Prelink - KDE - GCC

octapus · n00b Joined: 10 Mar 2003 Posts: 35

I recently updated basic packages in my gentoo system such as (order given)
gcc-4.1.1
binutils-2.16.93
libstdc++-v3-3.3.6
glibc-2.4-r3
kdelibs-3.5.3
kdebase-3.5.3

Before the update I had the system prelinked. When I tried to prelink it after the updates I get errors from every program I try to execute (including emerge). i.e :

# su:
9319: out of bounds madness ... 0x3ee34
bash: symbol lookup error: /lib/libncurses.so.5: undefined symbol: _nc_tparm_err

My prelink command was: prelink -amvR

Literally I couldn't start any new program at all (either command line or X) but the running programs were fine. Most programs displayed symbol lookup errors on
/usr/lib/libXdmcp.so.6
/usr/lib/libxml2.so.2
/lib/libncurses.so.5

I had to undo prelinking in order to have my system usable again (prelink -auv)

I tried compiling the corresponding packages ( libXdmcp, libxml2, ncurses) with gcc-4.1.1 and but I had the same problems.

Other than that my system is stable but I would appreciate the faster program execution startup that prelink gives.

Does anybody know what could be wrong?

kfiaciarka · Posted: Sat Jun 10, 2006 9:43 pm Post subject:

a first post your emerge --info or make.conf.

octapus · n00b Joined: 10 Mar 2003 Posts: 35

Yes, I have my CFLAGS & LDFLAGS overloaded but there must be a more specific reason for what I observerd.

I tried searching for similar cases but nothing relevant.

Here it is:

kfiaciarka · Posted: Mon Jun 12, 2006 9:25 am Post subject:

IMHO it may be because of hardcored ldflags (-hashvals, -zdynsort and cflags -Os (your amd64 has at least 512 kb cache, so you use -Os?), -fweb -funit-at-a-time -freorder-blocks-and-partition -fmerge-all-constants -ftree-vectorize. I don't use such flags:) exept -funit-at-a-time which is a part od -O2. If you have hardcored system, don't be supprised with such behavior of prelink:)

octapus · n00b Joined: 10 Mar 2003 Posts: 35

Most of the not-so-common CFLAGS are new for me too (2 weeks). Some of them are supported only for GCC 4.1.1.

But for LDFLAGS, I have been using those for a long time (3-6 months, I'm not quite sure) and I have prelinked many times without problems (but with GCC 3.4.5 compiled binaries).

The strangest thing is that I can't find any other cases for which display similar messages. And I can't imagine why CFLAGS would affect linking (as long as the programs run perfectly well before prelinking). Let me say again that even though you may consider these settings hardcore they cause no instability whatsoever.

Flameeyes · Posted: Mon Jun 12, 2006 9:54 am Post subject:

A part reiterating that --enable-new-dtags is pointless on Gentoo (it's forced in binutils already, and anyway it just makes the ELF files use a different internal structure, no change on performances), prelink has known problems with -hashvals. So if you use one, you shouldn't use the other, quite simple

_________________
You want to know what I'm working on right now? Just follow my blog.

vipernicus · Posted: Mon Jun 12, 2006 1:31 pm Post subject:

Michael Meek's patches (-Bdirect, -hashvals,-zdynsort) were meant to be an alternative to prelinking. I no longer use them, because of how they have been thrown out by the binutils devs.

Now, I use:

octapus · n00b Joined: 10 Mar 2003 Posts: 35

Thank you for your answers.

If LDFALGS option -hashvals and prelink are meant to do the same thing what would you recommend in terms of performance (or how good each of them is at achieving the goal)?

On other aspects (other than peformance) we have the following:
hashvals needs to be done once during compilation for each program while prelink needs a lot of time for the whole system after each library upgrade.
hashvals is more of a hack compared to the way prelink works (?) and needs patched binutils.

One more thing: I'm not using prelink at all right now and things seem to be pretty fast. Do the below values have any meaning without prelink? I mean can I interpret them somehow to measure how well hashvals work (other than human perceived interactiveness)? :
LD_DEBUG=statistics kedit
22930:
22930: runtime linker statistics:
22930: total startup time in dynamic loader: 115512154 clock cycles
22930: time needed for relocation: 108494928 clock cycles (93.9%)
22930: number of relocations: 20010
22930: number of relocations from cache: 60914
22930: number of relative relocations: 40896
22930: time needed to load objects: 6560602 clock cycles (5.6%)
22930:
22930: runtime linker statistics:
22930: final number of relocations: 24654
22930: final number of relocations from cache: 60948