X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FARM%2FARMConstantIslandPass.cpp;h=9aced30f0136e2e038e62558585ed70a63500de4;hb=4ee451de366474b9c228b4e5fa573795a715216d;hp=8fe8167abaf5f736356fc2e357a965e23fab291b;hpb=ce74de46dd2b1ca040bc5d2e17af64832bc0b633;p=oota-llvm.git diff --git a/lib/Target/ARM/ARMConstantIslandPass.cpp b/lib/Target/ARM/ARMConstantIslandPass.cpp index 8fe8167abaf..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. // //===----------------------------------------------------------------------===// // @@ -51,10 +51,14 @@ namespace { 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 @@ -108,7 +112,7 @@ 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; @@ -121,8 +125,12 @@ namespace { 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 { @@ -138,14 +146,22 @@ namespace { MachineBasicBlock *SplitBlockBeforeInstr(MachineInstr *MI); void UpdateForInsertedWaterBlock(MachineBasicBlock *NewBB); void AdjustBBOffsetsAfter(MachineBasicBlock *BB, int delta); - bool DecrementOldEntry(unsigned CPI, MachineInstr* CPEMI, unsigned Size); + bool DecrementOldEntry(unsigned CPI, MachineInstr* CPEMI); int LookForExistingCPEntry(CPUser& U, unsigned UserOffset); - bool HandleConstantPoolUser(MachineFunction &Fn, CPUser &U); + 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, - unsigned Disp); + CPUser &U); bool OffsetIsInRange(unsigned UserOffset, unsigned TrialOffset, unsigned Disp, bool NegativeOK); bool BBIsInRange(MachineInstr *MI, MachineBasicBlock *BB, unsigned Disp); @@ -155,7 +171,37 @@ namespace { bool UndoLRSpillRestore(); unsigned GetOffsetOf(MachineInstr *MI) 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 @@ -169,6 +215,7 @@ bool ARMConstantIslands::runOnMachineFunction(MachineFunction &Fn) { TII = Fn.getTarget().getInstrInfo(); AFI = Fn.getInfo(); + isThumb = AFI->isThumbFunction(); HasFarJump = false; @@ -176,11 +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(); @@ -191,23 +245,31 @@ 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(); @@ -236,7 +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. @@ -315,8 +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; @@ -360,6 +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; @@ -410,12 +483,12 @@ void ARMConstantIslands::InitialFunctionScan(MachineFunction &Fn, } } - // In thumb mode, if this block is a constpool island, pessimistically - // 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); @@ -435,6 +508,13 @@ unsigned ARMConstantIslands::GetOffsetOf(MachineInstr *MI) const { // it is in. 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) { assert(I != MBB->end() && "Didn't find MI in its own basic block?"); @@ -478,7 +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()); @@ -533,7 +612,8 @@ MachineBasicBlock *ARMConstantIslands::SplitBlockBeforeInstr(MachineInstr *MI) { else WaterList.insert(IP, OrigBB); - // Figure out how large the first NewMBB is. + // 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) @@ -558,10 +638,21 @@ MachineBasicBlock *ARMConstantIslands::SplitBlockBeforeInstr(MachineInstr *MI) { return NewBB; } -/// OffsetIsInRange - Checks whether UserOffset is within MaxDisp of -/// TrialOffset. +/// 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) @@ -577,17 +668,18 @@ bool ARMConstantIslands::OffsetIsInRange(unsigned UserOffset, /// Water (a basic block) will be in range for the specific MI. bool ARMConstantIslands::WaterIsInRange(unsigned UserOffset, - MachineBasicBlock* Water, unsigned MaxDisp) + MachineBasicBlock* Water, CPUser &U) { - bool isThumb = AFI->isThumbFunction(); + unsigned MaxDisp = U.MaxDisp; + MachineFunction::iterator I = next(MachineFunction::iterator(Water)); unsigned CPEOffset = BBOffsets[Water->getNumber()] + BBSizes[Water->getNumber()]; - // If the Water is a constpool island, it has already been aligned. - // If not, align it. - if (isThumb && - (Water->empty() || - Water->begin()->getOpcode() != ARM::CONSTPOOL_ENTRY)) - CPEOffset += 2; + + // 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); } @@ -597,11 +689,8 @@ bool ARMConstantIslands::WaterIsInRange(unsigned UserOffset, bool ARMConstantIslands::CPEIsInRange(MachineInstr *MI, unsigned UserOffset, MachineInstr *CPEMI, unsigned MaxDisp, bool DoDump) { - // In thumb mode, pessimistically assumes the .align 2 before the first CPE - // in the island adds two byte padding. - bool isThumb = AFI->isThumbFunction(); - unsigned AlignAdj = isThumb ? 2 : 0; - unsigned CPEOffset = GetOffsetOf(CPEMI) + AlignAdj; + unsigned CPEOffset = GetOffsetOf(CPEMI); + assert(CPEOffset%4 == 0 && "Misaligned CPE"); if (DoDump) { DOUT << "User of CPE#" << CPEMI->getOperand(0).getImm() @@ -628,11 +717,53 @@ static bool BBIsJumpedOver(MachineBasicBlock *MBB) { return false; } -void ARMConstantIslands::AdjustBBOffsetsAfter(MachineBasicBlock *BB, int delta) -{ - MachineFunction::iterator MBBI = BB->getParent()->end(); - for(unsigned i=BB->getNumber()+1; igetParent()->getNumBlockIDs(); i++) +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 @@ -640,34 +771,13 @@ void ARMConstantIslands::AdjustBBOffsetsAfter(MachineBasicBlock *BB, int delta) /// 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, - unsigned Size) { +bool ARMConstantIslands::DecrementOldEntry(unsigned CPI, MachineInstr *CPEMI) { // Find the old entry. Eliminate it if it is no longer used. - CPEntry *OldCPE = findConstPoolEntry(CPI, CPEMI); - assert(OldCPE && "Unexpected!"); - if (--OldCPE->RefCount == 0) { - MachineBasicBlock *OldCPEBB = OldCPE->CPEMI->getParent(); - if (OldCPEBB->empty()) { - // In thumb mode, the size of island is padded by two to compensate for - // the alignment requirement. Thus 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[OldCPEBB->getNumber()] != 0) { - AdjustBBOffsetsAfter(OldCPEBB, -BBSizes[OldCPEBB->getNumber()]); - BBSizes[OldCPEBB->getNumber()] = 0; - } - // 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(OldCPEBB) && "How did this happen?"); - // FIXME: remove the empty blocks after all the work is done? - } else { - BBSizes[OldCPEBB->getNumber()] -= Size; - // All succeeding offsets have the current size value added in, fix this. - AdjustBBOffsetsAfter(OldCPEBB, -Size); - } - OldCPE->CPEMI->eraseFromParent(); - OldCPE->CPEMI = NULL; + CPEntry *CPE = findConstPoolEntry(CPI, CPEMI); + assert(CPE && "Unexpected!"); + if (--CPE->RefCount == 0) { + RemoveDeadCPEMI(CPEMI); + CPE->CPEMI = NULL; NumCPEs--; return true; } @@ -715,26 +825,190 @@ int ARMConstantIslands::LookForExistingCPEntry(CPUser& U, unsigned UserOffset) 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. - unsigned Size = CPEMI->getOperand(2).getImm(); - return DecrementOldEntry(CPI, CPEMI, Size) ? 2 : 1; + return DecrementOldEntry(CPI, CPEMI) ? 2 : 1; } } return 0; } +/// 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)); +} + +/// 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. + // 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 { + // 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, CPUser &U){ +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(); - bool isThumb = AFI->isThumbFunction(); MachineBasicBlock *NewMBB; - // Compute this only once, it's expensive + // 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. @@ -749,67 +1023,11 @@ bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &Fn, CPUser &U){ // 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). - bool WaterFound = false; - if (!WaterList.empty()) { - for (std::vector::iterator IP = prior(WaterList.end()), - B = WaterList.begin();; --IP) { - MachineBasicBlock* WaterBB = *IP; - if (WaterIsInRange(UserOffset, WaterBB, U.MaxDisp)) { - WaterFound = true; - DOUT << "found water in range\n"; - // CPE goes before following block (NewMBB). - NewMBB = next(MachineFunction::iterator(WaterBB)); - // 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); - break; - } - if (IP == B) - break; - } - } - if (!WaterFound) { + if (!LookForWater(U, UserOffset, &NewMBB)) { // No water found. - DOUT << "No water found\n"; - MachineBasicBlock *UserMBB = UserMI->getParent(); - unsigned TrialOffset = BBOffsets[UserMBB->getNumber()] + - BBSizes[UserMBB->getNumber()] + - isThumb ? 2 : 4; /* for branch to be added */ - - // If the use is at the end of the block, or the end of the block - // is within range, make new water there. (If the block ends in - // an unconditional branch already, it is water, and is known to - // be out of range; so it's OK to assume above we'll be adding a Br.) - if (&UserMBB->back() == UserMI || - OffsetIsInRange(UserOffset, TrialOffset, U.MaxDisp, !isThumb)) { - 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); - int delta = isThumb ? 2 : 4; - BBSizes[UserMBB->getNumber()] += delta; - AdjustBBOffsetsAfter(UserMBB, delta); - } else { - // 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. - - // Solution of last resort: split the user's MBB right after the user - // and insert a clone of the CPE into the newly created water. - MachineInstr *NextMI = next(MachineBasicBlock::iterator(UserMI)); - NewMBB = SplitBlockBeforeInstr(NextMI); - } + CreateNewWater(CPUserIndex, UserOffset, &NewMBB); } // Okay, we know we can put an island before NewMBB now, do it! @@ -820,7 +1038,7 @@ bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &Fn, CPUser &U){ UpdateForInsertedWaterBlock(NewIsland); // Decrement the old entry, and remove it if refcount becomes 0. - DecrementOldEntry(CPI, CPEMI, Size); + DecrementOldEntry(CPI, CPEMI); // Now that we have an island to add the CPE to, clone the original CPE and // add it to the island. @@ -829,11 +1047,12 @@ bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &Fn, CPUser &U){ 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; - BBOffsets[NewIsland->getNumber()] = BBOffsets[NewMBB->getNumber()]; AdjustBBOffsetsAfter(NewIsland, Size); // Finally, change the CPI in the instruction operand to be ID. @@ -848,20 +1067,72 @@ bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &Fn, CPUser &U){ return true; } +/// 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 = BBOffsets[DestBB->getNumber()]; DOUT << "Branch of destination BB#" << DestBB->getNumber() << " from BB#" << MI->getParent()->getNumber() << " max delta=" << MaxDisp - << " at offset " << int(DestOffset-BrOffset) << "\t" << *MI; + << " from " << GetOffsetOf(MI) << " to " << DestOffset + << " offset " << int(DestOffset-BrOffset) << "\t" << *MI; - return OffsetIsInRange(BrOffset, DestOffset, MaxDisp, true); + if (BrOffset <= DestOffset) { + // Branch before the Dest. + if (DestOffset-BrOffset <= MaxDisp) + return true; + } else { + if (BrOffset-DestOffset <= MaxDisp) + return true; + } + return false; } /// FixUpImmediateBr - Fix up an immediate branch whose destination is too far @@ -887,7 +1158,7 @@ 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; @@ -902,12 +1173,6 @@ ARMConstantIslands::FixUpUnconditionalBr(MachineFunction &Fn, ImmBranch &Br) { 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; -} - /// 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. @@ -925,6 +1190,7 @@ 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, @@ -959,7 +1225,12 @@ 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)); @@ -967,18 +1238,23 @@ ARMConstantIslands::FixUpConditionalBr(MachineFunction &Fn, ImmBranch &Br) { << " 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. + // The net size change is an addition of one unconditional branch. int delta = ARM::GetInstSize(&MBB->back()); - BBSizes[MBB->getNumber()] += delta; AdjustBBOffsetsAfter(MBB, delta); return true; } @@ -989,12 +1265,12 @@ 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;