Build 32-bit with 64-bit llvm-gcc

Question

I have a 64-bit version of llvm-gcc, but I want to be able to build both 32-bit and 64-bit binaries. Is there a flag for this? I tried passing -m32 (which works on the regular gcc), but I get an error message like this:

[jay@andesite]$ llvm-gcc -m32 test.c -o test
Warning: Generation of 64-bit code for a 32-bit processor requested.
Warning: 64-bit processors all have at least SSE2.
/tmp/cchzYo9t.s: Assembler messages:
/tmp/cchzYo9t.s:8: Error: bad register name `%rbp'
/tmp/cchzYo9t.s:9: Error: bad register name `%rsp'
...

This is backwards; I want to generate 32-bit code for a 64-bit processor!

I'm running llvm-gcc 4.2, the one that comes with Ubuntu 9.04 x86-64.

---

EDIT: Here is the relevant part of the output when I run llvm-gcc with the -v flag:

[jay@andesite]$ llvm-gcc -v -m32 test.c -o test.bc
Using built-in specs.
Target: x86_64-linux-gnu
Configured with: ../llvm-gcc4.2-2.2.source/configure --host=x86_64-linux-gnu --build=x86_64-linux-gnu --prefix=/usr/lib/llvm/gcc-4.2 --enable-languages=c,c++ --program-prefix=llvm- --enable-llvm=/usr/lib/llvm --enable-threads --disable-nls --disable-shared --disable-multilib --disable-bootstrap
Thread model: posix
gcc version 4.2.1 (Based on Apple Inc. build 5546) (LLVM build)
 /usr/lib/llvm/gcc-4.2/libexec/gcc/x86_64-linux-gnu/4.2.1/cc1 -quiet -v -imultilib . test.c -quiet -dumpbase test.c -m32 -mtune=generic -auxbase test -version -o /tmp/ccw6TZY6.s

I looked in /usr/lib/llvm/gcc-4.2/libexec/gcc hoping to find another binary, but the only directory there is x86_64-linux-gnu. I will probably look at compiling llvm-gcc from source with appropriate options next.

Lietuva

The error message you're getting is really odd; can you compile with -v and post the full argument list that llvm-gcc is actually using? – Duomo

Answer 1

Try setting:

export CFLAGS="-m32"
export LDFLAGS="-m32"

before compiling...

Housemother

llvm-gcc ignores both of these and builds a 64-bit version. – Lietuva

Have you installed these: sudo apt-get install gcc-multilib libc6-i386 lib6-dev-i386? – Housemother

All of those packages are already installed. It's like llvm-gcc is hardcoded to produce 64-bit output. – Lietuva

Answer 2

Could you try this series of commands and see if it works? Theoretically if you provided llvm-gcc with the -m32 option these steps should be taken by llvm-gcc, but maybe it's not working correctly, so let's make all the steps explicit:

llvm-gcc -m32 -emit-llvm test.c -c -o test.bc
llc test.bc -march=x86 -o test.S
gcc test.S -m32 -o test

This should be the sequence of steps (or something similar) that llvm-gcc performs implicitly, but it looks like in your case it's emitting 64bit assembly for some reason, then trying to assemble and link it for 32bit.

Martainn

This almost does what I want, but not quite. Although it does in fact produce a 32-bit binary, my test program says sizeof(void*) is 8, which is very worrisome. – Lietuva

Oops, you also have to pass an m32 to llvm gcc too. Try the new set of commands. – Martainn

That's the problem though. llvm-gcc is ignoring -m32. – Lietuva

the initial -m32 is ignored, so test.bc has 64 bit items. From that point on the tools work they just want solutions for 64 bit functions. The gcc -m32 at the end will complain about a library for some 64 bit function not being found, depending on the target and compiler/cross compiler – Quagga

Answer 3

I had the same problem, llvm-gcc ignores the flags, the only solution that worked was to switch from llvm-gcc to clang which does respect the -m32. That or switch to a 32 bit operating system for llvm-gcc work.

Quagga

Hmm, I will try out clang tomorrow. – Lietuva

I didnt use the cutting edge clang, or at least dont think I did as I had problems with it, I have been using whatever release_26 stands for. svn co llvm.org/svn/llvm-project/llvm/branches/release_26 llvm cd llvm cd tools svn co llvm.org/svn/llvm-project/cfe/branches/release_26 clang I basically started here: clang.llvm.org/get_started.html – Quagga

hmmm, formmatting is hosed in the last comment, sorry – Quagga