initialize: keep the previous dispatch state when a policy throws - #95

(Written by Claude Code, on behalf of @jll63.)

Fixes #81.

write_global_data() patched every class' static_vptr, every overrider's next and each method's slots and strides to point into a local vector, then ran the policies' initialize, and only then swapped the vector into the registry's state. A policy that threw - fast_perfect_hash failing to find hash factors under throw_error_handler, or a plain bad_alloc - left all of those pointers dangling into a vector that unwinding had just freed, and clobbered the dispatch state of the previous initialize. Re-initialize after dlopen/dlclose is the documented flow, so this is reachable.

Change

write_global_data() is now stage → policies → commit, with an RAII guard and no try/catch:

• Stage: the dispatch data is built in a local vector, and each class' v-table pointer is staged in its class_ instead of being written to *ci->static_vptr.
• Policies run inside detail::registry_state_transaction<Registry>, which copies the registry state's policies tuple whole on construction and moves it back on destruction unless commit() was called. The class and method lists need no saving (initialize only reads them), and dispatch_data must not be restored from a copy: the previous static_vptrs point into the live buffer, which is why it is only ever replaced at commit.
• Commit, on a path where nothing can throw: slots and strides (and their propagation to other modules' copies), next pointers, static_vptr, then the dispatch_data swap.

A failed initialize() leaves the previous tables in place, complete and consistent, though not marked initialized: they do not reflect the registrations that prompted the call, and after a dlclose they may point into unloaded code, so require_initialized() keeps refusing to dispatch until a call succeeds.

The InitializeClass blueprint changes accordingly: vptr() returns the staged pointer by value, and a new static_vptr() returns the address that will receive it, for indirect_vptr policies. vptr_vector and vptr_map use it where they took &iter->vptr(); a user policy doing that now gets a compile error instead of a silently dangling pointer.

Test

test/test_initialize_transaction.cpp: two registries (vptr_vector; vptr_map + indirect_vptr) and a policy, placed after the vptr policy, whose initialize writes to its state then throws on demand. Asserts that the dispatch buffer, static_vptr<Dog/Cat>, next<poke_dog>, the hash range, the vptr policy's state and the throwing policy's own state are unchanged; that dispatch throws not_initialized; that a later initialize() succeeds; and that a failing first initialize() leaves everything null/empty and finalize() copes. Against the previous headers it fails 13 checks.

Verified with CMake (gcc, Debug: 148/148) and b2 (gcc: 87/87); docs rendered (Antora + MrDocs).

🤖 Generated with Claude Code

https://claude.ai/code/session_012QA4PCHE4oooe1ZtG9aKj7