X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FARM%2FARMConstantIslandPass.cpp;h=9aced30f0136e2e038e62558585ed70a63500de4;hb=4ee451de366474b9c228b4e5fa573795a715216d;hp=199782dcfe556f924d31c85a91a9279a31422caa;hpb=05cc424082f797c5820b19f29f398c9cce9b9928;p=oota-llvm.git diff --git a/lib/Target/ARM/ARMConstantIslandPass.cpp b/lib/Target/ARM/ARMConstantIslandPass.cpp index 199782dcfe5..9aced30f013 100644 --- a/lib/Target/ARM/ARMConstantIslandPass.cpp +++ b/lib/Target/ARM/ARMConstantIslandPass.cpp @@ -2,8 +2,8 @@ // // The LLVM Compiler Infrastructure // -// This file was developed by Chris Lattner 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. // //===----------------------------------------------------------------------===// // @@ -24,20 +24,21 @@ #include "llvm/Target/TargetMachine.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/Debug.h" +#include "llvm/ADT/SmallVector.h" #include "llvm/ADT/STLExtras.h" #include "llvm/ADT/Statistic.h" -#include using namespace llvm; +STATISTIC(NumCPEs, "Number of constpool entries"); STATISTIC(NumSplit, "Number of uncond branches inserted"); STATISTIC(NumCBrFixed, "Number of cond branches fixed"); STATISTIC(NumUBrFixed, "Number of uncond branches fixed"); namespace { - /// ARMConstantIslands - Due to limited pc-relative displacements, ARM + /// ARMConstantIslands - Due to limited PC-relative displacements, ARM /// requires constant pool entries to be scattered among the instructions /// inside a function. To do this, it completely ignores the normal LLVM - /// constant pool, instead, it places constants where-ever it feels like with + /// constant pool; instead, it places constants wherever it feels like with /// special instructions. /// /// The terminology used in this pass includes: @@ -48,16 +49,23 @@ namespace { class VISIBILITY_HIDDEN ARMConstantIslands : public MachineFunctionPass { /// NextUID - Assign unique ID's to CPE's. unsigned NextUID; - + /// BBSizes - The size of each MachineBasicBlock in bytes of code, indexed - /// by MBB Number. + /// by MBB Number. The two-byte pads required for Thumb alignment are + /// counted as part of the following block (i.e., the offset and size for + /// a padded block will both be ==2 mod 4). std::vector BBSizes; + /// BBOffsets - the offset of each MBB in bytes, starting from 0. + /// The two-byte pads required for Thumb alignment are counted as part of + /// the following block. + std::vector BBOffsets; + /// WaterList - A sorted list of basic blocks where islands could be placed /// (i.e. blocks that don't fall through to the following block, due /// to a return, unreachable, or unconditional branch). std::vector WaterList; - + /// CPUser - One user of a constant pool, keeping the machine instruction /// pointer, the constant pool being referenced, and the max displacement /// allowed from the instruction to the CP. @@ -73,6 +81,24 @@ namespace { /// constant pools and their max displacement. std::vector CPUsers; + /// CPEntry - One per constant pool entry, keeping the machine instruction + /// pointer, the constpool index, and the number of CPUser's which + /// reference this entry. + struct CPEntry { + MachineInstr *CPEMI; + unsigned CPI; + unsigned RefCount; + CPEntry(MachineInstr *cpemi, unsigned cpi, unsigned rc = 0) + : CPEMI(cpemi), CPI(cpi), RefCount(rc) {} + }; + + /// CPEntries - Keep track of all of the constant pool entry machine + /// instructions. For each original constpool index (i.e. those that + /// existed upon entry to this pass), it keeps a vector of entries. + /// Original elements are cloned as we go along; the clones are + /// put in the vector of the original element, but have distinct CPIs. + std::vector > CPEntries; + /// ImmBranch - One per immediate branch, keeping the machine instruction /// pointer, conditional or unconditional, the max displacement, /// and (if isCond is true) the corresponding unconditional branch @@ -86,21 +112,25 @@ namespace { : MI(mi), MaxDisp(maxdisp), isCond(cond), UncondBr(ubr) {} }; - /// Branches - Keep track of all the immediate branch instructions. + /// ImmBranches - Keep track of all the immediate branch instructions. /// std::vector ImmBranches; /// PushPopMIs - Keep track of all the Thumb push / pop instructions. /// - std::vector PushPopMIs; + SmallVector PushPopMIs; /// HasFarJump - True if any far jump instruction has been emitted during /// the branch fix up pass. bool HasFarJump; const TargetInstrInfo *TII; - const ARMFunctionInfo *AFI; + ARMFunctionInfo *AFI; + bool isThumb; public: + static char ID; + ARMConstantIslands() : MachineFunctionPass((intptr_t)&ID) {} + virtual bool runOnMachineFunction(MachineFunction &Fn); virtual const char *getPassName() const { @@ -110,12 +140,30 @@ namespace { private: void DoInitialPlacement(MachineFunction &Fn, std::vector &CPEMIs); + CPEntry *findConstPoolEntry(unsigned CPI, const MachineInstr *CPEMI); void InitialFunctionScan(MachineFunction &Fn, const std::vector &CPEMIs); MachineBasicBlock *SplitBlockBeforeInstr(MachineInstr *MI); void UpdateForInsertedWaterBlock(MachineBasicBlock *NewBB); - bool HandleConstantPoolUser(MachineFunction &Fn, CPUser &U); - bool CPEIsInRange(MachineInstr *MI, MachineInstr *CPEMI, unsigned Disp); + void AdjustBBOffsetsAfter(MachineBasicBlock *BB, int delta); + bool DecrementOldEntry(unsigned CPI, MachineInstr* CPEMI); + int LookForExistingCPEntry(CPUser& U, unsigned UserOffset); + bool LookForWater(CPUser&U, unsigned UserOffset, + MachineBasicBlock** NewMBB); + MachineBasicBlock* AcceptWater(MachineBasicBlock *WaterBB, + std::vector::iterator IP); + void CreateNewWater(unsigned CPUserIndex, unsigned UserOffset, + MachineBasicBlock** NewMBB); + bool HandleConstantPoolUser(MachineFunction &Fn, unsigned CPUserIndex); + void RemoveDeadCPEMI(MachineInstr *CPEMI); + bool RemoveUnusedCPEntries(); + bool CPEIsInRange(MachineInstr *MI, unsigned UserOffset, + MachineInstr *CPEMI, unsigned Disp, + bool DoDump); + bool WaterIsInRange(unsigned UserOffset, MachineBasicBlock *Water, + CPUser &U); + bool OffsetIsInRange(unsigned UserOffset, unsigned TrialOffset, + unsigned Disp, bool NegativeOK); bool BBIsInRange(MachineInstr *MI, MachineBasicBlock *BB, unsigned Disp); bool FixUpImmediateBr(MachineFunction &Fn, ImmBranch &Br); bool FixUpConditionalBr(MachineFunction &Fn, ImmBranch &Br); @@ -123,8 +171,37 @@ namespace { bool UndoLRSpillRestore(); unsigned GetOffsetOf(MachineInstr *MI) const; - unsigned GetOffsetOf(MachineBasicBlock *MBB) const; + void dumpBBs(); + void verify(MachineFunction &Fn); }; + char ARMConstantIslands::ID = 0; +} + +/// verify - check BBOffsets, BBSizes, alignment of islands +void ARMConstantIslands::verify(MachineFunction &Fn) { + assert(BBOffsets.size() == BBSizes.size()); + for (unsigned i = 1, e = BBOffsets.size(); i != e; ++i) + assert(BBOffsets[i-1]+BBSizes[i-1] == BBOffsets[i]); + if (isThumb) { + for (MachineFunction::iterator MBBI = Fn.begin(), E = Fn.end(); + MBBI != E; ++MBBI) { + MachineBasicBlock *MBB = MBBI; + if (!MBB->empty() && + MBB->begin()->getOpcode() == ARM::CONSTPOOL_ENTRY) + assert((BBOffsets[MBB->getNumber()]%4 == 0 && + BBSizes[MBB->getNumber()]%4 == 0) || + (BBOffsets[MBB->getNumber()]%4 != 0 && + BBSizes[MBB->getNumber()]%4 != 0)); + } + } +} + +/// print block size and offset information - debugging +void ARMConstantIslands::dumpBBs() { + for (unsigned J = 0, E = BBOffsets.size(); J !=E; ++J) { + DOUT << "block " << J << " offset " << BBOffsets[J] << + " size " << BBSizes[J] << "\n"; + } } /// createARMConstantIslandPass - returns an instance of the constpool @@ -138,6 +215,7 @@ bool ARMConstantIslands::runOnMachineFunction(MachineFunction &Fn) { TII = Fn.getTarget().getInstrInfo(); AFI = Fn.getInfo(); + isThumb = AFI->isThumbFunction(); HasFarJump = false; @@ -145,11 +223,18 @@ bool ARMConstantIslands::runOnMachineFunction(MachineFunction &Fn) { // the numbers agree with the position of the block in the function. Fn.RenumberBlocks(); + /// Thumb functions containing constant pools get 2-byte alignment. This is so + /// we can keep exact track of where the alignment padding goes. Set default. + AFI->setAlign(isThumb ? 1U : 2U); + // Perform the initial placement of the constant pool entries. To start with, // we put them all at the end of the function. std::vector CPEMIs; - if (!MCP.isEmpty()) + if (!MCP.isEmpty()) { DoInitialPlacement(Fn, CPEMIs); + if (isThumb) + AFI->setAlign(2U); + } /// The next UID to take is the first unused one. NextUID = CPEMIs.size(); @@ -160,29 +245,40 @@ bool ARMConstantIslands::runOnMachineFunction(MachineFunction &Fn) { InitialFunctionScan(Fn, CPEMIs); CPEMIs.clear(); + /// Remove dead constant pool entries. + RemoveUnusedCPEntries(); + // Iteratively place constant pool entries and fix up branches until there // is no change. bool MadeChange = false; while (true) { bool Change = false; for (unsigned i = 0, e = CPUsers.size(); i != e; ++i) - Change |= HandleConstantPoolUser(Fn, CPUsers[i]); + Change |= HandleConstantPoolUser(Fn, i); + DEBUG(dumpBBs()); for (unsigned i = 0, e = ImmBranches.size(); i != e; ++i) Change |= FixUpImmediateBr(Fn, ImmBranches[i]); + DEBUG(dumpBBs()); if (!Change) break; MadeChange = true; } - + + // After a while, this might be made debug-only, but it is not expensive. + verify(Fn); + // If LR has been forced spilled and no far jumps (i.e. BL) has been issued. // Undo the spill / restore of LR if possible. - if (!HasFarJump && AFI->isLRForceSpilled() && AFI->isThumbFunction()) + if (!HasFarJump && AFI->isLRSpilledForFarJump() && isThumb) MadeChange |= UndoLRSpillRestore(); BBSizes.clear(); + BBOffsets.clear(); WaterList.clear(); CPUsers.clear(); + CPEntries.clear(); ImmBranches.clear(); + PushPopMIs.clear(); return MadeChange; } @@ -190,7 +286,7 @@ bool ARMConstantIslands::runOnMachineFunction(MachineFunction &Fn) { /// DoInitialPlacement - Perform the initial placement of the constant pool /// entries. To start with, we put them all at the end of the function. void ARMConstantIslands::DoInitialPlacement(MachineFunction &Fn, - std::vector &CPEMIs){ + std::vector &CPEMIs){ // Create the basic block to hold the CPE's. MachineBasicBlock *BB = new MachineBasicBlock(); Fn.getBasicBlockList().push_back(BB); @@ -202,7 +298,7 @@ void ARMConstantIslands::DoInitialPlacement(MachineFunction &Fn, const TargetData &TD = *Fn.getTarget().getTargetData(); for (unsigned i = 0, e = CPs.size(); i != e; ++i) { - unsigned Size = TD.getTypeSize(CPs[i].getType()); + unsigned Size = TD.getABITypeSize(CPs[i].getType()); // Verify that all constant pool entries are a multiple of 4 bytes. If not, // we would have to pad them out or something so that instructions stay // aligned. @@ -211,12 +307,17 @@ void ARMConstantIslands::DoInitialPlacement(MachineFunction &Fn, BuildMI(BB, TII->get(ARM::CONSTPOOL_ENTRY)) .addImm(i).addConstantPoolIndex(i).addImm(Size); CPEMIs.push_back(CPEMI); - DEBUG(std::cerr << "Moved CPI#" << i << " to end of function as #" - << i << "\n"); + + // Add a new CPEntry, but no corresponding CPUser yet. + std::vector CPEs; + CPEs.push_back(CPEntry(CPEMI, i)); + CPEntries.push_back(CPEs); + NumCPEs++; + DOUT << "Moved CPI#" << i << " to end of function as #" << i << "\n"; } } -/// BBHasFallthrough - Return true of the specified basic block can fallthrough +/// BBHasFallthrough - Return true if the specified basic block can fallthrough /// into the block immediately after it. static bool BBHasFallthrough(MachineBasicBlock *MBB) { // Get the next machine basic block in the function. @@ -233,11 +334,27 @@ static bool BBHasFallthrough(MachineBasicBlock *MBB) { return false; } +/// findConstPoolEntry - Given the constpool index and CONSTPOOL_ENTRY MI, +/// look up the corresponding CPEntry. +ARMConstantIslands::CPEntry +*ARMConstantIslands::findConstPoolEntry(unsigned CPI, + const MachineInstr *CPEMI) { + std::vector &CPEs = CPEntries[CPI]; + // Number of entries per constpool index should be small, just do a + // linear search. + for (unsigned i = 0, e = CPEs.size(); i != e; ++i) { + if (CPEs[i].CPEMI == CPEMI) + return &CPEs[i]; + } + return NULL; +} + /// InitialFunctionScan - Do the initial scan of the function, building up /// information about the sizes of each block, the location of all the water, /// and finding all of the constant pool users. void ARMConstantIslands::InitialFunctionScan(MachineFunction &Fn, - const std::vector &CPEMIs) { + const std::vector &CPEMIs) { + unsigned Offset = 0; for (MachineFunction::iterator MBBI = Fn.begin(), E = Fn.end(); MBBI != E; ++MBBI) { MachineBasicBlock &MBB = *MBBI; @@ -260,8 +377,15 @@ void ARMConstantIslands::InitialFunctionScan(MachineFunction &Fn, unsigned Scale = 1; int UOpc = Opc; switch (Opc) { + case ARM::tBR_JTr: + // A Thumb table jump may involve padding; for the offsets to + // be right, functions containing these must be 4-byte aligned. + AFI->setAlign(2U); + if ((Offset+MBBSize)%4 != 0) + MBBSize += 2; // padding + continue; // Does not get an entry in ImmBranches default: - continue; // Ignore JT branches + continue; // Ignore other JT branches case ARM::Bcc: isCond = true; UOpc = ARM::B; @@ -283,7 +407,7 @@ void ARMConstantIslands::InitialFunctionScan(MachineFunction &Fn, } // Record this immediate branch. - unsigned MaxOffs = (1 << (Bits-1)) * Scale; + unsigned MaxOffs = ((1 << (Bits-1))-1) * Scale; ImmBranches.push_back(ImmBranch(I, MaxOffs, isCond, UOpc)); } @@ -305,6 +429,10 @@ void ARMConstantIslands::InitialFunctionScan(MachineFunction &Fn, // Constant pool entries can reach anything. if (I->getOpcode() == ARM::CONSTPOOL_ENTRY) continue; + if (I->getOpcode() == ARM::tLEApcrel) { + Bits = 8; // Taking the address of a CP entry. + break; + } assert(0 && "Unknown addressing mode for CP reference!"); case ARMII::AddrMode1: // AM1: 8 bits << 2 Bits = 8; @@ -339,9 +467,15 @@ void ARMConstantIslands::InitialFunctionScan(MachineFunction &Fn, } // Remember that this is a user of a CP entry. - MachineInstr *CPEMI =CPEMIs[I->getOperand(op).getConstantPoolIndex()]; + unsigned CPI = I->getOperand(op).getConstantPoolIndex(); + MachineInstr *CPEMI = CPEMIs[CPI]; unsigned MaxOffs = ((1 << Bits)-1) * Scale; CPUsers.push_back(CPUser(I, CPEMI, MaxOffs)); + + // Increment corresponding CPEntry reference count. + CPEntry *CPE = findConstPoolEntry(CPI, CPEMI); + assert(CPE && "Cannot find a corresponding CPEntry!"); + CPE->RefCount++; // Instructions can only use one CP entry, don't bother scanning the // rest of the operands. @@ -349,14 +483,17 @@ void ARMConstantIslands::InitialFunctionScan(MachineFunction &Fn, } } - // In thumb mode, if this block is a constpool island, pessmisticly assume - // it needs to be padded by two byte so it's aligned on 4 byte boundary. - if (AFI->isThumbFunction() && + // In thumb mode, if this block is a constpool island, we may need padding + // so it's aligned on 4 byte boundary. + if (isThumb && !MBB.empty() && - MBB.begin()->getOpcode() == ARM::CONSTPOOL_ENTRY) + MBB.begin()->getOpcode() == ARM::CONSTPOOL_ENTRY && + (Offset%4) != 0) MBBSize += 2; BBSizes.push_back(MBBSize); + BBOffsets.push_back(Offset); + Offset += MBBSize; } } @@ -369,11 +506,14 @@ unsigned ARMConstantIslands::GetOffsetOf(MachineInstr *MI) const { // The offset is composed of two things: the sum of the sizes of all MBB's // before this instruction's block, and the offset from the start of the block // it is in. - unsigned Offset = 0; - - // Sum block sizes before MBB. - for (unsigned BB = 0, e = MBB->getNumber(); BB != e; ++BB) - Offset += BBSizes[BB]; + unsigned Offset = BBOffsets[MBB->getNumber()]; + + // If we're looking for a CONSTPOOL_ENTRY in Thumb, see if this block has + // alignment padding, and compensate if so. + if (isThumb && + MI->getOpcode() == ARM::CONSTPOOL_ENTRY && + Offset%4 != 0) + Offset += 2; // Sum instructions before MI in MBB. for (MachineBasicBlock::iterator I = MBB->begin(); ; ++I) { @@ -383,18 +523,6 @@ unsigned ARMConstantIslands::GetOffsetOf(MachineInstr *MI) const { } } -/// GetOffsetOf - Return the current offset of the specified machine BB -/// from the start of the function. This offset changes as stuff is moved -/// around inside the function. -unsigned ARMConstantIslands::GetOffsetOf(MachineBasicBlock *MBB) const { - // Sum block sizes before MBB. - unsigned Offset = 0; - for (unsigned BB = 0, e = MBB->getNumber(); BB != e; ++BB) - Offset += BBSizes[BB]; - - return Offset; -} - /// CompareMBBNumbers - Little predicate function to sort the WaterList by MBB /// ID. static bool CompareMBBNumbers(const MachineBasicBlock *LHS, @@ -412,6 +540,9 @@ void ARMConstantIslands::UpdateForInsertedWaterBlock(MachineBasicBlock *NewBB) { // Insert a size into BBSizes to align it properly with the (newly // renumbered) block numbers. BBSizes.insert(BBSizes.begin()+NewBB->getNumber(), 0); + + // Likewise for BBOffsets. + BBOffsets.insert(BBOffsets.begin()+NewBB->getNumber(), 0); // Next, update WaterList. Specifically, we need to add NewMBB as having // available water after it. @@ -427,7 +558,6 @@ void ARMConstantIslands::UpdateForInsertedWaterBlock(MachineBasicBlock *NewBB) { /// account for this change and returns the newly created block. MachineBasicBlock *ARMConstantIslands::SplitBlockBeforeInstr(MachineInstr *MI) { MachineBasicBlock *OrigBB = MI->getParent(); - bool isThumb = AFI->isThumbFunction(); // Create a new MBB for the code after the OrigBB. MachineBasicBlock *NewBB = new MachineBasicBlock(OrigBB->getBasicBlock()); @@ -458,106 +588,472 @@ MachineBasicBlock *ARMConstantIslands::SplitBlockBeforeInstr(MachineInstr *MI) { OrigBB->addSuccessor(NewBB); // Update internal data structures to account for the newly inserted MBB. - UpdateForInsertedWaterBlock(NewBB); + // This is almost the same as UpdateForInsertedWaterBlock, except that + // the Water goes after OrigBB, not NewBB. + NewBB->getParent()->RenumberBlocks(NewBB); + + // Insert a size into BBSizes to align it properly with the (newly + // renumbered) block numbers. + BBSizes.insert(BBSizes.begin()+NewBB->getNumber(), 0); - // Figure out how large the first NewMBB is. + // Likewise for BBOffsets. + BBOffsets.insert(BBOffsets.begin()+NewBB->getNumber(), 0); + + // Next, update WaterList. Specifically, we need to add OrigMBB as having + // available water after it (but not if it's already there, which happens + // when splitting before a conditional branch that is followed by an + // unconditional branch - in that case we want to insert NewBB). + std::vector::iterator IP = + std::lower_bound(WaterList.begin(), WaterList.end(), OrigBB, + CompareMBBNumbers); + MachineBasicBlock* WaterBB = *IP; + if (WaterBB == OrigBB) + WaterList.insert(next(IP), NewBB); + else + WaterList.insert(IP, OrigBB); + + // Figure out how large the first NewMBB is. (It cannot + // contain a constpool_entry or tablejump.) unsigned NewBBSize = 0; for (MachineBasicBlock::iterator I = NewBB->begin(), E = NewBB->end(); I != E; ++I) NewBBSize += ARM::GetInstSize(I); + unsigned OrigBBI = OrigBB->getNumber(); + unsigned NewBBI = NewBB->getNumber(); // Set the size of NewBB in BBSizes. - BBSizes[NewBB->getNumber()] = NewBBSize; + BBSizes[NewBBI] = NewBBSize; // We removed instructions from UserMBB, subtract that off from its size. // Add 2 or 4 to the block to count the unconditional branch we added to it. - BBSizes[OrigBB->getNumber()] -= NewBBSize - (isThumb ? 2 : 4); + unsigned delta = isThumb ? 2 : 4; + BBSizes[OrigBBI] -= NewBBSize - delta; + + // ...and adjust BBOffsets for NewBB accordingly. + BBOffsets[NewBBI] = BBOffsets[OrigBBI] + BBSizes[OrigBBI]; + + // All BBOffsets following these blocks must be modified. + AdjustBBOffsetsAfter(NewBB, delta); return NewBB; } -/// CPEIsInRange - Returns true is the distance between specific MI and -/// specific ConstPool entry instruction can fit in MI's displacement field. -bool ARMConstantIslands::CPEIsInRange(MachineInstr *MI, MachineInstr *CPEMI, - unsigned MaxDisp) { - unsigned PCAdj = AFI->isThumbFunction() ? 4 : 8; - unsigned UserOffset = GetOffsetOf(MI) + PCAdj; - // In thumb mode, pessmisticly assumes the .align 2 before the first CPE - // in the island adds two byte padding. - unsigned AlignAdj = AFI->isThumbFunction() ? 2 : 0; - unsigned CPEOffset = GetOffsetOf(CPEMI) + AlignAdj; - - DEBUG(std::cerr << "User of CPE#" << CPEMI->getOperand(0).getImm() - << " max delta=" << MaxDisp - << " at offset " << int(UserOffset-CPEOffset) << "\t" - << *MI); - - if (UserOffset <= CPEOffset) { - // User before the CPE. - if (CPEOffset-UserOffset <= MaxDisp) +/// OffsetIsInRange - Checks whether UserOffset (the location of a constant pool +/// reference) is within MaxDisp of TrialOffset (a proposed location of a +/// constant pool entry). +bool ARMConstantIslands::OffsetIsInRange(unsigned UserOffset, + unsigned TrialOffset, unsigned MaxDisp, bool NegativeOK) { + // On Thumb offsets==2 mod 4 are rounded down by the hardware for + // purposes of the displacement computation; compensate for that here. + // Effectively, the valid range of displacements is 2 bytes smaller for such + // references. + if (isThumb && UserOffset%4 !=0) + UserOffset -= 2; + // CPEs will be rounded up to a multiple of 4. + if (isThumb && TrialOffset%4 != 0) + TrialOffset += 2; + + if (UserOffset <= TrialOffset) { + // User before the Trial. + if (TrialOffset-UserOffset <= MaxDisp) return true; - } else if (!AFI->isThumbFunction()) { - // Thumb LDR cannot encode negative offset. - if (UserOffset-CPEOffset <= MaxDisp) + } else if (NegativeOK) { + if (UserOffset-TrialOffset <= MaxDisp) return true; } return false; } -/// HandleConstantPoolUser - Analyze the specified user, checking to see if it -/// is out-of-range. If so, pick it up the constant pool value and move it some -/// place in-range. -bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &Fn, CPUser &U){ +/// WaterIsInRange - Returns true if a CPE placed after the specified +/// Water (a basic block) will be in range for the specific MI. + +bool ARMConstantIslands::WaterIsInRange(unsigned UserOffset, + MachineBasicBlock* Water, CPUser &U) +{ + unsigned MaxDisp = U.MaxDisp; + MachineFunction::iterator I = next(MachineFunction::iterator(Water)); + unsigned CPEOffset = BBOffsets[Water->getNumber()] + + BBSizes[Water->getNumber()]; + + // If the CPE is to be inserted before the instruction, that will raise + // the offset of the instruction. (Currently applies only to ARM, so + // no alignment compensation attempted here.) + if (CPEOffset < UserOffset) + UserOffset += U.CPEMI->getOperand(2).getImm(); + + return OffsetIsInRange (UserOffset, CPEOffset, MaxDisp, !isThumb); +} + +/// CPEIsInRange - Returns true if the distance between specific MI and +/// specific ConstPool entry instruction can fit in MI's displacement field. +bool ARMConstantIslands::CPEIsInRange(MachineInstr *MI, unsigned UserOffset, + MachineInstr *CPEMI, + unsigned MaxDisp, bool DoDump) { + unsigned CPEOffset = GetOffsetOf(CPEMI); + assert(CPEOffset%4 == 0 && "Misaligned CPE"); + + if (DoDump) { + DOUT << "User of CPE#" << CPEMI->getOperand(0).getImm() + << " max delta=" << MaxDisp + << " insn address=" << UserOffset + << " CPE address=" << CPEOffset + << " offset=" << int(CPEOffset-UserOffset) << "\t" << *MI; + } + + return OffsetIsInRange(UserOffset, CPEOffset, MaxDisp, !isThumb); +} + +/// BBIsJumpedOver - Return true of the specified basic block's only predecessor +/// unconditionally branches to its only successor. +static bool BBIsJumpedOver(MachineBasicBlock *MBB) { + if (MBB->pred_size() != 1 || MBB->succ_size() != 1) + return false; + + MachineBasicBlock *Succ = *MBB->succ_begin(); + MachineBasicBlock *Pred = *MBB->pred_begin(); + MachineInstr *PredMI = &Pred->back(); + if (PredMI->getOpcode() == ARM::B || PredMI->getOpcode() == ARM::tB) + return PredMI->getOperand(0).getMBB() == Succ; + return false; +} + +void ARMConstantIslands::AdjustBBOffsetsAfter(MachineBasicBlock *BB, + int delta) { + MachineFunction::iterator MBBI = BB; MBBI = next(MBBI); + for(unsigned i=BB->getNumber()+1; igetParent()->getNumBlockIDs(); i++) { + BBOffsets[i] += delta; + // If some existing blocks have padding, adjust the padding as needed, a + // bit tricky. delta can be negative so don't use % on that. + if (isThumb) { + MachineBasicBlock *MBB = MBBI; + if (!MBB->empty()) { + // Constant pool entries require padding. + if (MBB->begin()->getOpcode() == ARM::CONSTPOOL_ENTRY) { + unsigned oldOffset = BBOffsets[i] - delta; + if (oldOffset%4==0 && BBOffsets[i]%4!=0) { + // add new padding + BBSizes[i] += 2; + delta += 2; + } else if (oldOffset%4!=0 && BBOffsets[i]%4==0) { + // remove existing padding + BBSizes[i] -=2; + delta -= 2; + } + } + // Thumb jump tables require padding. They should be at the end; + // following unconditional branches are removed by AnalyzeBranch. + MachineInstr *ThumbJTMI = NULL; + if (prior(MBB->end())->getOpcode() == ARM::tBR_JTr) + ThumbJTMI = prior(MBB->end()); + if (ThumbJTMI) { + unsigned newMIOffset = GetOffsetOf(ThumbJTMI); + unsigned oldMIOffset = newMIOffset - delta; + if (oldMIOffset%4 == 0 && newMIOffset%4 != 0) { + // remove existing padding + BBSizes[i] -= 2; + delta -= 2; + } else if (oldMIOffset%4 != 0 && newMIOffset%4 == 0) { + // add new padding + BBSizes[i] += 2; + delta += 2; + } + } + if (delta==0) + return; + } + MBBI = next(MBBI); + } + } +} + +/// DecrementOldEntry - find the constant pool entry with index CPI +/// and instruction CPEMI, and decrement its refcount. If the refcount +/// becomes 0 remove the entry and instruction. Returns true if we removed +/// the entry, false if we didn't. + +bool ARMConstantIslands::DecrementOldEntry(unsigned CPI, MachineInstr *CPEMI) { + // Find the old entry. Eliminate it if it is no longer used. + CPEntry *CPE = findConstPoolEntry(CPI, CPEMI); + assert(CPE && "Unexpected!"); + if (--CPE->RefCount == 0) { + RemoveDeadCPEMI(CPEMI); + CPE->CPEMI = NULL; + NumCPEs--; + return true; + } + return false; +} + +/// LookForCPEntryInRange - see if the currently referenced CPE is in range; +/// if not, see if an in-range clone of the CPE is in range, and if so, +/// change the data structures so the user references the clone. Returns: +/// 0 = no existing entry found +/// 1 = entry found, and there were no code insertions or deletions +/// 2 = entry found, and there were code insertions or deletions +int ARMConstantIslands::LookForExistingCPEntry(CPUser& U, unsigned UserOffset) +{ MachineInstr *UserMI = U.MI; MachineInstr *CPEMI = U.CPEMI; // Check to see if the CPE is already in-range. - if (CPEIsInRange(UserMI, CPEMI, U.MaxDisp)) - return false; + if (CPEIsInRange(UserMI, UserOffset, CPEMI, U.MaxDisp, true)) { + DOUT << "In range\n"; + return 1; + } - // Solution guaranteed to work: split the user's MBB right after the user and - // insert a clone the CPE into the newly created water. + // No. Look for previously created clones of the CPE that are in range. + unsigned CPI = CPEMI->getOperand(1).getConstantPoolIndex(); + std::vector &CPEs = CPEntries[CPI]; + for (unsigned i = 0, e = CPEs.size(); i != e; ++i) { + // We already tried this one + if (CPEs[i].CPEMI == CPEMI) + continue; + // Removing CPEs can leave empty entries, skip + if (CPEs[i].CPEMI == NULL) + continue; + if (CPEIsInRange(UserMI, UserOffset, CPEs[i].CPEMI, U.MaxDisp, false)) { + DOUT << "Replacing CPE#" << CPI << " with CPE#" << CPEs[i].CPI << "\n"; + // Point the CPUser node to the replacement + U.CPEMI = CPEs[i].CPEMI; + // Change the CPI in the instruction operand to refer to the clone. + for (unsigned j = 0, e = UserMI->getNumOperands(); j != e; ++j) + if (UserMI->getOperand(j).isConstantPoolIndex()) { + UserMI->getOperand(j).setConstantPoolIndex(CPEs[i].CPI); + break; + } + // Adjust the refcount of the clone... + CPEs[i].RefCount++; + // ...and the original. If we didn't remove the old entry, none of the + // addresses changed, so we don't need another pass. + return DecrementOldEntry(CPI, CPEMI) ? 2 : 1; + } + } + return 0; +} - MachineBasicBlock *UserMBB = UserMI->getParent(); - MachineBasicBlock *NewMBB; +/// getUnconditionalBrDisp - Returns the maximum displacement that can fit in +/// the specific unconditional branch instruction. +static inline unsigned getUnconditionalBrDisp(int Opc) { + return (Opc == ARM::tB) ? ((1<<10)-1)*2 : ((1<<23)-1)*4; +} + +/// AcceptWater - Small amount of common code factored out of the following. + +MachineBasicBlock* ARMConstantIslands::AcceptWater(MachineBasicBlock *WaterBB, + std::vector::iterator IP) { + DOUT << "found water in range\n"; + // Remove the original WaterList entry; we want subsequent + // insertions in this vicinity to go after the one we're + // about to insert. This considerably reduces the number + // of times we have to move the same CPE more than once. + WaterList.erase(IP); + // CPE goes before following block (NewMBB). + return next(MachineFunction::iterator(WaterBB)); +} - // TODO: Search for the best place to split the code. In practice, using - // loop nesting information to insert these guys outside of loops would be - // sufficient. - bool isThumb = AFI->isThumbFunction(); - if (&UserMBB->back() == UserMI) { - assert(BBHasFallthrough(UserMBB) && "Expected a fallthrough BB!"); - NewMBB = next(MachineFunction::iterator(UserMBB)); +/// LookForWater - look for an existing entry in the WaterList in which +/// we can place the CPE referenced from U so it's within range of U's MI. +/// Returns true if found, false if not. If it returns true, *NewMBB +/// is set to the WaterList entry. +/// For ARM, we prefer the water that's farthest away. For Thumb, prefer +/// water that will not introduce padding to water that will; within each +/// group, prefer the water that's farthest away. + +bool ARMConstantIslands::LookForWater(CPUser &U, unsigned UserOffset, + MachineBasicBlock** NewMBB) { + std::vector::iterator IPThatWouldPad; + MachineBasicBlock* WaterBBThatWouldPad = NULL; + if (!WaterList.empty()) { + for (std::vector::iterator IP = prior(WaterList.end()), + B = WaterList.begin();; --IP) { + MachineBasicBlock* WaterBB = *IP; + if (WaterIsInRange(UserOffset, WaterBB, U)) { + if (isThumb && + (BBOffsets[WaterBB->getNumber()] + + BBSizes[WaterBB->getNumber()])%4 != 0) { + // This is valid Water, but would introduce padding. Remember + // it in case we don't find any Water that doesn't do this. + if (!WaterBBThatWouldPad) { + WaterBBThatWouldPad = WaterBB; + IPThatWouldPad = IP; + } + } else { + *NewMBB = AcceptWater(WaterBB, IP); + return true; + } + } + if (IP == B) + break; + } + } + if (isThumb && WaterBBThatWouldPad) { + *NewMBB = AcceptWater(WaterBBThatWouldPad, IPThatWouldPad); + return true; + } + return false; +} + +/// CreateNewWater - No existing WaterList entry will work for +/// CPUsers[CPUserIndex], so create a place to put the CPE. The end of the +/// block is used if in range, and the conditional branch munged so control +/// flow is correct. Otherwise the block is split to create a hole with an +/// unconditional branch around it. In either case *NewMBB is set to a +/// block following which the new island can be inserted (the WaterList +/// is not adjusted). + +void ARMConstantIslands::CreateNewWater(unsigned CPUserIndex, + unsigned UserOffset, MachineBasicBlock** NewMBB) { + CPUser &U = CPUsers[CPUserIndex]; + MachineInstr *UserMI = U.MI; + MachineInstr *CPEMI = U.CPEMI; + MachineBasicBlock *UserMBB = UserMI->getParent(); + unsigned OffsetOfNextBlock = BBOffsets[UserMBB->getNumber()] + + BBSizes[UserMBB->getNumber()]; + assert(OffsetOfNextBlock== BBOffsets[UserMBB->getNumber()+1]); + + // If the use is at the end of the block, or the end of the block + // is within range, make new water there. (The addition below is + // for the unconditional branch we will be adding: 4 bytes on ARM, + // 2 on Thumb. Possible Thumb alignment padding is allowed for + // inside OffsetIsInRange. + // If the block ends in an unconditional branch already, it is water, + // and is known to be out of range, so we'll always be adding a branch.) + if (&UserMBB->back() == UserMI || + OffsetIsInRange(UserOffset, OffsetOfNextBlock + (isThumb ? 2: 4), + U.MaxDisp, !isThumb)) { + DOUT << "Split at end of block\n"; + if (&UserMBB->back() == UserMI) + assert(BBHasFallthrough(UserMBB) && "Expected a fallthrough BB!"); + *NewMBB = next(MachineFunction::iterator(UserMBB)); // Add an unconditional branch from UserMBB to fallthrough block. - // Note the new unconditional branch is not being recorded. - BuildMI(UserMBB, TII->get(isThumb ? ARM::tB : ARM::B)).addMBB(NewMBB); - BBSizes[UserMBB->getNumber()] += isThumb ? 2 : 4; + // Record it for branch lengthening; this new branch will not get out of + // range, but if the preceding conditional branch is out of range, the + // targets will be exchanged, and the altered branch may be out of + // range, so the machinery has to know about it. + int UncondBr = isThumb ? ARM::tB : ARM::B; + BuildMI(UserMBB, TII->get(UncondBr)).addMBB(*NewMBB); + unsigned MaxDisp = getUnconditionalBrDisp(UncondBr); + ImmBranches.push_back(ImmBranch(&UserMBB->back(), + MaxDisp, false, UncondBr)); + int delta = isThumb ? 2 : 4; + BBSizes[UserMBB->getNumber()] += delta; + AdjustBBOffsetsAfter(UserMBB, delta); } else { - MachineInstr *NextMI = next(MachineBasicBlock::iterator(UserMI)); - NewMBB = SplitBlockBeforeInstr(NextMI); + // What a big block. Find a place within the block to split it. + // This is a little tricky on Thumb since instructions are 2 bytes + // and constant pool entries are 4 bytes: if instruction I references + // island CPE, and instruction I+1 references CPE', it will + // not work well to put CPE as far forward as possible, since then + // CPE' cannot immediately follow it (that location is 2 bytes + // farther away from I+1 than CPE was from I) and we'd need to create + // a new island. So, we make a first guess, then walk through the + // instructions between the one currently being looked at and the + // possible insertion point, and make sure any other instructions + // that reference CPEs will be able to use the same island area; + // if not, we back up the insertion point. + + // The 4 in the following is for the unconditional branch we'll be + // inserting (allows for long branch on Thumb). Alignment of the + // island is handled inside OffsetIsInRange. + unsigned BaseInsertOffset = UserOffset + U.MaxDisp -4; + // This could point off the end of the block if we've already got + // constant pool entries following this block; only the last one is + // in the water list. Back past any possible branches (allow for a + // conditional and a maximally long unconditional). + if (BaseInsertOffset >= BBOffsets[UserMBB->getNumber()+1]) + BaseInsertOffset = BBOffsets[UserMBB->getNumber()+1] - + (isThumb ? 6 : 8); + unsigned EndInsertOffset = BaseInsertOffset + + CPEMI->getOperand(2).getImm(); + MachineBasicBlock::iterator MI = UserMI; + ++MI; + unsigned CPUIndex = CPUserIndex+1; + for (unsigned Offset = UserOffset+ARM::GetInstSize(UserMI); + Offset < BaseInsertOffset; + Offset += ARM::GetInstSize(MI), + MI = next(MI)) { + if (CPUIndex < CPUsers.size() && CPUsers[CPUIndex].MI == MI) { + if (!OffsetIsInRange(Offset, EndInsertOffset, + CPUsers[CPUIndex].MaxDisp, !isThumb)) { + BaseInsertOffset -= (isThumb ? 2 : 4); + EndInsertOffset -= (isThumb ? 2 : 4); + } + // This is overly conservative, as we don't account for CPEMIs + // being reused within the block, but it doesn't matter much. + EndInsertOffset += CPUsers[CPUIndex].CPEMI->getOperand(2).getImm(); + CPUIndex++; + } + } + DOUT << "Split in middle of big block\n"; + *NewMBB = SplitBlockBeforeInstr(prior(MI)); + } +} + +/// HandleConstantPoolUser - Analyze the specified user, checking to see if it +/// is out-of-range. If so, pick it up the constant pool value and move it some +/// place in-range. Return true if we changed any addresses (thus must run +/// another pass of branch lengthening), false otherwise. +bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &Fn, + unsigned CPUserIndex){ + CPUser &U = CPUsers[CPUserIndex]; + MachineInstr *UserMI = U.MI; + MachineInstr *CPEMI = U.CPEMI; + unsigned CPI = CPEMI->getOperand(1).getConstantPoolIndex(); + unsigned Size = CPEMI->getOperand(2).getImm(); + MachineBasicBlock *NewMBB; + // Compute this only once, it's expensive. The 4 or 8 is the value the + // hardware keeps in the PC (2 insns ahead of the reference). + unsigned UserOffset = GetOffsetOf(UserMI) + (isThumb ? 4 : 8); + + // Special case: tLEApcrel are two instructions MI's. The actual user is the + // second instruction. + if (UserMI->getOpcode() == ARM::tLEApcrel) + UserOffset += 2; + + // See if the current entry is within range, or there is a clone of it + // in range. + int result = LookForExistingCPEntry(U, UserOffset); + if (result==1) return false; + else if (result==2) return true; + + // No existing clone of this CPE is within range. + // We will be generating a new clone. Get a UID for it. + unsigned ID = NextUID++; + + // Look for water where we can place this CPE. We look for the farthest one + // away that will work. Forward references only for now (although later + // we might find some that are backwards). + + if (!LookForWater(U, UserOffset, &NewMBB)) { + // No water found. + DOUT << "No water found\n"; + CreateNewWater(CPUserIndex, UserOffset, &NewMBB); } - // Okay, we know we can put an island before UserMBB now, do it! + // Okay, we know we can put an island before NewMBB now, do it! MachineBasicBlock *NewIsland = new MachineBasicBlock(); Fn.getBasicBlockList().insert(NewMBB, NewIsland); // Update internal data structures to account for the newly inserted MBB. UpdateForInsertedWaterBlock(NewIsland); + // Decrement the old entry, and remove it if refcount becomes 0. + DecrementOldEntry(CPI, CPEMI); + // Now that we have an island to add the CPE to, clone the original CPE and // add it to the island. - unsigned ID = NextUID++; - unsigned CPI = CPEMI->getOperand(1).getConstantPoolIndex(); - unsigned Size = CPEMI->getOperand(2).getImm(); - - // Build a new CPE for this user. U.CPEMI = BuildMI(NewIsland, TII->get(ARM::CONSTPOOL_ENTRY)) .addImm(ID).addConstantPoolIndex(CPI).addImm(Size); - + CPEntries[CPI].push_back(CPEntry(U.CPEMI, ID, 1)); + NumCPEs++; + + BBOffsets[NewIsland->getNumber()] = BBOffsets[NewMBB->getNumber()]; // Compensate for .align 2 in thumb mode. - if (isThumb) Size += 2; + if (isThumb && BBOffsets[NewIsland->getNumber()]%4 != 0) + Size += 2; // Increase the size of the island block to account for the new entry. BBSizes[NewIsland->getNumber()] += Size; + AdjustBBOffsetsAfter(NewIsland, Size); // Finally, change the CPI in the instruction operand to be ID. for (unsigned i = 0, e = UserMI->getNumOperands(); i != e; ++i) @@ -566,30 +1062,74 @@ bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &Fn, CPUser &U){ break; } - DEBUG(std::cerr << " Moved CPE to #" << ID << " CPI=" << CPI << "\t" - << *UserMI); + DOUT << " Moved CPE to #" << ID << " CPI=" << CPI << "\t" << *UserMI; return true; } -/// BBIsInRange - Returns true is the distance between specific MI and +/// RemoveDeadCPEMI - Remove a dead constant pool entry instruction. Update +/// sizes and offsets of impacted basic blocks. +void ARMConstantIslands::RemoveDeadCPEMI(MachineInstr *CPEMI) { + MachineBasicBlock *CPEBB = CPEMI->getParent(); + unsigned Size = CPEMI->getOperand(2).getImm(); + CPEMI->eraseFromParent(); + BBSizes[CPEBB->getNumber()] -= Size; + // All succeeding offsets have the current size value added in, fix this. + if (CPEBB->empty()) { + // In thumb mode, the size of island may be padded by two to compensate for + // the alignment requirement. Then it will now be 2 when the block is + // empty, so fix this. + // All succeeding offsets have the current size value added in, fix this. + if (BBSizes[CPEBB->getNumber()] != 0) { + Size += BBSizes[CPEBB->getNumber()]; + BBSizes[CPEBB->getNumber()] = 0; + } + } + AdjustBBOffsetsAfter(CPEBB, -Size); + // An island has only one predecessor BB and one successor BB. Check if + // this BB's predecessor jumps directly to this BB's successor. This + // shouldn't happen currently. + assert(!BBIsJumpedOver(CPEBB) && "How did this happen?"); + // FIXME: remove the empty blocks after all the work is done? +} + +/// RemoveUnusedCPEntries - Remove constant pool entries whose refcounts +/// are zero. +bool ARMConstantIslands::RemoveUnusedCPEntries() { + unsigned MadeChange = false; + for (unsigned i = 0, e = CPEntries.size(); i != e; ++i) { + std::vector &CPEs = CPEntries[i]; + for (unsigned j = 0, ee = CPEs.size(); j != ee; ++j) { + if (CPEs[j].RefCount == 0 && CPEs[j].CPEMI) { + RemoveDeadCPEMI(CPEs[j].CPEMI); + CPEs[j].CPEMI = NULL; + MadeChange = true; + } + } + } + return MadeChange; +} + +/// BBIsInRange - Returns true if the distance between specific MI and /// specific BB can fit in MI's displacement field. bool ARMConstantIslands::BBIsInRange(MachineInstr *MI,MachineBasicBlock *DestBB, unsigned MaxDisp) { - unsigned PCAdj = AFI->isThumbFunction() ? 4 : 8; + unsigned PCAdj = isThumb ? 4 : 8; unsigned BrOffset = GetOffsetOf(MI) + PCAdj; - unsigned DestOffset = GetOffsetOf(DestBB); + unsigned DestOffset = BBOffsets[DestBB->getNumber()]; - DEBUG(std::cerr << "Branch of destination BB#" << DestBB->getNumber() - << " max delta=" << MaxDisp - << " at offset " << int(BrOffset-DestOffset) << "\t" - << *MI); + DOUT << "Branch of destination BB#" << DestBB->getNumber() + << " from BB#" << MI->getParent()->getNumber() + << " max delta=" << MaxDisp + << " from " << GetOffsetOf(MI) << " to " << DestOffset + << " offset " << int(DestOffset-BrOffset) << "\t" << *MI; if (BrOffset <= DestOffset) { - if (DestOffset - BrOffset <= MaxDisp) + // Branch before the Dest. + if (DestOffset-BrOffset <= MaxDisp) return true; } else { - if (BrOffset - DestOffset <= MaxDisp) + if (BrOffset-DestOffset <= MaxDisp) return true; } return false; @@ -610,32 +1150,30 @@ bool ARMConstantIslands::FixUpImmediateBr(MachineFunction &Fn, ImmBranch &Br) { return FixUpConditionalBr(Fn, Br); } -/// FixUpUnconditionalBr - Fix up an unconditional branches whose destination is -/// too far away to fit in its displacement field. If LR register has been +/// FixUpUnconditionalBr - Fix up an unconditional branch whose destination is +/// too far away to fit in its displacement field. If the LR register has been /// spilled in the epilogue, then we can use BL to implement a far jump. -/// Otherwise, add a intermediate branch instruction to to a branch. +/// Otherwise, add an intermediate branch instruction to to a branch. bool ARMConstantIslands::FixUpUnconditionalBr(MachineFunction &Fn, ImmBranch &Br) { MachineInstr *MI = Br.MI; MachineBasicBlock *MBB = MI->getParent(); - assert(AFI->isThumbFunction() && "Expected a Thumb function!"); + assert(isThumb && "Expected a Thumb function!"); // Use BL to implement far jump. Br.MaxDisp = (1 << 21) * 2; MI->setInstrDescriptor(TII->get(ARM::tBfar)); BBSizes[MBB->getNumber()] += 2; + AdjustBBOffsetsAfter(MBB, 2); HasFarJump = true; NumUBrFixed++; - return true; -} -/// getUnconditionalBrDisp - Returns the maximum displacement that can fit in the -/// specific unconditional branch instruction. -static inline unsigned getUnconditionalBrDisp(int Opc) { - return (Opc == ARM::tB) ? (1<<10)*2 : (1<<23)*4; + DOUT << " Changed B to long jump " << *MI; + + return true; } -/// FixUpConditionalBr - Fix up a conditional branches whose destination is too +/// FixUpConditionalBr - Fix up a conditional branch whose destination is too /// far away to fit in its displacement field. It is converted to an inverse /// conditional branch + an unconditional branch to the destination. bool @@ -652,18 +1190,19 @@ ARMConstantIslands::FixUpConditionalBr(MachineFunction &Fn, ImmBranch &Br) { // L2: ARMCC::CondCodes CC = (ARMCC::CondCodes)MI->getOperand(1).getImmedValue(); CC = ARMCC::getOppositeCondition(CC); + unsigned CCReg = MI->getOperand(2).getReg(); // If the branch is at the end of its MBB and that has a fall-through block, // direct the updated conditional branch to the fall-through block. Otherwise, // split the MBB before the next instruction. MachineBasicBlock *MBB = MI->getParent(); - MachineInstr *BackMI = &MBB->back(); - bool NeedSplit = (BackMI != MI) || !BBHasFallthrough(MBB); + MachineInstr *BMI = &MBB->back(); + bool NeedSplit = (BMI != MI) || !BBHasFallthrough(MBB); NumCBrFixed++; - if (BackMI != MI) { + if (BMI != MI) { if (next(MachineBasicBlock::iterator(MI)) == MBB->back() && - BackMI->getOpcode() == Br.UncondBr) { + BMI->getOpcode() == Br.UncondBr) { // Last MI in the BB is a unconditional branch. Can we simply invert the // condition and swap destinations: // beq L1 @@ -671,9 +1210,10 @@ ARMConstantIslands::FixUpConditionalBr(MachineFunction &Fn, ImmBranch &Br) { // => // bne L2 // b L1 - MachineBasicBlock *NewDest = BackMI->getOperand(0).getMachineBasicBlock(); + MachineBasicBlock *NewDest = BMI->getOperand(0).getMachineBasicBlock(); if (BBIsInRange(MI, NewDest, Br.MaxDisp)) { - BackMI->getOperand(0).setMachineBasicBlock(DestBB); + DOUT << " Invert Bcc condition and swap its destination with " << *BMI; + BMI->getOperand(0).setMachineBasicBlock(DestBB); MI->getOperand(0).setMachineBasicBlock(NewDest); MI->getOperand(1).setImm(CC); return true; @@ -685,37 +1225,52 @@ ARMConstantIslands::FixUpConditionalBr(MachineFunction &Fn, ImmBranch &Br) { SplitBlockBeforeInstr(MI); // No need for the branch to the next block. We're adding a unconditional // branch to the destination. + int delta = ARM::GetInstSize(&MBB->back()); + BBSizes[MBB->getNumber()] -= delta; + MachineBasicBlock* SplitBB = next(MachineFunction::iterator(MBB)); + AdjustBBOffsetsAfter(SplitBB, -delta); MBB->back().eraseFromParent(); + // BBOffsets[SplitBB] is wrong temporarily, fixed below } MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); + + DOUT << " Insert B to BB#" << DestBB->getNumber() + << " also invert condition and change dest. to BB#" + << NextBB->getNumber() << "\n"; - // Insert a unconditional branch and replace the conditional branch. + // Insert a new conditional branch and a new unconditional branch. // Also update the ImmBranch as well as adding a new entry for the new branch. - BuildMI(MBB, TII->get(MI->getOpcode())).addMBB(NextBB).addImm(CC); + BuildMI(MBB, TII->get(MI->getOpcode())).addMBB(NextBB) + .addImm(CC).addReg(CCReg); Br.MI = &MBB->back(); + BBSizes[MBB->getNumber()] += ARM::GetInstSize(&MBB->back()); BuildMI(MBB, TII->get(Br.UncondBr)).addMBB(DestBB); + BBSizes[MBB->getNumber()] += ARM::GetInstSize(&MBB->back()); unsigned MaxDisp = getUnconditionalBrDisp(Br.UncondBr); ImmBranches.push_back(ImmBranch(&MBB->back(), MaxDisp, false, Br.UncondBr)); + + // Remove the old conditional branch. It may or may not still be in MBB. + BBSizes[MI->getParent()->getNumber()] -= ARM::GetInstSize(MI); MI->eraseFromParent(); - // Increase the size of MBB to account for the new unconditional branch. - BBSizes[MBB->getNumber()] += ARM::GetInstSize(&MBB->back()); + // The net size change is an addition of one unconditional branch. + int delta = ARM::GetInstSize(&MBB->back()); + AdjustBBOffsetsAfter(MBB, delta); return true; } - /// UndoLRSpillRestore - Remove Thumb push / pop instructions that only spills /// LR / restores LR to pc. bool ARMConstantIslands::UndoLRSpillRestore() { bool MadeChange = false; for (unsigned i = 0, e = PushPopMIs.size(); i != e; ++i) { MachineInstr *MI = PushPopMIs[i]; - if (MI->getNumOperands() == 1) { - if (MI->getOpcode() == ARM::tPOP_RET && - MI->getOperand(0).getReg() == ARM::PC) - BuildMI(MI->getParent(), TII->get(ARM::tBX_RET)); - MI->eraseFromParent(); - MadeChange = true; + if (MI->getOpcode() == ARM::tPOP_RET && + MI->getOperand(0).getReg() == ARM::PC && + MI->getNumExplicitOperands() == 1) { + BuildMI(MI->getParent(), TII->get(ARM::tBX_RET)); + MI->eraseFromParent(); + MadeChange = true; } } return MadeChange;