1
//===--- ScheduleDAGSDNodes.cpp - Implement the ScheduleDAGSDNodes class --===//
3
// The LLVM Compiler Infrastructure
5
// This file is distributed under the University of Illinois Open Source
6
// License. See LICENSE.TXT for details.
8
//===----------------------------------------------------------------------===//
10
// This implements the ScheduleDAG class, which is a base class used by
11
// scheduling implementation classes.
13
//===----------------------------------------------------------------------===//
15
#define DEBUG_TYPE "pre-RA-sched"
16
#include "SDNodeDbgValue.h"
17
#include "ScheduleDAGSDNodes.h"
18
#include "InstrEmitter.h"
19
#include "llvm/CodeGen/SelectionDAG.h"
20
#include "llvm/Target/TargetMachine.h"
21
#include "llvm/Target/TargetInstrInfo.h"
22
#include "llvm/Target/TargetLowering.h"
23
#include "llvm/Target/TargetRegisterInfo.h"
24
#include "llvm/Target/TargetSubtarget.h"
25
#include "llvm/ADT/DenseMap.h"
26
#include "llvm/ADT/SmallPtrSet.h"
27
#include "llvm/ADT/SmallSet.h"
28
#include "llvm/ADT/SmallVector.h"
29
#include "llvm/ADT/Statistic.h"
30
#include "llvm/Support/Debug.h"
31
#include "llvm/Support/raw_ostream.h"
34
STATISTIC(LoadsClustered, "Number of loads clustered together");
36
ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
40
/// Run - perform scheduling.
42
void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb,
43
MachineBasicBlock::iterator insertPos) {
45
ScheduleDAG::Run(bb, insertPos);
48
/// NewSUnit - Creates a new SUnit and return a ptr to it.
50
SUnit *ScheduleDAGSDNodes::NewSUnit(SDNode *N) {
52
const SUnit *Addr = 0;
56
SUnits.push_back(SUnit(N, (unsigned)SUnits.size()));
57
assert((Addr == 0 || Addr == &SUnits[0]) &&
58
"SUnits std::vector reallocated on the fly!");
59
SUnits.back().OrigNode = &SUnits.back();
60
SUnit *SU = &SUnits.back();
61
const TargetLowering &TLI = DAG->getTargetLoweringInfo();
63
(N->isMachineOpcode() &&
64
N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF))
65
SU->SchedulingPref = Sched::None;
67
SU->SchedulingPref = TLI.getSchedulingPreference(N);
71
SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) {
72
SUnit *SU = NewSUnit(Old->getNode());
73
SU->OrigNode = Old->OrigNode;
74
SU->Latency = Old->Latency;
75
SU->isTwoAddress = Old->isTwoAddress;
76
SU->isCommutable = Old->isCommutable;
77
SU->hasPhysRegDefs = Old->hasPhysRegDefs;
78
SU->hasPhysRegClobbers = Old->hasPhysRegClobbers;
79
SU->SchedulingPref = Old->SchedulingPref;
84
/// CheckForPhysRegDependency - Check if the dependency between def and use of
85
/// a specified operand is a physical register dependency. If so, returns the
86
/// register and the cost of copying the register.
87
static void CheckForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op,
88
const TargetRegisterInfo *TRI,
89
const TargetInstrInfo *TII,
90
unsigned &PhysReg, int &Cost) {
91
if (Op != 2 || User->getOpcode() != ISD::CopyToReg)
94
unsigned Reg = cast<RegisterSDNode>(User->getOperand(1))->getReg();
95
if (TargetRegisterInfo::isVirtualRegister(Reg))
98
unsigned ResNo = User->getOperand(2).getResNo();
99
if (Def->isMachineOpcode()) {
100
const TargetInstrDesc &II = TII->get(Def->getMachineOpcode());
101
if (ResNo >= II.getNumDefs() &&
102
II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) {
104
const TargetRegisterClass *RC =
105
TRI->getMinimalPhysRegClass(Reg, Def->getValueType(ResNo));
106
Cost = RC->getCopyCost();
111
static void AddFlags(SDNode *N, SDValue Flag, bool AddFlag,
113
SmallVector<EVT, 4> VTs;
114
SDNode *FlagDestNode = Flag.getNode();
116
// Don't add a flag from a node to itself.
117
if (FlagDestNode == N) return;
119
// Don't add a flag to something which already has a flag.
120
if (N->getValueType(N->getNumValues() - 1) == MVT::Flag) return;
122
for (unsigned I = 0, E = N->getNumValues(); I != E; ++I)
123
VTs.push_back(N->getValueType(I));
126
VTs.push_back(MVT::Flag);
128
SmallVector<SDValue, 4> Ops;
129
for (unsigned I = 0, E = N->getNumOperands(); I != E; ++I)
130
Ops.push_back(N->getOperand(I));
135
SDVTList VTList = DAG->getVTList(&VTs[0], VTs.size());
136
MachineSDNode::mmo_iterator Begin = 0, End = 0;
137
MachineSDNode *MN = dyn_cast<MachineSDNode>(N);
139
// Store memory references.
141
Begin = MN->memoperands_begin();
142
End = MN->memoperands_end();
145
DAG->MorphNodeTo(N, N->getOpcode(), VTList, &Ops[0], Ops.size());
147
// Reset the memory references
149
MN->setMemRefs(Begin, End);
152
/// ClusterNeighboringLoads - Force nearby loads together by "flagging" them.
153
/// This function finds loads of the same base and different offsets. If the
154
/// offsets are not far apart (target specific), it add MVT::Flag inputs and
155
/// outputs to ensure they are scheduled together and in order. This
156
/// optimization may benefit some targets by improving cache locality.
157
void ScheduleDAGSDNodes::ClusterNeighboringLoads(SDNode *Node) {
159
unsigned NumOps = Node->getNumOperands();
160
if (Node->getOperand(NumOps-1).getValueType() == MVT::Other)
161
Chain = Node->getOperand(NumOps-1).getNode();
165
// Look for other loads of the same chain. Find loads that are loading from
166
// the same base pointer and different offsets.
167
SmallPtrSet<SDNode*, 16> Visited;
168
SmallVector<int64_t, 4> Offsets;
169
DenseMap<long long, SDNode*> O2SMap; // Map from offset to SDNode.
170
bool Cluster = false;
172
for (SDNode::use_iterator I = Chain->use_begin(), E = Chain->use_end();
175
if (User == Node || !Visited.insert(User))
177
int64_t Offset1, Offset2;
178
if (!TII->areLoadsFromSameBasePtr(Base, User, Offset1, Offset2) ||
180
// FIXME: Should be ok if they addresses are identical. But earlier
181
// optimizations really should have eliminated one of the loads.
183
if (O2SMap.insert(std::make_pair(Offset1, Base)).second)
184
Offsets.push_back(Offset1);
185
O2SMap.insert(std::make_pair(Offset2, User));
186
Offsets.push_back(Offset2);
187
if (Offset2 < Offset1)
195
// Sort them in increasing order.
196
std::sort(Offsets.begin(), Offsets.end());
198
// Check if the loads are close enough.
199
SmallVector<SDNode*, 4> Loads;
200
unsigned NumLoads = 0;
201
int64_t BaseOff = Offsets[0];
202
SDNode *BaseLoad = O2SMap[BaseOff];
203
Loads.push_back(BaseLoad);
204
for (unsigned i = 1, e = Offsets.size(); i != e; ++i) {
205
int64_t Offset = Offsets[i];
206
SDNode *Load = O2SMap[Offset];
207
if (!TII->shouldScheduleLoadsNear(BaseLoad, Load, BaseOff, Offset,NumLoads))
208
break; // Stop right here. Ignore loads that are further away.
209
Loads.push_back(Load);
216
// Cluster loads by adding MVT::Flag outputs and inputs. This also
217
// ensure they are scheduled in order of increasing addresses.
218
SDNode *Lead = Loads[0];
219
AddFlags(Lead, SDValue(0, 0), true, DAG);
221
SDValue InFlag = SDValue(Lead, Lead->getNumValues() - 1);
222
for (unsigned I = 1, E = Loads.size(); I != E; ++I) {
223
bool OutFlag = I < E - 1;
224
SDNode *Load = Loads[I];
226
AddFlags(Load, InFlag, OutFlag, DAG);
229
InFlag = SDValue(Load, Load->getNumValues() - 1);
235
/// ClusterNodes - Cluster certain nodes which should be scheduled together.
237
void ScheduleDAGSDNodes::ClusterNodes() {
238
for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
239
E = DAG->allnodes_end(); NI != E; ++NI) {
241
if (!Node || !Node->isMachineOpcode())
244
unsigned Opc = Node->getMachineOpcode();
245
const TargetInstrDesc &TID = TII->get(Opc);
247
// Cluster loads from "near" addresses into combined SUnits.
248
ClusterNeighboringLoads(Node);
252
void ScheduleDAGSDNodes::BuildSchedUnits() {
253
// During scheduling, the NodeId field of SDNode is used to map SDNodes
254
// to their associated SUnits by holding SUnits table indices. A value
255
// of -1 means the SDNode does not yet have an associated SUnit.
256
unsigned NumNodes = 0;
257
for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
258
E = DAG->allnodes_end(); NI != E; ++NI) {
263
// Reserve entries in the vector for each of the SUnits we are creating. This
264
// ensure that reallocation of the vector won't happen, so SUnit*'s won't get
266
// FIXME: Multiply by 2 because we may clone nodes during scheduling.
267
// This is a temporary workaround.
268
SUnits.reserve(NumNodes * 2);
270
// Add all nodes in depth first order.
271
SmallVector<SDNode*, 64> Worklist;
272
SmallPtrSet<SDNode*, 64> Visited;
273
Worklist.push_back(DAG->getRoot().getNode());
274
Visited.insert(DAG->getRoot().getNode());
276
while (!Worklist.empty()) {
277
SDNode *NI = Worklist.pop_back_val();
279
// Add all operands to the worklist unless they've already been added.
280
for (unsigned i = 0, e = NI->getNumOperands(); i != e; ++i)
281
if (Visited.insert(NI->getOperand(i).getNode()))
282
Worklist.push_back(NI->getOperand(i).getNode());
284
if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
287
// If this node has already been processed, stop now.
288
if (NI->getNodeId() != -1) continue;
290
SUnit *NodeSUnit = NewSUnit(NI);
292
// See if anything is flagged to this node, if so, add them to flagged
293
// nodes. Nodes can have at most one flag input and one flag output. Flags
294
// are required to be the last operand and result of a node.
296
// Scan up to find flagged preds.
298
while (N->getNumOperands() &&
299
N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
300
N = N->getOperand(N->getNumOperands()-1).getNode();
301
assert(N->getNodeId() == -1 && "Node already inserted!");
302
N->setNodeId(NodeSUnit->NodeNum);
305
// Scan down to find any flagged succs.
307
while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
308
SDValue FlagVal(N, N->getNumValues()-1);
310
// There are either zero or one users of the Flag result.
311
bool HasFlagUse = false;
312
for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
314
if (FlagVal.isOperandOf(*UI)) {
316
assert(N->getNodeId() == -1 && "Node already inserted!");
317
N->setNodeId(NodeSUnit->NodeNum);
321
if (!HasFlagUse) break;
324
// If there are flag operands involved, N is now the bottom-most node
325
// of the sequence of nodes that are flagged together.
327
NodeSUnit->setNode(N);
328
assert(N->getNodeId() == -1 && "Node already inserted!");
329
N->setNodeId(NodeSUnit->NodeNum);
331
// Assign the Latency field of NodeSUnit using target-provided information.
332
ComputeLatency(NodeSUnit);
336
void ScheduleDAGSDNodes::AddSchedEdges() {
337
const TargetSubtarget &ST = TM.getSubtarget<TargetSubtarget>();
339
// Check to see if the scheduler cares about latencies.
340
bool UnitLatencies = ForceUnitLatencies();
342
// Pass 2: add the preds, succs, etc.
343
for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
344
SUnit *SU = &SUnits[su];
345
SDNode *MainNode = SU->getNode();
347
if (MainNode->isMachineOpcode()) {
348
unsigned Opc = MainNode->getMachineOpcode();
349
const TargetInstrDesc &TID = TII->get(Opc);
350
for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
351
if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
352
SU->isTwoAddress = true;
356
if (TID.isCommutable())
357
SU->isCommutable = true;
360
// Find all predecessors and successors of the group.
361
for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode()) {
362
if (N->isMachineOpcode() &&
363
TII->get(N->getMachineOpcode()).getImplicitDefs()) {
364
SU->hasPhysRegClobbers = true;
365
unsigned NumUsed = InstrEmitter::CountResults(N);
366
while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
367
--NumUsed; // Skip over unused values at the end.
368
if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
369
SU->hasPhysRegDefs = true;
372
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
373
SDNode *OpN = N->getOperand(i).getNode();
374
if (isPassiveNode(OpN)) continue; // Not scheduled.
375
SUnit *OpSU = &SUnits[OpN->getNodeId()];
376
assert(OpSU && "Node has no SUnit!");
377
if (OpSU == SU) continue; // In the same group.
379
EVT OpVT = N->getOperand(i).getValueType();
380
assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!");
381
bool isChain = OpVT == MVT::Other;
383
unsigned PhysReg = 0;
385
// Determine if this is a physical register dependency.
386
CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
387
assert((PhysReg == 0 || !isChain) &&
388
"Chain dependence via physreg data?");
389
// FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
390
// emits a copy from the physical register to a virtual register unless
391
// it requires a cross class copy (cost < 0). That means we are only
392
// treating "expensive to copy" register dependency as physical register
393
// dependency. This may change in the future though.
397
// If this is a ctrl dep, latency is 1.
398
unsigned OpLatency = isChain ? 1 : OpSU->Latency;
399
const SDep &dep = SDep(OpSU, isChain ? SDep::Order : SDep::Data,
401
if (!isChain && !UnitLatencies) {
402
ComputeOperandLatency(OpN, N, i, const_cast<SDep &>(dep));
403
ST.adjustSchedDependency(OpSU, SU, const_cast<SDep &>(dep));
412
/// BuildSchedGraph - Build the SUnit graph from the selection dag that we
413
/// are input. This SUnit graph is similar to the SelectionDAG, but
414
/// excludes nodes that aren't interesting to scheduling, and represents
415
/// flagged together nodes with a single SUnit.
416
void ScheduleDAGSDNodes::BuildSchedGraph(AliasAnalysis *AA) {
417
// Cluster certain nodes which should be scheduled together.
419
// Populate the SUnits array.
421
// Compute all the scheduling dependencies between nodes.
425
void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
426
// Check to see if the scheduler cares about latencies.
427
if (ForceUnitLatencies()) {
432
const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
433
if (InstrItins.isEmpty()) {
438
// Compute the latency for the node. We use the sum of the latencies for
439
// all nodes flagged together into this SUnit.
441
for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode())
442
if (N->isMachineOpcode()) {
443
SU->Latency += InstrItins.
444
getStageLatency(TII->get(N->getMachineOpcode()).getSchedClass());
448
void ScheduleDAGSDNodes::ComputeOperandLatency(SDNode *Def, SDNode *Use,
449
unsigned OpIdx, SDep& dep) const{
450
// Check to see if the scheduler cares about latencies.
451
if (ForceUnitLatencies())
454
const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
455
if (InstrItins.isEmpty())
458
if (dep.getKind() != SDep::Data)
461
unsigned DefIdx = Use->getOperand(OpIdx).getResNo();
462
if (Def->isMachineOpcode()) {
463
const TargetInstrDesc &II = TII->get(Def->getMachineOpcode());
464
if (DefIdx >= II.getNumDefs())
466
int DefCycle = InstrItins.getOperandCycle(II.getSchedClass(), DefIdx);
470
if (Use->isMachineOpcode()) {
471
const unsigned UseClass = TII->get(Use->getMachineOpcode()).getSchedClass();
472
UseCycle = InstrItins.getOperandCycle(UseClass, OpIdx);
475
int Latency = DefCycle - UseCycle + 1;
477
dep.setLatency(Latency);
482
void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const {
483
if (!SU->getNode()) {
484
dbgs() << "PHYS REG COPY\n";
488
SU->getNode()->dump(DAG);
490
SmallVector<SDNode *, 4> FlaggedNodes;
491
for (SDNode *N = SU->getNode()->getFlaggedNode(); N; N = N->getFlaggedNode())
492
FlaggedNodes.push_back(N);
493
while (!FlaggedNodes.empty()) {
495
FlaggedNodes.back()->dump(DAG);
497
FlaggedNodes.pop_back();
503
bool operator()(const std::pair<unsigned, MachineInstr*> &A,
504
const std::pair<unsigned, MachineInstr*> &B) {
505
return A.first < B.first;
510
// ProcessSourceNode - Process nodes with source order numbers. These are added
511
// to a vector which EmitSchedule uses to determine how to insert dbg_value
512
// instructions in the right order.
513
static void ProcessSourceNode(SDNode *N, SelectionDAG *DAG,
514
InstrEmitter &Emitter,
515
DenseMap<SDValue, unsigned> &VRBaseMap,
516
SmallVector<std::pair<unsigned, MachineInstr*>, 32> &Orders,
517
SmallSet<unsigned, 8> &Seen) {
518
unsigned Order = DAG->GetOrdering(N);
519
if (!Order || !Seen.insert(Order))
522
MachineBasicBlock *BB = Emitter.getBlock();
523
if (Emitter.getInsertPos() == BB->begin() || BB->back().isPHI()) {
524
// Did not insert any instruction.
525
Orders.push_back(std::make_pair(Order, (MachineInstr*)0));
529
Orders.push_back(std::make_pair(Order, prior(Emitter.getInsertPos())));
530
if (!N->getHasDebugValue())
532
// Opportunistically insert immediate dbg_value uses, i.e. those with source
533
// order number right after the N.
534
MachineBasicBlock::iterator InsertPos = Emitter.getInsertPos();
535
SmallVector<SDDbgValue*,2> &DVs = DAG->GetDbgValues(N);
536
for (unsigned i = 0, e = DVs.size(); i != e; ++i) {
537
if (DVs[i]->isInvalidated())
539
unsigned DVOrder = DVs[i]->getOrder();
540
if (DVOrder == ++Order) {
541
MachineInstr *DbgMI = Emitter.EmitDbgValue(DVs[i], VRBaseMap);
543
Orders.push_back(std::make_pair(DVOrder, DbgMI));
544
BB->insert(InsertPos, DbgMI);
546
DVs[i]->setIsInvalidated();
552
/// EmitSchedule - Emit the machine code in scheduled order.
553
MachineBasicBlock *ScheduleDAGSDNodes::EmitSchedule() {
554
InstrEmitter Emitter(BB, InsertPos);
555
DenseMap<SDValue, unsigned> VRBaseMap;
556
DenseMap<SUnit*, unsigned> CopyVRBaseMap;
557
SmallVector<std::pair<unsigned, MachineInstr*>, 32> Orders;
558
SmallSet<unsigned, 8> Seen;
559
bool HasDbg = DAG->hasDebugValues();
561
// If this is the first BB, emit byval parameter dbg_value's.
562
if (HasDbg && BB->getParent()->begin() == MachineFunction::iterator(BB)) {
563
SDDbgInfo::DbgIterator PDI = DAG->ByvalParmDbgBegin();
564
SDDbgInfo::DbgIterator PDE = DAG->ByvalParmDbgEnd();
565
for (; PDI != PDE; ++PDI) {
566
MachineInstr *DbgMI= Emitter.EmitDbgValue(*PDI, VRBaseMap);
568
BB->insert(InsertPos, DbgMI);
572
for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
573
SUnit *SU = Sequence[i];
575
// Null SUnit* is a noop.
580
// For pre-regalloc scheduling, create instructions corresponding to the
581
// SDNode and any flagged SDNodes and append them to the block.
582
if (!SU->getNode()) {
584
EmitPhysRegCopy(SU, CopyVRBaseMap);
588
SmallVector<SDNode *, 4> FlaggedNodes;
589
for (SDNode *N = SU->getNode()->getFlaggedNode(); N;
590
N = N->getFlaggedNode())
591
FlaggedNodes.push_back(N);
592
while (!FlaggedNodes.empty()) {
593
SDNode *N = FlaggedNodes.back();
594
Emitter.EmitNode(FlaggedNodes.back(), SU->OrigNode != SU, SU->isCloned,
596
// Remember the source order of the inserted instruction.
598
ProcessSourceNode(N, DAG, Emitter, VRBaseMap, Orders, Seen);
599
FlaggedNodes.pop_back();
601
Emitter.EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned,
603
// Remember the source order of the inserted instruction.
605
ProcessSourceNode(SU->getNode(), DAG, Emitter, VRBaseMap, Orders,
609
// Insert all the dbg_values which have not already been inserted in source
612
MachineBasicBlock::iterator BBBegin = BB->getFirstNonPHI();
614
// Sort the source order instructions and use the order to insert debug
616
std::sort(Orders.begin(), Orders.end(), OrderSorter());
618
SDDbgInfo::DbgIterator DI = DAG->DbgBegin();
619
SDDbgInfo::DbgIterator DE = DAG->DbgEnd();
620
// Now emit the rest according to source order.
621
unsigned LastOrder = 0;
622
for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) {
623
unsigned Order = Orders[i].first;
624
MachineInstr *MI = Orders[i].second;
625
// Insert all SDDbgValue's whose order(s) are before "Order".
629
unsigned LastDIOrder = 0;
632
(*DI)->getOrder() >= LastOrder && (*DI)->getOrder() < Order; ++DI) {
634
assert((*DI)->getOrder() >= LastDIOrder &&
635
"SDDbgValue nodes must be in source order!");
636
LastDIOrder = (*DI)->getOrder();
638
if ((*DI)->isInvalidated())
640
MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap);
643
// Insert to start of the BB (after PHIs).
644
BB->insert(BBBegin, DbgMI);
646
// Insert at the instruction, which may be in a different
647
// block, if the block was split by a custom inserter.
648
MachineBasicBlock::iterator Pos = MI;
649
MI->getParent()->insert(llvm::next(Pos), DbgMI);
655
// Add trailing DbgValue's before the terminator. FIXME: May want to add
656
// some of them before one or more conditional branches?
658
MachineBasicBlock *InsertBB = Emitter.getBlock();
659
MachineBasicBlock::iterator Pos= Emitter.getBlock()->getFirstTerminator();
660
if (!(*DI)->isInvalidated()) {
661
MachineInstr *DbgMI= Emitter.EmitDbgValue(*DI, VRBaseMap);
663
InsertBB->insert(Pos, DbgMI);
669
BB = Emitter.getBlock();
670
InsertPos = Emitter.getInsertPos();