[PATCH] Add a callback to FunctionPass to enable skipping execution on a per-function basis

[Chandler Carruth]

I'm suggesting rather than a decorator, we use an optional predicate
parameter when constructing the pass. To me, that seems somewhat cleaner.

On Wed, Apr 8, 2015 at 8:09 PM Eric Christopher <echristo at gmail.com> wrote:

> Right, I agree with this in general, but looking to avoid the weird
> subtarget flags that I mentioned. Perhaps we should revisit Akira's
> original idea of wrapping pass in a decorator if you want to pull it out of
> the pass manager machinery?
>
> On Wed, Apr 8, 2015, 8:04 PM Chandler Carruth <chandlerc at google.com>
> wrote:
>
>> Actually, even this could be done cleanly.
>>
>> You could change the *pass* to accept the generic predicate in this case,
>> and add one unpredicated version to the pipeline and add a predicated form
>> later.
>>
>> I'm essentially trying to lift the predicate logic out of the pass
>> management machinery and into the pass itself because that's where the
>> motivation for a predicate comes from.
>>
>> On Wed, Apr 8, 2015 at 8:03 PM Chandler Carruth <chandlerc at google.com>
>> wrote:
>>
>>> Is this a real use case or a hypothetical one? Because it seems somewhat
>>> contrived to me...
>>>
>>> If there really is some predicate that necessitates really radically
>>> different pass pipelines, I feel like they should be, well, two separate
>>> pass pipelines.
>>>
>>> On Wed, Apr 8, 2015 at 7:54 PM Eric Christopher <echristo at gmail.com>
>>> wrote:
>>>
>>>> Optnone, IMO, needs to be replaced by something less terrible.
>>>>
>>>> I'm not sure how this is going to work with the "I want to run the
>>>> first cfgcleanup unconditionally, but not the second" without tying the
>>>> subtargets to things like shouldRunCfgCleanup2().
>>>>
>>>> On Wed, Apr 8, 2015, 7:38 PM Sean Silva <chisophugis at gmail.com> wrote:
>>>>
>>>>> Adding Paul as this seems related to optnone.
>>>>>
>>>>> On Wed, Apr 8, 2015 at 7:31 PM, Chandler Carruth <chandlerc at gmail.com>
>>>>> wrote:
>>>>>
>>>>>> I feel like we could do something much simpler than this. This
>>>>>> feeling is predicated on one primary theory: most passes will run for most
>>>>>> subtargets. Put another way, there will only be a small number of passes
>>>>>> that we actually want to opt out of on a per-subtarget basis.
>>>>>>
>>>>>> If we think that's likely to be the case, here is an alternative
>>>>>> suggestion:
>>>>>>
>>>>>> - Add bool-returning predicates for each pass to the subtarget base
>>>>>> class (eg, "isIfConversionProfitable()") with the expected default ("true").
>>>>>> - Override these for the subtargets that want to opt out.
>>>>>> - Change the pass to directly get the subtarget, query it, and bail
>>>>>> without doing anything if it gets "false".
>>>>>>
>>>>>> From looking at and thinking about if-conversion at least, this seems
>>>>>> nicer to me. It makes someone working on the pass aware that there are
>>>>>> subtarget profitability concerns, and it makes it very clear that we are
>>>>>> *running* all of the passes, just that some have no effect on certain
>>>>>> subtargets.
>>>>>>
>>>>>> This also matches how an optimization pass should query the function
>>>>>> for the 'noopt' attribute and bail.
>>>>>>
>>>>>> Thoughts?
>>>>>>
>>>>>
>>>>>>
>>>>>> http://reviews.llvm.org/D8717
>>>>>>
>>>>>> EMAIL PREFERENCES
>>>>>>   http://reviews.llvm.org/settings/panel/emailpreferences/
>>>>>>
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> llvm-commits mailing list
>>>>>> llvm-commits at cs.uiuc.edu
>>>>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>>>>>
>>>>> _______________________________________________
>>>> llvm-commits mailing list
>>>> llvm-commits at cs.uiuc.edu
>>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>>>
>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150409/2a1ebc55/attachment.html>