Fix wrapping pointer arithmetic

crates/core_simd/src/elements/const_ptr.rs

@@ -1,11 +1,14 @@
 use super::sealed::Sealed;
-use crate::simd::{LaneCount, Mask, Simd, SimdPartialEq, SupportedLaneCount};
+use crate::simd::{intrinsics, LaneCount, Mask, Simd, SimdPartialEq, SupportedLaneCount};
 
 /// Operations on SIMD vectors of constant pointers.
 pub trait SimdConstPtr: Copy + Sealed {
-    /// Vector of usize with the same number of lanes.
+    /// Vector of `usize` with the same number of lanes.
     type Usize;
 
+    /// Vector of `isize` with the same number of lanes.
+    type Isize;
+
     /// Vector of mutable pointers to the same type.
     type MutPtr;
 
@@ -23,10 +26,20 @@ pub trait SimdConstPtr: Copy + Sealed {
     /// Equivalent to calling [`pointer::addr`] on each lane.
     fn addr(self) -> Self::Usize;
 
+    /// Calculates the offset from a pointer using wrapping arithmetic.
+    ///
+    /// Equivalent to calling [`pointer::wrapping_offset`] on each lane.
+    fn wrapping_offset(self, offset: Self::Isize) -> Self;
+
     /// Calculates the offset from a pointer using wrapping arithmetic.
     ///
     /// Equivalent to calling [`pointer::wrapping_add`] on each lane.
     fn wrapping_add(self, count: Self::Usize) -> Self;
+
+    /// Calculates the offset from a pointer using wrapping arithmetic.
+    ///
+    /// Equivalent to calling [`pointer::wrapping_add`] on each lane.
+    fn wrapping_sub(self, count: Self::Usize) -> Self;
 }
 
 impl<T, const LANES: usize> Sealed for Simd<*const T, LANES> where
@@ -39,6 +52,7 @@ impl<T, const LANES: usize> SimdConstPtr for Simd<*const T, LANES>
     LaneCount<LANES>: SupportedLaneCount,
 {
     type Usize = Simd<usize, LANES>;
+    type Isize = Simd<isize, LANES>;
     type MutPtr = Simd<*mut T, LANES>;
     type Mask = Mask<isize, LANES>;
 
@@ -57,10 +71,19 @@ fn addr(self) -> Self::Usize {
         self.cast()
     }
 
+    #[inline]
+    fn wrapping_offset(self, count: Self::Isize) -> Self {
+        // Safety: simd_arith_offset takes a vector of pointers and a vector of offsets
+        unsafe { intrinsics::simd_arith_offset(self, count) }
+    }
+
     #[inline]
     fn wrapping_add(self, count: Self::Usize) -> Self {
-        let addr = self.addr() + (count * Simd::splat(core::mem::size_of::<T>()));
-        // Safety: transmuting usize to pointers is safe, even if accessing those pointers isn't.
-        unsafe { core::mem::transmute_copy(&addr) }
+        self.wrapping_offset(count.cast())
+    }
+
+    #[inline]
+    fn wrapping_sub(self, count: Self::Usize) -> Self {
+        self.wrapping_offset(-count.cast::<isize>())
     }
 }

crates/core_simd/src/elements/mut_ptr.rs

@@ -1,11 +1,14 @@
 use super::sealed::Sealed;
-use crate::simd::{LaneCount, Mask, Simd, SimdPartialEq, SupportedLaneCount};
+use crate::simd::{intrinsics, LaneCount, Mask, Simd, SimdPartialEq, SupportedLaneCount};
 
 /// Operations on SIMD vectors of mutable pointers.
 pub trait SimdMutPtr: Copy + Sealed {
-    /// Vector of usize with the same number of lanes.
+    /// Vector of `usize` with the same number of lanes.
     type Usize;
 
+    /// Vector of `isize` with the same number of lanes.
+    type Isize;
+
     /// Vector of constant pointers to the same type.
     type ConstPtr;
 
@@ -23,10 +26,20 @@ pub trait SimdMutPtr: Copy + Sealed {
     /// Equivalent to calling [`pointer::addr`] on each lane.
     fn addr(self) -> Self::Usize;
 
+    /// Calculates the offset from a pointer using wrapping arithmetic.
+    ///
+    /// Equivalent to calling [`pointer::wrapping_offset`] on each lane.
+    fn wrapping_offset(self, offset: Self::Isize) -> Self;
+
     /// Calculates the offset from a pointer using wrapping arithmetic.
     ///
     /// Equivalent to calling [`pointer::wrapping_add`] on each lane.
     fn wrapping_add(self, count: Self::Usize) -> Self;
+
+    /// Calculates the offset from a pointer using wrapping arithmetic.
+    ///
+    /// Equivalent to calling [`pointer::wrapping_add`] on each lane.
+    fn wrapping_sub(self, count: Self::Usize) -> Self;
 }
 
 impl<T, const LANES: usize> Sealed for Simd<*mut T, LANES> where LaneCount<LANES>: SupportedLaneCount
@@ -37,6 +50,7 @@ impl<T, const LANES: usize> SimdMutPtr for Simd<*mut T, LANES>
     LaneCount<LANES>: SupportedLaneCount,
 {
     type Usize = Simd<usize, LANES>;
+    type Isize = Simd<isize, LANES>;
     type ConstPtr = Simd<*const T, LANES>;
     type Mask = Mask<isize, LANES>;
 
@@ -55,10 +69,19 @@ fn addr(self) -> Self::Usize {
         self.cast()
     }
 
+    #[inline]
+    fn wrapping_offset(self, count: Self::Isize) -> Self {
+        // Safety: simd_arith_offset takes a vector of pointers and a vector of offsets
+        unsafe { intrinsics::simd_arith_offset(self, count) }
+    }
+
     #[inline]
     fn wrapping_add(self, count: Self::Usize) -> Self {
-        let addr = self.addr() + (count * Simd::splat(core::mem::size_of::<T>()));
-        // Safety: transmuting usize to pointers is safe, even if accessing those pointers isn't.
-        unsafe { core::mem::transmute_copy(&addr) }
+        self.wrapping_offset(count.cast())
+    }
+
+    #[inline]
+    fn wrapping_sub(self, count: Self::Usize) -> Self {
+        self.wrapping_offset(-count.cast::<isize>())
     }
 }

crates/core_simd/src/intrinsics.rs

@@ -151,4 +151,7 @@
     pub(crate) fn simd_select<M, T>(m: M, yes: T, no: T) -> T;
     #[allow(unused)]
     pub(crate) fn simd_select_bitmask<M, T>(m: M, yes: T, no: T) -> T;
+
+    // equivalent to wrapping_offset
+    pub(crate) fn simd_arith_offset<T, U>(ptr: T, offset: U) -> T;
 }

crates/core_simd/tests/pointers.rs

@@ -21,6 +21,14 @@ fn addr<const LANES: usize>() {
                 );
             }
 
+            fn wrapping_offset<const LANES: usize>() {
+                test_helpers::test_binary_elementwise(
+                    &Simd::<*$constness (), LANES>::wrapping_offset,
+                    &<*$constness ()>::wrapping_offset,
+                    &|_, _| true,
+                );
+            }
+
             fn wrapping_add<const LANES: usize>() {
                 test_helpers::test_binary_elementwise(
                     &Simd::<*$constness (), LANES>::wrapping_add,
@@ -28,6 +36,14 @@ fn wrapping_add<const LANES: usize>() {
                     &|_, _| true,
                 );
             }
+
+            fn wrapping_sub<const LANES: usize>() {
+                test_helpers::test_binary_elementwise(
+                    &Simd::<*$constness (), LANES>::wrapping_sub,
+                    &<*$constness ()>::wrapping_sub,
+                    &|_, _| true,
+                );
+            }
         }
     }
 }