controlling exception type per target

controlling exception type per target

[Robert Lipe]

I've been looking at the dwarf2 exception stuff and have a timid
suggestion that I'd like some feedback on before I actually write
the trivial (?) patch to do it.

EGCS on OpenServer generates either COFF or ELF.   For ELF, we want
DWARF2 exception handling because that's what the people that know
about such things say we want.   For COFF, we run into two problems 
with the native assembler.

1) The section names are too long.   I have submitted a patch to cure
   this.

2) The DWARF2 stuff (reasonably) expects an assembler that can handle
   forward references.   

Since DWARF and COFF don't go together (right?) does it make more sense 
for us to just set 'exceptions_via_longjmp' if we're emitting COFF?   
Some preliminary poking around in the 924 snapshot shows that if I just
set this variable either via the debugger or via -fsjlj-exceptions while
generating COFF, we don't trigger the lameness in the assembler which 
eliminates both of the above problems.

To do this, we need a way that the target files (ala sco5.h) can whack
this variable at runtime.   

Would it be reasonable to provide a line in, say, init_eh() that could
override the constant initialization of exceptions_via_longjmp at the 
top of exception.c?   Then we could stick something like

#define DEFAULT_SJLJ_EXCEPTION \
	(TARGET_ELF) ? 0 : 1 

in the affected targets.

In an appropriate place (init_eh()?) we could then say

#if defined (DEFAULT_SJLJ_EXCEPTION)
	exceptions_via_longjmp = DEFAULT_SJLJ_EXCEPTION ;
#endif

We can't do it where this variable is now being set becuase it has
to be a constant.

Is there a better way to do this?    It looks like resetting it 
in init_eh() works for at least the trivial case.

Thanx in advance.

-- 
Robert Lipe       http://www.dgii.com/people/robertl       robertl@dgii.com

Re: controlling exception type per target

[Robert Lipe]

On Sep 25, Robert Lipe wrote:
> I've been looking at the dwarf2 exception stuff and have a timid
> suggestion that I'd like some feedback on before I actually write
> the trivial (?) patch to do it.
> 
[ explanation of difficulties encountered with current scheme deleted ] 
> 
> Would it be reasonable to provide a line in, say, init_eh() that could
> override the constant initialization of exceptions_via_longjmp at the 
> top of exception.c?   Then we could stick something like
> 
> #define DEFAULT_SJLJ_EXCEPTION \
> 	(TARGET_ELF) ? 0 : 1 
> 
> in the affected targets.
> 
> In an appropriate place (init_eh()?) we could then say
> 
> #if defined (DEFAULT_SJLJ_EXCEPTION)
> 	exceptions_via_longjmp = DEFAULT_SJLJ_EXCEPTION ;
> #endif

In the silence that ensued, I did implement a variation this.   This 
allowed me to get all the library stuff built for both COFF (sjlj) and ELF 
(dwarf2unwind) but I then ran into problem with the crtstuff.c file
where it tried to do dwarf things when it shouldn't have (on this target).

Should I expand this scheme on into crtstuff as well?

If someone could give me a little guidance on what would be the 
preferred direction to go with this (there are a couple) I don't 
mind doing the coding and testing.    I just don't want to have
to do it twice...

Thanx,
RJL

Re: controlling exception type per target

[Jason Merrill]

Sorry I haven't gotten back to you before.  I think this should be handled
by setting up exceptions_via_longjmp in main(), with a simple

#ifdef DWARF2_UNWIND_INFO
exceptions_via_longjmp = DWARF2_UNWIND_INFO;
#endif

> Should I expand this scheme on into crtstuff as well?

You can tell whether you are doing COFF or ELF with #ifdefs, right?  I
would think you could special-case the SCO tm.h so that when CRT_BEGIN or
CRT_END are defined, you use unconditional defines.  irix6.h does something
similar for n32/n64 multilibbing.

Jason

Patches for egcs/OSR5 Was: controlling exception type per target

[Robert Lipe]

> Sorry I haven't gotten back to you before.  

'Tis quite OK.  I'm glad I waited for your reply, becuase I was 
indeed headed down a bad path.    Your suggestions took me where
I needed to go.  Thanx, Jason!

> I think this should be handled
> by setting up exceptions_via_longjmp in main(), with a simple
> 
> #ifdef DWARF2_UNWIND_INFO
> exceptions_via_longjmp = DWARF2_UNWIND_INFO;
> #endif

The negative logic definitions took a while for me to grok.

I'll leave submitting/applying that patch to you.   I stuck it in 
my copy in what seemed like an OK place.  It is required.   

*** toplev.c_   Mon Sep 15 21:07:28 1997
--- toplev.c    Wed Oct  1 22:47:54 1997
***************
*** 4289,4294 ****
--- 4289,4298 ----
    OVERRIDE_OPTIONS;
  #endif

+ #ifdef DWARF2_UNWIND_INFO
+   exceptions_via_longjmp = DWARF2_UNWIND_INFO ;
+ #endif
+
    if (profile_block_flag == 3)
      {
        warning ("`-ax' and `-a' are conflicting options. `-a' ignored.");


> > Should I expand this scheme on into crtstuff as well?
> 
> You can tell whether you are doing COFF or ELF with #ifdefs, right?  I
> would think you could special-case the SCO tm.h so that when CRT_BEGIN or
> CRT_END are defined, you use unconditional defines.  irix6.h does something
> similar for n32/n64 multilibbing.

You could.   I think I've found that if I just let crtstuff spill
the eh sections, it's OK as long as I pick a name that makes the 
assembler not vomit.  It seems that they're never referenced.  I decided
to go ahead and put the scaffolding for EH_FRAMEs in for COFF in case
they come in handy in the future.

I haven't rerun the full make bootstrap/runtest with it, but I did
do a 'make bootstrap3 followed by a 

nice runtest --outdir ${COFFDIR} --tool g++  \
        --srcdir ${SRC}/gcc/testsuite \
        --target_board 'unix{-mcoff,-melf}' \
        GXX_UNDER_TEST=negcs \
        CXXFLAGS="-g -O" \
        old-deja.exp=eh*
 
and got really promising results:

Test Run By robertl on Thu Oct  2 00:41:57 1997
Native configuration is i586-pc-sco3.2v5.0.4

                === g++ tests ===

Running target unix/-mcoff
Running /play/egcs-970917/gcc/testsuite/g++.old-deja/old-deja.exp ...

                === g++ Summary for unix/-mcoff ===

# of expected passes            103
Running target unix/-melf
Running /play/egcs-970917/gcc/testsuite/g++.old-deja/old-deja.exp ...

                === g++ Summary for unix/-melf ===

# of expected passes            103

                === g++ Summary ===

# of expected passes            206
negcs version egcs-2.90.11 970929 (gcc2-970802 experimental)



So please consider this patch for inclusion:


Thu Oct  2 00:30:15 1997  Robert Lipe (robertl@dgii.com)

	* sco5.h: Let ELF use dwarf2 unwinding.  COFF uses sjlj
	(EH_FRAME_SECTION_ASM_OP, EH_FRAME_SECTION_ASM_OP_ELF): Defined.
	(EH_FRAME_SECTION_ASM_OP_COFF): Likewise.
	(DWARF2_UNWIND_INFO): Let this track object file format.
	(EXTRA_SECTIONS): Add in_eh.
	(EH_FRAME_SECTION_ASM_OP, EH_FRAME_SECTION_ASM_OP_ELF, EH_FRAME_SECTION_ASM_OP_COFF): defined


*** config/i386/sco5.h_	Wed Oct  1 22:35:32 1997
--- config/i386/sco5.h	Wed Oct  1 23:17:35 1997
***************
*** 76,81 ****
--- 76,92 ----
  #undef GLOBAL_ASM_OP
  #define GLOBAL_ASM_OP			"\t.globl"
  
+ #undef EH_FRAME_SECTION_ASM_OP
+ #define EH_FRAME_SECTION_ASM_OP_COFF	"\t.section\t.ehfram, \"x\""
+ #define EH_FRAME_SECTION_ASM_OP_ELF	"\t.section\t.eh_frame, \"aw\""
+ #define EH_FRAME_SECTION_ASM_OP	\
+   ((TARGET_ELF) ? EH_FRAME_SECTION_ASM_OP_ELF : EH_FRAME_SECTION_ASM_OP_COFF)
+ 
+ /* Avoid problems (long sectino names, forward assembler refs) with DWARF
+    exception unwinding when we're generating COFF */
+ #define DWARF2_UNWIND_INFO	\
+   ((TARGET_ELF) ? 1 : 0 )  
+ 
  #undef CONST_SECTION_ASM_OP
  #define CONST_SECTION_ASM_OP_COFF	"\t.section\t.rodata, \"x\""
  #define CONST_SECTION_ASM_OP_ELF	"\t.section\t.rodata"
***************
*** 563,569 ****
    ((TARGET_ELF) ? DWARF_DEBUG: SDB_DEBUG)
  
  #undef EXTRA_SECTIONS
! #define EXTRA_SECTIONS in_const, in_init, in_fini, in_ctors, in_dtors
  
  #undef EXTRA_SECTION_FUNCTIONS
  #define EXTRA_SECTION_FUNCTIONS						\
--- 574,580 ----
    ((TARGET_ELF) ? DWARF_DEBUG: SDB_DEBUG)
  
  #undef EXTRA_SECTIONS
! #define EXTRA_SECTIONS in_const, in_init, in_fini, in_ctors, in_dtors, in_eh
  
  #undef EXTRA_SECTION_FUNCTIONS
  #define EXTRA_SECTION_FUNCTIONS						\
***************
*** 901,906 ****
--- 912,918 ----
  # undef FINI_SECTION_ASM_OP
  # undef CTORS_SECTION_ASM_OP
  # undef DTORS_SECTION_ASM_OP
+ # undef EH_FRAME_SECTION_ASM_OP
  # undef CTOR_LIST_BEGIN
  # undef CTOR_LIST_END
  # undef DO_GLOBAL_CTORS_BODY
***************
*** 912,922 ****
--- 924,936 ----
  #  define FINI_SECTION_ASM_OP FINI_SECTION_ASM_OP_ELF
  #  define DTORS_SECTION_ASM_OP DTORS_SECTION_ASM_OP_ELF
  #  define CTORS_SECTION_ASM_OP CTORS_SECTION_ASM_OP_ELF
+ #  define EH_FRAME_SECTION_ASM_OP EH_FRAME_SECTION_ASM_OP_ELF
  # else /* ! _SCO_ELF */
  #  define INIT_SECTION_ASM_OP INIT_SECTION_ASM_OP_COFF
  #  define FINI_SECTION_ASM_OP FINI_SECTION_ASM_OP_COFF
  #  define DTORS_SECTION_ASM_OP DTORS_SECTION_ASM_OP_COFF
  #  define CTORS_SECTION_ASM_OP CTORS_SECTION_ASM_OP_COFF
+ #  define EH_FRAME_SECTION_ASM_OP ""
  #  define CTOR_LIST_BEGIN asm (INIT_SECTION_ASM_OP); asm ("pushl $0")
  #  define CTOR_LIST_END CTOR_LIST_BEGIN
  #  define DO_GLOBAL_CTORS_BODY						\

Re: Patches for egcs/OSR5 Was: controlling exception type per target

[Jason Merrill]

>>>>> Robert Lipe <robertl@dgii.com> writes:

>> #ifdef DWARF2_UNWIND_INFO
>> exceptions_via_longjmp = DWARF2_UNWIND_INFO;
>> #endif

> The negative logic definitions took a while for me to grok.

Er, you're right.  It should be !DWARF2_UNWIND_INFO.

> I'll leave submitting/applying that patch to you.

OK.

> ***************
> *** 563,569 ****
>     ((TARGET_ELF) ? DWARF_DEBUG: SDB_DEBUG)
  
>   #undef EXTRA_SECTIONS
> ! #define EXTRA_SECTIONS in_const, in_init, in_fini, in_ctors, in_dtors
  
>   #undef EXTRA_SECTION_FUNCTIONS
>   #define EXTRA_SECTION_FUNCTIONS						\
> --- 574,580 ----
>     ((TARGET_ELF) ? DWARF_DEBUG: SDB_DEBUG)
  
>   #undef EXTRA_SECTIONS
> ! #define EXTRA_SECTIONS in_const, in_init, in_fini, in_ctors, in_dtors, in_eh
  
>   #undef EXTRA_SECTION_FUNCTIONS
>   #define EXTRA_SECTION_FUNCTIONS						\

This hunk seems unnecessary.  Other than that, looks fine.

Jason