diff --git a/ci/dox.sh b/ci/dox.sh
index 9803f7e371..d65192688f 100755
--- a/ci/dox.sh
+++ b/ci/dox.sh
@@ -22,6 +22,7 @@ dox() {
     export CARGO_UNSTABLE_BUILD_STD=core
     # amdgpu needs a target-cpu, any is fine
     export RUSTFLAGS="${RUSTFLAGS} -Ctarget-cpu=gfx900"
+    export RUSTDOCFLAGS="${RUSTDOCFLAGS} -Ctarget-cpu=gfx900"
   fi
 
   cargo build --verbose --target "${1}" --manifest-path crates/core_arch/Cargo.toml
diff --git a/ci/run.sh b/ci/run.sh
index ea012b42f9..8c75df6afb 100755
--- a/ci/run.sh
+++ b/ci/run.sh
@@ -82,7 +82,7 @@ cargo_test() {
             else
               dir="debug"
             fi
-            export CARGO_TARGET_WASM32_WASIP1_RUNNER="wasmtime -Wexceptions --dir /checkout/target/wasm32-wasip1/$dir/deps::."
+            export CARGO_TARGET_WASM32_WASIP1_RUNNER="wasmtime -Wexceptions --dir /checkout/target/wasm32-wasip1/$dir/build::."
             cmd="$cmd --nocapture"
             ;;
     esac
diff --git a/crates/core_arch/src/aarch64/sve/mod.rs b/crates/core_arch/src/aarch64/sve/mod.rs
index f11ca660b1..41fd12d351 100644
--- a/crates/core_arch/src/aarch64/sve/mod.rs
+++ b/crates/core_arch/src/aarch64/sve/mod.rs
@@ -39,7 +39,7 @@ impl<T> SveInto<T> for T {
 macro_rules! impl_sve_type {
     ($(($v:vis, $elem_type:ty, $name:ident, $elt:literal))*) => ($(
         #[doc = concat!("Scalable vector of type ", stringify!($elem_type))]
-        #[derive(Clone, Copy, Debug)]
+        #[derive(Clone, Copy)]
         #[rustc_scalable_vector($elt)]
         #[unstable(feature = "stdarch_aarch64_sve", issue = "145052")]
         $v struct $name($elem_type);
@@ -52,21 +52,21 @@ macro_rules! impl_sve_tuple_type {
     )*);
     (@ ($v:vis, $vec_type:ty, 2, $name:ident)) => (
         #[doc = concat!("Two-element tuple of scalable vectors of type ", stringify!($vec_type))]
-        #[derive(Clone, Copy, Debug)]
+        #[derive(Clone, Copy)]
         #[rustc_scalable_vector]
         #[unstable(feature = "stdarch_aarch64_sve", issue = "145052")]
         $v struct $name($vec_type, $vec_type);
     );
     (@ ($v:vis, $vec_type:ty, 3, $name:ident)) => (
         #[doc = concat!("Three-element tuple of scalable vectors of type ", stringify!($vec_type))]
-        #[derive(Clone, Copy, Debug)]
+        #[derive(Clone, Copy)]
         #[rustc_scalable_vector]
         #[unstable(feature = "stdarch_aarch64_sve", issue = "145052")]
         $v struct $name($vec_type, $vec_type, $vec_type);
     );
     (@ ($v:vis, $vec_type:ty, 4, $name:ident)) => (
         #[doc = concat!("Four-element tuple of scalable vectors of type ", stringify!($vec_type))]
-        #[derive(Clone, Copy, Debug)]
+        #[derive(Clone, Copy)]
         #[rustc_scalable_vector]
         #[unstable(feature = "stdarch_aarch64_sve", issue = "145052")]
         $v struct $name($vec_type, $vec_type, $vec_type, $vec_type);
diff --git a/crates/core_arch/src/amdgpu/mod.rs b/crates/core_arch/src/amdgpu/mod.rs
index 374f582696..264240c505 100644
--- a/crates/core_arch/src/amdgpu/mod.rs
+++ b/crates/core_arch/src/amdgpu/mod.rs
@@ -360,6 +360,7 @@ pub unsafe fn sched_barrier<const MASK: u32>() {
 /// sched_group_barrier::<8, 5, 0>()
 /// ```
 ///
+#[doc(cfg(target_arch = "amdgpu"))]
 #[doc = include_str!("intrinsic_is_convergent.md")]
 #[inline]
 #[unstable(feature = "stdarch_amdgpu", issue = "149988")]
diff --git a/crates/core_arch/src/wasm32/simd128.rs b/crates/core_arch/src/wasm32/simd128.rs
index e1a3754965..b52702f44e 100644
--- a/crates/core_arch/src/wasm32/simd128.rs
+++ b/crates/core_arch/src/wasm32/simd128.rs
@@ -4323,7 +4323,7 @@ mod tests {
             let vec = i64x2_load_extend_i32x2(arr.as_ptr());
             compare_bytes(vec, i64x2(-1, 1));
             let vec = i64x2_load_extend_u32x2(arr.as_ptr() as *const u32);
-            compare_bytes(vec, i64x2(u32::max_value().into(), 1));
+            compare_bytes(vec, i64x2(u32::MAX.into(), 1));
         }
     }
 
diff --git a/crates/core_arch/src/x86/mod.rs b/crates/core_arch/src/x86/mod.rs
index fbf1002eab..fb6ecd3616 100644
--- a/crates/core_arch/src/x86/mod.rs
+++ b/crates/core_arch/src/x86/mod.rs
@@ -40,6 +40,7 @@ types! {
     /// # Examples
     ///
     /// ```
+    /// #![cfg(any(target_arch = "x86", target_arch = "x86_64"))]
     /// #[cfg(target_arch = "x86")]
     /// use std::arch::x86::*;
     /// #[cfg(target_arch = "x86_64")]
@@ -56,6 +57,7 @@ types! {
     /// # if is_x86_feature_detected!("sse2") { unsafe { foo() } }
     /// # }
     /// ```
+    #[doc(cfg(any(target_arch = "x86", target_arch = "x86_64")))]
     pub struct __m128i(2 x i64);
 
     /// 128-bit wide set of four `f32` types, x86-specific
@@ -83,6 +85,7 @@ types! {
     /// # Examples
     ///
     /// ```
+    /// #![cfg(any(target_arch = "x86", target_arch = "x86_64"))]
     /// #[cfg(target_arch = "x86")]
     /// use std::arch::x86::*;
     /// #[cfg(target_arch = "x86_64")]
@@ -99,6 +102,7 @@ types! {
     /// # if is_x86_feature_detected!("sse") { unsafe { foo() } }
     /// # }
     /// ```
+    #[doc(cfg(any(target_arch = "x86", target_arch = "x86_64")))]
     pub struct __m128(4 x f32);
 
     /// 128-bit wide set of two `f64` types, x86-specific
@@ -126,6 +130,7 @@ types! {
     /// # Examples
     ///
     /// ```
+    /// #![cfg(any(target_arch = "x86", target_arch = "x86_64"))]
     /// #[cfg(target_arch = "x86")]
     /// use std::arch::x86::*;
     /// #[cfg(target_arch = "x86_64")]
@@ -142,6 +147,7 @@ types! {
     /// # if is_x86_feature_detected!("sse2") { unsafe { foo() } }
     /// # }
     /// ```
+    #[doc(cfg(any(target_arch = "x86", target_arch = "x86_64")))]
     pub struct __m128d(2 x f64);
 
     /// 256-bit wide integer vector type, x86-specific
@@ -173,6 +179,7 @@ types! {
     /// # Examples
     ///
     /// ```
+    /// #![cfg(any(target_arch = "x86", target_arch = "x86_64"))]
     /// #[cfg(target_arch = "x86")]
     /// use std::arch::x86::*;
     /// #[cfg(target_arch = "x86_64")]
@@ -189,6 +196,7 @@ types! {
     /// # if is_x86_feature_detected!("avx") { unsafe { foo() } }
     /// # }
     /// ```
+    #[doc(cfg(any(target_arch = "x86", target_arch = "x86_64")))]
     pub struct __m256i(4 x i64);
 
     /// 256-bit wide set of eight `f32` types, x86-specific
@@ -216,6 +224,7 @@ types! {
     /// # Examples
     ///
     /// ```
+    /// #![cfg(any(target_arch = "x86", target_arch = "x86_64"))]
     /// #[cfg(target_arch = "x86")]
     /// use std::arch::x86::*;
     /// #[cfg(target_arch = "x86_64")]
@@ -232,6 +241,7 @@ types! {
     /// # if is_x86_feature_detected!("avx") { unsafe { foo() } }
     /// # }
     /// ```
+    #[doc(cfg(any(target_arch = "x86", target_arch = "x86_64")))]
     pub struct __m256(8 x f32);
 
     /// 256-bit wide set of four `f64` types, x86-specific
@@ -259,6 +269,7 @@ types! {
     /// # Examples
     ///
     /// ```
+    /// #![cfg(any(target_arch = "x86", target_arch = "x86_64"))]
     /// #[cfg(target_arch = "x86")]
     /// use std::arch::x86::*;
     /// #[cfg(target_arch = "x86_64")]
@@ -275,6 +286,7 @@ types! {
     /// # if is_x86_feature_detected!("avx") { unsafe { foo() } }
     /// # }
     /// ```
+    #[doc(cfg(any(target_arch = "x86", target_arch = "x86_64")))]
     pub struct __m256d(4 x f64);
 }
 
diff --git a/crates/core_arch/src/x86/sse2.rs b/crates/core_arch/src/x86/sse2.rs
index 5e8fe959c3..66ebec22e1 100644
--- a/crates/core_arch/src/x86/sse2.rs
+++ b/crates/core_arch/src/x86/sse2.rs
@@ -3332,7 +3332,7 @@ mod tests {
         core_arch::{simd::*, x86::*},
         hint::black_box,
     };
-    use std::{boxed, f32, f64, mem, ptr};
+    use std::{boxed, mem, ptr};
     use stdarch_test::simd_test;
 
     const NAN: f64 = f64::NAN;
diff --git a/crates/stdarch-test/src/wasm.rs b/crates/stdarch-test/src/wasm.rs
index bf411c1214..feadaa1215 100644
--- a/crates/stdarch-test/src/wasm.rs
+++ b/crates/stdarch-test/src/wasm.rs
@@ -2,6 +2,7 @@
 
 use crate::Function;
 use std::collections::HashSet;
+use std::path::Path;
 
 pub(crate) fn disassemble_myself() -> HashSet<Function> {
     // Use `std::env::args` to find the path to our executable. Assume the
@@ -11,6 +12,27 @@ pub(crate) fn disassemble_myself() -> HashSet<Function> {
     let me = std::env::args()
         .next()
         .expect("failed to find current wasm file");
+    let me = Path::new(&me);
+    let me = if me.exists() {
+        // Old build-dir layout
+        me.to_path_buf()
+    } else {
+        // Cargo's build-dir layout stores an artifact named
+        // `<crate>-<hash>.wasm` at `<crate>/<hash>/out/<crate>-<hash>.wasm`.
+        // The build directory is mounted as the WASI working directory, so
+        // reconstruct that guest-visible path from the executable name.
+        let file_name = me
+            .file_name()
+            .and_then(|name| name.to_str())
+            .expect("current wasm file has a file name");
+        let stem = file_name
+            .strip_suffix(".wasm")
+            .expect("current wasm file does not have a .wasm extension");
+        let (crate_name, hash) = stem
+            .rsplit_once('-')
+            .expect("current wasm file name does not contain an artifact hash");
+        Path::new(crate_name).join(hash).join("out").join(file_name)
+    };
     let output = wasmprinter::print_file(&me).unwrap();
 
     let mut ret: HashSet<Function> = HashSet::new();