#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
+#include "llvm/CodeGen/ScheduleDAG.h"
#include "llvm/CodeGen/SlotIndexes.h"
+#include "llvm/CodeGen/StackMaps.h"
#include "llvm/MC/MCAsmInfo.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/TargetRegistry.h"
#include "llvm/Support/raw_ostream.h"
+using namespace llvm;
+
+#define DEBUG_TYPE "ppc-instr-info"
+
#define GET_INSTRMAP_INFO
#define GET_INSTRINFO_CTOR_DTOR
#include "PPCGenInstrInfo.inc"
-using namespace llvm;
-
static cl::
opt<bool> DisableCTRLoopAnal("disable-ppc-ctrloop-analysis", cl::Hidden,
cl::desc("Disable analysis for CTR loops"));
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()) {}
+PPCInstrInfo::PPCInstrInfo(PPCSubtarget &STI)
+ : PPCGenInstrInfo(PPC::ADJCALLSTACKDOWN, PPC::ADJCALLSTACKUP),
+ Subtarget(STI), RI(STI) {}
/// CreateTargetHazardRecognizer - Return the hazard recognizer to use for
/// this target when scheduling the DAG.
-ScheduleHazardRecognizer *PPCInstrInfo::CreateTargetHazardRecognizer(
- const TargetMachine *TM,
- const ScheduleDAG *DAG) const {
- unsigned Directive = TM->getSubtarget<PPCSubtarget>().getDarwinDirective();
+ScheduleHazardRecognizer *
+PPCInstrInfo::CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
+ const ScheduleDAG *DAG) const {
+ unsigned Directive =
+ static_cast<const PPCSubtarget *>(STI)->getDarwinDirective();
if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2 ||
Directive == PPC::DIR_E500mc || Directive == PPC::DIR_E5500) {
- const InstrItineraryData *II = TM->getInstrItineraryData();
+ const InstrItineraryData *II =
+ static_cast<const PPCSubtarget *>(STI)->getInstrItineraryData();
return new ScoreboardHazardRecognizer(II, DAG);
}
- return TargetInstrInfo::CreateTargetHazardRecognizer(TM, DAG);
+ return TargetInstrInfo::CreateTargetHazardRecognizer(STI, DAG);
}
/// CreateTargetPostRAHazardRecognizer - Return the postRA hazard recognizer
/// to use for this target when scheduling the DAG.
-ScheduleHazardRecognizer *PPCInstrInfo::CreateTargetPostRAHazardRecognizer(
- const InstrItineraryData *II,
- const ScheduleDAG *DAG) const {
- unsigned Directive = TM.getSubtarget<PPCSubtarget>().getDarwinDirective();
+ScheduleHazardRecognizer *
+PPCInstrInfo::CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
+ const ScheduleDAG *DAG) const {
+ unsigned Directive =
+ DAG->MF.getSubtarget<PPCSubtarget>().getDarwinDirective();
- if (Directive == PPC::DIR_PWR7)
+ if (Directive == PPC::DIR_PWR7 || Directive == PPC::DIR_PWR8)
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) {
- assert(TM.getInstrInfo() && "No InstrInfo?");
+ assert(DAG->TII && "No InstrInfo?");
- return new PPCHazardRecognizer970(TM);
+ return new PPCHazardRecognizer970(*DAG);
}
return new ScoreboardHazardRecognizer(II, DAG);
// 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();
+ unsigned Directive = Subtarget.getDarwinDirective();
switch (Directive) {
default: break;
case PPC::DIR_7400:
case PPC::DIR_PWR6:
case PPC::DIR_PWR6X:
case PPC::DIR_PWR7:
+ case PPC::DIR_PWR8:
Latency += 2;
break;
}
// Normal instructions can be commuted the obvious way.
if (MI->getOpcode() != PPC::RLWIMI &&
- MI->getOpcode() != PPC::RLWIMIo &&
- MI->getOpcode() != PPC::RLWIMI8 &&
- MI->getOpcode() != PPC::RLWIMI8o)
+ MI->getOpcode() != PPC::RLWIMIo)
return TargetInstrInfo::commuteInstruction(MI, NewMI);
+ // Note that RLWIMI can be commuted as a 32-bit instruction, but not as a
+ // 64-bit instruction (so we don't handle PPC::RLWIMI8 here), because
+ // changing the relative order of the mask operands might change what happens
+ // to the high-bits of the mask (and, thus, the result).
// Cannot commute if it has a non-zero rotate count.
if (MI->getOperand(3).getImm() != 0)
- return 0;
+ return nullptr;
// If we have a zero rotate count, we have:
// M = mask(MB,ME)
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 Directive = Subtarget.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;
+ case PPC::DIR_PWR8: Opcode = PPC::NOP_GT_PWR7; break; /* FIXME: Update when P8 InstrScheduling model is ready */
}
DebugLoc DL;
BuildMI(MBB, MI, DL, get(Opcode));
}
+/// getNoopForMachoTarget - Return the noop instruction to use for a noop.
+void PPCInstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
+ NopInst.setOpcode(PPC::NOP);
+}
+
// 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.
MachineBasicBlock *&FBB,
SmallVectorImpl<MachineOperand> &Cond,
bool AllowModify) const {
- bool isPPC64 = TM.getSubtargetImpl()->isPPC64();
+ bool isPPC64 = Subtarget.isPPC64();
// If the block has no terminators, it just falls into the block after it.
MachineBasicBlock::iterator I = MBB.end();
assert((Cond.size() == 2 || Cond.size() == 0) &&
"PPC branch conditions have two components!");
- bool isPPC64 = TM.getSubtargetImpl()->isPPC64();
+ bool isPPC64 = Subtarget.isPPC64();
// One-way branch.
- if (FBB == 0) {
+ if (!FBB) {
if (Cond.empty()) // Unconditional branch
BuildMI(&MBB, DL, get(PPC::B)).addMBB(TBB);
else if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8)
const SmallVectorImpl<MachineOperand> &Cond,
unsigned TrueReg, unsigned FalseReg,
int &CondCycles, int &TrueCycles, int &FalseCycles) const {
- if (!TM.getSubtargetImpl()->hasISEL())
+ if (!Subtarget.hasISEL())
return false;
if (Cond.size() != 2)
assert(Cond.size() == 2 &&
"PPC branch conditions have two components!");
- assert(TM.getSubtargetImpl()->hasISEL() &&
+ assert(Subtarget.hasISEL() &&
"Cannot insert select on target without ISEL support");
// Get the register classes.
// 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
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() &&
+ assert(Subtarget.isDarwin() &&
"VRSAVE only needs spill/restore on Darwin");
NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_VRSAVE))
.addReg(SrcReg,
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() &&
+ assert(Subtarget.isDarwin() &&
"VRSAVE only needs spill/restore on Darwin");
NewMIs.push_back(addFrameReference(BuildMI(MF, DL,
get(PPC::RESTORE_VRSAVE),
unsigned ZeroReg;
if (UseInfo->isLookupPtrRegClass()) {
- bool isPPC64 = TM.getSubtargetImpl()->isPPC64();
+ bool isPPC64 = Subtarget.isPPC64();
ZeroReg = isPPC64 ? PPC::ZERO8 : PPC::ZERO;
} else {
ZeroReg = UseInfo->RegClass == PPC::G8RC_NOX0RegClassID ?
MachineInstr *MI,
const SmallVectorImpl<MachineOperand> &Pred) const {
unsigned OpC = MI->getOpcode();
- if (OpC == PPC::BLR) {
+ if (OpC == PPC::BLR || OpC == PPC::BLR8) {
if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) {
- bool isPPC64 = TM.getSubtargetImpl()->isPPC64();
+ bool isPPC64 = Subtarget.isPPC64();
MI->setDesc(get(Pred[0].getImm() ?
(isPPC64 ? PPC::BDNZLR8 : PPC::BDNZLR) :
(isPPC64 ? PPC::BDZLR8 : PPC::BDZLR)));
return true;
} else if (OpC == PPC::B) {
if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) {
- bool isPPC64 = TM.getSubtargetImpl()->isPPC64();
+ bool isPPC64 = Subtarget.isPPC64();
MI->setDesc(get(Pred[0].getImm() ?
(isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
(isPPC64 ? PPC::BDZ8 : PPC::BDZ)));
llvm_unreachable("Cannot predicate bctr[l] on the ctr register");
bool setLR = OpC == PPC::BCTRL || OpC == PPC::BCTRL8;
- bool isPPC64 = TM.getSubtargetImpl()->isPPC64();
+ bool isPPC64 = Subtarget.isPPC64();
if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8 : PPC::BCCTR8) :
return false;
case PPC::B:
case PPC::BLR:
+ case PPC::BLR8:
case PPC::BCTR:
case PPC::BCTR8:
case PPC::BCTRL:
// for equality checks (as those don't depend on the sign). On PPC64,
// we are restricted to equality for unsigned 64-bit comparisons and for
// signed 32-bit comparisons the applicability is more restricted.
- bool isPPC64 = TM.getSubtargetImpl()->isPPC64();
+ bool isPPC64 = Subtarget.isPPC64();
bool is32BitSignedCompare = OpC == PPC::CMPWI || OpC == PPC::CMPW;
bool is32BitUnsignedCompare = OpC == PPC::CMPLWI || OpC == PPC::CMPLW;
bool is64BitUnsignedCompare = OpC == PPC::CMPLDI || OpC == PPC::CMPLD;
// There are two possible candidates which can be changed to set CR[01].
// One is MI, the other is a SUB instruction.
// For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
- MachineInstr *Sub = NULL;
+ MachineInstr *Sub = nullptr;
if (SrcReg2 != 0)
// MI is not a candidate for CMPrr.
- MI = NULL;
+ MI = nullptr;
// FIXME: Conservatively refuse to convert an instruction which isn't in the
// same BB as the comparison. This is to allow the check below to avoid calls
// (and other explicit clobbers); instead we should really check for these
const MachineFunction *MF = MI->getParent()->getParent();
const char *AsmStr = MI->getOperand(0).getSymbolName();
return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
+ } else if (Opcode == TargetOpcode::STACKMAP) {
+ return MI->getOperand(1).getImm();
+ } else if (Opcode == TargetOpcode::PATCHPOINT) {
+ PatchPointOpers Opers(MI);
+ return Opers.getMetaOper(PatchPointOpers::NBytesPos).getImm();
} 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->getParent() != MI->getParent())
- continue;
-
- // The addend must be a full copy within the same register class.
-
- if (!AddendMI->isFullCopy())
- continue;
-
- if (MRI.getRegClass(AddendMI->getOperand(0).getReg()) !=
- MRI.getRegClass(AddendMI->getOperand(1).getReg()))
- 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-early-ret"
STATISTIC(NumBCLR, "Number of early conditional returns");
initializePPCEarlyReturnPass(*PassRegistry::getPassRegistry());
}
- const PPCTargetMachine *TM;
- const PPCInstrInfo *TII;
+ const TargetInstrInfo *TII;
protected:
bool processBlock(MachineBasicBlock &ReturnMBB) {
I = ReturnMBB.SkipPHIsAndLabels(I);
// The block must be essentially empty except for the blr.
- if (I == ReturnMBB.end() || I->getOpcode() != PPC::BLR ||
+ if (I == ReturnMBB.end() ||
+ (I->getOpcode() != PPC::BLR && I->getOpcode() != PPC::BLR8) ||
I != ReturnMBB.getLastNonDebugInstr())
return Changed;
PIE = ReturnMBB.pred_end(); PI != PIE; ++PI) {
bool OtherReference = false, BlockChanged = false;
for (MachineBasicBlock::iterator J = (*PI)->getLastNonDebugInstr();;) {
+ MachineInstrBuilder MIB;
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.
- BuildMI(**PI, J, J->getDebugLoc(), TII->get(PPC::BLR));
+ MIB =
+ BuildMI(**PI, J, J->getDebugLoc(), TII->get(I->getOpcode()));
+ MIB.copyImplicitOps(I);
MachineBasicBlock::iterator K = J--;
K->eraseFromParent();
BlockChanged = true;
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::BCCLR))
- .addImm(J->getOperand(0).getImm())
- .addReg(J->getOperand(1).getReg());
+ MIB = BuildMI(**PI, J, J->getDebugLoc(), TII->get(PPC::BCCLR))
+ .addImm(J->getOperand(0).getImm())
+ .addReg(J->getOperand(1).getReg());
+ MIB.copyImplicitOps(I);
MachineBasicBlock::iterator K = J--;
K->eraseFromParent();
BlockChanged = true;
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());
+ MIB = BuildMI(**PI, J, J->getDebugLoc(),
+ TII->get(J->getOpcode() == PPC::BC ?
+ PPC::BCLR : PPC::BCLRn))
+ .addReg(J->getOperand(0).getReg());
+ MIB.copyImplicitOps(I);
MachineBasicBlock::iterator K = J--;
K->eraseFromParent();
BlockChanged = true;
}
public:
- virtual bool runOnMachineFunction(MachineFunction &MF) {
- TM = static_cast<const PPCTargetMachine *>(&MF.getTarget());
- TII = TM->getInstrInfo();
+ bool runOnMachineFunction(MachineFunction &MF) override {
+ TII = MF.getSubtarget().getInstrInfo();
bool Changed = false;
return Changed;
}
- virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ void getAnalysisUsage(AnalysisUsage &AU) const override {
MachineFunctionPass::getAnalysisUsage(AU);
}
};