Run DCE after a LoopFlatten test to reduce spurious output [nfc]
[llvm-project.git] / llvm / lib / CodeGen / SelectionDAG / ScheduleDAGSDNodes.cpp
blob0579c1664d5c9abadebdcea06652f76f9b9c28e2
1 //===--- ScheduleDAGSDNodes.cpp - Implement the ScheduleDAGSDNodes class --===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This implements the ScheduleDAG class, which is a base class used by
10 // scheduling implementation classes.
12 //===----------------------------------------------------------------------===//
14 #include "ScheduleDAGSDNodes.h"
15 #include "InstrEmitter.h"
16 #include "SDNodeDbgValue.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/SelectionDAG.h"
25 #include "llvm/CodeGen/TargetInstrInfo.h"
26 #include "llvm/CodeGen/TargetLowering.h"
27 #include "llvm/CodeGen/TargetRegisterInfo.h"
28 #include "llvm/CodeGen/TargetSubtargetInfo.h"
29 #include "llvm/Config/llvm-config.h"
30 #include "llvm/MC/MCInstrItineraries.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetMachine.h"
35 using namespace llvm;
37 #define DEBUG_TYPE "pre-RA-sched"
39 STATISTIC(LoadsClustered, "Number of loads clustered together");
41 // This allows the latency-based scheduler to notice high latency instructions
42 // without a target itinerary. The choice of number here has more to do with
43 // balancing scheduler heuristics than with the actual machine latency.
44 static cl::opt<int> HighLatencyCycles(
45 "sched-high-latency-cycles", cl::Hidden, cl::init(10),
46 cl::desc("Roughly estimate the number of cycles that 'long latency'"
47 "instructions take for targets with no itinerary"));
49 ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
50 : ScheduleDAG(mf), InstrItins(mf.getSubtarget().getInstrItineraryData()) {}
52 /// Run - perform scheduling.
53 ///
54 void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb) {
55 BB = bb;
56 DAG = dag;
58 // Clear the scheduler's SUnit DAG.
59 ScheduleDAG::clearDAG();
60 Sequence.clear();
62 // Invoke the target's selection of scheduler.
63 Schedule();
66 /// NewSUnit - Creates a new SUnit and return a ptr to it.
67 ///
68 SUnit *ScheduleDAGSDNodes::newSUnit(SDNode *N) {
69 #ifndef NDEBUG
70 const SUnit *Addr = nullptr;
71 if (!SUnits.empty())
72 Addr = &SUnits[0];
73 #endif
74 SUnits.emplace_back(N, (unsigned)SUnits.size());
75 assert((Addr == nullptr || Addr == &SUnits[0]) &&
76 "SUnits std::vector reallocated on the fly!");
77 SUnits.back().OrigNode = &SUnits.back();
78 SUnit *SU = &SUnits.back();
79 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
80 if (!N ||
81 (N->isMachineOpcode() &&
82 N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF))
83 SU->SchedulingPref = Sched::None;
84 else
85 SU->SchedulingPref = TLI.getSchedulingPreference(N);
86 return SU;
89 SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) {
90 SUnit *SU = newSUnit(Old->getNode());
91 SU->OrigNode = Old->OrigNode;
92 SU->Latency = Old->Latency;
93 SU->isVRegCycle = Old->isVRegCycle;
94 SU->isCall = Old->isCall;
95 SU->isCallOp = Old->isCallOp;
96 SU->isTwoAddress = Old->isTwoAddress;
97 SU->isCommutable = Old->isCommutable;
98 SU->hasPhysRegDefs = Old->hasPhysRegDefs;
99 SU->hasPhysRegClobbers = Old->hasPhysRegClobbers;
100 SU->isScheduleHigh = Old->isScheduleHigh;
101 SU->isScheduleLow = Old->isScheduleLow;
102 SU->SchedulingPref = Old->SchedulingPref;
103 Old->isCloned = true;
104 return SU;
107 /// CheckForPhysRegDependency - Check if the dependency between def and use of
108 /// a specified operand is a physical register dependency. If so, returns the
109 /// register and the cost of copying the register.
110 static void CheckForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op,
111 const TargetRegisterInfo *TRI,
112 const TargetInstrInfo *TII,
113 const TargetLowering &TLI,
114 unsigned &PhysReg, int &Cost) {
115 if (Op != 2 || User->getOpcode() != ISD::CopyToReg)
116 return;
118 unsigned Reg = cast<RegisterSDNode>(User->getOperand(1))->getReg();
119 if (TLI.checkForPhysRegDependency(Def, User, Op, TRI, TII, PhysReg, Cost))
120 return;
122 if (Register::isVirtualRegister(Reg))
123 return;
125 unsigned ResNo = User->getOperand(2).getResNo();
126 if (Def->getOpcode() == ISD::CopyFromReg &&
127 cast<RegisterSDNode>(Def->getOperand(1))->getReg() == Reg) {
128 PhysReg = Reg;
129 } else if (Def->isMachineOpcode()) {
130 const MCInstrDesc &II = TII->get(Def->getMachineOpcode());
131 if (ResNo >= II.getNumDefs() && II.hasImplicitDefOfPhysReg(Reg))
132 PhysReg = Reg;
135 if (PhysReg != 0) {
136 const TargetRegisterClass *RC =
137 TRI->getMinimalPhysRegClass(Reg, Def->getSimpleValueType(ResNo));
138 Cost = RC->getCopyCost();
142 // Helper for AddGlue to clone node operands.
143 static void CloneNodeWithValues(SDNode *N, SelectionDAG *DAG, ArrayRef<EVT> VTs,
144 SDValue ExtraOper = SDValue()) {
145 SmallVector<SDValue, 8> Ops(N->op_begin(), N->op_end());
146 if (ExtraOper.getNode())
147 Ops.push_back(ExtraOper);
149 SDVTList VTList = DAG->getVTList(VTs);
150 MachineSDNode *MN = dyn_cast<MachineSDNode>(N);
152 // Store memory references.
153 SmallVector<MachineMemOperand *, 2> MMOs;
154 if (MN)
155 MMOs.assign(MN->memoperands_begin(), MN->memoperands_end());
157 DAG->MorphNodeTo(N, N->getOpcode(), VTList, Ops);
159 // Reset the memory references
160 if (MN)
161 DAG->setNodeMemRefs(MN, MMOs);
164 static bool AddGlue(SDNode *N, SDValue Glue, bool AddGlue, SelectionDAG *DAG) {
165 SDNode *GlueDestNode = Glue.getNode();
167 // Don't add glue from a node to itself.
168 if (GlueDestNode == N) return false;
170 // Don't add a glue operand to something that already uses glue.
171 if (GlueDestNode &&
172 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) {
173 return false;
175 // Don't add glue to something that already has a glue value.
176 if (N->getValueType(N->getNumValues() - 1) == MVT::Glue) return false;
178 SmallVector<EVT, 4> VTs(N->values());
179 if (AddGlue)
180 VTs.push_back(MVT::Glue);
182 CloneNodeWithValues(N, DAG, VTs, Glue);
184 return true;
187 // Cleanup after unsuccessful AddGlue. Use the standard method of morphing the
188 // node even though simply shrinking the value list is sufficient.
189 static void RemoveUnusedGlue(SDNode *N, SelectionDAG *DAG) {
190 assert((N->getValueType(N->getNumValues() - 1) == MVT::Glue &&
191 !N->hasAnyUseOfValue(N->getNumValues() - 1)) &&
192 "expected an unused glue value");
194 CloneNodeWithValues(N, DAG,
195 ArrayRef(N->value_begin(), N->getNumValues() - 1));
198 /// ClusterNeighboringLoads - Force nearby loads together by "gluing" them.
199 /// This function finds loads of the same base and different offsets. If the
200 /// offsets are not far apart (target specific), it add MVT::Glue inputs and
201 /// outputs to ensure they are scheduled together and in order. This
202 /// optimization may benefit some targets by improving cache locality.
203 void ScheduleDAGSDNodes::ClusterNeighboringLoads(SDNode *Node) {
204 SDValue Chain;
205 unsigned NumOps = Node->getNumOperands();
206 if (Node->getOperand(NumOps-1).getValueType() == MVT::Other)
207 Chain = Node->getOperand(NumOps-1);
208 if (!Chain)
209 return;
211 // Skip any load instruction that has a tied input. There may be an additional
212 // dependency requiring a different order than by increasing offsets, and the
213 // added glue may introduce a cycle.
214 auto hasTiedInput = [this](const SDNode *N) {
215 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
216 for (unsigned I = 0; I != MCID.getNumOperands(); ++I) {
217 if (MCID.getOperandConstraint(I, MCOI::TIED_TO) != -1)
218 return true;
221 return false;
224 // Look for other loads of the same chain. Find loads that are loading from
225 // the same base pointer and different offsets.
226 SmallPtrSet<SDNode*, 16> Visited;
227 SmallVector<int64_t, 4> Offsets;
228 DenseMap<long long, SDNode*> O2SMap; // Map from offset to SDNode.
229 bool Cluster = false;
230 SDNode *Base = Node;
232 if (hasTiedInput(Base))
233 return;
235 // This algorithm requires a reasonably low use count before finding a match
236 // to avoid uselessly blowing up compile time in large blocks.
237 unsigned UseCount = 0;
238 for (SDNode::use_iterator I = Chain->use_begin(), E = Chain->use_end();
239 I != E && UseCount < 100; ++I, ++UseCount) {
240 if (I.getUse().getResNo() != Chain.getResNo())
241 continue;
243 SDNode *User = *I;
244 if (User == Node || !Visited.insert(User).second)
245 continue;
246 int64_t Offset1, Offset2;
247 if (!TII->areLoadsFromSameBasePtr(Base, User, Offset1, Offset2) ||
248 Offset1 == Offset2 ||
249 hasTiedInput(User)) {
250 // FIXME: Should be ok if they addresses are identical. But earlier
251 // optimizations really should have eliminated one of the loads.
252 continue;
254 if (O2SMap.insert(std::make_pair(Offset1, Base)).second)
255 Offsets.push_back(Offset1);
256 O2SMap.insert(std::make_pair(Offset2, User));
257 Offsets.push_back(Offset2);
258 if (Offset2 < Offset1)
259 Base = User;
260 Cluster = true;
261 // Reset UseCount to allow more matches.
262 UseCount = 0;
265 if (!Cluster)
266 return;
268 // Sort them in increasing order.
269 llvm::sort(Offsets);
271 // Check if the loads are close enough.
272 SmallVector<SDNode*, 4> Loads;
273 unsigned NumLoads = 0;
274 int64_t BaseOff = Offsets[0];
275 SDNode *BaseLoad = O2SMap[BaseOff];
276 Loads.push_back(BaseLoad);
277 for (unsigned i = 1, e = Offsets.size(); i != e; ++i) {
278 int64_t Offset = Offsets[i];
279 SDNode *Load = O2SMap[Offset];
280 if (!TII->shouldScheduleLoadsNear(BaseLoad, Load, BaseOff, Offset,NumLoads))
281 break; // Stop right here. Ignore loads that are further away.
282 Loads.push_back(Load);
283 ++NumLoads;
286 if (NumLoads == 0)
287 return;
289 // Cluster loads by adding MVT::Glue outputs and inputs. This also
290 // ensure they are scheduled in order of increasing addresses.
291 SDNode *Lead = Loads[0];
292 SDValue InGlue;
293 if (AddGlue(Lead, InGlue, true, DAG))
294 InGlue = SDValue(Lead, Lead->getNumValues() - 1);
295 for (unsigned I = 1, E = Loads.size(); I != E; ++I) {
296 bool OutGlue = I < E - 1;
297 SDNode *Load = Loads[I];
299 // If AddGlue fails, we could leave an unsused glue value. This should not
300 // cause any
301 if (AddGlue(Load, InGlue, OutGlue, DAG)) {
302 if (OutGlue)
303 InGlue = SDValue(Load, Load->getNumValues() - 1);
305 ++LoadsClustered;
307 else if (!OutGlue && InGlue.getNode())
308 RemoveUnusedGlue(InGlue.getNode(), DAG);
312 /// ClusterNodes - Cluster certain nodes which should be scheduled together.
314 void ScheduleDAGSDNodes::ClusterNodes() {
315 for (SDNode &NI : DAG->allnodes()) {
316 SDNode *Node = &NI;
317 if (!Node || !Node->isMachineOpcode())
318 continue;
320 unsigned Opc = Node->getMachineOpcode();
321 const MCInstrDesc &MCID = TII->get(Opc);
322 if (MCID.mayLoad())
323 // Cluster loads from "near" addresses into combined SUnits.
324 ClusterNeighboringLoads(Node);
328 void ScheduleDAGSDNodes::BuildSchedUnits() {
329 // During scheduling, the NodeId field of SDNode is used to map SDNodes
330 // to their associated SUnits by holding SUnits table indices. A value
331 // of -1 means the SDNode does not yet have an associated SUnit.
332 unsigned NumNodes = 0;
333 for (SDNode &NI : DAG->allnodes()) {
334 NI.setNodeId(-1);
335 ++NumNodes;
338 // Reserve entries in the vector for each of the SUnits we are creating. This
339 // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
340 // invalidated.
341 // FIXME: Multiply by 2 because we may clone nodes during scheduling.
342 // This is a temporary workaround.
343 SUnits.reserve(NumNodes * 2);
345 // Add all nodes in depth first order.
346 SmallVector<SDNode*, 64> Worklist;
347 SmallPtrSet<SDNode*, 32> Visited;
348 Worklist.push_back(DAG->getRoot().getNode());
349 Visited.insert(DAG->getRoot().getNode());
351 SmallVector<SUnit*, 8> CallSUnits;
352 while (!Worklist.empty()) {
353 SDNode *NI = Worklist.pop_back_val();
355 // Add all operands to the worklist unless they've already been added.
356 for (const SDValue &Op : NI->op_values())
357 if (Visited.insert(Op.getNode()).second)
358 Worklist.push_back(Op.getNode());
360 if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
361 continue;
363 // If this node has already been processed, stop now.
364 if (NI->getNodeId() != -1) continue;
366 SUnit *NodeSUnit = newSUnit(NI);
368 // See if anything is glued to this node, if so, add them to glued
369 // nodes. Nodes can have at most one glue input and one glue output. Glue
370 // is required to be the last operand and result of a node.
372 // Scan up to find glued preds.
373 SDNode *N = NI;
374 while (N->getNumOperands() &&
375 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) {
376 N = N->getOperand(N->getNumOperands()-1).getNode();
377 assert(N->getNodeId() == -1 && "Node already inserted!");
378 N->setNodeId(NodeSUnit->NodeNum);
379 if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
380 NodeSUnit->isCall = true;
383 // Scan down to find any glued succs.
384 N = NI;
385 while (N->getValueType(N->getNumValues()-1) == MVT::Glue) {
386 SDValue GlueVal(N, N->getNumValues()-1);
388 // There are either zero or one users of the Glue result.
389 bool HasGlueUse = false;
390 for (SDNode *U : N->uses())
391 if (GlueVal.isOperandOf(U)) {
392 HasGlueUse = true;
393 assert(N->getNodeId() == -1 && "Node already inserted!");
394 N->setNodeId(NodeSUnit->NodeNum);
395 N = U;
396 if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
397 NodeSUnit->isCall = true;
398 break;
400 if (!HasGlueUse) break;
403 if (NodeSUnit->isCall)
404 CallSUnits.push_back(NodeSUnit);
406 // Schedule zero-latency TokenFactor below any nodes that may increase the
407 // schedule height. Otherwise, ancestors of the TokenFactor may appear to
408 // have false stalls.
409 if (NI->getOpcode() == ISD::TokenFactor)
410 NodeSUnit->isScheduleLow = true;
412 // If there are glue operands involved, N is now the bottom-most node
413 // of the sequence of nodes that are glued together.
414 // Update the SUnit.
415 NodeSUnit->setNode(N);
416 assert(N->getNodeId() == -1 && "Node already inserted!");
417 N->setNodeId(NodeSUnit->NodeNum);
419 // Compute NumRegDefsLeft. This must be done before AddSchedEdges.
420 InitNumRegDefsLeft(NodeSUnit);
422 // Assign the Latency field of NodeSUnit using target-provided information.
423 computeLatency(NodeSUnit);
426 // Find all call operands.
427 while (!CallSUnits.empty()) {
428 SUnit *SU = CallSUnits.pop_back_val();
429 for (const SDNode *SUNode = SU->getNode(); SUNode;
430 SUNode = SUNode->getGluedNode()) {
431 if (SUNode->getOpcode() != ISD::CopyToReg)
432 continue;
433 SDNode *SrcN = SUNode->getOperand(2).getNode();
434 if (isPassiveNode(SrcN)) continue; // Not scheduled.
435 SUnit *SrcSU = &SUnits[SrcN->getNodeId()];
436 SrcSU->isCallOp = true;
441 void ScheduleDAGSDNodes::AddSchedEdges() {
442 const TargetSubtargetInfo &ST = MF.getSubtarget();
444 // Check to see if the scheduler cares about latencies.
445 bool UnitLatencies = forceUnitLatencies();
447 // Pass 2: add the preds, succs, etc.
448 for (SUnit &SU : SUnits) {
449 SDNode *MainNode = SU.getNode();
451 if (MainNode->isMachineOpcode()) {
452 unsigned Opc = MainNode->getMachineOpcode();
453 const MCInstrDesc &MCID = TII->get(Opc);
454 for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
455 if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
456 SU.isTwoAddress = true;
457 break;
460 if (MCID.isCommutable())
461 SU.isCommutable = true;
464 // Find all predecessors and successors of the group.
465 for (SDNode *N = SU.getNode(); N; N = N->getGluedNode()) {
466 if (N->isMachineOpcode() &&
467 !TII->get(N->getMachineOpcode()).implicit_defs().empty()) {
468 SU.hasPhysRegClobbers = true;
469 unsigned NumUsed = InstrEmitter::CountResults(N);
470 while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
471 --NumUsed; // Skip over unused values at the end.
472 if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
473 SU.hasPhysRegDefs = true;
476 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
477 SDNode *OpN = N->getOperand(i).getNode();
478 unsigned DefIdx = N->getOperand(i).getResNo();
479 if (isPassiveNode(OpN)) continue; // Not scheduled.
480 SUnit *OpSU = &SUnits[OpN->getNodeId()];
481 assert(OpSU && "Node has no SUnit!");
482 if (OpSU == &SU)
483 continue; // In the same group.
485 EVT OpVT = N->getOperand(i).getValueType();
486 assert(OpVT != MVT::Glue && "Glued nodes should be in same sunit!");
487 bool isChain = OpVT == MVT::Other;
489 unsigned PhysReg = 0;
490 int Cost = 1;
491 // Determine if this is a physical register dependency.
492 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
493 CheckForPhysRegDependency(OpN, N, i, TRI, TII, TLI, PhysReg, Cost);
494 assert((PhysReg == 0 || !isChain) &&
495 "Chain dependence via physreg data?");
496 // FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
497 // emits a copy from the physical register to a virtual register unless
498 // it requires a cross class copy (cost < 0). That means we are only
499 // treating "expensive to copy" register dependency as physical register
500 // dependency. This may change in the future though.
501 if (Cost >= 0 && !StressSched)
502 PhysReg = 0;
504 // If this is a ctrl dep, latency is 1.
505 unsigned OpLatency = isChain ? 1 : OpSU->Latency;
506 // Special-case TokenFactor chains as zero-latency.
507 if(isChain && OpN->getOpcode() == ISD::TokenFactor)
508 OpLatency = 0;
510 SDep Dep = isChain ? SDep(OpSU, SDep::Barrier)
511 : SDep(OpSU, SDep::Data, PhysReg);
512 Dep.setLatency(OpLatency);
513 if (!isChain && !UnitLatencies) {
514 computeOperandLatency(OpN, N, i, Dep);
515 ST.adjustSchedDependency(OpSU, DefIdx, &SU, i, Dep);
518 if (!SU.addPred(Dep) && !Dep.isCtrl() && OpSU->NumRegDefsLeft > 1) {
519 // Multiple register uses are combined in the same SUnit. For example,
520 // we could have a set of glued nodes with all their defs consumed by
521 // another set of glued nodes. Register pressure tracking sees this as
522 // a single use, so to keep pressure balanced we reduce the defs.
524 // We can't tell (without more book-keeping) if this results from
525 // glued nodes or duplicate operands. As long as we don't reduce
526 // NumRegDefsLeft to zero, we handle the common cases well.
527 --OpSU->NumRegDefsLeft;
534 /// BuildSchedGraph - Build the SUnit graph from the selection dag that we
535 /// are input. This SUnit graph is similar to the SelectionDAG, but
536 /// excludes nodes that aren't interesting to scheduling, and represents
537 /// glued together nodes with a single SUnit.
538 void ScheduleDAGSDNodes::BuildSchedGraph(AAResults *AA) {
539 // Cluster certain nodes which should be scheduled together.
540 ClusterNodes();
541 // Populate the SUnits array.
542 BuildSchedUnits();
543 // Compute all the scheduling dependencies between nodes.
544 AddSchedEdges();
547 // Initialize NumNodeDefs for the current Node's opcode.
548 void ScheduleDAGSDNodes::RegDefIter::InitNodeNumDefs() {
549 // Check for phys reg copy.
550 if (!Node)
551 return;
553 if (!Node->isMachineOpcode()) {
554 if (Node->getOpcode() == ISD::CopyFromReg)
555 NodeNumDefs = 1;
556 else
557 NodeNumDefs = 0;
558 return;
560 unsigned POpc = Node->getMachineOpcode();
561 if (POpc == TargetOpcode::IMPLICIT_DEF) {
562 // No register need be allocated for this.
563 NodeNumDefs = 0;
564 return;
566 if (POpc == TargetOpcode::PATCHPOINT &&
567 Node->getValueType(0) == MVT::Other) {
568 // PATCHPOINT is defined to have one result, but it might really have none
569 // if we're not using CallingConv::AnyReg. Don't mistake the chain for a
570 // real definition.
571 NodeNumDefs = 0;
572 return;
574 unsigned NRegDefs = SchedDAG->TII->get(Node->getMachineOpcode()).getNumDefs();
575 // Some instructions define regs that are not represented in the selection DAG
576 // (e.g. unused flags). See tMOVi8. Make sure we don't access past NumValues.
577 NodeNumDefs = std::min(Node->getNumValues(), NRegDefs);
578 DefIdx = 0;
581 // Construct a RegDefIter for this SUnit and find the first valid value.
582 ScheduleDAGSDNodes::RegDefIter::RegDefIter(const SUnit *SU,
583 const ScheduleDAGSDNodes *SD)
584 : SchedDAG(SD), Node(SU->getNode()) {
585 InitNodeNumDefs();
586 Advance();
589 // Advance to the next valid value defined by the SUnit.
590 void ScheduleDAGSDNodes::RegDefIter::Advance() {
591 for (;Node;) { // Visit all glued nodes.
592 for (;DefIdx < NodeNumDefs; ++DefIdx) {
593 if (!Node->hasAnyUseOfValue(DefIdx))
594 continue;
595 ValueType = Node->getSimpleValueType(DefIdx);
596 ++DefIdx;
597 return; // Found a normal regdef.
599 Node = Node->getGluedNode();
600 if (!Node) {
601 return; // No values left to visit.
603 InitNodeNumDefs();
607 void ScheduleDAGSDNodes::InitNumRegDefsLeft(SUnit *SU) {
608 assert(SU->NumRegDefsLeft == 0 && "expect a new node");
609 for (RegDefIter I(SU, this); I.IsValid(); I.Advance()) {
610 assert(SU->NumRegDefsLeft < USHRT_MAX && "overflow is ok but unexpected");
611 ++SU->NumRegDefsLeft;
615 void ScheduleDAGSDNodes::computeLatency(SUnit *SU) {
616 SDNode *N = SU->getNode();
618 // TokenFactor operands are considered zero latency, and some schedulers
619 // (e.g. Top-Down list) may rely on the fact that operand latency is nonzero
620 // whenever node latency is nonzero.
621 if (N && N->getOpcode() == ISD::TokenFactor) {
622 SU->Latency = 0;
623 return;
626 // Check to see if the scheduler cares about latencies.
627 if (forceUnitLatencies()) {
628 SU->Latency = 1;
629 return;
632 if (!InstrItins || InstrItins->isEmpty()) {
633 if (N && N->isMachineOpcode() &&
634 TII->isHighLatencyDef(N->getMachineOpcode()))
635 SU->Latency = HighLatencyCycles;
636 else
637 SU->Latency = 1;
638 return;
641 // Compute the latency for the node. We use the sum of the latencies for
642 // all nodes glued together into this SUnit.
643 SU->Latency = 0;
644 for (SDNode *N = SU->getNode(); N; N = N->getGluedNode())
645 if (N->isMachineOpcode())
646 SU->Latency += TII->getInstrLatency(InstrItins, N);
649 void ScheduleDAGSDNodes::computeOperandLatency(SDNode *Def, SDNode *Use,
650 unsigned OpIdx, SDep& dep) const{
651 // Check to see if the scheduler cares about latencies.
652 if (forceUnitLatencies())
653 return;
655 if (dep.getKind() != SDep::Data)
656 return;
658 unsigned DefIdx = Use->getOperand(OpIdx).getResNo();
659 if (Use->isMachineOpcode())
660 // Adjust the use operand index by num of defs.
661 OpIdx += TII->get(Use->getMachineOpcode()).getNumDefs();
662 int Latency = TII->getOperandLatency(InstrItins, Def, DefIdx, Use, OpIdx);
663 if (Latency > 1 && Use->getOpcode() == ISD::CopyToReg &&
664 !BB->succ_empty()) {
665 unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
666 if (Register::isVirtualRegister(Reg))
667 // This copy is a liveout value. It is likely coalesced, so reduce the
668 // latency so not to penalize the def.
669 // FIXME: need target specific adjustment here?
670 Latency = Latency - 1;
672 if (Latency >= 0)
673 dep.setLatency(Latency);
676 void ScheduleDAGSDNodes::dumpNode(const SUnit &SU) const {
677 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
678 dumpNodeName(SU);
679 dbgs() << ": ";
681 if (!SU.getNode()) {
682 dbgs() << "PHYS REG COPY\n";
683 return;
686 SU.getNode()->dump(DAG);
687 dbgs() << "\n";
688 SmallVector<SDNode *, 4> GluedNodes;
689 for (SDNode *N = SU.getNode()->getGluedNode(); N; N = N->getGluedNode())
690 GluedNodes.push_back(N);
691 while (!GluedNodes.empty()) {
692 dbgs() << " ";
693 GluedNodes.back()->dump(DAG);
694 dbgs() << "\n";
695 GluedNodes.pop_back();
697 #endif
700 void ScheduleDAGSDNodes::dump() const {
701 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
702 if (EntrySU.getNode() != nullptr)
703 dumpNodeAll(EntrySU);
704 for (const SUnit &SU : SUnits)
705 dumpNodeAll(SU);
706 if (ExitSU.getNode() != nullptr)
707 dumpNodeAll(ExitSU);
708 #endif
711 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
712 void ScheduleDAGSDNodes::dumpSchedule() const {
713 for (const SUnit *SU : Sequence) {
714 if (SU)
715 dumpNode(*SU);
716 else
717 dbgs() << "**** NOOP ****\n";
720 #endif
722 #ifndef NDEBUG
723 /// VerifyScheduledSequence - Verify that all SUnits were scheduled and that
724 /// their state is consistent with the nodes listed in Sequence.
726 void ScheduleDAGSDNodes::VerifyScheduledSequence(bool isBottomUp) {
727 unsigned ScheduledNodes = ScheduleDAG::VerifyScheduledDAG(isBottomUp);
728 unsigned Noops = llvm::count(Sequence, nullptr);
729 assert(Sequence.size() - Noops == ScheduledNodes &&
730 "The number of nodes scheduled doesn't match the expected number!");
732 #endif // NDEBUG
734 /// ProcessSDDbgValues - Process SDDbgValues associated with this node.
735 static void
736 ProcessSDDbgValues(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter,
737 SmallVectorImpl<std::pair<unsigned, MachineInstr*> > &Orders,
738 DenseMap<SDValue, Register> &VRBaseMap, unsigned Order) {
739 if (!N->getHasDebugValue())
740 return;
742 /// Returns true if \p DV has any VReg operand locations which don't exist in
743 /// VRBaseMap.
744 auto HasUnknownVReg = [&VRBaseMap](SDDbgValue *DV) {
745 for (const SDDbgOperand &L : DV->getLocationOps()) {
746 if (L.getKind() == SDDbgOperand::SDNODE &&
747 VRBaseMap.count({L.getSDNode(), L.getResNo()}) == 0)
748 return true;
750 return false;
753 // Opportunistically insert immediate dbg_value uses, i.e. those with the same
754 // source order number as N.
755 MachineBasicBlock *BB = Emitter.getBlock();
756 MachineBasicBlock::iterator InsertPos = Emitter.getInsertPos();
757 for (auto *DV : DAG->GetDbgValues(N)) {
758 if (DV->isEmitted())
759 continue;
760 unsigned DVOrder = DV->getOrder();
761 if (Order != 0 && DVOrder != Order)
762 continue;
763 // If DV has any VReg location operands which haven't been mapped then
764 // either that node is no longer available or we just haven't visited the
765 // node yet. In the former case we should emit an undef dbg_value, but we
766 // can do it later. And for the latter we'll want to wait until all
767 // dependent nodes have been visited.
768 if (!DV->isInvalidated() && HasUnknownVReg(DV))
769 continue;
770 MachineInstr *DbgMI = Emitter.EmitDbgValue(DV, VRBaseMap);
771 if (!DbgMI)
772 continue;
773 Orders.push_back({DVOrder, DbgMI});
774 BB->insert(InsertPos, DbgMI);
778 // ProcessSourceNode - Process nodes with source order numbers. These are added
779 // to a vector which EmitSchedule uses to determine how to insert dbg_value
780 // instructions in the right order.
781 static void
782 ProcessSourceNode(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter,
783 DenseMap<SDValue, Register> &VRBaseMap,
784 SmallVectorImpl<std::pair<unsigned, MachineInstr *>> &Orders,
785 SmallSet<Register, 8> &Seen, MachineInstr *NewInsn) {
786 unsigned Order = N->getIROrder();
787 if (!Order || Seen.count(Order)) {
788 // Process any valid SDDbgValues even if node does not have any order
789 // assigned.
790 ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, 0);
791 return;
794 // If a new instruction was generated for this Order number, record it.
795 // Otherwise, leave this order number unseen: we will either find later
796 // instructions for it, or leave it unseen if there were no instructions at
797 // all.
798 if (NewInsn) {
799 Seen.insert(Order);
800 Orders.push_back({Order, NewInsn});
803 // Even if no instruction was generated, a Value may have become defined via
804 // earlier nodes. Try to process them now.
805 ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, Order);
808 void ScheduleDAGSDNodes::
809 EmitPhysRegCopy(SUnit *SU, DenseMap<SUnit*, Register> &VRBaseMap,
810 MachineBasicBlock::iterator InsertPos) {
811 for (const SDep &Pred : SU->Preds) {
812 if (Pred.isCtrl())
813 continue; // ignore chain preds
814 if (Pred.getSUnit()->CopyDstRC) {
815 // Copy to physical register.
816 DenseMap<SUnit *, Register>::iterator VRI =
817 VRBaseMap.find(Pred.getSUnit());
818 assert(VRI != VRBaseMap.end() && "Node emitted out of order - late");
819 // Find the destination physical register.
820 Register Reg;
821 for (const SDep &Succ : SU->Succs) {
822 if (Succ.isCtrl())
823 continue; // ignore chain preds
824 if (Succ.getReg()) {
825 Reg = Succ.getReg();
826 break;
829 BuildMI(*BB, InsertPos, DebugLoc(), TII->get(TargetOpcode::COPY), Reg)
830 .addReg(VRI->second);
831 } else {
832 // Copy from physical register.
833 assert(Pred.getReg() && "Unknown physical register!");
834 Register VRBase = MRI.createVirtualRegister(SU->CopyDstRC);
835 bool isNew = VRBaseMap.insert(std::make_pair(SU, VRBase)).second;
836 (void)isNew; // Silence compiler warning.
837 assert(isNew && "Node emitted out of order - early");
838 BuildMI(*BB, InsertPos, DebugLoc(), TII->get(TargetOpcode::COPY), VRBase)
839 .addReg(Pred.getReg());
841 break;
845 /// EmitSchedule - Emit the machine code in scheduled order. Return the new
846 /// InsertPos and MachineBasicBlock that contains this insertion
847 /// point. ScheduleDAGSDNodes holds a BB pointer for convenience, but this does
848 /// not necessarily refer to returned BB. The emitter may split blocks.
849 MachineBasicBlock *ScheduleDAGSDNodes::
850 EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
851 InstrEmitter Emitter(DAG->getTarget(), BB, InsertPos);
852 DenseMap<SDValue, Register> VRBaseMap;
853 DenseMap<SUnit*, Register> CopyVRBaseMap;
854 SmallVector<std::pair<unsigned, MachineInstr*>, 32> Orders;
855 SmallSet<Register, 8> Seen;
856 bool HasDbg = DAG->hasDebugValues();
858 // Emit a node, and determine where its first instruction is for debuginfo.
859 // Zero, one, or multiple instructions can be created when emitting a node.
860 auto EmitNode =
861 [&](SDNode *Node, bool IsClone, bool IsCloned,
862 DenseMap<SDValue, Register> &VRBaseMap) -> MachineInstr * {
863 // Fetch instruction prior to this, or end() if nonexistant.
864 auto GetPrevInsn = [&](MachineBasicBlock::iterator I) {
865 if (I == BB->begin())
866 return BB->end();
867 else
868 return std::prev(Emitter.getInsertPos());
871 MachineBasicBlock::iterator Before = GetPrevInsn(Emitter.getInsertPos());
872 Emitter.EmitNode(Node, IsClone, IsCloned, VRBaseMap);
873 MachineBasicBlock::iterator After = GetPrevInsn(Emitter.getInsertPos());
875 // If the iterator did not change, no instructions were inserted.
876 if (Before == After)
877 return nullptr;
879 MachineInstr *MI;
880 if (Before == BB->end()) {
881 // There were no prior instructions; the new ones must start at the
882 // beginning of the block.
883 MI = &Emitter.getBlock()->instr_front();
884 } else {
885 // Return first instruction after the pre-existing instructions.
886 MI = &*std::next(Before);
889 if (MI->isCandidateForCallSiteEntry() &&
890 DAG->getTarget().Options.EmitCallSiteInfo)
891 MF.addCallArgsForwardingRegs(MI, DAG->getCallSiteInfo(Node));
893 if (DAG->getNoMergeSiteInfo(Node)) {
894 MI->setFlag(MachineInstr::MIFlag::NoMerge);
897 if (MDNode *MD = DAG->getPCSections(Node))
898 MI->setPCSections(MF, MD);
900 return MI;
903 // If this is the first BB, emit byval parameter dbg_value's.
904 if (HasDbg && BB->getParent()->begin() == MachineFunction::iterator(BB)) {
905 SDDbgInfo::DbgIterator PDI = DAG->ByvalParmDbgBegin();
906 SDDbgInfo::DbgIterator PDE = DAG->ByvalParmDbgEnd();
907 for (; PDI != PDE; ++PDI) {
908 MachineInstr *DbgMI= Emitter.EmitDbgValue(*PDI, VRBaseMap);
909 if (DbgMI) {
910 BB->insert(InsertPos, DbgMI);
911 // We re-emit the dbg_value closer to its use, too, after instructions
912 // are emitted to the BB.
913 (*PDI)->clearIsEmitted();
918 for (SUnit *SU : Sequence) {
919 if (!SU) {
920 // Null SUnit* is a noop.
921 TII->insertNoop(*Emitter.getBlock(), InsertPos);
922 continue;
925 // For pre-regalloc scheduling, create instructions corresponding to the
926 // SDNode and any glued SDNodes and append them to the block.
927 if (!SU->getNode()) {
928 // Emit a copy.
929 EmitPhysRegCopy(SU, CopyVRBaseMap, InsertPos);
930 continue;
933 SmallVector<SDNode *, 4> GluedNodes;
934 for (SDNode *N = SU->getNode()->getGluedNode(); N; N = N->getGluedNode())
935 GluedNodes.push_back(N);
936 while (!GluedNodes.empty()) {
937 SDNode *N = GluedNodes.back();
938 auto NewInsn = EmitNode(N, SU->OrigNode != SU, SU->isCloned, VRBaseMap);
939 // Remember the source order of the inserted instruction.
940 if (HasDbg)
941 ProcessSourceNode(N, DAG, Emitter, VRBaseMap, Orders, Seen, NewInsn);
943 if (MDNode *MD = DAG->getHeapAllocSite(N))
944 if (NewInsn && NewInsn->isCall())
945 NewInsn->setHeapAllocMarker(MF, MD);
947 GluedNodes.pop_back();
949 auto NewInsn =
950 EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned, VRBaseMap);
951 // Remember the source order of the inserted instruction.
952 if (HasDbg)
953 ProcessSourceNode(SU->getNode(), DAG, Emitter, VRBaseMap, Orders, Seen,
954 NewInsn);
956 if (MDNode *MD = DAG->getHeapAllocSite(SU->getNode())) {
957 if (NewInsn && NewInsn->isCall())
958 NewInsn->setHeapAllocMarker(MF, MD);
962 // Insert all the dbg_values which have not already been inserted in source
963 // order sequence.
964 if (HasDbg) {
965 MachineBasicBlock::iterator BBBegin = BB->getFirstNonPHI();
967 // Sort the source order instructions and use the order to insert debug
968 // values. Use stable_sort so that DBG_VALUEs are inserted in the same order
969 // regardless of the host's implementation fo std::sort.
970 llvm::stable_sort(Orders, less_first());
971 std::stable_sort(DAG->DbgBegin(), DAG->DbgEnd(),
972 [](const SDDbgValue *LHS, const SDDbgValue *RHS) {
973 return LHS->getOrder() < RHS->getOrder();
976 SDDbgInfo::DbgIterator DI = DAG->DbgBegin();
977 SDDbgInfo::DbgIterator DE = DAG->DbgEnd();
978 // Now emit the rest according to source order.
979 unsigned LastOrder = 0;
980 for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) {
981 unsigned Order = Orders[i].first;
982 MachineInstr *MI = Orders[i].second;
983 // Insert all SDDbgValue's whose order(s) are before "Order".
984 assert(MI);
985 for (; DI != DE; ++DI) {
986 if ((*DI)->getOrder() < LastOrder || (*DI)->getOrder() >= Order)
987 break;
988 if ((*DI)->isEmitted())
989 continue;
991 MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap);
992 if (DbgMI) {
993 if (!LastOrder)
994 // Insert to start of the BB (after PHIs).
995 BB->insert(BBBegin, DbgMI);
996 else {
997 // Insert at the instruction, which may be in a different
998 // block, if the block was split by a custom inserter.
999 MachineBasicBlock::iterator Pos = MI;
1000 MI->getParent()->insert(Pos, DbgMI);
1004 LastOrder = Order;
1006 // Add trailing DbgValue's before the terminator. FIXME: May want to add
1007 // some of them before one or more conditional branches?
1008 SmallVector<MachineInstr*, 8> DbgMIs;
1009 for (; DI != DE; ++DI) {
1010 if ((*DI)->isEmitted())
1011 continue;
1012 assert((*DI)->getOrder() >= LastOrder &&
1013 "emitting DBG_VALUE out of order");
1014 if (MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap))
1015 DbgMIs.push_back(DbgMI);
1018 MachineBasicBlock *InsertBB = Emitter.getBlock();
1019 MachineBasicBlock::iterator Pos = InsertBB->getFirstTerminator();
1020 InsertBB->insert(Pos, DbgMIs.begin(), DbgMIs.end());
1022 SDDbgInfo::DbgLabelIterator DLI = DAG->DbgLabelBegin();
1023 SDDbgInfo::DbgLabelIterator DLE = DAG->DbgLabelEnd();
1024 // Now emit the rest according to source order.
1025 LastOrder = 0;
1026 for (const auto &InstrOrder : Orders) {
1027 unsigned Order = InstrOrder.first;
1028 MachineInstr *MI = InstrOrder.second;
1029 if (!MI)
1030 continue;
1032 // Insert all SDDbgLabel's whose order(s) are before "Order".
1033 for (; DLI != DLE &&
1034 (*DLI)->getOrder() >= LastOrder && (*DLI)->getOrder() < Order;
1035 ++DLI) {
1036 MachineInstr *DbgMI = Emitter.EmitDbgLabel(*DLI);
1037 if (DbgMI) {
1038 if (!LastOrder)
1039 // Insert to start of the BB (after PHIs).
1040 BB->insert(BBBegin, DbgMI);
1041 else {
1042 // Insert at the instruction, which may be in a different
1043 // block, if the block was split by a custom inserter.
1044 MachineBasicBlock::iterator Pos = MI;
1045 MI->getParent()->insert(Pos, DbgMI);
1049 if (DLI == DLE)
1050 break;
1052 LastOrder = Order;
1056 InsertPos = Emitter.getInsertPos();
1057 // In some cases, DBG_VALUEs might be inserted after the first terminator,
1058 // which results in an invalid MBB. If that happens, move the DBG_VALUEs
1059 // before the first terminator.
1060 MachineBasicBlock *InsertBB = Emitter.getBlock();
1061 auto FirstTerm = InsertBB->getFirstTerminator();
1062 if (FirstTerm != InsertBB->end()) {
1063 assert(!FirstTerm->isDebugValue() &&
1064 "first terminator cannot be a debug value");
1065 for (MachineInstr &MI : make_early_inc_range(
1066 make_range(std::next(FirstTerm), InsertBB->end()))) {
1067 // Only scan up to insertion point.
1068 if (&MI == InsertPos)
1069 break;
1071 if (!MI.isDebugValue())
1072 continue;
1074 // The DBG_VALUE was referencing a value produced by a terminator. By
1075 // moving the DBG_VALUE, the referenced value also needs invalidating.
1076 MI.getOperand(0).ChangeToRegister(0, false);
1077 MI.moveBefore(&*FirstTerm);
1080 return InsertBB;
1083 /// Return the basic block label.
1084 std::string ScheduleDAGSDNodes::getDAGName() const {
1085 return "sunit-dag." + BB->getFullName();