llvm dev - Mar 2009 - [LLVMdev] Adding Intrinsics for custom processor (frotend problem)

Chris Lattner wrote:
> As others have mentioned, hacking the front-end isn't that hard.  In  
> any case though, please be aware that a precompiled binary for a non- 
> MIPS target won't produce correct code if you are (for example) using  
> an x86 front-end and forcing llc to generate mips code with llc - 
> march=mips.  The front-end does type layout and knows very ABI things  
> that have to be right for mips.
>
> -Chris
>
>   
Thanks for your reply Chris.
I would still like to use the 'standard' precompiled binary for llvm-gcc
front end unless its really impossible. Actually I thought that the 
whole idea of LLVM was to have an abstract middle level so that front 
end does not have to know 'very ABI things' :-) . To make things clear, 
I am using the frontend only with -emit-llvm flag, and then run my llc, 
but I guess this is obvious, so I was wondering if you can expound on 
what kind of problems I may encounter.

Actually, the strategy of using only the precompiled binary front end 
worked for us satisfactorily in our previous processor iteration, a very 
non-standard stack machine architecture so I doubt we would have an 
insurmountable problem doing the same now, with the new little endian 
MIPS like architecture.

To put things in perspective: we only require a subset of "C" to be 
supported by the compiler. Our processor is in fact a low level DSP like 
sub-processor in the whole chip, which runs relatively small and simple 
programs (it has only 16K program memory after all), and all the 
existing code base is currently written by us in that subset. If 
unions/bitfields would not work , or passing structs by value, then 
fine, we'd have no problem officially not supporting this features. MIPS 
calling convention ABI, is also not relevant for us. We may use it if it 
works fine, but if not we might ditch it altogether for a simpler 
calling convention, and might even decide not to support a stack at all 
(Our current "C" subset does not allow recursion).

Anyways, as I noticed the MIPS backend, is not really finished and have 
some bugs still (I reported some to Bruno), but it is much easier work 
to have that "almost working" code as a starting point, than having to
write our backend from scratch ...

On Mar 4, 2009, at 1:03 AM, Gil Dogon wrote:
> Chris Lattner wrote:
>> As others have mentioned, hacking the front-end isn't that hard. 
In
>> any case though, please be aware that a precompiled binary for a non-
>> MIPS target won't produce correct code if you are (for example)
using
>> an x86 front-end and forcing llc to generate mips code with llc -
>> march=mips.  The front-end does type layout and knows very ABI things
>> that have to be right for mips.
>>
>> -Chris
>>
>>
> Thanks for your reply Chris.
> I would still like to use the 'standard' precompiled binary for
llvm-
> gcc
> front end unless its really impossible.
If you want 100% correctness, it is.
> Actually I thought that the
> whole idea of LLVM was to have an abstract middle level so that front
> end does not have to know 'very ABI things' :-) .
Yes it is.  LLVM continues to keep getting better in that respect, but  
at the end of the day, LLVM is defined to be independent of source  
language and ABIs are typically defined in terms of C.  There is a  
many to one mapping between C constructs and LLVM IR, so things that  
depending on the distinctions in C cannot currently be regurgitated by  
the code generator.
> Actually, the strategy of using only the precompiled binary front end
> worked for us satisfactorily in our previous processor iteration, a  
> very
> non-standard stack machine architecture so I doubt we would have an
> insurmountable problem doing the same now, with the new little endian
> MIPS like architecture.
If you don't care about the ABI (because you compile everything with  
llvm) and don't care about processor predefines (e.g. __i386__) then  
you can probably get a way with this.

-Chris