12:48 a.m.
[Should we still have bug-gcc in the CC list?]
[Warning: It is late at night here and I had a long day]
Zack Weinberg <zack(a)wolery.cumb.org> writes:
A lot of the code in the lib-src directory needed to be worked over
before it would compile. I can send the patch if anyone cares.
Please send the patch to xemacs-patches(a)xemacs.org anyway, if it isn't
much trouble. I am not sure whether it would be possible to apply it
all (at least some of the lib-src stuff is third part maintained). It
would be a waste to loose any work that you have already done.
Building XEmacs with g++ is definitely Martin's thing not mine,
however I do not see some of the symptoms you see.
The makefile attempts to link temacs by hand and gets it horribly
wrong, causing the built temacs to crash inside libc startup code.
You're doing something like gcc -nostdlib pre-crt0.o /usr/lib/crt1.o
/usr/lib/crti.o <objects> <X libraries> -ldl -lgcc /usr/lib/crtn.o
For me the link looks like (plain Redhat 6.0 with egcs 1.1.2)
gcc -nostdlib -L/usr/X11R6/lib -o temacs pre-crt0.o /usr/lib/crt1.o /usr/lib/crti.o
<objects> ../lwlib/liblw.a -lXaw3d -ltiff -lpng -ljpeg -lz -lcompface -lXpm -lXmu
-lXt -lXext -lX11 -lSM -lICE -ldl -ldb -lgpm -lncurses -L/usr/lib -lesd -laudiofile -lm
-lm -lgcc -lc -lgcc /usr/lib/crtn.o `gcc -print-libgcc-file-name`
Note the ugly "-lm -lm -lgcc -lc -lgcc" part. I am fairly sure the -lc
comes explicitly from
dnl Non-ordinary link usually requires -lc
test "$ordinary_link" = "no" -a -z "$libs_standard"
&& libs_standard="-lc"
First off, -nostdlib doesn't turn off automatic linking with crt1
in
all versions of gcc. You need -nostartfiles too.
Are you sure? I would call it a bug then. The phrase
`-nostdlib'
Do not use the standard system startup files or libraries when
linking. No startup files and only the libraries you specify will
be passed to the linker.
Is in gcc.info from gcc 2.6, egcs 1.1.2 and gcc 2.95
Anyway. I hate that stuff too (because it makes it harder to enable
profiling). However it is there for a reason, so it must have been
necessary at least in the past on some platform.
Jan
P.S. Martin. I think we need to have "--with-ordinary-link" (which
already 95% there it seems)
3:01 a.m.
New subject: egcs-20000131 c++ inline code generation broken
On Sat, Feb 05, 2000 at 01:48:20AM +0100, Jan Vroonhof wrote:
Zack Weinberg <zack(a)wolery.cumb.org> writes:
> A lot of the code in the lib-src directory needed to be worked over
> before it would compile. I can send the patch if anyone cares.
Please send the patch to xemacs-patches(a)xemacs.org anyway, if it isn't
much trouble. I am not sure whether it would be possible to apply it
all (at least some of the lib-src stuff is third part maintained). It
would be a waste to loose any work that you have already done.
I shall do so under separate cover.
Also, I recompiled with g++ in current CVS (roughly the 20000204
snapshot), and both temacs and the xemacs it dumps work correctly
(well, I didn't do anything with xemacs except start it up).
Given that Martin was running temacs under the debugger in batch
mode, I think the actual miscompilation was that message_internal was
calling echo_area_append when it should have been calling
write_string_to_stdio_stream.
Odds are that the bug was fixed by this patch:
2000-02-04 Richard Henderson <rth(a)cygnus.com>
* function.c (diddle_return_value): Rework to use a callback function.
Use current_function_return_rtx if it's been set up.
(do_clobber_return_reg, clobber_return_register): New.
(do_use_return_reg, use_return_register): New.
(expand_function_end): Use them.
* stmt.c (expand_null_return): Likewise.
* function.h: Declare them.
* flow.c (mark_regs_live_at_end): Use diddle_return_value.
(mark_reg): Change arguments as appropriate for callback.
* integrate.c (expand_inline_function): Revert 19 Jan change.
> The makefile attempts to link temacs by hand and gets it
horribly
> wrong, causing the built temacs to crash inside libc startup code.
>
> You're doing something like gcc -nostdlib pre-crt0.o /usr/lib/crt1.o
> /usr/lib/crti.o <objects> <X libraries> -ldl -lgcc /usr/lib/crtn.o
For me the link looks like (plain Redhat 6.0 with egcs 1.1.2)
gcc -nostdlib -L/usr/X11R6/lib -o temacs pre-crt0.o /usr/lib/crt1.o /usr/lib/crti.o
<objects> ../lwlib/liblw.a -lXaw3d -ltiff -lpng -ljpeg -lz -lcompface -lXpm -lXmu
-lXt -lXext -lX11 -lSM -lICE -ldl -ldb -lgpm -lncurses -L/usr/lib -lesd -laudiofile -lm
-lm -lgcc -lc -lgcc /usr/lib/crtn.o `gcc -print-libgcc-file-name`
Note the ugly "-lm -lm -lgcc -lc -lgcc" part. I am fairly sure the -lc
comes explicitly from
dnl Non-ordinary link usually requires -lc
test "$ordinary_link" = "no" -a -z "$libs_standard"
&& libs_standard="-lc"
Weird. Mine didn't do that, and it works if I do it that way. (The
key is the inclusion of /usr/lib/crtn.o.)
> First off, -nostdlib doesn't turn off automatic linking with
crt1 in
> all versions of gcc. You need -nostartfiles too.
Are you sure? I would call it a bug then. The phrase
`-nostdlib'
Do not use the standard system startup files or libraries when
linking. No startup files and only the libraries you specify will
be passed to the linker.
The versions I was thinking of are _ancient_. I only know about the
problem because of historical cruft in glibc's Makefiles. But a
little extra paranoia never hurt anyone.
Anyway. I hate that stuff too (because it makes it harder to enable
profiling). However it is there for a reason, so it must have been
necessary at least in the past on some platform.
Hrm. ORDINARY_LINK should be safe for LINUX_ELF, provided you define
DATA_START as well... My suggested patch follows.
zw
===================================================================
Index: src/s/linux.h
--- src/s/linux.h 1999/07/06 07:33:05 1.11.2.2
+++ src/s/linux.h 2000/02/05 02:42:06
＠＠ -84,20 +84,21 ＠＠ Boston, MA 02111-1307, USA. */
#endif /* !_IO_STDIO_H */
#endif /* C_CODE */
-/* Ask GCC where to find libgcc.a. */
-#define LIB_GCC "`$(CC) $(C_SWITCH_X_SITE) -print-libgcc-file-name`"
+#ifdef LINUX_ELF
+
+#define UNEXEC "unexelf.o"
+#define UNEXEC_USE_MAP_PRIVATE
+#define DATA_START ({ extern int data_start; (char *) &data_start; })
+#define ORDINARY_LINK
-#ifndef LINUX_ELF
+#else
+
/* Linux has crt0.o in a non-standard place */
#define START_FILES "pre-crt0.o /usr/lib/crt0.o"
-#else
-/**/
-#if defined(__linux__) && defined(powerpc) /*Added by Fukui*/
-#define START_FILES /*Added by Fukui*/
-#else /*Added by Fukui*/
+#define LIB_STANDARD "-lc" /* avoid -lPW */
+#define LIB_GCC "-lgcc"
+#define LINKER "$(CC) -nostdlib"
-#define START_FILES "pre-crt0.o /usr/lib/crt1.o /usr/lib/crti.o"
-#endif /*Added by Fukui*/
#endif
/* This is needed for sysdep.c */
＠＠ -112,14 +113,6 ＠＠ Boston, MA 02111-1307, USA. */
/* Best not to include -lg, unless it is last on the command line */
#define LIBS_DEBUG
#define LIBS_TERMCAP "-ltermcap -lcurses" /* save some space with shared
libs*/
-#ifndef LINUX_ELF
-#define LIB_STANDARD "-lc" /* avoid -lPW */
-#else
-/*#undef LIB_GCC
- #define LIB_GCC*/
-#define LIB_STANDARD "-lgcc -lc -lgcc /usr/lib/crtn.o"
-#define LINKER "$(CC) -nostdlib"
-#endif
#ifdef TERM
#define LIBS_SYSTEM "-lclient"
＠＠ -129,26 +122,6 ＠＠ Boston, MA 02111-1307, USA. */
since libc-4.4.1. So -lipc was deleted. */
#define LIBS_SYSTEM
#endif
-
-#ifdef LINUX_ELF
-#define UNEXEC "unexelf.o"
-#define UNEXEC_USE_MAP_PRIVATE
-#if 0
-/* mrb - Ordinary link is simple and effective */
-/* slb - Not any more ... :-( */
-#define ORDINARY_LINK
-#endif /* 0 */
-
-/* I still think ORDINARY_LINK should be the default, but since slb
- insists, ORDINARY_LINK will stay on until we expunge the dump code.
- However, the user (i.e. me!) should be able to specify ORDINARY_LINK via
- configure --cppflags=-DORDINARY_LINK ... */
-#ifdef ORDINARY_LINK
-#undef LIB_STANDARD
-#undef START_FILES
-#undef LIB_GCC
-#endif
-#endif /* LINUX_ELF */
#ifdef LINUX_QMAGIC
===================================================================
6:17 a.m.
New subject: egcs-20000131 c++ inline code generation broken
>>>> "ZW" == Zack Weinberg
<zack(a)wolery.cumb.org> writes:
ZW> On Sat, Feb 05, 2000 at 01:48:20AM +0100, Jan Vroonhof wrote:
> Zack Weinberg <zack(a)wolery.cumb.org> writes:
> > A lot of the code in the lib-src directory needed to be worked over
> > before it would compile. I can send the patch if anyone cares.
>
> Please send the patch to xemacs-review(a)xemacs.org anyway, if it isn't
> much trouble. I am not sure whether it would be possible to apply it
> all (at least some of the lib-src stuff is third part maintained). It
> would be a waste to loose any work that you have already done.
ZW> I shall do so under separate cover.
ZW> Also, I recompiled with g++ in current CVS (roughly the 20000204
ZW> snapshot), and both temacs and the xemacs it dumps work correctly
ZW> (well, I didn't do anything with xemacs except start it up).
ZW> Given that Martin was running temacs under the debugger in batch
ZW> mode, I think the actual miscompilation was that message_internal was
ZW> calling echo_area_append when it should have been calling
ZW> write_string_to_stdio_stream.
ZW> Odds are that the bug was fixed by this patch:
ZW> 2000-02-04 Richard Henderson <rth(a)cygnus.com>
ZW> * function.c (diddle_return_value): Rework to use a callback function.
ZW> Use current_function_return_rtx if it's been set up.
ZW> (do_clobber_return_reg, clobber_return_register): New.
ZW> (do_use_return_reg, use_return_register): New.
ZW> (expand_function_end): Use them.
ZW> * stmt.c (expand_null_return): Likewise.
ZW> * function.h: Declare them.
ZW> * flow.c (mark_regs_live_at_end): Use diddle_return_value.
ZW> (mark_reg): Change arguments as appropriate for callback.
ZW> * integrate.c (expand_inline_function): Revert 19 Jan change.
I will test again when the next snapshot is released. That patch sure
looks close.
> > The makefile attempts to link temacs by hand and gets it
horribly
> > wrong, causing the built temacs to crash inside libc startup code.
> >
> > You're doing something like gcc -nostdlib pre-crt0.o /usr/lib/crt1.o
> > /usr/lib/crti.o <objects> <X libraries> -ldl -lgcc /usr/lib/crtn.o
>
> For me the link looks like (plain Redhat 6.0 with egcs 1.1.2)
>
> gcc -nostdlib -L/usr/X11R6/lib -o temacs pre-crt0.o /usr/lib/crt1.o
/usr/lib/crti.o <objects> ../lwlib/liblw.a -lXaw3d -ltiff -lpng -ljpeg -lz
-lcompface -lXpm -lXmu -lXt -lXext -lX11 -lSM -lICE -ldl -ldb -lgpm -lncurses -L/usr/lib
-lesd -laudiofile -lm -lm -lgcc -lc -lgcc /usr/lib/crtn.o `gcc -print-libgcc-file-name`
>
> Note the ugly "-lm -lm -lgcc -lc -lgcc" part. I am fairly sure the -lc
> comes explicitly from
>
> dnl Non-ordinary link usually requires -lc
> test "$ordinary_link" = "no" -a -z "$libs_standard"
&& libs_standard="-lc"
Sometime in the past, I changed linking on Linux to use
ORDINARY_LINK. We then had a patch-of-war, with me and Steve Baur
reversing each other's changes. He prevailed.
I recommend just leaving the current dumping mechanism. Olivier
Galibert's new portable dumper promises to let us use a consistent
mechanism across systems that doesn't involve calling ld directly or
knowing the guts of the object file format on the system.
ZW> Hrm. ORDINARY_LINK should be safe for LINUX_ELF, provided you define
ZW> DATA_START as well... My suggested patch follows.
Zack, I like your spirit, and I like this kind of patch, and I've
created a similar one myself, but I'll veto it anyways, because we
have a better way out of this quagmire.
Martin
11:01 a.m.
New subject: egcs-20000131 c++ inline code generation broken
>>>> "ZW" == Zack Weinberg
<zack(a)wolery.cumb.org> writes:
ZW> Odds are that the bug was fixed by this patch:
ZW> 2000-02-04 Richard Henderson <rth(a)cygnus.com>
ZW> * integrate.c (expand_inline_function): Revert 19 Jan change.
I am sorry to report that this inlining bug is still there in
egcs-20000207. Again I offer access to my machine if provided with an
ssh public key. A reproducible test case is worth its weight in electrons.
Don't try reproducing it on CVS XEmacs 21.2. That has my egcs bug
workaround patch applied.
I also found another workaround that sheds light on the bug.
Recap: The original bug concerned C++ inlining a function with 6
arguments, one of which has its address taken. This is a rare enough
operation that one can see how the compiler can easily get it wrong -
it wants arguments to be in registers, maybe.
If I copy the argument to an auto variable, and take the argument of
_that_, the problem goes away. I really hope this is ringing some bells.
Index: minibuf.c
===================================================================
RCS file: /usr/CVSroot/XEmacs/xemacs/src/minibuf.c,v
retrieving revision 1.9.2.5
diff -u -U8 -r1.9.2.5 minibuf.c
--- minibuf.c 2000/02/07 07:59:46 1.9.2.5
+++ minibuf.c 2000/02/08 10:54:33
＠＠ -705,57 +705,52 ＠＠
echo_area_append (struct frame *f, CONST Bufbyte *nonreloc, Lisp_Object reloc,
Bytecount offset, Bytecount length,
Lisp_Object label)
{
/* This function can call lisp */
Lisp_Object obj;
struct gcpro gcpro1;
Lisp_Object frame;
+ Lisp_Object length_on_the_stack = length;
- /* There is an inlining bug in egcs-20000131 c++ that can be worked
- around as follows: */
-#if defined (__GNUC__) && defined (__cplusplus)
- alloca (4);
-#endif
-
/* some callers pass in a null string as a way of clearing the echo area.
check for length == 0 now; if this case, neither nonreloc nor reloc
may be valid. */
- if (length == 0)
+ if (length_on_the_stack == 0)
return;
- fixup_internal_substring (nonreloc, reloc, offset, &length);
+ fixup_internal_substring (nonreloc, reloc, offset, &length_on_the_stack);
/* also check it here, in case the string was really blank. */
- if (length == 0)
+ if (length_on_the_stack == 0)
return;
if (!NILP (Ffboundp (Qappend_message)))
{
- if (STRINGP (reloc) && offset == 0 && length == XSTRING_LENGTH
(reloc))
+ if (STRINGP (reloc) && offset == 0 && length_on_the_stack ==
XSTRING_LENGTH (reloc))
obj = reloc;
else
{
if (STRINGP (reloc))
nonreloc = XSTRING_DATA (reloc);
- obj = make_string (nonreloc + offset, length);
+ obj = make_string (nonreloc + offset, length_on_the_stack);
}
XSETFRAME (frame, f);
GCPRO1 (obj);
call4 (Qappend_message, label, obj, frame,
EQ (label, Qprint) ? Qt : Qnil);
UNGCPRO;
}
else
{
if (STRINGP (reloc))
nonreloc = XSTRING_DATA (reloc);
- write_string_to_stdio_stream (stderr, 0, nonreloc, offset, length,
+ write_string_to_stdio_stream (stderr, 0, nonreloc, offset, length_on_the_stack,
Qterminal);
}
}
12:10 p.m.
New subject: egcs-20000131 c++ inline code generation broken
At 12:01 08.02.00 , Martin Buchholz wrote:
>>>>> "ZW" == Zack Weinberg
<zack(a)wolery.cumb.org> writes:
ZW> Odds are that the bug was fixed by this patch:
ZW> 2000-02-04 Richard Henderson <rth(a)cygnus.com>
ZW> * integrate.c (expand_inline_function): Revert 19 Jan change.
I am sorry to report that this inlining bug is still there in
egcs-20000207. Again I offer access to my machine if provided with an
ssh public key. A reproducible test case is worth its weight in electrons.
Don't try reproducing it on CVS XEmacs 21.2. That has my egcs bug
workaround patch applied.
I also found another workaround that sheds light on the bug.
Recap: The original bug concerned C++ inlining a function with 6
arguments, one of which has its address taken. This is a rare enough
operation that one can see how the compiler can easily get it wrong -
it wants arguments to be in registers, maybe.
That sounds familiar to one bug I'm seeing on powerpc-linuc-gnu with C++.
Can you check if adding -fno-strict-aliasing fixes the bug? If yes, could
you try if the patch in
<http://gcc.gnu.org/ml/gcc-bugs/2000-01/msg00038.html> fixes the bug for
the default -fstrict-aliasing too? I would be very interested to get a C
testcase for that problem.
Franz.
9055
days inactive
9058
days old
4 comments
4 participants
participants (4)
-
Franz Sirl -
Jan Vroonhof -
Martin Buchholz -
Zack Weinberg