jextract for operating system API

[Sundararajan Athijegannathan]

Thanks for your experiments and comments. I'll go over and file bug(s) as needed.

-Sundar
________________________________
From: jextract-dev <jextract-dev-retn at openjdk.org> on behalf of Manuel Bleichenbacher <manuel.bleichenbacher at gmail.com>
Sent: 23 August 2022 02:24
To: jextract-dev at openjdk.org <jextract-dev at openjdk.org>
Subject: jextract for operating system API

Hi everybody

As promised, I've compiled my findings when using jextract for the Java Does USB project, which uses the foreign function & memory API to communicate with USB devices on Windows, macOS and Linux. After initial difficulties with jextract, I've written all the layouts, method and variable handles manually. But now I've given jextract another try. The commands I've used can be found in in the "jextract" branch of the project at https://github.com/manuelbl/JavaDoesUSB/tree/jextract/java-does-usb/jextract

Please note that I think that automatically generating code for some of the old Windows and POSIX APIs is almost impossible. My use of jextract (for OS APIs) is probably far from the main use with third-party libraries. And I might also have made mistakes, or there might be a simple solution to my issues. See my conclusions below for areas where I think jextract has still potential.


Linux:

Overall, I was able to generate most of what I needed with jextract. The issues were:

- Generating code for libsystemd (sd-device.h) failed with "unknown type name 'intmax_t'". I think this should be defined by inttypes.h, which is included. My guess is that clang headers and Linux headers are mixed up.
- When generating code for usbdevice_fs.h, it didn't generate the constants for USBDEVFS_CLAIMINTERFACE and similar ones. The reason might be that the defines look like functions even though they evaluate to a constant value.


macOS:

Major parts of macOS are provided in the form of frameworks and they are treated in a special way, both when it comes to the header files and dynamic libraries. As an example, my library uses several functions from the CoreFoundation framework. In C/CC++, you can simply include them like so:

#include <CoreFoundation/CoreFoundation.h>

That's surprising as the file actually resides in /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/CoreFoundation.framework/Versions/A/Headers/CoreFoundation.h.

Note that it is in a directory called "Headers" and not in "CoreFoundation" as the include statement suggests. The relative path "CoreFoundation/CoreFoundation.h" does not exist at all.

This wouldn't be a big problem if this was just the entry point. But this pattern is used recursively. CoreFoundation.h includes about 30 other header files like this, all with a relative path that does not exist.

Xcode magically resolves this. But jextract seems to be missing this magic. So it cannot be used for any macOS frameworks. In other words, it cannot be for the major part of the macOS APIs. My library uses a single non-framework functions. So using jextract for macOS is pointless without resolving this issue.


Windows:

On Windows I spent the most time -- with no success at all. I only partially understand the problems. Here is what I've encountered:

The first simple goal was to have code for CloseHandle() generated. I tried it three different ways:

  *   Generating code for handleapi.h generated 1000 classes. Code for CloseHandle() was generated in 2 classes, but none of them was publicly available.
  *   Generating code for handleapi.h restricted to the function CloseHandle() generated a small number of class with a public static method for CloseHandle(). However, at run-time the method produces the error "java.lang.UnsatisfiedLinkError: unresolved symbol: CloseHandle"
  *   Generating code for windows.h restricted to CloseHandle() also resulted in the UnsatisfiedLinkError.

I then tried to generate code for Winusb.h. Unfortunately, this header file is not independent. It expects that you have already processed other header files before it. So it fails with "error: unknown type name 'LARGE_INTEGER'". This behavior with be reproduced in Visual Studio.

Next up was Cfgmgr32.h. It also doesn't seem to be independent and fails with "error: unknown type name 'ULONG'". Setupapi.h and Usbioctl.h behave similarly.

Windows was the only system where I had problems with preprocessor macros. The macro _M_AMD64=100 is certainly needed. That got me further than without it. And I've defined two Unicode macros, copied from the Visual Studio project. But the macros might still be the source of some problems.



Conclusion:

At the moment, jextract doesn't help much for accessing OS APIs.

On Linux, the "unknown type name 'intmax_t'" is worth investigating. It could be helpful for many APIs.

On macOS, it's all about the frameworks and the magic behind resolving the include path. If it can be made to work, it hopefully opens up the world of framework (plus some challenges that are still hidden).

On Windows, more investigation is needed. Work on the required macros is needed, as well as more experiments with header files just built for code generation (to get beyond the problem of header files that are not independent).


Regards
Manuel

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/jextract-dev/attachments/20220823/9d289547/attachment.htm>