1 //===- lib/CodeGen/MachineTraceMetrics.cpp ----------------------*- C++ -*-===//
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 #include "llvm/CodeGen/MachineTraceMetrics.h"
11 #include "llvm/ADT/PostOrderIterator.h"
12 #include "llvm/ADT/SparseSet.h"
13 #include "llvm/CodeGen/MachineBasicBlock.h"
14 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
15 #include "llvm/CodeGen/MachineLoopInfo.h"
16 #include "llvm/CodeGen/MachineRegisterInfo.h"
17 #include "llvm/CodeGen/Passes.h"
18 #include "llvm/MC/MCSubtargetInfo.h"
19 #include "llvm/Support/Debug.h"
20 #include "llvm/Support/Format.h"
21 #include "llvm/Support/raw_ostream.h"
22 #include "llvm/Target/TargetInstrInfo.h"
23 #include "llvm/Target/TargetRegisterInfo.h"
24 #include "llvm/Target/TargetSubtargetInfo.h"
28 #define DEBUG_TYPE "machine-trace-metrics"
30 char MachineTraceMetrics::ID = 0;
31 char &llvm::MachineTraceMetricsID = MachineTraceMetrics::ID;
33 INITIALIZE_PASS_BEGIN(MachineTraceMetrics,
34 "machine-trace-metrics", "Machine Trace Metrics", false, true)
35 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
36 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
37 INITIALIZE_PASS_END(MachineTraceMetrics,
38 "machine-trace-metrics", "Machine Trace Metrics", false, true)
40 MachineTraceMetrics::MachineTraceMetrics()
41 : MachineFunctionPass(ID), MF(nullptr), TII(nullptr), TRI(nullptr),
42 MRI(nullptr), Loops(nullptr) {
43 std::fill(std::begin(Ensembles), std::end(Ensembles), nullptr);
46 void MachineTraceMetrics::getAnalysisUsage(AnalysisUsage &AU) const {
48 AU.addRequired<MachineBranchProbabilityInfo>();
49 AU.addRequired<MachineLoopInfo>();
50 MachineFunctionPass::getAnalysisUsage(AU);
53 bool MachineTraceMetrics::runOnMachineFunction(MachineFunction &Func) {
55 const TargetSubtargetInfo &ST = MF->getSubtarget();
56 TII = ST.getInstrInfo();
57 TRI = ST.getRegisterInfo();
58 MRI = &MF->getRegInfo();
59 Loops = &getAnalysis<MachineLoopInfo>();
60 SchedModel.init(ST.getSchedModel(), &ST, TII);
61 BlockInfo.resize(MF->getNumBlockIDs());
62 ProcResourceCycles.resize(MF->getNumBlockIDs() *
63 SchedModel.getNumProcResourceKinds());
67 void MachineTraceMetrics::releaseMemory() {
70 for (unsigned i = 0; i != TS_NumStrategies; ++i) {
72 Ensembles[i] = nullptr;
76 //===----------------------------------------------------------------------===//
77 // Fixed block information
78 //===----------------------------------------------------------------------===//
80 // The number of instructions in a basic block and the CPU resources used by
81 // those instructions don't depend on any given trace strategy.
83 /// Compute the resource usage in basic block MBB.
84 const MachineTraceMetrics::FixedBlockInfo*
85 MachineTraceMetrics::getResources(const MachineBasicBlock *MBB) {
86 assert(MBB && "No basic block");
87 FixedBlockInfo *FBI = &BlockInfo[MBB->getNumber()];
88 if (FBI->hasResources())
91 // Compute resource usage in the block.
92 FBI->HasCalls = false;
93 unsigned InstrCount = 0;
95 // Add up per-processor resource cycles as well.
96 unsigned PRKinds = SchedModel.getNumProcResourceKinds();
97 SmallVector<unsigned, 32> PRCycles(PRKinds);
99 for (const auto &MI : *MBB) {
100 if (MI.isTransient())
104 FBI->HasCalls = true;
106 // Count processor resources used.
107 if (!SchedModel.hasInstrSchedModel())
109 const MCSchedClassDesc *SC = SchedModel.resolveSchedClass(&MI);
113 for (TargetSchedModel::ProcResIter
114 PI = SchedModel.getWriteProcResBegin(SC),
115 PE = SchedModel.getWriteProcResEnd(SC); PI != PE; ++PI) {
116 assert(PI->ProcResourceIdx < PRKinds && "Bad processor resource kind");
117 PRCycles[PI->ProcResourceIdx] += PI->Cycles;
120 FBI->InstrCount = InstrCount;
122 // Scale the resource cycles so they are comparable.
123 unsigned PROffset = MBB->getNumber() * PRKinds;
124 for (unsigned K = 0; K != PRKinds; ++K)
125 ProcResourceCycles[PROffset + K] =
126 PRCycles[K] * SchedModel.getResourceFactor(K);
132 MachineTraceMetrics::getProcResourceCycles(unsigned MBBNum) const {
133 assert(BlockInfo[MBBNum].hasResources() &&
134 "getResources() must be called before getProcResourceCycles()");
135 unsigned PRKinds = SchedModel.getNumProcResourceKinds();
136 assert((MBBNum+1) * PRKinds <= ProcResourceCycles.size());
137 return makeArrayRef(ProcResourceCycles.data() + MBBNum * PRKinds, PRKinds);
141 //===----------------------------------------------------------------------===//
142 // Ensemble utility functions
143 //===----------------------------------------------------------------------===//
145 MachineTraceMetrics::Ensemble::Ensemble(MachineTraceMetrics *ct)
147 BlockInfo.resize(MTM.BlockInfo.size());
148 unsigned PRKinds = MTM.SchedModel.getNumProcResourceKinds();
149 ProcResourceDepths.resize(MTM.BlockInfo.size() * PRKinds);
150 ProcResourceHeights.resize(MTM.BlockInfo.size() * PRKinds);
153 // Virtual destructor serves as an anchor.
154 MachineTraceMetrics::Ensemble::~Ensemble() {}
157 MachineTraceMetrics::Ensemble::getLoopFor(const MachineBasicBlock *MBB) const {
158 return MTM.Loops->getLoopFor(MBB);
161 // Update resource-related information in the TraceBlockInfo for MBB.
162 // Only update resources related to the trace above MBB.
163 void MachineTraceMetrics::Ensemble::
164 computeDepthResources(const MachineBasicBlock *MBB) {
165 TraceBlockInfo *TBI = &BlockInfo[MBB->getNumber()];
166 unsigned PRKinds = MTM.SchedModel.getNumProcResourceKinds();
167 unsigned PROffset = MBB->getNumber() * PRKinds;
169 // Compute resources from trace above. The top block is simple.
172 TBI->Head = MBB->getNumber();
173 std::fill(ProcResourceDepths.begin() + PROffset,
174 ProcResourceDepths.begin() + PROffset + PRKinds, 0);
178 // Compute from the block above. A post-order traversal ensures the
179 // predecessor is always computed first.
180 unsigned PredNum = TBI->Pred->getNumber();
181 TraceBlockInfo *PredTBI = &BlockInfo[PredNum];
182 assert(PredTBI->hasValidDepth() && "Trace above has not been computed yet");
183 const FixedBlockInfo *PredFBI = MTM.getResources(TBI->Pred);
184 TBI->InstrDepth = PredTBI->InstrDepth + PredFBI->InstrCount;
185 TBI->Head = PredTBI->Head;
187 // Compute per-resource depths.
188 ArrayRef<unsigned> PredPRDepths = getProcResourceDepths(PredNum);
189 ArrayRef<unsigned> PredPRCycles = MTM.getProcResourceCycles(PredNum);
190 for (unsigned K = 0; K != PRKinds; ++K)
191 ProcResourceDepths[PROffset + K] = PredPRDepths[K] + PredPRCycles[K];
194 // Update resource-related information in the TraceBlockInfo for MBB.
195 // Only update resources related to the trace below MBB.
196 void MachineTraceMetrics::Ensemble::
197 computeHeightResources(const MachineBasicBlock *MBB) {
198 TraceBlockInfo *TBI = &BlockInfo[MBB->getNumber()];
199 unsigned PRKinds = MTM.SchedModel.getNumProcResourceKinds();
200 unsigned PROffset = MBB->getNumber() * PRKinds;
202 // Compute resources for the current block.
203 TBI->InstrHeight = MTM.getResources(MBB)->InstrCount;
204 ArrayRef<unsigned> PRCycles = MTM.getProcResourceCycles(MBB->getNumber());
206 // The trace tail is done.
208 TBI->Tail = MBB->getNumber();
209 std::copy(PRCycles.begin(), PRCycles.end(),
210 ProcResourceHeights.begin() + PROffset);
214 // Compute from the block below. A post-order traversal ensures the
215 // predecessor is always computed first.
216 unsigned SuccNum = TBI->Succ->getNumber();
217 TraceBlockInfo *SuccTBI = &BlockInfo[SuccNum];
218 assert(SuccTBI->hasValidHeight() && "Trace below has not been computed yet");
219 TBI->InstrHeight += SuccTBI->InstrHeight;
220 TBI->Tail = SuccTBI->Tail;
222 // Compute per-resource heights.
223 ArrayRef<unsigned> SuccPRHeights = getProcResourceHeights(SuccNum);
224 for (unsigned K = 0; K != PRKinds; ++K)
225 ProcResourceHeights[PROffset + K] = SuccPRHeights[K] + PRCycles[K];
228 // Check if depth resources for MBB are valid and return the TBI.
229 // Return NULL if the resources have been invalidated.
230 const MachineTraceMetrics::TraceBlockInfo*
231 MachineTraceMetrics::Ensemble::
232 getDepthResources(const MachineBasicBlock *MBB) const {
233 const TraceBlockInfo *TBI = &BlockInfo[MBB->getNumber()];
234 return TBI->hasValidDepth() ? TBI : nullptr;
237 // Check if height resources for MBB are valid and return the TBI.
238 // Return NULL if the resources have been invalidated.
239 const MachineTraceMetrics::TraceBlockInfo*
240 MachineTraceMetrics::Ensemble::
241 getHeightResources(const MachineBasicBlock *MBB) const {
242 const TraceBlockInfo *TBI = &BlockInfo[MBB->getNumber()];
243 return TBI->hasValidHeight() ? TBI : nullptr;
246 /// Get an array of processor resource depths for MBB. Indexed by processor
247 /// resource kind, this array contains the scaled processor resources consumed
248 /// by all blocks preceding MBB in its trace. It does not include instructions
251 /// Compare TraceBlockInfo::InstrDepth.
253 MachineTraceMetrics::Ensemble::
254 getProcResourceDepths(unsigned MBBNum) const {
255 unsigned PRKinds = MTM.SchedModel.getNumProcResourceKinds();
256 assert((MBBNum+1) * PRKinds <= ProcResourceDepths.size());
257 return makeArrayRef(ProcResourceDepths.data() + MBBNum * PRKinds, PRKinds);
260 /// Get an array of processor resource heights for MBB. Indexed by processor
261 /// resource kind, this array contains the scaled processor resources consumed
262 /// by this block and all blocks following it in its trace.
264 /// Compare TraceBlockInfo::InstrHeight.
266 MachineTraceMetrics::Ensemble::
267 getProcResourceHeights(unsigned MBBNum) const {
268 unsigned PRKinds = MTM.SchedModel.getNumProcResourceKinds();
269 assert((MBBNum+1) * PRKinds <= ProcResourceHeights.size());
270 return makeArrayRef(ProcResourceHeights.data() + MBBNum * PRKinds, PRKinds);
273 //===----------------------------------------------------------------------===//
274 // Trace Selection Strategies
275 //===----------------------------------------------------------------------===//
277 // A trace selection strategy is implemented as a sub-class of Ensemble. The
278 // trace through a block B is computed by two DFS traversals of the CFG
279 // starting from B. One upwards, and one downwards. During the upwards DFS,
280 // pickTracePred() is called on the post-ordered blocks. During the downwards
281 // DFS, pickTraceSucc() is called in a post-order.
284 // We never allow traces that leave loops, but we do allow traces to enter
285 // nested loops. We also never allow traces to contain back-edges.
287 // This means that a loop header can never appear above the center block of a
288 // trace, except as the trace head. Below the center block, loop exiting edges
291 // Return true if an edge from the From loop to the To loop is leaving a loop.
292 // Either of To and From can be null.
293 static bool isExitingLoop(const MachineLoop *From, const MachineLoop *To) {
294 return From && !From->contains(To);
297 // MinInstrCountEnsemble - Pick the trace that executes the least number of
300 class MinInstrCountEnsemble : public MachineTraceMetrics::Ensemble {
301 const char *getName() const override { return "MinInstr"; }
302 const MachineBasicBlock *pickTracePred(const MachineBasicBlock*) override;
303 const MachineBasicBlock *pickTraceSucc(const MachineBasicBlock*) override;
306 MinInstrCountEnsemble(MachineTraceMetrics *mtm)
307 : MachineTraceMetrics::Ensemble(mtm) {}
311 // Select the preferred predecessor for MBB.
312 const MachineBasicBlock*
313 MinInstrCountEnsemble::pickTracePred(const MachineBasicBlock *MBB) {
314 if (MBB->pred_empty())
316 const MachineLoop *CurLoop = getLoopFor(MBB);
317 // Don't leave loops, and never follow back-edges.
318 if (CurLoop && MBB == CurLoop->getHeader())
320 unsigned CurCount = MTM.getResources(MBB)->InstrCount;
321 const MachineBasicBlock *Best = nullptr;
322 unsigned BestDepth = 0;
323 for (MachineBasicBlock::const_pred_iterator
324 I = MBB->pred_begin(), E = MBB->pred_end(); I != E; ++I) {
325 const MachineBasicBlock *Pred = *I;
326 const MachineTraceMetrics::TraceBlockInfo *PredTBI =
327 getDepthResources(Pred);
328 // Ignore cycles that aren't natural loops.
331 // Pick the predecessor that would give this block the smallest InstrDepth.
332 unsigned Depth = PredTBI->InstrDepth + CurCount;
333 if (!Best || Depth < BestDepth)
334 Best = Pred, BestDepth = Depth;
339 // Select the preferred successor for MBB.
340 const MachineBasicBlock*
341 MinInstrCountEnsemble::pickTraceSucc(const MachineBasicBlock *MBB) {
342 if (MBB->pred_empty())
344 const MachineLoop *CurLoop = getLoopFor(MBB);
345 const MachineBasicBlock *Best = nullptr;
346 unsigned BestHeight = 0;
347 for (MachineBasicBlock::const_succ_iterator
348 I = MBB->succ_begin(), E = MBB->succ_end(); I != E; ++I) {
349 const MachineBasicBlock *Succ = *I;
350 // Don't consider back-edges.
351 if (CurLoop && Succ == CurLoop->getHeader())
353 // Don't consider successors exiting CurLoop.
354 if (isExitingLoop(CurLoop, getLoopFor(Succ)))
356 const MachineTraceMetrics::TraceBlockInfo *SuccTBI =
357 getHeightResources(Succ);
358 // Ignore cycles that aren't natural loops.
361 // Pick the successor that would give this block the smallest InstrHeight.
362 unsigned Height = SuccTBI->InstrHeight;
363 if (!Best || Height < BestHeight)
364 Best = Succ, BestHeight = Height;
369 // Get an Ensemble sub-class for the requested trace strategy.
370 MachineTraceMetrics::Ensemble *
371 MachineTraceMetrics::getEnsemble(MachineTraceMetrics::Strategy strategy) {
372 assert(strategy < TS_NumStrategies && "Invalid trace strategy enum");
373 Ensemble *&E = Ensembles[strategy];
377 // Allocate new Ensemble on demand.
379 case TS_MinInstrCount: return (E = new MinInstrCountEnsemble(this));
380 default: llvm_unreachable("Invalid trace strategy enum");
384 void MachineTraceMetrics::invalidate(const MachineBasicBlock *MBB) {
385 DEBUG(dbgs() << "Invalidate traces through BB#" << MBB->getNumber() << '\n');
386 BlockInfo[MBB->getNumber()].invalidate();
387 for (unsigned i = 0; i != TS_NumStrategies; ++i)
389 Ensembles[i]->invalidate(MBB);
392 void MachineTraceMetrics::verifyAnalysis() const {
396 assert(BlockInfo.size() == MF->getNumBlockIDs() && "Outdated BlockInfo size");
397 for (unsigned i = 0; i != TS_NumStrategies; ++i)
399 Ensembles[i]->verify();
403 //===----------------------------------------------------------------------===//
405 //===----------------------------------------------------------------------===//
407 // Traces are built by two CFG traversals. To avoid recomputing too much, use a
408 // set abstraction that confines the search to the current loop, and doesn't
413 MutableArrayRef<MachineTraceMetrics::TraceBlockInfo> Blocks;
414 SmallPtrSet<const MachineBasicBlock*, 8> Visited;
415 const MachineLoopInfo *Loops;
417 LoopBounds(MutableArrayRef<MachineTraceMetrics::TraceBlockInfo> blocks,
418 const MachineLoopInfo *loops)
419 : Blocks(blocks), Loops(loops), Downward(false) {}
423 // Specialize po_iterator_storage in order to prune the post-order traversal so
424 // it is limited to the current loop and doesn't traverse the loop back edges.
427 class po_iterator_storage<LoopBounds, true> {
430 po_iterator_storage(LoopBounds &lb) : LB(lb) {}
431 void finishPostorder(const MachineBasicBlock*) {}
433 bool insertEdge(const MachineBasicBlock *From, const MachineBasicBlock *To) {
434 // Skip already visited To blocks.
435 MachineTraceMetrics::TraceBlockInfo &TBI = LB.Blocks[To->getNumber()];
436 if (LB.Downward ? TBI.hasValidHeight() : TBI.hasValidDepth())
438 // From is null once when To is the trace center block.
440 if (const MachineLoop *FromLoop = LB.Loops->getLoopFor(From)) {
441 // Don't follow backedges, don't leave FromLoop when going upwards.
442 if ((LB.Downward ? To : From) == FromLoop->getHeader())
444 // Don't leave FromLoop.
445 if (isExitingLoop(FromLoop, LB.Loops->getLoopFor(To)))
449 // To is a new block. Mark the block as visited in case the CFG has cycles
450 // that MachineLoopInfo didn't recognize as a natural loop.
451 return LB.Visited.insert(To).second;
456 /// Compute the trace through MBB.
457 void MachineTraceMetrics::Ensemble::computeTrace(const MachineBasicBlock *MBB) {
458 DEBUG(dbgs() << "Computing " << getName() << " trace through BB#"
459 << MBB->getNumber() << '\n');
460 // Set up loop bounds for the backwards post-order traversal.
461 LoopBounds Bounds(BlockInfo, MTM.Loops);
463 // Run an upwards post-order search for the trace start.
464 Bounds.Downward = false;
465 Bounds.Visited.clear();
466 typedef ipo_ext_iterator<const MachineBasicBlock*, LoopBounds> UpwardPO;
467 for (UpwardPO I = ipo_ext_begin(MBB, Bounds), E = ipo_ext_end(MBB, Bounds);
469 DEBUG(dbgs() << " pred for BB#" << I->getNumber() << ": ");
470 TraceBlockInfo &TBI = BlockInfo[I->getNumber()];
471 // All the predecessors have been visited, pick the preferred one.
472 TBI.Pred = pickTracePred(*I);
475 dbgs() << "BB#" << TBI.Pred->getNumber() << '\n';
479 // The trace leading to I is now known, compute the depth resources.
480 computeDepthResources(*I);
483 // Run a downwards post-order search for the trace end.
484 Bounds.Downward = true;
485 Bounds.Visited.clear();
486 typedef po_ext_iterator<const MachineBasicBlock*, LoopBounds> DownwardPO;
487 for (DownwardPO I = po_ext_begin(MBB, Bounds), E = po_ext_end(MBB, Bounds);
489 DEBUG(dbgs() << " succ for BB#" << I->getNumber() << ": ");
490 TraceBlockInfo &TBI = BlockInfo[I->getNumber()];
491 // All the successors have been visited, pick the preferred one.
492 TBI.Succ = pickTraceSucc(*I);
495 dbgs() << "BB#" << TBI.Succ->getNumber() << '\n';
499 // The trace leaving I is now known, compute the height resources.
500 computeHeightResources(*I);
504 /// Invalidate traces through BadMBB.
506 MachineTraceMetrics::Ensemble::invalidate(const MachineBasicBlock *BadMBB) {
507 SmallVector<const MachineBasicBlock*, 16> WorkList;
508 TraceBlockInfo &BadTBI = BlockInfo[BadMBB->getNumber()];
510 // Invalidate height resources of blocks above MBB.
511 if (BadTBI.hasValidHeight()) {
512 BadTBI.invalidateHeight();
513 WorkList.push_back(BadMBB);
515 const MachineBasicBlock *MBB = WorkList.pop_back_val();
516 DEBUG(dbgs() << "Invalidate BB#" << MBB->getNumber() << ' ' << getName()
518 // Find any MBB predecessors that have MBB as their preferred successor.
519 // They are the only ones that need to be invalidated.
520 for (MachineBasicBlock::const_pred_iterator
521 I = MBB->pred_begin(), E = MBB->pred_end(); I != E; ++I) {
522 TraceBlockInfo &TBI = BlockInfo[(*I)->getNumber()];
523 if (!TBI.hasValidHeight())
525 if (TBI.Succ == MBB) {
526 TBI.invalidateHeight();
527 WorkList.push_back(*I);
530 // Verify that TBI.Succ is actually a *I successor.
531 assert((!TBI.Succ || (*I)->isSuccessor(TBI.Succ)) && "CFG changed");
533 } while (!WorkList.empty());
536 // Invalidate depth resources of blocks below MBB.
537 if (BadTBI.hasValidDepth()) {
538 BadTBI.invalidateDepth();
539 WorkList.push_back(BadMBB);
541 const MachineBasicBlock *MBB = WorkList.pop_back_val();
542 DEBUG(dbgs() << "Invalidate BB#" << MBB->getNumber() << ' ' << getName()
544 // Find any MBB successors that have MBB as their preferred predecessor.
545 // They are the only ones that need to be invalidated.
546 for (MachineBasicBlock::const_succ_iterator
547 I = MBB->succ_begin(), E = MBB->succ_end(); I != E; ++I) {
548 TraceBlockInfo &TBI = BlockInfo[(*I)->getNumber()];
549 if (!TBI.hasValidDepth())
551 if (TBI.Pred == MBB) {
552 TBI.invalidateDepth();
553 WorkList.push_back(*I);
556 // Verify that TBI.Pred is actually a *I predecessor.
557 assert((!TBI.Pred || (*I)->isPredecessor(TBI.Pred)) && "CFG changed");
559 } while (!WorkList.empty());
562 // Clear any per-instruction data. We only have to do this for BadMBB itself
563 // because the instructions in that block may change. Other blocks may be
564 // invalidated, but their instructions will stay the same, so there is no
565 // need to erase the Cycle entries. They will be overwritten when we
567 for (const auto &I : *BadMBB)
571 void MachineTraceMetrics::Ensemble::verify() const {
573 assert(BlockInfo.size() == MTM.MF->getNumBlockIDs() &&
574 "Outdated BlockInfo size");
575 for (unsigned Num = 0, e = BlockInfo.size(); Num != e; ++Num) {
576 const TraceBlockInfo &TBI = BlockInfo[Num];
577 if (TBI.hasValidDepth() && TBI.Pred) {
578 const MachineBasicBlock *MBB = MTM.MF->getBlockNumbered(Num);
579 assert(MBB->isPredecessor(TBI.Pred) && "CFG doesn't match trace");
580 assert(BlockInfo[TBI.Pred->getNumber()].hasValidDepth() &&
581 "Trace is broken, depth should have been invalidated.");
582 const MachineLoop *Loop = getLoopFor(MBB);
583 assert(!(Loop && MBB == Loop->getHeader()) && "Trace contains backedge");
585 if (TBI.hasValidHeight() && TBI.Succ) {
586 const MachineBasicBlock *MBB = MTM.MF->getBlockNumbered(Num);
587 assert(MBB->isSuccessor(TBI.Succ) && "CFG doesn't match trace");
588 assert(BlockInfo[TBI.Succ->getNumber()].hasValidHeight() &&
589 "Trace is broken, height should have been invalidated.");
590 const MachineLoop *Loop = getLoopFor(MBB);
591 const MachineLoop *SuccLoop = getLoopFor(TBI.Succ);
592 assert(!(Loop && Loop == SuccLoop && TBI.Succ == Loop->getHeader()) &&
593 "Trace contains backedge");
599 //===----------------------------------------------------------------------===//
601 //===----------------------------------------------------------------------===//
603 // Compute the depth and height of each instruction based on data dependencies
604 // and instruction latencies. These cycle numbers assume that the CPU can issue
605 // an infinite number of instructions per cycle as long as their dependencies
608 // A data dependency is represented as a defining MI and operand numbers on the
609 // defining and using MI.
612 const MachineInstr *DefMI;
616 DataDep(const MachineInstr *DefMI, unsigned DefOp, unsigned UseOp)
617 : DefMI(DefMI), DefOp(DefOp), UseOp(UseOp) {}
619 /// Create a DataDep from an SSA form virtual register.
620 DataDep(const MachineRegisterInfo *MRI, unsigned VirtReg, unsigned UseOp)
622 assert(TargetRegisterInfo::isVirtualRegister(VirtReg));
623 MachineRegisterInfo::def_iterator DefI = MRI->def_begin(VirtReg);
624 assert(!DefI.atEnd() && "Register has no defs");
625 DefMI = DefI->getParent();
626 DefOp = DefI.getOperandNo();
627 assert((++DefI).atEnd() && "Register has multiple defs");
632 // Get the input data dependencies that must be ready before UseMI can issue.
633 // Return true if UseMI has any physreg operands.
634 static bool getDataDeps(const MachineInstr *UseMI,
635 SmallVectorImpl<DataDep> &Deps,
636 const MachineRegisterInfo *MRI) {
637 bool HasPhysRegs = false;
638 for (ConstMIOperands MO(UseMI); MO.isValid(); ++MO) {
641 unsigned Reg = MO->getReg();
644 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
648 // Collect virtual register reads.
650 Deps.push_back(DataDep(MRI, Reg, MO.getOperandNo()));
655 // Get the input data dependencies of a PHI instruction, using Pred as the
656 // preferred predecessor.
657 // This will add at most one dependency to Deps.
658 static void getPHIDeps(const MachineInstr *UseMI,
659 SmallVectorImpl<DataDep> &Deps,
660 const MachineBasicBlock *Pred,
661 const MachineRegisterInfo *MRI) {
662 // No predecessor at the beginning of a trace. Ignore dependencies.
665 assert(UseMI->isPHI() && UseMI->getNumOperands() % 2 && "Bad PHI");
666 for (unsigned i = 1; i != UseMI->getNumOperands(); i += 2) {
667 if (UseMI->getOperand(i + 1).getMBB() == Pred) {
668 unsigned Reg = UseMI->getOperand(i).getReg();
669 Deps.push_back(DataDep(MRI, Reg, i));
675 // Keep track of physreg data dependencies by recording each live register unit.
676 // Associate each regunit with an instruction operand. Depending on the
677 // direction instructions are scanned, it could be the operand that defined the
678 // regunit, or the highest operand to read the regunit.
683 const MachineInstr *MI;
686 unsigned getSparseSetIndex() const { return RegUnit; }
688 LiveRegUnit(unsigned RU) : RegUnit(RU), Cycle(0), MI(nullptr), Op(0) {}
692 // Identify physreg dependencies for UseMI, and update the live regunit
693 // tracking set when scanning instructions downwards.
694 static void updatePhysDepsDownwards(const MachineInstr *UseMI,
695 SmallVectorImpl<DataDep> &Deps,
696 SparseSet<LiveRegUnit> &RegUnits,
697 const TargetRegisterInfo *TRI) {
698 SmallVector<unsigned, 8> Kills;
699 SmallVector<unsigned, 8> LiveDefOps;
701 for (ConstMIOperands MO(UseMI); MO.isValid(); ++MO) {
704 unsigned Reg = MO->getReg();
705 if (!TargetRegisterInfo::isPhysicalRegister(Reg))
707 // Track live defs and kills for updating RegUnits.
710 Kills.push_back(Reg);
712 LiveDefOps.push_back(MO.getOperandNo());
713 } else if (MO->isKill())
714 Kills.push_back(Reg);
715 // Identify dependencies.
718 for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units) {
719 SparseSet<LiveRegUnit>::iterator I = RegUnits.find(*Units);
720 if (I == RegUnits.end())
722 Deps.push_back(DataDep(I->MI, I->Op, MO.getOperandNo()));
727 // Update RegUnits to reflect live registers after UseMI.
729 for (unsigned i = 0, e = Kills.size(); i != e; ++i)
730 for (MCRegUnitIterator Units(Kills[i], TRI); Units.isValid(); ++Units)
731 RegUnits.erase(*Units);
733 // Second, live defs.
734 for (unsigned i = 0, e = LiveDefOps.size(); i != e; ++i) {
735 unsigned DefOp = LiveDefOps[i];
736 for (MCRegUnitIterator Units(UseMI->getOperand(DefOp).getReg(), TRI);
737 Units.isValid(); ++Units) {
738 LiveRegUnit &LRU = RegUnits[*Units];
745 /// The length of the critical path through a trace is the maximum of two path
748 /// 1. The maximum height+depth over all instructions in the trace center block.
750 /// 2. The longest cross-block dependency chain. For small blocks, it is
751 /// possible that the critical path through the trace doesn't include any
752 /// instructions in the block.
754 /// This function computes the second number from the live-in list of the
756 unsigned MachineTraceMetrics::Ensemble::
757 computeCrossBlockCriticalPath(const TraceBlockInfo &TBI) {
758 assert(TBI.HasValidInstrDepths && "Missing depth info");
759 assert(TBI.HasValidInstrHeights && "Missing height info");
761 for (unsigned i = 0, e = TBI.LiveIns.size(); i != e; ++i) {
762 const LiveInReg &LIR = TBI.LiveIns[i];
763 if (!TargetRegisterInfo::isVirtualRegister(LIR.Reg))
765 const MachineInstr *DefMI = MTM.MRI->getVRegDef(LIR.Reg);
766 // Ignore dependencies outside the current trace.
767 const TraceBlockInfo &DefTBI = BlockInfo[DefMI->getParent()->getNumber()];
768 if (!DefTBI.isUsefulDominator(TBI))
770 unsigned Len = LIR.Height + Cycles[DefMI].Depth;
771 MaxLen = std::max(MaxLen, Len);
776 /// Compute instruction depths for all instructions above or in MBB in its
777 /// trace. This assumes that the trace through MBB has already been computed.
778 void MachineTraceMetrics::Ensemble::
779 computeInstrDepths(const MachineBasicBlock *MBB) {
780 // The top of the trace may already be computed, and HasValidInstrDepths
781 // implies Head->HasValidInstrDepths, so we only need to start from the first
782 // block in the trace that needs to be recomputed.
783 SmallVector<const MachineBasicBlock*, 8> Stack;
785 TraceBlockInfo &TBI = BlockInfo[MBB->getNumber()];
786 assert(TBI.hasValidDepth() && "Incomplete trace");
787 if (TBI.HasValidInstrDepths)
789 Stack.push_back(MBB);
793 // FIXME: If MBB is non-null at this point, it is the last pre-computed block
794 // in the trace. We should track any live-out physregs that were defined in
795 // the trace. This is quite rare in SSA form, typically created by CSE
796 // hoisting a compare.
797 SparseSet<LiveRegUnit> RegUnits;
798 RegUnits.setUniverse(MTM.TRI->getNumRegUnits());
800 // Go through trace blocks in top-down order, stopping after the center block.
801 SmallVector<DataDep, 8> Deps;
802 while (!Stack.empty()) {
803 MBB = Stack.pop_back_val();
804 DEBUG(dbgs() << "\nDepths for BB#" << MBB->getNumber() << ":\n");
805 TraceBlockInfo &TBI = BlockInfo[MBB->getNumber()];
806 TBI.HasValidInstrDepths = true;
807 TBI.CriticalPath = 0;
809 // Print out resource depths here as well.
811 dbgs() << format("%7u Instructions\n", TBI.InstrDepth);
812 ArrayRef<unsigned> PRDepths = getProcResourceDepths(MBB->getNumber());
813 for (unsigned K = 0; K != PRDepths.size(); ++K)
815 unsigned Factor = MTM.SchedModel.getResourceFactor(K);
816 dbgs() << format("%6uc @ ", MTM.getCycles(PRDepths[K]))
817 << MTM.SchedModel.getProcResource(K)->Name << " ("
818 << PRDepths[K]/Factor << " ops x" << Factor << ")\n";
822 // Also compute the critical path length through MBB when possible.
823 if (TBI.HasValidInstrHeights)
824 TBI.CriticalPath = computeCrossBlockCriticalPath(TBI);
826 for (const auto &UseMI : *MBB) {
827 // Collect all data dependencies.
830 getPHIDeps(&UseMI, Deps, TBI.Pred, MTM.MRI);
831 else if (getDataDeps(&UseMI, Deps, MTM.MRI))
832 updatePhysDepsDownwards(&UseMI, Deps, RegUnits, MTM.TRI);
834 // Filter and process dependencies, computing the earliest issue cycle.
836 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
837 const DataDep &Dep = Deps[i];
838 const TraceBlockInfo&DepTBI =
839 BlockInfo[Dep.DefMI->getParent()->getNumber()];
840 // Ignore dependencies from outside the current trace.
841 if (!DepTBI.isUsefulDominator(TBI))
843 assert(DepTBI.HasValidInstrDepths && "Inconsistent dependency");
844 unsigned DepCycle = Cycles.lookup(Dep.DefMI).Depth;
845 // Add latency if DefMI is a real instruction. Transients get latency 0.
846 if (!Dep.DefMI->isTransient())
847 DepCycle += MTM.SchedModel
848 .computeOperandLatency(Dep.DefMI, Dep.DefOp, &UseMI, Dep.UseOp);
849 Cycle = std::max(Cycle, DepCycle);
851 // Remember the instruction depth.
852 InstrCycles &MICycles = Cycles[&UseMI];
853 MICycles.Depth = Cycle;
855 if (!TBI.HasValidInstrHeights) {
856 DEBUG(dbgs() << Cycle << '\t' << UseMI);
859 // Update critical path length.
860 TBI.CriticalPath = std::max(TBI.CriticalPath, Cycle + MICycles.Height);
861 DEBUG(dbgs() << TBI.CriticalPath << '\t' << Cycle << '\t' << UseMI);
866 // Identify physreg dependencies for MI when scanning instructions upwards.
867 // Return the issue height of MI after considering any live regunits.
868 // Height is the issue height computed from virtual register dependencies alone.
869 static unsigned updatePhysDepsUpwards(const MachineInstr *MI, unsigned Height,
870 SparseSet<LiveRegUnit> &RegUnits,
871 const TargetSchedModel &SchedModel,
872 const TargetInstrInfo *TII,
873 const TargetRegisterInfo *TRI) {
874 SmallVector<unsigned, 8> ReadOps;
875 for (ConstMIOperands MO(MI); MO.isValid(); ++MO) {
878 unsigned Reg = MO->getReg();
879 if (!TargetRegisterInfo::isPhysicalRegister(Reg))
882 ReadOps.push_back(MO.getOperandNo());
885 // This is a def of Reg. Remove corresponding entries from RegUnits, and
886 // update MI Height to consider the physreg dependencies.
887 for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units) {
888 SparseSet<LiveRegUnit>::iterator I = RegUnits.find(*Units);
889 if (I == RegUnits.end())
891 unsigned DepHeight = I->Cycle;
892 if (!MI->isTransient()) {
893 // We may not know the UseMI of this dependency, if it came from the
894 // live-in list. SchedModel can handle a NULL UseMI.
895 DepHeight += SchedModel
896 .computeOperandLatency(MI, MO.getOperandNo(), I->MI, I->Op);
898 Height = std::max(Height, DepHeight);
899 // This regunit is dead above MI.
904 // Now we know the height of MI. Update any regunits read.
905 for (unsigned i = 0, e = ReadOps.size(); i != e; ++i) {
906 unsigned Reg = MI->getOperand(ReadOps[i]).getReg();
907 for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units) {
908 LiveRegUnit &LRU = RegUnits[*Units];
909 // Set the height to the highest reader of the unit.
910 if (LRU.Cycle <= Height && LRU.MI != MI) {
922 typedef DenseMap<const MachineInstr *, unsigned> MIHeightMap;
924 // Push the height of DefMI upwards if required to match UseMI.
925 // Return true if this is the first time DefMI was seen.
926 static bool pushDepHeight(const DataDep &Dep,
927 const MachineInstr *UseMI, unsigned UseHeight,
928 MIHeightMap &Heights,
929 const TargetSchedModel &SchedModel,
930 const TargetInstrInfo *TII) {
931 // Adjust height by Dep.DefMI latency.
932 if (!Dep.DefMI->isTransient())
933 UseHeight += SchedModel.computeOperandLatency(Dep.DefMI, Dep.DefOp,
936 // Update Heights[DefMI] to be the maximum height seen.
937 MIHeightMap::iterator I;
939 std::tie(I, New) = Heights.insert(std::make_pair(Dep.DefMI, UseHeight));
943 // DefMI has been pushed before. Give it the max height.
944 if (I->second < UseHeight)
945 I->second = UseHeight;
949 /// Assuming that the virtual register defined by DefMI:DefOp was used by
950 /// Trace.back(), add it to the live-in lists of all the blocks in Trace. Stop
951 /// when reaching the block that contains DefMI.
952 void MachineTraceMetrics::Ensemble::
953 addLiveIns(const MachineInstr *DefMI, unsigned DefOp,
954 ArrayRef<const MachineBasicBlock*> Trace) {
955 assert(!Trace.empty() && "Trace should contain at least one block");
956 unsigned Reg = DefMI->getOperand(DefOp).getReg();
957 assert(TargetRegisterInfo::isVirtualRegister(Reg));
958 const MachineBasicBlock *DefMBB = DefMI->getParent();
960 // Reg is live-in to all blocks in Trace that follow DefMBB.
961 for (unsigned i = Trace.size(); i; --i) {
962 const MachineBasicBlock *MBB = Trace[i-1];
965 TraceBlockInfo &TBI = BlockInfo[MBB->getNumber()];
966 // Just add the register. The height will be updated later.
967 TBI.LiveIns.push_back(Reg);
971 /// Compute instruction heights in the trace through MBB. This updates MBB and
972 /// the blocks below it in the trace. It is assumed that the trace has already
974 void MachineTraceMetrics::Ensemble::
975 computeInstrHeights(const MachineBasicBlock *MBB) {
976 // The bottom of the trace may already be computed.
977 // Find the blocks that need updating.
978 SmallVector<const MachineBasicBlock*, 8> Stack;
980 TraceBlockInfo &TBI = BlockInfo[MBB->getNumber()];
981 assert(TBI.hasValidHeight() && "Incomplete trace");
982 if (TBI.HasValidInstrHeights)
984 Stack.push_back(MBB);
989 // As we move upwards in the trace, keep track of instructions that are
990 // required by deeper trace instructions. Map MI -> height required so far.
993 // For physregs, the def isn't known when we see the use.
994 // Instead, keep track of the highest use of each regunit.
995 SparseSet<LiveRegUnit> RegUnits;
996 RegUnits.setUniverse(MTM.TRI->getNumRegUnits());
998 // If the bottom of the trace was already precomputed, initialize heights
999 // from its live-in list.
1000 // MBB is the highest precomputed block in the trace.
1002 TraceBlockInfo &TBI = BlockInfo[MBB->getNumber()];
1003 for (unsigned i = 0, e = TBI.LiveIns.size(); i != e; ++i) {
1004 LiveInReg LI = TBI.LiveIns[i];
1005 if (TargetRegisterInfo::isVirtualRegister(LI.Reg)) {
1006 // For virtual registers, the def latency is included.
1007 unsigned &Height = Heights[MTM.MRI->getVRegDef(LI.Reg)];
1008 if (Height < LI.Height)
1011 // For register units, the def latency is not included because we don't
1012 // know the def yet.
1013 RegUnits[LI.Reg].Cycle = LI.Height;
1018 // Go through the trace blocks in bottom-up order.
1019 SmallVector<DataDep, 8> Deps;
1020 for (;!Stack.empty(); Stack.pop_back()) {
1022 DEBUG(dbgs() << "Heights for BB#" << MBB->getNumber() << ":\n");
1023 TraceBlockInfo &TBI = BlockInfo[MBB->getNumber()];
1024 TBI.HasValidInstrHeights = true;
1025 TBI.CriticalPath = 0;
1028 dbgs() << format("%7u Instructions\n", TBI.InstrHeight);
1029 ArrayRef<unsigned> PRHeights = getProcResourceHeights(MBB->getNumber());
1030 for (unsigned K = 0; K != PRHeights.size(); ++K)
1032 unsigned Factor = MTM.SchedModel.getResourceFactor(K);
1033 dbgs() << format("%6uc @ ", MTM.getCycles(PRHeights[K]))
1034 << MTM.SchedModel.getProcResource(K)->Name << " ("
1035 << PRHeights[K]/Factor << " ops x" << Factor << ")\n";
1039 // Get dependencies from PHIs in the trace successor.
1040 const MachineBasicBlock *Succ = TBI.Succ;
1041 // If MBB is the last block in the trace, and it has a back-edge to the
1042 // loop header, get loop-carried dependencies from PHIs in the header. For
1043 // that purpose, pretend that all the loop header PHIs have height 0.
1045 if (const MachineLoop *Loop = getLoopFor(MBB))
1046 if (MBB->isSuccessor(Loop->getHeader()))
1047 Succ = Loop->getHeader();
1050 for (const auto &PHI : *Succ) {
1054 getPHIDeps(&PHI, Deps, MBB, MTM.MRI);
1055 if (!Deps.empty()) {
1056 // Loop header PHI heights are all 0.
1057 unsigned Height = TBI.Succ ? Cycles.lookup(&PHI).Height : 0;
1058 DEBUG(dbgs() << "pred\t" << Height << '\t' << PHI);
1059 if (pushDepHeight(Deps.front(), &PHI, Height,
1060 Heights, MTM.SchedModel, MTM.TII))
1061 addLiveIns(Deps.front().DefMI, Deps.front().DefOp, Stack);
1066 // Go through the block backwards.
1067 for (MachineBasicBlock::const_iterator BI = MBB->end(), BB = MBB->begin();
1069 const MachineInstr *MI = --BI;
1071 // Find the MI height as determined by virtual register uses in the
1074 MIHeightMap::iterator HeightI = Heights.find(MI);
1075 if (HeightI != Heights.end()) {
1076 Cycle = HeightI->second;
1077 // We won't be seeing any more MI uses.
1078 Heights.erase(HeightI);
1081 // Don't process PHI deps. They depend on the specific predecessor, and
1082 // we'll get them when visiting the predecessor.
1084 bool HasPhysRegs = !MI->isPHI() && getDataDeps(MI, Deps, MTM.MRI);
1086 // There may also be regunit dependencies to include in the height.
1088 Cycle = updatePhysDepsUpwards(MI, Cycle, RegUnits,
1089 MTM.SchedModel, MTM.TII, MTM.TRI);
1091 // Update the required height of any virtual registers read by MI.
1092 for (unsigned i = 0, e = Deps.size(); i != e; ++i)
1093 if (pushDepHeight(Deps[i], MI, Cycle, Heights, MTM.SchedModel, MTM.TII))
1094 addLiveIns(Deps[i].DefMI, Deps[i].DefOp, Stack);
1096 InstrCycles &MICycles = Cycles[MI];
1097 MICycles.Height = Cycle;
1098 if (!TBI.HasValidInstrDepths) {
1099 DEBUG(dbgs() << Cycle << '\t' << *MI);
1102 // Update critical path length.
1103 TBI.CriticalPath = std::max(TBI.CriticalPath, Cycle + MICycles.Depth);
1104 DEBUG(dbgs() << TBI.CriticalPath << '\t' << Cycle << '\t' << *MI);
1107 // Update virtual live-in heights. They were added by addLiveIns() with a 0
1108 // height because the final height isn't known until now.
1109 DEBUG(dbgs() << "BB#" << MBB->getNumber() << " Live-ins:");
1110 for (unsigned i = 0, e = TBI.LiveIns.size(); i != e; ++i) {
1111 LiveInReg &LIR = TBI.LiveIns[i];
1112 const MachineInstr *DefMI = MTM.MRI->getVRegDef(LIR.Reg);
1113 LIR.Height = Heights.lookup(DefMI);
1114 DEBUG(dbgs() << ' ' << PrintReg(LIR.Reg) << '@' << LIR.Height);
1117 // Transfer the live regunits to the live-in list.
1118 for (SparseSet<LiveRegUnit>::const_iterator
1119 RI = RegUnits.begin(), RE = RegUnits.end(); RI != RE; ++RI) {
1120 TBI.LiveIns.push_back(LiveInReg(RI->RegUnit, RI->Cycle));
1121 DEBUG(dbgs() << ' ' << PrintRegUnit(RI->RegUnit, MTM.TRI)
1122 << '@' << RI->Cycle);
1124 DEBUG(dbgs() << '\n');
1126 if (!TBI.HasValidInstrDepths)
1128 // Add live-ins to the critical path length.
1129 TBI.CriticalPath = std::max(TBI.CriticalPath,
1130 computeCrossBlockCriticalPath(TBI));
1131 DEBUG(dbgs() << "Critical path: " << TBI.CriticalPath << '\n');
1135 MachineTraceMetrics::Trace
1136 MachineTraceMetrics::Ensemble::getTrace(const MachineBasicBlock *MBB) {
1137 // FIXME: Check cache tags, recompute as needed.
1139 computeInstrDepths(MBB);
1140 computeInstrHeights(MBB);
1141 return Trace(*this, BlockInfo[MBB->getNumber()]);
1145 MachineTraceMetrics::Trace::getInstrSlack(const MachineInstr *MI) const {
1146 assert(MI && "Not an instruction.");
1147 assert(getBlockNum() == unsigned(MI->getParent()->getNumber()) &&
1148 "MI must be in the trace center block");
1149 InstrCycles Cyc = getInstrCycles(MI);
1150 return getCriticalPath() - (Cyc.Depth + Cyc.Height);
1154 MachineTraceMetrics::Trace::getPHIDepth(const MachineInstr *PHI) const {
1155 const MachineBasicBlock *MBB = TE.MTM.MF->getBlockNumbered(getBlockNum());
1156 SmallVector<DataDep, 1> Deps;
1157 getPHIDeps(PHI, Deps, MBB, TE.MTM.MRI);
1158 assert(Deps.size() == 1 && "PHI doesn't have MBB as a predecessor");
1159 DataDep &Dep = Deps.front();
1160 unsigned DepCycle = getInstrCycles(Dep.DefMI).Depth;
1161 // Add latency if DefMI is a real instruction. Transients get latency 0.
1162 if (!Dep.DefMI->isTransient())
1163 DepCycle += TE.MTM.SchedModel
1164 .computeOperandLatency(Dep.DefMI, Dep.DefOp, PHI, Dep.UseOp);
1168 /// When bottom is set include instructions in current block in estimate.
1169 unsigned MachineTraceMetrics::Trace::getResourceDepth(bool Bottom) const {
1170 // Find the limiting processor resource.
1171 // Numbers have been pre-scaled to be comparable.
1173 ArrayRef<unsigned> PRDepths = TE.getProcResourceDepths(getBlockNum());
1175 ArrayRef<unsigned> PRCycles = TE.MTM.getProcResourceCycles(getBlockNum());
1176 for (unsigned K = 0; K != PRDepths.size(); ++K)
1177 PRMax = std::max(PRMax, PRDepths[K] + PRCycles[K]);
1179 for (unsigned K = 0; K != PRDepths.size(); ++K)
1180 PRMax = std::max(PRMax, PRDepths[K]);
1182 // Convert to cycle count.
1183 PRMax = TE.MTM.getCycles(PRMax);
1185 /// All instructions before current block
1186 unsigned Instrs = TBI.InstrDepth;
1187 // plus instructions in current block
1189 Instrs += TE.MTM.BlockInfo[getBlockNum()].InstrCount;
1190 if (unsigned IW = TE.MTM.SchedModel.getIssueWidth())
1192 // Assume issue width 1 without a schedule model.
1193 return std::max(Instrs, PRMax);
1196 unsigned MachineTraceMetrics::Trace::getResourceLength(
1197 ArrayRef<const MachineBasicBlock *> Extrablocks,
1198 ArrayRef<const MCSchedClassDesc *> ExtraInstrs,
1199 ArrayRef<const MCSchedClassDesc *> RemoveInstrs) const {
1200 // Add up resources above and below the center block.
1201 ArrayRef<unsigned> PRDepths = TE.getProcResourceDepths(getBlockNum());
1202 ArrayRef<unsigned> PRHeights = TE.getProcResourceHeights(getBlockNum());
1205 // Capture computing cycles from extra instructions
1206 auto extraCycles = [this](ArrayRef<const MCSchedClassDesc *> Instrs,
1207 unsigned ResourceIdx)
1209 unsigned Cycles = 0;
1210 for (unsigned I = 0; I != Instrs.size(); ++I) {
1211 const MCSchedClassDesc *SC = Instrs[I];
1214 for (TargetSchedModel::ProcResIter
1215 PI = TE.MTM.SchedModel.getWriteProcResBegin(SC),
1216 PE = TE.MTM.SchedModel.getWriteProcResEnd(SC);
1218 if (PI->ProcResourceIdx != ResourceIdx)
1221 (PI->Cycles * TE.MTM.SchedModel.getResourceFactor(ResourceIdx));
1227 for (unsigned K = 0; K != PRDepths.size(); ++K) {
1228 unsigned PRCycles = PRDepths[K] + PRHeights[K];
1229 for (unsigned I = 0; I != Extrablocks.size(); ++I)
1230 PRCycles += TE.MTM.getProcResourceCycles(Extrablocks[I]->getNumber())[K];
1231 PRCycles += extraCycles(ExtraInstrs, K);
1232 PRCycles -= extraCycles(RemoveInstrs, K);
1233 PRMax = std::max(PRMax, PRCycles);
1235 // Convert to cycle count.
1236 PRMax = TE.MTM.getCycles(PRMax);
1238 // Instrs: #instructions in current trace outside current block.
1239 unsigned Instrs = TBI.InstrDepth + TBI.InstrHeight;
1240 // Add instruction count from the extra blocks.
1241 for (unsigned i = 0, e = Extrablocks.size(); i != e; ++i)
1242 Instrs += TE.MTM.getResources(Extrablocks[i])->InstrCount;
1243 Instrs += ExtraInstrs.size();
1244 Instrs -= RemoveInstrs.size();
1245 if (unsigned IW = TE.MTM.SchedModel.getIssueWidth())
1247 // Assume issue width 1 without a schedule model.
1248 return std::max(Instrs, PRMax);
1251 bool MachineTraceMetrics::Trace::isDepInTrace(const MachineInstr *DefMI,
1252 const MachineInstr *UseMI) const {
1253 if (DefMI->getParent() == UseMI->getParent())
1256 const TraceBlockInfo &DepTBI = TE.BlockInfo[DefMI->getParent()->getNumber()];
1257 const TraceBlockInfo &TBI = TE.BlockInfo[UseMI->getParent()->getNumber()];
1259 return DepTBI.isUsefulDominator(TBI);
1262 void MachineTraceMetrics::Ensemble::print(raw_ostream &OS) const {
1263 OS << getName() << " ensemble:\n";
1264 for (unsigned i = 0, e = BlockInfo.size(); i != e; ++i) {
1265 OS << " BB#" << i << '\t';
1266 BlockInfo[i].print(OS);
1271 void MachineTraceMetrics::TraceBlockInfo::print(raw_ostream &OS) const {
1272 if (hasValidDepth()) {
1273 OS << "depth=" << InstrDepth;
1275 OS << " pred=BB#" << Pred->getNumber();
1278 OS << " head=BB#" << Head;
1279 if (HasValidInstrDepths)
1282 OS << "depth invalid";
1284 if (hasValidHeight()) {
1285 OS << "height=" << InstrHeight;
1287 OS << " succ=BB#" << Succ->getNumber();
1290 OS << " tail=BB#" << Tail;
1291 if (HasValidInstrHeights)
1294 OS << "height invalid";
1295 if (HasValidInstrDepths && HasValidInstrHeights)
1296 OS << ", crit=" << CriticalPath;
1299 void MachineTraceMetrics::Trace::print(raw_ostream &OS) const {
1300 unsigned MBBNum = &TBI - &TE.BlockInfo[0];
1302 OS << TE.getName() << " trace BB#" << TBI.Head << " --> BB#" << MBBNum
1303 << " --> BB#" << TBI.Tail << ':';
1304 if (TBI.hasValidHeight() && TBI.hasValidDepth())
1305 OS << ' ' << getInstrCount() << " instrs.";
1306 if (TBI.HasValidInstrDepths && TBI.HasValidInstrHeights)
1307 OS << ' ' << TBI.CriticalPath << " cycles.";
1309 const MachineTraceMetrics::TraceBlockInfo *Block = &TBI;
1310 OS << "\nBB#" << MBBNum;
1311 while (Block->hasValidDepth() && Block->Pred) {
1312 unsigned Num = Block->Pred->getNumber();
1313 OS << " <- BB#" << Num;
1314 Block = &TE.BlockInfo[Num];
1319 while (Block->hasValidHeight() && Block->Succ) {
1320 unsigned Num = Block->Succ->getNumber();
1321 OS << " -> BB#" << Num;
1322 Block = &TE.BlockInfo[Num];