The problem is that llvm-mc always compiles "jmp" this as short jump, no matter where the target is. Hence my question. I dont know if there is any way to change this behavior. Looks like a bug to me so far.
It isn't. It's just created some assembly which, when assembled, may end up as a near jump or a short jump.
Craig said there is a trick with object file, but I dont know how to do that. Also, why compiling to object file changes the result?
Here's an object file (although I don't know why it's producing ELF rather than Mach-o):
$ echo "jmp _label"|llvm-mc-mp-3.7 -assemble -triple=i386 -filetype=obj > a.o
$ x86_64-pc-elf-objdump -d a.o
a.o: file format elf32-i386
Disassembly of section .text:
0: e9 fc ff ff ff jmp 0x1
And here's assembly which is then run through the compiler.
$ echo "jmp _label"|llvm-mc-mp-3.7 -assemble -triple=i386 -show-encoding |tee b.s
jmp _label # encoding: [0xeb,A]
# fixup A - offset: 1, value: _label-1, kind: FK_PCRel_1
$ clang -c b.s
$ otool -tvj b.o
0000000000000000 e900000000 jmp 0x5
As you can see, both produced near jumps. (I assume the difference between the offsets has to do with a difference in relocations between ELF and Mach-o.)