#define DEBUG_TYPE "t2-reduce-size"
#include "ARM.h"
-#include "ARMAddressingModes.h"
-#include "ARMBaseRegisterInfo.h"
#include "ARMBaseInstrInfo.h"
+#include "ARMBaseRegisterInfo.h"
#include "ARMSubtarget.h"
+#include "MCTargetDesc/ARMAddressingModes.h"
#include "Thumb2InstrInfo.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
-#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/Statistic.h"
using namespace llvm;
STATISTIC(NumNarrows, "Number of 32-bit instrs reduced to 16-bit ones");
/// ReduceTable - A static table with information on mapping from wide
/// opcodes to narrow
struct ReduceEntry {
- unsigned WideOpc; // Wide opcode
- unsigned NarrowOpc1; // Narrow opcode to transform to
- unsigned NarrowOpc2; // Narrow opcode when it's two-address
+ uint16_t WideOpc; // Wide opcode
+ uint16_t NarrowOpc1; // Narrow opcode to transform to
+ uint16_t NarrowOpc2; // Narrow opcode when it's two-address
uint8_t Imm1Limit; // Limit of immediate field (bits)
uint8_t Imm2Limit; // Limit of immediate field when it's two-address
unsigned LowRegs1 : 1; // Only possible if low-registers are used
{ ARM::t2BICrr, 0, ARM::tBIC, 0, 0, 0, 1, 0,0, 1,0 },
//FIXME: Disable CMN, as CCodes are backwards from compare expectations
//{ ARM::t2CMNrr, ARM::tCMN, 0, 0, 0, 1, 0, 2,0, 0,0 },
+ { ARM::t2CMNzrr, ARM::tCMNz, 0, 0, 0, 1, 0, 2,0, 0,0 },
{ ARM::t2CMPri, ARM::tCMPi8, 0, 8, 0, 1, 0, 2,0, 0,0 },
{ ARM::t2CMPrr, ARM::tCMPhir, 0, 0, 0, 0, 0, 2,0, 0,1 },
{ ARM::t2EORrr, 0, ARM::tEOR, 0, 0, 0, 1, 0,0, 1,0 },
{ ARM::t2SUBrr, ARM::tSUBrr, 0, 0, 0, 1, 0, 0,0, 0,0 },
{ ARM::t2SUBSri,ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 2,2, 0,0 },
{ ARM::t2SUBSrr,ARM::tSUBrr, 0, 0, 0, 1, 0, 2,0, 0,0 },
- { ARM::t2SXTBr, ARM::tSXTB, 0, 0, 0, 1, 0, 1,0, 0,0 },
- { ARM::t2SXTHr, ARM::tSXTH, 0, 0, 0, 1, 0, 1,0, 0,0 },
+ { ARM::t2SXTB, ARM::tSXTB, 0, 0, 0, 1, 0, 1,0, 0,1 },
+ { ARM::t2SXTH, ARM::tSXTH, 0, 0, 0, 1, 0, 1,0, 0,1 },
{ ARM::t2TSTrr, ARM::tTST, 0, 0, 0, 1, 0, 2,0, 0,0 },
- { ARM::t2UXTBr, ARM::tUXTB, 0, 0, 0, 1, 0, 1,0, 0,0 },
- { ARM::t2UXTHr, ARM::tUXTH, 0, 0, 0, 1, 0, 1,0, 0,0 },
+ { ARM::t2UXTB, ARM::tUXTB, 0, 0, 0, 1, 0, 1,0, 0,1 },
+ { ARM::t2UXTH, ARM::tUXTH, 0, 0, 0, 1, 0, 1,0, 0,1 },
// FIXME: Clean this up after splitting each Thumb load / store opcode
// into multiple ones.
/// ReduceOpcodeMap - Maps wide opcode to index of entry in ReduceTable.
DenseMap<unsigned, unsigned> ReduceOpcodeMap;
- bool canAddPseudoFlagDep(MachineInstr *Def, MachineInstr *Use);
+ bool canAddPseudoFlagDep(MachineInstr *Def, MachineInstr *Use,
+ bool IsSelfLoop);
bool VerifyPredAndCC(MachineInstr *MI, const ReduceEntry &Entry,
bool is2Addr, ARMCC::CondCodes Pred,
bool ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI,
const ReduceEntry &Entry, bool LiveCPSR,
- MachineInstr *CPSRDef);
+ MachineInstr *CPSRDef, bool IsSelfLoop);
/// ReduceTo2Addr - Reduce a 32-bit instruction to a 16-bit two-address
/// instruction.
bool ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI,
const ReduceEntry &Entry,
- bool LiveCPSR, MachineInstr *CPSRDef);
+ bool LiveCPSR, MachineInstr *CPSRDef,
+ bool IsSelfLoop);
/// ReduceToNarrow - Reduce a 32-bit instruction to a 16-bit
/// non-two-address instruction.
bool ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI,
const ReduceEntry &Entry,
- bool LiveCPSR, MachineInstr *CPSRDef);
+ bool LiveCPSR, MachineInstr *CPSRDef,
+ bool IsSelfLoop);
/// ReduceMBB - Reduce width of instructions in the specified basic block.
bool ReduceMBB(MachineBasicBlock &MBB);
}
static bool HasImplicitCPSRDef(const MCInstrDesc &MCID) {
- for (const unsigned *Regs = MCID.ImplicitDefs; *Regs; ++Regs)
+ for (const uint16_t *Regs = MCID.getImplicitDefs(); *Regs; ++Regs)
if (*Regs == ARM::CPSR)
return true;
return false;
/// In this case it would have been ok to narrow the mul.w to muls since there
/// are indirect RAW dependency between the muls and the mul.w
bool
-Thumb2SizeReduce::canAddPseudoFlagDep(MachineInstr *Def, MachineInstr *Use) {
- if (!Def || !STI->avoidCPSRPartialUpdate())
+Thumb2SizeReduce::canAddPseudoFlagDep(MachineInstr *Def, MachineInstr *Use,
+ bool FirstInSelfLoop) {
+ // FIXME: Disable check for -Oz (aka OptimizeForSizeHarder).
+ if (!STI->avoidCPSRPartialUpdate())
return false;
+ if (!Def)
+ // If this BB loops back to itself, conservatively avoid narrowing the
+ // first instruction that does partial flag update.
+ return FirstInSelfLoop;
+
SmallSet<unsigned, 2> Defs;
for (unsigned i = 0, e = Def->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = Def->getOperand(i);
// Add the 16-bit load / store instruction.
DebugLoc dl = MI->getDebugLoc();
- MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, TII->get(Opc));
+ MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, TII->get(Opc));
if (!isLdStMul) {
MIB.addOperand(MI->getOperand(0));
MIB.addOperand(MI->getOperand(1));
DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
- MBB.erase(MI);
+ MBB.erase_instr(MI);
++NumLdSts;
return true;
}
bool
Thumb2SizeReduce::ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI,
const ReduceEntry &Entry,
- bool LiveCPSR, MachineInstr *CPSRDef) {
+ bool LiveCPSR, MachineInstr *CPSRDef,
+ bool IsSelfLoop) {
unsigned Opc = MI->getOpcode();
if (Opc == ARM::t2ADDri) {
// If the source register is SP, try to reduce to tADDrSPi, otherwise
// it's a normal reduce.
if (MI->getOperand(1).getReg() != ARM::SP) {
- if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, CPSRDef))
+ if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, CPSRDef, IsSelfLoop))
return true;
- return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef);
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef, IsSelfLoop);
}
// Try to reduce to tADDrSPi.
unsigned Imm = MI->getOperand(2).getImm();
MI->getOperand(MCID.getNumOperands()-1).getReg() == ARM::CPSR)
return false;
- MachineInstrBuilder MIB = BuildMI(MBB, *MI, MI->getDebugLoc(),
+ MachineInstrBuilder MIB = BuildMI(MBB, MI, MI->getDebugLoc(),
TII->get(ARM::tADDrSPi))
.addOperand(MI->getOperand(0))
.addOperand(MI->getOperand(1))
.addImm(Imm / 4); // The tADDrSPi has an implied scale by four.
+ AddDefaultPred(MIB);
// Transfer MI flags.
MIB.setMIFlags(MI->getFlags());
DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " <<*MIB);
- MBB.erase(MI);
+ MBB.erase_instr(MI);
++NumNarrows;
return true;
}
if (Entry.LowRegs1 && !VerifyLowRegs(MI))
return false;
- const MCInstrDesc &MCID = MI->getDesc();
- if (MCID.mayLoad() || MCID.mayStore())
+ if (MI->mayLoad() || MI->mayStore())
return ReduceLoadStore(MBB, MI, Entry);
switch (Opc) {
switch (Opc) {
default: break;
case ARM::t2ADDSri: {
- if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, CPSRDef))
+ if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, CPSRDef, IsSelfLoop))
return true;
// fallthrough
}
case ARM::t2ADDSrr:
- return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef);
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef, IsSelfLoop);
}
}
break;
}
case ARM::t2RSBri:
case ARM::t2RSBSri:
+ case ARM::t2SXTB:
+ case ARM::t2SXTH:
+ case ARM::t2UXTB:
+ case ARM::t2UXTH:
if (MI->getOperand(2).getImm() == 0)
- return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef);
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef, IsSelfLoop);
break;
case ARM::t2MOVi16:
// Can convert only 'pure' immediate operands, not immediates obtained as
// globals' addresses.
if (MI->getOperand(1).isImm())
- return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef);
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef, IsSelfLoop);
break;
case ARM::t2CMPrr: {
// Try to reduce to the lo-reg only version first. Why there are two
// source insn opcode. So for now, we hack a local entry record to use.
static const ReduceEntry NarrowEntry =
{ ARM::t2CMPrr,ARM::tCMPr, 0, 0, 0, 1, 1,2, 0, 0,1 };
- if (ReduceToNarrow(MBB, MI, NarrowEntry, LiveCPSR, CPSRDef))
+ if (ReduceToNarrow(MBB, MI, NarrowEntry, LiveCPSR, CPSRDef, IsSelfLoop))
return true;
- return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef);
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef, IsSelfLoop);
}
}
return false;
bool
Thumb2SizeReduce::ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI,
const ReduceEntry &Entry,
- bool LiveCPSR, MachineInstr *CPSRDef) {
+ bool LiveCPSR, MachineInstr *CPSRDef,
+ bool IsSelfLoop) {
if (ReduceLimit2Addr != -1 && ((int)Num2Addrs >= ReduceLimit2Addr))
return false;
unsigned Reg0 = MI->getOperand(0).getReg();
unsigned Reg1 = MI->getOperand(1).getReg();
- if (Reg0 != Reg1) {
+ // t2MUL is "special". The tied source operand is second, not first.
+ if (MI->getOpcode() == ARM::t2MUL) {
+ unsigned Reg2 = MI->getOperand(2).getReg();
+ // Early exit if the regs aren't all low regs.
+ if (!isARMLowRegister(Reg0) || !isARMLowRegister(Reg1)
+ || !isARMLowRegister(Reg2))
+ return false;
+ if (Reg0 != Reg2) {
+ // If the other operand also isn't the same as the destination, we
+ // can't reduce.
+ if (Reg1 != Reg0)
+ return false;
+ // Try to commute the operands to make it a 2-address instruction.
+ MachineInstr *CommutedMI = TII->commuteInstruction(MI);
+ if (!CommutedMI)
+ return false;
+ }
+ } else if (Reg0 != Reg1) {
// Try to commute the operands to make it a 2-address instruction.
unsigned CommOpIdx1, CommOpIdx2;
if (!TII->findCommutedOpIndices(MI, CommOpIdx1, CommOpIdx2) ||
// Avoid adding a false dependency on partial flag update by some 16-bit
// instructions which has the 's' bit set.
if (Entry.PartFlag && NewMCID.hasOptionalDef() && HasCC &&
- canAddPseudoFlagDep(CPSRDef, MI))
+ canAddPseudoFlagDep(CPSRDef, MI, IsSelfLoop))
return false;
// Add the 16-bit instruction.
DebugLoc dl = MI->getDebugLoc();
- MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, NewMCID);
+ MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, NewMCID);
MIB.addOperand(MI->getOperand(0));
if (NewMCID.hasOptionalDef()) {
if (HasCC)
DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
- MBB.erase(MI);
+ MBB.erase_instr(MI);
++Num2Addrs;
return true;
}
bool
Thumb2SizeReduce::ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI,
const ReduceEntry &Entry,
- bool LiveCPSR, MachineInstr *CPSRDef) {
+ bool LiveCPSR, MachineInstr *CPSRDef,
+ bool IsSelfLoop) {
if (ReduceLimit != -1 && ((int)NumNarrows >= ReduceLimit))
return false;
// Avoid adding a false dependency on partial flag update by some 16-bit
// instructions which has the 's' bit set.
if (Entry.PartFlag && NewMCID.hasOptionalDef() && HasCC &&
- canAddPseudoFlagDep(CPSRDef, MI))
+ canAddPseudoFlagDep(CPSRDef, MI, IsSelfLoop))
return false;
// Add the 16-bit instruction.
DebugLoc dl = MI->getDebugLoc();
- MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, NewMCID);
+ MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, NewMCID);
MIB.addOperand(MI->getOperand(0));
if (NewMCID.hasOptionalDef()) {
if (HasCC)
if (i < NumOps && MCID.OpInfo[i].isOptionalDef())
continue;
if ((MCID.getOpcode() == ARM::t2RSBSri ||
- MCID.getOpcode() == ARM::t2RSBri) && i == 2)
+ MCID.getOpcode() == ARM::t2RSBri ||
+ MCID.getOpcode() == ARM::t2SXTB ||
+ MCID.getOpcode() == ARM::t2SXTH ||
+ MCID.getOpcode() == ARM::t2UXTB ||
+ MCID.getOpcode() == ARM::t2UXTH) && i == 2)
// Skip the zero immediate operand, it's now implicit.
continue;
bool isPred = (i < NumOps && MCID.OpInfo[i].isPredicate());
DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
- MBB.erase(MI);
+ MBB.erase_instr(MI);
++NumNarrows;
return true;
}
// Yes, CPSR could be livein.
bool LiveCPSR = MBB.isLiveIn(ARM::CPSR);
MachineInstr *CPSRDef = 0;
+ MachineInstr *BundleMI = 0;
- MachineBasicBlock::iterator MII = MBB.begin(), E = MBB.end();
- MachineBasicBlock::iterator NextMII;
+ // If this BB loops back to itself, conservatively avoid narrowing the
+ // first instruction that does partial flag update.
+ bool IsSelfLoop = MBB.isSuccessor(&MBB);
+ MachineBasicBlock::instr_iterator MII = MBB.instr_begin(),E = MBB.instr_end();
+ MachineBasicBlock::instr_iterator NextMII;
for (; MII != E; MII = NextMII) {
NextMII = llvm::next(MII);
MachineInstr *MI = &*MII;
+ if (MI->isBundle()) {
+ BundleMI = MI;
+ continue;
+ }
+
LiveCPSR = UpdateCPSRUse(*MI, LiveCPSR);
unsigned Opcode = MI->getOpcode();
const ReduceEntry &Entry = ReduceTable[OPI->second];
// Ignore "special" cases for now.
if (Entry.Special) {
- if (ReduceSpecial(MBB, MI, Entry, LiveCPSR, CPSRDef)) {
+ if (ReduceSpecial(MBB, MI, Entry, LiveCPSR, CPSRDef, IsSelfLoop)) {
Modified = true;
- MachineBasicBlock::iterator I = prior(NextMII);
+ MachineBasicBlock::instr_iterator I = prior(NextMII);
MI = &*I;
}
goto ProcessNext;
// Try to transform to a 16-bit two-address instruction.
if (Entry.NarrowOpc2 &&
- ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, CPSRDef)) {
+ ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, CPSRDef, IsSelfLoop)) {
Modified = true;
- MachineBasicBlock::iterator I = prior(NextMII);
+ MachineBasicBlock::instr_iterator I = prior(NextMII);
MI = &*I;
goto ProcessNext;
}
// Try to transform to a 16-bit non-two-address instruction.
if (Entry.NarrowOpc1 &&
- ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef)) {
+ ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef, IsSelfLoop)) {
Modified = true;
- MachineBasicBlock::iterator I = prior(NextMII);
+ MachineBasicBlock::instr_iterator I = prior(NextMII);
MI = &*I;
}
}
ProcessNext:
+ if (NextMII != E && MI->isInsideBundle() && !NextMII->isInsideBundle()) {
+ // FIXME: Since post-ra scheduler operates on bundles, the CPSR kill
+ // marker is only on the BUNDLE instruction. Process the BUNDLE
+ // instruction as we finish with the bundled instruction to work around
+ // the inconsistency.
+ if (BundleMI->killsRegister(ARM::CPSR))
+ LiveCPSR = false;
+ MachineOperand *MO = BundleMI->findRegisterDefOperand(ARM::CPSR);
+ if (MO && !MO->isDead())
+ LiveCPSR = true;
+ }
+
bool DefCPSR = false;
LiveCPSR = UpdateCPSRDef(*MI, LiveCPSR, DefCPSR);
- if (MI->getDesc().isCall())
+ if (MI->isCall()) {
// Calls don't really set CPSR.
CPSRDef = 0;
- else if (DefCPSR)
+ IsSelfLoop = false;
+ } else if (DefCPSR) {
// This is the last CPSR defining instruction.
CPSRDef = MI;
+ IsSelfLoop = false;
+ }
}
return Modified;