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 "SystemZInstrBuilder.h"
16 #include "llvm/CodeGen/MachineRegisterInfo.h"
17 #include "llvm/Target/TargetMachine.h"
19 #define GET_INSTRINFO_CTOR
20 #define GET_INSTRMAP_INFO
21 #include "SystemZGenInstrInfo.inc"
25 SystemZInstrInfo::SystemZInstrInfo(SystemZTargetMachine &tm)
26 : SystemZGenInstrInfo(SystemZ::ADJCALLSTACKDOWN, SystemZ::ADJCALLSTACKUP),
30 // MI is a 128-bit load or store. Split it into two 64-bit loads or stores,
31 // each having the opcode given by NewOpcode.
32 void SystemZInstrInfo::splitMove(MachineBasicBlock::iterator MI,
33 unsigned NewOpcode) const {
34 MachineBasicBlock *MBB = MI->getParent();
35 MachineFunction &MF = *MBB->getParent();
37 // Get two load or store instructions. Use the original instruction for one
38 // of them (arbitarily the second here) and create a clone for the other.
39 MachineInstr *EarlierMI = MF.CloneMachineInstr(MI);
40 MBB->insert(MI, EarlierMI);
42 // Set up the two 64-bit registers.
43 MachineOperand &HighRegOp = EarlierMI->getOperand(0);
44 MachineOperand &LowRegOp = MI->getOperand(0);
45 HighRegOp.setReg(RI.getSubReg(HighRegOp.getReg(), SystemZ::subreg_high));
46 LowRegOp.setReg(RI.getSubReg(LowRegOp.getReg(), SystemZ::subreg_low));
48 // The address in the first (high) instruction is already correct.
49 // Adjust the offset in the second (low) instruction.
50 MachineOperand &HighOffsetOp = EarlierMI->getOperand(2);
51 MachineOperand &LowOffsetOp = MI->getOperand(2);
52 LowOffsetOp.setImm(LowOffsetOp.getImm() + 8);
55 unsigned HighOpcode = getOpcodeForOffset(NewOpcode, HighOffsetOp.getImm());
56 unsigned LowOpcode = getOpcodeForOffset(NewOpcode, LowOffsetOp.getImm());
57 assert(HighOpcode && LowOpcode && "Both offsets should be in range");
59 EarlierMI->setDesc(get(HighOpcode));
60 MI->setDesc(get(LowOpcode));
63 // Split ADJDYNALLOC instruction MI.
64 void SystemZInstrInfo::splitAdjDynAlloc(MachineBasicBlock::iterator MI) const {
65 MachineBasicBlock *MBB = MI->getParent();
66 MachineFunction &MF = *MBB->getParent();
67 MachineFrameInfo *MFFrame = MF.getFrameInfo();
68 MachineOperand &OffsetMO = MI->getOperand(2);
70 uint64_t Offset = (MFFrame->getMaxCallFrameSize() +
71 SystemZMC::CallFrameSize +
73 unsigned NewOpcode = getOpcodeForOffset(SystemZ::LA, Offset);
74 assert(NewOpcode && "No support for huge argument lists yet");
75 MI->setDesc(get(NewOpcode));
76 OffsetMO.setImm(Offset);
79 // If MI is a simple load or store for a frame object, return the register
80 // it loads or stores and set FrameIndex to the index of the frame object.
81 // Return 0 otherwise.
83 // Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores.
84 static int isSimpleMove(const MachineInstr *MI, int &FrameIndex,
86 const MCInstrDesc &MCID = MI->getDesc();
87 if ((MCID.TSFlags & Flag) &&
88 MI->getOperand(1).isFI() &&
89 MI->getOperand(2).getImm() == 0 &&
90 MI->getOperand(3).getReg() == 0) {
91 FrameIndex = MI->getOperand(1).getIndex();
92 return MI->getOperand(0).getReg();
97 unsigned SystemZInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
98 int &FrameIndex) const {
99 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXLoad);
102 unsigned SystemZInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
103 int &FrameIndex) const {
104 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXStore);
107 bool SystemZInstrInfo::isStackSlotCopy(const MachineInstr *MI,
109 int &SrcFrameIndex) const {
110 // Check for MVC 0(Length,FI1),0(FI2)
111 const MachineFrameInfo *MFI = MI->getParent()->getParent()->getFrameInfo();
112 if (MI->getOpcode() != SystemZ::MVC ||
113 !MI->getOperand(0).isFI() ||
114 MI->getOperand(1).getImm() != 0 ||
115 !MI->getOperand(3).isFI() ||
116 MI->getOperand(4).getImm() != 0)
119 // Check that Length covers the full slots.
120 int64_t Length = MI->getOperand(2).getImm();
121 unsigned FI1 = MI->getOperand(0).getIndex();
122 unsigned FI2 = MI->getOperand(3).getIndex();
123 if (MFI->getObjectSize(FI1) != Length ||
124 MFI->getObjectSize(FI2) != Length)
127 DestFrameIndex = FI1;
132 bool SystemZInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
133 MachineBasicBlock *&TBB,
134 MachineBasicBlock *&FBB,
135 SmallVectorImpl<MachineOperand> &Cond,
136 bool AllowModify) const {
137 // Most of the code and comments here are boilerplate.
139 // Start from the bottom of the block and work up, examining the
140 // terminator instructions.
141 MachineBasicBlock::iterator I = MBB.end();
142 while (I != MBB.begin()) {
144 if (I->isDebugValue())
147 // Working from the bottom, when we see a non-terminator instruction, we're
149 if (!isUnpredicatedTerminator(I))
152 // A terminator that isn't a branch can't easily be handled by this
157 // Can't handle indirect branches.
158 SystemZII::Branch Branch(getBranchInfo(I));
159 if (!Branch.Target->isMBB())
162 // Punt on compound branches.
163 if (Branch.Type != SystemZII::BranchNormal)
166 if (Branch.CCMask == SystemZ::CCMASK_ANY) {
167 // Handle unconditional branches.
169 TBB = Branch.Target->getMBB();
173 // If the block has any instructions after a JMP, delete them.
174 while (llvm::next(I) != MBB.end())
175 llvm::next(I)->eraseFromParent();
180 // Delete the JMP if it's equivalent to a fall-through.
181 if (MBB.isLayoutSuccessor(Branch.Target->getMBB())) {
183 I->eraseFromParent();
188 // TBB is used to indicate the unconditinal destination.
189 TBB = Branch.Target->getMBB();
193 // Working from the bottom, handle the first conditional branch.
195 // FIXME: add X86-style branch swap
197 TBB = Branch.Target->getMBB();
198 Cond.push_back(MachineOperand::CreateImm(Branch.CCMask));
202 // Handle subsequent conditional branches.
203 assert(Cond.size() == 1);
206 // Only handle the case where all conditional branches branch to the same
208 if (TBB != Branch.Target->getMBB())
211 // If the conditions are the same, we can leave them alone.
212 unsigned OldCond = Cond[0].getImm();
213 if (OldCond == Branch.CCMask)
216 // FIXME: Try combining conditions like X86 does. Should be easy on Z!
222 unsigned SystemZInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
223 // Most of the code and comments here are boilerplate.
224 MachineBasicBlock::iterator I = MBB.end();
227 while (I != MBB.begin()) {
229 if (I->isDebugValue())
233 if (!getBranchInfo(I).Target->isMBB())
235 // Remove the branch.
236 I->eraseFromParent();
245 SystemZInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
246 MachineBasicBlock *FBB,
247 const SmallVectorImpl<MachineOperand> &Cond,
249 // In this function we output 32-bit branches, which should always
250 // have enough range. They can be shortened and relaxed by later code
251 // in the pipeline, if desired.
253 // Shouldn't be a fall through.
254 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
255 assert((Cond.size() == 1 || Cond.size() == 0) &&
256 "SystemZ branch conditions have one component!");
259 // Unconditional branch?
260 assert(!FBB && "Unconditional branch with multiple successors!");
261 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(TBB);
265 // Conditional branch.
267 unsigned CC = Cond[0].getImm();
268 BuildMI(&MBB, DL, get(SystemZ::BRC)).addImm(CC).addMBB(TBB);
272 // Two-way Conditional branch. Insert the second branch.
273 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(FBB);
280 SystemZInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
281 MachineBasicBlock::iterator MBBI, DebugLoc DL,
282 unsigned DestReg, unsigned SrcReg,
283 bool KillSrc) const {
284 // Split 128-bit GPR moves into two 64-bit moves. This handles ADDR128 too.
285 if (SystemZ::GR128BitRegClass.contains(DestReg, SrcReg)) {
286 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_high),
287 RI.getSubReg(SrcReg, SystemZ::subreg_high), KillSrc);
288 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_low),
289 RI.getSubReg(SrcReg, SystemZ::subreg_low), KillSrc);
293 // Everything else needs only one instruction.
295 if (SystemZ::GR32BitRegClass.contains(DestReg, SrcReg))
296 Opcode = SystemZ::LR;
297 else if (SystemZ::GR64BitRegClass.contains(DestReg, SrcReg))
298 Opcode = SystemZ::LGR;
299 else if (SystemZ::FP32BitRegClass.contains(DestReg, SrcReg))
300 Opcode = SystemZ::LER;
301 else if (SystemZ::FP64BitRegClass.contains(DestReg, SrcReg))
302 Opcode = SystemZ::LDR;
303 else if (SystemZ::FP128BitRegClass.contains(DestReg, SrcReg))
304 Opcode = SystemZ::LXR;
306 llvm_unreachable("Impossible reg-to-reg copy");
308 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg)
309 .addReg(SrcReg, getKillRegState(KillSrc));
313 SystemZInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
314 MachineBasicBlock::iterator MBBI,
315 unsigned SrcReg, bool isKill,
317 const TargetRegisterClass *RC,
318 const TargetRegisterInfo *TRI) const {
319 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
321 // Callers may expect a single instruction, so keep 128-bit moves
322 // together for now and lower them after register allocation.
323 unsigned LoadOpcode, StoreOpcode;
324 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode);
325 addFrameReference(BuildMI(MBB, MBBI, DL, get(StoreOpcode))
326 .addReg(SrcReg, getKillRegState(isKill)), FrameIdx);
330 SystemZInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
331 MachineBasicBlock::iterator MBBI,
332 unsigned DestReg, int FrameIdx,
333 const TargetRegisterClass *RC,
334 const TargetRegisterInfo *TRI) const {
335 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
337 // Callers may expect a single instruction, so keep 128-bit moves
338 // together for now and lower them after register allocation.
339 unsigned LoadOpcode, StoreOpcode;
340 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode);
341 addFrameReference(BuildMI(MBB, MBBI, DL, get(LoadOpcode), DestReg),
345 // Return true if MI is a simple load or store with a 12-bit displacement
346 // and no index. Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores.
347 static bool isSimpleBD12Move(const MachineInstr *MI, unsigned Flag) {
348 const MCInstrDesc &MCID = MI->getDesc();
349 return ((MCID.TSFlags & Flag) &&
350 isUInt<12>(MI->getOperand(2).getImm()) &&
351 MI->getOperand(3).getReg() == 0);
355 SystemZInstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
357 const SmallVectorImpl<unsigned> &Ops,
358 int FrameIndex) const {
359 const MachineFrameInfo *MFI = MF.getFrameInfo();
360 unsigned Size = MFI->getObjectSize(FrameIndex);
362 // Eary exit for cases we don't care about
366 unsigned OpNum = Ops[0];
367 assert(Size == MF.getRegInfo()
368 .getRegClass(MI->getOperand(OpNum).getReg())->getSize() &&
369 "Invalid size combination");
371 unsigned Opcode = MI->getOpcode();
372 if (Opcode == SystemZ::LGDR || Opcode == SystemZ::LDGR) {
373 bool Op0IsGPR = (Opcode == SystemZ::LGDR);
374 bool Op1IsGPR = (Opcode == SystemZ::LDGR);
375 // If we're spilling the destination of an LDGR or LGDR, store the
376 // source register instead.
378 unsigned StoreOpcode = Op1IsGPR ? SystemZ::STG : SystemZ::STD;
379 return BuildMI(MF, MI->getDebugLoc(), get(StoreOpcode))
380 .addOperand(MI->getOperand(1)).addFrameIndex(FrameIndex)
381 .addImm(0).addReg(0);
383 // If we're spilling the source of an LDGR or LGDR, load the
384 // destination register instead.
386 unsigned LoadOpcode = Op0IsGPR ? SystemZ::LG : SystemZ::LD;
387 unsigned Dest = MI->getOperand(0).getReg();
388 return BuildMI(MF, MI->getDebugLoc(), get(LoadOpcode), Dest)
389 .addFrameIndex(FrameIndex).addImm(0).addReg(0);
393 // Look for cases where the source of a simple store or the destination
394 // of a simple load is being spilled. Try to use MVC instead.
396 // Although MVC is in practice a fast choice in these cases, it is still
397 // logically a bytewise copy. This means that we cannot use it if the
398 // load or store is volatile. It also means that the transformation is
399 // not valid in cases where the two memories partially overlap; however,
400 // that is not a problem here, because we know that one of the memories
401 // is a full frame index.
402 if (OpNum == 0 && MI->hasOneMemOperand()) {
403 MachineMemOperand *MMO = *MI->memoperands_begin();
404 if (MMO->getSize() == Size && !MMO->isVolatile()) {
405 // Handle conversion of loads.
406 if (isSimpleBD12Move(MI, SystemZII::SimpleBDXLoad)) {
407 return BuildMI(MF, MI->getDebugLoc(), get(SystemZ::MVC))
408 .addFrameIndex(FrameIndex).addImm(0).addImm(Size)
409 .addOperand(MI->getOperand(1)).addImm(MI->getOperand(2).getImm())
412 // Handle conversion of stores.
413 if (isSimpleBD12Move(MI, SystemZII::SimpleBDXStore)) {
414 return BuildMI(MF, MI->getDebugLoc(), get(SystemZ::MVC))
415 .addOperand(MI->getOperand(1)).addImm(MI->getOperand(2).getImm())
416 .addImm(Size).addFrameIndex(FrameIndex).addImm(0)
422 // If the spilled operand is the final one, try to change <INSN>R
424 int MemOpcode = SystemZ::getMemOpcode(Opcode);
425 if (MemOpcode >= 0) {
426 unsigned NumOps = MI->getNumExplicitOperands();
427 if (OpNum == NumOps - 1) {
428 const MCInstrDesc &MemDesc = get(MemOpcode);
429 uint64_t AccessBytes = SystemZII::getAccessSize(MemDesc.TSFlags);
430 assert(AccessBytes != 0 && "Size of access should be known");
431 assert(AccessBytes <= Size && "Access outside the frame index");
432 uint64_t Offset = Size - AccessBytes;
433 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(MemOpcode));
434 for (unsigned I = 0; I < OpNum; ++I)
435 MIB.addOperand(MI->getOperand(I));
436 MIB.addFrameIndex(FrameIndex).addImm(Offset);
437 if (MemDesc.TSFlags & SystemZII::HasIndex)
447 SystemZInstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr* MI,
448 const SmallVectorImpl<unsigned> &Ops,
449 MachineInstr* LoadMI) const {
454 SystemZInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
455 switch (MI->getOpcode()) {
457 splitMove(MI, SystemZ::LG);
461 splitMove(MI, SystemZ::STG);
465 splitMove(MI, SystemZ::LD);
469 splitMove(MI, SystemZ::STD);
472 case SystemZ::ADJDYNALLOC:
473 splitAdjDynAlloc(MI);
481 bool SystemZInstrInfo::
482 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
483 assert(Cond.size() == 1 && "Invalid branch condition!");
484 Cond[0].setImm(Cond[0].getImm() ^ SystemZ::CCMASK_ANY);
488 uint64_t SystemZInstrInfo::getInstSizeInBytes(const MachineInstr *MI) const {
489 if (MI->getOpcode() == TargetOpcode::INLINEASM) {
490 const MachineFunction *MF = MI->getParent()->getParent();
491 const char *AsmStr = MI->getOperand(0).getSymbolName();
492 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
494 return MI->getDesc().getSize();
498 SystemZInstrInfo::getBranchInfo(const MachineInstr *MI) const {
499 switch (MI->getOpcode()) {
503 return SystemZII::Branch(SystemZII::BranchNormal, SystemZ::CCMASK_ANY,
508 return SystemZII::Branch(SystemZII::BranchNormal,
509 MI->getOperand(0).getImm(), &MI->getOperand(1));
513 return SystemZII::Branch(SystemZII::BranchC, MI->getOperand(2).getImm(),
518 return SystemZII::Branch(SystemZII::BranchCG, MI->getOperand(2).getImm(),
522 llvm_unreachable("Unrecognized branch opcode");
526 void SystemZInstrInfo::getLoadStoreOpcodes(const TargetRegisterClass *RC,
527 unsigned &LoadOpcode,
528 unsigned &StoreOpcode) const {
529 if (RC == &SystemZ::GR32BitRegClass || RC == &SystemZ::ADDR32BitRegClass) {
530 LoadOpcode = SystemZ::L;
531 StoreOpcode = SystemZ::ST32;
532 } else if (RC == &SystemZ::GR64BitRegClass ||
533 RC == &SystemZ::ADDR64BitRegClass) {
534 LoadOpcode = SystemZ::LG;
535 StoreOpcode = SystemZ::STG;
536 } else if (RC == &SystemZ::GR128BitRegClass ||
537 RC == &SystemZ::ADDR128BitRegClass) {
538 LoadOpcode = SystemZ::L128;
539 StoreOpcode = SystemZ::ST128;
540 } else if (RC == &SystemZ::FP32BitRegClass) {
541 LoadOpcode = SystemZ::LE;
542 StoreOpcode = SystemZ::STE;
543 } else if (RC == &SystemZ::FP64BitRegClass) {
544 LoadOpcode = SystemZ::LD;
545 StoreOpcode = SystemZ::STD;
546 } else if (RC == &SystemZ::FP128BitRegClass) {
547 LoadOpcode = SystemZ::LX;
548 StoreOpcode = SystemZ::STX;
550 llvm_unreachable("Unsupported regclass to load or store");
553 unsigned SystemZInstrInfo::getOpcodeForOffset(unsigned Opcode,
554 int64_t Offset) const {
555 const MCInstrDesc &MCID = get(Opcode);
556 int64_t Offset2 = (MCID.TSFlags & SystemZII::Is128Bit ? Offset + 8 : Offset);
557 if (isUInt<12>(Offset) && isUInt<12>(Offset2)) {
558 // Get the instruction to use for unsigned 12-bit displacements.
559 int Disp12Opcode = SystemZ::getDisp12Opcode(Opcode);
560 if (Disp12Opcode >= 0)
563 // All address-related instructions can use unsigned 12-bit
567 if (isInt<20>(Offset) && isInt<20>(Offset2)) {
568 // Get the instruction to use for signed 20-bit displacements.
569 int Disp20Opcode = SystemZ::getDisp20Opcode(Opcode);
570 if (Disp20Opcode >= 0)
573 // Check whether Opcode allows signed 20-bit displacements.
574 if (MCID.TSFlags & SystemZII::Has20BitOffset)
580 unsigned SystemZInstrInfo::getCompareAndBranch(unsigned Opcode,
581 const MachineInstr *MI) const {
586 return SystemZ::CGRJ;
588 return MI && isInt<8>(MI->getOperand(1).getImm()) ? SystemZ::CIJ : 0;
590 return MI && isInt<8>(MI->getOperand(1).getImm()) ? SystemZ::CGIJ : 0;
596 void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB,
597 MachineBasicBlock::iterator MBBI,
598 unsigned Reg, uint64_t Value) const {
599 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
601 if (isInt<16>(Value))
602 Opcode = SystemZ::LGHI;
603 else if (SystemZ::isImmLL(Value))
604 Opcode = SystemZ::LLILL;
605 else if (SystemZ::isImmLH(Value)) {
606 Opcode = SystemZ::LLILH;
609 assert(isInt<32>(Value) && "Huge values not handled yet");
610 Opcode = SystemZ::LGFI;
612 BuildMI(MBB, MBBI, DL, get(Opcode), Reg).addImm(Value);