[LLVMdev] an OS around LLVM

[John Criswell]

On 12/17/13 8:03 AM, mindmachine at free.fr wrote:
> Hi all,
>
>
> If it's not the right place to ask, please forgive me.
>
> Currently I'm working on a new operating system concept, called "Om".
>
> The first feature would be Android-like apps, coming in *.opk files that would
> contain all needed resources and source-code expressed in LLVM-IR assembly
> language.
>
> http://sett.com/openminded-os/uid/88508
>
> How does it sound ?

I've seen several suggestions made on this thread on topics related to 
what you want to do.  To be honest, I don't really have a good idea what 
your goal is, and I didn't have sufficient patience to wade through your 
blog to find out.  As a suggestion for the future, you should describe 
your goal or goals more specifically in the email to the list.

What you need really depends on what your goals are.  If you want some 
sort of language in which application code is represented so that it is 
portable across difference architectures, then I think LLVM would be a 
good choice.  That said, there are several issues that LLVM doesn't deal 
with that you would need to address.  The biggest one is 
architecture-specific details that are exposed in source-languages like 
C/C++ and are often encoded in the TargetData information within an LLVM 
IR bitcode file.  Examples include pointer size, endianness, memory 
alignment requirements, etc.  Note also that you'll have to deal with 
how these details are exposed at the source language level.  Languages 
like C make these details visible, so the programmer can write 
non-portable code.  Languages like Java can keep these details hidden, 
allowing Java programs to be much more portable.

I believe the PNaCL work addresses some of these issues.

If you want *everything,* include the OS kernel, to be represented in 
LLVM IR, then you should take a look at the LLVA/SVA work.  In that 
work, we removed the inline assembly code in the OS and replaced it with 
high-level instructions that we (logically) added to the LLVM 
instruction set.  It still isn't portable in the way that PNaCL is, but 
it could probably be extended to do so.  You can find the SVA 
publications at http://sva.cs.illinois.edu/pubs.

As an aside, my (admittedly biased) opinion is that SVA is a very good 
platform for doing novel things with operating systems and compilers.  
We've used it to protect Linux and FreeBSD from memory safety errors 
using SAFECode-like techniques and simpler control-flow integrity 
approaches, and we'll have a paper out in March on using it to protect 
applications from compromised operating system kernels.  We haven't 
open-sourced the code yet, but if you're up for some low-level coding, 
you can write a new implementation of the instructions.

Finally, regarding TCG, I'm not sure that TCG will be that much better a 
fit than LLVM IR.  TCG doesn't have pointer sizes the way LLVM does, but 
all of your loads and stores will still have a size argument (8, 16, 32, 
64 bits).  It is not clear to me that a TCG program can be converted 
from working with 32-bit pointers to 64-bit pointers automatically.  
Also, I think LLVM provides many more optimizations than TCG and makes 
writing optimizations and code transformations easier.

Hope some of this helps,

-- John T.

>
>
> Julien
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev