[rfa:dwarf2] Use frame_unwind_address_in_block

[rfa:dwarf2] Use frame_unwind_address_in_block

[Andrew Cagney]

[-- Attachment #1: Type: text/plain, Size: 378 bytes --]

Hello,

This, finally, is the tweak to the dwarf2-unwind code, to use the 
frame's address-in-block, instead of the frame's PC to select the CFI 
info.  Two test cases stop failing:

1gdb.base/corefile.exp: backtrace in corefile.exp FAIL PASS 
2gdb.base/gdb1250.exp: backtrace from abort KFAIL PASS

Ok for the mainline?
Ok to pull something equivalent into 6.0 branch?

Andrew

[-- Attachment #2: diffs --]
[-- Type: text/plain, Size: 2565 bytes --]

2003-07-16  Andrew Cagney  <cagney@redhat.com>

	* dwarf2-frame.c (dwarf2_frame_sniffer): Use
	frame_unwind_address_in_block, instead of frame_pc_unwind.
	(dwarf2_frame_cache): Ditto.

Index: dwarf2-frame.c
===================================================================
RCS file: /cvs/src/src/gdb/dwarf2-frame.c,v
retrieving revision 1.9
diff -u -r1.9 dwarf2-frame.c
--- dwarf2-frame.c	16 Jul 2003 22:29:13 -0000	1.9
+++ dwarf2-frame.c	16 Jul 2003 23:13:29 -0000
@@ -491,15 +491,12 @@
      done for "normal" frames and not for resume-type frames (signal
      handlers, sentinel frames, dummy frames).
 
-     We don't do what GCC's does here (yet).  It's not clear how
-     reliable the method is.  There's also a problem with finding the
-     right FDE; see the comment in dwarf_frame_p.  If dwarf_frame_p
-     selected this frame unwinder because it found the FDE for the
-     next function, using the adjusted return address might not yield
-     a FDE at all.  The problem isn't specific to DWARF CFI; other
-     unwinders loose in similar ways.  Therefore it's probably
-     acceptable to leave things slightly broken for now.  */
-  fs->pc = frame_pc_unwind (next_frame);
+     frame_unwind_address_in_block does just this.
+
+     It's not clear how reliable the method is though - there is the
+     potential for the register state pre-call being different to that
+     on return.  */
+  fs->pc = frame_unwind_address_in_block (next_frame);
 
   /* Find the correct FDE.  */
   fde = dwarf2_frame_find_fde (&fs->pc);
@@ -710,15 +707,11 @@
 const struct frame_unwind *
 dwarf2_frame_sniffer (struct frame_info *next_frame)
 {
-  CORE_ADDR pc = frame_pc_unwind (next_frame);
-  /* The way GDB works, this function can be called with PC just after
-     the last instruction of the function we're supposed to return the
-     unwind methods for.  In that case we won't find the correct FDE;
-     instead we find the FDE for the next function, or we won't find
-     an FDE at all.  There is a possible solution (see the comment in
-     dwarf2_frame_cache), GDB doesn't pass us enough information to
-     implement it.  */
-  if (dwarf2_frame_find_fde (&pc))
+  /* Grab an address that is guarenteed to reside somewhere within the
+     function.  frame_pc_unwind(), for a no-return next function, can
+     end up returning something past the end of this function's body.  */
+  CORE_ADDR block_addr = frame_unwind_address_in_block (next_frame);
+  if (dwarf2_frame_find_fde (&block_addr))
     return &dwarf2_frame_unwind;
 
   return NULL;

Re: [rfa:dwarf2] Use frame_unwind_address_in_block

[Andrew Cagney]

[-- Attachment #1: Type: text/plain, Size: 446 bytes --]

> Hello,
> 
> This, finally, is the tweak to the dwarf2-unwind code, to use the frame's address-in-block, instead of the frame's PC to select the CFI info.  Two test cases stop failing:
> 
> 1gdb.base/corefile.exp: backtrace in corefile.exp FAIL PASS 2gdb.base/gdb1250.exp: backtrace from abort KFAIL PASS
> 
> Ok for the mainline?
> Ok to pull something equivalent into 6.0 branch?

FYI,

The attached is the equivalent for the branch.

Andrew


[-- Attachment #2: diffs --]
[-- Type: text/plain, Size: 4813 bytes --]

2003-07-18  Andrew Cagney  <cagney@redhat.com>

	* dwarf2-frame.h (dwarf2_frame_sniffer): Replace "dwarf2_frame_p".
	* dwarf2-frame.c (dwarf2_frame_sniffer): Replace "dwarf2_frame_p".
	(dwarf2_frame_cache): Use frame_unwind_address_in_block, instead
	of frame_pc_unwind.
	* i386-tdep.c (i386_gdbarch_init): Update.
	* alpha-tdep.c (alpha_dwarf2_init_abi): Update.

Index: alpha-tdep.c
===================================================================
RCS file: /cvs/src/src/gdb/alpha-tdep.c,v
retrieving revision 1.116
diff -u -r1.116 alpha-tdep.c
--- alpha-tdep.c	14 Jun 2003 22:35:23 -0000	1.116
+++ alpha-tdep.c	18 Jul 2003 15:12:50 -0000
@@ -1583,7 +1583,7 @@
 void
 alpha_dwarf2_init_abi (struct gdbarch_info info, struct gdbarch *gdbarch)
 {
-  frame_unwind_append_predicate (gdbarch, dwarf2_frame_p);
+  frame_unwind_append_sniffer (gdbarch, dwarf2_frame_sniffer);
   frame_base_append_predicate (gdbarch, dwarf2_frame_base_p);
   set_gdbarch_dwarf2_build_frame_info (gdbarch, dwarf2_build_frame_info);
 }
Index: dwarf2-frame.c
===================================================================
RCS file: /cvs/src/src/gdb/dwarf2-frame.c,v
retrieving revision 1.7.4.1
diff -u -r1.7.4.1 dwarf2-frame.c
--- dwarf2-frame.c	11 Jul 2003 16:56:09 -0000	1.7.4.1
+++ dwarf2-frame.c	18 Jul 2003 15:12:51 -0000
@@ -491,15 +491,12 @@
      done for "normal" frames and not for resume-type frames (signal
      handlers, sentinel frames, dummy frames).
 
-     We don't do what GCC's does here (yet).  It's not clear how
-     reliable the method is.  There's also a problem with finding the
-     right FDE; see the comment in dwarf_frame_p.  If dwarf_frame_p
-     selected this frame unwinder because it found the FDE for the
-     next function, using the adjusted return address might not yield
-     a FDE at all.  The problem isn't specific to DWARF CFI; other
-     unwinders loose in similar ways.  Therefore it's probably
-     acceptable to leave things slightly broken for now.  */
-  fs->pc = frame_pc_unwind (next_frame);
+     frame_unwind_address_in_block does just this.
+
+     It's not clear how reliable the method is though - there is the
+     potential for the register state pre-call being different to that
+     on return.  */
+  fs->pc = frame_unwind_address_in_block (next_frame);
 
   /* Find the correct FDE.  */
   fde = dwarf2_frame_find_fde (&fs->pc);
@@ -708,16 +705,13 @@
 };
 
 const struct frame_unwind *
-dwarf2_frame_p (CORE_ADDR pc)
+dwarf2_frame_sniffer (struct frame_info *next_frame)
 {
-  /* The way GDB works, this function can be called with PC just after
-     the last instruction of the function we're supposed to return the
-     unwind methods for.  In that case we won't find the correct FDE;
-     instead we find the FDE for the next function, or we won't find
-     an FDE at all.  There is a possible solution (see the comment in
-     dwarf2_frame_cache), GDB doesn't pass us enough information to
-     implement it.  */
-  if (dwarf2_frame_find_fde (&pc))
+  /* Grab an address that is guarenteed to reside somewhere within the
+     function.  frame_pc_unwind(), for a no-return next function, can
+     end up returning something past the end of this function's body.  */
+  CORE_ADDR block_addr = frame_unwind_address_in_block (next_frame);
+  if (dwarf2_frame_find_fde (&block_addr))
     return &dwarf2_frame_unwind;
 
   return NULL;
Index: dwarf2-frame.h
===================================================================
RCS file: /cvs/src/src/gdb/dwarf2-frame.h,v
retrieving revision 1.1
diff -u -r1.1 dwarf2-frame.h
--- dwarf2-frame.h	31 May 2003 19:18:05 -0000	1.1
+++ dwarf2-frame.h	18 Jul 2003 15:12:51 -0000
@@ -29,7 +29,7 @@
 /* Return the frame unwind methods for the function that contains PC,
    or NULL if it can't be handled by DWARF CFI frame unwinder.  */
 
-const struct frame_unwind *dwarf2_frame_p (CORE_ADDR pc);
+const struct frame_unwind *dwarf2_frame_sniffer (struct frame_info *next_frame);
 
 /* Return the frame base methods for the function that contains PC, or
    NULL if it can't be handled by the DWARF CFI frame unwinder.  */
Index: i386-tdep.c
===================================================================
RCS file: /cvs/src/src/gdb/i386-tdep.c,v
retrieving revision 1.158.2.2
diff -u -r1.158.2.2 i386-tdep.c
--- i386-tdep.c	2 Jul 2003 17:15:54 -0000	1.158.2.2
+++ i386-tdep.c	18 Jul 2003 15:12:52 -0000
@@ -1811,7 +1811,7 @@
   set_gdbarch_fetch_pointer_argument (gdbarch, i386_fetch_pointer_argument);
 
   /* Hook in the DWARF CFI frame unwinder.  */
-  frame_unwind_append_predicate (gdbarch, dwarf2_frame_p);
+  frame_unwind_append_sniffer (gdbarch, dwarf2_frame_sniffer);
   set_gdbarch_dwarf2_build_frame_info (gdbarch, dwarf2_build_frame_info);
 
   frame_base_set_default (gdbarch, &i386_frame_base);

Re: [rfa:dwarf2] Use frame_unwind_address_in_block

[Elena Zannoni]

Andrew Cagney writes:
 > Hello,
 > 
 > This, finally, is the tweak to the dwarf2-unwind code, to use the 
 > frame's address-in-block, instead of the frame's PC to select the CFI 
 > info.  Two test cases stop failing:
 > 
 > 1gdb.base/corefile.exp: backtrace in corefile.exp FAIL PASS 
 > 2gdb.base/gdb1250.exp: backtrace from abort KFAIL PASS
 > 
 > Ok for the mainline?
 > Ok to pull something equivalent into 6.0 branch?

seems sane.

elena

 > 
 > Andrew
 > 2003-07-16  Andrew Cagney  <cagney@redhat.com>
 > 
 > 	* dwarf2-frame.c (dwarf2_frame_sniffer): Use
 > 	frame_unwind_address_in_block, instead of frame_pc_unwind.
 > 	(dwarf2_frame_cache): Ditto.
 > 
 > Index: dwarf2-frame.c
 > ===================================================================
 > RCS file: /cvs/src/src/gdb/dwarf2-frame.c,v
 > retrieving revision 1.9
 > diff -u -r1.9 dwarf2-frame.c
 > --- dwarf2-frame.c	16 Jul 2003 22:29:13 -0000	1.9
 > +++ dwarf2-frame.c	16 Jul 2003 23:13:29 -0000
 > @@ -491,15 +491,12 @@
 >       done for "normal" frames and not for resume-type frames (signal
 >       handlers, sentinel frames, dummy frames).
 >  
 > -     We don't do what GCC's does here (yet).  It's not clear how
 > -     reliable the method is.  There's also a problem with finding the
 > -     right FDE; see the comment in dwarf_frame_p.  If dwarf_frame_p
 > -     selected this frame unwinder because it found the FDE for the
 > -     next function, using the adjusted return address might not yield
 > -     a FDE at all.  The problem isn't specific to DWARF CFI; other
 > -     unwinders loose in similar ways.  Therefore it's probably
 > -     acceptable to leave things slightly broken for now.  */
 > -  fs->pc = frame_pc_unwind (next_frame);
 > +     frame_unwind_address_in_block does just this.
 > +
 > +     It's not clear how reliable the method is though - there is the
 > +     potential for the register state pre-call being different to that
 > +     on return.  */
 > +  fs->pc = frame_unwind_address_in_block (next_frame);
 >  
 >    /* Find the correct FDE.  */
 >    fde = dwarf2_frame_find_fde (&fs->pc);
 > @@ -710,15 +707,11 @@
 >  const struct frame_unwind *
 >  dwarf2_frame_sniffer (struct frame_info *next_frame)
 >  {
 > -  CORE_ADDR pc = frame_pc_unwind (next_frame);
 > -  /* The way GDB works, this function can be called with PC just after
 > -     the last instruction of the function we're supposed to return the
 > -     unwind methods for.  In that case we won't find the correct FDE;
 > -     instead we find the FDE for the next function, or we won't find
 > -     an FDE at all.  There is a possible solution (see the comment in
 > -     dwarf2_frame_cache), GDB doesn't pass us enough information to
 > -     implement it.  */
 > -  if (dwarf2_frame_find_fde (&pc))
 > +  /* Grab an address that is guarenteed to reside somewhere within the
 > +     function.  frame_pc_unwind(), for a no-return next function, can
 > +     end up returning something past the end of this function's body.  */
 > +  CORE_ADDR block_addr = frame_unwind_address_in_block (next_frame);
 > +  if (dwarf2_frame_find_fde (&block_addr))
 >      return &dwarf2_frame_unwind;
 >  
 >    return NULL;

Re: [rfa:dwarf2] Use frame_unwind_address_in_block

[Andrew Cagney]

> Andrew Cagney writes:
>  > Hello,
>  > 
>  > This, finally, is the tweak to the dwarf2-unwind code, to use the 
>  > frame's address-in-block, instead of the frame's PC to select the CFI 
>  > info.  Two test cases stop failing:
>  > 
>  > 1gdb.base/corefile.exp: backtrace in corefile.exp FAIL PASS 
>  > 2gdb.base/gdb1250.exp: backtrace from abort KFAIL PASS
>  > 
>  > Ok for the mainline?
>  > Ok to pull something equivalent into 6.0 branch?
> 
> seems sane.

Committed to mainline.  I'll leave 6.0 until next week.

Andrew

Re: [rfa:dwarf2] Use frame_unwind_address_in_block

[Michael Elizabeth Chastain]

Works for me with dwarf-2, but stabs+ is still broken.
No regressions though.

I tested this in mainline with:

  native i686-pc-linux-gnu
  gcc 2.95.3, gcc 3.3, gcc HEAD
  dwarf-2 and stabs+

Michael C

===

2003-07-16  Andrew Cagney  <cagney@redhat.com>

	* dwarf2-frame.c (dwarf2_frame_sniffer): Use
	frame_unwind_address_in_block, instead of frame_pc_unwind.
	(dwarf2_frame_cache): Ditto.