Darwin, libjava testing takes for ever. The issue is that loading
gij (and
the native programs) is slow and most of the time is spent dyld (the
dynamic
loader, like ld.so). If we link with -single_module, the issue goes
away
because the dynamic loader have less symbols to resolve at load time.
I stole the code in libobjc and libgfortran which I wrote to fix
implement this.
K? Bootstrapped and tested on powerpc-darwin with no regressions.
and testing libjava takes now 2.5 hours instead of 4 hours or so.
Thanks,
Andrew Pinski
ChangeLog:
* Makefile.am (extra_ldflags_libjava): New variable to
substitute in.
(LIBLINK): Add extra_ldflags_libjava to it.
* configure.ac (extra_ldflags_libgfortran): Set for *-darwin* to
"-Wl,-single_module".

Aug 18, 2005, at 11:26 PM, Andrew Pinski wrote:

and testing libjava takes now 2.5 hours instead of 4 hours or so.

it is more than 4 hours, it was more like 6 hours, I was counting
from the wrong time when libjava testing started.

Also this should speed up the regression tester so we get a smaller
set of tests tested.
-- Pinski

Andrew Pinski writes:

Darwin, libjava testing takes for ever. The issue is that
loading gij (and the native programs) is slow and most of the time
is spent dyld (the dynamic loader, like ld.so). If we link with
-single_module, the issue goes away because the dynamic loader have
less symbols to resolve at load time.

I stole the code in libobjc and libgfortran which I wrote to fix
implement this.

K? Bootstrapped and tested on powerpc-darwin with no regressions.

Looks good. Is -single_module Darwinese for -Bsymbolic?

Andrew.

Aug 19, 2005, at 5:21 AM, Andrew Haley wrote:
>K? Bootstrapped and tested on powerpc-darwin with no regressions.
>
Looks good. Is -single_module Darwinese for -Bsymbolic?

In a way yes. The way functions call other functions on Darwin is via
a stub which was created by the compiler, -signal_module tells the
linker
to get rid of all those stubs for the functions inside that module.
So for an example:
if you have the following TUs:
int g(void);
int f(void) { return f();}

int g(void) { return 0; }

Normally in the first TU, the compiler creates a stub to call g.
And then in a shared library linked with -single_module, the stub is
removed and a direct call is created so you end up with no way to
over ride the function g at all out side the shared library.
So you end up with less relocations in the shared library for the
loader to deal with.

For executables, it is about the same as shared libraries with
-single_module.

-- Pinski

"Andrew" == Andrew Pinski <pinskia (AT) physics (DOT) uc.eduwrites:

Andrew* Makefile.am (extra_ldflags_libjava): New variable to
Andrewsubstitute in.
Andrew(LIBLINK): Add extra_ldflags_libjava to it.

You used automake 1.9.4 while the tree as a whole (excepting
libgfortran) is currently using 1.9.3. I think it would be fine for
all directories to move forward, but IM it is better to do this in
sync (and update install.texi) instead of at random.

Tom

Mike Stump writes:
Thursday, August 18, 2005, at 08:26 PM, Andrew Pinski wrote:
Darwin, libjava testing takes for ever. The issue is that loading
gij (and
the native programs) is slow and most of the time is spent dyld (the
dynamic
loader, like ld.so). If we link with -single_module, the issue goes
away
because the dynamic loader have less symbols to resolve at load time.

Seems reasonable to me, as long as everyone likes the semantics (cannot
override/replace funcs from that lib).

That's the right thing for the java core runtime library.

Andrew.

"Andrew" == Andrew Haley <aph (AT) redhat (DOT) comwrites:

>Seems reasonable to me, as long as everyone likes the semantics (cannot
>override/replace funcs from that lib).

AndrewThat's the right thing for the java core runtime library.

reason we haven't done this for Linux builds is that Anthony has a
jvmpi-using LD_PRELAD hack he likes to use

Tom

Tom Tromey writes:
"Andrew" == Andrew Haley <aph (AT) redhat (DOT) comwrites:

>Seems reasonable to me, as long as everyone likes the semantics (cannot
>override/replace funcs from that lib).

AndrewThat's the right thing for the java core runtime library.

reason we haven't done this for Linux builds is that Anthony has a
jvmpi-using LD_PRELAD hack he likes to use

This is pretty lame, really. We really should fix it for 4.1.

Andrew.