#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/SmallVector.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/Statistic.h"
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");
namespace {
/// ARMConstantIslands - Due to limited PC-relative displacements, ARM
/// to a return, unreachable, or unconditional branch).
std::vector<MachineBasicBlock*> WaterList;
+ 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.
///
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;
+ const ARMSubtarget *STI;
ARMFunctionInfo *AFI;
bool isThumb;
bool isThumb1;
static char 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 InitialFunctionScan(MachineFunction &MF,
const std::vector<MachineInstr*> &CPEMIs);
MachineBasicBlock *SplitBlockBeforeInstr(MachineInstr *MI);
void UpdateForInsertedWaterBlock(MachineBasicBlock *NewBB);
bool LookForWater(CPUser&U, unsigned UserOffset,
MachineBasicBlock** NewMBB);
MachineBasicBlock* AcceptWater(MachineBasicBlock *WaterBB,
- std::vector<MachineBasicBlock*>::iterator IP);
+ water_iterator IP);
void CreateNewWater(unsigned CPUserIndex, unsigned UserOffset,
MachineBasicBlock** NewMBB);
- bool HandleConstantPoolUser(MachineFunction &Fn, unsigned CPUserIndex);
+ bool HandleConstantPoolUser(MachineFunction &MF, unsigned CPUserIndex);
void RemoveDeadCPEMI(MachineInstr *CPEMI);
bool RemoveUnusedCPEntries();
bool CPEIsInRange(MachineInstr *MI, unsigned UserOffset,
bool OffsetIsInRange(unsigned UserOffset, unsigned TrialOffset,
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 OptimizeThumb2JumpTables(MachineFunction &MF);
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)
return;
#ifndef NDEBUG
- 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 (!MBB->empty() &&
/// 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();
+bool ARMConstantIslands::runOnMachineFunction(MachineFunction &MF) {
+ MachineConstantPool &MCP = *MF.getConstantPool();
+
+ TII = MF.getTarget().getInstrInfo();
+ AFI = MF.getInfo<ARMFunctionInfo>();
+ STI = &MF.getTarget().getSubtarget<ARMSubtarget>();
- TII = Fn.getTarget().getInstrInfo();
- AFI = Fn.getInfo<ARMFunctionInfo>();
isThumb = AFI->isThumbFunction();
isThumb1 = AFI->isThumb1OnlyFunction();
isThumb2 = AFI->isThumb2Function();
// 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();
- // Thumb1 functions containing constant pools get 2-byte alignment.
+ // Thumb1 functions containing constant pools get 4-byte alignment.
// This is so we can keep exact track of where the alignment padding goes.
- // Set default. Thumb1 function is 1-byte aligned, ARM and Thumb2 are 2-byte
+ // Set default. Thumb1 function is 2-byte aligned, ARM and Thumb2 are 4-byte
// aligned.
AFI->setAlign(isThumb1 ? 1U : 2U);
// we put them all at the end of the function.
std::vector<MachineInstr*> CPEMIs;
if (!MCP.isEmpty()) {
- DoInitialPlacement(Fn, CPEMIs);
+ DoInitialPlacement(MF, CPEMIs);
if (isThumb1)
AFI->setAlign(2U);
}
// 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();
/// Remove dead constant pool entries.
// 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());
+
+ 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 (isThumb && !HasFarJump && AFI->isLRSpilledForFarJump())
MadeChange |= UndoLRSpillRestore();
BBSizes.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,
+void ARMConstantIslands::DoInitialPlacement(MachineFunction &MF,
std::vector<MachineInstr*> &CPEMIs) {
// Create the basic block to hold the CPE's.
- MachineBasicBlock *BB = Fn.CreateMachineBasicBlock();
- Fn.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();
+ MF.getConstantPool()->getConstants();
- const TargetData &TD = *Fn.getTarget().getTargetData();
+ const TargetData &TD = *MF.getTarget().getTargetData();
for (unsigned i = 0, e = CPs.size(); i != e; ++i) {
unsigned Size = TD.getTypeAllocSize(CPs[i].getType());
// Verify that all constant pool entries are a multiple of 4 bytes. If not,
CPEs.push_back(CPEntry(CPEMI, i));
CPEntries.push_back(CPEs);
NumCPEs++;
- DOUT << "Moved CPI#" << i << " to end of function as #" << i << "\n";
+ DEBUG(errs() << "Moved CPI#" << i << " to end of function as #" << i
+ << "\n");
}
}
/// 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) {
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;
unsigned Scale = 1;
int UOpc = Opc;
switch (Opc) {
+ default:
+ continue; // Ignore other JT branches
case ARM::tBR_JTr:
// 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)
+ // 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;
// 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);
// 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;
// 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 = isThumb1 ? 2 : 4;
+ int delta = isThumb1 ? 2 : 4;
BBSizes[OrigBBI] -= NewBBSize - delta;
// ...and adjust BBOffsets for NewBB accordingly.
// 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.
assert(CPEOffset%4 == 0 && "Misaligned CPE");
if (DoDump) {
- DOUT << "User of CPE#" << CPEMI->getOperand(0).getImm()
- << " max delta=" << MaxDisp
- << " insn address=" << UserOffset
- << " CPE address=" << CPEOffset
- << " offset=" << int(CPEOffset-UserOffset) << "\t" << *MI;
+ 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, NegOk);
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) {
+ 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) {
+ } else if ((OldOffset%4) != 0 && (BBOffsets[i]%4) == 0) {
// remove existing padding
- BBSizes[i] -=2;
+ BBSizes[i] -= 2;
delta -= 2;
}
}
// following unconditional branches are removed by AnalyzeBranch.
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) {
+ 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) {
+ } else if ((OldMIOffset%4) != 0 && (NewMIOffset%4) == 0) {
// add new padding
BBSizes[i] += 2;
delta += 2;
// Check to see if the CPE is already in-range.
if (CPEIsInRange(UserMI, UserOffset, CPEMI, U.MaxDisp, U.NegOk, true)) {
- DOUT << "In range\n";
+ DEBUG(errs() << "In range\n");
return 1;
}
if (CPEs[i].CPEMI == NULL)
continue;
if (CPEIsInRange(UserMI, UserOffset, CPEs[i].CPEMI, U.MaxDisp, U.NegOk)) {
- DOUT << "Replacing CPE#" << CPI << " with CPE#" << CPEs[i].CPI << "\n";
+ 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.
default:
break;
}
-
+
return ((1<<23)-1)*4;
}
/// 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";
+ water_iterator IP) {
+ DEBUG(errs() << "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
/// group, prefer the water that's farthest away.
bool ARMConstantIslands::LookForWater(CPUser &U, unsigned UserOffset,
MachineBasicBlock** NewMBB) {
- std::vector<MachineBasicBlock*>::iterator IPThatWouldPad;
+ water_iterator IPThatWouldPad;
MachineBasicBlock* WaterBBThatWouldPad = NULL;
if (!WaterList.empty()) {
- for (std::vector<MachineBasicBlock*>::iterator IP = prior(WaterList.end()),
+ for (water_iterator IP = prior(WaterList.end()),
B = WaterList.begin();; --IP) {
MachineBasicBlock* WaterBB = *IP;
if (WaterIsInRange(UserOffset, WaterBB, U)) {
if (&UserMBB->back() == UserMI ||
OffsetIsInRange(UserOffset, OffsetOfNextBlock + (isThumb1 ? 2: 4),
U.MaxDisp, U.NegOk, U.IsSoImm)) {
- DOUT << "Split at end of block\n";
+ DEBUG(errs() << "Split at end of block\n");
if (&UserMBB->back() == UserMI)
assert(BBHasFallthrough(UserMBB) && "Expected a fallthrough BB!");
*NewMBB = next(MachineFunction::iterator(UserMBB));
CPUIndex++;
}
}
- DOUT << "Split in middle of big block\n";
+ DEBUG(errs() << "Split in middle of big block\n");
*NewMBB = SplitBlockBeforeInstr(prior(MI));
}
}
/// 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,
+bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &MF,
unsigned CPUserIndex) {
CPUser &U = CPUsers[CPUserIndex];
MachineInstr *UserMI = U.MI;
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);
// See if the current entry is within range, or there is a clone of it
if (!LookForWater(U, UserOffset, &NewMBB)) {
// No water found.
- DOUT << "No water found\n";
+ DEBUG(errs() << "No water found\n");
CreateNewWater(CPUserIndex, UserOffset, &NewMBB);
}
// Okay, we know we can put an island before NewMBB now, do it!
- MachineBasicBlock *NewIsland = Fn.CreateMachineBasicBlock();
- Fn.insert(NewMBB, NewIsland);
+ MachineBasicBlock *NewIsland = MF.CreateMachineBasicBlock();
+ MF.insert(NewMBB, NewIsland);
// Update internal data structures to account for the newly inserted MBB.
UpdateForInsertedWaterBlock(NewIsland);
break;
}
- DOUT << " Moved CPE to #" << ID << " CPI=" << CPI << "\t" << *UserMI;
+ DEBUG(errs() << " Moved CPE to #" << ID << " CPI=" << CPI
+ << '\t' << *UserMI);
return true;
}
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).getMBB();
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
/// 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 &Fn, ImmBranch &Br) {
+ARMConstantIslands::FixUpUnconditionalBr(MachineFunction &MF, ImmBranch &Br) {
MachineInstr *MI = Br.MI;
MachineBasicBlock *MBB = MI->getParent();
- assert(isThumb && !isThumb2 && "Expected a Thumb1 function!");
+ if (!isThumb1)
+ llvm_unreachable("FixUpUnconditionalBr is Thumb1 only!");
// Use BL to implement far jump.
Br.MaxDisp = (1 << 21) * 2;
HasFarJump = true;
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).getMBB();
// b L1
MachineBasicBlock *NewDest = BMI->getOperand(0).getMBB();
if (BBIsInRange(MI, NewDest, Br.MaxDisp)) {
- DOUT << " Invert Bcc condition and swap its destination with " << *BMI;
+ 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);
}
MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
- DOUT << " Insert B to BB#" << DestBB->getNumber()
- << " also invert condition and change dest. to BB#"
- << NextBB->getNumber() << "\n";
+ 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.
}
/// 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, the third is a zero since there
+ // is no writeback.
if (MI->getOpcode() == ARM::tPOP_RET &&
- MI->getOperand(0).getReg() == ARM::PC &&
- MI->getNumExplicitOperands() == 1) {
+ MI->getOperand(3).getReg() == ARM::PC &&
+ MI->getNumExplicitOperands() == 4) {
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)
+ continue;
+
+ 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;
+ }
+ }
+
+ 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?
+ const MachineJumpTableInfo *MJTI = MF.getJumpTableInfo();
+ 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();
+ MachineBasicBlock::iterator PrevI = MI;
+ if (PrevI == MBB->begin())
+ continue;
+
+ MachineInstr *AddrMI = --PrevI;
+ bool OptOk = true;
+ // Examine the instruction that calculate the jumptable entry address.
+ // If it's not the one just before the t2BR_JT, we won't delete it, then
+ // it's not worth doing the optimization.
+ 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;
+
+ // The previous 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;
+}