X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTarget%2FARM%2FARMConstantIslandPass.cpp;h=a50ddcdd182ed504ea2582b17b6605ead6fa24a3;hp=65abff0b950139a7dbf36cd0351777548f08d12c;hb=d628f19f5df9e4033adce5af969049e90a90ae5d;hpb=b813f924a749396ffb4a4bd087ee1dbb678551f3 diff --git a/lib/Target/ARM/ARMConstantIslandPass.cpp b/lib/Target/ARM/ARMConstantIslandPass.cpp index 65abff0b950..a50ddcdd182 100644 --- a/lib/Target/ARM/ARMConstantIslandPass.cpp +++ b/lib/Target/ARM/ARMConstantIslandPass.cpp @@ -16,23 +16,23 @@ #define DEBUG_TYPE "arm-cp-islands" #include "ARM.h" #include "ARMMachineFunctionInfo.h" -#include "ARMInstrInfo.h" -#include "Thumb2InstrInfo.h" #include "MCTargetDesc/ARMAddressingModes.h" +#include "Thumb2InstrInfo.h" +#include "llvm/ADT/STLExtras.h" +#include "llvm/ADT/SmallSet.h" +#include "llvm/ADT/SmallVector.h" +#include "llvm/ADT/Statistic.h" #include "llvm/CodeGen/MachineConstantPool.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineJumpTableInfo.h" -#include "llvm/Target/TargetData.h" -#include "llvm/Target/TargetMachine.h" +#include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/IR/DataLayout.h" +#include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/Format.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/ADT/SmallSet.h" -#include "llvm/ADT/SmallVector.h" -#include "llvm/ADT/STLExtras.h" -#include "llvm/ADT/Statistic.h" -#include "llvm/Support/CommandLine.h" +#include "llvm/Target/TargetMachine.h" #include using namespace llvm; @@ -52,8 +52,9 @@ static cl::opt AdjustJumpTableBlocks("arm-adjust-jump-tables", cl::Hidden, cl::init(true), cl::desc("Adjust basic block layout to better use TB[BH]")); +// FIXME: This option should be removed once it has received sufficient testing. static cl::opt -AlignConstantIslands("arm-align-constant-island", cl::Hidden, +AlignConstantIslands("arm-align-constant-islands", cl::Hidden, cl::init(true), cl::desc("Align constant islands in code")); /// UnknownPadding - Return the worst case padding that could result from @@ -68,27 +69,6 @@ static inline unsigned UnknownPadding(unsigned LogAlign, unsigned KnownBits) { return 0; } -/// WorstCaseAlign - Assuming only the low KnownBits bits in Offset are exact, -/// add padding such that: -/// -/// 1. The result is aligned to 1 << LogAlign. -/// -/// 2. No other value of the unknown bits would require more padding. -/// -/// This may add more padding than is required to satisfy just one of the -/// constraints. It is necessary to compute alignment this way to guarantee -/// that we don't underestimate the padding before an aligned block. If the -/// real padding before a block is larger than we think, constant pool entries -/// may go out of range. -static inline unsigned WorstCaseAlign(unsigned Offset, unsigned LogAlign, - unsigned KnownBits) { - // Add the worst possible padding that the unknown bits could cause. - Offset += UnknownPadding(LogAlign, KnownBits); - - // Then align the result. - return RoundUpToAlignment(Offset, 1u << LogAlign); -} - namespace { /// ARMConstantIslands - Due to limited PC-relative displacements, ARM /// requires constant pool entries to be scattered among the instructions @@ -108,7 +88,12 @@ namespace { /// Offset - Distance from the beginning of the function to the beginning /// of this basic block. /// - /// The offset is always aligned as required by the basic block. + /// Offsets are computed assuming worst case padding before an aligned + /// block. This means that subtracting basic block offsets always gives a + /// conservative estimate of the real distance which may be smaller. + /// + /// Because worst case padding is used, the computed offset of an aligned + /// block may not actually be aligned. unsigned Offset; /// Size - Size of the basic block in bytes. If the block contains @@ -139,24 +124,34 @@ namespace { /// This number should be used to predict worst case padding when /// splitting the block. unsigned internalKnownBits() const { - return Unalign ? Unalign : KnownBits; + unsigned Bits = Unalign ? Unalign : KnownBits; + // If the block size isn't a multiple of the known bits, assume the + // worst case padding. + if (Size & ((1u << Bits) - 1)) + Bits = countTrailingZeros(Size); + return Bits; } - /// Compute the offset immediately following this block. - unsigned postOffset() const { + /// Compute the offset immediately following this block. If LogAlign is + /// specified, return the offset the successor block will get if it has + /// this alignment. + unsigned postOffset(unsigned LogAlign = 0) const { unsigned PO = Offset + Size; - if (!PostAlign) + unsigned LA = std::max(unsigned(PostAlign), LogAlign); + if (!LA) return PO; // Add alignment padding from the terminator. - return WorstCaseAlign(PO, PostAlign, internalKnownBits()); + return PO + UnknownPadding(LA, internalKnownBits()); } /// Compute the number of known low bits of postOffset. If this block /// contains inline asm, the number of known bits drops to the /// instruction alignment. An aligned terminator may increase the number /// of know bits. - unsigned postKnownBits() const { - return std::max(unsigned(PostAlign), internalKnownBits()); + /// If LogAlign is given, also consider the alignment of the next block. + unsigned postKnownBits(unsigned LogAlign = 0) const { + return std::max(std::max(unsigned(PostAlign), LogAlign), + internalKnownBits()); } }; @@ -189,14 +184,24 @@ namespace { MachineInstr *MI; MachineInstr *CPEMI; MachineBasicBlock *HighWaterMark; + private: unsigned MaxDisp; + public: bool NegOk; bool IsSoImm; + bool KnownAlignment; CPUser(MachineInstr *mi, MachineInstr *cpemi, unsigned maxdisp, bool neg, bool soimm) - : MI(mi), CPEMI(cpemi), MaxDisp(maxdisp), NegOk(neg), IsSoImm(soimm) { + : MI(mi), CPEMI(cpemi), MaxDisp(maxdisp), NegOk(neg), IsSoImm(soimm), + KnownAlignment(false) { HighWaterMark = CPEMI->getParent(); } + /// getMaxDisp - Returns the maximum displacement supported by MI. + /// Correct for unknown alignment. + /// Conservatively subtract 2 bytes to handle weird alignment effects. + unsigned getMaxDisp() const { + return (KnownAlignment ? MaxDisp : MaxDisp - 2) - 2; + } }; /// CPUsers - Keep track of all of the machine instructions that use various @@ -249,7 +254,9 @@ namespace { /// the branch fix up pass. bool HasFarJump; - const ARMInstrInfo *TII; + MachineFunction *MF; + MachineConstantPool *MCP; + const ARMBaseInstrInfo *TII; const ARMSubtarget *STI; ARMFunctionInfo *AFI; bool isThumb; @@ -266,74 +273,82 @@ namespace { } private: - void DoInitialPlacement(MachineFunction &MF, - std::vector &CPEMIs); + void doInitialPlacement(std::vector &CPEMIs); CPEntry *findConstPoolEntry(unsigned CPI, const MachineInstr *CPEMI); - void JumpTableFunctionScan(MachineFunction &MF); - void InitialFunctionScan(MachineFunction &MF, - const std::vector &CPEMIs); - MachineBasicBlock *SplitBlockBeforeInstr(MachineInstr *MI); - void UpdateForInsertedWaterBlock(MachineBasicBlock *NewBB); - void AdjustBBOffsetsAfter(MachineBasicBlock *BB); - bool DecrementOldEntry(unsigned CPI, MachineInstr* CPEMI); - int LookForExistingCPEntry(CPUser& U, unsigned UserOffset); - bool LookForWater(CPUser&U, unsigned UserOffset, water_iterator &WaterIter); - void CreateNewWater(unsigned CPUserIndex, unsigned UserOffset, + unsigned getCPELogAlign(const MachineInstr *CPEMI); + void scanFunctionJumpTables(); + void initializeFunctionInfo(const std::vector &CPEMIs); + MachineBasicBlock *splitBlockBeforeInstr(MachineInstr *MI); + void updateForInsertedWaterBlock(MachineBasicBlock *NewBB); + void adjustBBOffsetsAfter(MachineBasicBlock *BB); + bool decrementCPEReferenceCount(unsigned CPI, MachineInstr* CPEMI); + int findInRangeCPEntry(CPUser& U, unsigned UserOffset); + bool findAvailableWater(CPUser&U, unsigned UserOffset, + water_iterator &WaterIter); + void createNewWater(unsigned CPUserIndex, unsigned UserOffset, MachineBasicBlock *&NewMBB); - bool HandleConstantPoolUser(MachineFunction &MF, unsigned CPUserIndex); - void RemoveDeadCPEMI(MachineInstr *CPEMI); - bool RemoveUnusedCPEntries(); - bool CPEIsInRange(MachineInstr *MI, unsigned UserOffset, - MachineInstr *CPEMI, unsigned Disp, bool NegOk, - bool DoDump = false); - bool WaterIsInRange(unsigned UserOffset, MachineBasicBlock *Water, - CPUser &U); - bool BBIsInRange(MachineInstr *MI, MachineBasicBlock *BB, unsigned Disp); - bool FixUpImmediateBr(MachineFunction &MF, ImmBranch &Br); - bool FixUpConditionalBr(MachineFunction &MF, ImmBranch &Br); - bool FixUpUnconditionalBr(MachineFunction &MF, ImmBranch &Br); - bool UndoLRSpillRestore(); - bool OptimizeThumb2Instructions(MachineFunction &MF); - bool OptimizeThumb2Branches(MachineFunction &MF); - bool ReorderThumb2JumpTables(MachineFunction &MF); - bool OptimizeThumb2JumpTables(MachineFunction &MF); - MachineBasicBlock *AdjustJTTargetBlockForward(MachineBasicBlock *BB, + bool handleConstantPoolUser(unsigned CPUserIndex); + void removeDeadCPEMI(MachineInstr *CPEMI); + bool removeUnusedCPEntries(); + bool isCPEntryInRange(MachineInstr *MI, unsigned UserOffset, + MachineInstr *CPEMI, unsigned Disp, bool NegOk, + bool DoDump = false); + bool isWaterInRange(unsigned UserOffset, MachineBasicBlock *Water, + CPUser &U, unsigned &Growth); + bool isBBInRange(MachineInstr *MI, MachineBasicBlock *BB, unsigned Disp); + bool fixupImmediateBr(ImmBranch &Br); + bool fixupConditionalBr(ImmBranch &Br); + bool fixupUnconditionalBr(ImmBranch &Br); + bool undoLRSpillRestore(); + bool mayOptimizeThumb2Instruction(const MachineInstr *MI) const; + bool optimizeThumb2Instructions(); + bool optimizeThumb2Branches(); + bool reorderThumb2JumpTables(); + bool optimizeThumb2JumpTables(); + MachineBasicBlock *adjustJTTargetBlockForward(MachineBasicBlock *BB, MachineBasicBlock *JTBB); - void ComputeBlockSize(MachineBasicBlock *MBB); - unsigned GetOffsetOf(MachineInstr *MI) const; + void computeBlockSize(MachineBasicBlock *MBB); + unsigned getOffsetOf(MachineInstr *MI) const; + unsigned getUserOffset(CPUser&) const; void dumpBBs(); - void verify(MachineFunction &MF); + void verify(); - bool OffsetIsInRange(unsigned UserOffset, unsigned TrialOffset, + bool isOffsetInRange(unsigned UserOffset, unsigned TrialOffset, unsigned Disp, bool NegativeOK, bool IsSoImm = false); - bool OffsetIsInRange(unsigned UserOffset, unsigned TrialOffset, + bool isOffsetInRange(unsigned UserOffset, unsigned TrialOffset, const CPUser &U) { - return OffsetIsInRange(UserOffset, TrialOffset, - U.MaxDisp, U.NegOk, U.IsSoImm); + return isOffsetInRange(UserOffset, TrialOffset, + U.getMaxDisp(), U.NegOk, U.IsSoImm); } }; char ARMConstantIslands::ID = 0; } /// verify - check BBOffsets, BBSizes, alignment of islands -void ARMConstantIslands::verify(MachineFunction &MF) { +void ARMConstantIslands::verify() { #ifndef NDEBUG - for (MachineFunction::iterator MBBI = MF.begin(), E = MF.end(); + for (MachineFunction::iterator MBBI = MF->begin(), E = MF->end(); MBBI != E; ++MBBI) { MachineBasicBlock *MBB = MBBI; - unsigned Align = MBB->getAlignment(); unsigned MBBId = MBB->getNumber(); - assert(BBInfo[MBBId].Offset % (1u << Align) == 0); assert(!MBBId || BBInfo[MBBId - 1].postOffset() <= BBInfo[MBBId].Offset); } + DEBUG(dbgs() << "Verifying " << CPUsers.size() << " CP users.\n"); for (unsigned i = 0, e = CPUsers.size(); i != e; ++i) { CPUser &U = CPUsers[i]; - unsigned UserOffset = GetOffsetOf(U.MI) + (isThumb ? 4 : 8); - unsigned CPEOffset = GetOffsetOf(U.CPEMI); - unsigned Disp = UserOffset < CPEOffset ? CPEOffset - UserOffset : - UserOffset - CPEOffset; - assert(Disp <= U.MaxDisp || "Constant pool entry out of range!"); + unsigned UserOffset = getUserOffset(U); + // Verify offset using the real max displacement without the safety + // adjustment. + if (isCPEntryInRange(U.MI, UserOffset, U.CPEMI, U.getMaxDisp()+2, U.NegOk, + /* DoDump = */ true)) { + DEBUG(dbgs() << "OK\n"); + continue; + } + DEBUG(dbgs() << "Out of range.\n"); + dumpBBs(); + DEBUG(MF->dump()); + llvm_unreachable("Constant pool entry out of range!"); } #endif } @@ -358,16 +373,17 @@ FunctionPass *llvm::createARMConstantIslandPass() { return new ARMConstantIslands(); } -bool ARMConstantIslands::runOnMachineFunction(MachineFunction &MF) { - MachineConstantPool &MCP = *MF.getConstantPool(); +bool ARMConstantIslands::runOnMachineFunction(MachineFunction &mf) { + MF = &mf; + MCP = mf.getConstantPool(); DEBUG(dbgs() << "***** ARMConstantIslands: " - << MCP.getConstants().size() << " CP entries, aligned to " - << MCP.getConstantPoolAlignment() << " bytes *****\n"); + << MCP->getConstants().size() << " CP entries, aligned to " + << MCP->getConstantPoolAlignment() << " bytes *****\n"); - TII = (const ARMInstrInfo*)MF.getTarget().getInstrInfo(); - AFI = MF.getInfo(); - STI = &MF.getTarget().getSubtarget(); + TII = (const ARMBaseInstrInfo*)MF->getTarget().getInstrInfo(); + AFI = MF->getInfo(); + STI = &MF->getTarget().getSubtarget(); isThumb = AFI->isThumbFunction(); isThumb1 = AFI->isThumb1OnlyFunction(); @@ -375,20 +391,23 @@ bool ARMConstantIslands::runOnMachineFunction(MachineFunction &MF) { HasFarJump = false; + // This pass invalidates liveness information when it splits basic blocks. + MF->getRegInfo().invalidateLiveness(); + // Renumber all of the machine basic blocks in the function, guaranteeing that // the numbers agree with the position of the block in the function. - MF.RenumberBlocks(); + MF->RenumberBlocks(); // Try to reorder and otherwise adjust the block layout to make good use // of the TB[BH] instructions. bool MadeChange = false; if (isThumb2 && AdjustJumpTableBlocks) { - JumpTableFunctionScan(MF); - MadeChange |= ReorderThumb2JumpTables(MF); + scanFunctionJumpTables(); + MadeChange |= reorderThumb2JumpTables(); // Data is out of date, so clear it. It'll be re-computed later. T2JumpTables.clear(); // Blocks may have shifted around. Keep the numbering up to date. - MF.RenumberBlocks(); + MF->RenumberBlocks(); } // Thumb1 functions containing constant pools get 4-byte alignment. @@ -396,16 +415,13 @@ bool ARMConstantIslands::runOnMachineFunction(MachineFunction &MF) { // ARM and Thumb2 functions need to be 4-byte aligned. if (!isThumb1) - MF.EnsureAlignment(2); // 2 = log2(4) + MF->ensureAlignment(2); // 2 = log2(4) // 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()) { - DoInitialPlacement(MF, CPEMIs); - if (isThumb1) - MF.EnsureAlignment(2); // 2 = log2(4) - } + if (!MCP->isEmpty()) + doInitialPlacement(CPEMIs); /// The next UID to take is the first unused one. AFI->initPICLabelUId(CPEMIs.size()); @@ -413,13 +429,13 @@ bool ARMConstantIslands::runOnMachineFunction(MachineFunction &MF) { // 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. - InitialFunctionScan(MF, CPEMIs); + initializeFunctionInfo(CPEMIs); CPEMIs.clear(); DEBUG(dumpBBs()); /// Remove dead constant pool entries. - MadeChange |= RemoveUnusedCPEntries(); + MadeChange |= removeUnusedCPEntries(); // Iteratively place constant pool entries and fix up branches until there // is no change. @@ -428,9 +444,9 @@ bool ARMConstantIslands::runOnMachineFunction(MachineFunction &MF) { DEBUG(dbgs() << "Beginning CP iteration #" << NoCPIters << '\n'); bool CPChange = false; for (unsigned i = 0, e = CPUsers.size(); i != e; ++i) - CPChange |= HandleConstantPoolUser(MF, i); + CPChange |= handleConstantPoolUser(i); if (CPChange && ++NoCPIters > 30) - llvm_unreachable("Constant Island pass failed to converge!"); + report_fatal_error("Constant Island pass failed to converge!"); DEBUG(dumpBBs()); // Clear NewWaterList now. If we split a block for branches, it should @@ -440,9 +456,9 @@ bool ARMConstantIslands::runOnMachineFunction(MachineFunction &MF) { DEBUG(dbgs() << "Beginning BR iteration #" << NoBRIters << '\n'); bool BRChange = false; for (unsigned i = 0, e = ImmBranches.size(); i != e; ++i) - BRChange |= FixUpImmediateBr(MF, ImmBranches[i]); + BRChange |= fixupImmediateBr(ImmBranches[i]); if (BRChange && ++NoBRIters > 30) - llvm_unreachable("Branch Fix Up pass failed to converge!"); + report_fatal_error("Branch Fix Up pass failed to converge!"); DEBUG(dumpBBs()); if (!CPChange && !BRChange) @@ -452,15 +468,15 @@ bool ARMConstantIslands::runOnMachineFunction(MachineFunction &MF) { // Shrink 32-bit Thumb2 branch, load, and store instructions. if (isThumb2 && !STI->prefers32BitThumb()) - MadeChange |= OptimizeThumb2Instructions(MF); + MadeChange |= optimizeThumb2Instructions(); // After a while, this might be made debug-only, but it is not expensive. - verify(MF); + verify(); // If LR has been forced spilled and no far jump (i.e. BL) has been issued, // undo the spill / restore of LR if possible. if (isThumb && !HasFarJump && AFI->isLRSpilledForFarJump()) - MadeChange |= UndoLRSpillRestore(); + MadeChange |= undoLRSpillRestore(); // Save the mapping between original and cloned constpool entries. for (unsigned i = 0, e = CPEntries.size(); i != e; ++i) { @@ -483,21 +499,25 @@ bool ARMConstantIslands::runOnMachineFunction(MachineFunction &MF) { return MadeChange; } -/// DoInitialPlacement - Perform the initial placement of the constant pool +/// 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 &MF, - std::vector &CPEMIs) { +void +ARMConstantIslands::doInitialPlacement(std::vector &CPEMIs) { // Create the basic block to hold the CPE's. - MachineBasicBlock *BB = MF.CreateMachineBasicBlock(); - MF.push_back(BB); + MachineBasicBlock *BB = MF->CreateMachineBasicBlock(); + MF->push_back(BB); // MachineConstantPool measures alignment in bytes. We measure in log2(bytes). - unsigned MaxAlign = Log2_32(MF.getConstantPool()->getConstantPoolAlignment()); + unsigned MaxAlign = Log2_32(MCP->getConstantPoolAlignment()); // Mark the basic block as required by the const-pool. // If AlignConstantIslands isn't set, use 4-byte alignment for everything. BB->setAlignment(AlignConstantIslands ? MaxAlign : 2); + // The function needs to be as aligned as the basic blocks. The linker may + // move functions around based on their alignment. + MF->ensureAlignment(BB->getAlignment()); + // Order the entries in BB by descending alignment. That ensures correct // alignment of all entries as long as BB is sufficiently aligned. Keep // track of the insertion point for each alignment. We are going to bucket @@ -506,10 +526,9 @@ void ARMConstantIslands::DoInitialPlacement(MachineFunction &MF, // Add all of the constants from the constant pool to the end block, use an // identity mapping of CPI's to CPE's. - const std::vector &CPs = - MF.getConstantPool()->getConstants(); + const std::vector &CPs = MCP->getConstants(); - const TargetData &TD = *MF.getTarget().getTargetData(); + const DataLayout &TD = *MF->getTarget().getDataLayout(); for (unsigned i = 0, e = CPs.size(); i != e; ++i) { unsigned Size = TD.getTypeAllocSize(CPs[i].getType()); assert(Size >= 4 && "Too small constant pool entry"); @@ -529,7 +548,7 @@ void ARMConstantIslands::DoInitialPlacement(MachineFunction &MF, // Ensure that future entries with higher alignment get inserted before // CPEMI. This is bucket sort with iterators. - for (unsigned a = LogAlign + 1; a < MaxAlign; ++a) + for (unsigned a = LogAlign + 1; a <= MaxAlign; ++a) if (InsPoint[a] == InsAt) InsPoint[a] = CPEMI; @@ -538,7 +557,8 @@ void ARMConstantIslands::DoInitialPlacement(MachineFunction &MF, CPEs.push_back(CPEntry(CPEMI, i)); CPEntries.push_back(CPEs); ++NumCPEs; - DEBUG(dbgs() << "Moved CPI#" << i << " to end of function\n"); + DEBUG(dbgs() << "Moved CPI#" << i << " to end of function, size = " + << Size << ", align = " << Align <<'\n'); } DEBUG(BB->dump()); } @@ -549,10 +569,10 @@ static bool BBHasFallthrough(MachineBasicBlock *MBB) { // Get the next machine basic block in the function. MachineFunction::iterator MBBI = MBB; // Can't fall off end of function. - if (llvm::next(MBBI) == MBB->getParent()->end()) + if (std::next(MBBI) == MBB->getParent()->end()) return false; - MachineBasicBlock *NextBB = llvm::next(MBBI); + MachineBasicBlock *NextBB = std::next(MBBI); for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(), E = MBB->succ_end(); I != E; ++I) if (*I == NextBB) @@ -576,11 +596,27 @@ ARMConstantIslands::CPEntry return NULL; } -/// JumpTableFunctionScan - Do a scan of the function, building up +/// getCPELogAlign - Returns the required alignment of the constant pool entry +/// represented by CPEMI. Alignment is measured in log2(bytes) units. +unsigned ARMConstantIslands::getCPELogAlign(const MachineInstr *CPEMI) { + assert(CPEMI && CPEMI->getOpcode() == ARM::CONSTPOOL_ENTRY); + + // Everything is 4-byte aligned unless AlignConstantIslands is set. + if (!AlignConstantIslands) + return 2; + + unsigned CPI = CPEMI->getOperand(1).getIndex(); + assert(CPI < MCP->getConstants().size() && "Invalid constant pool index."); + unsigned Align = MCP->getConstants()[CPI].getAlignment(); + assert(isPowerOf2_32(Align) && "Invalid CPE alignment"); + return Log2_32(Align); +} + +/// scanFunctionJumpTables - Do a scan of the function, building up /// information about the sizes of each block and the locations of all /// the jump tables. -void ARMConstantIslands::JumpTableFunctionScan(MachineFunction &MF) { - for (MachineFunction::iterator MBBI = MF.begin(), E = MF.end(); +void ARMConstantIslands::scanFunctionJumpTables() { + for (MachineFunction::iterator MBBI = MF->begin(), E = MF->end(); MBBI != E; ++MBBI) { MachineBasicBlock &MBB = *MBBI; @@ -591,30 +627,30 @@ void ARMConstantIslands::JumpTableFunctionScan(MachineFunction &MF) { } } -/// InitialFunctionScan - Do the initial scan of the function, building up +/// initializeFunctionInfo - 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 &MF, - const std::vector &CPEMIs) { +void ARMConstantIslands:: +initializeFunctionInfo(const std::vector &CPEMIs) { BBInfo.clear(); - BBInfo.resize(MF.getNumBlockIDs()); + BBInfo.resize(MF->getNumBlockIDs()); // First thing, compute the size of all basic blocks, and see if the function // has any inline assembly in it. If so, we have to be conservative about // alignment assumptions, as we don't know for sure the size of any // instructions in the inline assembly. - for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) - ComputeBlockSize(I); + for (MachineFunction::iterator I = MF->begin(), E = MF->end(); I != E; ++I) + computeBlockSize(I); // The known bits of the entry block offset are determined by the function // alignment. - BBInfo.front().KnownBits = MF.getAlignment(); + BBInfo.front().KnownBits = MF->getAlignment(); // Compute block offsets and known bits. - AdjustBBOffsetsAfter(MF.begin()); + adjustBBOffsetsAfter(MF->begin()); // Now go back through the instructions and build up our data structures. - for (MachineFunction::iterator MBBI = MF.begin(), E = MF.end(); + for (MachineFunction::iterator MBBI = MF->begin(), E = MF->end(); MBBI != E; ++MBBI) { MachineBasicBlock &MBB = *MBBI; @@ -696,7 +732,6 @@ void ARMConstantIslands::InitialFunctionScan(MachineFunction &MF, switch (Opc) { default: llvm_unreachable("Unknown addressing mode for CP reference!"); - break; // Taking the address of a CP entry. case ARM::LEApcrel: @@ -718,6 +753,7 @@ void ARMConstantIslands::InitialFunctionScan(MachineFunction &MF, Scale = 4; break; + case ARM::LDRBi12: case ARM::LDRi12: case ARM::LDRcp: case ARM::t2LDRpci: @@ -757,9 +793,9 @@ void ARMConstantIslands::InitialFunctionScan(MachineFunction &MF, } } -/// ComputeBlockSize - Compute the size and some alignment information for MBB. +/// computeBlockSize - Compute the size and some alignment information for MBB. /// This function updates BBInfo directly. -void ARMConstantIslands::ComputeBlockSize(MachineBasicBlock *MBB) { +void ARMConstantIslands::computeBlockSize(MachineBasicBlock *MBB) { BasicBlockInfo &BBI = BBInfo[MBB->getNumber()]; BBI.Size = 0; BBI.Unalign = 0; @@ -772,19 +808,22 @@ void ARMConstantIslands::ComputeBlockSize(MachineBasicBlock *MBB) { // The actual size may be smaller, but still a multiple of the instr size. if (I->isInlineAsm()) BBI.Unalign = isThumb ? 1 : 2; + // Also consider instructions that may be shrunk later. + else if (isThumb && mayOptimizeThumb2Instruction(I)) + BBI.Unalign = 1; } // tBR_JTr contains a .align 2 directive. if (!MBB->empty() && MBB->back().getOpcode() == ARM::tBR_JTr) { BBI.PostAlign = 2; - MBB->getParent()->EnsureAlignment(2); + MBB->getParent()->ensureAlignment(2); } } -/// GetOffsetOf - Return the current offset of the specified machine instruction +/// getOffsetOf - Return the current offset of the specified machine instruction /// from the start of the function. This offset changes as stuff is moved /// around inside the function. -unsigned ARMConstantIslands::GetOffsetOf(MachineInstr *MI) const { +unsigned ARMConstantIslands::getOffsetOf(MachineInstr *MI) const { MachineBasicBlock *MBB = MI->getParent(); // The offset is composed of two things: the sum of the sizes of all MBB's @@ -793,11 +832,11 @@ unsigned ARMConstantIslands::GetOffsetOf(MachineInstr *MI) const { unsigned Offset = BBInfo[MBB->getNumber()].Offset; // Sum instructions before MI in MBB. - for (MachineBasicBlock::iterator I = MBB->begin(); ; ++I) { + for (MachineBasicBlock::iterator I = MBB->begin(); &*I != MI; ++I) { assert(I != MBB->end() && "Didn't find MI in its own basic block?"); - if (&*I == MI) return Offset; Offset += TII->GetInstSizeInBytes(I); } + return Offset; } /// CompareMBBNumbers - Little predicate function to sort the WaterList by MBB @@ -807,10 +846,10 @@ static bool CompareMBBNumbers(const MachineBasicBlock *LHS, return LHS->getNumber() < RHS->getNumber(); } -/// UpdateForInsertedWaterBlock - When a block is newly inserted into the +/// updateForInsertedWaterBlock - When a block is newly inserted into the /// machine function, it upsets all of the block numbers. Renumber the blocks /// and update the arrays that parallel this numbering. -void ARMConstantIslands::UpdateForInsertedWaterBlock(MachineBasicBlock *NewBB) { +void ARMConstantIslands::updateForInsertedWaterBlock(MachineBasicBlock *NewBB) { // Renumber the MBB's to keep them consecutive. NewBB->getParent()->RenumberBlocks(NewBB); @@ -830,15 +869,14 @@ void ARMConstantIslands::UpdateForInsertedWaterBlock(MachineBasicBlock *NewBB) { /// Split the basic block containing MI into two blocks, which are joined by /// an unconditional branch. Update data structures and renumber blocks to /// account for this change and returns the newly created block. -MachineBasicBlock *ARMConstantIslands::SplitBlockBeforeInstr(MachineInstr *MI) { +MachineBasicBlock *ARMConstantIslands::splitBlockBeforeInstr(MachineInstr *MI) { MachineBasicBlock *OrigBB = MI->getParent(); - MachineFunction &MF = *OrigBB->getParent(); // Create a new MBB for the code after the OrigBB. MachineBasicBlock *NewBB = - MF.CreateMachineBasicBlock(OrigBB->getBasicBlock()); + MF->CreateMachineBasicBlock(OrigBB->getBasicBlock()); MachineFunction::iterator MBBI = OrigBB; ++MBBI; - MF.insert(MBBI, NewBB); + MF->insert(MBBI, NewBB); // Splice the instructions starting with MI over to NewBB. NewBB->splice(NewBB->end(), OrigBB, MI, OrigBB->end()); @@ -862,9 +900,9 @@ MachineBasicBlock *ARMConstantIslands::SplitBlockBeforeInstr(MachineInstr *MI) { OrigBB->addSuccessor(NewBB); // Update internal data structures to account for the newly inserted MBB. - // This is almost the same as UpdateForInsertedWaterBlock, except that + // This is almost the same as updateForInsertedWaterBlock, except that // the Water goes after OrigBB, not NewBB. - MF.RenumberBlocks(NewBB); + MF->RenumberBlocks(NewBB); // Insert an entry into BBInfo to align it properly with the (newly // renumbered) block numbers. @@ -879,7 +917,7 @@ MachineBasicBlock *ARMConstantIslands::SplitBlockBeforeInstr(MachineInstr *MI) { CompareMBBNumbers); MachineBasicBlock* WaterBB = *IP; if (WaterBB == OrigBB) - WaterList.insert(llvm::next(IP), NewBB); + WaterList.insert(std::next(IP), NewBB); else WaterList.insert(IP, OrigBB); NewWaterList.insert(OrigBB); @@ -889,46 +927,51 @@ MachineBasicBlock *ARMConstantIslands::SplitBlockBeforeInstr(MachineInstr *MI) { // the new jump we added. (It should be possible to do this without // recounting everything, but it's very confusing, and this is rarely // executed.) - ComputeBlockSize(OrigBB); + computeBlockSize(OrigBB); // Figure out how large the NewMBB is. As the second half of the original // block, it may contain a tablejump. - ComputeBlockSize(NewBB); + computeBlockSize(NewBB); // All BBOffsets following these blocks must be modified. - AdjustBBOffsetsAfter(OrigBB); + adjustBBOffsetsAfter(OrigBB); return NewBB; } -/// OffsetIsInRange - Checks whether UserOffset (the location of a constant pool +/// getUserOffset - Compute the offset of U.MI as seen by the hardware +/// displacement computation. Update U.KnownAlignment to match its current +/// basic block location. +unsigned ARMConstantIslands::getUserOffset(CPUser &U) const { + unsigned UserOffset = getOffsetOf(U.MI); + const BasicBlockInfo &BBI = BBInfo[U.MI->getParent()->getNumber()]; + unsigned KnownBits = BBI.internalKnownBits(); + + // The value read from PC is offset from the actual instruction address. + UserOffset += (isThumb ? 4 : 8); + + // Because of inline assembly, we may not know the alignment (mod 4) of U.MI. + // Make sure U.getMaxDisp() returns a constrained range. + U.KnownAlignment = (KnownBits >= 2); + + // On Thumb, offsets==2 mod 4 are rounded down by the hardware for + // purposes of the displacement computation; compensate for that here. + // For unknown alignments, getMaxDisp() constrains the range instead. + if (isThumb && U.KnownAlignment) + UserOffset &= ~3u; + + return UserOffset; +} + +/// isOffsetInRange - 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, +/// UserOffset is computed by getUserOffset above to include PC adjustments. If +/// the mod 4 alignment of UserOffset is not known, the uncertainty must be +/// subtracted from MaxDisp instead. CPUser::getMaxDisp() does that. +bool ARMConstantIslands::isOffsetInRange(unsigned UserOffset, unsigned TrialOffset, unsigned MaxDisp, bool NegativeOK, bool IsSoImm) { - // 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. - unsigned TotalAdj = 0; - if (isThumb && UserOffset%4 !=0) { - UserOffset -= 2; - TotalAdj = 2; - } - // CPEs will be rounded up to a multiple of 4. - if (isThumb && TrialOffset%4 != 0) { - TrialOffset += 2; - TotalAdj += 2; - } - - // In Thumb2 mode, later branch adjustments can shift instructions up and - // cause alignment change. In the worst case scenario this can cause the - // user's effective address to be subtracted by 2 and the CPE's address to - // be plus 2. - if (isThumb2 && TotalAdj != 4) - MaxDisp -= (4 - TotalAdj); - if (UserOffset <= TrialOffset) { // User before the Trial. if (TrialOffset - UserOffset <= MaxDisp) @@ -942,28 +985,54 @@ bool ARMConstantIslands::OffsetIsInRange(unsigned UserOffset, return false; } -/// WaterIsInRange - Returns true if a CPE placed after the specified +/// isWaterInRange - 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 CPEOffset = BBInfo[Water->getNumber()].postOffset(); - - // If the CPE is to be inserted before the instruction, that will raise - // the offset of the instruction. - if (CPEOffset < UserOffset) - UserOffset += U.CPEMI->getOperand(2).getImm(); - - return OffsetIsInRange(UserOffset, CPEOffset, U); +/// +/// Compute how much the function will grow by inserting a CPE after Water. +bool ARMConstantIslands::isWaterInRange(unsigned UserOffset, + MachineBasicBlock* Water, CPUser &U, + unsigned &Growth) { + unsigned CPELogAlign = getCPELogAlign(U.CPEMI); + unsigned CPEOffset = BBInfo[Water->getNumber()].postOffset(CPELogAlign); + unsigned NextBlockOffset, NextBlockAlignment; + MachineFunction::const_iterator NextBlock = Water; + if (++NextBlock == MF->end()) { + NextBlockOffset = BBInfo[Water->getNumber()].postOffset(); + NextBlockAlignment = 0; + } else { + NextBlockOffset = BBInfo[NextBlock->getNumber()].Offset; + NextBlockAlignment = NextBlock->getAlignment(); + } + unsigned Size = U.CPEMI->getOperand(2).getImm(); + unsigned CPEEnd = CPEOffset + Size; + + // The CPE may be able to hide in the alignment padding before the next + // block. It may also cause more padding to be required if it is more aligned + // that the next block. + if (CPEEnd > NextBlockOffset) { + Growth = CPEEnd - NextBlockOffset; + // Compute the padding that would go at the end of the CPE to align the next + // block. + Growth += OffsetToAlignment(CPEEnd, 1u << NextBlockAlignment); + + // If the CPE is to be inserted before the instruction, that will raise + // the offset of the instruction. Also account for unknown alignment padding + // in blocks between CPE and the user. + if (CPEOffset < UserOffset) + UserOffset += Growth + UnknownPadding(MF->getAlignment(), CPELogAlign); + } else + // CPE fits in existing padding. + Growth = 0; + + return isOffsetInRange(UserOffset, CPEOffset, U); } -/// CPEIsInRange - Returns true if the distance between specific MI and +/// isCPEntryInRange - 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, +bool ARMConstantIslands::isCPEntryInRange(MachineInstr *MI, unsigned UserOffset, MachineInstr *CPEMI, unsigned MaxDisp, bool NegOk, bool DoDump) { - unsigned CPEOffset = GetOffsetOf(CPEMI); - assert(CPEOffset % 4 == 0 && "Misaligned CPE"); + unsigned CPEOffset = getOffsetOf(CPEMI); if (DoDump) { DEBUG({ @@ -979,7 +1048,7 @@ bool ARMConstantIslands::CPEIsInRange(MachineInstr *MI, unsigned UserOffset, }); } - return OffsetIsInRange(UserOffset, CPEOffset, MaxDisp, NegOk); + return isOffsetInRange(UserOffset, CPEOffset, MaxDisp, NegOk); } #ifndef NDEBUG @@ -999,36 +1068,40 @@ static bool BBIsJumpedOver(MachineBasicBlock *MBB) { } #endif // NDEBUG -void ARMConstantIslands::AdjustBBOffsetsAfter(MachineBasicBlock *BB) { - MachineFunction *MF = BB->getParent(); - for(unsigned i = BB->getNumber() + 1, e = MF->getNumBlockIDs(); i < e; ++i) { +void ARMConstantIslands::adjustBBOffsetsAfter(MachineBasicBlock *BB) { + unsigned BBNum = BB->getNumber(); + for(unsigned i = BBNum + 1, e = MF->getNumBlockIDs(); i < e; ++i) { // Get the offset and known bits at the end of the layout predecessor. - unsigned Offset = BBInfo[i - 1].postOffset(); - unsigned KnownBits = BBInfo[i - 1].postKnownBits(); - - // Add padding before an aligned block. This may teach us more bits. - if (unsigned Align = MF->getBlockNumbered(i)->getAlignment()) { - Offset = WorstCaseAlign(Offset, Align, KnownBits); - KnownBits = std::max(KnownBits, Align); - } + // Include the alignment of the current block. + unsigned LogAlign = MF->getBlockNumbered(i)->getAlignment(); + unsigned Offset = BBInfo[i - 1].postOffset(LogAlign); + unsigned KnownBits = BBInfo[i - 1].postKnownBits(LogAlign); + + // This is where block i begins. Stop if the offset is already correct, + // and we have updated 2 blocks. This is the maximum number of blocks + // changed before calling this function. + if (i > BBNum + 2 && + BBInfo[i].Offset == Offset && + BBInfo[i].KnownBits == KnownBits) + break; - // This is where block i begins. BBInfo[i].Offset = Offset; BBInfo[i].KnownBits = KnownBits; } } -/// DecrementOldEntry - find the constant pool entry with index CPI +/// decrementCPEReferenceCount - 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) { +bool ARMConstantIslands::decrementCPEReferenceCount(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); + removeDeadCPEMI(CPEMI); CPE->CPEMI = NULL; --NumCPEs; return true; @@ -1042,13 +1115,14 @@ bool ARMConstantIslands::DecrementOldEntry(unsigned CPI, MachineInstr *CPEMI) { /// 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) +int ARMConstantIslands::findInRangeCPEntry(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, UserOffset, CPEMI, U.MaxDisp, U.NegOk, true)) { + if (isCPEntryInRange(UserMI, UserOffset, CPEMI, U.getMaxDisp(), U.NegOk, + true)) { DEBUG(dbgs() << "In range\n"); return 1; } @@ -1063,7 +1137,8 @@ int ARMConstantIslands::LookForExistingCPEntry(CPUser& U, unsigned UserOffset) // Removing CPEs can leave empty entries, skip if (CPEs[i].CPEMI == NULL) continue; - if (CPEIsInRange(UserMI, UserOffset, CPEs[i].CPEMI, U.MaxDisp, U.NegOk)) { + if (isCPEntryInRange(UserMI, UserOffset, CPEs[i].CPEMI, U.getMaxDisp(), + U.NegOk)) { DEBUG(dbgs() << "Replacing CPE#" << CPI << " with CPE#" << CPEs[i].CPI << "\n"); // Point the CPUser node to the replacement @@ -1078,7 +1153,7 @@ 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. - return DecrementOldEntry(CPI, CPEMI) ? 2 : 1; + return decrementCPEReferenceCount(CPI, CPEMI) ? 2 : 1; } } return 0; @@ -1099,7 +1174,7 @@ static inline unsigned getUnconditionalBrDisp(int Opc) { return ((1<<23)-1)*4; } -/// LookForWater - Look for an existing entry in the WaterList in which +/// findAvailableWater - 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, WaterIter /// is set to the WaterList entry. For Thumb, prefer water that will not @@ -1107,15 +1182,14 @@ static inline unsigned getUnconditionalBrDisp(int Opc) { /// terminates, the CPE location for a particular CPUser is only allowed to /// move to a lower address, so search backward from the end of the list and /// prefer the first water that is in range. -bool ARMConstantIslands::LookForWater(CPUser &U, unsigned UserOffset, +bool ARMConstantIslands::findAvailableWater(CPUser &U, unsigned UserOffset, water_iterator &WaterIter) { if (WaterList.empty()) return false; - bool FoundWaterThatWouldPad = false; - water_iterator IPThatWouldPad; - for (water_iterator IP = prior(WaterList.end()), - B = WaterList.begin();; --IP) { + unsigned BestGrowth = ~0u; + for (water_iterator IP = std::prev(WaterList.end()), B = WaterList.begin();; + --IP) { MachineBasicBlock* WaterBB = *IP; // Check if water is in range and is either at a lower address than the // current "high water mark" or a new water block that was created since @@ -1125,182 +1199,177 @@ bool ARMConstantIslands::LookForWater(CPUser &U, unsigned UserOffset, // should be relatively uncommon and when it does happen, we want to be // sure to take advantage of it for all the CPEs near that block, so that // we don't insert more branches than necessary. - if (WaterIsInRange(UserOffset, WaterBB, U) && + unsigned Growth; + if (isWaterInRange(UserOffset, WaterBB, U, Growth) && (WaterBB->getNumber() < U.HighWaterMark->getNumber() || - NewWaterList.count(WaterBB))) { - unsigned WBBId = WaterBB->getNumber(); - if (isThumb && BBInfo[WBBId].postOffset()%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 (!FoundWaterThatWouldPad) { - FoundWaterThatWouldPad = true; - IPThatWouldPad = IP; - } - } else { - WaterIter = IP; + NewWaterList.count(WaterBB)) && Growth < BestGrowth) { + // This is the least amount of required padding seen so far. + BestGrowth = Growth; + WaterIter = IP; + DEBUG(dbgs() << "Found water after BB#" << WaterBB->getNumber() + << " Growth=" << Growth << '\n'); + + // Keep looking unless it is perfect. + if (BestGrowth == 0) return true; - } } if (IP == B) break; } - if (FoundWaterThatWouldPad) { - WaterIter = IPThatWouldPad; - return true; - } - return false; + return BestGrowth != ~0u; } -/// CreateNewWater - No existing WaterList entry will work for +/// 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, +void ARMConstantIslands::createNewWater(unsigned CPUserIndex, unsigned UserOffset, MachineBasicBlock *&NewMBB) { CPUser &U = CPUsers[CPUserIndex]; MachineInstr *UserMI = U.MI; MachineInstr *CPEMI = U.CPEMI; + unsigned CPELogAlign = getCPELogAlign(CPEMI); MachineBasicBlock *UserMBB = UserMI->getParent(); const BasicBlockInfo &UserBBI = BBInfo[UserMBB->getNumber()]; - unsigned OffsetOfNextBlock = UserBBI.postOffset(); // If the block does not end in an unconditional branch already, and if 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 + - // Thumb2, 2 on Thumb1. Possible Thumb1 alignment padding is allowed for - // inside OffsetIsInRange. - if (BBHasFallthrough(UserMBB) && - OffsetIsInRange(UserOffset, OffsetOfNextBlock + (isThumb1 ? 2: 4), U)) { - DEBUG(dbgs() << "Split at end of block\n"); - if (&UserMBB->back() == UserMI) - assert(BBHasFallthrough(UserMBB) && "Expected a fallthrough BB!"); - NewMBB = llvm::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 ? ((isThumb2) ? ARM::t2B : ARM::tB) : ARM::B; - if (!isThumb) - BuildMI(UserMBB, DebugLoc(), TII->get(UncondBr)).addMBB(NewMBB); - else - BuildMI(UserMBB, DebugLoc(), TII->get(UncondBr)).addMBB(NewMBB) - .addImm(ARMCC::AL).addReg(0); - unsigned MaxDisp = getUnconditionalBrDisp(UncondBr); - ImmBranches.push_back(ImmBranch(&UserMBB->back(), - MaxDisp, false, UncondBr)); - int delta = isThumb1 ? 2 : 4; - BBInfo[UserMBB->getNumber()].Size += delta; - AdjustBBOffsetsAfter(UserMBB); - } else { - // What a big block. Find a place within the block to split it. - // This is a little tricky on Thumb1 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. - - // Try to split the block so it's fully aligned. Compute the latest split - // point where we can add a 4-byte branch instruction, and then - // WorstCaseAlign to LogAlign. - unsigned LogAlign = UserMBB->getParent()->getAlignment(); - unsigned KnownBits = UserBBI.internalKnownBits(); - unsigned UPad = UnknownPadding(LogAlign, KnownBits); - unsigned BaseInsertOffset = UserOffset + U.MaxDisp; - DEBUG(dbgs() << format("Split in middle of big block before %#x", - BaseInsertOffset)); - - // Account for alignment and unknown padding. - BaseInsertOffset &= ~((1u << LogAlign) - 1); - BaseInsertOffset -= UPad; - - // The 4 in the following is for the unconditional branch we'll be - // inserting (allows for long branch on Thumb1). Alignment of the - // island is handled inside OffsetIsInRange. - BaseInsertOffset -= 4; - - DEBUG(dbgs() << format(", adjusted to %#x", BaseInsertOffset) - << " la=" << LogAlign - << " kb=" << KnownBits - << " up=" << UPad << '\n'); - - // 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 >= BBInfo[UserMBB->getNumber()+1].Offset) - BaseInsertOffset = BBInfo[UserMBB->getNumber()+1].Offset - - (isThumb1 ? 6 : 8); - unsigned EndInsertOffset = - WorstCaseAlign(BaseInsertOffset + 4, LogAlign, KnownBits) + - CPEMI->getOperand(2).getImm(); - MachineBasicBlock::iterator MI = UserMI; - ++MI; - unsigned CPUIndex = CPUserIndex+1; - unsigned NumCPUsers = CPUsers.size(); - MachineInstr *LastIT = 0; - for (unsigned Offset = UserOffset+TII->GetInstSizeInBytes(UserMI); - Offset < BaseInsertOffset; - Offset += TII->GetInstSizeInBytes(MI), - MI = llvm::next(MI)) { - if (CPUIndex < NumCPUsers && CPUsers[CPUIndex].MI == MI) { - CPUser &U = CPUsers[CPUIndex]; - if (!OffsetIsInRange(Offset, EndInsertOffset, U)) { - BaseInsertOffset -= 1u << LogAlign; - EndInsertOffset -= 1u << LogAlign; - } - // 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++; - } + // Thumb2, 2 on Thumb1. + if (BBHasFallthrough(UserMBB)) { + // Size of branch to insert. + unsigned Delta = isThumb1 ? 2 : 4; + // Compute the offset where the CPE will begin. + unsigned CPEOffset = UserBBI.postOffset(CPELogAlign) + Delta; + + if (isOffsetInRange(UserOffset, CPEOffset, U)) { + DEBUG(dbgs() << "Split at end of BB#" << UserMBB->getNumber() + << format(", expected CPE offset %#x\n", CPEOffset)); + NewMBB = std::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 ? ((isThumb2) ? ARM::t2B : ARM::tB) : ARM::B; + if (!isThumb) + BuildMI(UserMBB, DebugLoc(), TII->get(UncondBr)).addMBB(NewMBB); + else + BuildMI(UserMBB, DebugLoc(), TII->get(UncondBr)).addMBB(NewMBB) + .addImm(ARMCC::AL).addReg(0); + unsigned MaxDisp = getUnconditionalBrDisp(UncondBr); + ImmBranches.push_back(ImmBranch(&UserMBB->back(), + MaxDisp, false, UncondBr)); + BBInfo[UserMBB->getNumber()].Size += Delta; + adjustBBOffsetsAfter(UserMBB); + return; + } + } - // Remember the last IT instruction. - if (MI->getOpcode() == ARM::t2IT) - LastIT = MI; + // What a big block. Find a place within the block to split it. This is a + // little tricky on Thumb1 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. + + // Try to split the block so it's fully aligned. Compute the latest split + // point where we can add a 4-byte branch instruction, and then align to + // LogAlign which is the largest possible alignment in the function. + unsigned LogAlign = MF->getAlignment(); + assert(LogAlign >= CPELogAlign && "Over-aligned constant pool entry"); + unsigned KnownBits = UserBBI.internalKnownBits(); + unsigned UPad = UnknownPadding(LogAlign, KnownBits); + unsigned BaseInsertOffset = UserOffset + U.getMaxDisp() - UPad; + DEBUG(dbgs() << format("Split in middle of big block before %#x", + BaseInsertOffset)); + + // The 4 in the following is for the unconditional branch we'll be inserting + // (allows for long branch on Thumb1). Alignment of the island is handled + // inside isOffsetInRange. + BaseInsertOffset -= 4; + + DEBUG(dbgs() << format(", adjusted to %#x", BaseInsertOffset) + << " la=" << LogAlign + << " kb=" << KnownBits + << " up=" << UPad << '\n'); + + // 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 + 8 >= UserBBI.postOffset()) { + BaseInsertOffset = UserBBI.postOffset() - UPad - 8; + DEBUG(dbgs() << format("Move inside block: %#x\n", BaseInsertOffset)); + } + unsigned EndInsertOffset = BaseInsertOffset + 4 + UPad + + CPEMI->getOperand(2).getImm(); + MachineBasicBlock::iterator MI = UserMI; + ++MI; + unsigned CPUIndex = CPUserIndex+1; + unsigned NumCPUsers = CPUsers.size(); + MachineInstr *LastIT = 0; + for (unsigned Offset = UserOffset+TII->GetInstSizeInBytes(UserMI); + Offset < BaseInsertOffset; + Offset += TII->GetInstSizeInBytes(MI), MI = std::next(MI)) { + assert(MI != UserMBB->end() && "Fell off end of block"); + if (CPUIndex < NumCPUsers && CPUsers[CPUIndex].MI == MI) { + CPUser &U = CPUsers[CPUIndex]; + if (!isOffsetInRange(Offset, EndInsertOffset, U)) { + // Shift intertion point by one unit of alignment so it is within reach. + BaseInsertOffset -= 1u << LogAlign; + EndInsertOffset -= 1u << LogAlign; + } + // This is overly conservative, as we don't account for CPEMIs being + // reused within the block, but it doesn't matter much. Also assume CPEs + // are added in order with alignment padding. We may eventually be able + // to pack the aligned CPEs better. + EndInsertOffset += U.CPEMI->getOperand(2).getImm(); + CPUIndex++; } - --MI; + // Remember the last IT instruction. + if (MI->getOpcode() == ARM::t2IT) + LastIT = MI; + } - // Avoid splitting an IT block. - if (LastIT) { - unsigned PredReg = 0; - ARMCC::CondCodes CC = llvm::getITInstrPredicate(MI, PredReg); - if (CC != ARMCC::AL) - MI = LastIT; - } - NewMBB = SplitBlockBeforeInstr(MI); + --MI; + + // Avoid splitting an IT block. + if (LastIT) { + unsigned PredReg = 0; + ARMCC::CondCodes CC = getITInstrPredicate(MI, PredReg); + if (CC != ARMCC::AL) + MI = LastIT; } + NewMBB = splitBlockBeforeInstr(MI); } -/// HandleConstantPoolUser - Analyze the specified user, checking to see if it +/// handleConstantPoolUser - Analyze the specified user, checking to see if it /// is out-of-range. If so, pick 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 &MF, - unsigned CPUserIndex) { +bool ARMConstantIslands::handleConstantPoolUser(unsigned CPUserIndex) { CPUser &U = CPUsers[CPUserIndex]; MachineInstr *UserMI = U.MI; MachineInstr *CPEMI = U.CPEMI; unsigned CPI = CPEMI->getOperand(1).getIndex(); unsigned Size = CPEMI->getOperand(2).getImm(); - // Compute this only once, it's expensive. The 4 or 8 is the value the - // hardware keeps in the PC. - unsigned UserOffset = GetOffsetOf(UserMI) + (isThumb ? 4 : 8); + // Compute this only once, it's expensive. + unsigned UserOffset = getUserOffset(U); // See if the current entry is within range, or there is a clone of it // in range. - int result = LookForExistingCPEntry(U, UserOffset); + int result = findInRangeCPEntry(U, UserOffset); if (result==1) return false; else if (result==2) return true; @@ -1309,34 +1378,33 @@ bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &MF, unsigned ID = AFI->createPICLabelUId(); // Look for water where we can place this CPE. - MachineBasicBlock *NewIsland = MF.CreateMachineBasicBlock(); + MachineBasicBlock *NewIsland = MF->CreateMachineBasicBlock(); MachineBasicBlock *NewMBB; water_iterator IP; - if (LookForWater(U, UserOffset, IP)) { + if (findAvailableWater(U, UserOffset, IP)) { DEBUG(dbgs() << "Found water in range\n"); MachineBasicBlock *WaterBB = *IP; // If the original WaterList entry was "new water" on this iteration, // propagate that to the new island. This is just keeping NewWaterList // updated to match the WaterList, which will be updated below. - if (NewWaterList.count(WaterBB)) { - NewWaterList.erase(WaterBB); + if (NewWaterList.erase(WaterBB)) NewWaterList.insert(NewIsland); - } + // The new CPE goes before the following block (NewMBB). - NewMBB = llvm::next(MachineFunction::iterator(WaterBB)); + NewMBB = std::next(MachineFunction::iterator(WaterBB)); } else { // No water found. DEBUG(dbgs() << "No water found\n"); - CreateNewWater(CPUserIndex, UserOffset, NewMBB); + createNewWater(CPUserIndex, UserOffset, NewMBB); - // SplitBlockBeforeInstr adds to WaterList, which is important when it is + // splitBlockBeforeInstr adds to WaterList, which is important when it is // called while handling branches so that the water will be seen on the // next iteration for constant pools, but in this context, we don't want // it. Check for this so it will be removed from the WaterList. // Also remove any entry from NewWaterList. - MachineBasicBlock *WaterBB = prior(MachineFunction::iterator(NewMBB)); + MachineBasicBlock *WaterBB = std::prev(MachineFunction::iterator(NewMBB)); IP = std::find(WaterList.begin(), WaterList.end(), WaterBB); if (IP != WaterList.end()) NewWaterList.erase(WaterBB); @@ -1353,13 +1421,13 @@ bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &MF, WaterList.erase(IP); // Okay, we know we can put an island before NewMBB now, do it! - MF.insert(NewMBB, NewIsland); + MF->insert(NewMBB, NewIsland); // Update internal data structures to account for the newly inserted MBB. - UpdateForInsertedWaterBlock(NewIsland); + updateForInsertedWaterBlock(NewIsland); // Decrement the old entry, and remove it if refcount becomes 0. - DecrementOldEntry(CPI, CPEMI); + decrementCPEReferenceCount(CPI, CPEMI); // Now that we have an island to add the CPE to, clone the original CPE and // add it to the island. @@ -1369,12 +1437,12 @@ bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &MF, CPEntries[CPI].push_back(CPEntry(U.CPEMI, ID, 1)); ++NumCPEs; - // Mark the basic block as 4-byte aligned as required by the const-pool entry. - NewIsland->setAlignment(2); + // Mark the basic block as aligned as required by the const-pool entry. + NewIsland->setAlignment(getCPELogAlign(U.CPEMI)); // Increase the size of the island block to account for the new entry. BBInfo[NewIsland->getNumber()].Size += Size; - AdjustBBOffsetsAfter(llvm::prior(MachineFunction::iterator(NewIsland))); + adjustBBOffsetsAfter(std::prev(MachineFunction::iterator(NewIsland))); // Finally, change the CPI in the instruction operand to be ID. for (unsigned i = 0, e = UserMI->getNumOperands(); i != e; ++i) @@ -1389,28 +1457,24 @@ bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &MF, return true; } -/// RemoveDeadCPEMI - Remove a dead constant pool entry instruction. Update +/// removeDeadCPEMI - Remove a dead constant pool entry instruction. Update /// sizes and offsets of impacted basic blocks. -void ARMConstantIslands::RemoveDeadCPEMI(MachineInstr *CPEMI) { +void ARMConstantIslands::removeDeadCPEMI(MachineInstr *CPEMI) { MachineBasicBlock *CPEBB = CPEMI->getParent(); unsigned Size = CPEMI->getOperand(2).getImm(); CPEMI->eraseFromParent(); BBInfo[CPEBB->getNumber()].Size -= Size; // All succeeding offsets have the current size value added in, fix this. if (CPEBB->empty()) { - // In thumb1 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 (BBInfo[CPEBB->getNumber()].Size != 0) { - Size += BBInfo[CPEBB->getNumber()].Size; - BBInfo[CPEBB->getNumber()].Size = 0; - } + BBInfo[CPEBB->getNumber()].Size = 0; - // This block no longer needs to be aligned. . + // This block no longer needs to be aligned. CPEBB->setAlignment(0); - } - AdjustBBOffsetsAfter(CPEBB); + } else + // Entries are sorted by descending alignment, so realign from the front. + CPEBB->setAlignment(getCPELogAlign(CPEBB->begin())); + + adjustBBOffsetsAfter(CPEBB); // 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. @@ -1418,15 +1482,15 @@ void ARMConstantIslands::RemoveDeadCPEMI(MachineInstr *CPEMI) { // FIXME: remove the empty blocks after all the work is done? } -/// RemoveUnusedCPEntries - Remove constant pool entries whose refcounts +/// removeUnusedCPEntries - Remove constant pool entries whose refcounts /// are zero. -bool ARMConstantIslands::RemoveUnusedCPEntries() { +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); + removeDeadCPEMI(CPEs[j].CPEMI); CPEs[j].CPEMI = NULL; MadeChange = true; } @@ -1435,18 +1499,18 @@ bool ARMConstantIslands::RemoveUnusedCPEntries() { return MadeChange; } -/// BBIsInRange - Returns true if the distance between specific MI and +/// isBBInRange - 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, +bool ARMConstantIslands::isBBInRange(MachineInstr *MI,MachineBasicBlock *DestBB, unsigned MaxDisp) { unsigned PCAdj = isThumb ? 4 : 8; - unsigned BrOffset = GetOffsetOf(MI) + PCAdj; + unsigned BrOffset = getOffsetOf(MI) + PCAdj; unsigned DestOffset = BBInfo[DestBB->getNumber()].Offset; DEBUG(dbgs() << "Branch of destination BB#" << DestBB->getNumber() << " from BB#" << MI->getParent()->getNumber() << " max delta=" << MaxDisp - << " from " << GetOffsetOf(MI) << " to " << DestOffset + << " from " << getOffsetOf(MI) << " to " << DestOffset << " offset " << int(DestOffset-BrOffset) << "\t" << *MI); if (BrOffset <= DestOffset) { @@ -1460,37 +1524,37 @@ bool ARMConstantIslands::BBIsInRange(MachineInstr *MI,MachineBasicBlock *DestBB, return false; } -/// FixUpImmediateBr - Fix up an immediate branch whose destination is too far +/// fixupImmediateBr - Fix up an immediate branch whose destination is too far /// away to fit in its displacement field. -bool ARMConstantIslands::FixUpImmediateBr(MachineFunction &MF, ImmBranch &Br) { +bool ARMConstantIslands::fixupImmediateBr(ImmBranch &Br) { MachineInstr *MI = Br.MI; MachineBasicBlock *DestBB = MI->getOperand(0).getMBB(); // Check to see if the DestBB is already in-range. - if (BBIsInRange(MI, DestBB, Br.MaxDisp)) + if (isBBInRange(MI, DestBB, Br.MaxDisp)) return false; if (!Br.isCond) - return FixUpUnconditionalBr(MF, Br); - return FixUpConditionalBr(MF, Br); + return fixupUnconditionalBr(Br); + return fixupConditionalBr(Br); } -/// FixUpUnconditionalBr - Fix up an unconditional branch whose destination is +/// 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 an intermediate branch instruction to a branch. bool -ARMConstantIslands::FixUpUnconditionalBr(MachineFunction &MF, ImmBranch &Br) { +ARMConstantIslands::fixupUnconditionalBr(ImmBranch &Br) { MachineInstr *MI = Br.MI; MachineBasicBlock *MBB = MI->getParent(); if (!isThumb1) - llvm_unreachable("FixUpUnconditionalBr is Thumb1 only!"); + llvm_unreachable("fixupUnconditionalBr is Thumb1 only!"); // Use BL to implement far jump. Br.MaxDisp = (1 << 21) * 2; MI->setDesc(TII->get(ARM::tBfar)); BBInfo[MBB->getNumber()].Size += 2; - AdjustBBOffsetsAfter(MBB); + adjustBBOffsetsAfter(MBB); HasFarJump = true; ++NumUBrFixed; @@ -1499,11 +1563,11 @@ ARMConstantIslands::FixUpUnconditionalBr(MachineFunction &MF, ImmBranch &Br) { return true; } -/// FixUpConditionalBr - Fix up a conditional branch 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 -ARMConstantIslands::FixUpConditionalBr(MachineFunction &MF, ImmBranch &Br) { +ARMConstantIslands::fixupConditionalBr(ImmBranch &Br) { MachineInstr *MI = Br.MI; MachineBasicBlock *DestBB = MI->getOperand(0).getMBB(); @@ -1527,7 +1591,7 @@ ARMConstantIslands::FixUpConditionalBr(MachineFunction &MF, ImmBranch &Br) { ++NumCBrFixed; if (BMI != MI) { - if (llvm::next(MachineBasicBlock::iterator(MI)) == prior(MBB->end()) && + if (std::next(MachineBasicBlock::iterator(MI)) == std::prev(MBB->end()) && BMI->getOpcode() == Br.UncondBr) { // Last MI in the BB is an unconditional branch. Can we simply invert the // condition and swap destinations: @@ -1537,7 +1601,7 @@ ARMConstantIslands::FixUpConditionalBr(MachineFunction &MF, ImmBranch &Br) { // bne L2 // b L1 MachineBasicBlock *NewDest = BMI->getOperand(0).getMBB(); - if (BBIsInRange(MI, NewDest, Br.MaxDisp)) { + if (isBBInRange(MI, NewDest, Br.MaxDisp)) { DEBUG(dbgs() << " Invert Bcc condition and swap its destination with " << *BMI); BMI->getOperand(0).setMBB(DestBB); @@ -1549,7 +1613,7 @@ ARMConstantIslands::FixUpConditionalBr(MachineFunction &MF, ImmBranch &Br) { } if (NeedSplit) { - SplitBlockBeforeInstr(MI); + splitBlockBeforeInstr(MI); // No need for the branch to the next block. We're adding an unconditional // branch to the destination. int delta = TII->GetInstSizeInBytes(&MBB->back()); @@ -1557,7 +1621,7 @@ ARMConstantIslands::FixUpConditionalBr(MachineFunction &MF, ImmBranch &Br) { MBB->back().eraseFromParent(); // BBInfo[SplitBB].Offset is wrong temporarily, fixed below } - MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB)); + MachineBasicBlock *NextBB = std::next(MachineFunction::iterator(MBB)); DEBUG(dbgs() << " Insert B to BB#" << DestBB->getNumber() << " also invert condition and change dest. to BB#" @@ -1581,14 +1645,14 @@ ARMConstantIslands::FixUpConditionalBr(MachineFunction &MF, ImmBranch &Br) { // Remove the old conditional branch. It may or may not still be in MBB. BBInfo[MI->getParent()->getNumber()].Size -= TII->GetInstSizeInBytes(MI); MI->eraseFromParent(); - AdjustBBOffsetsAfter(MBB); + adjustBBOffsetsAfter(MBB); return true; } -/// UndoLRSpillRestore - Remove Thumb push / pop instructions that only spills +/// undoLRSpillRestore - Remove Thumb push / pop instructions that only spills /// LR / restores LR to pc. FIXME: This is done here because it's only possible /// to do this if tBfar is not used. -bool ARMConstantIslands::UndoLRSpillRestore() { +bool ARMConstantIslands::undoLRSpillRestore() { bool MadeChange = false; for (unsigned i = 0, e = PushPopMIs.size(); i != e; ++i) { MachineInstr *MI = PushPopMIs[i]; @@ -1607,7 +1671,26 @@ bool ARMConstantIslands::UndoLRSpillRestore() { return MadeChange; } -bool ARMConstantIslands::OptimizeThumb2Instructions(MachineFunction &MF) { +// mayOptimizeThumb2Instruction - Returns true if optimizeThumb2Instructions +// below may shrink MI. +bool +ARMConstantIslands::mayOptimizeThumb2Instruction(const MachineInstr *MI) const { + switch(MI->getOpcode()) { + // optimizeThumb2Instructions. + case ARM::t2LEApcrel: + case ARM::t2LDRpci: + // optimizeThumb2Branches. + case ARM::t2B: + case ARM::t2Bcc: + case ARM::tBcc: + // optimizeThumb2JumpTables. + case ARM::t2BR_JT: + return true; + } + return false; +} + +bool ARMConstantIslands::optimizeThumb2Instructions() { bool MadeChange = false; // Shrink ADR and LDR from constantpool. @@ -1638,25 +1721,31 @@ bool ARMConstantIslands::OptimizeThumb2Instructions(MachineFunction &MF) { if (!NewOpc) continue; - unsigned UserOffset = GetOffsetOf(U.MI) + 4; + unsigned UserOffset = getUserOffset(U); unsigned MaxOffs = ((1 << Bits) - 1) * Scale; + + // Be conservative with inline asm. + if (!U.KnownAlignment) + MaxOffs -= 2; + // FIXME: Check if offset is multiple of scale if scale is not 4. - if (CPEIsInRange(U.MI, UserOffset, U.CPEMI, MaxOffs, false, true)) { + if (isCPEntryInRange(U.MI, UserOffset, U.CPEMI, MaxOffs, false, true)) { + DEBUG(dbgs() << "Shrink: " << *U.MI); U.MI->setDesc(TII->get(NewOpc)); MachineBasicBlock *MBB = U.MI->getParent(); BBInfo[MBB->getNumber()].Size -= 2; - AdjustBBOffsetsAfter(MBB); + adjustBBOffsetsAfter(MBB); ++NumT2CPShrunk; MadeChange = true; } } - MadeChange |= OptimizeThumb2Branches(MF); - MadeChange |= OptimizeThumb2JumpTables(MF); + MadeChange |= optimizeThumb2Branches(); + MadeChange |= optimizeThumb2JumpTables(); return MadeChange; } -bool ARMConstantIslands::OptimizeThumb2Branches(MachineFunction &MF) { +bool ARMConstantIslands::optimizeThumb2Branches() { bool MadeChange = false; for (unsigned i = 0, e = ImmBranches.size(); i != e; ++i) { @@ -1682,11 +1771,12 @@ bool ARMConstantIslands::OptimizeThumb2Branches(MachineFunction &MF) { if (NewOpc) { unsigned MaxOffs = ((1 << (Bits-1))-1) * Scale; MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB(); - if (BBIsInRange(Br.MI, DestBB, MaxOffs)) { + if (isBBInRange(Br.MI, DestBB, MaxOffs)) { + DEBUG(dbgs() << "Shrink branch: " << *Br.MI); Br.MI->setDesc(TII->get(NewOpc)); MachineBasicBlock *MBB = Br.MI->getParent(); BBInfo[MBB->getNumber()].Size -= 2; - AdjustBBOffsetsAfter(MBB); + adjustBBOffsetsAfter(MBB); ++NumT2BrShrunk; MadeChange = true; } @@ -1696,9 +1786,14 @@ bool ARMConstantIslands::OptimizeThumb2Branches(MachineFunction &MF) { if (Opcode != ARM::tBcc) continue; + // If the conditional branch doesn't kill CPSR, then CPSR can be liveout + // so this transformation is not safe. + if (!Br.MI->killsRegister(ARM::CPSR)) + continue; + NewOpc = 0; unsigned PredReg = 0; - ARMCC::CondCodes Pred = llvm::getInstrPredicate(Br.MI, PredReg); + ARMCC::CondCodes Pred = getInstrPredicate(Br.MI, PredReg); if (Pred == ARMCC::EQ) NewOpc = ARM::tCBZ; else if (Pred == ARMCC::NE) @@ -1708,7 +1803,7 @@ bool ARMConstantIslands::OptimizeThumb2Branches(MachineFunction &MF) { MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB(); // Check if the distance is within 126. Subtract starting offset by 2 // because the cmp will be eliminated. - unsigned BrOffset = GetOffsetOf(Br.MI) + 4 - 2; + unsigned BrOffset = getOffsetOf(Br.MI) + 4 - 2; unsigned DestOffset = BBInfo[DestBB->getNumber()].Offset; if (BrOffset < DestOffset && (DestOffset - BrOffset) <= 126) { MachineBasicBlock::iterator CmpMI = Br.MI; @@ -1716,11 +1811,12 @@ bool ARMConstantIslands::OptimizeThumb2Branches(MachineFunction &MF) { --CmpMI; if (CmpMI->getOpcode() == ARM::tCMPi8) { unsigned Reg = CmpMI->getOperand(0).getReg(); - Pred = llvm::getInstrPredicate(CmpMI, PredReg); + Pred = getInstrPredicate(CmpMI, PredReg); if (Pred == ARMCC::AL && CmpMI->getOperand(1).getImm() == 0 && isARMLowRegister(Reg)) { MachineBasicBlock *MBB = Br.MI->getParent(); + DEBUG(dbgs() << "Fold: " << *CmpMI << " and: " << *Br.MI); MachineInstr *NewBR = BuildMI(*MBB, CmpMI, Br.MI->getDebugLoc(), TII->get(NewOpc)) .addReg(Reg).addMBB(DestBB,Br.MI->getOperand(0).getTargetFlags()); @@ -1728,7 +1824,7 @@ bool ARMConstantIslands::OptimizeThumb2Branches(MachineFunction &MF) { Br.MI->eraseFromParent(); Br.MI = NewBR; BBInfo[MBB->getNumber()].Size -= 2; - AdjustBBOffsetsAfter(MBB); + adjustBBOffsetsAfter(MBB); ++NumCBZ; MadeChange = true; } @@ -1740,14 +1836,14 @@ bool ARMConstantIslands::OptimizeThumb2Branches(MachineFunction &MF) { return MadeChange; } -/// OptimizeThumb2JumpTables - Use tbb / tbh instructions to generate smaller +/// optimizeThumb2JumpTables - Use tbb / tbh instructions to generate smaller /// jumptables when it's possible. -bool ARMConstantIslands::OptimizeThumb2JumpTables(MachineFunction &MF) { +bool ARMConstantIslands::optimizeThumb2JumpTables() { bool MadeChange = false; // FIXME: After the tables are shrunk, can we get rid some of the // constantpool tables? - MachineJumpTableInfo *MJTI = MF.getJumpTableInfo(); + MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); if (MJTI == 0) return false; const std::vector &JT = MJTI->getJumpTables(); @@ -1762,7 +1858,7 @@ bool ARMConstantIslands::OptimizeThumb2JumpTables(MachineFunction &MF) { bool ByteOk = true; bool HalfWordOk = true; - unsigned JTOffset = GetOffsetOf(MI) + 4; + unsigned JTOffset = getOffsetOf(MI) + 4; const std::vector &JTBBs = JT[JTI].MBBs; for (unsigned j = 0, ee = JTBBs.size(); j != ee; ++j) { MachineBasicBlock *MBB = JTBBs[j]; @@ -1837,11 +1933,14 @@ bool ARMConstantIslands::OptimizeThumb2JumpTables(MachineFunction &MF) { if (!OptOk) continue; + DEBUG(dbgs() << "Shrink JT: " << *MI << " addr: " << *AddrMI + << " lea: " << *LeaMI); unsigned Opc = ByteOk ? ARM::t2TBB_JT : ARM::t2TBH_JT; MachineInstr *NewJTMI = BuildMI(MBB, MI->getDebugLoc(), TII->get(Opc)) .addReg(IdxReg, getKillRegState(IdxRegKill)) .addJumpTableIndex(JTI, JTOP.getTargetFlags()) .addImm(MI->getOperand(JTOpIdx+1).getImm()); + DEBUG(dbgs() << "BB#" << MBB->getNumber() << ": " << *NewJTMI); // FIXME: Insert an "ALIGN" instruction to ensure the next instruction // is 2-byte aligned. For now, asm printer will fix it up. unsigned NewSize = TII->GetInstSizeInBytes(NewJTMI); @@ -1855,7 +1954,7 @@ bool ARMConstantIslands::OptimizeThumb2JumpTables(MachineFunction &MF) { int delta = OrigSize - NewSize; BBInfo[MBB->getNumber()].Size -= delta; - AdjustBBOffsetsAfter(MBB); + adjustBBOffsetsAfter(MBB); ++NumTBs; MadeChange = true; @@ -1865,12 +1964,12 @@ bool ARMConstantIslands::OptimizeThumb2JumpTables(MachineFunction &MF) { return MadeChange; } -/// ReorderThumb2JumpTables - Adjust the function's block layout to ensure that +/// reorderThumb2JumpTables - Adjust the function's block layout to ensure that /// jump tables always branch forwards, since that's what tbb and tbh need. -bool ARMConstantIslands::ReorderThumb2JumpTables(MachineFunction &MF) { +bool ARMConstantIslands::reorderThumb2JumpTables() { bool MadeChange = false; - MachineJumpTableInfo *MJTI = MF.getJumpTableInfo(); + MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); if (MJTI == 0) return false; const std::vector &JT = MJTI->getJumpTables(); @@ -1896,7 +1995,7 @@ bool ARMConstantIslands::ReorderThumb2JumpTables(MachineFunction &MF) { // The destination precedes the switch. Try to move the block forward // so we have a positive offset. MachineBasicBlock *NewBB = - AdjustJTTargetBlockForward(MBB, MI->getParent()); + adjustJTTargetBlockForward(MBB, MI->getParent()); if (NewBB) MJTI->ReplaceMBBInJumpTable(JTI, JTBBs[j], NewBB); MadeChange = true; @@ -1908,10 +2007,7 @@ bool ARMConstantIslands::ReorderThumb2JumpTables(MachineFunction &MF) { } MachineBasicBlock *ARMConstantIslands:: -AdjustJTTargetBlockForward(MachineBasicBlock *BB, MachineBasicBlock *JTBB) -{ - MachineFunction &MF = *BB->getParent(); - +adjustJTTargetBlockForward(MachineBasicBlock *BB, MachineBasicBlock *JTBB) { // If the destination block is terminated by an unconditional branch, // try to move it; otherwise, create a new block following the jump // table that branches back to the actual target. This is a very simple @@ -1920,7 +2016,7 @@ AdjustJTTargetBlockForward(MachineBasicBlock *BB, MachineBasicBlock *JTBB) SmallVector Cond; SmallVector CondPrior; MachineFunction::iterator BBi = BB; - MachineFunction::iterator OldPrior = prior(BBi); + MachineFunction::iterator OldPrior = std::prev(BBi); // If the block terminator isn't analyzable, don't try to move the block bool B = TII->AnalyzeBranch(*BB, TBB, FBB, Cond); @@ -1928,22 +2024,22 @@ AdjustJTTargetBlockForward(MachineBasicBlock *BB, MachineBasicBlock *JTBB) // If the block ends in an unconditional branch, move it. The prior block // has to have an analyzable terminator for us to move this one. Be paranoid // and make sure we're not trying to move the entry block of the function. - if (!B && Cond.empty() && BB != MF.begin() && + if (!B && Cond.empty() && BB != MF->begin() && !TII->AnalyzeBranch(*OldPrior, TBB, FBB, CondPrior)) { BB->moveAfter(JTBB); OldPrior->updateTerminator(); BB->updateTerminator(); // Update numbering to account for the block being moved. - MF.RenumberBlocks(); + MF->RenumberBlocks(); ++NumJTMoved; return NULL; } // Create a new MBB for the code after the jump BB. MachineBasicBlock *NewBB = - MF.CreateMachineBasicBlock(JTBB->getBasicBlock()); + MF->CreateMachineBasicBlock(JTBB->getBasicBlock()); MachineFunction::iterator MBBI = JTBB; ++MBBI; - MF.insert(MBBI, NewBB); + MF->insert(MBBI, NewBB); // Add an unconditional branch from NewBB to BB. // There doesn't seem to be meaningful DebugInfo available; this doesn't @@ -1953,7 +2049,7 @@ AdjustJTTargetBlockForward(MachineBasicBlock *BB, MachineBasicBlock *JTBB) .addImm(ARMCC::AL).addReg(0); // Update internal data structures to account for the newly inserted MBB. - MF.RenumberBlocks(NewBB); + MF->RenumberBlocks(NewBB); // Update the CFG. NewBB->addSuccessor(BB);