#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 <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(NumTBs, "Number of table branches generated");
+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");
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 {
+ 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
/// 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;
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) {}
+ : 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
bool HasFarJump;
const TargetInstrInfo *TII;
+ const ARMSubtarget *STI;
ARMFunctionInfo *AFI;
bool isThumb;
bool isThumb1;
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);
+ MachineBasicBlock *&NewMBB);
bool HandleConstantPoolUser(MachineFunction &MF, unsigned CPUserIndex);
void RemoveDeadCPEMI(MachineInstr *CPEMI);
bool RemoveUnusedCPEntries();
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;
/// 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");
}
}
TII = MF.getTarget().getInstrInfo();
AFI = MF.getInfo<ARMFunctionInfo>();
+ STI = &MF.getTarget().getSubtarget<ARMSubtarget>();
+
isThumb = AFI->isThumbFunction();
isThumb1 = AFI->isThumb1OnlyFunction();
isThumb2 = AFI->isThumb2Function();
// the numbers agree with the position of the block in the function.
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);
// 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(MF, 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(MF, 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(MF);
if (isThumb && !HasFarJump && AFI->isLRSpilledForFarJump())
MadeChange |= UndoLRSpillRestore();
- // Let's see if we can use tbb / tbh to do jump tables.
- MadeChange |= OptimizeThumb2JumpTables(MF);
-
BBSizes.clear();
BBOffsets.clear();
WaterList.clear();
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");
}
}
// 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();
// 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;
WaterList.insert(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.)
// 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.
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();
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";
- // 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)) {
- 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 (!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;
/// 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) {
+ unsigned UserOffset,
+ MachineBasicBlock *&NewMBB) {
CPUser &U = CPUsers[CPUserIndex];
MachineInstr *UserMI = U.MI;
MachineInstr *CPEMI = U.CPEMI;
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 + Thumb2,
- // 2 on Thumb1. Possible Thumb1 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 ||
+ if (BBHasFallthrough(UserMBB) &&
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));
+ NewMBB = next(MachineFunction::iterator(UserMBB));
// Add an unconditional branch from UserMBB to fallthrough block.
// Record it for branch lengthening; this new branch will not get out of
// range, but if the preceding conditional branch is out of range, the
// range, so the machinery has to know about it.
int UncondBr = isThumb ? ((isThumb2) ? ARM::t2B : ARM::tB) : ARM::B;
BuildMI(UserMBB, DebugLoc::getUnknownLoc(),
- TII->get(UncondBr)).addMBB(*NewMBB);
+ TII->get(UncondBr)).addMBB(NewMBB);
unsigned MaxDisp = getUnconditionalBrDisp(UncondBr);
ImmBranches.push_back(ImmBranch(&UserMBB->back(),
MaxDisp, false, UncondBr));
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));
}
}
MachineInstr *CPEMI = U.CPEMI;
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);
// See if the current entry is within range, or there is a clone of it
// We will be generating a new clone. Get a UID for it.
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 = 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 = MF.CreateMachineBasicBlock();
MF.insert(NewMBB, NewIsland);
// Update internal data structures to account for the newly inserted MBB.
// Now that we have an island to add the CPE to, clone the original CPE and
// add it to the island.
+ U.HighWaterMark = NewIsland;
U.CPEMI = BuildMI(NewIsland, DebugLoc::getUnknownLoc(),
TII->get(ARM::CONSTPOOL_ENTRY))
.addImm(ID).addConstantPoolIndex(CPI).addImm(Size);
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.
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;
}
// 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) {
+ 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;
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 && !OffsetIsInRange(JTOffset, DstOffset, (1<<8)-1, false))
+ if (ByteOk && (DstOffset - JTOffset) > ((1<<8)-1)*2)
ByteOk = false;
- if (HalfWordOk &&
- !OffsetIsInRange(JTOffset, DstOffset, (1<<16)-1, false))
+ unsigned TBHLimit = ((1<<16)-1)*2;
+ if (HalfWordOk && (DstOffset - JTOffset) > TBHLimit)
HalfWordOk = false;
if (!ByteOk && !HalfWordOk)
break;
if (!OptOk)
continue;
- // The previous instruction should be a t2LEApcrelJT, we want to delete
- // it as well.
+ // The previous instruction should be a tLEApcrel or t2LEApcrelJT, we want
+ // to delete it as well.
MachineInstr *LeaMI = --PrevI;
- if (LeaMI->getOpcode() != ARM::t2LEApcrelJT ||
+ if ((LeaMI->getOpcode() != ARM::tLEApcrelJT &&
+ LeaMI->getOpcode() != ARM::t2LEApcrelJT) ||
LeaMI->getOperand(0).getReg() != BaseReg)
- LeaMI = 0;
-
- if (OptOk) {
- unsigned Opc = ByteOk ? ARM::t2TBB : ARM::t2TBH;
- AddDefaultPred(BuildMI(MBB, MI->getDebugLoc(), TII->get(Opc))
- .addReg(IdxReg, getKillRegState(IdxRegKill))
- .addJumpTableIndex(JTI, JTOP.getTargetFlags())
- .addImm(MI->getOperand(JTOpIdx+1).getImm()));
-
- AddrMI->eraseFromParent();
- if (LeaMI)
- LeaMI->eraseFromParent();
- MI->eraseFromParent();
- ++NumTBs;
- MadeChange = true;
- }
+ 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;
}
}