Re: [PATCH 7/7] gdb: add special handling for frame level 0 in frame_info_ptr

[Simon Marchi <simon.marchi@polymtl.ca>]

On 11/8/22 05:40, Bruno Larsen wrote:
> On 07/11/2022 16:53, Simon Marchi via Gdb-patches wrote:
>> From: Simon Marchi <simon.marchi@efficios.com>
>>
>> I noticed this problem while preparing the initial submission for the
>> ROCm GDB port.  One particularity of this patch set is that it does not
>> support unwinding frames, that requires support of some DWARF extensions
>> that will come later.  It was still possible to run to a breakpoint and
>> print frame #0, though.
>>
>> When rebasing on top of the frame_info_ptr work, GDB started tripping on
>> a prepare_reinflate call, making it not possible anymore to event print
>> the frame when stopping on a breakpoint.  One thing to know about frame
>> 0 is that its id is lazily computed when something requests it through
>> get_frame_id.  See:
>>
>>    https://gitlab.com/gnutools/binutils-gdb/-/blob/23912acd402f5af9caf91b257e5209ec4c58a09c/gdb/frame.c#L2070-2080
>>
>> So, up to that prepare_reinflate call, frame 0's id was not computed,
>> and prepare_reinflate, calling get_frame_id, forces it to be computed.
>> Computing the frame id generally requires unwinding the previous frame,
>> which with my ROCm GDB patch fails.  An exception is thrown and the
>> printing of the frame is simply abandonned.
>>
>> Regardless of this ROCm GDB problem (which is admittedly temporary, it
>> will be possible to unwind with subsequent patches), we want to avoid
>> prepare_reinflate to force the computing of the frame id, for the same
>> reasons we lazily compute it in the first place.
>>
>> In addition, frame 0's id is subject to change across a frame cache
>> reset.  This is why save_selected_frame and restore_selected_frame have
>> special handling for frame 0:
>>
>>    https://gitlab.com/gnutools/binutils-gdb/-/blob/23912acd402f5af9caf91b257e5209ec4c58a09c/gdb/frame.c#L1841-1863
>>
>> For this last reason, we also need to handle frame 0 specially in
>> prepare_reinflate / reinflate.  Because the frame id of frame 0 can
>> change across a frame cache reset, we must not rely on the frame id from
>> that frame to reinflate it.  We should instead just re-fetch the current
>> frame at that point.
>>
>> This patch adds a frame_info_ptr::m_cached_level field, set in
>> frame_info_ptr::prepare_reinflate, so we can tell if a frame is frame 0.
>> There are cases where a frame_info_ptr object wraps a sentinel frame,
>> for which frame_relative_level returns -1, so I have chosen the value -2
>> to represent "invalid frame level", for when the frame_info_ptr object
>> is empty.
>>
>> In frame_info_ptr::prepare_reinflate, only cache the frame id if the
>> frame level is not 0.  It's fine to cache the frame id for the sentinel
>> frame, it will be properly handled by frame_find_by_id later.
>>
>> In frame_info_ptr::reinflate, if the frame level is 0, call
>> get_current_frame to get the target's current frame.  Otherwise, use
>> frame_find_by_id just as before.
>>
>> This patch should not have user-visible changes with upstream GDB.  But
>> it will avoid forcing the computation of frame 0's when calling
>> prepare_reinflate.  And, well, it fixes the upcoming ROCm GDB patch
>> series.
>>
>> Change-Id: I176ed7ee9317ddbb190acee8366e087e08e4d266
> 
> This all makes sense. I have a small style preference below, but even if you dislike it, the code is still fine.
> 
> Reviewed-By: Bruno Larsen <blarsen@redhat.com>
> 
>> ---
>>   gdb/frame-info.c | 24 ++++++++++++++++++++----
>>   gdb/frame-info.h | 20 ++++++++++++++++++--
>>   2 files changed, 38 insertions(+), 6 deletions(-)
>>
>> diff --git a/gdb/frame-info.c b/gdb/frame-info.c
>> index 584222dc490f..e3ee9f8174e1 100644
>> --- a/gdb/frame-info.c
>> +++ b/gdb/frame-info.c
>> @@ -31,7 +31,11 @@ intrusive_list<frame_info_ptr> frame_info_ptr::frame_list;
>>   void
>>   frame_info_ptr::prepare_reinflate ()
>>   {
>> -  m_cached_id = get_frame_id (*this);
>> +  m_cached_level = frame_relative_level (*this);
>> +  gdb_assert (m_cached_level >= -1);
> 
> Since you have declared invalid_level = -2 for this class, I feel like it would be more better to have the assert be
> 
> gdb_assert (m_cached_level > invalid_level);

This form assumes that invalid_level is -2, defeating the purpose to
have the abstraction in the first place.  If we changed invalid_level to
be INT_MAX, for intsance, the assertion wouldn't be right anymore.

> This way there is no need to wonder why -1 is a valid level, and makes it easier to grep for the comment in the file, should someone want to know.

In my vision of things, the sentinel frame having level -1 is well
known, because it's the frame just below the current frame, which is
known to have level 0.  So while it looks like a magic random value,
it's not really.  The numerical value has a meaning.  We wouldn't want
to change the sentinel frame's level value to be any other arbitrary
numerical value.

Here, I can just drop the assert.  It's basically just checking that
frame_relative_level didn't return something that doesn't make sense.
But there's no reason for frame_relative_level to return something that
doesn't make sense in the first place.  Other callers of
frame_relative_level don't do this assert, they just trust that
frame_relative_level returns something that makes sense, nothing really
different here.

>> +
>> +  if (m_cached_level != 0)
>> +    m_cached_id = get_frame_id (*this);
>>   }
>>     /* See frame-info-ptr.h.  */
>> @@ -39,9 +43,21 @@ frame_info_ptr::prepare_reinflate ()
>>   void
>>   frame_info_ptr::reinflate ()
>>   {
>> -  gdb_assert (frame_id_p (m_cached_id));
>> +  gdb_assert (m_cached_level >= -1);
> Likewise

Here, I could add a comment like:

  /* Ensure we have a valid frame level, indicating prepare_reinflate
     was called.  */

Simon