2fae 100644
--- a/src/geometry/point_coordinates.rs
+++ b/src/geometry/point_coordinates.rs
@@ -1,7 +1,7 @@
 use std::ops::{Deref, DerefMut};
 
 use crate::base::coordinates::{X, XY, XYZ, XYZW, XYZWA, XYZWAB};
-use crate::base::{U1, U2, U3, U4, U5, U6};
+use crate::base::{Scalar, U1, U2, U3, U4, U5, U6};
 
 use crate::geometry::OPoint;
 
@@ -13,7 +13,7 @@ use crate::geometry::OPoint;
 
 macro_rules! deref_impl(
     ($D: ty, $Target: ident $(, $comps: ident)*) => {
-        impl<T> Deref for OPoint<T, $D>
+        impl<T: Scalar> Deref for OPoint<T, $D>
         {
             type Target = $Target<T>;
 
@@ -23,7 +23,7 @@ macro_rules! deref_impl(
             }
         }
 
-        impl<T> DerefMut for OPoint<T, $D>
+        impl<T: Scalar> DerefMut for OPoint<T, $D>
         {
             #[inline]
             fn deref_mut(&mut self) -> &mut Self::Target {
diff --git a/src/geometry/point_ops.rs b/src/geometry/point_ops.rs
index 72d91ff3..5b019a9d 100644
--- a/src/geometry/point_ops.rs
+++ b/src/geometry/point_ops.rs
@@ -21,7 +21,7 @@ use crate::DefaultAllocator;
  * Indexing.
  *
  */
-impl<T, D: DimName> Index<usize> for OPoint<T, D>
+impl<T: Scalar, D: DimName> Index<usize> for OPoint<T, D>
 where
     DefaultAllocator: Allocator<T, D>,
 {
@@ -33,7 +33,7 @@ where
     }
 }
 
-impl<T, D: DimName> IndexMut<usize> for OPoint<T, D>
+impl<T: Scalar, D: DimName> IndexMut<usize> for OPoint<T, D>
 where
     DefaultAllocator: Allocator<T, D>,
 {
diff --git a/src/geometry/point_simba.rs b/src/geometry/point_simba.rs
index aa630adf..ad7433af 100644
--- a/src/geometry/point_simba.rs
+++ b/src/geometry/point_simba.rs
@@ -1,8 +1,8 @@
 use simba::simd::SimdValue;
 
-use crate::base::OVector;
+use crate::base::{OVector, Scalar};
+
 use crate::geometry::Point;
-use crate::Scalar;
 
 impl<T: Scalar + SimdValue, const D: usize> SimdValue for Point<T, D>
 where
diff --git a/src/geometry/quaternion.rs b/src/geometry/quaternion.rs
index 26bb8d97..cd248c94 100755
--- a/src/geometry/quaternion.rs
+++ b/src/geometry/quaternion.rs
@@ -6,7 +6,7 @@ use std::hash::{Hash, Hasher};
 use std::io::{Result as IOResult, Write};
 
 #[cfg(feature = "serde-serialize-no-std")]
-use crate::base::storage::InnerOwned;
+use crate::base::storage::Owned;
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
 
@@ -26,29 +26,29 @@ use crate::geometry::{Point3, Rotation};
 
 /// A quaternion. See the type alias `UnitQuaternion = Unit<Quaternion>` for a quaternion
 /// that may be used as a rotation.
-#[repr(transparent)]
+#[repr(C)]
 #[derive(Debug, Copy, Clone)]
 pub struct Quaternion<T> {
     /// This quaternion as a 4D vector of coordinates in the `[ x, y, z, w ]` storage order.
     pub coords: Vector4<T>,
 }
 
-impl<T: Hash> Hash for Quaternion<T> {
+impl<T: Scalar + Hash> Hash for Quaternion<T> {
     fn hash<H: Hasher>(&self, state: &mut H) {
         self.coords.hash(state)
     }
 }
 
-impl<T: Eq> Eq for Quaternion<T> {}
+impl<T: Scalar + Eq> Eq for Quaternion<T> {}
 
-impl<T: PartialEq> PartialEq for Quaternion<T> {
+impl<T: Scalar> PartialEq for Quaternion<T> {
     #[inline]
     fn eq(&self, right: &Self) -> bool {
         self.coords == right.coords
     }
 }
 
-impl<T: Zero + Clone> Default for Quaternion<T> {
+impl<T: Scalar + Zero> Default for Quaternion<T> {
     fn default() -> Self {
         Quaternion {
             coords: Vector4::zeros(),
@@ -57,10 +57,10 @@ impl<T: Zero + Clone> Default for Quaternion<T> {
 }
 
 #[cfg(feature = "bytemuck")]
-unsafe impl<T> bytemuck::Zeroable for Quaternion<T> where Vector4<T>: bytemuck::Zeroable {}
+unsafe impl<T: Scalar> bytemuck::Zeroable for Quaternion<T> where Vector4<T>: bytemuck::Zeroable {}
 
 #[cfg(feature = "bytemuck")]
-unsafe impl<T> bytemuck::Pod for Quaternion<T>
+unsafe impl<T: Scalar> bytemuck::Pod for Quaternion<T>
 where
     Vector4<T>: bytemuck::Pod,
     T: Copy,
@@ -68,7 +68,7 @@ where
 }
 
 #[cfg(feature = "abomonation-serialize")]
-impl<T> Abomonation for Quaternion<T>
+impl<T: Scalar> Abomonation for Quaternion<T>
 where
     Vector4<T>: Abomonation,
 {
@@ -86,7 +86,7 @@ where
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<T> Serialize for Quaternion<T>
+impl<T: Scalar> Serialize for Quaternion<T>
 where
     Owned<T, U4>: Serialize,
 {
@@ -99,7 +99,7 @@ where
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<'a, T> Deserialize<'a> for Quaternion<T>
+impl<'a, T: Scalar> Deserialize<'a> for Quaternion<T>
 where
     Owned<T, U4>: Deserialize<'a>,
 {
@@ -1045,8 +1045,8 @@ impl<T: RealField + UlpsEq<Epsilon = T>> UlpsEq for Quaternion<T> {
     }
 }
 
-impl<T: fmt::Display> fmt::Display for Quaternion<T> {
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+impl<T: RealField + fmt::Display> fmt::Display for Quaternion<T> {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         write!(
             f,
             "Quaternion {} − ({}, {}, {})",
@@ -1097,7 +1097,7 @@ impl<T: SimdRealField> UnitQuaternion<T>
 where
     T::Element: SimdRealField,
 {
-    /// The rotation angle in \[0; pi\] of this unit quaternion.
+    /// The rotation angle in [0; pi] of this unit quaternion.
     ///
     /// # Example
     /// ```
diff --git a/src/geometry/quaternion_construction.rs b/src/geometry/quaternion_construction.rs
index 451d5d55..61b1fe3e 100644
--- a/src/geometry/quaternion_construction.rs
+++ b/src/geometry/quaternion_construction.rs
@@ -1,7 +1,7 @@
 #[cfg(feature = "arbitrary")]
 use crate::base::dimension::U4;
 #[cfg(feature = "arbitrary")]
-use crate::base::storage::InnerOwned;
+use crate::base::storage::Owned;
 #[cfg(feature = "arbitrary")]
 use quickcheck::{Arbitrary, Gen};
 
@@ -179,7 +179,7 @@ where
 #[cfg(feature = "arbitrary")]
 impl<T: SimdRealField + Arbitrary> Arbitrary for Quaternion<T>
 where
-    InnerOwned<T, U4>: Send,
+    Owned<T, U4>: Send,
 {
     #[inline]
     fn arbitrary(g: &mut Gen) -> Self {
@@ -881,8 +881,8 @@ where
 #[cfg(feature = "arbitrary")]
 impl<T: RealField + Arbitrary> Arbitrary for UnitQuaternion<T>
 where
-    InnerOwned<T, U4>: Send,
-    InnerOwned<T, U3>: Send,
+    Owned<T, U4>: Send,
+    Owned<T, U3>: Send,
 {
     #[inline]
     fn arbitrary(g: &mut Gen) -> Self {
diff --git a/src/geometry/quaternion_conversion.rs b/src/geometry/quaternion_conversion.rs
index d12797d2..6dfbfbc6 100644
--- a/src/geometry/quaternion_conversion.rs
+++ b/src/geometry/quaternion_conversion.rs
@@ -28,7 +28,8 @@ use crate::geometry::{
 
 impl<T1, T2> SubsetOf<Quaternion<T2>> for Quaternion<T1>
 where
-    T2: SupersetOf<T1>,
+    T1: Scalar,
+    T2: Scalar + SupersetOf<T1>,
 {
     #[inline]
     fn to_superset(&self) -> Quaternion<T2> {
@@ -50,7 +51,8 @@ where
 
 impl<T1, T2> SubsetOf<UnitQuaternion<T2>> for UnitQuaternion<T1>
 where
-    T2: SupersetOf<T1>,
+    T1: Scalar,
+    T2: Scalar + SupersetOf<T1>,
 {
     #[inline]
     fn to_superset(&self) -> UnitQuaternion<T2> {
@@ -237,14 +239,14 @@ where
     }
 }
 
-impl<T> From<Vector4<T>> for Quaternion<T> {
+impl<T: Scalar> From<Vector4<T>> for Quaternion<T> {
     #[inline]
     fn from(coords: Vector4<T>) -> Self {
         Self { coords }
     }
 }
 
-impl<T> From<[T; 4]> for Quaternion<T> {
+impl<T: Scalar> From<[T; 4]> for Quaternion<T> {
     #[inline]
     fn from(coords: [T; 4]) -> Self {
         Self {
diff --git a/src/geometry/quaternion_coordinates.rs b/src/geometry/quaternion_coordinates.rs
index 40d8ca84..cb16e59e 100644
--- a/src/geometry/quaternion_coordinates.rs
+++ b/src/geometry/quaternion_coordinates.rs
@@ -12,14 +12,13 @@ impl<T: Scalar + SimdValue> Deref for Quaternion<T> {
 
     #[inline]
     fn deref(&self) -> &Self::Target {
-        // Safety: Self and IJKW are both stored as contiguous coordinates.
-        unsafe { &*(self as *const _ as *const _) }
+        unsafe { &*(self as *const Self as *const Self::Target) }
     }
 }
 
 impl<T: Scalar + SimdValue> DerefMut for Quaternion<T> {
     #[inline]
     fn deref_mut(&mut self) -> &mut Self::Target {
-        unsafe { &mut *(self as *mut _ as *mut _) }
+        unsafe { &mut *(self as *mut Self as *mut Self::Target) }
     }
 }
diff --git a/src/geometry/quaternion_ops.rs b/src/geometry/quaternion_ops.rs
index 12c371c2..eb7a15cd 100644
--- a/src/geometry/quaternion_ops.rs
+++ b/src/geometry/quaternion_ops.rs
@@ -59,12 +59,12 @@ use std::ops::{
 
 use crate::base::dimension::U3;
 use crate::base::storage::Storage;
-use crate::base::{Const, Unit, Vector, Vector3};
+use crate::base::{Const, Scalar, Unit, Vector, Vector3};
 use crate::SimdRealField;
 
 use crate::geometry::{Point3, Quaternion, Rotation, UnitQuaternion};
 
-impl<T> Index<usize> for Quaternion<T> {
+impl<T: Scalar> Index<usize> for Quaternion<T> {
     type Output = T;
 
     #[inline]
@@ -73,7 +73,7 @@ impl<T> Index<usize> for Quaternion<T> {
     }
 }
 
-impl<T> IndexMut<usize> for Quaternion<T> {
+impl<T: Scalar> IndexMut<usize> for Quaternion<T> {
     #[inline]
     fn index_mut(&mut self, i: usize) -> &mut T {
         &mut self.coords[i]
@@ -371,12 +371,12 @@ quaternion_op_impl!(
     ;
     self: Rotation<T, 3>, rhs: UnitQuaternion<T>,
     Output = UnitQuaternion<T>;
-    UnitQuaternion::<T>::from_rotation_matrix(&self) / rhs;);
+    UnitQuaternion::<T>::from_rotation_matrix(&self) / rhs; );
 
 // UnitQuaternion × Vector
 quaternion_op_impl!(
     Mul, mul;
-    SB: Storage<T, Const<3>>;
+    SB: Storage<T, Const<3>> ;
     self: &'a UnitQuaternion<T>, rhs: &'b Vector<T, Const<3>, SB>,
     Output = Vector3<T>;
     {
diff --git a/src/geometry/reflection.rs b/src/geometry/reflection.rs
index cc12594a..a48b8024 100644
--- a/src/geometry/reflection.rs
+++ b/src/geometry/reflection.rs
@@ -1,5 +1,3 @@
-use std::mem::MaybeUninit;
-
 use crate::base::constraint::{AreMultipliable, DimEq, SameNumberOfRows, ShapeConstraint};
 use crate::base::{Const, Matrix, Unit, Vector};
 use crate::dimension::{Dim, U1};
@@ -9,7 +7,7 @@ use simba::scalar::ComplexField;
 use crate::geometry::Point;
 
 /// A reflection wrt. a plane.
-pub struct Reflection<T, D: Dim, S> {
+pub struct Reflection<T, D, S> {
     axis: Vector<T, D, S>,
     bias: T,
 }
@@ -88,40 +86,40 @@ impl<T: ComplexField, D: Dim, S: Storage<T, D>> Reflection<T, D, S> {
     pub fn reflect_rows<R2: Dim, C2: Dim, S2, S3>(
         &self,
         lhs: &mut Matrix<T, R2, C2, S2>,
-        work: &mut Vector<MaybeUninit<T>, R2, S3>,
+        work: &mut Vector<T, R2, S3>,
     ) where
         S2: StorageMut<T, R2, C2>,
-        S3: StorageMut<MaybeUninit<T>, R2>,
+        S3: StorageMut<T, R2>,
         ShapeConstraint: DimEq<C2, D> + AreMultipliable<R2, C2, D, U1>,
     {
-        let mut work = lhs.mul_to(&self.axis, work);
+        lhs.mul_to(&self.axis, work);
 
         if !self.bias.is_zero() {
             work.add_scalar_mut(-self.bias);
         }
 
         let m_two: T = crate::convert(-2.0f64);
-        lhs.gerc(m_two, &work, &self.axis, T::one());
+        lhs.gerc(m_two, work, &self.axis, T::one());
     }
 
     /// Applies the reflection to the rows of `lhs`.
     pub fn reflect_rows_with_sign<R2: Dim, C2: Dim, S2, S3>(
         &self,
         lhs: &mut Matrix<T, R2, C2, S2>,
-        work: &mut Vector<MaybeUninit<T>, R2, S3>,
+        work: &mut Vector<T, R2, S3>,
         sign: T,
     ) where
         S2: StorageMut<T, R2, C2>,
-        S3: StorageMut<MaybeUninit<T>, R2>,
+        S3: StorageMut<T, R2>,
         ShapeConstraint: DimEq<C2, D> + AreMultipliable<R2, C2, D, U1>,
     {
-        let mut work = lhs.mul_to(&self.axis, work);
+        lhs.mul_to(&self.axis, work);
 
         if !self.bias.is_zero() {
             work.add_scalar_mut(-self.bias);
         }
 
         let m_two = sign.scale(crate::convert(-2.0f64));
-        lhs.gerc(m_two, &work, &self.axis, sign);
+        lhs.gerc(m_two, work, &self.axis, sign);
     }
 }
diff --git a/src/geometry/rotation.rs b/src/geometry/rotation.rs
index 7cde243a..33e42dda 100755
--- a/src/geometry/rotation.rs
+++ b/src/geometry/rotation.rs
@@ -9,8 +9,7 @@ use std::io::{Result as IOResult, Write};
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
 
 #[cfg(feature = "serde-serialize-no-std")]
-use crate::base::storage::InnerOwned;
-use crate::storage::InnerOwned;
+use crate::base::storage::Owned;
 
 #[cfg(feature = "abomonation-serialize")]
 use abomonation::Abomonation;
@@ -54,26 +53,29 @@ use crate::geometry::Point;
 /// # Conversion
 /// * [Conversion to a matrix <span style="float:right;">`matrix`, `to_homogeneous`…</span>](#conversion-to-a-matrix)
 ///
-#[repr(transparent)]
+#[repr(C)]
 #[derive(Debug)]
 pub struct Rotation<T, const D: usize> {
     matrix: SMatrix<T, D, D>,
 }
 
-impl<T: hash::Hash, const D: usize> hash::Hash for Rotation<T, D>
+impl<T: Scalar + hash::Hash, const D: usize> hash::Hash for Rotation<T, D>
 where
-    InnerOwned<T, Const<D>, Const<D>>: hash::Hash,
+    <DefaultAllocator as Allocator<T, Const<D>, Const<D>>>::Buffer: hash::Hash,
 {
     fn hash<H: hash::Hasher>(&self, state: &mut H) {
         self.matrix.hash(state)
     }
 }
 
-impl<T: Copy, const D: usize> Copy for Rotation<T, D> where InnerOwned<T, Const<D>, Const<D>>: Copy {}
+impl<T: Scalar + Copy, const D: usize> Copy for Rotation<T, D> where
+    <DefaultAllocator as Allocator<T, Const<D>, Const<D>>>::Buffer: Copy
+{
+}
 
-impl<T: Clone, const D: usize> Clone for Rotation<T, D>
+impl<T: Scalar, const D: usize> Clone for Rotation<T, D>
 where
-    InnerOwned<T, Const<D>, Const<D>>: Clone,
+    <DefaultAllocator as Allocator<T, Const<D>, Const<D>>>::Buffer: Clone,
 {
     #[inline]
     fn clone(&self) -> Self {
@@ -100,6 +102,7 @@ where
 #[cfg(feature = "abomonation-serialize")]
 impl<T, const D: usize> Abomonation for Rotation<T, D>
 where
+    T: Scalar,
     SMatrix<T, D, D>: Abomonation,
 {
     unsafe fn entomb<W: Write>(&self, writer: &mut W) -> IOResult<()> {
@@ -118,7 +121,7 @@ where
 #[cfg(feature = "serde-serialize-no-std")]
 impl<T: Scalar, const D: usize> Serialize for Rotation<T, D>
 where
-    InnerOwned<T, Const<D>, Const<D>>: Serialize,
+    Owned<T, Const<D>, Const<D>>: Serialize,
 {
     fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
     where
@@ -129,9 +132,9 @@ where
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<'a, T, const D: usize> Deserialize<'a> for Rotation<T, D>
+impl<'a, T: Scalar, const D: usize> Deserialize<'a> for Rotation<T, D>
 where
-    InnerOwned<T, Const<D>, Const<D>>: Deserialize<'a>,
+    Owned<T, Const<D>, Const<D>>: Deserialize<'a>,
 {
     fn deserialize<Des>(deserializer: Des) -> Result<Self, Des::Error>
     where
@@ -173,7 +176,7 @@ impl<T, const D: usize> Rotation<T, D> {
 }
 
 /// # Conversion to a matrix
-impl<T, const D: usize> Rotation<T, D> {
+impl<T: Scalar, const D: usize> Rotation<T, D> {
     /// A reference to the underlying matrix representation of this rotation.
     ///
     /// # Example
@@ -201,7 +204,7 @@ impl<T, const D: usize> Rotation<T, D> {
     /// A mutable reference to the underlying matrix representation of this rotation.
     #[inline]
     #[deprecated(note = "Use `.matrix_mut_unchecked()` instead.")]
-    pub  fn matrix_mut(&mut self) -> &mut SMatrix<T, D, D> {
+    pub unsafe fn matrix_mut(&mut self) -> &mut SMatrix<T, D, D> {
         &mut self.matrix
     }
 
@@ -274,7 +277,7 @@ impl<T, const D: usize> Rotation<T, D> {
     #[must_use]
     pub fn to_homogeneous(&self) -> OMatrix<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>
     where
-        T: Zero + One + Scalar,
+        T: Zero + One,
         Const<D>: DimNameAdd<U1>,
         DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
     {
diff --git a/src/geometry/rotation_specialization.rs b/src/geometry/rotation_specialization.rs
index 397f5bf6..5cd44119 100644
--- a/src/geometry/rotation_specialization.rs
+++ b/src/geometry/rotation_specialization.rs
@@ -1,5 +1,5 @@
 #[cfg(feature = "arbitrary")]
-use crate::base::storage::InnerOwned;
+use crate::base::storage::Owned;
 #[cfg(feature = "arbitrary")]
 use quickcheck::{Arbitrary, Gen};
 
@@ -284,7 +284,7 @@ where
 impl<T: SimdRealField + Arbitrary> Arbitrary for Rotation2<T>
 where
     T::Element: SimdRealField,
-    InnerOwned<T, U2, U2>: Send,
+    Owned<T, U2, U2>: Send,
 {
     #[inline]
     fn arbitrary(g: &mut Gen) -> Self {
@@ -976,8 +976,8 @@ where
 impl<T: SimdRealField + Arbitrary> Arbitrary for Rotation3<T>
 where
     T::Element: SimdRealField,
-    InnerOwned<T, U3, U3>: Send,
-    InnerOwned<T, U3>: Send,
+    Owned<T, U3, U3>: Send,
+    Owned<T, U3>: Send,
 {
     #[inline]
     fn arbitrary(g: &mut Gen) -> Self {
diff --git a/src/geometry/similarity.rs b/src/geometry/similarity.rs
index 506c0896..32a19772 100755
--- a/src/geometry/similarity.rs
+++ b/src/geometry/similarity.rs
@@ -17,11 +17,12 @@ use simba::simd::SimdRealField;
 
 use crate::base::allocator::Allocator;
 use crate::base::dimension::{DimNameAdd, DimNameSum, U1};
-use crate::base::storage::InnerOwned;
+use crate::base::storage::Owned;
 use crate::base::{Const, DefaultAllocator, OMatrix, SVector, Scalar};
 use crate::geometry::{AbstractRotation, Isometry, Point, Translation};
 
 /// A similarity, i.e., an uniform scaling, followed by a rotation, followed by a translation.
+#[repr(C)]
 #[derive(Debug)]
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
 #[cfg_attr(
@@ -64,7 +65,7 @@ where
 
 impl<T: Scalar + hash::Hash, R: hash::Hash, const D: usize> hash::Hash for Similarity<T, R, D>
 where
-    InnerOwned<T, Const<D>>: hash::Hash,
+    Owned<T, Const<D>>: hash::Hash,
 {
     fn hash<H: hash::Hasher>(&self, state: &mut H) {
         self.isometry.hash(state);
@@ -75,7 +76,7 @@ where
 impl<T: Scalar + Copy + Zero, R: AbstractRotation<T, D> + Copy, const D: usize> Copy
     for Similarity<T, R, D>
 where
-    InnerOwned<T, Const<D>>: Copy,
+    Owned<T, Const<D>>: Copy,
 {
 }
 
diff --git a/src/geometry/similarity_construction.rs b/src/geometry/similarity_construction.rs
index 1e2a29a0..feb5719b 100644
--- a/src/geometry/similarity_construction.rs
+++ b/src/geometry/similarity_construction.rs
@@ -1,5 +1,5 @@
 #[cfg(feature = "arbitrary")]
-use crate::base::storage::InnerOwned;
+use crate::base::storage::Owned;
 #[cfg(feature = "arbitrary")]
 use quickcheck::{Arbitrary, Gen};
 
@@ -109,7 +109,7 @@ where
     T: crate::RealField + Arbitrary + Send,
     T::Element: crate::RealField,
     R: AbstractRotation<T, D> + Arbitrary + Send,
-    InnerOwned<T, crate::Const<D>>: Send,
+    Owned<T, crate::Const<D>>: Send,
 {
     #[inline]
     fn arbitrary(rng: &mut Gen) -> Self {
diff --git a/src/geometry/transform.rs b/src/geometry/transform.rs
index a39ed75c..71544b59 100755
--- a/src/geometry/transform.rs
+++ b/src/geometry/transform.rs
@@ -1,6 +1,5 @@
 use approx::{AbsDiffEq, RelativeEq, UlpsEq};
 use std::any::Any;
-use std::fmt;
 use std::fmt::Debug;
 use std::hash;
 use std::marker::PhantomData;
@@ -8,11 +7,11 @@ use std::marker::PhantomData;
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
 
-use simba::scalar::{ComplexField, RealField};
+use simba::scalar::RealField;
 
 use crate::base::allocator::Allocator;
 use crate::base::dimension::{DimNameAdd, DimNameSum, U1};
-use crate::base::storage::InnerOwned;
+use crate::base::storage::Owned;
 use crate::base::{Const, DefaultAllocator, DimName, OMatrix, SVector};
 
 use crate::geometry::Point;
@@ -120,7 +119,7 @@ macro_rules! category_mul_impl(
     )*}
 );
 
-// We require stability upon multiplication.
+// We require stability uppon multiplication.
 impl<T: TCategory> TCategoryMul<T> for T {
     type Representative = T;
 }
@@ -157,8 +156,9 @@ super_tcategory_impl!(
 ///
 /// It is stored as a matrix with dimensions `(D + 1, D + 1)`, e.g., it stores a 4x4 matrix for a
 /// 3D transformation.
-#[repr(transparent)]
-pub struct Transform<T, C: TCategory, const D: usize>
+#[repr(C)]
+#[derive(Debug)]
+pub struct Transform<T: RealField, C: TCategory, const D: usize>
 where
     Const<D>: DimNameAdd<U1>,
     DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
@@ -167,32 +167,29 @@ where
     _phantom: PhantomData<C>,
 }
 
-impl<T: hash::Hash, C: TCategory, const D: usize> hash::Hash for Transform<T, C, D>
+impl<T: RealField + hash::Hash, C: TCategory, const D: usize> hash::Hash for Transform<T, C, D>
 where
     Const<D>: DimNameAdd<U1>,
     DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
-    InnerOwned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: hash::Hash,
+    Owned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: hash::Hash,
 {
     fn hash<H: hash::Hasher>(&self, state: &mut H) {
         self.matrix.hash(state);
     }
 }
 
-/*
-impl<T: Copy, C: TCategory, const D: usize> Copy for Transform<T, C, D>
+impl<T: RealField, C: TCategory, const D: usize> Copy for Transform<T, C, D>
 where
     Const<D>: DimNameAdd<U1>,
     DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
-    InnerOwned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: Copy,
+    Owned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: Copy,
 {
 }
-*/
 
-impl<T: Clone, C: TCategory, const D: usize> Clone for Transform<T, C, D>
+impl<T: RealField, C: TCategory, const D: usize> Clone for Transform<T, C, D>
 where
     Const<D>: DimNameAdd<U1>,
     DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
-    InnerOwned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: Clone,
 {
     #[inline]
     fn clone(&self) -> Self {
@@ -200,25 +197,33 @@ where
     }
 }
 
-impl<T: Debug, C: TCategory, const D: usize> Debug for Transform<T, C, D>
+#[cfg(feature = "bytemuck")]
+unsafe impl<T, C: TCategory, const D: usize> bytemuck::Zeroable for Transform<T, C, D>
 where
+    T: RealField + bytemuck::Zeroable,
     Const<D>: DimNameAdd<U1>,
     DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
-    InnerOwned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: Debug,
+    OMatrix<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: bytemuck::Zeroable,
+{
+}
+
+#[cfg(feature = "bytemuck")]
+unsafe impl<T, C: TCategory, const D: usize> bytemuck::Pod for Transform<T, C, D>
+where
+    T: RealField + bytemuck::Pod,
+    Const<D>: DimNameAdd<U1>,
+    DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
+    OMatrix<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: bytemuck::Pod,
+    Owned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: Copy,
 {
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("Transform")
-            .field("matrix", &self.matrix)
-            .finish()
-    }
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<T, C: TCategory, const D: usize> Serialize for Transform<T, C, D>
+impl<T: RealField, C: TCategory, const D: usize> Serialize for Transform<T, C, D>
 where
     Const<D>: DimNameAdd<U1>,
     DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
-    InnerOwned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: Serialize,
+    Owned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: Serialize,
 {
     fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
     where
@@ -229,11 +234,11 @@ where
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<'a, T, C: TCategory, const D: usize> Deserialize<'a> for Transform<T, C, D>
+impl<'a, T: RealField, C: TCategory, const D: usize> Deserialize<'a> for Transform<T, C, D>
 where
     Const<D>: DimNameAdd<U1>,
     DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
-    InnerOwned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: Deserialize<'a>,
+    Owned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: Deserialize<'a>,
 {
     fn deserialize<Des>(deserializer: Des) -> Result<Self, Des::Error>
     where
@@ -247,14 +252,14 @@ where
     }
 }
 
-impl<T: Eq, C: TCategory, const D: usize> Eq for Transform<T, C, D>
+impl<T: RealField + Eq, C: TCategory, const D: usize> Eq for Transform<T, C, D>
 where
     Const<D>: DimNameAdd<U1>,
     DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
 {
 }
 
-impl<T: PartialEq, C: TCategory, const D: usize> PartialEq for Transform<T, C, D>
+impl<T: RealField, C: TCategory, const D: usize> PartialEq for Transform<T, C, D>
 where
     Const<D>: DimNameAdd<U1>,
     DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
@@ -265,7 +270,7 @@ where
     }
 }
 
-impl<T, C: TCategory, const D: usize> Transform<T, C, D>
+impl<T: RealField, C: TCategory, const D: usize> Transform<T, C, D>
 where
     Const<D>: DimNameAdd<U1>,
     DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
@@ -370,10 +375,7 @@ where
     #[deprecated(
         note = "This method is redundant with automatic `Copy` and the `.clone()` method and will be removed in a future release."
     )]
-    pub fn clone_owned(&self) -> Transform<T, C, D>
-    where
-        T: Clone,
-    {
+    pub fn clone_owned(&self) -> Transform<T, C, D> {
         Transform::from_matrix_unchecked(self.matrix.clone_owned())
     }
 
@@ -391,10 +393,7 @@ where
     /// ```
     #[inline]
     #[must_use]
-    pub fn to_homogeneous(&self) -> OMatrix<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>
-    where
-        T: Clone,
-    {
+    pub fn to_homogeneous(&self) -> OMatrix<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>> {
         self.matrix().clone_owned()
     }
 
@@ -423,10 +422,7 @@ where
     /// ```
     #[inline]
     #[must_use = "Did you mean to use try_inverse_mut()?"]
-    pub fn try_inverse(self) -> Option<Transform<T, C, D>>
-    where
-        T: ComplexField,
-    {
+    pub fn try_inverse(self) -> Option<Transform<T, C, D>> {
         self.matrix
             .try_inverse()
             .map(Transform::from_matrix_unchecked)
@@ -452,7 +448,6 @@ where
     #[must_use = "Did you mean to use inverse_mut()?"]
     pub fn inverse(self) -> Transform<T, C, D>
     where
-        T: ComplexField,
         C: SubTCategoryOf<TProjective>,
     {
         // TODO: specialize for TAffine?
@@ -484,10 +479,7 @@ where
     /// assert!(!t.try_inverse_mut());
     /// ```
     #[inline]
-    pub fn try_inverse_mut(&mut self) -> bool
-    where
-        T: ComplexField,
-    {
+    pub fn try_inverse_mut(&mut self) -> bool {
         self.matrix.try_inverse_mut()
     }
 
@@ -511,7 +503,6 @@ where
     #[inline]
     pub fn inverse_mut(&mut self)
     where
-        T: ComplexField,
         C: SubTCategoryOf<TProjective>,
     {
         let _ = self.matrix.try_inverse_mut();
@@ -552,8 +543,8 @@ where
     Const<D>: DimNameAdd<U1>,
     C: SubTCategoryOf<TProjective>,
     DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>
-        + Allocator<T, DimNameSum<Const<D>, U1>>,
-    InnerOwned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: Clone,
+        + Allocator<T, DimNameSum<Const<D>, U1>>, // + Allocator<T, D, D>
+                                                  // + Allocator<T, D>
 {
     /// Transform the given point by the inverse of this transformation.
     /// This may be cheaper than inverting the transformation and transforming
diff --git a/src/geometry/transform_ops.rs b/src/geometry/transform_ops.rs
index 8b4be18f..94ef4ab3 100644
--- a/src/geometry/transform_ops.rs
+++ b/src/geometry/transform_ops.rs
@@ -9,7 +9,6 @@ use simba::scalar::{ClosedAdd, ClosedMul, RealField, SubsetOf};
 use crate::base::allocator::Allocator;
 use crate::base::dimension::{DimNameAdd, DimNameSum, U1};
 use crate::base::{Const, DefaultAllocator, OMatrix, SVector, Scalar};
-use crate::storage::InnerOwned;
 
 use crate::geometry::{
     Isometry, Point, Rotation, Similarity, SubTCategoryOf, SuperTCategoryOf, TAffine, TCategory,
@@ -373,8 +372,7 @@ md_impl_all!(
     const D;
     for CA, CB;
     where Const<D>: DimNameAdd<U1>, CA: TCategoryMul<CB>, CB: SubTCategoryOf<TProjective>,
-          DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
-          Transform<T, CB, D>: Clone; // There's probably a better bound here.
+          DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>;
     self: Transform<T, CA, D>, rhs: Transform<T, CB, D>, Output = Transform<T, CA::Representative, D>;
     [val val] => #[allow(clippy::suspicious_arithmetic_impl)] { self * rhs.inverse() };
     [ref val] => #[allow(clippy::suspicious_arithmetic_impl)] { self * rhs.inverse() };
@@ -628,8 +626,7 @@ md_assign_impl_all!(
     const D;
     for CA, CB;
     where Const<D>: DimNameAdd<U1>, CA: SuperTCategoryOf<CB>, CB: SubTCategoryOf<TProjective>,
-          DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
-          InnerOwned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: Clone;
+          DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>;
     self: Transform<T, CA, D>, rhs: Transform<T, CB, D>;
     [val] => #[allow(clippy::suspicious_op_assign_impl)] { *self *= rhs.inverse() };
     [ref] => #[allow(clippy::suspicious_op_assign_impl)] { *self *= rhs.clone().inverse() };
diff --git a/src/geometry/translation.rs b/src/geometry/translation.rs
index 6f983fec..1dd6f6d5 100755
--- a/src/geometry/translation.rs
+++ b/src/geometry/translation.rs
@@ -15,13 +15,13 @@ use simba::scalar::{ClosedAdd, ClosedNeg, ClosedSub};
 
 use crate::base::allocator::Allocator;
 use crate::base::dimension::{DimNameAdd, DimNameSum, U1};
-use crate::base::storage::InnerOwned;
+use crate::base::storage::Owned;
 use crate::base::{Const, DefaultAllocator, OMatrix, SVector, Scalar};
 
 use crate::geometry::Point;
 
 /// A translation.
-#[repr(transparent)]
+#[repr(C)]
 #[derive(Debug)]
 pub struct Translation<T, const D: usize> {
     /// The translation coordinates, i.e., how much is added to a point's coordinates when it is
@@ -29,20 +29,20 @@ pub struct Translation<T, const D: usize> {
     pub vector: SVector<T, D>,
 }
 
-impl<T: hash::Hash, const D: usize> hash::Hash for Translation<T, D>
+impl<T: Scalar + hash::Hash, const D: usize> hash::Hash for Translation<T, D>
 where
-    InnerOwned<T, Const<D>>: hash::Hash,
+    Owned<T, Const<D>>: hash::Hash,
 {
     fn hash<H: hash::Hasher>(&self, state: &mut H) {
         self.vector.hash(state)
     }
 }
 
-impl<T: Copy, const D: usize> Copy for Translation<T, D> {}
+impl<T: Scalar + Copy, const D: usize> Copy for Translation<T, D> {}
 
-impl<T: Clone, const D: usize> Clone for Translation<T, D>
+impl<T: Scalar, const D: usize> Clone for Translation<T, D>
 where
-    InnerOwned<T, Const<D>>: Clone,
+    Owned<T, Const<D>>: Clone,
 {
     #[inline]
     fn clone(&self) -> Self {
@@ -69,6 +69,7 @@ where
 #[cfg(feature = "abomonation-serialize")]
 impl<T, const D: usize> Abomonation for Translation<T, D>
 where
+    T: Scalar,
     SVector<T, D>: Abomonation,
 {
     unsafe fn entomb<W: Write>(&self, writer: &mut W) -> IOResult<()> {
@@ -85,9 +86,9 @@ where
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<T, const D: usize> Serialize for Translation<T, D>
+impl<T: Scalar, const D: usize> Serialize for Translation<T, D>
 where
-    InnerOwned<T, Const<D>>: Serialize,
+    Owned<T, Const<D>>: Serialize,
 {
     fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
     where
@@ -98,9 +99,9 @@ where
 }
 
 #[cfg(feature = "serde-serialize-no-std")]
-impl<'a, T, const D: usize> Deserialize<'a> for Translation<T, D>
+impl<'a, T: Scalar, const D: usize> Deserialize<'a> for Translation<T, D>
 where
-    InnerOwned<T, Const<D>>: Deserialize<'a>,
+    Owned<T, Const<D>>: Deserialize<'a>,
 {
     fn deserialize<Des>(deserializer: Des) -> Result<Self, Des::Error>
     where
@@ -155,7 +156,7 @@ mod rkyv_impl {
     }
 }
 
-impl<T, const D: usize> Translation<T, D> {
+impl<T: Scalar, const D: usize> Translation<T, D> {
     /// Creates a new translation from the given vector.
     #[inline]
     #[deprecated(note = "Use `::from` instead.")]
@@ -181,7 +182,7 @@ impl<T, const D: usize> Translation<T, D> {
     #[must_use = "Did you mean to use inverse_mut()?"]
     pub fn inverse(&self) -> Translation<T, D>
     where
-        T: ClosedNeg + Scalar,
+        T: ClosedNeg,
     {
         Translation::from(-&self.vector)
     }
@@ -208,7 +209,7 @@ impl<T, const D: usize> Translation<T, D> {
     #[must_use]
     pub fn to_homogeneous(&self) -> OMatrix<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>
     where
-        T: Zero + One + Scalar,
+        T: Zero + One,
         Const<D>: DimNameAdd<U1>,
         DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
     {
@@ -239,7 +240,7 @@ impl<T, const D: usize> Translation<T, D> {
     #[inline]
     pub fn inverse_mut(&mut self)
     where
-        T: ClosedNeg + Scalar,
+        T: ClosedNeg,
     {
         self.vector.neg_mut()
     }
@@ -279,16 +280,16 @@ impl<T: Scalar + ClosedSub, const D: usize> Translation<T, D> {
     }
 }
 
-impl<T: Eq, const D: usize> Eq for Translation<T, D> {}
+impl<T: Scalar + Eq, const D: usize> Eq for Translation<T, D> {}
 
-impl<T: PartialEq, const D: usize> PartialEq for Translation<T, D> {
+impl<T: Scalar + PartialEq, const D: usize> PartialEq for Translation<T, D> {
     #[inline]
     fn eq(&self, right: &Translation<T, D>) -> bool {
         self.vector == right.vector
     }
 }
 
-impl<T: AbsDiffEq, const D: usize> AbsDiffEq for Translation<T, D>
+impl<T: Scalar + AbsDiffEq, const D: usize> AbsDiffEq for Translation<T, D>
 where
     T::Epsilon: Copy,
 {
@@ -305,7 +306,7 @@ where
     }
 }
 
-impl<T: RelativeEq, const D: usize> RelativeEq for Translation<T, D>
+impl<T: Scalar + RelativeEq, const D: usize> RelativeEq for Translation<T, D>
 where
     T::Epsilon: Copy,
 {
@@ -326,7 +327,7 @@ where
     }
 }
 
-impl<T: UlpsEq, const D: usize> UlpsEq for Translation<T, D>
+impl<T: Scalar + UlpsEq, const D: usize> UlpsEq for Translation<T, D>
 where
     T::Epsilon: Copy,
 {
diff --git a/src/geometry/translation_construction.rs b/src/geometry/translation_construction.rs
index a9f501be..5371b648 100644
--- a/src/geometry/translation_construction.rs
+++ b/src/geometry/translation_construction.rs
@@ -1,5 +1,5 @@
 #[cfg(feature = "arbitrary")]
-use crate::base::storage::InnerOwned;
+use crate::base::storage::Owned;
 #[cfg(feature = "arbitrary")]
 use quickcheck::{Arbitrary, Gen};
 
@@ -77,7 +77,7 @@ where
 #[cfg(feature = "arbitrary")]
 impl<T: Scalar + Arbitrary + Send, const D: usize> Arbitrary for Translation<T, D>
 where
-    InnerOwned<T, crate::Const<D>>: Send,
+    Owned<T, crate::Const<D>>: Send,
 {
     #[inline]
     fn arbitrary(rng: &mut Gen) -> Self {
diff --git a/src/geometry/translation_conversion.rs b/src/geometry/translation_conversion.rs
index bed39f7a..d443a2f4 100644
--- a/src/geometry/translation_conversion.rs
+++ b/src/geometry/translation_conversion.rs
@@ -27,7 +27,8 @@ use crate::Point;
 
 impl<T1, T2, const D: usize> SubsetOf<Translation<T2, D>> for Translation<T1, D>
 where
-    T2: SupersetOf<T1>,
+    T1: Scalar,
+    T2: Scalar + SupersetOf<T1>,
 {
     #[inline]
     fn to_superset(&self) -> Translation<T2, D> {
@@ -192,14 +193,14 @@ where
     }
 }
 
-impl<T, const D: usize> From<OVector<T, Const<D>>> for Translation<T, D> {
+impl<T: Scalar, const D: usize> From<OVector<T, Const<D>>> for Translation<T, D> {
     #[inline]
     fn from(vector: OVector<T, Const<D>>) -> Self {
         Translation { vector }
     }
 }
 
-impl<T, const D: usize> From<[T; D]> for Translation<T, D> {
+impl<T: Scalar, const D: usize> From<[T; D]> for Translation<T, D> {
     #[inline]
     fn from(coords: [T; D]) -> Self {
         Translation {
@@ -208,17 +209,14 @@ impl<T, const D: usize> From<[T; D]> for Translation<T, D> {
     }
 }
 
-impl<T, const D: usize> From<Point<T, D>> for Translation<T, D> {
+impl<T: Scalar, const D: usize> From<Point<T, D>> for Translation<T, D> {
     #[inline]
     fn from(pt: Point<T, D>) -> Self {
         Translation { vector: pt.coords }
     }
 }
 
-impl<T, const D: usize> From<Translation<T, D>> for [T; D]
-where
-    T: Clone,
-{
+impl<T: Scalar, const D: usize> From<Translation<T, D>> for [T; D] {
     #[inline]
     fn from(t: Translation<T, D>) -> Self {
         t.vector.into()
diff --git a/src/geometry/translation_coordinates.rs b/src/geometry/translation_coordinates.rs
index bda57f59..80267e06 100644
--- a/src/geometry/translation_coordinates.rs
+++ b/src/geometry/translation_coordinates.rs
@@ -18,14 +18,14 @@ macro_rules! deref_impl(
 
             #[inline]
             fn deref(&self) -> &Self::Target {
-                unsafe { &*(self as *const _ as *const _) }
+                unsafe { &*(self as *const Translation<T, $D> as *const Self::Target) }
             }
         }
 
         impl<T: Scalar> DerefMut for Translation<T, $D> {
             #[inline]
             fn deref_mut(&mut self) -> &mut Self::Target {
-                unsafe { &mut *(self as *mut _ as *mut _) }
+                unsafe { &mut *(self as *mut Translation<T, $D> as *mut Self::Target) }
             }
         }
     }
diff --git a/src/lib.rs b/src/lib.rs
index e21f0709..650a601a 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -77,12 +77,12 @@ an optimized set of tools for computer graphics and physics. Those features incl
     unused_parens,
     unused_qualifications,
     unused_results,
-    missing_docs,
     rust_2018_idioms,
     rust_2018_compatibility,
     future_incompatible,
     missing_copy_implementations
 )]
+// #![deny(missing_docs)] // XXX: deny that
 #![doc(
     html_favicon_url = "https://nalgebra.org/img/favicon.ico",
     html_root_url = "https://docs.rs/nalgebra/0.25.0"
diff --git a/src/linalg/balancing.rs b/src/linalg/balancing.rs
index f4f8b659..15679e2b 100644
--- a/src/linalg/balancing.rs
+++ b/src/linalg/balancing.rs
@@ -5,7 +5,6 @@ use std::ops::{DivAssign, MulAssign};
 
 use crate::allocator::Allocator;
 use crate::base::dimension::Dim;
-use crate::base::storage::Storage;
 use crate::base::{Const, DefaultAllocator, OMatrix, OVector};
 
 /// Applies in-place a modified Parlett and Reinsch matrix balancing with 2-norm to the matrix and returns
@@ -18,7 +17,7 @@ where
 {
     assert!(matrix.is_square(), "Unable to balance a non-square matrix.");
 
-    let dim = matrix.data.shape().0;
+    let dim = matrix.shape_generic().0;
     let radix: T = crate::convert(2.0f64);
     let mut d = OVector::from_element_generic(dim, Const::<1>, T::one());
 
diff --git a/src/linalg/bidiagonal.rs b/src/linalg/bidiagonal.rs
index d4b6a1e3..e269b4a0 100644
--- a/src/linalg/bidiagonal.rs
+++ b/src/linalg/bidiagonal.rs
@@ -1,17 +1,14 @@
-use std::fmt;
-
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Serialize};
 
 use crate::allocator::Allocator;
 use crate::base::{DefaultAllocator, Matrix, OMatrix, OVector, Unit};
 use crate::dimension::{Const, Dim, DimDiff, DimMin, DimMinimum, DimSub, U1};
-use crate::storage::{InnerOwned, Storage};
-use crate::Dynamic;
 use simba::scalar::ComplexField;
 
 use crate::geometry::Reflection;
 use crate::linalg::householder;
+use std::mem::MaybeUninit;
 
 /// The bidiagonalization of a general matrix.
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
@@ -35,6 +32,7 @@ use crate::linalg::householder;
          OVector<T, DimMinimum<R, C>>: Deserialize<'de>,
          OVector<T, DimDiff<DimMinimum<R, C>, U1>>: Deserialize<'de>"))
 )]
+#[derive(Clone, Debug)]
 pub struct Bidiagonal<T: ComplexField, R: DimMin<C>, C: Dim>
 where
     DimMinimum<R, C>: DimSub<U1>,
@@ -52,59 +50,17 @@ where
     upper_diagonal: bool,
 }
 
-impl<T: ComplexField, R: DimMin<C>, C: Dim> Clone for Bidiagonal<T, R, C>
-where
-    DimMinimum<R, C>: DimSub<U1>,
-    DefaultAllocator: Allocator<T, R, C>
-        + Allocator<T, DimMinimum<R, C>>
-        + Allocator<T, DimDiff<DimMinimum<R, C>, U1>>,
-    InnerOwned<T, R, C>: Clone,
-    InnerOwned<T, DimMinimum<R, C>>: Clone,
-    InnerOwned<T, DimDiff<DimMinimum<R, C>, U1>>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self {
-            uv: self.uv.clone(),
-            diagonal: self.diagonal.clone(),
-            off_diagonal: self.off_diagonal.clone(),
-            upper_diagonal: self.upper_diagonal,
-        }
-    }
-}
-
-/*
 impl<T: ComplexField, R: DimMin<C>, C: Dim> Copy for Bidiagonal<T, R, C>
 where
     DimMinimum<R, C>: DimSub<U1>,
     DefaultAllocator: Allocator<T, R, C>
         + Allocator<T, DimMinimum<R, C>>
         + Allocator<T, DimDiff<DimMinimum<R, C>, U1>>,
-    InnerOwned<T, R, C>: Copy,
-    InnerOwned<T, DimMinimum<R, C>>: Copy,
-    InnerOwned<T, DimDiff<DimMinimum<R, C>, U1>>: Copy,
+    OMatrix<T, R, C>: Copy,
+    OVector<T, DimMinimum<R, C>>: Copy,
+    OVector<T, DimDiff<DimMinimum<R, C>, U1>>: Copy,
 {
 }
-*/
-
-impl<T: ComplexField, R: DimMin<C>, C: Dim> fmt::Debug for Bidiagonal<T, R, C>
-where
-    DimMinimum<R, C>: DimSub<U1>,
-    DefaultAllocator: Allocator<T, R, C>
-        + Allocator<T, DimMinimum<R, C>>
-        + Allocator<T, DimDiff<DimMinimum<R, C>, U1>>,
-    InnerOwned<T, R, C>: fmt::Debug,
-    InnerOwned<T, DimMinimum<R, C>>: fmt::Debug,
-    InnerOwned<T, DimDiff<DimMinimum<R, C>, U1>>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("Bidiagonal")
-            .field("uv", &self.uv)
-            .field("diagonal", &self.diagonal)
-            .field("off_diagonal", &self.off_diagonal)
-            .field("upper_diagonal", &self.upper_diagonal)
-            .finish()
-    }
-}
 
 impl<T: ComplexField, R: DimMin<C>, C: Dim> Bidiagonal<T, R, C>
 where
@@ -117,7 +73,7 @@ where
 {
     /// Computes the Bidiagonal decomposition using householder reflections.
     pub fn new(mut matrix: OMatrix<T, R, C>) -> Self {
-        let (nrows, ncols) = matrix.data.shape();
+        let (nrows, ncols) = matrix.shape_generic();
         let min_nrows_ncols = nrows.min(ncols);
         let dim = min_nrows_ncols.value();
         assert!(
@@ -125,80 +81,70 @@ where
             "Cannot compute the bidiagonalization of an empty matrix."
         );
 
-        let mut diagonal = Matrix::new_uninitialized_generic(min_nrows_ncols, Const::<1>);
-        let mut off_diagonal =
-            Matrix::new_uninitialized_generic(min_nrows_ncols.sub(Const::<1>), Const::<1>);
-        let mut axis_packed = Matrix::new_uninitialized_generic(ncols, Const::<1>);
-        let mut work = Matrix::new_uninitialized_generic(nrows, Const::<1>);
+        let mut diagonal = Matrix::uninit(min_nrows_ncols, Const::<1>);
+        let mut off_diagonal = Matrix::uninit(min_nrows_ncols.sub(Const::<1>), Const::<1>);
+        let mut axis_packed = Matrix::zeros_generic(ncols, Const::<1>);
+        let mut work = Matrix::zeros_generic(nrows, Const::<1>);
 
         let upper_diagonal = nrows.value() >= ncols.value();
-
-        // Safety: all pointers involved are valid for writes, aligned, and uninitialized.
-        unsafe {
-            if upper_diagonal {
-                for ite in 0..dim - 1 {
-                    householder::clear_column_unchecked(
-                        &mut matrix,
-                        diagonal[ite].as_mut_ptr(),
-                        ite,
-                        0,
-                        None,
-                    );
-                    householder::clear_row_unchecked(
-                        &mut matrix,
-                        off_diagonal[ite].as_mut_ptr(),
-                        &mut axis_packed,
-                        &mut work,
-                        ite,
-                        1,
-                    );
-                }
-
-                householder::clear_column_unchecked(
+        if upper_diagonal {
+            for ite in 0..dim - 1 {
+                diagonal[ite] = MaybeUninit::new(householder::clear_column_unchecked(
                     &mut matrix,
-                    diagonal[dim - 1].as_mut_ptr(),
-                    dim - 1,
+                    ite,
                     0,
                     None,
-                );
-            } else {
-                for ite in 0..dim - 1 {
-                    householder::clear_row_unchecked(
-                        &mut matrix,
-                        diagonal[ite].as_mut_ptr(),
-                        &mut axis_packed,
-                        &mut work,
-                        ite,
-                        0,
-                    );
-                    householder::clear_column_unchecked(
-                        &mut matrix,
-                        off_diagonal[ite].as_mut_ptr(),
-                        ite,
-                        1,
-                        None,
-                    );
-                }
-
-                householder::clear_row_unchecked(
+                ));
+                off_diagonal[ite] = MaybeUninit::new(householder::clear_row_unchecked(
                     &mut matrix,
-                    diagonal[dim - 1].as_mut_ptr(),
                     &mut axis_packed,
                     &mut work,
-                    dim - 1,
-                    0,
-                );
+                    ite,
+                    1,
+                ));
             }
+
+            diagonal[dim - 1] = MaybeUninit::new(householder::clear_column_unchecked(
+                &mut matrix,
+                dim - 1,
+                0,
+                None,
+            ));
+        } else {
+            for ite in 0..dim - 1 {
+                diagonal[ite] = MaybeUninit::new(householder::clear_row_unchecked(
+                    &mut matrix,
+                    &mut axis_packed,
+                    &mut work,
+                    ite,
+                    0,
+                ));
+                off_diagonal[ite] = MaybeUninit::new(householder::clear_column_unchecked(
+                    &mut matrix,
+                    ite,
+                    1,
+                    None,
+                ));
+            }
+
+            diagonal[dim - 1] = MaybeUninit::new(householder::clear_row_unchecked(
+                &mut matrix,
+                &mut axis_packed,
+                &mut work,
+                dim - 1,
+                0,
+            ));
         }
 
-        // Safety: all values have been initialized.
-        unsafe {
-            Bidiagonal {
-                uv: matrix,
-                diagonal: diagonal.assume_init(),
-                off_diagonal: off_diagonal.assume_init(),
-                upper_diagonal,
-            }
+        // Safety: diagonal and off_diagonal have been fully initialized.
+        let (diagonal, off_diagonal) =
+            unsafe { (diagonal.assume_init(), off_diagonal.assume_init()) };
+
+        Bidiagonal {
+            uv: matrix,
+            diagonal,
+            off_diagonal,
+            upper_diagonal,
         }
     }
 
@@ -245,7 +191,7 @@ where
     where
         DefaultAllocator: Allocator<T, DimMinimum<R, C>, DimMinimum<R, C>>,
     {
-        let (nrows, ncols) = self.uv.data.shape();
+        let (nrows, ncols) = self.uv.shape_generic();
 
         let d = nrows.min(ncols);
         let mut res = OMatrix::identity_generic(d, d);
@@ -265,7 +211,7 @@ where
     where
         DefaultAllocator: Allocator<T, R, DimMinimum<R, C>>,
     {
-        let (nrows, ncols) = self.uv.data.shape();
+        let (nrows, ncols) = self.uv.shape_generic();
 
         let mut res = Matrix::identity_generic(nrows, nrows.min(ncols));
         let dim = self.diagonal.len();
@@ -294,23 +240,21 @@ where
     #[must_use]
     pub fn v_t(&self) -> OMatrix<T, DimMinimum<R, C>, C>
     where
-        DefaultAllocator: Allocator<T, DimMinimum<R, C>, C> + Allocator<T, Dynamic, U1>,
+        DefaultAllocator: Allocator<T, DimMinimum<R, C>, C>,
     {
-        let (nrows, ncols) = self.uv.data.shape();
+        let (nrows, ncols) = self.uv.shape_generic();
         let min_nrows_ncols = nrows.min(ncols);
 
         let mut res = Matrix::identity_generic(min_nrows_ncols, ncols);
-        let mut work = Matrix::new_uninitialized_generic(min_nrows_ncols, Const::<1>);
-        let mut axis_packed = Matrix::new_uninitialized_generic(ncols, Const::<1>);
+        let mut work = Matrix::zeros_generic(min_nrows_ncols, Const::<1>);
+        let mut axis_packed = Matrix::zeros_generic(ncols, Const::<1>);
 
         let shift = self.axis_shift().1;
 
         for i in (0..min_nrows_ncols.value() - shift).rev() {
             let axis = self.uv.slice_range(i, i + shift..);
             let mut axis_packed = axis_packed.rows_range_mut(i + shift..);
-            axis_packed.tr_copy_init_from(&axis);
-            let axis_packed = unsafe { axis_packed.slice_assume_init() };
-
+            axis_packed.tr_copy_from(&axis);
             // TODO: sometimes, the axis might have a zero magnitude.
             let refl = Reflection::new(Unit::new_unchecked(axis_packed), T::zero());
 
@@ -404,7 +348,7 @@ where
 //         assert!(self.uv.is_square(), "Bidiagonal inverse: unable to compute the inverse of a non-square matrix.");
 //
 //         // TODO: is there a less naive method ?
-//         let (nrows, ncols) = self.uv.data.shape();
+//         let (nrows, ncols) = self.uv.shape_generic();
 //         let mut res = OMatrix::identity_generic(nrows, ncols);
 //         self.solve_mut(&mut res);
 //         res
diff --git a/src/linalg/cholesky.rs b/src/linalg/cholesky.rs
index 2abd8242..47939311 100644
--- a/src/linalg/cholesky.rs
+++ b/src/linalg/cholesky.rs
@@ -1,6 +1,3 @@
-use std::fmt;
-use std::mem::MaybeUninit;
-
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Serialize};
 
@@ -12,7 +9,7 @@ use crate::allocator::Allocator;
 use crate::base::{Const, DefaultAllocator, Matrix, OMatrix, Vector};
 use crate::constraint::{SameNumberOfRows, ShapeConstraint};
 use crate::dimension::{Dim, DimAdd, DimDiff, DimSub, DimSum, U1};
-use crate::storage::{InnerOwned, Storage, StorageMut};
+use crate::storage::{Storage, StorageMut};
 
 /// The Cholesky decomposition of a symmetric-definite-positive matrix.
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
@@ -26,6 +23,7 @@ use crate::storage::{InnerOwned, Storage, StorageMut};
     serde(bound(deserialize = "DefaultAllocator: Allocator<T, D>,
          OMatrix<T, D, D>: Deserialize<'de>"))
 )]
+#[derive(Clone, Debug)]
 pub struct Cholesky<T: SimdComplexField, D: Dim>
 where
     DefaultAllocator: Allocator<T, D, D>,
@@ -33,38 +31,12 @@ where
     chol: OMatrix<T, D, D>,
 }
 
-/*
 impl<T: SimdComplexField, D: Dim> Copy for Cholesky<T, D>
 where
     DefaultAllocator: Allocator<T, D, D>,
-    InnerOwned<T, D, D>: Copy,
+    OMatrix<T, D, D>: Copy,
 {
 }
-*/
-
-impl<T: SimdComplexField, D: Dim> Clone for Cholesky<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D>,
-    InnerOwned<T, D, D>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self {
-            chol: self.chol.clone(),
-        }
-    }
-}
-
-impl<T: SimdComplexField, D: Dim> fmt::Debug for Cholesky<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D>,
-    InnerOwned<T, D, D>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("Cholesky")
-            .field("chol", &self.chol)
-            .finish()
-    }
-}
 
 impl<T: SimdComplexField, D: Dim> Cholesky<T, D>
 where
@@ -164,7 +136,7 @@ where
     /// Computes the inverse of the decomposed matrix.
     #[must_use]
     pub fn inverse(&self) -> OMatrix<T, D, D> {
-        let shape = self.chol.data.shape();
+        let shape = self.chol.shape_generic();
         let mut res = OMatrix::identity_generic(shape.0, shape.1);
 
         self.solve_mut(&mut res);
@@ -254,8 +226,6 @@ where
         DefaultAllocator: Allocator<T, DimSum<D, U1>, DimSum<D, U1>> + Allocator<T, R2>,
         ShapeConstraint: SameNumberOfRows<R2, DimSum<D, U1>>,
     {
-        // TODO: check that MaybeUninit manipulations are sound!
-
         let mut col = col.into_owned();
         // for an explanation of the formulas, see https://en.wikipedia.org/wiki/Cholesky_decomposition#Updating_the_decomposition
         let n = col.nrows();
@@ -267,20 +237,19 @@ where
         assert!(j < n, "j needs to be within the bound of the new matrix.");
 
         // loads the data into a new matrix with an additional jth row/column
-        let mut chol = Matrix::new_uninitialized_generic(
-            self.chol.data.shape().0.add(Const::<1>),
-            self.chol.data.shape().1.add(Const::<1>),
+        // TODO: would it be worth it to avoid the zero-initialization?
+        let mut chol = Matrix::zeros_generic(
+            self.chol.shape_generic().0.add(Const::<1>),
+            self.chol.shape_generic().1.add(Const::<1>),
         );
-
-        // TODO: checked that every entry is initialized EXACTLY once.
         chol.slice_range_mut(..j, ..j)
-            .copy_init_from(&self.chol.slice_range(..j, ..j));
+            .copy_from(&self.chol.slice_range(..j, ..j));
         chol.slice_range_mut(..j, j + 1..)
-            .copy_init_from(&self.chol.slice_range(..j, j..));
+            .copy_from(&self.chol.slice_range(..j, j..));
         chol.slice_range_mut(j + 1.., ..j)
-            .copy_init_from(&self.chol.slice_range(j.., ..j));
+            .copy_from(&self.chol.slice_range(j.., ..j));
         chol.slice_range_mut(j + 1.., j + 1..)
-            .copy_init_from(&self.chol.slice_range(j.., j..));
+            .copy_from(&self.chol.slice_range(j.., j..));
 
         // update the jth row
         let top_left_corner = self.chol.slice_range(..j, ..j);
@@ -296,7 +265,7 @@ where
 
         // update the center element
         let center_element = T::sqrt(col_j - T::from_real(new_rowj_adjoint.norm_squared()));
-        chol[(j, j)] = MaybeUninit::new(center_element);
+        chol[(j, j)] = center_element;
 
         // update the jth column
         let bottom_left_corner = self.chol.slice_range(j.., ..j);
@@ -307,9 +276,7 @@ where
             &new_rowj_adjoint,
             T::one() / center_element,
         );
-        chol.slice_range_mut(j + 1.., j).copy_init_from(&new_colj);
-
-        let mut chol = unsafe { chol.assume_init() };
+        chol.slice_range_mut(j + 1.., j).copy_from(&new_colj);
 
         // update the bottom right corner
         let mut bottom_right_corner = chol.slice_range_mut(j + 1.., j + 1..);
@@ -330,27 +297,24 @@ where
         D: DimSub<U1>,
         DefaultAllocator: Allocator<T, DimDiff<D, U1>, DimDiff<D, U1>> + Allocator<T, D>,
     {
-        // TODO: check that MaybeUninit manipulations are sound!
-
         let n = self.chol.nrows();
         assert!(n > 0, "The matrix needs at least one column.");
         assert!(j < n, "j needs to be within the bound of the matrix.");
 
         // loads the data into a new matrix except for the jth row/column
-        let mut chol = Matrix::new_uninitialized_generic(
-            self.chol.data.shape().0.sub(Const::<1>),
-            self.chol.data.shape().1.sub(Const::<1>),
+        // TODO: would it be worth it to avoid this zero initialization?
+        let mut chol = Matrix::zeros_generic(
+            self.chol.shape_generic().0.sub(Const::<1>),
+            self.chol.shape_generic().1.sub(Const::<1>),
         );
-
         chol.slice_range_mut(..j, ..j)
-            .copy_init_from(&self.chol.slice_range(..j, ..j));
+            .copy_from(&self.chol.slice_range(..j, ..j));
         chol.slice_range_mut(..j, j..)
-            .copy_init_from(&self.chol.slice_range(..j, j + 1..));
+            .copy_from(&self.chol.slice_range(..j, j + 1..));
         chol.slice_range_mut(j.., ..j)
-            .copy_init_from(&self.chol.slice_range(j + 1.., ..j));
+            .copy_from(&self.chol.slice_range(j + 1.., ..j));
         chol.slice_range_mut(j.., j..)
-            .copy_init_from(&self.chol.slice_range(j + 1.., j + 1..));
-        let mut chol = unsafe { chol.assume_init() };
+            .copy_from(&self.chol.slice_range(j + 1.., j + 1..));
 
         // updates the bottom right corner
         let mut bottom_right_corner = chol.slice_range_mut(j.., j..);
@@ -366,12 +330,14 @@ where
     ///
     /// This helper method is called by `rank_one_update` but also `insert_column` and `remove_column`
     /// where it is used on a square slice of the decomposition
-    fn xx_rank_one_update<Dm: Dim, Sm, Rx: Dim, Sx>(
+    fn xx_rank_one_update<Dm, Sm, Rx, Sx>(
         chol: &mut Matrix<T, Dm, Dm, Sm>,
         x: &mut Vector<T, Rx, Sx>,
         sigma: T::RealField,
     ) where
         //T: ComplexField,
+        Dm: Dim,
+        Rx: Dim,
         Sm: StorageMut<T, Dm, Dm>,
         Sx: StorageMut<T, Rx, U1>,
     {
diff --git a/src/linalg/col_piv_qr.rs b/src/linalg/col_piv_qr.rs
index 438ee83a..f5c61336 100644
--- a/src/linalg/col_piv_qr.rs
+++ b/src/linalg/col_piv_qr.rs
@@ -6,11 +6,12 @@ use crate::allocator::{Allocator, Reallocator};
 use crate::base::{Const, DefaultAllocator, Matrix, OMatrix, OVector, Unit};
 use crate::constraint::{SameNumberOfRows, ShapeConstraint};
 use crate::dimension::{Dim, DimMin, DimMinimum};
-use crate::storage::{Storage, StorageMut};
+use crate::storage::StorageMut;
 use crate::ComplexField;
 
 use crate::geometry::Reflection;
 use crate::linalg::{householder, PermutationSequence};
+use std::mem::MaybeUninit;
 
 /// The QR decomposition (with column pivoting) of a general matrix.
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
@@ -30,6 +31,7 @@ use crate::linalg::{householder, PermutationSequence};
          PermutationSequence<DimMinimum<R, C>>: Deserialize<'de>,
          OVector<T, DimMinimum<R, C>>: Deserialize<'de>"))
 )]
+#[derive(Clone, Debug)]
 pub struct ColPivQR<T: ComplexField, R: DimMin<C>, C: Dim>
 where
     DefaultAllocator: Allocator<T, R, C>
@@ -52,24 +54,6 @@ where
 {
 }
 
-impl<T: ComplexField, R: DimMin<C>, C: Dim> Clone for ColPivQR<T, R, C>
-where
-    DefaultAllocator: Allocator<T, R, C>
-        + Allocator<T, DimMinimum<R, C>>
-        + Allocator<(usize, usize), DimMinimum<R, C>>,
-    OMatrix<T, R, C>: Clone,
-    PermutationSequence<DimMinimum<R, C>>: Clone,
-    OVector<T, DimMinimum<R, C>>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self {
-            col_piv_qr: self.col_piv_qr.clone(),
-            p: self.p.clone(),
-            diag: self.diag.clone(),
-        }
-    }
-}
-
 impl<T: ComplexField, R: DimMin<C>, C: Dim> ColPivQR<T, R, C>
 where
     DefaultAllocator: Allocator<T, R, C>
@@ -79,42 +63,37 @@ where
 {
     /// Computes the `ColPivQR` decomposition using householder reflections.
     pub fn new(mut matrix: OMatrix<T, R, C>) -> Self {
-        let (nrows, ncols) = matrix.data.shape();
+        let (nrows, ncols) = matrix.shape_generic();
         let min_nrows_ncols = nrows.min(ncols);
         let mut p = PermutationSequence::identity_generic(min_nrows_ncols);
 
-        let mut diag = Matrix::new_uninitialized_generic(min_nrows_ncols, Const::<1>);
-
         if min_nrows_ncols.value() == 0 {
-            // Safety: there's no (uninitialized) values.
-            unsafe {
-                return ColPivQR {
-                    col_piv_qr: matrix,
-                    p,
-                    diag: diag.assume_init(),
-                };
+            return ColPivQR {
+                col_piv_qr: matrix,
+                p,
+                diag: Matrix::zeros_generic(min_nrows_ncols, Const::<1>),
             };
         }
 
+        let mut diag = Matrix::uninit(min_nrows_ncols, Const::<1>);
+
         for i in 0..min_nrows_ncols.value() {
             let piv = matrix.slice_range(i.., i..).icamax_full();
             let col_piv = piv.1 + i;
             matrix.swap_columns(i, col_piv);
             p.append_permutation(i, col_piv);
 
-            // Safety: the pointer is valid for writes, aligned, and uninitialized.
-            unsafe {
-                householder::clear_column_unchecked(&mut matrix, diag[i].as_mut_ptr(), i, 0, None);
-            }
+            diag[i] =
+                MaybeUninit::new(householder::clear_column_unchecked(&mut matrix, i, 0, None));
         }
 
-        // Safety: all values have been initialized.
-        unsafe {
-            ColPivQR {
-                col_piv_qr: matrix,
-                p,
-                diag: diag.assume_init(),
-            }
+        // Safety: diag is now fully initialized.
+        let diag = unsafe { diag.assume_init() };
+
+        ColPivQR {
+            col_piv_qr: matrix,
+            p,
+            diag,
         }
     }
 
@@ -125,7 +104,7 @@ where
     where
         DefaultAllocator: Allocator<T, DimMinimum<R, C>, C>,
     {
-        let (nrows, ncols) = self.col_piv_qr.data.shape();
+        let (nrows, ncols) = self.col_piv_qr.shape_generic();
         let mut res = self
             .col_piv_qr
             .rows_generic(0, nrows.min(ncols))
@@ -142,7 +121,7 @@ where
     where
         DefaultAllocator: Reallocator<T, R, C, DimMinimum<R, C>, C>,
     {
-        let (nrows, ncols) = self.col_piv_qr.data.shape();
+        let (nrows, ncols) = self.col_piv_qr.shape_generic();
         let mut res = self
             .col_piv_qr
             .resize_generic(nrows.min(ncols), ncols, T::zero());
@@ -157,7 +136,7 @@ where
     where
         DefaultAllocator: Allocator<T, R, DimMinimum<R, C>>,
     {
-        let (nrows, ncols) = self.col_piv_qr.data.shape();
+        let (nrows, ncols) = self.col_piv_qr.shape_generic();
 
         // NOTE: we could build the identity matrix and call q_mul on it.
         // Instead we don't so that we take in account the matrix sparseness.
@@ -320,7 +299,7 @@ where
         );
 
         // TODO: is there a less naive method ?
-        let (nrows, ncols) = self.col_piv_qr.data.shape();
+        let (nrows, ncols) = self.col_piv_qr.shape_generic();
         let mut res = OMatrix::identity_generic(nrows, ncols);
 
         if self.solve_mut(&mut res) {
diff --git a/src/linalg/convolution.rs b/src/linalg/convolution.rs
index 36cea3a0..21a32dbc 100644
--- a/src/linalg/convolution.rs
+++ b/src/linalg/convolution.rs
@@ -38,7 +38,7 @@ impl<T: RealField, D1: Dim, S1: Storage<T, D1>> Vector<T, D1, S1> {
             .data
             .shape()
             .0
-            .add(kernel.data.shape().0)
+            .add(kernel.shape_generic().0)
             .sub(Const::<1>);
         let mut conv = OVector::zeros_generic(result_len, Const::<1>);
 
@@ -92,7 +92,7 @@ impl<T: RealField, D1: Dim, S1: Storage<T, D1>> Vector<T, D1, S1> {
             .shape()
             .0
             .add(Const::<1>)
-            .sub(kernel.data.shape().0);
+            .sub(kernel.shape_generic().0);
         let mut conv = OVector::zeros_generic(result_len, Const::<1>);
 
         for i in 0..(vec - ker + 1) {
@@ -126,7 +126,7 @@ impl<T: RealField, D1: Dim, S1: Storage<T, D1>> Vector<T, D1, S1> {
             panic!("convolve_same expects `self.len() >= kernel.len() > 0`, received {} and {} respectively.",vec,ker);
         }
 
-        let mut conv = OVector::zeros_generic(self.data.shape().0, Const::<1>);
+        let mut conv = OVector::zeros_generic(self.shape_generic().0, Const::<1>);
 
         for i in 0..vec {
             for j in 0..ker {
diff --git a/src/linalg/exp.rs b/src/linalg/exp.rs
index 76e2ddf5..e7751af2 100644
--- a/src/linalg/exp.rs
+++ b/src/linalg/exp.rs
@@ -4,12 +4,9 @@ use crate::{
     base::{
         allocator::Allocator,
         dimension::{Const, Dim, DimMin, DimMinimum},
-        storage::Storage,
         DefaultAllocator,
     },
-    convert,
-    storage::InnerOwned,
-    try_convert, ComplexField, OMatrix, RealField,
+    convert, try_convert, ComplexField, OMatrix, RealField,
 };
 
 use crate::num::Zero;
@@ -49,7 +46,7 @@ where
     DefaultAllocator: Allocator<T, D, D> + Allocator<(usize, usize), DimMinimum<D, D>>,
 {
     fn new(a: OMatrix<T, D, D>, use_exact_norm: bool) -> Self {
-        let (nrows, ncols) = a.data.shape();
+        let (nrows, ncols) = a.shape_generic();
         ExpmPadeHelper {
             use_exact_norm,
             ident: OMatrix::<T, D, D>::identity_generic(nrows, ncols),
@@ -350,7 +347,7 @@ where
     D: Dim,
     DefaultAllocator: Allocator<T, D, D> + Allocator<T, D>,
 {
-    let nrows = a.data.shape().0;
+    let nrows = a.shape_generic().0;
     let mut v = crate::OVector::<T, D>::repeat_generic(nrows, Const::<1>, convert(1.0));
     let m = a.transpose();
 
@@ -435,7 +432,6 @@ where
         + Allocator<T, D>
         + Allocator<T::RealField, D>
         + Allocator<T::RealField, D, D>,
-    InnerOwned<T, D, D>: Clone,
 {
     /// Computes exponential of this matrix
     #[must_use]
diff --git a/src/linalg/full_piv_lu.rs b/src/linalg/full_piv_lu.rs
index 71e0755e..20033c3c 100644
--- a/src/linalg/full_piv_lu.rs
+++ b/src/linalg/full_piv_lu.rs
@@ -1,5 +1,3 @@
-use std::fmt;
-
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Serialize};
 
@@ -29,7 +27,8 @@ use crate::linalg::PermutationSequence;
          OMatrix<T, R, C>: Deserialize<'de>,
          PermutationSequence<DimMinimum<R, C>>: Deserialize<'de>"))
 )]
-pub struct FullPivLU<T, R: DimMin<C>, C: Dim>
+#[derive(Clone, Debug)]
+pub struct FullPivLU<T: ComplexField, R: DimMin<C>, C: Dim>
 where
     DefaultAllocator: Allocator<T, R, C> + Allocator<(usize, usize), DimMinimum<R, C>>,
 {
@@ -41,41 +40,11 @@ where
 impl<T: ComplexField, R: DimMin<C>, C: Dim> Copy for FullPivLU<T, R, C>
 where
     DefaultAllocator: Allocator<T, R, C> + Allocator<(usize, usize), DimMinimum<R, C>>,
-    PermutationSequence<DimMinimum<R, C>>: Copy,
     OMatrix<T, R, C>: Copy,
+    PermutationSequence<DimMinimum<R, C>>: Copy,
 {
 }
 
-impl<T: ComplexField, R: DimMin<C>, C: Dim> Clone for FullPivLU<T, R, C>
-where
-    DefaultAllocator: Allocator<T, R, C> + Allocator<(usize, usize), DimMinimum<R, C>>,
-    PermutationSequence<DimMinimum<R, C>>: Clone,
-    OMatrix<T, R, C>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self {
-            lu: self.lu.clone(),
-            p: self.p.clone(),
-            q: self.q.clone(),
-        }
-    }
-}
-
-impl<T: ComplexField, R: DimMin<C>, C: Dim> fmt::Debug for FullPivLU<T, R, C>
-where
-    DefaultAllocator: Allocator<T, R, C> + Allocator<(usize, usize), DimMinimum<R, C>>,
-    PermutationSequence<DimMinimum<R, C>>: fmt::Debug,
-    OMatrix<T, R, C>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("FullPivLU")
-            .field("lu", &self.lu)
-            .field("p", &self.p)
-            .field("q", &self.q)
-            .finish()
-    }
-}
-
 impl<T: ComplexField, R: DimMin<C>, C: Dim> FullPivLU<T, R, C>
 where
     DefaultAllocator: Allocator<T, R, C> + Allocator<(usize, usize), DimMinimum<R, C>>,
@@ -84,7 +53,7 @@ where
     ///
     /// This effectively computes `P, L, U, Q` such that `P * matrix * Q = LU`.
     pub fn new(mut matrix: OMatrix<T, R, C>) -> Self {
-        let (nrows, ncols) = matrix.data.shape();
+        let (nrows, ncols) = matrix.shape_generic();
         let min_nrows_ncols = nrows.min(ncols);
 
         let mut p = PermutationSequence::identity_generic(min_nrows_ncols);
@@ -132,7 +101,7 @@ where
     where
         DefaultAllocator: Allocator<T, R, DimMinimum<R, C>>,
     {
-        let (nrows, ncols) = self.lu.data.shape();
+        let (nrows, ncols) = self.lu.shape_generic();
         let mut m = self.lu.columns_generic(0, nrows.min(ncols)).into_owned();
         m.fill_upper_triangle(T::zero(), 1);
         m.fill_diagonal(T::one());
@@ -146,7 +115,7 @@ where
     where
         DefaultAllocator: Allocator<T, DimMinimum<R, C>, C>,
     {
-        let (nrows, ncols) = self.lu.data.shape();
+        let (nrows, ncols) = self.lu.shape_generic();
         self.lu.rows_generic(0, nrows.min(ncols)).upper_triangle()
     }
 
@@ -253,7 +222,7 @@ where
             "FullPivLU inverse: unable to compute the inverse of a non-square matrix."
         );
 
-        let (nrows, ncols) = self.lu.data.shape();
+        let (nrows, ncols) = self.lu.shape_generic();
 
         let mut res = OMatrix::identity_generic(nrows, ncols);
         if self.solve_mut(&mut res) {
diff --git a/src/linalg/hessenberg.rs b/src/linalg/hessenberg.rs
index 3874bf77..1e266b16 100644
--- a/src/linalg/hessenberg.rs
+++ b/src/linalg/hessenberg.rs
@@ -1,17 +1,14 @@
-use std::fmt;
-use std::mem::MaybeUninit;
-
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Serialize};
 
 use crate::allocator::Allocator;
 use crate::base::{DefaultAllocator, OMatrix, OVector};
 use crate::dimension::{Const, DimDiff, DimSub, U1};
-use crate::storage::{InnerOwned, Storage};
-use crate::Matrix;
 use simba::scalar::ComplexField;
 
 use crate::linalg::householder;
+use crate::Matrix;
+use std::mem::MaybeUninit;
 
 /// Hessenberg decomposition of a general matrix.
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
@@ -29,6 +26,7 @@ use crate::linalg::householder;
          OMatrix<T, D, D>: Deserialize<'de>,
          OVector<T, DimDiff<D, U1>>: Deserialize<'de>"))
 )]
+#[derive(Clone, Debug)]
 pub struct Hessenberg<T: ComplexField, D: DimSub<U1>>
 where
     DefaultAllocator: Allocator<T, D, D> + Allocator<T, DimDiff<D, U1>>,
@@ -37,43 +35,13 @@ where
     subdiag: OVector<T, DimDiff<D, U1>>,
 }
 
-/*
 impl<T: ComplexField, D: DimSub<U1>> Copy for Hessenberg<T, D>
 where
     DefaultAllocator: Allocator<T, D, D> + Allocator<T, DimDiff<D, U1>>,
-    InnerOwned<T, D, D>: Copy,
-    InnerOwned<T, DimDiff<D, U1>>: Copy,
+    OMatrix<T, D, D>: Copy,
+    OVector<T, DimDiff<D, U1>>: Copy,
 {
 }
-*/
-
-impl<T: ComplexField, D: DimSub<U1>> Clone for Hessenberg<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D> + Allocator<T, DimDiff<D, U1>>,
-    InnerOwned<T, D, D>: Clone,
-    InnerOwned<T, DimDiff<D, U1>>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self {
-            hess: self.hess.clone(),
-            subdiag: self.subdiag.clone(),
-        }
-    }
-}
-
-impl<T: ComplexField, D: DimSub<U1>> fmt::Debug for Hessenberg<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D> + Allocator<T, DimDiff<D, U1>>,
-    InnerOwned<T, D, D>: fmt::Debug,
-    InnerOwned<T, DimDiff<D, U1>>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("Hessenberg")
-            .field("hess", &self.hess)
-            .field("subdiag", &self.subdiag)
-            .finish()
-    }
-}
 
 impl<T: ComplexField, D: DimSub<U1>> Hessenberg<T, D>
 where
@@ -81,7 +49,7 @@ where
 {
     /// Computes the Hessenberg decomposition using householder reflections.
     pub fn new(hess: OMatrix<T, D, D>) -> Self {
-        let mut work = OVector::new_uninitialized_generic(hess.data.shape().0, Const::<1>);
+        let mut work = Matrix::zeros_generic(hess.shape_generic().0, Const::<1>);
         Self::new_with_workspace(hess, &mut work)
     }
 
@@ -89,16 +57,13 @@ where
     ///
     /// The workspace containing `D` elements must be provided but its content does not have to be
     /// initialized.
-    pub fn new_with_workspace(
-        mut hess: OMatrix<T, D, D>,
-        work: &mut OVector<MaybeUninit<T>, D>,
-    ) -> Self {
+    pub fn new_with_workspace(mut hess: OMatrix<T, D, D>, work: &mut OVector<T, D>) -> Self {
         assert!(
             hess.is_square(),
             "Cannot compute the hessenberg decomposition of a non-square matrix."
         );
 
-        let dim = hess.data.shape().0;
+        let dim = hess.shape_generic().0;
 
         assert!(
             dim.value() != 0,
@@ -110,38 +75,27 @@ where
             "Hessenberg: invalid workspace size."
         );
 
-        let mut subdiag = Matrix::new_uninitialized_generic(dim.sub(Const::<1>), Const::<1>);
-
         if dim.value() == 0 {
-            // Safety: there's no (uninitialized) values.
-            unsafe {
-                return Self {
-                    hess,
-                    subdiag: subdiag.assume_init(),
-                };
-            }
+            return Hessenberg {
+                hess,
+                subdiag: Matrix::zeros_generic(dim.sub(Const::<1>), Const::<1>),
+            };
         }
 
+        let mut subdiag = Matrix::uninit(dim.sub(Const::<1>), Const::<1>);
+
         for ite in 0..dim.value() - 1 {
-            // Safety: the pointer is valid for writes, aligned, and uninitialized.
-            unsafe {
-                householder::clear_column_unchecked(
-                    &mut hess,
-                    subdiag[ite].as_mut_ptr(),
-                    ite,
-                    1,
-                    Some(work),
-                );
-            }
+            subdiag[ite] = MaybeUninit::new(householder::clear_column_unchecked(
+                &mut hess,
+                ite,
+                1,
+                Some(work),
+            ));
         }
 
-        // Safety: all values have been initialized.
-        unsafe {
-            Self {
-                hess,
-                subdiag: subdiag.assume_init(),
-            }
-        }
+        // Safety: subdiag is now fully initialized.
+        let subdiag = unsafe { subdiag.assume_init() };
+        Hessenberg { hess, subdiag }
     }
 
     /// Retrieves `(q, h)` with `q` the orthogonal matrix of this decomposition and `h` the
@@ -170,10 +124,7 @@ where
     /// This is less efficient than `.unpack_h()` as it allocates a new matrix.
     #[inline]
     #[must_use]
-    pub fn h(&self) -> OMatrix<T, D, D>
-    where
-        InnerOwned<T, D, D>: Clone,
-    {
+    pub fn h(&self) -> OMatrix<T, D, D> {
         let dim = self.hess.nrows();
         let mut res = self.hess.clone();
         res.fill_lower_triangle(T::zero(), 2);
diff --git a/src/linalg/householder.rs b/src/linalg/householder.rs
index 06a50d8e..6d20205d 100644
--- a/src/linalg/householder.rs
+++ b/src/linalg/householder.rs
@@ -1,11 +1,9 @@
 //! Construction of householder elementary reflections.
 
-use std::mem::MaybeUninit;
-
 use crate::allocator::Allocator;
 use crate::base::{DefaultAllocator, OMatrix, OVector, Unit, Vector};
 use crate::dimension::Dim;
-use crate::storage::{Storage, StorageMut};
+use crate::storage::StorageMut;
 use num::Zero;
 use simba::scalar::ComplexField;
 
@@ -46,29 +44,22 @@ pub fn reflection_axis_mut<T: ComplexField, D: Dim, S: StorageMut<T, D>>(
 /// Uses an householder reflection to zero out the `icol`-th column, starting with the `shift + 1`-th
 /// subdiagonal element.
 ///
-/// # Safety
-/// Behavior is undefined if any of the following conditions are violated:
-///
-/// - `diag_elt` must be valid for writes.
-/// - `diag_elt` must be properly aligned.
-///
-/// Furthermore, if `diag_elt` was previously initialized, this method will leak
-/// its data.
+/// Returns the signed norm of the column.
 #[doc(hidden)]
-pub unsafe fn clear_column_unchecked<T: ComplexField, R: Dim, C: Dim>(
+#[must_use]
+pub fn clear_column_unchecked<T: ComplexField, R: Dim, C: Dim>(
     matrix: &mut OMatrix<T, R, C>,
-    diag_elt: *mut T,
     icol: usize,
     shift: usize,
-    bilateral: Option<&mut OVector<MaybeUninit<T>, R>>,
-) where
+    bilateral: Option<&mut OVector<T, R>>,
+) -> T
+where
     DefaultAllocator: Allocator<T, R, C> + Allocator<T, R>,
 {
     let (mut left, mut right) = matrix.columns_range_pair_mut(icol, icol + 1..);
     let mut axis = left.rows_range_mut(icol + shift..);
 
     let (reflection_norm, not_zero) = reflection_axis_mut(&mut axis);
-    diag_elt.write(reflection_norm);
 
     if not_zero {
         let refl = Reflection::new(Unit::new_unchecked(axis), T::zero());
@@ -78,38 +69,32 @@ pub unsafe fn clear_column_unchecked<T: ComplexField, R: Dim, C: Dim>(
         }
         refl.reflect_with_sign(&mut right.rows_range_mut(icol + shift..), sign.conjugate());
     }
+
+    reflection_norm
 }
 
 /// Uses an householder reflection to zero out the `irow`-th row, ending before the `shift + 1`-th
 /// superdiagonal element.
 ///
-/// # Safety
-/// Behavior is undefined if any of the following conditions are violated:
-///
-/// - `diag_elt` must be valid for writes.
-/// - `diag_elt` must be properly aligned.
-///
-/// Furthermore, if `diag_elt` was previously initialized, this method will leak
-/// its data.
+/// Returns the signed norm of the column.
 #[doc(hidden)]
-pub unsafe fn clear_row_unchecked<T: ComplexField, R: Dim, C: Dim>(
+#[must_use]
+pub fn clear_row_unchecked<T: ComplexField, R: Dim, C: Dim>(
     matrix: &mut OMatrix<T, R, C>,
-    diag_elt: *mut T,
-    axis_packed: &mut OVector<MaybeUninit<T>, C>,
-    work: &mut OVector<MaybeUninit<T>, R>,
+    axis_packed: &mut OVector<T, C>,
+    work: &mut OVector<T, R>,
     irow: usize,
     shift: usize,
-) where
+) -> T
+where
     DefaultAllocator: Allocator<T, R, C> + Allocator<T, R> + Allocator<T, C>,
 {
     let (mut top, mut bottom) = matrix.rows_range_pair_mut(irow, irow + 1..);
     let mut axis = axis_packed.rows_range_mut(irow + shift..);
-    axis.tr_copy_init_from(&top.columns_range(irow + shift..));
-    let mut axis = axis.assume_init_mut();
+    axis.tr_copy_from(&top.columns_range(irow + shift..));
 
     let (reflection_norm, not_zero) = reflection_axis_mut(&mut axis);
     axis.conjugate_mut(); // So that reflect_rows actually cancels the first row.
-    diag_elt.write(reflection_norm);
 
     if not_zero {
         let refl = Reflection::new(Unit::new_unchecked(axis), T::zero());
@@ -123,6 +108,8 @@ pub unsafe fn clear_row_unchecked<T: ComplexField, R: Dim, C: Dim>(
     } else {
         top.columns_range_mut(irow + shift..).tr_copy_from(&axis);
     }
+
+    reflection_norm
 }
 
 /// Computes the orthogonal transformation described by the elementary reflector axii stored on
@@ -134,7 +121,7 @@ where
     DefaultAllocator: Allocator<T, D, D>,
 {
     assert!(m.is_square());
-    let dim = m.data.shape().0;
+    let dim = m.shape_generic().0;
 
     // NOTE: we could build the identity matrix and call p_mult on it.
     // Instead we don't so that we take in account the matrix sparseness.
diff --git a/src/linalg/lu.rs b/src/linalg/lu.rs
index 6fc0d9fa..0e3be559 100644
--- a/src/linalg/lu.rs
+++ b/src/linalg/lu.rs
@@ -1,6 +1,3 @@
-use std::fmt;
-use std::mem;
-
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Serialize};
 
@@ -8,8 +5,9 @@ use crate::allocator::{Allocator, Reallocator};
 use crate::base::{DefaultAllocator, Matrix, OMatrix, Scalar};
 use crate::constraint::{SameNumberOfRows, ShapeConstraint};
 use crate::dimension::{Dim, DimMin, DimMinimum};
-use crate::storage::{InnerOwned, Storage, StorageMut};
+use crate::storage::{Storage, StorageMut};
 use simba::scalar::{ComplexField, Field};
+use std::mem;
 
 use crate::linalg::PermutationSequence;
 
@@ -29,7 +27,8 @@ use crate::linalg::PermutationSequence;
          OMatrix<T, R, C>: Deserialize<'de>,
          PermutationSequence<DimMinimum<R, C>>: Deserialize<'de>"))
 )]
-pub struct LU<T, R: DimMin<C>, C: Dim>
+#[derive(Clone, Debug)]
+pub struct LU<T: ComplexField, R: DimMin<C>, C: Dim>
 where
     DefaultAllocator: Allocator<T, R, C> + Allocator<(usize, usize), DimMinimum<R, C>>,
 {
@@ -37,43 +36,13 @@ where
     p: PermutationSequence<DimMinimum<R, C>>,
 }
 
-/*
-impl<T: Copy, R: DimMin<C>, C: Dim> Copy for LU<T, R, C>
+impl<T: ComplexField, R: DimMin<C>, C: Dim> Copy for LU<T, R, C>
 where
     DefaultAllocator: Allocator<T, R, C> + Allocator<(usize, usize), DimMinimum<R, C>>,
+    OMatrix<T, R, C>: Copy,
     PermutationSequence<DimMinimum<R, C>>: Copy,
-    InnerOwned<T, R, C>: Copy,
 {
 }
-*/
-
-impl<T: Clone, R: DimMin<C>, C: Dim> Clone for LU<T, R, C>
-where
-    DefaultAllocator: Allocator<T, R, C> + Allocator<(usize, usize), DimMinimum<R, C>>,
-    PermutationSequence<DimMinimum<R, C>>: Clone,
-    InnerOwned<T, R, C>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self {
-            lu: self.lu.clone(),
-            p: self.p.clone(),
-        }
-    }
-}
-
-impl<T: ComplexField, R: DimMin<C>, C: Dim> fmt::Debug for LU<T, R, C>
-where
-    DefaultAllocator: Allocator<T, R, C> + Allocator<(usize, usize), DimMinimum<R, C>>,
-    PermutationSequence<DimMinimum<R, C>>: fmt::Debug,
-    InnerOwned<T, R, C>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("LU")
-            .field("lu", &self.lu)
-            .field("p", &self.p)
-            .finish()
-    }
-}
 
 /// Performs a LU decomposition to overwrite `out` with the inverse of `matrix`.
 ///
@@ -121,7 +90,7 @@ where
 {
     /// Computes the LU decomposition with partial (row) pivoting of `matrix`.
     pub fn new(mut matrix: OMatrix<T, R, C>) -> Self {
-        let (nrows, ncols) = matrix.data.shape();
+        let (nrows, ncols) = matrix.shape_generic();
         let min_nrows_ncols = nrows.min(ncols);
 
         let mut p = PermutationSequence::identity_generic(min_nrows_ncols);
@@ -163,7 +132,7 @@ where
     where
         DefaultAllocator: Allocator<T, R, DimMinimum<R, C>>,
     {
-        let (nrows, ncols) = self.lu.data.shape();
+        let (nrows, ncols) = self.lu.shape_generic();
         let mut m = self.lu.columns_generic(0, nrows.min(ncols)).into_owned();
         m.fill_upper_triangle(T::zero(), 1);
         m.fill_diagonal(T::one());
@@ -180,7 +149,7 @@ where
     where
         DefaultAllocator: Reallocator<T, R, C, R, DimMinimum<R, C>>,
     {
-        let (nrows, ncols) = self.lu.data.shape();
+        let (nrows, ncols) = self.lu.shape_generic();
         let mut m = self.lu.resize_generic(nrows, nrows.min(ncols), T::zero());
         m.fill_upper_triangle(T::zero(), 1);
         m.fill_diagonal(T::one());
@@ -193,7 +162,7 @@ where
     where
         DefaultAllocator: Reallocator<T, R, C, R, DimMinimum<R, C>>,
     {
-        let (nrows, ncols) = self.lu.data.shape();
+        let (nrows, ncols) = self.lu.shape_generic();
         let mut m = self.lu.resize_generic(nrows, nrows.min(ncols), T::zero());
         m.fill_upper_triangle(T::zero(), 1);
         m.fill_diagonal(T::one());
@@ -207,7 +176,7 @@ where
     where
         DefaultAllocator: Allocator<T, DimMinimum<R, C>, C>,
     {
-        let (nrows, ncols) = self.lu.data.shape();
+        let (nrows, ncols) = self.lu.shape_generic();
         self.lu.rows_generic(0, nrows.min(ncols)).upper_triangle()
     }
 
@@ -299,7 +268,7 @@ where
             "LU inverse: unable to compute the inverse of a non-square matrix."
         );
 
-        let (nrows, ncols) = self.lu.data.shape();
+        let (nrows, ncols) = self.lu.shape_generic();
         let mut res = OMatrix::identity_generic(nrows, ncols);
         if self.try_inverse_to(&mut res) {
             Some(res)
diff --git a/src/linalg/permutation_sequence.rs b/src/linalg/permutation_sequence.rs
index 14ff718d..f4521988 100644
--- a/src/linalg/permutation_sequence.rs
+++ b/src/linalg/permutation_sequence.rs
@@ -1,6 +1,3 @@
-use std::fmt;
-use std::mem::MaybeUninit;
-
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Serialize};
 
@@ -11,10 +8,8 @@ use crate::allocator::Allocator;
 use crate::base::{DefaultAllocator, Matrix, OVector, Scalar};
 #[cfg(any(feature = "std", feature = "alloc"))]
 use crate::dimension::Dynamic;
-use crate::dimension::{Dim, DimName};
-use crate::iter::MatrixIter;
-use crate::storage::{InnerOwned, StorageMut};
-use crate::{Const, U1};
+use crate::dimension::{Const, Dim, DimName};
+use crate::storage::StorageMut;
 
 /// A sequence of row or column permutations.
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
@@ -28,47 +23,22 @@ use crate::{Const, U1};
     serde(bound(deserialize = "DefaultAllocator: Allocator<(usize, usize), D>,
          OVector<(usize, usize), D>: Deserialize<'de>"))
 )]
+#[derive(Clone, Debug)]
 pub struct PermutationSequence<D: Dim>
 where
     DefaultAllocator: Allocator<(usize, usize), D>,
 {
     len: usize,
-    ipiv: OVector<MaybeUninit<(usize, usize)>, D>,
+    ipiv: OVector<(usize, usize), D>,
 }
 
 impl<D: Dim> Copy for PermutationSequence<D>
 where
     DefaultAllocator: Allocator<(usize, usize), D>,
-    OVector<MaybeUninit<(usize, usize)>, D>: Copy,
+    OVector<(usize, usize), D>: Copy,
 {
 }
 
-impl<D: Dim> Clone for PermutationSequence<D>
-where
-    DefaultAllocator: Allocator<(usize, usize), D>,
-    OVector<MaybeUninit<(usize, usize)>, D>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self {
-            len: self.len,
-            ipiv: self.ipiv.clone(),
-        }
-    }
-}
-
-impl<D: Dim> fmt::Debug for PermutationSequence<D>
-where
-    DefaultAllocator: Allocator<(usize, usize), D>,
-    OVector<MaybeUninit<(usize, usize)>, D>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("PermutationSequence")
-            .field("len", &self.len)
-            .field("ipiv", &self.ipiv)
-            .finish()
-    }
-}
-
 impl<D: DimName> PermutationSequence<D>
 where
     DefaultAllocator: Allocator<(usize, usize), D>,
@@ -101,7 +71,9 @@ where
     pub fn identity_generic(dim: D) -> Self {
         Self {
             len: 0,
-            ipiv: OVector::new_uninitialized_generic(dim, Const::<1>),
+            // TODO: using a uninitialized matrix would save some computation, but
+            //       that loos difficult to setup with MaybeUninit.
+            ipiv: Matrix::repeat_generic(dim, Const::<1>, (0, 0)),
         }
     }
 
@@ -114,7 +86,7 @@ where
                 self.len < self.ipiv.len(),
                 "Maximum number of permutations exceeded."
             );
-            self.ipiv[self.len] = MaybeUninit::new((i, i2));
+            self.ipiv[self.len] = (i, i2);
             self.len += 1;
         }
     }
@@ -125,8 +97,8 @@ where
     where
         S2: StorageMut<T, R2, C2>,
     {
-        for perm in self.iter() {
-            rhs.swap_rows(perm.0, perm.1)
+        for i in self.ipiv.rows_range(..self.len).iter() {
+            rhs.swap_rows(i.0, i.1)
         }
     }
 
@@ -136,8 +108,8 @@ where
     where
         S2: StorageMut<T, R2, C2>,
     {
-        for perm in self.iter().rev() {
-            let (i1, i2) = perm;
+        for i in 0..self.len {
+            let (i1, i2) = self.ipiv[self.len - i - 1];
             rhs.swap_rows(i1, i2)
         }
     }
@@ -148,8 +120,8 @@ where
     where
         S2: StorageMut<T, R2, C2>,
     {
-        for perm in self.iter() {
-            rhs.swap_columns(perm.0, perm.1)
+        for i in self.ipiv.rows_range(..self.len).iter() {
+            rhs.swap_columns(i.0, i.1)
         }
     }
 
@@ -161,8 +133,8 @@ where
     ) where
         S2: StorageMut<T, R2, C2>,
     {
-        for perm in self.iter().rev() {
-            let (i1, i2) = perm;
+        for i in 0..self.len {
+            let (i1, i2) = self.ipiv[self.len - i - 1];
             rhs.swap_columns(i1, i2)
         }
     }
@@ -189,27 +161,4 @@ where
             -T::one()
         }
     }
-
-    /// Iterates over the permutations that have been initialized.
-    pub fn iter(
-        &self,
-    ) -> std::iter::Map<
-        std::iter::Copied<
-            std::iter::Take<
-                MatrixIter<
-                    MaybeUninit<(usize, usize)>,
-                    D,
-                    U1,
-                    InnerOwned<MaybeUninit<(usize, usize)>, D, U1>,
-                >,
-            >,
-        >,
-        impl FnMut(MaybeUninit<(usize, usize)>) -> (usize, usize),
-    > {
-        self.ipiv
-            .iter()
-            .take(self.len)
-            .copied()
-            .map(|e| unsafe { e.assume_init() })
-    }
 }
diff --git a/src/linalg/pow.rs b/src/linalg/pow.rs
index 000dc8b8..df513643 100644
--- a/src/linalg/pow.rs
+++ b/src/linalg/pow.rs
@@ -40,31 +40,19 @@ where
 
         // We use the buffer to hold the result of multiplier ^ 2, thus avoiding
         // extra allocations.
-        let (nrows, ncols) = self.data.shape();
         let mut multiplier = self.clone_owned();
-        let mut buf = Matrix::new_uninitialized_generic(nrows, ncols);
+        let mut buf = self.clone_owned();
 
         // Exponentiation by squares.
         loop {
             if e % two == one {
-                let init_buf = self.mul_to(&multiplier, &mut buf);
-                self.copy_from(&init_buf);
-
-                // Safety: `mul_to` leaves `buf` completely initialized.
-                unsafe {
-                    buf.reinitialize();
-                }
+                self.mul_to(&multiplier, &mut buf);
+                self.copy_from(&buf);
             }
 
             e /= two;
-
-            let init_buf = multiplier.mul_to(&multiplier, &mut buf);
-            multiplier.copy_from(&init_buf);
-
-            // Safety: `mul_to` leaves `buf` completely initialized.
-            unsafe {
-                buf.reinitialize();
-            }
+            multiplier.mul_to(&multiplier, &mut buf);
+            multiplier.copy_from(&buf);
 
             if e == zero {
                 return true;
diff --git a/src/linalg/qr.rs b/src/linalg/qr.rs
index e4a4911b..e2f8e0c3 100644
--- a/src/linalg/qr.rs
+++ b/src/linalg/qr.rs
@@ -1,5 +1,3 @@
-use std::fmt;
-
 use num::Zero;
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Serialize};
@@ -8,11 +6,12 @@ use crate::allocator::{Allocator, Reallocator};
 use crate::base::{DefaultAllocator, Matrix, OMatrix, OVector, Unit};
 use crate::constraint::{SameNumberOfRows, ShapeConstraint};
 use crate::dimension::{Const, Dim, DimMin, DimMinimum};
-use crate::storage::{InnerOwned, Storage, StorageMut};
+use crate::storage::{Storage, StorageMut};
 use simba::scalar::ComplexField;
 
 use crate::geometry::Reflection;
 use crate::linalg::householder;
+use std::mem::MaybeUninit;
 
 /// The QR decomposition of a general matrix.
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
@@ -30,8 +29,8 @@ use crate::linalg::householder;
          OMatrix<T, R, C>: Deserialize<'de>,
          OVector<T, DimMinimum<R, C>>: Deserialize<'de>"))
 )]
-
-pub struct QR<T, R: DimMin<C>, C: Dim>
+#[derive(Clone, Debug)]
+pub struct QR<T: ComplexField, R: DimMin<C>, C: Dim>
 where
     DefaultAllocator: Allocator<T, R, C> + Allocator<T, DimMinimum<R, C>>,
 {
@@ -39,43 +38,13 @@ where
     diag: OVector<T, DimMinimum<R, C>>,
 }
 
-/*
-impl<T: Copy, R: DimMin<C>, C: Dim> Copy for QR<T, R, C>
+impl<T: ComplexField, R: DimMin<C>, C: Dim> Copy for QR<T, R, C>
 where
     DefaultAllocator: Allocator<T, R, C> + Allocator<T, DimMinimum<R, C>>,
-    InnerOwned<T, R, C>: Copy,
-    InnerOwned<T, DimMinimum<R, C>>: Copy,
+    OMatrix<T, R, C>: Copy,
+    OVector<T, DimMinimum<R, C>>: Copy,
 {
 }
-*/
-
-impl<T: Clone, R: DimMin<C>, C: Dim> Clone for QR<T, R, C>
-where
-    DefaultAllocator: Allocator<T, R, C> + Allocator<T, DimMinimum<R, C>>,
-    InnerOwned<T, R, C>: Clone,
-    InnerOwned<T, DimMinimum<R, C>>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self {
-            qr: self.qr.clone(),
-            diag: self.diag.clone(),
-        }
-    }
-}
-
-impl<T: fmt::Debug, R: DimMin<C>, C: Dim> fmt::Debug for QR<T, R, C>
-where
-    DefaultAllocator: Allocator<T, R, C> + Allocator<T, DimMinimum<R, C>>,
-    InnerOwned<T, R, C>: fmt::Debug,
-    InnerOwned<T, DimMinimum<R, C>>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("QR")
-            .field("qr", &self.qr)
-            .field("diag", &self.diag)
-            .finish()
-    }
-}
 
 impl<T: ComplexField, R: DimMin<C>, C: Dim> QR<T, R, C>
 where
@@ -83,32 +52,26 @@ where
 {
     /// Computes the QR decomposition using householder reflections.
     pub fn new(mut matrix: OMatrix<T, R, C>) -> Self {
-        let (nrows, ncols) = matrix.data.shape();
+        let (nrows, ncols) = matrix.shape_generic();
         let min_nrows_ncols = nrows.min(ncols);
 
-        let mut diag = Matrix::new_uninitialized_generic(min_nrows_ncols, Const::<1>);
-
         if min_nrows_ncols.value() == 0 {
-            return Self {
+            return QR {
                 qr: matrix,
-                diag: unsafe { diag.assume_init() },
+                diag: Matrix::zeros_generic(min_nrows_ncols, Const::<1>),
             };
         }
 
+        let mut diag = Matrix::uninit(min_nrows_ncols, Const::<1>);
+
         for i in 0..min_nrows_ncols.value() {
-            // Safety: the pointer is valid for writes, aligned, and uninitialized.
-            unsafe {
-                householder::clear_column_unchecked(&mut matrix, diag[i].as_mut_ptr(), i, 0, None);
-            }
+            diag[i] =
+                MaybeUninit::new(householder::clear_column_unchecked(&mut matrix, i, 0, None));
         }
 
-        // Safety: all values have been initialized.
-        unsafe {
-            Self {
-                qr: matrix,
-                diag: diag.assume_init(),
-            }
-        }
+        // Safety: diag is now fully initialized.
+        let diag = unsafe { diag.assume_init() };
+        QR { qr: matrix, diag }
     }
 
     /// Retrieves the upper trapezoidal submatrix `R` of this decomposition.
@@ -118,7 +81,7 @@ where
     where
         DefaultAllocator: Allocator<T, DimMinimum<R, C>, C>,
     {
-        let (nrows, ncols) = self.qr.data.shape();
+        let (nrows, ncols) = self.qr.shape_generic();
         let mut res = self.qr.rows_generic(0, nrows.min(ncols)).upper_triangle();
         res.set_partial_diagonal(self.diag.iter().map(|e| T::from_real(e.modulus())));
         res
@@ -132,7 +95,7 @@ where
     where
         DefaultAllocator: Reallocator<T, R, C, DimMinimum<R, C>, C>,
     {
-        let (nrows, ncols) = self.qr.data.shape();
+        let (nrows, ncols) = self.qr.shape_generic();
         let mut res = self.qr.resize_generic(nrows.min(ncols), ncols, T::zero());
         res.fill_lower_triangle(T::zero(), 1);
         res.set_partial_diagonal(self.diag.iter().map(|e| T::from_real(e.modulus())));
@@ -145,7 +108,7 @@ where
     where
         DefaultAllocator: Allocator<T, R, DimMinimum<R, C>>,
     {
-        let (nrows, ncols) = self.qr.data.shape();
+        let (nrows, ncols) = self.qr.shape_generic();
 
         // NOTE: we could build the identity matrix and call q_mul on it.
         // Instead we don't so that we take in account the matrix sparseness.
@@ -297,7 +260,7 @@ where
         );
 
         // TODO: is there a less naive method ?
-        let (nrows, ncols) = self.qr.data.shape();
+        let (nrows, ncols) = self.qr.shape_generic();
         let mut res = OMatrix::identity_generic(nrows, ncols);
 
         if self.solve_mut(&mut res) {
diff --git a/src/linalg/schur.rs b/src/linalg/schur.rs
index d4ee2446..953e9953 100644
--- a/src/linalg/schur.rs
+++ b/src/linalg/schur.rs
@@ -1,25 +1,23 @@
 #![allow(clippy::suspicious_operation_groupings)]
-use std::cmp;
-use std::mem::MaybeUninit;
-
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Serialize};
 
 use approx::AbsDiffEq;
 use num_complex::Complex as NumComplex;
 use simba::scalar::{ComplexField, RealField};
+use std::cmp;
 
 use crate::allocator::Allocator;
-use crate::base::dimension::{Const, Dim, DimDiff,  DimSub, Dynamic, U1, U2};
-use crate::base::storage::{InnerOwned, Storage};
-use crate::base::{
-    DefaultAllocator, OMatrix, OVector,  SquareMatrix, Unit, Vector2, Vector3,
-};
+use crate::base::dimension::{Const, Dim, DimDiff, DimSub, Dynamic, U1, U2};
+use crate::base::storage::Storage;
+use crate::base::{DefaultAllocator, OMatrix, OVector, SquareMatrix, Unit, Vector2, Vector3};
 
 use crate::geometry::Reflection;
 use crate::linalg::givens::GivensRotation;
 use crate::linalg::householder;
 use crate::linalg::Hessenberg;
+use crate::{Matrix, UninitVector};
+use std::mem::MaybeUninit;
 
 /// Schur decomposition of a square matrix.
 ///
@@ -36,7 +34,7 @@ use crate::linalg::Hessenberg;
          OMatrix<T, D, D>: Deserialize<'de>"))
 )]
 #[derive(Clone, Debug)]
-pub struct Schur<T, D: Dim>
+pub struct Schur<T: ComplexField, D: Dim>
 where
     DefaultAllocator: Allocator<T, D, D>,
 {
@@ -44,10 +42,10 @@ where
     t: OMatrix<T, D, D>,
 }
 
-impl<T: Copy, D: Dim> Copy for Schur<T, D>
+impl<T: ComplexField, D: Dim> Copy for Schur<T, D>
 where
     DefaultAllocator: Allocator<T, D, D>,
-    InnerOwned<T, D, D>: Copy,
+    OMatrix<T, D, D>: Copy,
 {
 }
 
@@ -76,7 +74,7 @@ where
     /// number of iteration is exceeded, `None` is returned. If `niter == 0`, then the algorithm
     /// continues indefinitely until convergence.
     pub fn try_new(m: OMatrix<T, D, D>, eps: T::RealField, max_niter: usize) -> Option<Self> {
-        let mut work = OVector::new_uninitialized_generic(m.data.shape().0, Const::<1>);
+        let mut work = Matrix::zeros_generic(m.shape_generic().0, Const::<1>);
 
         Self::do_decompose(m, &mut work, eps, max_niter, true)
             .map(|(q, t)| Schur { q: q.unwrap(), t })
@@ -84,7 +82,7 @@ where
 
     fn do_decompose(
         mut m: OMatrix<T, D, D>,
-        work: &mut OVector<MaybeUninit<T>, D>,
+        work: &mut OVector<T, D>,
         eps: T::RealField,
         max_niter: usize,
         compute_q: bool,
@@ -94,7 +92,7 @@ where
             "Unable to compute the eigenvectors and eigenvalues of a non-square matrix."
         );
 
-        let dim = m.data.shape().0;
+        let dim = m.shape_generic().0;
 
         // Specialization would make this easier.
         if dim.value() == 0 {
@@ -273,9 +271,7 @@ where
     }
 
     /// Computes the eigenvalues of the decomposed matrix.
-    fn do_eigenvalues(t: &OMatrix<T, D, D>, out: &mut OVector<MaybeUninit<T>, D>) -> bool {
-        // TODO: check dropping stuff.
-
+    fn do_eigenvalues(t: &OMatrix<T, D, D>, out: &mut OVector<T, D>) -> bool {
         let dim = t.nrows();
         let mut m = 0;
 
@@ -283,7 +279,7 @@ where
             let n = m + 1;
 
             if t[(n, m)].is_zero() {
-                out[m] = MaybeUninit::new(t[(m, m)]);
+                out[m] = t[(m, m)];
                 m += 1;
             } else {
                 // Complex eigenvalue.
@@ -292,22 +288,18 @@ where
         }
 
         if m == dim - 1 {
-            out[m] = MaybeUninit::new(t[(m, m)]);
+            out[m] = t[(m, m)];
         }
 
         true
     }
 
     /// Computes the complex eigenvalues of the decomposed matrix.
-    fn do_complex_eigenvalues(
-        t: &OMatrix<T, D, D>,
-        out: &mut OVector<MaybeUninit<NumComplex<T>>, D>,
-    ) where
+    fn do_complex_eigenvalues(t: &OMatrix<T, D, D>, out: &mut UninitVector<NumComplex<T>, D>)
+    where
         T: RealField,
         DefaultAllocator: Allocator<NumComplex<T>, D>,
     {
-        // TODO: check for dropping behavior.
-
         let dim = t.nrows();
         let mut m = 0;
 
@@ -397,9 +389,9 @@ where
     /// Return `None` if some eigenvalues are complex.
     #[must_use]
     pub fn eigenvalues(&self) -> Option<OVector<T, D>> {
-        let mut out = OVector::new_uninitialized_generic(self.t.data.shape().0, Const::<1>);
+        let mut out = Matrix::zeros_generic(self.t.shape_generic().0, Const::<1>);
         if Self::do_eigenvalues(&self.t, &mut out) {
-            Some(unsafe { out.assume_init() })
+            Some(out)
         } else {
             None
         }
@@ -412,8 +404,9 @@ where
         T: RealField,
         DefaultAllocator: Allocator<NumComplex<T>, D>,
     {
-        let mut out = OVector::new_uninitialized_generic(self.t.data.shape().0, Const::<1>);
+        let mut out = Matrix::uninit(self.t.shape_generic().0, Const::<1>);
         Self::do_complex_eigenvalues(&self.t, &mut out);
+        // Safety: out has been fully initialized by do_complex_eigenvalues.
         unsafe { out.assume_init() }
     }
 }
@@ -425,7 +418,7 @@ fn decompose_2x2<T: ComplexField, D: Dim>(
 where
     DefaultAllocator: Allocator<T, D, D>,
 {
-    let dim = m.data.shape().0;
+    let dim = m.shape_generic().0;
     let mut q = None;
     match compute_2x2_basis(&m.fixed_slice::<2, 2>(0, 0)) {
         Some(rot) => {
@@ -519,14 +512,12 @@ where
     /// Computes the eigenvalues of this matrix.
     #[must_use]
     pub fn eigenvalues(&self) -> Option<OVector<T, D>> {
-        // TODO: check drop stuff.
-
         assert!(
             self.is_square(),
             "Unable to compute eigenvalues of a non-square matrix."
         );
 
-        let mut work = OVector::new_uninitialized_generic(self.data.shape().0, Const::<1>);
+        let mut work = Matrix::zeros_generic(self.shape_generic().0, Const::<1>);
 
         // Special case for 2x2 matrices.
         if self.nrows() == 2 {
@@ -535,9 +526,9 @@ where
             let me = self.fixed_slice::<2, 2>(0, 0);
             return match compute_2x2_eigvals(&me) {
                 Some((a, b)) => {
-                    work[0] = MaybeUninit::new(a);
-                    work[1] = MaybeUninit::new(b);
-                    Some(unsafe { work.assume_init() })
+                    work[0] = a;
+                    work[1] = b;
+                    Some(work)
                 }
                 None => None,
             };
@@ -552,8 +543,9 @@ where
             false,
         )
         .unwrap();
+
         if Schur::do_eigenvalues(&schur.1, &mut work) {
-            Some(unsafe { work.assume_init() })
+            Some(work)
         } else {
             None
         }
@@ -567,8 +559,8 @@ where
         T: RealField,
         DefaultAllocator: Allocator<NumComplex<T>, D>,
     {
-        let dim = self.data.shape().0;
-        let mut work = OVector::new_uninitialized_generic(dim, Const::<1>);
+        let dim = self.shape_generic().0;
+        let mut work = Matrix::zeros_generic(dim, Const::<1>);
 
         let schur = Schur::do_decompose(
             self.clone_owned(),
@@ -578,8 +570,9 @@ where
             false,
         )
         .unwrap();
-        let mut eig = OVector::new_uninitialized_generic(dim, Const::<1>);
+        let mut eig = Matrix::uninit(dim, Const::<1>);
         Schur::do_complex_eigenvalues(&schur.1, &mut eig);
+        // Safety: eig has been fully initialized by do_complex_eigenvalues.
         unsafe { eig.assume_init() }
     }
 }
diff --git a/src/linalg/svd.rs b/src/linalg/svd.rs
index 355d1569..0b50fc9b 100644
--- a/src/linalg/svd.rs
+++ b/src/linalg/svd.rs
@@ -7,8 +7,8 @@ use num::{One, Zero};
 use crate::allocator::Allocator;
 use crate::base::{DefaultAllocator, Matrix, Matrix2x3, OMatrix, OVector, Vector2};
 use crate::constraint::{SameNumberOfRows, ShapeConstraint};
-use crate::dimension::{Dim, DimDiff, DimMin, DimMinimum, DimName, DimSub, U1};
-use crate::storage::{InnerOwned, Storage};
+use crate::dimension::{Dim, DimDiff, DimMin, DimMinimum, DimSub, U1};
+use crate::storage::Storage;
 use simba::scalar::{ComplexField, RealField};
 
 use crate::linalg::givens::GivensRotation;
@@ -54,14 +54,14 @@ where
     pub singular_values: OVector<T::RealField, DimMinimum<R, C>>,
 }
 
-impl<T: ComplexField + Copy, R: DimName + DimMin<C>, C: DimName> Copy for SVD<T, R, C>
+impl<T: ComplexField, R: DimMin<C>, C: Dim> Copy for SVD<T, R, C>
 where
     DefaultAllocator: Allocator<T, DimMinimum<R, C>, C>
         + Allocator<T, R, DimMinimum<R, C>>
         + Allocator<T::RealField, DimMinimum<R, C>>,
-    InnerOwned<T, R, DimMinimum<R, C>>: Copy,
-    InnerOwned<T, DimMinimum<R, C>, C>: Copy,
-    InnerOwned<T::RealField, DimMinimum<R, C>>: Copy,
+    OMatrix<T, R, DimMinimum<R, C>>: Copy,
+    OMatrix<T, DimMinimum<R, C>, C>: Copy,
+    OVector<T::RealField, DimMinimum<R, C>>: Copy,
 {
 }
 
@@ -111,7 +111,7 @@ where
             !matrix.is_empty(),
             "Cannot compute the SVD of an empty matrix."
         );
-        let (nrows, ncols) = matrix.data.shape();
+        let (nrows, ncols) = matrix.shape_generic();
         let min_nrows_ncols = nrows.min(ncols);
         let dim = min_nrows_ncols.value();
 
diff --git a/src/linalg/symmetric_eigen.rs b/src/linalg/symmetric_eigen.rs
index df32cdac..5ac6d5da 100644
--- a/src/linalg/symmetric_eigen.rs
+++ b/src/linalg/symmetric_eigen.rs
@@ -1,5 +1,3 @@
-use std::fmt;
-
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Serialize};
 
@@ -8,8 +6,8 @@ use num::Zero;
 
 use crate::allocator::Allocator;
 use crate::base::{DefaultAllocator, Matrix2, OMatrix, OVector, SquareMatrix, Vector2};
-use crate::dimension::{Dim, DimDiff, DimName, DimSub, U1};
-use crate::storage::{InnerOwned, Storage};
+use crate::dimension::{Dim, DimDiff, DimSub, U1};
+use crate::storage::Storage;
 use simba::scalar::ComplexField;
 
 use crate::linalg::givens::GivensRotation;
@@ -31,6 +29,7 @@ use crate::linalg::SymmetricTridiagonal;
          OVector<T::RealField, D>: Deserialize<'de>,
          OMatrix<T, D, D>: Deserialize<'de>"))
 )]
+#[derive(Clone, Debug)]
 pub struct SymmetricEigen<T: ComplexField, D: Dim>
 where
     DefaultAllocator: Allocator<T, D, D> + Allocator<T::RealField, D>,
@@ -42,42 +41,14 @@ where
     pub eigenvalues: OVector<T::RealField, D>,
 }
 
-impl<T: ComplexField, D: DimName> Copy for SymmetricEigen<T, D>
+impl<T: ComplexField, D: Dim> Copy for SymmetricEigen<T, D>
 where
     DefaultAllocator: Allocator<T, D, D> + Allocator<T::RealField, D>,
-    InnerOwned<T, D, D>: Copy,
-    InnerOwned<T::RealField, D>: Copy,
+    OMatrix<T, D, D>: Copy,
+    OVector<T::RealField, D>: Copy,
 {
 }
 
-impl<T: ComplexField, D: Dim> Clone for SymmetricEigen<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D> + Allocator<T::RealField, D>,
-    InnerOwned<T, D, D>: Clone,
-    InnerOwned<T::RealField, D>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self {
-            eigenvectors: self.eigenvectors.clone(),
-            eigenvalues: self.eigenvalues.clone(),
-        }
-    }
-}
-
-impl<T: ComplexField, D: Dim> fmt::Debug for SymmetricEigen<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D> + Allocator<T::RealField, D>,
-    InnerOwned<T, D, D>: fmt::Debug,
-    InnerOwned<T::RealField, D>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("SymmetricEigen")
-            .field("eigenvectors", &self.eigenvectors)
-            .field("eigenvalues", &self.eigenvalues)
-            .finish()
-    }
-}
-
 impl<T: ComplexField, D: Dim> SymmetricEigen<T, D>
 where
     DefaultAllocator: Allocator<T, D, D> + Allocator<T::RealField, D>,
@@ -299,10 +270,7 @@ where
     ///
     /// This is useful if some of the eigenvalues have been manually modified.
     #[must_use]
-    pub fn recompose(&self) -> OMatrix<T, D, D>
-    where
-        InnerOwned<T, D, D>: Clone,
-    {
+    pub fn recompose(&self) -> OMatrix<T, D, D> {
         let mut u_t = self.eigenvectors.clone();
         for i in 0..self.eigenvalues.len() {
             let val = self.eigenvalues[i];
diff --git a/src/linalg/symmetric_tridiagonal.rs b/src/linalg/symmetric_tridiagonal.rs
index f074b0eb..e071a916 100644
--- a/src/linalg/symmetric_tridiagonal.rs
+++ b/src/linalg/symmetric_tridiagonal.rs
@@ -1,16 +1,14 @@
-use std::fmt;
-use std::mem::MaybeUninit;
-
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Serialize};
 
 use crate::allocator::Allocator;
 use crate::base::{DefaultAllocator, OMatrix, OVector};
-use crate::dimension::{Const, DimDiff, DimName, DimSub, U1};
-use crate::storage::{InnerOwned, Storage};
+use crate::dimension::{Const, DimDiff, DimSub, U1};
 use simba::scalar::ComplexField;
 
 use crate::linalg::householder;
+use crate::Matrix;
+use std::mem::MaybeUninit;
 
 /// Tridiagonalization of a symmetric matrix.
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
@@ -28,7 +26,8 @@ use crate::linalg::householder;
          OMatrix<T, D, D>: Deserialize<'de>,
          OVector<T, DimDiff<D, U1>>: Deserialize<'de>"))
 )]
-pub struct SymmetricTridiagonal<T, D: DimSub<U1>>
+#[derive(Clone, Debug)]
+pub struct SymmetricTridiagonal<T: ComplexField, D: DimSub<U1>>
 where
     DefaultAllocator: Allocator<T, D, D> + Allocator<T, DimDiff<D, U1>>,
 {
@@ -36,42 +35,14 @@ where
     off_diagonal: OVector<T, DimDiff<D, U1>>,
 }
 
-impl<T: Copy, D: DimSub<U1> + DimName> Copy for SymmetricTridiagonal<T, D>
+impl<T: ComplexField, D: DimSub<U1>> Copy for SymmetricTridiagonal<T, D>
 where
     DefaultAllocator: Allocator<T, D, D> + Allocator<T, DimDiff<D, U1>>,
-    InnerOwned<T, D, D>: Copy,
-    InnerOwned<T, DimDiff<D, U1>>: Copy,
+    OMatrix<T, D, D>: Copy,
+    OVector<T, DimDiff<D, U1>>: Copy,
 {
 }
 
-impl<T: Clone, D: DimSub<U1>> Clone for SymmetricTridiagonal<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D> + Allocator<T, DimDiff<D, U1>>,
-    InnerOwned<T, D, D>: Clone,
-    InnerOwned<T, DimDiff<D, U1>>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self {
-            tri: self.tri.clone(),
-            off_diagonal: self.off_diagonal.clone(),
-        }
-    }
-}
-
-impl<T: fmt::Debug, D: DimSub<U1>> fmt::Debug for SymmetricTridiagonal<T, D>
-where
-    DefaultAllocator: Allocator<T, D, D> + Allocator<T, DimDiff<D, U1>>,
-    InnerOwned<T, D, D>: fmt::Debug,
-    InnerOwned<T, DimDiff<D, U1>>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("SymmetricTridiagonal")
-            .field("tri", &self.tri)
-            .field("off_diagonal", &self.off_diagonal)
-            .finish()
-    }
-}
-
 impl<T: ComplexField, D: DimSub<U1>> SymmetricTridiagonal<T, D>
 where
     DefaultAllocator: Allocator<T, D, D> + Allocator<T, DimDiff<D, U1>>,
@@ -80,7 +51,7 @@ where
     ///
     /// Only the lower-triangular part (including the diagonal) of `m` is read.
     pub fn new(mut m: OMatrix<T, D, D>) -> Self {
-        let dim = m.data.shape().0;
+        let dim = m.shape_generic().0;
 
         assert!(
             m.is_square(),
@@ -91,8 +62,8 @@ where
             "Unable to compute the symmetric tridiagonal decomposition of an empty matrix."
         );
 
-        let mut off_diagonal = OVector::new_uninitialized_generic(dim.sub(Const::<1>), Const::<1>);
-        let mut p = OVector::new_uninitialized_generic(dim.sub(Const::<1>), Const::<1>);
+        let mut off_diagonal = Matrix::uninit(dim.sub(Const::<1>), Const::<1>);
+        let mut p = Matrix::zeros_generic(dim.sub(Const::<1>), Const::<1>);
 
         for i in 0..dim.value() - 1 {
             let mut m = m.rows_range_mut(i + 1..);
@@ -104,8 +75,7 @@ where
             if not_zero {
                 let mut p = p.rows_range_mut(i..);
 
-                p.hegemv_z(crate::convert(2.0), &m, &axis);
-                let p = unsafe { p.slice_assume_init() };
+                p.hegemv(crate::convert(2.0), &m, &axis, T::zero());
 
                 let dot = axis.dotc(&p);
                 m.hegerc(-T::one(), &p, &axis, T::one());
@@ -114,9 +84,11 @@ where
             }
         }
 
+        // Safety: off_diagonal has been fully initialized.
+        let off_diagonal = unsafe { off_diagonal.assume_init() };
         Self {
             tri: m,
-            off_diagonal: unsafe { off_diagonal.assume_init() },
+            off_diagonal,
         }
     }
 
diff --git a/src/linalg/udu.rs b/src/linalg/udu.rs
index 5d78951b..546fa95a 100644
--- a/src/linalg/udu.rs
+++ b/src/linalg/udu.rs
@@ -1,12 +1,9 @@
-use std::fmt;
-
 #[cfg(feature = "serde-serialize-no-std")]
 use serde::{Deserialize, Serialize};
 
 use crate::allocator::Allocator;
 use crate::base::{Const, DefaultAllocator, OMatrix, OVector};
-use crate::dimension::{Dim, DimName};
-use crate::storage::{InnerOwned, Storage};
+use crate::dimension::Dim;
 use simba::scalar::RealField;
 
 /// UDU factorization.
@@ -21,7 +18,8 @@ use simba::scalar::RealField;
         deserialize = "OVector<T, D>: Deserialize<'de>, OMatrix<T, D, D>: Deserialize<'de>"
     ))
 )]
-pub struct UDU<T, D: Dim>
+#[derive(Clone, Debug)]
+pub struct UDU<T: RealField, D: Dim>
 where
     DefaultAllocator: Allocator<T, D> + Allocator<T, D, D>,
 {
@@ -31,42 +29,14 @@ where
     pub d: OVector<T, D>,
 }
 
-impl<T: Copy, D: DimName> Copy for UDU<T, D>
+impl<T: RealField, D: Dim> Copy for UDU<T, D>
 where
     DefaultAllocator: Allocator<T, D> + Allocator<T, D, D>,
-    InnerOwned<T, D>: Copy,
-    InnerOwned<T, D, D>: Copy,
+    OVector<T, D>: Copy,
+    OMatrix<T, D, D>: Copy,
 {
 }
 
-impl<T: Clone, D: Dim> Clone for UDU<T, D>
-where
-    DefaultAllocator: Allocator<T, D> + Allocator<T, D, D>,
-    InnerOwned<T, D>: Clone,
-    InnerOwned<T, D, D>: Clone,
-{
-    fn clone(&self) -> Self {
-        Self {
-            u: self.u.clone(),
-            d: self.d.clone(),
-        }
-    }
-}
-
-impl<T: fmt::Debug, D: Dim> fmt::Debug for UDU<T, D>
-where
-    DefaultAllocator: Allocator<T, D> + Allocator<T, D, D>,
-    InnerOwned<T, D>: fmt::Debug,
-    InnerOwned<T, D, D>: fmt::Debug,
-{
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.debug_struct("UDU")
-            .field("u", &self.u)
-            .field("d", &self.d)
-            .finish()
-    }
-}
-
 impl<T: RealField, D: Dim> UDU<T, D>
 where
     DefaultAllocator: Allocator<T, D> + Allocator<T, D, D>,
@@ -79,7 +49,7 @@ where
     /// Ref.: "Optimal control and estimation-Dover Publications", Robert F. Stengel, (1994) page 360
     pub fn new(p: OMatrix<T, D, D>) -> Option<Self> {
         let n = p.ncols();
-        let n_dim = p.data.shape().1;
+        let n_dim = p.shape_generic().1;
 
         let mut d = OVector::zeros_generic(n_dim, Const::<1>);
         let mut u = OMatrix::zeros_generic(n_dim, n_dim);
diff --git a/src/proptest/mod.rs b/src/proptest/mod.rs
index e910bdf0..a7cbe549 100644
--- a/src/proptest/mod.rs
+++ b/src/proptest/mod.rs
@@ -263,7 +263,7 @@ where
 }
 
 /// Same as `matrix`, but without the additional anonymous generic types
-fn matrix_<R: Dim, C: Dim, ScalarStrategy>(
+fn matrix_<R, C, ScalarStrategy>(
     value_strategy: ScalarStrategy,
     rows: DimRange<R>,
     cols: DimRange<C>,
@@ -271,6 +271,8 @@ fn matrix_<R: Dim, C: Dim, ScalarStrategy>(
 where
     ScalarStrategy: Strategy + Clone + 'static,
     ScalarStrategy::Value: Scalar,
+    R: Dim,
+    C: Dim,
     DefaultAllocator: Allocator<ScalarStrategy::Value, R, C>,
 {
     let nrows = rows.lower_bound().value()..=rows.upper_bound().value();
@@ -330,7 +332,12 @@ where
     matrix_(value_strategy, length.into(), Const::<1>.into())
 }
 
-impl<NParameters: Default, R: DimName, C: DimName> Default for MatrixParameters<NParameters, R, C> {
+impl<NParameters, R, C> Default for MatrixParameters<NParameters, R, C>
+where
+    NParameters: Default,
+    R: DimName,
+    C: DimName,
+{
     fn default() -> Self {
         Self {
             rows: DimRange::from(R::name()),
diff --git a/src/sparse/cs_matrix.rs b/src/sparse/cs_matrix.rs
index 4a1a3f83..c717e90e 100644
--- a/src/sparse/cs_matrix.rs
+++ b/src/sparse/cs_matrix.rs
@@ -7,7 +7,7 @@ use std::slice;
 
 use crate::allocator::Allocator;
 use crate::sparse::cs_utils;
-use crate::{Const, DefaultAllocator, Dim, Dynamic, OVector, Scalar, Vector, U1};
+use crate::{Const, DefaultAllocator, Dim, Dynamic, Matrix, OVector, Scalar, Vector, U1};
 
 pub struct ColumnEntries<'a, T> {
     curr: usize,
@@ -263,10 +263,6 @@ where
     /// `nvals` possible non-zero values.
     pub fn new_uninitialized_generic(nrows: R, ncols: C, nvals: usize) -> Self {
         let mut i = Vec::with_capacity(nvals);
-
-        // IMPORTANT TODO: this method is still UB, and we should decide how to
-        // update the API to take it into account.
-
         unsafe {
             i.set_len(nvals);
         }
@@ -474,7 +470,7 @@ where
     {
         // Size = R
         let nrows = self.data.shape().0;
-        let mut workspace = CsMatrix::new_uninitialized_generic(nrows, Const::<1>);
+        let mut workspace = Matrix::zeros_generic(nrows, Const::<1>);
         self.sort_with_workspace(workspace.as_mut_slice());
     }
 
diff --git a/src/sparse/cs_matrix_cholesky.rs b/src/sparse/cs_matrix_cholesky.rs
index cd8bf975..ff9ca023 100644
--- a/src/sparse/cs_matrix_cholesky.rs
+++ b/src/sparse/cs_matrix_cholesky.rs
@@ -3,7 +3,7 @@ use std::mem;
 
 use crate::allocator::Allocator;
 use crate::sparse::{CsMatrix, CsStorage, CsStorageIter, CsStorageIterMut, CsVecStorage};
-use crate::{Const, DefaultAllocator, Dim, OVector, RealField};
+use crate::{Const, DefaultAllocator, Dim, Matrix, OVector, RealField};
 
 /// The cholesky decomposition of a column compressed sparse matrix.
 pub struct CsCholesky<T: RealField, D: Dim>
@@ -48,8 +48,8 @@ where
         let (l, u) = Self::nonzero_pattern(m);
 
         // Workspaces.
-        let work_x = Matrix::new_uninitialized_generic(m.data.shape().0, Const::<1>);
-        let work_c = Matrix::new_uninitialized_generic(m.data.shape().1, Const::<1>);
+        let work_x = Matrix::zeros_generic(m.data.shape().0, Const::<1>);
+        let work_c = Matrix::zeros_generic(m.data.shape().1, Const::<1>);
         let mut original_p = m.data.p.as_slice().to_vec();
         original_p.push(m.data.i.len());
 
@@ -292,7 +292,7 @@ where
         let etree = Self::elimination_tree(m);
         let (nrows, ncols) = m.data.shape();
         let mut rows = Vec::with_capacity(m.len());
-        let mut cols = Matrix::new_uninitialized_generic(m.data.shape().0, Const::<1>);
+        let mut cols = Matrix::zeros_generic(m.data.shape().0, Const::<1>);
         let mut marks = Vec::new();
 
         // NOTE: the following will actually compute the non-zero pattern of
diff --git a/src/sparse/cs_matrix_ops.rs b/src/sparse/cs_matrix_ops.rs
index 2170f5d2..fba5d41b 100644
--- a/src/sparse/cs_matrix_ops.rs
+++ b/src/sparse/cs_matrix_ops.rs
@@ -6,7 +6,7 @@ use crate::allocator::Allocator;
 use crate::constraint::{AreMultipliable, DimEq, ShapeConstraint};
 use crate::sparse::{CsMatrix, CsStorage, CsStorageMut, CsVector};
 use crate::storage::StorageMut;
-use crate::{Const, DefaultAllocator, Dim, OVector, Scalar, Vector};
+use crate::{Const, DefaultAllocator, Dim, Matrix, OVector, Scalar, Vector};
 
 impl<T: Scalar, R: Dim, C: Dim, S: CsStorage<T, R, C>> CsMatrix<T, R, C, S> {
     fn scatter<R2: Dim, C2: Dim>(
@@ -242,7 +242,7 @@ where
 
         let mut res = CsMatrix::new_uninitialized_generic(nrows1, ncols2, self.len() + rhs.len());
         let mut timestamps = OVector::zeros_generic(nrows1, Const::<1>);
-        let mut workspace = Matrix::new_uninitialized_generic(nrows1, Const::<1>);
+        let mut workspace = Matrix::zeros_generic(nrows1, Const::<1>);
         let mut nz = 0;
 
         for j in 0..ncols2.value() {
diff --git a/src/sparse/cs_matrix_solve.rs b/src/sparse/cs_matrix_solve.rs
index 092ad15b..6136a0f8 100644
--- a/src/sparse/cs_matrix_solve.rs
+++ b/src/sparse/cs_matrix_solve.rs
@@ -152,7 +152,7 @@ impl<T: RealField, D: Dim, S: CsStorage<T, D, D>> CsMatrix<T, D, D, S> {
         self.lower_triangular_reach(b, &mut reach);
         // We sort the reach so the result matrix has sorted indices.
         reach.sort_unstable();
-        let mut workspace = Matrix::new_uninitialized_generic(b.data.shape().0, Const::<1>);
+        let mut workspace = Matrix::zeros_generic(b.data.shape().0, Const::<1>);
 
         for i in reach.iter().cloned() {
             workspace[i] = T::zero();
diff --git a/src/third_party/alga/alga_matrix.rs b/src/third_party/alga/alga_matrix.rs
index f80b021a..6a4cb982 100644
--- a/src/third_party/alga/alga_matrix.rs
+++ b/src/third_party/alga/alga_matrix.rs
@@ -15,8 +15,9 @@ use alga::linear::{
 
 use crate::base::allocator::Allocator;
 use crate::base::dimension::{Dim, DimName};
-use crate::base::storage::{Storage, StorageMut};
-use crate::base::{DefaultAllocator, OMatrix, Scalar};
+use crate::base::storage::{RawStorage, RawStorageMut};
+use crate::base::{DefaultAllocator, Matrix, OMatrix, Scalar};
+use std::mem::MaybeUninit;
 
 /*
  *
@@ -427,14 +428,14 @@ where
 {
     #[inline]
     fn meet_join(&self, other: &Self) -> (Self, Self) {
-        let shape = self.data.shape();
+        let shape = self.shape_generic();
         assert!(
-            shape == other.data.shape(),
+            shape == other.shape_generic(),
             "Matrix meet/join error: mismatched dimensions."
         );
 
-        let mut mres = Matrix::new_uninitialized_generic(shape.0, shape.1);
-        let mut jres = Matrix::new_uninitialized_generic(shape.0, shape.1);
+        let mut mres = Matrix::uninit(shape.0, shape.1);
+        let mut jres = Matrix::uninit(shape.0, shape.1);
 
         for i in 0..shape.0.value() * shape.1.value() {
             unsafe {
@@ -442,11 +443,12 @@ where
                     .data
                     .get_unchecked_linear(i)
                     .meet_join(other.data.get_unchecked_linear(i));
-                *mres.data.get_unchecked_linear_mut(i) = mj.0;
-                *jres.data.get_unchecked_linear_mut(i) = mj.1;
+                *mres.data.get_unchecked_linear_mut(i) = MaybeUninit::new(mj.0);
+                *jres.data.get_unchecked_linear_mut(i) = MaybeUninit::new(mj.1);
             }
         }
 
-        (mres, jres)
+        // Safety: both mres and jres are now completely initialized.
+        unsafe { (mres.assume_init(), jres.assume_init()) }
     }
 }
diff --git a/src/third_party/glam/common/glam_matrix.rs b/src/third_party/glam/common/glam_matrix.rs
index 77b68b5e..80f88054 100644
--- a/src/third_party/glam/common/glam_matrix.rs
+++ b/src/third_party/glam/common/glam_matrix.rs
@@ -2,7 +2,7 @@ use super::glam::{
     BVec2, BVec3, BVec4, DMat2, DMat3, DMat4, DVec2, DVec3, DVec4, IVec2, IVec3, IVec4, Mat2, Mat3,
     Mat4, UVec2, UVec3, UVec4, Vec2, Vec3, Vec3A, Vec4,
 };
-use crate::storage::Storage;
+use crate::storage::RawStorage;
 use crate::{Matrix, Matrix2, Matrix3, Matrix4, Vector, Vector2, Vector3, Vector4, U2, U3, U4};
 
 macro_rules! impl_vec_conversion(
@@ -16,7 +16,7 @@ macro_rules! impl_vec_conversion(
 
         impl<S> From<Vector<$N, U2, S>> for $Vec2
         where
-            S: Storage<$N, U2>,
+            S: RawStorage<$N, U2>,
         {
             #[inline]
             fn from(e: Vector<$N, U2, S>) -> $Vec2 {
@@ -33,7 +33,7 @@ macro_rules! impl_vec_conversion(
 
         impl<S> From<Vector<$N, U3, S>> for $Vec3
         where
-            S: Storage<$N, U3>,
+            S: RawStorage<$N, U3>,
         {
             #[inline]
             fn from(e: Vector<$N, U3, S>) -> $Vec3 {
@@ -50,7 +50,7 @@ macro_rules! impl_vec_conversion(
 
         impl<S> From<Vector<$N, U4, S>> for $Vec4
         where
-            S: Storage<$N, U4>,
+            S: RawStorage<$N, U4>,
         {
             #[inline]
             fn from(e: Vector<$N, U4, S>) -> $Vec4 {
@@ -75,7 +75,7 @@ impl From<Vec3A> for Vector3<f32> {
 
 impl<S> From<Vector<f32, U3, S>> for Vec3A
 where
-    S: Storage<f32, U3>,
+    S: RawStorage<f32, U3>,
 {
     #[inline]
     fn from(e: Vector<f32, U3, S>) -> Vec3A {
@@ -92,7 +92,7 @@ impl From<Mat2> for Matrix2<f32> {
 
 impl<S> From<Matrix<f32, U2, U2, S>> for Mat2
 where
-    S: Storage<f32, U2, U2>,
+    S: RawStorage<f32, U2, U2>,
 {
     #[inline]
     fn from(e: Matrix<f32, U2, U2, S>) -> Mat2 {
@@ -112,7 +112,7 @@ impl From<Mat3> for Matrix3<f32> {
 
 impl<S> From<Matrix<f32, U3, U3, S>> for Mat3
 where
-    S: Storage<f32, U3, U3>,
+    S: RawStorage<f32, U3, U3>,
 {
     #[inline]
     fn from(e: Matrix<f32, U3, U3, S>) -> Mat3 {
@@ -133,7 +133,7 @@ impl From<Mat4> for Matrix4<f32> {
 
 impl<S> From<Matrix<f32, U4, U4, S>> for Mat4
 where
-    S: Storage<f32, U4, U4>,
+    S: RawStorage<f32, U4, U4>,
 {
     #[inline]
     fn from(e: Matrix<f32, U4, U4, S>) -> Mat4 {
@@ -155,7 +155,7 @@ impl From<DMat2> for Matrix2<f64> {
 
 impl<S> From<Matrix<f64, U2, U2, S>> for DMat2
 where
-    S: Storage<f64, U2, U2>,
+    S: RawStorage<f64, U2, U2>,
 {
     #[inline]
     fn from(e: Matrix<f64, U2, U2, S>) -> DMat2 {
@@ -175,7 +175,7 @@ impl From<DMat3> for Matrix3<f64> {
 
 impl<S> From<Matrix<f64, U3, U3, S>> for DMat3
 where
-    S: Storage<f64, U3, U3>,
+    S: RawStorage<f64, U3, U3>,
 {
     #[inline]
     fn from(e: Matrix<f64, U3, U3, S>) -> DMat3 {
@@ -196,7 +196,7 @@ impl From<DMat4> for Matrix4<f64> {
 
 impl<S> From<Matrix<f64, U4, U4, S>> for DMat4
 where
-    S: Storage<f64, U4, U4>,
+    S: RawStorage<f64, U4, U4>,
 {
     #[inline]
     fn from(e: Matrix<f64, U4, U4, S>) -> DMat4 {
diff --git a/src/third_party/mint/mint_matrix.rs b/src/third_party/mint/mint_matrix.rs
index 1e0a4d54..73d0a936 100644
--- a/src/third_party/mint/mint_matrix.rs
+++ b/src/third_party/mint/mint_matrix.rs
@@ -4,7 +4,7 @@ use std::ptr;
 
 use crate::base::allocator::Allocator;
 use crate::base::dimension::{U1, U2, U3, U4};
-use crate::base::storage::{ContiguousStorage, ContiguousStorageMut, Storage, StorageMut};
+use crate::base::storage::{IsContiguous, RawStorage, RawStorageMut};
 use crate::base::{DefaultAllocator, Matrix, OMatrix, Scalar};
 
 macro_rules! impl_from_into_mint_1D(
@@ -25,9 +25,10 @@ macro_rules! impl_from_into_mint_1D(
 
         impl<T, S> Into<mint::$VT<T>> for Matrix<T, $NRows, U1, S>
         where T: Scalar,
-              S: ContiguousStorage<T, $NRows, U1> {
+              S: RawStorage<T, $NRows, U1> + IsContiguous {
             #[inline]
             fn into(self) -> mint::$VT<T> {
+                // SAFETY: this is OK thanks to the IsContiguous bound.
                 unsafe {
                     let mut res: mint::$VT<T> = mem::MaybeUninit::uninit().assume_init();
                     ptr::copy_nonoverlapping(self.data.ptr(), &mut res.x, $SZ);
@@ -38,9 +39,10 @@ macro_rules! impl_from_into_mint_1D(
 
         impl<T, S> AsRef<mint::$VT<T>> for Matrix<T, $NRows, U1, S>
         where T: Scalar,
-              S: ContiguousStorage<T, $NRows, U1> {
+              S: RawStorage<T, $NRows, U1> + IsContiguous {
             #[inline]
             fn as_ref(&self) -> &mint::$VT<T> {
+                // SAFETY: this is OK thanks to the IsContiguous bound.
                 unsafe {
                     mem::transmute(self.data.ptr())
                 }
@@ -49,9 +51,10 @@ macro_rules! impl_from_into_mint_1D(
 
         impl<T, S> AsMut<mint::$VT<T>> for Matrix<T, $NRows, U1, S>
         where T: Scalar,
-              S: ContiguousStorageMut<T, $NRows, U1> {
+              S: RawStorageMut<T, $NRows, U1> + IsContiguous {
             #[inline]
             fn as_mut(&mut self) -> &mut mint::$VT<T> {
+                // SAFETY: this is OK thanks to the IsContiguous bound.
                 unsafe {
                     mem::transmute(self.data.ptr_mut())
                 }
diff --git a/src/third_party/mint/mint_point.rs b/src/third_party/mint/mint_point.rs
index fbce1c88..45f85e3c 100644
--- a/src/third_party/mint/mint_point.rs
+++ b/src/third_party/mint/mint_point.rs
@@ -1,4 +1,4 @@
-use crate::base::storage::{Storage, StorageMut};
+use crate::base::storage::{RawStorage, RawStorageMut};
 use crate::{OVector, Point, Scalar};
 use std::convert::{AsMut, AsRef};
 
diff --git a/src/third_party/mint/mint_quaternion.rs b/src/third_party/mint/mint_quaternion.rs
index 49b99f04..f41815ce 100644
--- a/src/third_party/mint/mint_quaternion.rs
+++ b/src/third_party/mint/mint_quaternion.rs
@@ -1,6 +1,6 @@
 use crate::{Quaternion, Scalar, SimdValue, UnitQuaternion};
 
-impl<T> From<mint::Quaternion<T>> for Quaternion<T> {
+impl<T: Scalar> From<mint::Quaternion<T>> for Quaternion<T> {
     fn from(q: mint::Quaternion<T>) -> Self {
         Self::new(q.s, q.v.x, q.v.y, q.v.z)
     }
diff --git a/tests/core/matrix.rs b/tests/core/matrix.rs
index eaa252db..4a35fb20 100644
--- a/tests/core/matrix.rs
+++ b/tests/core/matrix.rs
@@ -447,7 +447,7 @@ fn apply() {
         1.0, 2.0, 3.0, 4.0, 6.0, 7.0, 8.0, 9.0, 10.0, 9.0, 8.0, 7.0, 6.0, 4.0, 3.0, 2.0,
     );
 
-    a.apply(|e| e.round());
+    a.apply(|e| *e = e.round());
 
     assert_eq!(a, expected);
 }