-//===-- ARMLoadStoreOptimizer.cpp - ARM load / store opt. pass ----*- C++ -*-=//
+//===-- ARMLoadStoreOptimizer.cpp - ARM load / store opt. pass ------------===//
//
// The LLVM Compiler Infrastructure
//
#define DEBUG_TYPE "arm-ldst-opt"
#include "ARM.h"
#include "ARMBaseInstrInfo.h"
+#include "ARMBaseRegisterInfo.h"
#include "ARMMachineFunctionInfo.h"
-#include "ARMRegisterInfo.h"
#include "MCTargetDesc/ARMAddressingModes.h"
-#include "llvm/DerivedTypes.h"
-#include "llvm/Function.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/RegisterScavenging.h"
#include "llvm/CodeGen/SelectionDAGNodes.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/DataLayout.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Function.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetRegisterInfo.h"
-#include "llvm/Support/ErrorHandling.h"
-#include "llvm/Support/Debug.h"
-#include "llvm/Support/raw_ostream.h"
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/STLExtras.h"
-#include "llvm/ADT/SmallPtrSet.h"
-#include "llvm/ADT/SmallSet.h"
-#include "llvm/ADT/SmallVector.h"
-#include "llvm/ADT/Statistic.h"
using namespace llvm;
STATISTIC(NumLDMGened , "Number of ldm instructions generated");
bool MergeOps(MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
int Offset, unsigned Base, bool BaseKill, int Opcode,
ARMCC::CondCodes Pred, unsigned PredReg, unsigned Scratch,
- DebugLoc dl, SmallVector<std::pair<unsigned, bool>, 8> &Regs);
+ DebugLoc dl,
+ ArrayRef<std::pair<unsigned, bool> > Regs,
+ ArrayRef<unsigned> ImpDefs);
void MergeOpsUpdate(MachineBasicBlock &MBB,
MemOpQueue &MemOps,
unsigned memOpsBegin,
case ARM_AM::db: return ARM::LDMDB;
case ARM_AM::ib: return ARM::LDMIB;
}
- break;
case ARM::STRi12:
++NumSTMGened;
switch (Mode) {
case ARM_AM::db: return ARM::STMDB;
case ARM_AM::ib: return ARM::STMIB;
}
- break;
case ARM::t2LDRi8:
case ARM::t2LDRi12:
++NumLDMGened;
case ARM_AM::ia: return ARM::t2LDMIA;
case ARM_AM::db: return ARM::t2LDMDB;
}
- break;
case ARM::t2STRi8:
case ARM::t2STRi12:
++NumSTMGened;
case ARM_AM::ia: return ARM::t2STMIA;
case ARM_AM::db: return ARM::t2STMDB;
}
- break;
case ARM::VLDRS:
++NumVLDMGened;
switch (Mode) {
case ARM_AM::ia: return ARM::VLDMSIA;
case ARM_AM::db: return 0; // Only VLDMSDB_UPD exists.
}
- break;
case ARM::VSTRS:
++NumVSTMGened;
switch (Mode) {
case ARM_AM::ia: return ARM::VSTMSIA;
case ARM_AM::db: return 0; // Only VSTMSDB_UPD exists.
}
- break;
case ARM::VLDRD:
++NumVLDMGened;
switch (Mode) {
case ARM_AM::ia: return ARM::VLDMDIA;
case ARM_AM::db: return 0; // Only VLDMDDB_UPD exists.
}
- break;
case ARM::VSTRD:
++NumVSTMGened;
switch (Mode) {
case ARM_AM::ia: return ARM::VSTMDIA;
case ARM_AM::db: return 0; // Only VSTMDDB_UPD exists.
}
- break;
}
-
- return 0;
}
namespace llvm {
case ARM::STMIB_UPD:
return ARM_AM::ib;
}
-
- return ARM_AM::bad_am_submode;
}
} // end namespace ARM_AM
int Offset, unsigned Base, bool BaseKill,
int Opcode, ARMCC::CondCodes Pred,
unsigned PredReg, unsigned Scratch, DebugLoc dl,
- SmallVector<std::pair<unsigned, bool>, 8> &Regs) {
+ ArrayRef<std::pair<unsigned, bool> > Regs,
+ ArrayRef<unsigned> ImpDefs) {
// Only a single register to load / store. Don't bother.
unsigned NumRegs = Regs.size();
if (NumRegs <= 1)
MIB = MIB.addReg(Regs[i].first, getDefRegState(isDef)
| getKillRegState(Regs[i].second));
+ // Add implicit defs for super-registers.
+ for (unsigned i = 0, e = ImpDefs.size(); i != e; ++i)
+ MIB.addReg(ImpDefs[i], RegState::ImplicitDefine);
+
return true;
}
}
SmallVector<std::pair<unsigned, bool>, 8> Regs;
+ SmallVector<unsigned, 8> ImpDefs;
for (unsigned i = memOpsBegin; i < memOpsEnd; ++i) {
unsigned Reg = memOps[i].Reg;
// If we are inserting the merged operation after an operation that
// uses the same register, make sure to transfer any kill flag.
bool isKill = memOps[i].isKill || KilledRegs.count(Reg);
Regs.push_back(std::make_pair(Reg, isKill));
+
+ // Collect any implicit defs of super-registers. They must be preserved.
+ for (MIOperands MO(memOps[i].MBBI); MO.isValid(); ++MO) {
+ if (!MO->isReg() || !MO->isDef() || !MO->isImplicit() || MO->isDead())
+ continue;
+ unsigned DefReg = MO->getReg();
+ if (std::find(ImpDefs.begin(), ImpDefs.end(), DefReg) == ImpDefs.end())
+ ImpDefs.push_back(DefReg);
+ }
}
// Try to do the merge.
MachineBasicBlock::iterator Loc = memOps[insertAfter].MBBI;
++Loc;
if (!MergeOps(MBB, Loc, Offset, Base, BaseKill, Opcode,
- Pred, PredReg, Scratch, dl, Regs))
+ Pred, PredReg, Scratch, dl, Regs, ImpDefs))
return;
// Merge succeeded, update records.
DebugLoc dl = Loc->getDebugLoc();
const MachineOperand &PMO = Loc->getOperand(0);
unsigned PReg = PMO.getReg();
- unsigned PRegNum = PMO.isUndef() ? UINT_MAX
- : getARMRegisterNumbering(PReg);
+ unsigned PRegNum = PMO.isUndef() ? UINT_MAX : TRI->getEncodingValue(PReg);
unsigned Count = 1;
unsigned Limit = ~0U;
int NewOffset = MemOps[i].Offset;
const MachineOperand &MO = MemOps[i].MBBI->getOperand(0);
unsigned Reg = MO.getReg();
- unsigned RegNum = MO.isUndef() ? UINT_MAX
- : getARMRegisterNumbering(Reg);
+ unsigned RegNum = MO.isUndef() ? UINT_MAX : TRI->getEncodingValue(Reg);
// Register numbers must be in ascending order. For VFP / NEON load and
// store multiples, the registers must also be consecutive and within the
// limit on the number of registers per instruction.
return;
}
-static inline bool isMatchingDecrement(MachineInstr *MI, unsigned Base,
- unsigned Bytes, unsigned Limit,
- ARMCC::CondCodes Pred, unsigned PredReg){
+static bool definesCPSR(MachineInstr *MI) {
+ for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = MI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ if (MO.isDef() && MO.getReg() == ARM::CPSR && !MO.isDead())
+ // If the instruction has live CPSR def, then it's not safe to fold it
+ // into load / store.
+ return true;
+ }
+
+ return false;
+}
+
+static bool isMatchingDecrement(MachineInstr *MI, unsigned Base,
+ unsigned Bytes, unsigned Limit,
+ ARMCC::CondCodes Pred, unsigned PredReg) {
unsigned MyPredReg = 0;
if (!MI)
return false;
- if (MI->getOpcode() != ARM::t2SUBri &&
- MI->getOpcode() != ARM::tSUBspi &&
- MI->getOpcode() != ARM::SUBri)
- return false;
+
+ bool CheckCPSRDef = false;
+ switch (MI->getOpcode()) {
+ default: return false;
+ case ARM::t2SUBri:
+ case ARM::SUBri:
+ CheckCPSRDef = true;
+ // fallthrough
+ case ARM::tSUBspi:
+ break;
+ }
// Make sure the offset fits in 8 bits.
if (Bytes == 0 || (Limit && Bytes >= Limit))
return false;
unsigned Scale = (MI->getOpcode() == ARM::tSUBspi) ? 4 : 1; // FIXME
- return (MI->getOperand(0).getReg() == Base &&
- MI->getOperand(1).getReg() == Base &&
- (MI->getOperand(2).getImm()*Scale) == Bytes &&
- llvm::getInstrPredicate(MI, MyPredReg) == Pred &&
- MyPredReg == PredReg);
+ if (!(MI->getOperand(0).getReg() == Base &&
+ MI->getOperand(1).getReg() == Base &&
+ (MI->getOperand(2).getImm()*Scale) == Bytes &&
+ getInstrPredicate(MI, MyPredReg) == Pred &&
+ MyPredReg == PredReg))
+ return false;
+
+ return CheckCPSRDef ? !definesCPSR(MI) : true;
}
-static inline bool isMatchingIncrement(MachineInstr *MI, unsigned Base,
- unsigned Bytes, unsigned Limit,
- ARMCC::CondCodes Pred, unsigned PredReg){
+static bool isMatchingIncrement(MachineInstr *MI, unsigned Base,
+ unsigned Bytes, unsigned Limit,
+ ARMCC::CondCodes Pred, unsigned PredReg) {
unsigned MyPredReg = 0;
if (!MI)
return false;
- if (MI->getOpcode() != ARM::t2ADDri &&
- MI->getOpcode() != ARM::tADDspi &&
- MI->getOpcode() != ARM::ADDri)
- return false;
+
+ bool CheckCPSRDef = false;
+ switch (MI->getOpcode()) {
+ default: return false;
+ case ARM::t2ADDri:
+ case ARM::ADDri:
+ CheckCPSRDef = true;
+ // fallthrough
+ case ARM::tADDspi:
+ break;
+ }
if (Bytes == 0 || (Limit && Bytes >= Limit))
// Make sure the offset fits in 8 bits.
return false;
unsigned Scale = (MI->getOpcode() == ARM::tADDspi) ? 4 : 1; // FIXME
- return (MI->getOperand(0).getReg() == Base &&
- MI->getOperand(1).getReg() == Base &&
- (MI->getOperand(2).getImm()*Scale) == Bytes &&
- llvm::getInstrPredicate(MI, MyPredReg) == Pred &&
- MyPredReg == PredReg);
+ if (!(MI->getOperand(0).getReg() == Base &&
+ MI->getOperand(1).getReg() == Base &&
+ (MI->getOperand(2).getImm()*Scale) == Bytes &&
+ getInstrPredicate(MI, MyPredReg) == Pred &&
+ MyPredReg == PredReg))
+ return false;
+
+ return CheckCPSRDef ? !definesCPSR(MI) : true;
}
static inline unsigned getLSMultipleTransferSize(MachineInstr *MI) {
case ARM_AM::da: return ARM::LDMDA_UPD;
case ARM_AM::db: return ARM::LDMDB_UPD;
}
- break;
case ARM::STMIA:
case ARM::STMDA:
case ARM::STMDB:
case ARM_AM::da: return ARM::STMDA_UPD;
case ARM_AM::db: return ARM::STMDB_UPD;
}
- break;
case ARM::t2LDMIA:
case ARM::t2LDMDB:
switch (Mode) {
case ARM_AM::ia: return ARM::t2LDMIA_UPD;
case ARM_AM::db: return ARM::t2LDMDB_UPD;
}
- break;
case ARM::t2STMIA:
case ARM::t2STMDB:
switch (Mode) {
case ARM_AM::ia: return ARM::t2STMIA_UPD;
case ARM_AM::db: return ARM::t2STMDB_UPD;
}
- break;
case ARM::VLDMSIA:
switch (Mode) {
default: llvm_unreachable("Unhandled submode!");
case ARM_AM::ia: return ARM::VLDMSIA_UPD;
case ARM_AM::db: return ARM::VLDMSDB_UPD;
}
- break;
case ARM::VLDMDIA:
switch (Mode) {
default: llvm_unreachable("Unhandled submode!");
case ARM_AM::ia: return ARM::VLDMDIA_UPD;
case ARM_AM::db: return ARM::VLDMDDB_UPD;
}
- break;
case ARM::VSTMSIA:
switch (Mode) {
default: llvm_unreachable("Unhandled submode!");
case ARM_AM::ia: return ARM::VSTMSIA_UPD;
case ARM_AM::db: return ARM::VSTMSDB_UPD;
}
- break;
case ARM::VSTMDIA:
switch (Mode) {
default: llvm_unreachable("Unhandled submode!");
case ARM_AM::ia: return ARM::VSTMDIA_UPD;
case ARM_AM::db: return ARM::VSTMDDB_UPD;
}
- break;
}
-
- return 0;
}
/// MergeBaseUpdateLSMultiple - Fold proceeding/trailing inc/dec of base
bool BaseKill = MI->getOperand(0).isKill();
unsigned Bytes = getLSMultipleTransferSize(MI);
unsigned PredReg = 0;
- ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
+ ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
int Opcode = MI->getOpcode();
DebugLoc dl = MI->getDebugLoc();
return ARM::t2STR_PRE;
default: llvm_unreachable("Unhandled opcode!");
}
- return 0;
}
static unsigned getPostIndexedLoadStoreOpcode(unsigned Opc,
return ARM::t2STR_POST;
default: llvm_unreachable("Unhandled opcode!");
}
- return 0;
}
/// MergeBaseUpdateLoadStore - Fold proceeding/trailing inc/dec of base
return false;
unsigned PredReg = 0;
- ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
+ ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
bool DoMerge = false;
ARM_AM::AddrOpc AddSub = ARM_AM::add;
unsigned NewOpc = 0;
bool OffUndef = isT2 ? false : MI->getOperand(3).isUndef();
int OffImm = getMemoryOpOffset(MI);
unsigned PredReg = 0;
- ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
+ ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
if (OddRegNum > EvenRegNum && OffImm == 0) {
// Ascending register numbers and no offset. It's safe to change it to a
unsigned NewOpc = (isLd)
? (isT2 ? (OffImm < 0 ? ARM::t2LDRi8 : ARM::t2LDRi12) : ARM::LDRi12)
: (isT2 ? (OffImm < 0 ? ARM::t2STRi8 : ARM::t2STRi12) : ARM::STRi12);
+ // Be extra careful for thumb2. t2LDRi8 can't reference a zero offset,
+ // so adjust and use t2LDRi12 here for that.
+ unsigned NewOpc2 = (isLd)
+ ? (isT2 ? (OffImm+4 < 0 ? ARM::t2LDRi8 : ARM::t2LDRi12) : ARM::LDRi12)
+ : (isT2 ? (OffImm+4 < 0 ? ARM::t2STRi8 : ARM::t2STRi12) : ARM::STRi12);
DebugLoc dl = MBBI->getDebugLoc();
// If this is a load and base register is killed, it may have been
// re-defed by the load, make sure the first load does not clobber it.
(BaseKill || OffKill) &&
(TRI->regsOverlap(EvenReg, BaseReg))) {
assert(!TRI->regsOverlap(OddReg, BaseReg));
- InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc,
+ InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc2,
OddReg, OddDeadKill, false,
BaseReg, false, BaseUndef, false, OffUndef,
Pred, PredReg, TII, isT2);
EvenDeadKill = false;
OddDeadKill = true;
}
+ // Never kill the base register in the first instruction.
+ // <rdar://problem/11101911>
+ if (EvenReg == BaseReg)
+ EvenDeadKill = false;
InsertLDR_STR(MBB, MBBI, OffImm, isLd, dl, NewOpc,
EvenReg, EvenDeadKill, EvenUndef,
BaseReg, false, BaseUndef, false, OffUndef,
Pred, PredReg, TII, isT2);
NewBBI = llvm::prior(MBBI);
- InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc,
+ InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc2,
OddReg, OddDeadKill, OddUndef,
BaseReg, BaseKill, BaseUndef, OffKill, OffUndef,
Pred, PredReg, TII, isT2);
bool isKill = MO.isDef() ? false : MO.isKill();
unsigned Base = MBBI->getOperand(1).getReg();
unsigned PredReg = 0;
- ARMCC::CondCodes Pred = llvm::getInstrPredicate(MBBI, PredReg);
+ ARMCC::CondCodes Pred = getInstrPredicate(MBBI, PredReg);
int Offset = getMemoryOpOffset(MBBI);
// Watch out for:
// r4 := ldr [r5]
// First advance to the instruction just before the start of the chain.
AdvanceRS(MBB, MemOps);
// Find a scratch register.
- unsigned Scratch = RS->FindUnusedReg(ARM::GPRRegisterClass);
+ unsigned Scratch = RS->FindUnusedReg(&ARM::GPRRegClass);
// Process the load / store instructions.
RS->forward(prior(MBBI));
static char ID;
ARMPreAllocLoadStoreOpt() : MachineFunctionPass(ID) {}
- const TargetData *TD;
+ const DataLayout *TD;
const TargetInstrInfo *TII;
const TargetRegisterInfo *TRI;
const ARMSubtarget *STI;
}
bool ARMPreAllocLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
- TD = Fn.getTarget().getTargetData();
+ TD = Fn.getTarget().getDataLayout();
TII = Fn.getTarget().getInstrInfo();
TRI = Fn.getTarget().getRegisterInfo();
STI = &Fn.getTarget().getSubtarget<ARMSubtarget>();
while (++I != E) {
if (I->isDebugValue() || MemOps.count(&*I))
continue;
- const MCInstrDesc &MCID = I->getDesc();
- if (MCID.isCall() || MCID.isTerminator() || I->hasUnmodeledSideEffects())
+ if (I->isCall() || I->isTerminator() || I->hasUnmodeledSideEffects())
return false;
- if (isLd && MCID.mayStore())
+ if (isLd && I->mayStore())
return false;
if (!isLd) {
- if (MCID.mayLoad())
+ if (I->mayLoad())
return false;
// It's not safe to move the first 'str' down.
// str r1, [r0]
// strh r5, [r0]
// str r4, [r0, #+4]
- if (MCID.mayStore())
+ if (I->mayStore())
return false;
}
for (unsigned j = 0, NumOps = I->getNumOperands(); j != NumOps; ++j) {
if (EvenReg == OddReg)
return false;
BaseReg = Op0->getOperand(1).getReg();
- Pred = llvm::getInstrPredicate(Op0, PredReg);
+ Pred = getInstrPredicate(Op0, PredReg);
dl = Op0->getDebugLoc();
return true;
}
LastOp = Op;
}
- unsigned Opcode = Op->getOpcode();
- if (LastOpcode && Opcode != LastOpcode)
+ unsigned LSMOpcode
+ = getLoadStoreMultipleOpcode(Op->getOpcode(), ARM_AM::ia);
+ if (LastOpcode && LSMOpcode != LastOpcode)
break;
int Offset = getMemoryOpOffset(Op);
}
LastOffset = Offset;
LastBytes = Bytes;
- LastOpcode = Opcode;
+ LastOpcode = LSMOpcode;
if (++NumMove == 8) // FIXME: Tune this limit.
break;
}
Ops.pop_back();
const MCInstrDesc &MCID = TII->get(NewOpc);
- const TargetRegisterClass *TRC = TII->getRegClass(MCID, 0, TRI);
+ const TargetRegisterClass *TRC = TII->getRegClass(MCID, 0, TRI, *MF);
MRI->constrainRegClass(EvenReg, TRC);
MRI->constrainRegClass(OddReg, TRC);
while (MBBI != E) {
for (; MBBI != E; ++MBBI) {
MachineInstr *MI = MBBI;
- const MCInstrDesc &MCID = MI->getDesc();
- if (MCID.isCall() || MCID.isTerminator()) {
+ if (MI->isCall() || MI->isTerminator()) {
// Stop at barriers.
++MBBI;
break;
if (!isMemoryOp(MI))
continue;
unsigned PredReg = 0;
- if (llvm::getInstrPredicate(MI, PredReg) != ARMCC::AL)
+ if (getInstrPredicate(MI, PredReg) != ARMCC::AL)
continue;
int Opc = MI->getOpcode();