Performance Issue with Union

[danghvu]

I have a Union inside a Struct (with other primitive fields). The struct I implements Structure.ByReference and initialize it with a Pointer.

Inside the Union there is a few Structure with different size.

It works correctly when I do a read() on the outer Structure after setting the Union to the correct type. However the performance is 2-3x slower than when I process the Union separately.

By separately I mean I need to break the Union into normal Structure and dispatch them based on input information and init with a Pointer and an offset. The faster method however is not good since the offset in 32-bit machine and 64-bit machine is different.

So I guess the problem is the offset calculation internally of JNA. I am not sure if this is a known problem or I was doing it the wrong way. I tested with JNA 4.0.0 from the download link.

[twall]

Can you provide some specific code examples? Generally it’s pretty straightforward to perform a calculation that works properly on both 32- and 64-bit platforms.

On Nov 28, 2013, at 5:20 PM, Hoang-Vu Dang notifications@github.com wrote:

I have a Union inside a Struct (with other primitive fields). The struct I implements Structure.ByReference and initialize it with a Pointer.

Inside the Union there is a few Structure with different size.

It works correctly when I do a read() on the outer Structure after setting the Union to the correct type. However the performance is 3x - 4x slower than when I fixed the offset to Union (find out by sizeof() the C struct) and process the Union separately.

The faster method however is not good since the offset in 32-bit machine and 64-bit machine is different.

So I guess the problem is the offset calculation internally of JNA. I am not sure if this is a known problem or I was doing it the wrong way. I'm using JNA 3.5.1.

—
Reply to this email directly or view it on GitHub.

[danghvu]

This is the fine that I am maintaining here. The line is where I have to manually perform the trick to get the size.

I am not sure if you can make sense of the file since it's highly hacky in order to get a useful performance. Basically I have to separate the Union part from _cs_insn and manually dispatch them.

The reference struct

[twall]

Have you run the problematic code under a profiler? That should indicate pretty rapidly where JNA is spending its time, and whether it's doing redundant work or work that could be optimized.

[twall]

I would recommend changing your _cs_insn constructor to the following:

public _cs_insn(Pointer p) {
    super(p);
    // TODO: set union type here as appropriate, *then* perform a read
    read();
}

[twall]

Slight correction.

Add a Pointer-based constructor to _cs_detail:

public _cs_detail(Pointer p) {
    super(p);
    // TODO: initialize `arch` union field's type, *then* perform a read
    read();
}