1//===---- ScheduleDAGInstrs.cpp - MachineInstr Rescheduling ---------------===//
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/// \file This implements the ScheduleDAGInstrs class, which implements
10/// re-scheduling of MachineInstrs.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/CodeGen/ScheduleDAGInstrs.h"
15
16#include "llvm/ADT/IntEqClasses.h"
17#include "llvm/ADT/MapVector.h"
18#include "llvm/ADT/SmallVector.h"
19#include "llvm/ADT/SparseSet.h"
20#include "llvm/ADT/iterator_range.h"
21#include "llvm/Analysis/AliasAnalysis.h"
22#include "llvm/Analysis/ValueTracking.h"
23#include "llvm/CodeGen/LiveIntervals.h"
24#include "llvm/CodeGen/LivePhysRegs.h"
25#include "llvm/CodeGen/MachineBasicBlock.h"
26#include "llvm/CodeGen/MachineFrameInfo.h"
27#include "llvm/CodeGen/MachineFunction.h"
28#include "llvm/CodeGen/MachineInstr.h"
29#include "llvm/CodeGen/MachineInstrBundle.h"
30#include "llvm/CodeGen/MachineMemOperand.h"
31#include "llvm/CodeGen/MachineOperand.h"
32#include "llvm/CodeGen/MachineRegisterInfo.h"
33#include "llvm/CodeGen/PseudoSourceValue.h"
34#include "llvm/CodeGen/RegisterPressure.h"
35#include "llvm/CodeGen/ScheduleDAG.h"
36#include "llvm/CodeGen/ScheduleDFS.h"
37#include "llvm/CodeGen/SlotIndexes.h"
38#include "llvm/CodeGen/TargetRegisterInfo.h"
39#include "llvm/CodeGen/TargetSubtargetInfo.h"
40#include "llvm/Config/llvm-config.h"
41#include "llvm/IR/Constants.h"
42#include "llvm/IR/Function.h"
43#include "llvm/IR/Type.h"
44#include "llvm/IR/Value.h"
45#include "llvm/MC/LaneBitmask.h"
46#include "llvm/MC/MCRegisterInfo.h"
47#include "llvm/Support/Casting.h"
48#include "llvm/Support/CommandLine.h"
49#include "llvm/Support/Compiler.h"
50#include "llvm/Support/Debug.h"
51#include "llvm/Support/ErrorHandling.h"
52#include "llvm/Support/Format.h"
53#include "llvm/Support/raw_ostream.h"
54#include <algorithm>
55#include <cassert>
56#include <iterator>
57#include <utility>
58#include <vector>
59
60using namespace llvm;
61
62#define DEBUG_TYPE "machine-scheduler"
63
64static cl::opt<bool>
65 EnableAASchedMI("enable-aa-sched-mi", cl::Hidden,
66 cl::desc("Enable use of AA during MI DAG construction"));
67
68static cl::opt<bool> UseTBAA("use-tbaa-in-sched-mi", cl::Hidden,
69 cl::init(Val: true), cl::desc("Enable use of TBAA during MI DAG construction"));
70
71// Note: the two options below might be used in tuning compile time vs
72// output quality. Setting HugeRegion so large that it will never be
73// reached means best-effort, but may be slow.
74
75// When Stores and Loads maps (or NonAliasStores and NonAliasLoads)
76// together hold this many SUs, a reduction of maps will be done.
77static cl::opt<unsigned> HugeRegion("dag-maps-huge-region", cl::Hidden,
78 cl::init(Val: 1000), cl::desc("The limit to use while constructing the DAG "
79 "prior to scheduling, at which point a trade-off "
80 "is made to avoid excessive compile time."));
81
82static cl::opt<unsigned> ReductionSize(
83 "dag-maps-reduction-size", cl::Hidden,
84 cl::desc("A huge scheduling region will have maps reduced by this many "
85 "nodes at a time. Defaults to HugeRegion / 2."));
86
87#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
88static cl::opt<bool> SchedPrintCycles(
89 "sched-print-cycles", cl::Hidden, cl::init(false),
90 cl::desc("Report top/bottom cycles when dumping SUnit instances"));
91#endif
92
93static unsigned getReductionSize() {
94 // Always reduce a huge region with half of the elements, except
95 // when user sets this number explicitly.
96 if (ReductionSize.getNumOccurrences() == 0)
97 return HugeRegion / 2;
98 return ReductionSize;
99}
100
101static void dumpSUList(const ScheduleDAGInstrs::SUList &L) {
102#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
103 dbgs() << "{ ";
104 for (const SUnit *SU : L) {
105 dbgs() << "SU(" << SU->NodeNum << ")";
106 if (SU != L.back())
107 dbgs() << ", ";
108 }
109 dbgs() << "}\n";
110#endif
111}
112
113ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf,
114 const MachineLoopInfo *mli,
115 bool RemoveKillFlags)
116 : ScheduleDAG(mf), MLI(mli), MFI(mf.getFrameInfo()),
117 RemoveKillFlags(RemoveKillFlags),
118 UnknownValue(UndefValue::get(
119 T: Type::getVoidTy(C&: mf.getFunction().getContext()))), Topo(SUnits, &ExitSU) {
120 DbgValues.clear();
121
122 const TargetSubtargetInfo &ST = mf.getSubtarget();
123 SchedModel.init(TSInfo: &ST);
124}
125
126/// If this machine instr has memory reference information and it can be
127/// tracked to a normal reference to a known object, return the Value
128/// for that object. This function returns false the memory location is
129/// unknown or may alias anything.
130static bool getUnderlyingObjectsForInstr(const MachineInstr *MI,
131 const MachineFrameInfo &MFI,
132 UnderlyingObjectsVector &Objects,
133 const DataLayout &DL) {
134 auto AllMMOsOkay = [&]() {
135 for (const MachineMemOperand *MMO : MI->memoperands()) {
136 // TODO: Figure out whether isAtomic is really necessary (see D57601).
137 if (MMO->isVolatile() || MMO->isAtomic())
138 return false;
139
140 if (const PseudoSourceValue *PSV = MMO->getPseudoValue()) {
141 // Function that contain tail calls don't have unique PseudoSourceValue
142 // objects. Two PseudoSourceValues might refer to the same or
143 // overlapping locations. The client code calling this function assumes
144 // this is not the case. So return a conservative answer of no known
145 // object.
146 if (MFI.hasTailCall())
147 return false;
148
149 // For now, ignore PseudoSourceValues which may alias LLVM IR values
150 // because the code that uses this function has no way to cope with
151 // such aliases.
152 if (PSV->isAliased(&MFI))
153 return false;
154
155 bool MayAlias = PSV->mayAlias(&MFI);
156 Objects.emplace_back(Args&: PSV, Args&: MayAlias);
157 } else if (const Value *V = MMO->getValue()) {
158 SmallVector<Value *, 4> Objs;
159 if (!getUnderlyingObjectsForCodeGen(V, Objects&: Objs))
160 return false;
161
162 for (Value *V : Objs) {
163 assert(isIdentifiedObject(V));
164 Objects.emplace_back(Args&: V, Args: true);
165 }
166 } else
167 return false;
168 }
169 return true;
170 };
171
172 if (!AllMMOsOkay()) {
173 Objects.clear();
174 return false;
175 }
176
177 return true;
178}
179
180void ScheduleDAGInstrs::startBlock(MachineBasicBlock *bb) {
181 BB = bb;
182}
183
184void ScheduleDAGInstrs::finishBlock() {
185 // Subclasses should no longer refer to the old block.
186 BB = nullptr;
187}
188
189void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
190 MachineBasicBlock::iterator begin,
191 MachineBasicBlock::iterator end,
192 unsigned regioninstrs) {
193 assert(bb == BB && "startBlock should set BB");
194 RegionBegin = begin;
195 RegionEnd = end;
196 NumRegionInstrs = regioninstrs;
197}
198
199void ScheduleDAGInstrs::exitRegion() {
200 // Nothing to do.
201}
202
203void ScheduleDAGInstrs::addSchedBarrierDeps() {
204 MachineInstr *ExitMI =
205 RegionEnd != BB->end()
206 ? &*skipDebugInstructionsBackward(It: RegionEnd, Begin: RegionBegin)
207 : nullptr;
208 ExitSU.setInstr(ExitMI);
209 // Add dependencies on the defs and uses of the instruction.
210 if (ExitMI) {
211 for (const MachineOperand &MO : ExitMI->all_uses()) {
212 Register Reg = MO.getReg();
213 if (Reg.isPhysical()) {
214 for (MCRegUnit Unit : TRI->regunits(Reg))
215 Uses.insert(Val: PhysRegSUOper(&ExitSU, -1, Unit));
216 } else if (Reg.isVirtual() && MO.readsReg()) {
217 addVRegUseDeps(SU: &ExitSU, OperIdx: MO.getOperandNo());
218 }
219 }
220 }
221 if (!ExitMI || (!ExitMI->isCall() && !ExitMI->isBarrier())) {
222 // For others, e.g. fallthrough, conditional branch, assume the exit
223 // uses all the registers that are livein to the successor blocks.
224 for (const MachineBasicBlock *Succ : BB->successors()) {
225 for (const auto &LI : Succ->liveins()) {
226 for (MCRegUnitMaskIterator U(LI.PhysReg, TRI); U.isValid(); ++U) {
227 auto [Unit, Mask] = *U;
228 if ((Mask & LI.LaneMask).any() && !Uses.contains(Key: Unit))
229 Uses.insert(Val: PhysRegSUOper(&ExitSU, -1, Unit));
230 }
231 }
232 }
233 }
234}
235
236/// MO is an operand of SU's instruction that defines a physical register. Adds
237/// data dependencies from SU to any uses of the physical register.
238void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
239 const MachineOperand &MO = SU->getInstr()->getOperand(i: OperIdx);
240 assert(MO.isDef() && "expect physreg def");
241 Register Reg = MO.getReg();
242
243 // Ask the target if address-backscheduling is desirable, and if so how much.
244 const TargetSubtargetInfo &ST = MF.getSubtarget();
245
246 // Only use any non-zero latency for real defs/uses, in contrast to
247 // "fake" operands added by regalloc.
248 const MCInstrDesc &DefMIDesc = SU->getInstr()->getDesc();
249 bool ImplicitPseudoDef = (OperIdx >= DefMIDesc.getNumOperands() &&
250 !DefMIDesc.hasImplicitDefOfPhysReg(Reg));
251 for (MCRegUnit Unit : TRI->regunits(Reg)) {
252 for (RegUnit2SUnitsMap::iterator I = Uses.find(Key: Unit); I != Uses.end();
253 ++I) {
254 SUnit *UseSU = I->SU;
255 if (UseSU == SU)
256 continue;
257
258 // Adjust the dependence latency using operand def/use information,
259 // then allow the target to perform its own adjustments.
260 MachineInstr *UseInstr = nullptr;
261 int UseOpIdx = I->OpIdx;
262 bool ImplicitPseudoUse = false;
263 SDep Dep;
264 if (UseOpIdx < 0) {
265 Dep = SDep(SU, SDep::Artificial);
266 } else {
267 // Set the hasPhysRegDefs only for physreg defs that have a use within
268 // the scheduling region.
269 SU->hasPhysRegDefs = true;
270
271 UseInstr = UseSU->getInstr();
272 Register UseReg = UseInstr->getOperand(i: UseOpIdx).getReg();
273 const MCInstrDesc &UseMIDesc = UseInstr->getDesc();
274 ImplicitPseudoUse = UseOpIdx >= ((int)UseMIDesc.getNumOperands()) &&
275 !UseMIDesc.hasImplicitUseOfPhysReg(Reg: UseReg);
276
277 Dep = SDep(SU, SDep::Data, UseReg);
278 }
279 if (!ImplicitPseudoDef && !ImplicitPseudoUse) {
280 Dep.setLatency(SchedModel.computeOperandLatency(DefMI: SU->getInstr(), DefOperIdx: OperIdx,
281 UseMI: UseInstr, UseOperIdx: UseOpIdx));
282 } else {
283 Dep.setLatency(0);
284 }
285 ST.adjustSchedDependency(Def: SU, DefOpIdx: OperIdx, Use: UseSU, UseOpIdx, Dep, SchedModel: &SchedModel);
286 UseSU->addPred(D: Dep);
287 }
288 }
289}
290
291/// Adds register dependencies (data, anti, and output) from this SUnit
292/// to following instructions in the same scheduling region that depend the
293/// physical register referenced at OperIdx.
294void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
295 MachineInstr *MI = SU->getInstr();
296 MachineOperand &MO = MI->getOperand(i: OperIdx);
297 Register Reg = MO.getReg();
298 // We do not need to track any dependencies for constant registers.
299 if (MRI.isConstantPhysReg(PhysReg: Reg))
300 return;
301
302 const TargetSubtargetInfo &ST = MF.getSubtarget();
303
304 // Optionally add output and anti dependencies. For anti
305 // dependencies we use a latency of 0 because for a multi-issue
306 // target we want to allow the defining instruction to issue
307 // in the same cycle as the using instruction.
308 // TODO: Using a latency of 1 here for output dependencies assumes
309 // there's no cost for reusing registers.
310 SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
311 for (MCRegUnit Unit : TRI->regunits(Reg)) {
312 for (RegUnit2SUnitsMap::iterator I = Defs.find(Key: Unit); I != Defs.end();
313 ++I) {
314 SUnit *DefSU = I->SU;
315 if (DefSU == &ExitSU)
316 continue;
317 MachineInstr *DefInstr = DefSU->getInstr();
318 MachineOperand &DefMO = DefInstr->getOperand(i: I->OpIdx);
319 if (DefSU != SU &&
320 (Kind != SDep::Output || !MO.isDead() || !DefMO.isDead())) {
321 SDep Dep(SU, Kind, DefMO.getReg());
322 if (Kind != SDep::Anti) {
323 Dep.setLatency(
324 SchedModel.computeOutputLatency(DefMI: MI, DefOperIdx: OperIdx, DepMI: DefInstr));
325 }
326 ST.adjustSchedDependency(Def: SU, DefOpIdx: OperIdx, Use: DefSU, UseOpIdx: I->OpIdx, Dep,
327 SchedModel: &SchedModel);
328 DefSU->addPred(D: Dep);
329 }
330 }
331 }
332
333 if (MO.isUse()) {
334 SU->hasPhysRegUses = true;
335 // Either insert a new Reg2SUnits entry with an empty SUnits list, or
336 // retrieve the existing SUnits list for this register's uses.
337 // Push this SUnit on the use list.
338 for (MCRegUnit Unit : TRI->regunits(Reg))
339 Uses.insert(Val: PhysRegSUOper(SU, OperIdx, Unit));
340 if (RemoveKillFlags)
341 MO.setIsKill(false);
342 } else {
343 addPhysRegDataDeps(SU, OperIdx);
344
345 // Clear previous uses and defs of this register and its subregisters.
346 for (MCRegUnit Unit : TRI->regunits(Reg)) {
347 Uses.eraseAll(K: Unit);
348 if (!MO.isDead())
349 Defs.eraseAll(K: Unit);
350 }
351
352 if (MO.isDead() && SU->isCall) {
353 // Calls will not be reordered because of chain dependencies (see
354 // below). Since call operands are dead, calls may continue to be added
355 // to the DefList making dependence checking quadratic in the size of
356 // the block. Instead, we leave only one call at the back of the
357 // DefList.
358 for (MCRegUnit Unit : TRI->regunits(Reg)) {
359 RegUnit2SUnitsMap::RangePair P = Defs.equal_range(K: Unit);
360 RegUnit2SUnitsMap::iterator B = P.first;
361 RegUnit2SUnitsMap::iterator I = P.second;
362 for (bool isBegin = I == B; !isBegin; /* empty */) {
363 isBegin = (--I) == B;
364 if (!I->SU->isCall)
365 break;
366 I = Defs.erase(I);
367 }
368 }
369 }
370
371 // Defs are pushed in the order they are visited and never reordered.
372 for (MCRegUnit Unit : TRI->regunits(Reg))
373 Defs.insert(Val: PhysRegSUOper(SU, OperIdx, Unit));
374 }
375}
376
377LaneBitmask ScheduleDAGInstrs::getLaneMaskForMO(const MachineOperand &MO) const
378{
379 Register Reg = MO.getReg();
380 // No point in tracking lanemasks if we don't have interesting subregisters.
381 const TargetRegisterClass &RC = *MRI.getRegClass(Reg);
382 if (!RC.HasDisjunctSubRegs)
383 return LaneBitmask::getAll();
384
385 unsigned SubReg = MO.getSubReg();
386 if (SubReg == 0)
387 return RC.getLaneMask();
388 return TRI->getSubRegIndexLaneMask(SubIdx: SubReg);
389}
390
391bool ScheduleDAGInstrs::deadDefHasNoUse(const MachineOperand &MO) {
392 auto RegUse = CurrentVRegUses.find(Key: MO.getReg());
393 if (RegUse == CurrentVRegUses.end())
394 return true;
395 return (RegUse->LaneMask & getLaneMaskForMO(MO)).none();
396}
397
398/// Adds register output and data dependencies from this SUnit to instructions
399/// that occur later in the same scheduling region if they read from or write to
400/// the virtual register defined at OperIdx.
401///
402/// TODO: Hoist loop induction variable increments. This has to be
403/// reevaluated. Generally, IV scheduling should be done before coalescing.
404void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
405 MachineInstr *MI = SU->getInstr();
406 MachineOperand &MO = MI->getOperand(i: OperIdx);
407 Register Reg = MO.getReg();
408
409 LaneBitmask DefLaneMask;
410 LaneBitmask KillLaneMask;
411 if (TrackLaneMasks) {
412 bool IsKill = MO.getSubReg() == 0 || MO.isUndef();
413 DefLaneMask = getLaneMaskForMO(MO);
414 // If we have a <read-undef> flag, none of the lane values comes from an
415 // earlier instruction.
416 KillLaneMask = IsKill ? LaneBitmask::getAll() : DefLaneMask;
417
418 if (MO.getSubReg() != 0 && MO.isUndef()) {
419 // There may be other subregister defs on the same instruction of the same
420 // register in later operands. The lanes of other defs will now be live
421 // after this instruction, so these should not be treated as killed by the
422 // instruction even though they appear to be killed in this one operand.
423 for (const MachineOperand &OtherMO :
424 llvm::drop_begin(RangeOrContainer: MI->operands(), N: OperIdx + 1))
425 if (OtherMO.isReg() && OtherMO.isDef() && OtherMO.getReg() == Reg)
426 KillLaneMask &= ~getLaneMaskForMO(MO: OtherMO);
427 }
428
429 // Clear undef flag, we'll re-add it later once we know which subregister
430 // Def is first.
431 MO.setIsUndef(false);
432 } else {
433 DefLaneMask = LaneBitmask::getAll();
434 KillLaneMask = LaneBitmask::getAll();
435 }
436
437 if (MO.isDead()) {
438 assert(deadDefHasNoUse(MO) && "Dead defs should have no uses");
439 } else {
440 // Add data dependence to all uses we found so far.
441 const TargetSubtargetInfo &ST = MF.getSubtarget();
442 for (VReg2SUnitOperIdxMultiMap::iterator I = CurrentVRegUses.find(Key: Reg),
443 E = CurrentVRegUses.end(); I != E; /*empty*/) {
444 LaneBitmask LaneMask = I->LaneMask;
445 // Ignore uses of other lanes.
446 if ((LaneMask & KillLaneMask).none()) {
447 ++I;
448 continue;
449 }
450
451 if ((LaneMask & DefLaneMask).any()) {
452 SUnit *UseSU = I->SU;
453 MachineInstr *Use = UseSU->getInstr();
454 SDep Dep(SU, SDep::Data, Reg);
455 Dep.setLatency(SchedModel.computeOperandLatency(DefMI: MI, DefOperIdx: OperIdx, UseMI: Use,
456 UseOperIdx: I->OperandIndex));
457 ST.adjustSchedDependency(Def: SU, DefOpIdx: OperIdx, Use: UseSU, UseOpIdx: I->OperandIndex, Dep,
458 SchedModel: &SchedModel);
459 UseSU->addPred(D: Dep);
460 }
461
462 LaneMask &= ~KillLaneMask;
463 // If we found a Def for all lanes of this use, remove it from the list.
464 if (LaneMask.any()) {
465 I->LaneMask = LaneMask;
466 ++I;
467 } else
468 I = CurrentVRegUses.erase(I);
469 }
470 }
471
472 // Shortcut: Singly defined vregs do not have output/anti dependencies.
473 if (MRI.hasOneDef(RegNo: Reg))
474 return;
475
476 // Add output dependence to the next nearest defs of this vreg.
477 //
478 // Unless this definition is dead, the output dependence should be
479 // transitively redundant with antidependencies from this definition's
480 // uses. We're conservative for now until we have a way to guarantee the uses
481 // are not eliminated sometime during scheduling. The output dependence edge
482 // is also useful if output latency exceeds def-use latency.
483 LaneBitmask LaneMask = DefLaneMask;
484 for (VReg2SUnit &V2SU : make_range(x: CurrentVRegDefs.find(Key: Reg),
485 y: CurrentVRegDefs.end())) {
486 // Ignore defs for other lanes.
487 if ((V2SU.LaneMask & LaneMask).none())
488 continue;
489 // Add an output dependence.
490 SUnit *DefSU = V2SU.SU;
491 // Ignore additional defs of the same lanes in one instruction. This can
492 // happen because lanemasks are shared for targets with too many
493 // subregisters. We also use some representration tricks/hacks where we
494 // add super-register defs/uses, to imply that although we only access parts
495 // of the reg we care about the full one.
496 if (DefSU == SU)
497 continue;
498 SDep Dep(SU, SDep::Output, Reg);
499 Dep.setLatency(
500 SchedModel.computeOutputLatency(DefMI: MI, DefOperIdx: OperIdx, DepMI: DefSU->getInstr()));
501 DefSU->addPred(D: Dep);
502
503 // Update current definition. This can get tricky if the def was about a
504 // bigger lanemask before. We then have to shrink it and create a new
505 // VReg2SUnit for the non-overlapping part.
506 LaneBitmask OverlapMask = V2SU.LaneMask & LaneMask;
507 LaneBitmask NonOverlapMask = V2SU.LaneMask & ~LaneMask;
508 V2SU.SU = SU;
509 V2SU.LaneMask = OverlapMask;
510 if (NonOverlapMask.any())
511 CurrentVRegDefs.insert(Val: VReg2SUnit(Reg, NonOverlapMask, DefSU));
512 }
513 // If there was no CurrentVRegDefs entry for some lanes yet, create one.
514 if (LaneMask.any())
515 CurrentVRegDefs.insert(Val: VReg2SUnit(Reg, LaneMask, SU));
516}
517
518/// Adds a register data dependency if the instruction that defines the
519/// virtual register used at OperIdx is mapped to an SUnit. Add a register
520/// antidependency from this SUnit to instructions that occur later in the same
521/// scheduling region if they write the virtual register.
522///
523/// TODO: Handle ExitSU "uses" properly.
524void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
525 const MachineInstr *MI = SU->getInstr();
526 assert(!MI->isDebugOrPseudoInstr());
527
528 const MachineOperand &MO = MI->getOperand(i: OperIdx);
529 Register Reg = MO.getReg();
530
531 // Remember the use. Data dependencies will be added when we find the def.
532 LaneBitmask LaneMask = TrackLaneMasks ? getLaneMaskForMO(MO)
533 : LaneBitmask::getAll();
534 CurrentVRegUses.insert(Val: VReg2SUnitOperIdx(Reg, LaneMask, OperIdx, SU));
535
536 // Add antidependences to the following defs of the vreg.
537 for (VReg2SUnit &V2SU : make_range(x: CurrentVRegDefs.find(Key: Reg),
538 y: CurrentVRegDefs.end())) {
539 // Ignore defs for unrelated lanes.
540 LaneBitmask PrevDefLaneMask = V2SU.LaneMask;
541 if ((PrevDefLaneMask & LaneMask).none())
542 continue;
543 if (V2SU.SU == SU)
544 continue;
545
546 V2SU.SU->addPred(D: SDep(SU, SDep::Anti, Reg));
547 }
548}
549
550/// Returns true if MI is an instruction we are unable to reason about
551/// (like a call or something with unmodeled side effects).
552static inline bool isGlobalMemoryObject(MachineInstr *MI) {
553 return MI->isCall() || MI->hasUnmodeledSideEffects() ||
554 (MI->hasOrderedMemoryRef() && !MI->isDereferenceableInvariantLoad());
555}
556
557void ScheduleDAGInstrs::addChainDependency (SUnit *SUa, SUnit *SUb,
558 unsigned Latency) {
559 if (SUa->getInstr()->mayAlias(AA: AAForDep, Other: *SUb->getInstr(), UseTBAA)) {
560 SDep Dep(SUa, SDep::MayAliasMem);
561 Dep.setLatency(Latency);
562 SUb->addPred(D: Dep);
563 }
564}
565
566/// Creates an SUnit for each real instruction, numbered in top-down
567/// topological order. The instruction order A < B, implies that no edge exists
568/// from B to A.
569///
570/// Map each real instruction to its SUnit.
571///
572/// After initSUnits, the SUnits vector cannot be resized and the scheduler may
573/// hang onto SUnit pointers. We may relax this in the future by using SUnit IDs
574/// instead of pointers.
575///
576/// MachineScheduler relies on initSUnits numbering the nodes by their order in
577/// the original instruction list.
578void ScheduleDAGInstrs::initSUnits() {
579 // We'll be allocating one SUnit for each real instruction in the region,
580 // which is contained within a basic block.
581 SUnits.reserve(n: NumRegionInstrs);
582
583 for (MachineInstr &MI : make_range(x: RegionBegin, y: RegionEnd)) {
584 if (MI.isDebugOrPseudoInstr())
585 continue;
586
587 SUnit *SU = newSUnit(MI: &MI);
588 MISUnitMap[&MI] = SU;
589
590 SU->isCall = MI.isCall();
591 SU->isCommutable = MI.isCommutable();
592
593 // Assign the Latency field of SU using target-provided information.
594 SU->Latency = SchedModel.computeInstrLatency(MI: SU->getInstr());
595
596 // If this SUnit uses a reserved or unbuffered resource, mark it as such.
597 //
598 // Reserved resources block an instruction from issuing and stall the
599 // entire pipeline. These are identified by BufferSize=0.
600 //
601 // Unbuffered resources prevent execution of subsequent instructions that
602 // require the same resources. This is used for in-order execution pipelines
603 // within an out-of-order core. These are identified by BufferSize=1.
604 if (SchedModel.hasInstrSchedModel()) {
605 const MCSchedClassDesc *SC = getSchedClass(SU);
606 for (const MCWriteProcResEntry &PRE :
607 make_range(x: SchedModel.getWriteProcResBegin(SC),
608 y: SchedModel.getWriteProcResEnd(SC))) {
609 switch (SchedModel.getProcResource(PIdx: PRE.ProcResourceIdx)->BufferSize) {
610 case 0:
611 SU->hasReservedResource = true;
612 break;
613 case 1:
614 SU->isUnbuffered = true;
615 break;
616 default:
617 break;
618 }
619 }
620 }
621 }
622}
623
624class ScheduleDAGInstrs::Value2SUsMap : public MapVector<ValueType, SUList> {
625 /// Current total number of SUs in map.
626 unsigned NumNodes = 0;
627
628 /// 1 for loads, 0 for stores. (see comment in SUList)
629 unsigned TrueMemOrderLatency;
630
631public:
632 Value2SUsMap(unsigned lat = 0) : TrueMemOrderLatency(lat) {}
633
634 /// To keep NumNodes up to date, insert() is used instead of
635 /// this operator w/ push_back().
636 ValueType &operator[](const SUList &Key) {
637 llvm_unreachable("Don't use. Use insert() instead."); };
638
639 /// Adds SU to the SUList of V. If Map grows huge, reduce its size by calling
640 /// reduce().
641 void inline insert(SUnit *SU, ValueType V) {
642 MapVector::operator[](Key: V).push_back(x: SU);
643 NumNodes++;
644 }
645
646 /// Clears the list of SUs mapped to V.
647 void inline clearList(ValueType V) {
648 iterator Itr = find(Key: V);
649 if (Itr != end()) {
650 assert(NumNodes >= Itr->second.size());
651 NumNodes -= Itr->second.size();
652
653 Itr->second.clear();
654 }
655 }
656
657 /// Clears map from all contents.
658 void clear() {
659 MapVector<ValueType, SUList>::clear();
660 NumNodes = 0;
661 }
662
663 unsigned inline size() const { return NumNodes; }
664
665 /// Counts the number of SUs in this map after a reduction.
666 void reComputeSize() {
667 NumNodes = 0;
668 for (auto &I : *this)
669 NumNodes += I.second.size();
670 }
671
672 unsigned inline getTrueMemOrderLatency() const {
673 return TrueMemOrderLatency;
674 }
675
676 void dump();
677};
678
679void ScheduleDAGInstrs::addChainDependencies(SUnit *SU,
680 Value2SUsMap &Val2SUsMap) {
681 for (auto &I : Val2SUsMap)
682 addChainDependencies(SU, SUs&: I.second,
683 Latency: Val2SUsMap.getTrueMemOrderLatency());
684}
685
686void ScheduleDAGInstrs::addChainDependencies(SUnit *SU,
687 Value2SUsMap &Val2SUsMap,
688 ValueType V) {
689 Value2SUsMap::iterator Itr = Val2SUsMap.find(Key: V);
690 if (Itr != Val2SUsMap.end())
691 addChainDependencies(SU, SUs&: Itr->second,
692 Latency: Val2SUsMap.getTrueMemOrderLatency());
693}
694
695void ScheduleDAGInstrs::addBarrierChain(Value2SUsMap &map) {
696 assert(BarrierChain != nullptr);
697
698 for (auto &[V, SUs] : map) {
699 (void)V;
700 for (auto *SU : SUs)
701 SU->addPredBarrier(SU: BarrierChain);
702 }
703 map.clear();
704}
705
706void ScheduleDAGInstrs::insertBarrierChain(Value2SUsMap &map) {
707 assert(BarrierChain != nullptr);
708
709 // Go through all lists of SUs.
710 for (Value2SUsMap::iterator I = map.begin(), EE = map.end(); I != EE;) {
711 Value2SUsMap::iterator CurrItr = I++;
712 SUList &sus = CurrItr->second;
713 SUList::iterator SUItr = sus.begin(), SUEE = sus.end();
714 for (; SUItr != SUEE; ++SUItr) {
715 // Stop on BarrierChain or any instruction above it.
716 if ((*SUItr)->NodeNum <= BarrierChain->NodeNum)
717 break;
718
719 (*SUItr)->addPredBarrier(SU: BarrierChain);
720 }
721
722 // Remove also the BarrierChain from list if present.
723 if (SUItr != SUEE && *SUItr == BarrierChain)
724 SUItr++;
725
726 // Remove all SUs that are now successors of BarrierChain.
727 if (SUItr != sus.begin())
728 sus.erase(first: sus.begin(), last: SUItr);
729 }
730
731 // Remove all entries with empty su lists.
732 map.remove_if(Pred: [&](std::pair<ValueType, SUList> &mapEntry) {
733 return (mapEntry.second.empty()); });
734
735 // Recompute the size of the map (NumNodes).
736 map.reComputeSize();
737}
738
739void ScheduleDAGInstrs::buildSchedGraph(AAResults *AA,
740 RegPressureTracker *RPTracker,
741 PressureDiffs *PDiffs,
742 LiveIntervals *LIS,
743 bool TrackLaneMasks) {
744 const TargetSubtargetInfo &ST = MF.getSubtarget();
745 bool UseAA = EnableAASchedMI.getNumOccurrences() > 0 ? EnableAASchedMI
746 : ST.useAA();
747 AAForDep = UseAA ? AA : nullptr;
748
749 BarrierChain = nullptr;
750
751 this->TrackLaneMasks = TrackLaneMasks;
752 MISUnitMap.clear();
753 ScheduleDAG::clearDAG();
754
755 // Create an SUnit for each real instruction.
756 initSUnits();
757
758 if (PDiffs)
759 PDiffs->init(N: SUnits.size());
760
761 // We build scheduling units by walking a block's instruction list
762 // from bottom to top.
763
764 // Each MIs' memory operand(s) is analyzed to a list of underlying
765 // objects. The SU is then inserted in the SUList(s) mapped from the
766 // Value(s). Each Value thus gets mapped to lists of SUs depending
767 // on it, stores and loads kept separately. Two SUs are trivially
768 // non-aliasing if they both depend on only identified Values and do
769 // not share any common Value.
770 Value2SUsMap Stores, Loads(1 /*TrueMemOrderLatency*/);
771
772 // Certain memory accesses are known to not alias any SU in Stores
773 // or Loads, and have therefore their own 'NonAlias'
774 // domain. E.g. spill / reload instructions never alias LLVM I/R
775 // Values. It would be nice to assume that this type of memory
776 // accesses always have a proper memory operand modelling, and are
777 // therefore never unanalyzable, but this is conservatively not
778 // done.
779 Value2SUsMap NonAliasStores, NonAliasLoads(1 /*TrueMemOrderLatency*/);
780
781 // Track all instructions that may raise floating-point exceptions.
782 // These do not depend on one other (or normal loads or stores), but
783 // must not be rescheduled across global barriers. Note that we don't
784 // really need a "map" here since we don't track those MIs by value;
785 // using the same Value2SUsMap data type here is simply a matter of
786 // convenience.
787 Value2SUsMap FPExceptions;
788
789 // Remove any stale debug info; sometimes BuildSchedGraph is called again
790 // without emitting the info from the previous call.
791 DbgValues.clear();
792 FirstDbgValue = nullptr;
793
794 assert(Defs.empty() && Uses.empty() &&
795 "Only BuildGraph should update Defs/Uses");
796 Defs.setUniverse(TRI->getNumRegs());
797 Uses.setUniverse(TRI->getNumRegs());
798
799 assert(CurrentVRegDefs.empty() && "nobody else should use CurrentVRegDefs");
800 assert(CurrentVRegUses.empty() && "nobody else should use CurrentVRegUses");
801 unsigned NumVirtRegs = MRI.getNumVirtRegs();
802 CurrentVRegDefs.setUniverse(NumVirtRegs);
803 CurrentVRegUses.setUniverse(NumVirtRegs);
804
805 // Model data dependencies between instructions being scheduled and the
806 // ExitSU.
807 addSchedBarrierDeps();
808
809 // Walk the list of instructions, from bottom moving up.
810 MachineInstr *DbgMI = nullptr;
811 for (MachineBasicBlock::iterator MII = RegionEnd, MIE = RegionBegin;
812 MII != MIE; --MII) {
813 MachineInstr &MI = *std::prev(x: MII);
814 if (DbgMI) {
815 DbgValues.emplace_back(args&: DbgMI, args: &MI);
816 DbgMI = nullptr;
817 }
818
819 if (MI.isDebugValue() || MI.isDebugPHI()) {
820 DbgMI = &MI;
821 continue;
822 }
823
824 if (MI.isDebugLabel() || MI.isDebugRef() || MI.isPseudoProbe())
825 continue;
826
827 SUnit *SU = MISUnitMap[&MI];
828 assert(SU && "No SUnit mapped to this MI");
829
830 if (RPTracker) {
831 RegisterOperands RegOpers;
832 RegOpers.collect(MI, TRI: *TRI, MRI, TrackLaneMasks, IgnoreDead: false);
833 if (TrackLaneMasks) {
834 SlotIndex SlotIdx = LIS->getInstructionIndex(Instr: MI);
835 RegOpers.adjustLaneLiveness(LIS: *LIS, MRI, Pos: SlotIdx);
836 }
837 if (PDiffs != nullptr)
838 PDiffs->addInstruction(Idx: SU->NodeNum, RegOpers, MRI);
839
840 if (RPTracker->getPos() == RegionEnd || &*RPTracker->getPos() != &MI)
841 RPTracker->recedeSkipDebugValues();
842 assert(&*RPTracker->getPos() == &MI && "RPTracker in sync");
843 RPTracker->recede(RegOpers);
844 }
845
846 assert(
847 (CanHandleTerminators || (!MI.isTerminator() && !MI.isPosition())) &&
848 "Cannot schedule terminators or labels!");
849
850 // Add register-based dependencies (data, anti, and output).
851 // For some instructions (calls, returns, inline-asm, etc.) there can
852 // be explicit uses and implicit defs, in which case the use will appear
853 // on the operand list before the def. Do two passes over the operand
854 // list to make sure that defs are processed before any uses.
855 bool HasVRegDef = false;
856 for (unsigned j = 0, n = MI.getNumOperands(); j != n; ++j) {
857 const MachineOperand &MO = MI.getOperand(i: j);
858 if (!MO.isReg() || !MO.isDef())
859 continue;
860 Register Reg = MO.getReg();
861 if (Reg.isPhysical()) {
862 addPhysRegDeps(SU, OperIdx: j);
863 } else if (Reg.isVirtual()) {
864 HasVRegDef = true;
865 addVRegDefDeps(SU, OperIdx: j);
866 }
867 }
868 // Now process all uses.
869 for (unsigned j = 0, n = MI.getNumOperands(); j != n; ++j) {
870 const MachineOperand &MO = MI.getOperand(i: j);
871 // Only look at use operands.
872 // We do not need to check for MO.readsReg() here because subsequent
873 // subregister defs will get output dependence edges and need no
874 // additional use dependencies.
875 if (!MO.isReg() || !MO.isUse())
876 continue;
877 Register Reg = MO.getReg();
878 if (Reg.isPhysical()) {
879 addPhysRegDeps(SU, OperIdx: j);
880 } else if (Reg.isVirtual() && MO.readsReg()) {
881 addVRegUseDeps(SU, OperIdx: j);
882 }
883 }
884
885 // If we haven't seen any uses in this scheduling region, create a
886 // dependence edge to ExitSU to model the live-out latency. This is required
887 // for vreg defs with no in-region use, and prefetches with no vreg def.
888 //
889 // FIXME: NumDataSuccs would be more precise than NumSuccs here. This
890 // check currently relies on being called before adding chain deps.
891 if (SU->NumSuccs == 0 && SU->Latency > 1 && (HasVRegDef || MI.mayLoad())) {
892 SDep Dep(SU, SDep::Artificial);
893 Dep.setLatency(SU->Latency - 1);
894 ExitSU.addPred(D: Dep);
895 }
896
897 // Add memory dependencies (Note: isStoreToStackSlot and
898 // isLoadFromStackSLot are not usable after stack slots are lowered to
899 // actual addresses).
900
901 // This is a barrier event that acts as a pivotal node in the DAG.
902 if (isGlobalMemoryObject(MI: &MI)) {
903
904 // Become the barrier chain.
905 if (BarrierChain)
906 BarrierChain->addPredBarrier(SU);
907 BarrierChain = SU;
908
909 LLVM_DEBUG(dbgs() << "Global memory object and new barrier chain: SU("
910 << BarrierChain->NodeNum << ").\n";);
911
912 // Add dependencies against everything below it and clear maps.
913 addBarrierChain(map&: Stores);
914 addBarrierChain(map&: Loads);
915 addBarrierChain(map&: NonAliasStores);
916 addBarrierChain(map&: NonAliasLoads);
917 addBarrierChain(map&: FPExceptions);
918
919 continue;
920 }
921
922 // Instructions that may raise FP exceptions may not be moved
923 // across any global barriers.
924 if (MI.mayRaiseFPException()) {
925 if (BarrierChain)
926 BarrierChain->addPredBarrier(SU);
927
928 FPExceptions.insert(SU, V: UnknownValue);
929
930 if (FPExceptions.size() >= HugeRegion) {
931 LLVM_DEBUG(dbgs() << "Reducing FPExceptions map.\n";);
932 Value2SUsMap empty;
933 reduceHugeMemNodeMaps(stores&: FPExceptions, loads&: empty, N: getReductionSize());
934 }
935 }
936
937 // If it's not a store or a variant load, we're done.
938 if (!MI.mayStore() &&
939 !(MI.mayLoad() && !MI.isDereferenceableInvariantLoad()))
940 continue;
941
942 // Always add dependecy edge to BarrierChain if present.
943 if (BarrierChain)
944 BarrierChain->addPredBarrier(SU);
945
946 // Find the underlying objects for MI. The Objs vector is either
947 // empty, or filled with the Values of memory locations which this
948 // SU depends on.
949 UnderlyingObjectsVector Objs;
950 bool ObjsFound = getUnderlyingObjectsForInstr(MI: &MI, MFI, Objects&: Objs,
951 DL: MF.getDataLayout());
952
953 if (MI.mayStore()) {
954 if (!ObjsFound) {
955 // An unknown store depends on all stores and loads.
956 addChainDependencies(SU, Val2SUsMap&: Stores);
957 addChainDependencies(SU, Val2SUsMap&: NonAliasStores);
958 addChainDependencies(SU, Val2SUsMap&: Loads);
959 addChainDependencies(SU, Val2SUsMap&: NonAliasLoads);
960
961 // Map this store to 'UnknownValue'.
962 Stores.insert(SU, V: UnknownValue);
963 } else {
964 // Add precise dependencies against all previously seen memory
965 // accesses mapped to the same Value(s).
966 for (const UnderlyingObject &UnderlObj : Objs) {
967 ValueType V = UnderlObj.getValue();
968 bool ThisMayAlias = UnderlObj.mayAlias();
969
970 // Add dependencies to previous stores and loads mapped to V.
971 addChainDependencies(SU, Val2SUsMap&: (ThisMayAlias ? Stores : NonAliasStores), V);
972 addChainDependencies(SU, Val2SUsMap&: (ThisMayAlias ? Loads : NonAliasLoads), V);
973 }
974 // Update the store map after all chains have been added to avoid adding
975 // self-loop edge if multiple underlying objects are present.
976 for (const UnderlyingObject &UnderlObj : Objs) {
977 ValueType V = UnderlObj.getValue();
978 bool ThisMayAlias = UnderlObj.mayAlias();
979
980 // Map this store to V.
981 (ThisMayAlias ? Stores : NonAliasStores).insert(SU, V);
982 }
983 // The store may have dependencies to unanalyzable loads and
984 // stores.
985 addChainDependencies(SU, Val2SUsMap&: Loads, V: UnknownValue);
986 addChainDependencies(SU, Val2SUsMap&: Stores, V: UnknownValue);
987 }
988 } else { // SU is a load.
989 if (!ObjsFound) {
990 // An unknown load depends on all stores.
991 addChainDependencies(SU, Val2SUsMap&: Stores);
992 addChainDependencies(SU, Val2SUsMap&: NonAliasStores);
993
994 Loads.insert(SU, V: UnknownValue);
995 } else {
996 for (const UnderlyingObject &UnderlObj : Objs) {
997 ValueType V = UnderlObj.getValue();
998 bool ThisMayAlias = UnderlObj.mayAlias();
999
1000 // Add precise dependencies against all previously seen stores
1001 // mapping to the same Value(s).
1002 addChainDependencies(SU, Val2SUsMap&: (ThisMayAlias ? Stores : NonAliasStores), V);
1003
1004 // Map this load to V.
1005 (ThisMayAlias ? Loads : NonAliasLoads).insert(SU, V);
1006 }
1007 // The load may have dependencies to unanalyzable stores.
1008 addChainDependencies(SU, Val2SUsMap&: Stores, V: UnknownValue);
1009 }
1010 }
1011
1012 // Reduce maps if they grow huge.
1013 if (Stores.size() + Loads.size() >= HugeRegion) {
1014 LLVM_DEBUG(dbgs() << "Reducing Stores and Loads maps.\n";);
1015 reduceHugeMemNodeMaps(stores&: Stores, loads&: Loads, N: getReductionSize());
1016 }
1017 if (NonAliasStores.size() + NonAliasLoads.size() >= HugeRegion) {
1018 LLVM_DEBUG(
1019 dbgs() << "Reducing NonAliasStores and NonAliasLoads maps.\n";);
1020 reduceHugeMemNodeMaps(stores&: NonAliasStores, loads&: NonAliasLoads, N: getReductionSize());
1021 }
1022 }
1023
1024 if (DbgMI)
1025 FirstDbgValue = DbgMI;
1026
1027 Defs.clear();
1028 Uses.clear();
1029 CurrentVRegDefs.clear();
1030 CurrentVRegUses.clear();
1031
1032 Topo.MarkDirty();
1033}
1034
1035raw_ostream &llvm::operator<<(raw_ostream &OS, const PseudoSourceValue* PSV) {
1036 PSV->printCustom(O&: OS);
1037 return OS;
1038}
1039
1040void ScheduleDAGInstrs::Value2SUsMap::dump() {
1041 for (const auto &[ValType, SUs] : *this) {
1042 if (isa<const Value *>(Val: ValType)) {
1043 const Value *V = cast<const Value *>(Val: ValType);
1044 if (isa<UndefValue>(Val: V))
1045 dbgs() << "Unknown";
1046 else
1047 V->printAsOperand(O&: dbgs());
1048 } else if (isa<const PseudoSourceValue *>(Val: ValType))
1049 dbgs() << cast<const PseudoSourceValue *>(Val: ValType);
1050 else
1051 llvm_unreachable("Unknown Value type.");
1052
1053 dbgs() << " : ";
1054 dumpSUList(L: SUs);
1055 }
1056}
1057
1058void ScheduleDAGInstrs::reduceHugeMemNodeMaps(Value2SUsMap &stores,
1059 Value2SUsMap &loads, unsigned N) {
1060 LLVM_DEBUG(dbgs() << "Before reduction:\nStoring SUnits:\n"; stores.dump();
1061 dbgs() << "Loading SUnits:\n"; loads.dump());
1062
1063 // Insert all SU's NodeNums into a vector and sort it.
1064 std::vector<unsigned> NodeNums;
1065 NodeNums.reserve(n: stores.size() + loads.size());
1066 for (const auto &[V, SUs] : stores) {
1067 (void)V;
1068 for (const auto *SU : SUs)
1069 NodeNums.push_back(x: SU->NodeNum);
1070 }
1071 for (const auto &[V, SUs] : loads) {
1072 (void)V;
1073 for (const auto *SU : SUs)
1074 NodeNums.push_back(x: SU->NodeNum);
1075 }
1076 llvm::sort(C&: NodeNums);
1077
1078 // The N last elements in NodeNums will be removed, and the SU with
1079 // the lowest NodeNum of them will become the new BarrierChain to
1080 // let the not yet seen SUs have a dependency to the removed SUs.
1081 assert(N <= NodeNums.size());
1082 SUnit *newBarrierChain = &SUnits[*(NodeNums.end() - N)];
1083 if (BarrierChain) {
1084 // The aliasing and non-aliasing maps reduce independently of each
1085 // other, but share a common BarrierChain. Check if the
1086 // newBarrierChain is above the former one. If it is not, it may
1087 // introduce a loop to use newBarrierChain, so keep the old one.
1088 if (newBarrierChain->NodeNum < BarrierChain->NodeNum) {
1089 BarrierChain->addPredBarrier(SU: newBarrierChain);
1090 BarrierChain = newBarrierChain;
1091 LLVM_DEBUG(dbgs() << "Inserting new barrier chain: SU("
1092 << BarrierChain->NodeNum << ").\n";);
1093 }
1094 else
1095 LLVM_DEBUG(dbgs() << "Keeping old barrier chain: SU("
1096 << BarrierChain->NodeNum << ").\n";);
1097 }
1098 else
1099 BarrierChain = newBarrierChain;
1100
1101 insertBarrierChain(map&: stores);
1102 insertBarrierChain(map&: loads);
1103
1104 LLVM_DEBUG(dbgs() << "After reduction:\nStoring SUnits:\n"; stores.dump();
1105 dbgs() << "Loading SUnits:\n"; loads.dump());
1106}
1107
1108static void toggleKills(const MachineRegisterInfo &MRI, LiveRegUnits &LiveRegs,
1109 MachineInstr &MI, bool addToLiveRegs) {
1110 for (MachineOperand &MO : MI.operands()) {
1111 if (!MO.isReg() || !MO.readsReg())
1112 continue;
1113 Register Reg = MO.getReg();
1114 if (!Reg)
1115 continue;
1116
1117 // Things that are available after the instruction are killed by it.
1118 bool IsKill = LiveRegs.available(Reg);
1119
1120 // Exception: Do not kill reserved registers
1121 MO.setIsKill(IsKill && !MRI.isReserved(PhysReg: Reg));
1122 if (addToLiveRegs)
1123 LiveRegs.addReg(Reg);
1124 }
1125}
1126
1127void ScheduleDAGInstrs::fixupKills(MachineBasicBlock &MBB) {
1128 LLVM_DEBUG(dbgs() << "Fixup kills for " << printMBBReference(MBB) << '\n');
1129
1130 LiveRegs.init(TRI: *TRI);
1131 LiveRegs.addLiveOuts(MBB);
1132
1133 // Examine block from end to start...
1134 for (MachineInstr &MI : llvm::reverse(C&: MBB)) {
1135 if (MI.isDebugOrPseudoInstr())
1136 continue;
1137
1138 // Update liveness. Registers that are defed but not used in this
1139 // instruction are now dead. Mark register and all subregs as they
1140 // are completely defined.
1141 for (ConstMIBundleOperands O(MI); O.isValid(); ++O) {
1142 const MachineOperand &MO = *O;
1143 if (MO.isReg()) {
1144 if (!MO.isDef())
1145 continue;
1146 Register Reg = MO.getReg();
1147 if (!Reg)
1148 continue;
1149 LiveRegs.removeReg(Reg);
1150 } else if (MO.isRegMask()) {
1151 LiveRegs.removeRegsNotPreserved(RegMask: MO.getRegMask());
1152 }
1153 }
1154
1155 // If there is a bundle header fix it up first.
1156 if (!MI.isBundled()) {
1157 toggleKills(MRI, LiveRegs, MI, addToLiveRegs: true);
1158 } else {
1159 MachineBasicBlock::instr_iterator Bundle = MI.getIterator();
1160 if (MI.isBundle())
1161 toggleKills(MRI, LiveRegs, MI, addToLiveRegs: false);
1162
1163 // Some targets make the (questionable) assumtion that the instructions
1164 // inside the bundle are ordered and consequently only the last use of
1165 // a register inside the bundle can kill it.
1166 MachineBasicBlock::instr_iterator I = std::next(x: Bundle);
1167 while (I->isBundledWithSucc())
1168 ++I;
1169 do {
1170 if (!I->isDebugOrPseudoInstr())
1171 toggleKills(MRI, LiveRegs, MI&: *I, addToLiveRegs: true);
1172 --I;
1173 } while (I != Bundle);
1174 }
1175 }
1176}
1177
1178void ScheduleDAGInstrs::dumpNode(const SUnit &SU) const {
1179#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1180 dumpNodeName(SU);
1181 if (SchedPrintCycles)
1182 dbgs() << " [TopReadyCycle = " << SU.TopReadyCycle
1183 << ", BottomReadyCycle = " << SU.BotReadyCycle << "]";
1184 dbgs() << ": ";
1185 SU.getInstr()->dump();
1186#endif
1187}
1188
1189void ScheduleDAGInstrs::dump() const {
1190#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1191 if (EntrySU.getInstr() != nullptr)
1192 dumpNodeAll(EntrySU);
1193 for (const SUnit &SU : SUnits)
1194 dumpNodeAll(SU);
1195 if (ExitSU.getInstr() != nullptr)
1196 dumpNodeAll(ExitSU);
1197#endif
1198}
1199
1200std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
1201 std::string s;
1202 raw_string_ostream oss(s);
1203 if (SU == &EntrySU)
1204 oss << "<entry>";
1205 else if (SU == &ExitSU)
1206 oss << "<exit>";
1207 else
1208 SU->getInstr()->print(OS&: oss, /*IsStandalone=*/true);
1209 return s;
1210}
1211
1212/// Return the basic block label. It is not necessarilly unique because a block
1213/// contains multiple scheduling regions. But it is fine for visualization.
1214std::string ScheduleDAGInstrs::getDAGName() const {
1215 return "dag." + BB->getFullName();
1216}
1217
1218bool ScheduleDAGInstrs::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
1219 return SuccSU == &ExitSU || !Topo.IsReachable(SU: PredSU, TargetSU: SuccSU);
1220}
1221
1222bool ScheduleDAGInstrs::addEdge(SUnit *SuccSU, const SDep &PredDep) {
1223 if (SuccSU != &ExitSU) {
1224 // Do not use WillCreateCycle, it assumes SD scheduling.
1225 // If Pred is reachable from Succ, then the edge creates a cycle.
1226 if (Topo.IsReachable(SU: PredDep.getSUnit(), TargetSU: SuccSU))
1227 return false;
1228 Topo.AddPredQueued(Y: SuccSU, X: PredDep.getSUnit());
1229 }
1230 SuccSU->addPred(D: PredDep, /*Required=*/!PredDep.isArtificial());
1231 // Return true regardless of whether a new edge needed to be inserted.
1232 return true;
1233}
1234
1235//===----------------------------------------------------------------------===//
1236// SchedDFSResult Implementation
1237//===----------------------------------------------------------------------===//
1238
1239namespace llvm {
1240
1241/// Internal state used to compute SchedDFSResult.
1242class SchedDFSImpl {
1243 SchedDFSResult &R;
1244
1245 /// Join DAG nodes into equivalence classes by their subtree.
1246 IntEqClasses SubtreeClasses;
1247 /// List PredSU, SuccSU pairs that represent data edges between subtrees.
1248 std::vector<std::pair<const SUnit *, const SUnit*>> ConnectionPairs;
1249
1250 struct RootData {
1251 unsigned NodeID;
1252 unsigned ParentNodeID; ///< Parent node (member of the parent subtree).
1253 unsigned SubInstrCount = 0; ///< Instr count in this tree only, not
1254 /// children.
1255
1256 RootData(unsigned id): NodeID(id),
1257 ParentNodeID(SchedDFSResult::InvalidSubtreeID) {}
1258
1259 unsigned getSparseSetIndex() const { return NodeID; }
1260 };
1261
1262 SparseSet<RootData> RootSet;
1263
1264public:
1265 SchedDFSImpl(SchedDFSResult &r): R(r), SubtreeClasses(R.DFSNodeData.size()) {
1266 RootSet.setUniverse(R.DFSNodeData.size());
1267 }
1268
1269 /// Returns true if this node been visited by the DFS traversal.
1270 ///
1271 /// During visitPostorderNode the Node's SubtreeID is assigned to the Node
1272 /// ID. Later, SubtreeID is updated but remains valid.
1273 bool isVisited(const SUnit *SU) const {
1274 return R.DFSNodeData[SU->NodeNum].SubtreeID
1275 != SchedDFSResult::InvalidSubtreeID;
1276 }
1277
1278 /// Initializes this node's instruction count. We don't need to flag the node
1279 /// visited until visitPostorder because the DAG cannot have cycles.
1280 void visitPreorder(const SUnit *SU) {
1281 R.DFSNodeData[SU->NodeNum].InstrCount =
1282 SU->getInstr()->isTransient() ? 0 : 1;
1283 }
1284
1285 /// Called once for each node after all predecessors are visited. Revisit this
1286 /// node's predecessors and potentially join them now that we know the ILP of
1287 /// the other predecessors.
1288 void visitPostorderNode(const SUnit *SU) {
1289 // Mark this node as the root of a subtree. It may be joined with its
1290 // successors later.
1291 R.DFSNodeData[SU->NodeNum].SubtreeID = SU->NodeNum;
1292 RootData RData(SU->NodeNum);
1293 RData.SubInstrCount = SU->getInstr()->isTransient() ? 0 : 1;
1294
1295 // If any predecessors are still in their own subtree, they either cannot be
1296 // joined or are large enough to remain separate. If this parent node's
1297 // total instruction count is not greater than a child subtree by at least
1298 // the subtree limit, then try to join it now since splitting subtrees is
1299 // only useful if multiple high-pressure paths are possible.
1300 unsigned InstrCount = R.DFSNodeData[SU->NodeNum].InstrCount;
1301 for (const SDep &PredDep : SU->Preds) {
1302 if (PredDep.getKind() != SDep::Data)
1303 continue;
1304 unsigned PredNum = PredDep.getSUnit()->NodeNum;
1305 if ((InstrCount - R.DFSNodeData[PredNum].InstrCount) < R.SubtreeLimit)
1306 joinPredSubtree(PredDep, Succ: SU, /*CheckLimit=*/false);
1307
1308 // Either link or merge the TreeData entry from the child to the parent.
1309 if (R.DFSNodeData[PredNum].SubtreeID == PredNum) {
1310 // If the predecessor's parent is invalid, this is a tree edge and the
1311 // current node is the parent.
1312 if (RootSet[PredNum].ParentNodeID == SchedDFSResult::InvalidSubtreeID)
1313 RootSet[PredNum].ParentNodeID = SU->NodeNum;
1314 }
1315 else if (RootSet.count(Key: PredNum)) {
1316 // The predecessor is not a root, but is still in the root set. This
1317 // must be the new parent that it was just joined to. Note that
1318 // RootSet[PredNum].ParentNodeID may either be invalid or may still be
1319 // set to the original parent.
1320 RData.SubInstrCount += RootSet[PredNum].SubInstrCount;
1321 RootSet.erase(Key: PredNum);
1322 }
1323 }
1324 RootSet[SU->NodeNum] = RData;
1325 }
1326
1327 /// Called once for each tree edge after calling visitPostOrderNode on
1328 /// the predecessor. Increment the parent node's instruction count and
1329 /// preemptively join this subtree to its parent's if it is small enough.
1330 void visitPostorderEdge(const SDep &PredDep, const SUnit *Succ) {
1331 R.DFSNodeData[Succ->NodeNum].InstrCount
1332 += R.DFSNodeData[PredDep.getSUnit()->NodeNum].InstrCount;
1333 joinPredSubtree(PredDep, Succ);
1334 }
1335
1336 /// Adds a connection for cross edges.
1337 void visitCrossEdge(const SDep &PredDep, const SUnit *Succ) {
1338 ConnectionPairs.emplace_back(args: PredDep.getSUnit(), args&: Succ);
1339 }
1340
1341 /// Sets each node's subtree ID to the representative ID and record
1342 /// connections between trees.
1343 void finalize() {
1344 SubtreeClasses.compress();
1345 R.DFSTreeData.resize(N: SubtreeClasses.getNumClasses());
1346 assert(SubtreeClasses.getNumClasses() == RootSet.size()
1347 && "number of roots should match trees");
1348 for (const RootData &Root : RootSet) {
1349 unsigned TreeID = SubtreeClasses[Root.NodeID];
1350 if (Root.ParentNodeID != SchedDFSResult::InvalidSubtreeID)
1351 R.DFSTreeData[TreeID].ParentTreeID = SubtreeClasses[Root.ParentNodeID];
1352 R.DFSTreeData[TreeID].SubInstrCount = Root.SubInstrCount;
1353 // Note that SubInstrCount may be greater than InstrCount if we joined
1354 // subtrees across a cross edge. InstrCount will be attributed to the
1355 // original parent, while SubInstrCount will be attributed to the joined
1356 // parent.
1357 }
1358 R.SubtreeConnections.resize(new_size: SubtreeClasses.getNumClasses());
1359 R.SubtreeConnectLevels.resize(new_size: SubtreeClasses.getNumClasses());
1360 LLVM_DEBUG(dbgs() << R.getNumSubtrees() << " subtrees:\n");
1361 for (unsigned Idx = 0, End = R.DFSNodeData.size(); Idx != End; ++Idx) {
1362 R.DFSNodeData[Idx].SubtreeID = SubtreeClasses[Idx];
1363 LLVM_DEBUG(dbgs() << " SU(" << Idx << ") in tree "
1364 << R.DFSNodeData[Idx].SubtreeID << '\n');
1365 }
1366 for (const auto &[Pred, Succ] : ConnectionPairs) {
1367 unsigned PredTree = SubtreeClasses[Pred->NodeNum];
1368 unsigned SuccTree = SubtreeClasses[Succ->NodeNum];
1369 if (PredTree == SuccTree)
1370 continue;
1371 unsigned Depth = Pred->getDepth();
1372 addConnection(FromTree: PredTree, ToTree: SuccTree, Depth);
1373 addConnection(FromTree: SuccTree, ToTree: PredTree, Depth);
1374 }
1375 }
1376
1377protected:
1378 /// Joins the predecessor subtree with the successor that is its DFS parent.
1379 /// Applies some heuristics before joining.
1380 bool joinPredSubtree(const SDep &PredDep, const SUnit *Succ,
1381 bool CheckLimit = true) {
1382 assert(PredDep.getKind() == SDep::Data && "Subtrees are for data edges");
1383
1384 // Check if the predecessor is already joined.
1385 const SUnit *PredSU = PredDep.getSUnit();
1386 unsigned PredNum = PredSU->NodeNum;
1387 if (R.DFSNodeData[PredNum].SubtreeID != PredNum)
1388 return false;
1389
1390 // Four is the magic number of successors before a node is considered a
1391 // pinch point.
1392 unsigned NumDataSucs = 0;
1393 for (const SDep &SuccDep : PredSU->Succs) {
1394 if (SuccDep.getKind() == SDep::Data) {
1395 if (++NumDataSucs >= 4)
1396 return false;
1397 }
1398 }
1399 if (CheckLimit && R.DFSNodeData[PredNum].InstrCount > R.SubtreeLimit)
1400 return false;
1401 R.DFSNodeData[PredNum].SubtreeID = Succ->NodeNum;
1402 SubtreeClasses.join(a: Succ->NodeNum, b: PredNum);
1403 return true;
1404 }
1405
1406 /// Called by finalize() to record a connection between trees.
1407 void addConnection(unsigned FromTree, unsigned ToTree, unsigned Depth) {
1408 if (!Depth)
1409 return;
1410
1411 do {
1412 SmallVectorImpl<SchedDFSResult::Connection> &Connections =
1413 R.SubtreeConnections[FromTree];
1414 for (SchedDFSResult::Connection &C : Connections) {
1415 if (C.TreeID == ToTree) {
1416 C.Level = std::max(a: C.Level, b: Depth);
1417 return;
1418 }
1419 }
1420 Connections.push_back(Elt: SchedDFSResult::Connection(ToTree, Depth));
1421 FromTree = R.DFSTreeData[FromTree].ParentTreeID;
1422 } while (FromTree != SchedDFSResult::InvalidSubtreeID);
1423 }
1424};
1425
1426} // end namespace llvm
1427
1428namespace {
1429
1430/// Manage the stack used by a reverse depth-first search over the DAG.
1431class SchedDAGReverseDFS {
1432 std::vector<std::pair<const SUnit *, SUnit::const_pred_iterator>> DFSStack;
1433
1434public:
1435 bool isComplete() const { return DFSStack.empty(); }
1436
1437 void follow(const SUnit *SU) {
1438 DFSStack.emplace_back(args&: SU, args: SU->Preds.begin());
1439 }
1440 void advance() { ++DFSStack.back().second; }
1441
1442 const SDep *backtrack() {
1443 DFSStack.pop_back();
1444 return DFSStack.empty() ? nullptr : std::prev(x: DFSStack.back().second);
1445 }
1446
1447 const SUnit *getCurr() const { return DFSStack.back().first; }
1448
1449 SUnit::const_pred_iterator getPred() const { return DFSStack.back().second; }
1450
1451 SUnit::const_pred_iterator getPredEnd() const {
1452 return getCurr()->Preds.end();
1453 }
1454};
1455
1456} // end anonymous namespace
1457
1458static bool hasDataSucc(const SUnit *SU) {
1459 for (const SDep &SuccDep : SU->Succs) {
1460 if (SuccDep.getKind() == SDep::Data &&
1461 !SuccDep.getSUnit()->isBoundaryNode())
1462 return true;
1463 }
1464 return false;
1465}
1466
1467/// Computes an ILP metric for all nodes in the subDAG reachable via depth-first
1468/// search from this root.
1469void SchedDFSResult::compute(ArrayRef<SUnit> SUnits) {
1470 if (!IsBottomUp)
1471 llvm_unreachable("Top-down ILP metric is unimplemented");
1472
1473 SchedDFSImpl Impl(*this);
1474 for (const SUnit &SU : SUnits) {
1475 if (Impl.isVisited(SU: &SU) || hasDataSucc(SU: &SU))
1476 continue;
1477
1478 SchedDAGReverseDFS DFS;
1479 Impl.visitPreorder(SU: &SU);
1480 DFS.follow(SU: &SU);
1481 while (true) {
1482 // Traverse the leftmost path as far as possible.
1483 while (DFS.getPred() != DFS.getPredEnd()) {
1484 const SDep &PredDep = *DFS.getPred();
1485 DFS.advance();
1486 // Ignore non-data edges.
1487 if (PredDep.getKind() != SDep::Data
1488 || PredDep.getSUnit()->isBoundaryNode()) {
1489 continue;
1490 }
1491 // An already visited edge is a cross edge, assuming an acyclic DAG.
1492 if (Impl.isVisited(SU: PredDep.getSUnit())) {
1493 Impl.visitCrossEdge(PredDep, Succ: DFS.getCurr());
1494 continue;
1495 }
1496 Impl.visitPreorder(SU: PredDep.getSUnit());
1497 DFS.follow(SU: PredDep.getSUnit());
1498 }
1499 // Visit the top of the stack in postorder and backtrack.
1500 const SUnit *Child = DFS.getCurr();
1501 const SDep *PredDep = DFS.backtrack();
1502 Impl.visitPostorderNode(SU: Child);
1503 if (PredDep)
1504 Impl.visitPostorderEdge(PredDep: *PredDep, Succ: DFS.getCurr());
1505 if (DFS.isComplete())
1506 break;
1507 }
1508 }
1509 Impl.finalize();
1510}
1511
1512/// The root of the given SubtreeID was just scheduled. For all subtrees
1513/// connected to this tree, record the depth of the connection so that the
1514/// nearest connected subtrees can be prioritized.
1515void SchedDFSResult::scheduleTree(unsigned SubtreeID) {
1516 for (const Connection &C : SubtreeConnections[SubtreeID]) {
1517 SubtreeConnectLevels[C.TreeID] =
1518 std::max(a: SubtreeConnectLevels[C.TreeID], b: C.Level);
1519 LLVM_DEBUG(dbgs() << " Tree: " << C.TreeID << " @"
1520 << SubtreeConnectLevels[C.TreeID] << '\n');
1521 }
1522}
1523
1524#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1525LLVM_DUMP_METHOD void ILPValue::print(raw_ostream &OS) const {
1526 OS << InstrCount << " / " << Length << " = ";
1527 if (!Length)
1528 OS << "BADILP";
1529 else
1530 OS << format("%g", ((double)InstrCount / Length));
1531}
1532
1533LLVM_DUMP_METHOD void ILPValue::dump() const {
1534 dbgs() << *this << '\n';
1535}
1536
1537namespace llvm {
1538
1539LLVM_ATTRIBUTE_UNUSED
1540raw_ostream &operator<<(raw_ostream &OS, const ILPValue &Val) {
1541 Val.print(OS);
1542 return OS;
1543}
1544
1545} // end namespace llvm
1546
1547#endif
1548