1 //===-- SystemZInstrInfo.cpp - SystemZ instruction information ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the SystemZ implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "SystemZInstrInfo.h"
15 #include "MCTargetDesc/SystemZMCTargetDesc.h"
17 #include "SystemZInstrBuilder.h"
18 #include "SystemZSubtarget.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/CodeGen/LiveInterval.h"
21 #include "llvm/CodeGen/LiveIntervals.h"
22 #include "llvm/CodeGen/LiveVariables.h"
23 #include "llvm/CodeGen/MachineBasicBlock.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstr.h"
27 #include "llvm/CodeGen/MachineMemOperand.h"
28 #include "llvm/CodeGen/MachineOperand.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/SlotIndexes.h"
31 #include "llvm/CodeGen/TargetInstrInfo.h"
32 #include "llvm/CodeGen/TargetSubtargetInfo.h"
33 #include "llvm/MC/MCInstrDesc.h"
34 #include "llvm/MC/MCRegisterInfo.h"
35 #include "llvm/Support/BranchProbability.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Target/TargetMachine.h"
45 #define GET_INSTRINFO_CTOR_DTOR
46 #define GET_INSTRMAP_INFO
47 #include "SystemZGenInstrInfo.inc"
49 #define DEBUG_TYPE "systemz-II"
50 STATISTIC(LOCRMuxJumps, "Number of LOCRMux jump-sequences (lower is better)");
52 // Return a mask with Count low bits set.
53 static uint64_t allOnes(unsigned int Count) {
54 return Count == 0 ? 0 : (uint64_t(1) << (Count - 1) << 1) - 1;
57 // Reg should be a 32-bit GPR. Return true if it is a high register rather
58 // than a low register.
59 static bool isHighReg(unsigned int Reg) {
60 if (SystemZ::GRH32BitRegClass.contains(Reg))
62 assert(SystemZ::GR32BitRegClass.contains(Reg) && "Invalid GRX32");
66 // Pin the vtable to this file.
67 void SystemZInstrInfo::anchor() {}
69 SystemZInstrInfo::SystemZInstrInfo(SystemZSubtarget &sti)
70 : SystemZGenInstrInfo(SystemZ::ADJCALLSTACKDOWN, SystemZ::ADJCALLSTACKUP),
74 // MI is a 128-bit load or store. Split it into two 64-bit loads or stores,
75 // each having the opcode given by NewOpcode.
76 void SystemZInstrInfo::splitMove(MachineBasicBlock::iterator MI,
77 unsigned NewOpcode) const {
78 MachineBasicBlock *MBB = MI->getParent();
79 MachineFunction &MF = *MBB->getParent();
81 // Get two load or store instructions. Use the original instruction for one
82 // of them (arbitrarily the second here) and create a clone for the other.
83 MachineInstr *EarlierMI = MF.CloneMachineInstr(&*MI);
84 MBB->insert(MI, EarlierMI);
86 // Set up the two 64-bit registers and remember super reg and its flags.
87 MachineOperand &HighRegOp = EarlierMI->getOperand(0);
88 MachineOperand &LowRegOp = MI->getOperand(0);
89 unsigned Reg128 = LowRegOp.getReg();
90 unsigned Reg128Killed = getKillRegState(LowRegOp.isKill());
91 unsigned Reg128Undef = getUndefRegState(LowRegOp.isUndef());
92 HighRegOp.setReg(RI.getSubReg(HighRegOp.getReg(), SystemZ::subreg_h64));
93 LowRegOp.setReg(RI.getSubReg(LowRegOp.getReg(), SystemZ::subreg_l64));
96 // Add implicit uses of the super register in case one of the subregs is
97 // undefined. We could track liveness and skip storing an undefined
98 // subreg, but this is hopefully rare (discovered with llvm-stress).
99 // If Reg128 was killed, set kill flag on MI.
100 unsigned Reg128UndefImpl = (Reg128Undef | RegState::Implicit);
101 MachineInstrBuilder(MF, EarlierMI).addReg(Reg128, Reg128UndefImpl);
102 MachineInstrBuilder(MF, MI).addReg(Reg128, (Reg128UndefImpl | Reg128Killed));
105 // The address in the first (high) instruction is already correct.
106 // Adjust the offset in the second (low) instruction.
107 MachineOperand &HighOffsetOp = EarlierMI->getOperand(2);
108 MachineOperand &LowOffsetOp = MI->getOperand(2);
109 LowOffsetOp.setImm(LowOffsetOp.getImm() + 8);
111 // Clear the kill flags on the registers in the first instruction.
112 if (EarlierMI->getOperand(0).isReg() && EarlierMI->getOperand(0).isUse())
113 EarlierMI->getOperand(0).setIsKill(false);
114 EarlierMI->getOperand(1).setIsKill(false);
115 EarlierMI->getOperand(3).setIsKill(false);
118 unsigned HighOpcode = getOpcodeForOffset(NewOpcode, HighOffsetOp.getImm());
119 unsigned LowOpcode = getOpcodeForOffset(NewOpcode, LowOffsetOp.getImm());
120 assert(HighOpcode && LowOpcode && "Both offsets should be in range");
122 EarlierMI->setDesc(get(HighOpcode));
123 MI->setDesc(get(LowOpcode));
126 // Split ADJDYNALLOC instruction MI.
127 void SystemZInstrInfo::splitAdjDynAlloc(MachineBasicBlock::iterator MI) const {
128 MachineBasicBlock *MBB = MI->getParent();
129 MachineFunction &MF = *MBB->getParent();
130 MachineFrameInfo &MFFrame = MF.getFrameInfo();
131 MachineOperand &OffsetMO = MI->getOperand(2);
133 uint64_t Offset = (MFFrame.getMaxCallFrameSize() +
134 SystemZMC::CallFrameSize +
136 unsigned NewOpcode = getOpcodeForOffset(SystemZ::LA, Offset);
137 assert(NewOpcode && "No support for huge argument lists yet");
138 MI->setDesc(get(NewOpcode));
139 OffsetMO.setImm(Offset);
142 // MI is an RI-style pseudo instruction. Replace it with LowOpcode
143 // if the first operand is a low GR32 and HighOpcode if the first operand
144 // is a high GR32. ConvertHigh is true if LowOpcode takes a signed operand
145 // and HighOpcode takes an unsigned 32-bit operand. In those cases,
146 // MI has the same kind of operand as LowOpcode, so needs to be converted
147 // if HighOpcode is used.
148 void SystemZInstrInfo::expandRIPseudo(MachineInstr &MI, unsigned LowOpcode,
150 bool ConvertHigh) const {
151 unsigned Reg = MI.getOperand(0).getReg();
152 bool IsHigh = isHighReg(Reg);
153 MI.setDesc(get(IsHigh ? HighOpcode : LowOpcode));
154 if (IsHigh && ConvertHigh)
155 MI.getOperand(1).setImm(uint32_t(MI.getOperand(1).getImm()));
158 // MI is a three-operand RIE-style pseudo instruction. Replace it with
159 // LowOpcodeK if the registers are both low GR32s, otherwise use a move
160 // followed by HighOpcode or LowOpcode, depending on whether the target
161 // is a high or low GR32.
162 void SystemZInstrInfo::expandRIEPseudo(MachineInstr &MI, unsigned LowOpcode,
164 unsigned HighOpcode) const {
165 unsigned DestReg = MI.getOperand(0).getReg();
166 unsigned SrcReg = MI.getOperand(1).getReg();
167 bool DestIsHigh = isHighReg(DestReg);
168 bool SrcIsHigh = isHighReg(SrcReg);
169 if (!DestIsHigh && !SrcIsHigh)
170 MI.setDesc(get(LowOpcodeK));
172 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), DestReg, SrcReg,
173 SystemZ::LR, 32, MI.getOperand(1).isKill(),
174 MI.getOperand(1).isUndef());
175 MI.setDesc(get(DestIsHigh ? HighOpcode : LowOpcode));
176 MI.getOperand(1).setReg(DestReg);
177 MI.tieOperands(0, 1);
181 // MI is an RXY-style pseudo instruction. Replace it with LowOpcode
182 // if the first operand is a low GR32 and HighOpcode if the first operand
184 void SystemZInstrInfo::expandRXYPseudo(MachineInstr &MI, unsigned LowOpcode,
185 unsigned HighOpcode) const {
186 unsigned Reg = MI.getOperand(0).getReg();
187 unsigned Opcode = getOpcodeForOffset(isHighReg(Reg) ? HighOpcode : LowOpcode,
188 MI.getOperand(2).getImm());
189 MI.setDesc(get(Opcode));
192 // MI is a load-on-condition pseudo instruction with a single register
193 // (source or destination) operand. Replace it with LowOpcode if the
194 // register is a low GR32 and HighOpcode if the register is a high GR32.
195 void SystemZInstrInfo::expandLOCPseudo(MachineInstr &MI, unsigned LowOpcode,
196 unsigned HighOpcode) const {
197 unsigned Reg = MI.getOperand(0).getReg();
198 unsigned Opcode = isHighReg(Reg) ? HighOpcode : LowOpcode;
199 MI.setDesc(get(Opcode));
202 // MI is a load-register-on-condition pseudo instruction. Replace it with
203 // LowOpcode if source and destination are both low GR32s and HighOpcode if
204 // source and destination are both high GR32s.
205 void SystemZInstrInfo::expandLOCRPseudo(MachineInstr &MI, unsigned LowOpcode,
206 unsigned HighOpcode) const {
207 unsigned DestReg = MI.getOperand(0).getReg();
208 unsigned SrcReg = MI.getOperand(2).getReg();
209 bool DestIsHigh = isHighReg(DestReg);
210 bool SrcIsHigh = isHighReg(SrcReg);
212 if (!DestIsHigh && !SrcIsHigh)
213 MI.setDesc(get(LowOpcode));
214 else if (DestIsHigh && SrcIsHigh)
215 MI.setDesc(get(HighOpcode));
219 // If we were unable to implement the pseudo with a single instruction, we
220 // need to convert it back into a branch sequence. This cannot be done here
221 // since the caller of expandPostRAPseudo does not handle changes to the CFG
222 // correctly. This change is defered to the SystemZExpandPseudo pass.
225 // MI is an RR-style pseudo instruction that zero-extends the low Size bits
226 // of one GRX32 into another. Replace it with LowOpcode if both operands
227 // are low registers, otherwise use RISB[LH]G.
228 void SystemZInstrInfo::expandZExtPseudo(MachineInstr &MI, unsigned LowOpcode,
229 unsigned Size) const {
230 MachineInstrBuilder MIB =
231 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(),
232 MI.getOperand(0).getReg(), MI.getOperand(1).getReg(), LowOpcode,
233 Size, MI.getOperand(1).isKill(), MI.getOperand(1).isUndef());
235 // Keep the remaining operands as-is.
236 for (unsigned I = 2; I < MI.getNumOperands(); ++I)
237 MIB.add(MI.getOperand(I));
239 MI.eraseFromParent();
242 void SystemZInstrInfo::expandLoadStackGuard(MachineInstr *MI) const {
243 MachineBasicBlock *MBB = MI->getParent();
244 MachineFunction &MF = *MBB->getParent();
245 const unsigned Reg64 = MI->getOperand(0).getReg();
246 const unsigned Reg32 = RI.getSubReg(Reg64, SystemZ::subreg_l32);
248 // EAR can only load the low subregister so us a shift for %a0 to produce
249 // the GR containing %a0 and %a1.
252 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::EAR), Reg32)
254 .addReg(Reg64, RegState::ImplicitDefine);
256 // sllg <reg>, <reg>, 32
257 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::SLLG), Reg64)
263 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::EAR), Reg32)
264 .addReg(SystemZ::A1);
266 // lg <reg>, 40(<reg>)
267 MI->setDesc(get(SystemZ::LG));
268 MachineInstrBuilder(MF, MI).addReg(Reg64).addImm(40).addReg(0);
271 // Emit a zero-extending move from 32-bit GPR SrcReg to 32-bit GPR
272 // DestReg before MBBI in MBB. Use LowLowOpcode when both DestReg and SrcReg
273 // are low registers, otherwise use RISB[LH]G. Size is the number of bits
274 // taken from the low end of SrcReg (8 for LLCR, 16 for LLHR and 32 for LR).
275 // KillSrc is true if this move is the last use of SrcReg.
277 SystemZInstrInfo::emitGRX32Move(MachineBasicBlock &MBB,
278 MachineBasicBlock::iterator MBBI,
279 const DebugLoc &DL, unsigned DestReg,
280 unsigned SrcReg, unsigned LowLowOpcode,
281 unsigned Size, bool KillSrc,
282 bool UndefSrc) const {
284 bool DestIsHigh = isHighReg(DestReg);
285 bool SrcIsHigh = isHighReg(SrcReg);
286 if (DestIsHigh && SrcIsHigh)
287 Opcode = SystemZ::RISBHH;
288 else if (DestIsHigh && !SrcIsHigh)
289 Opcode = SystemZ::RISBHL;
290 else if (!DestIsHigh && SrcIsHigh)
291 Opcode = SystemZ::RISBLH;
293 return BuildMI(MBB, MBBI, DL, get(LowLowOpcode), DestReg)
294 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc));
296 unsigned Rotate = (DestIsHigh != SrcIsHigh ? 32 : 0);
297 return BuildMI(MBB, MBBI, DL, get(Opcode), DestReg)
298 .addReg(DestReg, RegState::Undef)
299 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc))
300 .addImm(32 - Size).addImm(128 + 31).addImm(Rotate);
303 MachineInstr *SystemZInstrInfo::commuteInstructionImpl(MachineInstr &MI,
306 unsigned OpIdx2) const {
307 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & {
309 return *MI.getParent()->getParent()->CloneMachineInstr(&MI);
313 switch (MI.getOpcode()) {
314 case SystemZ::LOCRMux:
315 case SystemZ::LOCFHR:
317 case SystemZ::LOCGR: {
318 auto &WorkingMI = cloneIfNew(MI);
320 unsigned CCValid = WorkingMI.getOperand(3).getImm();
321 unsigned CCMask = WorkingMI.getOperand(4).getImm();
322 WorkingMI.getOperand(4).setImm(CCMask ^ CCValid);
323 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
327 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
331 // If MI is a simple load or store for a frame object, return the register
332 // it loads or stores and set FrameIndex to the index of the frame object.
333 // Return 0 otherwise.
335 // Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores.
336 static int isSimpleMove(const MachineInstr &MI, int &FrameIndex,
338 const MCInstrDesc &MCID = MI.getDesc();
339 if ((MCID.TSFlags & Flag) && MI.getOperand(1).isFI() &&
340 MI.getOperand(2).getImm() == 0 && MI.getOperand(3).getReg() == 0) {
341 FrameIndex = MI.getOperand(1).getIndex();
342 return MI.getOperand(0).getReg();
347 unsigned SystemZInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
348 int &FrameIndex) const {
349 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXLoad);
352 unsigned SystemZInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
353 int &FrameIndex) const {
354 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXStore);
357 bool SystemZInstrInfo::isStackSlotCopy(const MachineInstr &MI,
359 int &SrcFrameIndex) const {
360 // Check for MVC 0(Length,FI1),0(FI2)
361 const MachineFrameInfo &MFI = MI.getParent()->getParent()->getFrameInfo();
362 if (MI.getOpcode() != SystemZ::MVC || !MI.getOperand(0).isFI() ||
363 MI.getOperand(1).getImm() != 0 || !MI.getOperand(3).isFI() ||
364 MI.getOperand(4).getImm() != 0)
367 // Check that Length covers the full slots.
368 int64_t Length = MI.getOperand(2).getImm();
369 unsigned FI1 = MI.getOperand(0).getIndex();
370 unsigned FI2 = MI.getOperand(3).getIndex();
371 if (MFI.getObjectSize(FI1) != Length ||
372 MFI.getObjectSize(FI2) != Length)
375 DestFrameIndex = FI1;
380 bool SystemZInstrInfo::analyzeBranch(MachineBasicBlock &MBB,
381 MachineBasicBlock *&TBB,
382 MachineBasicBlock *&FBB,
383 SmallVectorImpl<MachineOperand> &Cond,
384 bool AllowModify) const {
385 // Most of the code and comments here are boilerplate.
387 // Start from the bottom of the block and work up, examining the
388 // terminator instructions.
389 MachineBasicBlock::iterator I = MBB.end();
390 while (I != MBB.begin()) {
392 if (I->isDebugInstr())
395 // Working from the bottom, when we see a non-terminator instruction, we're
397 if (!isUnpredicatedTerminator(*I))
400 // A terminator that isn't a branch can't easily be handled by this
405 // Can't handle indirect branches.
406 SystemZII::Branch Branch(getBranchInfo(*I));
407 if (!Branch.Target->isMBB())
410 // Punt on compound branches.
411 if (Branch.Type != SystemZII::BranchNormal)
414 if (Branch.CCMask == SystemZ::CCMASK_ANY) {
415 // Handle unconditional branches.
417 TBB = Branch.Target->getMBB();
421 // If the block has any instructions after a JMP, delete them.
422 while (std::next(I) != MBB.end())
423 std::next(I)->eraseFromParent();
428 // Delete the JMP if it's equivalent to a fall-through.
429 if (MBB.isLayoutSuccessor(Branch.Target->getMBB())) {
431 I->eraseFromParent();
436 // TBB is used to indicate the unconditinal destination.
437 TBB = Branch.Target->getMBB();
441 // Working from the bottom, handle the first conditional branch.
443 // FIXME: add X86-style branch swap
445 TBB = Branch.Target->getMBB();
446 Cond.push_back(MachineOperand::CreateImm(Branch.CCValid));
447 Cond.push_back(MachineOperand::CreateImm(Branch.CCMask));
451 // Handle subsequent conditional branches.
452 assert(Cond.size() == 2 && TBB && "Should have seen a conditional branch");
454 // Only handle the case where all conditional branches branch to the same
456 if (TBB != Branch.Target->getMBB())
459 // If the conditions are the same, we can leave them alone.
460 unsigned OldCCValid = Cond[0].getImm();
461 unsigned OldCCMask = Cond[1].getImm();
462 if (OldCCValid == Branch.CCValid && OldCCMask == Branch.CCMask)
465 // FIXME: Try combining conditions like X86 does. Should be easy on Z!
472 unsigned SystemZInstrInfo::removeBranch(MachineBasicBlock &MBB,
473 int *BytesRemoved) const {
474 assert(!BytesRemoved && "code size not handled");
476 // Most of the code and comments here are boilerplate.
477 MachineBasicBlock::iterator I = MBB.end();
480 while (I != MBB.begin()) {
482 if (I->isDebugInstr())
486 if (!getBranchInfo(*I).Target->isMBB())
488 // Remove the branch.
489 I->eraseFromParent();
497 bool SystemZInstrInfo::
498 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
499 assert(Cond.size() == 2 && "Invalid condition");
500 Cond[1].setImm(Cond[1].getImm() ^ Cond[0].getImm());
504 unsigned SystemZInstrInfo::insertBranch(MachineBasicBlock &MBB,
505 MachineBasicBlock *TBB,
506 MachineBasicBlock *FBB,
507 ArrayRef<MachineOperand> Cond,
509 int *BytesAdded) const {
510 // In this function we output 32-bit branches, which should always
511 // have enough range. They can be shortened and relaxed by later code
512 // in the pipeline, if desired.
514 // Shouldn't be a fall through.
515 assert(TBB && "insertBranch must not be told to insert a fallthrough");
516 assert((Cond.size() == 2 || Cond.size() == 0) &&
517 "SystemZ branch conditions have one component!");
518 assert(!BytesAdded && "code size not handled");
521 // Unconditional branch?
522 assert(!FBB && "Unconditional branch with multiple successors!");
523 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(TBB);
527 // Conditional branch.
529 unsigned CCValid = Cond[0].getImm();
530 unsigned CCMask = Cond[1].getImm();
531 BuildMI(&MBB, DL, get(SystemZ::BRC))
532 .addImm(CCValid).addImm(CCMask).addMBB(TBB);
536 // Two-way Conditional branch. Insert the second branch.
537 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(FBB);
543 bool SystemZInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
544 unsigned &SrcReg2, int &Mask,
546 assert(MI.isCompare() && "Caller should have checked for a comparison");
548 if (MI.getNumExplicitOperands() == 2 && MI.getOperand(0).isReg() &&
549 MI.getOperand(1).isImm()) {
550 SrcReg = MI.getOperand(0).getReg();
552 Value = MI.getOperand(1).getImm();
560 bool SystemZInstrInfo::canInsertSelect(const MachineBasicBlock &MBB,
561 ArrayRef<MachineOperand> Pred,
562 unsigned TrueReg, unsigned FalseReg,
563 int &CondCycles, int &TrueCycles,
564 int &FalseCycles) const {
565 // Not all subtargets have LOCR instructions.
566 if (!STI.hasLoadStoreOnCond())
568 if (Pred.size() != 2)
571 // Check register classes.
572 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
573 const TargetRegisterClass *RC =
574 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
578 // We have LOCR instructions for 32 and 64 bit general purpose registers.
579 if ((STI.hasLoadStoreOnCond2() &&
580 SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) ||
581 SystemZ::GR32BitRegClass.hasSubClassEq(RC) ||
582 SystemZ::GR64BitRegClass.hasSubClassEq(RC)) {
589 // Can't do anything else.
593 void SystemZInstrInfo::insertSelect(MachineBasicBlock &MBB,
594 MachineBasicBlock::iterator I,
595 const DebugLoc &DL, unsigned DstReg,
596 ArrayRef<MachineOperand> Pred,
598 unsigned FalseReg) const {
599 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
600 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
602 assert(Pred.size() == 2 && "Invalid condition");
603 unsigned CCValid = Pred[0].getImm();
604 unsigned CCMask = Pred[1].getImm();
607 if (SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) {
608 if (STI.hasLoadStoreOnCond2())
609 Opc = SystemZ::LOCRMux;
612 MRI.constrainRegClass(DstReg, &SystemZ::GR32BitRegClass);
613 unsigned TReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass);
614 unsigned FReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass);
615 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), TReg).addReg(TrueReg);
616 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), FReg).addReg(FalseReg);
620 } else if (SystemZ::GR64BitRegClass.hasSubClassEq(RC))
621 Opc = SystemZ::LOCGR;
623 llvm_unreachable("Invalid register class");
625 BuildMI(MBB, I, DL, get(Opc), DstReg)
626 .addReg(FalseReg).addReg(TrueReg)
627 .addImm(CCValid).addImm(CCMask);
630 bool SystemZInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
632 MachineRegisterInfo *MRI) const {
633 unsigned DefOpc = DefMI.getOpcode();
634 if (DefOpc != SystemZ::LHIMux && DefOpc != SystemZ::LHI &&
635 DefOpc != SystemZ::LGHI)
637 if (DefMI.getOperand(0).getReg() != Reg)
639 int32_t ImmVal = (int32_t)DefMI.getOperand(1).getImm();
641 unsigned UseOpc = UseMI.getOpcode();
646 case SystemZ::LOCRMux:
647 if (!STI.hasLoadStoreOnCond2())
649 NewUseOpc = SystemZ::LOCHIMux;
650 if (UseMI.getOperand(2).getReg() == Reg)
652 else if (UseMI.getOperand(1).getReg() == Reg)
653 UseIdx = 2, CommuteIdx = 1;
658 if (!STI.hasLoadStoreOnCond2())
660 NewUseOpc = SystemZ::LOCGHI;
661 if (UseMI.getOperand(2).getReg() == Reg)
663 else if (UseMI.getOperand(1).getReg() == Reg)
664 UseIdx = 2, CommuteIdx = 1;
672 if (CommuteIdx != -1)
673 if (!commuteInstruction(UseMI, false, CommuteIdx, UseIdx))
676 bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
677 UseMI.setDesc(get(NewUseOpc));
678 UseMI.getOperand(UseIdx).ChangeToImmediate(ImmVal);
680 DefMI.eraseFromParent();
685 bool SystemZInstrInfo::isPredicable(const MachineInstr &MI) const {
686 unsigned Opcode = MI.getOpcode();
687 if (Opcode == SystemZ::Return ||
688 Opcode == SystemZ::Trap ||
689 Opcode == SystemZ::CallJG ||
690 Opcode == SystemZ::CallBR)
695 bool SystemZInstrInfo::
696 isProfitableToIfCvt(MachineBasicBlock &MBB,
697 unsigned NumCycles, unsigned ExtraPredCycles,
698 BranchProbability Probability) const {
699 // Avoid using conditional returns at the end of a loop (since then
700 // we'd need to emit an unconditional branch to the beginning anyway,
701 // making the loop body longer). This doesn't apply for low-probability
702 // loops (eg. compare-and-swap retry), so just decide based on branch
703 // probability instead of looping structure.
704 // However, since Compare and Trap instructions cost the same as a regular
705 // Compare instruction, we should allow the if conversion to convert this
706 // into a Conditional Compare regardless of the branch probability.
707 if (MBB.getLastNonDebugInstr()->getOpcode() != SystemZ::Trap &&
708 MBB.succ_empty() && Probability < BranchProbability(1, 8))
710 // For now only convert single instructions.
711 return NumCycles == 1;
714 bool SystemZInstrInfo::
715 isProfitableToIfCvt(MachineBasicBlock &TMBB,
716 unsigned NumCyclesT, unsigned ExtraPredCyclesT,
717 MachineBasicBlock &FMBB,
718 unsigned NumCyclesF, unsigned ExtraPredCyclesF,
719 BranchProbability Probability) const {
720 // For now avoid converting mutually-exclusive cases.
724 bool SystemZInstrInfo::
725 isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
726 BranchProbability Probability) const {
727 // For now only duplicate single instructions.
728 return NumCycles == 1;
731 bool SystemZInstrInfo::PredicateInstruction(
732 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const {
733 assert(Pred.size() == 2 && "Invalid condition");
734 unsigned CCValid = Pred[0].getImm();
735 unsigned CCMask = Pred[1].getImm();
736 assert(CCMask > 0 && CCMask < 15 && "Invalid predicate");
737 unsigned Opcode = MI.getOpcode();
738 if (Opcode == SystemZ::Trap) {
739 MI.setDesc(get(SystemZ::CondTrap));
740 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
741 .addImm(CCValid).addImm(CCMask)
742 .addReg(SystemZ::CC, RegState::Implicit);
745 if (Opcode == SystemZ::Return) {
746 MI.setDesc(get(SystemZ::CondReturn));
747 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
748 .addImm(CCValid).addImm(CCMask)
749 .addReg(SystemZ::CC, RegState::Implicit);
752 if (Opcode == SystemZ::CallJG) {
753 MachineOperand FirstOp = MI.getOperand(0);
754 const uint32_t *RegMask = MI.getOperand(1).getRegMask();
757 MI.setDesc(get(SystemZ::CallBRCL));
758 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
763 .addReg(SystemZ::CC, RegState::Implicit);
766 if (Opcode == SystemZ::CallBR) {
767 const uint32_t *RegMask = MI.getOperand(0).getRegMask();
769 MI.setDesc(get(SystemZ::CallBCR));
770 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
771 .addImm(CCValid).addImm(CCMask)
773 .addReg(SystemZ::CC, RegState::Implicit);
779 void SystemZInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
780 MachineBasicBlock::iterator MBBI,
781 const DebugLoc &DL, unsigned DestReg,
782 unsigned SrcReg, bool KillSrc) const {
783 // Split 128-bit GPR moves into two 64-bit moves. Add implicit uses of the
784 // super register in case one of the subregs is undefined.
785 // This handles ADDR128 too.
786 if (SystemZ::GR128BitRegClass.contains(DestReg, SrcReg)) {
787 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_h64),
788 RI.getSubReg(SrcReg, SystemZ::subreg_h64), KillSrc);
789 MachineInstrBuilder(*MBB.getParent(), std::prev(MBBI))
790 .addReg(SrcReg, RegState::Implicit);
791 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_l64),
792 RI.getSubReg(SrcReg, SystemZ::subreg_l64), KillSrc);
793 MachineInstrBuilder(*MBB.getParent(), std::prev(MBBI))
794 .addReg(SrcReg, (getKillRegState(KillSrc) | RegState::Implicit));
798 if (SystemZ::GRX32BitRegClass.contains(DestReg, SrcReg)) {
799 emitGRX32Move(MBB, MBBI, DL, DestReg, SrcReg, SystemZ::LR, 32, KillSrc,
804 // Move 128-bit floating-point values between VR128 and FP128.
805 if (SystemZ::VR128BitRegClass.contains(DestReg) &&
806 SystemZ::FP128BitRegClass.contains(SrcReg)) {
808 RI.getMatchingSuperReg(RI.getSubReg(SrcReg, SystemZ::subreg_h64),
809 SystemZ::subreg_h64, &SystemZ::VR128BitRegClass);
811 RI.getMatchingSuperReg(RI.getSubReg(SrcReg, SystemZ::subreg_l64),
812 SystemZ::subreg_h64, &SystemZ::VR128BitRegClass);
814 BuildMI(MBB, MBBI, DL, get(SystemZ::VMRHG), DestReg)
815 .addReg(SrcRegHi, getKillRegState(KillSrc))
816 .addReg(SrcRegLo, getKillRegState(KillSrc));
819 if (SystemZ::FP128BitRegClass.contains(DestReg) &&
820 SystemZ::VR128BitRegClass.contains(SrcReg)) {
822 RI.getMatchingSuperReg(RI.getSubReg(DestReg, SystemZ::subreg_h64),
823 SystemZ::subreg_h64, &SystemZ::VR128BitRegClass);
825 RI.getMatchingSuperReg(RI.getSubReg(DestReg, SystemZ::subreg_l64),
826 SystemZ::subreg_h64, &SystemZ::VR128BitRegClass);
828 if (DestRegHi != SrcReg)
829 copyPhysReg(MBB, MBBI, DL, DestRegHi, SrcReg, false);
830 BuildMI(MBB, MBBI, DL, get(SystemZ::VREPG), DestRegLo)
831 .addReg(SrcReg, getKillRegState(KillSrc)).addImm(1);
835 // Move CC value from/to a GR32.
836 if (SrcReg == SystemZ::CC) {
837 auto MIB = BuildMI(MBB, MBBI, DL, get(SystemZ::IPM), DestReg);
839 const MachineFunction *MF = MBB.getParent();
840 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
841 MIB->addRegisterKilled(SrcReg, TRI);
845 if (DestReg == SystemZ::CC) {
846 BuildMI(MBB, MBBI, DL, get(SystemZ::TMLH))
847 .addReg(SrcReg, getKillRegState(KillSrc))
848 .addImm(3 << (SystemZ::IPM_CC - 16));
852 // Everything else needs only one instruction.
854 if (SystemZ::GR64BitRegClass.contains(DestReg, SrcReg))
855 Opcode = SystemZ::LGR;
856 else if (SystemZ::FP32BitRegClass.contains(DestReg, SrcReg))
857 // For z13 we prefer LDR over LER to avoid partial register dependencies.
858 Opcode = STI.hasVector() ? SystemZ::LDR32 : SystemZ::LER;
859 else if (SystemZ::FP64BitRegClass.contains(DestReg, SrcReg))
860 Opcode = SystemZ::LDR;
861 else if (SystemZ::FP128BitRegClass.contains(DestReg, SrcReg))
862 Opcode = SystemZ::LXR;
863 else if (SystemZ::VR32BitRegClass.contains(DestReg, SrcReg))
864 Opcode = SystemZ::VLR32;
865 else if (SystemZ::VR64BitRegClass.contains(DestReg, SrcReg))
866 Opcode = SystemZ::VLR64;
867 else if (SystemZ::VR128BitRegClass.contains(DestReg, SrcReg))
868 Opcode = SystemZ::VLR;
869 else if (SystemZ::AR32BitRegClass.contains(DestReg, SrcReg))
870 Opcode = SystemZ::CPYA;
871 else if (SystemZ::AR32BitRegClass.contains(DestReg) &&
872 SystemZ::GR32BitRegClass.contains(SrcReg))
873 Opcode = SystemZ::SAR;
874 else if (SystemZ::GR32BitRegClass.contains(DestReg) &&
875 SystemZ::AR32BitRegClass.contains(SrcReg))
876 Opcode = SystemZ::EAR;
878 llvm_unreachable("Impossible reg-to-reg copy");
880 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg)
881 .addReg(SrcReg, getKillRegState(KillSrc));
884 void SystemZInstrInfo::storeRegToStackSlot(
885 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned SrcReg,
886 bool isKill, int FrameIdx, const TargetRegisterClass *RC,
887 const TargetRegisterInfo *TRI) const {
888 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
890 // Callers may expect a single instruction, so keep 128-bit moves
891 // together for now and lower them after register allocation.
892 unsigned LoadOpcode, StoreOpcode;
893 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode);
894 addFrameReference(BuildMI(MBB, MBBI, DL, get(StoreOpcode))
895 .addReg(SrcReg, getKillRegState(isKill)),
899 void SystemZInstrInfo::loadRegFromStackSlot(
900 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned DestReg,
901 int FrameIdx, const TargetRegisterClass *RC,
902 const TargetRegisterInfo *TRI) const {
903 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
905 // Callers may expect a single instruction, so keep 128-bit moves
906 // together for now and lower them after register allocation.
907 unsigned LoadOpcode, StoreOpcode;
908 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode);
909 addFrameReference(BuildMI(MBB, MBBI, DL, get(LoadOpcode), DestReg),
913 // Return true if MI is a simple load or store with a 12-bit displacement
914 // and no index. Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores.
915 static bool isSimpleBD12Move(const MachineInstr *MI, unsigned Flag) {
916 const MCInstrDesc &MCID = MI->getDesc();
917 return ((MCID.TSFlags & Flag) &&
918 isUInt<12>(MI->getOperand(2).getImm()) &&
919 MI->getOperand(3).getReg() == 0);
926 LogicOp(unsigned regSize, unsigned immLSB, unsigned immSize)
927 : RegSize(regSize), ImmLSB(immLSB), ImmSize(immSize) {}
929 explicit operator bool() const { return RegSize; }
931 unsigned RegSize = 0;
933 unsigned ImmSize = 0;
936 } // end anonymous namespace
938 static LogicOp interpretAndImmediate(unsigned Opcode) {
940 case SystemZ::NILMux: return LogicOp(32, 0, 16);
941 case SystemZ::NIHMux: return LogicOp(32, 16, 16);
942 case SystemZ::NILL64: return LogicOp(64, 0, 16);
943 case SystemZ::NILH64: return LogicOp(64, 16, 16);
944 case SystemZ::NIHL64: return LogicOp(64, 32, 16);
945 case SystemZ::NIHH64: return LogicOp(64, 48, 16);
946 case SystemZ::NIFMux: return LogicOp(32, 0, 32);
947 case SystemZ::NILF64: return LogicOp(64, 0, 32);
948 case SystemZ::NIHF64: return LogicOp(64, 32, 32);
949 default: return LogicOp();
953 static void transferDeadCC(MachineInstr *OldMI, MachineInstr *NewMI) {
954 if (OldMI->registerDefIsDead(SystemZ::CC)) {
955 MachineOperand *CCDef = NewMI->findRegisterDefOperand(SystemZ::CC);
956 if (CCDef != nullptr)
957 CCDef->setIsDead(true);
961 // Used to return from convertToThreeAddress after replacing two-address
962 // instruction OldMI with three-address instruction NewMI.
963 static MachineInstr *finishConvertToThreeAddress(MachineInstr *OldMI,
967 unsigned NumOps = OldMI->getNumOperands();
968 for (unsigned I = 1; I < NumOps; ++I) {
969 MachineOperand &Op = OldMI->getOperand(I);
970 if (Op.isReg() && Op.isKill())
971 LV->replaceKillInstruction(Op.getReg(), *OldMI, *NewMI);
974 transferDeadCC(OldMI, NewMI);
978 MachineInstr *SystemZInstrInfo::convertToThreeAddress(
979 MachineFunction::iterator &MFI, MachineInstr &MI, LiveVariables *LV) const {
980 MachineBasicBlock *MBB = MI.getParent();
981 MachineFunction *MF = MBB->getParent();
982 MachineRegisterInfo &MRI = MF->getRegInfo();
984 unsigned Opcode = MI.getOpcode();
985 unsigned NumOps = MI.getNumOperands();
987 // Try to convert something like SLL into SLLK, if supported.
988 // We prefer to keep the two-operand form where possible both
989 // because it tends to be shorter and because some instructions
990 // have memory forms that can be used during spilling.
991 if (STI.hasDistinctOps()) {
992 MachineOperand &Dest = MI.getOperand(0);
993 MachineOperand &Src = MI.getOperand(1);
994 unsigned DestReg = Dest.getReg();
995 unsigned SrcReg = Src.getReg();
996 // AHIMux is only really a three-operand instruction when both operands
997 // are low registers. Try to constrain both operands to be low if
999 if (Opcode == SystemZ::AHIMux &&
1000 TargetRegisterInfo::isVirtualRegister(DestReg) &&
1001 TargetRegisterInfo::isVirtualRegister(SrcReg) &&
1002 MRI.getRegClass(DestReg)->contains(SystemZ::R1L) &&
1003 MRI.getRegClass(SrcReg)->contains(SystemZ::R1L)) {
1004 MRI.constrainRegClass(DestReg, &SystemZ::GR32BitRegClass);
1005 MRI.constrainRegClass(SrcReg, &SystemZ::GR32BitRegClass);
1007 int ThreeOperandOpcode = SystemZ::getThreeOperandOpcode(Opcode);
1008 if (ThreeOperandOpcode >= 0) {
1009 // Create three address instruction without adding the implicit
1010 // operands. Those will instead be copied over from the original
1011 // instruction by the loop below.
1012 MachineInstrBuilder MIB(
1013 *MF, MF->CreateMachineInstr(get(ThreeOperandOpcode), MI.getDebugLoc(),
1014 /*NoImplicit=*/true));
1016 // Keep the kill state, but drop the tied flag.
1017 MIB.addReg(Src.getReg(), getKillRegState(Src.isKill()), Src.getSubReg());
1018 // Keep the remaining operands as-is.
1019 for (unsigned I = 2; I < NumOps; ++I)
1020 MIB.add(MI.getOperand(I));
1021 MBB->insert(MI, MIB);
1022 return finishConvertToThreeAddress(&MI, MIB, LV);
1026 // Try to convert an AND into an RISBG-type instruction.
1027 if (LogicOp And = interpretAndImmediate(Opcode)) {
1028 uint64_t Imm = MI.getOperand(2).getImm() << And.ImmLSB;
1029 // AND IMMEDIATE leaves the other bits of the register unchanged.
1030 Imm |= allOnes(And.RegSize) & ~(allOnes(And.ImmSize) << And.ImmLSB);
1031 unsigned Start, End;
1032 if (isRxSBGMask(Imm, And.RegSize, Start, End)) {
1034 if (And.RegSize == 64) {
1035 NewOpcode = SystemZ::RISBG;
1036 // Prefer RISBGN if available, since it does not clobber CC.
1037 if (STI.hasMiscellaneousExtensions())
1038 NewOpcode = SystemZ::RISBGN;
1040 NewOpcode = SystemZ::RISBMux;
1044 MachineOperand &Dest = MI.getOperand(0);
1045 MachineOperand &Src = MI.getOperand(1);
1046 MachineInstrBuilder MIB =
1047 BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpcode))
1050 .addReg(Src.getReg(), getKillRegState(Src.isKill()),
1055 return finishConvertToThreeAddress(&MI, MIB, LV);
1061 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl(
1062 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
1063 MachineBasicBlock::iterator InsertPt, int FrameIndex,
1064 LiveIntervals *LIS) const {
1065 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1066 const MachineFrameInfo &MFI = MF.getFrameInfo();
1067 unsigned Size = MFI.getObjectSize(FrameIndex);
1068 unsigned Opcode = MI.getOpcode();
1070 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
1071 if (LIS != nullptr && (Opcode == SystemZ::LA || Opcode == SystemZ::LAY) &&
1072 isInt<8>(MI.getOperand(2).getImm()) && !MI.getOperand(3).getReg()) {
1074 // Check CC liveness, since new instruction introduces a dead
1076 MCRegUnitIterator CCUnit(SystemZ::CC, TRI);
1077 LiveRange &CCLiveRange = LIS->getRegUnit(*CCUnit);
1079 assert(!CCUnit.isValid() && "CC only has one reg unit.");
1081 LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot();
1082 if (!CCLiveRange.liveAt(MISlot)) {
1083 // LA(Y) %reg, CONST(%reg) -> AGSI %mem, CONST
1084 MachineInstr *BuiltMI = BuildMI(*InsertPt->getParent(), InsertPt,
1085 MI.getDebugLoc(), get(SystemZ::AGSI))
1086 .addFrameIndex(FrameIndex)
1088 .addImm(MI.getOperand(2).getImm());
1089 BuiltMI->findRegisterDefOperand(SystemZ::CC)->setIsDead(true);
1090 CCLiveRange.createDeadDef(MISlot, LIS->getVNInfoAllocator());
1097 // All other cases require a single operand.
1098 if (Ops.size() != 1)
1101 unsigned OpNum = Ops[0];
1103 TRI->getRegSizeInBits(*MF.getRegInfo()
1104 .getRegClass(MI.getOperand(OpNum).getReg())) &&
1105 "Invalid size combination");
1107 if ((Opcode == SystemZ::AHI || Opcode == SystemZ::AGHI) && OpNum == 0 &&
1108 isInt<8>(MI.getOperand(2).getImm())) {
1109 // A(G)HI %reg, CONST -> A(G)SI %mem, CONST
1110 Opcode = (Opcode == SystemZ::AHI ? SystemZ::ASI : SystemZ::AGSI);
1111 MachineInstr *BuiltMI =
1112 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode))
1113 .addFrameIndex(FrameIndex)
1115 .addImm(MI.getOperand(2).getImm());
1116 transferDeadCC(&MI, BuiltMI);
1120 if ((Opcode == SystemZ::ALFI && OpNum == 0 &&
1121 isInt<8>((int32_t)MI.getOperand(2).getImm())) ||
1122 (Opcode == SystemZ::ALGFI && OpNum == 0 &&
1123 isInt<8>((int64_t)MI.getOperand(2).getImm()))) {
1124 // AL(G)FI %reg, CONST -> AL(G)SI %mem, CONST
1125 Opcode = (Opcode == SystemZ::ALFI ? SystemZ::ALSI : SystemZ::ALGSI);
1126 MachineInstr *BuiltMI =
1127 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode))
1128 .addFrameIndex(FrameIndex)
1130 .addImm((int8_t)MI.getOperand(2).getImm());
1131 transferDeadCC(&MI, BuiltMI);
1135 if ((Opcode == SystemZ::SLFI && OpNum == 0 &&
1136 isInt<8>((int32_t)-MI.getOperand(2).getImm())) ||
1137 (Opcode == SystemZ::SLGFI && OpNum == 0 &&
1138 isInt<8>((int64_t)-MI.getOperand(2).getImm()))) {
1139 // SL(G)FI %reg, CONST -> AL(G)SI %mem, -CONST
1140 Opcode = (Opcode == SystemZ::SLFI ? SystemZ::ALSI : SystemZ::ALGSI);
1141 MachineInstr *BuiltMI =
1142 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode))
1143 .addFrameIndex(FrameIndex)
1145 .addImm((int8_t)-MI.getOperand(2).getImm());
1146 transferDeadCC(&MI, BuiltMI);
1150 if (Opcode == SystemZ::LGDR || Opcode == SystemZ::LDGR) {
1151 bool Op0IsGPR = (Opcode == SystemZ::LGDR);
1152 bool Op1IsGPR = (Opcode == SystemZ::LDGR);
1153 // If we're spilling the destination of an LDGR or LGDR, store the
1154 // source register instead.
1156 unsigned StoreOpcode = Op1IsGPR ? SystemZ::STG : SystemZ::STD;
1157 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(),
1159 .add(MI.getOperand(1))
1160 .addFrameIndex(FrameIndex)
1164 // If we're spilling the source of an LDGR or LGDR, load the
1165 // destination register instead.
1167 unsigned LoadOpcode = Op0IsGPR ? SystemZ::LG : SystemZ::LD;
1168 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(),
1170 .add(MI.getOperand(0))
1171 .addFrameIndex(FrameIndex)
1177 // Look for cases where the source of a simple store or the destination
1178 // of a simple load is being spilled. Try to use MVC instead.
1180 // Although MVC is in practice a fast choice in these cases, it is still
1181 // logically a bytewise copy. This means that we cannot use it if the
1182 // load or store is volatile. We also wouldn't be able to use MVC if
1183 // the two memories partially overlap, but that case cannot occur here,
1184 // because we know that one of the memories is a full frame index.
1186 // For performance reasons, we also want to avoid using MVC if the addresses
1187 // might be equal. We don't worry about that case here, because spill slot
1188 // coloring happens later, and because we have special code to remove
1189 // MVCs that turn out to be redundant.
1190 if (OpNum == 0 && MI.hasOneMemOperand()) {
1191 MachineMemOperand *MMO = *MI.memoperands_begin();
1192 if (MMO->getSize() == Size && !MMO->isVolatile()) {
1193 // Handle conversion of loads.
1194 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXLoad)) {
1195 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(),
1197 .addFrameIndex(FrameIndex)
1200 .add(MI.getOperand(1))
1201 .addImm(MI.getOperand(2).getImm())
1202 .addMemOperand(MMO);
1204 // Handle conversion of stores.
1205 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXStore)) {
1206 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(),
1208 .add(MI.getOperand(1))
1209 .addImm(MI.getOperand(2).getImm())
1211 .addFrameIndex(FrameIndex)
1213 .addMemOperand(MMO);
1218 // If the spilled operand is the final one, try to change <INSN>R
1220 int MemOpcode = SystemZ::getMemOpcode(Opcode);
1221 if (MemOpcode >= 0) {
1222 unsigned NumOps = MI.getNumExplicitOperands();
1223 if (OpNum == NumOps - 1) {
1224 const MCInstrDesc &MemDesc = get(MemOpcode);
1225 uint64_t AccessBytes = SystemZII::getAccessSize(MemDesc.TSFlags);
1226 assert(AccessBytes != 0 && "Size of access should be known");
1227 assert(AccessBytes <= Size && "Access outside the frame index");
1228 uint64_t Offset = Size - AccessBytes;
1229 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt,
1230 MI.getDebugLoc(), get(MemOpcode));
1231 for (unsigned I = 0; I < OpNum; ++I)
1232 MIB.add(MI.getOperand(I));
1233 MIB.addFrameIndex(FrameIndex).addImm(Offset);
1234 if (MemDesc.TSFlags & SystemZII::HasIndex)
1236 transferDeadCC(&MI, MIB);
1244 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl(
1245 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
1246 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI,
1247 LiveIntervals *LIS) const {
1251 bool SystemZInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
1252 switch (MI.getOpcode()) {
1254 splitMove(MI, SystemZ::LG);
1257 case SystemZ::ST128:
1258 splitMove(MI, SystemZ::STG);
1262 splitMove(MI, SystemZ::LD);
1266 splitMove(MI, SystemZ::STD);
1269 case SystemZ::LBMux:
1270 expandRXYPseudo(MI, SystemZ::LB, SystemZ::LBH);
1273 case SystemZ::LHMux:
1274 expandRXYPseudo(MI, SystemZ::LH, SystemZ::LHH);
1277 case SystemZ::LLCRMux:
1278 expandZExtPseudo(MI, SystemZ::LLCR, 8);
1281 case SystemZ::LLHRMux:
1282 expandZExtPseudo(MI, SystemZ::LLHR, 16);
1285 case SystemZ::LLCMux:
1286 expandRXYPseudo(MI, SystemZ::LLC, SystemZ::LLCH);
1289 case SystemZ::LLHMux:
1290 expandRXYPseudo(MI, SystemZ::LLH, SystemZ::LLHH);
1294 expandRXYPseudo(MI, SystemZ::L, SystemZ::LFH);
1297 case SystemZ::LOCMux:
1298 expandLOCPseudo(MI, SystemZ::LOC, SystemZ::LOCFH);
1301 case SystemZ::LOCHIMux:
1302 expandLOCPseudo(MI, SystemZ::LOCHI, SystemZ::LOCHHI);
1305 case SystemZ::LOCRMux:
1306 expandLOCRPseudo(MI, SystemZ::LOCR, SystemZ::LOCFHR);
1309 case SystemZ::STCMux:
1310 expandRXYPseudo(MI, SystemZ::STC, SystemZ::STCH);
1313 case SystemZ::STHMux:
1314 expandRXYPseudo(MI, SystemZ::STH, SystemZ::STHH);
1317 case SystemZ::STMux:
1318 expandRXYPseudo(MI, SystemZ::ST, SystemZ::STFH);
1321 case SystemZ::STOCMux:
1322 expandLOCPseudo(MI, SystemZ::STOC, SystemZ::STOCFH);
1325 case SystemZ::LHIMux:
1326 expandRIPseudo(MI, SystemZ::LHI, SystemZ::IIHF, true);
1329 case SystemZ::IIFMux:
1330 expandRIPseudo(MI, SystemZ::IILF, SystemZ::IIHF, false);
1333 case SystemZ::IILMux:
1334 expandRIPseudo(MI, SystemZ::IILL, SystemZ::IIHL, false);
1337 case SystemZ::IIHMux:
1338 expandRIPseudo(MI, SystemZ::IILH, SystemZ::IIHH, false);
1341 case SystemZ::NIFMux:
1342 expandRIPseudo(MI, SystemZ::NILF, SystemZ::NIHF, false);
1345 case SystemZ::NILMux:
1346 expandRIPseudo(MI, SystemZ::NILL, SystemZ::NIHL, false);
1349 case SystemZ::NIHMux:
1350 expandRIPseudo(MI, SystemZ::NILH, SystemZ::NIHH, false);
1353 case SystemZ::OIFMux:
1354 expandRIPseudo(MI, SystemZ::OILF, SystemZ::OIHF, false);
1357 case SystemZ::OILMux:
1358 expandRIPseudo(MI, SystemZ::OILL, SystemZ::OIHL, false);
1361 case SystemZ::OIHMux:
1362 expandRIPseudo(MI, SystemZ::OILH, SystemZ::OIHH, false);
1365 case SystemZ::XIFMux:
1366 expandRIPseudo(MI, SystemZ::XILF, SystemZ::XIHF, false);
1369 case SystemZ::TMLMux:
1370 expandRIPseudo(MI, SystemZ::TMLL, SystemZ::TMHL, false);
1373 case SystemZ::TMHMux:
1374 expandRIPseudo(MI, SystemZ::TMLH, SystemZ::TMHH, false);
1377 case SystemZ::AHIMux:
1378 expandRIPseudo(MI, SystemZ::AHI, SystemZ::AIH, false);
1381 case SystemZ::AHIMuxK:
1382 expandRIEPseudo(MI, SystemZ::AHI, SystemZ::AHIK, SystemZ::AIH);
1385 case SystemZ::AFIMux:
1386 expandRIPseudo(MI, SystemZ::AFI, SystemZ::AIH, false);
1389 case SystemZ::CHIMux:
1390 expandRIPseudo(MI, SystemZ::CHI, SystemZ::CIH, false);
1393 case SystemZ::CFIMux:
1394 expandRIPseudo(MI, SystemZ::CFI, SystemZ::CIH, false);
1397 case SystemZ::CLFIMux:
1398 expandRIPseudo(MI, SystemZ::CLFI, SystemZ::CLIH, false);
1402 expandRXYPseudo(MI, SystemZ::C, SystemZ::CHF);
1405 case SystemZ::CLMux:
1406 expandRXYPseudo(MI, SystemZ::CL, SystemZ::CLHF);
1409 case SystemZ::RISBMux: {
1410 bool DestIsHigh = isHighReg(MI.getOperand(0).getReg());
1411 bool SrcIsHigh = isHighReg(MI.getOperand(2).getReg());
1412 if (SrcIsHigh == DestIsHigh)
1413 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHH : SystemZ::RISBLL));
1415 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHL : SystemZ::RISBLH));
1416 MI.getOperand(5).setImm(MI.getOperand(5).getImm() ^ 32);
1421 case SystemZ::ADJDYNALLOC:
1422 splitAdjDynAlloc(MI);
1425 case TargetOpcode::LOAD_STACK_GUARD:
1426 expandLoadStackGuard(&MI);
1434 unsigned SystemZInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
1435 if (MI.getOpcode() == TargetOpcode::INLINEASM) {
1436 const MachineFunction *MF = MI.getParent()->getParent();
1437 const char *AsmStr = MI.getOperand(0).getSymbolName();
1438 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
1440 return MI.getDesc().getSize();
1444 SystemZInstrInfo::getBranchInfo(const MachineInstr &MI) const {
1445 switch (MI.getOpcode()) {
1450 return SystemZII::Branch(SystemZII::BranchNormal, SystemZ::CCMASK_ANY,
1451 SystemZ::CCMASK_ANY, &MI.getOperand(0));
1455 return SystemZII::Branch(SystemZII::BranchNormal, MI.getOperand(0).getImm(),
1456 MI.getOperand(1).getImm(), &MI.getOperand(2));
1459 case SystemZ::BRCTH:
1460 return SystemZII::Branch(SystemZII::BranchCT, SystemZ::CCMASK_ICMP,
1461 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2));
1463 case SystemZ::BRCTG:
1464 return SystemZII::Branch(SystemZII::BranchCTG, SystemZ::CCMASK_ICMP,
1465 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2));
1469 return SystemZII::Branch(SystemZII::BranchC, SystemZ::CCMASK_ICMP,
1470 MI.getOperand(2).getImm(), &MI.getOperand(3));
1474 return SystemZII::Branch(SystemZII::BranchCL, SystemZ::CCMASK_ICMP,
1475 MI.getOperand(2).getImm(), &MI.getOperand(3));
1479 return SystemZII::Branch(SystemZII::BranchCG, SystemZ::CCMASK_ICMP,
1480 MI.getOperand(2).getImm(), &MI.getOperand(3));
1482 case SystemZ::CLGIJ:
1483 case SystemZ::CLGRJ:
1484 return SystemZII::Branch(SystemZII::BranchCLG, SystemZ::CCMASK_ICMP,
1485 MI.getOperand(2).getImm(), &MI.getOperand(3));
1488 llvm_unreachable("Unrecognized branch opcode");
1492 void SystemZInstrInfo::getLoadStoreOpcodes(const TargetRegisterClass *RC,
1493 unsigned &LoadOpcode,
1494 unsigned &StoreOpcode) const {
1495 if (RC == &SystemZ::GR32BitRegClass || RC == &SystemZ::ADDR32BitRegClass) {
1496 LoadOpcode = SystemZ::L;
1497 StoreOpcode = SystemZ::ST;
1498 } else if (RC == &SystemZ::GRH32BitRegClass) {
1499 LoadOpcode = SystemZ::LFH;
1500 StoreOpcode = SystemZ::STFH;
1501 } else if (RC == &SystemZ::GRX32BitRegClass) {
1502 LoadOpcode = SystemZ::LMux;
1503 StoreOpcode = SystemZ::STMux;
1504 } else if (RC == &SystemZ::GR64BitRegClass ||
1505 RC == &SystemZ::ADDR64BitRegClass) {
1506 LoadOpcode = SystemZ::LG;
1507 StoreOpcode = SystemZ::STG;
1508 } else if (RC == &SystemZ::GR128BitRegClass ||
1509 RC == &SystemZ::ADDR128BitRegClass) {
1510 LoadOpcode = SystemZ::L128;
1511 StoreOpcode = SystemZ::ST128;
1512 } else if (RC == &SystemZ::FP32BitRegClass) {
1513 LoadOpcode = SystemZ::LE;
1514 StoreOpcode = SystemZ::STE;
1515 } else if (RC == &SystemZ::FP64BitRegClass) {
1516 LoadOpcode = SystemZ::LD;
1517 StoreOpcode = SystemZ::STD;
1518 } else if (RC == &SystemZ::FP128BitRegClass) {
1519 LoadOpcode = SystemZ::LX;
1520 StoreOpcode = SystemZ::STX;
1521 } else if (RC == &SystemZ::VR32BitRegClass) {
1522 LoadOpcode = SystemZ::VL32;
1523 StoreOpcode = SystemZ::VST32;
1524 } else if (RC == &SystemZ::VR64BitRegClass) {
1525 LoadOpcode = SystemZ::VL64;
1526 StoreOpcode = SystemZ::VST64;
1527 } else if (RC == &SystemZ::VF128BitRegClass ||
1528 RC == &SystemZ::VR128BitRegClass) {
1529 LoadOpcode = SystemZ::VL;
1530 StoreOpcode = SystemZ::VST;
1532 llvm_unreachable("Unsupported regclass to load or store");
1535 unsigned SystemZInstrInfo::getOpcodeForOffset(unsigned Opcode,
1536 int64_t Offset) const {
1537 const MCInstrDesc &MCID = get(Opcode);
1538 int64_t Offset2 = (MCID.TSFlags & SystemZII::Is128Bit ? Offset + 8 : Offset);
1539 if (isUInt<12>(Offset) && isUInt<12>(Offset2)) {
1540 // Get the instruction to use for unsigned 12-bit displacements.
1541 int Disp12Opcode = SystemZ::getDisp12Opcode(Opcode);
1542 if (Disp12Opcode >= 0)
1543 return Disp12Opcode;
1545 // All address-related instructions can use unsigned 12-bit
1549 if (isInt<20>(Offset) && isInt<20>(Offset2)) {
1550 // Get the instruction to use for signed 20-bit displacements.
1551 int Disp20Opcode = SystemZ::getDisp20Opcode(Opcode);
1552 if (Disp20Opcode >= 0)
1553 return Disp20Opcode;
1555 // Check whether Opcode allows signed 20-bit displacements.
1556 if (MCID.TSFlags & SystemZII::Has20BitOffset)
1562 unsigned SystemZInstrInfo::getLoadAndTest(unsigned Opcode) const {
1564 case SystemZ::L: return SystemZ::LT;
1565 case SystemZ::LY: return SystemZ::LT;
1566 case SystemZ::LG: return SystemZ::LTG;
1567 case SystemZ::LGF: return SystemZ::LTGF;
1568 case SystemZ::LR: return SystemZ::LTR;
1569 case SystemZ::LGFR: return SystemZ::LTGFR;
1570 case SystemZ::LGR: return SystemZ::LTGR;
1571 case SystemZ::LER: return SystemZ::LTEBR;
1572 case SystemZ::LDR: return SystemZ::LTDBR;
1573 case SystemZ::LXR: return SystemZ::LTXBR;
1574 case SystemZ::LCDFR: return SystemZ::LCDBR;
1575 case SystemZ::LPDFR: return SystemZ::LPDBR;
1576 case SystemZ::LNDFR: return SystemZ::LNDBR;
1577 case SystemZ::LCDFR_32: return SystemZ::LCEBR;
1578 case SystemZ::LPDFR_32: return SystemZ::LPEBR;
1579 case SystemZ::LNDFR_32: return SystemZ::LNEBR;
1580 // On zEC12 we prefer to use RISBGN. But if there is a chance to
1581 // actually use the condition code, we may turn it back into RISGB.
1582 // Note that RISBG is not really a "load-and-test" instruction,
1583 // but sets the same condition code values, so is OK to use here.
1584 case SystemZ::RISBGN: return SystemZ::RISBG;
1589 // Return true if Mask matches the regexp 0*1+0*, given that zero masks
1590 // have already been filtered out. Store the first set bit in LSB and
1591 // the number of set bits in Length if so.
1592 static bool isStringOfOnes(uint64_t Mask, unsigned &LSB, unsigned &Length) {
1593 unsigned First = findFirstSet(Mask);
1594 uint64_t Top = (Mask >> First) + 1;
1595 if ((Top & -Top) == Top) {
1597 Length = findFirstSet(Top);
1603 bool SystemZInstrInfo::isRxSBGMask(uint64_t Mask, unsigned BitSize,
1604 unsigned &Start, unsigned &End) const {
1605 // Reject trivial all-zero masks.
1606 Mask &= allOnes(BitSize);
1610 // Handle the 1+0+ or 0+1+0* cases. Start then specifies the index of
1611 // the msb and End specifies the index of the lsb.
1612 unsigned LSB, Length;
1613 if (isStringOfOnes(Mask, LSB, Length)) {
1614 Start = 63 - (LSB + Length - 1);
1619 // Handle the wrap-around 1+0+1+ cases. Start then specifies the msb
1620 // of the low 1s and End specifies the lsb of the high 1s.
1621 if (isStringOfOnes(Mask ^ allOnes(BitSize), LSB, Length)) {
1622 assert(LSB > 0 && "Bottom bit must be set");
1623 assert(LSB + Length < BitSize && "Top bit must be set");
1624 Start = 63 - (LSB - 1);
1625 End = 63 - (LSB + Length);
1632 unsigned SystemZInstrInfo::getFusedCompare(unsigned Opcode,
1633 SystemZII::FusedCompareType Type,
1634 const MachineInstr *MI) const {
1638 if (!(MI && isInt<8>(MI->getOperand(1).getImm())))
1642 case SystemZ::CLGFI:
1643 if (!(MI && isUInt<8>(MI->getOperand(1).getImm())))
1648 if (!STI.hasMiscellaneousExtensions())
1650 if (!(MI && MI->getOperand(3).getReg() == 0))
1655 case SystemZII::CompareAndBranch:
1658 return SystemZ::CRJ;
1660 return SystemZ::CGRJ;
1662 return SystemZ::CIJ;
1664 return SystemZ::CGIJ;
1666 return SystemZ::CLRJ;
1668 return SystemZ::CLGRJ;
1670 return SystemZ::CLIJ;
1671 case SystemZ::CLGFI:
1672 return SystemZ::CLGIJ;
1676 case SystemZII::CompareAndReturn:
1679 return SystemZ::CRBReturn;
1681 return SystemZ::CGRBReturn;
1683 return SystemZ::CIBReturn;
1685 return SystemZ::CGIBReturn;
1687 return SystemZ::CLRBReturn;
1689 return SystemZ::CLGRBReturn;
1691 return SystemZ::CLIBReturn;
1692 case SystemZ::CLGFI:
1693 return SystemZ::CLGIBReturn;
1697 case SystemZII::CompareAndSibcall:
1700 return SystemZ::CRBCall;
1702 return SystemZ::CGRBCall;
1704 return SystemZ::CIBCall;
1706 return SystemZ::CGIBCall;
1708 return SystemZ::CLRBCall;
1710 return SystemZ::CLGRBCall;
1712 return SystemZ::CLIBCall;
1713 case SystemZ::CLGFI:
1714 return SystemZ::CLGIBCall;
1718 case SystemZII::CompareAndTrap:
1721 return SystemZ::CRT;
1723 return SystemZ::CGRT;
1725 return SystemZ::CIT;
1727 return SystemZ::CGIT;
1729 return SystemZ::CLRT;
1731 return SystemZ::CLGRT;
1733 return SystemZ::CLFIT;
1734 case SystemZ::CLGFI:
1735 return SystemZ::CLGIT;
1737 return SystemZ::CLT;
1739 return SystemZ::CLGT;
1747 unsigned SystemZInstrInfo::getLoadAndTrap(unsigned Opcode) const {
1748 if (!STI.hasLoadAndTrap())
1753 return SystemZ::LAT;
1755 return SystemZ::LGAT;
1757 return SystemZ::LFHAT;
1759 return SystemZ::LLGFAT;
1761 return SystemZ::LLGTAT;
1766 void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB,
1767 MachineBasicBlock::iterator MBBI,
1768 unsigned Reg, uint64_t Value) const {
1769 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
1771 if (isInt<16>(Value))
1772 Opcode = SystemZ::LGHI;
1773 else if (SystemZ::isImmLL(Value))
1774 Opcode = SystemZ::LLILL;
1775 else if (SystemZ::isImmLH(Value)) {
1776 Opcode = SystemZ::LLILH;
1779 assert(isInt<32>(Value) && "Huge values not handled yet");
1780 Opcode = SystemZ::LGFI;
1782 BuildMI(MBB, MBBI, DL, get(Opcode), Reg).addImm(Value);
1785 bool SystemZInstrInfo::
1786 areMemAccessesTriviallyDisjoint(MachineInstr &MIa, MachineInstr &MIb,
1787 AliasAnalysis *AA) const {
1789 if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand())
1792 // If mem-operands show that the same address Value is used by both
1793 // instructions, check for non-overlapping offsets and widths. Not
1794 // sure if a register based analysis would be an improvement...
1796 MachineMemOperand *MMOa = *MIa.memoperands_begin();
1797 MachineMemOperand *MMOb = *MIb.memoperands_begin();
1798 const Value *VALa = MMOa->getValue();
1799 const Value *VALb = MMOb->getValue();
1800 bool SameVal = (VALa && VALb && (VALa == VALb));
1802 const PseudoSourceValue *PSVa = MMOa->getPseudoValue();
1803 const PseudoSourceValue *PSVb = MMOb->getPseudoValue();
1804 if (PSVa && PSVb && (PSVa == PSVb))
1808 int OffsetA = MMOa->getOffset(), OffsetB = MMOb->getOffset();
1809 int WidthA = MMOa->getSize(), WidthB = MMOb->getSize();
1810 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB;
1811 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA;
1812 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
1813 if (LowOffset + LowWidth <= HighOffset)