[AMDGPU][Scheduler] Delete RegionsWithMinOcc bitvector from scheduler (NFC) (#142361)

The `GCNScheduleDAGMILive`'s `RegionsWithMinOcc` bitvector is only used
by the `UnclusteredHighRPStage`. Its presence in the scheduler's state
forces us to maintain its value throughout scheduling even though it is
of no use to the iterative scheduling process itself. At any point
during scheduling it is possible to cheaply compute the occupancy
induced by a particular register pressure. Furthermore, the field
doesn't appear to be updated correctly throughout scheduling i.e., bits
corresponding to regions at minimum occupancy are not always set in the
vector.

This removes the bitvector from `GCNScheduleDAGMILive`.
`UnclusteredHighRPStage::initGCNRegion` now directly computes the
occupancy of possibly reschedulable regions instead of querying the
vector. Since it is the most expensive check, it is done last in the
list.

Author: lucas-rami

llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp

@@ -942,11 +942,9 @@ void GCNScheduleDAGMILive::finalizeSchedule() {
   Pressure.resize(Regions.size());
   RegionsWithHighRP.resize(Regions.size());
   RegionsWithExcessRP.resize(Regions.size());
-  RegionsWithMinOcc.resize(Regions.size());
   RegionsWithIGLPInstrs.resize(Regions.size());
   RegionsWithHighRP.reset();
   RegionsWithExcessRP.reset();
-  RegionsWithMinOcc.reset();
   RegionsWithIGLPInstrs.reset();
 
   runSchedStages();
@@ -1096,8 +1094,7 @@ bool PreRARematStage::initGCNSchedStage() {
   // fixed if there is another pass after this pass.
   assert(!S.hasNextStage());
 
-  if (!GCNSchedStage::initGCNSchedStage() || DAG.RegionsWithMinOcc.none() ||
-      DAG.Regions.size() == 1)
+  if (!GCNSchedStage::initGCNSchedStage() || DAG.Regions.size() == 1)
     return false;
 
   // Before performing any IR modification record the parent region of each MI
@@ -1139,11 +1136,6 @@ void UnclusteredHighRPStage::finalizeGCNSchedStage() {
   SavedMutations.swap(DAG.Mutations);
   S.SGPRLimitBias = S.VGPRLimitBias = 0;
   if (DAG.MinOccupancy > InitialOccupancy) {
-    for (unsigned IDX = 0; IDX < DAG.Pressure.size(); ++IDX)
-      DAG.RegionsWithMinOcc[IDX] =
-          DAG.Pressure[IDX].getOccupancy(
-              DAG.ST, DAG.MFI.getDynamicVGPRBlockSize()) == DAG.MinOccupancy;
-
     LLVM_DEBUG(dbgs() << StageID
                       << " stage successfully increased occupancy to "
                       << DAG.MinOccupancy << '\n');
@@ -1215,11 +1207,15 @@ bool GCNSchedStage::initGCNRegion() {
 }
 
 bool UnclusteredHighRPStage::initGCNRegion() {
-  // Only reschedule regions with the minimum occupancy or regions that may have
-  // spilling (excess register pressure).
-  if ((!DAG.RegionsWithMinOcc[RegionIdx] ||
-       DAG.MinOccupancy <= InitialOccupancy) &&
-      !DAG.RegionsWithExcessRP[RegionIdx])
+  // Only reschedule regions that have excess register pressure (i.e. spilling)
+  // or had minimum occupancy at the beginning of the stage (as long as
+  // rescheduling of previous regions did not make occupancy drop back down to
+  // the initial minimum).
+  unsigned DynamicVGPRBlockSize = DAG.MFI.getDynamicVGPRBlockSize();
+  if (!DAG.RegionsWithExcessRP[RegionIdx] &&
+      (DAG.MinOccupancy <= InitialOccupancy ||
+       DAG.Pressure[RegionIdx].getOccupancy(ST, DynamicVGPRBlockSize) !=
+           InitialOccupancy))
     return false;
 
   return GCNSchedStage::initGCNRegion();
@@ -1284,9 +1280,6 @@ void GCNSchedStage::checkScheduling() {
   if (PressureAfter.getSGPRNum() <= S.SGPRCriticalLimit &&
       PressureAfter.getVGPRNum(ST.hasGFX90AInsts()) <= S.VGPRCriticalLimit) {
     DAG.Pressure[RegionIdx] = PressureAfter;
-    DAG.RegionsWithMinOcc[RegionIdx] =
-        PressureAfter.getOccupancy(ST, DynamicVGPRBlockSize) ==
-        DAG.MinOccupancy;
 
     // Early out if we have achieved the occupancy target.
     LLVM_DEBUG(dbgs() << "Pressure in desired limits, done.\n");
@@ -1320,7 +1313,6 @@ void GCNSchedStage::checkScheduling() {
   if (NewOccupancy < DAG.MinOccupancy) {
     DAG.MinOccupancy = NewOccupancy;
     MFI.limitOccupancy(DAG.MinOccupancy);
-    DAG.RegionsWithMinOcc.reset();
     LLVM_DEBUG(dbgs() << "Occupancy lowered for the function to "
                       << DAG.MinOccupancy << ".\n");
   }
@@ -1342,14 +1334,10 @@ void GCNSchedStage::checkScheduling() {
 
   // Revert if this region's schedule would cause a drop in occupancy or
   // spilling.
-  if (shouldRevertScheduling(WavesAfter)) {
+  if (shouldRevertScheduling(WavesAfter))
     revertScheduling();
-  } else {
+  else
     DAG.Pressure[RegionIdx] = PressureAfter;
-    DAG.RegionsWithMinOcc[RegionIdx] =
-        PressureAfter.getOccupancy(ST, DynamicVGPRBlockSize) ==
-        DAG.MinOccupancy;
-  }
 }
 
 unsigned
@@ -1579,9 +1567,6 @@ bool GCNSchedStage::mayCauseSpilling(unsigned WavesAfter) {
 }
 
 void GCNSchedStage::revertScheduling() {
-  DAG.RegionsWithMinOcc[RegionIdx] =
-      PressureBefore.getOccupancy(ST, DAG.MFI.getDynamicVGPRBlockSize()) ==
-      DAG.MinOccupancy;
   LLVM_DEBUG(dbgs() << "Attempting to revert scheduling.\n");
   DAG.RegionEnd = DAG.RegionBegin;
   int SkippedDebugInstr = 0;

llvm/lib/Target/AMDGPU/GCNSchedStrategy.h

@@ -250,9 +250,6 @@ class GCNScheduleDAGMILive final : public ScheduleDAGMILive {
   // limit. Register pressure in these regions usually will result in spilling.
   BitVector RegionsWithExcessRP;
 
-  // Regions that has the same occupancy as the latest MinOccupancy
-  BitVector RegionsWithMinOcc;
-
   // Regions that have IGLP instructions (SCHED_GROUP_BARRIER or IGLP_OPT).
   BitVector RegionsWithIGLPInstrs;