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/LiveIntervals.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"
49 STATISTIC(NumStoreSPILLVSRRCAsVec,
50 "Number of spillvsrrc spilled to stack as vec");
51 STATISTIC(NumStoreSPILLVSRRCAsGpr,
52 "Number of spillvsrrc spilled to stack as gpr");
53 STATISTIC(NumGPRtoVSRSpill, "Number of gpr spills to spillvsrrc");
54 STATISTIC(CmpIselsConverted,
55 "Number of ISELs that depend on comparison of constants converted");
56 STATISTIC(MissedConvertibleImmediateInstrs,
57 "Number of compare-immediate instructions fed by constants");
58 STATISTIC(NumRcRotatesConvertedToRcAnd,
59 "Number of record-form rotates converted to record-form andi");
62 opt<bool> DisableCTRLoopAnal("disable-ppc-ctrloop-analysis", cl::Hidden,
63 cl::desc("Disable analysis for CTR loops"));
65 static cl::opt<bool> DisableCmpOpt("disable-ppc-cmp-opt",
66 cl::desc("Disable compare instruction optimization"), cl::Hidden);
68 static cl::opt<bool> VSXSelfCopyCrash("crash-on-ppc-vsx-self-copy",
69 cl::desc("Causes the backend to crash instead of generating a nop VSX copy"),
73 UseOldLatencyCalc("ppc-old-latency-calc", cl::Hidden,
74 cl::desc("Use the old (incorrect) instruction latency calculation"));
76 // Index into the OpcodesForSpill array.
86 SOK_VectorFloat8Spill,
87 SOK_VectorFloat4Spill,
95 SOK_LastOpcodeSpill // This must be last on the enum.
98 // Pin the vtable to this file.
99 void PPCInstrInfo::anchor() {}
101 PPCInstrInfo::PPCInstrInfo(PPCSubtarget &STI)
102 : PPCGenInstrInfo(PPC::ADJCALLSTACKDOWN, PPC::ADJCALLSTACKUP,
103 /* CatchRetOpcode */ -1,
104 STI.isPPC64() ? PPC::BLR8 : PPC::BLR),
105 Subtarget(STI), RI(STI.getTargetMachine()) {}
107 /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for
108 /// this target when scheduling the DAG.
109 ScheduleHazardRecognizer *
110 PPCInstrInfo::CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
111 const ScheduleDAG *DAG) const {
113 static_cast<const PPCSubtarget *>(STI)->getDarwinDirective();
114 if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2 ||
115 Directive == PPC::DIR_E500mc || Directive == PPC::DIR_E5500) {
116 const InstrItineraryData *II =
117 static_cast<const PPCSubtarget *>(STI)->getInstrItineraryData();
118 return new ScoreboardHazardRecognizer(II, DAG);
121 return TargetInstrInfo::CreateTargetHazardRecognizer(STI, DAG);
124 /// CreateTargetPostRAHazardRecognizer - Return the postRA hazard recognizer
125 /// to use for this target when scheduling the DAG.
126 ScheduleHazardRecognizer *
127 PPCInstrInfo::CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
128 const ScheduleDAG *DAG) const {
130 DAG->MF.getSubtarget<PPCSubtarget>().getDarwinDirective();
132 // FIXME: Leaving this as-is until we have POWER9 scheduling info
133 if (Directive == PPC::DIR_PWR7 || Directive == PPC::DIR_PWR8)
134 return new PPCDispatchGroupSBHazardRecognizer(II, DAG);
136 // Most subtargets use a PPC970 recognizer.
137 if (Directive != PPC::DIR_440 && Directive != PPC::DIR_A2 &&
138 Directive != PPC::DIR_E500mc && Directive != PPC::DIR_E5500) {
139 assert(DAG->TII && "No InstrInfo?");
141 return new PPCHazardRecognizer970(*DAG);
144 return new ScoreboardHazardRecognizer(II, DAG);
147 unsigned PPCInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
148 const MachineInstr &MI,
149 unsigned *PredCost) const {
150 if (!ItinData || UseOldLatencyCalc)
151 return PPCGenInstrInfo::getInstrLatency(ItinData, MI, PredCost);
153 // The default implementation of getInstrLatency calls getStageLatency, but
154 // getStageLatency does not do the right thing for us. While we have
155 // itinerary, most cores are fully pipelined, and so the itineraries only
156 // express the first part of the pipeline, not every stage. Instead, we need
157 // to use the listed output operand cycle number (using operand 0 here, which
160 unsigned Latency = 1;
161 unsigned DefClass = MI.getDesc().getSchedClass();
162 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
163 const MachineOperand &MO = MI.getOperand(i);
164 if (!MO.isReg() || !MO.isDef() || MO.isImplicit())
167 int Cycle = ItinData->getOperandCycle(DefClass, i);
171 Latency = std::max(Latency, (unsigned) Cycle);
177 int PPCInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
178 const MachineInstr &DefMI, unsigned DefIdx,
179 const MachineInstr &UseMI,
180 unsigned UseIdx) const {
181 int Latency = PPCGenInstrInfo::getOperandLatency(ItinData, DefMI, DefIdx,
184 if (!DefMI.getParent())
187 const MachineOperand &DefMO = DefMI.getOperand(DefIdx);
188 unsigned Reg = DefMO.getReg();
191 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
192 const MachineRegisterInfo *MRI =
193 &DefMI.getParent()->getParent()->getRegInfo();
194 IsRegCR = MRI->getRegClass(Reg)->hasSuperClassEq(&PPC::CRRCRegClass) ||
195 MRI->getRegClass(Reg)->hasSuperClassEq(&PPC::CRBITRCRegClass);
197 IsRegCR = PPC::CRRCRegClass.contains(Reg) ||
198 PPC::CRBITRCRegClass.contains(Reg);
201 if (UseMI.isBranch() && IsRegCR) {
203 Latency = getInstrLatency(ItinData, DefMI);
205 // On some cores, there is an additional delay between writing to a condition
206 // register, and using it from a branch.
207 unsigned Directive = Subtarget.getDarwinDirective();
221 // FIXME: Is this needed for POWER9?
230 // This function does not list all associative and commutative operations, but
231 // only those worth feeding through the machine combiner in an attempt to
232 // reduce the critical path. Mostly, this means floating-point operations,
233 // because they have high latencies (compared to other operations, such and
234 // and/or, which are also associative and commutative, but have low latencies).
235 bool PPCInstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst) const {
236 switch (Inst.getOpcode()) {
269 bool PPCInstrInfo::getMachineCombinerPatterns(
271 SmallVectorImpl<MachineCombinerPattern> &Patterns) const {
272 // Using the machine combiner in this way is potentially expensive, so
273 // restrict to when aggressive optimizations are desired.
274 if (Subtarget.getTargetMachine().getOptLevel() != CodeGenOpt::Aggressive)
277 // FP reassociation is only legal when we don't need strict IEEE semantics.
278 if (!Root.getParent()->getParent()->getTarget().Options.UnsafeFPMath)
281 return TargetInstrInfo::getMachineCombinerPatterns(Root, Patterns);
284 // Detect 32 -> 64-bit extensions where we may reuse the low sub-register.
285 bool PPCInstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
286 unsigned &SrcReg, unsigned &DstReg,
287 unsigned &SubIdx) const {
288 switch (MI.getOpcode()) {
289 default: return false;
292 case PPC::EXTSW_32_64:
293 SrcReg = MI.getOperand(1).getReg();
294 DstReg = MI.getOperand(0).getReg();
295 SubIdx = PPC::sub_32;
300 unsigned PPCInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
301 int &FrameIndex) const {
302 unsigned Opcode = MI.getOpcode();
303 const unsigned *OpcodesForSpill = getLoadOpcodesForSpillArray();
304 const unsigned *End = OpcodesForSpill + SOK_LastOpcodeSpill;
306 if (End != std::find(OpcodesForSpill, End, Opcode)) {
307 // Check for the operands added by addFrameReference (the immediate is the
308 // offset which defaults to 0).
309 if (MI.getOperand(1).isImm() && !MI.getOperand(1).getImm() &&
310 MI.getOperand(2).isFI()) {
311 FrameIndex = MI.getOperand(2).getIndex();
312 return MI.getOperand(0).getReg();
318 // For opcodes with the ReMaterializable flag set, this function is called to
319 // verify the instruction is really rematable.
320 bool PPCInstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
321 AliasAnalysis *AA) const {
322 switch (MI.getOpcode()) {
324 // This function should only be called for opcodes with the ReMaterializable
326 llvm_unreachable("Unknown rematerializable operation!");
333 case PPC::ADDIStocHA:
335 case PPC::LOAD_STACK_GUARD:
341 unsigned PPCInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
342 int &FrameIndex) const {
343 unsigned Opcode = MI.getOpcode();
344 const unsigned *OpcodesForSpill = getStoreOpcodesForSpillArray();
345 const unsigned *End = OpcodesForSpill + SOK_LastOpcodeSpill;
347 if (End != std::find(OpcodesForSpill, End, Opcode)) {
348 if (MI.getOperand(1).isImm() && !MI.getOperand(1).getImm() &&
349 MI.getOperand(2).isFI()) {
350 FrameIndex = MI.getOperand(2).getIndex();
351 return MI.getOperand(0).getReg();
357 MachineInstr *PPCInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
359 unsigned OpIdx2) const {
360 MachineFunction &MF = *MI.getParent()->getParent();
362 // Normal instructions can be commuted the obvious way.
363 if (MI.getOpcode() != PPC::RLWIMI && MI.getOpcode() != PPC::RLWIMIo)
364 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
365 // Note that RLWIMI can be commuted as a 32-bit instruction, but not as a
366 // 64-bit instruction (so we don't handle PPC::RLWIMI8 here), because
367 // changing the relative order of the mask operands might change what happens
368 // to the high-bits of the mask (and, thus, the result).
370 // Cannot commute if it has a non-zero rotate count.
371 if (MI.getOperand(3).getImm() != 0)
374 // If we have a zero rotate count, we have:
376 // Op0 = (Op1 & ~M) | (Op2 & M)
378 // M = mask((ME+1)&31, (MB-1)&31)
379 // Op0 = (Op2 & ~M) | (Op1 & M)
382 assert(((OpIdx1 == 1 && OpIdx2 == 2) || (OpIdx1 == 2 && OpIdx2 == 1)) &&
383 "Only the operands 1 and 2 can be swapped in RLSIMI/RLWIMIo.");
384 unsigned Reg0 = MI.getOperand(0).getReg();
385 unsigned Reg1 = MI.getOperand(1).getReg();
386 unsigned Reg2 = MI.getOperand(2).getReg();
387 unsigned SubReg1 = MI.getOperand(1).getSubReg();
388 unsigned SubReg2 = MI.getOperand(2).getSubReg();
389 bool Reg1IsKill = MI.getOperand(1).isKill();
390 bool Reg2IsKill = MI.getOperand(2).isKill();
391 bool ChangeReg0 = false;
392 // If machine instrs are no longer in two-address forms, update
393 // destination register as well.
395 // Must be two address instruction!
396 assert(MI.getDesc().getOperandConstraint(0, MCOI::TIED_TO) &&
397 "Expecting a two-address instruction!");
398 assert(MI.getOperand(0).getSubReg() == SubReg1 && "Tied subreg mismatch");
404 unsigned MB = MI.getOperand(4).getImm();
405 unsigned ME = MI.getOperand(5).getImm();
407 // We can't commute a trivial mask (there is no way to represent an all-zero
409 if (MB == 0 && ME == 31)
413 // Create a new instruction.
414 unsigned Reg0 = ChangeReg0 ? Reg2 : MI.getOperand(0).getReg();
415 bool Reg0IsDead = MI.getOperand(0).isDead();
416 return BuildMI(MF, MI.getDebugLoc(), MI.getDesc())
417 .addReg(Reg0, RegState::Define | getDeadRegState(Reg0IsDead))
418 .addReg(Reg2, getKillRegState(Reg2IsKill))
419 .addReg(Reg1, getKillRegState(Reg1IsKill))
420 .addImm((ME + 1) & 31)
421 .addImm((MB - 1) & 31);
425 MI.getOperand(0).setReg(Reg2);
426 MI.getOperand(0).setSubReg(SubReg2);
428 MI.getOperand(2).setReg(Reg1);
429 MI.getOperand(1).setReg(Reg2);
430 MI.getOperand(2).setSubReg(SubReg1);
431 MI.getOperand(1).setSubReg(SubReg2);
432 MI.getOperand(2).setIsKill(Reg1IsKill);
433 MI.getOperand(1).setIsKill(Reg2IsKill);
435 // Swap the mask around.
436 MI.getOperand(4).setImm((ME + 1) & 31);
437 MI.getOperand(5).setImm((MB - 1) & 31);
441 bool PPCInstrInfo::findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1,
442 unsigned &SrcOpIdx2) const {
443 // For VSX A-Type FMA instructions, it is the first two operands that can be
444 // commuted, however, because the non-encoded tied input operand is listed
445 // first, the operands to swap are actually the second and third.
447 int AltOpc = PPC::getAltVSXFMAOpcode(MI.getOpcode());
449 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
451 // The commutable operand indices are 2 and 3. Return them in SrcOpIdx1
453 return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 2, 3);
456 void PPCInstrInfo::insertNoop(MachineBasicBlock &MBB,
457 MachineBasicBlock::iterator MI) const {
458 // This function is used for scheduling, and the nop wanted here is the type
459 // that terminates dispatch groups on the POWER cores.
460 unsigned Directive = Subtarget.getDarwinDirective();
463 default: Opcode = PPC::NOP; break;
464 case PPC::DIR_PWR6: Opcode = PPC::NOP_GT_PWR6; break;
465 case PPC::DIR_PWR7: Opcode = PPC::NOP_GT_PWR7; break;
466 case PPC::DIR_PWR8: Opcode = PPC::NOP_GT_PWR7; break; /* FIXME: Update when P8 InstrScheduling model is ready */
467 // FIXME: Update when POWER9 scheduling model is ready.
468 case PPC::DIR_PWR9: Opcode = PPC::NOP_GT_PWR7; break;
472 BuildMI(MBB, MI, DL, get(Opcode));
475 /// Return the noop instruction to use for a noop.
476 void PPCInstrInfo::getNoop(MCInst &NopInst) const {
477 NopInst.setOpcode(PPC::NOP);
481 // Note: If the condition register is set to CTR or CTR8 then this is a
482 // BDNZ (imm == 1) or BDZ (imm == 0) branch.
483 bool PPCInstrInfo::analyzeBranch(MachineBasicBlock &MBB,
484 MachineBasicBlock *&TBB,
485 MachineBasicBlock *&FBB,
486 SmallVectorImpl<MachineOperand> &Cond,
487 bool AllowModify) const {
488 bool isPPC64 = Subtarget.isPPC64();
490 // If the block has no terminators, it just falls into the block after it.
491 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
495 if (!isUnpredicatedTerminator(*I))
499 // If the BB ends with an unconditional branch to the fallthrough BB,
500 // we eliminate the branch instruction.
501 if (I->getOpcode() == PPC::B &&
502 MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
503 I->eraseFromParent();
505 // We update iterator after deleting the last branch.
506 I = MBB.getLastNonDebugInstr();
507 if (I == MBB.end() || !isUnpredicatedTerminator(*I))
512 // Get the last instruction in the block.
513 MachineInstr &LastInst = *I;
515 // If there is only one terminator instruction, process it.
516 if (I == MBB.begin() || !isUnpredicatedTerminator(*--I)) {
517 if (LastInst.getOpcode() == PPC::B) {
518 if (!LastInst.getOperand(0).isMBB())
520 TBB = LastInst.getOperand(0).getMBB();
522 } else if (LastInst.getOpcode() == PPC::BCC) {
523 if (!LastInst.getOperand(2).isMBB())
525 // Block ends with fall-through condbranch.
526 TBB = LastInst.getOperand(2).getMBB();
527 Cond.push_back(LastInst.getOperand(0));
528 Cond.push_back(LastInst.getOperand(1));
530 } else if (LastInst.getOpcode() == PPC::BC) {
531 if (!LastInst.getOperand(1).isMBB())
533 // Block ends with fall-through condbranch.
534 TBB = LastInst.getOperand(1).getMBB();
535 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET));
536 Cond.push_back(LastInst.getOperand(0));
538 } else if (LastInst.getOpcode() == PPC::BCn) {
539 if (!LastInst.getOperand(1).isMBB())
541 // Block ends with fall-through condbranch.
542 TBB = LastInst.getOperand(1).getMBB();
543 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET));
544 Cond.push_back(LastInst.getOperand(0));
546 } else if (LastInst.getOpcode() == PPC::BDNZ8 ||
547 LastInst.getOpcode() == PPC::BDNZ) {
548 if (!LastInst.getOperand(0).isMBB())
550 if (DisableCTRLoopAnal)
552 TBB = LastInst.getOperand(0).getMBB();
553 Cond.push_back(MachineOperand::CreateImm(1));
554 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR,
557 } else if (LastInst.getOpcode() == PPC::BDZ8 ||
558 LastInst.getOpcode() == PPC::BDZ) {
559 if (!LastInst.getOperand(0).isMBB())
561 if (DisableCTRLoopAnal)
563 TBB = LastInst.getOperand(0).getMBB();
564 Cond.push_back(MachineOperand::CreateImm(0));
565 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR,
570 // Otherwise, don't know what this is.
574 // Get the instruction before it if it's a terminator.
575 MachineInstr &SecondLastInst = *I;
577 // If there are three terminators, we don't know what sort of block this is.
578 if (I != MBB.begin() && isUnpredicatedTerminator(*--I))
581 // If the block ends with PPC::B and PPC:BCC, handle it.
582 if (SecondLastInst.getOpcode() == PPC::BCC &&
583 LastInst.getOpcode() == PPC::B) {
584 if (!SecondLastInst.getOperand(2).isMBB() ||
585 !LastInst.getOperand(0).isMBB())
587 TBB = SecondLastInst.getOperand(2).getMBB();
588 Cond.push_back(SecondLastInst.getOperand(0));
589 Cond.push_back(SecondLastInst.getOperand(1));
590 FBB = LastInst.getOperand(0).getMBB();
592 } else if (SecondLastInst.getOpcode() == PPC::BC &&
593 LastInst.getOpcode() == PPC::B) {
594 if (!SecondLastInst.getOperand(1).isMBB() ||
595 !LastInst.getOperand(0).isMBB())
597 TBB = SecondLastInst.getOperand(1).getMBB();
598 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET));
599 Cond.push_back(SecondLastInst.getOperand(0));
600 FBB = LastInst.getOperand(0).getMBB();
602 } else if (SecondLastInst.getOpcode() == PPC::BCn &&
603 LastInst.getOpcode() == PPC::B) {
604 if (!SecondLastInst.getOperand(1).isMBB() ||
605 !LastInst.getOperand(0).isMBB())
607 TBB = SecondLastInst.getOperand(1).getMBB();
608 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET));
609 Cond.push_back(SecondLastInst.getOperand(0));
610 FBB = LastInst.getOperand(0).getMBB();
612 } else if ((SecondLastInst.getOpcode() == PPC::BDNZ8 ||
613 SecondLastInst.getOpcode() == PPC::BDNZ) &&
614 LastInst.getOpcode() == PPC::B) {
615 if (!SecondLastInst.getOperand(0).isMBB() ||
616 !LastInst.getOperand(0).isMBB())
618 if (DisableCTRLoopAnal)
620 TBB = SecondLastInst.getOperand(0).getMBB();
621 Cond.push_back(MachineOperand::CreateImm(1));
622 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR,
624 FBB = LastInst.getOperand(0).getMBB();
626 } else if ((SecondLastInst.getOpcode() == PPC::BDZ8 ||
627 SecondLastInst.getOpcode() == PPC::BDZ) &&
628 LastInst.getOpcode() == PPC::B) {
629 if (!SecondLastInst.getOperand(0).isMBB() ||
630 !LastInst.getOperand(0).isMBB())
632 if (DisableCTRLoopAnal)
634 TBB = SecondLastInst.getOperand(0).getMBB();
635 Cond.push_back(MachineOperand::CreateImm(0));
636 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR,
638 FBB = LastInst.getOperand(0).getMBB();
642 // If the block ends with two PPC:Bs, handle it. The second one is not
643 // executed, so remove it.
644 if (SecondLastInst.getOpcode() == PPC::B && LastInst.getOpcode() == PPC::B) {
645 if (!SecondLastInst.getOperand(0).isMBB())
647 TBB = SecondLastInst.getOperand(0).getMBB();
650 I->eraseFromParent();
654 // Otherwise, can't handle this.
658 unsigned PPCInstrInfo::removeBranch(MachineBasicBlock &MBB,
659 int *BytesRemoved) const {
660 assert(!BytesRemoved && "code size not handled");
662 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
666 if (I->getOpcode() != PPC::B && I->getOpcode() != PPC::BCC &&
667 I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn &&
668 I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ &&
669 I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ)
672 // Remove the branch.
673 I->eraseFromParent();
677 if (I == MBB.begin()) return 1;
679 if (I->getOpcode() != PPC::BCC &&
680 I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn &&
681 I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ &&
682 I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ)
685 // Remove the branch.
686 I->eraseFromParent();
690 unsigned PPCInstrInfo::insertBranch(MachineBasicBlock &MBB,
691 MachineBasicBlock *TBB,
692 MachineBasicBlock *FBB,
693 ArrayRef<MachineOperand> Cond,
695 int *BytesAdded) const {
696 // Shouldn't be a fall through.
697 assert(TBB && "insertBranch must not be told to insert a fallthrough");
698 assert((Cond.size() == 2 || Cond.size() == 0) &&
699 "PPC branch conditions have two components!");
700 assert(!BytesAdded && "code size not handled");
702 bool isPPC64 = Subtarget.isPPC64();
706 if (Cond.empty()) // Unconditional branch
707 BuildMI(&MBB, DL, get(PPC::B)).addMBB(TBB);
708 else if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8)
709 BuildMI(&MBB, DL, get(Cond[0].getImm() ?
710 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
711 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB);
712 else if (Cond[0].getImm() == PPC::PRED_BIT_SET)
713 BuildMI(&MBB, DL, get(PPC::BC)).add(Cond[1]).addMBB(TBB);
714 else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET)
715 BuildMI(&MBB, DL, get(PPC::BCn)).add(Cond[1]).addMBB(TBB);
716 else // Conditional branch
717 BuildMI(&MBB, DL, get(PPC::BCC))
718 .addImm(Cond[0].getImm())
724 // Two-way Conditional Branch.
725 if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8)
726 BuildMI(&MBB, DL, get(Cond[0].getImm() ?
727 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
728 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB);
729 else if (Cond[0].getImm() == PPC::PRED_BIT_SET)
730 BuildMI(&MBB, DL, get(PPC::BC)).add(Cond[1]).addMBB(TBB);
731 else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET)
732 BuildMI(&MBB, DL, get(PPC::BCn)).add(Cond[1]).addMBB(TBB);
734 BuildMI(&MBB, DL, get(PPC::BCC))
735 .addImm(Cond[0].getImm())
738 BuildMI(&MBB, DL, get(PPC::B)).addMBB(FBB);
743 bool PPCInstrInfo::canInsertSelect(const MachineBasicBlock &MBB,
744 ArrayRef<MachineOperand> Cond,
745 unsigned TrueReg, unsigned FalseReg,
746 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
747 if (Cond.size() != 2)
750 // If this is really a bdnz-like condition, then it cannot be turned into a
752 if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8)
755 // Check register classes.
756 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
757 const TargetRegisterClass *RC =
758 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
762 // isel is for regular integer GPRs only.
763 if (!PPC::GPRCRegClass.hasSubClassEq(RC) &&
764 !PPC::GPRC_NOR0RegClass.hasSubClassEq(RC) &&
765 !PPC::G8RCRegClass.hasSubClassEq(RC) &&
766 !PPC::G8RC_NOX0RegClass.hasSubClassEq(RC))
769 // FIXME: These numbers are for the A2, how well they work for other cores is
770 // an open question. On the A2, the isel instruction has a 2-cycle latency
771 // but single-cycle throughput. These numbers are used in combination with
772 // the MispredictPenalty setting from the active SchedMachineModel.
780 void PPCInstrInfo::insertSelect(MachineBasicBlock &MBB,
781 MachineBasicBlock::iterator MI,
782 const DebugLoc &dl, unsigned DestReg,
783 ArrayRef<MachineOperand> Cond, unsigned TrueReg,
784 unsigned FalseReg) const {
785 assert(Cond.size() == 2 &&
786 "PPC branch conditions have two components!");
788 // Get the register classes.
789 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
790 const TargetRegisterClass *RC =
791 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
792 assert(RC && "TrueReg and FalseReg must have overlapping register classes");
794 bool Is64Bit = PPC::G8RCRegClass.hasSubClassEq(RC) ||
795 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC);
797 PPC::GPRCRegClass.hasSubClassEq(RC) ||
798 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) &&
799 "isel is for regular integer GPRs only");
801 unsigned OpCode = Is64Bit ? PPC::ISEL8 : PPC::ISEL;
802 auto SelectPred = static_cast<PPC::Predicate>(Cond[0].getImm());
805 bool SwapOps = false;
806 switch (SelectPred) {
808 case PPC::PRED_EQ_MINUS:
809 case PPC::PRED_EQ_PLUS:
810 SubIdx = PPC::sub_eq; SwapOps = false; break;
812 case PPC::PRED_NE_MINUS:
813 case PPC::PRED_NE_PLUS:
814 SubIdx = PPC::sub_eq; SwapOps = true; break;
816 case PPC::PRED_LT_MINUS:
817 case PPC::PRED_LT_PLUS:
818 SubIdx = PPC::sub_lt; SwapOps = false; break;
820 case PPC::PRED_GE_MINUS:
821 case PPC::PRED_GE_PLUS:
822 SubIdx = PPC::sub_lt; SwapOps = true; break;
824 case PPC::PRED_GT_MINUS:
825 case PPC::PRED_GT_PLUS:
826 SubIdx = PPC::sub_gt; SwapOps = false; break;
828 case PPC::PRED_LE_MINUS:
829 case PPC::PRED_LE_PLUS:
830 SubIdx = PPC::sub_gt; SwapOps = true; break;
832 case PPC::PRED_UN_MINUS:
833 case PPC::PRED_UN_PLUS:
834 SubIdx = PPC::sub_un; SwapOps = false; break;
836 case PPC::PRED_NU_MINUS:
837 case PPC::PRED_NU_PLUS:
838 SubIdx = PPC::sub_un; SwapOps = true; break;
839 case PPC::PRED_BIT_SET: SubIdx = 0; SwapOps = false; break;
840 case PPC::PRED_BIT_UNSET: SubIdx = 0; SwapOps = true; break;
843 unsigned FirstReg = SwapOps ? FalseReg : TrueReg,
844 SecondReg = SwapOps ? TrueReg : FalseReg;
846 // The first input register of isel cannot be r0. If it is a member
847 // of a register class that can be r0, then copy it first (the
848 // register allocator should eliminate the copy).
849 if (MRI.getRegClass(FirstReg)->contains(PPC::R0) ||
850 MRI.getRegClass(FirstReg)->contains(PPC::X0)) {
851 const TargetRegisterClass *FirstRC =
852 MRI.getRegClass(FirstReg)->contains(PPC::X0) ?
853 &PPC::G8RC_NOX0RegClass : &PPC::GPRC_NOR0RegClass;
854 unsigned OldFirstReg = FirstReg;
855 FirstReg = MRI.createVirtualRegister(FirstRC);
856 BuildMI(MBB, MI, dl, get(TargetOpcode::COPY), FirstReg)
857 .addReg(OldFirstReg);
860 BuildMI(MBB, MI, dl, get(OpCode), DestReg)
861 .addReg(FirstReg).addReg(SecondReg)
862 .addReg(Cond[1].getReg(), 0, SubIdx);
865 static unsigned getCRBitValue(unsigned CRBit) {
867 if (CRBit == PPC::CR0LT || CRBit == PPC::CR1LT ||
868 CRBit == PPC::CR2LT || CRBit == PPC::CR3LT ||
869 CRBit == PPC::CR4LT || CRBit == PPC::CR5LT ||
870 CRBit == PPC::CR6LT || CRBit == PPC::CR7LT)
872 if (CRBit == PPC::CR0GT || CRBit == PPC::CR1GT ||
873 CRBit == PPC::CR2GT || CRBit == PPC::CR3GT ||
874 CRBit == PPC::CR4GT || CRBit == PPC::CR5GT ||
875 CRBit == PPC::CR6GT || CRBit == PPC::CR7GT)
877 if (CRBit == PPC::CR0EQ || CRBit == PPC::CR1EQ ||
878 CRBit == PPC::CR2EQ || CRBit == PPC::CR3EQ ||
879 CRBit == PPC::CR4EQ || CRBit == PPC::CR5EQ ||
880 CRBit == PPC::CR6EQ || CRBit == PPC::CR7EQ)
882 if (CRBit == PPC::CR0UN || CRBit == PPC::CR1UN ||
883 CRBit == PPC::CR2UN || CRBit == PPC::CR3UN ||
884 CRBit == PPC::CR4UN || CRBit == PPC::CR5UN ||
885 CRBit == PPC::CR6UN || CRBit == PPC::CR7UN)
888 assert(Ret != 4 && "Invalid CR bit register");
892 void PPCInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
893 MachineBasicBlock::iterator I,
894 const DebugLoc &DL, unsigned DestReg,
895 unsigned SrcReg, bool KillSrc) const {
896 // We can end up with self copies and similar things as a result of VSX copy
897 // legalization. Promote them here.
898 const TargetRegisterInfo *TRI = &getRegisterInfo();
899 if (PPC::F8RCRegClass.contains(DestReg) &&
900 PPC::VSRCRegClass.contains(SrcReg)) {
902 TRI->getMatchingSuperReg(DestReg, PPC::sub_64, &PPC::VSRCRegClass);
904 if (VSXSelfCopyCrash && SrcReg == SuperReg)
905 llvm_unreachable("nop VSX copy");
908 } else if (PPC::F8RCRegClass.contains(SrcReg) &&
909 PPC::VSRCRegClass.contains(DestReg)) {
911 TRI->getMatchingSuperReg(SrcReg, PPC::sub_64, &PPC::VSRCRegClass);
913 if (VSXSelfCopyCrash && DestReg == SuperReg)
914 llvm_unreachable("nop VSX copy");
919 // Different class register copy
920 if (PPC::CRBITRCRegClass.contains(SrcReg) &&
921 PPC::GPRCRegClass.contains(DestReg)) {
922 unsigned CRReg = getCRFromCRBit(SrcReg);
923 BuildMI(MBB, I, DL, get(PPC::MFOCRF), DestReg).addReg(CRReg);
924 getKillRegState(KillSrc);
925 // Rotate the CR bit in the CR fields to be the least significant bit and
926 // then mask with 0x1 (MB = ME = 31).
927 BuildMI(MBB, I, DL, get(PPC::RLWINM), DestReg)
928 .addReg(DestReg, RegState::Kill)
929 .addImm(TRI->getEncodingValue(CRReg) * 4 + (4 - getCRBitValue(SrcReg)))
933 } else if (PPC::CRRCRegClass.contains(SrcReg) &&
934 PPC::G8RCRegClass.contains(DestReg)) {
935 BuildMI(MBB, I, DL, get(PPC::MFOCRF8), DestReg).addReg(SrcReg);
936 getKillRegState(KillSrc);
938 } else if (PPC::CRRCRegClass.contains(SrcReg) &&
939 PPC::GPRCRegClass.contains(DestReg)) {
940 BuildMI(MBB, I, DL, get(PPC::MFOCRF), DestReg).addReg(SrcReg);
941 getKillRegState(KillSrc);
943 } else if (PPC::G8RCRegClass.contains(SrcReg) &&
944 PPC::VSFRCRegClass.contains(DestReg)) {
945 BuildMI(MBB, I, DL, get(PPC::MTVSRD), DestReg).addReg(SrcReg);
947 getKillRegState(KillSrc);
949 } else if (PPC::VSFRCRegClass.contains(SrcReg) &&
950 PPC::G8RCRegClass.contains(DestReg)) {
951 BuildMI(MBB, I, DL, get(PPC::MFVSRD), DestReg).addReg(SrcReg);
952 getKillRegState(KillSrc);
954 } else if (PPC::SPERCRegClass.contains(SrcReg) &&
955 PPC::SPE4RCRegClass.contains(DestReg)) {
956 BuildMI(MBB, I, DL, get(PPC::EFSCFD), DestReg).addReg(SrcReg);
957 getKillRegState(KillSrc);
959 } else if (PPC::SPE4RCRegClass.contains(SrcReg) &&
960 PPC::SPERCRegClass.contains(DestReg)) {
961 BuildMI(MBB, I, DL, get(PPC::EFDCFS), DestReg).addReg(SrcReg);
962 getKillRegState(KillSrc);
968 if (PPC::GPRCRegClass.contains(DestReg, SrcReg))
970 else if (PPC::G8RCRegClass.contains(DestReg, SrcReg))
972 else if (PPC::F4RCRegClass.contains(DestReg, SrcReg))
974 else if (PPC::CRRCRegClass.contains(DestReg, SrcReg))
976 else if (PPC::VRRCRegClass.contains(DestReg, SrcReg))
978 else if (PPC::VSRCRegClass.contains(DestReg, SrcReg))
979 // There are two different ways this can be done:
980 // 1. xxlor : This has lower latency (on the P7), 2 cycles, but can only
981 // issue in VSU pipeline 0.
982 // 2. xmovdp/xmovsp: This has higher latency (on the P7), 6 cycles, but
983 // can go to either pipeline.
984 // We'll always use xxlor here, because in practically all cases where
985 // copies are generated, they are close enough to some use that the
986 // lower-latency form is preferable.
988 else if (PPC::VSFRCRegClass.contains(DestReg, SrcReg) ||
989 PPC::VSSRCRegClass.contains(DestReg, SrcReg))
990 Opc = (Subtarget.hasP9Vector()) ? PPC::XSCPSGNDP : PPC::XXLORf;
991 else if (PPC::QFRCRegClass.contains(DestReg, SrcReg))
993 else if (PPC::QSRCRegClass.contains(DestReg, SrcReg))
995 else if (PPC::QBRCRegClass.contains(DestReg, SrcReg))
997 else if (PPC::CRBITRCRegClass.contains(DestReg, SrcReg))
999 else if (PPC::SPERCRegClass.contains(DestReg, SrcReg))
1002 llvm_unreachable("Impossible reg-to-reg copy");
1004 const MCInstrDesc &MCID = get(Opc);
1005 if (MCID.getNumOperands() == 3)
1006 BuildMI(MBB, I, DL, MCID, DestReg)
1007 .addReg(SrcReg).addReg(SrcReg, getKillRegState(KillSrc));
1009 BuildMI(MBB, I, DL, MCID, DestReg).addReg(SrcReg, getKillRegState(KillSrc));
1012 unsigned PPCInstrInfo::getStoreOpcodeForSpill(unsigned Reg,
1013 const TargetRegisterClass *RC)
1015 const unsigned *OpcodesForSpill = getStoreOpcodesForSpillArray();
1016 int OpcodeIndex = 0;
1018 if (RC != nullptr) {
1019 if (PPC::GPRCRegClass.hasSubClassEq(RC) ||
1020 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) {
1021 OpcodeIndex = SOK_Int4Spill;
1022 } else if (PPC::G8RCRegClass.hasSubClassEq(RC) ||
1023 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) {
1024 OpcodeIndex = SOK_Int8Spill;
1025 } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) {
1026 OpcodeIndex = SOK_Float8Spill;
1027 } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) {
1028 OpcodeIndex = SOK_Float4Spill;
1029 } else if (PPC::SPERCRegClass.hasSubClassEq(RC)) {
1030 OpcodeIndex = SOK_SPESpill;
1031 } else if (PPC::SPE4RCRegClass.hasSubClassEq(RC)) {
1032 OpcodeIndex = SOK_SPE4Spill;
1033 } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) {
1034 OpcodeIndex = SOK_CRSpill;
1035 } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) {
1036 OpcodeIndex = SOK_CRBitSpill;
1037 } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) {
1038 OpcodeIndex = SOK_VRVectorSpill;
1039 } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) {
1040 OpcodeIndex = SOK_VSXVectorSpill;
1041 } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) {
1042 OpcodeIndex = SOK_VectorFloat8Spill;
1043 } else if (PPC::VSSRCRegClass.hasSubClassEq(RC)) {
1044 OpcodeIndex = SOK_VectorFloat4Spill;
1045 } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) {
1046 OpcodeIndex = SOK_VRSaveSpill;
1047 } else if (PPC::QFRCRegClass.hasSubClassEq(RC)) {
1048 OpcodeIndex = SOK_QuadFloat8Spill;
1049 } else if (PPC::QSRCRegClass.hasSubClassEq(RC)) {
1050 OpcodeIndex = SOK_QuadFloat4Spill;
1051 } else if (PPC::QBRCRegClass.hasSubClassEq(RC)) {
1052 OpcodeIndex = SOK_QuadBitSpill;
1053 } else if (PPC::SPILLTOVSRRCRegClass.hasSubClassEq(RC)) {
1054 OpcodeIndex = SOK_SpillToVSR;
1056 llvm_unreachable("Unknown regclass!");
1059 if (PPC::GPRCRegClass.contains(Reg) ||
1060 PPC::GPRC_NOR0RegClass.contains(Reg)) {
1061 OpcodeIndex = SOK_Int4Spill;
1062 } else if (PPC::G8RCRegClass.contains(Reg) ||
1063 PPC::G8RC_NOX0RegClass.contains(Reg)) {
1064 OpcodeIndex = SOK_Int8Spill;
1065 } else if (PPC::F8RCRegClass.contains(Reg)) {
1066 OpcodeIndex = SOK_Float8Spill;
1067 } else if (PPC::F4RCRegClass.contains(Reg)) {
1068 OpcodeIndex = SOK_Float4Spill;
1069 } else if (PPC::CRRCRegClass.contains(Reg)) {
1070 OpcodeIndex = SOK_CRSpill;
1071 } else if (PPC::CRBITRCRegClass.contains(Reg)) {
1072 OpcodeIndex = SOK_CRBitSpill;
1073 } else if (PPC::VRRCRegClass.contains(Reg)) {
1074 OpcodeIndex = SOK_VRVectorSpill;
1075 } else if (PPC::VSRCRegClass.contains(Reg)) {
1076 OpcodeIndex = SOK_VSXVectorSpill;
1077 } else if (PPC::VSFRCRegClass.contains(Reg)) {
1078 OpcodeIndex = SOK_VectorFloat8Spill;
1079 } else if (PPC::VSSRCRegClass.contains(Reg)) {
1080 OpcodeIndex = SOK_VectorFloat4Spill;
1081 } else if (PPC::VRSAVERCRegClass.contains(Reg)) {
1082 OpcodeIndex = SOK_VRSaveSpill;
1083 } else if (PPC::QFRCRegClass.contains(Reg)) {
1084 OpcodeIndex = SOK_QuadFloat8Spill;
1085 } else if (PPC::QSRCRegClass.contains(Reg)) {
1086 OpcodeIndex = SOK_QuadFloat4Spill;
1087 } else if (PPC::QBRCRegClass.contains(Reg)) {
1088 OpcodeIndex = SOK_QuadBitSpill;
1089 } else if (PPC::SPILLTOVSRRCRegClass.contains(Reg)) {
1090 OpcodeIndex = SOK_SpillToVSR;
1092 llvm_unreachable("Unknown regclass!");
1095 return OpcodesForSpill[OpcodeIndex];
1099 PPCInstrInfo::getLoadOpcodeForSpill(unsigned Reg,
1100 const TargetRegisterClass *RC) const {
1101 const unsigned *OpcodesForSpill = getLoadOpcodesForSpillArray();
1102 int OpcodeIndex = 0;
1104 if (RC != nullptr) {
1105 if (PPC::GPRCRegClass.hasSubClassEq(RC) ||
1106 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) {
1107 OpcodeIndex = SOK_Int4Spill;
1108 } else if (PPC::G8RCRegClass.hasSubClassEq(RC) ||
1109 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) {
1110 OpcodeIndex = SOK_Int8Spill;
1111 } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) {
1112 OpcodeIndex = SOK_Float8Spill;
1113 } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) {
1114 OpcodeIndex = SOK_Float4Spill;
1115 } else if (PPC::SPERCRegClass.hasSubClassEq(RC)) {
1116 OpcodeIndex = SOK_SPESpill;
1117 } else if (PPC::SPE4RCRegClass.hasSubClassEq(RC)) {
1118 OpcodeIndex = SOK_SPE4Spill;
1119 } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) {
1120 OpcodeIndex = SOK_CRSpill;
1121 } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) {
1122 OpcodeIndex = SOK_CRBitSpill;
1123 } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) {
1124 OpcodeIndex = SOK_VRVectorSpill;
1125 } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) {
1126 OpcodeIndex = SOK_VSXVectorSpill;
1127 } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) {
1128 OpcodeIndex = SOK_VectorFloat8Spill;
1129 } else if (PPC::VSSRCRegClass.hasSubClassEq(RC)) {
1130 OpcodeIndex = SOK_VectorFloat4Spill;
1131 } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) {
1132 OpcodeIndex = SOK_VRSaveSpill;
1133 } else if (PPC::QFRCRegClass.hasSubClassEq(RC)) {
1134 OpcodeIndex = SOK_QuadFloat8Spill;
1135 } else if (PPC::QSRCRegClass.hasSubClassEq(RC)) {
1136 OpcodeIndex = SOK_QuadFloat4Spill;
1137 } else if (PPC::QBRCRegClass.hasSubClassEq(RC)) {
1138 OpcodeIndex = SOK_QuadBitSpill;
1139 } else if (PPC::SPILLTOVSRRCRegClass.hasSubClassEq(RC)) {
1140 OpcodeIndex = SOK_SpillToVSR;
1142 llvm_unreachable("Unknown regclass!");
1145 if (PPC::GPRCRegClass.contains(Reg) ||
1146 PPC::GPRC_NOR0RegClass.contains(Reg)) {
1147 OpcodeIndex = SOK_Int4Spill;
1148 } else if (PPC::G8RCRegClass.contains(Reg) ||
1149 PPC::G8RC_NOX0RegClass.contains(Reg)) {
1150 OpcodeIndex = SOK_Int8Spill;
1151 } else if (PPC::F8RCRegClass.contains(Reg)) {
1152 OpcodeIndex = SOK_Float8Spill;
1153 } else if (PPC::F4RCRegClass.contains(Reg)) {
1154 OpcodeIndex = SOK_Float4Spill;
1155 } else if (PPC::CRRCRegClass.contains(Reg)) {
1156 OpcodeIndex = SOK_CRSpill;
1157 } else if (PPC::CRBITRCRegClass.contains(Reg)) {
1158 OpcodeIndex = SOK_CRBitSpill;
1159 } else if (PPC::VRRCRegClass.contains(Reg)) {
1160 OpcodeIndex = SOK_VRVectorSpill;
1161 } else if (PPC::VSRCRegClass.contains(Reg)) {
1162 OpcodeIndex = SOK_VSXVectorSpill;
1163 } else if (PPC::VSFRCRegClass.contains(Reg)) {
1164 OpcodeIndex = SOK_VectorFloat8Spill;
1165 } else if (PPC::VSSRCRegClass.contains(Reg)) {
1166 OpcodeIndex = SOK_VectorFloat4Spill;
1167 } else if (PPC::VRSAVERCRegClass.contains(Reg)) {
1168 OpcodeIndex = SOK_VRSaveSpill;
1169 } else if (PPC::QFRCRegClass.contains(Reg)) {
1170 OpcodeIndex = SOK_QuadFloat8Spill;
1171 } else if (PPC::QSRCRegClass.contains(Reg)) {
1172 OpcodeIndex = SOK_QuadFloat4Spill;
1173 } else if (PPC::QBRCRegClass.contains(Reg)) {
1174 OpcodeIndex = SOK_QuadBitSpill;
1175 } else if (PPC::SPILLTOVSRRCRegClass.contains(Reg)) {
1176 OpcodeIndex = SOK_SpillToVSR;
1178 llvm_unreachable("Unknown regclass!");
1181 return OpcodesForSpill[OpcodeIndex];
1184 void PPCInstrInfo::StoreRegToStackSlot(
1185 MachineFunction &MF, unsigned SrcReg, bool isKill, int FrameIdx,
1186 const TargetRegisterClass *RC,
1187 SmallVectorImpl<MachineInstr *> &NewMIs) const {
1188 unsigned Opcode = getStoreOpcodeForSpill(PPC::NoRegister, RC);
1191 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1192 FuncInfo->setHasSpills();
1194 NewMIs.push_back(addFrameReference(
1195 BuildMI(MF, DL, get(Opcode)).addReg(SrcReg, getKillRegState(isKill)),
1198 if (PPC::CRRCRegClass.hasSubClassEq(RC) ||
1199 PPC::CRBITRCRegClass.hasSubClassEq(RC))
1200 FuncInfo->setSpillsCR();
1202 if (PPC::VRSAVERCRegClass.hasSubClassEq(RC))
1203 FuncInfo->setSpillsVRSAVE();
1205 if (isXFormMemOp(Opcode))
1206 FuncInfo->setHasNonRISpills();
1209 void PPCInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
1210 MachineBasicBlock::iterator MI,
1211 unsigned SrcReg, bool isKill,
1213 const TargetRegisterClass *RC,
1214 const TargetRegisterInfo *TRI) const {
1215 MachineFunction &MF = *MBB.getParent();
1216 SmallVector<MachineInstr *, 4> NewMIs;
1218 // We need to avoid a situation in which the value from a VRRC register is
1219 // spilled using an Altivec instruction and reloaded into a VSRC register
1220 // using a VSX instruction. The issue with this is that the VSX
1221 // load/store instructions swap the doublewords in the vector and the Altivec
1222 // ones don't. The register classes on the spill/reload may be different if
1223 // the register is defined using an Altivec instruction and is then used by a
1227 StoreRegToStackSlot(MF, SrcReg, isKill, FrameIdx, RC, NewMIs);
1229 for (unsigned i = 0, e = NewMIs.size(); i != e; ++i)
1230 MBB.insert(MI, NewMIs[i]);
1232 const MachineFrameInfo &MFI = MF.getFrameInfo();
1233 MachineMemOperand *MMO = MF.getMachineMemOperand(
1234 MachinePointerInfo::getFixedStack(MF, FrameIdx),
1235 MachineMemOperand::MOStore, MFI.getObjectSize(FrameIdx),
1236 MFI.getObjectAlignment(FrameIdx));
1237 NewMIs.back()->addMemOperand(MF, MMO);
1240 void PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, const DebugLoc &DL,
1241 unsigned DestReg, int FrameIdx,
1242 const TargetRegisterClass *RC,
1243 SmallVectorImpl<MachineInstr *> &NewMIs)
1245 unsigned Opcode = getLoadOpcodeForSpill(PPC::NoRegister, RC);
1246 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(Opcode), DestReg),
1248 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1250 if (PPC::CRRCRegClass.hasSubClassEq(RC) ||
1251 PPC::CRBITRCRegClass.hasSubClassEq(RC))
1252 FuncInfo->setSpillsCR();
1254 if (PPC::VRSAVERCRegClass.hasSubClassEq(RC))
1255 FuncInfo->setSpillsVRSAVE();
1257 if (isXFormMemOp(Opcode))
1258 FuncInfo->setHasNonRISpills();
1262 PPCInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1263 MachineBasicBlock::iterator MI,
1264 unsigned DestReg, int FrameIdx,
1265 const TargetRegisterClass *RC,
1266 const TargetRegisterInfo *TRI) const {
1267 MachineFunction &MF = *MBB.getParent();
1268 SmallVector<MachineInstr*, 4> NewMIs;
1270 if (MI != MBB.end()) DL = MI->getDebugLoc();
1272 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1273 FuncInfo->setHasSpills();
1275 // We need to avoid a situation in which the value from a VRRC register is
1276 // spilled using an Altivec instruction and reloaded into a VSRC register
1277 // using a VSX instruction. The issue with this is that the VSX
1278 // load/store instructions swap the doublewords in the vector and the Altivec
1279 // ones don't. The register classes on the spill/reload may be different if
1280 // the register is defined using an Altivec instruction and is then used by a
1282 if (Subtarget.hasVSX() && RC == &PPC::VRRCRegClass)
1283 RC = &PPC::VSRCRegClass;
1285 LoadRegFromStackSlot(MF, DL, DestReg, FrameIdx, RC, NewMIs);
1287 for (unsigned i = 0, e = NewMIs.size(); i != e; ++i)
1288 MBB.insert(MI, NewMIs[i]);
1290 const MachineFrameInfo &MFI = MF.getFrameInfo();
1291 MachineMemOperand *MMO = MF.getMachineMemOperand(
1292 MachinePointerInfo::getFixedStack(MF, FrameIdx),
1293 MachineMemOperand::MOLoad, MFI.getObjectSize(FrameIdx),
1294 MFI.getObjectAlignment(FrameIdx));
1295 NewMIs.back()->addMemOperand(MF, MMO);
1299 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
1300 assert(Cond.size() == 2 && "Invalid PPC branch opcode!");
1301 if (Cond[1].getReg() == PPC::CTR8 || Cond[1].getReg() == PPC::CTR)
1302 Cond[0].setImm(Cond[0].getImm() == 0 ? 1 : 0);
1304 // Leave the CR# the same, but invert the condition.
1305 Cond[0].setImm(PPC::InvertPredicate((PPC::Predicate)Cond[0].getImm()));
1309 bool PPCInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
1310 unsigned Reg, MachineRegisterInfo *MRI) const {
1311 // For some instructions, it is legal to fold ZERO into the RA register field.
1312 // A zero immediate should always be loaded with a single li.
1313 unsigned DefOpc = DefMI.getOpcode();
1314 if (DefOpc != PPC::LI && DefOpc != PPC::LI8)
1316 if (!DefMI.getOperand(1).isImm())
1318 if (DefMI.getOperand(1).getImm() != 0)
1321 // Note that we cannot here invert the arguments of an isel in order to fold
1322 // a ZERO into what is presented as the second argument. All we have here
1323 // is the condition bit, and that might come from a CR-logical bit operation.
1325 const MCInstrDesc &UseMCID = UseMI.getDesc();
1327 // Only fold into real machine instructions.
1328 if (UseMCID.isPseudo())
1332 for (UseIdx = 0; UseIdx < UseMI.getNumOperands(); ++UseIdx)
1333 if (UseMI.getOperand(UseIdx).isReg() &&
1334 UseMI.getOperand(UseIdx).getReg() == Reg)
1337 assert(UseIdx < UseMI.getNumOperands() && "Cannot find Reg in UseMI");
1338 assert(UseIdx < UseMCID.getNumOperands() && "No operand description for Reg");
1340 const MCOperandInfo *UseInfo = &UseMCID.OpInfo[UseIdx];
1342 // We can fold the zero if this register requires a GPRC_NOR0/G8RC_NOX0
1343 // register (which might also be specified as a pointer class kind).
1344 if (UseInfo->isLookupPtrRegClass()) {
1345 if (UseInfo->RegClass /* Kind */ != 1)
1348 if (UseInfo->RegClass != PPC::GPRC_NOR0RegClassID &&
1349 UseInfo->RegClass != PPC::G8RC_NOX0RegClassID)
1353 // Make sure this is not tied to an output register (or otherwise
1354 // constrained). This is true for ST?UX registers, for example, which
1355 // are tied to their output registers.
1356 if (UseInfo->Constraints != 0)
1360 if (UseInfo->isLookupPtrRegClass()) {
1361 bool isPPC64 = Subtarget.isPPC64();
1362 ZeroReg = isPPC64 ? PPC::ZERO8 : PPC::ZERO;
1364 ZeroReg = UseInfo->RegClass == PPC::G8RC_NOX0RegClassID ?
1365 PPC::ZERO8 : PPC::ZERO;
1368 bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
1369 UseMI.getOperand(UseIdx).setReg(ZeroReg);
1372 DefMI.eraseFromParent();
1377 static bool MBBDefinesCTR(MachineBasicBlock &MBB) {
1378 for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end();
1380 if (I->definesRegister(PPC::CTR) || I->definesRegister(PPC::CTR8))
1385 // We should make sure that, if we're going to predicate both sides of a
1386 // condition (a diamond), that both sides don't define the counter register. We
1387 // can predicate counter-decrement-based branches, but while that predicates
1388 // the branching, it does not predicate the counter decrement. If we tried to
1389 // merge the triangle into one predicated block, we'd decrement the counter
1391 bool PPCInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
1392 unsigned NumT, unsigned ExtraT,
1393 MachineBasicBlock &FMBB,
1394 unsigned NumF, unsigned ExtraF,
1395 BranchProbability Probability) const {
1396 return !(MBBDefinesCTR(TMBB) && MBBDefinesCTR(FMBB));
1400 bool PPCInstrInfo::isPredicated(const MachineInstr &MI) const {
1401 // The predicated branches are identified by their type, not really by the
1402 // explicit presence of a predicate. Furthermore, some of them can be
1403 // predicated more than once. Because if conversion won't try to predicate
1404 // any instruction which already claims to be predicated (by returning true
1405 // here), always return false. In doing so, we let isPredicable() be the
1406 // final word on whether not the instruction can be (further) predicated.
1411 bool PPCInstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const {
1412 if (!MI.isTerminator())
1415 // Conditional branch is a special case.
1416 if (MI.isBranch() && !MI.isBarrier())
1419 return !isPredicated(MI);
1422 bool PPCInstrInfo::PredicateInstruction(MachineInstr &MI,
1423 ArrayRef<MachineOperand> Pred) const {
1424 unsigned OpC = MI.getOpcode();
1425 if (OpC == PPC::BLR || OpC == PPC::BLR8) {
1426 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) {
1427 bool isPPC64 = Subtarget.isPPC64();
1428 MI.setDesc(get(Pred[0].getImm() ? (isPPC64 ? PPC::BDNZLR8 : PPC::BDNZLR)
1429 : (isPPC64 ? PPC::BDZLR8 : PPC::BDZLR)));
1430 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
1431 MI.setDesc(get(PPC::BCLR));
1432 MachineInstrBuilder(*MI.getParent()->getParent(), MI).add(Pred[1]);
1433 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
1434 MI.setDesc(get(PPC::BCLRn));
1435 MachineInstrBuilder(*MI.getParent()->getParent(), MI).add(Pred[1]);
1437 MI.setDesc(get(PPC::BCCLR));
1438 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1439 .addImm(Pred[0].getImm())
1444 } else if (OpC == PPC::B) {
1445 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) {
1446 bool isPPC64 = Subtarget.isPPC64();
1447 MI.setDesc(get(Pred[0].getImm() ? (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ)
1448 : (isPPC64 ? PPC::BDZ8 : PPC::BDZ)));
1449 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
1450 MachineBasicBlock *MBB = MI.getOperand(0).getMBB();
1451 MI.RemoveOperand(0);
1453 MI.setDesc(get(PPC::BC));
1454 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1457 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
1458 MachineBasicBlock *MBB = MI.getOperand(0).getMBB();
1459 MI.RemoveOperand(0);
1461 MI.setDesc(get(PPC::BCn));
1462 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1466 MachineBasicBlock *MBB = MI.getOperand(0).getMBB();
1467 MI.RemoveOperand(0);
1469 MI.setDesc(get(PPC::BCC));
1470 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1471 .addImm(Pred[0].getImm())
1477 } else if (OpC == PPC::BCTR || OpC == PPC::BCTR8 || OpC == PPC::BCTRL ||
1478 OpC == PPC::BCTRL8) {
1479 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR)
1480 llvm_unreachable("Cannot predicate bctr[l] on the ctr register");
1482 bool setLR = OpC == PPC::BCTRL || OpC == PPC::BCTRL8;
1483 bool isPPC64 = Subtarget.isPPC64();
1485 if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
1486 MI.setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8 : PPC::BCCTR8)
1487 : (setLR ? PPC::BCCTRL : PPC::BCCTR)));
1488 MachineInstrBuilder(*MI.getParent()->getParent(), MI).add(Pred[1]);
1490 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
1491 MI.setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8n : PPC::BCCTR8n)
1492 : (setLR ? PPC::BCCTRLn : PPC::BCCTRn)));
1493 MachineInstrBuilder(*MI.getParent()->getParent(), MI).add(Pred[1]);
1497 MI.setDesc(get(isPPC64 ? (setLR ? PPC::BCCCTRL8 : PPC::BCCCTR8)
1498 : (setLR ? PPC::BCCCTRL : PPC::BCCCTR)));
1499 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1500 .addImm(Pred[0].getImm())
1508 bool PPCInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
1509 ArrayRef<MachineOperand> Pred2) const {
1510 assert(Pred1.size() == 2 && "Invalid PPC first predicate");
1511 assert(Pred2.size() == 2 && "Invalid PPC second predicate");
1513 if (Pred1[1].getReg() == PPC::CTR8 || Pred1[1].getReg() == PPC::CTR)
1515 if (Pred2[1].getReg() == PPC::CTR8 || Pred2[1].getReg() == PPC::CTR)
1518 // P1 can only subsume P2 if they test the same condition register.
1519 if (Pred1[1].getReg() != Pred2[1].getReg())
1522 PPC::Predicate P1 = (PPC::Predicate) Pred1[0].getImm();
1523 PPC::Predicate P2 = (PPC::Predicate) Pred2[0].getImm();
1528 // Does P1 subsume P2, e.g. GE subsumes GT.
1529 if (P1 == PPC::PRED_LE &&
1530 (P2 == PPC::PRED_LT || P2 == PPC::PRED_EQ))
1532 if (P1 == PPC::PRED_GE &&
1533 (P2 == PPC::PRED_GT || P2 == PPC::PRED_EQ))
1539 bool PPCInstrInfo::DefinesPredicate(MachineInstr &MI,
1540 std::vector<MachineOperand> &Pred) const {
1541 // Note: At the present time, the contents of Pred from this function is
1542 // unused by IfConversion. This implementation follows ARM by pushing the
1543 // CR-defining operand. Because the 'DZ' and 'DNZ' count as types of
1544 // predicate, instructions defining CTR or CTR8 are also included as
1545 // predicate-defining instructions.
1547 const TargetRegisterClass *RCs[] =
1548 { &PPC::CRRCRegClass, &PPC::CRBITRCRegClass,
1549 &PPC::CTRRCRegClass, &PPC::CTRRC8RegClass };
1552 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1553 const MachineOperand &MO = MI.getOperand(i);
1554 for (unsigned c = 0; c < array_lengthof(RCs) && !Found; ++c) {
1555 const TargetRegisterClass *RC = RCs[c];
1557 if (MO.isDef() && RC->contains(MO.getReg())) {
1561 } else if (MO.isRegMask()) {
1562 for (TargetRegisterClass::iterator I = RC->begin(),
1563 IE = RC->end(); I != IE; ++I)
1564 if (MO.clobbersPhysReg(*I)) {
1575 bool PPCInstrInfo::isPredicable(const MachineInstr &MI) const {
1576 unsigned OpC = MI.getOpcode();
1591 bool PPCInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
1592 unsigned &SrcReg2, int &Mask,
1594 unsigned Opc = MI.getOpcode();
1597 default: return false;
1602 SrcReg = MI.getOperand(1).getReg();
1604 Value = MI.getOperand(2).getImm();
1613 SrcReg = MI.getOperand(1).getReg();
1614 SrcReg2 = MI.getOperand(2).getReg();
1621 bool PPCInstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
1622 unsigned SrcReg2, int Mask, int Value,
1623 const MachineRegisterInfo *MRI) const {
1627 int OpC = CmpInstr.getOpcode();
1628 unsigned CRReg = CmpInstr.getOperand(0).getReg();
1630 // FP record forms set CR1 based on the exception status bits, not a
1631 // comparison with zero.
1632 if (OpC == PPC::FCMPUS || OpC == PPC::FCMPUD)
1635 // The record forms set the condition register based on a signed comparison
1636 // with zero (so says the ISA manual). This is not as straightforward as it
1637 // seems, however, because this is always a 64-bit comparison on PPC64, even
1638 // for instructions that are 32-bit in nature (like slw for example).
1639 // So, on PPC32, for unsigned comparisons, we can use the record forms only
1640 // for equality checks (as those don't depend on the sign). On PPC64,
1641 // we are restricted to equality for unsigned 64-bit comparisons and for
1642 // signed 32-bit comparisons the applicability is more restricted.
1643 bool isPPC64 = Subtarget.isPPC64();
1644 bool is32BitSignedCompare = OpC == PPC::CMPWI || OpC == PPC::CMPW;
1645 bool is32BitUnsignedCompare = OpC == PPC::CMPLWI || OpC == PPC::CMPLW;
1646 bool is64BitUnsignedCompare = OpC == PPC::CMPLDI || OpC == PPC::CMPLD;
1648 // Get the unique definition of SrcReg.
1649 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
1650 if (!MI) return false;
1652 bool equalityOnly = false;
1655 if (is32BitSignedCompare) {
1656 // We can perform this optimization only if MI is sign-extending.
1657 if (isSignExtended(*MI))
1661 } else if (is32BitUnsignedCompare) {
1662 // We can perform this optimization, equality only, if MI is
1664 if (isZeroExtended(*MI)) {
1666 equalityOnly = true;
1670 equalityOnly = is64BitUnsignedCompare;
1672 equalityOnly = is32BitUnsignedCompare;
1675 // We need to check the uses of the condition register in order to reject
1676 // non-equality comparisons.
1677 for (MachineRegisterInfo::use_instr_iterator
1678 I = MRI->use_instr_begin(CRReg), IE = MRI->use_instr_end();
1680 MachineInstr *UseMI = &*I;
1681 if (UseMI->getOpcode() == PPC::BCC) {
1682 PPC::Predicate Pred = (PPC::Predicate)UseMI->getOperand(0).getImm();
1683 unsigned PredCond = PPC::getPredicateCondition(Pred);
1684 // We ignore hint bits when checking for non-equality comparisons.
1685 if (PredCond != PPC::PRED_EQ && PredCond != PPC::PRED_NE)
1687 } else if (UseMI->getOpcode() == PPC::ISEL ||
1688 UseMI->getOpcode() == PPC::ISEL8) {
1689 unsigned SubIdx = UseMI->getOperand(3).getSubReg();
1690 if (SubIdx != PPC::sub_eq)
1697 MachineBasicBlock::iterator I = CmpInstr;
1699 // Scan forward to find the first use of the compare.
1700 for (MachineBasicBlock::iterator EL = CmpInstr.getParent()->end(); I != EL;
1702 bool FoundUse = false;
1703 for (MachineRegisterInfo::use_instr_iterator
1704 J = MRI->use_instr_begin(CRReg), JE = MRI->use_instr_end();
1715 SmallVector<std::pair<MachineOperand*, PPC::Predicate>, 4> PredsToUpdate;
1716 SmallVector<std::pair<MachineOperand*, unsigned>, 4> SubRegsToUpdate;
1718 // There are two possible candidates which can be changed to set CR[01].
1719 // One is MI, the other is a SUB instruction.
1720 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
1721 MachineInstr *Sub = nullptr;
1723 // MI is not a candidate for CMPrr.
1725 // FIXME: Conservatively refuse to convert an instruction which isn't in the
1726 // same BB as the comparison. This is to allow the check below to avoid calls
1727 // (and other explicit clobbers); instead we should really check for these
1728 // more explicitly (in at least a few predecessors).
1729 else if (MI->getParent() != CmpInstr.getParent())
1731 else if (Value != 0) {
1732 // The record-form instructions set CR bit based on signed comparison
1733 // against 0. We try to convert a compare against 1 or -1 into a compare
1734 // against 0 to exploit record-form instructions. For example, we change
1735 // the condition "greater than -1" into "greater than or equal to 0"
1736 // and "less than 1" into "less than or equal to 0".
1738 // Since we optimize comparison based on a specific branch condition,
1739 // we don't optimize if condition code is used by more than once.
1740 if (equalityOnly || !MRI->hasOneUse(CRReg))
1743 MachineInstr *UseMI = &*MRI->use_instr_begin(CRReg);
1744 if (UseMI->getOpcode() != PPC::BCC)
1747 PPC::Predicate Pred = (PPC::Predicate)UseMI->getOperand(0).getImm();
1748 PPC::Predicate NewPred = Pred;
1749 unsigned PredCond = PPC::getPredicateCondition(Pred);
1750 unsigned PredHint = PPC::getPredicateHint(Pred);
1751 int16_t Immed = (int16_t)Value;
1753 // When modifying the condition in the predicate, we propagate hint bits
1754 // from the original predicate to the new one.
1755 if (Immed == -1 && PredCond == PPC::PRED_GT)
1756 // We convert "greater than -1" into "greater than or equal to 0",
1757 // since we are assuming signed comparison by !equalityOnly
1758 NewPred = PPC::getPredicate(PPC::PRED_GE, PredHint);
1759 else if (Immed == -1 && PredCond == PPC::PRED_LE)
1760 // We convert "less than or equal to -1" into "less than 0".
1761 NewPred = PPC::getPredicate(PPC::PRED_LT, PredHint);
1762 else if (Immed == 1 && PredCond == PPC::PRED_LT)
1763 // We convert "less than 1" into "less than or equal to 0".
1764 NewPred = PPC::getPredicate(PPC::PRED_LE, PredHint);
1765 else if (Immed == 1 && PredCond == PPC::PRED_GE)
1766 // We convert "greater than or equal to 1" into "greater than 0".
1767 NewPred = PPC::getPredicate(PPC::PRED_GT, PredHint);
1771 PredsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(0)),
1776 const TargetRegisterInfo *TRI = &getRegisterInfo();
1779 // Get ready to iterate backward from CmpInstr.
1780 MachineBasicBlock::iterator E = MI, B = CmpInstr.getParent()->begin();
1782 for (; I != E && !noSub; --I) {
1783 const MachineInstr &Instr = *I;
1784 unsigned IOpC = Instr.getOpcode();
1786 if (&*I != &CmpInstr && (Instr.modifiesRegister(PPC::CR0, TRI) ||
1787 Instr.readsRegister(PPC::CR0, TRI)))
1788 // This instruction modifies or uses the record condition register after
1789 // the one we want to change. While we could do this transformation, it
1790 // would likely not be profitable. This transformation removes one
1791 // instruction, and so even forcing RA to generate one move probably
1792 // makes it unprofitable.
1795 // Check whether CmpInstr can be made redundant by the current instruction.
1796 if ((OpC == PPC::CMPW || OpC == PPC::CMPLW ||
1797 OpC == PPC::CMPD || OpC == PPC::CMPLD) &&
1798 (IOpC == PPC::SUBF || IOpC == PPC::SUBF8) &&
1799 ((Instr.getOperand(1).getReg() == SrcReg &&
1800 Instr.getOperand(2).getReg() == SrcReg2) ||
1801 (Instr.getOperand(1).getReg() == SrcReg2 &&
1802 Instr.getOperand(2).getReg() == SrcReg))) {
1808 // The 'and' is below the comparison instruction.
1812 // Return false if no candidates exist.
1816 // The single candidate is called MI.
1820 int MIOpC = MI->getOpcode();
1821 if (MIOpC == PPC::ANDIo || MIOpC == PPC::ANDIo8 ||
1822 MIOpC == PPC::ANDISo || MIOpC == PPC::ANDISo8)
1825 NewOpC = PPC::getRecordFormOpcode(MIOpC);
1826 if (NewOpC == -1 && PPC::getNonRecordFormOpcode(MIOpC) != -1)
1830 // FIXME: On the non-embedded POWER architectures, only some of the record
1831 // forms are fast, and we should use only the fast ones.
1833 // The defining instruction has a record form (or is already a record
1834 // form). It is possible, however, that we'll need to reverse the condition
1835 // code of the users.
1839 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based on CMP
1840 // needs to be updated to be based on SUB. Push the condition code
1841 // operands to OperandsToUpdate. If it is safe to remove CmpInstr, the
1842 // condition code of these operands will be modified.
1843 // Here, Value == 0 means we haven't converted comparison against 1 or -1 to
1844 // comparison against 0, which may modify predicate.
1845 bool ShouldSwap = false;
1846 if (Sub && Value == 0) {
1847 ShouldSwap = SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
1848 Sub->getOperand(2).getReg() == SrcReg;
1850 // The operands to subf are the opposite of sub, so only in the fixed-point
1851 // case, invert the order.
1852 ShouldSwap = !ShouldSwap;
1856 for (MachineRegisterInfo::use_instr_iterator
1857 I = MRI->use_instr_begin(CRReg), IE = MRI->use_instr_end();
1859 MachineInstr *UseMI = &*I;
1860 if (UseMI->getOpcode() == PPC::BCC) {
1861 PPC::Predicate Pred = (PPC::Predicate) UseMI->getOperand(0).getImm();
1862 unsigned PredCond = PPC::getPredicateCondition(Pred);
1863 assert((!equalityOnly ||
1864 PredCond == PPC::PRED_EQ || PredCond == PPC::PRED_NE) &&
1865 "Invalid predicate for equality-only optimization");
1866 (void)PredCond; // To suppress warning in release build.
1867 PredsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(0)),
1868 PPC::getSwappedPredicate(Pred)));
1869 } else if (UseMI->getOpcode() == PPC::ISEL ||
1870 UseMI->getOpcode() == PPC::ISEL8) {
1871 unsigned NewSubReg = UseMI->getOperand(3).getSubReg();
1872 assert((!equalityOnly || NewSubReg == PPC::sub_eq) &&
1873 "Invalid CR bit for equality-only optimization");
1875 if (NewSubReg == PPC::sub_lt)
1876 NewSubReg = PPC::sub_gt;
1877 else if (NewSubReg == PPC::sub_gt)
1878 NewSubReg = PPC::sub_lt;
1880 SubRegsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(3)),
1882 } else // We need to abort on a user we don't understand.
1885 assert(!(Value != 0 && ShouldSwap) &&
1886 "Non-zero immediate support and ShouldSwap"
1887 "may conflict in updating predicate");
1889 // Create a new virtual register to hold the value of the CR set by the
1890 // record-form instruction. If the instruction was not previously in
1891 // record form, then set the kill flag on the CR.
1892 CmpInstr.eraseFromParent();
1894 MachineBasicBlock::iterator MII = MI;
1895 BuildMI(*MI->getParent(), std::next(MII), MI->getDebugLoc(),
1896 get(TargetOpcode::COPY), CRReg)
1897 .addReg(PPC::CR0, MIOpC != NewOpC ? RegState::Kill : 0);
1899 // Even if CR0 register were dead before, it is alive now since the
1900 // instruction we just built uses it.
1901 MI->clearRegisterDeads(PPC::CR0);
1903 if (MIOpC != NewOpC) {
1904 // We need to be careful here: we're replacing one instruction with
1905 // another, and we need to make sure that we get all of the right
1906 // implicit uses and defs. On the other hand, the caller may be holding
1907 // an iterator to this instruction, and so we can't delete it (this is
1908 // specifically the case if this is the instruction directly after the
1911 // Rotates are expensive instructions. If we're emitting a record-form
1912 // rotate that can just be an andi/andis, we should just emit that.
1913 if (MIOpC == PPC::RLWINM || MIOpC == PPC::RLWINM8) {
1914 unsigned GPRRes = MI->getOperand(0).getReg();
1915 int64_t SH = MI->getOperand(2).getImm();
1916 int64_t MB = MI->getOperand(3).getImm();
1917 int64_t ME = MI->getOperand(4).getImm();
1918 // We can only do this if both the start and end of the mask are in the
1920 bool MBInLoHWord = MB >= 16;
1921 bool MEInLoHWord = ME >= 16;
1922 uint64_t Mask = ~0LLU;
1924 if (MB <= ME && MBInLoHWord == MEInLoHWord && SH == 0) {
1925 Mask = ((1LLU << (32 - MB)) - 1) & ~((1LLU << (31 - ME)) - 1);
1926 // The mask value needs to shift right 16 if we're emitting andis.
1927 Mask >>= MBInLoHWord ? 0 : 16;
1928 NewOpC = MIOpC == PPC::RLWINM ?
1929 (MBInLoHWord ? PPC::ANDIo : PPC::ANDISo) :
1930 (MBInLoHWord ? PPC::ANDIo8 :PPC::ANDISo8);
1931 } else if (MRI->use_empty(GPRRes) && (ME == 31) &&
1932 (ME - MB + 1 == SH) && (MB >= 16)) {
1933 // If we are rotating by the exact number of bits as are in the mask
1934 // and the mask is in the least significant bits of the register,
1935 // that's just an andis. (as long as the GPR result has no uses).
1936 Mask = ((1LLU << 32) - 1) & ~((1LLU << (32 - SH)) - 1);
1938 NewOpC = MIOpC == PPC::RLWINM ? PPC::ANDISo :PPC::ANDISo8;
1940 // If we've set the mask, we can transform.
1941 if (Mask != ~0LLU) {
1942 MI->RemoveOperand(4);
1943 MI->RemoveOperand(3);
1944 MI->getOperand(2).setImm(Mask);
1945 NumRcRotatesConvertedToRcAnd++;
1947 } else if (MIOpC == PPC::RLDICL && MI->getOperand(2).getImm() == 0) {
1948 int64_t MB = MI->getOperand(3).getImm();
1950 uint64_t Mask = (1LLU << (63 - MB + 1)) - 1;
1951 NewOpC = PPC::ANDIo8;
1952 MI->RemoveOperand(3);
1953 MI->getOperand(2).setImm(Mask);
1954 NumRcRotatesConvertedToRcAnd++;
1958 const MCInstrDesc &NewDesc = get(NewOpC);
1959 MI->setDesc(NewDesc);
1961 if (NewDesc.ImplicitDefs)
1962 for (const MCPhysReg *ImpDefs = NewDesc.getImplicitDefs();
1963 *ImpDefs; ++ImpDefs)
1964 if (!MI->definesRegister(*ImpDefs))
1965 MI->addOperand(*MI->getParent()->getParent(),
1966 MachineOperand::CreateReg(*ImpDefs, true, true));
1967 if (NewDesc.ImplicitUses)
1968 for (const MCPhysReg *ImpUses = NewDesc.getImplicitUses();
1969 *ImpUses; ++ImpUses)
1970 if (!MI->readsRegister(*ImpUses))
1971 MI->addOperand(*MI->getParent()->getParent(),
1972 MachineOperand::CreateReg(*ImpUses, false, true));
1974 assert(MI->definesRegister(PPC::CR0) &&
1975 "Record-form instruction does not define cr0?");
1977 // Modify the condition code of operands in OperandsToUpdate.
1978 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to
1979 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
1980 for (unsigned i = 0, e = PredsToUpdate.size(); i < e; i++)
1981 PredsToUpdate[i].first->setImm(PredsToUpdate[i].second);
1983 for (unsigned i = 0, e = SubRegsToUpdate.size(); i < e; i++)
1984 SubRegsToUpdate[i].first->setSubReg(SubRegsToUpdate[i].second);
1989 /// GetInstSize - Return the number of bytes of code the specified
1990 /// instruction may be. This returns the maximum number of bytes.
1992 unsigned PPCInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
1993 unsigned Opcode = MI.getOpcode();
1995 if (Opcode == PPC::INLINEASM) {
1996 const MachineFunction *MF = MI.getParent()->getParent();
1997 const char *AsmStr = MI.getOperand(0).getSymbolName();
1998 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
1999 } else if (Opcode == TargetOpcode::STACKMAP) {
2000 StackMapOpers Opers(&MI);
2001 return Opers.getNumPatchBytes();
2002 } else if (Opcode == TargetOpcode::PATCHPOINT) {
2003 PatchPointOpers Opers(&MI);
2004 return Opers.getNumPatchBytes();
2006 return get(Opcode).getSize();
2010 std::pair<unsigned, unsigned>
2011 PPCInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
2012 const unsigned Mask = PPCII::MO_ACCESS_MASK;
2013 return std::make_pair(TF & Mask, TF & ~Mask);
2016 ArrayRef<std::pair<unsigned, const char *>>
2017 PPCInstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
2018 using namespace PPCII;
2019 static const std::pair<unsigned, const char *> TargetFlags[] = {
2022 {MO_TPREL_LO, "ppc-tprel-lo"},
2023 {MO_TPREL_HA, "ppc-tprel-ha"},
2024 {MO_DTPREL_LO, "ppc-dtprel-lo"},
2025 {MO_TLSLD_LO, "ppc-tlsld-lo"},
2026 {MO_TOC_LO, "ppc-toc-lo"},
2027 {MO_TLS, "ppc-tls"}};
2028 return makeArrayRef(TargetFlags);
2031 ArrayRef<std::pair<unsigned, const char *>>
2032 PPCInstrInfo::getSerializableBitmaskMachineOperandTargetFlags() const {
2033 using namespace PPCII;
2034 static const std::pair<unsigned, const char *> TargetFlags[] = {
2035 {MO_PLT, "ppc-plt"},
2036 {MO_PIC_FLAG, "ppc-pic"},
2037 {MO_NLP_FLAG, "ppc-nlp"},
2038 {MO_NLP_HIDDEN_FLAG, "ppc-nlp-hidden"}};
2039 return makeArrayRef(TargetFlags);
2042 // Expand VSX Memory Pseudo instruction to either a VSX or a FP instruction.
2043 // The VSX versions have the advantage of a full 64-register target whereas
2044 // the FP ones have the advantage of lower latency and higher throughput. So
2045 // what we are after is using the faster instructions in low register pressure
2046 // situations and using the larger register file in high register pressure
2048 bool PPCInstrInfo::expandVSXMemPseudo(MachineInstr &MI) const {
2049 unsigned UpperOpcode, LowerOpcode;
2050 switch (MI.getOpcode()) {
2051 case PPC::DFLOADf32:
2052 UpperOpcode = PPC::LXSSP;
2053 LowerOpcode = PPC::LFS;
2055 case PPC::DFLOADf64:
2056 UpperOpcode = PPC::LXSD;
2057 LowerOpcode = PPC::LFD;
2059 case PPC::DFSTOREf32:
2060 UpperOpcode = PPC::STXSSP;
2061 LowerOpcode = PPC::STFS;
2063 case PPC::DFSTOREf64:
2064 UpperOpcode = PPC::STXSD;
2065 LowerOpcode = PPC::STFD;
2067 case PPC::XFLOADf32:
2068 UpperOpcode = PPC::LXSSPX;
2069 LowerOpcode = PPC::LFSX;
2071 case PPC::XFLOADf64:
2072 UpperOpcode = PPC::LXSDX;
2073 LowerOpcode = PPC::LFDX;
2075 case PPC::XFSTOREf32:
2076 UpperOpcode = PPC::STXSSPX;
2077 LowerOpcode = PPC::STFSX;
2079 case PPC::XFSTOREf64:
2080 UpperOpcode = PPC::STXSDX;
2081 LowerOpcode = PPC::STFDX;
2084 UpperOpcode = PPC::LXSIWAX;
2085 LowerOpcode = PPC::LFIWAX;
2088 UpperOpcode = PPC::LXSIWZX;
2089 LowerOpcode = PPC::LFIWZX;
2092 UpperOpcode = PPC::STXSIWX;
2093 LowerOpcode = PPC::STFIWX;
2096 llvm_unreachable("Unknown Operation!");
2099 unsigned TargetReg = MI.getOperand(0).getReg();
2101 if ((TargetReg >= PPC::F0 && TargetReg <= PPC::F31) ||
2102 (TargetReg >= PPC::VSL0 && TargetReg <= PPC::VSL31))
2103 Opcode = LowerOpcode;
2105 Opcode = UpperOpcode;
2106 MI.setDesc(get(Opcode));
2110 static bool isAnImmediateOperand(const MachineOperand &MO) {
2111 return MO.isCPI() || MO.isGlobal() || MO.isImm();
2114 bool PPCInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
2115 auto &MBB = *MI.getParent();
2116 auto DL = MI.getDebugLoc();
2118 switch (MI.getOpcode()) {
2119 case TargetOpcode::LOAD_STACK_GUARD: {
2120 assert(Subtarget.isTargetLinux() &&
2121 "Only Linux target is expected to contain LOAD_STACK_GUARD");
2122 const int64_t Offset = Subtarget.isPPC64() ? -0x7010 : -0x7008;
2123 const unsigned Reg = Subtarget.isPPC64() ? PPC::X13 : PPC::R2;
2124 MI.setDesc(get(Subtarget.isPPC64() ? PPC::LD : PPC::LWZ));
2125 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
2130 case PPC::DFLOADf32:
2131 case PPC::DFLOADf64:
2132 case PPC::DFSTOREf32:
2133 case PPC::DFSTOREf64: {
2134 assert(Subtarget.hasP9Vector() &&
2135 "Invalid D-Form Pseudo-ops on Pre-P9 target.");
2136 assert(MI.getOperand(2).isReg() &&
2137 isAnImmediateOperand(MI.getOperand(1)) &&
2138 "D-form op must have register and immediate operands");
2139 return expandVSXMemPseudo(MI);
2141 case PPC::XFLOADf32:
2142 case PPC::XFSTOREf32:
2146 assert(Subtarget.hasP8Vector() &&
2147 "Invalid X-Form Pseudo-ops on Pre-P8 target.");
2148 assert(MI.getOperand(2).isReg() && MI.getOperand(1).isReg() &&
2149 "X-form op must have register and register operands");
2150 return expandVSXMemPseudo(MI);
2152 case PPC::XFLOADf64:
2153 case PPC::XFSTOREf64: {
2154 assert(Subtarget.hasVSX() &&
2155 "Invalid X-Form Pseudo-ops on target that has no VSX.");
2156 assert(MI.getOperand(2).isReg() && MI.getOperand(1).isReg() &&
2157 "X-form op must have register and register operands");
2158 return expandVSXMemPseudo(MI);
2160 case PPC::SPILLTOVSR_LD: {
2161 unsigned TargetReg = MI.getOperand(0).getReg();
2162 if (PPC::VSFRCRegClass.contains(TargetReg)) {
2163 MI.setDesc(get(PPC::DFLOADf64));
2164 return expandPostRAPseudo(MI);
2167 MI.setDesc(get(PPC::LD));
2170 case PPC::SPILLTOVSR_ST: {
2171 unsigned SrcReg = MI.getOperand(0).getReg();
2172 if (PPC::VSFRCRegClass.contains(SrcReg)) {
2173 NumStoreSPILLVSRRCAsVec++;
2174 MI.setDesc(get(PPC::DFSTOREf64));
2175 return expandPostRAPseudo(MI);
2177 NumStoreSPILLVSRRCAsGpr++;
2178 MI.setDesc(get(PPC::STD));
2182 case PPC::SPILLTOVSR_LDX: {
2183 unsigned TargetReg = MI.getOperand(0).getReg();
2184 if (PPC::VSFRCRegClass.contains(TargetReg))
2185 MI.setDesc(get(PPC::LXSDX));
2187 MI.setDesc(get(PPC::LDX));
2190 case PPC::SPILLTOVSR_STX: {
2191 unsigned SrcReg = MI.getOperand(0).getReg();
2192 if (PPC::VSFRCRegClass.contains(SrcReg)) {
2193 NumStoreSPILLVSRRCAsVec++;
2194 MI.setDesc(get(PPC::STXSDX));
2196 NumStoreSPILLVSRRCAsGpr++;
2197 MI.setDesc(get(PPC::STDX));
2202 case PPC::CFENCE8: {
2203 auto Val = MI.getOperand(0).getReg();
2204 BuildMI(MBB, MI, DL, get(PPC::CMPD), PPC::CR7).addReg(Val).addReg(Val);
2205 BuildMI(MBB, MI, DL, get(PPC::CTRL_DEP))
2206 .addImm(PPC::PRED_NE_MINUS)
2209 MI.setDesc(get(PPC::ISYNC));
2210 MI.RemoveOperand(0);
2217 // Essentially a compile-time implementation of a compare->isel sequence.
2218 // It takes two constants to compare, along with the true/false registers
2219 // and the comparison type (as a subreg to a CR field) and returns one
2220 // of the true/false registers, depending on the comparison results.
2221 static unsigned selectReg(int64_t Imm1, int64_t Imm2, unsigned CompareOpc,
2222 unsigned TrueReg, unsigned FalseReg,
2223 unsigned CRSubReg) {
2224 // Signed comparisons. The immediates are assumed to be sign-extended.
2225 if (CompareOpc == PPC::CMPWI || CompareOpc == PPC::CMPDI) {
2227 default: llvm_unreachable("Unknown integer comparison type.");
2229 return Imm1 < Imm2 ? TrueReg : FalseReg;
2231 return Imm1 > Imm2 ? TrueReg : FalseReg;
2233 return Imm1 == Imm2 ? TrueReg : FalseReg;
2236 // Unsigned comparisons.
2237 else if (CompareOpc == PPC::CMPLWI || CompareOpc == PPC::CMPLDI) {
2239 default: llvm_unreachable("Unknown integer comparison type.");
2241 return (uint64_t)Imm1 < (uint64_t)Imm2 ? TrueReg : FalseReg;
2243 return (uint64_t)Imm1 > (uint64_t)Imm2 ? TrueReg : FalseReg;
2245 return Imm1 == Imm2 ? TrueReg : FalseReg;
2248 return PPC::NoRegister;
2251 void PPCInstrInfo::replaceInstrOperandWithImm(MachineInstr &MI,
2253 int64_t Imm) const {
2254 assert(MI.getOperand(OpNo).isReg() && "Operand must be a REG");
2255 // Replace the REG with the Immediate.
2256 unsigned InUseReg = MI.getOperand(OpNo).getReg();
2257 MI.getOperand(OpNo).ChangeToImmediate(Imm);
2259 if (empty(MI.implicit_operands()))
2262 // We need to make sure that the MI didn't have any implicit use
2263 // of this REG any more.
2264 const TargetRegisterInfo *TRI = &getRegisterInfo();
2265 int UseOpIdx = MI.findRegisterUseOperandIdx(InUseReg, false, TRI);
2266 if (UseOpIdx >= 0) {
2267 MachineOperand &MO = MI.getOperand(UseOpIdx);
2268 if (MO.isImplicit())
2269 // The operands must always be in the following order:
2270 // - explicit reg defs,
2271 // - other explicit operands (reg uses, immediates, etc.),
2272 // - implicit reg defs
2273 // - implicit reg uses
2274 // Therefore, removing the implicit operand won't change the explicit
2276 MI.RemoveOperand(UseOpIdx);
2280 // Replace an instruction with one that materializes a constant (and sets
2281 // CR0 if the original instruction was a record-form instruction).
2282 void PPCInstrInfo::replaceInstrWithLI(MachineInstr &MI,
2283 const LoadImmediateInfo &LII) const {
2284 // Remove existing operands.
2285 int OperandToKeep = LII.SetCR ? 1 : 0;
2286 for (int i = MI.getNumOperands() - 1; i > OperandToKeep; i--)
2287 MI.RemoveOperand(i);
2289 // Replace the instruction.
2291 MI.setDesc(get(LII.Is64Bit ? PPC::ANDIo8 : PPC::ANDIo));
2292 // Set the immediate.
2293 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
2294 .addImm(LII.Imm).addReg(PPC::CR0, RegState::ImplicitDefine);
2298 MI.setDesc(get(LII.Is64Bit ? PPC::LI8 : PPC::LI));
2300 // Set the immediate.
2301 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
2305 MachineInstr *PPCInstrInfo::getForwardingDefMI(
2307 unsigned &OpNoForForwarding,
2308 bool &SeenIntermediateUse) const {
2309 OpNoForForwarding = ~0U;
2310 MachineInstr *DefMI = nullptr;
2311 MachineRegisterInfo *MRI = &MI.getParent()->getParent()->getRegInfo();
2312 const TargetRegisterInfo *TRI = &getRegisterInfo();
2313 // If we're in SSA, get the defs through the MRI. Otherwise, only look
2314 // within the basic block to see if the register is defined using an LI/LI8.
2316 for (int i = 1, e = MI.getNumOperands(); i < e; i++) {
2317 if (!MI.getOperand(i).isReg())
2319 unsigned Reg = MI.getOperand(i).getReg();
2320 if (!TargetRegisterInfo::isVirtualRegister(Reg))
2322 unsigned TrueReg = TRI->lookThruCopyLike(Reg, MRI);
2323 if (TargetRegisterInfo::isVirtualRegister(TrueReg)) {
2324 DefMI = MRI->getVRegDef(TrueReg);
2325 if (DefMI->getOpcode() == PPC::LI || DefMI->getOpcode() == PPC::LI8) {
2326 OpNoForForwarding = i;
2332 // Looking back through the definition for each operand could be expensive,
2333 // so exit early if this isn't an instruction that either has an immediate
2334 // form or is already an immediate form that we can handle.
2336 unsigned Opc = MI.getOpcode();
2337 bool ConvertibleImmForm =
2338 Opc == PPC::CMPWI || Opc == PPC::CMPLWI ||
2339 Opc == PPC::CMPDI || Opc == PPC::CMPLDI ||
2340 Opc == PPC::ADDI || Opc == PPC::ADDI8 ||
2341 Opc == PPC::ORI || Opc == PPC::ORI8 ||
2342 Opc == PPC::XORI || Opc == PPC::XORI8 ||
2343 Opc == PPC::RLDICL || Opc == PPC::RLDICLo ||
2344 Opc == PPC::RLDICL_32 || Opc == PPC::RLDICL_32_64 ||
2345 Opc == PPC::RLWINM || Opc == PPC::RLWINMo ||
2346 Opc == PPC::RLWINM8 || Opc == PPC::RLWINM8o;
2347 if (!instrHasImmForm(MI, III, true) && !ConvertibleImmForm)
2350 // Don't convert or %X, %Y, %Y since that's just a register move.
2351 if ((Opc == PPC::OR || Opc == PPC::OR8) &&
2352 MI.getOperand(1).getReg() == MI.getOperand(2).getReg())
2354 for (int i = 1, e = MI.getNumOperands(); i < e; i++) {
2355 MachineOperand &MO = MI.getOperand(i);
2356 SeenIntermediateUse = false;
2357 if (MO.isReg() && MO.isUse() && !MO.isImplicit()) {
2358 MachineBasicBlock::reverse_iterator E = MI.getParent()->rend(), It = MI;
2360 unsigned Reg = MI.getOperand(i).getReg();
2361 // MachineInstr::readsRegister only returns true if the machine
2362 // instruction reads the exact register or its super-register. It
2363 // does not consider uses of sub-registers which seems like strange
2364 // behaviour. Nonetheless, if we end up with a 64-bit register here,
2365 // get the corresponding 32-bit register to check.
2366 if (PPC::G8RCRegClass.contains(Reg))
2367 Reg = Reg - PPC::X0 + PPC::R0;
2369 // Is this register defined by some form of add-immediate (including
2370 // load-immediate) within this basic block?
2371 for ( ; It != E; ++It) {
2372 if (It->modifiesRegister(Reg, &getRegisterInfo())) {
2373 switch (It->getOpcode()) {
2380 OpNoForForwarding = i;
2384 } else if (It->readsRegister(Reg, &getRegisterInfo()))
2385 // If we see another use of this reg between the def and the MI,
2386 // we want to flat it so the def isn't deleted.
2387 SeenIntermediateUse = true;
2392 return OpNoForForwarding == ~0U ? nullptr : DefMI;
2395 const unsigned *PPCInstrInfo::getStoreOpcodesForSpillArray() const {
2396 static const unsigned OpcodesForSpill[2][SOK_LastOpcodeSpill] = {
2398 {PPC::STW, PPC::STD, PPC::STFD, PPC::STFS, PPC::SPILL_CR,
2399 PPC::SPILL_CRBIT, PPC::STVX, PPC::STXVD2X, PPC::STXSDX, PPC::STXSSPX,
2400 PPC::SPILL_VRSAVE, PPC::QVSTFDX, PPC::QVSTFSXs, PPC::QVSTFDXb,
2401 PPC::SPILLTOVSR_ST, PPC::EVSTDD, PPC::SPESTW},
2403 {PPC::STW, PPC::STD, PPC::STFD, PPC::STFS, PPC::SPILL_CR,
2404 PPC::SPILL_CRBIT, PPC::STVX, PPC::STXV, PPC::DFSTOREf64, PPC::DFSTOREf32,
2405 PPC::SPILL_VRSAVE, PPC::QVSTFDX, PPC::QVSTFSXs, PPC::QVSTFDXb,
2406 PPC::SPILLTOVSR_ST}};
2408 return OpcodesForSpill[(Subtarget.hasP9Vector()) ? 1 : 0];
2411 const unsigned *PPCInstrInfo::getLoadOpcodesForSpillArray() const {
2412 static const unsigned OpcodesForSpill[2][SOK_LastOpcodeSpill] = {
2414 {PPC::LWZ, PPC::LD, PPC::LFD, PPC::LFS, PPC::RESTORE_CR,
2415 PPC::RESTORE_CRBIT, PPC::LVX, PPC::LXVD2X, PPC::LXSDX, PPC::LXSSPX,
2416 PPC::RESTORE_VRSAVE, PPC::QVLFDX, PPC::QVLFSXs, PPC::QVLFDXb,
2417 PPC::SPILLTOVSR_LD, PPC::EVLDD, PPC::SPELWZ},
2419 {PPC::LWZ, PPC::LD, PPC::LFD, PPC::LFS, PPC::RESTORE_CR,
2420 PPC::RESTORE_CRBIT, PPC::LVX, PPC::LXV, PPC::DFLOADf64, PPC::DFLOADf32,
2421 PPC::RESTORE_VRSAVE, PPC::QVLFDX, PPC::QVLFSXs, PPC::QVLFDXb,
2422 PPC::SPILLTOVSR_LD}};
2424 return OpcodesForSpill[(Subtarget.hasP9Vector()) ? 1 : 0];
2427 // If this instruction has an immediate form and one of its operands is a
2428 // result of a load-immediate or an add-immediate, convert it to
2429 // the immediate form if the constant is in range.
2430 bool PPCInstrInfo::convertToImmediateForm(MachineInstr &MI,
2431 MachineInstr **KilledDef) const {
2432 MachineFunction *MF = MI.getParent()->getParent();
2433 MachineRegisterInfo *MRI = &MF->getRegInfo();
2434 bool PostRA = !MRI->isSSA();
2435 bool SeenIntermediateUse = true;
2436 unsigned ForwardingOperand = ~0U;
2437 MachineInstr *DefMI = getForwardingDefMI(MI, ForwardingOperand,
2438 SeenIntermediateUse);
2441 assert(ForwardingOperand < MI.getNumOperands() &&
2442 "The forwarding operand needs to be valid at this point");
2443 bool KillFwdDefMI = !SeenIntermediateUse &&
2444 MI.getOperand(ForwardingOperand).isKill();
2445 if (KilledDef && KillFwdDefMI)
2449 bool HasImmForm = instrHasImmForm(MI, III, PostRA);
2450 // If this is a reg+reg instruction that has a reg+imm form,
2451 // and one of the operands is produced by an add-immediate,
2452 // try to convert it.
2453 if (HasImmForm && transformToImmFormFedByAdd(MI, III, ForwardingOperand,
2454 *DefMI, KillFwdDefMI))
2457 if ((DefMI->getOpcode() != PPC::LI && DefMI->getOpcode() != PPC::LI8) ||
2458 !DefMI->getOperand(1).isImm())
2461 int64_t Immediate = DefMI->getOperand(1).getImm();
2462 // Sign-extend to 64-bits.
2463 int64_t SExtImm = ((uint64_t)Immediate & ~0x7FFFuLL) != 0 ?
2464 (Immediate | 0xFFFFFFFFFFFF0000) : Immediate;
2466 // If this is a reg+reg instruction that has a reg+imm form,
2467 // and one of the operands is produced by LI, convert it now.
2469 return transformToImmFormFedByLI(MI, III, ForwardingOperand, SExtImm);
2471 bool ReplaceWithLI = false;
2472 bool Is64BitLI = false;
2475 unsigned Opc = MI.getOpcode();
2477 default: return false;
2479 // FIXME: Any branches conditional on such a comparison can be made
2480 // unconditional. At this time, this happens too infrequently to be worth
2481 // the implementation effort, but if that ever changes, we could convert
2482 // such a pattern here.
2487 // Doing this post-RA would require dataflow analysis to reliably find uses
2488 // of the CR register set by the compare.
2491 // If a compare-immediate is fed by an immediate and is itself an input of
2492 // an ISEL (the most common case) into a COPY of the correct register.
2493 bool Changed = false;
2494 unsigned DefReg = MI.getOperand(0).getReg();
2495 int64_t Comparand = MI.getOperand(2).getImm();
2496 int64_t SExtComparand = ((uint64_t)Comparand & ~0x7FFFuLL) != 0 ?
2497 (Comparand | 0xFFFFFFFFFFFF0000) : Comparand;
2499 for (auto &CompareUseMI : MRI->use_instructions(DefReg)) {
2500 unsigned UseOpc = CompareUseMI.getOpcode();
2501 if (UseOpc != PPC::ISEL && UseOpc != PPC::ISEL8)
2503 unsigned CRSubReg = CompareUseMI.getOperand(3).getSubReg();
2504 unsigned TrueReg = CompareUseMI.getOperand(1).getReg();
2505 unsigned FalseReg = CompareUseMI.getOperand(2).getReg();
2506 unsigned RegToCopy = selectReg(SExtImm, SExtComparand, Opc, TrueReg,
2507 FalseReg, CRSubReg);
2508 if (RegToCopy == PPC::NoRegister)
2510 // Can't use PPC::COPY to copy PPC::ZERO[8]. Convert it to LI[8] 0.
2511 if (RegToCopy == PPC::ZERO || RegToCopy == PPC::ZERO8) {
2512 CompareUseMI.setDesc(get(UseOpc == PPC::ISEL8 ? PPC::LI8 : PPC::LI));
2513 replaceInstrOperandWithImm(CompareUseMI, 1, 0);
2514 CompareUseMI.RemoveOperand(3);
2515 CompareUseMI.RemoveOperand(2);
2519 dbgs() << "Found LI -> CMPI -> ISEL, replacing with a copy.\n");
2520 LLVM_DEBUG(DefMI->dump(); MI.dump(); CompareUseMI.dump());
2521 LLVM_DEBUG(dbgs() << "Is converted to:\n");
2522 // Convert to copy and remove unneeded operands.
2523 CompareUseMI.setDesc(get(PPC::COPY));
2524 CompareUseMI.RemoveOperand(3);
2525 CompareUseMI.RemoveOperand(RegToCopy == TrueReg ? 2 : 1);
2526 CmpIselsConverted++;
2528 LLVM_DEBUG(CompareUseMI.dump());
2532 // This may end up incremented multiple times since this function is called
2533 // during a fixed-point transformation, but it is only meant to indicate the
2534 // presence of this opportunity.
2535 MissedConvertibleImmediateInstrs++;
2539 // Immediate forms - may simply be convertable to an LI.
2542 // Does the sum fit in a 16-bit signed field?
2543 int64_t Addend = MI.getOperand(2).getImm();
2544 if (isInt<16>(Addend + SExtImm)) {
2545 ReplaceWithLI = true;
2546 Is64BitLI = Opc == PPC::ADDI8;
2547 NewImm = Addend + SExtImm;
2554 case PPC::RLDICL_32:
2555 case PPC::RLDICL_32_64: {
2556 // Use APInt's rotate function.
2557 int64_t SH = MI.getOperand(2).getImm();
2558 int64_t MB = MI.getOperand(3).getImm();
2559 APInt InVal((Opc == PPC::RLDICL || Opc == PPC::RLDICLo) ?
2560 64 : 32, SExtImm, true);
2561 InVal = InVal.rotl(SH);
2562 uint64_t Mask = (1LLU << (63 - MB + 1)) - 1;
2564 // Can't replace negative values with an LI as that will sign-extend
2565 // and not clear the left bits. If we're setting the CR bit, we will use
2566 // ANDIo which won't sign extend, so that's safe.
2567 if (isUInt<15>(InVal.getSExtValue()) ||
2568 (Opc == PPC::RLDICLo && isUInt<16>(InVal.getSExtValue()))) {
2569 ReplaceWithLI = true;
2570 Is64BitLI = Opc != PPC::RLDICL_32;
2571 NewImm = InVal.getSExtValue();
2572 SetCR = Opc == PPC::RLDICLo;
2580 case PPC::RLWINM8o: {
2581 int64_t SH = MI.getOperand(2).getImm();
2582 int64_t MB = MI.getOperand(3).getImm();
2583 int64_t ME = MI.getOperand(4).getImm();
2584 APInt InVal(32, SExtImm, true);
2585 InVal = InVal.rotl(SH);
2586 // Set the bits ( MB + 32 ) to ( ME + 32 ).
2587 uint64_t Mask = ((1LLU << (32 - MB)) - 1) & ~((1LLU << (31 - ME)) - 1);
2589 // Can't replace negative values with an LI as that will sign-extend
2590 // and not clear the left bits. If we're setting the CR bit, we will use
2591 // ANDIo which won't sign extend, so that's safe.
2592 bool ValueFits = isUInt<15>(InVal.getSExtValue());
2593 ValueFits |= ((Opc == PPC::RLWINMo || Opc == PPC::RLWINM8o) &&
2594 isUInt<16>(InVal.getSExtValue()));
2596 ReplaceWithLI = true;
2597 Is64BitLI = Opc == PPC::RLWINM8 || Opc == PPC::RLWINM8o;
2598 NewImm = InVal.getSExtValue();
2599 SetCR = Opc == PPC::RLWINMo || Opc == PPC::RLWINM8o;
2608 int64_t LogicalImm = MI.getOperand(2).getImm();
2610 if (Opc == PPC::ORI || Opc == PPC::ORI8)
2611 Result = LogicalImm | SExtImm;
2613 Result = LogicalImm ^ SExtImm;
2614 if (isInt<16>(Result)) {
2615 ReplaceWithLI = true;
2616 Is64BitLI = Opc == PPC::ORI8 || Opc == PPC::XORI8;
2624 if (ReplaceWithLI) {
2625 // We need to be careful with CR-setting instructions we're replacing.
2627 // We don't know anything about uses when we're out of SSA, so only
2628 // replace if the new immediate will be reproduced.
2629 bool ImmChanged = (SExtImm & NewImm) != NewImm;
2630 if (PostRA && ImmChanged)
2634 // If the defining load-immediate has no other uses, we can just replace
2635 // the immediate with the new immediate.
2636 if (MRI->hasOneUse(DefMI->getOperand(0).getReg()))
2637 DefMI->getOperand(1).setImm(NewImm);
2639 // If we're not using the GPR result of the CR-setting instruction, we
2640 // just need to and with zero/non-zero depending on the new immediate.
2641 else if (MRI->use_empty(MI.getOperand(0).getReg())) {
2643 assert(Immediate && "Transformation converted zero to non-zero?");
2647 else if (ImmChanged)
2652 LLVM_DEBUG(dbgs() << "Replacing instruction:\n");
2653 LLVM_DEBUG(MI.dump());
2654 LLVM_DEBUG(dbgs() << "Fed by:\n");
2655 LLVM_DEBUG(DefMI->dump());
2656 LoadImmediateInfo LII;
2658 LII.Is64Bit = Is64BitLI;
2660 // If we're setting the CR, the original load-immediate must be kept (as an
2661 // operand to ANDIo/ANDI8o).
2662 if (KilledDef && SetCR)
2663 *KilledDef = nullptr;
2664 replaceInstrWithLI(MI, LII);
2665 LLVM_DEBUG(dbgs() << "With:\n");
2666 LLVM_DEBUG(MI.dump());
2672 static bool isVFReg(unsigned Reg) {
2673 return PPC::VFRCRegClass.contains(Reg);
2676 bool PPCInstrInfo::instrHasImmForm(const MachineInstr &MI,
2677 ImmInstrInfo &III, bool PostRA) const {
2678 unsigned Opc = MI.getOpcode();
2679 // The vast majority of the instructions would need their operand 2 replaced
2680 // with an immediate when switching to the reg+imm form. A marked exception
2681 // are the update form loads/stores for which a constant operand 2 would need
2682 // to turn into a displacement and move operand 1 to the operand 2 position.
2684 III.OpNoForForwarding = 2;
2686 III.ImmMustBeMultipleOf = 1;
2687 III.TruncateImmTo = 0;
2688 III.IsSummingOperands = false;
2690 default: return false;
2693 III.SignedImm = true;
2694 III.ZeroIsSpecialOrig = 0;
2695 III.ZeroIsSpecialNew = 1;
2696 III.IsCommutative = true;
2697 III.IsSummingOperands = true;
2698 III.ImmOpcode = Opc == PPC::ADD4 ? PPC::ADDI : PPC::ADDI8;
2702 III.SignedImm = true;
2703 III.ZeroIsSpecialOrig = 0;
2704 III.ZeroIsSpecialNew = 0;
2705 III.IsCommutative = true;
2706 III.IsSummingOperands = true;
2707 III.ImmOpcode = Opc == PPC::ADDC ? PPC::ADDIC : PPC::ADDIC8;
2710 III.SignedImm = true;
2711 III.ZeroIsSpecialOrig = 0;
2712 III.ZeroIsSpecialNew = 0;
2713 III.IsCommutative = true;
2714 III.IsSummingOperands = true;
2715 III.ImmOpcode = PPC::ADDICo;
2719 III.SignedImm = true;
2720 III.ZeroIsSpecialOrig = 0;
2721 III.ZeroIsSpecialNew = 0;
2722 III.IsCommutative = false;
2723 III.ImmOpcode = Opc == PPC::SUBFC ? PPC::SUBFIC : PPC::SUBFIC8;
2727 III.SignedImm = true;
2728 III.ZeroIsSpecialOrig = 0;
2729 III.ZeroIsSpecialNew = 0;
2730 III.IsCommutative = false;
2731 III.ImmOpcode = Opc == PPC::CMPW ? PPC::CMPWI : PPC::CMPDI;
2735 III.SignedImm = false;
2736 III.ZeroIsSpecialOrig = 0;
2737 III.ZeroIsSpecialNew = 0;
2738 III.IsCommutative = false;
2739 III.ImmOpcode = Opc == PPC::CMPLW ? PPC::CMPLWI : PPC::CMPLDI;
2747 III.SignedImm = false;
2748 III.ZeroIsSpecialOrig = 0;
2749 III.ZeroIsSpecialNew = 0;
2750 III.IsCommutative = true;
2752 default: llvm_unreachable("Unknown opcode");
2753 case PPC::ANDo: III.ImmOpcode = PPC::ANDIo; break;
2754 case PPC::AND8o: III.ImmOpcode = PPC::ANDIo8; break;
2755 case PPC::OR: III.ImmOpcode = PPC::ORI; break;
2756 case PPC::OR8: III.ImmOpcode = PPC::ORI8; break;
2757 case PPC::XOR: III.ImmOpcode = PPC::XORI; break;
2758 case PPC::XOR8: III.ImmOpcode = PPC::XORI8; break;
2775 III.SignedImm = false;
2776 III.ZeroIsSpecialOrig = 0;
2777 III.ZeroIsSpecialNew = 0;
2778 III.IsCommutative = false;
2779 // This isn't actually true, but the instructions ignore any of the
2780 // upper bits, so any immediate loaded with an LI is acceptable.
2781 // This does not apply to shift right algebraic because a value
2782 // out of range will produce a -1/0.
2784 if (Opc == PPC::RLWNM || Opc == PPC::RLWNM8 ||
2785 Opc == PPC::RLWNMo || Opc == PPC::RLWNM8o)
2786 III.TruncateImmTo = 5;
2788 III.TruncateImmTo = 6;
2790 default: llvm_unreachable("Unknown opcode");
2791 case PPC::RLWNM: III.ImmOpcode = PPC::RLWINM; break;
2792 case PPC::RLWNM8: III.ImmOpcode = PPC::RLWINM8; break;
2793 case PPC::RLWNMo: III.ImmOpcode = PPC::RLWINMo; break;
2794 case PPC::RLWNM8o: III.ImmOpcode = PPC::RLWINM8o; break;
2795 case PPC::SLW: III.ImmOpcode = PPC::RLWINM; break;
2796 case PPC::SLW8: III.ImmOpcode = PPC::RLWINM8; break;
2797 case PPC::SLWo: III.ImmOpcode = PPC::RLWINMo; break;
2798 case PPC::SLW8o: III.ImmOpcode = PPC::RLWINM8o; break;
2799 case PPC::SRW: III.ImmOpcode = PPC::RLWINM; break;
2800 case PPC::SRW8: III.ImmOpcode = PPC::RLWINM8; break;
2801 case PPC::SRWo: III.ImmOpcode = PPC::RLWINMo; break;
2802 case PPC::SRW8o: III.ImmOpcode = PPC::RLWINM8o; break;
2805 III.TruncateImmTo = 0;
2806 III.ImmOpcode = PPC::SRAWI;
2810 III.TruncateImmTo = 0;
2811 III.ImmOpcode = PPC::SRAWIo;
2825 III.SignedImm = false;
2826 III.ZeroIsSpecialOrig = 0;
2827 III.ZeroIsSpecialNew = 0;
2828 III.IsCommutative = false;
2829 // This isn't actually true, but the instructions ignore any of the
2830 // upper bits, so any immediate loaded with an LI is acceptable.
2831 // This does not apply to shift right algebraic because a value
2832 // out of range will produce a -1/0.
2834 if (Opc == PPC::RLDCL || Opc == PPC::RLDCLo ||
2835 Opc == PPC::RLDCR || Opc == PPC::RLDCRo)
2836 III.TruncateImmTo = 6;
2838 III.TruncateImmTo = 7;
2840 default: llvm_unreachable("Unknown opcode");
2841 case PPC::RLDCL: III.ImmOpcode = PPC::RLDICL; break;
2842 case PPC::RLDCLo: III.ImmOpcode = PPC::RLDICLo; break;
2843 case PPC::RLDCR: III.ImmOpcode = PPC::RLDICR; break;
2844 case PPC::RLDCRo: III.ImmOpcode = PPC::RLDICRo; break;
2845 case PPC::SLD: III.ImmOpcode = PPC::RLDICR; break;
2846 case PPC::SLDo: III.ImmOpcode = PPC::RLDICRo; break;
2847 case PPC::SRD: III.ImmOpcode = PPC::RLDICL; break;
2848 case PPC::SRDo: III.ImmOpcode = PPC::RLDICLo; break;
2851 III.TruncateImmTo = 0;
2852 III.ImmOpcode = PPC::SRADI;
2856 III.TruncateImmTo = 0;
2857 III.ImmOpcode = PPC::SRADIo;
2861 // Loads and stores:
2883 III.SignedImm = true;
2884 III.ZeroIsSpecialOrig = 1;
2885 III.ZeroIsSpecialNew = 2;
2886 III.IsCommutative = true;
2887 III.IsSummingOperands = true;
2889 III.OpNoForForwarding = 2;
2891 default: llvm_unreachable("Unknown opcode");
2892 case PPC::LBZX: III.ImmOpcode = PPC::LBZ; break;
2893 case PPC::LBZX8: III.ImmOpcode = PPC::LBZ8; break;
2894 case PPC::LHZX: III.ImmOpcode = PPC::LHZ; break;
2895 case PPC::LHZX8: III.ImmOpcode = PPC::LHZ8; break;
2896 case PPC::LHAX: III.ImmOpcode = PPC::LHA; break;
2897 case PPC::LHAX8: III.ImmOpcode = PPC::LHA8; break;
2898 case PPC::LWZX: III.ImmOpcode = PPC::LWZ; break;
2899 case PPC::LWZX8: III.ImmOpcode = PPC::LWZ8; break;
2901 III.ImmOpcode = PPC::LWA;
2902 III.ImmMustBeMultipleOf = 4;
2904 case PPC::LDX: III.ImmOpcode = PPC::LD; III.ImmMustBeMultipleOf = 4; break;
2905 case PPC::LFSX: III.ImmOpcode = PPC::LFS; break;
2906 case PPC::LFDX: III.ImmOpcode = PPC::LFD; break;
2907 case PPC::STBX: III.ImmOpcode = PPC::STB; break;
2908 case PPC::STBX8: III.ImmOpcode = PPC::STB8; break;
2909 case PPC::STHX: III.ImmOpcode = PPC::STH; break;
2910 case PPC::STHX8: III.ImmOpcode = PPC::STH8; break;
2911 case PPC::STWX: III.ImmOpcode = PPC::STW; break;
2912 case PPC::STWX8: III.ImmOpcode = PPC::STW8; break;
2914 III.ImmOpcode = PPC::STD;
2915 III.ImmMustBeMultipleOf = 4;
2917 case PPC::STFSX: III.ImmOpcode = PPC::STFS; break;
2918 case PPC::STFDX: III.ImmOpcode = PPC::STFD; break;
2941 III.SignedImm = true;
2942 III.ZeroIsSpecialOrig = 2;
2943 III.ZeroIsSpecialNew = 3;
2944 III.IsCommutative = false;
2945 III.IsSummingOperands = true;
2947 III.OpNoForForwarding = 3;
2949 default: llvm_unreachable("Unknown opcode");
2950 case PPC::LBZUX: III.ImmOpcode = PPC::LBZU; break;
2951 case PPC::LBZUX8: III.ImmOpcode = PPC::LBZU8; break;
2952 case PPC::LHZUX: III.ImmOpcode = PPC::LHZU; break;
2953 case PPC::LHZUX8: III.ImmOpcode = PPC::LHZU8; break;
2954 case PPC::LHAUX: III.ImmOpcode = PPC::LHAU; break;
2955 case PPC::LHAUX8: III.ImmOpcode = PPC::LHAU8; break;
2956 case PPC::LWZUX: III.ImmOpcode = PPC::LWZU; break;
2957 case PPC::LWZUX8: III.ImmOpcode = PPC::LWZU8; break;
2959 III.ImmOpcode = PPC::LDU;
2960 III.ImmMustBeMultipleOf = 4;
2962 case PPC::LFSUX: III.ImmOpcode = PPC::LFSU; break;
2963 case PPC::LFDUX: III.ImmOpcode = PPC::LFDU; break;
2964 case PPC::STBUX: III.ImmOpcode = PPC::STBU; break;
2965 case PPC::STBUX8: III.ImmOpcode = PPC::STBU8; break;
2966 case PPC::STHUX: III.ImmOpcode = PPC::STHU; break;
2967 case PPC::STHUX8: III.ImmOpcode = PPC::STHU8; break;
2968 case PPC::STWUX: III.ImmOpcode = PPC::STWU; break;
2969 case PPC::STWUX8: III.ImmOpcode = PPC::STWU8; break;
2971 III.ImmOpcode = PPC::STDU;
2972 III.ImmMustBeMultipleOf = 4;
2974 case PPC::STFSUX: III.ImmOpcode = PPC::STFSU; break;
2975 case PPC::STFDUX: III.ImmOpcode = PPC::STFDU; break;
2978 // Power9 and up only. For some of these, the X-Form version has access to all
2979 // 64 VSR's whereas the D-Form only has access to the VR's. We replace those
2980 // with pseudo-ops pre-ra and for post-ra, we check that the register loaded
2981 // into or stored from is one of the VR registers.
2988 case PPC::XFLOADf32:
2989 case PPC::XFLOADf64:
2990 case PPC::XFSTOREf32:
2991 case PPC::XFSTOREf64:
2992 if (!Subtarget.hasP9Vector())
2994 III.SignedImm = true;
2995 III.ZeroIsSpecialOrig = 1;
2996 III.ZeroIsSpecialNew = 2;
2997 III.IsCommutative = true;
2998 III.IsSummingOperands = true;
3000 III.OpNoForForwarding = 2;
3001 III.ImmMustBeMultipleOf = 4;
3003 default: llvm_unreachable("Unknown opcode");
3005 III.ImmOpcode = PPC::LXV;
3006 III.ImmMustBeMultipleOf = 16;
3010 if (isVFReg(MI.getOperand(0).getReg()))
3011 III.ImmOpcode = PPC::LXSSP;
3013 III.ImmOpcode = PPC::LFS;
3014 III.ImmMustBeMultipleOf = 1;
3019 case PPC::XFLOADf32:
3020 III.ImmOpcode = PPC::DFLOADf32;
3024 if (isVFReg(MI.getOperand(0).getReg()))
3025 III.ImmOpcode = PPC::LXSD;
3027 III.ImmOpcode = PPC::LFD;
3028 III.ImmMustBeMultipleOf = 1;
3033 case PPC::XFLOADf64:
3034 III.ImmOpcode = PPC::DFLOADf64;
3037 III.ImmOpcode = PPC::STXV;
3038 III.ImmMustBeMultipleOf = 16;
3042 if (isVFReg(MI.getOperand(0).getReg()))
3043 III.ImmOpcode = PPC::STXSSP;
3045 III.ImmOpcode = PPC::STFS;
3046 III.ImmMustBeMultipleOf = 1;
3051 case PPC::XFSTOREf32:
3052 III.ImmOpcode = PPC::DFSTOREf32;
3056 if (isVFReg(MI.getOperand(0).getReg()))
3057 III.ImmOpcode = PPC::STXSD;
3059 III.ImmOpcode = PPC::STFD;
3060 III.ImmMustBeMultipleOf = 1;
3065 case PPC::XFSTOREf64:
3066 III.ImmOpcode = PPC::DFSTOREf64;
3074 // Utility function for swaping two arbitrary operands of an instruction.
3075 static void swapMIOperands(MachineInstr &MI, unsigned Op1, unsigned Op2) {
3076 assert(Op1 != Op2 && "Cannot swap operand with itself.");
3078 unsigned MaxOp = std::max(Op1, Op2);
3079 unsigned MinOp = std::min(Op1, Op2);
3080 MachineOperand MOp1 = MI.getOperand(MinOp);
3081 MachineOperand MOp2 = MI.getOperand(MaxOp);
3082 MI.RemoveOperand(std::max(Op1, Op2));
3083 MI.RemoveOperand(std::min(Op1, Op2));
3085 // If the operands we are swapping are the two at the end (the common case)
3086 // we can just remove both and add them in the opposite order.
3087 if (MaxOp - MinOp == 1 && MI.getNumOperands() == MinOp) {
3088 MI.addOperand(MOp2);
3089 MI.addOperand(MOp1);
3091 // Store all operands in a temporary vector, remove them and re-add in the
3093 SmallVector<MachineOperand, 2> MOps;
3094 unsigned TotalOps = MI.getNumOperands() + 2; // We've already removed 2 ops.
3095 for (unsigned i = MI.getNumOperands() - 1; i >= MinOp; i--) {
3096 MOps.push_back(MI.getOperand(i));
3097 MI.RemoveOperand(i);
3099 // MOp2 needs to be added next.
3100 MI.addOperand(MOp2);
3101 // Now add the rest.
3102 for (unsigned i = MI.getNumOperands(); i < TotalOps; i++) {
3104 MI.addOperand(MOp1);
3106 MI.addOperand(MOps.back());
3113 // Check if the 'MI' that has the index OpNoForForwarding
3114 // meets the requirement described in the ImmInstrInfo.
3115 bool PPCInstrInfo::isUseMIElgibleForForwarding(MachineInstr &MI,
3116 const ImmInstrInfo &III,
3117 unsigned OpNoForForwarding
3119 // As the algorithm of checking for PPC::ZERO/PPC::ZERO8
3120 // would not work pre-RA, we can only do the check post RA.
3121 MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
3125 // Cannot do the transform if MI isn't summing the operands.
3126 if (!III.IsSummingOperands)
3129 // The instruction we are trying to replace must have the ZeroIsSpecialOrig set.
3130 if (!III.ZeroIsSpecialOrig)
3133 // We cannot do the transform if the operand we are trying to replace
3134 // isn't the same as the operand the instruction allows.
3135 if (OpNoForForwarding != III.OpNoForForwarding)
3138 // Check if the instruction we are trying to transform really has
3139 // the special zero register as its operand.
3140 if (MI.getOperand(III.ZeroIsSpecialOrig).getReg() != PPC::ZERO &&
3141 MI.getOperand(III.ZeroIsSpecialOrig).getReg() != PPC::ZERO8)
3144 // This machine instruction is convertible if it is,
3145 // 1. summing the operands.
3146 // 2. one of the operands is special zero register.
3147 // 3. the operand we are trying to replace is allowed by the MI.
3151 // Check if the DefMI is the add inst and set the ImmMO and RegMO
3153 bool PPCInstrInfo::isDefMIElgibleForForwarding(MachineInstr &DefMI,
3154 const ImmInstrInfo &III,
3155 MachineOperand *&ImmMO,
3156 MachineOperand *&RegMO) const {
3157 unsigned Opc = DefMI.getOpcode();
3158 if (Opc != PPC::ADDItocL && Opc != PPC::ADDI && Opc != PPC::ADDI8)
3161 assert(DefMI.getNumOperands() >= 3 &&
3162 "Add inst must have at least three operands");
3163 RegMO = &DefMI.getOperand(1);
3164 ImmMO = &DefMI.getOperand(2);
3166 // This DefMI is elgible for forwarding if it is:
3168 // 2. one of the operands is Imm/CPI/Global.
3169 return isAnImmediateOperand(*ImmMO);
3172 bool PPCInstrInfo::isRegElgibleForForwarding(const MachineOperand &RegMO,
3173 const MachineInstr &DefMI,
3174 const MachineInstr &MI,
3179 // z = lfdx 0, x -> z = lfd imm(y)
3180 // The Reg "y" can be forwarded to the MI(z) only when there is no DEF
3181 // of "y" between the DEF of "x" and "z".
3182 // The query is only valid post RA.
3183 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
3187 // MachineInstr::readsRegister only returns true if the machine
3188 // instruction reads the exact register or its super-register. It
3189 // does not consider uses of sub-registers which seems like strange
3190 // behaviour. Nonetheless, if we end up with a 64-bit register here,
3191 // get the corresponding 32-bit register to check.
3192 unsigned Reg = RegMO.getReg();
3193 if (PPC::G8RCRegClass.contains(Reg))
3194 Reg = Reg - PPC::X0 + PPC::R0;
3196 // Walking the inst in reverse(MI-->DefMI) to get the last DEF of the Reg.
3197 MachineBasicBlock::const_reverse_iterator It = MI;
3198 MachineBasicBlock::const_reverse_iterator E = MI.getParent()->rend();
3200 for (; It != E; ++It) {
3201 if (It->modifiesRegister(Reg, &getRegisterInfo()) && (&*It) != &DefMI)
3203 // Made it to DefMI without encountering a clobber.
3204 if ((&*It) == &DefMI)
3207 assert((&*It) == &DefMI && "DefMI is missing");
3209 // If DefMI also uses the register to be forwarded, we can only forward it
3210 // if DefMI is being erased.
3211 if (DefMI.readsRegister(Reg, &getRegisterInfo()))
3217 bool PPCInstrInfo::isImmElgibleForForwarding(const MachineOperand &ImmMO,
3218 const MachineInstr &DefMI,
3219 const ImmInstrInfo &III,
3220 int64_t &Imm) const {
3221 assert(isAnImmediateOperand(ImmMO) && "ImmMO is NOT an immediate");
3222 if (DefMI.getOpcode() == PPC::ADDItocL) {
3223 // The operand for ADDItocL is CPI, which isn't imm at compiling time,
3224 // However, we know that, it is 16-bit width, and has the alignment of 4.
3225 // Check if the instruction met the requirement.
3226 if (III.ImmMustBeMultipleOf > 4 ||
3227 III.TruncateImmTo || III.ImmWidth != 16)
3230 // Going from XForm to DForm loads means that the displacement needs to be
3231 // not just an immediate but also a multiple of 4, or 16 depending on the
3232 // load. A DForm load cannot be represented if it is a multiple of say 2.
3233 // XForm loads do not have this restriction.
3234 if (ImmMO.isGlobal() &&
3235 ImmMO.getGlobal()->getAlignment() < III.ImmMustBeMultipleOf)
3241 if (ImmMO.isImm()) {
3242 // It is Imm, we need to check if the Imm fit the range.
3243 int64_t Immediate = ImmMO.getImm();
3244 // Sign-extend to 64-bits.
3245 Imm = ((uint64_t)Immediate & ~0x7FFFuLL) != 0 ?
3246 (Immediate | 0xFFFFFFFFFFFF0000) : Immediate;
3248 if (Imm % III.ImmMustBeMultipleOf)
3250 if (III.TruncateImmTo)
3251 Imm &= ((1 << III.TruncateImmTo) - 1);
3252 if (III.SignedImm) {
3253 APInt ActualValue(64, Imm, true);
3254 if (!ActualValue.isSignedIntN(III.ImmWidth))
3257 uint64_t UnsignedMax = (1 << III.ImmWidth) - 1;
3258 if ((uint64_t)Imm > UnsignedMax)
3265 // This ImmMO is forwarded if it meets the requriement describle
3270 // If an X-Form instruction is fed by an add-immediate and one of its operands
3271 // is the literal zero, attempt to forward the source of the add-immediate to
3272 // the corresponding D-Form instruction with the displacement coming from
3273 // the immediate being added.
3274 bool PPCInstrInfo::transformToImmFormFedByAdd(MachineInstr &MI,
3275 const ImmInstrInfo &III,
3276 unsigned OpNoForForwarding,
3277 MachineInstr &DefMI,
3278 bool KillDefMI) const {
3281 // x = addi reg, imm <----- DefMI
3282 // y = op 0 , x <----- MI
3284 // OpNoForForwarding
3285 // Check if the MI meet the requirement described in the III.
3286 if (!isUseMIElgibleForForwarding(MI, III, OpNoForForwarding))
3289 // Check if the DefMI meet the requirement
3290 // described in the III. If yes, set the ImmMO and RegMO accordingly.
3291 MachineOperand *ImmMO = nullptr;
3292 MachineOperand *RegMO = nullptr;
3293 if (!isDefMIElgibleForForwarding(DefMI, III, ImmMO, RegMO))
3295 assert(ImmMO && RegMO && "Imm and Reg operand must have been set");
3297 // As we get the Imm operand now, we need to check if the ImmMO meet
3298 // the requirement described in the III. If yes set the Imm.
3300 if (!isImmElgibleForForwarding(*ImmMO, DefMI, III, Imm))
3303 // Check if the RegMO can be forwarded to MI.
3304 if (!isRegElgibleForForwarding(*RegMO, DefMI, MI, KillDefMI))
3307 // We know that, the MI and DefMI both meet the pattern, and
3308 // the Imm also meet the requirement with the new Imm-form.
3309 // It is safe to do the transformation now.
3310 LLVM_DEBUG(dbgs() << "Replacing instruction:\n");
3311 LLVM_DEBUG(MI.dump());
3312 LLVM_DEBUG(dbgs() << "Fed by:\n");
3313 LLVM_DEBUG(DefMI.dump());
3315 // Update the base reg first.
3316 MI.getOperand(III.OpNoForForwarding).ChangeToRegister(RegMO->getReg(),
3320 // Then, update the imm.
3321 if (ImmMO->isImm()) {
3322 // If the ImmMO is Imm, change the operand that has ZERO to that Imm
3324 replaceInstrOperandWithImm(MI, III.ZeroIsSpecialOrig, Imm);
3327 // Otherwise, it is Constant Pool Index(CPI) or Global,
3328 // which is relocation in fact. We need to replace the special zero
3329 // register with ImmMO.
3330 // Before that, we need to fixup the target flags for imm.
3331 // For some reason, we miss to set the flag for the ImmMO if it is CPI.
3332 if (DefMI.getOpcode() == PPC::ADDItocL)
3333 ImmMO->setTargetFlags(PPCII::MO_TOC_LO);
3335 // MI didn't have the interface such as MI.setOperand(i) though
3336 // it has MI.getOperand(i). To repalce the ZERO MachineOperand with
3337 // ImmMO, we need to remove ZERO operand and all the operands behind it,
3338 // and, add the ImmMO, then, move back all the operands behind ZERO.
3339 SmallVector<MachineOperand, 2> MOps;
3340 for (unsigned i = MI.getNumOperands() - 1; i >= III.ZeroIsSpecialOrig; i--) {
3341 MOps.push_back(MI.getOperand(i));
3342 MI.RemoveOperand(i);
3345 // Remove the last MO in the list, which is ZERO operand in fact.
3347 // Add the imm operand.
3348 MI.addOperand(*ImmMO);
3349 // Now add the rest back.
3350 for (auto &MO : MOps)
3354 // Update the opcode.
3355 MI.setDesc(get(III.ImmOpcode));
3357 LLVM_DEBUG(dbgs() << "With:\n");
3358 LLVM_DEBUG(MI.dump());
3363 bool PPCInstrInfo::transformToImmFormFedByLI(MachineInstr &MI,
3364 const ImmInstrInfo &III,
3365 unsigned ConstantOpNo,
3366 int64_t Imm) const {
3367 MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
3368 bool PostRA = !MRI.isSSA();
3369 // Exit early if we can't convert this.
3370 if ((ConstantOpNo != III.OpNoForForwarding) && !III.IsCommutative)
3372 if (Imm % III.ImmMustBeMultipleOf)
3374 if (III.TruncateImmTo)
3375 Imm &= ((1 << III.TruncateImmTo) - 1);
3376 if (III.SignedImm) {
3377 APInt ActualValue(64, Imm, true);
3378 if (!ActualValue.isSignedIntN(III.ImmWidth))
3381 uint64_t UnsignedMax = (1 << III.ImmWidth) - 1;
3382 if ((uint64_t)Imm > UnsignedMax)
3386 // If we're post-RA, the instructions don't agree on whether register zero is
3387 // special, we can transform this as long as the register operand that will
3388 // end up in the location where zero is special isn't R0.
3389 if (PostRA && III.ZeroIsSpecialOrig != III.ZeroIsSpecialNew) {
3390 unsigned PosForOrigZero = III.ZeroIsSpecialOrig ? III.ZeroIsSpecialOrig :
3391 III.ZeroIsSpecialNew + 1;
3392 unsigned OrigZeroReg = MI.getOperand(PosForOrigZero).getReg();
3393 unsigned NewZeroReg = MI.getOperand(III.ZeroIsSpecialNew).getReg();
3394 // If R0 is in the operand where zero is special for the new instruction,
3395 // it is unsafe to transform if the constant operand isn't that operand.
3396 if ((NewZeroReg == PPC::R0 || NewZeroReg == PPC::X0) &&
3397 ConstantOpNo != III.ZeroIsSpecialNew)
3399 if ((OrigZeroReg == PPC::R0 || OrigZeroReg == PPC::X0) &&
3400 ConstantOpNo != PosForOrigZero)
3404 unsigned Opc = MI.getOpcode();
3405 bool SpecialShift32 =
3406 Opc == PPC::SLW || Opc == PPC::SLWo || Opc == PPC::SRW || Opc == PPC::SRWo;
3407 bool SpecialShift64 =
3408 Opc == PPC::SLD || Opc == PPC::SLDo || Opc == PPC::SRD || Opc == PPC::SRDo;
3409 bool SetCR = Opc == PPC::SLWo || Opc == PPC::SRWo ||
3410 Opc == PPC::SLDo || Opc == PPC::SRDo;
3412 Opc == PPC::SRW || Opc == PPC::SRWo || Opc == PPC::SRD || Opc == PPC::SRDo;
3414 MI.setDesc(get(III.ImmOpcode));
3415 if (ConstantOpNo == III.OpNoForForwarding) {
3416 // Converting shifts to immediate form is a bit tricky since they may do
3417 // one of three things:
3418 // 1. If the shift amount is between OpSize and 2*OpSize, the result is zero
3419 // 2. If the shift amount is zero, the result is unchanged (save for maybe
3421 // 3. If the shift amount is in [1, OpSize), it's just a shift
3422 if (SpecialShift32 || SpecialShift64) {
3423 LoadImmediateInfo LII;
3426 LII.Is64Bit = SpecialShift64;
3427 uint64_t ShAmt = Imm & (SpecialShift32 ? 0x1F : 0x3F);
3428 if (Imm & (SpecialShift32 ? 0x20 : 0x40))
3429 replaceInstrWithLI(MI, LII);
3430 // Shifts by zero don't change the value. If we don't need to set CR0,
3431 // just convert this to a COPY. Can't do this post-RA since we've already
3432 // cleaned up the copies.
3433 else if (!SetCR && ShAmt == 0 && !PostRA) {
3434 MI.RemoveOperand(2);
3435 MI.setDesc(get(PPC::COPY));
3437 // The 32 bit and 64 bit instructions are quite different.
3438 if (SpecialShift32) {
3439 // Left shifts use (N, 0, 31-N), right shifts use (32-N, N, 31).
3440 uint64_t SH = RightShift ? 32 - ShAmt : ShAmt;
3441 uint64_t MB = RightShift ? ShAmt : 0;
3442 uint64_t ME = RightShift ? 31 : 31 - ShAmt;
3443 replaceInstrOperandWithImm(MI, III.OpNoForForwarding, SH);
3444 MachineInstrBuilder(*MI.getParent()->getParent(), MI).addImm(MB)
3447 // Left shifts use (N, 63-N), right shifts use (64-N, N).
3448 uint64_t SH = RightShift ? 64 - ShAmt : ShAmt;
3449 uint64_t ME = RightShift ? ShAmt : 63 - ShAmt;
3450 replaceInstrOperandWithImm(MI, III.OpNoForForwarding, SH);
3451 MachineInstrBuilder(*MI.getParent()->getParent(), MI).addImm(ME);
3455 replaceInstrOperandWithImm(MI, ConstantOpNo, Imm);
3457 // Convert commutative instructions (switch the operands and convert the
3458 // desired one to an immediate.
3459 else if (III.IsCommutative) {
3460 replaceInstrOperandWithImm(MI, ConstantOpNo, Imm);
3461 swapMIOperands(MI, ConstantOpNo, III.OpNoForForwarding);
3463 llvm_unreachable("Should have exited early!");
3465 // For instructions for which the constant register replaces a different
3466 // operand than where the immediate goes, we need to swap them.
3467 if (III.OpNoForForwarding != III.ImmOpNo)
3468 swapMIOperands(MI, III.OpNoForForwarding, III.ImmOpNo);
3470 // If the special R0/X0 register index are different for original instruction
3471 // and new instruction, we need to fix up the register class in new
3473 if (!PostRA && III.ZeroIsSpecialOrig != III.ZeroIsSpecialNew) {
3474 if (III.ZeroIsSpecialNew) {
3475 // If operand at III.ZeroIsSpecialNew is physical reg(eg: ZERO/ZERO8), no
3476 // need to fix up register class.
3477 unsigned RegToModify = MI.getOperand(III.ZeroIsSpecialNew).getReg();
3478 if (TargetRegisterInfo::isVirtualRegister(RegToModify)) {
3479 const TargetRegisterClass *NewRC =
3480 MRI.getRegClass(RegToModify)->hasSuperClassEq(&PPC::GPRCRegClass) ?
3481 &PPC::GPRC_and_GPRC_NOR0RegClass : &PPC::G8RC_and_G8RC_NOX0RegClass;
3482 MRI.setRegClass(RegToModify, NewRC);
3489 const TargetRegisterClass *
3490 PPCInstrInfo::updatedRC(const TargetRegisterClass *RC) const {
3491 if (Subtarget.hasVSX() && RC == &PPC::VRRCRegClass)
3492 return &PPC::VSRCRegClass;
3496 int PPCInstrInfo::getRecordFormOpcode(unsigned Opcode) {
3497 return PPC::getRecordFormOpcode(Opcode);
3500 // This function returns true if the machine instruction
3501 // always outputs a value by sign-extending a 32 bit value,
3502 // i.e. 0 to 31-th bits are same as 32-th bit.
3503 static bool isSignExtendingOp(const MachineInstr &MI) {
3504 int Opcode = MI.getOpcode();
3505 if (Opcode == PPC::LI || Opcode == PPC::LI8 ||
3506 Opcode == PPC::LIS || Opcode == PPC::LIS8 ||
3507 Opcode == PPC::SRAW || Opcode == PPC::SRAWo ||
3508 Opcode == PPC::SRAWI || Opcode == PPC::SRAWIo ||
3509 Opcode == PPC::LWA || Opcode == PPC::LWAX ||
3510 Opcode == PPC::LWA_32 || Opcode == PPC::LWAX_32 ||
3511 Opcode == PPC::LHA || Opcode == PPC::LHAX ||
3512 Opcode == PPC::LHA8 || Opcode == PPC::LHAX8 ||
3513 Opcode == PPC::LBZ || Opcode == PPC::LBZX ||
3514 Opcode == PPC::LBZ8 || Opcode == PPC::LBZX8 ||
3515 Opcode == PPC::LBZU || Opcode == PPC::LBZUX ||
3516 Opcode == PPC::LBZU8 || Opcode == PPC::LBZUX8 ||
3517 Opcode == PPC::LHZ || Opcode == PPC::LHZX ||
3518 Opcode == PPC::LHZ8 || Opcode == PPC::LHZX8 ||
3519 Opcode == PPC::LHZU || Opcode == PPC::LHZUX ||
3520 Opcode == PPC::LHZU8 || Opcode == PPC::LHZUX8 ||
3521 Opcode == PPC::EXTSB || Opcode == PPC::EXTSBo ||
3522 Opcode == PPC::EXTSH || Opcode == PPC::EXTSHo ||
3523 Opcode == PPC::EXTSB8 || Opcode == PPC::EXTSH8 ||
3524 Opcode == PPC::EXTSW || Opcode == PPC::EXTSWo ||
3525 Opcode == PPC::SETB || Opcode == PPC::SETB8 ||
3526 Opcode == PPC::EXTSH8_32_64 || Opcode == PPC::EXTSW_32_64 ||
3527 Opcode == PPC::EXTSB8_32_64)
3530 if (Opcode == PPC::RLDICL && MI.getOperand(3).getImm() >= 33)
3533 if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINMo ||
3534 Opcode == PPC::RLWNM || Opcode == PPC::RLWNMo) &&
3535 MI.getOperand(3).getImm() > 0 &&
3536 MI.getOperand(3).getImm() <= MI.getOperand(4).getImm())
3542 // This function returns true if the machine instruction
3543 // always outputs zeros in higher 32 bits.
3544 static bool isZeroExtendingOp(const MachineInstr &MI) {
3545 int Opcode = MI.getOpcode();
3546 // The 16-bit immediate is sign-extended in li/lis.
3547 // If the most significant bit is zero, all higher bits are zero.
3548 if (Opcode == PPC::LI || Opcode == PPC::LI8 ||
3549 Opcode == PPC::LIS || Opcode == PPC::LIS8) {
3550 int64_t Imm = MI.getOperand(1).getImm();
3551 if (((uint64_t)Imm & ~0x7FFFuLL) == 0)
3555 // We have some variations of rotate-and-mask instructions
3556 // that clear higher 32-bits.
3557 if ((Opcode == PPC::RLDICL || Opcode == PPC::RLDICLo ||
3558 Opcode == PPC::RLDCL || Opcode == PPC::RLDCLo ||
3559 Opcode == PPC::RLDICL_32_64) &&
3560 MI.getOperand(3).getImm() >= 32)
3563 if ((Opcode == PPC::RLDIC || Opcode == PPC::RLDICo) &&
3564 MI.getOperand(3).getImm() >= 32 &&
3565 MI.getOperand(3).getImm() <= 63 - MI.getOperand(2).getImm())
3568 if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINMo ||
3569 Opcode == PPC::RLWNM || Opcode == PPC::RLWNMo ||
3570 Opcode == PPC::RLWINM8 || Opcode == PPC::RLWNM8) &&
3571 MI.getOperand(3).getImm() <= MI.getOperand(4).getImm())
3574 // There are other instructions that clear higher 32-bits.
3575 if (Opcode == PPC::CNTLZW || Opcode == PPC::CNTLZWo ||
3576 Opcode == PPC::CNTTZW || Opcode == PPC::CNTTZWo ||
3577 Opcode == PPC::CNTLZW8 || Opcode == PPC::CNTTZW8 ||
3578 Opcode == PPC::CNTLZD || Opcode == PPC::CNTLZDo ||
3579 Opcode == PPC::CNTTZD || Opcode == PPC::CNTTZDo ||
3580 Opcode == PPC::POPCNTD || Opcode == PPC::POPCNTW ||
3581 Opcode == PPC::SLW || Opcode == PPC::SLWo ||
3582 Opcode == PPC::SRW || Opcode == PPC::SRWo ||
3583 Opcode == PPC::SLW8 || Opcode == PPC::SRW8 ||
3584 Opcode == PPC::SLWI || Opcode == PPC::SLWIo ||
3585 Opcode == PPC::SRWI || Opcode == PPC::SRWIo ||
3586 Opcode == PPC::LWZ || Opcode == PPC::LWZX ||
3587 Opcode == PPC::LWZU || Opcode == PPC::LWZUX ||
3588 Opcode == PPC::LWBRX || Opcode == PPC::LHBRX ||
3589 Opcode == PPC::LHZ || Opcode == PPC::LHZX ||
3590 Opcode == PPC::LHZU || Opcode == PPC::LHZUX ||
3591 Opcode == PPC::LBZ || Opcode == PPC::LBZX ||
3592 Opcode == PPC::LBZU || Opcode == PPC::LBZUX ||
3593 Opcode == PPC::LWZ8 || Opcode == PPC::LWZX8 ||
3594 Opcode == PPC::LWZU8 || Opcode == PPC::LWZUX8 ||
3595 Opcode == PPC::LWBRX8 || Opcode == PPC::LHBRX8 ||
3596 Opcode == PPC::LHZ8 || Opcode == PPC::LHZX8 ||
3597 Opcode == PPC::LHZU8 || Opcode == PPC::LHZUX8 ||
3598 Opcode == PPC::LBZ8 || Opcode == PPC::LBZX8 ||
3599 Opcode == PPC::LBZU8 || Opcode == PPC::LBZUX8 ||
3600 Opcode == PPC::ANDIo || Opcode == PPC::ANDISo ||
3601 Opcode == PPC::ROTRWI || Opcode == PPC::ROTRWIo ||
3602 Opcode == PPC::EXTLWI || Opcode == PPC::EXTLWIo ||
3603 Opcode == PPC::MFVSRWZ)
3609 // This function returns true if the input MachineInstr is a TOC save
3611 bool PPCInstrInfo::isTOCSaveMI(const MachineInstr &MI) const {
3612 if (!MI.getOperand(1).isImm() || !MI.getOperand(2).isReg())
3614 unsigned TOCSaveOffset = Subtarget.getFrameLowering()->getTOCSaveOffset();
3615 unsigned StackOffset = MI.getOperand(1).getImm();
3616 unsigned StackReg = MI.getOperand(2).getReg();
3617 if (StackReg == PPC::X1 && StackOffset == TOCSaveOffset)
3623 // We limit the max depth to track incoming values of PHIs or binary ops
3624 // (e.g. AND) to avoid excessive cost.
3625 const unsigned MAX_DEPTH = 1;
3628 PPCInstrInfo::isSignOrZeroExtended(const MachineInstr &MI, bool SignExt,
3629 const unsigned Depth) const {
3630 const MachineFunction *MF = MI.getParent()->getParent();
3631 const MachineRegisterInfo *MRI = &MF->getRegInfo();
3633 // If we know this instruction returns sign- or zero-extended result,
3635 if (SignExt ? isSignExtendingOp(MI):
3636 isZeroExtendingOp(MI))
3639 switch (MI.getOpcode()) {
3641 unsigned SrcReg = MI.getOperand(1).getReg();
3643 // In both ELFv1 and v2 ABI, method parameters and the return value
3644 // are sign- or zero-extended.
3645 if (MF->getSubtarget<PPCSubtarget>().isSVR4ABI()) {
3646 const PPCFunctionInfo *FuncInfo = MF->getInfo<PPCFunctionInfo>();
3647 // We check the ZExt/SExt flags for a method parameter.
3648 if (MI.getParent()->getBasicBlock() ==
3649 &MF->getFunction().getEntryBlock()) {
3650 unsigned VReg = MI.getOperand(0).getReg();
3651 if (MF->getRegInfo().isLiveIn(VReg))
3652 return SignExt ? FuncInfo->isLiveInSExt(VReg) :
3653 FuncInfo->isLiveInZExt(VReg);
3656 // For a method return value, we check the ZExt/SExt flags in attribute.
3657 // We assume the following code sequence for method call.
3658 // ADJCALLSTACKDOWN 32, implicit dead %r1, implicit %r1
3659 // BL8_NOP @func,...
3660 // ADJCALLSTACKUP 32, 0, implicit dead %r1, implicit %r1
3661 // %5 = COPY %x3; G8RC:%5
3662 if (SrcReg == PPC::X3) {
3663 const MachineBasicBlock *MBB = MI.getParent();
3664 MachineBasicBlock::const_instr_iterator II =
3665 MachineBasicBlock::const_instr_iterator(&MI);
3666 if (II != MBB->instr_begin() &&
3667 (--II)->getOpcode() == PPC::ADJCALLSTACKUP) {
3668 const MachineInstr &CallMI = *(--II);
3669 if (CallMI.isCall() && CallMI.getOperand(0).isGlobal()) {
3670 const Function *CalleeFn =
3671 dyn_cast<Function>(CallMI.getOperand(0).getGlobal());
3674 const IntegerType *IntTy =
3675 dyn_cast<IntegerType>(CalleeFn->getReturnType());
3676 const AttributeSet &Attrs =
3677 CalleeFn->getAttributes().getRetAttributes();
3678 if (IntTy && IntTy->getBitWidth() <= 32)
3679 return Attrs.hasAttribute(SignExt ? Attribute::SExt :
3686 // If this is a copy from another register, we recursively check source.
3687 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
3689 const MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
3691 return isSignOrZeroExtended(*SrcMI, SignExt, Depth);
3708 // logical operation with 16-bit immediate does not change the upper bits.
3709 // So, we track the operand register as we do for register copy.
3710 unsigned SrcReg = MI.getOperand(1).getReg();
3711 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
3713 const MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
3715 return isSignOrZeroExtended(*SrcMI, SignExt, Depth);
3720 // If all incoming values are sign-/zero-extended,
3721 // the output of OR, ISEL or PHI is also sign-/zero-extended.
3726 if (Depth >= MAX_DEPTH)
3729 // The input registers for PHI are operand 1, 3, ...
3730 // The input registers for others are operand 1 and 2.
3731 unsigned E = 3, D = 1;
3732 if (MI.getOpcode() == PPC::PHI) {
3733 E = MI.getNumOperands();
3737 for (unsigned I = 1; I != E; I += D) {
3738 if (MI.getOperand(I).isReg()) {
3739 unsigned SrcReg = MI.getOperand(I).getReg();
3740 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
3742 const MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
3743 if (SrcMI == NULL || !isSignOrZeroExtended(*SrcMI, SignExt, Depth+1))
3752 // If at least one of the incoming values of an AND is zero extended
3753 // then the output is also zero-extended. If both of the incoming values
3754 // are sign-extended then the output is also sign extended.
3757 if (Depth >= MAX_DEPTH)
3760 assert(MI.getOperand(1).isReg() && MI.getOperand(2).isReg());
3762 unsigned SrcReg1 = MI.getOperand(1).getReg();
3763 unsigned SrcReg2 = MI.getOperand(2).getReg();
3765 if (!TargetRegisterInfo::isVirtualRegister(SrcReg1) ||
3766 !TargetRegisterInfo::isVirtualRegister(SrcReg2))
3769 const MachineInstr *MISrc1 = MRI->getVRegDef(SrcReg1);
3770 const MachineInstr *MISrc2 = MRI->getVRegDef(SrcReg2);
3771 if (!MISrc1 || !MISrc2)
3775 return isSignOrZeroExtended(*MISrc1, SignExt, Depth+1) &&
3776 isSignOrZeroExtended(*MISrc2, SignExt, Depth+1);
3778 return isSignOrZeroExtended(*MISrc1, SignExt, Depth+1) ||
3779 isSignOrZeroExtended(*MISrc2, SignExt, Depth+1);