[LLVMdev] enabling Debian x86_64 for llvm 1.7

[Chris Lattner]

On Mon, 10 Jul 2006, Al Stone wrote:
> In trying to package up LLVM for Debian, it appears that x86_64 is no
> longer a supported architecture -- so, my first question is, is that
> correct?

Yes and no :(.  It's not that it is unsupported, it is just that noone 
has tested it and made it work.

> Best I can tell, the only thing that's supposed to work for
> x86_64 is the C backend.

Right.  This isn't something new unfortunately, the X86 backend doesn't 
support x86-64 yet.

> For Debian, I need to build everything from scratch.  When trying to
> build llvm-gcc4 from source, though, I get part way through the build
> and am told that x86_64-unknown-linux-pc is an unsupported architecture.
> What I have tried so far is this quick and dirty patch to
> gcc/Makefile.in:

This makes sense.

> Yes, it is a hack, but it does let me get further.  I can get to this
> point in compiling the C front-end:
> Writer.cpp:(.text+0x379): undefined reference to
> `llvm::FindUsedTypes::stub()'
> /home/ahs3/tmp/llvm-1.7/llvm/Release/lib/LLVMCBackend.o: In function

Okay, there are two problems with this.

1. llvm-gcc4, as of llvm-1.7, hardcoded in a list of libraries to link in,
    which apparently don't satisfy the dependencies needed by the C
    backend.  This is easily solvable.  If you use current llvm-gcc4 from
    svn, available here: http://llvm.org/docs/CFEBuildInstrs.html, a
    similar patch to the makefile will work better.  The makefile will
    automatically infer that an extra lib is needed and link it in.

2. More seriously, this will cause 'llvm-gcc x.c -S -o x.s' to put C code
    into x.s, which will cause the assembler to barf (e.g. when -c is
    used).  OTOH, "llvm-gcc -emit-llvm x.c -S -o x.ll" will work.
    Unfortunately, the former is used by the GCC build process.

I don't know if the workarounds for #2 are really worth it at this point.

> My first question is: should I even attempt to get things working
> for x86_64, or has it been tried and abandoned?  The second question
> is, is this even a reasonable approach?  (I have to claim an apparent
> lack of detailed understanding of the build system, at this point).
> I'm more than willing to believe I've started going about this all
> wrong...

My only suggestion is that it might be worth trying to map x86_64 to the 
X86 backend.  It would "fix" #2 (x86 assembly would end up in .s files, 
though it probably won't work right because it is 32-bit).  Some minor 
hacks might be needed in the x86 backend to accept the x86_64 target 
triple, but that would probably be easier.

We're currently in crunch mode right now, but I'd like to build a proper 
x86-64 backend for LLVM one day when time permits.  Perhaps in a couple of 
months I'll have more time to work on it.  In the meantime, it would be 
quite reasonable to use llvm-gcc3 for the llvm1.7 debian package.

Thanks Al!

-Chris

-- 
http://nondot.org/sabre/
http://llvm.org/