Re: [PATCH v2 1/2] Allow subtarget customization of CC1_SPEC

[Iain Sandoe <idsandoe@googlemail.com>]

Hi

> On 7 Dec 2022, at 07:54, Sebastian Huber <sebastian.huber@embedded-brains.de> wrote:
> 
> 
> 
> On 07.12.22 08:10, Thomas Schwinge wrote:
>> Hi!
>> On 2022-12-07T07:04:10+0100, Sebastian Huber <sebastian.huber@embedded-brains.de> wrote:
>>> On 06.12.22 22:06, Thomas Schwinge wrote:
>>> I suppose I just fail to see some detail here, but:
>>> 
>>>> On 2022-11-21T08:25:25+0100, Sebastian Huber<sebastian.huber@embedded-brains.de>  wrote:
>>>>> gcc/ChangeLog:
>>>>> 
>>>>>        * gcc.cc (SUBTARGET_CC1_SPEC): Define if not defined.
>>>>>        (cc1_spec): Append SUBTARGET_CC1_SPEC.
>>>>> ---
>>>>> v2: Append SUBTARGET_CC1_SPEC directly to cc1_spec and not through CC1_SPEC.
>>>>>      This avoids having to modify all the CC1_SPEC definitions in the targets.
>>>>> 
>>>>>   gcc/gcc.cc | 9 ++++++++-
>>>>>   1 file changed, 8 insertions(+), 1 deletion(-)
>>>>> 
>>>>> diff --git a/gcc/gcc.cc b/gcc/gcc.cc
>>>>> index 830ab88701f..4e1574a4df1 100644
>>>>> --- a/gcc/gcc.cc
>>>>> +++ b/gcc/gcc.cc
>>>>> @@ -706,6 +706,13 @@ proper position among the other output files.  */
>>>>>   #define CPP_SPEC ""
>>>>>   #endif
>>>>> 
>>>>> +/* Subtargets can define SUBTARGET_CC1_SPEC to provide extra args to cc1 and
>>>>> +   cc1plus or extra switch-translations.  The SUBTARGET_CC1_SPEC is appended
>>>>> +   to CC1_SPEC.  */
>>>>> +#ifndef SUBTARGET_CC1_SPEC
>>>>> +#define SUBTARGET_CC1_SPEC ""
>>>>> +#endif
>>>>> +
>>>>>   /* config.h can define CC1_SPEC to provide extra args to cc1 and cc1plus
>>>>>      or extra switch-translations.  */
>>>>>   #ifndef CC1_SPEC
>>>>> @@ -1174,7 +1181,7 @@ proper position among the other output files.  */
>>>>>   static const char *asm_debug = ASM_DEBUG_SPEC;
>>>>>   static const char *asm_debug_option = ASM_DEBUG_OPTION_SPEC;
>>>>>   static const char *cpp_spec = CPP_SPEC;
>>>>> -static const char *cc1_spec = CC1_SPEC;
>>>>> +static const char *cc1_spec = CC1_SPEC SUBTARGET_CC1_SPEC;
>>>>>   static const char *cc1plus_spec = CC1PLUS_SPEC;
>>>>>   static const char *link_gcc_c_sequence_spec = LINK_GCC_C_SEQUENCE_SPEC;
>>>>>   static const char *link_ssp_spec = LINK_SSP_SPEC;
>>>> 
>>>> ... doesn't this (at least potentially?) badly interact with any existing
>>>> 'SUBTARGET_CC1_SPEC' definitions -- which pe rabove get appended to
>>>> 'cc1_spec'?
>>>> 
>>>>      gcc/config/loongarch/gnu-user.h-   and provides this hook instead.  */
>>>>      gcc/config/loongarch/gnu-user.h:#undef SUBTARGET_CC1_SPEC
>>>>      gcc/config/loongarch/gnu-user.h:#define SUBTARGET_CC1_SPEC GNU_USER_TARGET_CC1_SPEC
>>>>      gcc/config/loongarch/gnu-user.h-
>>>>      --
>>>>      gcc/config/loongarch/loongarch.h-#define EXTRA_SPECS \
>>>>      gcc/config/loongarch/loongarch.h:  {"subtarget_cc1_spec", SUBTARGET_CC1_SPEC}, \
>>>>      gcc/config/loongarch/loongarch.h-  {"subtarget_cpp_spec", SUBTARGET_CPP_SPEC}, \
>>>>      --
>>>>      gcc/config/mips/gnu-user.h-   and provides this hook instead.  */
>>>>      gcc/config/mips/gnu-user.h:#undef SUBTARGET_CC1_SPEC
>>>>      gcc/config/mips/gnu-user.h:#define SUBTARGET_CC1_SPEC GNU_USER_TARGET_CC1_SPEC
>>>>      gcc/config/mips/gnu-user.h-
>>>>      --
>>>>      gcc/config/mips/linux-common.h-
>>>>      gcc/config/mips/linux-common.h:#undef  SUBTARGET_CC1_SPEC
>>>>      gcc/config/mips/linux-common.h:#define SUBTARGET_CC1_SPEC                                           \
>>>>      gcc/config/mips/linux-common.h-  LINUX_OR_ANDROID_CC (GNU_USER_TARGET_CC1_SPEC,                     \
>>>>      --
>>>>      gcc/config/mips/mips.h-
>>>>      gcc/config/mips/mips.h:/* SUBTARGET_CC1_SPEC is passed to the compiler proper.  It may be
>>>>      gcc/config/mips/mips.h-   overridden by subtargets.  */
>>>>      gcc/config/mips/mips.h:#ifndef SUBTARGET_CC1_SPEC
>>>>      gcc/config/mips/mips.h:#define SUBTARGET_CC1_SPEC ""
>>>>      gcc/config/mips/mips.h-#endif
>>>>      --
>>>>      gcc/config/mips/mips.h-#define EXTRA_SPECS                                                  \
>>>>      gcc/config/mips/mips.h:  { "subtarget_cc1_spec", SUBTARGET_CC1_SPEC },                              \
>>>>      gcc/config/mips/mips.h-  { "subtarget_cpp_spec", SUBTARGET_CPP_SPEC },                              \
>>>>      --
>>>>      gcc/config/mips/r3900.h-/* By default (if not mips-something-else) produce code for the r3900 */
>>>>      gcc/config/mips/r3900.h:#undef SUBTARGET_CC1_SPEC
>>>>      gcc/config/mips/r3900.h:#define SUBTARGET_CC1_SPEC "\
>>>>      gcc/config/mips/r3900.h-%{mhard-float:%e-mhard-float not supported} \
>>> 
>>> Oh, I came up with the name SUBTARGET_CC1_SPEC after a discussion on the
>>> mailing list
>> I've put Iain in CC.
>>> and I have to admit that I didn't check that it was
>>> actually already in use.
>> Always one of the first things I do.  ;-)
>>> What about renaming the loongarch/mips define
>>> to LOONGARCH_CC1_SPEC and MIPS_CC1_SPEC?
>> Also in use are a number of other 'SUBTARGET_[...]_SPEC' and
>> corresponding 'subtarget_[...]_spec' in 'EXTRA_SPECS', for example:
>>     gcc/config/mips/mips.h-#define EXTRA_SPECS                                                      \
>>     gcc/config/mips/mips.h:  { "subtarget_cc1_spec", SUBTARGET_CC1_SPEC },                          \
>>     gcc/config/mips/mips.h:  { "subtarget_cpp_spec", SUBTARGET_CPP_SPEC },                          \
>>     gcc/config/mips/mips.h:  { "subtarget_asm_debugging_spec", SUBTARGET_ASM_DEBUGGING_SPEC },      \
>>     gcc/config/mips/mips.h:  { "subtarget_asm_spec", SUBTARGET_ASM_SPEC },                          \
>>     gcc/config/mips/mips.h-  { "asm_abi_default_spec", "-" MULTILIB_ABI_DEFAULT },                  \
>>     gcc/config/mips/mips.h-  { "endian_spec", ENDIAN_SPEC },                                        \
>>     gcc/config/mips/mips.h:  SUBTARGET_EXTRA_SPECS
>> Do we need/want to keep the association of same-name
>> upper-case/lower-case variants; in your proposal you'd then get
>> '{ "subtarget_cc1_spec", MIPS_CC1_SPEC }', for example?  (I didn't
>> quickly grok all 'EXTRA_SPECS' usage.)
>> Alternatively, what about renaming your 'SUBTARGET_CC1_SPEC' to
>> 'CC1_SPEC_EXTRA' -- if that makes sense?
>>     static const char *cc1_spec = CC1_SPEC CC1_SPEC_EXTRA;
> 
> I was told that an operating system is the subtarget in this context. So from the name SUBTARGET_CC1_SPEC is is clear who is in charge. This is not clear from CC1_SPEC_EXTRA.

Perhaps I was not precise enough (or misunderstood the full requirement).

Having reloaded some state on this thread…

 * The initial point was that target specs were an alternate mechanism for doing what was required
   which I understood to be that there was "an OS-specific common spec that needed to be added to multiple ports".

 * There are existing cases (at least Darwin is one) where this is done -  multiple ports (e.g. x86, rs6000) refer to common specs in the gcc/config/  
   - so, for example, the specs in gcc/config/i386/darwin.h pull in common specs from gcc/config/darwin.h

  * there are also cases where this is specifically done in port code (sometimes with “SUBSUBTARGET_xxxx”)

===

    I made the comment that, IMO this is initially confusing when the OS might be considered the target and the ISA the sub-target - however the status quo in specs naming is the opposite way round.

( I do not wish add further confusion here - but attempting to clarify - these styles of provision might not be suitable for the specific case, I guess - anyway, at this point, I’ve got nothing more to add :) ) 

Iain

> 
>> But doesn't somehow this whole thing feel a bit like "chating the
>> system"?  ;-)
>> Can't you actually achieve your thing (TLS model) via (new) 'EXTRA_SPECS'
>> in 'gcc/config/rtems.h', for example?
> 
> The EXTRA_SPECS definition seems to be target-specific. Not all targets let an operating system define SUBTARGET_EXTRA_SPECS. The SUBTARGET_EXTRA_SPECS would need to get propagated to the corresponding specs, which seems to be also target-specific, for example for mips we have:
> 
> #undef CC1_SPEC
> #define CC1_SPEC "\
> %{G*} %{EB:-meb} %{EL:-mel} %{EB:%{EL:%emay not use both -EB and -EL}} \
> %(subtarget_cc1_spec)"
> 
> I think going this route would lead to a lot of changes affecting all targets.
> 
> -- 
> embedded brains GmbH
> Herr Sebastian HUBER
> Dornierstr. 4
> 82178 Puchheim
> Germany
> email: sebastian.huber@embedded-brains.de
> phone: +49-89-18 94 741 - 16
> fax:   +49-89-18 94 741 - 08
> 
> Registergericht: Amtsgericht München
> Registernummer: HRB 157899
> Vertretungsberechtigte Geschäftsführer: Peter Rasmussen, Thomas Dörfler
> Unsere Datenschutzerklärung finden Sie hier:
> https://embedded-brains.de/datenschutzerklaerung/