#include "PPCInstrBuilder.h"
#include "PPCMachineFunctionInfo.h"
#include "PPCTargetMachine.h"
-#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/LiveIntervalAnalysis.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
+#include "llvm/CodeGen/SlotIndexes.h"
#include "llvm/MC/MCAsmInfo.h"
#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/TargetRegistry.h"
#include "llvm/Support/raw_ostream.h"
#define GET_INSTRMAP_INFO
-#define GET_INSTRINFO_CTOR
+#define GET_INSTRINFO_CTOR_DTOR
#include "PPCGenInstrInfo.inc"
using namespace llvm;
static cl::opt<bool> DisableCmpOpt("disable-ppc-cmp-opt",
cl::desc("Disable compare instruction optimization"), cl::Hidden);
+static cl::opt<bool> DisableVSXFMAMutate("disable-ppc-vsx-fma-mutation",
+cl::desc("Disable VSX FMA instruction mutation"), cl::Hidden);
+
+static cl::opt<bool> VSXSelfCopyCrash("crash-on-ppc-vsx-self-copy",
+cl::desc("Causes the backend to crash instead of generating a nop VSX copy"),
+cl::Hidden);
+
+// Pin the vtable to this file.
+void PPCInstrInfo::anchor() {}
+
PPCInstrInfo::PPCInstrInfo(PPCTargetMachine &tm)
: PPCGenInstrInfo(PPC::ADJCALLSTACKDOWN, PPC::ADJCALLSTACKUP),
TM(tm), RI(*TM.getSubtargetImpl()) {}
if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2 ||
Directive == PPC::DIR_E500mc || Directive == PPC::DIR_E5500) {
const InstrItineraryData *II = TM->getInstrItineraryData();
- return new PPCScoreboardHazardRecognizer(II, DAG);
+ return new ScoreboardHazardRecognizer(II, DAG);
}
return TargetInstrInfo::CreateTargetHazardRecognizer(TM, DAG);
const ScheduleDAG *DAG) const {
unsigned Directive = TM.getSubtarget<PPCSubtarget>().getDarwinDirective();
+ if (Directive == PPC::DIR_PWR7)
+ return new PPCDispatchGroupSBHazardRecognizer(II, DAG);
+
// Most subtargets use a PPC970 recognizer.
if (Directive != PPC::DIR_440 && Directive != PPC::DIR_A2 &&
Directive != PPC::DIR_E500mc && Directive != PPC::DIR_E5500) {
return new PPCHazardRecognizer970(TM);
}
- return new PPCScoreboardHazardRecognizer(II, DAG);
+ return new ScoreboardHazardRecognizer(II, DAG);
+}
+
+
+int PPCInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
+ const MachineInstr *DefMI, unsigned DefIdx,
+ const MachineInstr *UseMI,
+ unsigned UseIdx) const {
+ int Latency = PPCGenInstrInfo::getOperandLatency(ItinData, DefMI, DefIdx,
+ UseMI, UseIdx);
+
+ const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
+ unsigned Reg = DefMO.getReg();
+
+ const TargetRegisterInfo *TRI = &getRegisterInfo();
+ bool IsRegCR;
+ if (TRI->isVirtualRegister(Reg)) {
+ const MachineRegisterInfo *MRI =
+ &DefMI->getParent()->getParent()->getRegInfo();
+ IsRegCR = MRI->getRegClass(Reg)->hasSuperClassEq(&PPC::CRRCRegClass) ||
+ MRI->getRegClass(Reg)->hasSuperClassEq(&PPC::CRBITRCRegClass);
+ } else {
+ IsRegCR = PPC::CRRCRegClass.contains(Reg) ||
+ PPC::CRBITRCRegClass.contains(Reg);
+ }
+
+ if (UseMI->isBranch() && IsRegCR) {
+ if (Latency < 0)
+ Latency = getInstrLatency(ItinData, DefMI);
+
+ // On some cores, there is an additional delay between writing to a condition
+ // register, and using it from a branch.
+ unsigned Directive = TM.getSubtarget<PPCSubtarget>().getDarwinDirective();
+ switch (Directive) {
+ default: break;
+ case PPC::DIR_7400:
+ case PPC::DIR_750:
+ case PPC::DIR_970:
+ case PPC::DIR_E5500:
+ case PPC::DIR_PWR4:
+ case PPC::DIR_PWR5:
+ case PPC::DIR_PWR5X:
+ case PPC::DIR_PWR6:
+ case PPC::DIR_PWR6X:
+ case PPC::DIR_PWR7:
+ Latency += 2;
+ break;
+ }
+ }
+
+ return Latency;
}
// Detect 32 -> 64-bit extensions where we may reuse the low sub-register.
case PPC::LFS:
case PPC::LFD:
case PPC::RESTORE_CR:
+ case PPC::RESTORE_CRBIT:
case PPC::LVX:
+ case PPC::LXVD2X:
case PPC::RESTORE_VRSAVE:
// Check for the operands added by addFrameReference (the immediate is the
// offset which defaults to 0).
case PPC::STFS:
case PPC::STFD:
case PPC::SPILL_CR:
+ case PPC::SPILL_CRBIT:
case PPC::STVX:
+ case PPC::STXVD2X:
case PPC::SPILL_VRSAVE:
// Check for the operands added by addFrameReference (the immediate is the
// offset which defaults to 0).
// Normal instructions can be commuted the obvious way.
if (MI->getOpcode() != PPC::RLWIMI &&
- MI->getOpcode() != PPC::RLWIMIo)
+ MI->getOpcode() != PPC::RLWIMIo &&
+ MI->getOpcode() != PPC::RLWIMI8 &&
+ MI->getOpcode() != PPC::RLWIMI8o)
return TargetInstrInfo::commuteInstruction(MI, NewMI);
// Cannot commute if it has a non-zero rotate count.
unsigned Reg0 = MI->getOperand(0).getReg();
unsigned Reg1 = MI->getOperand(1).getReg();
unsigned Reg2 = MI->getOperand(2).getReg();
+ unsigned SubReg1 = MI->getOperand(1).getSubReg();
+ unsigned SubReg2 = MI->getOperand(2).getSubReg();
bool Reg1IsKill = MI->getOperand(1).isKill();
bool Reg2IsKill = MI->getOperand(2).isKill();
bool ChangeReg0 = false;
// Must be two address instruction!
assert(MI->getDesc().getOperandConstraint(0, MCOI::TIED_TO) &&
"Expecting a two-address instruction!");
+ assert(MI->getOperand(0).getSubReg() == SubReg1 && "Tied subreg mismatch");
Reg2IsKill = false;
ChangeReg0 = true;
}
.addImm((MB-1) & 31);
}
- if (ChangeReg0)
+ if (ChangeReg0) {
MI->getOperand(0).setReg(Reg2);
+ MI->getOperand(0).setSubReg(SubReg2);
+ }
MI->getOperand(2).setReg(Reg1);
MI->getOperand(1).setReg(Reg2);
+ MI->getOperand(2).setSubReg(SubReg1);
+ MI->getOperand(1).setSubReg(SubReg2);
MI->getOperand(2).setIsKill(Reg1IsKill);
MI->getOperand(1).setIsKill(Reg2IsKill);
return MI;
}
+bool PPCInstrInfo::findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
+ unsigned &SrcOpIdx2) const {
+ // For VSX A-Type FMA instructions, it is the first two operands that can be
+ // commuted, however, because the non-encoded tied input operand is listed
+ // first, the operands to swap are actually the second and third.
+
+ int AltOpc = PPC::getAltVSXFMAOpcode(MI->getOpcode());
+ if (AltOpc == -1)
+ return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
+
+ SrcOpIdx1 = 2;
+ SrcOpIdx2 = 3;
+ return true;
+}
+
void PPCInstrInfo::insertNoop(MachineBasicBlock &MBB,
MachineBasicBlock::iterator MI) const {
+ // This function is used for scheduling, and the nop wanted here is the type
+ // that terminates dispatch groups on the POWER cores.
+ unsigned Directive = TM.getSubtarget<PPCSubtarget>().getDarwinDirective();
+ unsigned Opcode;
+ switch (Directive) {
+ default: Opcode = PPC::NOP; break;
+ case PPC::DIR_PWR6: Opcode = PPC::NOP_GT_PWR6; break;
+ case PPC::DIR_PWR7: Opcode = PPC::NOP_GT_PWR7; break;
+ }
+
DebugLoc DL;
- BuildMI(MBB, MI, DL, get(PPC::NOP));
+ BuildMI(MBB, MI, DL, get(Opcode));
}
-
// Branch analysis.
// Note: If the condition register is set to CTR or CTR8 then this is a
// BDNZ (imm == 1) or BDZ (imm == 0) branch.
Cond.push_back(LastInst->getOperand(0));
Cond.push_back(LastInst->getOperand(1));
return false;
+ } else if (LastInst->getOpcode() == PPC::BC) {
+ if (!LastInst->getOperand(1).isMBB())
+ return true;
+ // Block ends with fall-through condbranch.
+ TBB = LastInst->getOperand(1).getMBB();
+ Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET));
+ Cond.push_back(LastInst->getOperand(0));
+ return false;
+ } else if (LastInst->getOpcode() == PPC::BCn) {
+ if (!LastInst->getOperand(1).isMBB())
+ return true;
+ // Block ends with fall-through condbranch.
+ TBB = LastInst->getOperand(1).getMBB();
+ Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET));
+ Cond.push_back(LastInst->getOperand(0));
+ return false;
} else if (LastInst->getOpcode() == PPC::BDNZ8 ||
LastInst->getOpcode() == PPC::BDNZ) {
if (!LastInst->getOperand(0).isMBB())
Cond.push_back(SecondLastInst->getOperand(1));
FBB = LastInst->getOperand(0).getMBB();
return false;
+ } else if (SecondLastInst->getOpcode() == PPC::BC &&
+ LastInst->getOpcode() == PPC::B) {
+ if (!SecondLastInst->getOperand(1).isMBB() ||
+ !LastInst->getOperand(0).isMBB())
+ return true;
+ TBB = SecondLastInst->getOperand(1).getMBB();
+ Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET));
+ Cond.push_back(SecondLastInst->getOperand(0));
+ FBB = LastInst->getOperand(0).getMBB();
+ return false;
+ } else if (SecondLastInst->getOpcode() == PPC::BCn &&
+ LastInst->getOpcode() == PPC::B) {
+ if (!SecondLastInst->getOperand(1).isMBB() ||
+ !LastInst->getOperand(0).isMBB())
+ return true;
+ TBB = SecondLastInst->getOperand(1).getMBB();
+ Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET));
+ Cond.push_back(SecondLastInst->getOperand(0));
+ FBB = LastInst->getOperand(0).getMBB();
+ return false;
} else if ((SecondLastInst->getOpcode() == PPC::BDNZ8 ||
SecondLastInst->getOpcode() == PPC::BDNZ) &&
LastInst->getOpcode() == PPC::B) {
--I;
}
if (I->getOpcode() != PPC::B && I->getOpcode() != PPC::BCC &&
+ I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn &&
I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ &&
I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ)
return 0;
if (I == MBB.begin()) return 1;
--I;
if (I->getOpcode() != PPC::BCC &&
+ I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn &&
I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ &&
I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ)
return 1;
BuildMI(&MBB, DL, get(Cond[0].getImm() ?
(isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
(isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB);
+ else if (Cond[0].getImm() == PPC::PRED_BIT_SET)
+ BuildMI(&MBB, DL, get(PPC::BC)).addOperand(Cond[1]).addMBB(TBB);
+ else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET)
+ BuildMI(&MBB, DL, get(PPC::BCn)).addOperand(Cond[1]).addMBB(TBB);
else // Conditional branch
BuildMI(&MBB, DL, get(PPC::BCC))
- .addImm(Cond[0].getImm()).addReg(Cond[1].getReg()).addMBB(TBB);
+ .addImm(Cond[0].getImm()).addOperand(Cond[1]).addMBB(TBB);
return 1;
}
BuildMI(&MBB, DL, get(Cond[0].getImm() ?
(isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
(isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB);
+ else if (Cond[0].getImm() == PPC::PRED_BIT_SET)
+ BuildMI(&MBB, DL, get(PPC::BC)).addOperand(Cond[1]).addMBB(TBB);
+ else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET)
+ BuildMI(&MBB, DL, get(PPC::BCn)).addOperand(Cond[1]).addMBB(TBB);
else
BuildMI(&MBB, DL, get(PPC::BCC))
- .addImm(Cond[0].getImm()).addReg(Cond[1].getReg()).addMBB(TBB);
+ .addImm(Cond[0].getImm()).addOperand(Cond[1]).addMBB(TBB);
BuildMI(&MBB, DL, get(PPC::B)).addMBB(FBB);
return 2;
}
// isel is for regular integer GPRs only.
if (!PPC::GPRCRegClass.hasSubClassEq(RC) &&
- !PPC::G8RCRegClass.hasSubClassEq(RC))
+ !PPC::GPRC_NOR0RegClass.hasSubClassEq(RC) &&
+ !PPC::G8RCRegClass.hasSubClassEq(RC) &&
+ !PPC::G8RC_NOX0RegClass.hasSubClassEq(RC))
return false;
// FIXME: These numbers are for the A2, how well they work for other cores is
const TargetRegisterClass *RC =
RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
assert(RC && "TrueReg and FalseReg must have overlapping register classes");
- assert((PPC::GPRCRegClass.hasSubClassEq(RC) ||
- PPC::G8RCRegClass.hasSubClassEq(RC)) &&
+
+ bool Is64Bit = PPC::G8RCRegClass.hasSubClassEq(RC) ||
+ PPC::G8RC_NOX0RegClass.hasSubClassEq(RC);
+ assert((Is64Bit ||
+ PPC::GPRCRegClass.hasSubClassEq(RC) ||
+ PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) &&
"isel is for regular integer GPRs only");
- unsigned OpCode =
- PPC::GPRCRegClass.hasSubClassEq(RC) ? PPC::ISEL : PPC::ISEL8;
+ unsigned OpCode = Is64Bit ? PPC::ISEL8 : PPC::ISEL;
unsigned SelectPred = Cond[0].getImm();
unsigned SubIdx;
case PPC::PRED_LE: SubIdx = PPC::sub_gt; SwapOps = true; break;
case PPC::PRED_UN: SubIdx = PPC::sub_un; SwapOps = false; break;
case PPC::PRED_NU: SubIdx = PPC::sub_un; SwapOps = true; break;
+ case PPC::PRED_BIT_SET: SubIdx = 0; SwapOps = false; break;
+ case PPC::PRED_BIT_UNSET: SubIdx = 0; SwapOps = true; break;
}
unsigned FirstReg = SwapOps ? FalseReg : TrueReg,
MachineBasicBlock::iterator I, DebugLoc DL,
unsigned DestReg, unsigned SrcReg,
bool KillSrc) const {
+ // We can end up with self copies and similar things as a result of VSX copy
+ // legalization. Promote them here.
+ const TargetRegisterInfo *TRI = &getRegisterInfo();
+ if (PPC::F8RCRegClass.contains(DestReg) &&
+ PPC::VSLRCRegClass.contains(SrcReg)) {
+ unsigned SuperReg =
+ TRI->getMatchingSuperReg(DestReg, PPC::sub_64, &PPC::VSRCRegClass);
+
+ if (VSXSelfCopyCrash && SrcReg == SuperReg)
+ llvm_unreachable("nop VSX copy");
+
+ DestReg = SuperReg;
+ } else if (PPC::VRRCRegClass.contains(DestReg) &&
+ PPC::VSHRCRegClass.contains(SrcReg)) {
+ unsigned SuperReg =
+ TRI->getMatchingSuperReg(DestReg, PPC::sub_128, &PPC::VSRCRegClass);
+
+ if (VSXSelfCopyCrash && SrcReg == SuperReg)
+ llvm_unreachable("nop VSX copy");
+
+ DestReg = SuperReg;
+ } else if (PPC::F8RCRegClass.contains(SrcReg) &&
+ PPC::VSLRCRegClass.contains(DestReg)) {
+ unsigned SuperReg =
+ TRI->getMatchingSuperReg(SrcReg, PPC::sub_64, &PPC::VSRCRegClass);
+
+ if (VSXSelfCopyCrash && DestReg == SuperReg)
+ llvm_unreachable("nop VSX copy");
+
+ SrcReg = SuperReg;
+ } else if (PPC::VRRCRegClass.contains(SrcReg) &&
+ PPC::VSHRCRegClass.contains(DestReg)) {
+ unsigned SuperReg =
+ TRI->getMatchingSuperReg(SrcReg, PPC::sub_128, &PPC::VSRCRegClass);
+
+ if (VSXSelfCopyCrash && DestReg == SuperReg)
+ llvm_unreachable("nop VSX copy");
+
+ SrcReg = SuperReg;
+ }
+
unsigned Opc;
if (PPC::GPRCRegClass.contains(DestReg, SrcReg))
Opc = PPC::OR;
Opc = PPC::MCRF;
else if (PPC::VRRCRegClass.contains(DestReg, SrcReg))
Opc = PPC::VOR;
+ else if (PPC::VSRCRegClass.contains(DestReg, SrcReg))
+ // There are two different ways this can be done:
+ // 1. xxlor : This has lower latency (on the P7), 2 cycles, but can only
+ // issue in VSU pipeline 0.
+ // 2. xmovdp/xmovsp: This has higher latency (on the P7), 6 cycles, but
+ // can go to either pipeline.
+ // We'll always use xxlor here, because in practically all cases where
+ // copies are generated, they are close enough to some use that the
+ // lower-latency form is preferable.
+ Opc = PPC::XXLOR;
+ else if (PPC::VSFRCRegClass.contains(DestReg, SrcReg))
+ Opc = PPC::XXLORf;
else if (PPC::CRBITRCRegClass.contains(DestReg, SrcReg))
Opc = PPC::CROR;
else
// update isStoreToStackSlot.
DebugLoc DL;
- if (PPC::GPRCRegClass.hasSubClassEq(RC)) {
+ if (PPC::GPRCRegClass.hasSubClassEq(RC) ||
+ PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) {
NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STW))
.addReg(SrcReg,
getKillRegState(isKill)),
FrameIdx));
- } else if (PPC::G8RCRegClass.hasSubClassEq(RC)) {
+ } else if (PPC::G8RCRegClass.hasSubClassEq(RC) ||
+ PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) {
NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STD))
.addReg(SrcReg,
getKillRegState(isKill)),
FrameIdx));
return true;
} else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) {
- // FIXME: We use CRi here because there is no mtcrf on a bit. Since the
- // backend currently only uses CR1EQ as an individual bit, this should
- // not cause any bug. If we need other uses of CR bits, the following
- // code may be invalid.
- unsigned Reg = 0;
- if (SrcReg == PPC::CR0LT || SrcReg == PPC::CR0GT ||
- SrcReg == PPC::CR0EQ || SrcReg == PPC::CR0UN)
- Reg = PPC::CR0;
- else if (SrcReg == PPC::CR1LT || SrcReg == PPC::CR1GT ||
- SrcReg == PPC::CR1EQ || SrcReg == PPC::CR1UN)
- Reg = PPC::CR1;
- else if (SrcReg == PPC::CR2LT || SrcReg == PPC::CR2GT ||
- SrcReg == PPC::CR2EQ || SrcReg == PPC::CR2UN)
- Reg = PPC::CR2;
- else if (SrcReg == PPC::CR3LT || SrcReg == PPC::CR3GT ||
- SrcReg == PPC::CR3EQ || SrcReg == PPC::CR3UN)
- Reg = PPC::CR3;
- else if (SrcReg == PPC::CR4LT || SrcReg == PPC::CR4GT ||
- SrcReg == PPC::CR4EQ || SrcReg == PPC::CR4UN)
- Reg = PPC::CR4;
- else if (SrcReg == PPC::CR5LT || SrcReg == PPC::CR5GT ||
- SrcReg == PPC::CR5EQ || SrcReg == PPC::CR5UN)
- Reg = PPC::CR5;
- else if (SrcReg == PPC::CR6LT || SrcReg == PPC::CR6GT ||
- SrcReg == PPC::CR6EQ || SrcReg == PPC::CR6UN)
- Reg = PPC::CR6;
- else if (SrcReg == PPC::CR7LT || SrcReg == PPC::CR7GT ||
- SrcReg == PPC::CR7EQ || SrcReg == PPC::CR7UN)
- Reg = PPC::CR7;
-
- return StoreRegToStackSlot(MF, Reg, isKill, FrameIdx,
- &PPC::CRRCRegClass, NewMIs, NonRI, SpillsVRS);
-
+ NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_CRBIT))
+ .addReg(SrcReg,
+ getKillRegState(isKill)),
+ FrameIdx));
+ return true;
} else if (PPC::VRRCRegClass.hasSubClassEq(RC)) {
NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STVX))
.addReg(SrcReg,
getKillRegState(isKill)),
FrameIdx));
NonRI = true;
+ } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) {
+ NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXVD2X))
+ .addReg(SrcReg,
+ getKillRegState(isKill)),
+ FrameIdx));
+ NonRI = true;
+ } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) {
+ NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXSDX))
+ .addReg(SrcReg,
+ getKillRegState(isKill)),
+ FrameIdx));
+ NonRI = true;
} else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) {
assert(TM.getSubtargetImpl()->isDarwin() &&
"VRSAVE only needs spill/restore on Darwin");
// Note: If additional load instructions are added here,
// update isLoadFromStackSlot.
- if (PPC::GPRCRegClass.hasSubClassEq(RC)) {
+ if (PPC::GPRCRegClass.hasSubClassEq(RC) ||
+ PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) {
NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LWZ),
DestReg), FrameIdx));
- } else if (PPC::G8RCRegClass.hasSubClassEq(RC)) {
+ } else if (PPC::G8RCRegClass.hasSubClassEq(RC) ||
+ PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) {
NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LD), DestReg),
FrameIdx));
} else if (PPC::F8RCRegClass.hasSubClassEq(RC)) {
FrameIdx));
return true;
} else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) {
-
- unsigned Reg = 0;
- if (DestReg == PPC::CR0LT || DestReg == PPC::CR0GT ||
- DestReg == PPC::CR0EQ || DestReg == PPC::CR0UN)
- Reg = PPC::CR0;
- else if (DestReg == PPC::CR1LT || DestReg == PPC::CR1GT ||
- DestReg == PPC::CR1EQ || DestReg == PPC::CR1UN)
- Reg = PPC::CR1;
- else if (DestReg == PPC::CR2LT || DestReg == PPC::CR2GT ||
- DestReg == PPC::CR2EQ || DestReg == PPC::CR2UN)
- Reg = PPC::CR2;
- else if (DestReg == PPC::CR3LT || DestReg == PPC::CR3GT ||
- DestReg == PPC::CR3EQ || DestReg == PPC::CR3UN)
- Reg = PPC::CR3;
- else if (DestReg == PPC::CR4LT || DestReg == PPC::CR4GT ||
- DestReg == PPC::CR4EQ || DestReg == PPC::CR4UN)
- Reg = PPC::CR4;
- else if (DestReg == PPC::CR5LT || DestReg == PPC::CR5GT ||
- DestReg == PPC::CR5EQ || DestReg == PPC::CR5UN)
- Reg = PPC::CR5;
- else if (DestReg == PPC::CR6LT || DestReg == PPC::CR6GT ||
- DestReg == PPC::CR6EQ || DestReg == PPC::CR6UN)
- Reg = PPC::CR6;
- else if (DestReg == PPC::CR7LT || DestReg == PPC::CR7GT ||
- DestReg == PPC::CR7EQ || DestReg == PPC::CR7UN)
- Reg = PPC::CR7;
-
- return LoadRegFromStackSlot(MF, DL, Reg, FrameIdx,
- &PPC::CRRCRegClass, NewMIs, NonRI, SpillsVRS);
-
+ NewMIs.push_back(addFrameReference(BuildMI(MF, DL,
+ get(PPC::RESTORE_CRBIT), DestReg),
+ FrameIdx));
+ return true;
} else if (PPC::VRRCRegClass.hasSubClassEq(RC)) {
NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LVX), DestReg),
FrameIdx));
NonRI = true;
+ } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) {
+ NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXVD2X), DestReg),
+ FrameIdx));
+ NonRI = true;
+ } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) {
+ NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXSDX), DestReg),
+ FrameIdx));
+ NonRI = true;
} else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) {
assert(TM.getSubtargetImpl()->isDarwin() &&
"VRSAVE only needs spill/restore on Darwin");
NewMIs.back()->addMemOperand(MF, MMO);
}
-MachineInstr*
-PPCInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
- int FrameIx, uint64_t Offset,
- const MDNode *MDPtr,
- DebugLoc DL) const {
- MachineInstrBuilder MIB = BuildMI(MF, DL, get(PPC::DBG_VALUE));
- addFrameReference(MIB, FrameIx, 0, false).addImm(Offset).addMetadata(MDPtr);
- return &*MIB;
-}
-
bool PPCInstrInfo::
ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
assert(Cond.size() == 2 && "Invalid PPC branch opcode!");
MI->setDesc(get(Pred[0].getImm() ?
(isPPC64 ? PPC::BDNZLR8 : PPC::BDNZLR) :
(isPPC64 ? PPC::BDZLR8 : PPC::BDZLR)));
- } else {
+ } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
MI->setDesc(get(PPC::BCLR));
MachineInstrBuilder(*MI->getParent()->getParent(), MI)
+ .addReg(Pred[1].getReg());
+ } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
+ MI->setDesc(get(PPC::BCLRn));
+ MachineInstrBuilder(*MI->getParent()->getParent(), MI)
+ .addReg(Pred[1].getReg());
+ } else {
+ MI->setDesc(get(PPC::BCCLR));
+ MachineInstrBuilder(*MI->getParent()->getParent(), MI)
.addImm(Pred[0].getImm())
.addReg(Pred[1].getReg());
}
MI->setDesc(get(Pred[0].getImm() ?
(isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
(isPPC64 ? PPC::BDZ8 : PPC::BDZ)));
+ } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
+ MachineBasicBlock *MBB = MI->getOperand(0).getMBB();
+ MI->RemoveOperand(0);
+
+ MI->setDesc(get(PPC::BC));
+ MachineInstrBuilder(*MI->getParent()->getParent(), MI)
+ .addReg(Pred[1].getReg())
+ .addMBB(MBB);
+ } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
+ MachineBasicBlock *MBB = MI->getOperand(0).getMBB();
+ MI->RemoveOperand(0);
+
+ MI->setDesc(get(PPC::BCn));
+ MachineInstrBuilder(*MI->getParent()->getParent(), MI)
+ .addReg(Pred[1].getReg())
+ .addMBB(MBB);
} else {
MachineBasicBlock *MBB = MI->getOperand(0).getMBB();
MI->RemoveOperand(0);
bool setLR = OpC == PPC::BCTRL || OpC == PPC::BCTRL8;
bool isPPC64 = TM.getSubtargetImpl()->isPPC64();
- MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8 : PPC::BCCTR8) :
- (setLR ? PPC::BCCTRL : PPC::BCCTR)));
+
+ if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
+ MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8 : PPC::BCCTR8) :
+ (setLR ? PPC::BCCTRL : PPC::BCCTR)));
+ MachineInstrBuilder(*MI->getParent()->getParent(), MI)
+ .addReg(Pred[1].getReg());
+ return true;
+ } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
+ MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8n : PPC::BCCTR8n) :
+ (setLR ? PPC::BCCTRLn : PPC::BCCTRn)));
+ MachineInstrBuilder(*MI->getParent()->getParent(), MI)
+ .addReg(Pred[1].getReg());
+ return true;
+ }
+
+ MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCCTRL8 : PPC::BCCCTR8) :
+ (setLR ? PPC::BCCCTRL : PPC::BCCCTR)));
MachineInstrBuilder(*MI->getParent()->getParent(), MI)
.addImm(Pred[0].getImm())
.addReg(Pred[1].getReg());
if (Pred2[1].getReg() == PPC::CTR8 || Pred2[1].getReg() == PPC::CTR)
return false;
+ // P1 can only subsume P2 if they test the same condition register.
+ if (Pred1[1].getReg() != Pred2[1].getReg())
+ return false;
+
PPC::Predicate P1 = (PPC::Predicate) Pred1[0].getImm();
PPC::Predicate P2 = (PPC::Predicate) Pred2[0].getImm();
if (equalityOnly) {
// We need to check the uses of the condition register in order to reject
// non-equality comparisons.
- for (MachineRegisterInfo::use_iterator I = MRI->use_begin(CRReg),
- IE = MRI->use_end(); I != IE; ++I) {
+ for (MachineRegisterInfo::use_instr_iterator I =MRI->use_instr_begin(CRReg),
+ IE = MRI->use_instr_end(); I != IE; ++I) {
MachineInstr *UseMI = &*I;
if (UseMI->getOpcode() == PPC::BCC) {
unsigned Pred = UseMI->getOperand(0).getImm();
for (MachineBasicBlock::iterator EL = CmpInstr->getParent()->end();
I != EL; ++I) {
bool FoundUse = false;
- for (MachineRegisterInfo::use_iterator J = MRI->use_begin(CRReg),
- JE = MRI->use_end(); J != JE; ++J)
+ for (MachineRegisterInfo::use_instr_iterator J =MRI->use_instr_begin(CRReg),
+ JE = MRI->use_instr_end(); J != JE; ++J)
if (&*J == &*I) {
FoundUse = true;
break;
}
if (ShouldSwap)
- for (MachineRegisterInfo::use_iterator I = MRI->use_begin(CRReg),
- IE = MRI->use_end(); I != IE; ++I) {
+ for (MachineRegisterInfo::use_instr_iterator
+ I = MRI->use_instr_begin(CRReg), IE = MRI->use_instr_end();
+ I != IE; ++I) {
MachineInstr *UseMI = &*I;
if (UseMI->getOpcode() == PPC::BCC) {
PPC::Predicate Pred = (PPC::Predicate) UseMI->getOperand(0).getImm();
assert((!equalityOnly ||
Pred == PPC::PRED_EQ || Pred == PPC::PRED_NE) &&
"Invalid predicate for equality-only optimization");
- PredsToUpdate.push_back(std::make_pair(&((*I).getOperand(0)),
+ PredsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(0)),
PPC::getSwappedPredicate(Pred)));
} else if (UseMI->getOpcode() == PPC::ISEL ||
UseMI->getOpcode() == PPC::ISEL8) {
else if (NewSubReg == PPC::sub_gt)
NewSubReg = PPC::sub_lt;
- SubRegsToUpdate.push_back(std::make_pair(&((*I).getOperand(3)),
+ SubRegsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(3)),
NewSubReg));
} else // We need to abort on a user we don't understand.
return false;
CmpInstr->eraseFromParent();
MachineBasicBlock::iterator MII = MI;
- BuildMI(*MI->getParent(), llvm::next(MII), MI->getDebugLoc(),
+ BuildMI(*MI->getParent(), std::next(MII), MI->getDebugLoc(),
get(TargetOpcode::COPY), CRReg)
.addReg(PPC::CR0, MIOpC != NewOpC ? RegState::Kill : 0);
/// instruction may be. This returns the maximum number of bytes.
///
unsigned PPCInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
- switch (MI->getOpcode()) {
- case PPC::INLINEASM: { // Inline Asm: Variable size.
+ unsigned Opcode = MI->getOpcode();
+
+ if (Opcode == PPC::INLINEASM) {
const MachineFunction *MF = MI->getParent()->getParent();
const char *AsmStr = MI->getOperand(0).getSymbolName();
return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
- }
- case PPC::PROLOG_LABEL:
- case PPC::EH_LABEL:
- case PPC::GC_LABEL:
- case PPC::DBG_VALUE:
- return 0;
- case PPC::BL8_NOP:
- case PPC::BLA8_NOP:
- return 8;
- default:
- return 4; // PowerPC instructions are all 4 bytes
+ } else {
+ const MCInstrDesc &Desc = get(Opcode);
+ return Desc.getSize();
}
}
#undef DEBUG_TYPE
+#define DEBUG_TYPE "ppc-vsx-fma-mutate"
+
+namespace {
+ // PPCVSXFMAMutate pass - For copies between VSX registers and non-VSX registers
+ // (Altivec and scalar floating-point registers), we need to transform the
+ // copies into subregister copies with other restrictions.
+ struct PPCVSXFMAMutate : public MachineFunctionPass {
+ static char ID;
+ PPCVSXFMAMutate() : MachineFunctionPass(ID) {
+ initializePPCVSXFMAMutatePass(*PassRegistry::getPassRegistry());
+ }
+
+ LiveIntervals *LIS;
+
+ const PPCTargetMachine *TM;
+ const PPCInstrInfo *TII;
+
+protected:
+ bool processBlock(MachineBasicBlock &MBB) {
+ bool Changed = false;
+
+ MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
+ for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end();
+ I != IE; ++I) {
+ MachineInstr *MI = I;
+
+ // The default (A-type) VSX FMA form kills the addend (it is taken from
+ // the target register, which is then updated to reflect the result of
+ // the FMA). If the instruction, however, kills one of the registers
+ // used for the product, then we can use the M-form instruction (which
+ // will take that value from the to-be-defined register).
+
+ int AltOpc = PPC::getAltVSXFMAOpcode(MI->getOpcode());
+ if (AltOpc == -1)
+ continue;
+
+ // This pass is run after register coalescing, and so we're looking for
+ // a situation like this:
+ // ...
+ // %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9
+ // %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16,
+ // %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16
+ // ...
+ // %vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19,
+ // %RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19
+ // ...
+ // Where we can eliminate the copy by changing from the A-type to the
+ // M-type instruction. Specifically, for this example, this means:
+ // %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16,
+ // %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16
+ // is replaced by:
+ // %vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9,
+ // %RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9
+ // and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9
+
+ SlotIndex FMAIdx = LIS->getInstructionIndex(MI);
+
+ VNInfo *AddendValNo =
+ LIS->getInterval(MI->getOperand(1).getReg()).Query(FMAIdx).valueIn();
+ MachineInstr *AddendMI = LIS->getInstructionFromIndex(AddendValNo->def);
+
+ // The addend and this instruction must be in the same block.
+
+ if (!AddendMI || AddendMI->getParent() != MI->getParent())
+ continue;
+
+ // The addend must be a full copy within the same register class.
+
+ if (!AddendMI->isFullCopy())
+ continue;
+
+ unsigned AddendSrcReg = AddendMI->getOperand(1).getReg();
+ if (TargetRegisterInfo::isVirtualRegister(AddendSrcReg)) {
+ if (MRI.getRegClass(AddendMI->getOperand(0).getReg()) !=
+ MRI.getRegClass(AddendSrcReg))
+ continue;
+ } else {
+ // If AddendSrcReg is a physical register, make sure the destination
+ // register class contains it.
+ if (!MRI.getRegClass(AddendMI->getOperand(0).getReg())
+ ->contains(AddendSrcReg))
+ continue;
+ }
+
+ // In theory, there could be other uses of the addend copy before this
+ // fma. We could deal with this, but that would require additional
+ // logic below and I suspect it will not occur in any relevant
+ // situations.
+ bool OtherUsers = false;
+ for (auto J = std::prev(I), JE = MachineBasicBlock::iterator(AddendMI);
+ J != JE; --J)
+ if (J->readsVirtualRegister(AddendMI->getOperand(0).getReg())) {
+ OtherUsers = true;
+ break;
+ }
+
+ if (OtherUsers)
+ continue;
+
+ // Find one of the product operands that is killed by this instruction.
+
+ unsigned KilledProdOp = 0, OtherProdOp = 0;
+ if (LIS->getInterval(MI->getOperand(2).getReg())
+ .Query(FMAIdx).isKill()) {
+ KilledProdOp = 2;
+ OtherProdOp = 3;
+ } else if (LIS->getInterval(MI->getOperand(3).getReg())
+ .Query(FMAIdx).isKill()) {
+ KilledProdOp = 3;
+ OtherProdOp = 2;
+ }
+
+ // If there are no killed product operands, then this transformation is
+ // likely not profitable.
+ if (!KilledProdOp)
+ continue;
+
+ // In order to replace the addend here with the source of the copy,
+ // it must still be live here.
+ if (!LIS->getInterval(AddendMI->getOperand(1).getReg()).liveAt(FMAIdx))
+ continue;
+
+ // Transform: (O2 * O3) + O1 -> (O2 * O1) + O3.
+
+ unsigned AddReg = AddendMI->getOperand(1).getReg();
+ unsigned KilledProdReg = MI->getOperand(KilledProdOp).getReg();
+ unsigned OtherProdReg = MI->getOperand(OtherProdOp).getReg();
+
+ unsigned AddSubReg = AddendMI->getOperand(1).getSubReg();
+ unsigned KilledProdSubReg = MI->getOperand(KilledProdOp).getSubReg();
+ unsigned OtherProdSubReg = MI->getOperand(OtherProdOp).getSubReg();
+
+ bool AddRegKill = AddendMI->getOperand(1).isKill();
+ bool KilledProdRegKill = MI->getOperand(KilledProdOp).isKill();
+ bool OtherProdRegKill = MI->getOperand(OtherProdOp).isKill();
+
+ bool AddRegUndef = AddendMI->getOperand(1).isUndef();
+ bool KilledProdRegUndef = MI->getOperand(KilledProdOp).isUndef();
+ bool OtherProdRegUndef = MI->getOperand(OtherProdOp).isUndef();
+
+ unsigned OldFMAReg = MI->getOperand(0).getReg();
+
+ assert(OldFMAReg == AddendMI->getOperand(0).getReg() &&
+ "Addend copy not tied to old FMA output!");
+
+ DEBUG(dbgs() << "VSX FMA Mutation:\n " << *MI;);
+
+ MI->getOperand(0).setReg(KilledProdReg);
+ MI->getOperand(1).setReg(KilledProdReg);
+ MI->getOperand(3).setReg(AddReg);
+ MI->getOperand(2).setReg(OtherProdReg);
+
+ MI->getOperand(0).setSubReg(KilledProdSubReg);
+ MI->getOperand(1).setSubReg(KilledProdSubReg);
+ MI->getOperand(3).setSubReg(AddSubReg);
+ MI->getOperand(2).setSubReg(OtherProdSubReg);
+
+ MI->getOperand(1).setIsKill(KilledProdRegKill);
+ MI->getOperand(3).setIsKill(AddRegKill);
+ MI->getOperand(2).setIsKill(OtherProdRegKill);
+
+ MI->getOperand(1).setIsUndef(KilledProdRegUndef);
+ MI->getOperand(3).setIsUndef(AddRegUndef);
+ MI->getOperand(2).setIsUndef(OtherProdRegUndef);
+
+ MI->setDesc(TII->get(AltOpc));
+
+ DEBUG(dbgs() << " -> " << *MI);
+
+ // The killed product operand was killed here, so we can reuse it now
+ // for the result of the fma.
+
+ LiveInterval &FMAInt = LIS->getInterval(OldFMAReg);
+ VNInfo *FMAValNo = FMAInt.getVNInfoAt(FMAIdx.getRegSlot());
+ for (auto UI = MRI.reg_nodbg_begin(OldFMAReg), UE = MRI.reg_nodbg_end();
+ UI != UE;) {
+ MachineOperand &UseMO = *UI;
+ MachineInstr *UseMI = UseMO.getParent();
+ ++UI;
+
+ // Don't replace the result register of the copy we're about to erase.
+ if (UseMI == AddendMI)
+ continue;
+
+ UseMO.setReg(KilledProdReg);
+ UseMO.setSubReg(KilledProdSubReg);
+ }
+
+ // Extend the live intervals of the killed product operand to hold the
+ // fma result.
+
+ LiveInterval &NewFMAInt = LIS->getInterval(KilledProdReg);
+ for (LiveInterval::iterator AI = FMAInt.begin(), AE = FMAInt.end();
+ AI != AE; ++AI) {
+ // Don't add the segment that corresponds to the original copy.
+ if (AI->valno == AddendValNo)
+ continue;
+
+ VNInfo *NewFMAValNo =
+ NewFMAInt.getNextValue(AI->start,
+ LIS->getVNInfoAllocator());
+
+ NewFMAInt.addSegment(LiveInterval::Segment(AI->start, AI->end,
+ NewFMAValNo));
+ }
+ DEBUG(dbgs() << " extended: " << NewFMAInt << '\n');
+
+ FMAInt.removeValNo(FMAValNo);
+ DEBUG(dbgs() << " trimmed: " << FMAInt << '\n');
+
+ // Remove the (now unused) copy.
+
+ DEBUG(dbgs() << " removing: " << *AddendMI << '\n');
+ LIS->RemoveMachineInstrFromMaps(AddendMI);
+ AddendMI->eraseFromParent();
+
+ Changed = true;
+ }
+
+ return Changed;
+ }
+
+public:
+ virtual bool runOnMachineFunction(MachineFunction &MF) {
+ LIS = &getAnalysis<LiveIntervals>();
+
+ TM = static_cast<const PPCTargetMachine *>(&MF.getTarget());
+ TII = TM->getInstrInfo();
+
+ bool Changed = false;
+
+ if (DisableVSXFMAMutate)
+ return Changed;
+
+ for (MachineFunction::iterator I = MF.begin(); I != MF.end();) {
+ MachineBasicBlock &B = *I++;
+ if (processBlock(B))
+ Changed = true;
+ }
+
+ return Changed;
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<LiveIntervals>();
+ AU.addPreserved<LiveIntervals>();
+ AU.addRequired<SlotIndexes>();
+ AU.addPreserved<SlotIndexes>();
+ MachineFunctionPass::getAnalysisUsage(AU);
+ }
+ };
+}
+
+INITIALIZE_PASS_BEGIN(PPCVSXFMAMutate, DEBUG_TYPE,
+ "PowerPC VSX FMA Mutation", false, false)
+INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
+INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
+INITIALIZE_PASS_END(PPCVSXFMAMutate, DEBUG_TYPE,
+ "PowerPC VSX FMA Mutation", false, false)
+
+char &llvm::PPCVSXFMAMutateID = PPCVSXFMAMutate::ID;
+
+char PPCVSXFMAMutate::ID = 0;
+FunctionPass*
+llvm::createPPCVSXFMAMutatePass() { return new PPCVSXFMAMutate(); }
+
+#undef DEBUG_TYPE
+#define DEBUG_TYPE "ppc-vsx-copy"
+
+namespace llvm {
+ void initializePPCVSXCopyPass(PassRegistry&);
+}
+
+namespace {
+ // PPCVSXCopy pass - For copies between VSX registers and non-VSX registers
+ // (Altivec and scalar floating-point registers), we need to transform the
+ // copies into subregister copies with other restrictions.
+ struct PPCVSXCopy : public MachineFunctionPass {
+ static char ID;
+ PPCVSXCopy() : MachineFunctionPass(ID) {
+ initializePPCVSXCopyPass(*PassRegistry::getPassRegistry());
+ }
+
+ const PPCTargetMachine *TM;
+ const PPCInstrInfo *TII;
+
+ bool IsRegInClass(unsigned Reg, const TargetRegisterClass *RC,
+ MachineRegisterInfo &MRI) {
+ if (TargetRegisterInfo::isVirtualRegister(Reg)) {
+ return RC->hasSubClassEq(MRI.getRegClass(Reg));
+ } else if (RC->contains(Reg)) {
+ return true;
+ }
+
+ return false;
+ }
+
+ bool IsVSReg(unsigned Reg, MachineRegisterInfo &MRI) {
+ return IsRegInClass(Reg, &PPC::VSRCRegClass, MRI);
+ }
+
+ bool IsVRReg(unsigned Reg, MachineRegisterInfo &MRI) {
+ return IsRegInClass(Reg, &PPC::VRRCRegClass, MRI);
+ }
+
+ bool IsF8Reg(unsigned Reg, MachineRegisterInfo &MRI) {
+ return IsRegInClass(Reg, &PPC::F8RCRegClass, MRI);
+ }
+
+protected:
+ bool processBlock(MachineBasicBlock &MBB) {
+ bool Changed = false;
+
+ MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
+ for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end();
+ I != IE; ++I) {
+ MachineInstr *MI = I;
+ if (!MI->isFullCopy())
+ continue;
+
+ MachineOperand &DstMO = MI->getOperand(0);
+ MachineOperand &SrcMO = MI->getOperand(1);
+
+ if ( IsVSReg(DstMO.getReg(), MRI) &&
+ !IsVSReg(SrcMO.getReg(), MRI)) {
+ // This is a copy *to* a VSX register from a non-VSX register.
+ Changed = true;
+
+ const TargetRegisterClass *SrcRC =
+ IsVRReg(SrcMO.getReg(), MRI) ? &PPC::VSHRCRegClass :
+ &PPC::VSLRCRegClass;
+ assert((IsF8Reg(SrcMO.getReg(), MRI) ||
+ IsVRReg(SrcMO.getReg(), MRI)) &&
+ "Unknown source for a VSX copy");
+
+ unsigned NewVReg = MRI.createVirtualRegister(SrcRC);
+ BuildMI(MBB, MI, MI->getDebugLoc(),
+ TII->get(TargetOpcode::SUBREG_TO_REG), NewVReg)
+ .addImm(1) // add 1, not 0, because there is no implicit clearing
+ // of the high bits.
+ .addOperand(SrcMO)
+ .addImm(IsVRReg(SrcMO.getReg(), MRI) ? PPC::sub_128 :
+ PPC::sub_64);
+
+ // The source of the original copy is now the new virtual register.
+ SrcMO.setReg(NewVReg);
+ } else if (!IsVSReg(DstMO.getReg(), MRI) &&
+ IsVSReg(SrcMO.getReg(), MRI)) {
+ // This is a copy *from* a VSX register to a non-VSX register.
+ Changed = true;
+
+ const TargetRegisterClass *DstRC =
+ IsVRReg(DstMO.getReg(), MRI) ? &PPC::VSHRCRegClass :
+ &PPC::VSLRCRegClass;
+ assert((IsF8Reg(DstMO.getReg(), MRI) ||
+ IsVRReg(DstMO.getReg(), MRI)) &&
+ "Unknown destination for a VSX copy");
+
+ // Copy the VSX value into a new VSX register of the correct subclass.
+ unsigned NewVReg = MRI.createVirtualRegister(DstRC);
+ BuildMI(MBB, MI, MI->getDebugLoc(),
+ TII->get(TargetOpcode::COPY), NewVReg)
+ .addOperand(SrcMO);
+
+ // Transform the original copy into a subregister extraction copy.
+ SrcMO.setReg(NewVReg);
+ SrcMO.setSubReg(IsVRReg(DstMO.getReg(), MRI) ? PPC::sub_128 :
+ PPC::sub_64);
+ }
+ }
+
+ return Changed;
+ }
+
+public:
+ virtual bool runOnMachineFunction(MachineFunction &MF) {
+ TM = static_cast<const PPCTargetMachine *>(&MF.getTarget());
+ TII = TM->getInstrInfo();
+
+ bool Changed = false;
+
+ for (MachineFunction::iterator I = MF.begin(); I != MF.end();) {
+ MachineBasicBlock &B = *I++;
+ if (processBlock(B))
+ Changed = true;
+ }
+
+ return Changed;
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ MachineFunctionPass::getAnalysisUsage(AU);
+ }
+ };
+}
+
+INITIALIZE_PASS(PPCVSXCopy, DEBUG_TYPE,
+ "PowerPC VSX Copy Legalization", false, false)
+
+char PPCVSXCopy::ID = 0;
+FunctionPass*
+llvm::createPPCVSXCopyPass() { return new PPCVSXCopy(); }
+
+#undef DEBUG_TYPE
+#define DEBUG_TYPE "ppc-vsx-copy-cleanup"
+
+namespace llvm {
+ void initializePPCVSXCopyCleanupPass(PassRegistry&);
+}
+
+namespace {
+ // PPCVSXCopyCleanup pass - We sometimes end up generating self copies of VSX
+ // registers (mostly because the ABI code still places all values into the
+ // "traditional" floating-point and vector registers). Remove them here.
+ struct PPCVSXCopyCleanup : public MachineFunctionPass {
+ static char ID;
+ PPCVSXCopyCleanup() : MachineFunctionPass(ID) {
+ initializePPCVSXCopyCleanupPass(*PassRegistry::getPassRegistry());
+ }
+
+ const PPCTargetMachine *TM;
+ const PPCInstrInfo *TII;
+
+protected:
+ bool processBlock(MachineBasicBlock &MBB) {
+ bool Changed = false;
+
+ SmallVector<MachineInstr *, 4> ToDelete;
+ for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end();
+ I != IE; ++I) {
+ MachineInstr *MI = I;
+ if (MI->getOpcode() == PPC::XXLOR &&
+ MI->getOperand(0).getReg() == MI->getOperand(1).getReg() &&
+ MI->getOperand(0).getReg() == MI->getOperand(2).getReg())
+ ToDelete.push_back(MI);
+ }
+
+ if (!ToDelete.empty())
+ Changed = true;
+
+ for (unsigned i = 0, ie = ToDelete.size(); i != ie; ++i) {
+ DEBUG(dbgs() << "Removing VSX self-copy: " << *ToDelete[i]);
+ ToDelete[i]->eraseFromParent();
+ }
+
+ return Changed;
+ }
+
+public:
+ virtual bool runOnMachineFunction(MachineFunction &MF) {
+ TM = static_cast<const PPCTargetMachine *>(&MF.getTarget());
+ TII = TM->getInstrInfo();
+
+ bool Changed = false;
+
+ for (MachineFunction::iterator I = MF.begin(); I != MF.end();) {
+ MachineBasicBlock &B = *I++;
+ if (processBlock(B))
+ Changed = true;
+ }
+
+ return Changed;
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ MachineFunctionPass::getAnalysisUsage(AU);
+ }
+ };
+}
+
+INITIALIZE_PASS(PPCVSXCopyCleanup, DEBUG_TYPE,
+ "PowerPC VSX Copy Cleanup", false, false)
+
+char PPCVSXCopyCleanup::ID = 0;
+FunctionPass*
+llvm::createPPCVSXCopyCleanupPass() { return new PPCVSXCopyCleanup(); }
+
+#undef DEBUG_TYPE
#define DEBUG_TYPE "ppc-early-ret"
STATISTIC(NumBCLR, "Number of early conditional returns");
STATISTIC(NumBLR, "Number of early returns");
if (J->getOpcode() == PPC::B) {
if (J->getOperand(0).getMBB() == &ReturnMBB) {
// This is an unconditional branch to the return. Replace the
- // branch with a blr.
+ // branch with a blr.
BuildMI(**PI, J, J->getDebugLoc(), TII->get(PPC::BLR));
MachineBasicBlock::iterator K = J--;
K->eraseFromParent();
if (J->getOperand(2).getMBB() == &ReturnMBB) {
// This is a conditional branch to the return. Replace the branch
// with a bclr.
- BuildMI(**PI, J, J->getDebugLoc(), TII->get(PPC::BCLR))
+ BuildMI(**PI, J, J->getDebugLoc(), TII->get(PPC::BCCLR))
.addImm(J->getOperand(0).getImm())
.addReg(J->getOperand(1).getReg());
MachineBasicBlock::iterator K = J--;
++NumBCLR;
continue;
}
+ } else if (J->getOpcode() == PPC::BC || J->getOpcode() == PPC::BCn) {
+ if (J->getOperand(1).getMBB() == &ReturnMBB) {
+ // This is a conditional branch to the return. Replace the branch
+ // with a bclr.
+ BuildMI(**PI, J, J->getDebugLoc(),
+ TII->get(J->getOpcode() == PPC::BC ?
+ PPC::BCLR : PPC::BCLRn))
+ .addReg(J->getOperand(0).getReg());
+ MachineBasicBlock::iterator K = J--;
+ K->eraseFromParent();
+ BlockChanged = true;
+ ++NumBCLR;
+ continue;
+ }
} else if (J->isBranch()) {
if (J->isIndirectBranch()) {
if (ReturnMBB.hasAddressTaken())
if ((*PI)->canFallThrough() && (*PI)->isLayoutSuccessor(&ReturnMBB))
OtherReference = true;
- // Predecessors are stored in a vector and can't be removed here.
+ // Predecessors are stored in a vector and can't be removed here.
if (!OtherReference && BlockChanged) {
PredToRemove.push_back(*PI);
}
return Changed;
for (MachineFunction::iterator I = MF.begin(); I != MF.end();) {
- MachineBasicBlock &B = *I++;
+ MachineBasicBlock &B = *I++;
if (processBlock(B))
Changed = true;
}
char PPCEarlyReturn::ID = 0;
FunctionPass*
llvm::createPPCEarlyReturnPass() { return new PPCEarlyReturn(); }
-
OSDN Git Service
