[PATCH] Make frame saves identical for setjmp and non_local_goto in function.c

[PATCH] Make frame saves identical for setjmp and non_local_goto in function.c

[Andy H]

[-- Attachment #1: Type: text/plain, Size: 1862 bytes --]

This very simple  patch corrects an anomaly in the generation of 
non-local goto information.

The patch makes the saved frame value identical for both  nested 
function gotos and setjmps.

As no gcc hosted target (yet) uses TARGET_BUILTIN_SETJMP_FRAME_VALUE, 
the change has no effect.
However, it is a prerequisite  to correct AVR setjmp problems with an 
optimal solution.

If this is acceptable change, please approve for 4.4 mainline.

best regards


Background
------------

To perform either setjmp or goto from a nested function, the frame and 
stack pointers are saved in buffer.

This pointer save code is created in two places.

1) For setjmp  it is created by expand_builtin_setjmp_setup
2) For nested function goto  it is created by: expand_function_start

For setjmp the frame pointer value is saved using:

    emit_move_insn (mem, targetm.builtin_setjmp_frame_value ())

However, expand_function_start uses:

      emit_move_insn (r_save, virtual_stack_vars_rtx)

As both use same goto/receiver code they should  to be identical. This 
works only because of the default :
for TARGET_BUILTIN_SETJMP_FRAME_VALUE  is  virtual_stack_vars_rtx

This target hook  allows the true value of the frame_pointer to be saved 
and restored. Whereas
virtual_stack_vars_rtx will incur addition and subtraction overheads if  
STARTING_FRAME_OFFSET !=0.
However, because of the mismatch any  attempt  to utilise this hook, 
will cause either goto or setjmp to fail.

In the AVR target, this is the case and  I wish to set

setjmp_frame_value = virtual_stack_vars_rtx - STARTING_FRAME_OFFSET

which eliminate add/sub operations on frame pointer and minimizes the 
receiver code to nothing.
(and allow me to fix another bug)



2008-05-22    Andy Hutchinson   <hutchinsonandy@aim.com>

  *  gcc/function.c: Include target hook for nonlocal_goto frame value.









[-- Attachment #2: funcFrame.patch --]
[-- Type: text/plain, Size: 481 bytes --]

Index: function.c
===================================================================
--- function.c  (revision 135702)
+++ function.c  (working copy)
@@ -4231,7 +4231,7 @@
       r_save = expand_expr (t_save, NULL_RTX, VOIDmode, EXPAND_WRITE);
       r_save = convert_memory_address (Pmode, r_save);
 
-      emit_move_insn (r_save, virtual_stack_vars_rtx);
+      emit_move_insn (r_save, targetm.builtin_setjmp_frame_value ());
       update_nonlocal_goto_save_area ();
     }
 

Re: [PATCH] Make frame saves identical for setjmp and non_local_goto in function.c

[Richard Guenther]

On Thu, 22 May 2008, Andy H wrote:

> This very simple  patch corrects an anomaly in the generation of non-local
> goto information.
> 
> The patch makes the saved frame value identical for both  nested function
> gotos and setjmps.
> 
> As no gcc hosted target (yet) uses TARGET_BUILTIN_SETJMP_FRAME_VALUE, the
> change has no effect.
> However, it is a prerequisite  to correct AVR setjmp problems with an optimal
> solution.
> 
> If this is acceptable change, please approve for 4.4 mainline.
> 
> best regards
> 
> 
> Background
> ------------
> 
> To perform either setjmp or goto from a nested function, the frame and stack
> pointers are saved in buffer.
> 
> This pointer save code is created in two places.
> 
> 1) For setjmp  it is created by expand_builtin_setjmp_setup
> 2) For nested function goto  it is created by: expand_function_start
> 
> For setjmp the frame pointer value is saved using:
> 
>    emit_move_insn (mem, targetm.builtin_setjmp_frame_value ())
> 
> However, expand_function_start uses:
> 
>      emit_move_insn (r_save, virtual_stack_vars_rtx)
> 
> As both use same goto/receiver code they should  to be identical. This works
> only because of the default :
> for TARGET_BUILTIN_SETJMP_FRAME_VALUE  is  virtual_stack_vars_rtx
> 
> This target hook  allows the true value of the frame_pointer to be saved and
> restored. Whereas
> virtual_stack_vars_rtx will incur addition and subtraction overheads if
> STARTING_FRAME_OFFSET !=0.
> However, because of the mismatch any  attempt  to utilise this hook, will
> cause either goto or setjmp to fail.
> 
> In the AVR target, this is the case and  I wish to set
> 
> setjmp_frame_value = virtual_stack_vars_rtx - STARTING_FRAME_OFFSET
> 
> which eliminate add/sub operations on frame pointer and minimizes the receiver
> code to nothing.
> (and allow me to fix another bug)

And you tested this on ... ?

Otherwise this is ok for mainline.

Thanks,
Richard.


> 2008-05-22    Andy Hutchinson   <hutchinsonandy@aim.com>
> 
>  *  gcc/function.c: Include target hook for nonlocal_goto frame value.

Re: [PATCH] Make frame saves identical for setjmp and non_local_goto in function.c

[Andy H]

Sorry for missing that out.

it was tested with/without patch for all gcc tests on Native 
configuration is i686-pc-linux-gnu

No differences in result.

Thank for prompt reply!
Andy
> And you tested this on ... ?
>
> Otherwise this is ok for mainline.
>
> Thanks,
> Richard.
>
>
> > 2008-05-22    Andy Hutchinson   <hutchinsonandy@aim.com>
> > 
> >  *  gcc/function.c: Include target hook for nonlocal_goto frame value.
>