[RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[Joel Brobecker]

[-- Attachment #1: Type: text/plain, Size: 2756 bytes --]

Hello,

I noticed the following SEGV in our testsuite. Here is below how to
reproduce it using the store.c sources in testsuite/gdb.base:

        % gcc -c -g store.c
        % gcc -o store store.o

The following transcript shows how to cause the SEGV:

        (gdb) b wack_doublest
        Breakpoint 1 at 0x1000256c: file store.c, line 125.
        (gdb) run
        Starting program: /[...]/gdb.base/store

        Breakpoint 1, wack_doublest (u=Unhandled dwarf expression opcode 0x93
        ) at store.c:125
        125       register doublest l = u, r = v;
        (gdb) n
        warning: GDB can't find the start of the function at 0x100108d4.

            GDB is unable to find the start of the function at 0x100108d4
        and thus can't determine the size of that function's stack frame.
        This means that GDB may be unable to access that stack frame, or
        the frames below it.
            This problem is most likely caused by an invalid program counter or
        stack pointer.
            However, if you think GDB should simply search farther back
        from 0x100108d4 for code which looks like the beginning of a
        function, you can increase the range of the search using the `set
        heuristic-fence-post' command.
        126       l = add_doublest (l, r);
        (gdb) p l
        zsh: 6356790 segmentation fault (core dumped)  ../../gdb store

We'll ignore the 2 warnings for now ("opcode 0x93" = OP_piece, and
"can't find start of function at 0x..."), and focus on the SEGV.

The problem is that GDB thinks that long double variables are 64 bits
long, but they are in fact 128 bits long. The discrepancy causes GDB
to fail to find the associated format for type doublest (which is
a typedef of long double), and then call floatformat_is_valid() with
a NULL format... See values.c:unpack_double():

      if (!floatformat_is_valid (floatformat_from_type (type), valaddr))
        {
          *invp = 1;
          return 0.0;
        }

The attached patch should fixes the problem (on 032, the SGI compiler
claims that "long double" is not supported on implicitly changes them
into plain "double" types (64bit), while GCC says they are 64bit).

2004-07-21  Joel Brobecker  <brobecker@gnat.com>

        * mips-irix-tdep.c: #include "mips-tdep.h".
        (mips_irix_init_abi): Set size of long double to 128 bits
        for N32 and N64 ABIs.
        * Makefile.in (mips-irix-tdep.o): Add dependency on mips-tdep.h.

Tested on mips-irix, fixes the problem above and many other FAILs.
OK to commit?

As for the 6.2 branch, my guess is that it would need this fix as well.
But I haven't looked at it, because I am already too overloaded with
work. There is also the bfd patch I just submitted, too.

-- 
Joel

[-- Attachment #2: irix.diff --]
[-- Type: text/plain, Size: 1832 bytes --]

Index: mips-irix-tdep.c
===================================================================
RCS file: /cvs/src/src/gdb/mips-irix-tdep.c,v
retrieving revision 1.2
diff -u -p -r1.2 mips-irix-tdep.c
--- mips-irix-tdep.c	21 Dec 2002 19:58:07 -0000	1.2
+++ mips-irix-tdep.c	21 Jul 2004 21:18:45 -0000
@@ -24,6 +24,7 @@
 #include "osabi.h"
 
 #include "elf-bfd.h"
+#include "mips-tdep.h"
 
 static void
 mips_irix_elf_osabi_sniff_abi_tag_sections (bfd *abfd, asection *sect,
@@ -80,6 +81,20 @@ static void
 mips_irix_init_abi (struct gdbarch_info info,
                     struct gdbarch *gdbarch)
 {
+  struct gdbarch_tdep *tdep = gdbarch_tdep (gdbarch);
+  enum mips_abi abi = mips_abi (gdbarch);
+
+  switch (abi)
+    {
+      case MIPS_ABI_O32:
+         break;
+      case MIPS_ABI_N32:
+      case MIPS_ABI_N64:
+         set_gdbarch_long_double_bit (gdbarch, 128);
+         break;
+      default:
+         internal_error (__FILE__, __LINE__, "unexpected ABI");
+    }
 }
 
 void
Index: Makefile.in
===================================================================
RCS file: /cvs/src/src/gdb/Makefile.in,v
retrieving revision 1.595
diff -u -p -r1.595 Makefile.in
--- Makefile.in	17 Jul 2004 11:03:46 -0000	1.595
+++ Makefile.in	21 Jul 2004 21:18:56 -0000
@@ -2113,7 +2113,8 @@ mem-break.o: mem-break.c $(defs_h) $(sym
 	$(target_h)
 minsyms.o: minsyms.c $(defs_h) $(gdb_string_h) $(symtab_h) $(bfd_h) \
 	$(symfile_h) $(objfiles_h) $(demangle_h) $(value_h) $(cp_abi_h)
-mips-irix-tdep.o: mips-irix-tdep.c $(defs_h) $(osabi_h) $(elf_bfd_h)
+mips-irix-tdep.o: mips-irix-tdep.c $(defs_h) $(osabi_h) $(elf_bfd_h) \
+	$(mips_tdep_h)
 mips-linux-nat.o: mips-linux-nat.c $(defs_h) $(mips_tdep_h)
 mips-linux-tdep.o: mips-linux-tdep.c $(defs_h) $(gdbcore_h) $(target_h) \
 	$(solib_svr4_h) $(osabi_h) $(mips_tdep_h) $(gdb_string_h) \

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[Daniel Jacobowitz]

On Wed, Jul 21, 2004 at 02:25:22PM -0700, Joel Brobecker wrote:
> The problem is that GDB thinks that long double variables are 64 bits
> long, but they are in fact 128 bits long. The discrepancy causes GDB
> to fail to find the associated format for type doublest (which is
> a typedef of long double), and then call floatformat_is_valid() with
> a NULL format... See values.c:unpack_double():
> 
>       if (!floatformat_is_valid (floatformat_from_type (type), valaddr))
>         {
>           *invp = 1;
>           return 0.0;
>         }
> 
> The attached patch should fixes the problem (on 032, the SGI compiler
> claims that "long double" is not supported on implicitly changes them
> into plain "double" types (64bit), while GCC says they are 64bit).
> 
> 2004-07-21  Joel Brobecker  <brobecker@gnat.com>
> 
>         * mips-irix-tdep.c: #include "mips-tdep.h".
>         (mips_irix_init_abi): Set size of long double to 128 bits
>         for N32 and N64 ABIs.
>         * Makefile.in (mips-irix-tdep.o): Add dependency on mips-tdep.h.
> 
> Tested on mips-irix, fixes the problem above and many other FAILs.
> OK to commit?

I can't find the right CVS tree at the moment to double-check, but I
think this is true for mips-linux and embedded MIPS using N32/N64 also.

-- 
Daniel Jacobowitz

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[Joel Brobecker]

> > 2004-07-21  Joel Brobecker  <brobecker@gnat.com>
> > 
> >         * mips-irix-tdep.c: #include "mips-tdep.h".
> >         (mips_irix_init_abi): Set size of long double to 128 bits
> >         for N32 and N64 ABIs.
> >         * Makefile.in (mips-irix-tdep.o): Add dependency on mips-tdep.h.
> > 
> > Tested on mips-irix, fixes the problem above and many other FAILs.
> > OK to commit?
> 
> I can't find the right CVS tree at the moment to double-check, but I
> think this is true for mips-linux and embedded MIPS using N32/N64 also.

Do you think this is going to affect this patch? Or are you suggesting
that we should fix mips-linux and embedded N32/N64 as well?

-- 
Joel

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[Daniel Jacobowitz]

On Wed, Jul 21, 2004 at 04:33:10PM -0700, Joel Brobecker wrote:
> > > 2004-07-21  Joel Brobecker  <brobecker@gnat.com>
> > > 
> > >         * mips-irix-tdep.c: #include "mips-tdep.h".
> > >         (mips_irix_init_abi): Set size of long double to 128 bits
> > >         for N32 and N64 ABIs.
> > >         * Makefile.in (mips-irix-tdep.o): Add dependency on mips-tdep.h.
> > > 
> > > Tested on mips-irix, fixes the problem above and many other FAILs.
> > > OK to commit?
> > 
> > I can't find the right CVS tree at the moment to double-check, but I
> > think this is true for mips-linux and embedded MIPS using N32/N64 also.
> 
> Do you think this is going to affect this patch? Or are you suggesting
> that we should fix mips-linux and embedded N32/N64 as well?

I'm suggesting that this change belongs in mips-tdep.c rather than in
an OSABI handler.  Unfortunately I'm not set up to test on mips64-linux
at the moment (and results were pretty dismal, though not quite as bad
as Irix).

-- 
Daniel Jacobowitz

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[Joel Brobecker]

> I'm suggesting that this change belongs in mips-tdep.c rather than in
> an OSABI handler.  Unfortunately I'm not set up to test on mips64-linux
> at the moment (and results were pretty dismal, though not quite as bad
> as Irix).

But would this change be applicable to NBSD as well? I am happy to
install this change mips-tdep if you think long doubles are 16 bytes
long on all mips targets. The only problem is that I won't be able to
test it on anything but IRIX.

-- 
Joel

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[Daniel Jacobowitz]

On Wed, Jul 21, 2004 at 05:46:40PM -0700, Joel Brobecker wrote:
> > I'm suggesting that this change belongs in mips-tdep.c rather than in
> > an OSABI handler.  Unfortunately I'm not set up to test on mips64-linux
> > at the moment (and results were pretty dismal, though not quite as bad
> > as Irix).
> 
> But would this change be applicable to NBSD as well? I am happy to
> install this change mips-tdep if you think long doubles are 16 bytes
> long on all mips targets. The only problem is that I won't be able to
> test it on anything but IRIX.

I don't think NetBSD even supports mips64.  Not in GCC HEAD, at least.

According to GCC long double is 128-bit for all n64 and n32 targets.

-- 
Daniel Jacobowitz

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[cgd]

At Thu, 22 Jul 2004 00:49:31 +0000 (UTC), "Daniel Jacobowitz" wrote:
> I don't think NetBSD even supports mips64.  Not in GCC HEAD, at least.

NetBSD doesn't support n32 or n64, or have any real support for
'mips64' at all, at this time (AFAIK).


> According to GCC long double is 128-bit for all n64 and n32 targets.

Note that the actual 'long double' format is *different* between
Linux and IRIX, AFAIK.

On Linux (w/ glibc), it's 128 bits in the typical IEEE format -- I
don't recall what the exact terminology or specification for 128-bit
IEEE-style value is, tho.  (like normal double, but wider mantissa and
wider exponent.)

On IRIX, if i recall correctly, 'long double' is implemented as a
*pair of doubles*, which allows one to get somewhat-extended
precision, but doesn't allow the complete range of 128-bit IEEE-style.
(I imagine the benefit is, you can do at least some of the 'long
double' ops w/ 'double' FP ops in HW.  So, better speed... maybe.)



cgd

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[Daniel Jacobowitz]

On Wed, Jul 21, 2004 at 06:19:34PM -0700, cgd@broadcom.com wrote:
> > According to GCC long double is 128-bit for all n64 and n32 targets.
> 
> Note that the actual 'long double' format is *different* between
> Linux and IRIX, AFAIK.
> 
> On Linux (w/ glibc), it's 128 bits in the typical IEEE format -- I
> don't recall what the exact terminology or specification for 128-bit
> IEEE-style value is, tho.  (like normal double, but wider mantissa and
> wider exponent.)
> 
> On IRIX, if i recall correctly, 'long double' is implemented as a
> *pair of doubles*, which allows one to get somewhat-extended
> precision, but doesn't allow the complete range of 128-bit IEEE-style.
> (I imagine the benefit is, you can do at least some of the 'long
> double' ops w/ 'double' FP ops in HW.  So, better speed... maybe.)

Oh, you're right.  We probably need to set the floatformat in one of
mips-linux or mips-irix, and the size in mips-tdep...

-- 
Daniel Jacobowitz

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[Joel Brobecker]

> According to GCC long double is 128-bit for all n64 and n32 targets.

OK. So I am planning on setting the size to 128-bit for the following
ABIs, is this the correct list:
   . O64 (??? Didn't see any mention of this in the IRIX man pages)
   . EABI32, EABI64
   . N32/N64
I will add the same set_gdbarch_...() call in each section of the case
statement on mips_abi.

Would that be the correct patch?

I am making extra sure before doing the testing because running the
testsuite takes several hours even after my fixes (it was even worse
before, it was taking 9-10 hours, and I think this is because of bugs
in expect/dejagnu that don't detect when the debugger dies or when
the debugger expects an answer from the user after having reported
an internal error - I just don't want to go there and fix the problems
in GDB instead). And to make things worse, the run sometimes get
screwed up for some reason I don't understand.  So I want to do
the testing overnight if I can and try to avoid any unecessary run.

Thanks,
-- 
Joel

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[Daniel Jacobowitz]

On Wed, Jul 21, 2004 at 07:08:54PM -0700, Joel Brobecker wrote:
> > According to GCC long double is 128-bit for all n64 and n32 targets.
> 
> OK. So I am planning on setting the size to 128-bit for the following
> ABIs, is this the correct list:
>    . O64 (??? Didn't see any mention of this in the IRIX man pages)
>    . EABI32, EABI64
>    . N32/N64
> I will add the same set_gdbarch_...() call in each section of the case
> statement on mips_abi.
> 
> Would that be the correct patch?

Ehm; no, I don't think so.  I only said N32 and N64, and according to
GCC that's all it is.  The older ABIs don't have 128-bit long double.

I think the best thing would be to make this change in mips-tdep; some
additional changes for Irix and/or Linux may be necessary to get the
format right, though.

-- 
Daniel Jacobowitz

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[Joel Brobecker]

> Ehm; no, I don't think so.  I only said N32 and N64, and according to
> GCC that's all it is.  The older ABIs don't have 128-bit long double.

Gotcha, N32 and N64. I thought you confirmed that Embedded N32&N64
where also 128bit long doubles, but I was mistaken. And I ventured
that 064 would have better precision long doubles. Oh well.

> I think the best thing would be to make this change in mips-tdep; some
> additional changes for Irix and/or Linux may be necessary to get the
> format right, though.

Yes. I'll look at this when the most urgent problems are fixed.

The unwinder seems a bit shaky for instance :-(.

Thanks Daniel, hopefully I'll be able to send a patch late tonight
or tomorrow morning.

-- 
Joel

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[Daniel Jacobowitz]

On Wed, Jul 21, 2004 at 07:34:14PM -0700, Joel Brobecker wrote:
> > Ehm; no, I don't think so.  I only said N32 and N64, and according to
> > GCC that's all it is.  The older ABIs don't have 128-bit long double.
> 
> Gotcha, N32 and N64. I thought you confirmed that Embedded N32&N64
> where also 128bit long doubles, but I was mistaken. And I ventured
> that 064 would have better precision long doubles. Oh well.

Embedded n32 and n64 -> mips64-elf -mabi=n32 or -mabi=n64.  EABI64 and
N64 have no relation.

> > I think the best thing would be to make this change in mips-tdep; some
> > additional changes for Irix and/or Linux may be necessary to get the
> > format right, though.
> 
> Yes. I'll look at this when the most urgent problems are fixed.
> 
> The unwinder seems a bit shaky for instance :-(.

Yeah, I've noticed some similar problems on mips64-linux but not had
time to go back to them.

-- 
Daniel Jacobowitz

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[David Anderson]

>OK. So I am planning on setting the size to 128-bit for the following
>ABIs, is this the correct list:
>   . O64 (??? Didn't see any mention of this in the IRIX man pages)

That's because there is no such thing as IRIX O64 (that I
ever heard of). O64 must be non-IRIX?

David B. Anderson davea at sgi dot com http://reality.sgiweb.org/davea

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[Andrew Cagney]

>>OK. So I am planning on setting the size to 128-bit for the following
>>>ABIs, is this the correct list:
>>>   . O64 (??? Didn't see any mention of this in the IRIX man pages)
> 
> 
> That's because there is no such thing as IRIX O64 (that I
> ever heard of). O64 must be non-IRIX?

It was a widely used cygnus creation (I don't longer remember why it was 
added).  I'm left wondering if it is still needed.

Andrew

Re: [RFA] mips-irix SEGV: long doubles are 128 bits long on IRIX

[cgd]

At Mon, 26 Jul 2004 22:19:59 +0000 (UTC), "Andrew Cagney" wrote:
> >>OK. So I am planning on setting the size to 128-bit for the following
> >>>ABIs, is this the correct list:
> >>>   . O64 (??? Didn't see any mention of this in the IRIX man pages)
> > That's because there is no such thing as IRIX O64 (that I
> > ever heard of). O64 must be non-IRIX?
> 
> It was a widely used cygnus creation (I don't longer remember why it
> was added).  I'm left wondering if it is still needed.

FWIW, it is still in somewhat-wide use for embedded purposes.  In
addition to being used by some 'random' embedded toolchains, IIRC
VxWorks also uses it for 64-bit MIPS platforms right now.

I seem to recall that some people have also attempted to use it to get
64-bit-ish userland (incl. shlibs) for some open-source unix-like
OSes.  (That makes me ... sad.  8-)


That doesn't mean that it's a good idea.  Just saying that it is, in
fact, still fairly widely used.



chris