This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
Re: PATCH: Extend gdb remote protocol for AVX
> Date: Tue, 7 Oct 2008 12:21:19 -0700
> From: "H.J. Lu" <hjl.tools@gmail.com>
>
> On Mon, Oct 6, 2008 at 2:35 PM, Mark Kettenis <mark.kettenis@xs4all.nl> wrote:
> >> Date: Sun, 5 Oct 2008 07:37:13 -0700
> >> From: "H.J. Lu" <hjl.tools@gmail.com>
> >>
> >> On Sat, Oct 4, 2008 at 3:13 PM, Daniel Jacobowitz <dan@codesourcery.com> wrote:
> >> >> I have no objection to the changes you proposed for the remote
> >> >> protocol. But your diff also touches the core register stuff, and
> >> >> that needs a bit more thought to make sure we don't surprise our
> >> >> users. At that point, it may be easier to use the same model for the
> >> >> remote protocol, where you transfer the top 128 bits of the %ymm
> >> >> registers in addition to the %xmm registers. Adter all this is how
> >> >> the hardware does it too (xsave is just an extension of fxsave).
> >> >
> >> > One way would be to transfer the xmm registers and then the remaining
> >> > bits as unnamed registers; another, probably easier way would be to
> >> > use an architecture specification or an actual register description to
> >> > transfer just the ymm registers and let GDB know about that fact, so
> >> > it can synthesize the xmm registers.
> >> >
> >> > (I don't remember the original patch, that may be what you're talking
> >> > about already.)
> >> >
> >>
> >> My proposal transfers the whole 256bit registers. We can display
> >> xmm registers as the lower 128bit ymm registers if we can
> >> display al/ax/eax.
> >
> > That certainly is a possibility, but if it is the right choice depends
> > on quite a bit more things that just ease of implementation of the
> > remote protocol. An important thing to check is what register numbers
> > compilers (and GCC in particular) use for these registers. Are there
> > compilers that already implement support for these new AVX
> > instructions?
>
> Gcc 4.4 supports AVX. Gcc uses the same register number internally
> for xmm0 and ymm0. It has no problems with register allocator.
I guess that makes sense. Otherwise the register allocator might
think it can use xmm0 independently from ymm0.
> > We really should make sure the DWARF register number mapping in the
> > AMD64 psABI gets updated for these new registers. Given the way the
> > current mapping us defined for %stN and %mmN, it probably makes sense
> > to give %ymmN their own numbers.
> >
>
> The AVX x86-64 psABI extension is at
>
> http://gcc.gnu.org/ml/gcc/2008-06/msg00408.html
>
> The ia32 psABI extension is similar to SSE. The only difference is
> __m256 requires 32byte alignment, instead of 16btye.
And it doesn't mention the DWARF debug register mapping. That's sad.
> I removed AVX register map:
>
> http://sourceware.org/ml/binutils/2008-08/msg00010.html
>
> since xmm/ymm registers are all caller-saved and there are
> no xmm/ymm registers in unwind info.
That's a mistake. Unwind info for signal handlers should mention
xmm/ymm registers. Although I'm not sure how one would express the
xsave layout in DWARF.
> For debug info, we can tell ymm from xmm with data size.
But what should "info address" display for a variable stored in a %ymm
register? If we can't distinguish between the two, we'll probably end
up printing %xmm.