So, I did some experiments locally.
I still can't see 10x or more slowdown, all I can observe is about 2x
the difference.
I can see that certain programs/binaries triggers this slowdown quite
a lot. For example, luatex, which is run like
/usr/bin/luatex -ini -jobname=luatex -progname=luatex luatex.ini
during the texlive init procedure.
But what is the most interesting here. When I turn pauth off, the
speed difference becomes non-existant.
For example, initing all tex formats for a minimal texlive-base install
takes:
bookworm trixie
pauth=impdef 0m13.049s 0m23.527s
pauth=off 0m12.616s 0m12.303s
(I set QEMU_CPU=max,pauth=off for the 2nd row; pauth=impdef is the
default)
So, it looks like, trixie toolchain started using pauth much more
aggressively. Actually, pauth was my very first suspect.
When installing glibc from trixie on a bookworm system and re-running
the test, pauth=impdef case time increases a little bit, to about 15s.
Now, we already had this discussion in qemu-devel ML, - should we enable
pauth and in which way. Obviously, a stronger variant of pauth is not
an option by default, because it becomes impractically slow. But now,
even with pauth=impdef, things are too slow too.. hwell..
Thanks,
/mjt
