miri: implement some llvm.x86.sse.* intrinsics and add tests

src/tools/miri/src/shims/foreign_items.rs

@@ -928,6 +928,28 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
                 this.write_scalar(Scalar::from_f64(res), dest)?;
             }
 
+            "llvm.prefetch" => {
+                let [_, rw, loc, ty] =
+                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;
+
+                let rw = this.read_scalar(rw)?.to_i32()?;
+                let loc = this.read_scalar(loc)?.to_i32()?;
+                let ty = this.read_scalar(ty)?.to_i32()?;
+
+                if ty == 1 {
+                    // Data cache prefetch
+
+                    if !matches!(rw, 0 | 1) {
+                        throw_ub_format!("invalid `rw` value passed to `llvm.prefetch`: {}", rw);
+                    }
+                    if !matches!(loc, 0..=3) {
+                        throw_ub_format!("invalid `loc` value passed to `llvm.prefetch`: {}", loc);
+                    }
+                } else {
+                    throw_unsup_format!("unsupported `llvm.prefetch` type argument: {}", ty);
+                }
+            }
+
             // Architecture-specific shims
             "llvm.x86.addcarry.64" if this.tcx.sess.target.arch == "x86_64" => {
                 // Computes u8+u64+u64, returning tuple (u8,u64) comprising the output carry and truncated sum.
@@ -980,6 +1002,12 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
                 }
             }
 
+            name if name.starts_with("llvm.x86.sse.") => {
+                return shims::x86::sse::EvalContextExt::emulate_x86_sse_intrinsic(
+                    this, link_name, abi, args, dest,
+                );
+            }
+
             // Platform-specific shims
             _ =>
                 return match this.tcx.sess.target.os.as_ref() {

src/tools/miri/src/shims/intrinsics/mod.rs

@@ -120,6 +120,12 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
                 this.write_bytes_ptr(ptr, iter::repeat(val_byte).take(byte_count.bytes_usize()))?;
             }
 
+            "nontemporal_store" => {
+                let [place, dest] = check_arg_count(args)?;
+                let place = this.deref_operand(place)?;
+                this.copy_op(dest, &place.into(), /*allow_transmute*/ false)?;
+            }
+
             "ptr_mask" => {
                 let [ptr, mask] = check_arg_count(args)?;
 

src/tools/miri/src/shims/mod.rs

@@ -7,6 +7,7 @@ pub mod foreign_items;
 pub mod intrinsics;
 pub mod unix;
 pub mod windows;
+mod x86;
 
 pub mod dlsym;
 pub mod env;

src/tools/miri/src/shims/x86/mod.rs

@@ -0,0 +1 @@
+pub(super) mod sse;