This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
Re: [rfc] Autoselect x86_64 or i386 based on the remote g packet size
- From: "Mark Kettenis" <mark dot kettenis at xs4all dot nl>
- To: gdb-patches at sourceware dot org
- Date: Thu, 9 Nov 2006 22:36:29 +0100 (CET)
- Subject: Re: [rfc] Autoselect x86_64 or i386 based on the remote g packet size
- References: <20061109205755.GA18755@nevyn.them.org>
> The operating principle is that a target is free to send back a "short"
> g packet, but some sizes are plausible and others are not. So I
> enhanced the i386 backend to register the sizes of likely register
> sets: core registers, core + i387, core + i387 + SSE. The Linux
> backend adds the %orig_eax and %orig_rax pseudo-"registers" too
> (which, if I were inventing them today, might be target objects like
> the sparc WCOOKIE, but are currently in the g packet).
>
> So if you have a GDB which defaults to i386-linux, and it connects to
> an amd64 target and gets exactly the right number of bytes to be
> a 64-bit register set, it'll assume that's what it's got.
>
> There's deliberately no mechanism to say "more than X bytes must be
> amd64",
> because I feel that's unsafe. The current registered guesses are
> high-confidence, since e.g the amd64 guesses would be quite odd sizes for
> an i386 target to return. For instance, the amd64 non-FP registers cut
> off between fioff and foseg if interpreted as a 32-bit register set: not
> very likely.
>
> All comments welcome! Tested on x86_64-pc-linux-gnu, using gdbserver for
> the testsuite and by hand.
Hi Daniel,
I can't say I'm very enthousiastic about this patch. It feels like a
kludge to avoid adding proper support to the remote protocol that allows
gdb to interrogate the target about this. A very clever kludge, but still
a kludge.
Mark