On Wed, Mar 18, 2015 at 6:02 PM, Tom de Vries <tom_devr...@mentor.com> wrote:
> On 18-03-15 12:18, Richard Biener wrote:
>>
>> On Wed, Mar 18, 2015 at 12:03 PM, Tom de Vries <tom_devr...@mentor.com>
>> wrote:
>>>
>>> On 18-03-15 11:16, Richard Biener wrote:
>>>>
>>>>
>>>> On Fri, Mar 13, 2015 at 4:28 PM, Tom de Vries <tom_devr...@mentor.com>
>>>> wrote:
>>>>>
>>>>>
>>>>> On 13-03-15 13:36, Richard Biener wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On Fri, Mar 13, 2015 at 1:07 PM, Jakub Jelinek <ja...@redhat.com>
>>>>>> wrote:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On Fri, Mar 13, 2015 at 01:04:57PM +0100, Richard Biener wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Not really (I don't like -fdump-passes ...), but we need to make
>>>>>>>> sure
>>>>>>>> that -fdump-passes doesn't crash (because it runs very early and
>>>>>>>> with cfun == NULL I think)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> If it runs with cfun == NULL, then supposedly the gates that are
>>>>>>> dependent
>>>>>>> on current function should for -fdump-passes purposes also return
>>>>>>> true
>>>>>>> if cfun == NULL (well, of course do all the unconditional checks).
>>>>>>> Though of course, with optimize/target attributes this is harder, as
>>>>>>> different functions can use different options.
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> Yes, one reason why I think -fdump-passes is just broken
>>>>>> implementation-wise.
>>>>>>
>>>>>
>>>>> Atm fdump-passes doesn't run with cfun == NULL.
>>>>>
>>>>> From pass_manager::dump_passes:
>>>>> ...
>>>>> FOR_EACH_FUNCTION (n)
>>>>> if (DECL_STRUCT_FUNCTION (n->decl))
>>>>> {
>>>>> node = n;
>>>>> break;
>>>>> }
>>>>>
>>>>> if (!node)
>>>>> return;
>>>>>
>>>>> push_cfun (DECL_STRUCT_FUNCTION (node->decl));
>>>>
>>>>
>>>>
>>>> Um - this now picks a random function which may be one with
>>>> an optimize or target attribute associated to it.
>>>>
>>>
>>> Indeed.
>>>
>>> Attached patch removes that code, and runs the gates with cfun == NULL
>>> for
>>> -fdump-passes. It at least builds, and allows us to compile
>>> src/gcc/testsuite/gcc.dg/dump-pass.c with -O2 -fdump-passes.
>>>
>>> Should I bootstrap and reg-test, or do you see a problem with this
>>> approach?
>>
>>
>> Yeah - it makes the -fdump-passes "hack" more pervasive throughout
>> the compiler.
>>
>> I suppose it should instead build & push a "dummy" sturct function.
>>
>
> Like this?
Looks good to me.
Richard.
>> Well, or simply don't care for it's brokeness.
>
>
> I'm afraid leaving it broken just means we'll keep coming back to it. So I'd
> prefer either fixing or removing.
>
> Thanks,
> - Tom
>
