//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "ifcvt"
-#include "BranchFolding.h"
-#include "llvm/Function.h"
#include "llvm/CodeGen/Passes.h"
-#include "llvm/CodeGen/MachineModuleInfo.h"
+#include "BranchFolding.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/LivePhysRegs.h"
+#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
+#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
-#include "llvm/Target/TargetInstrInfo.h"
-#include "llvm/Target/TargetLowering.h"
-#include "llvm/Target/TargetMachine.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineModuleInfo.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/TargetSchedule.h"
+#include "llvm/MC/MCInstrItineraries.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
-#include "llvm/ADT/DepthFirstIterator.h"
-#include "llvm/ADT/Statistic.h"
-#include "llvm/ADT/STLExtras.h"
+#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Target/TargetLowering.h"
+#include "llvm/Target/TargetRegisterInfo.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
+
using namespace llvm;
+#define DEBUG_TYPE "ifcvt"
+
// Hidden options for help debugging.
static cl::opt<int> IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden);
static cl::opt<int> IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden);
static cl::opt<int> IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden);
-static cl::opt<bool> DisableSimple("disable-ifcvt-simple",
+static cl::opt<bool> DisableSimple("disable-ifcvt-simple",
cl::init(false), cl::Hidden);
-static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false",
+static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false",
cl::init(false), cl::Hidden);
-static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle",
+static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle",
cl::init(false), cl::Hidden);
-static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev",
+static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev",
cl::init(false), cl::Hidden);
-static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false",
+static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false",
cl::init(false), cl::Hidden);
-static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev",
+static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev",
cl::init(false), cl::Hidden);
-static cl::opt<bool> DisableDiamond("disable-ifcvt-diamond",
+static cl::opt<bool> DisableDiamond("disable-ifcvt-diamond",
cl::init(false), cl::Hidden);
+static cl::opt<bool> IfCvtBranchFold("ifcvt-branch-fold",
+ cl::init(true), cl::Hidden);
STATISTIC(NumSimple, "Number of simple if-conversions performed");
STATISTIC(NumSimpleFalse, "Number of simple (F) if-conversions performed");
STATISTIC(NumDiamonds, "Number of diamond if-conversions performed");
STATISTIC(NumIfConvBBs, "Number of if-converted blocks");
STATISTIC(NumDupBBs, "Number of duplicated blocks");
+STATISTIC(NumUnpred, "Number of true blocks of diamonds unpredicated");
namespace {
class IfConverter : public MachineFunctionPass {
/// ClobbersPred - True if BB could modify predicates (e.g. has
/// cmp, call, etc.)
/// NonPredSize - Number of non-predicated instructions.
+ /// ExtraCost - Extra cost for multi-cycle instructions.
+ /// ExtraCost2 - Some instructions are slower when predicated
/// BB - Corresponding MachineBasicBlock.
/// TrueBB / FalseBB- See AnalyzeBranch().
/// BrCond - Conditions for end of block conditional branches.
bool CannotBeCopied : 1;
bool ClobbersPred : 1;
unsigned NonPredSize;
+ unsigned ExtraCost;
+ unsigned ExtraCost2;
MachineBasicBlock *BB;
MachineBasicBlock *TrueBB;
MachineBasicBlock *FalseBB;
IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false),
HasFallThrough(false), IsUnpredicable(false),
CannotBeCopied(false), ClobbersPred(false), NonPredSize(0),
- BB(0), TrueBB(0), FalseBB(0) {}
+ ExtraCost(0), ExtraCost2(0), BB(nullptr), TrueBB(nullptr),
+ FalseBB(nullptr) {}
};
- /// IfcvtToken - Record information about pending if-conversions to attemp:
+ /// IfcvtToken - Record information about pending if-conversions to attempt:
/// BBI - Corresponding BBInfo.
/// Kind - Type of block. See IfcvtKind.
/// NeedSubsumption - True if the to-be-predicated BB has already been
: BBI(b), Kind(k), NeedSubsumption(s), NumDups(d), NumDups2(d2) {}
};
- /// Roots - Basic blocks that do not have successors. These are the starting
- /// points of Graph traversal.
- std::vector<MachineBasicBlock*> Roots;
-
/// BBAnalysis - Results of if-conversion feasibility analysis indexed by
/// basic block number.
std::vector<BBInfo> BBAnalysis;
+ TargetSchedModel SchedModel;
- const TargetLowering *TLI;
+ const TargetLoweringBase *TLI;
const TargetInstrInfo *TII;
+ const TargetRegisterInfo *TRI;
+ const MachineBlockFrequencyInfo *MBFI;
+ const MachineBranchProbabilityInfo *MBPI;
+ MachineRegisterInfo *MRI;
+
+ LivePhysRegs Redefs;
+ LivePhysRegs DontKill;
+
+ bool PreRegAlloc;
bool MadeChange;
int FnNum;
public:
static char ID;
- IfConverter() : MachineFunctionPass(&ID), FnNum(-1) {}
+ IfConverter() : MachineFunctionPass(ID), FnNum(-1) {
+ initializeIfConverterPass(*PassRegistry::getPassRegistry());
+ }
+
+ void getAnalysisUsage(AnalysisUsage &AU) const override {
+ AU.addRequired<MachineBlockFrequencyInfo>();
+ AU.addRequired<MachineBranchProbabilityInfo>();
+ MachineFunctionPass::getAnalysisUsage(AU);
+ }
- virtual bool runOnMachineFunction(MachineFunction &MF);
- virtual const char *getPassName() const { return "If Converter"; }
+ bool runOnMachineFunction(MachineFunction &MF) override;
private:
bool ReverseBranchCondition(BBInfo &BBI);
- bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const;
+ bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups,
+ const BranchProbability &Prediction) const;
bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI,
- bool FalseBranch, unsigned &Dups) const;
+ bool FalseBranch, unsigned &Dups,
+ const BranchProbability &Prediction) const;
bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI,
unsigned &Dups1, unsigned &Dups2) const;
void ScanInstructions(BBInfo &BBI);
std::vector<IfcvtToken*> &Tokens);
bool FeasibilityAnalysis(BBInfo &BBI, SmallVectorImpl<MachineOperand> &Cond,
bool isTriangle = false, bool RevBranch = false);
- bool AnalyzeBlocks(MachineFunction &MF,
- std::vector<IfcvtToken*> &Tokens);
+ void AnalyzeBlocks(MachineFunction &MF, std::vector<IfcvtToken*> &Tokens);
void InvalidatePreds(MachineBasicBlock *BB);
void RemoveExtraEdges(BBInfo &BBI);
bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind);
unsigned NumDups1, unsigned NumDups2);
void PredicateBlock(BBInfo &BBI,
MachineBasicBlock::iterator E,
- SmallVectorImpl<MachineOperand> &Cond);
+ SmallVectorImpl<MachineOperand> &Cond,
+ SmallSet<unsigned, 4> *LaterRedefs = nullptr);
void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
SmallVectorImpl<MachineOperand> &Cond,
bool IgnoreBr = false);
- void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI);
+ void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges = true);
+
+ bool MeetIfcvtSizeLimit(MachineBasicBlock &BB,
+ unsigned Cycle, unsigned Extra,
+ const BranchProbability &Prediction) const {
+ return Cycle > 0 && TII->isProfitableToIfCvt(BB, Cycle, Extra,
+ Prediction);
+ }
- bool MeetIfcvtSizeLimit(unsigned Size) const {
- return Size > 0 && Size <= TLI->getIfCvtBlockSizeLimit();
+ bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB,
+ unsigned TCycle, unsigned TExtra,
+ MachineBasicBlock &FBB,
+ unsigned FCycle, unsigned FExtra,
+ const BranchProbability &Prediction) const {
+ return TCycle > 0 && FCycle > 0 &&
+ TII->isProfitableToIfCvt(TBB, TCycle, TExtra, FBB, FCycle, FExtra,
+ Prediction);
}
// blockAlwaysFallThrough - Block ends without a terminator.
bool blockAlwaysFallThrough(BBInfo &BBI) const {
- return BBI.IsBrAnalyzable && BBI.TrueBB == NULL;
+ return BBI.IsBrAnalyzable && BBI.TrueBB == nullptr;
}
// IfcvtTokenCmp - Used to sort if-conversion candidates.
char IfConverter::ID = 0;
}
-static RegisterPass<IfConverter>
-X("if-converter", "If Converter");
+char &llvm::IfConverterID = IfConverter::ID;
-FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); }
+INITIALIZE_PASS_BEGIN(IfConverter, "if-converter", "If Converter", false, false)
+INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
+INITIALIZE_PASS_END(IfConverter, "if-converter", "If Converter", false, false)
bool IfConverter::runOnMachineFunction(MachineFunction &MF) {
- TLI = MF.getTarget().getTargetLowering();
- TII = MF.getTarget().getInstrInfo();
+ const TargetSubtargetInfo &ST = MF.getSubtarget();
+ TLI = ST.getTargetLowering();
+ TII = ST.getInstrInfo();
+ TRI = ST.getRegisterInfo();
+ MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
+ MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
+ MRI = &MF.getRegInfo();
+ SchedModel.init(ST.getSchedModel(), &ST, TII);
+
if (!TII) return false;
- DEBUG(errs() << "\nIfcvt: function (" << ++FnNum << ") \'"
- << MF.getFunction()->getName() << "\'");
+ PreRegAlloc = MRI->isSSA();
+
+ bool BFChange = false;
+ if (!PreRegAlloc) {
+ // Tail merge tend to expose more if-conversion opportunities.
+ BranchFolder BF(true, false, *MBFI, *MBPI);
+ BFChange = BF.OptimizeFunction(MF, TII, ST.getRegisterInfo(),
+ getAnalysisIfAvailable<MachineModuleInfo>());
+ }
+
+ DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum << ") \'"
+ << MF.getName() << "\'");
if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) {
- DEBUG(errs() << " skipped\n");
+ DEBUG(dbgs() << " skipped\n");
return false;
}
- DEBUG(errs() << "\n");
+ DEBUG(dbgs() << "\n");
MF.RenumberBlocks();
BBAnalysis.resize(MF.getNumBlockIDs());
- // Look for root nodes, i.e. blocks without successors.
- for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
- if (I->succ_empty())
- Roots.push_back(I);
-
std::vector<IfcvtToken*> Tokens;
MadeChange = false;
unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle +
while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) {
// Do an initial analysis for each basic block and find all the potential
// candidates to perform if-conversion.
- bool Change = AnalyzeBlocks(MF, Tokens);
+ bool Change = false;
+ AnalyzeBlocks(MF, Tokens);
while (!Tokens.empty()) {
IfcvtToken *Token = Tokens.back();
Tokens.pop_back();
bool RetVal = false;
switch (Kind) {
- default: assert(false && "Unexpected!");
- break;
+ default: llvm_unreachable("Unexpected!");
case ICSimple:
case ICSimpleFalse: {
bool isFalse = Kind == ICSimpleFalse;
if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break;
- DEBUG(errs() << "Ifcvt (Simple" << (Kind == ICSimpleFalse ? " false" :"")
+ DEBUG(dbgs() << "Ifcvt (Simple" << (Kind == ICSimpleFalse ?
+ " false" : "")
<< "): BB#" << BBI.BB->getNumber() << " ("
<< ((Kind == ICSimpleFalse)
? BBI.FalseBB->getNumber()
: BBI.TrueBB->getNumber()) << ") ");
RetVal = IfConvertSimple(BBI, Kind);
- DEBUG(errs() << (RetVal ? "succeeded!" : "failed!") << "\n");
+ DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n");
if (RetVal) {
- if (isFalse) NumSimpleFalse++;
- else NumSimple++;
+ if (isFalse) ++NumSimpleFalse;
+ else ++NumSimple;
}
break;
}
if (DisableTriangleR && !isFalse && isRev) break;
if (DisableTriangleF && isFalse && !isRev) break;
if (DisableTriangleFR && isFalse && isRev) break;
- DEBUG(errs() << "Ifcvt (Triangle");
+ DEBUG(dbgs() << "Ifcvt (Triangle");
if (isFalse)
- DEBUG(errs() << " false");
+ DEBUG(dbgs() << " false");
if (isRev)
- DEBUG(errs() << " rev");
- DEBUG(errs() << "): BB#" << BBI.BB->getNumber() << " (T:"
+ DEBUG(dbgs() << " rev");
+ DEBUG(dbgs() << "): BB#" << BBI.BB->getNumber() << " (T:"
<< BBI.TrueBB->getNumber() << ",F:"
<< BBI.FalseBB->getNumber() << ") ");
RetVal = IfConvertTriangle(BBI, Kind);
- DEBUG(errs() << (RetVal ? "succeeded!" : "failed!") << "\n");
+ DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n");
if (RetVal) {
if (isFalse) {
- if (isRev) NumTriangleFRev++;
- else NumTriangleFalse++;
+ if (isRev) ++NumTriangleFRev;
+ else ++NumTriangleFalse;
} else {
- if (isRev) NumTriangleRev++;
- else NumTriangle++;
+ if (isRev) ++NumTriangleRev;
+ else ++NumTriangle;
}
}
break;
}
case ICDiamond: {
if (DisableDiamond) break;
- DEBUG(errs() << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:"
+ DEBUG(dbgs() << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:"
<< BBI.TrueBB->getNumber() << ",F:"
<< BBI.FalseBB->getNumber() << ") ");
RetVal = IfConvertDiamond(BBI, Kind, NumDups, NumDups2);
- DEBUG(errs() << (RetVal ? "succeeded!" : "failed!") << "\n");
- if (RetVal) NumDiamonds++;
+ DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n");
+ if (RetVal) ++NumDiamonds;
break;
}
}
}
Tokens.clear();
- Roots.clear();
BBAnalysis.clear();
- if (MadeChange) {
- BranchFolder BF(false);
- BF.OptimizeFunction(MF, TII,
- MF.getTarget().getRegisterInfo(),
+ if (MadeChange && IfCvtBranchFold) {
+ BranchFolder BF(false, false, *MBFI, *MBPI);
+ BF.OptimizeFunction(MF, TII, MF.getSubtarget().getRegisterInfo(),
getAnalysisIfAvailable<MachineModuleInfo>());
}
+ MadeChange |= BFChange;
return MadeChange;
}
if (SuccBB != TrueBB)
return SuccBB;
}
- return NULL;
+ return nullptr;
}
/// ReverseBranchCondition - Reverse the condition of the end of the block
/// branch. Swap block's 'true' and 'false' successors.
bool IfConverter::ReverseBranchCondition(BBInfo &BBI) {
+ DebugLoc dl; // FIXME: this is nowhere
if (!TII->ReverseBranchCondition(BBI.BrCond)) {
TII->RemoveBranch(*BBI.BB);
- TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond);
+ TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond, dl);
std::swap(BBI.TrueBB, BBI.FalseBB);
return true;
}
MachineFunction::iterator I = BB;
MachineFunction::iterator E = BB->getParent()->end();
if (++I == E)
- return NULL;
+ return nullptr;
return I;
}
/// predecessor) forms a valid simple shape for ifcvt. It also returns the
/// number of instructions that the ifcvt would need to duplicate if performed
/// in Dups.
-bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const {
+bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups,
+ const BranchProbability &Prediction) const {
Dups = 0;
if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
return false;
if (TrueBBI.BB->pred_size() > 1) {
if (TrueBBI.CannotBeCopied ||
- TrueBBI.NonPredSize > TLI->getIfCvtDupBlockSizeLimit())
+ !TII->isProfitableToDupForIfCvt(*TrueBBI.BB, TrueBBI.NonPredSize,
+ Prediction))
return false;
Dups = TrueBBI.NonPredSize;
}
/// ValidTriangle - Returns true if the 'true' and 'false' blocks (along
/// with their common predecessor) forms a valid triangle shape for ifcvt.
/// If 'FalseBranch' is true, it checks if 'true' block's false branch
-/// branches to the false branch rather than the other way around. It also
+/// branches to the 'false' block rather than the other way around. It also
/// returns the number of instructions that the ifcvt would need to duplicate
/// if performed in 'Dups'.
bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI,
- bool FalseBranch, unsigned &Dups) const {
+ bool FalseBranch, unsigned &Dups,
+ const BranchProbability &Prediction) const {
Dups = 0;
if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
return false;
++Size;
}
}
- if (Size > TLI->getIfCvtDupBlockSizeLimit())
+ if (!TII->isProfitableToDupForIfCvt(*TrueBBI.BB, Size, Prediction))
return false;
Dups = Size;
}
return TExit && TExit == FalseBBI.BB;
}
-static
-MachineBasicBlock::iterator firstNonBranchInst(MachineBasicBlock *BB,
- const TargetInstrInfo *TII) {
- MachineBasicBlock::iterator I = BB->end();
- while (I != BB->begin()) {
- --I;
- if (!I->getDesc().isBranch())
- break;
- }
- return I;
-}
-
/// ValidDiamond - Returns true if the 'true' and 'false' blocks (along
/// with their common predecessor) forms a valid diamond shape for ifcvt.
bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI,
FT = getNextBlock(FalseBBI.BB);
if (TT != FT)
return false;
- if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable))
+ if (!TT && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable))
return false;
if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1)
return false;
(TrueBBI.ClobbersPred && FalseBBI.ClobbersPred))
return false;
- MachineBasicBlock::iterator TI = TrueBBI.BB->begin();
- MachineBasicBlock::iterator FI = FalseBBI.BB->begin();
- while (TI != TrueBBI.BB->end() && FI != FalseBBI.BB->end()) {
- if (!TI->isIdenticalTo(FI))
+ // Count duplicate instructions at the beginning of the true and false blocks.
+ MachineBasicBlock::iterator TIB = TrueBBI.BB->begin();
+ MachineBasicBlock::iterator FIB = FalseBBI.BB->begin();
+ MachineBasicBlock::iterator TIE = TrueBBI.BB->end();
+ MachineBasicBlock::iterator FIE = FalseBBI.BB->end();
+ while (TIB != TIE && FIB != FIE) {
+ // Skip dbg_value instructions. These do not count.
+ if (TIB->isDebugValue()) {
+ while (TIB != TIE && TIB->isDebugValue())
+ ++TIB;
+ if (TIB == TIE)
+ break;
+ }
+ if (FIB->isDebugValue()) {
+ while (FIB != FIE && FIB->isDebugValue())
+ ++FIB;
+ if (FIB == FIE)
+ break;
+ }
+ if (!TIB->isIdenticalTo(FIB))
break;
++Dups1;
- ++TI;
- ++FI;
+ ++TIB;
+ ++FIB;
}
- TI = firstNonBranchInst(TrueBBI.BB, TII);
- FI = firstNonBranchInst(FalseBBI.BB, TII);
- while (TI != TrueBBI.BB->begin() && FI != FalseBBI.BB->begin()) {
- if (!TI->isIdenticalTo(FI))
+ // Now, in preparation for counting duplicate instructions at the ends of the
+ // blocks, move the end iterators up past any branch instructions.
+ while (TIE != TIB) {
+ --TIE;
+ if (!TIE->isBranch())
+ break;
+ }
+ while (FIE != FIB) {
+ --FIE;
+ if (!FIE->isBranch())
+ break;
+ }
+
+ // If Dups1 includes all of a block, then don't count duplicate
+ // instructions at the end of the blocks.
+ if (TIB == TIE || FIB == FIE)
+ return true;
+
+ // Count duplicate instructions at the ends of the blocks.
+ while (TIE != TIB && FIE != FIB) {
+ // Skip dbg_value instructions. These do not count.
+ if (TIE->isDebugValue()) {
+ while (TIE != TIB && TIE->isDebugValue())
+ --TIE;
+ if (TIE == TIB)
+ break;
+ }
+ if (FIE->isDebugValue()) {
+ while (FIE != FIB && FIE->isDebugValue())
+ --FIE;
+ if (FIE == FIB)
+ break;
+ }
+ if (!TIE->isIdenticalTo(FIE))
break;
++Dups2;
- --TI;
- --FI;
+ --TIE;
+ --FIE;
}
return true;
if (BBI.IsDone)
return;
- bool AlreadyPredicated = BBI.Predicate.size() > 0;
+ bool AlreadyPredicated = !BBI.Predicate.empty();
// First analyze the end of BB branches.
- BBI.TrueBB = BBI.FalseBB = NULL;
+ BBI.TrueBB = BBI.FalseBB = nullptr;
BBI.BrCond.clear();
BBI.IsBrAnalyzable =
!TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond);
- BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL;
+ BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == nullptr;
if (BBI.BrCond.size()) {
// No false branch. This BB must end with a conditional branch and a
// fallthrough.
if (!BBI.FalseBB)
- BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB);
+ BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB);
if (!BBI.FalseBB) {
// Malformed bcc? True and false blocks are the same?
BBI.IsUnpredicable = true;
// Then scan all the instructions.
BBI.NonPredSize = 0;
+ BBI.ExtraCost = 0;
+ BBI.ExtraCost2 = 0;
BBI.ClobbersPred = false;
for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end();
I != E; ++I) {
- const TargetInstrDesc &TID = I->getDesc();
- if (TID.isNotDuplicable())
+ if (I->isDebugValue())
+ continue;
+
+ if (I->isNotDuplicable())
BBI.CannotBeCopied = true;
bool isPredicated = TII->isPredicated(I);
- bool isCondBr = BBI.IsBrAnalyzable && TID.isConditionalBranch();
-
- if (!isCondBr) {
- if (!isPredicated)
- BBI.NonPredSize++;
- else if (!AlreadyPredicated) {
- // FIXME: This instruction is already predicated before the
- // if-conversion pass. It's probably something like a conditional move.
- // Mark this block unpredicable for now.
- BBI.IsUnpredicable = true;
- return;
- }
+ bool isCondBr = BBI.IsBrAnalyzable && I->isConditionalBranch();
+
+ // A conditional branch is not predicable, but it may be eliminated.
+ if (isCondBr)
+ continue;
+
+ if (!isPredicated) {
+ BBI.NonPredSize++;
+ unsigned ExtraPredCost = TII->getPredicationCost(&*I);
+ unsigned NumCycles = SchedModel.computeInstrLatency(&*I, false);
+ if (NumCycles > 1)
+ BBI.ExtraCost += NumCycles-1;
+ BBI.ExtraCost2 += ExtraPredCost;
+ } else if (!AlreadyPredicated) {
+ // FIXME: This instruction is already predicated before the
+ // if-conversion pass. It's probably something like a conditional move.
+ // Mark this block unpredicable for now.
+ BBI.IsUnpredicable = true;
+ return;
}
if (BBI.ClobbersPred && !isPredicated) {
// Predicate modification instruction should end the block (except for
// already predicated instructions and end of block branches).
- if (isCondBr) {
- // A conditional branch is not predicable, but it may be eliminated.
- continue;
- }
-
// Predicate may have been modified, the subsequent (currently)
// unpredicated instructions cannot be correctly predicated.
BBI.IsUnpredicable = true;
if (TII->DefinesPredicate(I, PredDefs))
BBI.ClobbersPred = true;
- if (!TID.isPredicable()) {
+ if (!TII->isPredicable(I)) {
BBI.IsUnpredicable = true;
return;
}
if (BBI.IsDone || BBI.IsUnpredicable)
return false;
- // If it is already predicated, check if its predicate subsumes the new
- // predicate.
- if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred))
+ // If it is already predicated, check if the new predicate subsumes
+ // its predicate.
+ if (BBI.Predicate.size() && !TII->SubsumesPredicate(Pred, BBI.Predicate))
return false;
if (BBI.BrCond.size()) {
ScanInstructions(BBI);
- // Unanalyzable or ends with fallthrough or unconditional branch.
- if (!BBI.IsBrAnalyzable || BBI.BrCond.empty()) {
+ // Unanalyzable or ends with fallthrough or unconditional branch, or if is not
+ // considered for ifcvt anymore.
+ if (!BBI.IsBrAnalyzable || BBI.BrCond.empty() || BBI.IsDone) {
BBI.IsBeingAnalyzed = false;
BBI.IsAnalyzed = true;
return BBI;
unsigned Dups = 0;
unsigned Dups2 = 0;
- bool TNeedSub = TrueBBI.Predicate.size() > 0;
- bool FNeedSub = FalseBBI.Predicate.size() > 0;
+ bool TNeedSub = !TrueBBI.Predicate.empty();
+ bool FNeedSub = !FalseBBI.Predicate.empty();
bool Enqueued = false;
+
+ BranchProbability Prediction = MBPI->getEdgeProbability(BB, TrueBBI.BB);
+
if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) &&
- MeetIfcvtSizeLimit(TrueBBI.NonPredSize - (Dups + Dups2)) &&
- MeetIfcvtSizeLimit(FalseBBI.NonPredSize - (Dups + Dups2)) &&
+ MeetIfcvtSizeLimit(*TrueBBI.BB, (TrueBBI.NonPredSize - (Dups + Dups2) +
+ TrueBBI.ExtraCost), TrueBBI.ExtraCost2,
+ *FalseBBI.BB, (FalseBBI.NonPredSize - (Dups + Dups2) +
+ FalseBBI.ExtraCost),FalseBBI.ExtraCost2,
+ Prediction) &&
FeasibilityAnalysis(TrueBBI, BBI.BrCond) &&
FeasibilityAnalysis(FalseBBI, RevCond)) {
// Diamond:
Enqueued = true;
}
- if (ValidTriangle(TrueBBI, FalseBBI, false, Dups) &&
- MeetIfcvtSizeLimit(TrueBBI.NonPredSize) &&
+ if (ValidTriangle(TrueBBI, FalseBBI, false, Dups, Prediction) &&
+ MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost,
+ TrueBBI.ExtraCost2, Prediction) &&
FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) {
// Triangle:
// EBB
Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups));
Enqueued = true;
}
-
- if (ValidTriangle(TrueBBI, FalseBBI, true, Dups) &&
- MeetIfcvtSizeLimit(TrueBBI.NonPredSize) &&
+
+ if (ValidTriangle(TrueBBI, FalseBBI, true, Dups, Prediction) &&
+ MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost,
+ TrueBBI.ExtraCost2, Prediction) &&
FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) {
Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups));
Enqueued = true;
}
- if (ValidSimple(TrueBBI, Dups) &&
- MeetIfcvtSizeLimit(TrueBBI.NonPredSize) &&
+ if (ValidSimple(TrueBBI, Dups, Prediction) &&
+ MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost,
+ TrueBBI.ExtraCost2, Prediction) &&
FeasibilityAnalysis(TrueBBI, BBI.BrCond)) {
// Simple (split, no rejoin):
// EBB
// | \_
// | |
// | TBB---> exit
- // |
+ // |
// FBB
Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups));
Enqueued = true;
if (CanRevCond) {
// Try the other path...
- if (ValidTriangle(FalseBBI, TrueBBI, false, Dups) &&
- MeetIfcvtSizeLimit(FalseBBI.NonPredSize) &&
+ if (ValidTriangle(FalseBBI, TrueBBI, false, Dups,
+ Prediction.getCompl()) &&
+ MeetIfcvtSizeLimit(*FalseBBI.BB,
+ FalseBBI.NonPredSize + FalseBBI.ExtraCost,
+ FalseBBI.ExtraCost2, Prediction.getCompl()) &&
FeasibilityAnalysis(FalseBBI, RevCond, true)) {
Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups));
Enqueued = true;
}
- if (ValidTriangle(FalseBBI, TrueBBI, true, Dups) &&
- MeetIfcvtSizeLimit(FalseBBI.NonPredSize) &&
+ if (ValidTriangle(FalseBBI, TrueBBI, true, Dups,
+ Prediction.getCompl()) &&
+ MeetIfcvtSizeLimit(*FalseBBI.BB,
+ FalseBBI.NonPredSize + FalseBBI.ExtraCost,
+ FalseBBI.ExtraCost2, Prediction.getCompl()) &&
FeasibilityAnalysis(FalseBBI, RevCond, true, true)) {
Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups));
Enqueued = true;
}
- if (ValidSimple(FalseBBI, Dups) &&
- MeetIfcvtSizeLimit(FalseBBI.NonPredSize) &&
+ if (ValidSimple(FalseBBI, Dups, Prediction.getCompl()) &&
+ MeetIfcvtSizeLimit(*FalseBBI.BB,
+ FalseBBI.NonPredSize + FalseBBI.ExtraCost,
+ FalseBBI.ExtraCost2, Prediction.getCompl()) &&
FeasibilityAnalysis(FalseBBI, RevCond)) {
Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups));
Enqueued = true;
}
/// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion
-/// candidates. It returns true if any CFG restructuring is done to expose more
-/// if-conversion opportunities.
-bool IfConverter::AnalyzeBlocks(MachineFunction &MF,
+/// candidates.
+void IfConverter::AnalyzeBlocks(MachineFunction &MF,
std::vector<IfcvtToken*> &Tokens) {
- bool Change = false;
- std::set<MachineBasicBlock*> Visited;
- for (unsigned i = 0, e = Roots.size(); i != e; ++i) {
- for (idf_ext_iterator<MachineBasicBlock*> I=idf_ext_begin(Roots[i],Visited),
- E = idf_ext_end(Roots[i], Visited); I != E; ++I) {
- MachineBasicBlock *BB = *I;
- AnalyzeBlock(BB, Tokens);
- }
+ for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
+ MachineBasicBlock *BB = I;
+ AnalyzeBlock(BB, Tokens);
}
// Sort to favor more complex ifcvt scheme.
std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp);
-
- return Change;
}
/// canFallThroughTo - Returns true either if ToBB is the next block after BB or
/// that all the intervening blocks are empty (given BB can fall through to its
/// next block).
static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) {
- MachineFunction::iterator I = BB;
+ MachineFunction::iterator PI = BB;
+ MachineFunction::iterator I = std::next(PI);
MachineFunction::iterator TI = ToBB;
MachineFunction::iterator E = BB->getParent()->end();
- while (++I != TI)
- if (I == E || !I->empty())
+ while (I != TI) {
+ // Check isSuccessor to avoid case where the next block is empty, but
+ // it's not a successor.
+ if (I == E || !I->empty() || !PI->isSuccessor(I))
return false;
+ PI = I++;
+ }
return true;
}
/// to determine if it can be if-converted. If predecessor is already enqueued,
/// dequeue it!
void IfConverter::InvalidatePreds(MachineBasicBlock *BB) {
- for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(),
- E = BB->pred_end(); PI != E; ++PI) {
- BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()];
+ for (const auto &Predecessor : BB->predecessors()) {
+ BBInfo &PBBI = BBAnalysis[Predecessor->getNumber()];
if (PBBI.IsDone || PBBI.BB == BB)
continue;
PBBI.IsAnalyzed = false;
///
static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB,
const TargetInstrInfo *TII) {
+ DebugLoc dl; // FIXME: this is nowhere
SmallVector<MachineOperand, 0> NoCond;
- TII->InsertBranch(*BB, ToBB, NULL, NoCond);
+ TII->InsertBranch(*BB, ToBB, nullptr, NoCond, dl);
}
/// RemoveExtraEdges - Remove true / false edges if either / both are no longer
/// successors.
void IfConverter::RemoveExtraEdges(BBInfo &BBI) {
- MachineBasicBlock *TBB = NULL, *FBB = NULL;
+ MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
SmallVector<MachineOperand, 4> Cond;
if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond))
BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
}
+/// Behaves like LiveRegUnits::StepForward() but also adds implicit uses to all
+/// values defined in MI which are not live/used by MI.
+static void UpdatePredRedefs(MachineInstr *MI, LivePhysRegs &Redefs) {
+ for (ConstMIBundleOperands Ops(MI); Ops.isValid(); ++Ops) {
+ if (!Ops->isReg() || !Ops->isKill())
+ continue;
+ unsigned Reg = Ops->getReg();
+ if (Reg == 0)
+ continue;
+ Redefs.removeReg(Reg);
+ }
+ for (MIBundleOperands Ops(MI); Ops.isValid(); ++Ops) {
+ if (!Ops->isReg() || !Ops->isDef())
+ continue;
+ unsigned Reg = Ops->getReg();
+ if (Reg == 0 || Redefs.contains(Reg))
+ continue;
+ Redefs.addReg(Reg);
+
+ MachineOperand &Op = *Ops;
+ MachineInstr *MI = Op.getParent();
+ MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
+ MIB.addReg(Reg, RegState::Implicit | RegState::Undef);
+ }
+}
+
+/**
+ * Remove kill flags from operands with a registers in the @p DontKill set.
+ */
+static void RemoveKills(MachineInstr &MI, const LivePhysRegs &DontKill) {
+ for (MIBundleOperands O(&MI); O.isValid(); ++O) {
+ if (!O->isReg() || !O->isKill())
+ continue;
+ if (DontKill.contains(O->getReg()))
+ O->setIsKill(false);
+ }
+}
+
+/**
+ * Walks a range of machine instructions and removes kill flags for registers
+ * in the @p DontKill set.
+ */
+static void RemoveKills(MachineBasicBlock::iterator I,
+ MachineBasicBlock::iterator E,
+ const LivePhysRegs &DontKill,
+ const MCRegisterInfo &MCRI) {
+ for ( ; I != E; ++I)
+ RemoveKills(*I, DontKill);
+}
+
/// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG.
///
bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) {
return false;
}
+ if (CvtBBI->BB->hasAddressTaken())
+ // Conservatively abort if-conversion if BB's address is taken.
+ return false;
+
if (Kind == ICSimpleFalse)
if (TII->ReverseBranchCondition(Cond))
- assert(false && "Unable to reverse branch condition!");
+ llvm_unreachable("Unable to reverse branch condition!");
+
+ // Initialize liveins to the first BB. These are potentiall redefined by
+ // predicated instructions.
+ Redefs.init(TRI);
+ Redefs.addLiveIns(CvtBBI->BB);
+ Redefs.addLiveIns(NextBBI->BB);
+
+ // Compute a set of registers which must not be killed by instructions in
+ // BB1: This is everything live-in to BB2.
+ DontKill.init(TRI);
+ DontKill.addLiveIns(NextBBI->BB);
if (CvtBBI->BB->pred_size() > 1) {
BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
// Copy instructions in the true block, predicate them, and add them to
// the entry block.
CopyAndPredicateBlock(BBI, *CvtBBI, Cond);
+
+ // RemoveExtraEdges won't work if the block has an unanalyzable branch, so
+ // explicitly remove CvtBBI as a successor.
+ BBI.BB->removeSuccessor(CvtBBI->BB);
} else {
+ RemoveKills(CvtBBI->BB->begin(), CvtBBI->BB->end(), DontKill, *TRI);
PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond);
// Merge converted block into entry block.
return true;
}
+/// Scale down weights to fit into uint32_t. NewTrue is the new weight
+/// for successor TrueBB, and NewFalse is the new weight for successor
+/// FalseBB.
+static void ScaleWeights(uint64_t NewTrue, uint64_t NewFalse,
+ MachineBasicBlock *MBB,
+ const MachineBasicBlock *TrueBB,
+ const MachineBasicBlock *FalseBB,
+ const MachineBranchProbabilityInfo *MBPI) {
+ uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
+ uint32_t Scale = (NewMax / UINT32_MAX) + 1;
+ for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
+ SE = MBB->succ_end();
+ SI != SE; ++SI) {
+ if (*SI == TrueBB)
+ MBB->setSuccWeight(SI, (uint32_t)(NewTrue / Scale));
+ else if (*SI == FalseBB)
+ MBB->setSuccWeight(SI, (uint32_t)(NewFalse / Scale));
+ else
+ MBB->setSuccWeight(SI, MBPI->getEdgeWeight(MBB, SI) / Scale);
+ }
+}
+
/// IfConvertTriangle - If convert a triangle sub-CFG.
///
bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) {
BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
BBInfo *CvtBBI = &TrueBBI;
BBInfo *NextBBI = &FalseBBI;
+ DebugLoc dl; // FIXME: this is nowhere
SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end());
if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
return false;
}
+ if (CvtBBI->BB->hasAddressTaken())
+ // Conservatively abort if-conversion if BB's address is taken.
+ return false;
+
if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
if (TII->ReverseBranchCondition(Cond))
- assert(false && "Unable to reverse branch condition!");
+ llvm_unreachable("Unable to reverse branch condition!");
if (Kind == ICTriangleRev || Kind == ICTriangleFRev) {
if (ReverseBranchCondition(*CvtBBI)) {
}
}
- bool HasEarlyExit = CvtBBI->FalseBB != NULL;
- bool DupBB = CvtBBI->BB->pred_size() > 1;
- if (DupBB) {
+ // Initialize liveins to the first BB. These are potentially redefined by
+ // predicated instructions.
+ Redefs.init(TRI);
+ Redefs.addLiveIns(CvtBBI->BB);
+ Redefs.addLiveIns(NextBBI->BB);
+
+ DontKill.clear();
+
+ bool HasEarlyExit = CvtBBI->FalseBB != nullptr;
+ uint64_t CvtNext = 0, CvtFalse = 0, BBNext = 0, BBCvt = 0, SumWeight = 0;
+ uint32_t WeightScale = 0;
+
+ if (HasEarlyExit) {
+ // Get weights before modifying CvtBBI->BB and BBI.BB.
+ CvtNext = MBPI->getEdgeWeight(CvtBBI->BB, NextBBI->BB);
+ CvtFalse = MBPI->getEdgeWeight(CvtBBI->BB, CvtBBI->FalseBB);
+ BBNext = MBPI->getEdgeWeight(BBI.BB, NextBBI->BB);
+ BBCvt = MBPI->getEdgeWeight(BBI.BB, CvtBBI->BB);
+ SumWeight = MBPI->getSumForBlock(CvtBBI->BB, WeightScale);
+ }
+
+ if (CvtBBI->BB->pred_size() > 1) {
BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
// Copy instructions in the true block, predicate them, and add them to
// the entry block.
CopyAndPredicateBlock(BBI, *CvtBBI, Cond, true);
+
+ // RemoveExtraEdges won't work if the block has an unanalyzable branch, so
+ // explicitly remove CvtBBI as a successor.
+ BBI.BB->removeSuccessor(CvtBBI->BB);
} else {
// Predicate the 'true' block after removing its branch.
CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB);
// Now merge the entry of the triangle with the true block.
BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
- MergeBlocks(BBI, *CvtBBI);
+ MergeBlocks(BBI, *CvtBBI, false);
}
// If 'true' block has a 'false' successor, add an exit branch to it.
SmallVector<MachineOperand, 4> RevCond(CvtBBI->BrCond.begin(),
CvtBBI->BrCond.end());
if (TII->ReverseBranchCondition(RevCond))
- assert(false && "Unable to reverse branch condition!");
- TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond);
+ llvm_unreachable("Unable to reverse branch condition!");
+ TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, nullptr, RevCond, dl);
BBI.BB->addSuccessor(CvtBBI->FalseBB);
+ // Update the edge weight for both CvtBBI->FalseBB and NextBBI.
+ // New_Weight(BBI.BB, NextBBI->BB) =
+ // Weight(BBI.BB, NextBBI->BB) * getSumForBlock(CvtBBI->BB) +
+ // Weight(BBI.BB, CvtBBI->BB) * Weight(CvtBBI->BB, NextBBI->BB)
+ // New_Weight(BBI.BB, CvtBBI->FalseBB) =
+ // Weight(BBI.BB, CvtBBI->BB) * Weight(CvtBBI->BB, CvtBBI->FalseBB)
+
+ uint64_t NewNext = BBNext * SumWeight + (BBCvt * CvtNext) / WeightScale;
+ uint64_t NewFalse = (BBCvt * CvtFalse) / WeightScale;
+ // We need to scale down all weights of BBI.BB to fit uint32_t.
+ // Here BBI.BB is connected to CvtBBI->FalseBB and will fall through to
+ // the next block.
+ ScaleWeights(NewNext, NewFalse, BBI.BB, getNextBlock(BBI.BB),
+ CvtBBI->FalseBB, MBPI);
}
// Merge in the 'false' block if the 'false' block has no other
// block. By not merging them, we make it possible to iteratively
// ifcvt the blocks.
if (!HasEarlyExit &&
- NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) {
+ NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough &&
+ !NextBBI->BB->hasAddressTaken()) {
MergeBlocks(BBI, *NextBBI);
FalseBBDead = true;
} else {
RemoveExtraEdges(BBI);
// Update block info. BB can be iteratively if-converted.
- if (!IterIfcvt)
+ if (!IterIfcvt)
BBI.IsDone = true;
InvalidatePreds(BBI.BB);
CvtBBI->IsDone = true;
return false;
}
- // Merge the 'true' and 'false' blocks by copying the instructions
- // from the 'false' block to the 'true' block. That is, unless the true
- // block would clobber the predicate, in that case, do the opposite.
+ if (TrueBBI.BB->hasAddressTaken() || FalseBBI.BB->hasAddressTaken())
+ // Conservatively abort if-conversion if either BB has its address taken.
+ return false;
+
+ // Put the predicated instructions from the 'true' block before the
+ // instructions from the 'false' block, unless the true block would clobber
+ // the predicate, in which case, do the opposite.
BBInfo *BBI1 = &TrueBBI;
BBInfo *BBI2 = &FalseBBI;
SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end());
if (TII->ReverseBranchCondition(RevCond))
- assert(false && "Unable to reverse branch condition!");
+ llvm_unreachable("Unable to reverse branch condition!");
SmallVector<MachineOperand, 4> *Cond1 = &BBI.BrCond;
SmallVector<MachineOperand, 4> *Cond2 = &RevCond;
// Remove the conditional branch from entry to the blocks.
BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
+ // Initialize liveins to the first BB. These are potentially redefined by
+ // predicated instructions.
+ Redefs.init(TRI);
+ Redefs.addLiveIns(BBI1->BB);
+
// Remove the duplicated instructions at the beginnings of both paths.
MachineBasicBlock::iterator DI1 = BBI1->BB->begin();
MachineBasicBlock::iterator DI2 = BBI2->BB->begin();
+ MachineBasicBlock::iterator DIE1 = BBI1->BB->end();
+ MachineBasicBlock::iterator DIE2 = BBI2->BB->end();
+ // Skip dbg_value instructions
+ while (DI1 != DIE1 && DI1->isDebugValue())
+ ++DI1;
+ while (DI2 != DIE2 && DI2->isDebugValue())
+ ++DI2;
BBI1->NonPredSize -= NumDups1;
BBI2->NonPredSize -= NumDups1;
+
+ // Skip past the dups on each side separately since there may be
+ // differing dbg_value entries.
+ for (unsigned i = 0; i < NumDups1; ++DI1) {
+ if (!DI1->isDebugValue())
+ ++i;
+ }
while (NumDups1 != 0) {
- ++DI1;
++DI2;
- --NumDups1;
+ if (!DI2->isDebugValue())
+ --NumDups1;
+ }
+
+ // Compute a set of registers which must not be killed by instructions in BB1:
+ // This is everything used+live in BB2 after the duplicated instructions. We
+ // can compute this set by simulating liveness backwards from the end of BB2.
+ DontKill.init(TRI);
+ for (MachineBasicBlock::reverse_iterator I = BBI2->BB->rbegin(),
+ E = MachineBasicBlock::reverse_iterator(DI2); I != E; ++I) {
+ DontKill.stepBackward(*I);
+ }
+
+ for (MachineBasicBlock::const_iterator I = BBI1->BB->begin(), E = DI1; I != E;
+ ++I) {
+ Redefs.stepForward(*I);
}
BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1);
BBI2->BB->erase(BBI2->BB->begin(), DI2);
- // Predicate the 'true' block after removing its branch.
+ // Remove branch from 'true' block and remove duplicated instructions.
BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB);
DI1 = BBI1->BB->end();
- for (unsigned i = 0; i != NumDups2; ++i)
+ for (unsigned i = 0; i != NumDups2; ) {
+ // NumDups2 only counted non-dbg_value instructions, so this won't
+ // run off the head of the list.
+ assert (DI1 != BBI1->BB->begin());
--DI1;
+ // skip dbg_value instructions
+ if (!DI1->isDebugValue())
+ ++i;
+ }
BBI1->BB->erase(DI1, BBI1->BB->end());
- PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1);
- // Predicate the 'false' block.
+ // Kill flags in the true block for registers living into the false block
+ // must be removed.
+ RemoveKills(BBI1->BB->begin(), BBI1->BB->end(), DontKill, *TRI);
+
+ // Remove 'false' block branch and find the last instruction to predicate.
BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB);
DI2 = BBI2->BB->end();
while (NumDups2 != 0) {
+ // NumDups2 only counted non-dbg_value instructions, so this won't
+ // run off the head of the list.
+ assert (DI2 != BBI2->BB->begin());
--DI2;
- --NumDups2;
+ // skip dbg_value instructions
+ if (!DI2->isDebugValue())
+ --NumDups2;
+ }
+
+ // Remember which registers would later be defined by the false block.
+ // This allows us not to predicate instructions in the true block that would
+ // later be re-defined. That is, rather than
+ // subeq r0, r1, #1
+ // addne r0, r1, #1
+ // generate:
+ // sub r0, r1, #1
+ // addne r0, r1, #1
+ SmallSet<unsigned, 4> RedefsByFalse;
+ SmallSet<unsigned, 4> ExtUses;
+ if (TII->isProfitableToUnpredicate(*BBI1->BB, *BBI2->BB)) {
+ for (MachineBasicBlock::iterator FI = BBI2->BB->begin(); FI != DI2; ++FI) {
+ if (FI->isDebugValue())
+ continue;
+ SmallVector<unsigned, 4> Defs;
+ for (unsigned i = 0, e = FI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = FI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned Reg = MO.getReg();
+ if (!Reg)
+ continue;
+ if (MO.isDef()) {
+ Defs.push_back(Reg);
+ } else if (!RedefsByFalse.count(Reg)) {
+ // These are defined before ctrl flow reach the 'false' instructions.
+ // They cannot be modified by the 'true' instructions.
+ for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
+ SubRegs.isValid(); ++SubRegs)
+ ExtUses.insert(*SubRegs);
+ }
+ }
+
+ for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
+ unsigned Reg = Defs[i];
+ if (!ExtUses.count(Reg)) {
+ for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
+ SubRegs.isValid(); ++SubRegs)
+ RedefsByFalse.insert(*SubRegs);
+ }
+ }
+ }
}
+
+ // Predicate the 'true' block.
+ PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, &RedefsByFalse);
+
+ // Predicate the 'false' block.
PredicateBlock(*BBI2, DI2, *Cond2);
// Merge the true block into the entry of the diamond.
- MergeBlocks(BBI, *BBI1);
- MergeBlocks(BBI, *BBI2);
+ MergeBlocks(BBI, *BBI1, TailBB == nullptr);
+ MergeBlocks(BBI, *BBI2, TailBB == nullptr);
// If the if-converted block falls through or unconditionally branches into
// the tail block, and the tail block does not have other predecessors, then
// fold the tail block in as well. Otherwise, unless it falls through to the
// tail, add a unconditional branch to it.
if (TailBB) {
- BBInfo TailBBI = BBAnalysis[TailBB->getNumber()];
- if (TailBB->pred_size() == 1 && !TailBBI.HasFallThrough) {
- BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
+ BBInfo &TailBBI = BBAnalysis[TailBB->getNumber()];
+ bool CanMergeTail = !TailBBI.HasFallThrough &&
+ !TailBBI.BB->hasAddressTaken();
+ // There may still be a fall-through edge from BBI1 or BBI2 to TailBB;
+ // check if there are any other predecessors besides those.
+ unsigned NumPreds = TailBB->pred_size();
+ if (NumPreds > 1)
+ CanMergeTail = false;
+ else if (NumPreds == 1 && CanMergeTail) {
+ MachineBasicBlock::pred_iterator PI = TailBB->pred_begin();
+ if (*PI != BBI1->BB && *PI != BBI2->BB)
+ CanMergeTail = false;
+ }
+ if (CanMergeTail) {
MergeBlocks(BBI, TailBBI);
TailBBI.IsDone = true;
} else {
+ BBI.BB->addSuccessor(TailBB);
InsertUncondBranch(BBI.BB, TailBB, TII);
BBI.HasFallThrough = false;
}
}
+ // RemoveExtraEdges won't work if the block has an unanalyzable branch,
+ // which can happen here if TailBB is unanalyzable and is merged, so
+ // explicitly remove BBI1 and BBI2 as successors.
+ BBI.BB->removeSuccessor(BBI1->BB);
+ BBI.BB->removeSuccessor(BBI2->BB);
RemoveExtraEdges(BBI);
// Update block info.
return true;
}
+static bool MaySpeculate(const MachineInstr *MI,
+ SmallSet<unsigned, 4> &LaterRedefs,
+ const TargetInstrInfo *TII) {
+ bool SawStore = true;
+ if (!MI->isSafeToMove(TII, nullptr, SawStore))
+ return false;
+
+ for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = MI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned Reg = MO.getReg();
+ if (!Reg)
+ continue;
+ if (MO.isDef() && !LaterRedefs.count(Reg))
+ return false;
+ }
+
+ return true;
+}
+
/// PredicateBlock - Predicate instructions from the start of the block to the
/// specified end with the specified condition.
void IfConverter::PredicateBlock(BBInfo &BBI,
MachineBasicBlock::iterator E,
- SmallVectorImpl<MachineOperand> &Cond) {
+ SmallVectorImpl<MachineOperand> &Cond,
+ SmallSet<unsigned, 4> *LaterRedefs) {
+ bool AnyUnpred = false;
+ bool MaySpec = LaterRedefs != nullptr;
for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) {
- if (TII->isPredicated(I))
+ if (I->isDebugValue() || TII->isPredicated(I))
+ continue;
+ // It may be possible not to predicate an instruction if it's the 'true'
+ // side of a diamond and the 'false' side may re-define the instruction's
+ // defs.
+ if (MaySpec && MaySpeculate(I, *LaterRedefs, TII)) {
+ AnyUnpred = true;
continue;
+ }
+ // If any instruction is predicated, then every instruction after it must
+ // be predicated.
+ MaySpec = false;
if (!TII->PredicateInstruction(I, Cond)) {
#ifndef NDEBUG
- errs() << "Unable to predicate " << *I << "!\n";
+ dbgs() << "Unable to predicate " << *I << "!\n";
#endif
- llvm_unreachable(0);
+ llvm_unreachable(nullptr);
}
+
+ // If the predicated instruction now redefines a register as the result of
+ // if-conversion, add an implicit kill.
+ UpdatePredRedefs(I, Redefs);
}
std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate));
BBI.IsAnalyzed = false;
BBI.NonPredSize = 0;
- NumIfConvBBs++;
+ ++NumIfConvBBs;
+ if (AnyUnpred)
+ ++NumUnpred;
}
/// CopyAndPredicateBlock - Copy and predicate instructions from source BB to
for (MachineBasicBlock::iterator I = FromBBI.BB->begin(),
E = FromBBI.BB->end(); I != E; ++I) {
- const TargetInstrDesc &TID = I->getDesc();
- bool isPredicated = TII->isPredicated(I);
// Do not copy the end of the block branches.
- if (IgnoreBr && !isPredicated && TID.isBranch())
+ if (IgnoreBr && I->isBranch())
break;
MachineInstr *MI = MF.CloneMachineInstr(I);
ToBBI.BB->insert(ToBBI.BB->end(), MI);
ToBBI.NonPredSize++;
+ unsigned ExtraPredCost = TII->getPredicationCost(&*I);
+ unsigned NumCycles = SchedModel.computeInstrLatency(&*I, false);
+ if (NumCycles > 1)
+ ToBBI.ExtraCost += NumCycles-1;
+ ToBBI.ExtraCost2 += ExtraPredCost;
- if (!isPredicated)
+ if (!TII->isPredicated(I) && !MI->isDebugValue()) {
if (!TII->PredicateInstruction(MI, Cond)) {
#ifndef NDEBUG
- errs() << "Unable to predicate " << *I << "!\n";
+ dbgs() << "Unable to predicate " << *I << "!\n";
#endif
- llvm_unreachable(0);
+ llvm_unreachable(nullptr);
}
+ }
+
+ // If the predicated instruction now redefines a register as the result of
+ // if-conversion, add an implicit kill.
+ UpdatePredRedefs(MI, Redefs);
+
+ // Some kill flags may not be correct anymore.
+ if (!DontKill.empty())
+ RemoveKills(*MI, DontKill);
}
- std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
- FromBBI.BB->succ_end());
- MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
- MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
+ if (!IgnoreBr) {
+ std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
+ FromBBI.BB->succ_end());
+ MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
+ MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : nullptr;
- for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
- MachineBasicBlock *Succ = Succs[i];
- // Fallthrough edge can't be transferred.
- if (Succ == FallThrough)
- continue;
- ToBBI.BB->addSuccessor(Succ);
+ for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
+ MachineBasicBlock *Succ = Succs[i];
+ // Fallthrough edge can't be transferred.
+ if (Succ == FallThrough)
+ continue;
+ ToBBI.BB->addSuccessor(Succ);
+ }
}
std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
ToBBI.IsAnalyzed = false;
- NumDupBBs++;
+ ++NumDupBBs;
}
/// MergeBlocks - Move all instructions from FromBB to the end of ToBB.
-///
-void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI) {
+/// This will leave FromBB as an empty block, so remove all of its
+/// successor edges except for the fall-through edge. If AddEdges is true,
+/// i.e., when FromBBI's branch is being moved, add those successor edges to
+/// ToBBI.
+void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges) {
+ assert(!FromBBI.BB->hasAddressTaken() &&
+ "Removing a BB whose address is taken!");
+
ToBBI.BB->splice(ToBBI.BB->end(),
FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end());
- // Redirect all branches to FromBB to ToBB.
- std::vector<MachineBasicBlock *> Preds(FromBBI.BB->pred_begin(),
- FromBBI.BB->pred_end());
- for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
- MachineBasicBlock *Pred = Preds[i];
- if (Pred == ToBBI.BB)
- continue;
- Pred->ReplaceUsesOfBlockWith(FromBBI.BB, ToBBI.BB);
- }
-
std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
FromBBI.BB->succ_end());
MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
- MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
+ MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : nullptr;
for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
MachineBasicBlock *Succ = Succs[i];
if (Succ == FallThrough)
continue;
FromBBI.BB->removeSuccessor(Succ);
- ToBBI.BB->addSuccessor(Succ);
+ if (AddEdges && !ToBBI.BB->isSuccessor(Succ))
+ ToBBI.BB->addSuccessor(Succ);
}
// Now FromBBI always falls through to the next block!
FromBBI.Predicate.clear();
ToBBI.NonPredSize += FromBBI.NonPredSize;
+ ToBBI.ExtraCost += FromBBI.ExtraCost;
+ ToBBI.ExtraCost2 += FromBBI.ExtraCost2;
FromBBI.NonPredSize = 0;
+ FromBBI.ExtraCost = 0;
+ FromBBI.ExtraCost2 = 0;
ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
ToBBI.HasFallThrough = FromBBI.HasFallThrough;