Re: AltiVec register ptrace support

[Daniel Jacobowitz <drow@mvista.com>]

On Fri, Dec 07, 2001 at 03:23:02PM -0700, Kevin Buettner wrote:
> On Dec 7,  2:57pm, Kumar Gala wrote:
>
> > I have two different patches to the ptrace mechanism to add support
> > for AltiVec registers.
> >
> > linux-2.4.8-altivec-ptrace.patch:  Adds support similar to existing
> > mechanisms to get/set registers via PEEK/POKE calls extending the FPU
> > model.
> >
> > linux-2.4.16-altivec-ptrace.patch: Adds support for new ptrace commands
> > that match sparc/x86 PTRACE_{GET,SET}*REGS.  These dump the full register
> > state in a single call.
> >
> > Personally, I would like to see the PTRACE_{GET,SET}*REGS method adopted
> > for 2.4.x.  RedHat is trying to push out some GDB changes for AltiVec that
> > require closure on this matter.
>
> I would like to better understand your reasons for preferring
> PTRACE_{GET,SET}*REGS.  Is it just because that's what x86 does
> or do you think that this mechanism improves GDB's performance?

I think that it improves performance and that it is generally cleaner.

> My personal opinion is that GETREGS/SETREGS does not greatly enhance
> performance.  Try running strace on gdb debugging itself on x86 and on
> PPC and compare the number of PTRACE_PEEKUSR calls on PPC vs.
> PTRACE_????  calls on x86.  (The ????  is printed because strace
> doesn't know about the various PTRACE_{GET,SET}*REGS calls.) When I
> tried it just a moment ago using gdb to debug itself and running to a
> breakpoint set on main(), I saw _more_ PTRACE_???? calls on x86 than
> PEEKUSR/POKUSR calls on PPC.  Now, I admit that my testing wasn't very
> exhaustive, but even if the number of PEEKUSR/POKEUSR calls were
> higher, I think you'd find that calls to PEEKTEXT (for prologue
> analysis) would dominate.  I.e, the majority of the ptrace() traffic
> is due to reading memory, not reading registers.

You get more because there are three sets, and we gratuitously fetch
all registers instead of just the needed type of register.  I'd bet a
lot that a third of the 18 ????'s I see are for SSE registers and a
third for FP registers.  That would bring it down to 6 vs the 16 on PPC
using PEEKUSER.

Also, while I think _GETREGS is better than PEEKUSER, we're talking
here specifically about VRREGS.  It's four ptrace calls per vector
register, since ptrace() can only transfer a word at a time (so far at
least; I'm contemplating proposing a change to that).  And when you
want one vector register the odds are very good that one wants to get
another.

Also, while single stepping there ought to be no PEEKTEXT calls, only
PEEKUSER, and at least two of them on PPC (in fact we do a lot of
gratuitous poking around in the text segment).

> Furthermore, I think that introducing GETREGS/SETREGS will make
> ppc-linux-nat.c (in the GDB sources) more complicated.  We'll need
> compile time tests to check for the presence of GETREGS/SETREGS and
> use these mechanisms if they exist.  If they don't, this code will
> have to fall back to using the old PEEKUSR/POKEUSR mechanism.  Also,
> it may be necessary to have runtime tests which attempt to use
> GETREGS/SETREGS and fall back to using PEEKUSR/POKEUSR.  In order to
> see just how messy it can get, take a look at i386-linux-nat.c.

This part is definitely true.

--
Daniel Jacobowitz                           Carnegie Mellon University
MontaVista Software                         Debian GNU/Linux Developer

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/