Re: [RFA] amd64 displaced stepping support

[Doug Evans <dje at google dot com>]

On Tue, Jan 27, 2009 at 9:27 AM, Daniel Jacobowitz <drow@false.org> wrote:
> On Mon, Jan 26, 2009 at 03:00:12PM -0800, Doug Evans wrote:
>> Using the disassembler to compute instruction lengths is awkward, I know.
>> It's needed in order to compute the address of rip-relative addressing.
>> The address is %rip + address-of-next-insn + displacement,
>> and the displacement is only 32 bits so it's not guaranteed to be enough
>> to cover the distance between the original instruction and its copy.
>> To compensate I compute an unused integer reg, set it to
>> %rip + address-of-next-insn, and rewrite the insn to use base+disp addressing.
>> I think the GNU tools need a general-purpose library of ISA-related tools.
>> Until then, I went with the disassembler.  The code is laid out such that
>> when a better implementation of computing insn lengths comes along, it
>> can be easily dropped in.
>
> IMO, "the disassembler" means a bit of GDB interface glue, and
> libopcodes.  Libopcodes is the obvious place for a library about
> opcodes.  It can export more information; there's an example of this
> at the very end of struct disassemble_info, though it probably needs
> more granularity.

I don't disagree that libopcodes is a reasonable place.

I'm assuming by "example of this" you're refering to
branch_delay_insns, target, target2, etc.  Right?  This is great
stuff, but it's in a struct named "disassemble_info". :-)  IWBN if
libopcodes could provide a more generic interface, and long term an
interface more suitable to general use (i.e. not restricted to the
confines of gdb/binutils releases).