Re: [RFC 00/15] Remove regcache::m_readonly_p

[Simon Marchi <simon.marchi@ericsson.com>]

On 2018-01-16 11:18 AM, Yao Qi wrote:
> Yao Qi <qiyaoltc@gmail.com> writes:
> 
>> regcache is used in many places in gdb in different ways, so regcache
>> becomes a flat and fat object.  That exposes some unnecessary APIs to
>> different part, and some APIs are misused or abused:
>>
>>  1) gdbarch methods pseudo_register_read and pseudo_register_read_value
>>  have a parameter 'regcache *', but these two gdbarch methods only need
>>  raw_read* and cooked_read* methods.  So it is better to pass a class
>>  which only has raw_read* and cooked_read* methods, and other regcache
>>  methods are invisible to each gdbarch implementation.
>>
>>  2) target_ops methods to_fetch_registers and to_store_registers have a
>>  parameter 'regcache *', but these two target_ops methods only need
>>  raw_supply and raw_collect methods, because raw registers are from target
>>  layer, pseudo registers are "composed" or "created" by gdbarch.
>>
>>  3) jit.c uses regcache in an odd way, and record-full.c should use
>>  a simple version regcache instead of an array (see patch 11)
>>
>> Beside these api issues, one issue in regcache is that there is no
>> type or class for readonly regcache.  We use a flag field m_readonly_p
>> to indicate the regcache is readonly or not, so some regcache apis have
>> assert that this regcache is or is not readonly.  The better way to do
>> this is to create a new class for readonly regcache which doesn't have
>> write methods at all.
>>
>> This patch series fixes all of the problems above except 2) (I had a
>> patch to fix 2 in my tree, but still need more time to polish it.) by
>> designing a class hierarchy about regcache, like this,
>>
>>                       reg_buffer
>>                          ^
>>                          |
>>                    ------+-----
>>                    ^
>>                    |
>>             regcache_read
>>                  ^
>>                  |
>>            ------+------
>>            ^           ^
>>            |           |
>>     reg_buffer_rw regcache_readonly
>>           ^
>>           |
>>       regcache
>>
>> Class reg_buffer is a simple class, having register contents and status
>> (in patch 7).  regcache_read is an abstract class only having raw_read*
>> and cooked_read* methods (in patch 8).  reg_buffer_rw is a class which
>> has read and write methods, but it disconnects from target, IOW, the
>> write doesn't go through.  Class regcache_readonly is the readonly
>> regcache, created from regcache::save method.
>>
>> This patch series is tested on {x86_64, aarch64, ppc64}-linux.  It is
>> an RFC, want to get comments.  I write them in at least seven different
>> ways, and this one is satisfying.  I don't push them in until 8.1 is
>> branched.

Hi Yao,

The design you propose makes sense to me in general, but I have some
questions.

Which of these types will have actual instances of them, and which ones
are only interfaces?

I have some problem understanding the difference between regcache_read
and regcache_readonly.  I think the fact that the name is so similar
doesn't help.  Would there be a better name for regcache_readonly?

Simon