//
// 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.
//
//===----------------------------------------------------------------------===//
//
#define DEBUG_TYPE "arm-cp-islands"
#include "ARM.h"
+#include "ARMAddressingModes.h"
#include "ARMMachineFunctionInfo.h"
#include "ARMInstrInfo.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetMachine.h"
-#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.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 <algorithm>
using namespace llvm;
-STATISTIC(NumCPEs, "Number of constpool entries");
-STATISTIC(NumSplit, "Number of uncond branches inserted");
-STATISTIC(NumCBrFixed, "Number of cond branches fixed");
-STATISTIC(NumUBrFixed, "Number of uncond branches fixed");
+STATISTIC(NumCPEs, "Number of constpool entries");
+STATISTIC(NumSplit, "Number of uncond branches inserted");
+STATISTIC(NumCBrFixed, "Number of cond branches fixed");
+STATISTIC(NumUBrFixed, "Number of uncond branches fixed");
+STATISTIC(NumTBs, "Number of table branches generated");
+STATISTIC(NumT2CPShrunk, "Number of Thumb2 constantpool instructions shrunk");
+STATISTIC(NumT2BrShrunk, "Number of Thumb2 immediate branches shrunk");
+STATISTIC(NumCBZ, "Number of CBZ / CBNZ formed");
+STATISTIC(NumJTMoved, "Number of jump table destination blocks moved");
+STATISTIC(NumJTInserted, "Number of jump table intermediate blocks inserted");
+
+
+static cl::opt<bool>
+AdjustJumpTableBlocks("arm-adjust-jump-tables", cl::Hidden, cl::init(true),
+ cl::desc("Adjust basic block layout to better use TB[BH]"));
namespace {
/// ARMConstantIslands - Due to limited PC-relative displacements, ARM
/// Water - Potential places where an island could be formed.
/// CPE - A constant pool entry that has been placed somewhere, which
/// tracks a list of users.
- class VISIBILITY_HIDDEN ARMConstantIslands : public MachineFunctionPass {
- /// NextUID - Assign unique ID's to CPE's.
- unsigned NextUID;
-
+ class ARMConstantIslands : public MachineFunctionPass {
/// BBSizes - The size of each MachineBasicBlock in bytes of code, indexed
/// 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<unsigned> 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.
/// to a return, unreachable, or unconditional branch).
std::vector<MachineBasicBlock*> WaterList;
+ /// NewWaterList - The subset of WaterList that was created since the
+ /// previous iteration by inserting unconditional branches.
+ SmallSet<MachineBasicBlock*, 4> NewWaterList;
+
+ typedef std::vector<MachineBasicBlock*>::iterator water_iterator;
+
/// CPUser - One user of a constant pool, keeping the machine instruction
/// pointer, the constant pool being referenced, and the max displacement
- /// allowed from the instruction to the CP.
+ /// allowed from the instruction to the CP. The HighWaterMark records the
+ /// highest basic block where a new CPEntry can be placed. To ensure this
+ /// pass terminates, the CP entries are initially placed at the end of the
+ /// function and then move monotonically to lower addresses. The
+ /// exception to this rule is when the current CP entry for a particular
+ /// CPUser is out of range, but there is another CP entry for the same
+ /// constant value in range. We want to use the existing in-range CP
+ /// entry, but if it later moves out of range, the search for new water
+ /// should resume where it left off. The HighWaterMark is used to record
+ /// that point.
struct CPUser {
MachineInstr *MI;
MachineInstr *CPEMI;
+ MachineBasicBlock *HighWaterMark;
unsigned MaxDisp;
- CPUser(MachineInstr *mi, MachineInstr *cpemi, unsigned maxdisp)
- : MI(mi), CPEMI(cpemi), MaxDisp(maxdisp) {}
+ bool NegOk;
+ bool IsSoImm;
+ CPUser(MachineInstr *mi, MachineInstr *cpemi, unsigned maxdisp,
+ bool neg, bool soimm)
+ : MI(mi), CPEMI(cpemi), MaxDisp(maxdisp), NegOk(neg), IsSoImm(soimm) {
+ HighWaterMark = CPEMI->getParent();
+ }
};
-
+
/// CPUsers - Keep track of all of the machine instructions that use various
/// constant pools and their max displacement.
std::vector<CPUser> CPUsers;
-
+
/// CPEntry - One per constant pool entry, keeping the machine instruction
/// pointer, the constpool index, and the number of CPUser's which
/// reference this entry.
/// Original elements are cloned as we go along; the clones are
/// put in the vector of the original element, but have distinct CPIs.
std::vector<std::vector<CPEntry> > CPEntries;
-
+
/// ImmBranch - One per immediate branch, keeping the machine instruction
/// pointer, conditional or unconditional, the max displacement,
/// and (if isCond is true) the corresponding unconditional branch
///
SmallVector<MachineInstr*, 4> PushPopMIs;
+ /// T2JumpTables - Keep track of all the Thumb2 jumptable instructions.
+ SmallVector<MachineInstr*, 4> T2JumpTables;
+
/// HasFarJump - True if any far jump instruction has been emitted during
/// the branch fix up pass.
bool HasFarJump;
- const TargetInstrInfo *TII;
+ /// HasInlineAsm - True if the function contains inline assembly.
+ bool HasInlineAsm;
+
+ const ARMInstrInfo *TII;
+ const ARMSubtarget *STI;
ARMFunctionInfo *AFI;
bool isThumb;
+ bool isThumb1;
+ bool isThumb2;
public:
static char ID;
- ARMConstantIslands() : MachineFunctionPass((intptr_t)&ID) {}
+ ARMConstantIslands() : MachineFunctionPass(ID) {}
- virtual bool runOnMachineFunction(MachineFunction &Fn);
+ virtual bool runOnMachineFunction(MachineFunction &MF);
virtual const char *getPassName() const {
return "ARM constant island placement and branch shortening pass";
}
-
+
private:
- void DoInitialPlacement(MachineFunction &Fn,
+ void DoInitialPlacement(MachineFunction &MF,
std::vector<MachineInstr*> &CPEMIs);
CPEntry *findConstPoolEntry(unsigned CPI, const MachineInstr *CPEMI);
- void InitialFunctionScan(MachineFunction &Fn,
+ void JumpTableFunctionScan(MachineFunction &MF);
+ void InitialFunctionScan(MachineFunction &MF,
const std::vector<MachineInstr*> &CPEMIs);
MachineBasicBlock *SplitBlockBeforeInstr(MachineInstr *MI);
void UpdateForInsertedWaterBlock(MachineBasicBlock *NewBB);
void AdjustBBOffsetsAfter(MachineBasicBlock *BB, int delta);
bool DecrementOldEntry(unsigned CPI, MachineInstr* CPEMI);
int LookForExistingCPEntry(CPUser& U, unsigned UserOffset);
- bool LookForWater(CPUser&U, unsigned UserOffset,
- MachineBasicBlock** NewMBB);
- MachineBasicBlock* AcceptWater(MachineBasicBlock *WaterBB,
- std::vector<MachineBasicBlock*>::iterator IP);
+ bool LookForWater(CPUser&U, unsigned UserOffset, water_iterator &WaterIter);
void CreateNewWater(unsigned CPUserIndex, unsigned UserOffset,
- MachineBasicBlock** NewMBB);
- bool HandleConstantPoolUser(MachineFunction &Fn, unsigned CPUserIndex);
+ 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 DoDump);
+ bool CPEIsInRange(MachineInstr *MI, unsigned UserOffset,
+ MachineInstr *CPEMI, unsigned Disp, bool NegOk,
+ bool DoDump = false);
bool WaterIsInRange(unsigned UserOffset, MachineBasicBlock *Water,
CPUser &U);
bool OffsetIsInRange(unsigned UserOffset, unsigned TrialOffset,
- unsigned Disp, bool NegativeOK);
+ unsigned Disp, bool NegativeOK, bool IsSoImm = false);
bool BBIsInRange(MachineInstr *MI, MachineBasicBlock *BB, unsigned Disp);
- bool FixUpImmediateBr(MachineFunction &Fn, ImmBranch &Br);
- bool FixUpConditionalBr(MachineFunction &Fn, ImmBranch &Br);
- bool FixUpUnconditionalBr(MachineFunction &Fn, ImmBranch &Br);
+ 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,
+ MachineBasicBlock *JTBB);
unsigned GetOffsetOf(MachineInstr *MI) const;
void dumpBBs();
- void verify(MachineFunction &Fn);
+ void verify(MachineFunction &MF);
};
char ARMConstantIslands::ID = 0;
}
/// verify - check BBOffsets, BBSizes, alignment of islands
-void ARMConstantIslands::verify(MachineFunction &Fn) {
+void ARMConstantIslands::verify(MachineFunction &MF) {
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));
+ if (!isThumb)
+ return;
+#ifndef NDEBUG
+ for (MachineFunction::iterator MBBI = MF.begin(), E = MF.end();
+ MBBI != E; ++MBBI) {
+ MachineBasicBlock *MBB = MBBI;
+ if (!MBB->empty() &&
+ MBB->begin()->getOpcode() == ARM::CONSTPOOL_ENTRY) {
+ unsigned MBBId = MBB->getNumber();
+ assert(HasInlineAsm ||
+ (BBOffsets[MBBId]%4 == 0 && BBSizes[MBBId]%4 == 0) ||
+ (BBOffsets[MBBId]%4 != 0 && BBSizes[MBBId]%4 != 0));
}
}
+ 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!");
+ }
+#endif
}
/// 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";
+ DEBUG(errs() << "block " << J << " offset " << BBOffsets[J]
+ << " size " << BBSizes[J] << "\n");
}
}
return new ARMConstantIslands();
}
-bool ARMConstantIslands::runOnMachineFunction(MachineFunction &Fn) {
- MachineConstantPool &MCP = *Fn.getConstantPool();
-
- TII = Fn.getTarget().getInstrInfo();
- AFI = Fn.getInfo<ARMFunctionInfo>();
+bool ARMConstantIslands::runOnMachineFunction(MachineFunction &MF) {
+ MachineConstantPool &MCP = *MF.getConstantPool();
+
+ TII = (const ARMInstrInfo*)MF.getTarget().getInstrInfo();
+ AFI = MF.getInfo<ARMFunctionInfo>();
+ STI = &MF.getTarget().getSubtarget<ARMSubtarget>();
+
isThumb = AFI->isThumbFunction();
+ isThumb1 = AFI->isThumb1OnlyFunction();
+ isThumb2 = AFI->isThumb2Function();
HasFarJump = false;
+ HasInlineAsm = false;
// Renumber all of the machine basic blocks in the function, guaranteeing that
// the numbers agree with the position of the block in the function.
- Fn.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);
+ // 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();
+ }
+
+ // Thumb1 functions containing constant pools get 4-byte alignment.
+ // This is so we can keep exact track of where the alignment padding goes.
- /// 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);
+ // ARM and Thumb2 functions need to be 4-byte aligned.
+ if (!isThumb1)
+ 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<MachineInstr*> CPEMIs;
if (!MCP.isEmpty()) {
- DoInitialPlacement(Fn, CPEMIs);
- if (isThumb)
- AFI->setAlign(2U);
+ DoInitialPlacement(MF, CPEMIs);
+ if (isThumb1)
+ MF.EnsureAlignment(2); // 2 = log2(4)
}
-
+
/// The next UID to take is the first unused one.
- NextUID = CPEMIs.size();
-
+ AFI->initConstPoolEntryUId(CPEMIs.size());
+
// 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(Fn, CPEMIs);
+ InitialFunctionScan(MF, CPEMIs);
CPEMIs.clear();
-
+ DEBUG(dumpBBs());
+
+
/// Remove dead constant pool entries.
RemoveUnusedCPEntries();
// Iteratively place constant pool entries and fix up branches until there
// is no change.
- bool MadeChange = false;
+ unsigned NoCPIters = 0, NoBRIters = 0;
while (true) {
- bool Change = false;
+ bool CPChange = false;
for (unsigned i = 0, e = CPUsers.size(); i != e; ++i)
- Change |= HandleConstantPoolUser(Fn, i);
+ CPChange |= HandleConstantPoolUser(MF, i);
+ if (CPChange && ++NoCPIters > 30)
+ llvm_unreachable("Constant Island pass failed to converge!");
DEBUG(dumpBBs());
+
+ // Clear NewWaterList now. If we split a block for branches, it should
+ // appear as "new water" for the next iteration of constant pool placement.
+ NewWaterList.clear();
+
+ bool BRChange = false;
for (unsigned i = 0, e = ImmBranches.size(); i != e; ++i)
- Change |= FixUpImmediateBr(Fn, ImmBranches[i]);
+ BRChange |= FixUpImmediateBr(MF, ImmBranches[i]);
+ if (BRChange && ++NoBRIters > 30)
+ llvm_unreachable("Branch Fix Up pass failed to converge!");
DEBUG(dumpBBs());
- if (!Change)
+
+ if (!CPChange && !BRChange)
break;
MadeChange = true;
}
+ // Shrink 32-bit Thumb2 branch, load, and store instructions.
+ if (isThumb2)
+ MadeChange |= OptimizeThumb2Instructions(MF);
+
// After a while, this might be made debug-only, but it is not expensive.
- verify(Fn);
+ verify(MF);
- // 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->isLRSpilledForFarJump() && isThumb)
+ // 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();
+ DEBUG(errs() << '\n'; dumpBBs());
+
BBSizes.clear();
BBOffsets.clear();
WaterList.clear();
CPEntries.clear();
ImmBranches.clear();
PushPopMIs.clear();
+ T2JumpTables.clear();
return MadeChange;
}
/// DoInitialPlacement - Perform the initial placement of the constant pool
/// entries. To start with, we put them all at the end of the function.
-void ARMConstantIslands::DoInitialPlacement(MachineFunction &Fn,
- std::vector<MachineInstr*> &CPEMIs){
+void ARMConstantIslands::DoInitialPlacement(MachineFunction &MF,
+ std::vector<MachineInstr*> &CPEMIs) {
// Create the basic block to hold the CPE's.
- MachineBasicBlock *BB = new MachineBasicBlock();
- Fn.getBasicBlockList().push_back(BB);
-
+ MachineBasicBlock *BB = MF.CreateMachineBasicBlock();
+ MF.push_back(BB);
+
// 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<MachineConstantPoolEntry> &CPs =
- Fn.getConstantPool()->getConstants();
-
- const TargetData &TD = *Fn.getTarget().getTargetData();
+ MF.getConstantPool()->getConstants();
+
+ const TargetData &TD = *MF.getTarget().getTargetData();
for (unsigned i = 0, e = CPs.size(); i != e; ++i) {
- unsigned Size = TD.getTypeSize(CPs[i].getType());
+ unsigned Size = TD.getTypeAllocSize(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.
assert((Size & 3) == 0 && "CP Entry not multiple of 4 bytes!");
MachineInstr *CPEMI =
- BuildMI(BB, TII->get(ARM::CONSTPOOL_ENTRY))
- .addImm(i).addConstantPoolIndex(i).addImm(Size);
+ BuildMI(BB, DebugLoc(), TII->get(ARM::CONSTPOOL_ENTRY))
+ .addImm(i).addConstantPoolIndex(i).addImm(Size);
CPEMIs.push_back(CPEMI);
// Add a new CPEntry, but no corresponding CPUser yet.
std::vector<CPEntry> CPEs;
CPEs.push_back(CPEntry(CPEMI, i));
CPEntries.push_back(CPEs);
- NumCPEs++;
- DOUT << "Moved CPI#" << i << " to end of function as #" << i << "\n";
+ ++NumCPEs;
+ DEBUG(errs() << "Moved CPI#" << i << " to end of function as #" << i
+ << "\n");
}
}
static bool BBHasFallthrough(MachineBasicBlock *MBB) {
// Get the next machine basic block in the function.
MachineFunction::iterator MBBI = MBB;
- if (next(MBBI) == MBB->getParent()->end()) // Can't fall off end of function.
+ // Can't fall off end of function.
+ if (llvm::next(MBBI) == MBB->getParent()->end())
return false;
-
- MachineBasicBlock *NextBB = next(MBBI);
+
+ MachineBasicBlock *NextBB = llvm::next(MBBI);
for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
E = MBB->succ_end(); I != E; ++I)
if (*I == NextBB)
return true;
-
+
return false;
}
return NULL;
}
+/// JumpTableFunctionScan - 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();
+ MBBI != E; ++MBBI) {
+ MachineBasicBlock &MBB = *MBBI;
+
+ for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
+ I != E; ++I)
+ if (I->getDesc().isBranch() && I->getOpcode() == ARM::t2BR_JT)
+ T2JumpTables.push_back(I);
+ }
+}
+
/// InitialFunctionScan - Do the initial scan of the function, building up
/// information about the sizes of each block, the location of all the water,
/// and finding all of the constant pool users.
-void ARMConstantIslands::InitialFunctionScan(MachineFunction &Fn,
+void ARMConstantIslands::InitialFunctionScan(MachineFunction &MF,
const std::vector<MachineInstr*> &CPEMIs) {
+ // First thing, 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 MBBI = MF.begin(), E = MF.end();
+ MBBI != E; ++MBBI) {
+ MachineBasicBlock &MBB = *MBBI;
+ for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
+ I != E; ++I)
+ if (I->getOpcode() == ARM::INLINEASM)
+ HasInlineAsm = true;
+ }
+
+ // Now go back through the instructions and build up our data structures
unsigned Offset = 0;
- for (MachineFunction::iterator MBBI = Fn.begin(), E = Fn.end();
+ for (MachineFunction::iterator MBBI = MF.begin(), E = MF.end();
MBBI != E; ++MBBI) {
MachineBasicBlock &MBB = *MBBI;
-
+
// If this block doesn't fall through into the next MBB, then this is
// 'water' that a constant pool island could be placed.
if (!BBHasFallthrough(&MBB))
WaterList.push_back(&MBB);
-
+
unsigned MBBSize = 0;
for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
I != E; ++I) {
+ if (I->isDebugValue())
+ continue;
// Add instruction size to MBBSize.
- MBBSize += ARM::GetInstSize(I);
+ MBBSize += TII->GetInstSizeInBytes(I);
int Opc = I->getOpcode();
- if (TII->isBranch(Opc)) {
+ if (I->getDesc().isBranch()) {
bool isCond = false;
unsigned Bits = 0;
unsigned Scale = 1;
int UOpc = Opc;
switch (Opc) {
+ default:
+ continue; // Ignore other JT branches
case ARM::tBR_JTr:
- // A Thumb table jump may involve padding; for the offsets to
+ // A Thumb1 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)
+ // tBR_JTr expands to a mov pc followed by .align 2 and then the jump
+ // table entries. So this code checks whether offset of tBR_JTr + 2
+ // is aligned. That is held in Offset+MBBSize, which already has
+ // 2 added in for the size of the mov pc instruction.
+ MF.EnsureAlignment(2U);
+ if ((Offset+MBBSize)%4 != 0 || HasInlineAsm)
+ // FIXME: Add a pseudo ALIGN instruction instead.
MBBSize += 2; // padding
continue; // Does not get an entry in ImmBranches
- default:
- continue; // Ignore other JT branches
+ case ARM::t2BR_JT:
+ T2JumpTables.push_back(I);
+ continue; // Does not get an entry in ImmBranches
case ARM::Bcc:
isCond = true;
UOpc = ARM::B;
Bits = 11;
Scale = 2;
break;
+ case ARM::t2Bcc:
+ isCond = true;
+ UOpc = ARM::t2B;
+ Bits = 20;
+ Scale = 2;
+ break;
+ case ARM::t2B:
+ Bits = 24;
+ Scale = 2;
+ break;
}
// Record this immediate branch.
if (Opc == ARM::tPUSH || Opc == ARM::tPOP_RET)
PushPopMIs.push_back(I);
+ if (Opc == ARM::CONSTPOOL_ENTRY)
+ continue;
+
// Scan the instructions for constant pool operands.
for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op)
- if (I->getOperand(op).isConstantPoolIndex()) {
+ if (I->getOperand(op).isCPI()) {
// We found one. The addressing mode tells us the max displacement
// from the PC that this instruction permits.
-
+
// Basic size info comes from the TSFlags field.
unsigned Bits = 0;
unsigned Scale = 1;
- unsigned TSFlags = I->getInstrDescriptor()->TSFlags;
- switch (TSFlags & ARMII::AddrModeMask) {
- default:
- // 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;
- Scale = 4; // Taking the address of a CP entry.
- break;
- case ARMII::AddrMode2:
- Bits = 12; // +-offset_12
- break;
- case ARMII::AddrMode3:
- Bits = 8; // +-offset_8
+ bool NegOk = false;
+ bool IsSoImm = false;
+
+ switch (Opc) {
+ default:
+ llvm_unreachable("Unknown addressing mode for CP reference!");
break;
- // addrmode4 has no immediate offset.
- case ARMII::AddrMode5:
+
+ // Taking the address of a CP entry.
+ case ARM::LEApcrel:
+ // This takes a SoImm, which is 8 bit immediate rotated. We'll
+ // pretend the maximum offset is 255 * 4. Since each instruction
+ // 4 byte wide, this is always correct. We'll check for other
+ // displacements that fits in a SoImm as well.
Bits = 8;
- Scale = 4; // +-(offset_8*4)
+ Scale = 4;
+ NegOk = true;
+ IsSoImm = true;
break;
- case ARMII::AddrModeT1:
- Bits = 5; // +offset_5
+ case ARM::t2LEApcrel:
+ Bits = 12;
+ NegOk = true;
break;
- case ARMII::AddrModeT2:
- Bits = 5;
- Scale = 2; // +(offset_5*2)
+ case ARM::tLEApcrel:
+ Bits = 8;
+ Scale = 4;
break;
- case ARMII::AddrModeT4:
- Bits = 5;
- Scale = 4; // +(offset_5*4)
+
+ case ARM::LDR:
+ case ARM::LDRcp:
+ case ARM::t2LDRpci:
+ Bits = 12; // +-offset_12
+ NegOk = true;
break;
- case ARMII::AddrModeTs:
+
+ case ARM::tLDRpci:
+ case ARM::tLDRcp:
Bits = 8;
Scale = 4; // +(offset_8*4)
break;
+
+ case ARM::VLDRD:
+ case ARM::VLDRS:
+ Bits = 8;
+ Scale = 4; // +-(offset_8*4)
+ NegOk = true;
+ break;
}
// Remember that this is a user of a CP entry.
- unsigned CPI = I->getOperand(op).getConstantPoolIndex();
+ unsigned CPI = I->getOperand(op).getIndex();
MachineInstr *CPEMI = CPEMIs[CPI];
- unsigned MaxOffs = ((1 << Bits)-1) * Scale;
- CPUsers.push_back(CPUser(I, CPEMI, MaxOffs));
+ unsigned MaxOffs = ((1 << Bits)-1) * Scale;
+ CPUsers.push_back(CPUser(I, CPEMI, MaxOffs, NegOk, IsSoImm));
// Increment corresponding CPEntry reference count.
CPEntry *CPE = findConstPoolEntry(CPI, CPEMI);
assert(CPE && "Cannot find a corresponding CPEntry!");
CPE->RefCount++;
-
+
// Instructions can only use one CP entry, don't bother scanning the
// rest of the operands.
break;
if (isThumb &&
!MBB.empty() &&
MBB.begin()->getOpcode() == ARM::CONSTPOOL_ENTRY &&
- (Offset%4) != 0)
+ ((Offset%4) != 0 || HasInlineAsm))
MBBSize += 2;
BBSizes.push_back(MBBSize);
/// around inside the function.
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
// before this instruction's block, and the offset from the start of the block
// it is in.
// 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)
+ if (isThumb &&
+ MI->getOpcode() == ARM::CONSTPOOL_ENTRY &&
+ (Offset%4 != 0 || HasInlineAsm))
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?");
if (&*I == MI) return Offset;
- Offset += ARM::GetInstSize(I);
+ Offset += TII->GetInstSizeInBytes(I);
}
}
void ARMConstantIslands::UpdateForInsertedWaterBlock(MachineBasicBlock *NewBB) {
// Renumber the MBB's to keep them consequtive.
NewBB->getParent()->RenumberBlocks(NewBB);
-
+
// Insert a size into BBSizes to align it properly with the (newly
// renumbered) block numbers.
BBSizes.insert(BBSizes.begin()+NewBB->getNumber(), 0);
// Likewise for BBOffsets.
BBOffsets.insert(BBOffsets.begin()+NewBB->getNumber(), 0);
-
- // Next, update WaterList. Specifically, we need to add NewMBB as having
+
+ // Next, update WaterList. Specifically, we need to add NewMBB as having
// available water after it.
- std::vector<MachineBasicBlock*>::iterator IP =
+ water_iterator IP =
std::lower_bound(WaterList.begin(), WaterList.end(), NewBB,
CompareMBBNumbers);
WaterList.insert(IP, NewBB);
/// Split the basic block containing MI into two blocks, which are joined by
-/// an unconditional branch. Update datastructures and renumber blocks to
+/// 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 *OrigBB = MI->getParent();
+ MachineFunction &MF = *OrigBB->getParent();
// Create a new MBB for the code after the OrigBB.
- MachineBasicBlock *NewBB = new MachineBasicBlock(OrigBB->getBasicBlock());
+ MachineBasicBlock *NewBB =
+ MF.CreateMachineBasicBlock(OrigBB->getBasicBlock());
MachineFunction::iterator MBBI = OrigBB; ++MBBI;
- OrigBB->getParent()->getBasicBlockList().insert(MBBI, NewBB);
-
+ MF.insert(MBBI, NewBB);
+
// Splice the instructions starting with MI over to NewBB.
NewBB->splice(NewBB->end(), OrigBB, MI, OrigBB->end());
-
+
// Add an unconditional branch from OrigBB to NewBB.
// Note the new unconditional branch is not being recorded.
- BuildMI(OrigBB, TII->get(isThumb ? ARM::tB : ARM::B)).addMBB(NewBB);
- NumSplit++;
-
+ // There doesn't seem to be meaningful DebugInfo available; this doesn't
+ // correspond to anything in the source.
+ unsigned Opc = isThumb ? (isThumb2 ? ARM::t2B : ARM::tB) : ARM::B;
+ BuildMI(OrigBB, DebugLoc(), TII->get(Opc)).addMBB(NewBB);
+ ++NumSplit;
+
// Update the CFG. All succs of OrigBB are now succs of NewBB.
while (!OrigBB->succ_empty()) {
MachineBasicBlock *Succ = *OrigBB->succ_begin();
OrigBB->removeSuccessor(Succ);
NewBB->addSuccessor(Succ);
-
+
// This pass should be run after register allocation, so there should be no
// PHI nodes to update.
- assert((Succ->empty() || Succ->begin()->getOpcode() != TargetInstrInfo::PHI)
+ assert((Succ->empty() || !Succ->begin()->isPHI())
&& "PHI nodes should be eliminated by now!");
}
-
+
// OrigBB branches to NewBB.
OrigBB->addSuccessor(NewBB);
-
+
// Update internal data structures to account for the newly inserted MBB.
// This is almost the same as UpdateForInsertedWaterBlock, except that
// the Water goes after OrigBB, not NewBB.
- NewBB->getParent()->RenumberBlocks(NewBB);
-
+ MF.RenumberBlocks(NewBB);
+
// Insert a size into BBSizes to align it properly with the (newly
// renumbered) block numbers.
BBSizes.insert(BBSizes.begin()+NewBB->getNumber(), 0);
-
+
// Likewise for BBOffsets.
BBOffsets.insert(BBOffsets.begin()+NewBB->getNumber(), 0);
- // Next, update WaterList. Specifically, we need to add OrigMBB as having
+ // Next, update WaterList. Specifically, we need to add OrigMBB as having
// available water after it (but not if it's already there, which happens
// when splitting before a conditional branch that is followed by an
// unconditional branch - in that case we want to insert NewBB).
- std::vector<MachineBasicBlock*>::iterator IP =
+ water_iterator IP =
std::lower_bound(WaterList.begin(), WaterList.end(), OrigBB,
CompareMBBNumbers);
MachineBasicBlock* WaterBB = *IP;
if (WaterBB == OrigBB)
- WaterList.insert(next(IP), NewBB);
+ WaterList.insert(llvm::next(IP), NewBB);
else
WaterList.insert(IP, OrigBB);
+ NewWaterList.insert(OrigBB);
- // Figure out how large the first NewMBB is. (It cannot
- // contain a constpool_entry or tablejump.)
- unsigned NewBBSize = 0;
- for (MachineBasicBlock::iterator I = NewBB->begin(), E = NewBB->end();
- I != E; ++I)
- NewBBSize += ARM::GetInstSize(I);
-
unsigned OrigBBI = OrigBB->getNumber();
unsigned NewBBI = NewBB->getNumber();
- // Set the size of NewBB in BBSizes.
- BBSizes[NewBBI] = NewBBSize;
-
- // We removed instructions from UserMBB, subtract that off from its size.
- // Add 2 or 4 to the block to count the unconditional branch we added to it.
- unsigned delta = isThumb ? 2 : 4;
- BBSizes[OrigBBI] -= NewBBSize - delta;
+
+ int delta = isThumb1 ? 2 : 4;
+
+ // Figure out how large the OrigBB is. As the first half of the original
+ // block, it cannot contain a tablejump. The size includes
+ // 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.)
+ unsigned OrigBBSize = 0;
+ for (MachineBasicBlock::iterator I = OrigBB->begin(), E = OrigBB->end();
+ I != E; ++I)
+ OrigBBSize += TII->GetInstSizeInBytes(I);
+ BBSizes[OrigBBI] = OrigBBSize;
// ...and adjust BBOffsets for NewBB accordingly.
BBOffsets[NewBBI] = BBOffsets[OrigBBI] + BBSizes[OrigBBI];
+ // Figure out how large the NewMBB is. As the second half of the original
+ // block, it may contain a tablejump.
+ unsigned NewBBSize = 0;
+ for (MachineBasicBlock::iterator I = NewBB->begin(), E = NewBB->end();
+ I != E; ++I)
+ NewBBSize += TII->GetInstSizeInBytes(I);
+ // Set the size of NewBB in BBSizes. It does not include any padding now.
+ BBSizes[NewBBI] = NewBBSize;
+
+ MachineInstr* ThumbJTMI = prior(NewBB->end());
+ if (ThumbJTMI->getOpcode() == ARM::tBR_JTr) {
+ // We've added another 2-byte instruction before this tablejump, which
+ // means we will always need padding if we didn't before, and vice versa.
+
+ // The original offset of the jump instruction was:
+ unsigned OrigOffset = BBOffsets[OrigBBI] + BBSizes[OrigBBI] - delta;
+ if (OrigOffset%4 == 0) {
+ // We had padding before and now we don't. No net change in code size.
+ delta = 0;
+ } else {
+ // We didn't have padding before and now we do.
+ BBSizes[NewBBI] += 2;
+ delta = 4;
+ }
+ }
+
// All BBOffsets following these blocks must be modified.
- AdjustBBOffsetsAfter(NewBB, delta);
+ if (delta)
+ AdjustBBOffsetsAfter(NewBB, delta);
return NewBB;
}
/// OffsetIsInRange - Checks whether UserOffset (the location of a constant pool
-/// reference) is within MaxDisp of TrialOffset (a proposed location of a
+/// 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.
+bool ARMConstantIslands::OffsetIsInRange(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.
- if (isThumb && UserOffset%4 !=0)
+ 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)
+ 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)
+ if (TrialOffset - UserOffset <= MaxDisp)
return true;
+ // FIXME: Make use full range of soimm values.
} else if (NegativeOK) {
- if (UserOffset-TrialOffset <= MaxDisp)
+ if (UserOffset - TrialOffset <= MaxDisp)
return true;
+ // FIXME: Make use full range of soimm values.
}
return false;
}
/// Water (a basic block) will be in range for the specific MI.
bool ARMConstantIslands::WaterIsInRange(unsigned UserOffset,
- MachineBasicBlock* Water, CPUser &U)
-{
+ MachineBasicBlock* Water, CPUser &U) {
unsigned MaxDisp = U.MaxDisp;
- MachineFunction::iterator I = next(MachineFunction::iterator(Water));
- unsigned CPEOffset = BBOffsets[Water->getNumber()] +
+ unsigned CPEOffset = BBOffsets[Water->getNumber()] +
BBSizes[Water->getNumber()];
// If the CPE is to be inserted before the instruction, that will raise
- // the offset of the instruction. (Currently applies only to ARM, so
- // no alignment compensation attempted here.)
+ // the offset of the instruction.
if (CPEOffset < UserOffset)
UserOffset += U.CPEMI->getOperand(2).getImm();
- return OffsetIsInRange (UserOffset, CPEOffset, MaxDisp, !isThumb);
+ return OffsetIsInRange(UserOffset, CPEOffset, MaxDisp, U.NegOk, U.IsSoImm);
}
/// CPEIsInRange - Returns true if the distance between specific MI and
/// specific ConstPool entry instruction can fit in MI's displacement field.
bool ARMConstantIslands::CPEIsInRange(MachineInstr *MI, unsigned UserOffset,
- MachineInstr *CPEMI,
- unsigned MaxDisp, bool DoDump) {
+ MachineInstr *CPEMI, unsigned MaxDisp,
+ bool NegOk, bool DoDump) {
unsigned CPEOffset = GetOffsetOf(CPEMI);
- assert(CPEOffset%4 == 0 && "Misaligned CPE");
+ assert((CPEOffset%4 == 0 || HasInlineAsm) && "Misaligned CPE");
if (DoDump) {
- DOUT << "User of CPE#" << CPEMI->getOperand(0).getImm()
- << " max delta=" << MaxDisp
- << " insn address=" << UserOffset
- << " CPE address=" << CPEOffset
- << " offset=" << int(CPEOffset-UserOffset) << "\t" << *MI;
+ DEBUG(errs() << "User of CPE#" << CPEMI->getOperand(0).getImm()
+ << " max delta=" << MaxDisp
+ << " insn address=" << UserOffset
+ << " CPE address=" << CPEOffset
+ << " offset=" << int(CPEOffset-UserOffset) << "\t" << *MI);
}
- return OffsetIsInRange(UserOffset, CPEOffset, MaxDisp, !isThumb);
+ return OffsetIsInRange(UserOffset, CPEOffset, MaxDisp, NegOk);
}
+#ifndef NDEBUG
/// BBIsJumpedOver - Return true of the specified basic block's only predecessor
/// unconditionally branches to its only successor.
static bool BBIsJumpedOver(MachineBasicBlock *MBB) {
MachineBasicBlock *Succ = *MBB->succ_begin();
MachineBasicBlock *Pred = *MBB->pred_begin();
MachineInstr *PredMI = &Pred->back();
- if (PredMI->getOpcode() == ARM::B || PredMI->getOpcode() == ARM::tB)
+ if (PredMI->getOpcode() == ARM::B || PredMI->getOpcode() == ARM::tB
+ || PredMI->getOpcode() == ARM::t2B)
return PredMI->getOperand(0).getMBB() == Succ;
return false;
}
+#endif // NDEBUG
-void ARMConstantIslands::AdjustBBOffsetsAfter(MachineBasicBlock *BB,
+void ARMConstantIslands::AdjustBBOffsetsAfter(MachineBasicBlock *BB,
int delta) {
- MachineFunction::iterator MBBI = BB; MBBI = next(MBBI);
- for(unsigned i=BB->getNumber()+1; i<BB->getParent()->getNumBlockIDs(); i++) {
+ MachineFunction::iterator MBBI = BB; MBBI = llvm::next(MBBI);
+ for(unsigned i = BB->getNumber()+1, e = BB->getParent()->getNumBlockIDs();
+ i < e; ++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;
- }
+ if (!isThumb)
+ continue;
+ MachineBasicBlock *MBB = MBBI;
+ if (!MBB->empty() && !HasInlineAsm) {
+ // 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;
- }
+ }
+ // Thumb1 jump tables require padding. They should be at the end;
+ // following unconditional branches are removed by AnalyzeBranch.
+ // tBR_JTr expands to a mov pc followed by .align 2 and then the jump
+ // table entries. So this code checks whether offset of tBR_JTr
+ // is aligned; if it is, the offset of the jump table following the
+ // instruction will not be aligned, and we need padding.
+ MachineInstr *ThumbJTMI = prior(MBB->end());
+ if (ThumbJTMI->getOpcode() == ARM::tBR_JTr) {
+ 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);
+ if (delta==0)
+ return;
}
+ MBBI = llvm::next(MBBI);
}
}
/// DecrementOldEntry - find the constant pool entry with index CPI
/// and instruction CPEMI, and decrement its refcount. If the refcount
-/// becomes 0 remove the entry and instruction. Returns true if we removed
+/// 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) {
if (--CPE->RefCount == 0) {
RemoveDeadCPEMI(CPEMI);
CPE->CPEMI = NULL;
- NumCPEs--;
+ --NumCPEs;
return true;
}
return false;
MachineInstr *CPEMI = U.CPEMI;
// Check to see if the CPE is already in-range.
- if (CPEIsInRange(UserMI, UserOffset, CPEMI, U.MaxDisp, true)) {
- DOUT << "In range\n";
+ if (CPEIsInRange(UserMI, UserOffset, CPEMI, U.MaxDisp, U.NegOk, true)) {
+ DEBUG(errs() << "In range\n");
return 1;
}
// No. Look for previously created clones of the CPE that are in range.
- unsigned CPI = CPEMI->getOperand(1).getConstantPoolIndex();
+ unsigned CPI = CPEMI->getOperand(1).getIndex();
std::vector<CPEntry> &CPEs = CPEntries[CPI];
for (unsigned i = 0, e = CPEs.size(); i != e; ++i) {
// We already tried this one
// Removing CPEs can leave empty entries, skip
if (CPEs[i].CPEMI == NULL)
continue;
- if (CPEIsInRange(UserMI, UserOffset, CPEs[i].CPEMI, U.MaxDisp, false)) {
- DOUT << "Replacing CPE#" << CPI << " with CPE#" << CPEs[i].CPI << "\n";
+ if (CPEIsInRange(UserMI, UserOffset, CPEs[i].CPEMI, U.MaxDisp, U.NegOk)) {
+ DEBUG(errs() << "Replacing CPE#" << CPI << " with CPE#"
+ << CPEs[i].CPI << "\n");
// Point the CPUser node to the replacement
U.CPEMI = CPEs[i].CPEMI;
// Change the CPI in the instruction operand to refer to the clone.
for (unsigned j = 0, e = UserMI->getNumOperands(); j != e; ++j)
- if (UserMI->getOperand(j).isConstantPoolIndex()) {
- UserMI->getOperand(j).setConstantPoolIndex(CPEs[i].CPI);
+ if (UserMI->getOperand(j).isCPI()) {
+ UserMI->getOperand(j).setIndex(CPEs[i].CPI);
break;
}
// Adjust the refcount of the clone...
/// 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;
-}
+ switch (Opc) {
+ case ARM::tB:
+ return ((1<<10)-1)*2;
+ case ARM::t2B:
+ return ((1<<23)-1)*2;
+ default:
+ break;
+ }
-/// AcceptWater - Small amount of common code factored out of the following.
-
-MachineBasicBlock* ARMConstantIslands::AcceptWater(MachineBasicBlock *WaterBB,
- std::vector<MachineBasicBlock*>::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));
+ return ((1<<23)-1)*4;
}
-/// LookForWater - look for an existing entry in the WaterList in which
+/// 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.
-
+/// 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
+/// introduce padding to water that will. To ensure that this pass
+/// 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,
- MachineBasicBlock** NewMBB) {
- std::vector<MachineBasicBlock*>::iterator IPThatWouldPad;
- MachineBasicBlock* WaterBBThatWouldPad = NULL;
- if (!WaterList.empty()) {
- for (std::vector<MachineBasicBlock*>::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;
+ 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) {
+ 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
+ // the previous iteration by inserting an unconditional branch. In the
+ // latter case, we want to allow resetting the high water mark back to
+ // this new water since we haven't seen it before. Inserting branches
+ // 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) &&
+ (WaterBB->getNumber() < U.HighWaterMark->getNumber() ||
+ NewWaterList.count(WaterBB))) {
+ unsigned WBBId = WaterBB->getNumber();
+ if (isThumb &&
+ (BBOffsets[WBBId] + BBSizes[WBBId])%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;
+ return true;
+ }
}
- if (IP == B)
- break;
- }
+ if (IP == B)
+ break;
}
- if (isThumb && WaterBBThatWouldPad) {
- *NewMBB = AcceptWater(WaterBBThatWouldPad, IPThatWouldPad);
+ if (FoundWaterThatWouldPad) {
+ WaterIter = IPThatWouldPad;
return true;
}
return false;
}
-/// 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
+/// 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) {
+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()] +
+ 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
+ // 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 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 (BBHasFallthrough(UserMBB) &&
+ OffsetIsInRange(UserOffset, OffsetOfNextBlock + (isThumb1 ? 2: 4),
+ U.MaxDisp, U.NegOk, U.IsSoImm)) {
+ DEBUG(errs() << "Split at end of block\n");
if (&UserMBB->back() == UserMI)
assert(BBHasFallthrough(UserMBB) && "Expected a fallthrough BB!");
- *NewMBB = next(MachineFunction::iterator(UserMBB));
+ 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 ? ARM::tB : ARM::B;
- BuildMI(UserMBB, TII->get(UncondBr)).addMBB(*NewMBB);
+ int UncondBr = isThumb ? ((isThumb2) ? ARM::t2B : ARM::tB) : ARM::B;
+ BuildMI(UserMBB, DebugLoc(), TII->get(UncondBr)).addMBB(NewMBB);
unsigned MaxDisp = getUnconditionalBrDisp(UncondBr);
- ImmBranches.push_back(ImmBranch(&UserMBB->back(),
+ ImmBranches.push_back(ImmBranch(&UserMBB->back(),
MaxDisp, false, UncondBr));
- int delta = isThumb ? 2 : 4;
+ int delta = isThumb1 ? 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
+ // 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
// 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
+ // inserting (allows for long branch on Thumb1). 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
// 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);
+ BaseInsertOffset = BBOffsets[UserMBB->getNumber()+1] -
+ (isThumb1 ? 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);
+ for (unsigned Offset = UserOffset+TII->GetInstSizeInBytes(UserMI);
Offset < BaseInsertOffset;
- Offset += ARM::GetInstSize(MI),
- MI = next(MI)) {
+ Offset += TII->GetInstSizeInBytes(MI),
+ MI = llvm::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);
+ CPUser &U = CPUsers[CPUIndex];
+ if (!OffsetIsInRange(Offset, EndInsertOffset,
+ U.MaxDisp, U.NegOk, U.IsSoImm)) {
+ BaseInsertOffset -= (isThumb1 ? 2 : 4);
+ EndInsertOffset -= (isThumb1 ? 2 : 4);
}
// This is overly conservative, as we don't account for CPEMIs
// being reused within the block, but it doesn't matter much.
CPUIndex++;
}
}
- DOUT << "Split in middle of big block\n";
- *NewMBB = SplitBlockBeforeInstr(prior(MI));
+ DEBUG(errs() << "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
+/// 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 &Fn,
- unsigned CPUserIndex){
+bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &MF,
+ unsigned CPUserIndex) {
CPUser &U = CPUsers[CPUserIndex];
MachineInstr *UserMI = U.MI;
MachineInstr *CPEMI = U.CPEMI;
- unsigned CPI = CPEMI->getOperand(1).getConstantPoolIndex();
+ unsigned CPI = CPEMI->getOperand(1).getIndex();
unsigned Size = CPEMI->getOperand(2).getImm();
- MachineBasicBlock *NewMBB;
// Compute this only once, it's expensive. The 4 or 8 is the value the
- // hardware keeps in the PC (2 insns ahead of the reference).
+ // hardware keeps in the PC.
unsigned UserOffset = GetOffsetOf(UserMI) + (isThumb ? 4 : 8);
- // Special case: tLEApcrel are two instructions MI's. The actual user is the
- // second instruction.
- if (UserMI->getOpcode() == ARM::tLEApcrel)
- UserOffset += 2;
-
// See if the current entry is within range, or there is a clone of it
// in range.
int result = LookForExistingCPEntry(U, UserOffset);
// No existing clone of this CPE is within range.
// We will be generating a new clone. Get a UID for it.
- unsigned ID = NextUID++;
+ unsigned ID = AFI->createConstPoolEntryUId();
- // 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).
+ // Look for water where we can place this CPE.
+ MachineBasicBlock *NewIsland = MF.CreateMachineBasicBlock();
+ MachineBasicBlock *NewMBB;
+ water_iterator IP;
+ if (LookForWater(U, UserOffset, IP)) {
+ DEBUG(errs() << "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);
+ NewWaterList.insert(NewIsland);
+ }
+ // The new CPE goes before the following block (NewMBB).
+ NewMBB = llvm::next(MachineFunction::iterator(WaterBB));
- if (!LookForWater(U, UserOffset, &NewMBB)) {
+ } else {
// No water found.
- DOUT << "No water found\n";
- CreateNewWater(CPUserIndex, UserOffset, &NewMBB);
+ DEBUG(errs() << "No water found\n");
+ CreateNewWater(CPUserIndex, UserOffset, NewMBB);
+
+ // 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));
+ IP = std::find(WaterList.begin(), WaterList.end(), WaterBB);
+ if (IP != WaterList.end())
+ NewWaterList.erase(WaterBB);
+
+ // We are adding new water. Update NewWaterList.
+ NewWaterList.insert(NewIsland);
}
+ // 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 and is also important to ensure the algorithm terminates.
+ if (IP != WaterList.end())
+ WaterList.erase(IP);
+
// Okay, we know we can put an island before NewMBB now, do it!
- MachineBasicBlock *NewIsland = new MachineBasicBlock();
- Fn.getBasicBlockList().insert(NewMBB, NewIsland);
+ MF.insert(NewMBB, NewIsland);
// Update internal data structures to account for the newly inserted MBB.
UpdateForInsertedWaterBlock(NewIsland);
// Now that we have an island to add the CPE to, clone the original CPE and
// add it to the island.
- U.CPEMI = BuildMI(NewIsland, TII->get(ARM::CONSTPOOL_ENTRY))
+ U.HighWaterMark = NewIsland;
+ U.CPEMI = BuildMI(NewIsland, DebugLoc(), TII->get(ARM::CONSTPOOL_ENTRY))
.addImm(ID).addConstantPoolIndex(CPI).addImm(Size);
CPEntries[CPI].push_back(CPEntry(U.CPEMI, ID, 1));
- NumCPEs++;
+ ++NumCPEs;
BBOffsets[NewIsland->getNumber()] = BBOffsets[NewMBB->getNumber()];
// Compensate for .align 2 in thumb mode.
- if (isThumb && BBOffsets[NewIsland->getNumber()]%4 != 0)
+ if (isThumb && (BBOffsets[NewIsland->getNumber()]%4 != 0 || HasInlineAsm))
Size += 2;
// Increase the size of the island block to account for the new entry.
BBSizes[NewIsland->getNumber()] += Size;
AdjustBBOffsetsAfter(NewIsland, Size);
-
+
// Finally, change the CPI in the instruction operand to be ID.
for (unsigned i = 0, e = UserMI->getNumOperands(); i != e; ++i)
- if (UserMI->getOperand(i).isConstantPoolIndex()) {
- UserMI->getOperand(i).setConstantPoolIndex(ID);
+ if (UserMI->getOperand(i).isCPI()) {
+ UserMI->getOperand(i).setIndex(ID);
break;
}
-
- DOUT << " Moved CPE to #" << ID << " CPI=" << CPI << "\t" << *UserMI;
-
+
+ DEBUG(errs() << " Moved CPE to #" << ID << " CPI=" << CPI
+ << '\t' << *UserMI);
+
return true;
}
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
+ // 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.
MadeChange = true;
}
}
- }
+ }
return MadeChange;
}
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
- << " from " << GetOffsetOf(MI) << " to " << DestOffset
- << " offset " << int(DestOffset-BrOffset) << "\t" << *MI;
+ DEBUG(errs() << "Branch of destination BB#" << DestBB->getNumber()
+ << " from BB#" << MI->getParent()->getNumber()
+ << " max delta=" << MaxDisp
+ << " from " << GetOffsetOf(MI) << " to " << DestOffset
+ << " offset " << int(DestOffset-BrOffset) << "\t" << *MI);
if (BrOffset <= DestOffset) {
// Branch before the Dest.
/// FixUpImmediateBr - Fix up an immediate branch whose destination is too far
/// away to fit in its displacement field.
-bool ARMConstantIslands::FixUpImmediateBr(MachineFunction &Fn, ImmBranch &Br) {
+bool ARMConstantIslands::FixUpImmediateBr(MachineFunction &MF, ImmBranch &Br) {
MachineInstr *MI = Br.MI;
- MachineBasicBlock *DestBB = MI->getOperand(0).getMachineBasicBlock();
+ MachineBasicBlock *DestBB = MI->getOperand(0).getMBB();
// Check to see if the DestBB is already in-range.
if (BBIsInRange(MI, DestBB, Br.MaxDisp))
return false;
if (!Br.isCond)
- return FixUpUnconditionalBr(Fn, Br);
- return FixUpConditionalBr(Fn, Br);
+ return FixUpUnconditionalBr(MF, Br);
+ return FixUpConditionalBr(MF, Br);
}
/// 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 to a branch.
+/// Otherwise, add an intermediate branch instruction to a branch.
bool
-ARMConstantIslands::FixUpUnconditionalBr(MachineFunction &Fn, ImmBranch &Br) {
+ARMConstantIslands::FixUpUnconditionalBr(MachineFunction &MF, ImmBranch &Br) {
MachineInstr *MI = Br.MI;
MachineBasicBlock *MBB = MI->getParent();
- assert(isThumb && "Expected a Thumb function!");
+ if (!isThumb1)
+ llvm_unreachable("FixUpUnconditionalBr is Thumb1 only!");
// Use BL to implement far jump.
Br.MaxDisp = (1 << 21) * 2;
- MI->setInstrDescriptor(TII->get(ARM::tBfar));
+ MI->setDesc(TII->get(ARM::tBfar));
BBSizes[MBB->getNumber()] += 2;
AdjustBBOffsetsAfter(MBB, 2);
HasFarJump = true;
- NumUBrFixed++;
+ ++NumUBrFixed;
- DOUT << " Changed B to long jump " << *MI;
+ DEBUG(errs() << " Changed B to long jump " << *MI);
return true;
}
/// 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 &Fn, ImmBranch &Br) {
+ARMConstantIslands::FixUpConditionalBr(MachineFunction &MF, ImmBranch &Br) {
MachineInstr *MI = Br.MI;
- MachineBasicBlock *DestBB = MI->getOperand(0).getMachineBasicBlock();
+ MachineBasicBlock *DestBB = MI->getOperand(0).getMBB();
- // Add a unconditional branch to the destination and invert the branch
+ // Add an unconditional branch to the destination and invert the branch
// condition to jump over it:
// blt L1
// =>
// bge L2
// b L1
// L2:
- ARMCC::CondCodes CC = (ARMCC::CondCodes)MI->getOperand(1).getImmedValue();
+ ARMCC::CondCodes CC = (ARMCC::CondCodes)MI->getOperand(1).getImm();
CC = ARMCC::getOppositeCondition(CC);
unsigned CCReg = MI->getOperand(2).getReg();
MachineInstr *BMI = &MBB->back();
bool NeedSplit = (BMI != MI) || !BBHasFallthrough(MBB);
- NumCBrFixed++;
+ ++NumCBrFixed;
if (BMI != MI) {
- if (next(MachineBasicBlock::iterator(MI)) == MBB->back() &&
+ if (llvm::next(MachineBasicBlock::iterator(MI)) == prior(MBB->end()) &&
BMI->getOpcode() == Br.UncondBr) {
- // Last MI in the BB is a unconditional branch. Can we simply invert the
+ // Last MI in the BB is an unconditional branch. Can we simply invert the
// condition and swap destinations:
// beq L1
// b L2
// =>
// bne L2
// b L1
- MachineBasicBlock *NewDest = BMI->getOperand(0).getMachineBasicBlock();
+ MachineBasicBlock *NewDest = BMI->getOperand(0).getMBB();
if (BBIsInRange(MI, NewDest, Br.MaxDisp)) {
- DOUT << " Invert Bcc condition and swap its destination with " << *BMI;
- BMI->getOperand(0).setMachineBasicBlock(DestBB);
- MI->getOperand(0).setMachineBasicBlock(NewDest);
+ DEBUG(errs() << " Invert Bcc condition and swap its destination with "
+ << *BMI);
+ BMI->getOperand(0).setMBB(DestBB);
+ MI->getOperand(0).setMBB(NewDest);
MI->getOperand(1).setImm(CC);
return true;
}
if (NeedSplit) {
SplitBlockBeforeInstr(MI);
- // No need for the branch to the next block. We're adding a unconditional
+ // No need for the branch to the next block. We're adding an unconditional
// branch to the destination.
- int delta = ARM::GetInstSize(&MBB->back());
+ int delta = TII->GetInstSizeInBytes(&MBB->back());
BBSizes[MBB->getNumber()] -= delta;
- MachineBasicBlock* SplitBB = next(MachineFunction::iterator(MBB));
+ MachineBasicBlock* SplitBB = llvm::next(MachineFunction::iterator(MBB));
AdjustBBOffsetsAfter(SplitBB, -delta);
MBB->back().eraseFromParent();
// BBOffsets[SplitBB] is wrong temporarily, fixed below
}
- MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
-
- DOUT << " Insert B to BB#" << DestBB->getNumber()
- << " also invert condition and change dest. to BB#"
- << NextBB->getNumber() << "\n";
+ MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
+
+ DEBUG(errs() << " Insert B to BB#" << DestBB->getNumber()
+ << " also invert condition and change dest. to BB#"
+ << NextBB->getNumber() << "\n");
// 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).addReg(CCReg);
+ BuildMI(MBB, DebugLoc(), 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());
+ BBSizes[MBB->getNumber()] += TII->GetInstSizeInBytes(&MBB->back());
+ BuildMI(MBB, DebugLoc(), TII->get(Br.UncondBr)).addMBB(DestBB);
+ BBSizes[MBB->getNumber()] += TII->GetInstSizeInBytes(&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);
+ BBSizes[MI->getParent()->getNumber()] -= TII->GetInstSizeInBytes(MI);
MI->eraseFromParent();
// The net size change is an addition of one unconditional branch.
- int delta = ARM::GetInstSize(&MBB->back());
+ int delta = TII->GetInstSizeInBytes(&MBB->back());
AdjustBBOffsetsAfter(MBB, delta);
return true;
}
/// UndoLRSpillRestore - Remove Thumb push / pop instructions that only spills
-/// LR / restores LR to pc.
+/// 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 MadeChange = false;
for (unsigned i = 0, e = PushPopMIs.size(); i != e; ++i) {
MachineInstr *MI = PushPopMIs[i];
+ // First two operands are predicates.
if (MI->getOpcode() == ARM::tPOP_RET &&
- MI->getOperand(0).getReg() == ARM::PC &&
- MI->getNumExplicitOperands() == 1) {
- BuildMI(MI->getParent(), TII->get(ARM::tBX_RET));
+ MI->getOperand(2).getReg() == ARM::PC &&
+ MI->getNumExplicitOperands() == 3) {
+ BuildMI(MI->getParent(), MI->getDebugLoc(), TII->get(ARM::tBX_RET));
+ MI->eraseFromParent();
+ MadeChange = true;
+ }
+ }
+ return MadeChange;
+}
+
+bool ARMConstantIslands::OptimizeThumb2Instructions(MachineFunction &MF) {
+ bool MadeChange = false;
+
+ // Shrink ADR and LDR from constantpool.
+ for (unsigned i = 0, e = CPUsers.size(); i != e; ++i) {
+ CPUser &U = CPUsers[i];
+ unsigned Opcode = U.MI->getOpcode();
+ unsigned NewOpc = 0;
+ unsigned Scale = 1;
+ unsigned Bits = 0;
+ switch (Opcode) {
+ default: break;
+ case ARM::t2LEApcrel:
+ if (isARMLowRegister(U.MI->getOperand(0).getReg())) {
+ NewOpc = ARM::tLEApcrel;
+ Bits = 8;
+ Scale = 4;
+ }
+ break;
+ case ARM::t2LDRpci:
+ if (isARMLowRegister(U.MI->getOperand(0).getReg())) {
+ NewOpc = ARM::tLDRpci;
+ Bits = 8;
+ Scale = 4;
+ }
+ break;
+ }
+
+ if (!NewOpc)
+ continue;
+
+ unsigned UserOffset = GetOffsetOf(U.MI) + 4;
+ unsigned MaxOffs = ((1 << Bits) - 1) * Scale;
+ // FIXME: Check if offset is multiple of scale if scale is not 4.
+ if (CPEIsInRange(U.MI, UserOffset, U.CPEMI, MaxOffs, false, true)) {
+ U.MI->setDesc(TII->get(NewOpc));
+ MachineBasicBlock *MBB = U.MI->getParent();
+ BBSizes[MBB->getNumber()] -= 2;
+ AdjustBBOffsetsAfter(MBB, -2);
+ ++NumT2CPShrunk;
+ MadeChange = true;
+ }
+ }
+
+ MadeChange |= OptimizeThumb2Branches(MF);
+ MadeChange |= OptimizeThumb2JumpTables(MF);
+ return MadeChange;
+}
+
+bool ARMConstantIslands::OptimizeThumb2Branches(MachineFunction &MF) {
+ bool MadeChange = false;
+
+ for (unsigned i = 0, e = ImmBranches.size(); i != e; ++i) {
+ ImmBranch &Br = ImmBranches[i];
+ unsigned Opcode = Br.MI->getOpcode();
+ unsigned NewOpc = 0;
+ unsigned Scale = 1;
+ unsigned Bits = 0;
+ switch (Opcode) {
+ default: break;
+ case ARM::t2B:
+ NewOpc = ARM::tB;
+ Bits = 11;
+ Scale = 2;
+ break;
+ case ARM::t2Bcc: {
+ NewOpc = ARM::tBcc;
+ Bits = 8;
+ Scale = 2;
+ break;
+ }
+ }
+ if (NewOpc) {
+ unsigned MaxOffs = ((1 << (Bits-1))-1) * Scale;
+ MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB();
+ if (BBIsInRange(Br.MI, DestBB, MaxOffs)) {
+ Br.MI->setDesc(TII->get(NewOpc));
+ MachineBasicBlock *MBB = Br.MI->getParent();
+ BBSizes[MBB->getNumber()] -= 2;
+ AdjustBBOffsetsAfter(MBB, -2);
+ ++NumT2BrShrunk;
+ MadeChange = true;
+ }
+ }
+
+ Opcode = Br.MI->getOpcode();
+ if (Opcode != ARM::tBcc)
+ continue;
+
+ NewOpc = 0;
+ unsigned PredReg = 0;
+ ARMCC::CondCodes Pred = llvm::getInstrPredicate(Br.MI, PredReg);
+ if (Pred == ARMCC::EQ)
+ NewOpc = ARM::tCBZ;
+ else if (Pred == ARMCC::NE)
+ NewOpc = ARM::tCBNZ;
+ if (!NewOpc)
+ continue;
+ 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 DestOffset = BBOffsets[DestBB->getNumber()];
+ if (BrOffset < DestOffset && (DestOffset - BrOffset) <= 126) {
+ MachineBasicBlock::iterator CmpMI = Br.MI; --CmpMI;
+ if (CmpMI->getOpcode() == ARM::tCMPzi8) {
+ unsigned Reg = CmpMI->getOperand(0).getReg();
+ Pred = llvm::getInstrPredicate(CmpMI, PredReg);
+ if (Pred == ARMCC::AL &&
+ CmpMI->getOperand(1).getImm() == 0 &&
+ isARMLowRegister(Reg)) {
+ MachineBasicBlock *MBB = Br.MI->getParent();
+ MachineInstr *NewBR =
+ BuildMI(*MBB, CmpMI, Br.MI->getDebugLoc(), TII->get(NewOpc))
+ .addReg(Reg).addMBB(DestBB, Br.MI->getOperand(0).getTargetFlags());
+ CmpMI->eraseFromParent();
+ Br.MI->eraseFromParent();
+ Br.MI = NewBR;
+ BBSizes[MBB->getNumber()] -= 2;
+ AdjustBBOffsetsAfter(MBB, -2);
+ ++NumCBZ;
+ MadeChange = true;
+ }
+ }
+ }
+ }
+
+ return MadeChange;
+}
+
+/// OptimizeThumb2JumpTables - Use tbb / tbh instructions to generate smaller
+/// jumptables when it's possible.
+bool ARMConstantIslands::OptimizeThumb2JumpTables(MachineFunction &MF) {
+ bool MadeChange = false;
+
+ // FIXME: After the tables are shrunk, can we get rid some of the
+ // constantpool tables?
+ MachineJumpTableInfo *MJTI = MF.getJumpTableInfo();
+ if (MJTI == 0) return false;
+
+ const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
+ for (unsigned i = 0, e = T2JumpTables.size(); i != e; ++i) {
+ MachineInstr *MI = T2JumpTables[i];
+ const TargetInstrDesc &TID = MI->getDesc();
+ unsigned NumOps = TID.getNumOperands();
+ unsigned JTOpIdx = NumOps - (TID.isPredicable() ? 3 : 2);
+ MachineOperand JTOP = MI->getOperand(JTOpIdx);
+ unsigned JTI = JTOP.getIndex();
+ assert(JTI < JT.size());
+
+ bool ByteOk = true;
+ bool HalfWordOk = true;
+ unsigned JTOffset = GetOffsetOf(MI) + 4;
+ const std::vector<MachineBasicBlock*> &JTBBs = JT[JTI].MBBs;
+ for (unsigned j = 0, ee = JTBBs.size(); j != ee; ++j) {
+ MachineBasicBlock *MBB = JTBBs[j];
+ unsigned DstOffset = BBOffsets[MBB->getNumber()];
+ // Negative offset is not ok. FIXME: We should change BB layout to make
+ // sure all the branches are forward.
+ if (ByteOk && (DstOffset - JTOffset) > ((1<<8)-1)*2)
+ ByteOk = false;
+ unsigned TBHLimit = ((1<<16)-1)*2;
+ if (HalfWordOk && (DstOffset - JTOffset) > TBHLimit)
+ HalfWordOk = false;
+ if (!ByteOk && !HalfWordOk)
+ break;
+ }
+
+ if (ByteOk || HalfWordOk) {
+ MachineBasicBlock *MBB = MI->getParent();
+ unsigned BaseReg = MI->getOperand(0).getReg();
+ bool BaseRegKill = MI->getOperand(0).isKill();
+ if (!BaseRegKill)
+ continue;
+ unsigned IdxReg = MI->getOperand(1).getReg();
+ bool IdxRegKill = MI->getOperand(1).isKill();
+
+ // Scan backwards to find the instruction that defines the base
+ // register. Due to post-RA scheduling, we can't count on it
+ // immediately preceding the branch instruction.
+ MachineBasicBlock::iterator PrevI = MI;
+ MachineBasicBlock::iterator B = MBB->begin();
+ while (PrevI != B && !PrevI->definesRegister(BaseReg))
+ --PrevI;
+
+ // If for some reason we didn't find it, we can't do anything, so
+ // just skip this one.
+ if (!PrevI->definesRegister(BaseReg))
+ continue;
+
+ MachineInstr *AddrMI = PrevI;
+ bool OptOk = true;
+ // Examine the instruction that calculates the jumptable entry address.
+ // Make sure it only defines the base register and kills any uses
+ // other than the index register.
+ for (unsigned k = 0, eee = AddrMI->getNumOperands(); k != eee; ++k) {
+ const MachineOperand &MO = AddrMI->getOperand(k);
+ if (!MO.isReg() || !MO.getReg())
+ continue;
+ if (MO.isDef() && MO.getReg() != BaseReg) {
+ OptOk = false;
+ break;
+ }
+ if (MO.isUse() && !MO.isKill() && MO.getReg() != IdxReg) {
+ OptOk = false;
+ break;
+ }
+ }
+ if (!OptOk)
+ continue;
+
+ // Now scan back again to find the tLEApcrel or t2LEApcrelJT instruction
+ // that gave us the initial base register definition.
+ for (--PrevI; PrevI != B && !PrevI->definesRegister(BaseReg); --PrevI)
+ ;
+
+ // The instruction should be a tLEApcrel or t2LEApcrelJT; we want
+ // to delete it as well.
+ MachineInstr *LeaMI = PrevI;
+ if ((LeaMI->getOpcode() != ARM::tLEApcrelJT &&
+ LeaMI->getOpcode() != ARM::t2LEApcrelJT) ||
+ LeaMI->getOperand(0).getReg() != BaseReg)
+ OptOk = false;
+
+ if (!OptOk)
+ continue;
+
+ unsigned Opc = ByteOk ? ARM::t2TBB : ARM::t2TBH;
+ MachineInstr *NewJTMI = BuildMI(MBB, MI->getDebugLoc(), TII->get(Opc))
+ .addReg(IdxReg, getKillRegState(IdxRegKill))
+ .addJumpTableIndex(JTI, JTOP.getTargetFlags())
+ .addImm(MI->getOperand(JTOpIdx+1).getImm());
+ // 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);
+ unsigned OrigSize = TII->GetInstSizeInBytes(AddrMI);
+ OrigSize += TII->GetInstSizeInBytes(LeaMI);
+ OrigSize += TII->GetInstSizeInBytes(MI);
+
+ AddrMI->eraseFromParent();
+ LeaMI->eraseFromParent();
MI->eraseFromParent();
+
+ int delta = OrigSize - NewSize;
+ BBSizes[MBB->getNumber()] -= delta;
+ AdjustBBOffsetsAfter(MBB, -delta);
+
+ ++NumTBs;
MadeChange = true;
}
}
+
+ return MadeChange;
+}
+
+/// 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 MadeChange = false;
+
+ MachineJumpTableInfo *MJTI = MF.getJumpTableInfo();
+ if (MJTI == 0) return false;
+
+ const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
+ for (unsigned i = 0, e = T2JumpTables.size(); i != e; ++i) {
+ MachineInstr *MI = T2JumpTables[i];
+ const TargetInstrDesc &TID = MI->getDesc();
+ unsigned NumOps = TID.getNumOperands();
+ unsigned JTOpIdx = NumOps - (TID.isPredicable() ? 3 : 2);
+ MachineOperand JTOP = MI->getOperand(JTOpIdx);
+ unsigned JTI = JTOP.getIndex();
+ assert(JTI < JT.size());
+
+ // We prefer if target blocks for the jump table come after the jump
+ // instruction so we can use TB[BH]. Loop through the target blocks
+ // and try to adjust them such that that's true.
+ int JTNumber = MI->getParent()->getNumber();
+ const std::vector<MachineBasicBlock*> &JTBBs = JT[JTI].MBBs;
+ for (unsigned j = 0, ee = JTBBs.size(); j != ee; ++j) {
+ MachineBasicBlock *MBB = JTBBs[j];
+ int DTNumber = MBB->getNumber();
+
+ if (DTNumber < JTNumber) {
+ // The destination precedes the switch. Try to move the block forward
+ // so we have a positive offset.
+ MachineBasicBlock *NewBB =
+ AdjustJTTargetBlockForward(MBB, MI->getParent());
+ if (NewBB)
+ MJTI->ReplaceMBBInJumpTable(JTI, JTBBs[j], NewBB);
+ MadeChange = true;
+ }
+ }
+ }
+
return MadeChange;
}
+
+MachineBasicBlock *ARMConstantIslands::
+AdjustJTTargetBlockForward(MachineBasicBlock *BB, MachineBasicBlock *JTBB)
+{
+ MachineFunction &MF = *BB->getParent();
+
+ // 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
+ // heuristic. FIXME: We can definitely improve it.
+ MachineBasicBlock *TBB = 0, *FBB = 0;
+ SmallVector<MachineOperand, 4> Cond;
+ SmallVector<MachineOperand, 4> CondPrior;
+ MachineFunction::iterator BBi = BB;
+ MachineFunction::iterator OldPrior = prior(BBi);
+
+ // If the block terminator isn't analyzable, don't try to move the block
+ bool B = TII->AnalyzeBranch(*BB, TBB, FBB, Cond);
+
+ // 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() &&
+ !TII->AnalyzeBranch(*OldPrior, TBB, FBB, CondPrior)) {
+ BB->moveAfter(JTBB);
+ OldPrior->updateTerminator();
+ BB->updateTerminator();
+ // Update numbering to account for the block being moved.
+ MF.RenumberBlocks();
+ ++NumJTMoved;
+ return NULL;
+ }
+
+ // Create a new MBB for the code after the jump BB.
+ MachineBasicBlock *NewBB =
+ MF.CreateMachineBasicBlock(JTBB->getBasicBlock());
+ MachineFunction::iterator MBBI = JTBB; ++MBBI;
+ MF.insert(MBBI, NewBB);
+
+ // Add an unconditional branch from NewBB to BB.
+ // There doesn't seem to be meaningful DebugInfo available; this doesn't
+ // correspond directly to anything in the source.
+ assert (isThumb2 && "Adjusting for TB[BH] but not in Thumb2?");
+ BuildMI(NewBB, DebugLoc(), TII->get(ARM::t2B)).addMBB(BB);
+
+ // Update internal data structures to account for the newly inserted MBB.
+ MF.RenumberBlocks(NewBB);
+
+ // Update the CFG.
+ NewBB->addSuccessor(BB);
+ JTBB->removeSuccessor(BB);
+ JTBB->addSuccessor(NewBB);
+
+ ++NumJTInserted;
+ return NewBB;
+}