1 //===-- PPCInstrInfo.cpp - PowerPC 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 PowerPC implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "PPCInstrInfo.h"
15 #include "MCTargetDesc/PPCPredicates.h"
17 #include "PPCHazardRecognizers.h"
18 #include "PPCInstrBuilder.h"
19 #include "PPCMachineFunctionInfo.h"
20 #include "PPCTargetMachine.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunctionPass.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineMemOperand.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/PseudoSourceValue.h"
30 #include "llvm/CodeGen/ScheduleDAG.h"
31 #include "llvm/CodeGen/SlotIndexes.h"
32 #include "llvm/CodeGen/StackMaps.h"
33 #include "llvm/MC/MCAsmInfo.h"
34 #include "llvm/MC/MCInst.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/TargetRegistry.h"
39 #include "llvm/Support/raw_ostream.h"
43 #define DEBUG_TYPE "ppc-instr-info"
45 #define GET_INSTRMAP_INFO
46 #define GET_INSTRINFO_CTOR_DTOR
47 #include "PPCGenInstrInfo.inc"
50 opt<bool> DisableCTRLoopAnal("disable-ppc-ctrloop-analysis", cl::Hidden,
51 cl::desc("Disable analysis for CTR loops"));
53 static cl::opt<bool> DisableCmpOpt("disable-ppc-cmp-opt",
54 cl::desc("Disable compare instruction optimization"), cl::Hidden);
56 static cl::opt<bool> VSXSelfCopyCrash("crash-on-ppc-vsx-self-copy",
57 cl::desc("Causes the backend to crash instead of generating a nop VSX copy"),
60 // Pin the vtable to this file.
61 void PPCInstrInfo::anchor() {}
63 PPCInstrInfo::PPCInstrInfo(PPCSubtarget &STI)
64 : PPCGenInstrInfo(PPC::ADJCALLSTACKDOWN, PPC::ADJCALLSTACKUP),
65 Subtarget(STI), RI(STI.getTargetMachine()) {}
67 /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for
68 /// this target when scheduling the DAG.
69 ScheduleHazardRecognizer *
70 PPCInstrInfo::CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
71 const ScheduleDAG *DAG) const {
73 static_cast<const PPCSubtarget *>(STI)->getDarwinDirective();
74 if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2 ||
75 Directive == PPC::DIR_E500mc || Directive == PPC::DIR_E5500) {
76 const InstrItineraryData *II =
77 static_cast<const PPCSubtarget *>(STI)->getInstrItineraryData();
78 return new ScoreboardHazardRecognizer(II, DAG);
81 return TargetInstrInfo::CreateTargetHazardRecognizer(STI, DAG);
84 /// CreateTargetPostRAHazardRecognizer - Return the postRA hazard recognizer
85 /// to use for this target when scheduling the DAG.
86 ScheduleHazardRecognizer *
87 PPCInstrInfo::CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
88 const ScheduleDAG *DAG) const {
90 DAG->MF.getSubtarget<PPCSubtarget>().getDarwinDirective();
92 if (Directive == PPC::DIR_PWR7 || Directive == PPC::DIR_PWR8)
93 return new PPCDispatchGroupSBHazardRecognizer(II, DAG);
95 // Most subtargets use a PPC970 recognizer.
96 if (Directive != PPC::DIR_440 && Directive != PPC::DIR_A2 &&
97 Directive != PPC::DIR_E500mc && Directive != PPC::DIR_E5500) {
98 assert(DAG->TII && "No InstrInfo?");
100 return new PPCHazardRecognizer970(*DAG);
103 return new ScoreboardHazardRecognizer(II, DAG);
107 int PPCInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
108 const MachineInstr *DefMI, unsigned DefIdx,
109 const MachineInstr *UseMI,
110 unsigned UseIdx) const {
111 int Latency = PPCGenInstrInfo::getOperandLatency(ItinData, DefMI, DefIdx,
114 const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
115 unsigned Reg = DefMO.getReg();
118 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
119 const MachineRegisterInfo *MRI =
120 &DefMI->getParent()->getParent()->getRegInfo();
121 IsRegCR = MRI->getRegClass(Reg)->hasSuperClassEq(&PPC::CRRCRegClass) ||
122 MRI->getRegClass(Reg)->hasSuperClassEq(&PPC::CRBITRCRegClass);
124 IsRegCR = PPC::CRRCRegClass.contains(Reg) ||
125 PPC::CRBITRCRegClass.contains(Reg);
128 if (UseMI->isBranch() && IsRegCR) {
130 Latency = getInstrLatency(ItinData, DefMI);
132 // On some cores, there is an additional delay between writing to a condition
133 // register, and using it from a branch.
134 unsigned Directive = Subtarget.getDarwinDirective();
156 // Detect 32 -> 64-bit extensions where we may reuse the low sub-register.
157 bool PPCInstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
158 unsigned &SrcReg, unsigned &DstReg,
159 unsigned &SubIdx) const {
160 switch (MI.getOpcode()) {
161 default: return false;
163 case PPC::EXTSW_32_64:
164 SrcReg = MI.getOperand(1).getReg();
165 DstReg = MI.getOperand(0).getReg();
166 SubIdx = PPC::sub_32;
171 unsigned PPCInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
172 int &FrameIndex) const {
173 // Note: This list must be kept consistent with LoadRegFromStackSlot.
174 switch (MI->getOpcode()) {
180 case PPC::RESTORE_CR:
181 case PPC::RESTORE_CRBIT:
187 case PPC::RESTORE_VRSAVE:
188 // Check for the operands added by addFrameReference (the immediate is the
189 // offset which defaults to 0).
190 if (MI->getOperand(1).isImm() && !MI->getOperand(1).getImm() &&
191 MI->getOperand(2).isFI()) {
192 FrameIndex = MI->getOperand(2).getIndex();
193 return MI->getOperand(0).getReg();
200 unsigned PPCInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
201 int &FrameIndex) const {
202 // Note: This list must be kept consistent with StoreRegToStackSlot.
203 switch (MI->getOpcode()) {
210 case PPC::SPILL_CRBIT:
216 case PPC::SPILL_VRSAVE:
217 // Check for the operands added by addFrameReference (the immediate is the
218 // offset which defaults to 0).
219 if (MI->getOperand(1).isImm() && !MI->getOperand(1).getImm() &&
220 MI->getOperand(2).isFI()) {
221 FrameIndex = MI->getOperand(2).getIndex();
222 return MI->getOperand(0).getReg();
229 // commuteInstruction - We can commute rlwimi instructions, but only if the
230 // rotate amt is zero. We also have to munge the immediates a bit.
232 PPCInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
233 MachineFunction &MF = *MI->getParent()->getParent();
235 // Normal instructions can be commuted the obvious way.
236 if (MI->getOpcode() != PPC::RLWIMI &&
237 MI->getOpcode() != PPC::RLWIMIo)
238 return TargetInstrInfo::commuteInstruction(MI, NewMI);
239 // Note that RLWIMI can be commuted as a 32-bit instruction, but not as a
240 // 64-bit instruction (so we don't handle PPC::RLWIMI8 here), because
241 // changing the relative order of the mask operands might change what happens
242 // to the high-bits of the mask (and, thus, the result).
244 // Cannot commute if it has a non-zero rotate count.
245 if (MI->getOperand(3).getImm() != 0)
248 // If we have a zero rotate count, we have:
250 // Op0 = (Op1 & ~M) | (Op2 & M)
252 // M = mask((ME+1)&31, (MB-1)&31)
253 // Op0 = (Op2 & ~M) | (Op1 & M)
256 unsigned Reg0 = MI->getOperand(0).getReg();
257 unsigned Reg1 = MI->getOperand(1).getReg();
258 unsigned Reg2 = MI->getOperand(2).getReg();
259 unsigned SubReg1 = MI->getOperand(1).getSubReg();
260 unsigned SubReg2 = MI->getOperand(2).getSubReg();
261 bool Reg1IsKill = MI->getOperand(1).isKill();
262 bool Reg2IsKill = MI->getOperand(2).isKill();
263 bool ChangeReg0 = false;
264 // If machine instrs are no longer in two-address forms, update
265 // destination register as well.
267 // Must be two address instruction!
268 assert(MI->getDesc().getOperandConstraint(0, MCOI::TIED_TO) &&
269 "Expecting a two-address instruction!");
270 assert(MI->getOperand(0).getSubReg() == SubReg1 && "Tied subreg mismatch");
276 unsigned MB = MI->getOperand(4).getImm();
277 unsigned ME = MI->getOperand(5).getImm();
280 // Create a new instruction.
281 unsigned Reg0 = ChangeReg0 ? Reg2 : MI->getOperand(0).getReg();
282 bool Reg0IsDead = MI->getOperand(0).isDead();
283 return BuildMI(MF, MI->getDebugLoc(), MI->getDesc())
284 .addReg(Reg0, RegState::Define | getDeadRegState(Reg0IsDead))
285 .addReg(Reg2, getKillRegState(Reg2IsKill))
286 .addReg(Reg1, getKillRegState(Reg1IsKill))
288 .addImm((MB-1) & 31);
292 MI->getOperand(0).setReg(Reg2);
293 MI->getOperand(0).setSubReg(SubReg2);
295 MI->getOperand(2).setReg(Reg1);
296 MI->getOperand(1).setReg(Reg2);
297 MI->getOperand(2).setSubReg(SubReg1);
298 MI->getOperand(1).setSubReg(SubReg2);
299 MI->getOperand(2).setIsKill(Reg1IsKill);
300 MI->getOperand(1).setIsKill(Reg2IsKill);
302 // Swap the mask around.
303 MI->getOperand(4).setImm((ME+1) & 31);
304 MI->getOperand(5).setImm((MB-1) & 31);
308 bool PPCInstrInfo::findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
309 unsigned &SrcOpIdx2) const {
310 // For VSX A-Type FMA instructions, it is the first two operands that can be
311 // commuted, however, because the non-encoded tied input operand is listed
312 // first, the operands to swap are actually the second and third.
314 int AltOpc = PPC::getAltVSXFMAOpcode(MI->getOpcode());
316 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
323 void PPCInstrInfo::insertNoop(MachineBasicBlock &MBB,
324 MachineBasicBlock::iterator MI) const {
325 // This function is used for scheduling, and the nop wanted here is the type
326 // that terminates dispatch groups on the POWER cores.
327 unsigned Directive = Subtarget.getDarwinDirective();
330 default: Opcode = PPC::NOP; break;
331 case PPC::DIR_PWR6: Opcode = PPC::NOP_GT_PWR6; break;
332 case PPC::DIR_PWR7: Opcode = PPC::NOP_GT_PWR7; break;
333 case PPC::DIR_PWR8: Opcode = PPC::NOP_GT_PWR7; break; /* FIXME: Update when P8 InstrScheduling model is ready */
337 BuildMI(MBB, MI, DL, get(Opcode));
340 /// getNoopForMachoTarget - Return the noop instruction to use for a noop.
341 void PPCInstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
342 NopInst.setOpcode(PPC::NOP);
346 // Note: If the condition register is set to CTR or CTR8 then this is a
347 // BDNZ (imm == 1) or BDZ (imm == 0) branch.
348 bool PPCInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
349 MachineBasicBlock *&FBB,
350 SmallVectorImpl<MachineOperand> &Cond,
351 bool AllowModify) const {
352 bool isPPC64 = Subtarget.isPPC64();
354 // If the block has no terminators, it just falls into the block after it.
355 MachineBasicBlock::iterator I = MBB.end();
356 if (I == MBB.begin())
359 while (I->isDebugValue()) {
360 if (I == MBB.begin())
364 if (!isUnpredicatedTerminator(I))
367 // Get the last instruction in the block.
368 MachineInstr *LastInst = I;
370 // If there is only one terminator instruction, process it.
371 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
372 if (LastInst->getOpcode() == PPC::B) {
373 if (!LastInst->getOperand(0).isMBB())
375 TBB = LastInst->getOperand(0).getMBB();
377 } else if (LastInst->getOpcode() == PPC::BCC) {
378 if (!LastInst->getOperand(2).isMBB())
380 // Block ends with fall-through condbranch.
381 TBB = LastInst->getOperand(2).getMBB();
382 Cond.push_back(LastInst->getOperand(0));
383 Cond.push_back(LastInst->getOperand(1));
385 } else if (LastInst->getOpcode() == PPC::BC) {
386 if (!LastInst->getOperand(1).isMBB())
388 // Block ends with fall-through condbranch.
389 TBB = LastInst->getOperand(1).getMBB();
390 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET));
391 Cond.push_back(LastInst->getOperand(0));
393 } else if (LastInst->getOpcode() == PPC::BCn) {
394 if (!LastInst->getOperand(1).isMBB())
396 // Block ends with fall-through condbranch.
397 TBB = LastInst->getOperand(1).getMBB();
398 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET));
399 Cond.push_back(LastInst->getOperand(0));
401 } else if (LastInst->getOpcode() == PPC::BDNZ8 ||
402 LastInst->getOpcode() == PPC::BDNZ) {
403 if (!LastInst->getOperand(0).isMBB())
405 if (DisableCTRLoopAnal)
407 TBB = LastInst->getOperand(0).getMBB();
408 Cond.push_back(MachineOperand::CreateImm(1));
409 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR,
412 } else if (LastInst->getOpcode() == PPC::BDZ8 ||
413 LastInst->getOpcode() == PPC::BDZ) {
414 if (!LastInst->getOperand(0).isMBB())
416 if (DisableCTRLoopAnal)
418 TBB = LastInst->getOperand(0).getMBB();
419 Cond.push_back(MachineOperand::CreateImm(0));
420 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR,
425 // Otherwise, don't know what this is.
429 // Get the instruction before it if it's a terminator.
430 MachineInstr *SecondLastInst = I;
432 // If there are three terminators, we don't know what sort of block this is.
433 if (SecondLastInst && I != MBB.begin() &&
434 isUnpredicatedTerminator(--I))
437 // If the block ends with PPC::B and PPC:BCC, handle it.
438 if (SecondLastInst->getOpcode() == PPC::BCC &&
439 LastInst->getOpcode() == PPC::B) {
440 if (!SecondLastInst->getOperand(2).isMBB() ||
441 !LastInst->getOperand(0).isMBB())
443 TBB = SecondLastInst->getOperand(2).getMBB();
444 Cond.push_back(SecondLastInst->getOperand(0));
445 Cond.push_back(SecondLastInst->getOperand(1));
446 FBB = LastInst->getOperand(0).getMBB();
448 } else if (SecondLastInst->getOpcode() == PPC::BC &&
449 LastInst->getOpcode() == PPC::B) {
450 if (!SecondLastInst->getOperand(1).isMBB() ||
451 !LastInst->getOperand(0).isMBB())
453 TBB = SecondLastInst->getOperand(1).getMBB();
454 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET));
455 Cond.push_back(SecondLastInst->getOperand(0));
456 FBB = LastInst->getOperand(0).getMBB();
458 } else if (SecondLastInst->getOpcode() == PPC::BCn &&
459 LastInst->getOpcode() == PPC::B) {
460 if (!SecondLastInst->getOperand(1).isMBB() ||
461 !LastInst->getOperand(0).isMBB())
463 TBB = SecondLastInst->getOperand(1).getMBB();
464 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET));
465 Cond.push_back(SecondLastInst->getOperand(0));
466 FBB = LastInst->getOperand(0).getMBB();
468 } else if ((SecondLastInst->getOpcode() == PPC::BDNZ8 ||
469 SecondLastInst->getOpcode() == PPC::BDNZ) &&
470 LastInst->getOpcode() == PPC::B) {
471 if (!SecondLastInst->getOperand(0).isMBB() ||
472 !LastInst->getOperand(0).isMBB())
474 if (DisableCTRLoopAnal)
476 TBB = SecondLastInst->getOperand(0).getMBB();
477 Cond.push_back(MachineOperand::CreateImm(1));
478 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR,
480 FBB = LastInst->getOperand(0).getMBB();
482 } else if ((SecondLastInst->getOpcode() == PPC::BDZ8 ||
483 SecondLastInst->getOpcode() == PPC::BDZ) &&
484 LastInst->getOpcode() == PPC::B) {
485 if (!SecondLastInst->getOperand(0).isMBB() ||
486 !LastInst->getOperand(0).isMBB())
488 if (DisableCTRLoopAnal)
490 TBB = SecondLastInst->getOperand(0).getMBB();
491 Cond.push_back(MachineOperand::CreateImm(0));
492 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR,
494 FBB = LastInst->getOperand(0).getMBB();
498 // If the block ends with two PPC:Bs, handle it. The second one is not
499 // executed, so remove it.
500 if (SecondLastInst->getOpcode() == PPC::B &&
501 LastInst->getOpcode() == PPC::B) {
502 if (!SecondLastInst->getOperand(0).isMBB())
504 TBB = SecondLastInst->getOperand(0).getMBB();
507 I->eraseFromParent();
511 // Otherwise, can't handle this.
515 unsigned PPCInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
516 MachineBasicBlock::iterator I = MBB.end();
517 if (I == MBB.begin()) return 0;
519 while (I->isDebugValue()) {
520 if (I == MBB.begin())
524 if (I->getOpcode() != PPC::B && I->getOpcode() != PPC::BCC &&
525 I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn &&
526 I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ &&
527 I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ)
530 // Remove the branch.
531 I->eraseFromParent();
535 if (I == MBB.begin()) return 1;
537 if (I->getOpcode() != PPC::BCC &&
538 I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn &&
539 I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ &&
540 I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ)
543 // Remove the branch.
544 I->eraseFromParent();
549 PPCInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
550 MachineBasicBlock *FBB,
551 ArrayRef<MachineOperand> Cond,
553 // Shouldn't be a fall through.
554 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
555 assert((Cond.size() == 2 || Cond.size() == 0) &&
556 "PPC branch conditions have two components!");
558 bool isPPC64 = Subtarget.isPPC64();
562 if (Cond.empty()) // Unconditional branch
563 BuildMI(&MBB, DL, get(PPC::B)).addMBB(TBB);
564 else if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8)
565 BuildMI(&MBB, DL, get(Cond[0].getImm() ?
566 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
567 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB);
568 else if (Cond[0].getImm() == PPC::PRED_BIT_SET)
569 BuildMI(&MBB, DL, get(PPC::BC)).addOperand(Cond[1]).addMBB(TBB);
570 else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET)
571 BuildMI(&MBB, DL, get(PPC::BCn)).addOperand(Cond[1]).addMBB(TBB);
572 else // Conditional branch
573 BuildMI(&MBB, DL, get(PPC::BCC))
574 .addImm(Cond[0].getImm()).addOperand(Cond[1]).addMBB(TBB);
578 // Two-way Conditional Branch.
579 if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8)
580 BuildMI(&MBB, DL, get(Cond[0].getImm() ?
581 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
582 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB);
583 else if (Cond[0].getImm() == PPC::PRED_BIT_SET)
584 BuildMI(&MBB, DL, get(PPC::BC)).addOperand(Cond[1]).addMBB(TBB);
585 else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET)
586 BuildMI(&MBB, DL, get(PPC::BCn)).addOperand(Cond[1]).addMBB(TBB);
588 BuildMI(&MBB, DL, get(PPC::BCC))
589 .addImm(Cond[0].getImm()).addOperand(Cond[1]).addMBB(TBB);
590 BuildMI(&MBB, DL, get(PPC::B)).addMBB(FBB);
595 bool PPCInstrInfo::canInsertSelect(const MachineBasicBlock &MBB,
596 ArrayRef<MachineOperand> Cond,
597 unsigned TrueReg, unsigned FalseReg,
598 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
599 if (!Subtarget.hasISEL())
602 if (Cond.size() != 2)
605 // If this is really a bdnz-like condition, then it cannot be turned into a
607 if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8)
610 // Check register classes.
611 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
612 const TargetRegisterClass *RC =
613 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
617 // isel is for regular integer GPRs only.
618 if (!PPC::GPRCRegClass.hasSubClassEq(RC) &&
619 !PPC::GPRC_NOR0RegClass.hasSubClassEq(RC) &&
620 !PPC::G8RCRegClass.hasSubClassEq(RC) &&
621 !PPC::G8RC_NOX0RegClass.hasSubClassEq(RC))
624 // FIXME: These numbers are for the A2, how well they work for other cores is
625 // an open question. On the A2, the isel instruction has a 2-cycle latency
626 // but single-cycle throughput. These numbers are used in combination with
627 // the MispredictPenalty setting from the active SchedMachineModel.
635 void PPCInstrInfo::insertSelect(MachineBasicBlock &MBB,
636 MachineBasicBlock::iterator MI, DebugLoc dl,
637 unsigned DestReg, ArrayRef<MachineOperand> Cond,
638 unsigned TrueReg, unsigned FalseReg) const {
639 assert(Cond.size() == 2 &&
640 "PPC branch conditions have two components!");
642 assert(Subtarget.hasISEL() &&
643 "Cannot insert select on target without ISEL support");
645 // Get the register classes.
646 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
647 const TargetRegisterClass *RC =
648 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
649 assert(RC && "TrueReg and FalseReg must have overlapping register classes");
651 bool Is64Bit = PPC::G8RCRegClass.hasSubClassEq(RC) ||
652 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC);
654 PPC::GPRCRegClass.hasSubClassEq(RC) ||
655 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) &&
656 "isel is for regular integer GPRs only");
658 unsigned OpCode = Is64Bit ? PPC::ISEL8 : PPC::ISEL;
659 unsigned SelectPred = Cond[0].getImm();
663 switch (SelectPred) {
664 default: llvm_unreachable("invalid predicate for isel");
665 case PPC::PRED_EQ: SubIdx = PPC::sub_eq; SwapOps = false; break;
666 case PPC::PRED_NE: SubIdx = PPC::sub_eq; SwapOps = true; break;
667 case PPC::PRED_LT: SubIdx = PPC::sub_lt; SwapOps = false; break;
668 case PPC::PRED_GE: SubIdx = PPC::sub_lt; SwapOps = true; break;
669 case PPC::PRED_GT: SubIdx = PPC::sub_gt; SwapOps = false; break;
670 case PPC::PRED_LE: SubIdx = PPC::sub_gt; SwapOps = true; break;
671 case PPC::PRED_UN: SubIdx = PPC::sub_un; SwapOps = false; break;
672 case PPC::PRED_NU: SubIdx = PPC::sub_un; SwapOps = true; break;
673 case PPC::PRED_BIT_SET: SubIdx = 0; SwapOps = false; break;
674 case PPC::PRED_BIT_UNSET: SubIdx = 0; SwapOps = true; break;
677 unsigned FirstReg = SwapOps ? FalseReg : TrueReg,
678 SecondReg = SwapOps ? TrueReg : FalseReg;
680 // The first input register of isel cannot be r0. If it is a member
681 // of a register class that can be r0, then copy it first (the
682 // register allocator should eliminate the copy).
683 if (MRI.getRegClass(FirstReg)->contains(PPC::R0) ||
684 MRI.getRegClass(FirstReg)->contains(PPC::X0)) {
685 const TargetRegisterClass *FirstRC =
686 MRI.getRegClass(FirstReg)->contains(PPC::X0) ?
687 &PPC::G8RC_NOX0RegClass : &PPC::GPRC_NOR0RegClass;
688 unsigned OldFirstReg = FirstReg;
689 FirstReg = MRI.createVirtualRegister(FirstRC);
690 BuildMI(MBB, MI, dl, get(TargetOpcode::COPY), FirstReg)
691 .addReg(OldFirstReg);
694 BuildMI(MBB, MI, dl, get(OpCode), DestReg)
695 .addReg(FirstReg).addReg(SecondReg)
696 .addReg(Cond[1].getReg(), 0, SubIdx);
699 static unsigned getCRBitValue(unsigned CRBit) {
701 if (CRBit == PPC::CR0LT || CRBit == PPC::CR1LT ||
702 CRBit == PPC::CR2LT || CRBit == PPC::CR3LT ||
703 CRBit == PPC::CR4LT || CRBit == PPC::CR5LT ||
704 CRBit == PPC::CR6LT || CRBit == PPC::CR7LT)
706 if (CRBit == PPC::CR0GT || CRBit == PPC::CR1GT ||
707 CRBit == PPC::CR2GT || CRBit == PPC::CR3GT ||
708 CRBit == PPC::CR4GT || CRBit == PPC::CR5GT ||
709 CRBit == PPC::CR6GT || CRBit == PPC::CR7GT)
711 if (CRBit == PPC::CR0EQ || CRBit == PPC::CR1EQ ||
712 CRBit == PPC::CR2EQ || CRBit == PPC::CR3EQ ||
713 CRBit == PPC::CR4EQ || CRBit == PPC::CR5EQ ||
714 CRBit == PPC::CR6EQ || CRBit == PPC::CR7EQ)
716 if (CRBit == PPC::CR0UN || CRBit == PPC::CR1UN ||
717 CRBit == PPC::CR2UN || CRBit == PPC::CR3UN ||
718 CRBit == PPC::CR4UN || CRBit == PPC::CR5UN ||
719 CRBit == PPC::CR6UN || CRBit == PPC::CR7UN)
722 assert(Ret != 4 && "Invalid CR bit register");
726 void PPCInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
727 MachineBasicBlock::iterator I, DebugLoc DL,
728 unsigned DestReg, unsigned SrcReg,
729 bool KillSrc) const {
730 // We can end up with self copies and similar things as a result of VSX copy
731 // legalization. Promote them here.
732 const TargetRegisterInfo *TRI = &getRegisterInfo();
733 if (PPC::F8RCRegClass.contains(DestReg) &&
734 PPC::VSRCRegClass.contains(SrcReg)) {
736 TRI->getMatchingSuperReg(DestReg, PPC::sub_64, &PPC::VSRCRegClass);
738 if (VSXSelfCopyCrash && SrcReg == SuperReg)
739 llvm_unreachable("nop VSX copy");
742 } else if (PPC::VRRCRegClass.contains(DestReg) &&
743 PPC::VSRCRegClass.contains(SrcReg)) {
745 TRI->getMatchingSuperReg(DestReg, PPC::sub_128, &PPC::VSRCRegClass);
747 if (VSXSelfCopyCrash && SrcReg == SuperReg)
748 llvm_unreachable("nop VSX copy");
751 } else if (PPC::F8RCRegClass.contains(SrcReg) &&
752 PPC::VSRCRegClass.contains(DestReg)) {
754 TRI->getMatchingSuperReg(SrcReg, PPC::sub_64, &PPC::VSRCRegClass);
756 if (VSXSelfCopyCrash && DestReg == SuperReg)
757 llvm_unreachable("nop VSX copy");
760 } else if (PPC::VRRCRegClass.contains(SrcReg) &&
761 PPC::VSRCRegClass.contains(DestReg)) {
763 TRI->getMatchingSuperReg(SrcReg, PPC::sub_128, &PPC::VSRCRegClass);
765 if (VSXSelfCopyCrash && DestReg == SuperReg)
766 llvm_unreachable("nop VSX copy");
771 // Different class register copy
772 if (PPC::CRBITRCRegClass.contains(SrcReg) &&
773 PPC::GPRCRegClass.contains(DestReg)) {
774 unsigned CRReg = getCRFromCRBit(SrcReg);
775 BuildMI(MBB, I, DL, get(PPC::MFOCRF), DestReg)
776 .addReg(CRReg), getKillRegState(KillSrc);
777 // Rotate the CR bit in the CR fields to be the least significant bit and
778 // then mask with 0x1 (MB = ME = 31).
779 BuildMI(MBB, I, DL, get(PPC::RLWINM), DestReg)
780 .addReg(DestReg, RegState::Kill)
781 .addImm(TRI->getEncodingValue(CRReg) * 4 + (4 - getCRBitValue(SrcReg)))
785 } else if (PPC::CRRCRegClass.contains(SrcReg) &&
786 PPC::G8RCRegClass.contains(DestReg)) {
787 BuildMI(MBB, I, DL, get(PPC::MFOCRF8), DestReg)
788 .addReg(SrcReg), getKillRegState(KillSrc);
790 } else if (PPC::CRRCRegClass.contains(SrcReg) &&
791 PPC::GPRCRegClass.contains(DestReg)) {
792 BuildMI(MBB, I, DL, get(PPC::MFOCRF), DestReg)
793 .addReg(SrcReg), getKillRegState(KillSrc);
798 if (PPC::GPRCRegClass.contains(DestReg, SrcReg))
800 else if (PPC::G8RCRegClass.contains(DestReg, SrcReg))
802 else if (PPC::F4RCRegClass.contains(DestReg, SrcReg))
804 else if (PPC::CRRCRegClass.contains(DestReg, SrcReg))
806 else if (PPC::VRRCRegClass.contains(DestReg, SrcReg))
808 else if (PPC::VSRCRegClass.contains(DestReg, SrcReg))
809 // There are two different ways this can be done:
810 // 1. xxlor : This has lower latency (on the P7), 2 cycles, but can only
811 // issue in VSU pipeline 0.
812 // 2. xmovdp/xmovsp: This has higher latency (on the P7), 6 cycles, but
813 // can go to either pipeline.
814 // We'll always use xxlor here, because in practically all cases where
815 // copies are generated, they are close enough to some use that the
816 // lower-latency form is preferable.
818 else if (PPC::VSFRCRegClass.contains(DestReg, SrcReg) ||
819 PPC::VSSRCRegClass.contains(DestReg, SrcReg))
821 else if (PPC::QFRCRegClass.contains(DestReg, SrcReg))
823 else if (PPC::QSRCRegClass.contains(DestReg, SrcReg))
825 else if (PPC::QBRCRegClass.contains(DestReg, SrcReg))
827 else if (PPC::CRBITRCRegClass.contains(DestReg, SrcReg))
830 llvm_unreachable("Impossible reg-to-reg copy");
832 const MCInstrDesc &MCID = get(Opc);
833 if (MCID.getNumOperands() == 3)
834 BuildMI(MBB, I, DL, MCID, DestReg)
835 .addReg(SrcReg).addReg(SrcReg, getKillRegState(KillSrc));
837 BuildMI(MBB, I, DL, MCID, DestReg).addReg(SrcReg, getKillRegState(KillSrc));
840 // This function returns true if a CR spill is necessary and false otherwise.
842 PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF,
843 unsigned SrcReg, bool isKill,
845 const TargetRegisterClass *RC,
846 SmallVectorImpl<MachineInstr*> &NewMIs,
847 bool &NonRI, bool &SpillsVRS) const{
848 // Note: If additional store instructions are added here,
849 // update isStoreToStackSlot.
852 if (PPC::GPRCRegClass.hasSubClassEq(RC) ||
853 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) {
854 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STW))
856 getKillRegState(isKill)),
858 } else if (PPC::G8RCRegClass.hasSubClassEq(RC) ||
859 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) {
860 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STD))
862 getKillRegState(isKill)),
864 } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) {
865 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STFD))
867 getKillRegState(isKill)),
869 } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) {
870 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STFS))
872 getKillRegState(isKill)),
874 } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) {
875 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_CR))
877 getKillRegState(isKill)),
880 } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) {
881 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_CRBIT))
883 getKillRegState(isKill)),
886 } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) {
887 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STVX))
889 getKillRegState(isKill)),
892 } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) {
893 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXVD2X))
895 getKillRegState(isKill)),
898 } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) {
899 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXSDX))
901 getKillRegState(isKill)),
904 } else if (PPC::VSSRCRegClass.hasSubClassEq(RC)) {
905 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXSSPX))
907 getKillRegState(isKill)),
910 } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) {
911 assert(Subtarget.isDarwin() &&
912 "VRSAVE only needs spill/restore on Darwin");
913 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_VRSAVE))
915 getKillRegState(isKill)),
918 } else if (PPC::QFRCRegClass.hasSubClassEq(RC)) {
919 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVSTFDX))
921 getKillRegState(isKill)),
924 } else if (PPC::QSRCRegClass.hasSubClassEq(RC)) {
925 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVSTFSXs))
927 getKillRegState(isKill)),
930 } else if (PPC::QBRCRegClass.hasSubClassEq(RC)) {
931 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVSTFDXb))
933 getKillRegState(isKill)),
937 llvm_unreachable("Unknown regclass!");
944 PPCInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
945 MachineBasicBlock::iterator MI,
946 unsigned SrcReg, bool isKill, int FrameIdx,
947 const TargetRegisterClass *RC,
948 const TargetRegisterInfo *TRI) const {
949 MachineFunction &MF = *MBB.getParent();
950 SmallVector<MachineInstr*, 4> NewMIs;
952 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
953 FuncInfo->setHasSpills();
955 bool NonRI = false, SpillsVRS = false;
956 if (StoreRegToStackSlot(MF, SrcReg, isKill, FrameIdx, RC, NewMIs,
958 FuncInfo->setSpillsCR();
961 FuncInfo->setSpillsVRSAVE();
964 FuncInfo->setHasNonRISpills();
966 for (unsigned i = 0, e = NewMIs.size(); i != e; ++i)
967 MBB.insert(MI, NewMIs[i]);
969 const MachineFrameInfo &MFI = *MF.getFrameInfo();
970 MachineMemOperand *MMO =
971 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
972 MachineMemOperand::MOStore,
973 MFI.getObjectSize(FrameIdx),
974 MFI.getObjectAlignment(FrameIdx));
975 NewMIs.back()->addMemOperand(MF, MMO);
979 PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL,
980 unsigned DestReg, int FrameIdx,
981 const TargetRegisterClass *RC,
982 SmallVectorImpl<MachineInstr*> &NewMIs,
983 bool &NonRI, bool &SpillsVRS) const{
984 // Note: If additional load instructions are added here,
985 // update isLoadFromStackSlot.
987 if (PPC::GPRCRegClass.hasSubClassEq(RC) ||
988 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) {
989 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LWZ),
990 DestReg), FrameIdx));
991 } else if (PPC::G8RCRegClass.hasSubClassEq(RC) ||
992 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) {
993 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LD), DestReg),
995 } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) {
996 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFD), DestReg),
998 } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) {
999 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFS), DestReg),
1001 } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) {
1002 NewMIs.push_back(addFrameReference(BuildMI(MF, DL,
1003 get(PPC::RESTORE_CR), DestReg),
1006 } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) {
1007 NewMIs.push_back(addFrameReference(BuildMI(MF, DL,
1008 get(PPC::RESTORE_CRBIT), DestReg),
1011 } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) {
1012 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LVX), DestReg),
1015 } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) {
1016 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXVD2X), DestReg),
1019 } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) {
1020 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXSDX), DestReg),
1023 } else if (PPC::VSSRCRegClass.hasSubClassEq(RC)) {
1024 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXSSPX), DestReg),
1027 } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) {
1028 assert(Subtarget.isDarwin() &&
1029 "VRSAVE only needs spill/restore on Darwin");
1030 NewMIs.push_back(addFrameReference(BuildMI(MF, DL,
1031 get(PPC::RESTORE_VRSAVE),
1035 } else if (PPC::QFRCRegClass.hasSubClassEq(RC)) {
1036 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVLFDX), DestReg),
1039 } else if (PPC::QSRCRegClass.hasSubClassEq(RC)) {
1040 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVLFSXs), DestReg),
1043 } else if (PPC::QBRCRegClass.hasSubClassEq(RC)) {
1044 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVLFDXb), DestReg),
1048 llvm_unreachable("Unknown regclass!");
1055 PPCInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1056 MachineBasicBlock::iterator MI,
1057 unsigned DestReg, int FrameIdx,
1058 const TargetRegisterClass *RC,
1059 const TargetRegisterInfo *TRI) const {
1060 MachineFunction &MF = *MBB.getParent();
1061 SmallVector<MachineInstr*, 4> NewMIs;
1063 if (MI != MBB.end()) DL = MI->getDebugLoc();
1065 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1066 FuncInfo->setHasSpills();
1068 bool NonRI = false, SpillsVRS = false;
1069 if (LoadRegFromStackSlot(MF, DL, DestReg, FrameIdx, RC, NewMIs,
1071 FuncInfo->setSpillsCR();
1074 FuncInfo->setSpillsVRSAVE();
1077 FuncInfo->setHasNonRISpills();
1079 for (unsigned i = 0, e = NewMIs.size(); i != e; ++i)
1080 MBB.insert(MI, NewMIs[i]);
1082 const MachineFrameInfo &MFI = *MF.getFrameInfo();
1083 MachineMemOperand *MMO =
1084 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
1085 MachineMemOperand::MOLoad,
1086 MFI.getObjectSize(FrameIdx),
1087 MFI.getObjectAlignment(FrameIdx));
1088 NewMIs.back()->addMemOperand(MF, MMO);
1092 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
1093 assert(Cond.size() == 2 && "Invalid PPC branch opcode!");
1094 if (Cond[1].getReg() == PPC::CTR8 || Cond[1].getReg() == PPC::CTR)
1095 Cond[0].setImm(Cond[0].getImm() == 0 ? 1 : 0);
1097 // Leave the CR# the same, but invert the condition.
1098 Cond[0].setImm(PPC::InvertPredicate((PPC::Predicate)Cond[0].getImm()));
1102 bool PPCInstrInfo::FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI,
1103 unsigned Reg, MachineRegisterInfo *MRI) const {
1104 // For some instructions, it is legal to fold ZERO into the RA register field.
1105 // A zero immediate should always be loaded with a single li.
1106 unsigned DefOpc = DefMI->getOpcode();
1107 if (DefOpc != PPC::LI && DefOpc != PPC::LI8)
1109 if (!DefMI->getOperand(1).isImm())
1111 if (DefMI->getOperand(1).getImm() != 0)
1114 // Note that we cannot here invert the arguments of an isel in order to fold
1115 // a ZERO into what is presented as the second argument. All we have here
1116 // is the condition bit, and that might come from a CR-logical bit operation.
1118 const MCInstrDesc &UseMCID = UseMI->getDesc();
1120 // Only fold into real machine instructions.
1121 if (UseMCID.isPseudo())
1125 for (UseIdx = 0; UseIdx < UseMI->getNumOperands(); ++UseIdx)
1126 if (UseMI->getOperand(UseIdx).isReg() &&
1127 UseMI->getOperand(UseIdx).getReg() == Reg)
1130 assert(UseIdx < UseMI->getNumOperands() && "Cannot find Reg in UseMI");
1131 assert(UseIdx < UseMCID.getNumOperands() && "No operand description for Reg");
1133 const MCOperandInfo *UseInfo = &UseMCID.OpInfo[UseIdx];
1135 // We can fold the zero if this register requires a GPRC_NOR0/G8RC_NOX0
1136 // register (which might also be specified as a pointer class kind).
1137 if (UseInfo->isLookupPtrRegClass()) {
1138 if (UseInfo->RegClass /* Kind */ != 1)
1141 if (UseInfo->RegClass != PPC::GPRC_NOR0RegClassID &&
1142 UseInfo->RegClass != PPC::G8RC_NOX0RegClassID)
1146 // Make sure this is not tied to an output register (or otherwise
1147 // constrained). This is true for ST?UX registers, for example, which
1148 // are tied to their output registers.
1149 if (UseInfo->Constraints != 0)
1153 if (UseInfo->isLookupPtrRegClass()) {
1154 bool isPPC64 = Subtarget.isPPC64();
1155 ZeroReg = isPPC64 ? PPC::ZERO8 : PPC::ZERO;
1157 ZeroReg = UseInfo->RegClass == PPC::G8RC_NOX0RegClassID ?
1158 PPC::ZERO8 : PPC::ZERO;
1161 bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
1162 UseMI->getOperand(UseIdx).setReg(ZeroReg);
1165 DefMI->eraseFromParent();
1170 static bool MBBDefinesCTR(MachineBasicBlock &MBB) {
1171 for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end();
1173 if (I->definesRegister(PPC::CTR) || I->definesRegister(PPC::CTR8))
1178 // We should make sure that, if we're going to predicate both sides of a
1179 // condition (a diamond), that both sides don't define the counter register. We
1180 // can predicate counter-decrement-based branches, but while that predicates
1181 // the branching, it does not predicate the counter decrement. If we tried to
1182 // merge the triangle into one predicated block, we'd decrement the counter
1184 bool PPCInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
1185 unsigned NumT, unsigned ExtraT,
1186 MachineBasicBlock &FMBB,
1187 unsigned NumF, unsigned ExtraF,
1188 const BranchProbability &Probability) const {
1189 return !(MBBDefinesCTR(TMBB) && MBBDefinesCTR(FMBB));
1193 bool PPCInstrInfo::isPredicated(const MachineInstr *MI) const {
1194 // The predicated branches are identified by their type, not really by the
1195 // explicit presence of a predicate. Furthermore, some of them can be
1196 // predicated more than once. Because if conversion won't try to predicate
1197 // any instruction which already claims to be predicated (by returning true
1198 // here), always return false. In doing so, we let isPredicable() be the
1199 // final word on whether not the instruction can be (further) predicated.
1204 bool PPCInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
1205 if (!MI->isTerminator())
1208 // Conditional branch is a special case.
1209 if (MI->isBranch() && !MI->isBarrier())
1212 return !isPredicated(MI);
1215 bool PPCInstrInfo::PredicateInstruction(MachineInstr *MI,
1216 ArrayRef<MachineOperand> Pred) const {
1217 unsigned OpC = MI->getOpcode();
1218 if (OpC == PPC::BLR || OpC == PPC::BLR8) {
1219 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) {
1220 bool isPPC64 = Subtarget.isPPC64();
1221 MI->setDesc(get(Pred[0].getImm() ?
1222 (isPPC64 ? PPC::BDNZLR8 : PPC::BDNZLR) :
1223 (isPPC64 ? PPC::BDZLR8 : PPC::BDZLR)));
1224 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
1225 MI->setDesc(get(PPC::BCLR));
1226 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
1227 .addReg(Pred[1].getReg());
1228 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
1229 MI->setDesc(get(PPC::BCLRn));
1230 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
1231 .addReg(Pred[1].getReg());
1233 MI->setDesc(get(PPC::BCCLR));
1234 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
1235 .addImm(Pred[0].getImm())
1236 .addReg(Pred[1].getReg());
1240 } else if (OpC == PPC::B) {
1241 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) {
1242 bool isPPC64 = Subtarget.isPPC64();
1243 MI->setDesc(get(Pred[0].getImm() ?
1244 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
1245 (isPPC64 ? PPC::BDZ8 : PPC::BDZ)));
1246 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
1247 MachineBasicBlock *MBB = MI->getOperand(0).getMBB();
1248 MI->RemoveOperand(0);
1250 MI->setDesc(get(PPC::BC));
1251 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
1252 .addReg(Pred[1].getReg())
1254 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
1255 MachineBasicBlock *MBB = MI->getOperand(0).getMBB();
1256 MI->RemoveOperand(0);
1258 MI->setDesc(get(PPC::BCn));
1259 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
1260 .addReg(Pred[1].getReg())
1263 MachineBasicBlock *MBB = MI->getOperand(0).getMBB();
1264 MI->RemoveOperand(0);
1266 MI->setDesc(get(PPC::BCC));
1267 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
1268 .addImm(Pred[0].getImm())
1269 .addReg(Pred[1].getReg())
1274 } else if (OpC == PPC::BCTR || OpC == PPC::BCTR8 ||
1275 OpC == PPC::BCTRL || OpC == PPC::BCTRL8) {
1276 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR)
1277 llvm_unreachable("Cannot predicate bctr[l] on the ctr register");
1279 bool setLR = OpC == PPC::BCTRL || OpC == PPC::BCTRL8;
1280 bool isPPC64 = Subtarget.isPPC64();
1282 if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
1283 MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8 : PPC::BCCTR8) :
1284 (setLR ? PPC::BCCTRL : PPC::BCCTR)));
1285 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
1286 .addReg(Pred[1].getReg());
1288 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
1289 MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8n : PPC::BCCTR8n) :
1290 (setLR ? PPC::BCCTRLn : PPC::BCCTRn)));
1291 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
1292 .addReg(Pred[1].getReg());
1296 MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCCTRL8 : PPC::BCCCTR8) :
1297 (setLR ? PPC::BCCCTRL : PPC::BCCCTR)));
1298 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
1299 .addImm(Pred[0].getImm())
1300 .addReg(Pred[1].getReg());
1307 bool PPCInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
1308 ArrayRef<MachineOperand> Pred2) const {
1309 assert(Pred1.size() == 2 && "Invalid PPC first predicate");
1310 assert(Pred2.size() == 2 && "Invalid PPC second predicate");
1312 if (Pred1[1].getReg() == PPC::CTR8 || Pred1[1].getReg() == PPC::CTR)
1314 if (Pred2[1].getReg() == PPC::CTR8 || Pred2[1].getReg() == PPC::CTR)
1317 // P1 can only subsume P2 if they test the same condition register.
1318 if (Pred1[1].getReg() != Pred2[1].getReg())
1321 PPC::Predicate P1 = (PPC::Predicate) Pred1[0].getImm();
1322 PPC::Predicate P2 = (PPC::Predicate) Pred2[0].getImm();
1327 // Does P1 subsume P2, e.g. GE subsumes GT.
1328 if (P1 == PPC::PRED_LE &&
1329 (P2 == PPC::PRED_LT || P2 == PPC::PRED_EQ))
1331 if (P1 == PPC::PRED_GE &&
1332 (P2 == PPC::PRED_GT || P2 == PPC::PRED_EQ))
1338 bool PPCInstrInfo::DefinesPredicate(MachineInstr *MI,
1339 std::vector<MachineOperand> &Pred) const {
1340 // Note: At the present time, the contents of Pred from this function is
1341 // unused by IfConversion. This implementation follows ARM by pushing the
1342 // CR-defining operand. Because the 'DZ' and 'DNZ' count as types of
1343 // predicate, instructions defining CTR or CTR8 are also included as
1344 // predicate-defining instructions.
1346 const TargetRegisterClass *RCs[] =
1347 { &PPC::CRRCRegClass, &PPC::CRBITRCRegClass,
1348 &PPC::CTRRCRegClass, &PPC::CTRRC8RegClass };
1351 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1352 const MachineOperand &MO = MI->getOperand(i);
1353 for (unsigned c = 0; c < array_lengthof(RCs) && !Found; ++c) {
1354 const TargetRegisterClass *RC = RCs[c];
1356 if (MO.isDef() && RC->contains(MO.getReg())) {
1360 } else if (MO.isRegMask()) {
1361 for (TargetRegisterClass::iterator I = RC->begin(),
1362 IE = RC->end(); I != IE; ++I)
1363 if (MO.clobbersPhysReg(*I)) {
1374 bool PPCInstrInfo::isPredicable(MachineInstr *MI) const {
1375 unsigned OpC = MI->getOpcode();
1390 bool PPCInstrInfo::analyzeCompare(const MachineInstr *MI,
1391 unsigned &SrcReg, unsigned &SrcReg2,
1392 int &Mask, int &Value) const {
1393 unsigned Opc = MI->getOpcode();
1396 default: return false;
1401 SrcReg = MI->getOperand(1).getReg();
1403 Value = MI->getOperand(2).getImm();
1412 SrcReg = MI->getOperand(1).getReg();
1413 SrcReg2 = MI->getOperand(2).getReg();
1418 bool PPCInstrInfo::optimizeCompareInstr(MachineInstr *CmpInstr,
1419 unsigned SrcReg, unsigned SrcReg2,
1420 int Mask, int Value,
1421 const MachineRegisterInfo *MRI) const {
1425 int OpC = CmpInstr->getOpcode();
1426 unsigned CRReg = CmpInstr->getOperand(0).getReg();
1428 // FP record forms set CR1 based on the execption status bits, not a
1429 // comparison with zero.
1430 if (OpC == PPC::FCMPUS || OpC == PPC::FCMPUD)
1433 // The record forms set the condition register based on a signed comparison
1434 // with zero (so says the ISA manual). This is not as straightforward as it
1435 // seems, however, because this is always a 64-bit comparison on PPC64, even
1436 // for instructions that are 32-bit in nature (like slw for example).
1437 // So, on PPC32, for unsigned comparisons, we can use the record forms only
1438 // for equality checks (as those don't depend on the sign). On PPC64,
1439 // we are restricted to equality for unsigned 64-bit comparisons and for
1440 // signed 32-bit comparisons the applicability is more restricted.
1441 bool isPPC64 = Subtarget.isPPC64();
1442 bool is32BitSignedCompare = OpC == PPC::CMPWI || OpC == PPC::CMPW;
1443 bool is32BitUnsignedCompare = OpC == PPC::CMPLWI || OpC == PPC::CMPLW;
1444 bool is64BitUnsignedCompare = OpC == PPC::CMPLDI || OpC == PPC::CMPLD;
1446 // Get the unique definition of SrcReg.
1447 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
1448 if (!MI) return false;
1449 int MIOpC = MI->getOpcode();
1451 bool equalityOnly = false;
1454 if (is32BitSignedCompare) {
1455 // We can perform this optimization only if MI is sign-extending.
1456 if (MIOpC == PPC::SRAW || MIOpC == PPC::SRAWo ||
1457 MIOpC == PPC::SRAWI || MIOpC == PPC::SRAWIo ||
1458 MIOpC == PPC::EXTSB || MIOpC == PPC::EXTSBo ||
1459 MIOpC == PPC::EXTSH || MIOpC == PPC::EXTSHo ||
1460 MIOpC == PPC::EXTSW || MIOpC == PPC::EXTSWo) {
1464 } else if (is32BitUnsignedCompare) {
1465 // We can perform this optimization, equality only, if MI is
1467 if (MIOpC == PPC::CNTLZW || MIOpC == PPC::CNTLZWo ||
1468 MIOpC == PPC::SLW || MIOpC == PPC::SLWo ||
1469 MIOpC == PPC::SRW || MIOpC == PPC::SRWo) {
1471 equalityOnly = true;
1475 equalityOnly = is64BitUnsignedCompare;
1477 equalityOnly = is32BitUnsignedCompare;
1480 // We need to check the uses of the condition register in order to reject
1481 // non-equality comparisons.
1482 for (MachineRegisterInfo::use_instr_iterator I =MRI->use_instr_begin(CRReg),
1483 IE = MRI->use_instr_end(); I != IE; ++I) {
1484 MachineInstr *UseMI = &*I;
1485 if (UseMI->getOpcode() == PPC::BCC) {
1486 unsigned Pred = UseMI->getOperand(0).getImm();
1487 if (Pred != PPC::PRED_EQ && Pred != PPC::PRED_NE)
1489 } else if (UseMI->getOpcode() == PPC::ISEL ||
1490 UseMI->getOpcode() == PPC::ISEL8) {
1491 unsigned SubIdx = UseMI->getOperand(3).getSubReg();
1492 if (SubIdx != PPC::sub_eq)
1499 MachineBasicBlock::iterator I = CmpInstr;
1501 // Scan forward to find the first use of the compare.
1502 for (MachineBasicBlock::iterator EL = CmpInstr->getParent()->end();
1504 bool FoundUse = false;
1505 for (MachineRegisterInfo::use_instr_iterator J =MRI->use_instr_begin(CRReg),
1506 JE = MRI->use_instr_end(); J != JE; ++J)
1516 // There are two possible candidates which can be changed to set CR[01].
1517 // One is MI, the other is a SUB instruction.
1518 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
1519 MachineInstr *Sub = nullptr;
1521 // MI is not a candidate for CMPrr.
1523 // FIXME: Conservatively refuse to convert an instruction which isn't in the
1524 // same BB as the comparison. This is to allow the check below to avoid calls
1525 // (and other explicit clobbers); instead we should really check for these
1526 // more explicitly (in at least a few predecessors).
1527 else if (MI->getParent() != CmpInstr->getParent() || Value != 0) {
1528 // PPC does not have a record-form SUBri.
1533 const TargetRegisterInfo *TRI = &getRegisterInfo();
1536 // Get ready to iterate backward from CmpInstr.
1537 MachineBasicBlock::iterator E = MI,
1538 B = CmpInstr->getParent()->begin();
1540 for (; I != E && !noSub; --I) {
1541 const MachineInstr &Instr = *I;
1542 unsigned IOpC = Instr.getOpcode();
1544 if (&*I != CmpInstr && (
1545 Instr.modifiesRegister(PPC::CR0, TRI) ||
1546 Instr.readsRegister(PPC::CR0, TRI)))
1547 // This instruction modifies or uses the record condition register after
1548 // the one we want to change. While we could do this transformation, it
1549 // would likely not be profitable. This transformation removes one
1550 // instruction, and so even forcing RA to generate one move probably
1551 // makes it unprofitable.
1554 // Check whether CmpInstr can be made redundant by the current instruction.
1555 if ((OpC == PPC::CMPW || OpC == PPC::CMPLW ||
1556 OpC == PPC::CMPD || OpC == PPC::CMPLD) &&
1557 (IOpC == PPC::SUBF || IOpC == PPC::SUBF8) &&
1558 ((Instr.getOperand(1).getReg() == SrcReg &&
1559 Instr.getOperand(2).getReg() == SrcReg2) ||
1560 (Instr.getOperand(1).getReg() == SrcReg2 &&
1561 Instr.getOperand(2).getReg() == SrcReg))) {
1567 // The 'and' is below the comparison instruction.
1571 // Return false if no candidates exist.
1575 // The single candidate is called MI.
1579 MIOpC = MI->getOpcode();
1580 if (MIOpC == PPC::ANDIo || MIOpC == PPC::ANDIo8)
1583 NewOpC = PPC::getRecordFormOpcode(MIOpC);
1584 if (NewOpC == -1 && PPC::getNonRecordFormOpcode(MIOpC) != -1)
1588 // FIXME: On the non-embedded POWER architectures, only some of the record
1589 // forms are fast, and we should use only the fast ones.
1591 // The defining instruction has a record form (or is already a record
1592 // form). It is possible, however, that we'll need to reverse the condition
1593 // code of the users.
1597 SmallVector<std::pair<MachineOperand*, PPC::Predicate>, 4> PredsToUpdate;
1598 SmallVector<std::pair<MachineOperand*, unsigned>, 4> SubRegsToUpdate;
1600 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based on CMP
1601 // needs to be updated to be based on SUB. Push the condition code
1602 // operands to OperandsToUpdate. If it is safe to remove CmpInstr, the
1603 // condition code of these operands will be modified.
1604 bool ShouldSwap = false;
1606 ShouldSwap = SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
1607 Sub->getOperand(2).getReg() == SrcReg;
1609 // The operands to subf are the opposite of sub, so only in the fixed-point
1610 // case, invert the order.
1611 ShouldSwap = !ShouldSwap;
1615 for (MachineRegisterInfo::use_instr_iterator
1616 I = MRI->use_instr_begin(CRReg), IE = MRI->use_instr_end();
1618 MachineInstr *UseMI = &*I;
1619 if (UseMI->getOpcode() == PPC::BCC) {
1620 PPC::Predicate Pred = (PPC::Predicate) UseMI->getOperand(0).getImm();
1621 assert((!equalityOnly ||
1622 Pred == PPC::PRED_EQ || Pred == PPC::PRED_NE) &&
1623 "Invalid predicate for equality-only optimization");
1624 PredsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(0)),
1625 PPC::getSwappedPredicate(Pred)));
1626 } else if (UseMI->getOpcode() == PPC::ISEL ||
1627 UseMI->getOpcode() == PPC::ISEL8) {
1628 unsigned NewSubReg = UseMI->getOperand(3).getSubReg();
1629 assert((!equalityOnly || NewSubReg == PPC::sub_eq) &&
1630 "Invalid CR bit for equality-only optimization");
1632 if (NewSubReg == PPC::sub_lt)
1633 NewSubReg = PPC::sub_gt;
1634 else if (NewSubReg == PPC::sub_gt)
1635 NewSubReg = PPC::sub_lt;
1637 SubRegsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(3)),
1639 } else // We need to abort on a user we don't understand.
1643 // Create a new virtual register to hold the value of the CR set by the
1644 // record-form instruction. If the instruction was not previously in
1645 // record form, then set the kill flag on the CR.
1646 CmpInstr->eraseFromParent();
1648 MachineBasicBlock::iterator MII = MI;
1649 BuildMI(*MI->getParent(), std::next(MII), MI->getDebugLoc(),
1650 get(TargetOpcode::COPY), CRReg)
1651 .addReg(PPC::CR0, MIOpC != NewOpC ? RegState::Kill : 0);
1653 if (MIOpC != NewOpC) {
1654 // We need to be careful here: we're replacing one instruction with
1655 // another, and we need to make sure that we get all of the right
1656 // implicit uses and defs. On the other hand, the caller may be holding
1657 // an iterator to this instruction, and so we can't delete it (this is
1658 // specifically the case if this is the instruction directly after the
1661 const MCInstrDesc &NewDesc = get(NewOpC);
1662 MI->setDesc(NewDesc);
1664 if (NewDesc.ImplicitDefs)
1665 for (const uint16_t *ImpDefs = NewDesc.getImplicitDefs();
1666 *ImpDefs; ++ImpDefs)
1667 if (!MI->definesRegister(*ImpDefs))
1668 MI->addOperand(*MI->getParent()->getParent(),
1669 MachineOperand::CreateReg(*ImpDefs, true, true));
1670 if (NewDesc.ImplicitUses)
1671 for (const uint16_t *ImpUses = NewDesc.getImplicitUses();
1672 *ImpUses; ++ImpUses)
1673 if (!MI->readsRegister(*ImpUses))
1674 MI->addOperand(*MI->getParent()->getParent(),
1675 MachineOperand::CreateReg(*ImpUses, false, true));
1678 // Modify the condition code of operands in OperandsToUpdate.
1679 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to
1680 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
1681 for (unsigned i = 0, e = PredsToUpdate.size(); i < e; i++)
1682 PredsToUpdate[i].first->setImm(PredsToUpdate[i].second);
1684 for (unsigned i = 0, e = SubRegsToUpdate.size(); i < e; i++)
1685 SubRegsToUpdate[i].first->setSubReg(SubRegsToUpdate[i].second);
1690 /// GetInstSize - Return the number of bytes of code the specified
1691 /// instruction may be. This returns the maximum number of bytes.
1693 unsigned PPCInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
1694 unsigned Opcode = MI->getOpcode();
1696 if (Opcode == PPC::INLINEASM) {
1697 const MachineFunction *MF = MI->getParent()->getParent();
1698 const char *AsmStr = MI->getOperand(0).getSymbolName();
1699 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
1700 } else if (Opcode == TargetOpcode::STACKMAP) {
1701 return MI->getOperand(1).getImm();
1702 } else if (Opcode == TargetOpcode::PATCHPOINT) {
1703 PatchPointOpers Opers(MI);
1704 return Opers.getMetaOper(PatchPointOpers::NBytesPos).getImm();
1706 const MCInstrDesc &Desc = get(Opcode);
1707 return Desc.getSize();