fix(cuda_std): stop dynamic_shared_mem overlapping static shared memory

crates/cuda_std/src/shared.rs

@@ -20,7 +20,7 @@ pub fn dynamic_shared_mem<T>() -> *mut T {
         #[allow(improper_ctypes)]
         // mangle it a bit to make sure nobody makes the same thing
         #[link_name = "_Zcuda_std_dyn_shared"]
-        static DYN_SHARED: ::core::cell::UnsafeCell<u128>;
+        static DYN_SHARED: ::core::cell::UnsafeCell<[u128; 0]>;
     }
 
     // SAFETY: extern statics is how dynamic shared mem is done in CUDA. This will turn into

tests/compiletests/ui/dis/dynamic_shared_mem.rs

@@ -0,0 +1,16 @@
+// build-pass
+// compile-flags: -Cllvm-args=--disassemble-globals --error-format=human -Cdebuginfo=0
+// normalize-stderr-test "sm_\d+" -> "sm_XX"
+
+// Regression test: the extern behind `dynamic_shared_mem` must be unsized (`[]`) in the PTX,
+// like nvcc's `extern __shared__` arrays. A sized one is laid out among the kernel's static
+// shared variables, so the dynamic region overlaps them.
+
+use cuda_std::{kernel, shared};
+
+#[kernel]
+pub unsafe fn dynamic_shared(out: *mut u32) {
+    unsafe {
+        *out = *shared::dynamic_shared_mem::<u32>();
+    }
+}

tests/compiletests/ui/dis/dynamic_shared_mem.stderr

@@ -0,0 +1,7 @@
+.version 8.7
+.target sm_XX
+.address_size 64
+.extern .shared .align 16 .b8 _Zcuda_std_dyn_shared[];
+.visible .entry dynamic_shared( { ... }
+
+