Fix gdb 8.1 Solaris/SPARC compilation (PR build/22206)

  When testing my Solaris < 10 removal patch on Solaris/SPARC, I found
that gdb mainline is currently broken there due to the recent SPARC M7
ADI patches:

/vol/src/gnu/gdb/gdb/local/gdb/sparc64-tdep.c:1876:0: error: "PSR_ICC" redefined [-Werror]
 #define PSR_ICC  0x00f00000
 ^
In file included from /usr/include/v7/sys/privregs.h:24:0,
                 from /usr/include/sys/regset.h:420,
                 from /usr/include/sys/ucontext.h:21,
                 from /usr/include/sys/signal.h:231,
                 from /usr/include/sys/procset.h:23,
                 from /usr/include/sys/wait.h:25,
                 from /usr/include/stdlib.h:21,
                 from build-gnulib/import/stdlib.h:36,
                 from /vol/src/gnu/gdb/gdb/local/gdb/common/common-defs.h:53,
                 from /vol/src/gnu/gdb/gdb/local/gdb/defs.h:28,
                 from /vol/src/gnu/gdb/gdb/local/gdb/sparc64-tdep.c:20:
/usr/include/v7/sys/psr.h:35:0: note: this is the location of the previous definition
 #define PSR_ICC  0x00F00000 /* integer condition codes */
 ^
/vol/src/gnu/gdb/gdb/local/gdb/sparc64-tdep.c:1878:0: error: "PSR_IMPL" redefined [-Werror]
 #define PSR_IMPL 0xf0000000
 ^
In file included from /usr/include/v7/sys/privregs.h:24:0,
                 from /usr/include/sys/regset.h:420,
                 from /usr/include/sys/ucontext.h:21,
                 from /usr/include/sys/signal.h:231,
                 from /usr/include/sys/procset.h:23,
                 from /usr/include/sys/wait.h:25,
                 from /usr/include/stdlib.h:21,
                 from build-gnulib/import/stdlib.h:36,
                 from /vol/src/gnu/gdb/gdb/local/gdb/common/common-defs.h:53,
                 from /vol/src/gnu/gdb/gdb/local/gdb/defs.h:28,
                 from /vol/src/gnu/gdb/gdb/local/gdb/sparc64-tdep.c:20:
/usr/include/v7/sys/psr.h:41:0: note: this is the location of the previous definition
 #define PSR_IMPL 0xF0000000 /* implementation */
 ^

Comparing Solaris 11.4 <v7/sys/psr.h> and sparc64-tdep.c, there are more
inconsistencies:

<v7/sys/psr.h>:

#define	PSR_S		0x00000080	/* supervisor mode */
#define	PSR_ICC		0x00F00000	/* integer condition codes */
#define	PSR_VER		0x0F000000	/* mask version */
#define	PSR_IMPL	0xF0000000	/* implementation */
#define	PSR_RSV		0x000FC000	/* reserved */

sparc64-tdep.c:

#define PSR_S		0x00000080
#define PSR_ICC		0x00f00000
#define PSR_VERS	0x0f000000
#define PSR_IMPL	0xf0000000
#define PSR_V8PLUS	0xff000000
#define PSR_XCC		0x000f0000

Apart from the capitalization differences that trip g++, the names
differ (PSR_VER vs. PSR_VERS), PSR_XCC is included in Solaris' PSR_RSV,
and there's no PSR_V8PLUS on Solaris either.

/vol/src/gnu/gdb/gdb/local/gdb/sparc64-tdep.c: In function ‘int adi_tag_fd()’:
/vol/src/gnu/gdb/gdb/local/gdb/sparc64-tdep.c:296:63: error: format ‘%d’ expects argument of type ‘int’, but argument 4 has type ‘pid_t {aka long int}’ [-Werror=format=]
   snprintf (cl_name, sizeof(cl_name), "/proc/%d/adi/tags", pid);
                                                               ^
/vol/src/gnu/gdb/gdb/local/gdb/sparc64-tdep.c: In function ‘bool adi_is_addr_mapped(CORE_ADDR, std::size_t)’:
/vol/src/gnu/gdb/gdb/local/gdb/sparc64-tdep.c:314:64: error: format ‘%d’ expects argument of type ‘int’, but argument 4 has type ‘pid_t {aka long int}’ [-Werror=format=]
   snprintf (filename, sizeof filename, "/proc/%d/adi/maps", pid);
                                                                ^

You cannot always print a pid_t, which can be either int or long on
Solaris, as an int.

Obviously, the ADI patch which modifies code shared between all SPARC
targets, hasn't been tested on anything but Linux/SPARC.

The patch below includes the minimal fixes necessary to unbreak the
Solaris/SPARC build.

However, as detailed in the PR, there's more breakage here: apart from
not bothering to implement ADI support on Solaris, the code contains
several more changes to shared/common SPARC code that are simply wrong
on anything but Linux/SPARC.

The patch was tested on sparcv9-sun-solaris2.10 and
sparcv9-sun-solaris2.11.4 (build and gdb/gdb gdb/gdb smoke test only).

Ok for mainline?

	Rainer
  

On 09/26/2017 10:33 AM, Rainer Orth wrote:

> Apart from the capitalization differences that trip g++, the names
> differ (PSR_VER vs. PSR_VERS), PSR_XCC is included in Solaris' PSR_RSV,
> and there's no PSR_V8PLUS on Solaris either.

What you've done is fine with me to unbreak the build.  Though
I'd prefer if we renamed those to avoid ever relying on host
symbols, anywhere.  Like:

 - #define PSR_S ...
 + #define SPARC64_PSR_S ...

etc.

> 
> /vol/src/gnu/gdb/gdb/local/gdb/sparc64-tdep.c: In function ‘int adi_tag_fd()’:
> /vol/src/gnu/gdb/gdb/local/gdb/sparc64-tdep.c:296:63: error: format ‘%d’ expects argument of type ‘int’, but argument 4 has type ‘pid_t {aka long int}’ [-Werror=format=]
>    snprintf (cl_name, sizeof(cl_name), "/proc/%d/adi/tags", pid);
>                                                                ^
> /vol/src/gnu/gdb/gdb/local/gdb/sparc64-tdep.c: In function ‘bool adi_is_addr_mapped(CORE_ADDR, std::size_t)’:
> /vol/src/gnu/gdb/gdb/local/gdb/sparc64-tdep.c:314:64: error: format ‘%d’ expects argument of type ‘int’, but argument 4 has type ‘pid_t {aka long int}’ [-Werror=format=]
>    snprintf (filename, sizeof filename, "/proc/%d/adi/maps", pid);
>                                                                 ^
> 
> You cannot always print a pid_t, which can be either int or long on
> Solaris, as an int.
> 
> Obviously, the ADI patch which modifies code shared between all SPARC
> targets, hasn't been tested on anything but Linux/SPARC.
> 
> The patch below includes the minimal fixes necessary to unbreak the
> Solaris/SPARC build.
> 
> However, as detailed in the PR, there's more breakage here: apart from
> not bothering to implement ADI support on Solaris, the code contains
> several more changes to shared/common SPARC code that are simply wrong
> on anything but Linux/SPARC.
> 
> The patch was tested on sparcv9-sun-solaris2.10 and
> sparcv9-sun-solaris2.11.4 (build and gdb/gdb gdb/gdb smoke test only).
> 
> Ok for mainline?

OK.

Thanks,
Pedro Alves
  

Hi Pedro,

> On 09/26/2017 10:33 AM, Rainer Orth wrote:
>
>> Apart from the capitalization differences that trip g++, the names
>> differ (PSR_VER vs. PSR_VERS), PSR_XCC is included in Solaris' PSR_RSV,
>> and there's no PSR_V8PLUS on Solaris either.
>
> What you've done is fine with me to unbreak the build.  Though
> I'd prefer if we renamed those to avoid ever relying on host
> symbols, anywhere.  Like:
>
>  - #define PSR_S ...
>  + #define SPARC64_PSR_S ...
>
> etc.

agreed.  That's probably best done when figuring out how to properly
separate target-dependent and independent parts of ADI support.

	Rainer
  

On 9/26/2017 6:02 AM, Rainer Orth wrote:
> Hi Pedro,
>
>> On 09/26/2017 10:33 AM, Rainer Orth wrote:
>>
>>> Apart from the capitalization differences that trip g++, the names
>>> differ (PSR_VER vs. PSR_VERS), PSR_XCC is included in Solaris' PSR_RSV,
>>> and there's no PSR_V8PLUS on Solaris either.
>> What you've done is fine with me to unbreak the build.  Though
>> I'd prefer if we renamed those to avoid ever relying on host
>> symbols, anywhere.  Like:
>>
>>   - #define PSR_S ...
>>   + #define SPARC64_PSR_S ...
>>
>> etc.
> agreed.  That's probably best done when figuring out how to properly
> separate target-dependent and independent parts of ADI support.
>
> 	Rainer
>

Macros PSR_ICC/PSR_IMPL are not parts of ADI and were added to 
sparc64-tdep.c
long before the Linux ADI project. Still trying to figure out why these 
two macros,
pulled from /usr/include/v7/sys/psr.h, are causing problems now.
  

Hi Wei-min,

> On 9/26/2017 6:02 AM, Rainer Orth wrote:
>> Hi Pedro,
>>
>>> On 09/26/2017 10:33 AM, Rainer Orth wrote:
>>>
>>>> Apart from the capitalization differences that trip g++, the names
>>>> differ (PSR_VER vs. PSR_VERS), PSR_XCC is included in Solaris' PSR_RSV,
>>>> and there's no PSR_V8PLUS on Solaris either.
>>> What you've done is fine with me to unbreak the build.  Though
>>> I'd prefer if we renamed those to avoid ever relying on host
>>> symbols, anywhere.  Like:
>>>
>>>   - #define PSR_S ...
>>>   + #define SPARC64_PSR_S ...
>>>
>>> etc.
>> agreed.  That's probably best done when figuring out how to properly
>> separate target-dependent and independent parts of ADI support.
>>
>> 	Rainer
>>
>
> Macros PSR_ICC/PSR_IMPL are not parts of ADI and were added to
> sparc64-tdep.c
> long before the Linux ADI project. Still trying to figure out why these two
> macros,
> pulled from /usr/include/v7/sys/psr.h, are causing problems now.

I suspect there were changes to gdb/defs.h, gdb/common/common-defs.h or
gnulib that ultimately dragged that file in post gdb-8.1.  Seems I've
been barking up the wrong tree for this one, sorry.

Even so, the other problems with the ADI implemention mentioned both in
the patch submission and the PR are certainly new.

	Rainer
  

Hi Rainer,


On 9/27/2017 2:16 AM, Rainer Orth wrote:
> Hi Wei-min,
>
>> On 9/26/2017 6:02 AM, Rainer Orth wrote:
>>> Hi Pedro,
>>>
>>>> On 09/26/2017 10:33 AM, Rainer Orth wrote:
>>>>
>>>>> Apart from the capitalization differences that trip g++, the names
>>>>> differ (PSR_VER vs. PSR_VERS), PSR_XCC is included in Solaris' PSR_RSV,
>>>>> and there's no PSR_V8PLUS on Solaris either.
>>>> What you've done is fine with me to unbreak the build.  Though
>>>> I'd prefer if we renamed those to avoid ever relying on host
>>>> symbols, anywhere.  Like:
>>>>
>>>>    - #define PSR_S ...
>>>>    + #define SPARC64_PSR_S ...
>>>>
>>>> etc.
>>> agreed.  That's probably best done when figuring out how to properly
>>> separate target-dependent and independent parts of ADI support.
>>>
>>> 	Rainer
>>>
>> Macros PSR_ICC/PSR_IMPL are not parts of ADI and were added to
>> sparc64-tdep.c
>> long before the Linux ADI project. Still trying to figure out why these two
>> macros,
>> pulled from /usr/include/v7/sys/psr.h, are causing problems now.
> I suspect there were changes to gdb/defs.h, gdb/common/common-defs.h or
> gnulib that ultimately dragged that file in post gdb-8.1.  Seems I've
> been barking up the wrong tree for this one, sorry.
>
> Even so, the other problems with the ADI implemention mentioned both in
> the patch submission and the PR are certainly new.
>
> 	Rainer
>

Thanks for tracking down the marco problem.

As for the issues you raised on gdb ADI support for Solaris/SPARC in the 
bug report,
it's never intended to add that support for Solaris. While that Solaris 
provided a totally
different ADI API was a factor, for example, Linux made the information 
of ADI block
size available in the ELF aux record, one needed to make a syscall on 
Solaris to get that
piece of information, it's certainly doable.  It was never planned, 
however, and most likely
will never be.
  

Hi Wei-min,

> As for the issues you raised on gdb ADI support for Solaris/SPARC in the
> bug report,
> it's never intended to add that support for Solaris. While that Solaris
> provided a totally
> different ADI API was a factor, for example, Linux made the information of
> ADI block
> size available in the ELF aux record, one needed to make a syscall on
> Solaris to get that
> piece of information, it's certainly doable.  It was never planned,
> however, and most likely
> will never be.

this is extremely unfortunate, Oracle adding ADI support for one OS, but
not for its own ;-(

However, what's worse is the way in which ADI support was added: when
you add code to a file shared between different OSes like
sparc64-tdep.c, it's your responsibility to make sure that this code at
least doesn't break other SPARC targets.  Given how much of the code
there isn't actually shared (and not sharable, it seems), the approach
you've take seems wrong to me: it should have been properly factored
between shared (sparc64-tdep.c) and os-private
(e.g. sparc64-linux-tdep.c) files to avoid such breakage in the first
place.

	Rainer
  

On 09/28/2017 01:40 PM, Rainer Orth wrote:
> Hi Wei-min,
> 
>> As for the issues you raised on gdb ADI support for Solaris/SPARC in the
>> bug report,
>> it's never intended to add that support for Solaris. While that Solaris
>> provided a totally
>> different ADI API was a factor, for example, Linux made the information of
>> ADI block
>> size available in the ELF aux record, one needed to make a syscall on
>> Solaris to get that
>> piece of information, it's certainly doable.  It was never planned,
>> however, and most likely
>> will never be.
> 
> this is extremely unfortunate, Oracle adding ADI support for one OS, but
> not for its own ;-(
> 
> However, what's worse is the way in which ADI support was added: when
> you add code to a file shared between different OSes like
> sparc64-tdep.c, it's your responsibility to make sure that this code at
> least doesn't break other SPARC targets.  Given how much of the code
> there isn't actually shared (and not sharable, it seems), the approach
> you've take seems wrong to me: it should have been properly factored
> between shared (sparc64-tdep.c) and os-private
> (e.g. sparc64-linux-tdep.c) files to avoid such breakage in the first
> place.

To be fair, that is the sort of issue that should have been pointed out
in review.  When I pushed for moving the ADI support out of the nat 
files and into tdep files for cross debugging, I don't think I even
remembered Solaris was a thing...  Not that I have anything against
Solaris, to be clear.  It just didn't cross my mind.

To me, this indicates a few things:

 - It's great that the Solaris port is again seeing activity, and
   we should strive to make sure that SPARC changes consider it.

 - But also someone needs to continually keep an eye on Solaris
   lest it ends up forgotten and broken again.  We need an official
   Solaris port maintainer.  Any takers?  :-)

 - It'd be desirable to have a SPARC Solaris machine build slave
   in the GDB buildbot, so that we can offload part of the work to 
   bots.  I wonder whether Oracle can help with this?  Might be
   difficult with the whole Solaris situation...

Thanks,
Pedro Alves
  

Hi Pedro,

>> this is extremely unfortunate, Oracle adding ADI support for one OS, but
>> not for its own ;-(
>> 
>> However, what's worse is the way in which ADI support was added: when
>> you add code to a file shared between different OSes like
>> sparc64-tdep.c, it's your responsibility to make sure that this code at
>> least doesn't break other SPARC targets.  Given how much of the code
>> there isn't actually shared (and not sharable, it seems), the approach
>> you've take seems wrong to me: it should have been properly factored
>> between shared (sparc64-tdep.c) and os-private
>> (e.g. sparc64-linux-tdep.c) files to avoid such breakage in the first
>> place.
>
> To be fair, that is the sort of issue that should have been pointed out
> in review.  When I pushed for moving the ADI support out of the nat 
> files and into tdep files for cross debugging, I don't think I even
> remembered Solaris was a thing...  Not that I have anything against
> Solaris, to be clear.  It just didn't cross my mind.

understandable: it hasn't been exactly prominent until very recently...

> To me, this indicates a few things:
>
>  - It's great that the Solaris port is again seeing activity, and
>    we should strive to make sure that SPARC changes consider it.
>
>  - But also someone needs to continually keep an eye on Solaris
>    lest it ends up forgotten and broken again.  We need an official
>    Solaris port maintainer.  Any takers?  :-)

TBH, I'm a bit reluctant to take the position.  I've already too much on
my plate and am unsure if I can follow gdb development in any useful
way.  Right now, it's just testing gdb either when releases approach or
when major changes go in.

>  - It'd be desirable to have a SPARC Solaris machine build slave
>    in the GDB buildbot, so that we can offload part of the work to 
>    bots.  I wonder whether Oracle can help with this?  Might be
>    difficult with the whole Solaris situation...

Certainly from Oracle, but fortunately they provided me with a Netra
S7-2 as a long-term loan for my GCC work a few months ago.  I've even
set up a Solaris 11.3 zone to act as a build zone for Go and eventually
other FOSS developers to test on, and am considering to add it to the
GCC build farm if appropriate terms can be worked out.  That zone would
be a natural candidate for a GDB build slave provided that doesn't take
too many ressources.

	Rainer