[WIP]

.github/workflows/main.yml

@@ -16,7 +16,7 @@ env:
   CARGO_BUILD_INCREMENTAL: false
   # Rust's CI denies warnings. Deny them here too to ensure subtree syncs don't
   # fail because of warnings.
-  RUSTFLAGS: "-Dwarnings"
+  #RUSTFLAGS: "-Dwarnings"
 
 jobs:
   rustfmt:

example/std_example.rs

@@ -259,6 +259,8 @@ unsafe fn test_simd() {
     test_mm_insert_epi16();
     test_mm_shuffle_epi8();
 
+    test_mm_cmpestri();
+
     test_mm256_shuffle_epi8();
     test_mm256_permute2x128_si256();
     test_mm256_permutevar8x32_epi32();
@@ -430,6 +432,29 @@ 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(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];
+    ptr::copy_nonoverlapping(s.as_ptr(), slice.as_mut_ptr(), s.len());
+    _mm_loadu_si128(slice.as_ptr() as *const _)
+}
+
+#[cfg(target_arch = "x86_64")]
+#[simd_test(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() {

src/intrinsics/llvm_x86.rs

@@ -757,6 +757,8 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             let b = b.load_scalar(fx);
             let lb = lb.load_scalar(fx);
 
+            panic!();
+
             let imm8 =
                 if let Some(imm8) = crate::constant::mir_operand_get_const_val(fx, &args[4].node) {
                     imm8