Page faults and libgcj.so startup time

[Jeff Sturm]

On 3 Apr 2001, Per Bothner wrote:
> But it would be nice to understand better what R_386_RELATIVE is
> used for, especially since they seem to be the biggest chunk.

It is used whenever data is initialized to a text symbol, e.g.:

static void f() {}
void (*fp)() = f;

This needs a fixup because the image start address cannot be known at link
time.  The _Utf* symbols in libgcj link this way.  I have a hunch the
relocs here could be eliminated by replacing the static initializers with
code.

> Would compiling an entire package with a single compilation help?
> I would think it would remove some relocations, since they are now
> to the same .o file, and so may not need relocation at all,
> especially if -Bsymbolic is used.

In theory it could help, since relative addressing can be used between
text symbols within one object.  Also some redundant Utf* symbols could be
eliminated.  I'm not sure GNU ld is as smart as it can be about -Bsymbolic
though.

There are other benefits to compiling whole packages as well.  Has anybody
tried it with libgcj?

Jeff