[llvm] r176214 - The FastISEL should be fast. But when we record statistics we use atomic operations to increment the counters.

[Eli Bendersky]

>> Another problem is that it looks like a lot of the "make check" tests use
>> -stats.  For example: test/Analysis/RegionInfo/block_sort.ll.  I think there
>> are about 80 of these tests.  Do people run "make check" for Release builds
>> instead of a Release+Asserts build?  Otherwise, we'd have to rewrite each of
>> the tests to not depend on stats, and it's not clear how to do that for all
>> of the tests.
>>
>>
>> Yes; however, I think we can make those tests depend on +Asserts and not run
>> otherwise, or be XFAIL or something like that. I don't think we have to
>> rewrite them all.
>
> Yes, IMHO the best way to go here would be to ignore these tests
> ("unsupported") via a lit config setting. Longer term goal could be to
> rewrite as many tests as possible not to use stats.
>

Lit infrastructure already has config.enable_assertions setup in the
build-specific lit configurations. So the plan could be:

1. Isolate all tests relying on -stats into sub-directories in the
respective directories they live in
2. Add a clause that if config.enable_assertions = 0, these
lit.local.cfg's set config.unsupported = True and you're good to go

Eli