Growing jextract
Manuel Bleichenbacher
manuel.bleichenbacher at gmail.com
Sun Feb 18 13:31:19 UTC 2024
I don't fully understand the changes made between version 21 and 22. But in
my case, it was working in version 21 and for version 22 I had to add about
8 additional structs, mainly on Windows.
Jextract would certainly be much easier to use if it would automatically
generate code for all dependencies. That's any option you haven't mentioned.
It's not easy to track down all dependencies. As an example: The Windows
data structure USB_NODE_CONNECTION_INFORMATION_EX has 5 dependencies. One
of the fields looks like so:
USB_DEVICE_DESCRIPTOR DeviceDescriptor;
What needs to be included additionally?
--include-struct USB_DEVICE_DESCRIPTOR
or:
--include-typedef USB_DEVICE_DESCRIPTOR
None of these. It's actually:
--include-struct _USB_DEVICE_DESCRIPTOR
I'm sure it's documented somewhere that jextract resolves typedefs and
works with the underlying type. Each of these little decisions might make
sense. But in combination, jextract is hard to use. Users would benefit if
it was easier to use. And this is a feature that could make a difference.
-- Manuel
Am So., 18. Feb. 2024 um 12:05 Uhr schrieb Maurizio Cimadamore <
maurizio.cimadamore at oracle.com>:
>
> On 18/02/2024 08:53, Manuel Bleichenbacher wrote:
>
> Hi Jorn
>
> Regarding the dependencies: Do I get it correctly? If a struct, e.g.
> epoll_event, includes another struct, e.g. epoll_data, I have to explicitly
> include both of them?
>
> If this is the case, it's a bit annoying but doable. It should probably be
> better documented (or did I miss the documentation of this fact)?
>
> Yes, that's the intended behavior. The alternatives are:
>
> 1. fail with error (as we try to do)
> 2. issue warning and generate bad code
> 3. issue warning and emit padding when the outer struct refers to the
> nested missing struct
> 4. do nothing (and generate bad code)
>
> Before the changes, jextract used to do (4). Which means "your mileage
> might vary". I don't think that's great, and we did get some (internal)
> feedback that this was confusing. That said, while looking at this, we
> considered both (3) and (1) - in the end we opted for (1) because, if the
> user wants filtering, there might be some value in making sure the
> filtering is set up correctly (e.g. are we 100% sure that users are signing
> up for bad code or padding - or maybe they just "missed" a dependency?)
>
> We don't feel overly strong about this (mostly a mild preference). Perhaps
> something that can be reassessed with more feedback (after all the other
> issues have been fleshed out).
>
> This behavior is documented in the README page:
>
> https://github.com/openjdk/jextract?tab=readme-ov-file#filtering-symbols
>
> (see the end of this section). And, as Jorn said, these changes are still
> not included in the binaries.
>
>
> Anyhow, it leads to the next issue (likely a bug or limitation of
> jextract) with the generated code for epoll_event and epoll_data (from
> <sys/epoll.h>). When the generated class epoll_event is accessed for the
> first time, it crashes with:
>
> Caused by: java.lang.IllegalArgumentException: Invalid alignment
> constraint for member layout: [a8(ptr):[*:b1]|i4(fd)|i4(u32)|j8(u64)](data)
> at
> java.base/jdk.internal.foreign.layout.StructLayoutImpl.of(StructLayoutImpl.java:49)
> at
> java.base/java.lang.foreign.MemoryLayout.lambda$structLayout$1(MemoryLayout.java:1063)
> at java.base/jdk.internal.foreign.Utils.wrapOverflow(Utils.java:277)
> at
> java.base/java.lang.foreign.MemoryLayout.structLayout(MemoryLayout.java:1062)
> at
> net.codecrete.usb/net.codecrete.usb.linux.gen.epoll.epoll_event.<clinit>(epoll_event.java:29)
>
> epoll_event is a tricky structure. It looks like so:
>
> typedef union epoll_data
> {
> void *ptr;
> int fd;
> uint32_t u32;
> uint64_t u64;
> } epoll_data_t;
>
> struct epoll_event
> {
> uint32_t events;
> epoll_data_t data;
> } __EPOLL_PACKED;
>
> Because it's packed, the fields data.ptr and data.u64 are not aligned.
>
> We used to deal with this fine, but the code for emitting struct/union
> layouts has changed a lot (as a result of attempting to refer to
> struct/union by name). It seems like a new bug w.r.t. alignment has crept
> in. I will take a look. The way it _should_ work is that we should get the
> offset info from libclang, and then add "withAlignment" accordingly, to
> make sure the struct layout can be built.
>
> Maurizio
>
>
> It can be made to work if epoll_data.ptr and epoll_data.u64 are set to a
> byte alignment of 4. But that's not actually correct. epoll_data is not a
> packed or unaligned struct. It's only unaligned if used within another
> packed struct.
>
> Regards
> Manuel
>
>
>
>
>
> Am So., 18. Feb. 2024 um 09:34 Uhr schrieb Manuel Bleichenbacher <
> manuel.bleichenbacher at gmail.com>:
>
>> Hi Jorn
>>
>> Thanks for the prompt response. Using the pre-built binaries, I've made
>> some progress on Linux but run into new issues. I will provide the feedback
>> one by one.
>>
>> So let's start with the macOS. Here's the output run with verbose.
>>
>> I also propose to suppress the unrelated warning that occurs on all
>> platforms. It's just annoying. And it will send newcomers in the wrong
>> direction when they actually have to investigate real warnings or errors.
>>
>> Regards
>> Manuel
>>
>>
>> WARNING: A restricted method in java.lang.foreign.AddressLayout has been
>> called
>> WARNING: java.lang.foreign.AddressLayout::withTargetLayout has been
>> called by org.openjdk.jextract.clang.libclang.Index_h in module
>> org.openjdk.jextract
>> WARNING: Use --enable-native-access=org.openjdk.jextract to avoid a
>> warning for callers in this module
>> WARNING: Restricted methods will be blocked in a future release unless
>> native access is enabled
>>
>> Cannot invoke "org.openjdk.jextract.Declaration$Constant.name()" because
>> "<parameter2>" is null
>> java.lang.NullPointerException: Cannot invoke
>> "org.openjdk.jextract.Declaration$Constant.name()" because "<parameter2>"
>> is null
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.impl.TreeMaker.enumConstantString
>> (TreeMaker.java:532)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.impl.TreeMaker.lambda$createEnum$7
>> (TreeMaker.java:375)
>> at java.base/java.util.ArrayList.forEach(ArrayList.java:1597)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.impl.TreeMaker.createEnum
>> (TreeMaker.java:373)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.impl.TreeMaker.createTreeInternal
>> (TreeMaker.java:133)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.impl.TreeMaker.createTree
>> (TreeMaker.java:119)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.impl.Parser.lambda$parse$2
>> (Parser.java:86)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.clang.Cursor$CursorChildren.lambda$forEach$1
>> (Cursor.java:261)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.clang.Cursor$CursorChildren$Context.visit
>> (Cursor.java:234)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.clang.Cursor$CursorChildren.lambda$static$0
>> (Cursor.java:252)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.clang.libclang.Index_h.clang_visitChildren
>> (Index_h.java:8275)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.clang.Cursor$CursorChildren.forEachShortCircuit
>> (Cursor.java:271)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.clang.Cursor$CursorChildren.forEach
>> (Cursor.java:260)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.clang.Cursor.forEach
>> (Cursor.java:201)
>> at org.openjdk.jextract at 22/org.openjdk.jextract.impl.Parser.parse
>> (Parser.java:64)
>> at
>> org.openjdk.jextract at 22/org.openjdk.jextract.JextractTool.parse
>> (JextractTool.java:119)
>> at org.openjdk.jextract at 22/org.openjdk.jextract.JextractTool.run
>> (JextractTool.java:516)
>> at org.openjdk.jextract at 22/org.openjdk.jextract.JextractTool.main
>> (JextractTool.java:218)
>>
>>
>> Am Sa., 17. Feb. 2024 um 20:46 Uhr schrieb Jorn Vernee <
>> jorn.vernee at oracle.com>:
>>
>>> One more thing I realized about the warnings you get for
>>> /usr/include/libudev.h: the structs about which you get warnings are not
>>> being included (looking at your script, there are no --include-struct for
>>> them), so there shouldn't be any need for us to warn about them either.
>>>
>>> We currently check whether things are included first, and then mark
>>> things that aren't as 'skipped'. However, in the later pass which checks
>>> for unsupported types, we also issue warnings for those skipped elements. I
>>> think we should just skip issuing warnings for things that are marked as
>>> skipped by the earlier pass.
>>>
>>> Jorn
>>> On 17/02/2024 17:59, Jorn Vernee wrote:
>>>
>>> Hey Manuel,
>>>
>>> Thanks for giving jextract 22 a try, and sorry for the frustrating
>>> experience.
>>>
>>> - The NullPointerException is definitely a bug. Jextract should not
>>> just report an exception. We'll have a closer look at this on Monday, and
>>> see if we can find what code causes the issue. It would help if you could
>>> edit the jextract launch script, and add
>>> `JLINK_VM_OPTIONS=-Djextract.debug=true` (`JLINK_VM_OPTIONS=` should
>>> already be there), and then retry the failed extraction. This should print
>>> out more information, including the entire stack trace. that's the part I'm
>>> interested in.
>>>
>>>
>>> > I don't understand the error message. What does "xxx depends on yyy
>>> which has been excluded" mean? Have I used it incorrectly? The error
>>> message mostly mentions types (and possibly functions) that I don't use,
>>> neither directly nor indirectly.
>>>
>>> It is meant to tell the user that they've included something that
>>> depends on something else which was not included. This would previously
>>> lead to uncompileable code, since FunctionDescriptors and MemoryLayouts now
>>> refer directly to the layouts defined in struct classes. For example, if I
>>> have a function like `void foo(struct A)`, then the FunctionDescriptor for
>>> `foo` will depend directly on the class we generate for `struct A`:
>>>
>>> FunctionDescriptor.ofVoid(
>>> A.layout()
>>> );
>>>
>>> If `A` itself is not included, this code would be uncompileable. So, in
>>> the latest version we (try to) issue an error about that.
>>>
>>> - For the 'ERROR's due to missing dependencies you are seeing: this also
>>> looks like a bug. We are currently issuing errors if a _skipped_ symbols
>>> depends on a skipped symbol as well, which makes no sense. We can just
>>> ignore missing dependencies of skipped things. I've tried to implement a
>>> quick fix here [1]. Though, you might be better off using the pre-built
>>> binaries for now, which don't have this dependency scanning behavior in the
>>> first place (i.e the builds at jdk.java.net/jextract)
>>>
>>>
>>> > And why are udev, udev_list_entry, udev_device etc. skipped? They used
>>> to work with jextract 21.
>>>
>>> - I took a look at libudev.h, and the warnings about unsupported types
>>> seem correct to me. e.g. `struct udev` does not have a definition in
>>> libudev.h [2], so jextract can not generate a class for it (since the
>>> layout and fields are not known). That is all the the warning is trying to
>>> say. The other structs about which you get a warning seems to be the same
>>> in that they don't have a definition in that file. (FWIW this is one of the
>>> areas we put a lot of effort into, so it makes sense that you didn't seen a
>>> warning in JDK 21). I think we need to clarify the error message to say
>>> more than "is not supported" though.
>>>
>>>
>>> > /usr/include/linux/usbdevice_fs.h: The code generated for the struct
>>> usbdevfs_urb does not compile (it refers to a
>>> type usbdevfs_iso_packet_desc, which is missing).
>>>
>>> This is actually the situation we were trying to detect with the ERRORs
>>> you encountered. There should have been an error about
>>> usbdevfs_iso_packet_desc being excluded, but it seems like we are not
>>> scanning the element types of arrays at the moment (linked PR tries to fix
>>> this as well).
>>>
>>>
>>> And finally, yeah: we need to update the README on the master branch to
>>> reference JDK 22 instead of 21.
>>>
>>> Thanks for reporting back! While we try to test as much as possible,
>>> some issues slip through the cracks, and are only found through user
>>> feedback like yours.
>>> Jorn
>>>
>>> [1]: https://github.com/openjdk/jextract/pull/217
>>> [2]:
>>> https://github.com/systemd/systemd/blob/main/src/libudev/libudev.h#L20C1-L20C13
>>> On 17/02/2024 15:40, Manuel Bleichenbacher wrote:
>>>
>>> Hi Maurizio
>>>
>>> I have tried to upgrade the JavaDoesUSB library (
>>> https://github.com/manuelbl/JavaDoesUSB) from JDK 21 to 22, and it was
>>> very frustrating. It failed on macOS, Windows nor Linux. And the problem is
>>> jextract.
>>>
>>> (The header files I'm trying the process are operating system header
>>> files.)
>>>
>>>
>>> On macOS, it always crashes:
>>>
>>> FATAL: Unexpected exception java.lang.NullPointerException: Cannot
>>> invoke "org.openjdk.jextract.Declaration$Constant.name()" because
>>> "<parameter2>" is null occurred
>>>
>>> I've tried different branches (master, panama, jdk22) and the pre-built
>>> binaries. They all behave the same except the error message is sometimes
>>> shorter and starts at "Cannot invoke..."
>>>
>>> The full commands can be found here (remove "--source"):
>>> https://github.com/manuelbl/JavaDoesUSB/blob/main/java-does-usb/jextract/macos/gen_macos.sh
>>>
>>>
>>> On Linux, jextract worked partially. But for several files, it refused
>>> to generate anything or generated code that didn't compile:
>>>
>>> /usr/include/fcntl.h:
>>> ERROR: stat depends on timespec which has been excluded
>>> ERROR: stat depends on timespec which has been excluded
>>> ERROR: stat depends on timespec which has been excluded
>>>
>>> /usr/include/libudev.h:
>>> ERROR: __pthread_list_t depends on __pthread_internal_list which has
>>> been excluded
>>> ERROR: __pthread_slist_t depends on __pthread_internal_slist which has
>>> been excluded
>>> ERROR: __pthread_mutex_s depends on __pthread_internal_list which has
>>> been excluded
>>> WARNING: Skipping va_list (type <error: struct __va_list_tag> is not
>>> supported)
>>> WARNING: Skipping __gnuc_va_list (type <error: struct __va_list_tag> is
>>> not supported)
>>> WARNING: Skipping udev (type Declared(udev) is not supported)
>>> WARNING: Skipping udev_list_entry (type Declared(udev_list_entry) is not
>>> supported)
>>> WARNING: Skipping udev_device (type Declared(udev_device) is not
>>> supported)
>>> WARNING: Skipping udev_monitor (type Declared(udev_monitor) is not
>>> supported)
>>> WARNING: Skipping udev_enumerate (type Declared(udev_enumerate) is not
>>> supported)
>>> WARNING: Skipping udev_queue (type Declared(udev_queue) is not supported)
>>> WARNING: Skipping udev_hwdb (type Declared(udev_hwdb) is not supported)
>>>
>>> sys/epoll.h:
>>> ERROR: __pthread_list_t depends on __pthread_internal_list which has
>>> been excluded
>>> ERROR: __pthread_slist_t depends on __pthread_internal_slist which has
>>> been excluded
>>> ERROR: __pthread_mutex_s depends on __pthread_internal_list which has
>>> been excluded
>>> ERROR: epoll_data_t depends on epoll_data which has been excluded
>>> ERROR: epoll_event depends on epoll_data which has been excluded
>>>
>>> /usr/include/linux/usbdevice_fs.h:
>>> The code generated for the structusbdevfs_urb does not compile (it
>>> refers to a type usbdevfs_iso_packet_desc, which is missing).
>>>
>>> The full commands for jextract can be found here. I've only removed
>>> "--source" as it is no longer needed.
>>>
>>> https://github.com/manuelbl/JavaDoesUSB/blob/main/java-does-usb/jextract/linux/gen_linux.sh
>>>
>>> All these files could successfully be processed with jextract 21.
>>>
>>> I don't understand the error message. What does "xxx depends on yyy
>>> which has been excluded" mean? Have I used it incorrectly? The error
>>> message mostly mentions types (and possibly functions) that I don't use,
>>> neither directly nor indirectly.
>>>
>>> And why are udev, udev_list_entry, udev_device etc. skipped? They used
>>> to work with jextract 21.
>>>
>>>
>>> On Windows, it completely failed as well. jextract outputs more than
>>> 5000 lines of errors and warnings. Here is just a random selection:
>>>
>>> ERROR: mbstate_t depends on _Mbstatet which has been excluded
>>> ERROR: LIST_ENTRY depends on _LIST_ENTRY which has been excluded
>>> ERROR: depends on _M128A which has been excluded
>>> ERROR: WIN32_FIND_STREAM_DATA depends on _WIN32_FIND_STREAM_DATA which
>>> has been excluded
>>> ERROR: tagEXTLOGFONTW depends on tagLOGFONTW which has been excluded
>>> ERROR: MULTIKEYHELPW depends on tagMULTIKEYHELPW which has been excluded
>>> WARNING: Skipping MEM_EXTENDED_PARAMETER.Reserved (bitfields are not
>>> supported)
>>> WARNING: Skipping Flags.Reserved (bitfields are not supported)
>>>
>>> The full command can be found here (remove "--source"):
>>> https://github.com/manuelbl/JavaDoesUSB/blob/main/java-does-usb/jextract/windows/gen_win.cmd
>>>
>>>
>>>
>>> The REAMDE on https://github.com/openjdk/jextract could also profit
>>> from an update:
>>>
>>> The command line "$ sh ./gradlew -Pjdk21_home=<jdk21_home_dir> ..." sent
>>> me on search for the branch with the JDK 22 code. But it's just the README
>>> that's outdated.
>>>
>>> And the instruction "Run the Gradle build with a Java version
>>> appropriate for the Gradle version. For example, Gradle 7.5.1 supports JDK
>>> 21." sent me down another rabbit hole. I have the latest Gradle version
>>> installed, which nicely runs with JDK 21. But it failed anyway. Turns out
>>> that the installed gradle is irrelevant but the gradle wrapper relevant
>>> (see command line above). And the gradle wrapper is at version 7.3.3. So
>>> there is nothing to choose really. You must use JDK 17 (or even earlier).
>>> It has never worked with JDK 21.
>>>
>>>
>>> I would appreciate some feedback. Are these bugs? Or has jextract become
>>> so much more restrictive?
>>>
>>> Regards
>>> Manuel
>>>
>>>
>>>
>>> Am Fr., 16. Feb. 2024 um 16:50 Uhr schrieb Maurizio Cimadamore <
>>> maurizio.cimadamore at oracle.com>:
>>>
>>>> Hi,
>>>> with JDK 22 near us, we have spent some quality time with jextract, to
>>>> make sure the code it generates is as good as it can be ahead of the
>>>> finalizaton of the FFM API. This resulted in several changes, both in
>>>> the implementation (so, invisible to jextract users) and in the
>>>> generated code, as we cleaned up the translation strategy to better
>>>> adhere with the core principles behind the jextract tool. These changes
>>>> are captured in details in this document:
>>>>
>>>> https://cr.openjdk.org/~mcimadamore/panama/jextract_changes.html
>>>>
>>>> It might be a good time to take the latest jextract for a spin (using
>>>> your favourite C library!) and report back, in case we missed anything.
>>>> You can find the latest sources in this branch:
>>>>
>>>> https://github.com/openjdk/jextract/tree/panama
>>>>
>>>> Binary snapshots of this newer version are also available here (note
>>>> that MacOS/Arm64 builds is also supported now):
>>>>
>>>> https://jdk.java.net/jextract/
>>>>
>>>> Cheers
>>>> Maurizio
>>>>
>>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/jextract-dev/attachments/20240218/65be49b9/attachment-0001.htm>
More information about the jextract-dev
mailing list