On Thu, Aug 02, 2007 at 11:34:23PM +0100, Ken Moffat wrote:
>
> Anybody who follows my ramblings on clfs-dev will know that using
> an earlier kernel than the headers, particularly an -rc kernel, is
> "NOT a Good Idea" (TM) and I'm lucky I wasn't using glibc-2.6.
>
> I'll try to do a fresh by-the-book pair of builds (without
> non-toolchain tests) if I've got time.
>
I rather wish I hadn't bothered! This was 2007-07-31 with
glibc-2.5.1. The following files differed:
/lib/libc-2.5.1.so
/usr/bin/perl (also flagged as perl5.8.8 which is a hard link)
/usr/lib/gcc/i686-pc-linux-gnu/4.1.2/cc1 - generally expected
/usr/lib/gcc/i686-pc-linux-gnu/4.1.2/cc1plus - generally expected
/usr/lib/libstdc++.la - known difference
/usr/lib/libsupc++.la - known difference
/usr/share/doc/groff/1.18.1.4/meintro.ps impenetrable difference in numbers
/usr/share/doc/groff/1.18.1.4/meref.ps impenetrable difference in numbers
/usr/share/info/coreutils.info - known problem, second build used precompiled
info
Of these, the libtool archives add, or rather repeat, directories
in the first inplace build - libtool is like that, not a significant
issue. cc1 and cc1plus seem to differ for everybody, they should
probably be treated as 'expected to differ'.
The postscript files are impenetrable, the differences are in long
strings of numbers - this is the first time I've run farce on this
version of groff and I suspect these numbers are calculated, and
there is a small amount of error. I'm not worried about this.
That just leaves two somewhat important binaries. Looking at
'farce-extras' didn't show anything recognisable, so I can't get
away with changing one of the regexes to allow a different
date/time/version string. Time for strip --strip-all, but that
just seemed to show binary differences and strings moving slightly,
as on my previous build (UTF-8, for those wondering why this version
of groff is new to me, and unreliable because one kernel was
unreliable). This time, I made sure I was using 2.6.22.1 for both
builds. On to objdump -d followed by diff -u. For libc-2.5.1.so
there are various changes in addresses, enough to make me wonder if
address randomization is playing a part here:
-libc-2.5.1.so-0: file format elf32-i386
+libc-2.5.1.so-1: file format elf32-i386
Disassembly of section .plt:
@@ -310,7 +310,7 @@
158c6: 8d 76 00 lea 0x0(%esi),%esi
158c9: 8d bc 27 00 00 00 00 lea 0x0(%edi),%edi
158d0: 55 push %ebp
- 158d1: b8 e2 02 00 00 mov $0x2e2,%eax
+ 158d1: b8 de 02 00 00 mov $0x2de,%eax
158d6: 89 e5 mov %esp,%ebp
158d8: 53 push %ebx
158d9: e8 42 fd ff ff call 15620 <[EMAIL PROTECTED]>
@@ -482,7 +482,7 @@
15adc: 5d pop %ebp
15add: c3 ret
15ade: 89 f6 mov %esi,%esi
- 15ae0: 8d 83 cc 9a fe ff lea 0xfffe9acc(%ebx),%eax
+ 15ae0: 8d 83 ac 9a fe ff lea 0xfffe9aac(%ebx),%eax
15ae6: 89 04 24 mov %eax,(%esp)
15ae9: e8 82 89 04 00 call 5e470 <__libc_fatal>
15aee: 89 f6 mov %esi,%esi
@@ -1157,7 +1157,7 @@
1621f: 89 44 24 08 mov %eax,0x8(%esp)
16223: 8d 83 77 5f fe ff lea 0xfffe5f77(%ebx),%eax
16229: 89 44 24 04 mov %eax,0x4(%esp)
- 1622d: 8d 83 08 9b fe ff lea 0xfffe9b08(%ebx),%eax
+ 1622d: 8d 83 e8 9a fe ff lea 0xfffe9ae8(%ebx),%eax
16233: 89 04 24 mov %eax,(%esp)
16236: e8 85 bb 00 00 call 21dc0 <__assert_fail>
1623b: 8b 83 54 ff ff ff mov 0xffffff54(%ebx),%eax
@@ -1985,7 +1985,7 @@
16c59: 89 44 24 08 mov %eax,0x8(%esp)
16c5d: 8d 83 8c 5f fe ff lea 0xfffe5f8c(%ebx),%eax
16c63: 89 44 24 04 mov %eax,0x4(%esp)
- 16c67: 8d 83 2c 9b fe ff lea 0xfffe9b2c(%ebx),%eax
+ 16c67: 8d 83 0c 9b fe ff lea 0xfffe9b0c(%ebx),%eax
(stopped there, possibly there are other types of changes, but I'm
already out of my depth).
Perl is much nastier - the file sizes reported by 'ls -l' differ
even after 'strip --strip-all' : I know 'ls -l' is not totally
reliable for this, but the disassembly seems to indicate it has been
built differently:
--- objdump-d-perl-0 2007-08-04 23:23:45.000000000 +0100
+++ objdump-d-perl-1 2007-08-04 23:23:52.000000000 +0100
@@ -1,5 +1,5 @@
-perl-0: file format elf32-i386
+perl-1: file format elf32-i386
Disassembly of section .init:
@@ -7,9 +7,9 @@
805d804: 55 push %ebp
805d805: 89 e5 mov %esp,%ebp
805d807: 83 ec 08 sub $0x8,%esp
- 805d80a: e8 25 0e 00 00 call 805e634 <_start+0x24>
- 805d80f: e8 7c 0e 00 00 call 805e690 <_start+0x80>
- 805d814: e8 87 4f 0b 00 call 81127a0 <__libc_csu_init+0x220>
+ 805d80a: e8 25 0e 00 00 call 805e634 <call_gmon_start>
+ 805d80f: e8 7c 0e 00 00 call 805e690 <frame_dummy>
+ 805d814: e8 87 4f 0b 00 call 81127a0 <__do_global_ctors_aux>
805d819: c9 leave
805d81a: c3 ret
Disassembly of section .plt:
@@ -1148,16 +1148,18 @@
805e631: f4 hlt
805e632: 90 nop
805e633: 90 nop
+
+0805e634 <call_gmon_start>:
805e634: 55 push %ebp
805e635: 89 e5 mov %esp,%ebp
805e637: 53 push %ebx
805e638: 83 ec 04 sub $0x4,%esp
- 805e63b: e8 00 00 00 00 call 805e640 <_start+0x30>
+ 805e63b: e8 00 00 00 00 call 805e640 <call_gmon_start+0xc>
805e640: 5b pop %ebx
805e641: 81 c3 c8 9a 0c 00 add $0xc9ac8,%ebx
805e647: 8b 93 fc ff ff ff mov 0xfffffffc(%ebx),%edx
805e64d: 85 d2 test %edx,%edx
- 805e64f: 74 05 je 805e656 <_start+0x46>
+ 805e64f: 74 05 je 805e656 <call_gmon_start+0x22>
805e651: e8 76 ff ff ff call 805e5cc <[EMAIL PROTECTED]>
805e656: 58 pop %eax
805e657: 5b pop %ebx
@@ -1169,32 +1171,36 @@
805e65d: 90 nop
805e65e: 90 nop
805e65f: 90 nop
+
+0805e660 <__do_global_dtors_aux>:
805e660: 55 push %ebp
805e661: 89 e5 mov %esp,%ebp
805e663: 83 ec 08 sub $0x8,%esp
805e666: 80 3d 50 ae 12 08 00 cmpb $0x0,0x812ae50
- 805e66d: 74 0c je 805e67b <_start+0x6b>
- 805e66f: eb 1c jmp 805e68d <_start+0x7d>
+ 805e66d: 74 0c je 805e67b
<__do_global_dtors_aux+0x1b>
+ 805e66f: eb 1c jmp 805e68d
<__do_global_dtors_aux+0x2d>
805e671: 83 c0 04 add $0x4,%eax
805e674: a3 a8 84 12 08 mov %eax,0x81284a8
805e679: ff d2 call *%edx
and so forth. My perl build uses constant instructions across the
multiple builds, and deletes the directory, otherwise I'd blame it
on user error. Maybe glibc-2.5.1 is bringing something to the party
(what is this gmon_start ?), but in that case why doesn't it alter
both builds ?
Is something we build after perl suddenly causing a difference ?
To be clear, for the first run I built chapter 6 then followed
through installing bootscripts because I do them in the same
script. Then I saved it all, and later came back at the start
of chapter 6.
The differences in libc look to me like the thin end of the wedge
for farce analysis (if I'm right about randomization, somebody else
will have to develop a tool to analyze it), but these perl
differences baffle me.
ĸen
--
das eine Mal als Tragödie, das andere Mal als Farce
--
http://linuxfromscratch.org/mailman/listinfo/lfs-dev
FAQ: http://www.linuxfromscratch.org/faq/
Unsubscribe: See the above information page
