//
// The LLVM Compiler Infrastructure
//
-// This file was developed by the Evan Cheng and is distributed under
-// the University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "ifcvt"
-#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/MachineBranchProbabilityInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineModuleInfo.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/TargetSchedule.h"
+#include "llvm/CodeGen/LiveRegUnits.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/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
-#include "llvm/Support/CommandLine.h"
-#include "llvm/Support/Debug.h"
-#include "llvm/ADT/DepthFirstIterator.h"
-#include "llvm/ADT/Statistic.h"
+#include "llvm/Target/TargetRegisterInfo.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
+
using namespace llvm;
-namespace {
- // Hidden options for help debugging.
- cl::opt<int> IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden);
- cl::opt<int> IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden);
- cl::opt<int> IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden);
- cl::opt<bool> DisableSimple("disable-ifcvt-simple",
- cl::init(false), cl::Hidden);
- cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false",
- cl::init(false), cl::Hidden);
- cl::opt<bool> DisableTriangle("disable-ifcvt-triangle",
- cl::init(false), cl::Hidden);
- cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev",
- cl::init(false), cl::Hidden);
- cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false",
- cl::init(false), cl::Hidden);
- cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev",
- cl::init(false), cl::Hidden);
- cl::opt<bool> DisableDiamond("disable-ifcvt-diamond",
- cl::init(false), cl::Hidden);
-}
+// 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",
+ cl::init(false), cl::Hidden);
+static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false",
+ cl::init(false), cl::Hidden);
+static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle",
+ cl::init(false), cl::Hidden);
+static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev",
+ cl::init(false), cl::Hidden);
+static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false",
+ cl::init(false), cl::Hidden);
+static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev",
+ cl::init(false), cl::Hidden);
+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(NumTriangleFRev, "Number of triangle (F/R) 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 {
- enum BBICKind {
+ enum IfcvtKind {
ICNotClassfied, // BB data valid, but not classified.
- ICSimple, // BB is entry of an one split, no rejoin sub-CFG.
ICSimpleFalse, // Same as ICSimple, but on the false path.
- ICTriangle, // BB is entry of a triangle sub-CFG.
+ ICSimple, // BB is entry of an one split, no rejoin sub-CFG.
+ ICTriangleFRev, // Same as ICTriangleFalse, but false path rev condition.
ICTriangleRev, // Same as ICTriangle, but true path rev condition.
ICTriangleFalse, // Same as ICTriangle, but on the false path.
- ICTriangleFRev, // Same as ICTriangleFalse, but false path rev condition.
+ ICTriangle, // BB is entry of a triangle sub-CFG.
ICDiamond // BB is entry of a diamond sub-CFG.
};
/// diamond shape), its size, whether it's predicable, and whether any
/// instruction can clobber the 'would-be' predicate.
///
- /// Kind - Type of block. See BBICKind.
/// IsDone - True if BB is not to be considered for ifcvt.
/// IsBeingAnalyzed - True if BB is currently being analyzed.
/// IsAnalyzed - True if BB has been analyzed (info is still valid).
/// IsBrAnalyzable - True if AnalyzeBranch() returns false.
/// HasFallThrough - True if BB may fallthrough to the following BB.
/// IsUnpredicable - True if BB is known to be unpredicable.
- /// ClobbersPredicate- True if BB would modify the predicate (e.g. has
+ /// 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.
/// Predicate - Predicate used in the BB.
struct BBInfo {
- BBICKind Kind;
bool IsDone : 1;
bool IsBeingAnalyzed : 1;
bool IsAnalyzed : 1;
bool IsBrAnalyzable : 1;
bool HasFallThrough : 1;
bool IsUnpredicable : 1;
+ bool CannotBeCopied : 1;
bool ClobbersPred : 1;
unsigned NonPredSize;
+ unsigned ExtraCost;
+ unsigned ExtraCost2;
MachineBasicBlock *BB;
MachineBasicBlock *TrueBB;
MachineBasicBlock *FalseBB;
- MachineBasicBlock *TailBB;
- std::vector<MachineOperand> BrCond;
- std::vector<MachineOperand> Predicate;
- BBInfo() : Kind(ICNotClassfied), IsDone(false), IsBeingAnalyzed(false),
+ SmallVector<MachineOperand, 4> BrCond;
+ SmallVector<MachineOperand, 4> Predicate;
+ BBInfo() : IsDone(false), IsBeingAnalyzed(false),
IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false),
HasFallThrough(false), IsUnpredicable(false),
- ClobbersPred(false), NonPredSize(0),
- BB(0), TrueBB(0), FalseBB(0), TailBB(0) {}
+ CannotBeCopied(false), ClobbersPred(false), NonPredSize(0),
+ ExtraCost(0), ExtraCost2(0), BB(0), TrueBB(0), FalseBB(0) {}
};
- /// Roots - Basic blocks that do not have successors. These are the starting
- /// points of Graph traversal.
- std::vector<MachineBasicBlock*> Roots;
+ /// 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
+ /// predicated.
+ /// NumDups - Number of instructions that would be duplicated due
+ /// to this if-conversion. (For diamonds, the number of
+ /// identical instructions at the beginnings of both
+ /// paths).
+ /// NumDups2 - For diamonds, the number of identical instructions
+ /// at the ends of both paths.
+ struct IfcvtToken {
+ BBInfo &BBI;
+ IfcvtKind Kind;
+ bool NeedSubsumption;
+ unsigned NumDups;
+ unsigned NumDups2;
+ IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d, unsigned d2 = 0)
+ : BBI(b), Kind(k), NeedSubsumption(s), NumDups(d), NumDups2(d2) {}
+ };
/// 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 MachineBranchProbabilityInfo *MBPI;
+ MachineRegisterInfo *MRI;
+
+ bool PreRegAlloc;
bool MadeChange;
+ int FnNum;
public:
static char ID;
- IfConverter() : MachineFunctionPass((intptr_t)&ID) {}
+ IfConverter() : MachineFunctionPass(ID), FnNum(-1) {
+ initializeIfConverterPass(*PassRegistry::getPassRegistry());
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<MachineBranchProbabilityInfo>();
+ MachineFunctionPass::getAnalysisUsage(AU);
+ }
virtual bool runOnMachineFunction(MachineFunction &MF);
- virtual const char *getPassName() const { return "If converter"; }
private:
bool ReverseBranchCondition(BBInfo &BBI);
- bool ValidSimple(BBInfo &TrueBBI) const;
+ bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups,
+ const BranchProbability &Prediction) const;
bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI,
- bool FalseBranch = false) const;
- bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI) const;
+ bool FalseBranch, unsigned &Dups,
+ const BranchProbability &Prediction) const;
+ bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI,
+ unsigned &Dups1, unsigned &Dups2) const;
void ScanInstructions(BBInfo &BBI);
- BBInfo &AnalyzeBlock(MachineBasicBlock *BB);
- bool FeasibilityAnalysis(BBInfo &BBI, std::vector<MachineOperand> &Cond,
+ BBInfo &AnalyzeBlock(MachineBasicBlock *BB,
+ std::vector<IfcvtToken*> &Tokens);
+ bool FeasibilityAnalysis(BBInfo &BBI, SmallVectorImpl<MachineOperand> &Cond,
bool isTriangle = false, bool RevBranch = false);
- bool AttemptRestructuring(BBInfo &BBI);
- bool AnalyzeBlocks(MachineFunction &MF,
- std::vector<BBInfo*> &Candidates);
- void ReTryPreds(MachineBasicBlock *BB);
+ void AnalyzeBlocks(MachineFunction &MF, std::vector<IfcvtToken*> &Tokens);
+ void InvalidatePreds(MachineBasicBlock *BB);
void RemoveExtraEdges(BBInfo &BBI);
- bool IfConvertSimple(BBInfo &BBI);
- bool IfConvertTriangle(BBInfo &BBI);
- bool IfConvertDiamond(BBInfo &BBI);
+ bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind);
+ bool IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind);
+ bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind,
+ unsigned NumDups1, unsigned NumDups2);
void PredicateBlock(BBInfo &BBI,
- std::vector<MachineOperand> &Cond,
- bool IgnoreTerm = false);
- void MergeBlocks(BBInfo &TrueBBI, BBInfo &FalseBBI);
+ MachineBasicBlock::iterator E,
+ SmallVectorImpl<MachineOperand> &Cond,
+ LiveRegUnits &Redefs,
+ SmallSet<unsigned, 4> *LaterRedefs = 0);
+ void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
+ SmallVectorImpl<MachineOperand> &Cond,
+ LiveRegUnits &Redefs,
+ const LiveRegUnits *DontKill = 0,
+ bool IgnoreBr = false);
+ 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(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;
}
- // IfcvtCandidateCmp - Used to sort if-conversion candidates.
- static bool IfcvtCandidateCmp(BBInfo* C1, BBInfo* C2){
- // Favor diamond over triangle, etc.
- return (unsigned)C1->Kind < (unsigned)C2->Kind;
+ // IfcvtTokenCmp - Used to sort if-conversion candidates.
+ static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) {
+ int Incr1 = (C1->Kind == ICDiamond)
+ ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups;
+ int Incr2 = (C2->Kind == ICDiamond)
+ ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups;
+ if (Incr1 > Incr2)
+ return true;
+ else if (Incr1 == Incr2) {
+ // Favors subsumption.
+ if (C1->NeedSubsumption == false && C2->NeedSubsumption == true)
+ return true;
+ else if (C1->NeedSubsumption == C2->NeedSubsumption) {
+ // Favors diamond over triangle, etc.
+ if ((unsigned)C1->Kind < (unsigned)C2->Kind)
+ return true;
+ else if (C1->Kind == C2->Kind)
+ return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber();
+ }
+ }
+ return false;
}
};
+
char IfConverter::ID = 0;
}
-FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); }
+char &llvm::IfConverterID = IfConverter::ID;
+
+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();
+ TRI = MF.getTarget().getRegisterInfo();
+ MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
+ MRI = &MF.getRegInfo();
+
+ const TargetSubtargetInfo &ST =
+ MF.getTarget().getSubtarget<TargetSubtargetInfo>();
+ SchedModel.init(*ST.getSchedModel(), &ST, TII);
+
if (!TII) return false;
- static int FnNum = -1;
- DOUT << "\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);
+ BFChange = BF.OptimizeFunction(MF, TII,
+ MF.getTarget().getRegisterInfo(),
+ getAnalysisIfAvailable<MachineModuleInfo>());
+ }
+
+ DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum << ") \'"
+ << MF.getName() << "\'");
if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) {
- DOUT << " skipped\n";
+ DEBUG(dbgs() << " skipped\n");
return false;
}
- DOUT << "\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_size() == 0)
- Roots.push_back(I);
-
- std::vector<BBInfo*> Candidates;
+ std::vector<IfcvtToken*> Tokens;
MadeChange = false;
- while (IfCvtLimit == -1 || (int)NumIfConvBBs < IfCvtLimit) {
- // Do an intial analysis for each basic block and finding all the potential
- // candidates to perform if-convesion.
- bool Change = AnalyzeBlocks(MF, Candidates);
- while (!Candidates.empty()) {
- BBInfo &BBI = *Candidates.back();
- Candidates.pop_back();
+ unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle +
+ NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds;
+ 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 = false;
+ AnalyzeBlocks(MF, Tokens);
+ while (!Tokens.empty()) {
+ IfcvtToken *Token = Tokens.back();
+ Tokens.pop_back();
+ BBInfo &BBI = Token->BBI;
+ IfcvtKind Kind = Token->Kind;
+ unsigned NumDups = Token->NumDups;
+ unsigned NumDups2 = Token->NumDups2;
+
+ delete Token;
// If the block has been evicted out of the queue or it has already been
// marked dead (due to it being predicated), then skip it.
- if (!BBI.IsEnqueued || BBI.IsDone)
+ if (BBI.IsDone)
+ BBI.IsEnqueued = false;
+ if (!BBI.IsEnqueued)
continue;
+ BBI.IsEnqueued = false;
+
bool RetVal = false;
- switch (BBI.Kind) {
- default: assert(false && "Unexpected!");
- break;
+ switch (Kind) {
+ default: llvm_unreachable("Unexpected!");
case ICSimple:
case ICSimpleFalse: {
- bool isFalse = BBI.Kind == ICSimpleFalse;
+ bool isFalse = Kind == ICSimpleFalse;
if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break;
- DOUT << "Ifcvt (Simple" << (BBI.Kind == ICSimpleFalse ? " false" : "")
- << "): BB#" << BBI.BB->getNumber() << " ("
- << ((BBI.Kind == ICSimpleFalse)
- ? BBI.FalseBB->getNumber() : BBI.TrueBB->getNumber()) << ") ";
- RetVal = IfConvertSimple(BBI);
- DOUT << (RetVal ? "succeeded!" : "failed!") << "\n";
- if (RetVal)
- if (isFalse) NumSimpleFalse++;
- else NumSimple++;
+ DEBUG(dbgs() << "Ifcvt (Simple" << (Kind == ICSimpleFalse ?
+ " false" : "")
+ << "): BB#" << BBI.BB->getNumber() << " ("
+ << ((Kind == ICSimpleFalse)
+ ? BBI.FalseBB->getNumber()
+ : BBI.TrueBB->getNumber()) << ") ");
+ RetVal = IfConvertSimple(BBI, Kind);
+ DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n");
+ if (RetVal) {
+ if (isFalse) ++NumSimpleFalse;
+ else ++NumSimple;
+ }
break;
}
case ICTriangle:
case ICTriangleRev:
case ICTriangleFalse:
case ICTriangleFRev: {
- bool isFalse = BBI.Kind == ICTriangleFalse;
- bool isRev = (BBI.Kind == ICTriangleRev || BBI.Kind == ICTriangleFRev);
+ bool isFalse = Kind == ICTriangleFalse;
+ bool isRev = (Kind == ICTriangleRev || Kind == ICTriangleFRev);
if (DisableTriangle && !isFalse && !isRev) break;
if (DisableTriangleR && !isFalse && isRev) break;
if (DisableTriangleF && isFalse && !isRev) break;
if (DisableTriangleFR && isFalse && isRev) break;
- DOUT << "Ifcvt (Triangle";
+ DEBUG(dbgs() << "Ifcvt (Triangle");
if (isFalse)
- DOUT << " false";
+ DEBUG(dbgs() << " false");
if (isRev)
- DOUT << " rev";
- DOUT << "): BB#" << BBI.BB->getNumber() << " (T:"
- << BBI.TrueBB->getNumber() << ",F:" << BBI.FalseBB->getNumber()
- << ") ";
- RetVal = IfConvertTriangle(BBI);
- DOUT << (RetVal ? "succeeded!" : "failed!") << "\n";
+ DEBUG(dbgs() << " rev");
+ DEBUG(dbgs() << "): BB#" << BBI.BB->getNumber() << " (T:"
+ << BBI.TrueBB->getNumber() << ",F:"
+ << BBI.FalseBB->getNumber() << ") ");
+ RetVal = IfConvertTriangle(BBI, Kind);
+ 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:
+ case ICDiamond: {
if (DisableDiamond) break;
- DOUT << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:"
- << BBI.TrueBB->getNumber() << ",F:" << BBI.FalseBB->getNumber();
- if (BBI.TailBB)
- DOUT << "," << BBI.TailBB->getNumber() ;
- DOUT << ") ";
- RetVal = IfConvertDiamond(BBI);
- DOUT << (RetVal ? "succeeded!" : "failed!") << "\n";
- if (RetVal) NumDiamonds++;
+ DEBUG(dbgs() << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:"
+ << BBI.TrueBB->getNumber() << ",F:"
+ << BBI.FalseBB->getNumber() << ") ");
+ RetVal = IfConvertDiamond(BBI, Kind, NumDups, NumDups2);
+ DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n");
+ if (RetVal) ++NumDiamonds;
break;
}
+ }
+
Change |= RetVal;
- if (IfCvtLimit != -1 && (int)NumIfConvBBs >= IfCvtLimit)
+ NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + NumTriangleRev +
+ NumTriangleFalse + NumTriangleFRev + NumDiamonds;
+ if (IfCvtLimit != -1 && (int)NumIfCvts >= IfCvtLimit)
break;
}
MadeChange |= Change;
}
- Roots.clear();
+ // Delete tokens in case of early exit.
+ while (!Tokens.empty()) {
+ IfcvtToken *Token = Tokens.back();
+ Tokens.pop_back();
+ delete Token;
+ }
+
+ Tokens.clear();
BBAnalysis.clear();
+ if (MadeChange && IfCvtBranchFold) {
+ BranchFolder BF(false, false);
+ BF.OptimizeFunction(MF, TII,
+ MF.getTarget().getRegisterInfo(),
+ getAnalysisIfAvailable<MachineModuleInfo>());
+ }
+
+ MadeChange |= BFChange;
return MadeChange;
}
+/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
+/// its 'true' successor.
static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
MachineBasicBlock *TrueBB) {
for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
return NULL;
}
+/// 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;
}
}
/// ValidSimple - Returns true if the 'true' block (along with its
-/// predecessor) forms a valid simple shape for ifcvt.
-bool IfConverter::ValidSimple(BBInfo &TrueBBI) const {
- if (TrueBBI.IsBeingAnalyzed)
+/// 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 BranchProbability &Prediction) const {
+ Dups = 0;
+ if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
+ return false;
+
+ if (TrueBBI.IsBrAnalyzable)
return false;
- return !blockAlwaysFallThrough(TrueBBI) &&
- TrueBBI.BrCond.size() == 0 && TrueBBI.BB->pred_size() == 1;
+ if (TrueBBI.BB->pred_size() > 1) {
+ if (TrueBBI.CannotBeCopied ||
+ !TII->isProfitableToDupForIfCvt(*TrueBBI.BB, TrueBBI.NonPredSize,
+ Prediction))
+ return false;
+ Dups = TrueBBI.NonPredSize;
+ }
+
+ return true;
}
/// 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' 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) const {
- if (TrueBBI.IsBeingAnalyzed)
+ bool FalseBranch, unsigned &Dups,
+ const BranchProbability &Prediction) const {
+ Dups = 0;
+ if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone)
return false;
- if (TrueBBI.BB->pred_size() != 1)
- return false;
+ if (TrueBBI.BB->pred_size() > 1) {
+ if (TrueBBI.CannotBeCopied)
+ return false;
+
+ unsigned Size = TrueBBI.NonPredSize;
+ if (TrueBBI.IsBrAnalyzable) {
+ if (TrueBBI.TrueBB && TrueBBI.BrCond.empty())
+ // Ends with an unconditional branch. It will be removed.
+ --Size;
+ else {
+ MachineBasicBlock *FExit = FalseBranch
+ ? TrueBBI.TrueBB : TrueBBI.FalseBB;
+ if (FExit)
+ // Require a conditional branch
+ ++Size;
+ }
+ }
+ if (!TII->isProfitableToDupForIfCvt(*TrueBBI.BB, Size, Prediction))
+ return false;
+ Dups = Size;
+ }
MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB;
if (!TExit && blockAlwaysFallThrough(TrueBBI)) {
/// 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) const {
- if (TrueBBI.IsBeingAnalyzed || FalseBBI.IsBeingAnalyzed)
+bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI,
+ unsigned &Dups1, unsigned &Dups2) const {
+ Dups1 = Dups2 = 0;
+ if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone ||
+ FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone)
return false;
MachineBasicBlock *TT = TrueBBI.TrueBB;
return false;
if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable))
return false;
- // FIXME: Allow false block to have an early exit?
- return (TrueBBI.BB->pred_size() == 1 &&
- FalseBBI.BB->pred_size() == 1 &&
- !TrueBBI.FalseBB && !FalseBBI.FalseBB);
+ if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1)
+ return false;
+
+ // FIXME: Allow true block to have an early exit?
+ if (TrueBBI.FalseBB || FalseBBI.FalseBB ||
+ (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred))
+ return false;
+
+ // 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;
+ ++TIB;
+ ++FIB;
+ }
+
+ // 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;
+ --TIE;
+ --FIE;
+ }
+
+ return true;
}
/// ScanInstructions - Scan all the instructions in the block to determine if
/// the block is predicable. In most cases, that means all the instructions
-/// in the block has M_PREDICABLE flag. Also checks if the block contains any
+/// in the block are isPredicable(). Also checks if the block contains any
/// instruction which can clobber a predicate (e.g. condition code register).
/// If so, the block is not predicable unless it's the last instruction.
void IfConverter::ScanInstructions(BBInfo &BBI) {
if (BBI.IsDone)
return;
+ bool AlreadyPredicated = !BBI.Predicate.empty();
// First analyze the end of BB branches.
- BBI.TrueBB = BBI.FalseBB;
+ BBI.TrueBB = BBI.FalseBB = NULL;
BBI.BrCond.clear();
BBI.IsBrAnalyzable =
!TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond);
// No false branch. This BB must end with a conditional branch and a
// fallthrough.
if (!BBI.FalseBB)
- BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB);
- assert(BBI.FalseBB && "Expected to find the fallthrough block!");
+ BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB);
+ if (!BBI.FalseBB) {
+ // Malformed bcc? True and false blocks are the same?
+ BBI.IsUnpredicable = true;
+ return;
+ }
}
// Then scan all the instructions.
BBI.NonPredSize = 0;
+ BBI.ExtraCost = 0;
+ BBI.ExtraCost2 = 0;
BBI.ClobbersPred = false;
- bool SeenCondBr = false;
for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end();
I != E; ++I) {
- const TargetInstrDescriptor *TID = I->getInstrDescriptor();
+ if (I->isDebugValue())
+ continue;
+
+ if (I->isNotDuplicable())
+ BBI.CannotBeCopied = true;
+
bool isPredicated = TII->isPredicated(I);
- bool isCondBr = BBI.IsBrAnalyzable &&
- (TID->Flags & M_BRANCH_FLAG) != 0 && (TID->Flags & M_BARRIER_FLAG) == 0;
+ bool isCondBr = BBI.IsBrAnalyzable && I->isConditionalBranch();
+
+ // A conditional branch is not predicable, but it may be eliminated.
+ if (isCondBr)
+ continue;
- if (!isPredicated && !isCondBr)
+ 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) {
- SeenCondBr = true;
-
- // Conditional branches is not predicable. But it may be eliminated.
- continue;
- }
-
// Predicate may have been modified, the subsequent (currently)
- // unpredocated instructions cannot be correctly predicated.
+ // unpredicated instructions cannot be correctly predicated.
BBI.IsUnpredicable = true;
return;
}
- if (TID->Flags & M_CLOBBERS_PRED)
+ // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are
+ // still potentially predicable.
+ std::vector<MachineOperand> PredDefs;
+ if (TII->DefinesPredicate(I, PredDefs))
BBI.ClobbersPred = true;
- if (!I->isPredicable()) {
+ if (!TII->isPredicable(I)) {
BBI.IsUnpredicable = true;
return;
}
/// FeasibilityAnalysis - Determine if the block is a suitable candidate to be
/// predicated by the specified predicate.
bool IfConverter::FeasibilityAnalysis(BBInfo &BBI,
- std::vector<MachineOperand> &Pred,
+ SmallVectorImpl<MachineOperand> &Pred,
bool isTriangle, bool RevBranch) {
- // Forget about it if it's unpredicable.
- if (BBI.IsUnpredicable)
+ // If the block is dead or unpredicable, then it cannot be predicated.
+ if (BBI.IsDone || BBI.IsUnpredicable)
return false;
- // If the block is dead, or it is going to be the entry block of a sub-CFG
- // that will be if-converted, then it cannot be predicated.
- if (BBI.IsDone || BBI.IsEnqueued)
- return false;
-
- // Check predication threshold.
- if (BBI.NonPredSize == 0 || BBI.NonPredSize > TLI->getIfCvtBlockSizeLimit())
- 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()) {
if (!isTriangle)
return false;
- // Test predicate subsumsion.
- std::vector<MachineOperand> RevPred(Pred);
- std::vector<MachineOperand> Cond(BBI.BrCond);
+ // Test predicate subsumption.
+ SmallVector<MachineOperand, 4> RevPred(Pred.begin(), Pred.end());
+ SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end());
if (RevBranch) {
if (TII->ReverseBranchCondition(Cond))
return false;
/// AnalyzeBlock - Analyze the structure of the sub-CFG starting from
/// the specified block. Record its successors and whether it looks like an
/// if-conversion candidate.
-IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB) {
+IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
+ std::vector<IfcvtToken*> &Tokens) {
BBInfo &BBI = BBAnalysis[BB->getNumber()];
if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed)
BBI.BB = BB;
BBI.IsBeingAnalyzed = true;
- BBI.Kind = ICNotClassfied;
ScanInstructions(BBI);
- // Unanalyable or ends with fallthrough or unconditional branch.
- if (!BBI.IsBrAnalyzable || BBI.BrCond.size() == 0) {
+ // 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;
return BBI;
}
- BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB);
- BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB);
+ // Do not ifcvt if true and false fallthrough blocks are the same.
+ if (!BBI.FalseBB) {
+ BBI.IsBeingAnalyzed = false;
+ BBI.IsAnalyzed = true;
+ return BBI;
+ }
+
+ BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens);
+ BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens);
if (TrueBBI.IsDone && FalseBBI.IsDone) {
BBI.IsBeingAnalyzed = false;
return BBI;
}
- std::vector<MachineOperand> RevCond(BBI.BrCond);
+ SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end());
bool CanRevCond = !TII->ReverseBranchCondition(RevCond);
- if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI) &&
- !(TrueBBI.ClobbersPred && FalseBBI.ClobbersPred) &&
+ unsigned Dups = 0;
+ unsigned Dups2 = 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.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:
// \ /
// TailBB
// Note TailBB can be empty.
- BBI.Kind = ICDiamond;
- BBI.TailBB = TrueBBI.TrueBB;
- } else {
- // FIXME: Consider duplicating if BB is small.
- if (ValidTriangle(TrueBBI, FalseBBI) &&
- FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) {
- // Triangle:
- // EBB
- // | \_
- // | |
- // | TBB
- // | /
- // FBB
- BBI.Kind = ICTriangle;
- } else if (ValidTriangle(TrueBBI, FalseBBI, true) &&
- FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) {
- BBI.Kind = ICTriangleRev;
- } else if (ValidSimple(TrueBBI) &&
- FeasibilityAnalysis(TrueBBI, BBI.BrCond)) {
- // Simple (split, no rejoin):
- // EBB
- // | \_
- // | |
- // | TBB---> exit
- // |
- // FBB
- BBI.Kind = ICSimple;
- } else if (CanRevCond) {
- // Try the other path...
- if (ValidTriangle(FalseBBI, TrueBBI) &&
- FeasibilityAnalysis(FalseBBI, RevCond, true)) {
- BBI.Kind = ICTriangleFalse;
- } else if (ValidTriangle(FalseBBI, TrueBBI, true) &&
- FeasibilityAnalysis(FalseBBI, RevCond, true, true)) {
- BBI.Kind = ICTriangleFRev;
- } else if (ValidSimple(FalseBBI) &&
- FeasibilityAnalysis(FalseBBI, RevCond)) {
- BBI.Kind = ICSimpleFalse;
- }
+ Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups,
+ Dups2));
+ Enqueued = true;
+ }
+
+ if (ValidTriangle(TrueBBI, FalseBBI, false, Dups, Prediction) &&
+ MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost,
+ TrueBBI.ExtraCost2, Prediction) &&
+ FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) {
+ // Triangle:
+ // EBB
+ // | \_
+ // | |
+ // | TBB
+ // | /
+ // FBB
+ Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups));
+ Enqueued = true;
+ }
+
+ 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, 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,
+ 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,
+ 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, 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;
}
}
+ BBI.IsEnqueued = Enqueued;
BBI.IsBeingAnalyzed = false;
BBI.IsAnalyzed = true;
return BBI;
}
-/// AttemptRestructuring - Restructure the sub-CFG rooted in the given block to
-/// expose more if-conversion opportunities. e.g.
-///
-/// cmp
-/// b le BB1
-/// / \____
-/// / |
-/// cmp |
-/// b eq BB1 |
-/// / \____ |
-/// / \ |
-/// BB1
-/// ==>
-///
-/// cmp
-/// b eq BB1
-/// / \____
-/// / |
-/// cmp |
-/// b le BB1 |
-/// / \____ |
-/// / \ |
-/// BB1
-bool IfConverter::AttemptRestructuring(BBInfo &BBI) {
- return false;
-}
-
/// 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,
- std::vector<BBInfo*> &Candidates) {
- 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;
- BBInfo &BBI = AnalyzeBlock(BB);
- switch (BBI.Kind) {
- case ICSimple:
- case ICSimpleFalse:
- case ICTriangle:
- case ICTriangleRev:
- case ICTriangleFalse:
- case ICTriangleFRev:
- case ICDiamond:
- Candidates.push_back(&BBI);
- break;
- default:
- Change |= AttemptRestructuring(BBI);
- break;
- }
- }
+/// candidates.
+void IfConverter::AnalyzeBlocks(MachineFunction &MF,
+ std::vector<IfcvtToken*> &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(Candidates.begin(), Candidates.end(), IfcvtCandidateCmp);
-
- return Change;
+ std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp);
}
/// 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 = llvm::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;
}
-/// ReTryPreds - Invalidate predecessor BB info so it would be re-analyzed
-/// to determine if it can be if-converted.
-void IfConverter::ReTryPreds(MachineBasicBlock *BB) {
+/// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed
+/// 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()];
- if (!PBBI.IsDone && PBBI.Kind == ICNotClassfied) {
- assert(PBBI.IsEnqueued && "Unexpected");
- PBBI.IsAnalyzed = false;
- }
+ if (PBBI.IsDone || PBBI.BB == BB)
+ continue;
+ PBBI.IsAnalyzed = false;
+ PBBI.IsEnqueued = false;
}
}
///
static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB,
const TargetInstrInfo *TII) {
- std::vector<MachineOperand> NoCond;
- TII->InsertBranch(*BB, ToBB, NULL, NoCond);
+ DebugLoc dl; // FIXME: this is nowhere
+ SmallVector<MachineOperand, 0> NoCond;
+ TII->InsertBranch(*BB, ToBB, NULL, NoCond, dl);
}
/// RemoveExtraEdges - Remove true / false edges if either / both are no longer
/// successors.
void IfConverter::RemoveExtraEdges(BBInfo &BBI) {
MachineBasicBlock *TBB = NULL, *FBB = NULL;
- std::vector<MachineOperand> Cond;
- bool isAnalyzable = !TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond);
- bool CanFallthrough = isAnalyzable && (TBB == NULL || FBB == NULL);
- if (BBI.TrueBB && BBI.BB->isSuccessor(BBI.TrueBB))
- if (!(BBI.TrueBB == TBB || BBI.TrueBB == FBB ||
- (CanFallthrough && getNextBlock(BBI.BB) == BBI.TrueBB)))
- BBI.BB->removeSuccessor(BBI.TrueBB);
- if (BBI.FalseBB && BBI.BB->isSuccessor(BBI.FalseBB))
- if (!(BBI.FalseBB == TBB || BBI.FalseBB == FBB ||
- (CanFallthrough && getNextBlock(BBI.BB) == BBI.FalseBB)))
- BBI.BB->removeSuccessor(BBI.FalseBB);
+ 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, LiveRegUnits &Redefs,
+ const TargetRegisterInfo *TRI) {
+ for (ConstMIBundleOperands Ops(MI); Ops.isValid(); ++Ops) {
+ if (!Ops->isReg() || !Ops->isKill())
+ continue;
+ unsigned Reg = Ops->getReg();
+ if (Reg == 0)
+ continue;
+ Redefs.RemoveReg(Reg, *TRI);
+ }
+ for (MIBundleOperands Ops(MI); Ops.isValid(); ++Ops) {
+ if (!Ops->isReg() || !Ops->isDef())
+ continue;
+ unsigned Reg = Ops->getReg();
+ if (Reg == 0 || Redefs.Contains(Reg, *TRI))
+ continue;
+ Redefs.AddReg(Reg, *TRI);
+
+ 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 LiveRegUnits &DontKill,
+ const MCRegisterInfo &MCRI) {
+ for (MIBundleOperands O(&MI); O.isValid(); ++O) {
+ if (!O->isReg() || !O->isKill())
+ continue;
+ if (DontKill.Contains(O->getReg(), MCRI))
+ 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 LiveRegUnits &DontKill,
+ const MCRegisterInfo &MCRI) {
+ for ( ; I != E; ++I)
+ RemoveKills(*I, DontKill, MCRI);
}
/// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG.
///
-bool IfConverter::IfConvertSimple(BBInfo &BBI) {
+bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) {
BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
BBInfo *CvtBBI = &TrueBBI;
BBInfo *NextBBI = &FalseBBI;
- std::vector<MachineOperand> Cond(BBI.BrCond);
- if (BBI.Kind == ICSimpleFalse) {
+ SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end());
+ if (Kind == ICSimpleFalse)
std::swap(CvtBBI, NextBBI);
- TII->ReverseBranchCondition(Cond);
+
+ if (CvtBBI->IsDone ||
+ (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) {
+ // Something has changed. It's no longer safe to predicate this block.
+ BBI.IsAnalyzed = false;
+ CvtBBI->IsAnalyzed = false;
+ return false;
}
- PredicateBlock(*CvtBBI, Cond);
+ if (CvtBBI->BB->hasAddressTaken())
+ // Conservatively abort if-conversion if BB's address is taken.
+ return false;
- // Merge converted block into entry block.
- BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
- MergeBlocks(BBI, *CvtBBI);
+ if (Kind == ICSimpleFalse)
+ if (TII->ReverseBranchCondition(Cond))
+ llvm_unreachable("Unable to reverse branch condition!");
+
+ // Initialize liveins to the first BB. These are potentiall redefined by
+ // predicated instructions.
+ LiveRegUnits Redefs;
+ Redefs.AddLiveIns(*(CvtBBI->BB), *TRI);
+ Redefs.AddLiveIns(*(NextBBI->BB), *TRI);
+
+ // Compute a set of registers which must not be killed by instructions in
+ // BB1: This is everything live-in to BB2.
+ LiveRegUnits DontKill;
+ DontKill.AddLiveIns(*(NextBBI->BB), *TRI);
+
+ 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, Redefs, &DontKill);
+
+ // 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, Redefs);
+
+ // Merge converted block into entry block.
+ BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
+ MergeBlocks(BBI, *CvtBBI);
+ }
bool IterIfcvt = true;
if (!canFallThroughTo(BBI.BB, NextBBI->BB)) {
// Update block info. BB can be iteratively if-converted.
if (!IterIfcvt)
BBI.IsDone = true;
- ReTryPreds(BBI.BB);
+ InvalidatePreds(BBI.BB);
CvtBBI->IsDone = true;
// FIXME: Must maintain LiveIns.
/// IfConvertTriangle - If convert a triangle sub-CFG.
///
-bool IfConverter::IfConvertTriangle(BBInfo &BBI) {
+bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) {
BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
BBInfo *CvtBBI = &TrueBBI;
BBInfo *NextBBI = &FalseBBI;
+ DebugLoc dl; // FIXME: this is nowhere
- std::vector<MachineOperand> Cond(BBI.BrCond);
- if (BBI.Kind == ICTriangleFalse || BBI.Kind == ICTriangleFRev) {
+ SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end());
+ if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
std::swap(CvtBBI, NextBBI);
- TII->ReverseBranchCondition(Cond);
- }
- if (BBI.Kind == ICTriangleRev || BBI.Kind == ICTriangleFRev) {
- ReverseBranchCondition(*CvtBBI);
- // BB has been changed, modify its predecessors (except for this
- // one) so they don't get ifcvt'ed based on bad intel.
- for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(),
- E = CvtBBI->BB->pred_end(); PI != E; ++PI) {
- MachineBasicBlock *PBB = *PI;
- if (PBB == BBI.BB)
- continue;
- BBInfo &PBBI = BBAnalysis[PBB->getNumber()];
- if (PBBI.IsEnqueued)
- PBBI.IsEnqueued = false;
+
+ if (CvtBBI->IsDone ||
+ (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) {
+ // Something has changed. It's no longer safe to predicate this block.
+ BBI.IsAnalyzed = false;
+ CvtBBI->IsAnalyzed = false;
+ 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))
+ llvm_unreachable("Unable to reverse branch condition!");
+
+ if (Kind == ICTriangleRev || Kind == ICTriangleFRev) {
+ if (ReverseBranchCondition(*CvtBBI)) {
+ // BB has been changed, modify its predecessors (except for this
+ // one) so they don't get ifcvt'ed based on bad intel.
+ for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(),
+ E = CvtBBI->BB->pred_end(); PI != E; ++PI) {
+ MachineBasicBlock *PBB = *PI;
+ if (PBB == BBI.BB)
+ continue;
+ BBInfo &PBBI = BBAnalysis[PBB->getNumber()];
+ if (PBBI.IsEnqueued) {
+ PBBI.IsAnalyzed = false;
+ PBBI.IsEnqueued = false;
+ }
+ }
}
}
- // Predicate the 'true' block after removing its branch.
- CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB);
- PredicateBlock(*CvtBBI, Cond);
+ // Initialize liveins to the first BB. These are potentially redefined by
+ // predicated instructions.
+ LiveRegUnits Redefs;
+ Redefs.AddLiveIns(*(CvtBBI->BB), *TRI);
+ Redefs.AddLiveIns(*(NextBBI->BB), *TRI);
- // If 'true' block has a 'false' successor, add an exit branch to it.
bool HasEarlyExit = CvtBBI->FalseBB != NULL;
+ 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, Redefs, 0, 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);
+ PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs);
+
+ // Now merge the entry of the triangle with the true block.
+ BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
+ MergeBlocks(BBI, *CvtBBI, false);
+ }
+
+ // If 'true' block has a 'false' successor, add an exit branch to it.
if (HasEarlyExit) {
- std::vector<MachineOperand> RevCond(CvtBBI->BrCond);
+ SmallVector<MachineOperand, 4> RevCond(CvtBBI->BrCond.begin(),
+ CvtBBI->BrCond.end());
if (TII->ReverseBranchCondition(RevCond))
- assert(false && "Unable to reverse branch condition!");
- TII->InsertBranch(*CvtBBI->BB, CvtBBI->FalseBB, NULL, RevCond);
+ llvm_unreachable("Unable to reverse branch condition!");
+ TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond, dl);
+ BBI.BB->addSuccessor(CvtBBI->FalseBB);
}
- // Now merge the entry of the triangle with the true block.
- BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
- MergeBlocks(BBI, *CvtBBI);
-
// Merge in the 'false' block if the 'false' block has no other
- // predecessors. Otherwise, add a unconditional branch from to 'false'.
+ // predecessors. Otherwise, add an unconditional branch to 'false'.
bool FalseBBDead = false;
bool IterIfcvt = true;
bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB);
// Only merge them if the true block does not fallthrough to the false
// block. By not merging them, we make it possible to iteratively
// ifcvt the blocks.
- if (!HasEarlyExit && NextBBI->BB->pred_size() == 1) {
+ if (!HasEarlyExit &&
+ 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;
- ReTryPreds(BBI.BB);
+ InvalidatePreds(BBI.BB);
CvtBBI->IsDone = true;
if (FalseBBDead)
NextBBI->IsDone = true;
/// IfConvertDiamond - If convert a diamond sub-CFG.
///
-bool IfConverter::IfConvertDiamond(BBInfo &BBI) {
+bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind,
+ unsigned NumDups1, unsigned NumDups2) {
BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()];
BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()];
-
- SmallVector<MachineInstr*, 2> Dups;
- if (!BBI.TailBB) {
- // No common merge block. Check if the terminators (e.g. return) are
- // the same or predicable.
- MachineBasicBlock::iterator TT = BBI.TrueBB->getFirstTerminator();
- MachineBasicBlock::iterator FT = BBI.FalseBB->getFirstTerminator();
- while (TT != BBI.TrueBB->end() && FT != BBI.FalseBB->end()) {
- if (TT->isIdenticalTo(FT))
- Dups.push_back(TT); // Will erase these later.
- else if (!TT->isPredicable() && !FT->isPredicable())
- return false; // Can't if-convert. Abort!
- ++TT;
- ++FT;
- }
-
- // One of the two pathes have more terminators, make sure they are
- // all predicable.
- while (TT != BBI.TrueBB->end()) {
- if (!TT->isPredicable()) {
- return false; // Can't if-convert. Abort!
- }
- ++TT;
- }
- while (FT != BBI.FalseBB->end()) {
- if (!FT->isPredicable()) {
- return false; // Can't if-convert. Abort!
- }
- ++FT;
- }
+ MachineBasicBlock *TailBB = TrueBBI.TrueBB;
+ // True block must fall through or end with an unanalyzable terminator.
+ if (!TailBB) {
+ if (blockAlwaysFallThrough(TrueBBI))
+ TailBB = FalseBBI.TrueBB;
+ assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!");
}
- // Remove the duplicated instructions from the 'true' block.
- for (unsigned i = 0, e = Dups.size(); i != e; ++i) {
- Dups[i]->eraseFromParent();
- --TrueBBI.NonPredSize;
+ if (TrueBBI.IsDone || FalseBBI.IsDone ||
+ TrueBBI.BB->pred_size() > 1 ||
+ FalseBBI.BB->pred_size() > 1) {
+ // Something has changed. It's no longer safe to predicate these blocks.
+ BBI.IsAnalyzed = false;
+ TrueBBI.IsAnalyzed = false;
+ FalseBBI.IsAnalyzed = false;
+ 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;
- std::vector<MachineOperand> RevCond(BBI.BrCond);
- TII->ReverseBranchCondition(RevCond);
- std::vector<MachineOperand> *Cond1 = &BBI.BrCond;
- std::vector<MachineOperand> *Cond2 = &RevCond;
- // Check the 'true' and 'false' blocks if either isn't ended with a branch.
- // Either the block fallthrough to another block or it ends with a
- // return. If it's the former, add a branch to its successor.
- bool NeedBr1 = !BBI1->TrueBB && BBI1->BB->succ_size();
- bool NeedBr2 = !BBI2->TrueBB && BBI2->BB->succ_size();
-
- if ((TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred) ||
- (!TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred &&
- NeedBr1 && !NeedBr2)) {
+ SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end());
+ if (TII->ReverseBranchCondition(RevCond))
+ llvm_unreachable("Unable to reverse branch condition!");
+ SmallVector<MachineOperand, 4> *Cond1 = &BBI.BrCond;
+ SmallVector<MachineOperand, 4> *Cond2 = &RevCond;
+
+ // Figure out the more profitable ordering.
+ bool DoSwap = false;
+ if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred)
+ DoSwap = true;
+ else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) {
+ if (TrueBBI.NonPredSize > FalseBBI.NonPredSize)
+ DoSwap = true;
+ }
+ if (DoSwap) {
std::swap(BBI1, BBI2);
std::swap(Cond1, Cond2);
- std::swap(NeedBr1, NeedBr2);
}
- // Predicate the 'true' block after removing its branch.
- BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB);
- PredicateBlock(*BBI1, *Cond1);
+ // Remove the conditional branch from entry to the blocks.
+ BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
- // Add an early exit branch if needed.
- if (NeedBr1)
- TII->InsertBranch(*BBI1->BB, *BBI1->BB->succ_begin(), NULL, *Cond1);
+ // Initialize liveins to the first BB. These are potentially redefined by
+ // predicated instructions.
+ LiveRegUnits Redefs;
+ Redefs.AddLiveIns(*(BBI1->BB), *TRI);
+
+ // 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) {
+ ++DI2;
+ if (!DI2->isDebugValue())
+ --NumDups1;
+ }
- // Predicate the 'false' block.
- PredicateBlock(*BBI2, *Cond2, true);
-
- // Add an unconditional branch from 'false' to to 'false' successor if it
- // will not be the fallthrough block.
- if (NeedBr2 && !NeedBr1) {
- // If BBI2 isn't going to be merged in, then the existing fallthrough
- // or branch is fine.
- if (!canFallThroughTo(BBI.BB, *BBI2->BB->succ_begin())) {
- InsertUncondBranch(BBI2->BB, *BBI2->BB->succ_begin(), TII);
- BBI2->HasFallThrough = false;
- }
+ // 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.
+ LiveRegUnits DontKill;
+ for (MachineBasicBlock::reverse_instr_iterator I = BBI2->BB->rbegin(),
+ E = MachineBasicBlock::reverse_instr_iterator(&*DI2); I != E; ++I) {
+ DontKill.StepBackward(*I, *TRI);
}
- // Keep them as two separate blocks if there is an early exit.
- if (!NeedBr1)
- MergeBlocks(*BBI1, *BBI2);
+ for (MachineBasicBlock::const_iterator I = BBI1->BB->begin(), E = DI1; I != E;
+ ++I) {
+ Redefs.StepForward(*I, *TRI);
+ }
+ BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1);
+ BBI2->BB->erase(BBI2->BB->begin(), DI2);
- // Remove the conditional branch from entry to the blocks.
- BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB);
+ // 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; ) {
+ // 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());
+
+ // 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;
+ // skip dbg_value instructions
+ if (!DI2->isDebugValue())
+ --NumDups2;
+ }
- // Merge the combined block into the entry of the diamond.
- MergeBlocks(BBI, *BBI1);
+ // 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);
+ }
+ }
- // 'True' and 'false' aren't combined, see if we need to add a unconditional
- // branch to the 'false' block.
- if (NeedBr1 && !canFallThroughTo(BBI.BB, BBI2->BB)) {
- InsertUncondBranch(BBI.BB, BBI2->BB, TII);
- BBI1->HasFallThrough = false;
+ 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);
+ }
+ }
+ }
}
- // If the if-converted block fallthrough or unconditionally branch into the
- // tail block, and the tail block does not have other predecessors, then
- // fold the tail block in as well.
- BBInfo *CvtBBI = NeedBr1 ? BBI2 : &BBI;
- if (BBI.TailBB &&
- BBI.TailBB->pred_size() == 1 && CvtBBI->BB->succ_size() == 1) {
- CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB);
- BBInfo TailBBI = BBAnalysis[BBI.TailBB->getNumber()];
- MergeBlocks(*CvtBBI, TailBBI);
- TailBBI.IsDone = true;
+ // Predicate the 'true' block.
+ PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, Redefs, &RedefsByFalse);
+
+ // Predicate the 'false' block.
+ PredicateBlock(*BBI2, DI2, *Cond2, Redefs);
+
+ // Merge the true block into the entry of the diamond.
+ MergeBlocks(BBI, *BBI1, TailBB == 0);
+ MergeBlocks(BBI, *BBI2, TailBB == 0);
+
+ // 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()];
+ 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.
BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true;
+ InvalidatePreds(BBI.BB);
// FIXME: Must maintain LiveIns.
return true;
}
-/// PredicateBlock - Predicate every instruction in the block with the specified
-/// condition. If IgnoreTerm is true, skip over all terminator instructions.
+static bool MaySpeculate(const MachineInstr *MI,
+ SmallSet<unsigned, 4> &LaterRedefs,
+ const TargetInstrInfo *TII) {
+ bool SawStore = true;
+ if (!MI->isSafeToMove(TII, 0, 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,
- std::vector<MachineOperand> &Cond,
- bool IgnoreTerm) {
- for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end();
- I != E; ++I) {
- if (IgnoreTerm && TII->isTerminatorInstr(I->getOpcode()))
+ MachineBasicBlock::iterator E,
+ SmallVectorImpl<MachineOperand> &Cond,
+ LiveRegUnits &Redefs,
+ SmallSet<unsigned, 4> *LaterRedefs) {
+ bool AnyUnpred = false;
+ bool MaySpec = LaterRedefs != 0;
+ for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) {
+ if (I->isDebugValue() || TII->isPredicated(I))
continue;
- if (TII->isPredicated(I))
+ // 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)) {
- cerr << "Unable to predicate " << *I << "!\n";
- abort();
+#ifndef NDEBUG
+ dbgs() << "Unable to predicate " << *I << "!\n";
+#endif
+ llvm_unreachable(0);
}
+
+ // If the predicated instruction now redefines a register as the result of
+ // if-conversion, add an implicit kill.
+ UpdatePredRedefs(I, Redefs, TRI);
}
- BBI.NonPredSize = 0;
std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate));
- NumIfConvBBs++;
+ BBI.IsAnalyzed = false;
+ BBI.NonPredSize = 0;
+
+ ++NumIfConvBBs;
+ if (AnyUnpred)
+ ++NumUnpred;
+}
+
+/// CopyAndPredicateBlock - Copy and predicate instructions from source BB to
+/// the destination block. Skip end of block branches if IgnoreBr is true.
+void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
+ SmallVectorImpl<MachineOperand> &Cond,
+ LiveRegUnits &Redefs,
+ const LiveRegUnits *DontKill,
+ bool IgnoreBr) {
+ MachineFunction &MF = *ToBBI.BB->getParent();
+
+ for (MachineBasicBlock::iterator I = FromBBI.BB->begin(),
+ E = FromBBI.BB->end(); I != E; ++I) {
+ // Do not copy the end of the block branches.
+ 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 (!TII->isPredicated(I) && !MI->isDebugValue()) {
+ if (!TII->PredicateInstruction(MI, Cond)) {
+#ifndef NDEBUG
+ dbgs() << "Unable to predicate " << *I << "!\n";
+#endif
+ llvm_unreachable(0);
+ }
+ }
+
+ // If the predicated instruction now redefines a register as the result of
+ // if-conversion, add an implicit kill.
+ UpdatePredRedefs(MI, Redefs, TRI);
+
+ // Some kill flags may not be correct anymore.
+ if (DontKill != 0)
+ RemoveKills(*MI, *DontKill, *TRI);
+ }
+
+ if (!IgnoreBr) {
+ std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(),
+ FromBBI.BB->succ_end());
+ MachineBasicBlock *NBB = getNextBlock(FromBBI.BB);
+ MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL;
+
+ 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(),
+ std::back_inserter(ToBBI.Predicate));
+ std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate));
+
+ ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
+ ToBBI.IsAnalyzed = false;
+
+ ++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);
if (Succ == FallThrough)
continue;
FromBBI.BB->removeSuccessor(Succ);
- if (!ToBBI.BB->isSuccessor(Succ))
+ if (AddEdges && !ToBBI.BB->isSuccessor(Succ))
ToBBI.BB->addSuccessor(Succ);
}
- // Now FromBBI always fall through to the next block!
- if (NBB)
+ // Now FromBBI always falls through to the next block!
+ if (NBB && !FromBBI.BB->isSuccessor(NBB))
FromBBI.BB->addSuccessor(NBB);
+ std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
+ std::back_inserter(ToBBI.Predicate));
+ 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;
-
- std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),
- std::back_inserter(ToBBI.Predicate));
- FromBBI.Predicate.clear();
+ ToBBI.IsAnalyzed = false;
+ FromBBI.IsAnalyzed = false;
}
projects / oota-llvm.git / blobdiff