fixed sse2!

Author: nstilt1

chacha20/src/backends/sse2.rs

@@ -146,11 +146,10 @@ unsafe fn rounds<R: Rounds>(v: &[__m128i; 4]) -> [[__m128i; 4]; PAR_BLOCKS] {
     }
 
     for block in 0..PAR_BLOCKS {
-        for i in 0..4 {
+        for i in 0..3 {
             res[block][i] = _mm_add_epi32(res[block][i], v[i]);
         }
-        // add the counter since `v` is lacking updated counter values
-        res[block][3] = _mm_add_epi32(res[block][3], _mm_set_epi32(0, 0, 0, block as i32));
+        res[block][3] = _mm_add_epi32(res[block][3], _mm_add_epi64(v[3], _mm_set_epi64x(0, block as i64)));
     }
 
     res

chacha20/src/rng.rs

@@ -1193,7 +1193,7 @@ pub(crate) mod tests {
         rng.fill_bytes(&mut result);
         assert_eq!(first_blocks_end_word_pos, rng.get_word_pos());
         assert_eq!(first_blocks_end_block_counter, rng.get_block_pos() - 3);
-        
+
         if first_blocks[0..64 * 4].ne(&result[64..]) {
             for (i, (a, b)) in first_blocks.iter().zip(result.iter().skip(64)).enumerate() {
                 if a.ne(b) {