pgsql-server/src/makefiles Makefile.linux

Tom Lane writes:

Use -fPIC on Sparc, per Tom Callaway.

Then we might as well use -fPIC everywhere all the time, because if it
happens on Linux/Sparc then it's bound to happen on all platforms where it
makes a difference. (And where it doesn't make a difference, there's no
harm.)

--
Peter Eisentraut peter_e@gmx.net

Peter Eisentraut <peter_e@gmx.net> writes:

Tom Lane writes:

Use -fPIC on Sparc, per Tom Callaway.

Then we might as well use -fPIC everywhere all the time, because if it
happens on Linux/Sparc then it's bound to happen on all platforms where it
makes a difference.

No, it's not: Sparc has the smallest -fpic GOT size limit of all
platforms, and apparently it needs to generate more GOT entries than
others, too.

I'm not actually convinced that Sparc has a real need for -fPIC, either,
since Lamar could not reproduce Tom's problem, and the GOT sizes I can
measure on x86 are an order of magnitude smaller than the Sparc limit.
I pushed out the change quickly because of the impending 7.3.3 deadline,
but I'm more than half inclined to revert it.

I have verified that HPPA doesn't need -fPIC, and am planning to make
that change sometime. The reason Makefile.hpux is that way now is that
I was accustomed to needing -fPIC on a different project that generated
shared libraries far bigger than any of ours.

regards, tom lane

Tom Lane writes:

I'm not actually convinced that Sparc has a real need for -fPIC, either,
since Lamar could not reproduce Tom's problem, and the GOT sizes I can
measure on x86 are an order of magnitude smaller than the Sparc limit.
I pushed out the change quickly because of the impending 7.3.3 deadline,
but I'm more than half inclined to revert it.

If we don't have reproduceable evidence then we shouldn't play around with
these things in a minor release.

--
Peter Eisentraut peter_e@gmx.net

Peter Eisentraut <peter_e@gmx.net> writes:

Tom Lane writes:

I'm not actually convinced that Sparc has a real need for -fPIC, either,
since Lamar could not reproduce Tom's problem, and the GOT sizes I can
measure on x86 are an order of magnitude smaller than the Sparc limit.
I pushed out the change quickly because of the impending 7.3.3 deadline,
but I'm more than half inclined to revert it.

If we don't have reproduceable evidence then we shouldn't play around with
these things in a minor release.

<shrug> It's not like using -fPIC instead of -fpic could have any worse
consequences than losing a little performance. And Callaway's word does
carry some weight with me. I plan to bug him further about what problem
he actually saw before making a decision.

regards, tom lane