[polly] r262488 - Fix: Add pass manager barrier.

[Tobias Grosser via llvm-commits]

On 03/02/2016 08:03 PM, David Blaikie wrote:
>
>
> On Wed, Mar 2, 2016 at 10:54 AM, Michael Kruse <llvmdev at meinersbur.de
> <mailto:llvmdev at meinersbur.de>> wrote:
>
>     Hi,
>
>     this is only meant as a temporary workaround. Of course we want
>     analysis preservation fixed. At the moment I know just one bug that
>     needs to be fixed: llvm.org/PR6736 <http://llvm.org/PR6736>. We'll
>     gradually enable the
>     verification switches manually such that we eventually can remove this
>     workaround. We also plan to add the switches (-verify-scev
>     -verify-loop-info -verify-region-info -verify-dom-tree) to the Polly
>     correctness buildbots so I think no further action is necessary.
>
>     Maybe I just misunderstood. What do you mean by "normal test case"?
>     How can a single test case ensure all verification passes?
>
>
> I mean a lit test that's run for/by everyone, rather than "run the test
> suite in this custom way" (which is presumably where/how you found it).
>
> Doesn't necessarily have to test all the possible things this fixes (but
> it could, I suppose - add an extra run line to every one of the tests
> that was failing, but that's probably silly). But at least one test case
> that demonstrates this line of code is necessary.
>
> But that's also why I cc'd chandler - I'm not sure what the test
> strategy should look like for the pass manager transition. As things
> like this are fixed, how those fixes should be demonstrated/tracked, etc.

Dear David,

thanks for the comment. This commit is mostly a "hack" to hide a 
possibly larger number of issues, which we did not yet identify or 
reduce, where Polly forgot to update some analysis information. Hence, 
we do not have any good test cases for this. However, this hack allows 
us to get some of our buildbots green, which gives us a nice baseline to 
then gradually address the underlying problems.

When doing so we will identify the individual situations where Polly 
forgot to update some analysis information, fix the corresponding piece 
of code and add a test case that uses -verify-analysis-name to ensure 
this has indeed be fixed. Doing so a couple of times, should allow us to 
drop this hack again and to setup a buildbot that ensures we always pass 
the -verify-analysis checks.

Using -verify-analysis seems to be the best way to test analysis results 
today, ignoring the option to always verify all analysis which may 
indeed be too costly to use broadly.

Best,
Tobias