Re: [PATCH v3 1/2] gdb: dwarf2 generic implementation for caching function data

[Torbjorn SVENSSON <torbjorn.svensson@foss.st.com>]

On 2023-01-19 17:53, Simon Marchi wrote:
> 
> 
> On 1/19/23 05:29, Torbjörn SVENSSON via Gdb-patches wrote:
>> v2 -> v3:
>>
>> Addressed comments from Tom in
>> https://sourceware.org/pipermail/gdb-patches/2023-January/195882.html
>>
>>
>> ---
>>
>> When there is no dwarf2 data for a register, a function can be called
>> to provide the value of this register.  In some situations, it might
>> not be trivial to determine the value to return and it would cause a
>> performance bottleneck to do the computation each time.
>>
>> This patch allows the called function to have a "cache" object that it
>> can use to store some metadata between calls to reduce the performance
>> impact of the complex logic.
>>
>> The cache object is unique for each function and frame, so if there are
>> more than one function pointer stored in the dwarf2_frame_cache->reg
>> array, then the appropriate pointer will be supplied (the type is not
>> known by the dwarf2 implementation).
>>
>> dwarf2_frame_get_fn_data can be used to retrieve the function unique
>> cache object.
>> dwarf2_frame_allocate_fn_data can be used to allocate and retrieve the
>> function unique cache object.
>>
>> Signed-off-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com>
>> Signed-off-by: Yvan Roux <yvan.roux@foss.st.com>
> 
> Hi,
> 
> My main question is: this approach requires each custom prev_register
> function to set up and manage its own cache.  Did you consider caching
> values at a higher level?
> 
> Perhaps:
> 
>   1. In dwarf2_frame_prev_register, just for DWARF2_FRAME_REG_FN
>   2. In dwarf2_frame_prev_register, but for all kinds of registers
>   3. In frame_unwind_register_value, which would cache all register values
>      for all frames regardless of the unwinder
> 
> I don't know if other ways of unwinding would benefit that much from
> caching, but I guess it wouldn't hurt.  For instance, evaluating DWARF
> expressions is probably generally not super expensive, but it might
> still benefit from getting cached.  And it would be nice to write the
> caching code just once and have it work transparently for everything.

I suppose a generic cache might be useful in the long run, but it would 
have an impact on other code paths and I'm not comfortable of doing this 
major change now. Doing this change would imply that almost everything 
related to unwinding is impacted in one way or another and I suppose we 
would need to test that. With the solution I proposed, only Arm Cortex 
would be impacted and the scope for testing would be rather small.

I'm also aiming to resolve this for GDB13, so doing this major change is 
a bit late in the sprint, right?

If you have a better idea on how to cache everything, like in your 3 
ideas above, please don't hesitate do share the implementation. :)

> 
>> @@ -262,4 +264,35 @@ extern int dwarf2_fetch_cfa_info (struct gdbarch *gdbarch, CORE_ADDR pc,
>>   				  const gdb_byte **cfa_start_out,
>>   				  const gdb_byte **cfa_end_out);
>>   
>> +
>> +/* Allocate a new instance of the function unique data.
>> +
>> +   The main purpose of this custom function data object is to allow caching the
>> +   value of expensive lookups in the prev_register implementation.
> 
> I would replace "in the prev_register implementation" with "in
> prev_register implementations".  There isn't only one prev_register
> implementation AFAIK.

Will do that for v4 or when merging if no v4 is needed.

> 
> Simon


Kind regards,
Torbjörn