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))
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)
1433 .addReg(Pred[1].getReg());
1434 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
1435 MI.setDesc(get(PPC::BCLRn));
1436 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1437 .addReg(Pred[1].getReg());
1439 MI.setDesc(get(PPC::BCCLR));
1440 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1441 .addImm(Pred[0].getImm())
1442 .addReg(Pred[1].getReg());
1446 } else if (OpC == PPC::B) {
1447 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) {
1448 bool isPPC64 = Subtarget.isPPC64();
1449 MI.setDesc(get(Pred[0].getImm() ? (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ)
1450 : (isPPC64 ? PPC::BDZ8 : PPC::BDZ)));
1451 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
1452 MachineBasicBlock *MBB = MI.getOperand(0).getMBB();
1453 MI.RemoveOperand(0);
1455 MI.setDesc(get(PPC::BC));
1456 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1457 .addReg(Pred[1].getReg())
1459 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
1460 MachineBasicBlock *MBB = MI.getOperand(0).getMBB();
1461 MI.RemoveOperand(0);
1463 MI.setDesc(get(PPC::BCn));
1464 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1465 .addReg(Pred[1].getReg())
1468 MachineBasicBlock *MBB = MI.getOperand(0).getMBB();
1469 MI.RemoveOperand(0);
1471 MI.setDesc(get(PPC::BCC));
1472 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1473 .addImm(Pred[0].getImm())
1474 .addReg(Pred[1].getReg())
1479 } else if (OpC == PPC::BCTR || OpC == PPC::BCTR8 ||
1480 OpC == PPC::BCTRL || OpC == PPC::BCTRL8) {
1481 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR)
1482 llvm_unreachable("Cannot predicate bctr[l] on the ctr register");
1484 bool setLR = OpC == PPC::BCTRL || OpC == PPC::BCTRL8;
1485 bool isPPC64 = Subtarget.isPPC64();
1487 if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
1488 MI.setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8 : PPC::BCCTR8)
1489 : (setLR ? PPC::BCCTRL : PPC::BCCTR)));
1490 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1491 .addReg(Pred[1].getReg());
1493 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
1494 MI.setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8n : PPC::BCCTR8n)
1495 : (setLR ? PPC::BCCTRLn : PPC::BCCTRn)));
1496 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1497 .addReg(Pred[1].getReg());
1501 MI.setDesc(get(isPPC64 ? (setLR ? PPC::BCCCTRL8 : PPC::BCCCTR8)
1502 : (setLR ? PPC::BCCCTRL : PPC::BCCCTR)));
1503 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1504 .addImm(Pred[0].getImm())
1505 .addReg(Pred[1].getReg());
1512 bool PPCInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
1513 ArrayRef<MachineOperand> Pred2) const {
1514 assert(Pred1.size() == 2 && "Invalid PPC first predicate");
1515 assert(Pred2.size() == 2 && "Invalid PPC second predicate");
1517 if (Pred1[1].getReg() == PPC::CTR8 || Pred1[1].getReg() == PPC::CTR)
1519 if (Pred2[1].getReg() == PPC::CTR8 || Pred2[1].getReg() == PPC::CTR)
1522 // P1 can only subsume P2 if they test the same condition register.
1523 if (Pred1[1].getReg() != Pred2[1].getReg())
1526 PPC::Predicate P1 = (PPC::Predicate) Pred1[0].getImm();
1527 PPC::Predicate P2 = (PPC::Predicate) Pred2[0].getImm();
1532 // Does P1 subsume P2, e.g. GE subsumes GT.
1533 if (P1 == PPC::PRED_LE &&
1534 (P2 == PPC::PRED_LT || P2 == PPC::PRED_EQ))
1536 if (P1 == PPC::PRED_GE &&
1537 (P2 == PPC::PRED_GT || P2 == PPC::PRED_EQ))
1543 bool PPCInstrInfo::DefinesPredicate(MachineInstr &MI,
1544 std::vector<MachineOperand> &Pred) const {
1545 // Note: At the present time, the contents of Pred from this function is
1546 // unused by IfConversion. This implementation follows ARM by pushing the
1547 // CR-defining operand. Because the 'DZ' and 'DNZ' count as types of
1548 // predicate, instructions defining CTR or CTR8 are also included as
1549 // predicate-defining instructions.
1551 const TargetRegisterClass *RCs[] =
1552 { &PPC::CRRCRegClass, &PPC::CRBITRCRegClass,
1553 &PPC::CTRRCRegClass, &PPC::CTRRC8RegClass };
1556 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1557 const MachineOperand &MO = MI.getOperand(i);
1558 for (unsigned c = 0; c < array_lengthof(RCs) && !Found; ++c) {
1559 const TargetRegisterClass *RC = RCs[c];
1561 if (MO.isDef() && RC->contains(MO.getReg())) {
1565 } else if (MO.isRegMask()) {
1566 for (TargetRegisterClass::iterator I = RC->begin(),
1567 IE = RC->end(); I != IE; ++I)
1568 if (MO.clobbersPhysReg(*I)) {
1579 bool PPCInstrInfo::isPredicable(const MachineInstr &MI) const {
1580 unsigned OpC = MI.getOpcode();
1595 bool PPCInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
1596 unsigned &SrcReg2, int &Mask,
1598 unsigned Opc = MI.getOpcode();
1601 default: return false;
1606 SrcReg = MI.getOperand(1).getReg();
1608 Value = MI.getOperand(2).getImm();
1617 SrcReg = MI.getOperand(1).getReg();
1618 SrcReg2 = MI.getOperand(2).getReg();
1625 bool PPCInstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
1626 unsigned SrcReg2, int Mask, int Value,
1627 const MachineRegisterInfo *MRI) const {
1631 int OpC = CmpInstr.getOpcode();
1632 unsigned CRReg = CmpInstr.getOperand(0).getReg();
1634 // FP record forms set CR1 based on the exception status bits, not a
1635 // comparison with zero.
1636 if (OpC == PPC::FCMPUS || OpC == PPC::FCMPUD)
1639 // The record forms set the condition register based on a signed comparison
1640 // with zero (so says the ISA manual). This is not as straightforward as it
1641 // seems, however, because this is always a 64-bit comparison on PPC64, even
1642 // for instructions that are 32-bit in nature (like slw for example).
1643 // So, on PPC32, for unsigned comparisons, we can use the record forms only
1644 // for equality checks (as those don't depend on the sign). On PPC64,
1645 // we are restricted to equality for unsigned 64-bit comparisons and for
1646 // signed 32-bit comparisons the applicability is more restricted.
1647 bool isPPC64 = Subtarget.isPPC64();
1648 bool is32BitSignedCompare = OpC == PPC::CMPWI || OpC == PPC::CMPW;
1649 bool is32BitUnsignedCompare = OpC == PPC::CMPLWI || OpC == PPC::CMPLW;
1650 bool is64BitUnsignedCompare = OpC == PPC::CMPLDI || OpC == PPC::CMPLD;
1652 // Get the unique definition of SrcReg.
1653 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
1654 if (!MI) return false;
1656 bool equalityOnly = false;
1659 if (is32BitSignedCompare) {
1660 // We can perform this optimization only if MI is sign-extending.
1661 if (isSignExtended(*MI))
1665 } else if (is32BitUnsignedCompare) {
1666 // We can perform this optimization, equality only, if MI is
1668 if (isZeroExtended(*MI)) {
1670 equalityOnly = true;
1674 equalityOnly = is64BitUnsignedCompare;
1676 equalityOnly = is32BitUnsignedCompare;
1679 // We need to check the uses of the condition register in order to reject
1680 // non-equality comparisons.
1681 for (MachineRegisterInfo::use_instr_iterator
1682 I = MRI->use_instr_begin(CRReg), IE = MRI->use_instr_end();
1684 MachineInstr *UseMI = &*I;
1685 if (UseMI->getOpcode() == PPC::BCC) {
1686 PPC::Predicate Pred = (PPC::Predicate)UseMI->getOperand(0).getImm();
1687 unsigned PredCond = PPC::getPredicateCondition(Pred);
1688 // We ignore hint bits when checking for non-equality comparisons.
1689 if (PredCond != PPC::PRED_EQ && PredCond != PPC::PRED_NE)
1691 } else if (UseMI->getOpcode() == PPC::ISEL ||
1692 UseMI->getOpcode() == PPC::ISEL8) {
1693 unsigned SubIdx = UseMI->getOperand(3).getSubReg();
1694 if (SubIdx != PPC::sub_eq)
1701 MachineBasicBlock::iterator I = CmpInstr;
1703 // Scan forward to find the first use of the compare.
1704 for (MachineBasicBlock::iterator EL = CmpInstr.getParent()->end(); I != EL;
1706 bool FoundUse = false;
1707 for (MachineRegisterInfo::use_instr_iterator
1708 J = MRI->use_instr_begin(CRReg), JE = MRI->use_instr_end();
1719 SmallVector<std::pair<MachineOperand*, PPC::Predicate>, 4> PredsToUpdate;
1720 SmallVector<std::pair<MachineOperand*, unsigned>, 4> SubRegsToUpdate;
1722 // There are two possible candidates which can be changed to set CR[01].
1723 // One is MI, the other is a SUB instruction.
1724 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
1725 MachineInstr *Sub = nullptr;
1727 // MI is not a candidate for CMPrr.
1729 // FIXME: Conservatively refuse to convert an instruction which isn't in the
1730 // same BB as the comparison. This is to allow the check below to avoid calls
1731 // (and other explicit clobbers); instead we should really check for these
1732 // more explicitly (in at least a few predecessors).
1733 else if (MI->getParent() != CmpInstr.getParent())
1735 else if (Value != 0) {
1736 // The record-form instructions set CR bit based on signed comparison
1737 // against 0. We try to convert a compare against 1 or -1 into a compare
1738 // against 0 to exploit record-form instructions. For example, we change
1739 // the condition "greater than -1" into "greater than or equal to 0"
1740 // and "less than 1" into "less than or equal to 0".
1742 // Since we optimize comparison based on a specific branch condition,
1743 // we don't optimize if condition code is used by more than once.
1744 if (equalityOnly || !MRI->hasOneUse(CRReg))
1747 MachineInstr *UseMI = &*MRI->use_instr_begin(CRReg);
1748 if (UseMI->getOpcode() != PPC::BCC)
1751 PPC::Predicate Pred = (PPC::Predicate)UseMI->getOperand(0).getImm();
1752 PPC::Predicate NewPred = Pred;
1753 unsigned PredCond = PPC::getPredicateCondition(Pred);
1754 unsigned PredHint = PPC::getPredicateHint(Pred);
1755 int16_t Immed = (int16_t)Value;
1757 // When modifying the condition in the predicate, we propagate hint bits
1758 // from the original predicate to the new one.
1759 if (Immed == -1 && PredCond == PPC::PRED_GT)
1760 // We convert "greater than -1" into "greater than or equal to 0",
1761 // since we are assuming signed comparison by !equalityOnly
1762 NewPred = PPC::getPredicate(PPC::PRED_GE, PredHint);
1763 else if (Immed == -1 && PredCond == PPC::PRED_LE)
1764 // We convert "less than or equal to -1" into "less than 0".
1765 NewPred = PPC::getPredicate(PPC::PRED_LT, PredHint);
1766 else if (Immed == 1 && PredCond == PPC::PRED_LT)
1767 // We convert "less than 1" into "less than or equal to 0".
1768 NewPred = PPC::getPredicate(PPC::PRED_LE, PredHint);
1769 else if (Immed == 1 && PredCond == PPC::PRED_GE)
1770 // We convert "greater than or equal to 1" into "greater than 0".
1771 NewPred = PPC::getPredicate(PPC::PRED_GT, PredHint);
1775 PredsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(0)),
1780 const TargetRegisterInfo *TRI = &getRegisterInfo();
1783 // Get ready to iterate backward from CmpInstr.
1784 MachineBasicBlock::iterator E = MI, B = CmpInstr.getParent()->begin();
1786 for (; I != E && !noSub; --I) {
1787 const MachineInstr &Instr = *I;
1788 unsigned IOpC = Instr.getOpcode();
1790 if (&*I != &CmpInstr && (Instr.modifiesRegister(PPC::CR0, TRI) ||
1791 Instr.readsRegister(PPC::CR0, TRI)))
1792 // This instruction modifies or uses the record condition register after
1793 // the one we want to change. While we could do this transformation, it
1794 // would likely not be profitable. This transformation removes one
1795 // instruction, and so even forcing RA to generate one move probably
1796 // makes it unprofitable.
1799 // Check whether CmpInstr can be made redundant by the current instruction.
1800 if ((OpC == PPC::CMPW || OpC == PPC::CMPLW ||
1801 OpC == PPC::CMPD || OpC == PPC::CMPLD) &&
1802 (IOpC == PPC::SUBF || IOpC == PPC::SUBF8) &&
1803 ((Instr.getOperand(1).getReg() == SrcReg &&
1804 Instr.getOperand(2).getReg() == SrcReg2) ||
1805 (Instr.getOperand(1).getReg() == SrcReg2 &&
1806 Instr.getOperand(2).getReg() == SrcReg))) {
1812 // The 'and' is below the comparison instruction.
1816 // Return false if no candidates exist.
1820 // The single candidate is called MI.
1824 int MIOpC = MI->getOpcode();
1825 if (MIOpC == PPC::ANDIo || MIOpC == PPC::ANDIo8)
1828 NewOpC = PPC::getRecordFormOpcode(MIOpC);
1829 if (NewOpC == -1 && PPC::getNonRecordFormOpcode(MIOpC) != -1)
1833 // FIXME: On the non-embedded POWER architectures, only some of the record
1834 // forms are fast, and we should use only the fast ones.
1836 // The defining instruction has a record form (or is already a record
1837 // form). It is possible, however, that we'll need to reverse the condition
1838 // code of the users.
1842 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based on CMP
1843 // needs to be updated to be based on SUB. Push the condition code
1844 // operands to OperandsToUpdate. If it is safe to remove CmpInstr, the
1845 // condition code of these operands will be modified.
1846 // Here, Value == 0 means we haven't converted comparison against 1 or -1 to
1847 // comparison against 0, which may modify predicate.
1848 bool ShouldSwap = false;
1849 if (Sub && Value == 0) {
1850 ShouldSwap = SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
1851 Sub->getOperand(2).getReg() == SrcReg;
1853 // The operands to subf are the opposite of sub, so only in the fixed-point
1854 // case, invert the order.
1855 ShouldSwap = !ShouldSwap;
1859 for (MachineRegisterInfo::use_instr_iterator
1860 I = MRI->use_instr_begin(CRReg), IE = MRI->use_instr_end();
1862 MachineInstr *UseMI = &*I;
1863 if (UseMI->getOpcode() == PPC::BCC) {
1864 PPC::Predicate Pred = (PPC::Predicate) UseMI->getOperand(0).getImm();
1865 unsigned PredCond = PPC::getPredicateCondition(Pred);
1866 assert((!equalityOnly ||
1867 PredCond == PPC::PRED_EQ || PredCond == PPC::PRED_NE) &&
1868 "Invalid predicate for equality-only optimization");
1869 (void)PredCond; // To suppress warning in release build.
1870 PredsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(0)),
1871 PPC::getSwappedPredicate(Pred)));
1872 } else if (UseMI->getOpcode() == PPC::ISEL ||
1873 UseMI->getOpcode() == PPC::ISEL8) {
1874 unsigned NewSubReg = UseMI->getOperand(3).getSubReg();
1875 assert((!equalityOnly || NewSubReg == PPC::sub_eq) &&
1876 "Invalid CR bit for equality-only optimization");
1878 if (NewSubReg == PPC::sub_lt)
1879 NewSubReg = PPC::sub_gt;
1880 else if (NewSubReg == PPC::sub_gt)
1881 NewSubReg = PPC::sub_lt;
1883 SubRegsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(3)),
1885 } else // We need to abort on a user we don't understand.
1888 assert(!(Value != 0 && ShouldSwap) &&
1889 "Non-zero immediate support and ShouldSwap"
1890 "may conflict in updating predicate");
1892 // Create a new virtual register to hold the value of the CR set by the
1893 // record-form instruction. If the instruction was not previously in
1894 // record form, then set the kill flag on the CR.
1895 CmpInstr.eraseFromParent();
1897 MachineBasicBlock::iterator MII = MI;
1898 BuildMI(*MI->getParent(), std::next(MII), MI->getDebugLoc(),
1899 get(TargetOpcode::COPY), CRReg)
1900 .addReg(PPC::CR0, MIOpC != NewOpC ? RegState::Kill : 0);
1902 // Even if CR0 register were dead before, it is alive now since the
1903 // instruction we just built uses it.
1904 MI->clearRegisterDeads(PPC::CR0);
1906 if (MIOpC != NewOpC) {
1907 // We need to be careful here: we're replacing one instruction with
1908 // another, and we need to make sure that we get all of the right
1909 // implicit uses and defs. On the other hand, the caller may be holding
1910 // an iterator to this instruction, and so we can't delete it (this is
1911 // specifically the case if this is the instruction directly after the
1914 // Rotates are expensive instructions. If we're emitting a record-form
1915 // rotate that can just be an andi, we should just emit the andi.
1916 if ((MIOpC == PPC::RLWINM || MIOpC == PPC::RLWINM8) &&
1917 MI->getOperand(2).getImm() == 0) {
1918 int64_t MB = MI->getOperand(3).getImm();
1919 int64_t ME = MI->getOperand(4).getImm();
1920 if (MB < ME && MB >= 16) {
1921 uint64_t Mask = ((1LLU << (32 - MB)) - 1) & ~((1LLU << (31 - ME)) - 1);
1922 NewOpC = MIOpC == PPC::RLWINM ? PPC::ANDIo : PPC::ANDIo8;
1923 MI->RemoveOperand(4);
1924 MI->RemoveOperand(3);
1925 MI->getOperand(2).setImm(Mask);
1926 NumRcRotatesConvertedToRcAnd++;
1928 } else if (MIOpC == PPC::RLDICL && MI->getOperand(2).getImm() == 0) {
1929 int64_t MB = MI->getOperand(3).getImm();
1931 uint64_t Mask = (1LLU << (63 - MB + 1)) - 1;
1932 NewOpC = PPC::ANDIo8;
1933 MI->RemoveOperand(3);
1934 MI->getOperand(2).setImm(Mask);
1935 NumRcRotatesConvertedToRcAnd++;
1939 const MCInstrDesc &NewDesc = get(NewOpC);
1940 MI->setDesc(NewDesc);
1942 if (NewDesc.ImplicitDefs)
1943 for (const MCPhysReg *ImpDefs = NewDesc.getImplicitDefs();
1944 *ImpDefs; ++ImpDefs)
1945 if (!MI->definesRegister(*ImpDefs))
1946 MI->addOperand(*MI->getParent()->getParent(),
1947 MachineOperand::CreateReg(*ImpDefs, true, true));
1948 if (NewDesc.ImplicitUses)
1949 for (const MCPhysReg *ImpUses = NewDesc.getImplicitUses();
1950 *ImpUses; ++ImpUses)
1951 if (!MI->readsRegister(*ImpUses))
1952 MI->addOperand(*MI->getParent()->getParent(),
1953 MachineOperand::CreateReg(*ImpUses, false, true));
1955 assert(MI->definesRegister(PPC::CR0) &&
1956 "Record-form instruction does not define cr0?");
1958 // Modify the condition code of operands in OperandsToUpdate.
1959 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to
1960 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
1961 for (unsigned i = 0, e = PredsToUpdate.size(); i < e; i++)
1962 PredsToUpdate[i].first->setImm(PredsToUpdate[i].second);
1964 for (unsigned i = 0, e = SubRegsToUpdate.size(); i < e; i++)
1965 SubRegsToUpdate[i].first->setSubReg(SubRegsToUpdate[i].second);
1970 /// GetInstSize - Return the number of bytes of code the specified
1971 /// instruction may be. This returns the maximum number of bytes.
1973 unsigned PPCInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
1974 unsigned Opcode = MI.getOpcode();
1976 if (Opcode == PPC::INLINEASM) {
1977 const MachineFunction *MF = MI.getParent()->getParent();
1978 const char *AsmStr = MI.getOperand(0).getSymbolName();
1979 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
1980 } else if (Opcode == TargetOpcode::STACKMAP) {
1981 StackMapOpers Opers(&MI);
1982 return Opers.getNumPatchBytes();
1983 } else if (Opcode == TargetOpcode::PATCHPOINT) {
1984 PatchPointOpers Opers(&MI);
1985 return Opers.getNumPatchBytes();
1987 return get(Opcode).getSize();
1991 std::pair<unsigned, unsigned>
1992 PPCInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
1993 const unsigned Mask = PPCII::MO_ACCESS_MASK;
1994 return std::make_pair(TF & Mask, TF & ~Mask);
1997 ArrayRef<std::pair<unsigned, const char *>>
1998 PPCInstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
1999 using namespace PPCII;
2000 static const std::pair<unsigned, const char *> TargetFlags[] = {
2003 {MO_TPREL_LO, "ppc-tprel-lo"},
2004 {MO_TPREL_HA, "ppc-tprel-ha"},
2005 {MO_DTPREL_LO, "ppc-dtprel-lo"},
2006 {MO_TLSLD_LO, "ppc-tlsld-lo"},
2007 {MO_TOC_LO, "ppc-toc-lo"},
2008 {MO_TLS, "ppc-tls"}};
2009 return makeArrayRef(TargetFlags);
2012 ArrayRef<std::pair<unsigned, const char *>>
2013 PPCInstrInfo::getSerializableBitmaskMachineOperandTargetFlags() const {
2014 using namespace PPCII;
2015 static const std::pair<unsigned, const char *> TargetFlags[] = {
2016 {MO_PLT, "ppc-plt"},
2017 {MO_PIC_FLAG, "ppc-pic"},
2018 {MO_NLP_FLAG, "ppc-nlp"},
2019 {MO_NLP_HIDDEN_FLAG, "ppc-nlp-hidden"}};
2020 return makeArrayRef(TargetFlags);
2023 // Expand VSX Memory Pseudo instruction to either a VSX or a FP instruction.
2024 // The VSX versions have the advantage of a full 64-register target whereas
2025 // the FP ones have the advantage of lower latency and higher throughput. So
2026 // what we are after is using the faster instructions in low register pressure
2027 // situations and using the larger register file in high register pressure
2029 bool PPCInstrInfo::expandVSXMemPseudo(MachineInstr &MI) const {
2030 unsigned UpperOpcode, LowerOpcode;
2031 switch (MI.getOpcode()) {
2032 case PPC::DFLOADf32:
2033 UpperOpcode = PPC::LXSSP;
2034 LowerOpcode = PPC::LFS;
2036 case PPC::DFLOADf64:
2037 UpperOpcode = PPC::LXSD;
2038 LowerOpcode = PPC::LFD;
2040 case PPC::DFSTOREf32:
2041 UpperOpcode = PPC::STXSSP;
2042 LowerOpcode = PPC::STFS;
2044 case PPC::DFSTOREf64:
2045 UpperOpcode = PPC::STXSD;
2046 LowerOpcode = PPC::STFD;
2048 case PPC::XFLOADf32:
2049 UpperOpcode = PPC::LXSSPX;
2050 LowerOpcode = PPC::LFSX;
2052 case PPC::XFLOADf64:
2053 UpperOpcode = PPC::LXSDX;
2054 LowerOpcode = PPC::LFDX;
2056 case PPC::XFSTOREf32:
2057 UpperOpcode = PPC::STXSSPX;
2058 LowerOpcode = PPC::STFSX;
2060 case PPC::XFSTOREf64:
2061 UpperOpcode = PPC::STXSDX;
2062 LowerOpcode = PPC::STFDX;
2065 UpperOpcode = PPC::LXSIWAX;
2066 LowerOpcode = PPC::LFIWAX;
2069 UpperOpcode = PPC::LXSIWZX;
2070 LowerOpcode = PPC::LFIWZX;
2073 UpperOpcode = PPC::STXSIWX;
2074 LowerOpcode = PPC::STFIWX;
2077 llvm_unreachable("Unknown Operation!");
2080 unsigned TargetReg = MI.getOperand(0).getReg();
2082 if ((TargetReg >= PPC::F0 && TargetReg <= PPC::F31) ||
2083 (TargetReg >= PPC::VSL0 && TargetReg <= PPC::VSL31))
2084 Opcode = LowerOpcode;
2086 Opcode = UpperOpcode;
2087 MI.setDesc(get(Opcode));
2092 static bool isAnImmediateOperand(const MachineOperand &MO) {
2093 return MO.isCPI() || MO.isGlobal() || MO.isImm();
2097 bool PPCInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
2098 auto &MBB = *MI.getParent();
2099 auto DL = MI.getDebugLoc();
2101 switch (MI.getOpcode()) {
2102 case TargetOpcode::LOAD_STACK_GUARD: {
2103 assert(Subtarget.isTargetLinux() &&
2104 "Only Linux target is expected to contain LOAD_STACK_GUARD");
2105 const int64_t Offset = Subtarget.isPPC64() ? -0x7010 : -0x7008;
2106 const unsigned Reg = Subtarget.isPPC64() ? PPC::X13 : PPC::R2;
2107 MI.setDesc(get(Subtarget.isPPC64() ? PPC::LD : PPC::LWZ));
2108 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
2113 case PPC::DFLOADf32:
2114 case PPC::DFLOADf64:
2115 case PPC::DFSTOREf32:
2116 case PPC::DFSTOREf64: {
2117 assert(Subtarget.hasP9Vector() &&
2118 "Invalid D-Form Pseudo-ops on Pre-P9 target.");
2119 assert(MI.getOperand(2).isReg() &&
2120 isAnImmediateOperand(MI.getOperand(1)) &&
2121 "D-form op must have register and immediate operands");
2122 return expandVSXMemPseudo(MI);
2124 case PPC::XFLOADf32:
2125 case PPC::XFSTOREf32:
2129 assert(Subtarget.hasP8Vector() &&
2130 "Invalid X-Form Pseudo-ops on Pre-P8 target.");
2131 assert(MI.getOperand(2).isReg() && MI.getOperand(1).isReg() &&
2132 "X-form op must have register and register operands");
2133 return expandVSXMemPseudo(MI);
2135 case PPC::XFLOADf64:
2136 case PPC::XFSTOREf64: {
2137 assert(Subtarget.hasVSX() &&
2138 "Invalid X-Form Pseudo-ops on target that has no VSX.");
2139 assert(MI.getOperand(2).isReg() && MI.getOperand(1).isReg() &&
2140 "X-form op must have register and register operands");
2141 return expandVSXMemPseudo(MI);
2143 case PPC::SPILLTOVSR_LD: {
2144 unsigned TargetReg = MI.getOperand(0).getReg();
2145 if (PPC::VSFRCRegClass.contains(TargetReg)) {
2146 MI.setDesc(get(PPC::DFLOADf64));
2147 return expandPostRAPseudo(MI);
2150 MI.setDesc(get(PPC::LD));
2153 case PPC::SPILLTOVSR_ST: {
2154 unsigned SrcReg = MI.getOperand(0).getReg();
2155 if (PPC::VSFRCRegClass.contains(SrcReg)) {
2156 NumStoreSPILLVSRRCAsVec++;
2157 MI.setDesc(get(PPC::DFSTOREf64));
2158 return expandPostRAPseudo(MI);
2160 NumStoreSPILLVSRRCAsGpr++;
2161 MI.setDesc(get(PPC::STD));
2165 case PPC::SPILLTOVSR_LDX: {
2166 unsigned TargetReg = MI.getOperand(0).getReg();
2167 if (PPC::VSFRCRegClass.contains(TargetReg))
2168 MI.setDesc(get(PPC::LXSDX));
2170 MI.setDesc(get(PPC::LDX));
2173 case PPC::SPILLTOVSR_STX: {
2174 unsigned SrcReg = MI.getOperand(0).getReg();
2175 if (PPC::VSFRCRegClass.contains(SrcReg)) {
2176 NumStoreSPILLVSRRCAsVec++;
2177 MI.setDesc(get(PPC::STXSDX));
2179 NumStoreSPILLVSRRCAsGpr++;
2180 MI.setDesc(get(PPC::STDX));
2185 case PPC::CFENCE8: {
2186 auto Val = MI.getOperand(0).getReg();
2187 BuildMI(MBB, MI, DL, get(PPC::CMPD), PPC::CR7).addReg(Val).addReg(Val);
2188 BuildMI(MBB, MI, DL, get(PPC::CTRL_DEP))
2189 .addImm(PPC::PRED_NE_MINUS)
2192 MI.setDesc(get(PPC::ISYNC));
2193 MI.RemoveOperand(0);
2200 // Essentially a compile-time implementation of a compare->isel sequence.
2201 // It takes two constants to compare, along with the true/false registers
2202 // and the comparison type (as a subreg to a CR field) and returns one
2203 // of the true/false registers, depending on the comparison results.
2204 static unsigned selectReg(int64_t Imm1, int64_t Imm2, unsigned CompareOpc,
2205 unsigned TrueReg, unsigned FalseReg,
2206 unsigned CRSubReg) {
2207 // Signed comparisons. The immediates are assumed to be sign-extended.
2208 if (CompareOpc == PPC::CMPWI || CompareOpc == PPC::CMPDI) {
2210 default: llvm_unreachable("Unknown integer comparison type.");
2212 return Imm1 < Imm2 ? TrueReg : FalseReg;
2214 return Imm1 > Imm2 ? TrueReg : FalseReg;
2216 return Imm1 == Imm2 ? TrueReg : FalseReg;
2219 // Unsigned comparisons.
2220 else if (CompareOpc == PPC::CMPLWI || CompareOpc == PPC::CMPLDI) {
2222 default: llvm_unreachable("Unknown integer comparison type.");
2224 return (uint64_t)Imm1 < (uint64_t)Imm2 ? TrueReg : FalseReg;
2226 return (uint64_t)Imm1 > (uint64_t)Imm2 ? TrueReg : FalseReg;
2228 return Imm1 == Imm2 ? TrueReg : FalseReg;
2231 return PPC::NoRegister;
2234 // Replace an instruction with one that materializes a constant (and sets
2235 // CR0 if the original instruction was a record-form instruction).
2236 void PPCInstrInfo::replaceInstrWithLI(MachineInstr &MI,
2237 const LoadImmediateInfo &LII) const {
2238 // Remove existing operands.
2239 int OperandToKeep = LII.SetCR ? 1 : 0;
2240 for (int i = MI.getNumOperands() - 1; i > OperandToKeep; i--)
2241 MI.RemoveOperand(i);
2243 // Replace the instruction.
2245 MI.setDesc(get(LII.Is64Bit ? PPC::ANDIo8 : PPC::ANDIo));
2246 // Set the immediate.
2247 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
2248 .addImm(LII.Imm).addReg(PPC::CR0, RegState::ImplicitDefine);
2252 MI.setDesc(get(LII.Is64Bit ? PPC::LI8 : PPC::LI));
2254 // Set the immediate.
2255 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
2259 MachineInstr *PPCInstrInfo::getConstantDefMI(MachineInstr &MI,
2261 bool &SeenIntermediateUse) const {
2263 MachineInstr *DefMI = nullptr;
2264 MachineRegisterInfo *MRI = &MI.getParent()->getParent()->getRegInfo();
2265 const TargetRegisterInfo *TRI = &getRegisterInfo();
2266 // If we're in SSA, get the defs through the MRI. Otherwise, only look
2267 // within the basic block to see if the register is defined using an LI/LI8.
2269 for (int i = 1, e = MI.getNumOperands(); i < e; i++) {
2270 if (!MI.getOperand(i).isReg())
2272 unsigned Reg = MI.getOperand(i).getReg();
2273 if (!TargetRegisterInfo::isVirtualRegister(Reg))
2275 unsigned TrueReg = TRI->lookThruCopyLike(Reg, MRI);
2276 if (TargetRegisterInfo::isVirtualRegister(TrueReg)) {
2277 DefMI = MRI->getVRegDef(TrueReg);
2278 if (DefMI->getOpcode() == PPC::LI || DefMI->getOpcode() == PPC::LI8) {
2285 // Looking back through the definition for each operand could be expensive,
2286 // so exit early if this isn't an instruction that either has an immediate
2287 // form or is already an immediate form that we can handle.
2289 unsigned Opc = MI.getOpcode();
2290 bool ConvertibleImmForm =
2291 Opc == PPC::CMPWI || Opc == PPC::CMPLWI ||
2292 Opc == PPC::CMPDI || Opc == PPC::CMPLDI ||
2293 Opc == PPC::ADDI || Opc == PPC::ADDI8 ||
2294 Opc == PPC::ORI || Opc == PPC::ORI8 ||
2295 Opc == PPC::XORI || Opc == PPC::XORI8 ||
2296 Opc == PPC::RLDICL || Opc == PPC::RLDICLo ||
2297 Opc == PPC::RLDICL_32 || Opc == PPC::RLDICL_32_64 ||
2298 Opc == PPC::RLWINM || Opc == PPC::RLWINMo ||
2299 Opc == PPC::RLWINM8 || Opc == PPC::RLWINM8o;
2300 if (!instrHasImmForm(MI, III) && !ConvertibleImmForm)
2303 // Don't convert or %X, %Y, %Y since that's just a register move.
2304 if ((Opc == PPC::OR || Opc == PPC::OR8) &&
2305 MI.getOperand(1).getReg() == MI.getOperand(2).getReg())
2307 for (int i = 1, e = MI.getNumOperands(); i < e; i++) {
2308 MachineOperand &MO = MI.getOperand(i);
2309 SeenIntermediateUse = false;
2310 if (MO.isReg() && MO.isUse() && !MO.isImplicit()) {
2311 MachineBasicBlock::reverse_iterator E = MI.getParent()->rend(), It = MI;
2313 unsigned Reg = MI.getOperand(i).getReg();
2314 // MachineInstr::readsRegister only returns true if the machine
2315 // instruction reads the exact register or its super-register. It
2316 // does not consider uses of sub-registers which seems like strange
2317 // behaviour. Nonetheless, if we end up with a 64-bit register here,
2318 // get the corresponding 32-bit register to check.
2319 if (PPC::G8RCRegClass.contains(Reg))
2320 Reg = Reg - PPC::X0 + PPC::R0;
2322 // Is this register defined by a load-immediate in this block?
2323 for ( ; It != E; ++It) {
2324 if (It->modifiesRegister(Reg, &getRegisterInfo())) {
2325 if (It->getOpcode() == PPC::LI || It->getOpcode() == PPC::LI8) {
2330 } else if (It->readsRegister(Reg, &getRegisterInfo()))
2331 // If we see another use of this reg between the def and the MI,
2332 // we want to flat it so the def isn't deleted.
2333 SeenIntermediateUse = true;
2338 return ConstOp == ~0U ? nullptr : DefMI;
2341 const unsigned *PPCInstrInfo::getStoreOpcodesForSpillArray() const {
2342 static const unsigned OpcodesForSpill[2][SOK_LastOpcodeSpill] = {
2344 {PPC::STW, PPC::STD, PPC::STFD, PPC::STFS, PPC::SPILL_CR,
2345 PPC::SPILL_CRBIT, PPC::STVX, PPC::STXVD2X, PPC::STXSDX, PPC::STXSSPX,
2346 PPC::SPILL_VRSAVE, PPC::QVSTFDX, PPC::QVSTFSXs, PPC::QVSTFDXb,
2347 PPC::SPILLTOVSR_ST, PPC::EVSTDD, PPC::SPESTW},
2349 {PPC::STW, PPC::STD, PPC::STFD, PPC::STFS, PPC::SPILL_CR,
2350 PPC::SPILL_CRBIT, PPC::STVX, PPC::STXV, PPC::DFSTOREf64, PPC::DFSTOREf32,
2351 PPC::SPILL_VRSAVE, PPC::QVSTFDX, PPC::QVSTFSXs, PPC::QVSTFDXb,
2352 PPC::SPILLTOVSR_ST}};
2354 return OpcodesForSpill[(Subtarget.hasP9Vector()) ? 1 : 0];
2357 const unsigned *PPCInstrInfo::getLoadOpcodesForSpillArray() const {
2358 static const unsigned OpcodesForSpill[2][SOK_LastOpcodeSpill] = {
2360 {PPC::LWZ, PPC::LD, PPC::LFD, PPC::LFS, PPC::RESTORE_CR,
2361 PPC::RESTORE_CRBIT, PPC::LVX, PPC::LXVD2X, PPC::LXSDX, PPC::LXSSPX,
2362 PPC::RESTORE_VRSAVE, PPC::QVLFDX, PPC::QVLFSXs, PPC::QVLFDXb,
2363 PPC::SPILLTOVSR_LD, PPC::EVLDD, PPC::SPELWZ},
2365 {PPC::LWZ, PPC::LD, PPC::LFD, PPC::LFS, PPC::RESTORE_CR,
2366 PPC::RESTORE_CRBIT, PPC::LVX, PPC::LXV, PPC::DFLOADf64, PPC::DFLOADf32,
2367 PPC::RESTORE_VRSAVE, PPC::QVLFDX, PPC::QVLFSXs, PPC::QVLFDXb,
2368 PPC::SPILLTOVSR_LD}};
2370 return OpcodesForSpill[(Subtarget.hasP9Vector()) ? 1 : 0];
2373 // If this instruction has an immediate form and one of its operands is a
2374 // result of a load-immediate, convert it to the immediate form if the constant
2376 bool PPCInstrInfo::convertToImmediateForm(MachineInstr &MI,
2377 MachineInstr **KilledDef) const {
2378 MachineFunction *MF = MI.getParent()->getParent();
2379 MachineRegisterInfo *MRI = &MF->getRegInfo();
2380 bool PostRA = !MRI->isSSA();
2381 bool SeenIntermediateUse = true;
2382 unsigned ConstantOperand = ~0U;
2383 MachineInstr *DefMI = getConstantDefMI(MI, ConstantOperand,
2384 SeenIntermediateUse);
2385 if (!DefMI || !DefMI->getOperand(1).isImm())
2387 assert(ConstantOperand < MI.getNumOperands() &&
2388 "The constant operand needs to be valid at this point");
2390 int64_t Immediate = DefMI->getOperand(1).getImm();
2391 // Sign-extend to 64-bits.
2392 int64_t SExtImm = ((uint64_t)Immediate & ~0x7FFFuLL) != 0 ?
2393 (Immediate | 0xFFFFFFFFFFFF0000) : Immediate;
2395 if (KilledDef && MI.getOperand(ConstantOperand).isKill() &&
2396 !SeenIntermediateUse)
2399 // If this is a reg+reg instruction that has a reg+imm form, convert it now.
2401 if (instrHasImmForm(MI, III))
2402 return transformToImmForm(MI, III, ConstantOperand, SExtImm);
2404 bool ReplaceWithLI = false;
2405 bool Is64BitLI = false;
2408 unsigned Opc = MI.getOpcode();
2410 default: return false;
2412 // FIXME: Any branches conditional on such a comparison can be made
2413 // unconditional. At this time, this happens too infrequently to be worth
2414 // the implementation effort, but if that ever changes, we could convert
2415 // such a pattern here.
2420 // Doing this post-RA would require dataflow analysis to reliably find uses
2421 // of the CR register set by the compare.
2424 // If a compare-immediate is fed by an immediate and is itself an input of
2425 // an ISEL (the most common case) into a COPY of the correct register.
2426 bool Changed = false;
2427 unsigned DefReg = MI.getOperand(0).getReg();
2428 int64_t Comparand = MI.getOperand(2).getImm();
2429 int64_t SExtComparand = ((uint64_t)Comparand & ~0x7FFFuLL) != 0 ?
2430 (Comparand | 0xFFFFFFFFFFFF0000) : Comparand;
2432 for (auto &CompareUseMI : MRI->use_instructions(DefReg)) {
2433 unsigned UseOpc = CompareUseMI.getOpcode();
2434 if (UseOpc != PPC::ISEL && UseOpc != PPC::ISEL8)
2436 unsigned CRSubReg = CompareUseMI.getOperand(3).getSubReg();
2437 unsigned TrueReg = CompareUseMI.getOperand(1).getReg();
2438 unsigned FalseReg = CompareUseMI.getOperand(2).getReg();
2439 unsigned RegToCopy = selectReg(SExtImm, SExtComparand, Opc, TrueReg,
2440 FalseReg, CRSubReg);
2441 if (RegToCopy == PPC::NoRegister)
2443 // Can't use PPC::COPY to copy PPC::ZERO[8]. Convert it to LI[8] 0.
2444 if (RegToCopy == PPC::ZERO || RegToCopy == PPC::ZERO8) {
2445 CompareUseMI.setDesc(get(UseOpc == PPC::ISEL8 ? PPC::LI8 : PPC::LI));
2446 CompareUseMI.getOperand(1).ChangeToImmediate(0);
2447 CompareUseMI.RemoveOperand(3);
2448 CompareUseMI.RemoveOperand(2);
2452 dbgs() << "Found LI -> CMPI -> ISEL, replacing with a copy.\n");
2453 LLVM_DEBUG(DefMI->dump(); MI.dump(); CompareUseMI.dump());
2454 LLVM_DEBUG(dbgs() << "Is converted to:\n");
2455 // Convert to copy and remove unneeded operands.
2456 CompareUseMI.setDesc(get(PPC::COPY));
2457 CompareUseMI.RemoveOperand(3);
2458 CompareUseMI.RemoveOperand(RegToCopy == TrueReg ? 2 : 1);
2459 CmpIselsConverted++;
2461 LLVM_DEBUG(CompareUseMI.dump());
2465 // This may end up incremented multiple times since this function is called
2466 // during a fixed-point transformation, but it is only meant to indicate the
2467 // presence of this opportunity.
2468 MissedConvertibleImmediateInstrs++;
2472 // Immediate forms - may simply be convertable to an LI.
2475 // Does the sum fit in a 16-bit signed field?
2476 int64_t Addend = MI.getOperand(2).getImm();
2477 if (isInt<16>(Addend + SExtImm)) {
2478 ReplaceWithLI = true;
2479 Is64BitLI = Opc == PPC::ADDI8;
2480 NewImm = Addend + SExtImm;
2487 case PPC::RLDICL_32:
2488 case PPC::RLDICL_32_64: {
2489 // Use APInt's rotate function.
2490 int64_t SH = MI.getOperand(2).getImm();
2491 int64_t MB = MI.getOperand(3).getImm();
2492 APInt InVal((Opc == PPC::RLDICL || Opc == PPC::RLDICLo) ?
2493 64 : 32, SExtImm, true);
2494 InVal = InVal.rotl(SH);
2495 uint64_t Mask = (1LLU << (63 - MB + 1)) - 1;
2497 // Can't replace negative values with an LI as that will sign-extend
2498 // and not clear the left bits. If we're setting the CR bit, we will use
2499 // ANDIo which won't sign extend, so that's safe.
2500 if (isUInt<15>(InVal.getSExtValue()) ||
2501 (Opc == PPC::RLDICLo && isUInt<16>(InVal.getSExtValue()))) {
2502 ReplaceWithLI = true;
2503 Is64BitLI = Opc != PPC::RLDICL_32;
2504 NewImm = InVal.getSExtValue();
2505 SetCR = Opc == PPC::RLDICLo;
2513 case PPC::RLWINM8o: {
2514 int64_t SH = MI.getOperand(2).getImm();
2515 int64_t MB = MI.getOperand(3).getImm();
2516 int64_t ME = MI.getOperand(4).getImm();
2517 APInt InVal(32, SExtImm, true);
2518 InVal = InVal.rotl(SH);
2519 // Set the bits ( MB + 32 ) to ( ME + 32 ).
2520 uint64_t Mask = ((1LLU << (32 - MB)) - 1) & ~((1LLU << (31 - ME)) - 1);
2522 // Can't replace negative values with an LI as that will sign-extend
2523 // and not clear the left bits. If we're setting the CR bit, we will use
2524 // ANDIo which won't sign extend, so that's safe.
2525 bool ValueFits = isUInt<15>(InVal.getSExtValue());
2526 ValueFits |= ((Opc == PPC::RLWINMo || Opc == PPC::RLWINM8o) &&
2527 isUInt<16>(InVal.getSExtValue()));
2529 ReplaceWithLI = true;
2530 Is64BitLI = Opc == PPC::RLWINM8 || Opc == PPC::RLWINM8o;
2531 NewImm = InVal.getSExtValue();
2532 SetCR = Opc == PPC::RLWINMo || Opc == PPC::RLWINM8o;
2541 int64_t LogicalImm = MI.getOperand(2).getImm();
2543 if (Opc == PPC::ORI || Opc == PPC::ORI8)
2544 Result = LogicalImm | SExtImm;
2546 Result = LogicalImm ^ SExtImm;
2547 if (isInt<16>(Result)) {
2548 ReplaceWithLI = true;
2549 Is64BitLI = Opc == PPC::ORI8 || Opc == PPC::XORI8;
2557 if (ReplaceWithLI) {
2558 // We need to be careful with CR-setting instructions we're replacing.
2560 // We don't know anything about uses when we're out of SSA, so only
2561 // replace if the new immediate will be reproduced.
2562 bool ImmChanged = (SExtImm & NewImm) != NewImm;
2563 if (PostRA && ImmChanged)
2567 // If the defining load-immediate has no other uses, we can just replace
2568 // the immediate with the new immediate.
2569 if (MRI->hasOneUse(DefMI->getOperand(0).getReg()))
2570 DefMI->getOperand(1).setImm(NewImm);
2572 // If we're not using the GPR result of the CR-setting instruction, we
2573 // just need to and with zero/non-zero depending on the new immediate.
2574 else if (MRI->use_empty(MI.getOperand(0).getReg())) {
2576 assert(Immediate && "Transformation converted zero to non-zero?");
2580 else if (ImmChanged)
2585 LLVM_DEBUG(dbgs() << "Replacing instruction:\n");
2586 LLVM_DEBUG(MI.dump());
2587 LLVM_DEBUG(dbgs() << "Fed by:\n");
2588 LLVM_DEBUG(DefMI->dump());
2589 LoadImmediateInfo LII;
2591 LII.Is64Bit = Is64BitLI;
2593 // If we're setting the CR, the original load-immediate must be kept (as an
2594 // operand to ANDIo/ANDI8o).
2595 if (KilledDef && SetCR)
2596 *KilledDef = nullptr;
2597 replaceInstrWithLI(MI, LII);
2598 LLVM_DEBUG(dbgs() << "With:\n");
2599 LLVM_DEBUG(MI.dump());
2605 bool PPCInstrInfo::instrHasImmForm(const MachineInstr &MI,
2606 ImmInstrInfo &III) const {
2607 unsigned Opc = MI.getOpcode();
2608 // The vast majority of the instructions would need their operand 2 replaced
2609 // with an immediate when switching to the reg+imm form. A marked exception
2610 // are the update form loads/stores for which a constant operand 2 would need
2611 // to turn into a displacement and move operand 1 to the operand 2 position.
2613 III.ConstantOpNo = 2;
2615 III.ImmMustBeMultipleOf = 1;
2616 III.TruncateImmTo = 0;
2618 default: return false;
2621 III.SignedImm = true;
2622 III.ZeroIsSpecialOrig = 0;
2623 III.ZeroIsSpecialNew = 1;
2624 III.IsCommutative = true;
2625 III.ImmOpcode = Opc == PPC::ADD4 ? PPC::ADDI : PPC::ADDI8;
2629 III.SignedImm = true;
2630 III.ZeroIsSpecialOrig = 0;
2631 III.ZeroIsSpecialNew = 0;
2632 III.IsCommutative = true;
2633 III.ImmOpcode = Opc == PPC::ADDC ? PPC::ADDIC : PPC::ADDIC8;
2636 III.SignedImm = true;
2637 III.ZeroIsSpecialOrig = 0;
2638 III.ZeroIsSpecialNew = 0;
2639 III.IsCommutative = true;
2640 III.ImmOpcode = PPC::ADDICo;
2644 III.SignedImm = true;
2645 III.ZeroIsSpecialOrig = 0;
2646 III.ZeroIsSpecialNew = 0;
2647 III.IsCommutative = false;
2648 III.ImmOpcode = Opc == PPC::SUBFC ? PPC::SUBFIC : PPC::SUBFIC8;
2652 III.SignedImm = true;
2653 III.ZeroIsSpecialOrig = 0;
2654 III.ZeroIsSpecialNew = 0;
2655 III.IsCommutative = false;
2656 III.ImmOpcode = Opc == PPC::CMPW ? PPC::CMPWI : PPC::CMPDI;
2660 III.SignedImm = false;
2661 III.ZeroIsSpecialOrig = 0;
2662 III.ZeroIsSpecialNew = 0;
2663 III.IsCommutative = false;
2664 III.ImmOpcode = Opc == PPC::CMPLW ? PPC::CMPLWI : PPC::CMPLDI;
2672 III.SignedImm = false;
2673 III.ZeroIsSpecialOrig = 0;
2674 III.ZeroIsSpecialNew = 0;
2675 III.IsCommutative = true;
2677 default: llvm_unreachable("Unknown opcode");
2678 case PPC::ANDo: III.ImmOpcode = PPC::ANDIo; break;
2679 case PPC::AND8o: III.ImmOpcode = PPC::ANDIo8; break;
2680 case PPC::OR: III.ImmOpcode = PPC::ORI; break;
2681 case PPC::OR8: III.ImmOpcode = PPC::ORI8; break;
2682 case PPC::XOR: III.ImmOpcode = PPC::XORI; break;
2683 case PPC::XOR8: III.ImmOpcode = PPC::XORI8; break;
2700 III.SignedImm = false;
2701 III.ZeroIsSpecialOrig = 0;
2702 III.ZeroIsSpecialNew = 0;
2703 III.IsCommutative = false;
2704 // This isn't actually true, but the instructions ignore any of the
2705 // upper bits, so any immediate loaded with an LI is acceptable.
2706 // This does not apply to shift right algebraic because a value
2707 // out of range will produce a -1/0.
2709 if (Opc == PPC::RLWNM || Opc == PPC::RLWNM8 ||
2710 Opc == PPC::RLWNMo || Opc == PPC::RLWNM8o)
2711 III.TruncateImmTo = 5;
2713 III.TruncateImmTo = 6;
2715 default: llvm_unreachable("Unknown opcode");
2716 case PPC::RLWNM: III.ImmOpcode = PPC::RLWINM; break;
2717 case PPC::RLWNM8: III.ImmOpcode = PPC::RLWINM8; break;
2718 case PPC::RLWNMo: III.ImmOpcode = PPC::RLWINMo; break;
2719 case PPC::RLWNM8o: III.ImmOpcode = PPC::RLWINM8o; break;
2720 case PPC::SLW: III.ImmOpcode = PPC::RLWINM; break;
2721 case PPC::SLW8: III.ImmOpcode = PPC::RLWINM8; break;
2722 case PPC::SLWo: III.ImmOpcode = PPC::RLWINMo; break;
2723 case PPC::SLW8o: III.ImmOpcode = PPC::RLWINM8o; break;
2724 case PPC::SRW: III.ImmOpcode = PPC::RLWINM; break;
2725 case PPC::SRW8: III.ImmOpcode = PPC::RLWINM8; break;
2726 case PPC::SRWo: III.ImmOpcode = PPC::RLWINMo; break;
2727 case PPC::SRW8o: III.ImmOpcode = PPC::RLWINM8o; break;
2730 III.TruncateImmTo = 0;
2731 III.ImmOpcode = PPC::SRAWI;
2735 III.TruncateImmTo = 0;
2736 III.ImmOpcode = PPC::SRAWIo;
2750 III.SignedImm = false;
2751 III.ZeroIsSpecialOrig = 0;
2752 III.ZeroIsSpecialNew = 0;
2753 III.IsCommutative = false;
2754 // This isn't actually true, but the instructions ignore any of the
2755 // upper bits, so any immediate loaded with an LI is acceptable.
2756 // This does not apply to shift right algebraic because a value
2757 // out of range will produce a -1/0.
2759 if (Opc == PPC::RLDCL || Opc == PPC::RLDCLo ||
2760 Opc == PPC::RLDCR || Opc == PPC::RLDCRo)
2761 III.TruncateImmTo = 6;
2763 III.TruncateImmTo = 7;
2765 default: llvm_unreachable("Unknown opcode");
2766 case PPC::RLDCL: III.ImmOpcode = PPC::RLDICL; break;
2767 case PPC::RLDCLo: III.ImmOpcode = PPC::RLDICLo; break;
2768 case PPC::RLDCR: III.ImmOpcode = PPC::RLDICR; break;
2769 case PPC::RLDCRo: III.ImmOpcode = PPC::RLDICRo; break;
2770 case PPC::SLD: III.ImmOpcode = PPC::RLDICR; break;
2771 case PPC::SLDo: III.ImmOpcode = PPC::RLDICRo; break;
2772 case PPC::SRD: III.ImmOpcode = PPC::RLDICL; break;
2773 case PPC::SRDo: III.ImmOpcode = PPC::RLDICLo; break;
2776 III.TruncateImmTo = 0;
2777 III.ImmOpcode = PPC::SRADI;
2781 III.TruncateImmTo = 0;
2782 III.ImmOpcode = PPC::SRADIo;
2786 // Loads and stores:
2808 III.SignedImm = true;
2809 III.ZeroIsSpecialOrig = 1;
2810 III.ZeroIsSpecialNew = 2;
2811 III.IsCommutative = true;
2813 III.ConstantOpNo = 2;
2815 default: llvm_unreachable("Unknown opcode");
2816 case PPC::LBZX: III.ImmOpcode = PPC::LBZ; break;
2817 case PPC::LBZX8: III.ImmOpcode = PPC::LBZ8; break;
2818 case PPC::LHZX: III.ImmOpcode = PPC::LHZ; break;
2819 case PPC::LHZX8: III.ImmOpcode = PPC::LHZ8; break;
2820 case PPC::LHAX: III.ImmOpcode = PPC::LHA; break;
2821 case PPC::LHAX8: III.ImmOpcode = PPC::LHA8; break;
2822 case PPC::LWZX: III.ImmOpcode = PPC::LWZ; break;
2823 case PPC::LWZX8: III.ImmOpcode = PPC::LWZ8; break;
2825 III.ImmOpcode = PPC::LWA;
2826 III.ImmMustBeMultipleOf = 4;
2828 case PPC::LDX: III.ImmOpcode = PPC::LD; III.ImmMustBeMultipleOf = 4; break;
2829 case PPC::LFSX: III.ImmOpcode = PPC::LFS; break;
2830 case PPC::LFDX: III.ImmOpcode = PPC::LFD; break;
2831 case PPC::STBX: III.ImmOpcode = PPC::STB; break;
2832 case PPC::STBX8: III.ImmOpcode = PPC::STB8; break;
2833 case PPC::STHX: III.ImmOpcode = PPC::STH; break;
2834 case PPC::STHX8: III.ImmOpcode = PPC::STH8; break;
2835 case PPC::STWX: III.ImmOpcode = PPC::STW; break;
2836 case PPC::STWX8: III.ImmOpcode = PPC::STW8; break;
2838 III.ImmOpcode = PPC::STD;
2839 III.ImmMustBeMultipleOf = 4;
2841 case PPC::STFSX: III.ImmOpcode = PPC::STFS; break;
2842 case PPC::STFDX: III.ImmOpcode = PPC::STFD; break;
2865 III.SignedImm = true;
2866 III.ZeroIsSpecialOrig = 2;
2867 III.ZeroIsSpecialNew = 3;
2868 III.IsCommutative = false;
2870 III.ConstantOpNo = 3;
2872 default: llvm_unreachable("Unknown opcode");
2873 case PPC::LBZUX: III.ImmOpcode = PPC::LBZU; break;
2874 case PPC::LBZUX8: III.ImmOpcode = PPC::LBZU8; break;
2875 case PPC::LHZUX: III.ImmOpcode = PPC::LHZU; break;
2876 case PPC::LHZUX8: III.ImmOpcode = PPC::LHZU8; break;
2877 case PPC::LHAUX: III.ImmOpcode = PPC::LHAU; break;
2878 case PPC::LHAUX8: III.ImmOpcode = PPC::LHAU8; break;
2879 case PPC::LWZUX: III.ImmOpcode = PPC::LWZU; break;
2880 case PPC::LWZUX8: III.ImmOpcode = PPC::LWZU8; break;
2882 III.ImmOpcode = PPC::LDU;
2883 III.ImmMustBeMultipleOf = 4;
2885 case PPC::LFSUX: III.ImmOpcode = PPC::LFSU; break;
2886 case PPC::LFDUX: III.ImmOpcode = PPC::LFDU; break;
2887 case PPC::STBUX: III.ImmOpcode = PPC::STBU; break;
2888 case PPC::STBUX8: III.ImmOpcode = PPC::STBU8; break;
2889 case PPC::STHUX: III.ImmOpcode = PPC::STHU; break;
2890 case PPC::STHUX8: III.ImmOpcode = PPC::STHU8; break;
2891 case PPC::STWUX: III.ImmOpcode = PPC::STWU; break;
2892 case PPC::STWUX8: III.ImmOpcode = PPC::STWU8; break;
2894 III.ImmOpcode = PPC::STDU;
2895 III.ImmMustBeMultipleOf = 4;
2897 case PPC::STFSUX: III.ImmOpcode = PPC::STFSU; break;
2898 case PPC::STFDUX: III.ImmOpcode = PPC::STFDU; break;
2908 if (!Subtarget.hasP9Vector())
2910 III.SignedImm = true;
2911 III.ZeroIsSpecialOrig = 1;
2912 III.ZeroIsSpecialNew = 2;
2913 III.IsCommutative = true;
2915 III.ConstantOpNo = 2;
2917 default: llvm_unreachable("Unknown opcode");
2919 III.ImmOpcode = PPC::LXV;
2920 III.ImmMustBeMultipleOf = 16;
2923 III.ImmOpcode = PPC::LXSSP;
2924 III.ImmMustBeMultipleOf = 4;
2927 III.ImmOpcode = PPC::LXSD;
2928 III.ImmMustBeMultipleOf = 4;
2931 III.ImmOpcode = PPC::STXV;
2932 III.ImmMustBeMultipleOf = 16;
2935 III.ImmOpcode = PPC::STXSSP;
2936 III.ImmMustBeMultipleOf = 4;
2939 III.ImmOpcode = PPC::STXSD;
2940 III.ImmMustBeMultipleOf = 4;
2948 // Utility function for swaping two arbitrary operands of an instruction.
2949 static void swapMIOperands(MachineInstr &MI, unsigned Op1, unsigned Op2) {
2950 assert(Op1 != Op2 && "Cannot swap operand with itself.");
2952 unsigned MaxOp = std::max(Op1, Op2);
2953 unsigned MinOp = std::min(Op1, Op2);
2954 MachineOperand MOp1 = MI.getOperand(MinOp);
2955 MachineOperand MOp2 = MI.getOperand(MaxOp);
2956 MI.RemoveOperand(std::max(Op1, Op2));
2957 MI.RemoveOperand(std::min(Op1, Op2));
2959 // If the operands we are swapping are the two at the end (the common case)
2960 // we can just remove both and add them in the opposite order.
2961 if (MaxOp - MinOp == 1 && MI.getNumOperands() == MinOp) {
2962 MI.addOperand(MOp2);
2963 MI.addOperand(MOp1);
2965 // Store all operands in a temporary vector, remove them and re-add in the
2967 SmallVector<MachineOperand, 2> MOps;
2968 unsigned TotalOps = MI.getNumOperands() + 2; // We've already removed 2 ops.
2969 for (unsigned i = MI.getNumOperands() - 1; i >= MinOp; i--) {
2970 MOps.push_back(MI.getOperand(i));
2971 MI.RemoveOperand(i);
2973 // MOp2 needs to be added next.
2974 MI.addOperand(MOp2);
2975 // Now add the rest.
2976 for (unsigned i = MI.getNumOperands(); i < TotalOps; i++) {
2978 MI.addOperand(MOp1);
2980 MI.addOperand(MOps.back());
2987 bool PPCInstrInfo::transformToImmForm(MachineInstr &MI, const ImmInstrInfo &III,
2988 unsigned ConstantOpNo,
2989 int64_t Imm) const {
2990 MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
2991 bool PostRA = !MRI.isSSA();
2992 // Exit early if we can't convert this.
2993 if ((ConstantOpNo != III.ConstantOpNo) && !III.IsCommutative)
2995 if (Imm % III.ImmMustBeMultipleOf)
2997 if (III.TruncateImmTo)
2998 Imm &= ((1 << III.TruncateImmTo) - 1);
2999 if (III.SignedImm) {
3000 APInt ActualValue(64, Imm, true);
3001 if (!ActualValue.isSignedIntN(III.ImmWidth))
3004 uint64_t UnsignedMax = (1 << III.ImmWidth) - 1;
3005 if ((uint64_t)Imm > UnsignedMax)
3009 // If we're post-RA, the instructions don't agree on whether register zero is
3010 // special, we can transform this as long as the register operand that will
3011 // end up in the location where zero is special isn't R0.
3012 if (PostRA && III.ZeroIsSpecialOrig != III.ZeroIsSpecialNew) {
3013 unsigned PosForOrigZero = III.ZeroIsSpecialOrig ? III.ZeroIsSpecialOrig :
3014 III.ZeroIsSpecialNew + 1;
3015 unsigned OrigZeroReg = MI.getOperand(PosForOrigZero).getReg();
3016 unsigned NewZeroReg = MI.getOperand(III.ZeroIsSpecialNew).getReg();
3017 // If R0 is in the operand where zero is special for the new instruction,
3018 // it is unsafe to transform if the constant operand isn't that operand.
3019 if ((NewZeroReg == PPC::R0 || NewZeroReg == PPC::X0) &&
3020 ConstantOpNo != III.ZeroIsSpecialNew)
3022 if ((OrigZeroReg == PPC::R0 || OrigZeroReg == PPC::X0) &&
3023 ConstantOpNo != PosForOrigZero)
3027 unsigned Opc = MI.getOpcode();
3028 bool SpecialShift32 =
3029 Opc == PPC::SLW || Opc == PPC::SLWo || Opc == PPC::SRW || Opc == PPC::SRWo;
3030 bool SpecialShift64 =
3031 Opc == PPC::SLD || Opc == PPC::SLDo || Opc == PPC::SRD || Opc == PPC::SRDo;
3032 bool SetCR = Opc == PPC::SLWo || Opc == PPC::SRWo ||
3033 Opc == PPC::SLDo || Opc == PPC::SRDo;
3035 Opc == PPC::SRW || Opc == PPC::SRWo || Opc == PPC::SRD || Opc == PPC::SRDo;
3037 MI.setDesc(get(III.ImmOpcode));
3038 if (ConstantOpNo == III.ConstantOpNo) {
3039 // Converting shifts to immediate form is a bit tricky since they may do
3040 // one of three things:
3041 // 1. If the shift amount is between OpSize and 2*OpSize, the result is zero
3042 // 2. If the shift amount is zero, the result is unchanged (save for maybe
3044 // 3. If the shift amount is in [1, OpSize), it's just a shift
3045 if (SpecialShift32 || SpecialShift64) {
3046 LoadImmediateInfo LII;
3049 LII.Is64Bit = SpecialShift64;
3050 uint64_t ShAmt = Imm & (SpecialShift32 ? 0x1F : 0x3F);
3051 if (Imm & (SpecialShift32 ? 0x20 : 0x40))
3052 replaceInstrWithLI(MI, LII);
3053 // Shifts by zero don't change the value. If we don't need to set CR0,
3054 // just convert this to a COPY. Can't do this post-RA since we've already
3055 // cleaned up the copies.
3056 else if (!SetCR && ShAmt == 0 && !PostRA) {
3057 MI.RemoveOperand(2);
3058 MI.setDesc(get(PPC::COPY));
3060 // The 32 bit and 64 bit instructions are quite different.
3061 if (SpecialShift32) {
3062 // Left shifts use (N, 0, 31-N), right shifts use (32-N, N, 31).
3063 uint64_t SH = RightShift ? 32 - ShAmt : ShAmt;
3064 uint64_t MB = RightShift ? ShAmt : 0;
3065 uint64_t ME = RightShift ? 31 : 31 - ShAmt;
3066 MI.getOperand(III.ConstantOpNo).ChangeToImmediate(SH);
3067 MachineInstrBuilder(*MI.getParent()->getParent(), MI).addImm(MB)
3070 // Left shifts use (N, 63-N), right shifts use (64-N, N).
3071 uint64_t SH = RightShift ? 64 - ShAmt : ShAmt;
3072 uint64_t ME = RightShift ? ShAmt : 63 - ShAmt;
3073 MI.getOperand(III.ConstantOpNo).ChangeToImmediate(SH);
3074 MachineInstrBuilder(*MI.getParent()->getParent(), MI).addImm(ME);
3078 MI.getOperand(ConstantOpNo).ChangeToImmediate(Imm);
3080 // Convert commutative instructions (switch the operands and convert the
3081 // desired one to an immediate.
3082 else if (III.IsCommutative) {
3083 MI.getOperand(ConstantOpNo).ChangeToImmediate(Imm);
3084 swapMIOperands(MI, ConstantOpNo, III.ConstantOpNo);
3086 llvm_unreachable("Should have exited early!");
3088 // For instructions for which the constant register replaces a different
3089 // operand than where the immediate goes, we need to swap them.
3090 if (III.ConstantOpNo != III.ImmOpNo)
3091 swapMIOperands(MI, III.ConstantOpNo, III.ImmOpNo);
3093 // If the R0/X0 register is special for the original instruction and not for
3094 // the new instruction (or vice versa), we need to fix up the register class.
3095 if (!PostRA && III.ZeroIsSpecialOrig != III.ZeroIsSpecialNew) {
3096 if (!III.ZeroIsSpecialOrig) {
3097 unsigned RegToModify = MI.getOperand(III.ZeroIsSpecialNew).getReg();
3098 const TargetRegisterClass *NewRC =
3099 MRI.getRegClass(RegToModify)->hasSuperClassEq(&PPC::GPRCRegClass) ?
3100 &PPC::GPRC_and_GPRC_NOR0RegClass : &PPC::G8RC_and_G8RC_NOX0RegClass;
3101 MRI.setRegClass(RegToModify, NewRC);
3107 const TargetRegisterClass *
3108 PPCInstrInfo::updatedRC(const TargetRegisterClass *RC) const {
3109 if (Subtarget.hasVSX() && RC == &PPC::VRRCRegClass)
3110 return &PPC::VSRCRegClass;
3114 int PPCInstrInfo::getRecordFormOpcode(unsigned Opcode) {
3115 return PPC::getRecordFormOpcode(Opcode);
3118 // This function returns true if the machine instruction
3119 // always outputs a value by sign-extending a 32 bit value,
3120 // i.e. 0 to 31-th bits are same as 32-th bit.
3121 static bool isSignExtendingOp(const MachineInstr &MI) {
3122 int Opcode = MI.getOpcode();
3123 if (Opcode == PPC::LI || Opcode == PPC::LI8 ||
3124 Opcode == PPC::LIS || Opcode == PPC::LIS8 ||
3125 Opcode == PPC::SRAW || Opcode == PPC::SRAWo ||
3126 Opcode == PPC::SRAWI || Opcode == PPC::SRAWIo ||
3127 Opcode == PPC::LWA || Opcode == PPC::LWAX ||
3128 Opcode == PPC::LWA_32 || Opcode == PPC::LWAX_32 ||
3129 Opcode == PPC::LHA || Opcode == PPC::LHAX ||
3130 Opcode == PPC::LHA8 || Opcode == PPC::LHAX8 ||
3131 Opcode == PPC::LBZ || Opcode == PPC::LBZX ||
3132 Opcode == PPC::LBZ8 || Opcode == PPC::LBZX8 ||
3133 Opcode == PPC::LBZU || Opcode == PPC::LBZUX ||
3134 Opcode == PPC::LBZU8 || Opcode == PPC::LBZUX8 ||
3135 Opcode == PPC::LHZ || Opcode == PPC::LHZX ||
3136 Opcode == PPC::LHZ8 || Opcode == PPC::LHZX8 ||
3137 Opcode == PPC::LHZU || Opcode == PPC::LHZUX ||
3138 Opcode == PPC::LHZU8 || Opcode == PPC::LHZUX8 ||
3139 Opcode == PPC::EXTSB || Opcode == PPC::EXTSBo ||
3140 Opcode == PPC::EXTSH || Opcode == PPC::EXTSHo ||
3141 Opcode == PPC::EXTSB8 || Opcode == PPC::EXTSH8 ||
3142 Opcode == PPC::EXTSW || Opcode == PPC::EXTSWo ||
3143 Opcode == PPC::EXTSH8_32_64 || Opcode == PPC::EXTSW_32_64 ||
3144 Opcode == PPC::EXTSB8_32_64)
3147 if (Opcode == PPC::RLDICL && MI.getOperand(3).getImm() >= 33)
3150 if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINMo ||
3151 Opcode == PPC::RLWNM || Opcode == PPC::RLWNMo) &&
3152 MI.getOperand(3).getImm() > 0 &&
3153 MI.getOperand(3).getImm() <= MI.getOperand(4).getImm())
3159 // This function returns true if the machine instruction
3160 // always outputs zeros in higher 32 bits.
3161 static bool isZeroExtendingOp(const MachineInstr &MI) {
3162 int Opcode = MI.getOpcode();
3163 // The 16-bit immediate is sign-extended in li/lis.
3164 // If the most significant bit is zero, all higher bits are zero.
3165 if (Opcode == PPC::LI || Opcode == PPC::LI8 ||
3166 Opcode == PPC::LIS || Opcode == PPC::LIS8) {
3167 int64_t Imm = MI.getOperand(1).getImm();
3168 if (((uint64_t)Imm & ~0x7FFFuLL) == 0)
3172 // We have some variations of rotate-and-mask instructions
3173 // that clear higher 32-bits.
3174 if ((Opcode == PPC::RLDICL || Opcode == PPC::RLDICLo ||
3175 Opcode == PPC::RLDCL || Opcode == PPC::RLDCLo ||
3176 Opcode == PPC::RLDICL_32_64) &&
3177 MI.getOperand(3).getImm() >= 32)
3180 if ((Opcode == PPC::RLDIC || Opcode == PPC::RLDICo) &&
3181 MI.getOperand(3).getImm() >= 32 &&
3182 MI.getOperand(3).getImm() <= 63 - MI.getOperand(2).getImm())
3185 if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINMo ||
3186 Opcode == PPC::RLWNM || Opcode == PPC::RLWNMo ||
3187 Opcode == PPC::RLWINM8 || Opcode == PPC::RLWNM8) &&
3188 MI.getOperand(3).getImm() <= MI.getOperand(4).getImm())
3191 // There are other instructions that clear higher 32-bits.
3192 if (Opcode == PPC::CNTLZW || Opcode == PPC::CNTLZWo ||
3193 Opcode == PPC::CNTTZW || Opcode == PPC::CNTTZWo ||
3194 Opcode == PPC::CNTLZW8 || Opcode == PPC::CNTTZW8 ||
3195 Opcode == PPC::CNTLZD || Opcode == PPC::CNTLZDo ||
3196 Opcode == PPC::CNTTZD || Opcode == PPC::CNTTZDo ||
3197 Opcode == PPC::POPCNTD || Opcode == PPC::POPCNTW ||
3198 Opcode == PPC::SLW || Opcode == PPC::SLWo ||
3199 Opcode == PPC::SRW || Opcode == PPC::SRWo ||
3200 Opcode == PPC::SLW8 || Opcode == PPC::SRW8 ||
3201 Opcode == PPC::SLWI || Opcode == PPC::SLWIo ||
3202 Opcode == PPC::SRWI || Opcode == PPC::SRWIo ||
3203 Opcode == PPC::LWZ || Opcode == PPC::LWZX ||
3204 Opcode == PPC::LWZU || Opcode == PPC::LWZUX ||
3205 Opcode == PPC::LWBRX || Opcode == PPC::LHBRX ||
3206 Opcode == PPC::LHZ || Opcode == PPC::LHZX ||
3207 Opcode == PPC::LHZU || Opcode == PPC::LHZUX ||
3208 Opcode == PPC::LBZ || Opcode == PPC::LBZX ||
3209 Opcode == PPC::LBZU || Opcode == PPC::LBZUX ||
3210 Opcode == PPC::LWZ8 || Opcode == PPC::LWZX8 ||
3211 Opcode == PPC::LWZU8 || Opcode == PPC::LWZUX8 ||
3212 Opcode == PPC::LWBRX8 || Opcode == PPC::LHBRX8 ||
3213 Opcode == PPC::LHZ8 || Opcode == PPC::LHZX8 ||
3214 Opcode == PPC::LHZU8 || Opcode == PPC::LHZUX8 ||
3215 Opcode == PPC::LBZ8 || Opcode == PPC::LBZX8 ||
3216 Opcode == PPC::LBZU8 || Opcode == PPC::LBZUX8 ||
3217 Opcode == PPC::ANDIo || Opcode == PPC::ANDISo ||
3218 Opcode == PPC::ROTRWI || Opcode == PPC::ROTRWIo ||
3219 Opcode == PPC::EXTLWI || Opcode == PPC::EXTLWIo ||
3220 Opcode == PPC::MFVSRWZ)
3226 // This function returns true if the input MachineInstr is a TOC save
3228 bool PPCInstrInfo::isTOCSaveMI(const MachineInstr &MI) const {
3229 if (!MI.getOperand(1).isImm() || !MI.getOperand(2).isReg())
3231 unsigned TOCSaveOffset = Subtarget.getFrameLowering()->getTOCSaveOffset();
3232 unsigned StackOffset = MI.getOperand(1).getImm();
3233 unsigned StackReg = MI.getOperand(2).getReg();
3234 if (StackReg == PPC::X1 && StackOffset == TOCSaveOffset)
3240 // We limit the max depth to track incoming values of PHIs or binary ops
3241 // (e.g. AND) to avoid excessive cost.
3242 const unsigned MAX_DEPTH = 1;
3245 PPCInstrInfo::isSignOrZeroExtended(const MachineInstr &MI, bool SignExt,
3246 const unsigned Depth) const {
3247 const MachineFunction *MF = MI.getParent()->getParent();
3248 const MachineRegisterInfo *MRI = &MF->getRegInfo();
3250 // If we know this instruction returns sign- or zero-extended result,
3252 if (SignExt ? isSignExtendingOp(MI):
3253 isZeroExtendingOp(MI))
3256 switch (MI.getOpcode()) {
3258 unsigned SrcReg = MI.getOperand(1).getReg();
3260 // In both ELFv1 and v2 ABI, method parameters and the return value
3261 // are sign- or zero-extended.
3262 if (MF->getSubtarget<PPCSubtarget>().isSVR4ABI()) {
3263 const PPCFunctionInfo *FuncInfo = MF->getInfo<PPCFunctionInfo>();
3264 // We check the ZExt/SExt flags for a method parameter.
3265 if (MI.getParent()->getBasicBlock() ==
3266 &MF->getFunction().getEntryBlock()) {
3267 unsigned VReg = MI.getOperand(0).getReg();
3268 if (MF->getRegInfo().isLiveIn(VReg))
3269 return SignExt ? FuncInfo->isLiveInSExt(VReg) :
3270 FuncInfo->isLiveInZExt(VReg);
3273 // For a method return value, we check the ZExt/SExt flags in attribute.
3274 // We assume the following code sequence for method call.
3275 // ADJCALLSTACKDOWN 32, implicit dead %r1, implicit %r1
3276 // BL8_NOP @func,...
3277 // ADJCALLSTACKUP 32, 0, implicit dead %r1, implicit %r1
3278 // %5 = COPY %x3; G8RC:%5
3279 if (SrcReg == PPC::X3) {
3280 const MachineBasicBlock *MBB = MI.getParent();
3281 MachineBasicBlock::const_instr_iterator II =
3282 MachineBasicBlock::const_instr_iterator(&MI);
3283 if (II != MBB->instr_begin() &&
3284 (--II)->getOpcode() == PPC::ADJCALLSTACKUP) {
3285 const MachineInstr &CallMI = *(--II);
3286 if (CallMI.isCall() && CallMI.getOperand(0).isGlobal()) {
3287 const Function *CalleeFn =
3288 dyn_cast<Function>(CallMI.getOperand(0).getGlobal());
3291 const IntegerType *IntTy =
3292 dyn_cast<IntegerType>(CalleeFn->getReturnType());
3293 const AttributeSet &Attrs =
3294 CalleeFn->getAttributes().getRetAttributes();
3295 if (IntTy && IntTy->getBitWidth() <= 32)
3296 return Attrs.hasAttribute(SignExt ? Attribute::SExt :
3303 // If this is a copy from another register, we recursively check source.
3304 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
3306 const MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
3308 return isSignOrZeroExtended(*SrcMI, SignExt, Depth);
3325 // logical operation with 16-bit immediate does not change the upper bits.
3326 // So, we track the operand register as we do for register copy.
3327 unsigned SrcReg = MI.getOperand(1).getReg();
3328 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
3330 const MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
3332 return isSignOrZeroExtended(*SrcMI, SignExt, Depth);
3337 // If all incoming values are sign-/zero-extended,
3338 // the output of OR, ISEL or PHI is also sign-/zero-extended.
3343 if (Depth >= MAX_DEPTH)
3346 // The input registers for PHI are operand 1, 3, ...
3347 // The input registers for others are operand 1 and 2.
3348 unsigned E = 3, D = 1;
3349 if (MI.getOpcode() == PPC::PHI) {
3350 E = MI.getNumOperands();
3354 for (unsigned I = 1; I != E; I += D) {
3355 if (MI.getOperand(I).isReg()) {
3356 unsigned SrcReg = MI.getOperand(I).getReg();
3357 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
3359 const MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
3360 if (SrcMI == NULL || !isSignOrZeroExtended(*SrcMI, SignExt, Depth+1))
3369 // If at least one of the incoming values of an AND is zero extended
3370 // then the output is also zero-extended. If both of the incoming values
3371 // are sign-extended then the output is also sign extended.
3374 if (Depth >= MAX_DEPTH)
3377 assert(MI.getOperand(1).isReg() && MI.getOperand(2).isReg());
3379 unsigned SrcReg1 = MI.getOperand(1).getReg();
3380 unsigned SrcReg2 = MI.getOperand(2).getReg();
3382 if (!TargetRegisterInfo::isVirtualRegister(SrcReg1) ||
3383 !TargetRegisterInfo::isVirtualRegister(SrcReg2))
3386 const MachineInstr *MISrc1 = MRI->getVRegDef(SrcReg1);
3387 const MachineInstr *MISrc2 = MRI->getVRegDef(SrcReg2);
3388 if (!MISrc1 || !MISrc2)
3392 return isSignOrZeroExtended(*MISrc1, SignExt, Depth+1) &&
3393 isSignOrZeroExtended(*MISrc2, SignExt, Depth+1);
3395 return isSignOrZeroExtended(*MISrc1, SignExt, Depth+1) ||
3396 isSignOrZeroExtended(*MISrc2, SignExt, Depth+1);