Slow Xorg performance on dual Opteron + Radeon, Jessie 64-bit

[Chris Edwards <cedwards.debian@cedwards.geek.nz>]

Hi all,

I've recently installed Debian Jessie 64-bit on my (admittedly rather 
old) dual Opteron workstation, and I'm experiencing pretty bad 
performance in X11. Certain redrawing operations are extremely slow, 
with delays of half a second or more, and Xorg consumes a lot of CPU 
time. I've wondered if it's particularly a problem for older X clients 
using bitmapped fonts, as it's very noticeable when running xosview and 
when dragging tabs and invoking menus in Notion (window manager), but 
menu drawing in GIMP is also very slow.  The machine has 8 GiB of RAM 
and only a couple are being used.  It was formerly running Ubuntu 12 LTS 
and didn't have this problem.

`Xorg -version` reports the following:

X.Org X Server 1.16.4
Release Date: 2014-12-20
X Protocol Version 11, Revision 0
Build Operating System: Linux 3.16.0-4-amd64 x86_64 Debian
Current Operating System: Linux zaphod 3.16.0-4-amd64 #1 SMP Debian 
3.16.7-ckt11-1 (2015-05-24) x86_64
Kernel command line: BOOT_IMAGE=/boot/vmlinuz-3.16.0-4-amd64 
root=UUID=2cd952af-5971-442e-9254-bb15730dec1a ro quiet
Build Date: 11 February 2015  12:32:02AM
xorg-server 2:1.16.4-1 (http://www.debian.org/support)
Current version of pixman: 0.32.6


My video card is a Radeon X800 XT Platinum Edition AGP, and Xorg is 
using the RADEON driver.  It's running in dual-head mode.


`perf top` seems to indicate that the C library's memcpy implementation 
is the main culprit:

  44.29%  libc-2.19.so                   [.] __memcpy_sse2_unaligned


Here's the disassembly of the same, also courtesy of `perf`:
<<
__memcpy_sse2_unaligned  /lib/x86_64-linux-gnu/libc-2.19.so 

...
       │       lea    0x30(%r10),%rax 
        ▒
       │       movdqu (%rcx,%r10,1),%xmm8 
        ▒
 10.42 │       movdqa %xmm8,(%rcx) 
        ▒
  5.03 │       movdqu (%rcx,%r9,1),%xmm8 
        ▒
 10.35 │       movdqa %xmm8,0x10(%rcx) 
        ▒
 10.60 │       movdqu (%rcx,%r8,1),%xmm8 
        ▒
 14.00 │       movdqa %xmm8,0x20(%rcx) 
        ▒
  7.87 │       movdqu (%rcx,%rax,1),%xmm8 
        ▒
 12.53 │       movdqa %xmm8,0x30(%rcx) 
        ▒
  7.57 │       add    $0x40,%rcx 
        ▒
       │       cmp    %rcx,%rdx 
        ▒
 10.08 │     ↓ jne    780 
        ▒
       │     ↓ jmpq   6de 
        ▒
       │       cmp    %rsi,%rdi
>>

Could it be a performance regression due to recent enhancements to GLIBC 
for newer CPUs?  Any thoughts on how I can test this further?

Thanks,
Chris