Re: fail to get ra and bp

Re: fail to get ra and bp

[ligang]

gdb-owner@sourceware.org wrote on 2006-10-12 13:39:44:

> > I am porting GDB to a new target.
> 
> Would you mind telling us what this new target is? I believe this will
> greatly help us provide more accurate answers. In the meantime...

The new target is named Score ---- 32 bit core designed by Chinese IC 
company.
Both GCC for Score and Binutils for Score have entered GNU main tree just 
now.
In future, GDB for Score is willing to enter GNU official release.
 
> > The prologue has two insns ---- push ra and push bp.
> > When back tracing and printing variables, gdb will look for return 
address 
> > of this frame and frame pointer on stack.
> > Unfortunately, If compiling with -O2, the two insns are probably 
deleted. 
> > So, GDB will not get ra and bp value.
> 
> Typical problem. My recommendation, stay away from prologue analysis
> as much as you can, because as you know, higher levels of optimization
> will break the typical frame layout. What you want, is have this
> information provided by the compiler itself and dumped into the
> objects in some form.
> 
> One very popular form, at least in the GNU world, is DWARF CFI. There
> is also another popular form, which I don't really know much about.
> The information is stored in .eh_frame sections. But it essentially
> does the same: It tells you how to unwind the call stack.

Do you know which target use .eh_frame to handle call stack?
Would you please give me more info about .eh_frame or some documents 
describing it?
 
> This is what you want because this information allows you to avoid
> tricky prologue scanning and hairy heuristics to try to guess where
> things are.
> 
> Now that I think of it, if you have no choice but do prologue scanning,
> you might want to look at a framework that was recently contributed
> which should make your task easier. I think it's prologue-value.[hc].

Could I directly call the functions in prologue-value.[hc]?
 
> -- 
> Joel

Re: fail to get ra and bp

[Daniel Jacobowitz]

On Thu, Oct 12, 2006 at 03:17:06PM +0800, ligang@sunnorth.com.cn wrote:
> Do you know which target use .eh_frame to handle call stack?
> Would you please give me more info about .eh_frame or some documents 
> describing it?

Most targets supported by GCC use it.  It's the same as DWARF CFI, just
with a slightly different encoding to support use at runtime.

> Could I directly call the functions in prologue-value.[hc]?

Take a look at the one port in the tree (maybe two?) which already
uses this code.

-- 
Daniel Jacobowitz
CodeSourcery

fail to get ra and bp

[ligang]

Hi all,


I am porting GDB to a new target.

The prologue has two insns ---- push ra and push bp.
When back tracing and printing variables, gdb will look for return address 

of this frame and frame pointer on stack.
Unfortunately, If compiling with -O2, the two insns are probably deleted. 
So, GDB will not get ra and bp value.

It seems MIPS also fail to do this. 
Does it need to change ABI? Anyone has any thought?
 
Thanks.
Ligang

Re: fail to get ra and bp

[Joel Brobecker]

> I am porting GDB to a new target.

Would you mind telling us what this new target is? I believe this will
greatly help us provide more accurate answers. In the meantime...

> The prologue has two insns ---- push ra and push bp.
> When back tracing and printing variables, gdb will look for return address 
> of this frame and frame pointer on stack.
> Unfortunately, If compiling with -O2, the two insns are probably deleted. 
> So, GDB will not get ra and bp value.

Typical problem. My recommendation, stay away from prologue analysis
as much as you can, because as you know, higher levels of optimization
will break the typical frame layout. What you want, is have this
information provided by the compiler itself and dumped into the
objects in some form.

One very popular form, at least in the GNU world, is DWARF CFI. There
is also another popular form, which I don't really know much about.
The information is stored in .eh_frame sections. But it essentially
does the same: It tells you how to unwind the call stack.

This is what you want because this information allows you to avoid
tricky prologue scanning and hairy heuristics to try to guess where
things are.

Now that I think of it, if you have no choice but do prologue scanning,
you might want to look at a framework that was recently contributed
which should make your task easier. I think it's prologue-value.[hc].

-- 
Joel

fail to get ra and bp

[bjgnu]

Hi all,

The prologue has two insns ---- push ra and push bp.
When back tracing and printing variables, gdb will look for return address 
of this frame and frame pointer on stack.
Unfortunately, If compiling with -O2, the two insns are probably deleted. 
So, GDB will not get ra and bp value.

It seems MIPS also fail to do this.
Anyone has any thought?
 
Thanks.
ligang