Post-RA scheduler compile time fix. Quadratic computation of DAG node depth.

The post-ra scheduler was explicitly updating the depth of a node's
successors after scheduling it, regardless of whether the successor
was ready. This is quadratic for DAGs with transitively redundant
edges. I simply removed the useless update of depth, which is lazilly
computed later.
Fixes <rdar://problem/9044332> compiler takes way too long to build TextInput.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@130992 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/PostRASchedulerList.cpp b/lib/CodeGen/PostRASchedulerList.cpp
index 60c24b710792271053529c47d9b2aa4aaf59bd6b..848af2a6c32370cff895b7ace467ba5828bc64d4 100644 (file)
--- a/lib/CodeGen/PostRASchedulerList.cpp
+++ b/lib/CodeGen/PostRASchedulerList.cpp
@@ -540,10 +540,16 @@ void SchedulePostRATDList::ReleaseSucc(SUnit *SU, SDep *SuccEdge) {
 #endif
   --SuccSU->NumPredsLeft;
 
-  // Compute how many cycles it will be before this actually becomes
-  // available.  This is the max of the start time of all predecessors plus
-  // their latencies.
-  SuccSU->setDepthToAtLeast(SU->getDepth() + SuccEdge->getLatency());
+  // Standard scheduler algorithms will recomute the depth of the successor
+  // here as such:
+  //   SuccSU->setDepthToAtLeast(SU->getDepth() + SuccEdge->getLatency());
+  //
+  // However, we lazily compute node depth instead. Note that
+  // ScheduleNodeTopDown has already updated the depth of this node which causes
+  // all descendents to be marked dirty. Setting the successor depth explicitly
+  // here would cause depth to be recomputed for all its ancestors. If the
+  // successor is not yet ready (because of a transitively redundant edge) then
+  // this causes depth computation to be quadratic in the size of the DAG.
 
   // If all the node's predecessors are scheduled, this node is ready
   // to be scheduled. Ignore the special ExitSU node.