2 files changed, 71 insertions, 0 deletions

diff --git a/compiler/rustc_codegen_cranelift/example/neon.rs b/compiler/rustc_codegen_cranelift/example/neon.rs
index 00e437d7e54..69ce17d3d75 100644
--- a/compiler/rustc_codegen_cranelift/example/neon.rs
+++ b/compiler/rustc_codegen_cranelift/example/neon.rs
@@ -203,6 +203,44 @@ unsafe fn test_vqadd_u8() {
 }
 
 #[cfg(target_arch = "aarch64")]
+unsafe fn test_vmaxq_f32() {
+    // AArch64 llvm intrinsic: llvm.aarch64.neon.fmax.v4f32
+    let a = f32x4::from([0., -1., 2., -3.]);
+    let b = f32x4::from([-4., 5., -6., 7.]);
+    let e = f32x4::from([0., 5., 2., 7.]);
+    let r: f32x4 = transmute(vmaxq_f32(transmute(a), transmute(b)));
+    assert_eq!(r, e);
+}
+
+#[cfg(target_arch = "aarch64")]
+unsafe fn test_vminq_f32() {
+    // AArch64 llvm intrinsic: llvm.aarch64.neon.fmin.v4f32
+    let a = f32x4::from([0., -1., 2., -3.]);
+    let b = f32x4::from([-4., 5., -6., 7.]);
+    let e = f32x4::from([-4., -1., -6., -3.]);
+    let r: f32x4 = transmute(vminq_f32(transmute(a), transmute(b)));
+    assert_eq!(r, e);
+}
+
+#[cfg(target_arch = "aarch64")]
+unsafe fn test_vaddvq_f32() {
+    // AArch64 llvm intrinsic: llvm.aarch64.neon.faddv.f32.v4f32
+    let a = f32x4::from([0., 1., 2., 3.]);
+    let e = 6f32;
+    let r = vaddvq_f32(transmute(a));
+    assert_eq!(r, e);
+}
+
+#[cfg(target_arch = "aarch64")]
+unsafe fn test_vrndnq_f32() {
+    // AArch64 llvm intrinsic: llvm.aarch64.neon.frintn.v4f32
+    let a = f32x4::from([0.1, -1.9, 4.5, 5.5]);
+    let e = f32x4::from([0., -2., 4., 6.]);
+    let r: f32x4 = transmute(vrndnq_f32(transmute(a)));
+    assert_eq!(r, e);
+}
+
+#[cfg(target_arch = "aarch64")]
 fn main() {
     unsafe {
         test_vpmin_s8();
@@ -229,6 +267,11 @@ fn main() {
 
         test_vqsub_u8();
         test_vqadd_u8();
+
+        test_vmaxq_f32();
+        test_vminq_f32();
+        test_vaddvq_f32();
+        test_vrndnq_f32();
     }
 }
 
diff --git a/compiler/rustc_codegen_cranelift/example/std_example.rs b/compiler/rustc_codegen_cranelift/example/std_example.rs
index 602452d8276..ebaa9c69c81 100644
--- a/compiler/rustc_codegen_cranelift/example/std_example.rs
+++ b/compiler/rustc_codegen_cranelift/example/std_example.rs
@@ -258,6 +258,9 @@ unsafe fn test_simd() {
     test_mm_insert_epi16();
     test_mm_shuffle_epi8();
 
+    #[cfg(not(jit))]
+    test_mm_cmpestri();
+
     test_mm256_shuffle_epi8();
     test_mm256_permute2x128_si256();
     test_mm256_permutevar8x32_epi32();
@@ -429,6 +432,31 @@ unsafe fn test_mm_shuffle_epi8() {
     assert_eq_m128i(r, expected);
 }
 
+// Currently one cannot `load` a &[u8] that is less than 16
+// in length. This makes loading strings less than 16 in length
+// a bit difficult. Rather than `load` and mutate the __m128i,
+// it is easier to memcpy the given string to a local slice with
+// length 16 and `load` the local slice.
+#[cfg(not(jit))]
+#[cfg(target_arch = "x86_64")]
+#[target_feature(enable = "sse4.2")]
+unsafe fn str_to_m128i(s: &[u8]) -> __m128i {
+    assert!(s.len() <= 16);
+    let slice = &mut [0u8; 16];
+    std::ptr::copy_nonoverlapping(s.as_ptr(), slice.as_mut_ptr(), s.len());
+    _mm_loadu_si128(slice.as_ptr() as *const _)
+}
+
+#[cfg(not(jit))]
+#[cfg(target_arch = "x86_64")]
+#[target_feature(enable = "sse4.2")]
+unsafe fn test_mm_cmpestri() {
+    let a = str_to_m128i(b"bar - garbage");
+    let b = str_to_m128i(b"foobar");
+    let i = _mm_cmpestri::<_SIDD_CMP_EQUAL_ORDERED>(a, 3, b, 6);
+    assert_eq!(3, i);
+}
+
 #[cfg(target_arch = "x86_64")]
 #[target_feature(enable = "avx2")]
 unsafe fn test_mm256_shuffle_epi8() {