[llvm] [AMDGPU] Add AMDGPU specific tryCandidate() (PR #114957)

[via llvm-commits]

================
@@ -63,6 +63,10 @@ static cl::opt<bool> GCNTrackers(
     cl::desc("Use the AMDGPU specific RPTrackers during scheduling"),
     cl::init(false));
 
+static cl::opt<bool> UseAMDGPUScheduleHeuristic(
+    "amdgpu-use-amdgpu-schedule-heuristic", cl::Hidden,
----------------
ruiling wrote:

I think it is OK to be default on (the only difference would be lots of test changes). But I would like we could keep it off for the first submission (at least before we see some performance results on compute benchmarks). We may still need fine-tuning the heuristics after more extensive testing. We can turn it on later once this version's tryCandidate() becomes somewhat stable and do not regress much real world cases. I am not sure how such kind of performance testing is done in compute side. It would better we can get some performance numbers for the change. Please let me know if you still have concern on this.


> Name and description too general

Do you have any suggestion? Maybe `amdgpu-use-custom-try-candidate`?

https://github.com/llvm/llvm-project/pull/114957