x86: correct overflow checking for 16-bit PC-relative relocs

The only insn requiring a truly 16-bit PC-relative relocation outside of
16-bit mode is XBEGIN (with an operand size override). For it, the
relocation generated should behave similar to 8- and (for 64-bit) 32-bit
PC-relatives ones, i.e. be checked for a signed value to fit the field.
This same mode is also correct for 16-bit code. Outside of 16-bit code,
branches with operand size overrides act in a truly PC-relative way only
when living in the low 32k of address space, as they truncate rIP to 16
bits. This can't be expressed by a PC-relative relocation.

Putting in place a new testcase, I'd like to note that the two existing
ones (pcrel16 and pcrel16abs) appear to be pretty pointless: They don't
expect any error despite supposedly checking for overflow, and in fact
there can't possibly be any error for the
- former since gas doesn't emit any relocation in the first place there,
- latter because the way the relocation gets expressed by gas doesn't
  allow the linker to notice the overflow; it should be detected by gas
  if at all, but see above (an error would be reported here for x86-64
  afaict, but this test doesn't get re-used there).

bfd/elf32-i386.c

@@ -93,7 +93,7 @@ static reloc_howto_type elf_howto_table[]=
   HOWTO(R_386_16, 0, 1, 16, false, 0, complain_overflow_bitfield,
 	bfd_elf_generic_reloc, "R_386_16",
 	true, 0xffff, 0xffff, false),
-  HOWTO(R_386_PC16, 0, 1, 16, true, 0, complain_overflow_bitfield,
+  HOWTO(R_386_PC16, 0, 1, 16, true, 0, complain_overflow_signed,
 	bfd_elf_generic_reloc, "R_386_PC16",
 	true, 0xffff, 0xffff, true),
   HOWTO(R_386_8, 0, 0, 8, false, 0, complain_overflow_bitfield,