Re: [PATCH v10 04/10] Create empty common/linux-maps.[ch] and common/target-utils.[ch]

[Pedro Alves <palves at redhat dot com>]

On 07/16/2015 09:14 AM, Gary Benson wrote:
> Jan Kratochvil wrote:
>> On Mon, 08 Jun 2015 10:37:33 +0200, Gary Benson wrote:
>>> A bunch of stuff has changed in the way common code is laid out
>>> since May 2014, so while Tom approved this back then, it's not
>>> suitable any more.  Please update the series as follows:
>>>
>>>> 	* common/linux-maps.c: New file.
>>>> 	* common/linux-maps.h: New file.
>>>
>>> Nothing os-specific should be in common.  These files should be
>>> nat/linux-maps.[ch].
>>
>> While I have done so I do not share this opinion.  linux-maps.[ch]
>> is used from linux-tdep.c and not from linux-nat.c.  linux-tdep.c is
>> using common/* files a lot but it uses no function from nat/*:
>>
>> 	$ for i in $(for i in common/*.c;do nm `basename $i .c`.o;done|sed -n 's/^.* T //p');do grep -q "\<$i (" linux-tdep.c && echo $i;done|wc -l
>> 	17
>> 	$ for i in $(for i in    nat/*.c;do nm `basename $i .c`.o;done|sed -n 's/^.* T //p');do grep -q "\<$i (" linux-tdep.c && echo $i;done|wc -l
>> 	0
>> 	(commands above assume in-src-tree build)
>>
>> Additionally linux-maps.o should not be in config/**.mh::NATDEPFILES
>> (like linux-namespaces.o from nat/ ) but it needs to be in
>> configure.tgt instead.
>>
>> Should be linux-maps.[ch] therefore in nat/ or in common/ ?
>> If it matters, though.
> 
> I don't fully understand the tdep/non-tdep split very well.
> Pedro, could you comment?

TL;DR: roughly, nat/ holds native target_ops code, while tdep
files hold gdbarch code.

The idea of nat/ is to hold host-dependent code that implements
the native target (target_ops) backends.  Code in that directory makes use
of native system debug APIs, constants, etc..  Such code cannot be used
on a cross debugger that targets the OS the code is for, because it
e.g., relies on headers that will only exist on that target OS.
nat/ code is used by either gdb's native target_ops target, or, gdbserver's.
E.g., a --host=mingw32 --target=x86_64-linux gdb can't build the native
Linux code in nat/.

In contrast, the tdep files files contain code that _must_ be
host-independent.  The same --host=mingw32 --target=x86_64-linux gdb
build includes the linux-specific -tdep files.

A -tdep.c file including a "nat/foo.h" header thus raises
the abstraction-violation alarm bell.  It's at the same level of
a -tdep.c file including <sys/ptrace.h>.

So seems to me this maps code should not be in nat/.

Thanks,
Pedro Alves