[llvm-dev] [RFC] Add a new backend called LoongArch

[Renato Golin via llvm-dev]

Hi  Weining,

Welcome to the LLVM community!

I had a look at the documentation and your fork and the description you
gave matches what the community expects of new targets, so thanks for
making the effort to know what to do before proposing a new target!

Some more comments below...

On Thu, 16 Dec 2021 at 07:55, 陆伟宁 via llvm-dev <llvm-dev at lists.llvm.org>
wrote:

> 1. LoongArch intro
> LoongArch is a RISC style ISA which is independently designed by Loongson
> Technology in China. It is divided into two versions, the 32-bit version
> (LA32)
> and the 64-bit version (LA64). LA64 applications have application-level
> backward binary compatibility with LA32 applications. LoongArch is
> composed of
> a basic part (Loongson Base) and an expanded part. The expansion part
> includes
> Loongson Binary Translation (LBT), Loongson VirtualiZation (LVZ), Loongson
> SIMD
> EXtension (LSX) and Loongson Advanced SIMD EXtension(LASX).


That's an interesting target. It seems very sensible on its instruction
encoding, data sizes and alignment, etc.

I'm just curious about the binary translation unit. Is that to support
LA32/LA64 interchangeably (like x86_64 and AArch64 do with their 32-bit
counterparts), or is that for some other architecture (like MIPS or Arm)?

2. Conform to the policy
> According to
> https://llvm.org/docs/DeveloperPolicy.html#adding-a-new-target
> a) Of couse it will be an experimental target at first.
> b) I'd like to be the code owner of this target.
> c) There is an active community behind the target:
> https://github.com/loongson
>    And we will provide builbot support.


Sounds great!

d) Documentations:
>  - ISA:
>
> https://loongson.github.io/LoongArch-Documentation/LoongArch-Vol1-EN.html
>  - ABI:
>
> https://loongson.github.io/LoongArch-Documentation/LoongArch-ELF-ABI-EN.html
>  - More docs can be found at:
>    https://loongson.github.io/LoongArch-Documentation/README-EN.html


Some documents there (binary translation, vector extension) are TBD.

Given that the current llvm fork you have is only for basic support, I
imagine you'll tackle that after the basic support is merged. Before
bringing extensions, we'd need documents for those, too.

The one thing left is to know if there are existing implementations of the
hardware (as a product or a dev board) and/or if there are emulators freely
available.

This is important for people that want to test the back-end on your
hardware (for example, while debugging unrelated changes that break on your
target).

3. Status
> We started to implement an out-of-tree LoongArch LLVM port since last year
> based on llvm-8/11 and posted it to
> https://github.com/loongson/llvm-project
> last month. This port also adds supports to clang front-end and we finally
> pass
> 100% llvm-test-suite in O0/1/2/3 optimization levels.


This is actually really nice!

But in this port there
> are a few issues we must handle to get it upstreamed.
>  - The codebase is too old that means we may use some out-of-date
> interfaces.


Your fork is based on the tree as of Oct 2020, that's more than a year ago
and in LLVM's timeframes, an eternity.

That usually means you'll have to re-write a good portion of your existing
code, but I'm assuming you already knew that and is happy to proceed.

 - The test coverage is not broad enough.


Given that you pass the test-suite, I'm assuming you can generate a good
portion of the use cases.

But it is important to also have extensive LIT tests (IR-to-IR, MIR, DAG,
SRC-to-IR, etc) to avoid needing to run the test-suite for basic support
testing.

 - Coding standard is not met.


I'm sure you know this is a deal breaker. But since you're going to have to
re-write good part of the code, I'm also assuming you're happy making your
new code meet the standards. :)

The current status is that we have completed a series of 5 patches adding
> triple, ELF machine, basic interger instructions and registers definition.
> We
> will submit them for review later. Any comments are welcome and please do
> let
> me know if you'd like to be added as a reviewer to future patches.


This sounds like a good start. Let's get to lowering and parsing a function
with a few arguments, instructions and a return value, and their respective
tests.

Name those patches [X/N] with X being 1..N and N being the total number of
patches in the first series. They need to be reviewed all at the same time
and only when all are approved we can merge them all together.

This is important to make sure new targets start at the right place. From
them on, you won't need to number the patches, and can incrementally
develop your target to reach maturity.

So, overall, I think this looks promising. Your target seems to meet the
criteria we set for new targets, so looking forward to seeing the initial
patches!

Thanks!
Renato
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20211216/9f82ba5e/attachment.html>