Windows: release VirtualVector reservations only on FreeLibrary, not at process exit

[Copilot]

Since #773, ~VirtualVector releases all pagemap reservations during the C runtime's teardown. For an exe that is too early: fiber-local storage (FLS) callbacks run afterwards, inside ExitProcess (LdrShutdownProcess). That includes Rust's thread-local destructors, which then touch released memory and crash. This PR keeps the #773 behaviour (release when a DLL is unloaded while the process keeps running) and stops releasing at process exit, in a single build with no config flag.

Exit order on Windows

Module	C runtime statics destroyed	FLS callbacks / other DLLs' detach	RtlDllShutdownInProgress() in ~VirtualVector
exe	exit() → atexit list	later, in ExitProcess	FALSE
DLL, process exit	its DLL_PROCESS_DETACH	around the same time	TRUE
DLL, FreeLibrary	its DLL_PROCESS_DETACH	not affected	FALSE

The flag alone can't tell an exe apart from a DLL being unloaded by FreeLibrary, so the module type is checked as well.

Changes

• pal_windows.h
  • PALWindows::current_module_is_dll() reads IMAGE_FILE_DLL from the __ImageBase headers.
  • RtlDllShutdownInProgress is declared directly. MSDN documents it as exported by ntdll and not declared in SDK headers, so the caller must declare it.
  • ~VirtualVector returns early unless the module is a DLL and shutdown is not in progress:

    if (
      !snmalloc::PALWindows::current_module_is_dll() ||
      RtlDllShutdownInProgress())
      return;

  • Comments on init_seg and reservations are updated to describe this.
• Linking ntdll
  • MSVC/clang-cl: #pragma comment(lib, "ntdll.lib").
  • MinGW (which ignores the pragma): target_link_libraries(snmalloc INTERFACE ntdll).
  • snmalloc-sys/build.rs: rustc-link-lib=ntdll for both the MSVC and windows-gnu targets.
• Tests (Windows only, in TESTLIB_ONLY_TESTS)
  • func/fls_teardown reproduces the Rust pattern without Rust. It registers an FLS callback in an exe that touches and frees snmalloc memory at exit.
  • func/dll_teardown uses two helper DLLs:
    • alloc allocates and is unloaded with FreeLibrary; the test checks that the memory was released.
    • touch checks during DLL_PROCESS_DETACH at process exit that the allocation is still readable. If not, it calls TerminateProcess(1), because the loader swallows faults raised in DllMain.
  • Both tests fail without the early return.

For reviewers

• SDK import library: confirm that the Windows SDK ntdll.lib exports RtlDllShutdownInProgress for both x86 and x64.
• MinGW pins the DLL: MinGW keeps DLLs that use thread_local destructors loaded. The FreeLibrary release check in dll_teardown is skipped when the DLL is still loaded; confirm that under MSVC it actually runs.
• Behaviour change: executables no longer return their reservations explicitly; the OS reclaims them at process termination.

[Matthew Parkinson (mjp41)]

Damoy (@DamoyY) would you be able to confirm that this PR fixes the issue for you.

[Damoy (DamoyY)]

Matthew Parkinson (@mjp41) Yes I can confirm that this PR fixes the bug I reported
Thanks for the fix

[Matthew Parkinson (mjp41)]

Neil Monday (@NeilMonday) Sorry, I should have got you to comment before I merged. Can you confirm your use case is not affected. Could you confirm it is okay please. I would like to make a release with this fix.