Re: 3 gold testcases...

[David Miller <davem at davemloft dot net>]

From: Ian Lance Taylor <iant@google.com>
Date: Tue, 08 Apr 2008 16:03:54 -0700

> David Miller <davem@davemloft.net> writes:
> 
> > 1) justsyms puts a string at 0x10000 hoping that will be
> >    out of the way and not in some mapped region of the
> >    executable image so that it will generate a segfault.
> >
> >    This happens to be where the .text gets mapped so the
> >    segfault doesn't happen and the testcase fails.
> 
> Ouch, bad luck.  Really, though, the address can be anywhere.  What if
> we change it to 0x100?

Yes, that works for me.

> > 2) script_test_2 fails because of the assertion:
> >
> >    assert(reinterpret_cast<uintptr_t>(start_test_area_1 + 48 + 7 + 4)
> > 	 == reinterpret_cast<uintptr_t>(start_data));
> >
> >    It assumes that the string will only consume 7 bytes, but on
> >    Sparc strings are emitted by the compiler with an ".align 8"
> >    directive, and thus the size consumed by the string is rounded
> >    up to 8 bytes.
> 
> Would it work if we make both strings 8 characters long and adjust
> accordingly?

Yes, that works here.

> 
> > 3) script_test_3 fails because the PT_LOAD segments created overlap
> >    within a page, and the kernel thus refuses to load the binary.
> 
> I'm not sure I understand this one.  It's normally OK for PT_LOAD
> segments to overlap within a page in the file.  The segments here do
> not overlap in virtual address space, unless your pages are very
> large.

You're right, something else must be wrong, I'll try to figure out
what the problem is.

BTW, on x86 on current Fedora rawhide, script_test_2 won't even link
because:

gcctestdir/ld: internal error in segment_precedes, at ../../src/gold/layout.cc:1530

Both 'flags1' and 'flags2' are equal to '4'.  Perhaps it's some
side effect of what's in some system objects.