[PATCH] D24593: Standford/Bubble sort code restructure

[Evgeny Stupachenko via llvm-commits]

evstupac added a comment.

> I still think it is more honest to live with the performance swings (we can explain them after all, so no need to blame the compiler or your patches) rather than modifying our benchmarks so the results look nicer.


Honest - yes. More effective - No.
I've already spend a lot of time trying to prove that epilogue instead of prologue has no controlled influence on this test... if we leave it as is someone else will get performance improve because of some change and this will be ok. And after this someone else will get performance drop and spend a lot of time doing the same.

We should make a decision what for we are using these tests?

1. Regression control? Then history does not matter. We can apply changes.
2. Represent some widely used benchmark? Most likely no.
3. To prove that LLVM is moving forward? Release to release on flaky bubble sort? Maybe.
4. To represent some important customer application? Hmm... most likely no.
5. Something else?

What is the profit of leaving the benchmark as is?
The profit of changing is obvious - we'll safe time for future commits.


https://reviews.llvm.org/D24593