Windows LLVM-MinGW build, portable helper assertions and optional diagnostic UI.
Apply after upstream d4r patches 0002 through 0007, on ZLUDA ee2f25a.

diff --git a/ext/highs-sys/build.rs b/ext/highs-sys/build.rs
--- a/ext/highs-sys/build.rs
+++ b/ext/highs-sys/build.rs
@@ -66,11 +66,12 @@
     let linux = target.contains("linux");
     let mingw = target.contains("pc-windows-gnu");
     if apple {
         println!("cargo:rustc-link-lib=dylib=c++");
     } else if linux || mingw {
-        println!("cargo:rustc-link-lib=dylib=stdc++");
+        let cxx_library = env::var("CXXSTDLIB").unwrap_or_else(|_| "stdc++".to_owned());
+        println!("cargo:rustc-link-lib=dylib={cxx_library}");
     }
     println!("cargo:rerun-if-changed=../HiGHS/highs/interfaces/highs_c_api.h");
 
     true
 }
diff --git a/ext/hip_runtime-sys/build.rs b/ext/hip_runtime-sys/build.rs
--- a/ext/hip_runtime-sys/build.rs
+++ b/ext/hip_runtime-sys/build.rs
@@ -1,8 +1,12 @@
 fn main() {
     if cfg!(windows) {
-        println!("cargo:rustc-link-arg=/DELAYLOAD:amdhip64_7.dll");
+        if std::env::var("CARGO_CFG_TARGET_ENV").as_deref() == Ok("gnu") {
+            println!("cargo:rustc-link-arg=-Wl,--delayload=amdhip64_7.dll");
+        } else {
+            println!("cargo:rustc-link-arg=/DELAYLOAD:amdhip64_7.dll");
+        }
     } else {
         println!("cargo:rustc-link-lib=dylib=amdhip64");
         println!("cargo:rustc-link-search=native=/opt/rocm/lib/");
     }
 }
diff --git a/llvm_zluda/build.rs b/llvm_zluda/build.rs
--- a/llvm_zluda/build.rs
+++ b/llvm_zluda/build.rs
@@ -19,10 +19,18 @@
 
 fn main() {
     let bc_path1 = "src/device-libs/ockl.bc";
     println!("cargo:rerun-if-changed={}", bc_path1);
     check_lfs_file(bc_path1);
+
+    println!("cargo:rerun-if-env-changed=ZLUDA_LLVM_BUILD_DIR");
+    if let Some(directory) = env::var_os("ZLUDA_LLVM_BUILD_DIR") {
+        let llvm_build_path = PathBuf::from(directory);
+        println!("cargo:build-path={}", llvm_build_path.display());
+        link_prebuilt_llvm(&llvm_build_path);
+        return;
+    }
 
     let mut cmake = Config::new(r"../ext/llvm-project/llvm");
     try_use_sccache(&mut cmake);
     try_use_ninja(&mut cmake);
     cmake
@@ -57,16 +65,31 @@
         cmake.build();
     }
     let cmake_profile = cmake.get_profile();
     let llvm_build_path = get_llvm_build_path(&llvm_dir, cmake_profile);
     println!("cargo:build-path={}", llvm_build_path.display());
+    link_prebuilt_llvm(&llvm_build_path);
+}
+
+fn link_prebuilt_llvm(llvm_build_path: &PathBuf) {
     let (cxxflags, ldflags, libdir, lib_names, system_libs) =
         llvm_config(&llvm_build_path.join("bin/llvm-config")).unwrap();
     println!("cargo:rustc-link-arg={ldflags}");
     println!("cargo:rustc-link-search=native={libdir}");
     for lib in system_libs.split_ascii_whitespace() {
-        println!("cargo:rustc-link-arg={lib}");
+        if let Some(name) = lib.strip_prefix("-l") {
+            println!("cargo:rustc-link-lib={name}");
+        } else {
+            println!("cargo:rustc-link-arg={lib}");
+        }
+    }
+    // llvm-config omits system libraries on Windows. Propagate these through
+    // this rlib to its final cdylib consumer (link-arg alone cannot do that).
+    if env::var("CARGO_CFG_TARGET_OS").as_deref() == Ok("windows") {
+        for library in ["psapi", "shell32", "ole32", "uuid", "advapi32", "ws2_32", "ntdll"] {
+            println!("cargo:rustc-link-lib={library}");
+        }
     }
     link_lld_components();
     link_llvm_components(lib_names);
     compile_cxx_lib(cxxflags, &llvm_build_path);
 }
diff --git a/zluda/build.rs b/zluda/build.rs
--- a/zluda/build.rs
+++ b/zluda/build.rs
@@ -1,11 +1,15 @@
 use vergen_gix::{Emitter, Gix};
 
 fn main() {
     if cfg!(windows) {
-        println!("cargo:rustc-link-arg=delayimp.lib");
-        println!("cargo:rustc-link-arg=/DELAYLOAD:amdhip64_7.dll");
+        if std::env::var("CARGO_CFG_TARGET_ENV").as_deref() == Ok("gnu") {
+            println!("cargo:rustc-link-arg=-Wl,--delayload=amdhip64_7.dll");
+        } else {
+            println!("cargo:rustc-link-arg=delayimp.lib");
+            println!("cargo:rustc-link-arg=/DELAYLOAD:amdhip64_7.dll");
+        }
         let dll_path = "bin/nvcudart_hybrid64.dll";
         println!("cargo:rerun-if-changed={}", dll_path);
         check_lfs_file(dll_path);
     }
     let git = Gix::builder().sha(false).build();
diff --git a/zluda_windows/src/lib.rs b/zluda_windows/src/lib.rs
--- a/zluda_windows/src/lib.rs
+++ b/zluda_windows/src/lib.rs
@@ -11,27 +11,27 @@
 use uuid::uuid;
 use widestring::{u16str, U16Str, U16String};
 use windows::{
     core::{w, HRESULT, PCWSTR},
     Win32::{
-        Foundation::{HANDLE, HINSTANCE, HMODULE, HWND, LPARAM, S_OK, WPARAM},
+        Foundation::{FreeLibrary, HANDLE, HINSTANCE, HMODULE, HWND, LPARAM, S_OK, WPARAM},
         System::{
             JobObjects::{
                 AssignProcessToJobObject, JobObjectExtendedLimitInformation,
                 SetInformationJobObject, JOBOBJECT_EXTENDED_LIMIT_INFORMATION,
                 JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE,
             },
             LibraryLoader::{
-                GetModuleFileNameA, GetModuleFileNameW, GetModuleHandleExW, LoadLibraryExW,
+                GetModuleFileNameA, GetModuleFileNameW, GetModuleHandleExW, GetProcAddress, LoadLibraryExW,
                 GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS,
                 GET_MODULE_HANDLE_EX_FLAG_UNCHANGED_REFCOUNT, LOAD_LIBRARY_SEARCH_DEFAULT_DIRS,
-                LOAD_LIBRARY_SEARCH_DLL_LOAD_DIR,
+                LOAD_LIBRARY_SEARCH_DLL_LOAD_DIR, LOAD_LIBRARY_SEARCH_SYSTEM32,
             },
         },
         UI::{
             Controls::{
-                TaskDialogIndirect, TASKDIALOGCONFIG, TASKDIALOGCONFIG_0, TASKDIALOGCONFIG_1,
+                TASKDIALOGCONFIG, TASKDIALOGCONFIG_0, TASKDIALOGCONFIG_1,
                 TASKDIALOG_NOTIFICATIONS, TDCBF_CLOSE_BUTTON, TDF_ENABLE_HYPERLINKS,
                 TDN_HYPERLINK_CLICKED, TD_ERROR_ICON,
             },
             WindowsAndMessaging::IDCLOSE,
         },
@@ -487,11 +487,23 @@
             pszFooter: PCWSTR::null(),
             pfCallback: Some(task_dialog_callback),
             lpCallbackData: 0,
             cxWidth: 0,
         };
-        TaskDialogIndirect(&config, None, None, None).ok();
+        // A CUDA DLL may be loaded by an executable with no Common Controls
+        // v6 activation manifest. Do not make its diagnostic UI an eager DLL
+        // import: comctl32 v5 has no TaskDialogIndirect export.
+        if let Ok(controls) = LoadLibraryExW(w!("comctl32.dll"), None, LOAD_LIBRARY_SEARCH_SYSTEM32) {
+            if let Some(raw) = GetProcAddress(controls, windows::core::s!("TaskDialogIndirect")) {
+                let show: unsafe extern "system" fn(*const TASKDIALOGCONFIG, *mut i32, *mut i32, *mut i32) -> HRESULT =
+                    mem::transmute(raw);
+                let _ = show(&config, ptr::null_mut(), ptr::null_mut(), ptr::null_mut());
+            } else {
+                eprintln!("ZLUDA failed to load {}; set HIP_PATH to the selected HIP SDK", redirect_names[0]);
+            }
+            let _ = FreeLibrary(controls);
+        }
     }
     result
 }
 
 unsafe extern "system" fn task_dialog_callback(
diff --git a/ptx/lib/zluda_ptx_impl.cpp b/ptx/lib/zluda_ptx_impl.cpp
--- a/ptx/lib/zluda_ptx_impl.cpp
+++ b/ptx/lib/zluda_ptx_impl.cpp
@@ -405,11 +405,26 @@
                             uint32_t line,
                             uint64_t function,
                             uint64_t char_size)
     {
         (void)char_size;
+#if defined(_WIN32) || defined(_WIN64)
+        // HIP on Windows only supplies the wide _wassert entry point. Keep the
+        // CUDA assertion's ASCII diagnostics and trap through public OCKL APIs.
+        auto output = __ockl_fprintf_stderr_begin();
+        const char* parts[] = {"CUDA assertion: ", (const char*)message, " at ", (const char*)file,
+                               " in ", (const char*)function, "\n"};
+        for (const char* part : parts) {
+            const char* end = part;
+            while (*end) ++end;
+            output = __ockl_fprintf_append_string_n(output, part, end - part, 0);
+        }
+        __ockl_fprintf_append_args(output, 1, line, 0, 0, 0, 0, 0, 0, 1);
+        __builtin_trap();
+#else
         [[clang::noinline]] __assert_fail((const char *)message, (const char *)file, line, (const char *)function);
+#endif
     }
 
     // * Smallest denormal is 1.4 × 10^-45
     // * Smallest normal is ~1.175494351 × 10^(-38)
     // * Now, 1.175494351×10^-38 / 1.4 × 10^-45  = 8396388 + 31/140