summaryrefslogtreecommitdiffstats
path: root/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
diff options
context:
space:
mode:
authorNirav Dave <niravd@google.com>2017-05-31 18:43:17 +0000
committerNirav Dave <niravd@google.com>2017-05-31 18:43:17 +0000
commit3424373f303dbbfefa6c05660efb630850aa1baf (patch)
tree193b510e73e1d5da5c6efd5aacb320393f873f4f /llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
parentd6a36ae282ba22fbe2cd73ec40b660fb6c480c6b (diff)
downloadbcm5719-llvm-3424373f303dbbfefa6c05660efb630850aa1baf.tar.gz
bcm5719-llvm-3424373f303dbbfefa6c05660efb630850aa1baf.zip
[ScheduleDAG] Deal with already scheduled loads in ScheduleDAG.
Summary: If we attempt to unfold an SUnit in ScheduleDAG that results in finding an already scheduled load, we must should abort the unfold as it will not improve scheduling. This fixes PR32610. Reviewers: jmolloy, sunfish, bogner, spatel Subscribers: llvm-commits, MatzeB Differential Revision: https://reviews.llvm.org/D32911 llvm-svn: 304321
Diffstat (limited to 'llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp')
-rw-r--r--llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp278
1 files changed, 150 insertions, 128 deletions
diff --git a/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp b/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
index 4f4025d8ae6..579112c9bfc 100644
--- a/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
@@ -226,6 +226,7 @@ private:
void UnscheduleNodeBottomUp(SUnit*);
void RestoreHazardCheckerBottomUp();
void BacktrackBottomUp(SUnit*, SUnit*);
+ SUnit *TryUnfoldSU(SUnit *);
SUnit *CopyAndMoveSuccessors(SUnit*);
void InsertCopiesAndMoveSuccs(SUnit*, unsigned,
const TargetRegisterClass*,
@@ -780,7 +781,7 @@ void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU) {
}
/// CapturePred - This does the opposite of ReleasePred. Since SU is being
-/// unscheduled, incrcease the succ left count of its predecessors. Remove
+/// unscheduled, increase the succ left count of its predecessors. Remove
/// them from AvailableQueue if necessary.
void ScheduleDAGRRList::CapturePred(SDep *PredEdge) {
SUnit *PredSU = PredEdge->getSUnit();
@@ -934,6 +935,146 @@ static bool isOperandOf(const SUnit *SU, SDNode *N) {
return false;
}
+/// TryUnfold - Attempt to unfold
+SUnit *ScheduleDAGRRList::TryUnfoldSU(SUnit *SU) {
+ SDNode *N = SU->getNode();
+ // Use while over if to ease fall through.
+ SmallVector<SDNode *, 2> NewNodes;
+ if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
+ return nullptr;
+
+ // unfolding an x86 DEC64m operation results in store, dec, load which
+ // can't be handled here so quit
+ if (NewNodes.size() == 3)
+ return nullptr;
+
+ assert(NewNodes.size() == 2 && "Expected a load folding node!");
+
+ N = NewNodes[1];
+ SDNode *LoadNode = NewNodes[0];
+ unsigned NumVals = N->getNumValues();
+ unsigned OldNumVals = SU->getNode()->getNumValues();
+
+ // LoadNode may already exist. This can happen when there is another
+ // load from the same location and producing the same type of value
+ // but it has different alignment or volatileness.
+ bool isNewLoad = true;
+ SUnit *LoadSU;
+ if (LoadNode->getNodeId() != -1) {
+ LoadSU = &SUnits[LoadNode->getNodeId()];
+ // If LoadSU has already been scheduled, we should clone it but
+ // this would negate the benefit to unfolding so just return SU.
+ if (LoadSU->isScheduled)
+ return SU;
+ isNewLoad = false;
+ } else {
+ LoadSU = CreateNewSUnit(LoadNode);
+ LoadNode->setNodeId(LoadSU->NodeNum);
+
+ InitNumRegDefsLeft(LoadSU);
+ computeLatency(LoadSU);
+ }
+
+ DEBUG(dbgs() << "Unfolding SU #" << SU->NodeNum << "\n");
+
+ // Now that we are committed to unfolding replace DAG Uses.
+ for (unsigned i = 0; i != NumVals; ++i)
+ DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
+ DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals - 1),
+ SDValue(LoadNode, 1));
+
+ SUnit *NewSU = CreateNewSUnit(N);
+ assert(N->getNodeId() == -1 && "Node already inserted!");
+ N->setNodeId(NewSU->NodeNum);
+
+ const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
+ for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
+ if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
+ NewSU->isTwoAddress = true;
+ break;
+ }
+ }
+ if (MCID.isCommutable())
+ NewSU->isCommutable = true;
+
+ InitNumRegDefsLeft(NewSU);
+ computeLatency(NewSU);
+
+ // Record all the edges to and from the old SU, by category.
+ SmallVector<SDep, 4> ChainPreds;
+ SmallVector<SDep, 4> ChainSuccs;
+ SmallVector<SDep, 4> LoadPreds;
+ SmallVector<SDep, 4> NodePreds;
+ SmallVector<SDep, 4> NodeSuccs;
+ for (SDep &Pred : SU->Preds) {
+ if (Pred.isCtrl())
+ ChainPreds.push_back(Pred);
+ else if (isOperandOf(Pred.getSUnit(), LoadNode))
+ LoadPreds.push_back(Pred);
+ else
+ NodePreds.push_back(Pred);
+ }
+ for (SDep &Succ : SU->Succs) {
+ if (Succ.isCtrl())
+ ChainSuccs.push_back(Succ);
+ else
+ NodeSuccs.push_back(Succ);
+ }
+
+ // Now assign edges to the newly-created nodes.
+ for (const SDep &Pred : ChainPreds) {
+ RemovePred(SU, Pred);
+ if (isNewLoad)
+ AddPred(LoadSU, Pred);
+ }
+ for (const SDep &Pred : LoadPreds) {
+ RemovePred(SU, Pred);
+ if (isNewLoad)
+ AddPred(LoadSU, Pred);
+ }
+ for (const SDep &Pred : NodePreds) {
+ RemovePred(SU, Pred);
+ AddPred(NewSU, Pred);
+ }
+ for (SDep D : NodeSuccs) {
+ SUnit *SuccDep = D.getSUnit();
+ D.setSUnit(SU);
+ RemovePred(SuccDep, D);
+ D.setSUnit(NewSU);
+ AddPred(SuccDep, D);
+ // Balance register pressure.
+ if (AvailableQueue->tracksRegPressure() && SuccDep->isScheduled &&
+ !D.isCtrl() && NewSU->NumRegDefsLeft > 0)
+ --NewSU->NumRegDefsLeft;
+ }
+ for (SDep D : ChainSuccs) {
+ SUnit *SuccDep = D.getSUnit();
+ D.setSUnit(SU);
+ RemovePred(SuccDep, D);
+ if (isNewLoad) {
+ D.setSUnit(LoadSU);
+ AddPred(SuccDep, D);
+ }
+ }
+
+ // Add a data dependency to reflect that NewSU reads the value defined
+ // by LoadSU.
+ SDep D(LoadSU, SDep::Data, 0);
+ D.setLatency(LoadSU->Latency);
+ AddPred(NewSU, D);
+
+ if (isNewLoad)
+ AvailableQueue->addNode(LoadSU);
+ AvailableQueue->addNode(NewSU);
+
+ ++NumUnfolds;
+
+ if (NewSU->NumSuccsLeft == 0)
+ NewSU->isAvailable = true;
+
+ return NewSU;
+}
+
/// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
/// successors to the newly created node.
SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) {
@@ -959,135 +1100,16 @@ SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) {
return nullptr;
}
+ // If possible unfold instruction.
if (TryUnfold) {
- SmallVector<SDNode*, 2> NewNodes;
- if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
+ SUnit *UnfoldSU = TryUnfoldSU(SU);
+ if (!UnfoldSU)
return nullptr;
-
- // unfolding an x86 DEC64m operation results in store, dec, load which
- // can't be handled here so quit
- if (NewNodes.size() == 3)
- return nullptr;
-
- DEBUG(dbgs() << "Unfolding SU #" << SU->NodeNum << "\n");
- assert(NewNodes.size() == 2 && "Expected a load folding node!");
-
- N = NewNodes[1];
- SDNode *LoadNode = NewNodes[0];
- unsigned NumVals = N->getNumValues();
- unsigned OldNumVals = SU->getNode()->getNumValues();
- for (unsigned i = 0; i != NumVals; ++i)
- DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
- DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals-1),
- SDValue(LoadNode, 1));
-
- // LoadNode may already exist. This can happen when there is another
- // load from the same location and producing the same type of value
- // but it has different alignment or volatileness.
- bool isNewLoad = true;
- SUnit *LoadSU;
- if (LoadNode->getNodeId() != -1) {
- LoadSU = &SUnits[LoadNode->getNodeId()];
- isNewLoad = false;
- } else {
- LoadSU = CreateNewSUnit(LoadNode);
- LoadNode->setNodeId(LoadSU->NodeNum);
-
- InitNumRegDefsLeft(LoadSU);
- computeLatency(LoadSU);
- }
-
- SUnit *NewSU = CreateNewSUnit(N);
- assert(N->getNodeId() == -1 && "Node already inserted!");
- N->setNodeId(NewSU->NodeNum);
-
- const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
- for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
- if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
- NewSU->isTwoAddress = true;
- break;
- }
- }
- if (MCID.isCommutable())
- NewSU->isCommutable = true;
-
- InitNumRegDefsLeft(NewSU);
- computeLatency(NewSU);
-
- // Record all the edges to and from the old SU, by category.
- SmallVector<SDep, 4> ChainPreds;
- SmallVector<SDep, 4> ChainSuccs;
- SmallVector<SDep, 4> LoadPreds;
- SmallVector<SDep, 4> NodePreds;
- SmallVector<SDep, 4> NodeSuccs;
- for (SDep &Pred : SU->Preds) {
- if (Pred.isCtrl())
- ChainPreds.push_back(Pred);
- else if (isOperandOf(Pred.getSUnit(), LoadNode))
- LoadPreds.push_back(Pred);
- else
- NodePreds.push_back(Pred);
- }
- for (SDep &Succ : SU->Succs) {
- if (Succ.isCtrl())
- ChainSuccs.push_back(Succ);
- else
- NodeSuccs.push_back(Succ);
- }
-
- // Now assign edges to the newly-created nodes.
- for (const SDep &Pred : ChainPreds) {
- RemovePred(SU, Pred);
- if (isNewLoad)
- AddPred(LoadSU, Pred);
- }
- for (const SDep &Pred : LoadPreds) {
- RemovePred(SU, Pred);
- if (isNewLoad)
- AddPred(LoadSU, Pred);
- }
- for (const SDep &Pred : NodePreds) {
- RemovePred(SU, Pred);
- AddPred(NewSU, Pred);
- }
- for (SDep D : NodeSuccs) {
- SUnit *SuccDep = D.getSUnit();
- D.setSUnit(SU);
- RemovePred(SuccDep, D);
- D.setSUnit(NewSU);
- AddPred(SuccDep, D);
- // Balance register pressure.
- if (AvailableQueue->tracksRegPressure() && SuccDep->isScheduled
- && !D.isCtrl() && NewSU->NumRegDefsLeft > 0)
- --NewSU->NumRegDefsLeft;
- }
- for (SDep D : ChainSuccs) {
- SUnit *SuccDep = D.getSUnit();
- D.setSUnit(SU);
- RemovePred(SuccDep, D);
- if (isNewLoad) {
- D.setSUnit(LoadSU);
- AddPred(SuccDep, D);
- }
- }
-
- // Add a data dependency to reflect that NewSU reads the value defined
- // by LoadSU.
- SDep D(LoadSU, SDep::Data, 0);
- D.setLatency(LoadSU->Latency);
- AddPred(NewSU, D);
-
- if (isNewLoad)
- AvailableQueue->addNode(LoadSU);
- AvailableQueue->addNode(NewSU);
-
- ++NumUnfolds;
-
- if (NewSU->NumSuccsLeft == 0) {
- NewSU->isAvailable = true;
- return NewSU;
- }
- SU = NewSU;
+ SU = UnfoldSU;
+ N = SU->getNode();
+ // If this can be scheduled don't bother duplicating and just return
+ if (SU->NumSuccsLeft == 0)
+ return SU;
}
DEBUG(dbgs() << " Duplicating SU #" << SU->NodeNum << "\n");
OpenPOWER on IntegriCloud