#include "ARMAddressingModes.h"
#include "ARMMachineFunctionInfo.h"
#include "ARMInstrInfo.h"
+#include "Thumb2InstrInfo.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"
/// HasInlineAsm - True if the function contains inline assembly.
bool HasInlineAsm;
- const TargetInstrInfo *TII;
+ const ARMInstrInfo *TII;
const ARMSubtarget *STI;
ARMFunctionInfo *AFI;
bool isThumb;
bool isThumb2;
public:
static char ID;
- ARMConstantIslands() : MachineFunctionPass(&ID) {}
+ ARMConstantIslands() : MachineFunctionPass(ID) {}
virtual bool runOnMachineFunction(MachineFunction &MF);
bool ARMConstantIslands::runOnMachineFunction(MachineFunction &MF) {
MachineConstantPool &MCP = *MF.getConstantPool();
- TII = MF.getTarget().getInstrInfo();
+ TII = (const ARMInstrInfo*)MF.getTarget().getInstrInfo();
AFI = MF.getInfo<ARMFunctionInfo>();
STI = &MF.getTarget().getSubtarget<ARMSubtarget>();
// constant pool users.
InitialFunctionScan(MF, CPEMIs);
CPEMIs.clear();
+ DEBUG(dumpBBs());
+
/// Remove dead constant pool entries.
RemoveUnusedCPEntries();
}
// Shrink 32-bit Thumb2 branch, load, and store instructions.
- if (isThumb2)
+ if (isThumb2 && !STI->prefers32BitThumb())
MadeChange |= OptimizeThumb2Instructions(MF);
// After a while, this might be made debug-only, but it is not expensive.
// be right, functions containing these must be 4-byte aligned.
// 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.
+ // 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+2)%4 != 0 || HasInlineAsm)
+ if ((Offset+MBBSize)%4 != 0 || HasInlineAsm)
// FIXME: Add a pseudo ALIGN instruction instead.
MBBSize += 2; // padding
continue; // Does not get an entry in ImmBranches
Scale = 4;
break;
- case ARM::LDR:
+ case ARM::LDRi12:
case ARM::LDRcp:
- case ARM::t2LDRpci:
+ case ARM::t2LDRi12:
+ case ARM::t2LDRHi12:
+ case ARM::t2LDRBi12:
+ case ARM::t2LDRSHi12:
+ case ARM::t2LDRSBi12:
Bits = 12; // +-offset_12
NegOk = true;
break;
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 += TII->GetInstSizeInBytes(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.
int delta = isThumb1 ? 2 : 4;
- BBSizes[OrigBBI] -= NewBBSize - delta;
+
+ // 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;
}
// 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 + 2
- // is aligned.
+ // 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) + 2;
+ unsigned NewMIOffset = GetOffsetOf(ThumbJTMI);
unsigned OldMIOffset = NewMIOffset - delta;
if ((OldMIOffset%4) == 0 && (NewMIOffset%4) != 0) {
// remove existing padding
MachineBasicBlock::iterator MI = UserMI;
++MI;
unsigned CPUIndex = CPUserIndex+1;
+ unsigned NumCPUsers = CPUsers.size();
+ MachineInstr *LastIT = 0;
for (unsigned Offset = UserOffset+TII->GetInstSizeInBytes(UserMI);
Offset < BaseInsertOffset;
Offset += TII->GetInstSizeInBytes(MI),
- MI = llvm::next(MI)) {
- if (CPUIndex < CPUsers.size() && CPUsers[CPUIndex].MI == MI) {
+ MI = llvm::next(MI)) {
+ if (CPUIndex < NumCPUsers && CPUsers[CPUIndex].MI == MI) {
CPUser &U = CPUsers[CPUIndex];
if (!OffsetIsInRange(Offset, EndInsertOffset,
U.MaxDisp, U.NegOk, U.IsSoImm)) {
EndInsertOffset += CPUsers[CPUIndex].CPEMI->getOperand(2).getImm();
CPUIndex++;
}
+
+ // Remember the last IT instruction.
+ if (MI->getOpcode() == ARM::t2IT)
+ LastIT = MI;
}
+
DEBUG(errs() << "Split in middle of big block\n");
- NewMBB = SplitBlockBeforeInstr(prior(MI));
+ --MI;
+
+ // Avoid splitting an IT block.
+ if (LastIT) {
+ unsigned PredReg = 0;
+ ARMCC::CondCodes CC = llvm::getITInstrPredicate(MI, PredReg);
+ if (CC != ARMCC::AL)
+ MI = LastIT;
+ }
+ NewMBB = SplitBlockBeforeInstr(MI);
}
}
unsigned DestOffset = BBOffsets[DestBB->getNumber()];
if (BrOffset < DestOffset && (DestOffset - BrOffset) <= 126) {
MachineBasicBlock::iterator CmpMI = Br.MI; --CmpMI;
- if (CmpMI->getOpcode() == ARM::tCMPzi8) {
+ if (CmpMI->getOpcode() == ARM::tCMPi8) {
unsigned Reg = CmpMI->getOperand(0).getReg();
Pred = llvm::getInstrPredicate(CmpMI, PredReg);
if (Pred == ARMCC::AL &&
if (!OptOk)
continue;
- unsigned Opc = ByteOk ? ARM::t2TBB : ARM::t2TBH;
+ unsigned Opc = ByteOk ? ARM::t2TBB_JT : ARM::t2TBH_JT;
MachineInstr *NewJTMI = BuildMI(MBB, MI->getDebugLoc(), TII->get(Opc))
.addReg(IdxReg, getKillRegState(IdxRegKill))
.addJumpTableIndex(JTI, JTOP.getTargetFlags())
projects / oota-llvm.git / blobdiff