Add _mm_sfence to all non-temporal intrinsic tests

Author: sayantn

crates/core_arch/src/x86/avx.rs

@@ -4249,6 +4249,7 @@ mod tests {
         let a = _mm256_setr_epi64x(1, 2, 3, 4);
         let mut r = _mm256_undefined_si256();
         _mm256_stream_si256(ptr::addr_of_mut!(r), a);
+        _mm_sfence();
         assert_eq_m256i(r, a);
     }
 
@@ -4263,6 +4264,7 @@ mod tests {
         let mut mem = Memory { data: [-1.0; 4] };
 
         _mm256_stream_pd(ptr::addr_of_mut!(mem.data[0]), a);
+        _mm_sfence();
         for i in 0..4 {
             assert_eq!(mem.data[i], get_m256d(a, i));
         }
@@ -4279,6 +4281,7 @@ mod tests {
         let mut mem = Memory { data: [-1.0; 8] };
 
         _mm256_stream_ps(ptr::addr_of_mut!(mem.data[0]), a);
+        _mm_sfence();
         for i in 0..8 {
             assert_eq!(mem.data[i], get_m256(a, i));
         }

crates/core_arch/src/x86/avx512f.rs

@@ -56432,6 +56432,7 @@ mod tests {
         let mut mem = Memory { data: [-1.0; 16] };
 
         _mm512_stream_ps(&mut mem.data[0] as *mut f32, a);
+        _mm_sfence();
         for i in 0..16 {
             assert_eq!(mem.data[i], get_m512(a, i));
         }
@@ -56448,6 +56449,7 @@ mod tests {
         let mut mem = Memory { data: [-1.0; 8] };
 
         _mm512_stream_pd(&mut mem.data[0] as *mut f64, a);
+        _mm_sfence();
         for i in 0..8 {
             assert_eq!(mem.data[i], get_m512d(a, i));
         }
@@ -56464,6 +56466,7 @@ mod tests {
         let mut mem = Memory { data: [-1; 8] };
 
         _mm512_stream_si512(mem.data.as_mut_ptr().cast(), a);
+        _mm_sfence();
         for i in 0..8 {
             assert_eq!(mem.data[i], get_m512i(a, i));
         }

crates/core_arch/src/x86/sse.rs

@@ -3331,6 +3331,7 @@ mod tests {
         let mut mem = Memory { data: [-1.0; 4] };
 
         _mm_stream_ps(ptr::addr_of_mut!(mem.data[0]), a);
+        _mm_sfence();
         for i in 0..4 {
             assert_eq!(mem.data[i], get_m128(a, i));
         }

crates/core_arch/src/x86/sse2.rs

@@ -4066,6 +4066,7 @@ mod tests {
         );
         let mut r = _mm_set1_epi8(0);
         _mm_maskmoveu_si128(a, mask, ptr::addr_of_mut!(r) as *mut i8);
+        _mm_sfence();
         let e = _mm_set_epi8(0, 0, 9, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
         assert_eq_m128i(r, e);
     }
@@ -4102,6 +4103,7 @@ mod tests {
         let a = _mm_setr_epi32(1, 2, 3, 4);
         let mut r = _mm_undefined_si128();
         _mm_stream_si128(ptr::addr_of_mut!(r), a);
+        _mm_sfence();
         assert_eq_m128i(r, a);
     }
 
@@ -4113,6 +4115,7 @@ mod tests {
         let a: i32 = 7;
         let mut mem = boxed::Box::<i32>::new(-1);
         _mm_stream_si32(ptr::addr_of_mut!(*mem), a);
+        _mm_sfence();
         assert_eq!(a, *mem);
     }
 
@@ -4809,6 +4812,7 @@ mod tests {
         let mut mem = Memory { data: [-1.0; 2] };
 
         _mm_stream_pd(ptr::addr_of_mut!(mem.data[0]), a);
+        _mm_sfence();
         for i in 0..2 {
             assert_eq!(mem.data[i], get_m128d(a, i));
         }

crates/core_arch/src/x86/sse4a.rs

@@ -215,6 +215,7 @@ mod tests {
             let x = _mm_setr_pd(3.0, 4.0);
 
             _mm_stream_sd(d, x);
+            _mm_sfence();
         }
         assert_eq!(mem.data[0], 3.0);
         assert_eq!(mem.data[1], 2.0);
@@ -240,6 +241,7 @@ mod tests {
             let x = _mm_setr_ps(5.0, 6.0, 7.0, 8.0);
 
             _mm_stream_ss(d, x);
+            _mm_sfence();
         }
         assert_eq!(mem.data[0], 5.0);
         assert_eq!(mem.data[1], 2.0);

crates/core_arch/src/x86_64/sse2.rs

@@ -200,6 +200,7 @@ mod tests {
         let a: i64 = 7;
         let mut mem = boxed::Box::<i64>::new(-1);
         _mm_stream_si64(ptr::addr_of_mut!(*mem), a);
+        _mm_sfence();
         assert_eq!(a, *mem);
     }