refactor: replace load_32x4 with load_u32

src/intrinsics/aarch64.rs

@@ -3,16 +3,6 @@ use core::arch::asm;
 
 pub use core::arch::aarch64::uint8x16_t as AesBlock;
 
-#[inline]
-pub fn load_32x4(a: u32, b: u32, c: u32, d: u32) -> AesBlock {
-    unsafe {
-        vreinterpretq_u8_u32(vsetq_lane_u32::<3>(
-            d,
-            vsetq_lane_u32::<2>(c, vsetq_lane_u32::<1>(b, vmovq_n_u32(a))),
-        ))
-    }
-}
-
 #[inline]
 pub fn zero() -> AesBlock {
     unsafe { vmovq_n_u8(0) }
@@ -28,6 +18,11 @@ pub fn store(bytes: &mut [u8], block: AesBlock) {
     unsafe { vst1q_u8(bytes.as_mut_ptr(), block) };
 }
 
+#[inline]
+pub fn load_u32(words: &[u32]) -> AesBlock {
+    unsafe { vreinterpretq_u8_u32(vld1q_u32(words.as_ptr())) }
+}
+
 #[inline]
 pub fn store_u32(bytes: &mut [u32], block: AesBlock) {
     unsafe { vst1q_u32(bytes.as_mut_ptr(), vreinterpretq_u32_u8(block)) };

src/intrinsics/x86_64.rs

@@ -7,11 +7,6 @@ pub fn zero() -> AesBlock {
     unsafe { _mm_setzero_si128() }
 }
 
-#[inline]
-pub fn load_32x4(a: u32, b: u32, c: u32, d: u32) -> AesBlock {
-    unsafe { _mm_set_epi32(d as i32, c as i32, b as i32, a as i32) }
-}
-
 #[inline]
 pub fn load(bytes: &[u8]) -> AesBlock {
     unsafe { _mm_loadu_si128(bytes.as_ptr() as *const __m128i) }
@@ -22,6 +17,11 @@ pub fn store(bytes: &mut [u8], block: AesBlock) {
     unsafe { _mm_storeu_si128(bytes.as_mut_ptr() as *mut __m128i, block) };
 }
 
+#[inline]
+pub fn load_u32(words: &[u32]) -> AesBlock {
+    unsafe { _mm_loadu_si128(words.as_ptr() as *const __m128i) }
+}
+
 #[inline]
 pub fn store_u32(bytes: &mut [u32], block: AesBlock) {
     unsafe { _mm_storeu_si128(bytes.as_mut_ptr() as *mut __m128i, block) };

src/lib.rs

@@ -190,8 +190,8 @@ pub fn areion512_dm(
     store_u32(&mut x[12..], x3_p);
 
     (
-        load_32x4(x[2], x[3], x[6], x[7]),
-        load_32x4(x[8], x[9], x[12], x[13]),
+        load_u32(&[x[2], x[3], x[6], x[7]]),
+        load_u32(&[x[8], x[9], x[12], x[13]]),
     )
 }