fix: re-enable dune on older macos sdk's

[rgrinberg]

by disabling watch mode on such versions

Signed-off-by: Rudi Grinberg [email protected]

ps-id: dc8e96cc-61a9-4497-baac-0934b82efe6f

[rgrinberg]

@barracuda156 can you check if this works?

[barracuda156]

@barracuda156 can you check if this works?

Sure, gimme few min.

[barracuda156]

On a side note, and since the PR refers to SDKs, there is one more issue that has to be fixed for those: #6510 (large chunk of those functions in fsevents_stubs are 10.7+).

That one I can fix locally for myself, but it is perhaps desirable to fix it for everyone. I could have made a PR, but I hope that maybe someone can come up with a better fix than merely disabling the functions – like it was done for libuv: libuv/libuv@6602fca

[rgrinberg]

I tried making it work for < 10.7 but it seems to be difficult. There's no notification for file level notifications so it doesn't really work well for dune's needs.

[barracuda156]

I tried making it work for < 10.7 but it seems to be difficult. There's no notification for file level notifications so it doesn't really work well for dune's needs.

Original fix was a blank disabling: #5431
That would actually work for <10.7 (with obvious limitations), but this is perhaps a feasible and surely better option: https://en.wikipedia.org/wiki/Kqueue

[rgrinberg]

Kqueue isn't really better. The API is much lower level and incredibly inconvenient for watching entire trees recursively. We deal with a similar headache on inotify though, so it is feasible.

[barracuda156]

Kqueue isn't really better. The API is much lower level and incredibly inconvenient for watching entire trees recursively. We deal with a similar headache on inotify though, so it is feasible.

In Macports at least, ocaml-dune is a necessity to install a number of ports (only one build method can be supported per port), so even if some valuable functionality is not available on older OS, as long as it works to build packages, that is infinitely better than being just broken :)

(For a context, I am trying now to fix Stan for PPC and make the whole family of its ports available in Macports. Turned out, upstream uses OCaml interpreter to build models, and that interpreter depends upon a gazillion of tiny ports which all need Dune to build.)

[barracuda156]

@rgrinberg I have just now looked at the files changed preparing to build, and my assumption was wrong: you actually addressed the issue which I thought was pending.
However, this is broken: #6156
Unfortunately, I have no solution for that one, and while the failure looks trivial, I could not fix it (primarily because I cannot understand why it fails in the first place).

[rgrinberg]

Can you explain how to reproduce the error?

[barracuda156]

Can you explain how to reproduce the error?

I did not do anything special to get it. Whatever I tried to change, it persisted.

If versions are important:
ocamlc 4.14.0 (bytecode compiler)
ocaml-dune 3.6.0 / latest commit (same failure)
ocaml-csexp 1.5.1

This is not the case where I would tweak anything in an exotic manner :)

OCaml packages normally just build (including for Darwin PPC) with zero extra effort, the only thing I modify* as compared to Macports master, as far as OCaml goes, is updating to current versions (Macports is lazy or reluctant to update those).
I have tried at least three recent versions of ocaml-dune. None of those could build ocaml-dune-configurator (failing identically, unable to locate ocaml-csexp) , which I do require functional for a said reason (moving dozens of packages to another build system would be insane and cannot be portable to others).

I think Macports logic with these ports is simple and transparent. ocaml PG sets directories and flags, portfiles borrow those, everything works, aside of one instance.
As I mentioned, ocaml-csexp is not only installed de facto but also visible to external environment. Why ocaml-dune-configurator fails to find it is beyond me.

P. S. * Well, irrelevant to this discussion and packages in question, but for the sake of being precise, few OCaml packages do need a correct target being set manually in order ocamlopt not to be forced on PPC.

[barracuda156]

@rgrinberg Since I could not find a rational solution to the problem which needs to be solved somehow, I also rebuilt from scratch all three in question: OCaml itself, ocaml-dune, ocaml-csexp.

P. S. Since ocaml-findlib can see ocaml-csexp as installed without any issues, it is likely possible to manually re-write the whole build & install process without Dune, using ocamlc and ocamlbuild, but this is not something which I can sit and do without understanding two alien build systems well enough.

[barracuda156]

@rgrinberg Okay, I have fixed it for myself. Apparently sources are inconsistent: for w/e reason included csexp package is deleted by scripts, while the library is defined as local one. It was easier for me to fix usage of included csexp by restoring .opam file and patching few files accordingly. Now everything started working, finally.

I can test your SDK fix now too.

[barracuda156]

@rgrinberg Your patch works fine, the only thing, I believe, macro should be used after the header, as it is always done (otherwise you put the header inside the conditional which is determined by the header in question, that might work, but should not). I.e.:

--- src/fsevents/fsevents_stubs.c.orig	2022-11-14 22:11:03.000000000 +0800
+++ src/fsevents/fsevents_stubs.c	2022-11-21 12:46:25.000000000 +0800
@@ -7,8 +7,10 @@
 #include <caml/threads.h>
 
 #if defined(__APPLE__)
+#include <AvailabilityMacros.h>
+#endif
 
-#include <Availability.h>
+#if __MAC_OS_X_VERSION_MAX_ALLOWED >= 100700
 #include <CoreFoundation/CoreFoundation.h>
 #include <CoreServices/CoreServices.h>

(Wrapping is inconsequential, #endif can go to the end of file instead.)

P. S. Line 93 change does nothing or I miss something? It is inside a macro for new systems, so I cannot check it.

[rgrinberg]

@rgrinberg Your patch works fine, the only thing, I believe, macro should be used after the header, as it is always done (otherwise you put the header inside the conditional which is determined by the header in question, that might work, but should not). I.e.:

--- src/fsevents/fsevents_stubs.c.orig	2022-11-14 22:11:03.000000000 +0800
+++ src/fsevents/fsevents_stubs.c	2022-11-21 12:46:25.000000000 +0800
@@ -7,8 +7,10 @@
 #include <caml/threads.h>
 
 #if defined(__APPLE__)
+#include <AvailabilityMacros.h>
+#endif
 
-#include <Availability.h>
+#if __MAC_OS_X_VERSION_MAX_ALLOWED >= 100700
 #include <CoreFoundation/CoreFoundation.h>
 #include <CoreServices/CoreServices.h>

(Wrapping is inconsequential, #endif can go to the end of file instead.)

P. S. Line 93 change does nothing or I miss something? It is inside a macro for new systems, so I cannot check it.

Thanks. I applied your suggestion.