hey rod,
so i just filed:
6816972 check_rtime should not report libm_hwcap1.so.2 errors
i took a quick look at nightly.sh and check_rtime.pl and i wasn't quite
sure where the fix should go. both do some filtering of "unused
object=" errors, and i didn't see anything to indicate where this extra
filtering should belong.
originally i was also thinking that perhaps the solution should just be
to disable all HWCAP token expansion in check_rtime.pl, since that would
solve this problem and well as any future filter library problems. so i
spent some time staring at the linkers and libraries guide, but i didn't
find any strait forward way to disable the filter functionality.
ed
On Thu, Mar 12, 2009 at 01:26:59PM -0700, Rod Evans wrote:
> Edward Pilatowicz wrote:
>> hey guys,
>>
>> so whenever i do full nightly builds, i always use the -r option which
>> runs check_rtime. whenever i do this on kodiak (a sparc enterprise
>> M5000 build machine) i get LOTS of errors from:
>>
>> ==== Check ELF runtime attributes ====
>>
>> that all look like this:
>>
>> ./XXX: unused object=/lib/libm/libm_hwcap1.so.2 <remove lib or
>> -zignore?>
>>
>> this seems to be a runtime linker issue and not a problem associated
>> with the actual binaries built on disk.
>>
>> i say this because if i run ldd on any binaries on kodiak, even binaries
>> built by the gate machine, i see that libm_hwcap1.so.2 is being pulled
>> in. where as if i run ldd on the same binaries but i run it on the
>> sparc gate machine over nfs, then i don't get libm_hwcap1.so.2 pulled
>> in.
>>
>> so it seems to me that these error messages are bogus and should either
>> be filtered out by nightly.sh or check_rtime.pl.
>
> Looks like the sparc version of libm.so.2 has a filtee associated with it:
>
> oxpoly 406. elfdump -d /lib/libm.so.2
>
> Dynamic Section: .dynamic
> index tag value
> [0] NEEDED 0xd79 libc.so.1
> [1] SUNW_AUXILIARY 0xda2 $ORIGIN/libm/$HWCAP
> [2] SONAME 0x4 libm.so.2
>
> And the filtee kicks in on a machine that provides:
>
> /lib/libm/libm_hwcap1.so.2:
>
> Hardware/Software Capabilities Section: .SUNW_cap
> index tag value
> [0] CA_SUNW_HW_1 0x100 [ FMAF ]
>
> The only symbols this filtee offers are:
>
> /lib/libm/libm_hwcap1.so.2 - SUNW_1.1: fma;
> /lib/libm/libm_hwcap1.so.2 - SUNW_1.1: fmaf;
> /lib/libm/libm_hwcap1.so.2 - SUNW_1.1: __fma;
> /lib/libm/libm_hwcap1.so.2 - SUNW_1.1: __fmaf;
>
> so unless the application references these symbols, the filtee will look
> unused (neither libm.so.2 or libc.so.1 reference these symbols). Note,
> at runtime this filtee is only loaded if an attempt to bind to one of the
> fma symbols in libm.so.2 occurs. Check_rtime however forces all dependencies
> to be loaded (through ldd).
>
> So yes, it looks a check_rtime exception would be useful.
>
> --
>
> Rod.
