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 "MCTargetDesc/SystemZMCTargetDesc.h"
16 #include "SystemZInstrBuilder.h"
17 #include "SystemZInstrInfo.h"
18 #include "SystemZSubtarget.h"
19 #include "llvm/CodeGen/LiveInterval.h"
20 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
21 #include "llvm/CodeGen/LiveVariables.h"
22 #include "llvm/CodeGen/MachineBasicBlock.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineInstr.h"
26 #include "llvm/CodeGen/MachineMemOperand.h"
27 #include "llvm/CodeGen/MachineOperand.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/SlotIndexes.h"
30 #include "llvm/MC/MCInstrDesc.h"
31 #include "llvm/MC/MCRegisterInfo.h"
32 #include "llvm/Support/BranchProbability.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Target/TargetInstrInfo.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetSubtargetInfo.h"
44 #define GET_INSTRINFO_CTOR_DTOR
45 #define GET_INSTRMAP_INFO
46 #include "SystemZGenInstrInfo.inc"
48 // Return a mask with Count low bits set.
49 static uint64_t allOnes(unsigned int Count) {
50 return Count == 0 ? 0 : (uint64_t(1) << (Count - 1) << 1) - 1;
53 // Reg should be a 32-bit GPR. Return true if it is a high register rather
54 // than a low register.
55 static bool isHighReg(unsigned int Reg) {
56 if (SystemZ::GRH32BitRegClass.contains(Reg))
58 assert(SystemZ::GR32BitRegClass.contains(Reg) && "Invalid GRX32");
62 // Pin the vtable to this file.
63 void SystemZInstrInfo::anchor() {}
65 SystemZInstrInfo::SystemZInstrInfo(SystemZSubtarget &sti)
66 : SystemZGenInstrInfo(SystemZ::ADJCALLSTACKDOWN, SystemZ::ADJCALLSTACKUP),
70 // MI is a 128-bit load or store. Split it into two 64-bit loads or stores,
71 // each having the opcode given by NewOpcode.
72 void SystemZInstrInfo::splitMove(MachineBasicBlock::iterator MI,
73 unsigned NewOpcode) const {
74 MachineBasicBlock *MBB = MI->getParent();
75 MachineFunction &MF = *MBB->getParent();
77 // Get two load or store instructions. Use the original instruction for one
78 // of them (arbitrarily the second here) and create a clone for the other.
79 MachineInstr *EarlierMI = MF.CloneMachineInstr(&*MI);
80 MBB->insert(MI, EarlierMI);
82 // Set up the two 64-bit registers and remember super reg and its flags.
83 MachineOperand &HighRegOp = EarlierMI->getOperand(0);
84 MachineOperand &LowRegOp = MI->getOperand(0);
85 unsigned Reg128 = LowRegOp.getReg();
86 unsigned Reg128Killed = getKillRegState(LowRegOp.isKill());
87 unsigned Reg128Undef = getUndefRegState(LowRegOp.isUndef());
88 HighRegOp.setReg(RI.getSubReg(HighRegOp.getReg(), SystemZ::subreg_h64));
89 LowRegOp.setReg(RI.getSubReg(LowRegOp.getReg(), SystemZ::subreg_l64));
92 // Add implicit uses of the super register in case one of the subregs is
93 // undefined. We could track liveness and skip storing an undefined
94 // subreg, but this is hopefully rare (discovered with llvm-stress).
95 // If Reg128 was killed, set kill flag on MI.
96 unsigned Reg128UndefImpl = (Reg128Undef | RegState::Implicit);
97 MachineInstrBuilder(MF, EarlierMI).addReg(Reg128, Reg128UndefImpl);
98 MachineInstrBuilder(MF, MI).addReg(Reg128, (Reg128UndefImpl | Reg128Killed));
101 // The address in the first (high) instruction is already correct.
102 // Adjust the offset in the second (low) instruction.
103 MachineOperand &HighOffsetOp = EarlierMI->getOperand(2);
104 MachineOperand &LowOffsetOp = MI->getOperand(2);
105 LowOffsetOp.setImm(LowOffsetOp.getImm() + 8);
107 // Clear the kill flags on the registers in the first instruction.
108 if (EarlierMI->getOperand(0).isReg() && EarlierMI->getOperand(0).isUse())
109 EarlierMI->getOperand(0).setIsKill(false);
110 EarlierMI->getOperand(1).setIsKill(false);
111 EarlierMI->getOperand(3).setIsKill(false);
114 unsigned HighOpcode = getOpcodeForOffset(NewOpcode, HighOffsetOp.getImm());
115 unsigned LowOpcode = getOpcodeForOffset(NewOpcode, LowOffsetOp.getImm());
116 assert(HighOpcode && LowOpcode && "Both offsets should be in range");
118 EarlierMI->setDesc(get(HighOpcode));
119 MI->setDesc(get(LowOpcode));
122 // Split ADJDYNALLOC instruction MI.
123 void SystemZInstrInfo::splitAdjDynAlloc(MachineBasicBlock::iterator MI) const {
124 MachineBasicBlock *MBB = MI->getParent();
125 MachineFunction &MF = *MBB->getParent();
126 MachineFrameInfo &MFFrame = MF.getFrameInfo();
127 MachineOperand &OffsetMO = MI->getOperand(2);
129 uint64_t Offset = (MFFrame.getMaxCallFrameSize() +
130 SystemZMC::CallFrameSize +
132 unsigned NewOpcode = getOpcodeForOffset(SystemZ::LA, Offset);
133 assert(NewOpcode && "No support for huge argument lists yet");
134 MI->setDesc(get(NewOpcode));
135 OffsetMO.setImm(Offset);
138 // MI is an RI-style pseudo instruction. Replace it with LowOpcode
139 // if the first operand is a low GR32 and HighOpcode if the first operand
140 // is a high GR32. ConvertHigh is true if LowOpcode takes a signed operand
141 // and HighOpcode takes an unsigned 32-bit operand. In those cases,
142 // MI has the same kind of operand as LowOpcode, so needs to be converted
143 // if HighOpcode is used.
144 void SystemZInstrInfo::expandRIPseudo(MachineInstr &MI, unsigned LowOpcode,
146 bool ConvertHigh) const {
147 unsigned Reg = MI.getOperand(0).getReg();
148 bool IsHigh = isHighReg(Reg);
149 MI.setDesc(get(IsHigh ? HighOpcode : LowOpcode));
150 if (IsHigh && ConvertHigh)
151 MI.getOperand(1).setImm(uint32_t(MI.getOperand(1).getImm()));
154 // MI is a three-operand RIE-style pseudo instruction. Replace it with
155 // LowOpcodeK if the registers are both low GR32s, otherwise use a move
156 // followed by HighOpcode or LowOpcode, depending on whether the target
157 // is a high or low GR32.
158 void SystemZInstrInfo::expandRIEPseudo(MachineInstr &MI, unsigned LowOpcode,
160 unsigned HighOpcode) const {
161 unsigned DestReg = MI.getOperand(0).getReg();
162 unsigned SrcReg = MI.getOperand(1).getReg();
163 bool DestIsHigh = isHighReg(DestReg);
164 bool SrcIsHigh = isHighReg(SrcReg);
165 if (!DestIsHigh && !SrcIsHigh)
166 MI.setDesc(get(LowOpcodeK));
168 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), DestReg, SrcReg,
169 SystemZ::LR, 32, MI.getOperand(1).isKill(),
170 MI.getOperand(1).isUndef());
171 MI.setDesc(get(DestIsHigh ? HighOpcode : LowOpcode));
172 MI.getOperand(1).setReg(DestReg);
173 MI.tieOperands(0, 1);
177 // MI is an RXY-style pseudo instruction. Replace it with LowOpcode
178 // if the first operand is a low GR32 and HighOpcode if the first operand
180 void SystemZInstrInfo::expandRXYPseudo(MachineInstr &MI, unsigned LowOpcode,
181 unsigned HighOpcode) const {
182 unsigned Reg = MI.getOperand(0).getReg();
183 unsigned Opcode = getOpcodeForOffset(isHighReg(Reg) ? HighOpcode : LowOpcode,
184 MI.getOperand(2).getImm());
185 MI.setDesc(get(Opcode));
188 // MI is a load-on-condition pseudo instruction with a single register
189 // (source or destination) operand. Replace it with LowOpcode if the
190 // register is a low GR32 and HighOpcode if the register is a high GR32.
191 void SystemZInstrInfo::expandLOCPseudo(MachineInstr &MI, unsigned LowOpcode,
192 unsigned HighOpcode) const {
193 unsigned Reg = MI.getOperand(0).getReg();
194 unsigned Opcode = isHighReg(Reg) ? HighOpcode : LowOpcode;
195 MI.setDesc(get(Opcode));
198 // MI is a load-register-on-condition pseudo instruction. Replace it with
199 // LowOpcode if source and destination are both low GR32s and HighOpcode if
200 // source and destination are both high GR32s.
201 void SystemZInstrInfo::expandLOCRPseudo(MachineInstr &MI, unsigned LowOpcode,
202 unsigned HighOpcode) const {
203 unsigned DestReg = MI.getOperand(0).getReg();
204 unsigned SrcReg = MI.getOperand(2).getReg();
205 bool DestIsHigh = isHighReg(DestReg);
206 bool SrcIsHigh = isHighReg(SrcReg);
208 if (!DestIsHigh && !SrcIsHigh)
209 MI.setDesc(get(LowOpcode));
210 else if (DestIsHigh && SrcIsHigh)
211 MI.setDesc(get(HighOpcode));
213 // If we were unable to implement the pseudo with a single instruction, we
214 // need to convert it back into a branch sequence. This cannot be done here
215 // since the caller of expandPostRAPseudo does not handle changes to the CFG
216 // correctly. This change is defered to the SystemZExpandPseudo pass.
219 // MI is an RR-style pseudo instruction that zero-extends the low Size bits
220 // of one GRX32 into another. Replace it with LowOpcode if both operands
221 // are low registers, otherwise use RISB[LH]G.
222 void SystemZInstrInfo::expandZExtPseudo(MachineInstr &MI, unsigned LowOpcode,
223 unsigned Size) const {
224 MachineInstrBuilder MIB =
225 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(),
226 MI.getOperand(0).getReg(), MI.getOperand(1).getReg(), LowOpcode,
227 Size, MI.getOperand(1).isKill(), MI.getOperand(1).isUndef());
229 // Keep the remaining operands as-is.
230 for (unsigned I = 2; I < MI.getNumOperands(); ++I)
231 MIB.add(MI.getOperand(I));
233 MI.eraseFromParent();
236 void SystemZInstrInfo::expandLoadStackGuard(MachineInstr *MI) const {
237 MachineBasicBlock *MBB = MI->getParent();
238 MachineFunction &MF = *MBB->getParent();
239 const unsigned Reg = MI->getOperand(0).getReg();
241 // Conveniently, all 4 instructions are cloned from LOAD_STACK_GUARD,
242 // so they already have operand 0 set to reg.
245 MachineInstr *Ear1MI = MF.CloneMachineInstr(MI);
246 MBB->insert(MI, Ear1MI);
247 Ear1MI->setDesc(get(SystemZ::EAR));
248 MachineInstrBuilder(MF, Ear1MI).addReg(SystemZ::A0);
250 // sllg <reg>, <reg>, 32
251 MachineInstr *SllgMI = MF.CloneMachineInstr(MI);
252 MBB->insert(MI, SllgMI);
253 SllgMI->setDesc(get(SystemZ::SLLG));
254 MachineInstrBuilder(MF, SllgMI).addReg(Reg).addReg(0).addImm(32);
257 MachineInstr *Ear2MI = MF.CloneMachineInstr(MI);
258 MBB->insert(MI, Ear2MI);
259 Ear2MI->setDesc(get(SystemZ::EAR));
260 MachineInstrBuilder(MF, Ear2MI).addReg(SystemZ::A1);
262 // lg <reg>, 40(<reg>)
263 MI->setDesc(get(SystemZ::LG));
264 MachineInstrBuilder(MF, MI).addReg(Reg).addImm(40).addReg(0);
267 // Emit a zero-extending move from 32-bit GPR SrcReg to 32-bit GPR
268 // DestReg before MBBI in MBB. Use LowLowOpcode when both DestReg and SrcReg
269 // are low registers, otherwise use RISB[LH]G. Size is the number of bits
270 // taken from the low end of SrcReg (8 for LLCR, 16 for LLHR and 32 for LR).
271 // KillSrc is true if this move is the last use of SrcReg.
273 SystemZInstrInfo::emitGRX32Move(MachineBasicBlock &MBB,
274 MachineBasicBlock::iterator MBBI,
275 const DebugLoc &DL, unsigned DestReg,
276 unsigned SrcReg, unsigned LowLowOpcode,
277 unsigned Size, bool KillSrc,
278 bool UndefSrc) const {
280 bool DestIsHigh = isHighReg(DestReg);
281 bool SrcIsHigh = isHighReg(SrcReg);
282 if (DestIsHigh && SrcIsHigh)
283 Opcode = SystemZ::RISBHH;
284 else if (DestIsHigh && !SrcIsHigh)
285 Opcode = SystemZ::RISBHL;
286 else if (!DestIsHigh && SrcIsHigh)
287 Opcode = SystemZ::RISBLH;
289 return BuildMI(MBB, MBBI, DL, get(LowLowOpcode), DestReg)
290 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc));
292 unsigned Rotate = (DestIsHigh != SrcIsHigh ? 32 : 0);
293 return BuildMI(MBB, MBBI, DL, get(Opcode), DestReg)
294 .addReg(DestReg, RegState::Undef)
295 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc))
296 .addImm(32 - Size).addImm(128 + 31).addImm(Rotate);
299 MachineInstr *SystemZInstrInfo::commuteInstructionImpl(MachineInstr &MI,
302 unsigned OpIdx2) const {
303 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & {
305 return *MI.getParent()->getParent()->CloneMachineInstr(&MI);
309 switch (MI.getOpcode()) {
310 case SystemZ::LOCRMux:
311 case SystemZ::LOCFHR:
313 case SystemZ::LOCGR: {
314 auto &WorkingMI = cloneIfNew(MI);
316 unsigned CCValid = WorkingMI.getOperand(3).getImm();
317 unsigned CCMask = WorkingMI.getOperand(4).getImm();
318 WorkingMI.getOperand(4).setImm(CCMask ^ CCValid);
319 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
323 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
327 // If MI is a simple load or store for a frame object, return the register
328 // it loads or stores and set FrameIndex to the index of the frame object.
329 // Return 0 otherwise.
331 // Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores.
332 static int isSimpleMove(const MachineInstr &MI, int &FrameIndex,
334 const MCInstrDesc &MCID = MI.getDesc();
335 if ((MCID.TSFlags & Flag) && MI.getOperand(1).isFI() &&
336 MI.getOperand(2).getImm() == 0 && MI.getOperand(3).getReg() == 0) {
337 FrameIndex = MI.getOperand(1).getIndex();
338 return MI.getOperand(0).getReg();
343 unsigned SystemZInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
344 int &FrameIndex) const {
345 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXLoad);
348 unsigned SystemZInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
349 int &FrameIndex) const {
350 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXStore);
353 bool SystemZInstrInfo::isStackSlotCopy(const MachineInstr &MI,
355 int &SrcFrameIndex) const {
356 // Check for MVC 0(Length,FI1),0(FI2)
357 const MachineFrameInfo &MFI = MI.getParent()->getParent()->getFrameInfo();
358 if (MI.getOpcode() != SystemZ::MVC || !MI.getOperand(0).isFI() ||
359 MI.getOperand(1).getImm() != 0 || !MI.getOperand(3).isFI() ||
360 MI.getOperand(4).getImm() != 0)
363 // Check that Length covers the full slots.
364 int64_t Length = MI.getOperand(2).getImm();
365 unsigned FI1 = MI.getOperand(0).getIndex();
366 unsigned FI2 = MI.getOperand(3).getIndex();
367 if (MFI.getObjectSize(FI1) != Length ||
368 MFI.getObjectSize(FI2) != Length)
371 DestFrameIndex = FI1;
376 bool SystemZInstrInfo::analyzeBranch(MachineBasicBlock &MBB,
377 MachineBasicBlock *&TBB,
378 MachineBasicBlock *&FBB,
379 SmallVectorImpl<MachineOperand> &Cond,
380 bool AllowModify) const {
381 // Most of the code and comments here are boilerplate.
383 // Start from the bottom of the block and work up, examining the
384 // terminator instructions.
385 MachineBasicBlock::iterator I = MBB.end();
386 while (I != MBB.begin()) {
388 if (I->isDebugValue())
391 // Working from the bottom, when we see a non-terminator instruction, we're
393 if (!isUnpredicatedTerminator(*I))
396 // A terminator that isn't a branch can't easily be handled by this
401 // Can't handle indirect branches.
402 SystemZII::Branch Branch(getBranchInfo(*I));
403 if (!Branch.Target->isMBB())
406 // Punt on compound branches.
407 if (Branch.Type != SystemZII::BranchNormal)
410 if (Branch.CCMask == SystemZ::CCMASK_ANY) {
411 // Handle unconditional branches.
413 TBB = Branch.Target->getMBB();
417 // If the block has any instructions after a JMP, delete them.
418 while (std::next(I) != MBB.end())
419 std::next(I)->eraseFromParent();
424 // Delete the JMP if it's equivalent to a fall-through.
425 if (MBB.isLayoutSuccessor(Branch.Target->getMBB())) {
427 I->eraseFromParent();
432 // TBB is used to indicate the unconditinal destination.
433 TBB = Branch.Target->getMBB();
437 // Working from the bottom, handle the first conditional branch.
439 // FIXME: add X86-style branch swap
441 TBB = Branch.Target->getMBB();
442 Cond.push_back(MachineOperand::CreateImm(Branch.CCValid));
443 Cond.push_back(MachineOperand::CreateImm(Branch.CCMask));
447 // Handle subsequent conditional branches.
448 assert(Cond.size() == 2 && TBB && "Should have seen a conditional branch");
450 // Only handle the case where all conditional branches branch to the same
452 if (TBB != Branch.Target->getMBB())
455 // If the conditions are the same, we can leave them alone.
456 unsigned OldCCValid = Cond[0].getImm();
457 unsigned OldCCMask = Cond[1].getImm();
458 if (OldCCValid == Branch.CCValid && OldCCMask == Branch.CCMask)
461 // FIXME: Try combining conditions like X86 does. Should be easy on Z!
468 unsigned SystemZInstrInfo::removeBranch(MachineBasicBlock &MBB,
469 int *BytesRemoved) const {
470 assert(!BytesRemoved && "code size not handled");
472 // Most of the code and comments here are boilerplate.
473 MachineBasicBlock::iterator I = MBB.end();
476 while (I != MBB.begin()) {
478 if (I->isDebugValue())
482 if (!getBranchInfo(*I).Target->isMBB())
484 // Remove the branch.
485 I->eraseFromParent();
493 bool SystemZInstrInfo::
494 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
495 assert(Cond.size() == 2 && "Invalid condition");
496 Cond[1].setImm(Cond[1].getImm() ^ Cond[0].getImm());
500 unsigned SystemZInstrInfo::insertBranch(MachineBasicBlock &MBB,
501 MachineBasicBlock *TBB,
502 MachineBasicBlock *FBB,
503 ArrayRef<MachineOperand> Cond,
505 int *BytesAdded) const {
506 // In this function we output 32-bit branches, which should always
507 // have enough range. They can be shortened and relaxed by later code
508 // in the pipeline, if desired.
510 // Shouldn't be a fall through.
511 assert(TBB && "insertBranch must not be told to insert a fallthrough");
512 assert((Cond.size() == 2 || Cond.size() == 0) &&
513 "SystemZ branch conditions have one component!");
514 assert(!BytesAdded && "code size not handled");
517 // Unconditional branch?
518 assert(!FBB && "Unconditional branch with multiple successors!");
519 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(TBB);
523 // Conditional branch.
525 unsigned CCValid = Cond[0].getImm();
526 unsigned CCMask = Cond[1].getImm();
527 BuildMI(&MBB, DL, get(SystemZ::BRC))
528 .addImm(CCValid).addImm(CCMask).addMBB(TBB);
532 // Two-way Conditional branch. Insert the second branch.
533 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(FBB);
539 bool SystemZInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
540 unsigned &SrcReg2, int &Mask,
542 assert(MI.isCompare() && "Caller should have checked for a comparison");
544 if (MI.getNumExplicitOperands() == 2 && MI.getOperand(0).isReg() &&
545 MI.getOperand(1).isImm()) {
546 SrcReg = MI.getOperand(0).getReg();
548 Value = MI.getOperand(1).getImm();
556 // If Reg is a virtual register, return its definition, otherwise return null.
557 static MachineInstr *getDef(unsigned Reg,
558 const MachineRegisterInfo *MRI) {
559 if (TargetRegisterInfo::isPhysicalRegister(Reg))
561 return MRI->getUniqueVRegDef(Reg);
564 // Return true if MI is a shift of type Opcode by Imm bits.
565 static bool isShift(MachineInstr *MI, unsigned Opcode, int64_t Imm) {
566 return (MI->getOpcode() == Opcode &&
567 !MI->getOperand(2).getReg() &&
568 MI->getOperand(3).getImm() == Imm);
571 // If the destination of MI has no uses, delete it as dead.
572 static void eraseIfDead(MachineInstr *MI, const MachineRegisterInfo *MRI) {
573 if (MRI->use_nodbg_empty(MI->getOperand(0).getReg()))
574 MI->eraseFromParent();
577 // Compare compares SrcReg against zero. Check whether SrcReg contains
578 // the result of an IPM sequence whose input CC survives until Compare,
579 // and whether Compare is therefore redundant. Delete it and return
581 static bool removeIPMBasedCompare(MachineInstr &Compare, unsigned SrcReg,
582 const MachineRegisterInfo *MRI,
583 const TargetRegisterInfo *TRI) {
584 MachineInstr *LGFR = nullptr;
585 MachineInstr *RLL = getDef(SrcReg, MRI);
586 if (RLL && RLL->getOpcode() == SystemZ::LGFR) {
588 RLL = getDef(LGFR->getOperand(1).getReg(), MRI);
590 if (!RLL || !isShift(RLL, SystemZ::RLL, 31))
593 MachineInstr *SRL = getDef(RLL->getOperand(1).getReg(), MRI);
594 if (!SRL || !isShift(SRL, SystemZ::SRL, SystemZ::IPM_CC))
597 MachineInstr *IPM = getDef(SRL->getOperand(1).getReg(), MRI);
598 if (!IPM || IPM->getOpcode() != SystemZ::IPM)
601 // Check that there are no assignments to CC between the IPM and Compare,
602 if (IPM->getParent() != Compare.getParent())
604 MachineBasicBlock::iterator MBBI = IPM, MBBE = Compare.getIterator();
605 for (++MBBI; MBBI != MBBE; ++MBBI) {
606 MachineInstr &MI = *MBBI;
607 if (MI.modifiesRegister(SystemZ::CC, TRI))
611 Compare.eraseFromParent();
613 eraseIfDead(LGFR, MRI);
614 eraseIfDead(RLL, MRI);
615 eraseIfDead(SRL, MRI);
616 eraseIfDead(IPM, MRI);
621 bool SystemZInstrInfo::optimizeCompareInstr(
622 MachineInstr &Compare, unsigned SrcReg, unsigned SrcReg2, int Mask,
623 int Value, const MachineRegisterInfo *MRI) const {
624 assert(!SrcReg2 && "Only optimizing constant comparisons so far");
625 bool IsLogical = (Compare.getDesc().TSFlags & SystemZII::IsLogical) != 0;
626 return Value == 0 && !IsLogical &&
627 removeIPMBasedCompare(Compare, SrcReg, MRI, &RI);
630 bool SystemZInstrInfo::canInsertSelect(const MachineBasicBlock &MBB,
631 ArrayRef<MachineOperand> Pred,
632 unsigned TrueReg, unsigned FalseReg,
633 int &CondCycles, int &TrueCycles,
634 int &FalseCycles) const {
635 // Not all subtargets have LOCR instructions.
636 if (!STI.hasLoadStoreOnCond())
638 if (Pred.size() != 2)
641 // Check register classes.
642 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
643 const TargetRegisterClass *RC =
644 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
648 // We have LOCR instructions for 32 and 64 bit general purpose registers.
649 if ((STI.hasLoadStoreOnCond2() &&
650 SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) ||
651 SystemZ::GR32BitRegClass.hasSubClassEq(RC) ||
652 SystemZ::GR64BitRegClass.hasSubClassEq(RC)) {
659 // Can't do anything else.
663 void SystemZInstrInfo::insertSelect(MachineBasicBlock &MBB,
664 MachineBasicBlock::iterator I,
665 const DebugLoc &DL, unsigned DstReg,
666 ArrayRef<MachineOperand> Pred,
668 unsigned FalseReg) const {
669 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
670 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
672 assert(Pred.size() == 2 && "Invalid condition");
673 unsigned CCValid = Pred[0].getImm();
674 unsigned CCMask = Pred[1].getImm();
677 if (SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) {
678 if (STI.hasLoadStoreOnCond2())
679 Opc = SystemZ::LOCRMux;
682 MRI.constrainRegClass(DstReg, &SystemZ::GR32BitRegClass);
683 unsigned TReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass);
684 unsigned FReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass);
685 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), TReg).addReg(TrueReg);
686 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), FReg).addReg(FalseReg);
690 } else if (SystemZ::GR64BitRegClass.hasSubClassEq(RC))
691 Opc = SystemZ::LOCGR;
693 llvm_unreachable("Invalid register class");
695 BuildMI(MBB, I, DL, get(Opc), DstReg)
696 .addReg(FalseReg).addReg(TrueReg)
697 .addImm(CCValid).addImm(CCMask);
700 bool SystemZInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
702 MachineRegisterInfo *MRI) const {
703 unsigned DefOpc = DefMI.getOpcode();
704 if (DefOpc != SystemZ::LHIMux && DefOpc != SystemZ::LHI &&
705 DefOpc != SystemZ::LGHI)
707 if (DefMI.getOperand(0).getReg() != Reg)
709 int32_t ImmVal = (int32_t)DefMI.getOperand(1).getImm();
711 unsigned UseOpc = UseMI.getOpcode();
716 case SystemZ::LOCRMux:
717 if (!STI.hasLoadStoreOnCond2())
719 NewUseOpc = SystemZ::LOCHIMux;
720 if (UseMI.getOperand(2).getReg() == Reg)
722 else if (UseMI.getOperand(1).getReg() == Reg)
723 UseIdx = 2, CommuteIdx = 1;
728 if (!STI.hasLoadStoreOnCond2())
730 NewUseOpc = SystemZ::LOCGHI;
731 if (UseMI.getOperand(2).getReg() == Reg)
733 else if (UseMI.getOperand(1).getReg() == Reg)
734 UseIdx = 2, CommuteIdx = 1;
742 if (CommuteIdx != -1)
743 if (!commuteInstruction(UseMI, false, CommuteIdx, UseIdx))
746 bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
747 UseMI.setDesc(get(NewUseOpc));
748 UseMI.getOperand(UseIdx).ChangeToImmediate(ImmVal);
750 DefMI.eraseFromParent();
755 bool SystemZInstrInfo::isPredicable(const MachineInstr &MI) const {
756 unsigned Opcode = MI.getOpcode();
757 if (Opcode == SystemZ::Return ||
758 Opcode == SystemZ::Trap ||
759 Opcode == SystemZ::CallJG ||
760 Opcode == SystemZ::CallBR)
765 bool SystemZInstrInfo::
766 isProfitableToIfCvt(MachineBasicBlock &MBB,
767 unsigned NumCycles, unsigned ExtraPredCycles,
768 BranchProbability Probability) const {
769 // Avoid using conditional returns at the end of a loop (since then
770 // we'd need to emit an unconditional branch to the beginning anyway,
771 // making the loop body longer). This doesn't apply for low-probability
772 // loops (eg. compare-and-swap retry), so just decide based on branch
773 // probability instead of looping structure.
774 // However, since Compare and Trap instructions cost the same as a regular
775 // Compare instruction, we should allow the if conversion to convert this
776 // into a Conditional Compare regardless of the branch probability.
777 if (MBB.getLastNonDebugInstr()->getOpcode() != SystemZ::Trap &&
778 MBB.succ_empty() && Probability < BranchProbability(1, 8))
780 // For now only convert single instructions.
781 return NumCycles == 1;
784 bool SystemZInstrInfo::
785 isProfitableToIfCvt(MachineBasicBlock &TMBB,
786 unsigned NumCyclesT, unsigned ExtraPredCyclesT,
787 MachineBasicBlock &FMBB,
788 unsigned NumCyclesF, unsigned ExtraPredCyclesF,
789 BranchProbability Probability) const {
790 // For now avoid converting mutually-exclusive cases.
794 bool SystemZInstrInfo::
795 isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
796 BranchProbability Probability) const {
797 // For now only duplicate single instructions.
798 return NumCycles == 1;
801 bool SystemZInstrInfo::PredicateInstruction(
802 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const {
803 assert(Pred.size() == 2 && "Invalid condition");
804 unsigned CCValid = Pred[0].getImm();
805 unsigned CCMask = Pred[1].getImm();
806 assert(CCMask > 0 && CCMask < 15 && "Invalid predicate");
807 unsigned Opcode = MI.getOpcode();
808 if (Opcode == SystemZ::Trap) {
809 MI.setDesc(get(SystemZ::CondTrap));
810 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
811 .addImm(CCValid).addImm(CCMask)
812 .addReg(SystemZ::CC, RegState::Implicit);
815 if (Opcode == SystemZ::Return) {
816 MI.setDesc(get(SystemZ::CondReturn));
817 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
818 .addImm(CCValid).addImm(CCMask)
819 .addReg(SystemZ::CC, RegState::Implicit);
822 if (Opcode == SystemZ::CallJG) {
823 MachineOperand FirstOp = MI.getOperand(0);
824 const uint32_t *RegMask = MI.getOperand(1).getRegMask();
827 MI.setDesc(get(SystemZ::CallBRCL));
828 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
833 .addReg(SystemZ::CC, RegState::Implicit);
836 if (Opcode == SystemZ::CallBR) {
837 const uint32_t *RegMask = MI.getOperand(0).getRegMask();
839 MI.setDesc(get(SystemZ::CallBCR));
840 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
841 .addImm(CCValid).addImm(CCMask)
843 .addReg(SystemZ::CC, RegState::Implicit);
849 void SystemZInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
850 MachineBasicBlock::iterator MBBI,
851 const DebugLoc &DL, unsigned DestReg,
852 unsigned SrcReg, bool KillSrc) const {
853 // Split 128-bit GPR moves into two 64-bit moves. Add implicit uses of the
854 // super register in case one of the subregs is undefined.
855 // This handles ADDR128 too.
856 if (SystemZ::GR128BitRegClass.contains(DestReg, SrcReg)) {
857 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_h64),
858 RI.getSubReg(SrcReg, SystemZ::subreg_h64), KillSrc);
859 MachineInstrBuilder(*MBB.getParent(), std::prev(MBBI))
860 .addReg(SrcReg, RegState::Implicit);
861 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_l64),
862 RI.getSubReg(SrcReg, SystemZ::subreg_l64), KillSrc);
863 MachineInstrBuilder(*MBB.getParent(), std::prev(MBBI))
864 .addReg(SrcReg, (getKillRegState(KillSrc) | RegState::Implicit));
868 if (SystemZ::GRX32BitRegClass.contains(DestReg, SrcReg)) {
869 emitGRX32Move(MBB, MBBI, DL, DestReg, SrcReg, SystemZ::LR, 32, KillSrc,
874 // Everything else needs only one instruction.
876 if (SystemZ::GR64BitRegClass.contains(DestReg, SrcReg))
877 Opcode = SystemZ::LGR;
878 else if (SystemZ::FP32BitRegClass.contains(DestReg, SrcReg))
879 // For z13 we prefer LDR over LER to avoid partial register dependencies.
880 Opcode = STI.hasVector() ? SystemZ::LDR32 : SystemZ::LER;
881 else if (SystemZ::FP64BitRegClass.contains(DestReg, SrcReg))
882 Opcode = SystemZ::LDR;
883 else if (SystemZ::FP128BitRegClass.contains(DestReg, SrcReg))
884 Opcode = SystemZ::LXR;
885 else if (SystemZ::VR32BitRegClass.contains(DestReg, SrcReg))
886 Opcode = SystemZ::VLR32;
887 else if (SystemZ::VR64BitRegClass.contains(DestReg, SrcReg))
888 Opcode = SystemZ::VLR64;
889 else if (SystemZ::VR128BitRegClass.contains(DestReg, SrcReg))
890 Opcode = SystemZ::VLR;
891 else if (SystemZ::AR32BitRegClass.contains(DestReg, SrcReg))
892 Opcode = SystemZ::CPYA;
893 else if (SystemZ::AR32BitRegClass.contains(DestReg) &&
894 SystemZ::GR32BitRegClass.contains(SrcReg))
895 Opcode = SystemZ::SAR;
896 else if (SystemZ::GR32BitRegClass.contains(DestReg) &&
897 SystemZ::AR32BitRegClass.contains(SrcReg))
898 Opcode = SystemZ::EAR;
900 llvm_unreachable("Impossible reg-to-reg copy");
902 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg)
903 .addReg(SrcReg, getKillRegState(KillSrc));
906 void SystemZInstrInfo::storeRegToStackSlot(
907 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned SrcReg,
908 bool isKill, int FrameIdx, const TargetRegisterClass *RC,
909 const TargetRegisterInfo *TRI) const {
910 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
912 // Callers may expect a single instruction, so keep 128-bit moves
913 // together for now and lower them after register allocation.
914 unsigned LoadOpcode, StoreOpcode;
915 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode);
916 addFrameReference(BuildMI(MBB, MBBI, DL, get(StoreOpcode))
917 .addReg(SrcReg, getKillRegState(isKill)),
921 void SystemZInstrInfo::loadRegFromStackSlot(
922 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned DestReg,
923 int FrameIdx, const TargetRegisterClass *RC,
924 const TargetRegisterInfo *TRI) const {
925 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
927 // Callers may expect a single instruction, so keep 128-bit moves
928 // together for now and lower them after register allocation.
929 unsigned LoadOpcode, StoreOpcode;
930 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode);
931 addFrameReference(BuildMI(MBB, MBBI, DL, get(LoadOpcode), DestReg),
935 // Return true if MI is a simple load or store with a 12-bit displacement
936 // and no index. Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores.
937 static bool isSimpleBD12Move(const MachineInstr *MI, unsigned Flag) {
938 const MCInstrDesc &MCID = MI->getDesc();
939 return ((MCID.TSFlags & Flag) &&
940 isUInt<12>(MI->getOperand(2).getImm()) &&
941 MI->getOperand(3).getReg() == 0);
948 LogicOp(unsigned regSize, unsigned immLSB, unsigned immSize)
949 : RegSize(regSize), ImmLSB(immLSB), ImmSize(immSize) {}
951 explicit operator bool() const { return RegSize; }
953 unsigned RegSize = 0;
955 unsigned ImmSize = 0;
958 } // end anonymous namespace
960 static LogicOp interpretAndImmediate(unsigned Opcode) {
962 case SystemZ::NILMux: return LogicOp(32, 0, 16);
963 case SystemZ::NIHMux: return LogicOp(32, 16, 16);
964 case SystemZ::NILL64: return LogicOp(64, 0, 16);
965 case SystemZ::NILH64: return LogicOp(64, 16, 16);
966 case SystemZ::NIHL64: return LogicOp(64, 32, 16);
967 case SystemZ::NIHH64: return LogicOp(64, 48, 16);
968 case SystemZ::NIFMux: return LogicOp(32, 0, 32);
969 case SystemZ::NILF64: return LogicOp(64, 0, 32);
970 case SystemZ::NIHF64: return LogicOp(64, 32, 32);
971 default: return LogicOp();
975 static void transferDeadCC(MachineInstr *OldMI, MachineInstr *NewMI) {
976 if (OldMI->registerDefIsDead(SystemZ::CC)) {
977 MachineOperand *CCDef = NewMI->findRegisterDefOperand(SystemZ::CC);
978 if (CCDef != nullptr)
979 CCDef->setIsDead(true);
983 // Used to return from convertToThreeAddress after replacing two-address
984 // instruction OldMI with three-address instruction NewMI.
985 static MachineInstr *finishConvertToThreeAddress(MachineInstr *OldMI,
989 unsigned NumOps = OldMI->getNumOperands();
990 for (unsigned I = 1; I < NumOps; ++I) {
991 MachineOperand &Op = OldMI->getOperand(I);
992 if (Op.isReg() && Op.isKill())
993 LV->replaceKillInstruction(Op.getReg(), *OldMI, *NewMI);
996 transferDeadCC(OldMI, NewMI);
1000 MachineInstr *SystemZInstrInfo::convertToThreeAddress(
1001 MachineFunction::iterator &MFI, MachineInstr &MI, LiveVariables *LV) const {
1002 MachineBasicBlock *MBB = MI.getParent();
1003 MachineFunction *MF = MBB->getParent();
1004 MachineRegisterInfo &MRI = MF->getRegInfo();
1006 unsigned Opcode = MI.getOpcode();
1007 unsigned NumOps = MI.getNumOperands();
1009 // Try to convert something like SLL into SLLK, if supported.
1010 // We prefer to keep the two-operand form where possible both
1011 // because it tends to be shorter and because some instructions
1012 // have memory forms that can be used during spilling.
1013 if (STI.hasDistinctOps()) {
1014 MachineOperand &Dest = MI.getOperand(0);
1015 MachineOperand &Src = MI.getOperand(1);
1016 unsigned DestReg = Dest.getReg();
1017 unsigned SrcReg = Src.getReg();
1018 // AHIMux is only really a three-operand instruction when both operands
1019 // are low registers. Try to constrain both operands to be low if
1021 if (Opcode == SystemZ::AHIMux &&
1022 TargetRegisterInfo::isVirtualRegister(DestReg) &&
1023 TargetRegisterInfo::isVirtualRegister(SrcReg) &&
1024 MRI.getRegClass(DestReg)->contains(SystemZ::R1L) &&
1025 MRI.getRegClass(SrcReg)->contains(SystemZ::R1L)) {
1026 MRI.constrainRegClass(DestReg, &SystemZ::GR32BitRegClass);
1027 MRI.constrainRegClass(SrcReg, &SystemZ::GR32BitRegClass);
1029 int ThreeOperandOpcode = SystemZ::getThreeOperandOpcode(Opcode);
1030 if (ThreeOperandOpcode >= 0) {
1031 // Create three address instruction without adding the implicit
1032 // operands. Those will instead be copied over from the original
1033 // instruction by the loop below.
1034 MachineInstrBuilder MIB(
1035 *MF, MF->CreateMachineInstr(get(ThreeOperandOpcode), MI.getDebugLoc(),
1036 /*NoImplicit=*/true));
1038 // Keep the kill state, but drop the tied flag.
1039 MIB.addReg(Src.getReg(), getKillRegState(Src.isKill()), Src.getSubReg());
1040 // Keep the remaining operands as-is.
1041 for (unsigned I = 2; I < NumOps; ++I)
1042 MIB.add(MI.getOperand(I));
1043 MBB->insert(MI, MIB);
1044 return finishConvertToThreeAddress(&MI, MIB, LV);
1048 // Try to convert an AND into an RISBG-type instruction.
1049 if (LogicOp And = interpretAndImmediate(Opcode)) {
1050 uint64_t Imm = MI.getOperand(2).getImm() << And.ImmLSB;
1051 // AND IMMEDIATE leaves the other bits of the register unchanged.
1052 Imm |= allOnes(And.RegSize) & ~(allOnes(And.ImmSize) << And.ImmLSB);
1053 unsigned Start, End;
1054 if (isRxSBGMask(Imm, And.RegSize, Start, End)) {
1056 if (And.RegSize == 64) {
1057 NewOpcode = SystemZ::RISBG;
1058 // Prefer RISBGN if available, since it does not clobber CC.
1059 if (STI.hasMiscellaneousExtensions())
1060 NewOpcode = SystemZ::RISBGN;
1062 NewOpcode = SystemZ::RISBMux;
1066 MachineOperand &Dest = MI.getOperand(0);
1067 MachineOperand &Src = MI.getOperand(1);
1068 MachineInstrBuilder MIB =
1069 BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpcode))
1072 .addReg(Src.getReg(), getKillRegState(Src.isKill()),
1077 return finishConvertToThreeAddress(&MI, MIB, LV);
1083 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl(
1084 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
1085 MachineBasicBlock::iterator InsertPt, int FrameIndex,
1086 LiveIntervals *LIS) const {
1087 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1088 const MachineFrameInfo &MFI = MF.getFrameInfo();
1089 unsigned Size = MFI.getObjectSize(FrameIndex);
1090 unsigned Opcode = MI.getOpcode();
1092 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
1093 if (LIS != nullptr && (Opcode == SystemZ::LA || Opcode == SystemZ::LAY) &&
1094 isInt<8>(MI.getOperand(2).getImm()) && !MI.getOperand(3).getReg()) {
1096 // Check CC liveness, since new instruction introduces a dead
1098 MCRegUnitIterator CCUnit(SystemZ::CC, TRI);
1099 LiveRange &CCLiveRange = LIS->getRegUnit(*CCUnit);
1101 assert(!CCUnit.isValid() && "CC only has one reg unit.");
1103 LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot();
1104 if (!CCLiveRange.liveAt(MISlot)) {
1105 // LA(Y) %reg, CONST(%reg) -> AGSI %mem, CONST
1106 MachineInstr *BuiltMI = BuildMI(*InsertPt->getParent(), InsertPt,
1107 MI.getDebugLoc(), get(SystemZ::AGSI))
1108 .addFrameIndex(FrameIndex)
1110 .addImm(MI.getOperand(2).getImm());
1111 BuiltMI->findRegisterDefOperand(SystemZ::CC)->setIsDead(true);
1112 CCLiveRange.createDeadDef(MISlot, LIS->getVNInfoAllocator());
1119 // All other cases require a single operand.
1120 if (Ops.size() != 1)
1123 unsigned OpNum = Ops[0];
1125 TRI->getRegSizeInBits(*MF.getRegInfo()
1126 .getRegClass(MI.getOperand(OpNum).getReg())) &&
1127 "Invalid size combination");
1129 if ((Opcode == SystemZ::AHI || Opcode == SystemZ::AGHI) && OpNum == 0 &&
1130 isInt<8>(MI.getOperand(2).getImm())) {
1131 // A(G)HI %reg, CONST -> A(G)SI %mem, CONST
1132 Opcode = (Opcode == SystemZ::AHI ? SystemZ::ASI : SystemZ::AGSI);
1133 MachineInstr *BuiltMI =
1134 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode))
1135 .addFrameIndex(FrameIndex)
1137 .addImm(MI.getOperand(2).getImm());
1138 transferDeadCC(&MI, BuiltMI);
1142 if (Opcode == SystemZ::LGDR || Opcode == SystemZ::LDGR) {
1143 bool Op0IsGPR = (Opcode == SystemZ::LGDR);
1144 bool Op1IsGPR = (Opcode == SystemZ::LDGR);
1145 // If we're spilling the destination of an LDGR or LGDR, store the
1146 // source register instead.
1148 unsigned StoreOpcode = Op1IsGPR ? SystemZ::STG : SystemZ::STD;
1149 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(),
1151 .add(MI.getOperand(1))
1152 .addFrameIndex(FrameIndex)
1156 // If we're spilling the source of an LDGR or LGDR, load the
1157 // destination register instead.
1159 unsigned LoadOpcode = Op0IsGPR ? SystemZ::LG : SystemZ::LD;
1160 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(),
1162 .add(MI.getOperand(0))
1163 .addFrameIndex(FrameIndex)
1169 // Look for cases where the source of a simple store or the destination
1170 // of a simple load is being spilled. Try to use MVC instead.
1172 // Although MVC is in practice a fast choice in these cases, it is still
1173 // logically a bytewise copy. This means that we cannot use it if the
1174 // load or store is volatile. We also wouldn't be able to use MVC if
1175 // the two memories partially overlap, but that case cannot occur here,
1176 // because we know that one of the memories is a full frame index.
1178 // For performance reasons, we also want to avoid using MVC if the addresses
1179 // might be equal. We don't worry about that case here, because spill slot
1180 // coloring happens later, and because we have special code to remove
1181 // MVCs that turn out to be redundant.
1182 if (OpNum == 0 && MI.hasOneMemOperand()) {
1183 MachineMemOperand *MMO = *MI.memoperands_begin();
1184 if (MMO->getSize() == Size && !MMO->isVolatile()) {
1185 // Handle conversion of loads.
1186 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXLoad)) {
1187 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(),
1189 .addFrameIndex(FrameIndex)
1192 .add(MI.getOperand(1))
1193 .addImm(MI.getOperand(2).getImm())
1194 .addMemOperand(MMO);
1196 // Handle conversion of stores.
1197 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXStore)) {
1198 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(),
1200 .add(MI.getOperand(1))
1201 .addImm(MI.getOperand(2).getImm())
1203 .addFrameIndex(FrameIndex)
1205 .addMemOperand(MMO);
1210 // If the spilled operand is the final one, try to change <INSN>R
1212 int MemOpcode = SystemZ::getMemOpcode(Opcode);
1213 if (MemOpcode >= 0) {
1214 unsigned NumOps = MI.getNumExplicitOperands();
1215 if (OpNum == NumOps - 1) {
1216 const MCInstrDesc &MemDesc = get(MemOpcode);
1217 uint64_t AccessBytes = SystemZII::getAccessSize(MemDesc.TSFlags);
1218 assert(AccessBytes != 0 && "Size of access should be known");
1219 assert(AccessBytes <= Size && "Access outside the frame index");
1220 uint64_t Offset = Size - AccessBytes;
1221 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt,
1222 MI.getDebugLoc(), get(MemOpcode));
1223 for (unsigned I = 0; I < OpNum; ++I)
1224 MIB.add(MI.getOperand(I));
1225 MIB.addFrameIndex(FrameIndex).addImm(Offset);
1226 if (MemDesc.TSFlags & SystemZII::HasIndex)
1228 transferDeadCC(&MI, MIB);
1236 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl(
1237 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
1238 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI,
1239 LiveIntervals *LIS) const {
1243 bool SystemZInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
1244 switch (MI.getOpcode()) {
1246 splitMove(MI, SystemZ::LG);
1249 case SystemZ::ST128:
1250 splitMove(MI, SystemZ::STG);
1254 splitMove(MI, SystemZ::LD);
1258 splitMove(MI, SystemZ::STD);
1261 case SystemZ::LBMux:
1262 expandRXYPseudo(MI, SystemZ::LB, SystemZ::LBH);
1265 case SystemZ::LHMux:
1266 expandRXYPseudo(MI, SystemZ::LH, SystemZ::LHH);
1269 case SystemZ::LLCRMux:
1270 expandZExtPseudo(MI, SystemZ::LLCR, 8);
1273 case SystemZ::LLHRMux:
1274 expandZExtPseudo(MI, SystemZ::LLHR, 16);
1277 case SystemZ::LLCMux:
1278 expandRXYPseudo(MI, SystemZ::LLC, SystemZ::LLCH);
1281 case SystemZ::LLHMux:
1282 expandRXYPseudo(MI, SystemZ::LLH, SystemZ::LLHH);
1286 expandRXYPseudo(MI, SystemZ::L, SystemZ::LFH);
1289 case SystemZ::LOCMux:
1290 expandLOCPseudo(MI, SystemZ::LOC, SystemZ::LOCFH);
1293 case SystemZ::LOCHIMux:
1294 expandLOCPseudo(MI, SystemZ::LOCHI, SystemZ::LOCHHI);
1297 case SystemZ::LOCRMux:
1298 expandLOCRPseudo(MI, SystemZ::LOCR, SystemZ::LOCFHR);
1301 case SystemZ::STCMux:
1302 expandRXYPseudo(MI, SystemZ::STC, SystemZ::STCH);
1305 case SystemZ::STHMux:
1306 expandRXYPseudo(MI, SystemZ::STH, SystemZ::STHH);
1309 case SystemZ::STMux:
1310 expandRXYPseudo(MI, SystemZ::ST, SystemZ::STFH);
1313 case SystemZ::STOCMux:
1314 expandLOCPseudo(MI, SystemZ::STOC, SystemZ::STOCFH);
1317 case SystemZ::LHIMux:
1318 expandRIPseudo(MI, SystemZ::LHI, SystemZ::IIHF, true);
1321 case SystemZ::IIFMux:
1322 expandRIPseudo(MI, SystemZ::IILF, SystemZ::IIHF, false);
1325 case SystemZ::IILMux:
1326 expandRIPseudo(MI, SystemZ::IILL, SystemZ::IIHL, false);
1329 case SystemZ::IIHMux:
1330 expandRIPseudo(MI, SystemZ::IILH, SystemZ::IIHH, false);
1333 case SystemZ::NIFMux:
1334 expandRIPseudo(MI, SystemZ::NILF, SystemZ::NIHF, false);
1337 case SystemZ::NILMux:
1338 expandRIPseudo(MI, SystemZ::NILL, SystemZ::NIHL, false);
1341 case SystemZ::NIHMux:
1342 expandRIPseudo(MI, SystemZ::NILH, SystemZ::NIHH, false);
1345 case SystemZ::OIFMux:
1346 expandRIPseudo(MI, SystemZ::OILF, SystemZ::OIHF, false);
1349 case SystemZ::OILMux:
1350 expandRIPseudo(MI, SystemZ::OILL, SystemZ::OIHL, false);
1353 case SystemZ::OIHMux:
1354 expandRIPseudo(MI, SystemZ::OILH, SystemZ::OIHH, false);
1357 case SystemZ::XIFMux:
1358 expandRIPseudo(MI, SystemZ::XILF, SystemZ::XIHF, false);
1361 case SystemZ::TMLMux:
1362 expandRIPseudo(MI, SystemZ::TMLL, SystemZ::TMHL, false);
1365 case SystemZ::TMHMux:
1366 expandRIPseudo(MI, SystemZ::TMLH, SystemZ::TMHH, false);
1369 case SystemZ::AHIMux:
1370 expandRIPseudo(MI, SystemZ::AHI, SystemZ::AIH, false);
1373 case SystemZ::AHIMuxK:
1374 expandRIEPseudo(MI, SystemZ::AHI, SystemZ::AHIK, SystemZ::AIH);
1377 case SystemZ::AFIMux:
1378 expandRIPseudo(MI, SystemZ::AFI, SystemZ::AIH, false);
1381 case SystemZ::CHIMux:
1382 expandRIPseudo(MI, SystemZ::CHI, SystemZ::CIH, false);
1385 case SystemZ::CFIMux:
1386 expandRIPseudo(MI, SystemZ::CFI, SystemZ::CIH, false);
1389 case SystemZ::CLFIMux:
1390 expandRIPseudo(MI, SystemZ::CLFI, SystemZ::CLIH, false);
1394 expandRXYPseudo(MI, SystemZ::C, SystemZ::CHF);
1397 case SystemZ::CLMux:
1398 expandRXYPseudo(MI, SystemZ::CL, SystemZ::CLHF);
1401 case SystemZ::RISBMux: {
1402 bool DestIsHigh = isHighReg(MI.getOperand(0).getReg());
1403 bool SrcIsHigh = isHighReg(MI.getOperand(2).getReg());
1404 if (SrcIsHigh == DestIsHigh)
1405 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHH : SystemZ::RISBLL));
1407 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHL : SystemZ::RISBLH));
1408 MI.getOperand(5).setImm(MI.getOperand(5).getImm() ^ 32);
1413 case SystemZ::ADJDYNALLOC:
1414 splitAdjDynAlloc(MI);
1417 case TargetOpcode::LOAD_STACK_GUARD:
1418 expandLoadStackGuard(&MI);
1426 unsigned SystemZInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
1427 if (MI.getOpcode() == TargetOpcode::INLINEASM) {
1428 const MachineFunction *MF = MI.getParent()->getParent();
1429 const char *AsmStr = MI.getOperand(0).getSymbolName();
1430 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
1432 return MI.getDesc().getSize();
1436 SystemZInstrInfo::getBranchInfo(const MachineInstr &MI) const {
1437 switch (MI.getOpcode()) {
1441 return SystemZII::Branch(SystemZII::BranchNormal, SystemZ::CCMASK_ANY,
1442 SystemZ::CCMASK_ANY, &MI.getOperand(0));
1446 return SystemZII::Branch(SystemZII::BranchNormal, MI.getOperand(0).getImm(),
1447 MI.getOperand(1).getImm(), &MI.getOperand(2));
1450 case SystemZ::BRCTH:
1451 return SystemZII::Branch(SystemZII::BranchCT, SystemZ::CCMASK_ICMP,
1452 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2));
1454 case SystemZ::BRCTG:
1455 return SystemZII::Branch(SystemZII::BranchCTG, SystemZ::CCMASK_ICMP,
1456 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2));
1460 return SystemZII::Branch(SystemZII::BranchC, SystemZ::CCMASK_ICMP,
1461 MI.getOperand(2).getImm(), &MI.getOperand(3));
1465 return SystemZII::Branch(SystemZII::BranchCL, SystemZ::CCMASK_ICMP,
1466 MI.getOperand(2).getImm(), &MI.getOperand(3));
1470 return SystemZII::Branch(SystemZII::BranchCG, SystemZ::CCMASK_ICMP,
1471 MI.getOperand(2).getImm(), &MI.getOperand(3));
1473 case SystemZ::CLGIJ:
1474 case SystemZ::CLGRJ:
1475 return SystemZII::Branch(SystemZII::BranchCLG, SystemZ::CCMASK_ICMP,
1476 MI.getOperand(2).getImm(), &MI.getOperand(3));
1479 llvm_unreachable("Unrecognized branch opcode");
1483 void SystemZInstrInfo::getLoadStoreOpcodes(const TargetRegisterClass *RC,
1484 unsigned &LoadOpcode,
1485 unsigned &StoreOpcode) const {
1486 if (RC == &SystemZ::GR32BitRegClass || RC == &SystemZ::ADDR32BitRegClass) {
1487 LoadOpcode = SystemZ::L;
1488 StoreOpcode = SystemZ::ST;
1489 } else if (RC == &SystemZ::GRH32BitRegClass) {
1490 LoadOpcode = SystemZ::LFH;
1491 StoreOpcode = SystemZ::STFH;
1492 } else if (RC == &SystemZ::GRX32BitRegClass) {
1493 LoadOpcode = SystemZ::LMux;
1494 StoreOpcode = SystemZ::STMux;
1495 } else if (RC == &SystemZ::GR64BitRegClass ||
1496 RC == &SystemZ::ADDR64BitRegClass) {
1497 LoadOpcode = SystemZ::LG;
1498 StoreOpcode = SystemZ::STG;
1499 } else if (RC == &SystemZ::GR128BitRegClass ||
1500 RC == &SystemZ::ADDR128BitRegClass) {
1501 LoadOpcode = SystemZ::L128;
1502 StoreOpcode = SystemZ::ST128;
1503 } else if (RC == &SystemZ::FP32BitRegClass) {
1504 LoadOpcode = SystemZ::LE;
1505 StoreOpcode = SystemZ::STE;
1506 } else if (RC == &SystemZ::FP64BitRegClass) {
1507 LoadOpcode = SystemZ::LD;
1508 StoreOpcode = SystemZ::STD;
1509 } else if (RC == &SystemZ::FP128BitRegClass) {
1510 LoadOpcode = SystemZ::LX;
1511 StoreOpcode = SystemZ::STX;
1512 } else if (RC == &SystemZ::VR32BitRegClass) {
1513 LoadOpcode = SystemZ::VL32;
1514 StoreOpcode = SystemZ::VST32;
1515 } else if (RC == &SystemZ::VR64BitRegClass) {
1516 LoadOpcode = SystemZ::VL64;
1517 StoreOpcode = SystemZ::VST64;
1518 } else if (RC == &SystemZ::VF128BitRegClass ||
1519 RC == &SystemZ::VR128BitRegClass) {
1520 LoadOpcode = SystemZ::VL;
1521 StoreOpcode = SystemZ::VST;
1523 llvm_unreachable("Unsupported regclass to load or store");
1526 unsigned SystemZInstrInfo::getOpcodeForOffset(unsigned Opcode,
1527 int64_t Offset) const {
1528 const MCInstrDesc &MCID = get(Opcode);
1529 int64_t Offset2 = (MCID.TSFlags & SystemZII::Is128Bit ? Offset + 8 : Offset);
1530 if (isUInt<12>(Offset) && isUInt<12>(Offset2)) {
1531 // Get the instruction to use for unsigned 12-bit displacements.
1532 int Disp12Opcode = SystemZ::getDisp12Opcode(Opcode);
1533 if (Disp12Opcode >= 0)
1534 return Disp12Opcode;
1536 // All address-related instructions can use unsigned 12-bit
1540 if (isInt<20>(Offset) && isInt<20>(Offset2)) {
1541 // Get the instruction to use for signed 20-bit displacements.
1542 int Disp20Opcode = SystemZ::getDisp20Opcode(Opcode);
1543 if (Disp20Opcode >= 0)
1544 return Disp20Opcode;
1546 // Check whether Opcode allows signed 20-bit displacements.
1547 if (MCID.TSFlags & SystemZII::Has20BitOffset)
1553 unsigned SystemZInstrInfo::getLoadAndTest(unsigned Opcode) const {
1555 case SystemZ::L: return SystemZ::LT;
1556 case SystemZ::LY: return SystemZ::LT;
1557 case SystemZ::LG: return SystemZ::LTG;
1558 case SystemZ::LGF: return SystemZ::LTGF;
1559 case SystemZ::LR: return SystemZ::LTR;
1560 case SystemZ::LGFR: return SystemZ::LTGFR;
1561 case SystemZ::LGR: return SystemZ::LTGR;
1562 case SystemZ::LER: return SystemZ::LTEBR;
1563 case SystemZ::LDR: return SystemZ::LTDBR;
1564 case SystemZ::LXR: return SystemZ::LTXBR;
1565 case SystemZ::LCDFR: return SystemZ::LCDBR;
1566 case SystemZ::LPDFR: return SystemZ::LPDBR;
1567 case SystemZ::LNDFR: return SystemZ::LNDBR;
1568 case SystemZ::LCDFR_32: return SystemZ::LCEBR;
1569 case SystemZ::LPDFR_32: return SystemZ::LPEBR;
1570 case SystemZ::LNDFR_32: return SystemZ::LNEBR;
1571 // On zEC12 we prefer to use RISBGN. But if there is a chance to
1572 // actually use the condition code, we may turn it back into RISGB.
1573 // Note that RISBG is not really a "load-and-test" instruction,
1574 // but sets the same condition code values, so is OK to use here.
1575 case SystemZ::RISBGN: return SystemZ::RISBG;
1580 // Return true if Mask matches the regexp 0*1+0*, given that zero masks
1581 // have already been filtered out. Store the first set bit in LSB and
1582 // the number of set bits in Length if so.
1583 static bool isStringOfOnes(uint64_t Mask, unsigned &LSB, unsigned &Length) {
1584 unsigned First = findFirstSet(Mask);
1585 uint64_t Top = (Mask >> First) + 1;
1586 if ((Top & -Top) == Top) {
1588 Length = findFirstSet(Top);
1594 bool SystemZInstrInfo::isRxSBGMask(uint64_t Mask, unsigned BitSize,
1595 unsigned &Start, unsigned &End) const {
1596 // Reject trivial all-zero masks.
1597 Mask &= allOnes(BitSize);
1601 // Handle the 1+0+ or 0+1+0* cases. Start then specifies the index of
1602 // the msb and End specifies the index of the lsb.
1603 unsigned LSB, Length;
1604 if (isStringOfOnes(Mask, LSB, Length)) {
1605 Start = 63 - (LSB + Length - 1);
1610 // Handle the wrap-around 1+0+1+ cases. Start then specifies the msb
1611 // of the low 1s and End specifies the lsb of the high 1s.
1612 if (isStringOfOnes(Mask ^ allOnes(BitSize), LSB, Length)) {
1613 assert(LSB > 0 && "Bottom bit must be set");
1614 assert(LSB + Length < BitSize && "Top bit must be set");
1615 Start = 63 - (LSB - 1);
1616 End = 63 - (LSB + Length);
1623 unsigned SystemZInstrInfo::getFusedCompare(unsigned Opcode,
1624 SystemZII::FusedCompareType Type,
1625 const MachineInstr *MI) const {
1629 if (!(MI && isInt<8>(MI->getOperand(1).getImm())))
1633 case SystemZ::CLGFI:
1634 if (!(MI && isUInt<8>(MI->getOperand(1).getImm())))
1639 if (!STI.hasMiscellaneousExtensions())
1641 if (!(MI && MI->getOperand(3).getReg() == 0))
1646 case SystemZII::CompareAndBranch:
1649 return SystemZ::CRJ;
1651 return SystemZ::CGRJ;
1653 return SystemZ::CIJ;
1655 return SystemZ::CGIJ;
1657 return SystemZ::CLRJ;
1659 return SystemZ::CLGRJ;
1661 return SystemZ::CLIJ;
1662 case SystemZ::CLGFI:
1663 return SystemZ::CLGIJ;
1667 case SystemZII::CompareAndReturn:
1670 return SystemZ::CRBReturn;
1672 return SystemZ::CGRBReturn;
1674 return SystemZ::CIBReturn;
1676 return SystemZ::CGIBReturn;
1678 return SystemZ::CLRBReturn;
1680 return SystemZ::CLGRBReturn;
1682 return SystemZ::CLIBReturn;
1683 case SystemZ::CLGFI:
1684 return SystemZ::CLGIBReturn;
1688 case SystemZII::CompareAndSibcall:
1691 return SystemZ::CRBCall;
1693 return SystemZ::CGRBCall;
1695 return SystemZ::CIBCall;
1697 return SystemZ::CGIBCall;
1699 return SystemZ::CLRBCall;
1701 return SystemZ::CLGRBCall;
1703 return SystemZ::CLIBCall;
1704 case SystemZ::CLGFI:
1705 return SystemZ::CLGIBCall;
1709 case SystemZII::CompareAndTrap:
1712 return SystemZ::CRT;
1714 return SystemZ::CGRT;
1716 return SystemZ::CIT;
1718 return SystemZ::CGIT;
1720 return SystemZ::CLRT;
1722 return SystemZ::CLGRT;
1724 return SystemZ::CLFIT;
1725 case SystemZ::CLGFI:
1726 return SystemZ::CLGIT;
1728 return SystemZ::CLT;
1730 return SystemZ::CLGT;
1738 unsigned SystemZInstrInfo::getLoadAndTrap(unsigned Opcode) const {
1739 if (!STI.hasLoadAndTrap())
1744 return SystemZ::LAT;
1746 return SystemZ::LGAT;
1748 return SystemZ::LFHAT;
1750 return SystemZ::LLGFAT;
1752 return SystemZ::LLGTAT;
1757 void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB,
1758 MachineBasicBlock::iterator MBBI,
1759 unsigned Reg, uint64_t Value) const {
1760 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
1762 if (isInt<16>(Value))
1763 Opcode = SystemZ::LGHI;
1764 else if (SystemZ::isImmLL(Value))
1765 Opcode = SystemZ::LLILL;
1766 else if (SystemZ::isImmLH(Value)) {
1767 Opcode = SystemZ::LLILH;
1770 assert(isInt<32>(Value) && "Huge values not handled yet");
1771 Opcode = SystemZ::LGFI;
1773 BuildMI(MBB, MBBI, DL, get(Opcode), Reg).addImm(Value);
1776 bool SystemZInstrInfo::
1777 areMemAccessesTriviallyDisjoint(MachineInstr &MIa, MachineInstr &MIb,
1778 AliasAnalysis *AA) const {
1780 if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand())
1783 // If mem-operands show that the same address Value is used by both
1784 // instructions, check for non-overlapping offsets and widths. Not
1785 // sure if a register based analysis would be an improvement...
1787 MachineMemOperand *MMOa = *MIa.memoperands_begin();
1788 MachineMemOperand *MMOb = *MIb.memoperands_begin();
1789 const Value *VALa = MMOa->getValue();
1790 const Value *VALb = MMOb->getValue();
1791 bool SameVal = (VALa && VALb && (VALa == VALb));
1793 const PseudoSourceValue *PSVa = MMOa->getPseudoValue();
1794 const PseudoSourceValue *PSVb = MMOb->getPseudoValue();
1795 if (PSVa && PSVb && (PSVa == PSVb))
1799 int OffsetA = MMOa->getOffset(), OffsetB = MMOb->getOffset();
1800 int WidthA = MMOa->getSize(), WidthB = MMOb->getSize();
1801 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB;
1802 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA;
1803 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
1804 if (LowOffset + LowWidth <= HighOffset)