Ensure the entrypoint is taken into account when computing the export table

[dra27]

Presently on trunk and - as far as I can tell - since OCaml 3.11, in Cygwin64:

$ make natruntop
OCaml version 5.4.0+dev0-2024-08-25 - native toplevel
Enter #help;; for help.

# #load "unix.cmxs";;
Error while loading otherlibs/unix/unix.cmxs: error loading shared library: Failure("Cannot resolve caml_debugger_in_use").

This is strange, since ocaml/ocaml@27780d9#diff-3c8eb806a6be5e7baa5af9546b7e12cdfa1220daae562235a450f87b93dd06e7 in OCaml 3.12 ensures that debugger.o is linked into OCaml programs precisely to prevent this problem. The symbol is indeed present:

$ nm ocamlnat.exe | fgrep caml_debugger_in_use
0000000100f89818 B caml_debugger_in_use

the issue is that it is not included in the flexlink export table:

$ PATH=opt/bin:$PATH ./boot/ocamlrun.exe ./ocamlopt.exe -nostdlib -I ./stdlib -I otherlibs/dynlink -I utils -I parsing -I typing -I bytecomp -I file_formats -I lambda -I middle_end -I middle_end/closure -I middle_end/flambda -I middle_end/flambda/base_types -I asmcomp -I driver -I toplevel -I tools -I runtime -I otherlibs/dynlink -I otherlibs/str -I otherlibs/systhreads -I otherlibs/unix -I otherlibs/runtime_events -g -linkall -I toplevel/native   -o ocamlnat.exe compilerlibs/ocamlcommon.cmxa compilerlibs/ocamloptcomp.cmxa compilerlibs/ocamlbytecomp.cmxa otherlibs/dynlink/dynlink.cmxa compilerlibs/ocamltoplevel.cmxa  toplevel/topstart.cmx -cclib -show-exports | fgrep caml_debugger_in_use

The reason is that while -lasmrun is passed to flexlink, the objects passed on the command line give flexlink no reason to link startup_nat.n.o, which would then cause debugger.n.o to be considered and consequently these symbols to be added to the export table.

The fundamental fix here is that when linking an executable, flexlink now considers the entry point when closing over the required list of objects. For Cygwin, this fix is quite simple as it suffices to consider linking main. Cygwin's gcc does link crt0.o, but the full mechanism is technically a bit more involved. This doesn't matter, though, because being a Posix system, the entrypoint is always main (not wmain or WinMain/wWinMain, etc.).

For both MSVC and mingw-w64, the fix is a little more complex as, being non-Posix, there's not a strict requirement for the entrypoint to be main. For mingw-w64, the entrypoint is mainCRTStartup, but whether that calls main or wmain is determined by -municode which selects between using crt2.o or crt2u.o as the startup object. MSVC is affected both by the /ENTRY and /SUBSYSTEM linker options, which select between one of mainCRTStartup, wmainCRTStartup, WinMainCRTStartup or wWinMainCRTStartup. MSVCRT.lib contains each of those symbols in a separate object each of which then has an import reference to either main, wmain, WinMain or wWinMain as required. Since OCaml 4.06, OCaml therefore uses -municode for mingw-w64 and /ENTRY:wMainCRTStartup for MSVC.

This leads to the main part of this PR - the entrypoint for mingw-w64 and MSVC is therefore defined not in one of the objects or libraries passed on the command line (i.e. in libasmrun.a), but in one of the default libraries (either crt2u.o or MSVCRT.lib). The problem is that while these libraries are scanned, they are not analysed when closing over the objects which need to be linked. This is correct (the default libraries are assumed to be being linked in both executable and DLL) - but the objects should be considered in case they import symbols which do come from the objects and libraries passed on the command line. The changes here therefore mean that for mingw-w64, mainCRTStartup gets resolved to crt2u.o. That module imports wmain which then causes main.n.o from libasmrun.a to be linked and, consequently startup_nat.n.o and hence debugger.n.o. This PR happens to be an alternate fix (or, rather, workaround) for ocaml/ocaml#13520.

[nojb]

I only glanced at the code, but LGTM

[nojb]

Isn't -Wl,-e rewritten into -link arguments (cf

flexdll/cmdline.ml

Lines 281 to 285 in ab8161f

	| s :: rest when String.length s >= 6 && String.sub s 0 5 = "-Wl,-" ->
	let args =
	String.split_on_char ',' (String.sub s 4 (String.length s - 4))
	in
	List.fold_right (fun arg args -> "-link" :: arg :: args) args (tr rest)

) ?

[dra27]

Urgh - good catch, and unfortunately it's because this part is completely broken. The problem here is that when we link MSVC we actually call the linker, but when we link mingw-w64, we call the C compiler! I added this bit here for completeness - for MSVC, we of course select Unicode with -link /entry:wmainCRTStartup and that works because flexlink passes the /entry: directly to /link. For mingw-w64, we of course do -link -municode and that survives being passed to the C compiler to do the linking, but -Wl,-e,foo won't work being passed to the C compiler as -e foo. I only added this handling for -Wl,-e because I just written the /entry: part for MSVC (which I knew was needed).

Hmm. Given that you can't specify the entry point at the moment on mingw-w64, I'm tempted to remove this bit of code and just add a comment that you can't change the entrypoint on mingw-w64?

[nojb]

Sounds reasonable!

[dra27]

Oh - coffee kicks in! I remember how this works - the -Wl, translation to -link is a shim to allow common options to pass through. You can specify the entry point on mingw-w64, it's just that you have to say -link -Wl,-e,entrypoint at which point the code above does work as intended!

[MisterDA]

This should be moved in the section above.