builtin: fix Boehm interior pointers and modal GC aborts on Windows (…

[quaesitor-scientiam]

…fix #28896)

V reaches array data through pointers into the middle of a Boehm block: a managed array points one header past the block start, and a slice points anywhere inside its owner's block. Boehm honours such pointers from the stack in any case, but from heap objects only when interior-pointer recognition is on, or at registered displacements.

Every libgc V compiles from source enables ALL_INTERIOR_POINTERS. The prebuilt libgc.a linked on Windows with tcc does not (GC_get_all_interior_pointers() returns 0), so a collection freed array blocks that heap objects still used:

• -gc boehm_leak defines GC_DEBUG, which also puts a debug header in front of every object. The one displacement V registered (+ sizeof(voidptr), for the array header) went through the plain GC_register_displacement, so the real offset was never registered. Collections dropped live lifetime bookkeeping arrays in builtin.closure, and the next free() of one aborted with "Invalid pointer passed to free()" in a modal message box, hanging closure_lifetime_api_test.v (Windows + tcc: -gc boehm_leak aborts with "Invalid pointer passed to free()" in closure.Lifetime.dispose, and closure_lifetime_api_test.v hangs on the GC error dialog #28896).
• In every Boehm mode, a slice kept only by a heap object lost its owner's block at the next collection while the slice was still in use.

Call GC_set_all_interior_pointers(1) before GC_INIT, unless a host already initialized the collector, so every build behaves like the ones V compiles from source. Also register the array-header displacement through the GC_REGISTER_DISPLACEMENT macro, which under GC_DEBUG maps to GC_debug_register_displacement and covers the debug header too; gc.h says the debugging variant must be used when debugging allocation is done. That keeps V's arrays safe when a host initialized the collector without interior pointers.

On Windows, Boehm reports a fatal error in a modal "Fatal error in GC" message box and waits for someone to dismiss it, so any fatal GC error hung a console program or a test run instead of failing it. Its other output (error details, leak reports) went to an <exe>.gc.log file, not stderr. Make Windows behave like Linux and macOS:

• Build the bundled gc.c with NO_MSGBOX_ON_ERROR and CONSOLE_LOG, so the libgc V compiles from source (gcc, clang, msvc) writes everything to stderr and shows no message box.
• The prebuilt libgc.a linked with tcc cannot be rebuilt from here, so also replace Boehm's abort handler on Windows. V's handler prints the message on stderr and still calls the default handler with a nil message for the rest of its work (skipping the at-exit leak collection, GC_LOOP_ON_ABORT); Boehm then ends the process. It is not installed when a host initialized the collector first.

[medvednikov]

Reviewed all five changed files at 47d84822482b6504edf75efd678f5cb508076eba, ignoring CI as requested. One finding is attached inline: the new abort handler suppresses its fatal diagnostic when GC_LOOP_ON_ABORT is enabled.

Validation: built this PR on Linux; all three new Boehm regression tests passed. Also ran closure_lifetime_api_test.v, which failed in test_closure_lifetime_freestanding_no_std_object_compile because the compiler rejects -freestanding; that command and test are unchanged by this PR. Windows execution was not available. The inline finding was reproduced separately with the same callback ordering in a small C harness linked against Boehm.

[medvednikov]

[P2] Print and flush the fatal message before chaining to the saved handler

When GC_LOOP_ON_ABORT is set, Boehm's GC_default_on_abort(NULL) still enters its infinite debugging loop (thirdparty/libgc/gc.c:27419–27429). Consequently this call never reaches the fprintf/fflush below, so the new Windows handler suppresses the fatal reason precisely when someone enables the collector's debugging mode. Boehm's original handler prints the message before looping. Please move the stderr output before the saved-handler call; this preserves both the diagnostic and the intentional debugging loop. A C harness using this ordering timed out with empty stderr, while the print-first ordering emitted the message before timing out.

[quaesitor-scientiam]

Good catch. Fixed in commit bbb65c2.

internal_gc_abort_to_stderr now prints and flushes the message first and only then chains to the saved handler with a nil message. The at-exit-collection flag and GC_LOOP_ON_ABORT behave as before, just after the diagnostic, which is the same order GC_default_on_abort uses.

I reproduced it on Windows before changing anything. With GC_LOOP_ON_ABORT=1 the aborting child kept spinning with an empty stderr when built with tcc (the prebuilt libgc.a). Built with gcc (gc.c from source), stderr only had Boehm's own Invalid pointer passed to free(): <pointer> log line; the handler's line was missing in both cases.

gc_boehm_abort_reports_to_stderr_test.v has a new test for this. It runs the child with GC_LOOP_ON_ABORT=1, waits (with a deadline) for the handler's bare message line, then kills the spinning child. It failed with both compilers before the change and passes with both after it. It also passes on Linux, where V leaves Boehm's handler in place.