Don't use -G to build shared object on Solaris

[noloader]

Please don't use -G to build the shared object on Solaris. Instead, use -shared.

...
Generating hiredis.pc for pkgconfig...
gcc -G -o libhiredis.so -h libhiredis.so.0.14 -L/usr/local/lib -m64 -Wl,-R,'$ORIGIN/../lib' -Wl,-R,/usr/local/lib net.o hiredis.o sds.o async.o read.o
ar rcs libhiredis.a net.o hiredis.o sds.o async.o read.o
gcc -O3 -fPIC -g2 -O2 -m64 -march=native -fPIC -pthread -Wall -W -Wstrict-prototypes -Wwrite-strings -g -ggdb -o hiredis-test -L/usr/local/lib -m64 -Wl,-R,'$ORIGIN/../lib' -Wl,-R,/usr/local/lib -ldl -lnsl -lsocket test.o libhiredis.a

Here's the result of using -G:

ld.so.1: unittest: fatal: relocation error: R_AMD64_PC32: file /usr/local/lib/libhiredis.so.0.14: symbol _init: value 0x7f004e1390bd does not fit
gmake: *** [test] Killed

The failed self-test showed up when testing Unbound.

[michael-grunder]

Having done a bit of research it looks like you're right about this, but I don't have solaris boxes where I could test it.

The only thing I'm concerned about is whether the change might cause a regression in very old/obscure versions of Solaris. In other words, can we simply change -G to -shared within our SunOS uname branch, or do we need to be more specific about it.

Do you happen to know?

[noloader]

Hi Michael,

The only thing I'm concerned about ... very old/obscure versions of Solaris

Ah, OK. Just my 2 cents here... I help maintain some other libraries. My rule of thumb is, don't break the common case for the obscure case. Make the obscure case the special case so things "just work" for the common case. In this context, the common case is modern Solaris like 11.3 or 11.4, and the obscure case is Solaris 9.

As far as testing, I can give you SSH access to my test machines. For Solaris I use 11.3 i86pc (x86_64). It is a Core i5 4th gen, so you have AES, RDRAND and AVX. I need your authorized_keys file, if interested. Email me at noloader, gmail account.

OpenCSW provides free access to Solaris 9 through Solaris 11 on both Sparc and i86pc. See Upstream Maintainers to request an account.

I believe the GCC Compile Farm (Cfarm) provides access to the same machines via OpenCSW. gcc210 is Solaris 10/Sparc, and gcc211 is Solaris 11/Sparc. Accounts are free to FOSS developers. See Request an Account for the Cfarm.

I advise against running Solaris in a VM. Use real hardware instead. Solaris' virtual memory manager is different than Linux. The Solaris memory manager does not oversubscribe memory, and allocations are always backed by RAM. So a Solaris guest is a hog that causes the host's memory manager to work extra hard when Solaris allocates big chunks of RAM. The host's disk i/o goes wild when the guest is Solaris.

Do you happen to know?

Sorry, I'm not sure.

And I should probably say, use -shared when using GCC on Solaris. I have a sneaky suspicion -G may work as expected when using SunC compiler.

[noloader]

For what it is worth, here is how I patch the hiredis recipe in my build scripts:

IS_SUNC=$("$CC" -V 2>&1 | grep -i -c -E 'sun|studio')
IS_SOLARIS=$(uname -s | grep -i -c 'sunos')
...

# Awful Solaris 64-bit hack. Use -G for SunC, and -shared for GCC
if [[ "$IS_SOLARIS" -ne 0 && "$IS_SUNC" -eq 0 ]]; then
    sed 's/ -G / -shared /g' Makefile > Makefile.fixed
    mv Makefile.fixed Makefile; chmod +x Makefile
fi

[michael-grunder]

OK. Just my 2 cents here... I help maintain some other libraries. My rule of thumb is, don't break the common case for the obscure case

Yeah we're on the same page, and it certainly seems like -G is the obscure case now (especially given the line was last touched in 2011 via 608e29b).

Ideally we can make the change without breaking builds that are currently compiling although we have some leeway here since we're still pre 1.0.0.

[michael-grunder]

What do you think about the change in 23e3bd4.

I can't use grep -E because apparently that doesn't exist in Solaris? Or at least it threw an error in my Solaris VM.

I'm somewhat concerned about the portability of egrep but as far as I can tell this exists everywhere.

[michael-grunder]

@noloader Unless you have an objection to 23e3bd4 I'm going to get it merged. I tested building on Solaris both with SunStudio 12.6 and gcc and both worked for me.

Interestingly SunStudio does not like -ggdb but it didn't prevent building.

[michael-grunder]

Closing via #796