scalar rotates, rotate instr tests

Author: TheIronBorn

coresimd/ppsv/api/mod.rs

@@ -127,6 +127,8 @@ mod shifts;
 #[macro_use]
 mod rotates;
 
+pub use self::rotates::Rotate;
+
 /// Sealed trait used for constraining select implementations.
 pub trait Lanes<A> {}
 

coresimd/ppsv/api/rotates.rs

@@ -1,39 +1,59 @@
 //! Implements integer rotates.
 #![allow(unused)]
 
+/// Trait used for overloading rotates
+pub trait Rotate<T> {
+    /// Shifts the bits of each lane to the left by the specified amount in
+    /// the corresponding lane of `n`, wrapping the truncated bits to
+    /// the end of the resulting integer.
+    ///
+    /// Please note this isn't the same operation as `<<`!. Also note it
+    /// isn't equivalent to `slice::rotate_left`, it doesn't move the vector's
+    /// lanes around. (that can be implemented with vector shuffles).
+    fn rotate_left(self, n: T) -> Self;
+
+    /// Shifts the bits of each lane to the right by the specified amount in
+    /// the corresponding lane of `n`, wrapping the truncated bits to
+    /// the beginning of the resulting integer.
+    ///
+    /// Please note this isn't the same operation as `>>`!. Also note it
+    /// isn't similar to `slice::rotate_right`, it doesn't move the vector's
+    /// lanes around. (that can be implemented with vector shuffles).
+    fn rotate_right(self, n: T) -> Self;
+}
+
 macro_rules! impl_vector_rotates {
     ($id:ident, $elem_ty:ident) => {
-        impl $id {
-            /// Shifts the bits of each lane to the left by the specified amount in
-            /// the corresponding lane of `n`, wrapping the truncated bits to
-            /// the end of the resulting integer.
-            ///
-            /// Please note this isn't the same operation as `<<`!. Also note it
-            /// isn't equivalent to `slice::rotate_left` (that can be implemented
-            /// with vector shuffles).
+        impl super::api::Rotate<$id> for $id {
             #[inline]
-            pub fn rotate_left(self, n: $id) -> $id {
+            fn rotate_left(self, n: $id) -> $id {
                 const LANE_WIDTH: $elem_ty = ::mem::size_of::<$elem_ty>() as $elem_ty * 8;
                 // Protect against undefined behavior for over-long bit shifts
                 let n = n % LANE_WIDTH;
                 (self << n) | (self >> ((LANE_WIDTH - n) % LANE_WIDTH))
             }
 
-            /// Shifts the bits of each lane to the right by the specified amount in
-            /// the corresponding lane of `n`, wrapping the truncated bits to
-            /// the beginning of the resulting integer.
-            ///
-            /// Please note this isn't the same operation as `>>`!. Also note it
-            /// isn't similar to `slice::rotate_right`, it doesn't move the vector's
-            /// lanes around. (that can be implemented with vector shuffles).
             #[inline]
-            pub fn rotate_right(self, n: $id) -> $id {
+            fn rotate_right(self, n: $id) -> $id {
                 const LANE_WIDTH: $elem_ty = ::mem::size_of::<$elem_ty>() as $elem_ty * 8;
                 // Protect against undefined behavior for over-long bit shifts
                 let n = n % LANE_WIDTH;
                 (self >> n) | (self << ((LANE_WIDTH - n) % LANE_WIDTH))
             }
         }
+
+        impl super::api::Rotate<$elem_ty> for $id {
+            #[inline]
+            fn rotate_left(self, n: $elem_ty) -> $id {
+                self.rotate_left($id::splat(n))
+            }
+
+
+            #[inline]
+            fn rotate_right(self, n: $elem_ty) -> $id {
+                self.rotate_right($id::splat(n))
+            }
+        }
     };
 }
 
@@ -43,6 +63,7 @@ macro_rules! test_vector_rotate_ops {
         #[test]
         fn rotate_ops() {
             use coresimd::simd::$id;
+            use coresimd::simd::Rotate;;
             use std::mem;
             let z = $id::splat(0 as $elem_ty);
             let o = $id::splat(1 as $elem_ty);

coresimd/ppsv/mod.rs

@@ -48,6 +48,8 @@ pub use self::v32::*;
 pub use self::v512::*;
 pub use self::v64::*;
 
+pub use self::api::Rotate;
+
 /// Safe lossless bitwise conversion from `T` to `Self`.
 pub trait FromBits<T>: ::marker::Sized {
     /// Safe lossless bitwise from `T` to `Self`.

crates/coresimd/tests/rotate_tests.rs

@@ -0,0 +1,49 @@
+//! rotate instruction tests
+
+#![feature(stdsimd)]
+#![feature(proc_macro)]
+#![feature(avx512_target_feature)]
+#![feature(abi_vectorcall)]
+
+extern crate stdsimd_test;
+extern crate stdsimd;
+
+use stdsimd::simd::*;
+use stdsimd_test::assert_instr;
+
+// Verify that supported hardware compiles rotates into single instructions
+
+#[inline]
+#[cfg_attr(any(target_arch = "x86", target_arch = "x86_64"), target_feature(enable = "avx512f"))]
+#[cfg_attr(any(target_arch = "x86", target_arch = "x86_64"), assert_instr(vprorvq))]
+unsafe fn rotate_right_variable(x: u64x8) -> u64x8 {
+    x.rotate_right(u64x8::new(0, 1, 2, 3, 4, 5, 6, 7))
+}
+
+#[inline]
+#[cfg_attr(any(target_arch = "x86", target_arch = "x86_64"), target_feature(enable = "avx512f"))]
+#[cfg_attr(any(target_arch = "x86", target_arch = "x86_64"), assert_instr(vprolvq))]
+unsafe fn rotate_left_variable(x: u64x8) -> u64x8 {
+    x.rotate_left(u64x8::new(0, 1, 2, 3, 4, 5, 6, 7))
+}
+
+#[inline]
+#[cfg_attr(any(target_arch = "x86", target_arch = "x86_64"), target_feature(enable = "avx512f"))]
+#[cfg_attr(any(target_arch = "x86", target_arch = "x86_64"), assert_instr(vprorq))]
+unsafe fn rotate_right(x: u64x8) -> u64x8 {
+    x.rotate_right(12)
+}
+
+#[inline]
+#[cfg_attr(any(target_arch = "x86", target_arch = "x86_64"), target_feature(enable = "avx512f"))]
+#[cfg_attr(any(target_arch = "x86", target_arch = "x86_64"), assert_instr(vprolq))]
+unsafe fn rotate_left(x: u64x8) -> u64x8 {
+    x.rotate_left(12)
+}
+
+#[inline]
+#[cfg_attr(any(target_arch = "x86", target_arch = "x86_64"), target_feature(enable = "avx512f"))]
+#[cfg_attr(any(target_arch = "x86", target_arch = "x86_64"), assert_instr(vprolq))]
+unsafe fn rotate_left_x2(x: u64x2) -> u64x2 {
+    x.rotate_left(12)
+}