Bridj problems passing double under armhf

[jonhare]

Hi Olivier,

I've had a few reports of strange issues with the OpenIMAJ webcam library not working properly on linux boards using the armhf ABI (@sarxos also has a load of issues with his webcam project where the openimaj library is failing under armhf - see e.g. sarxos/webcam-capture#244).

I spent the day debugging, and have narrowed the problem down to the passing of double values. For some reason, if you try to pass a double value as a function/method argument from the Java side to the native side, the value will not be correctly passed, and the stack will be corrupted (noticeable in particular through subsequent arguments not having the correct values either, before the ensuing SEGFAULT).

I've made a very simple test case here: http://users.ecs.soton.ac.uk/jsh2/testcase.tgz
The test case contains a native library (sources in src/main/cpp) with a single method that takes a single double and prints it; the java code just wraps the native lib and calls the function with the double value 101.101. What we'd expect to happen is for 101.101000 to be printed, but instead we get 0.0. I've tested under an older armel JVM (using bridj 0.6.3-SNAPSHOT), and the code works as expected, so this definitely appears to be a specific problem with the armhf abi.

I don't know enough about the internals of bridj, but I wonder if the problem is related to the use of the wrong registers for passing the values (see this thread https://groups.google.com/forum/#!topic/avian/8nz4JZEPwZs, which was what got me looking at the double value in the first place).

Let me know if there's anything else I can do to help debug.

Jon

[jonhare]

After more digging, I came across this in the dyncall manual: "It is important to note, that dyncall supports the ARM architecture calling convention variant {\bf with floating point hardware disabled} (meaning that the FPA and the VFP (scalar mode) procedure call standards are not supported)."

Which kind of implies that modern (full hardfp as opposed to soft and softfp) arm linux builds won't work with dyncall at present...

[ochafik]

Hey @jonhare , thanks for digging this out!
I'll try to reach out to the dyncall guys, but we can also try and patch something ourselves (see Debian's hard-float page for a start, and some general info about vfp registers, and a vfp programming guide)

[ochafik]

Quick update: it looks like dyncall does support hard-float, but... not dyncallback (the part of dyncall that allows creating dynamic callbacks, which BridJ uses heavily). I'm having a look at how hard it would be to adding that support :-)

[ochafik]

This issue was moved to nativelibs4java/BridJ#40