[PATCH] D22836: Add option to diable Bitcode tests. Tests are enabled by default.

[Matthias Braun via llvm-commits]

MatzeB added a comment.

In https://reviews.llvm.org/D22836#497224, @mehdi_amini wrote:

> In https://reviews.llvm.org/D22836#497219, @asbirlea wrote:
>
> > Sounds good to me. 
> >  As Mehdi pointed out, the apple compiler will probably want to use these tests as well, so I'm guessing we'll revisit the flag when the released version parses the bitcode correctly.
>
>
> Well not really, the problem will stay forever (at some point our internal compiler will read the 3.9 bitcode but there might be some 4.0 bitcode). Also we're not the only downstream user, and other companies will have the same issue.


We really need a consistent story here that is not just "Use todays ToT clang compiler", that is not how the test-suite is designed (just imagine bisecting over a larger range of commits). We should have some default versioning checks that to the best of our knowledge only enable configuration that work. If there are ever newer compilers that don't read the bitcode anymore we have to extend the checks we have.

> The approach @MatzeB took is a different philosophy from the one I was suggesting, basically he implemented opt-in while I was suggesting opt-out: an out-of-tree clang will have to explicitly pass the flag to get the tests enabled.

> 

> Either way (opt-in or opt-out) are fine to me, but I don't think we need to have code to handle every proprietary version system in the test-suite config upstream.


I don't really understand the problem here. This is not about shutting out apple clang, this is about not having the test-suite fail out of the box which is a bad user experience.
If out of tree clang versions need an extra flag to enable the tests that is a price we should pay to not have the test-suite fail on other people who may have no clue about bitcode tests and versioning issues.


https://reviews.llvm.org/D22836