diff --git a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp index a6553083d722b..cb3c119b8a645 100644 --- a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp +++ b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp @@ -941,11 +941,9 @@ 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(); @@ -1095,8 +1093,7 @@ 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 @@ -1138,11 +1135,6 @@ 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'); @@ -1214,11 +1206,15 @@ 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(); @@ -1283,9 +1279,6 @@ 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"); @@ -1319,7 +1312,6 @@ 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"); } @@ -1341,14 +1333,10 @@ 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 @@ -1578,9 +1566,6 @@ 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; diff --git a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.h b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.h index 94cd795bbc8f6..32139a91c2428 100644 --- a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.h +++ b/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;