[_eh,dawin,Version2] Allow targets to suppress epilogues in _eh frames.

Hi Richard,

Well, of course your suggestions (and Jakub's comments in irc)   
resulted a better solution.

[the questions in a previous follow-up mail re. the GNU opcode still  
stand - but don't affect the base patch].

this has been looped around a few times on darwin9 & 10 and  
bootstrapped on x86_64-unk-linux-gnu.
Ok for trunk now?
Iain

On 14 Aug 2010, at 00:27, Richard Henderson wrote:

> On 08/13/2010 12:50 PM, IainS wrote:
>> +    DW_CFA_GNU_start_epilogue = 0x30
>
> I'd rather not place this in dwarf2.h since we don't plan to emit this
> as an actual opcode.  It only needs to be a marker in the dw_cfi_ref
> linked list for internal use of gcc.
>
> As such I think something like
>
> # define DW_CFA_INTERNAL_start_epilogue  0x100
>
> would be appropriate.  I believe you'll need to change
>
>  typedef struct GTY(()) dw_cfi_struct {
>    dw_cfi_ref dw_cfi_next;
> -   enum dwarf_call_frame_info dw_cfi_opc;
> +   unsigned int dw_cfi_opc;

done.
> in order for this to be legal.  Unfortunately that may
> require a host of other type changes or type casts in
> order to pass the C++ warning mode.
>
>> +/* True if start_epilogue should be emitted before following CFI  
>> directive.  */
>> +static bool emit_cfa_start_epilogue;
>
> I don't think this is needed...
>
>>   if (i == NULL)
>> -    return;
>> +    {
>> +      /* But we do mark the start of the epilogue to allow it to  
>> be skipped
>> +         in _eh frames.  */
>> +      emit_cfa_start_epilogue = true;
>> +      return;
>> +    }
>
> ... instead we want to add the cfi entry here, by hand.  The reason
> being that we want to avoid the DW_CFA_advance_loc4 that would be
> added by actually inserting this opcode via add_fde_cfi.

done

>> +/* Targets might not need epilogue information in dwarf2 _eh  
>> frames.  This
>> +   hook should return true if the epilogue should be suppressed in  
>> such frames.
>> +   Epilogues will still be emitted in _debug_frames.  */
>> +DEFHOOK_UNDOC
>> +(suppress_eh_epilogue_p,
>> + "",
>> + bool, (void),
>> + hook_bool_void_false)
>
> As discussed on IRC, I don't think we need a target hook for this.  We
> should key this off of !flag_asynchronous_unwind_tables.  That will
> allow all targets that only need unwind info for call sites omit this
> extra epilogue unwind info.

done.

> At that point you can adjust Darwin's override_options hook(s) to make
> sure that this flag is appropriately off.

not so simple - because toplev.c manipulates it unconditionally - I've  
adjusted toplev to set the flag only if it is unset.
and will make manipulation of default settings on darwin a separate  
patch (darwin does not yet use that hook).

>  It also seems like it would
> be a good idea to add a SPEC entry to disable compact unwind if the
> user explicitly uses -fasynchronous-unwind-info.

At present, I'm going to leave the compact unwinder disabled on  
Darwin10, it seems from the testing done by Jack that there are a  
bunch of other (probably non-gnu) issues to iron out on that yet.   
Ergo, the spec is not needed yet.

===

gcc:

	* gcc/dwarf2out.c: DW_CFA_INTERNAL_start_epilogue, CFI_T New.
	(struct dw_cfi_struct): Adjust type of dw_cfi_opc.
	(add_fde_cfi):  Handle DW_CFA_INTERNAL_start_epilogue.
	(dwarf2out_cfi_begin_epilogue): Insert epilogue marker.
	(dw_cfi_oprnd1_desc): Adjust argument type.
	(dw_cfi_oprnd2_desc): Ditto.
	(output_cfi_directive): Handle   DW_CFA_INTERNAL_start_epilogue.
	(emit_cfi_or_skip_epilogue): New.
	(output_fde): Use emit_cfi_or_skip_epilogue.
	*gcc/toplev.c (process_options): Do not set  
flag_asynchronous_unwind_tables
	unless it is unset by the target.  Set flag_unwind_tables for either
	flag_asynchronous_unwind_tables or flag_non_call_exceptions.
	* gcc/config/darwin.c (darwin_override_options): Warn the the target  
does
	not fully support flag_asynchronous_unwind_tables.  Switch  
flag_unwind_tables
	on for flag_non_call_exceptions  or flag_exceptions on darwin  
versions supporting
	_eh frames.  Ensure that all table flags are switched off for kernel  
code.

====

[_eh,dawin,Version2] Allow targets to suppress epilogues in _eh frames.

Commit Message

Comments

Patch