[PATCH] D35384: PGOInstrumentation: Move profile matching warnings to remarks

[David Blaikie via llvm-commits]

On Thu, Jul 13, 2017 at 10:53 PM Xinliang David Li <davidxl at google.com>
wrote:

> On Thu, Jul 13, 2017 at 10:41 PM, Chandler Carruth <chandlerc at google.com>
> wrote:
>
>> On Thu, Jul 13, 2017 at 7:29 PM Xinliang David Li via llvm-commits <
>> llvm-commits at lists.llvm.org> wrote:
>>
>>> On Thu, Jul 13, 2017 at 4:17 PM, David Blaikie <dblaikie at gmail.com>
>>> wrote:
>>>
>>>>
>>>>
>>>> On Thu, Jul 13, 2017 at 2:03 PM David Li via Phabricator <
>>>> reviews at reviews.llvm.org> wrote:
>>>>
>>>>> davidxl added a comment.
>>>>>
>>>>> Blindly downgrade the warning into remarks can be bad -- profile
>>>>> mismatch problems can go undetected. People may spend more time diagnosing
>>>>> performance regressions due to missed warnings. Worse, we may regress in
>>>>> compiler without noticing.
>>>>>
>>>>
>>>> Any other ideas?
>>>>
>>>> The diagnostic handling callbacks through to Clang aren't granular
>>>> enough to demote only this warning to -Wno-error
>>>>
>>>
>>> Can this be fixed?
>>>
>>
Maybe, though I imagine to do that would mean plumbing this diagnostic as a
first class/separately identified one in the callback system, which I would
guess/assume we try not to do because it increases the coupling/surface
area between LLVM and Clang and the optimizers and diagnostic
infrastructure.

In any case, even if we could, we'd still be left in the same place of
telling users basically "this diagnostic isn't high enough quality to
-Werror it, you should turn it back into a warning-not-error" and at that
point, we know from experience, people tend to to ignore them...


>
>>>
>>>
>>>> (also as a warning it undermines the confidence as warnings as
>>>> actionable things that Clang has helped provide users).
>>>>
>>>
>>>
>>> Warnings like this are usually actionable by users or compiler
>>> developers.
>>>
>>
>> Warnings are really to indicate a bug in user code.
>>
>> For example, we use some things like this to warn users when inline
>> assembly is likely to have a bug, because that is user code that can be
>> fixed.
>>
>> A profile mismatch seems much less like "user code has a bug" and so
>> doesn't seem to fit any well in the "warning" category.
>>
>
> It is not user code bug, but still be a user bug in how PGO is used.
>

Is it, though? It sounds like this is expected to some degree (& sounds
like internally we encourage users to -Wno-error to get around this? Which
really seems to undermine confidence in Clang's warning quality) - because
the intent is to apply profiles from previous versions of the program to
new versions of the program, so things will mismatch, right? And there's
nothing the user's really meant to do in that case - maybe if it were over
some threshold.


> The warning happens also to be a useful for catch potential compiler bug.
>

> David
>
>
>>
>> On the other hand, remarks were specifically designed to be non-bug
>> important information from optimization passes that users may need to act
>> on.
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170715/77765231/attachment.html>