[llvm] r179456 - Further generalize this scheduler test.

[Andrew Trick]

On Apr 13, 2013, at 4:16 AM, Rafael Espíndola <rafael.espindola at gmail.com> wrote:

>> Further generalize this scheduler test.
>> 
>> The order of copies depends on queue order, which is not very stable.
> 
> Non deterministic or just likely to change from commit to commit?

I hope it's deterministic. But if anything at all changes in the surrounding code because of an unrelated checkin or different triple, then everything is up for grabs.

This is a really minor example of a much bigger problem writing codegen tests. We need a way to test specific behavior independent of surrounding code and upstream/downstream passes.

I won't rant too much here because there's already a separate thread on this. But Jakob's CHECK-DAG proposal is very important if we're serious about testing codegen. Can you imagine rewriting dozens of tests just to make a trivial machine model fix? That's the reality now, and it means less things getting fixed, fewer things getting tested, and less incremental development.

I'll do anything within reason to make a pass more "stable". But I'm not going to add cost and complexity to the code just to solve one very rare instance of instability (rescheduled copies) if it does nothing to solve the larger problem.

-Andy