gdb/gdb-entryval-crash-1of3.patch
Jan Kratochvil a972d47b4c Rebase to FSF GDB 7.7.91.20140721 (pre-7.8 snapshot).
- Rebase the Intel VLA patchset.
- New fix of the optimized-out entry data values crash (BZ 1111910).
- [testsuite] Fix paginate-*.exp race for "read1".
2014-07-22 20:38:38 +02:00

195 lines
7.6 KiB
Diff

http://sourceware.org/ml/gdb-patches/2014-07/msg00530.html
Subject: [read_frame_arg patch] Handle partially optimized out values similarly to unavailable values (Re: [patchv2] Fix crash on optimized-out entry data values)
--V88s5gaDVPzZ0KCq
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
On Thu, 17 Jul 2014 14:23:06 +0200, Pedro Alves wrote:
> On 07/16/2014 10:58 PM, Jan Kratochvil wrote:
> > This patch is apparently not suitable for gdb-7.8 which is I guess often
> > crashing on -O2 -g entry values so there could be some rather minimal crash
> > avoiding patch instead.
>
> Yeah.
>
> So this was originally "caused" (more exposed) by 4f14910f:
>
> gdb/ChangeLog
> 2013-11-26 Andrew Burgess <aburgess@broadcom.com>
>
> * value.c (allocate_optimized_out_value): Mark value as non-lazy.
>
> I tried a few approaches in value_available_contents_eq
> today, and ended up thinking that the simplest should be to
> just revert that patch until we have the fuller fix in place.
OK, that seems as the best solution for 7.8 to me.
> While doing just that fixes the crash, it surprisingly causes
> one of your new tests to FAIL:
>
> (gdb) frame
> #0 bar (ref=ref@entry=@0x7fffffffd184: 10) at gdb.arch/amd64-entry-value-paramref.cc:23
> 23 vv++; /* break-here */
> (gdb) FAIL: gdb.arch/amd64-entry-value-paramref.exp: frame
There is a bug in that entry value code of mine, fix attached.
The testcase then PASSes with the reverted optimization by Andrew Burgess.
For the attached fix - if you nitpick the missing conditional case:
value_optimized_out (val_deref) && value_optimized_out (entryval_deref)
It is not detected there but that IMO does not matter much as
* It is for 7.8 only, for trunk it will get compared correctly thanks to the
new implementation of value_available_contents_eq()
called value_contents_eq().
* If the conditional
if (val != val_deref
&& !value_optimized_out (val_deref)
&& !value_optimized_out (entryval_deref)
&& value_available_contents_eq (val_deref, 0,
entryval_deref, 0,
TYPE_LENGTH (type_deref)))
val_equal = 1;
fails it may just print
bar (ref=@0x7fffffffd904: <optimized out>, ref@entry=@0x7fffffffd904: <optimized out>)
(or some variant with some partially optimized-out/unavailable parts)
instead of the more correct
bar (ref=ref@entry=@0x7fffffffd904: <optimized out>)
which is not much a bug.
The attached fix no longe makes sense after the new implementation
of value_available_contents_eq() called value_contents_eq() gets applied as it
handles all the optimized-out/unavailable values on its own, therefore the
attached patch is really only for 7.8.
> Turns out it's the code disabled in value_of_dwarf_reg_entry:
>
> target_val = dwarf_entry_parameter_to_value (parameter,
> TYPE_LENGTH (target_type),
> target_type, caller_frame,
> caller_per_cu);
>
> /* value_as_address dereferences TYPE_CODE_REF. */
> addr = extract_typed_address (value_contents (outer_val), checked_type);
>
> /* The target entry value has artificial address of the entry value
> reference. */
> VALUE_LVAL (target_val) = lval_memory;
> set_value_address (target_val, addr);
>
> It looks quite wrong to me to just change a value's lval like that.
>
> I ran the testsuite with that code disabled (like in the patch below),
> and that caused no regressions. I can't say I really understand the
> intention here though. What would we be missing if we removed that code?
I cannot reproduce any wrong case having the code above #if 0-ed.
I just do not find it correct to have it disabled. But at the same time I do
like much / I do not find correct the code myself. It is a bit problematic to
have struct value describing a memory content which is no longer present
there.
What happens there:
------------------------------------------------------------------------------
volatile int vv;
static __attribute__((noinline)) int
bar (int &ref) {
ref = 20;
vv++; /* break-here */
return ref;
}
int main (void) {
int var = 10;
return bar (var);
}
------------------------------------------------------------------------------
<4><c7>: Abbrev Number: 13 (DW_TAG_GNU_call_site_parameter)
<c8> DW_AT_location : 1 byte block: 55 (DW_OP_reg5 (rdi))
<ca> DW_AT_GNU_call_site_value: 2 byte block: 91 74 (DW_OP_fbreg: -12)
<cd> DW_AT_GNU_call_site_data_value: 1 byte block: 3a (DW_OP_lit10)
------------------------------------------------------------------------------
gdb -ex 'b value_addr' -ex r --args ../gdb ./1 -ex 'watch vv' -ex r -ex 'p &ref@entry'
->
6 return ref;
bar (ref=@0x7fffffffd944: 20, ref@entry=@0x7fffffffd944: 10) at 1.C:25
------------------------------------------------------------------------------
At /* break-here */ struct value variable 'ref' is TYPE_CODE_REF.
With FSF GDB HEAD:
(gdb) x/gx arg1.contents
0x6004000a4ad0: 0x00007fffffffd944
(gdb) p ((struct value *)arg1.location.computed.closure).lval
$1 = lval_memory
(gdb) p/x ((struct value *)arg1.location.computed.closure).location.address
$3 = 0x7fffffffd944
With your #if0-ed code:
(gdb) x/gx arg1.contents
0x6004000a4ad0: 0x00007fffffffd944
(gdb) p ((struct value *)arg1.location.computed.closure).lval
$8 = not_lval
(gdb) p/x ((struct value *)arg1.location.computed.closure).location.address
$9 = 0x0
I do not see how to access
((struct value *)arg1.location.computed.closure).location.address
from GDB CLI. Trying
(gdb) p &ref@entry
will invoke value_addr()'s:
if (TYPE_CODE (type) == TYPE_CODE_REF)
/* Copy the value, but change the type from (T&) to (T*). We
keep the same location information, which is efficient, and
allows &(&X) to get the location containing the reference. */
and therefore the address gets fetched already from
arg1.contents
and not from
((struct value *)arg1.location.computed.closure).location.address
.
And for any other type than TYPE_CODE_REF this code you #if 0-ed does not get
executed at all. This DW_AT_GNU_call_site_data_value DWARF was meant
primarily for Fortran but with -O0 entry values do not get produced
and with -Og and higher Fortran always optimizes out the passing by reference.
If you do not like the #if 0 code there I am OK with removing it as I do not
know how to make it's use reproducible for user anyway. In the worst case
- if there really is some way how to exploit it - one should just get
Attempt to take address of value not located in memory.
instead of some wrong value and it may be easy to fix then.
Thanks for the analysis,
Jan
--V88s5gaDVPzZ0KCq
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline; filename=1
gdb/
2014-07-20 Jan Kratochvil <jan.kratochvil@redhat.com>
* stack.c (read_frame_arg): Verify value_optimized_out before calling
value_available_contents_eq.
diff --git a/gdb/stack.c b/gdb/stack.c
index 0d6d8e7..4db5df5 100644
--- a/gdb/stack.c
+++ b/gdb/stack.c
@@ -413,6 +413,8 @@ read_frame_arg (struct symbol *sym, struct frame_info *frame,
/* If the reference addresses match but dereferenced
content does not match print them. */
if (val != val_deref
+ && !value_optimized_out (val_deref)
+ && !value_optimized_out (entryval_deref)
&& value_available_contents_eq (val_deref, 0,
entryval_deref, 0,
TYPE_LENGTH (type_deref)))
--V88s5gaDVPzZ0KCq--