Re: [RFA] Artifical dwarf2 debug info

[Daniel Jacobowitz <drow at mvista dot com>]

On Thu, Jan 02, 2003 at 11:05:37PM +0000, Andrew Cagney wrote:
> >On Thu, Jan 02, 2003 at 08:54:33PM +0000, Andrew Cagney wrote:
> >
> >>  fde = get_fde_for_addr (context->ra - 1);
> >
> >>> >+
> >>> >+  if (fde == NULL)
> >>> >+    fde = guess_generic_fde (context->ra - 1);
> >>> >
> >>> >   if (fde == NULL)
> >>> >     return;
> >
> >>
> >>Just to be clear.  The above is the change that I think is wrong.
> >>
> >>Instead of this function `guessing' the source of the FDE, the code 
> >>needs to be re-structured so that the caller always supplies a 
> >>pre-created FDE.
> >>
> >>That way a dwarf2 cfi frame can call the above function with an FDE 
> >>built from the object files debug info, while an artifical frame can 
> >>call it with an artifically created FDE.   There is no guessing involved.
> >
> >
> >Hold a second here; I don't think we're really communicating on what
> >this change is supposed to do.  Look at where that code is: it's in
> >frame_state_for.  Its inputs are a CONTEXT and FS (struct frame_state
> >*).  The first line in your quote is:
> >  fde = get_fde_for_addr (context->ra - 1);
> 
> Right.  That gets removed.  Instead that info gets passed to the CFI 
> code as a `parameter' (perhaphs explicitly, or perhaps implicitly as 
> part of a member of the CONTEXT object).

This I don't get at all.  Why?  What advantage is there?  Whose
business besides the CFI reader's is it whether or not there is an FDE?

> The latter can be further broken down into:
> 
> dwarf2 fde / cfi frames
> saved-register frames
> artifical fde / cfi frames
> ...

I still don't see the point of this distinction.  What differentiates a
saved-register frame from a dwarf2 fde / cfi frame, anyway?  There are
saved registers; we figure out where they are via CFI.

> >Maybe there will be others, but notice that all the above are
> >conceptually different kinds of things.  These "artifical" frames are
> >just normal frames, where we synthesize the debug information because
> >we didn't have any.  It's a mechanism to coalesce things like prologue
> >readers.  It is absolutely not a new type of frame.
> >
> >That's why I think this code is in exactly the right place, right now. 
> >Are you saying that the CFI code should just be returning, saying "no
> >idea, go away, don't talk to me", and leaving this be? 
> 
> Why was the dwarf2cfi code even called?  Since there is no dwarf2 cfi 
> that code path should not have been reached.  Per my comment below, this 
> would have happend because the caller (or something up the stack) failed 
> to check for an edge condition.  That change is patching things up after 
> the event.
> 
> Instead, during `struct frame_info' creation, if there isn't any dwarf2 
> info, and the architecture really wants to use the dwarf2cfi logic, it 
> should create an `artifical fde / cfi frame' that first fakes up the FDE 
> info and then supplies that to the dwarf2cfi logic.  Similarly, a dwarf2 
> cfi frame can first read the fde and then call the relevant code.

It sounds to me as if you want to move the call to the frame creation
logic somewhere else and that's it.  The place you want to move it
doesn't exist yet, as far as I can tell.  Am I right?

> > That's all well
> > and good but that way we end up duplicating the whole of the CFI
> > reader.  A good long term direction, with appropriate code factoring,
> > but it's hardly practical.
> 
> How does this result in the duplication of the CFI reader?

The entire CFA program execution and unwind logic would be exactly
identical to that of the "artificial frame" case.

> >>This is part of a long standing problem - it predates dwarf2cf by many 
> >>years.  Instead of using recursion, people modify debug/target dependent 
> >>frame code so that it attempts to directly handle all cases.  Cf all the 
> >>PC_IN_CALL_DUMMY(frame->next), PC_IN_SIGTRAMP(frame->next) and other 
> >>tests scattered through out the -tdep.c code; and the calls to 
> >>get_next_frame() in dwarf2cfi.c.
> >
> >
> >The one call to get_next_frame, which parallels init_frame_pc_default.
> 
> Right.  And init_frame_pc_default() is, again, typical of the problem. 
> It shouldn't need to refer to frame->next.

I can't claim to understand how that would work.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer