[External] : Re: Re: Panama-FFI API Porting for RISCV64

[何伟凯]

According to my understanding, if we want to fill upper 32 bits of date before it is transfered into return buffer, we have to add a Binding in UnboxBindingCalculator by using Binding.Builder, then BindingSpecializer will generate corresponding operations.

However, there is no Binding can extend a float value to a 64 bits value and fill its upper 32 bits. So, we have to modify shared code.

Is there any misunderstanding?

To avoid modifying of shared code, we take another approach that fills all bits of return buffer with 1s while creating upcallStub. Hence, floating-point data can be copied from return buffer into register by fld and integer data by ld.

link to the pr: https://www.github.com/openjdk/jdk/pull/11004




WeiKai He







-----原始邮件-----
发件人:"Jorn Vernee" <jorn.vernee at oracle.com>
发送时间:2022-11-06 06:16:33 (星期日)
收件人: "何伟凯" <weikai at isrc.iscas.ac.cn>
抄送: panama-dev at openjdk.org
主题: Re: [External] : Re: Re: Panama-FFI API Porting for RISCV64


Reply in line...
On 01/11/2022 13:02, 何伟凯 wrote:

Thanks for your attention.

flw will alter the bits being copied and it will be invalidated if we copy a float data with fld. When flw load a 32-bits float value into a 64-bits register, it will fill upper 32 bits of the register, however, fld will not. Hence, using fld or flw depend on whether float or double is placed in the return buffer.
What I suggest is to write the right bit pattern for the whole register to the buffer in the Java code (including the upper 32 bits for floats), and then just copy the bits from the buffer to the register in the VM code. Either with fld or another instruction that doesn't modify the bits.

Our modification is based on openjdk/jdk repo, because we would like to merge our code into the main stream. We have noticed changes in the panama-foreign repo, we also think the new VMStorage class will very helpful for solving the question. When the change merge into main stream of openjdk/jdk, we will follow it up.


We are in the processes of moving all the changes in the panama-foreign repo to the openjdk/jdk repo for JDK 20. I haven't made the PR with most of the VM changes yet, but it will be based on the main PR for the JEP [1]. There are also some last-minute changes that will likely be added to the main JEP PR [2]. So, to avoid having to review things twice, I suggest basing the riscv port on the code in the panama-foreign repo for now, and then integrate into the mainline repo after the changes from the panama-foreign repo have been moved there.


We look forward to your further suggestions.


I can take a look at the code once the PR goes up.

Jorn

[1]: https://github.com/openjdk/jdk/pull/10872
[2]: https://github.com/openjdk/panama-foreign/pull/750


WeiKai He







Hello,

The interface between Java and native is in general designed to communicate in register values (since it's intended to work with any foreign language in theory). Primitive types and the return buffer are use to transport the register values. The idea would be for the Java code to write the bits with the right semantics into the return buffer, for the entire width of the register (64-bits?), and then the native code could just copy the bits from the return buffer into the right register, without altering them.

But, it sounds like this doesn't work for flw and fld? i.e. they both alter the bits being copied in an incompatible way? Or is the problem that 2 floats can 'share' a register that would normally be taken up by a single double?

Also, please note that the latest state of development is in the openjdk/panama-foreign repo on the foreign-memaccess+abi branch. It looks like the code you have is based on the openjdk/jdk repo. In the latest version in the panama-foreign repo, the use of VMReg is replaced by a new class called VMStorage (mirroring the Java class), and it is possible to attach a register mask to it which can also be used to indicate the width of the register, if needed [1].

Jorn

[1]: https://github.com/openjdk/panama-foreign/blob/foreign-memaccess%2Babi/src/hotspot/share/prims/vmstorageBase.inline.hpp#L43

On 30/10/2022 13:44, 何伟凯 wrote:

## Summary

In recent, Panama FFI-API has been a preview feature. In many scenarios, the Panama FFI API can replace the JNI to implement native function access. The FFI API provides more secure and convenient access to native functions. The specific implementation of FFI API is related to ARCH and OS. In order to enable RISC V64 to use FFI API, porting is required. 

## Notable Things

Because there are different return value passing convention for special structures like `struct {int, float}` on the RISCV64[1]. 

When making an upcall, that's calling a Java method from a native function, return value of the Java method will be saved in a segment of stack memory called return buffer, and then riscv backend will transfer the data in memory to `a0` and `fa0`. 

The instructions used for data transfer are closely related to correctness. If the field containing in the special structure menteioned above is float, must use 'flw', otherwise, 'fld'. This requires a means to pass the width information of fields to the riscv backend.

Unfortunately, according to my understanding, the current interface does not provide a direct means to pass the width information of struct fields to the riscv backend, so the width information is encoded in other ways. Although this makes the riscv porting slightly different from other arch, it does not make the code more difficult to understand and maintain. For details, see comments in the code[2].

## Testing

All the tests in jtreg have been passed. Some tests under the user mode of the QEMU may fail, however they can pass on the development board.

## Reference

[1] https://github.com/riscv-non-isa/riscv-elf-psabi-doc/blob/master/riscv-cc.adoc [2] https://github.com/feilongjiang/jdk/tree/riscv-foreign-api
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/panama-dev/attachments/20221122/92561173/attachment-0001.htm>