Dwarf 2 Stack pointer unwind.

Dwarf 2 Stack pointer unwind.

[Con Bradley]

Hi,
   We have an ABI where the stack pointer is never saved to the stack
on a function call and there is NO frame pointer. Instead the stack is
maintained by stack pointer increments and decrements.
How does one describe this in Dwarf Call Frame Instructions? The only
rules available for register operations are to save on the stack or to
copy to another register. No rules to describe plus or negative
register adjustments.

  I find it difficult to understand why there is no support for this
in Dwarf 2 or Dwarf 3 as this must be a common idiom.

When we adjust the stack pointer we issue CFA_CFA_def_cfa operations
to describe the new rule to access the CFA, effectively issuing a CFA
rule to compensate for the stack pointer being changed. This is
sufficient to allow stack unwind to work. However, when we look up the
values of automatics in upper stack frames this fails. The reason
being that the address of these variables (in location expressions) is given
as an offset from the stack pointer but there appear to be no unwind
rules for the stack pointer register.

More correctly we'd like to say that the address of an automatic was
CFA plus offset but there doesn't appear to be any support for this.

How does one describe stack pointer changes which do not involve
saving on the stack or use of a frame pointer.

  Con Bradley.

Re: Dwarf 2 Stack pointer unwind.

[Daniel Jacobowitz]

On Tue, Feb 01, 2005 at 05:18:42PM +0000, Con Bradley wrote:
> Hi,
>    We have an ABI where the stack pointer is never saved to the stack
> on a function call and there is NO frame pointer. Instead the stack is
> maintained by stack pointer increments and decrements.
> How does one describe this in Dwarf Call Frame Instructions? The only
> rules available for register operations are to save on the stack or to
> copy to another register. No rules to describe plus or negative
> register adjustments.
> 
>   I find it difficult to understand why there is no support for this
> in Dwarf 2 or Dwarf 3 as this must be a common idiom.

I recommend you asking this on the dwarf2 mailing list instead of here.

> When we adjust the stack pointer we issue CFA_CFA_def_cfa operations
> to describe the new rule to access the CFA, effectively issuing a CFA
> rule to compensate for the stack pointer being changed. This is
> sufficient to allow stack unwind to work. However, when we look up the
> values of automatics in upper stack frames this fails. The reason
> being that the address of these variables (in location expressions) is given
> as an offset from the stack pointer but there appear to be no unwind
> rules for the stack pointer register.
> 
> More correctly we'd like to say that the address of an automatic was
> CFA plus offset but there doesn't appear to be any support for this.
> 
> How does one describe stack pointer changes which do not involve
> saving on the stack or use of a frame pointer.

A couple of ways to link the frame base (DW_AT_frame_base) to the CFA
have been proposed.  Nothing's been picked yet though.

To see how this is handled in GDB, look at
dwarf2_frame_default_init_reg.  By setting SP_REGNUM's rule to
DWARF2_FRAME_REG_CFA, we have the value of the stack pointer for the
frame.  Then you can define appropriate DW_AT_frame_base values and use
DW_OP_fbreg.

-- 
Daniel Jacobowitz