hello.ll

Here I am demonstrating several ways to implement Hello World program for OS X in LLVM Language.

1. Link with C standard library to get at least write() and exit() functions. That's the simplest way but adds unnecessary libc dependency. It's just boring =)
2. Implement your own minimal standard library (e.g. in assembly) and ask ld to link with it. Here we can fully bypass libc but system calls can't be inlined or otherwise optimized by LLVM.
3. Implement your own minimal standard library in LLVM Language and link with it using llvm-link. This is the most interesting way resulting in the tightest machine code. System calls could become a target to full-blown LLVM optimization and ld invocation becomes trivial.

How to play on your own

• Build LLVM and put everything to vendor/llvm directory (llc should be accessible via vendor/llvm/bin/llc)
• For the second option you also need nasm (could be installed via homebrew)
• Run code/build.osx.sh

 USAGE: ./code/build.osx.sh <arch> <mode>
 
 ARCHITECTURES:
   x86
   x86_64
 
 MODES:
   clean      Remove all build files
   cstd       Build using C stdlib (libSystem.dylib)
   minstd     Build using minimal stdlib (minstd.osx.<arch>.asm)
   llstd      Build using minimal stdlib defined in LLVM bitcode
              providing inlining possibility (minstd.osx.<arch>.ll)

• Run build/hello.out

 Hello World!

• Examine build/hello.out. E.g.

 $ otool -tv build/hello.out
 build/hello.out:
 (__TEXT,__text) section
 _start:
 0000000100000f80	leaq	0x1d(%rip), %rsi
 0000000100000f87	movl	$0x1, %edi
 0000000100000f8c	movl	$0xe, %edx
 0000000100000f91	movl	$0x2000004, %eax        ## imm = 0x2000004
 0000000100000f96	syscall
 0000000100000f98	movl	$0x2a, %edi
 0000000100000f9d	movl	$0x2000001, %eax        ## imm = 0x2000001
 0000000100000fa2	syscall