[Bug target/14202] New: [arm] Thumb __builtin_setjmp not interworking safe

[Bug target/14202] New: [arm] Thumb __builtin_setjmp not interworking safe

[drow at gcc dot gnu dot org]

There are two problems.  The first is simple, indirect jumps don't use bx.
arm_indirect_jump and thumb_indirect_jump are the only affected patterns,
the "fix" is obvious.

The second is triggered by that fix.  __builtin_setjmp works by saving
the address of a label, and later jumping to it in __builtin_longjmp.
The label has its low bit clear; if I fix the arm_indirect_jump to use bx, then
the bx will set us back into ARM mode.

So if the exception is thrown from ARM code and caught in Thumb code, we'll get
illegal instructions all over the place when we return to the Thumb code.  With
the use of bx this is probably true for exceptions within Thumb code also.

The obvious thing to do would be to mark the low bit of the label.  I couldn't
make gas do it; it explicitly rejects .thumb_func for local labels.  If I use
.thumb_set then it looks like the result will work, but the label becomes a symbol
and clutters up disassembly.

An alternative would be to explicitly or the lower bit of the label.  It could be
done at runtime in builtin_setjmp_setup or at compiletime by emitting
  .word L12 + 1
instead of
  .word L12
which would require a change to expand_builtin_setjmp.

Here's the trivial testcase I've been using.

int main ()
{
  try {
    throw 1;
  }
  catch (int)
    { return 0; }
  catch (...)
    { return 0; }
  return 1;
}

-- 
           Summary: [arm] Thumb __builtin_setjmp not interworking safe
           Product: gcc
           Version: 3.4.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: target
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: drow at gcc dot gnu dot org
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: i686-pc-linux-gnu
  GCC host triplet: i686-pc-linux-gnu
GCC target triplet: arm-elf


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14202

[Bug target/14202] [arm] Thumb __builtin_setjmp not interworking safe

[drow at gcc dot gnu dot org]

------- Additional Comments From drow at gcc dot gnu dot org  2004-02-19 00:15 -------
Created an attachment (id=5768)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=5768&action=view)
Ugly patch for Thumb interworking exceptions


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14202

[Bug target/14202] [arm] Thumb __builtin_setjmp not interworking safe

[pinskia at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
  GCC build triplet|i686-pc-linux-gnu           |
   GCC host triplet|i686-pc-linux-gnu           |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14202

[Bug target/14202] [arm] Thumb __builtin_setjmp not interworking safe

[rearnsha at gcc dot gnu dot org]

------- Additional Comments From rearnsha at gcc dot gnu dot org  2004-02-19 12:01 -------
I wonder whether it might be better to build unwind-sjlj.c with
DONT_USE_BUILTIN_SETJMP defined.  Not only should this fix the second problem
you are seeing, but it also fits in with the EABI philosopy, where only the
system library needs to know exactly which registers are available on the
machine at run time.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14202

[Bug target/14202] [arm] Thumb __builtin_setjmp not interworking safe

[rearnsha at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rearnsha at gcc dot gnu dot
                   |                            |org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14202

[Bug target/14202] [arm] Thumb __builtin_setjmp not interworking safe

[drow at gcc dot gnu dot org]

------- Additional Comments From drow at gcc dot gnu dot org  2004-02-19 14:56 -------
Subject: Re:  [arm] Thumb __builtin_setjmp not interworking safe

Won't that leave bits of GNAT broken?  I think it uses __builtin_setjmp
explicitly.

Also, it has occured to me that my solution does not fix the whole
problem.  Normally, emit_indirect_jump is used within a function (for
instance, TRY_FINALLY_EXPR, or computed goto).  For those mov pc is
appropriate.  But for __builtin_setjmp it isn't.  keep_stack_depressed
even has the hideous:
          jump_insn = emit_jump_insn (gen_indirect_jump (retaddr));



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14202

Re: [Bug target/14202] [arm] Thumb __builtin_setjmp not interworking safe

[Richard Earnshaw]

> 
> Won't that leave bits of GNAT broken?  I think it uses __builtin_setjmp
> explicitly.
> 

Possibly, but probably no more broken than it already is if it uses that 
routine.  How can the compiler know to restore registers it doesn't even 
know about?  The EABI has been carefully constructed to ensure that it 
needed care; but builtin_setjmp breaks that assumption.

The ultimate solution will be to convert the ARM port to use the EABI 
exception unwinding model, then most of this should be irrelevant from 
that point of view.

R.

[Bug target/14202] [arm] Thumb __builtin_setjmp not interworking safe

[rearnsha at arm dot com]

------- Additional Comments From rearnsha at arm dot com  2004-02-19 15:30 -------
Subject: Re:  [arm] Thumb __builtin_setjmp not 
 interworking safe

> 
> Won't that leave bits of GNAT broken?  I think it uses __builtin_setjmp
> explicitly.
> 

Possibly, but probably no more broken than it already is if it uses that 
routine.  How can the compiler know to restore registers it doesn't even 
know about?  The EABI has been carefully constructed to ensure that it 
needed care; but builtin_setjmp breaks that assumption.

The ultimate solution will be to convert the ARM port to use the EABI 
exception unwinding model, then most of this should be irrelevant from 
that point of view.

R.




-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14202

[Bug target/14202] [arm] Thumb __builtin_setjmp not interworking safe

[drow at gcc dot gnu dot org]

------- Additional Comments From drow at gcc dot gnu dot org  2004-02-19 17:51 -------
There are probably other affected bits of code.  As an alternate approach, I tried
emitting a +1 after all thumb-mode code labels output using arm_assemble_integer;
this results in working exceptions and working computed goto, and it even manages
not to mess up dwarf2 debug info (but only because dwarf2 bypasses
assemble_integer).
My assumption is that all such labels are being used to move into the PC
(in which case bx needs the bit and mov will ignore it); things in need of code
addresses
appear to go elsewhere.

I didn't test profiling, which may be broken by this approach.  I'd be amazed if
nothing
was.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14202

[Bug target/14202] [arm] Thumb __builtin_setjmp not interworking safe

[giovannibajo at libero dot it]

------- Additional Comments From giovannibajo at libero dot it  2004-10-11 10:57 -------
Daniel, Richard, is still bug still actual? Can we move this to confirmed 
status if you agree on the bug? Thanks.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14202

[Bug target/14202] [arm] Thumb __builtin_setjmp not interworking safe

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-07-02 01:31 -------
Confirmed.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
   Last reconfirmed|0000-00-00 00:00:00         |2005-07-02 01:31:50
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14202