blob: d1c5ddabb9755023d17b2a3ed158b46e378b11d0 [file] [log] [blame]
//===---- ScheduleDAGInstrs.cpp - MachineInstr Rescheduling ---------------===//
// The LLVM Compiler Infrastructure
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
/// \file This implements the ScheduleDAGInstrs class, which implements
/// re-scheduling of MachineInstrs.
#include "llvm/CodeGen/ScheduleDAGInstrs.h"
#include "llvm/ADT/IntEqClasses.h"
#include "llvm/ADT/MapVector.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/SparseSet.h"
#include "llvm/ADT/iterator_range.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/CodeGen/LiveIntervals.h"
#include "llvm/CodeGen/LivePhysRegs.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBundle.h"
#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
#include "llvm/CodeGen/RegisterPressure.h"
#include "llvm/CodeGen/ScheduleDAG.h"
#include "llvm/CodeGen/ScheduleDFS.h"
#include "llvm/CodeGen/SlotIndexes.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
#include "llvm/Config/llvm-config.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Operator.h"
#include "llvm/IR/Type.h"
#include "llvm/IR/Value.h"
#include "llvm/MC/LaneBitmask.h"
#include "llvm/MC/MCRegisterInfo.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/Format.h"
#include "llvm/Support/raw_ostream.h"
#include <algorithm>
#include <cassert>
#include <iterator>
#include <string>
#include <utility>
#include <vector>
using namespace llvm;
#define DEBUG_TYPE "machine-scheduler"
static cl::opt<bool> EnableAASchedMI("enable-aa-sched-mi", cl::Hidden,
cl::ZeroOrMore, cl::init(false),
cl::desc("Enable use of AA during MI DAG construction"));
static cl::opt<bool> UseTBAA("use-tbaa-in-sched-mi", cl::Hidden,
cl::init(true), cl::desc("Enable use of TBAA during MI DAG construction"));
// Note: the two options below might be used in tuning compile time vs
// output quality. Setting HugeRegion so large that it will never be
// reached means best-effort, but may be slow.
// When Stores and Loads maps (or NonAliasStores and NonAliasLoads)
// together hold this many SUs, a reduction of maps will be done.
static cl::opt<unsigned> HugeRegion("dag-maps-huge-region", cl::Hidden,
cl::init(1000), cl::desc("The limit to use while constructing the DAG "
"prior to scheduling, at which point a trade-off "
"is made to avoid excessive compile time."));
static cl::opt<unsigned> ReductionSize(
"dag-maps-reduction-size", cl::Hidden,
cl::desc("A huge scheduling region will have maps reduced by this many "
"nodes at a time. Defaults to HugeRegion / 2."));
static unsigned getReductionSize() {
// Always reduce a huge region with half of the elements, except
// when user sets this number explicitly.
if (ReductionSize.getNumOccurrences() == 0)
return HugeRegion / 2;
return ReductionSize;
static void dumpSUList(ScheduleDAGInstrs::SUList &L) {
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dbgs() << "{ ";
for (const SUnit *su : L) {
dbgs() << "SU(" << su->NodeNum << ")";
if (su != L.back())
dbgs() << ", ";
dbgs() << "}\n";
ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf,
const MachineLoopInfo *mli,
bool RemoveKillFlags)
: ScheduleDAG(mf), MLI(mli), MFI(mf.getFrameInfo()),
Type::getVoidTy(mf.getFunction().getContext()))) {
const TargetSubtargetInfo &ST = mf.getSubtarget();
/// If this machine instr has memory reference information and it can be
/// tracked to a normal reference to a known object, return the Value
/// for that object. This function returns false the memory location is
/// unknown or may alias anything.
static bool getUnderlyingObjectsForInstr(const MachineInstr *MI,
const MachineFrameInfo &MFI,
UnderlyingObjectsVector &Objects,
const DataLayout &DL) {
auto allMMOsOkay = [&]() {
for (const MachineMemOperand *MMO : MI->memoperands()) {
if (MMO->isVolatile())
return false;
if (const PseudoSourceValue *PSV = MMO->getPseudoValue()) {
// Function that contain tail calls don't have unique PseudoSourceValue
// objects. Two PseudoSourceValues might refer to the same or
// overlapping locations. The client code calling this function assumes
// this is not the case. So return a conservative answer of no known
// object.
if (MFI.hasTailCall())
return false;
// For now, ignore PseudoSourceValues which may alias LLVM IR values
// because the code that uses this function has no way to cope with
// such aliases.
if (PSV->isAliased(&MFI))
return false;
bool MayAlias = PSV->mayAlias(&MFI);
Objects.push_back(UnderlyingObjectsVector::value_type(PSV, MayAlias));
} else if (const Value *V = MMO->getValue()) {
SmallVector<Value *, 4> Objs;
if (!getUnderlyingObjectsForCodeGen(V, Objs, DL))
return false;
for (Value *V : Objs) {
Objects.push_back(UnderlyingObjectsVector::value_type(V, true));
} else
return false;
return true;
if (!allMMOsOkay()) {
return false;
return true;
void ScheduleDAGInstrs::startBlock(MachineBasicBlock *bb) {
BB = bb;
void ScheduleDAGInstrs::finishBlock() {
// Subclasses should no longer refer to the old block.
BB = nullptr;
void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
MachineBasicBlock::iterator begin,
MachineBasicBlock::iterator end,
unsigned regioninstrs) {
assert(bb == BB && "startBlock should set BB");
RegionBegin = begin;
RegionEnd = end;
NumRegionInstrs = regioninstrs;
void ScheduleDAGInstrs::exitRegion() {
// Nothing to do.
void ScheduleDAGInstrs::addSchedBarrierDeps() {
MachineInstr *ExitMI = RegionEnd != BB->end() ? &*RegionEnd : nullptr;
// Add dependencies on the defs and uses of the instruction.
if (ExitMI) {
for (const MachineOperand &MO : ExitMI->operands()) {
if (!MO.isReg() || MO.isDef()) continue;
unsigned Reg = MO.getReg();
if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
Uses.insert(PhysRegSUOper(&ExitSU, -1, Reg));
} else if (TargetRegisterInfo::isVirtualRegister(Reg) && MO.readsReg()) {
addVRegUseDeps(&ExitSU, ExitMI->getOperandNo(&MO));
if (!ExitMI || (!ExitMI->isCall() && !ExitMI->isBarrier())) {
// For others, e.g. fallthrough, conditional branch, assume the exit
// uses all the registers that are livein to the successor blocks.
for (const MachineBasicBlock *Succ : BB->successors()) {
for (const auto &LI : Succ->liveins()) {
if (!Uses.contains(LI.PhysReg))
Uses.insert(PhysRegSUOper(&ExitSU, -1, LI.PhysReg));
/// MO is an operand of SU's instruction that defines a physical register. Adds
/// data dependencies from SU to any uses of the physical register.
void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
const MachineOperand &MO = SU->getInstr()->getOperand(OperIdx);
assert(MO.isDef() && "expect physreg def");
// Ask the target if address-backscheduling is desirable, and if so how much.
const TargetSubtargetInfo &ST = MF.getSubtarget();
for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
Alias.isValid(); ++Alias) {
if (!Uses.contains(*Alias))
for (Reg2SUnitsMap::iterator I = Uses.find(*Alias); I != Uses.end(); ++I) {
SUnit *UseSU = I->SU;
if (UseSU == SU)
// Adjust the dependence latency using operand def/use information,
// then allow the target to perform its own adjustments.
int UseOp = I->OpIdx;
MachineInstr *RegUse = nullptr;
SDep Dep;
if (UseOp < 0)
Dep = SDep(SU, SDep::Artificial);
else {
// Set the hasPhysRegDefs only for physreg defs that have a use within
// the scheduling region.
SU->hasPhysRegDefs = true;
Dep = SDep(SU, SDep::Data, *Alias);
RegUse = UseSU->getInstr();
SchedModel.computeOperandLatency(SU->getInstr(), OperIdx, RegUse,
ST.adjustSchedDependency(SU, UseSU, Dep);
/// Adds register dependencies (data, anti, and output) from this SUnit
/// to following instructions in the same scheduling region that depend the
/// physical register referenced at OperIdx.
void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
MachineInstr *MI = SU->getInstr();
MachineOperand &MO = MI->getOperand(OperIdx);
unsigned Reg = MO.getReg();
// We do not need to track any dependencies for constant registers.
if (MRI.isConstantPhysReg(Reg))
// Optionally add output and anti dependencies. For anti
// dependencies we use a latency of 0 because for a multi-issue
// target we want to allow the defining instruction to issue
// in the same cycle as the using instruction.
// TODO: Using a latency of 1 here for output dependencies assumes
// there's no cost for reusing registers.
SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
for (MCRegAliasIterator Alias(Reg, TRI, true); Alias.isValid(); ++Alias) {
if (!Defs.contains(*Alias))
for (Reg2SUnitsMap::iterator I = Defs.find(*Alias); I != Defs.end(); ++I) {
SUnit *DefSU = I->SU;
if (DefSU == &ExitSU)
if (DefSU != SU &&
(Kind != SDep::Output || !MO.isDead() ||
!DefSU->getInstr()->registerDefIsDead(*Alias))) {
if (Kind == SDep::Anti)
DefSU->addPred(SDep(SU, Kind, /*Reg=*/*Alias));
else {
SDep Dep(SU, Kind, /*Reg=*/*Alias);
SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
if (!MO.isDef()) {
SU->hasPhysRegUses = true;
// Either insert a new Reg2SUnits entry with an empty SUnits list, or
// retrieve the existing SUnits list for this register's uses.
// Push this SUnit on the use list.
Uses.insert(PhysRegSUOper(SU, OperIdx, Reg));
if (RemoveKillFlags)
} else {
addPhysRegDataDeps(SU, OperIdx);
// Clear previous uses and defs of this register and its subergisters.
for (MCSubRegIterator SubReg(Reg, TRI, true); SubReg.isValid(); ++SubReg) {
if (Uses.contains(*SubReg))
if (!MO.isDead())
if (MO.isDead() && SU->isCall) {
// Calls will not be reordered because of chain dependencies (see
// below). Since call operands are dead, calls may continue to be added
// to the DefList making dependence checking quadratic in the size of
// the block. Instead, we leave only one call at the back of the
// DefList.
Reg2SUnitsMap::RangePair P = Defs.equal_range(Reg);
Reg2SUnitsMap::iterator B = P.first;
Reg2SUnitsMap::iterator I = P.second;
for (bool isBegin = I == B; !isBegin; /* empty */) {
isBegin = (--I) == B;
if (!I->SU->isCall)
I = Defs.erase(I);
// Defs are pushed in the order they are visited and never reordered.
Defs.insert(PhysRegSUOper(SU, OperIdx, Reg));
LaneBitmask ScheduleDAGInstrs::getLaneMaskForMO(const MachineOperand &MO) const
unsigned Reg = MO.getReg();
// No point in tracking lanemasks if we don't have interesting subregisters.
const TargetRegisterClass &RC = *MRI.getRegClass(Reg);
if (!RC.HasDisjunctSubRegs)
return LaneBitmask::getAll();
unsigned SubReg = MO.getSubReg();
if (SubReg == 0)
return RC.getLaneMask();
return TRI->getSubRegIndexLaneMask(SubReg);
/// Adds register output and data dependencies from this SUnit to instructions
/// that occur later in the same scheduling region if they read from or write to
/// the virtual register defined at OperIdx.
/// TODO: Hoist loop induction variable increments. This has to be
/// reevaluated. Generally, IV scheduling should be done before coalescing.
void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
MachineInstr *MI = SU->getInstr();
MachineOperand &MO = MI->getOperand(OperIdx);
unsigned Reg = MO.getReg();
LaneBitmask DefLaneMask;
LaneBitmask KillLaneMask;
if (TrackLaneMasks) {
bool IsKill = MO.getSubReg() == 0 || MO.isUndef();
DefLaneMask = getLaneMaskForMO(MO);
// If we have a <read-undef> flag, none of the lane values comes from an
// earlier instruction.
KillLaneMask = IsKill ? LaneBitmask::getAll() : DefLaneMask;
// Clear undef flag, we'll re-add it later once we know which subregister
// Def is first.
} else {
DefLaneMask = LaneBitmask::getAll();
KillLaneMask = LaneBitmask::getAll();
if (MO.isDead()) {
assert(CurrentVRegUses.find(Reg) == CurrentVRegUses.end() &&
"Dead defs should have no uses");
} else {
// Add data dependence to all uses we found so far.
const TargetSubtargetInfo &ST = MF.getSubtarget();
for (VReg2SUnitOperIdxMultiMap::iterator I = CurrentVRegUses.find(Reg),
E = CurrentVRegUses.end(); I != E; /*empty*/) {
LaneBitmask LaneMask = I->LaneMask;
// Ignore uses of other lanes.
if ((LaneMask & KillLaneMask).none()) {
if ((LaneMask & DefLaneMask).any()) {
SUnit *UseSU = I->SU;
MachineInstr *Use = UseSU->getInstr();
SDep Dep(SU, SDep::Data, Reg);
Dep.setLatency(SchedModel.computeOperandLatency(MI, OperIdx, Use,
ST.adjustSchedDependency(SU, UseSU, Dep);
LaneMask &= ~KillLaneMask;
// If we found a Def for all lanes of this use, remove it from the list.
if (LaneMask.any()) {
I->LaneMask = LaneMask;
} else
I = CurrentVRegUses.erase(I);
// Shortcut: Singly defined vregs do not have output/anti dependencies.
if (MRI.hasOneDef(Reg))
// Add output dependence to the next nearest defs of this vreg.
// Unless this definition is dead, the output dependence should be
// transitively redundant with antidependencies from this definition's
// uses. We're conservative for now until we have a way to guarantee the uses
// are not eliminated sometime during scheduling. The output dependence edge
// is also useful if output latency exceeds def-use latency.
LaneBitmask LaneMask = DefLaneMask;
for (VReg2SUnit &V2SU : make_range(CurrentVRegDefs.find(Reg),
CurrentVRegDefs.end())) {
// Ignore defs for other lanes.
if ((V2SU.LaneMask & LaneMask).none())
// Add an output dependence.
SUnit *DefSU = V2SU.SU;
// Ignore additional defs of the same lanes in one instruction. This can
// happen because lanemasks are shared for targets with too many
// subregisters. We also use some representration tricks/hacks where we
// add super-register defs/uses, to imply that although we only access parts
// of the reg we care about the full one.
if (DefSU == SU)
SDep Dep(SU, SDep::Output, Reg);
SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
// Update current definition. This can get tricky if the def was about a
// bigger lanemask before. We then have to shrink it and create a new
// VReg2SUnit for the non-overlapping part.
LaneBitmask OverlapMask = V2SU.LaneMask & LaneMask;
LaneBitmask NonOverlapMask = V2SU.LaneMask & ~LaneMask;
V2SU.LaneMask = OverlapMask;
if (NonOverlapMask.any())
CurrentVRegDefs.insert(VReg2SUnit(Reg, NonOverlapMask, DefSU));
// If there was no CurrentVRegDefs entry for some lanes yet, create one.
if (LaneMask.any())
CurrentVRegDefs.insert(VReg2SUnit(Reg, LaneMask, SU));
/// Adds a register data dependency if the instruction that defines the
/// virtual register used at OperIdx is mapped to an SUnit. Add a register
/// antidependency from this SUnit to instructions that occur later in the same
/// scheduling region if they write the virtual register.
/// TODO: Handle ExitSU "uses" properly.
void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
const MachineInstr *MI = SU->getInstr();
const MachineOperand &MO = MI->getOperand(OperIdx);
unsigned Reg = MO.getReg();
// Remember the use. Data dependencies will be added when we find the def.
LaneBitmask LaneMask = TrackLaneMasks ? getLaneMaskForMO(MO)
: LaneBitmask::getAll();
CurrentVRegUses.insert(VReg2SUnitOperIdx(Reg, LaneMask, OperIdx, SU));
// Add antidependences to the following defs of the vreg.
for (VReg2SUnit &V2SU : make_range(CurrentVRegDefs.find(Reg),
CurrentVRegDefs.end())) {
// Ignore defs for unrelated lanes.
LaneBitmask PrevDefLaneMask = V2SU.LaneMask;
if ((PrevDefLaneMask & LaneMask).none())
if (V2SU.SU == SU)
V2SU.SU->addPred(SDep(SU, SDep::Anti, Reg));
/// Returns true if MI is an instruction we are unable to reason about
/// (like a call or something with unmodeled side effects).
static inline bool isGlobalMemoryObject(AliasAnalysis *AA, MachineInstr *MI) {
return MI->isCall() || MI->hasUnmodeledSideEffects() ||
(MI->hasOrderedMemoryRef() && !MI->isDereferenceableInvariantLoad(AA));
void ScheduleDAGInstrs::addChainDependency (SUnit *SUa, SUnit *SUb,
unsigned Latency) {
if (SUa->getInstr()->mayAlias(AAForDep, *SUb->getInstr(), UseTBAA)) {
SDep Dep(SUa, SDep::MayAliasMem);
/// Creates an SUnit for each real instruction, numbered in top-down
/// topological order. The instruction order A < B, implies that no edge exists
/// from B to A.
/// Map each real instruction to its SUnit.
/// After initSUnits, the SUnits vector cannot be resized and the scheduler may
/// hang onto SUnit pointers. We may relax this in the future by using SUnit IDs
/// instead of pointers.
/// MachineScheduler relies on initSUnits numbering the nodes by their order in
/// the original instruction list.
void ScheduleDAGInstrs::initSUnits() {
// We'll be allocating one SUnit for each real instruction in the region,
// which is contained within a basic block.
for (MachineInstr &MI : make_range(RegionBegin, RegionEnd)) {
if (MI.isDebugInstr())
SUnit *SU = newSUnit(&MI);
MISUnitMap[&MI] = SU;
SU->isCall = MI.isCall();
SU->isCommutable = MI.isCommutable();
// Assign the Latency field of SU using target-provided information.
SU->Latency = SchedModel.computeInstrLatency(SU->getInstr());
// If this SUnit uses a reserved or unbuffered resource, mark it as such.
// Reserved resources block an instruction from issuing and stall the
// entire pipeline. These are identified by BufferSize=0.
// Unbuffered resources prevent execution of subsequent instructions that
// require the same resources. This is used for in-order execution pipelines
// within an out-of-order core. These are identified by BufferSize=1.
if (SchedModel.hasInstrSchedModel()) {
const MCSchedClassDesc *SC = getSchedClass(SU);
for (const MCWriteProcResEntry &PRE :
SchedModel.getWriteProcResEnd(SC))) {
switch (SchedModel.getProcResource(PRE.ProcResourceIdx)->BufferSize) {
case 0:
SU->hasReservedResource = true;
case 1:
SU->isUnbuffered = true;
class ScheduleDAGInstrs::Value2SUsMap : public MapVector<ValueType, SUList> {
/// Current total number of SUs in map.
unsigned NumNodes = 0;
/// 1 for loads, 0 for stores. (see comment in SUList)
unsigned TrueMemOrderLatency;
Value2SUsMap(unsigned lat = 0) : TrueMemOrderLatency(lat) {}
/// To keep NumNodes up to date, insert() is used instead of
/// this operator w/ push_back().
ValueType &operator[](const SUList &Key) {
llvm_unreachable("Don't use. Use insert() instead."); };
/// Adds SU to the SUList of V. If Map grows huge, reduce its size by calling
/// reduce().
void inline insert(SUnit *SU, ValueType V) {
/// Clears the list of SUs mapped to V.
void inline clearList(ValueType V) {
iterator Itr = find(V);
if (Itr != end()) {
assert(NumNodes >= Itr->second.size());
NumNodes -= Itr->second.size();
/// Clears map from all contents.
void clear() {
MapVector<ValueType, SUList>::clear();
NumNodes = 0;
unsigned inline size() const { return NumNodes; }
/// Counts the number of SUs in this map after a reduction.
void reComputeSize() {
NumNodes = 0;
for (auto &I : *this)
NumNodes += I.second.size();
unsigned inline getTrueMemOrderLatency() const {
return TrueMemOrderLatency;
void dump();
void ScheduleDAGInstrs::addChainDependencies(SUnit *SU,
Value2SUsMap &Val2SUsMap) {
for (auto &I : Val2SUsMap)
addChainDependencies(SU, I.second,
void ScheduleDAGInstrs::addChainDependencies(SUnit *SU,
Value2SUsMap &Val2SUsMap,
ValueType V) {
Value2SUsMap::iterator Itr = Val2SUsMap.find(V);
if (Itr != Val2SUsMap.end())
addChainDependencies(SU, Itr->second,
void ScheduleDAGInstrs::addBarrierChain(Value2SUsMap &map) {
assert(BarrierChain != nullptr);
for (auto &I : map) {
SUList &sus = I.second;
for (auto *SU : sus)
void ScheduleDAGInstrs::insertBarrierChain(Value2SUsMap &map) {
assert(BarrierChain != nullptr);
// Go through all lists of SUs.
for (Value2SUsMap::iterator I = map.begin(), EE = map.end(); I != EE;) {
Value2SUsMap::iterator CurrItr = I++;
SUList &sus = CurrItr->second;
SUList::iterator SUItr = sus.begin(), SUEE = sus.end();
for (; SUItr != SUEE; ++SUItr) {
// Stop on BarrierChain or any instruction above it.
if ((*SUItr)->NodeNum <= BarrierChain->NodeNum)
// Remove also the BarrierChain from list if present.
if (SUItr != SUEE && *SUItr == BarrierChain)
// Remove all SUs that are now successors of BarrierChain.
if (SUItr != sus.begin())
sus.erase(sus.begin(), SUItr);
// Remove all entries with empty su lists.
map.remove_if([&](std::pair<ValueType, SUList> &mapEntry) {
return (mapEntry.second.empty()); });
// Recompute the size of the map (NumNodes).
void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
RegPressureTracker *RPTracker,
PressureDiffs *PDiffs,
LiveIntervals *LIS,
bool TrackLaneMasks) {
const TargetSubtargetInfo &ST = MF.getSubtarget();
bool UseAA = EnableAASchedMI.getNumOccurrences() > 0 ? EnableAASchedMI
: ST.useAA();
AAForDep = UseAA ? AA : nullptr;
BarrierChain = nullptr;
this->TrackLaneMasks = TrackLaneMasks;
// Create an SUnit for each real instruction.
if (PDiffs)
// We build scheduling units by walking a block's instruction list
// from bottom to top.
// Each MIs' memory operand(s) is analyzed to a list of underlying
// objects. The SU is then inserted in the SUList(s) mapped from the
// Value(s). Each Value thus gets mapped to lists of SUs depending
// on it, stores and loads kept separately. Two SUs are trivially
// non-aliasing if they both depend on only identified Values and do
// not share any common Value.
Value2SUsMap Stores, Loads(1 /*TrueMemOrderLatency*/);
// Certain memory accesses are known to not alias any SU in Stores
// or Loads, and have therefore their own 'NonAlias'
// domain. E.g. spill / reload instructions never alias LLVM I/R
// Values. It would be nice to assume that this type of memory
// accesses always have a proper memory operand modelling, and are
// therefore never unanalyzable, but this is conservatively not
// done.
Value2SUsMap NonAliasStores, NonAliasLoads(1 /*TrueMemOrderLatency*/);
// Remove any stale debug info; sometimes BuildSchedGraph is called again
// without emitting the info from the previous call.
FirstDbgValue = nullptr;
assert(Defs.empty() && Uses.empty() &&
"Only BuildGraph should update Defs/Uses");
assert(CurrentVRegDefs.empty() && "nobody else should use CurrentVRegDefs");
assert(CurrentVRegUses.empty() && "nobody else should use CurrentVRegUses");
unsigned NumVirtRegs = MRI.getNumVirtRegs();
// Model data dependencies between instructions being scheduled and the
// ExitSU.
// Walk the list of instructions, from bottom moving up.
MachineInstr *DbgMI = nullptr;
for (MachineBasicBlock::iterator MII = RegionEnd, MIE = RegionBegin;
MII != MIE; --MII) {
MachineInstr &MI = *std::prev(MII);
if (DbgMI) {
DbgValues.push_back(std::make_pair(DbgMI, &MI));
DbgMI = nullptr;
if (MI.isDebugValue()) {
DbgMI = &MI;
if (MI.isDebugLabel())
SUnit *SU = MISUnitMap[&MI];
assert(SU && "No SUnit mapped to this MI");
if (RPTracker) {
RegisterOperands RegOpers;
RegOpers.collect(MI, *TRI, MRI, TrackLaneMasks, false);
if (TrackLaneMasks) {
SlotIndex SlotIdx = LIS->getInstructionIndex(MI);
RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx);
if (PDiffs != nullptr)
PDiffs->addInstruction(SU->NodeNum, RegOpers, MRI);
if (RPTracker->getPos() == RegionEnd || &*RPTracker->getPos() != &MI)
assert(&*RPTracker->getPos() == &MI && "RPTracker in sync");
(CanHandleTerminators || (!MI.isTerminator() && !MI.isPosition())) &&
"Cannot schedule terminators or labels!");
// Add register-based dependencies (data, anti, and output).
// For some instructions (calls, returns, inline-asm, etc.) there can
// be explicit uses and implicit defs, in which case the use will appear
// on the operand list before the def. Do two passes over the operand
// list to make sure that defs are processed before any uses.
bool HasVRegDef = false;
for (unsigned j = 0, n = MI.getNumOperands(); j != n; ++j) {
const MachineOperand &MO = MI.getOperand(j);
if (!MO.isReg() || !MO.isDef())
unsigned Reg = MO.getReg();
if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
addPhysRegDeps(SU, j);
} else if (TargetRegisterInfo::isVirtualRegister(Reg)) {
HasVRegDef = true;
addVRegDefDeps(SU, j);
// Now process all uses.
for (unsigned j = 0, n = MI.getNumOperands(); j != n; ++j) {
const MachineOperand &MO = MI.getOperand(j);
// Only look at use operands.
// We do not need to check for MO.readsReg() here because subsequent
// subregister defs will get output dependence edges and need no
// additional use dependencies.
if (!MO.isReg() || !MO.isUse())
unsigned Reg = MO.getReg();
if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
addPhysRegDeps(SU, j);
} else if (TargetRegisterInfo::isVirtualRegister(Reg) && MO.readsReg()) {
addVRegUseDeps(SU, j);
// If we haven't seen any uses in this scheduling region, create a
// dependence edge to ExitSU to model the live-out latency. This is required
// for vreg defs with no in-region use, and prefetches with no vreg def.
// FIXME: NumDataSuccs would be more precise than NumSuccs here. This
// check currently relies on being called before adding chain deps.
if (SU->NumSuccs == 0 && SU->Latency > 1 && (HasVRegDef || MI.mayLoad())) {
SDep Dep(SU, SDep::Artificial);
Dep.setLatency(SU->Latency - 1);
// Add memory dependencies (Note: isStoreToStackSlot and
// isLoadFromStackSLot are not usable after stack slots are lowered to
// actual addresses).
// This is a barrier event that acts as a pivotal node in the DAG.
if (isGlobalMemoryObject(AA, &MI)) {
// Become the barrier chain.
if (BarrierChain)
BarrierChain = SU;
LLVM_DEBUG(dbgs() << "Global memory object and new barrier chain: SU("
<< BarrierChain->NodeNum << ").\n";);
// Add dependencies against everything below it and clear maps.
// If it's not a store or a variant load, we're done.
if (!MI.mayStore() &&
!(MI.mayLoad() && !MI.isDereferenceableInvariantLoad(AA)))
// Always add dependecy edge to BarrierChain if present.
if (BarrierChain)
// Find the underlying objects for MI. The Objs vector is either
// empty, or filled with the Values of memory locations which this
// SU depends on.
UnderlyingObjectsVector Objs;
bool ObjsFound = getUnderlyingObjectsForInstr(&MI, MFI, Objs,
if (MI.mayStore()) {
if (!ObjsFound) {
// An unknown store depends on all stores and loads.
addChainDependencies(SU, Stores);
addChainDependencies(SU, NonAliasStores);
addChainDependencies(SU, Loads);
addChainDependencies(SU, NonAliasLoads);
// Map this store to 'UnknownValue'.
Stores.insert(SU, UnknownValue);
} else {
// Add precise dependencies against all previously seen memory
// accesses mapped to the same Value(s).
for (const UnderlyingObject &UnderlObj : Objs) {
ValueType V = UnderlObj.getValue();
bool ThisMayAlias = UnderlObj.mayAlias();
// Add dependencies to previous stores and loads mapped to V.
addChainDependencies(SU, (ThisMayAlias ? Stores : NonAliasStores), V);
addChainDependencies(SU, (ThisMayAlias ? Loads : NonAliasLoads), V);
// Update the store map after all chains have been added to avoid adding
// self-loop edge if multiple underlying objects are present.
for (const UnderlyingObject &UnderlObj : Objs) {
ValueType V = UnderlObj.getValue();
bool ThisMayAlias = UnderlObj.mayAlias();
// Map this store to V.
(ThisMayAlias ? Stores : NonAliasStores).insert(SU, V);
// The store may have dependencies to unanalyzable loads and
// stores.
addChainDependencies(SU, Loads, UnknownValue);
addChainDependencies(SU, Stores, UnknownValue);
} else { // SU is a load.
if (!ObjsFound) {
// An unknown load depends on all stores.
addChainDependencies(SU, Stores);
addChainDependencies(SU, NonAliasStores);
Loads.insert(SU, UnknownValue);
} else {
for (const UnderlyingObject &UnderlObj : Objs) {
ValueType V = UnderlObj.getValue();
bool ThisMayAlias = UnderlObj.mayAlias();
// Add precise dependencies against all previously seen stores
// mapping to the same Value(s).
addChainDependencies(SU, (ThisMayAlias ? Stores : NonAliasStores), V);
// Map this load to V.
(ThisMayAlias ? Loads : NonAliasLoads).insert(SU, V);
// The load may have dependencies to unanalyzable stores.
addChainDependencies(SU, Stores, UnknownValue);
// Reduce maps if they grow huge.
if (Stores.size() + Loads.size() >= HugeRegion) {
LLVM_DEBUG(dbgs() << "Reducing Stores and Loads maps.\n";);
reduceHugeMemNodeMaps(Stores, Loads, getReductionSize());
if (NonAliasStores.size() + NonAliasLoads.size() >= HugeRegion) {
dbgs() << "Reducing NonAliasStores and NonAliasLoads maps.\n";);
reduceHugeMemNodeMaps(NonAliasStores, NonAliasLoads, getReductionSize());
if (DbgMI)
FirstDbgValue = DbgMI;
raw_ostream &llvm::operator<<(raw_ostream &OS, const PseudoSourceValue* PSV) {
return OS;
void ScheduleDAGInstrs::Value2SUsMap::dump() {
for (auto &Itr : *this) {
if (<const Value*>()) {
const Value *V = Itr.first.get<const Value*>();
if (isa<UndefValue>(V))
dbgs() << "Unknown";
else if (<const PseudoSourceValue*>())
dbgs() << Itr.first.get<const PseudoSourceValue*>();
llvm_unreachable("Unknown Value type.");
dbgs() << " : ";
void ScheduleDAGInstrs::reduceHugeMemNodeMaps(Value2SUsMap &stores,
Value2SUsMap &loads, unsigned N) {
LLVM_DEBUG(dbgs() << "Before reduction:\nStoring SUnits:\n"; stores.dump();
dbgs() << "Loading SUnits:\n"; loads.dump());
// Insert all SU's NodeNums into a vector and sort it.
std::vector<unsigned> NodeNums;
NodeNums.reserve(stores.size() + loads.size());
for (auto &I : stores)
for (auto *SU : I.second)
for (auto &I : loads)
for (auto *SU : I.second)
llvm::sort(NodeNums.begin(), NodeNums.end());
// The N last elements in NodeNums will be removed, and the SU with
// the lowest NodeNum of them will become the new BarrierChain to
// let the not yet seen SUs have a dependency to the removed SUs.
assert(N <= NodeNums.size());
SUnit *newBarrierChain = &SUnits[*(NodeNums.end() - N)];
if (BarrierChain) {
// The aliasing and non-aliasing maps reduce independently of each
// other, but share a common BarrierChain. Check if the
// newBarrierChain is above the former one. If it is not, it may
// introduce a loop to use newBarrierChain, so keep the old one.
if (newBarrierChain->NodeNum < BarrierChain->NodeNum) {
BarrierChain = newBarrierChain;
LLVM_DEBUG(dbgs() << "Inserting new barrier chain: SU("
<< BarrierChain->NodeNum << ").\n";);
LLVM_DEBUG(dbgs() << "Keeping old barrier chain: SU("
<< BarrierChain->NodeNum << ").\n";);
BarrierChain = newBarrierChain;
LLVM_DEBUG(dbgs() << "After reduction:\nStoring SUnits:\n"; stores.dump();
dbgs() << "Loading SUnits:\n"; loads.dump());
static void toggleKills(const MachineRegisterInfo &MRI, LivePhysRegs &LiveRegs,
MachineInstr &MI, bool addToLiveRegs) {
for (MachineOperand &MO : MI.operands()) {
if (!MO.isReg() || !MO.readsReg())
unsigned Reg = MO.getReg();
if (!Reg)
// Things that are available after the instruction are killed by it.
bool IsKill = LiveRegs.available(MRI, Reg);
if (addToLiveRegs)
void ScheduleDAGInstrs::fixupKills(MachineBasicBlock &MBB) {
LLVM_DEBUG(dbgs() << "Fixup kills for " << printMBBReference(MBB) << '\n');
// Examine block from end to start...
for (MachineInstr &MI : make_range(MBB.rbegin(), MBB.rend())) {
if (MI.isDebugInstr())
// Update liveness. Registers that are defed but not used in this
// instruction are now dead. Mark register and all subregs as they
// are completely defined.
for (ConstMIBundleOperands O(MI); O.isValid(); ++O) {
const MachineOperand &MO = *O;
if (MO.isReg()) {
if (!MO.isDef())
unsigned Reg = MO.getReg();
if (!Reg)
} else if (MO.isRegMask()) {
// If there is a bundle header fix it up first.
if (!MI.isBundled()) {
toggleKills(MRI, LiveRegs, MI, true);
} else {
MachineBasicBlock::instr_iterator First = MI.getIterator();
if (MI.isBundle()) {
toggleKills(MRI, LiveRegs, MI, false);
// Some targets make the (questionable) assumtion that the instructions
// inside the bundle are ordered and consequently only the last use of
// a register inside the bundle can kill it.
MachineBasicBlock::instr_iterator I = std::next(First);
while (I->isBundledWithSucc())
do {
if (!I->isDebugInstr())
toggleKills(MRI, LiveRegs, *I, true);
} while(I != First);
void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const {
// Cannot completely remove virtual function even in release mode.
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
std::string s;
raw_string_ostream oss(s);
if (SU == &EntrySU)
oss << "<entry>";
else if (SU == &ExitSU)
oss << "<exit>";
SU->getInstr()->print(oss, /*SkipOpers=*/true);
return oss.str();
/// Return the basic block label. It is not necessarilly unique because a block
/// contains multiple scheduling regions. But it is fine for visualization.
std::string ScheduleDAGInstrs::getDAGName() const {
return "dag." + BB->getFullName();
// SchedDFSResult Implementation
namespace llvm {
/// Internal state used to compute SchedDFSResult.
class SchedDFSImpl {
SchedDFSResult &R;
/// Join DAG nodes into equivalence classes by their subtree.
IntEqClasses SubtreeClasses;
/// List PredSU, SuccSU pairs that represent data edges between subtrees.
std::vector<std::pair<const SUnit *, const SUnit*>> ConnectionPairs;
struct RootData {
unsigned NodeID;
unsigned ParentNodeID; ///< Parent node (member of the parent subtree).
unsigned SubInstrCount = 0; ///< Instr count in this tree only, not
/// children.
RootData(unsigned id): NodeID(id),
ParentNodeID(SchedDFSResult::InvalidSubtreeID) {}
unsigned getSparseSetIndex() const { return NodeID; }
SparseSet<RootData> RootSet;
SchedDFSImpl(SchedDFSResult &r): R(r), SubtreeClasses(R.DFSNodeData.size()) {
/// Returns true if this node been visited by the DFS traversal.
/// During visitPostorderNode the Node's SubtreeID is assigned to the Node
/// ID. Later, SubtreeID is updated but remains valid.
bool isVisited(const SUnit *SU) const {
return R.DFSNodeData[SU->NodeNum].SubtreeID
!= SchedDFSResult::InvalidSubtreeID;
/// Initializes this node's instruction count. We don't need to flag the node
/// visited until visitPostorder because the DAG cannot have cycles.
void visitPreorder(const SUnit *SU) {
R.DFSNodeData[SU->NodeNum].InstrCount =
SU->getInstr()->isTransient() ? 0 : 1;
/// Called once for each node after all predecessors are visited. Revisit this
/// node's predecessors and potentially join them now that we know the ILP of
/// the other predecessors.
void visitPostorderNode(const SUnit *SU) {
// Mark this node as the root of a subtree. It may be joined with its
// successors later.
R.DFSNodeData[SU->NodeNum].SubtreeID = SU->NodeNum;
RootData RData(SU->NodeNum);
RData.SubInstrCount = SU->getInstr()->isTransient() ? 0 : 1;
// If any predecessors are still in their own subtree, they either cannot be
// joined or are large enough to remain separate. If this parent node's
// total instruction count is not greater than a child subtree by at least
// the subtree limit, then try to join it now since splitting subtrees is
// only useful if multiple high-pressure paths are possible.
unsigned InstrCount = R.DFSNodeData[SU->NodeNum].InstrCount;
for (const SDep &PredDep : SU->Preds) {
if (PredDep.getKind() != SDep::Data)
unsigned PredNum = PredDep.getSUnit()->NodeNum;
if ((InstrCount - R.DFSNodeData[PredNum].InstrCount) < R.SubtreeLimit)
joinPredSubtree(PredDep, SU, /*CheckLimit=*/false);
// Either link or merge the TreeData entry from the child to the parent.
if (R.DFSNodeData[PredNum].SubtreeID == PredNum) {
// If the predecessor's parent is invalid, this is a tree edge and the
// current node is the parent.
if (RootSet[PredNum].ParentNodeID == SchedDFSResult::InvalidSubtreeID)
RootSet[PredNum].ParentNodeID = SU->NodeNum;
else if (RootSet.count(PredNum)) {
// The predecessor is not a root, but is still in the root set. This
// must be the new parent that it was just joined to. Note that
// RootSet[PredNum].ParentNodeID may either be invalid or may still be
// set to the original parent.
RData.SubInstrCount += RootSet[PredNum].SubInstrCount;
RootSet[SU->NodeNum] = RData;
/// Called once for each tree edge after calling visitPostOrderNode on
/// the predecessor. Increment the parent node's instruction count and
/// preemptively join this subtree to its parent's if it is small enough.
void visitPostorderEdge(const SDep &PredDep, const SUnit *Succ) {
+= R.DFSNodeData[PredDep.getSUnit()->NodeNum].InstrCount;
joinPredSubtree(PredDep, Succ);
/// Adds a connection for cross edges.
void visitCrossEdge(const SDep &PredDep, const SUnit *Succ) {
ConnectionPairs.push_back(std::make_pair(PredDep.getSUnit(), Succ));
/// Sets each node's subtree ID to the representative ID and record
/// connections between trees.
void finalize() {
assert(SubtreeClasses.getNumClasses() == RootSet.size()
&& "number of roots should match trees");
for (const RootData &Root : RootSet) {
unsigned TreeID = SubtreeClasses[Root.NodeID];
if (Root.ParentNodeID != SchedDFSResult::InvalidSubtreeID)
R.DFSTreeData[TreeID].ParentTreeID = SubtreeClasses[Root.ParentNodeID];
R.DFSTreeData[TreeID].SubInstrCount = Root.SubInstrCount;
// Note that SubInstrCount may be greater than InstrCount if we joined
// subtrees across a cross edge. InstrCount will be attributed to the
// original parent, while SubInstrCount will be attributed to the joined
// parent.
LLVM_DEBUG(dbgs() << R.getNumSubtrees() << " subtrees:\n");
for (unsigned Idx = 0, End = R.DFSNodeData.size(); Idx != End; ++Idx) {
R.DFSNodeData[Idx].SubtreeID = SubtreeClasses[Idx];
LLVM_DEBUG(dbgs() << " SU(" << Idx << ") in tree "
<< R.DFSNodeData[Idx].SubtreeID << '\n');
for (const std::pair<const SUnit*, const SUnit*> &P : ConnectionPairs) {
unsigned PredTree = SubtreeClasses[P.first->NodeNum];
unsigned SuccTree = SubtreeClasses[P.second->NodeNum];
if (PredTree == SuccTree)
unsigned Depth = P.first->getDepth();
addConnection(PredTree, SuccTree, Depth);
addConnection(SuccTree, PredTree, Depth);
/// Joins the predecessor subtree with the successor that is its DFS parent.
/// Applies some heuristics before joining.
bool joinPredSubtree(const SDep &PredDep, const SUnit *Succ,
bool CheckLimit = true) {
assert(PredDep.getKind() == SDep::Data && "Subtrees are for data edges");
// Check if the predecessor is already joined.
const SUnit *PredSU = PredDep.getSUnit();
unsigned PredNum = PredSU->NodeNum;
if (R.DFSNodeData[PredNum].SubtreeID != PredNum)
return false;
// Four is the magic number of successors before a node is considered a
// pinch point.
unsigned NumDataSucs = 0;
for (const SDep &SuccDep : PredSU->Succs) {
if (SuccDep.getKind() == SDep::Data) {
if (++NumDataSucs >= 4)
return false;
if (CheckLimit && R.DFSNodeData[PredNum].InstrCount > R.SubtreeLimit)
return false;
R.DFSNodeData[PredNum].SubtreeID = Succ->NodeNum;
SubtreeClasses.join(Succ->NodeNum, PredNum);
return true;
/// Called by finalize() to record a connection between trees.
void addConnection(unsigned FromTree, unsigned ToTree, unsigned Depth) {
if (!Depth)
do {
SmallVectorImpl<SchedDFSResult::Connection> &Connections =
for (SchedDFSResult::Connection &C : Connections) {
if (C.TreeID == ToTree) {
C.Level = std::max(C.Level, Depth);
Connections.push_back(SchedDFSResult::Connection(ToTree, Depth));
FromTree = R.DFSTreeData[FromTree].ParentTreeID;
} while (FromTree != SchedDFSResult::InvalidSubtreeID);
} // end namespace llvm
namespace {
/// Manage the stack used by a reverse depth-first search over the DAG.
class SchedDAGReverseDFS {
std::vector<std::pair<const SUnit *, SUnit::const_pred_iterator>> DFSStack;
bool isComplete() const { return DFSStack.empty(); }
void follow(const SUnit *SU) {
DFSStack.push_back(std::make_pair(SU, SU->Preds.begin()));
void advance() { ++DFSStack.back().second; }
const SDep *backtrack() {
return DFSStack.empty() ? nullptr : std::prev(DFSStack.back().second);
const SUnit *getCurr() const { return DFSStack.back().first; }
SUnit::const_pred_iterator getPred() const { return DFSStack.back().second; }
SUnit::const_pred_iterator getPredEnd() const {
return getCurr()->Preds.end();
} // end anonymous namespace
static bool hasDataSucc(const SUnit *SU) {
for (const SDep &SuccDep : SU->Succs) {
if (SuccDep.getKind() == SDep::Data &&
return true;
return false;
/// Computes an ILP metric for all nodes in the subDAG reachable via depth-first
/// search from this root.
void SchedDFSResult::compute(ArrayRef<SUnit> SUnits) {
if (!IsBottomUp)
llvm_unreachable("Top-down ILP metric is unimplemented");
SchedDFSImpl Impl(*this);
for (const SUnit &SU : SUnits) {
if (Impl.isVisited(&SU) || hasDataSucc(&SU))
SchedDAGReverseDFS DFS;
while (true) {
// Traverse the leftmost path as far as possible.
while (DFS.getPred() != DFS.getPredEnd()) {
const SDep &PredDep = *DFS.getPred();
// Ignore non-data edges.
if (PredDep.getKind() != SDep::Data
|| PredDep.getSUnit()->isBoundaryNode()) {
// An already visited edge is a cross edge, assuming an acyclic DAG.
if (Impl.isVisited(PredDep.getSUnit())) {
Impl.visitCrossEdge(PredDep, DFS.getCurr());
// Visit the top of the stack in postorder and backtrack.
const SUnit *Child = DFS.getCurr();
const SDep *PredDep = DFS.backtrack();
if (PredDep)
Impl.visitPostorderEdge(*PredDep, DFS.getCurr());
if (DFS.isComplete())
/// The root of the given SubtreeID was just scheduled. For all subtrees
/// connected to this tree, record the depth of the connection so that the
/// nearest connected subtrees can be prioritized.
void SchedDFSResult::scheduleTree(unsigned SubtreeID) {
for (const Connection &C : SubtreeConnections[SubtreeID]) {
SubtreeConnectLevels[C.TreeID] =
std::max(SubtreeConnectLevels[C.TreeID], C.Level);
LLVM_DEBUG(dbgs() << " Tree: " << C.TreeID << " @"
<< SubtreeConnectLevels[C.TreeID] << '\n');
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
LLVM_DUMP_METHOD void ILPValue::print(raw_ostream &OS) const {
OS << InstrCount << " / " << Length << " = ";
if (!Length)
OS << format("%g", ((double)InstrCount / Length));
LLVM_DUMP_METHOD void ILPValue::dump() const {
dbgs() << *this << '\n';
namespace llvm {
raw_ostream &operator<<(raw_ostream &OS, const ILPValue &Val) {
return OS;
} // end namespace llvm