mi-sched: smooth out the cyclicpath heuristic.

Arnold's idea.

I generally try to avoid stateful heuristics because it can make
debugging harder. However, we need a way to prevent the latency
priority from dominating, and it somewhat makes sense to schedule
aggressively for latency only within an issue group.

Swift in particular likes this, and it doesn't hurt anyone else:
| Benchmarks/MiBench/consumer-lame              |  10.39% |
| Benchmarks/Misc/himenobmtxpa                  |   9.63% |

llvm-svn: 190360

1 files changed, 4 insertions, 1 deletions

diff --git a/llvm/lib/CodeGen/MachineScheduler.cpp b/llvm/lib/CodeGen/MachineScheduler.cpp
index f79b39140c0..ef916174267 100644
--- a/llvm/lib/CodeGen/MachineScheduler.cpp
+++ b/llvm/lib/CodeGen/MachineScheduler.cpp
@@ -2492,7 +2492,10 @@ void ConvergingScheduler::tryCandidate(SchedCandidate &Cand,
     return;
 
   // For loops that are acyclic path limited, aggressively schedule for latency.
-  if (Rem.IsAcyclicLatencyLimited && tryLatency(TryCand, Cand, Zone))
+  // This can result in very long dependence chains scheduled in sequence, so
+  // once every cycle (when CurrMOps == 0), switch to normal heuristics.
+  if (Rem.IsAcyclicLatencyLimited && !Zone.CurrMOps
+      && tryLatency(TryCand, Cand, Zone))
     return;
 
   // Keep clustered nodes together to encourage downstream peephole