llvm dev - Jan 2013 - [LLVMdev] Initial thoughts on an LLVM backend for N-address generic assembly

On Thu, Jan 24, 2013 at 7:20 AM, Ahmed Bougacha <ahmed.bougacha at
gmail.com>wrote:
> On Thu, Jan 24, 2013 at 12:46 PM,  <nkavv at physics.auth.gr> wrote:
> > Hi all,
> >
> > i'm just starting out with LLVM (although i've been observing
its
> evolution
> > since that first release some years ago :)
> >
> > I would like to develop a backend for a generic assembly-like
language,
> > called NAC (N-Address Code). More info on NAC can be found here:
> > http://www.nkavvadias.com/hercules/nac-refman.html (HTML)
> > http://www.nkavvadias.com/hercules/nac-refman.pdf (PDF)
> >
> > You can consider NAC similar to an LLVM subset for hardware synthesis.
> > Although NAC was developed independently, certain decisions taken when
> > designing it, may or may not defer from other textual IRs like LLVM or
> PTX.
> >
> > I have a number of questions:
> >
> > 1) It seems to me that a template backend either the C backend or
NVPTX
> are
> > possible starting points. However, only NVPTX makes use of tblgen
which
> > removes some burder from the backend developer. I also think that a
> > tblgen-based backend like NVPTX is easier to maintain in the long
term.
> What
> > do you think?
>
> Hi Nikolaos,
>
> The C backend (that is by the way deprecated) is the exception rather
> than the rule: most targets use TableGen extensively (look at all the
> lib/Target/**/*.td).
>
It really depends on your goals and the abstractions provided by the target
IR.  For NVPTX, we use a traditional LLVM back-end approach (using TableGen
and the SelectionDAG/MachineInstr/MC infrastructure) because PTX is
low-level enough to be treated as an assembly language.  There are some
higher-level features available in the language, but we do not use those.
If your IR is more like LLVM IR, then it may make sense to do some kind of
direct translation from LLVM IR to your IR and bypass the traditional
codegen approach.

Another point to consider is how much optimization you want.  For PTX, we
have to walk a fine line compared to other back-ends because PTX is not the
machine ISA.  A separate assembler is used to convert to the real machine
ISA, and that assembler performs optimizations of its own!  We want to
optimize the PTX we generate, but we also do not want to prevent
optimizations that may be performed by the assembler.  If your IR is meant
for optimization passes, then it may not make sense to perform all of the
back-end optimizations available through LLVM, but just use the middle-end
optimizations performed on LLVM IR directly.

>
> As for your other questions, I'm no expert, but you might want to take
> a look at the available tutorials
> (http://llvm.org/docs/WritingAnLLVMBackend.html and
> http://jonathan2251.github.com/lbd/).
>
> -- Ahmed Bougacha
>
> > 2) NAC uses the following statement form (for any given statement):
> >
> > dst1, dst2, ..., dstm <= op src1, src2, ..., srcn;
> >
> > which expresses an operation op with n source operands and m
destination
> > operands. Do you think that tblgen supports such form or should I
> sanitize
> > it?
> >
> > 3) The NAC memory model uses separate address spaces per array. A
> > general-use stack/heap might also be supported. Should I use dot
> directives
> > to declare each address space in use for a given translation
unit/module?
> >
> > 4) What is the maximum that I can get from tblgen? Which C++ source
files
> > cannot be generated by .td files and always have to be coded by hand?
> >
> > In the end, source code for the LLVM->NAC backend and a NAC
interpreter
> will
> > be released, probably as third-party BSD-licensed tools.
> >
> >
> > Best regards
> > Nikolaos Kavvadias
> >
> >
> >
> > _______________________________________________
> > LLVM Developers mailing list
> > LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>


-- 

Thanks,

Justin Holewinski
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20130124/083b11df/attachment.html>

Hi Justin
> It really depends on your goals and the abstractions provided by the target
> IR.  For NVPTX, we use a traditional LLVM back-end approach (using TableGen
> and the SelectionDAG/MachineInstr/MC infrastructure) because PTX is
> low-level enough to be treated as an assembly language.
I consider NAC (N-Address Code) to be comparably low-level to PTX  
(AFAICS from PTX ver. 3.1 PDF).
> If your IR is more like LLVM IR, then it may make sense to do some kind of
> direct translation from LLVM IR to your IR and bypass the traditional
> codegen approach.
It seems to be that NAC has more resemblance to PTX than LLVM. That,  
there is e.g. a PHI instruction that works in the same way as in LLVM.

I gave thoughts of writing an external translator, but i'm afraid that  
it would be obsoleted if some dramatic change happens to the LLVM  
infrastructure. A backend would be maintainable in the longer term.
> For PTX, we have to walk a fine line compared to other back-ends  
> because PTX is not the machine ISA.
Similarly, NAC is not a specific ISA but a generic, assembly-like,  
input to a high-level synthesis process, mostly involving CDFG  
mapping, operation scheduling, state assignments, resource allocation  
and RTL code generation. The result is a hardware-only  
(non-programmable) architecture that implements the input program.

Regarding machine-independent optimizations, I would basically use  
whatever is supported by LLVM's "opt". Target-dependent
optimizations,
do happen, but as part of the high-level synthesis backend.
> If your IR is meant
> for optimization passes, then it may not make sense to perform all of the
> back-end optimizations available through LLVM, but just use the middle-end
> optimizations performed on LLVM IR directly.
I agree, this is my intention. The IR is meant as an LLVM  
simplification and subset for hardware synthesis.

Best regards
Nikolaos Kavvadias

Hi Justin

I just came across your presentation regarding the LLVM 3.0 (?) PTX backend:
http://llvm.org/devmtg/2011-11/Holewinski_PTXBackend.pdf

Is this the NVPTX backend or the predecessor (in PTX? directory) backend?

Best regards
Nikolaos Kavvadias

On Thu, Jan 24, 2013 at 11:41 AM, <nkavv at physics.auth.gr> wrote:
> Hi Justin
>
> I just came across your presentation regarding the LLVM 3.0 (?) PTX
> backend:
>
http://llvm.org/devmtg/2011-**11/Holewinski_PTXBackend.pdf<http://llvm.org/devmtg/2011-11/Holewinski_PTXBackend.pdf>
>
> Is this the NVPTX backend or the predecessor (in PTX? directory) backend?
>
It's the predecessor (removed in 3.1), but most of the concepts are the
same.  The short version of the history is that I developed the original
PTX back-end with a few other people (PTX back-end in 3.0/3.1) and it was
later replaced by the NVIDIA version of the back-end around May of 2012.

The NVIDIA version is a much more complete version, as it is actually used
in production compilers as opposed to being a small university project. :)

>
> Best regards
> Nikolaos Kavvadias
>
>
>
>

-- 

Thanks,

Justin Holewinski
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20130124/8fe15ee0/attachment.html>