1 //===-- PPCInstrInfo.cpp - PowerPC Instruction Information ----------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file contains the PowerPC implementation of the TargetInstrInfo class.
11 //===----------------------------------------------------------------------===//
13 #include "PPCInstrInfo.h"
14 #include "MCTargetDesc/PPCPredicates.h"
16 #include "PPCHazardRecognizers.h"
17 #include "PPCInstrBuilder.h"
18 #include "PPCMachineFunctionInfo.h"
19 #include "PPCTargetMachine.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/CodeGen/LiveIntervals.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineMemOperand.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/PseudoSourceValue.h"
29 #include "llvm/CodeGen/ScheduleDAG.h"
30 #include "llvm/CodeGen/SlotIndexes.h"
31 #include "llvm/CodeGen/StackMaps.h"
32 #include "llvm/MC/MCAsmInfo.h"
33 #include "llvm/MC/MCInst.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/TargetRegistry.h"
38 #include "llvm/Support/raw_ostream.h"
42 #define DEBUG_TYPE "ppc-instr-info"
44 #define GET_INSTRMAP_INFO
45 #define GET_INSTRINFO_CTOR_DTOR
46 #include "PPCGenInstrInfo.inc"
48 STATISTIC(NumStoreSPILLVSRRCAsVec,
49 "Number of spillvsrrc spilled to stack as vec");
50 STATISTIC(NumStoreSPILLVSRRCAsGpr,
51 "Number of spillvsrrc spilled to stack as gpr");
52 STATISTIC(NumGPRtoVSRSpill, "Number of gpr spills to spillvsrrc");
53 STATISTIC(CmpIselsConverted,
54 "Number of ISELs that depend on comparison of constants converted");
55 STATISTIC(MissedConvertibleImmediateInstrs,
56 "Number of compare-immediate instructions fed by constants");
57 STATISTIC(NumRcRotatesConvertedToRcAnd,
58 "Number of record-form rotates converted to record-form andi");
61 opt<bool> DisableCTRLoopAnal("disable-ppc-ctrloop-analysis", cl::Hidden,
62 cl::desc("Disable analysis for CTR loops"));
64 static cl::opt<bool> DisableCmpOpt("disable-ppc-cmp-opt",
65 cl::desc("Disable compare instruction optimization"), cl::Hidden);
67 static cl::opt<bool> VSXSelfCopyCrash("crash-on-ppc-vsx-self-copy",
68 cl::desc("Causes the backend to crash instead of generating a nop VSX copy"),
72 UseOldLatencyCalc("ppc-old-latency-calc", cl::Hidden,
73 cl::desc("Use the old (incorrect) instruction latency calculation"));
75 // Index into the OpcodesForSpill array.
85 SOK_VectorFloat8Spill,
86 SOK_VectorFloat4Spill,
94 SOK_LastOpcodeSpill // This must be last on the enum.
97 // Pin the vtable to this file.
98 void PPCInstrInfo::anchor() {}
100 PPCInstrInfo::PPCInstrInfo(PPCSubtarget &STI)
101 : PPCGenInstrInfo(PPC::ADJCALLSTACKDOWN, PPC::ADJCALLSTACKUP,
102 /* CatchRetOpcode */ -1,
103 STI.isPPC64() ? PPC::BLR8 : PPC::BLR),
104 Subtarget(STI), RI(STI.getTargetMachine()) {}
106 /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for
107 /// this target when scheduling the DAG.
108 ScheduleHazardRecognizer *
109 PPCInstrInfo::CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
110 const ScheduleDAG *DAG) const {
112 static_cast<const PPCSubtarget *>(STI)->getDarwinDirective();
113 if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2 ||
114 Directive == PPC::DIR_E500mc || Directive == PPC::DIR_E5500) {
115 const InstrItineraryData *II =
116 static_cast<const PPCSubtarget *>(STI)->getInstrItineraryData();
117 return new ScoreboardHazardRecognizer(II, DAG);
120 return TargetInstrInfo::CreateTargetHazardRecognizer(STI, DAG);
123 /// CreateTargetPostRAHazardRecognizer - Return the postRA hazard recognizer
124 /// to use for this target when scheduling the DAG.
125 ScheduleHazardRecognizer *
126 PPCInstrInfo::CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
127 const ScheduleDAG *DAG) const {
129 DAG->MF.getSubtarget<PPCSubtarget>().getDarwinDirective();
131 // FIXME: Leaving this as-is until we have POWER9 scheduling info
132 if (Directive == PPC::DIR_PWR7 || Directive == PPC::DIR_PWR8)
133 return new PPCDispatchGroupSBHazardRecognizer(II, DAG);
135 // Most subtargets use a PPC970 recognizer.
136 if (Directive != PPC::DIR_440 && Directive != PPC::DIR_A2 &&
137 Directive != PPC::DIR_E500mc && Directive != PPC::DIR_E5500) {
138 assert(DAG->TII && "No InstrInfo?");
140 return new PPCHazardRecognizer970(*DAG);
143 return new ScoreboardHazardRecognizer(II, DAG);
146 unsigned PPCInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
147 const MachineInstr &MI,
148 unsigned *PredCost) const {
149 if (!ItinData || UseOldLatencyCalc)
150 return PPCGenInstrInfo::getInstrLatency(ItinData, MI, PredCost);
152 // The default implementation of getInstrLatency calls getStageLatency, but
153 // getStageLatency does not do the right thing for us. While we have
154 // itinerary, most cores are fully pipelined, and so the itineraries only
155 // express the first part of the pipeline, not every stage. Instead, we need
156 // to use the listed output operand cycle number (using operand 0 here, which
159 unsigned Latency = 1;
160 unsigned DefClass = MI.getDesc().getSchedClass();
161 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
162 const MachineOperand &MO = MI.getOperand(i);
163 if (!MO.isReg() || !MO.isDef() || MO.isImplicit())
166 int Cycle = ItinData->getOperandCycle(DefClass, i);
170 Latency = std::max(Latency, (unsigned) Cycle);
176 int PPCInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
177 const MachineInstr &DefMI, unsigned DefIdx,
178 const MachineInstr &UseMI,
179 unsigned UseIdx) const {
180 int Latency = PPCGenInstrInfo::getOperandLatency(ItinData, DefMI, DefIdx,
183 if (!DefMI.getParent())
186 const MachineOperand &DefMO = DefMI.getOperand(DefIdx);
187 unsigned Reg = DefMO.getReg();
190 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
191 const MachineRegisterInfo *MRI =
192 &DefMI.getParent()->getParent()->getRegInfo();
193 IsRegCR = MRI->getRegClass(Reg)->hasSuperClassEq(&PPC::CRRCRegClass) ||
194 MRI->getRegClass(Reg)->hasSuperClassEq(&PPC::CRBITRCRegClass);
196 IsRegCR = PPC::CRRCRegClass.contains(Reg) ||
197 PPC::CRBITRCRegClass.contains(Reg);
200 if (UseMI.isBranch() && IsRegCR) {
202 Latency = getInstrLatency(ItinData, DefMI);
204 // On some cores, there is an additional delay between writing to a condition
205 // register, and using it from a branch.
206 unsigned Directive = Subtarget.getDarwinDirective();
220 // FIXME: Is this needed for POWER9?
229 // This function does not list all associative and commutative operations, but
230 // only those worth feeding through the machine combiner in an attempt to
231 // reduce the critical path. Mostly, this means floating-point operations,
232 // because they have high latencies (compared to other operations, such and
233 // and/or, which are also associative and commutative, but have low latencies).
234 bool PPCInstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst) const {
235 switch (Inst.getOpcode()) {
268 bool PPCInstrInfo::getMachineCombinerPatterns(
270 SmallVectorImpl<MachineCombinerPattern> &Patterns) const {
271 // Using the machine combiner in this way is potentially expensive, so
272 // restrict to when aggressive optimizations are desired.
273 if (Subtarget.getTargetMachine().getOptLevel() != CodeGenOpt::Aggressive)
276 // FP reassociation is only legal when we don't need strict IEEE semantics.
277 if (!Root.getParent()->getParent()->getTarget().Options.UnsafeFPMath)
280 return TargetInstrInfo::getMachineCombinerPatterns(Root, Patterns);
283 // Detect 32 -> 64-bit extensions where we may reuse the low sub-register.
284 bool PPCInstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
285 unsigned &SrcReg, unsigned &DstReg,
286 unsigned &SubIdx) const {
287 switch (MI.getOpcode()) {
288 default: return false;
291 case PPC::EXTSW_32_64:
292 SrcReg = MI.getOperand(1).getReg();
293 DstReg = MI.getOperand(0).getReg();
294 SubIdx = PPC::sub_32;
299 unsigned PPCInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
300 int &FrameIndex) const {
301 unsigned Opcode = MI.getOpcode();
302 const unsigned *OpcodesForSpill = getLoadOpcodesForSpillArray();
303 const unsigned *End = OpcodesForSpill + SOK_LastOpcodeSpill;
305 if (End != std::find(OpcodesForSpill, End, Opcode)) {
306 // Check for the operands added by addFrameReference (the immediate is the
307 // offset which defaults to 0).
308 if (MI.getOperand(1).isImm() && !MI.getOperand(1).getImm() &&
309 MI.getOperand(2).isFI()) {
310 FrameIndex = MI.getOperand(2).getIndex();
311 return MI.getOperand(0).getReg();
317 // For opcodes with the ReMaterializable flag set, this function is called to
318 // verify the instruction is really rematable.
319 bool PPCInstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
320 AliasAnalysis *AA) const {
321 switch (MI.getOpcode()) {
323 // This function should only be called for opcodes with the ReMaterializable
325 llvm_unreachable("Unknown rematerializable operation!");
332 case PPC::ADDIStocHA:
334 case PPC::LOAD_STACK_GUARD:
341 case PPC::V_SETALLONESB:
342 case PPC::V_SETALLONESH:
343 case PPC::V_SETALLONES:
351 unsigned PPCInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
352 int &FrameIndex) const {
353 unsigned Opcode = MI.getOpcode();
354 const unsigned *OpcodesForSpill = getStoreOpcodesForSpillArray();
355 const unsigned *End = OpcodesForSpill + SOK_LastOpcodeSpill;
357 if (End != std::find(OpcodesForSpill, End, Opcode)) {
358 if (MI.getOperand(1).isImm() && !MI.getOperand(1).getImm() &&
359 MI.getOperand(2).isFI()) {
360 FrameIndex = MI.getOperand(2).getIndex();
361 return MI.getOperand(0).getReg();
367 MachineInstr *PPCInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
369 unsigned OpIdx2) const {
370 MachineFunction &MF = *MI.getParent()->getParent();
372 // Normal instructions can be commuted the obvious way.
373 if (MI.getOpcode() != PPC::RLWIMI && MI.getOpcode() != PPC::RLWIMIo)
374 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
375 // Note that RLWIMI can be commuted as a 32-bit instruction, but not as a
376 // 64-bit instruction (so we don't handle PPC::RLWIMI8 here), because
377 // changing the relative order of the mask operands might change what happens
378 // to the high-bits of the mask (and, thus, the result).
380 // Cannot commute if it has a non-zero rotate count.
381 if (MI.getOperand(3).getImm() != 0)
384 // If we have a zero rotate count, we have:
386 // Op0 = (Op1 & ~M) | (Op2 & M)
388 // M = mask((ME+1)&31, (MB-1)&31)
389 // Op0 = (Op2 & ~M) | (Op1 & M)
392 assert(((OpIdx1 == 1 && OpIdx2 == 2) || (OpIdx1 == 2 && OpIdx2 == 1)) &&
393 "Only the operands 1 and 2 can be swapped in RLSIMI/RLWIMIo.");
394 Register Reg0 = MI.getOperand(0).getReg();
395 Register Reg1 = MI.getOperand(1).getReg();
396 Register Reg2 = MI.getOperand(2).getReg();
397 unsigned SubReg1 = MI.getOperand(1).getSubReg();
398 unsigned SubReg2 = MI.getOperand(2).getSubReg();
399 bool Reg1IsKill = MI.getOperand(1).isKill();
400 bool Reg2IsKill = MI.getOperand(2).isKill();
401 bool ChangeReg0 = false;
402 // If machine instrs are no longer in two-address forms, update
403 // destination register as well.
405 // Must be two address instruction!
406 assert(MI.getDesc().getOperandConstraint(0, MCOI::TIED_TO) &&
407 "Expecting a two-address instruction!");
408 assert(MI.getOperand(0).getSubReg() == SubReg1 && "Tied subreg mismatch");
414 unsigned MB = MI.getOperand(4).getImm();
415 unsigned ME = MI.getOperand(5).getImm();
417 // We can't commute a trivial mask (there is no way to represent an all-zero
419 if (MB == 0 && ME == 31)
423 // Create a new instruction.
424 Register Reg0 = ChangeReg0 ? Reg2 : MI.getOperand(0).getReg();
425 bool Reg0IsDead = MI.getOperand(0).isDead();
426 return BuildMI(MF, MI.getDebugLoc(), MI.getDesc())
427 .addReg(Reg0, RegState::Define | getDeadRegState(Reg0IsDead))
428 .addReg(Reg2, getKillRegState(Reg2IsKill))
429 .addReg(Reg1, getKillRegState(Reg1IsKill))
430 .addImm((ME + 1) & 31)
431 .addImm((MB - 1) & 31);
435 MI.getOperand(0).setReg(Reg2);
436 MI.getOperand(0).setSubReg(SubReg2);
438 MI.getOperand(2).setReg(Reg1);
439 MI.getOperand(1).setReg(Reg2);
440 MI.getOperand(2).setSubReg(SubReg1);
441 MI.getOperand(1).setSubReg(SubReg2);
442 MI.getOperand(2).setIsKill(Reg1IsKill);
443 MI.getOperand(1).setIsKill(Reg2IsKill);
445 // Swap the mask around.
446 MI.getOperand(4).setImm((ME + 1) & 31);
447 MI.getOperand(5).setImm((MB - 1) & 31);
451 bool PPCInstrInfo::findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1,
452 unsigned &SrcOpIdx2) const {
453 // For VSX A-Type FMA instructions, it is the first two operands that can be
454 // commuted, however, because the non-encoded tied input operand is listed
455 // first, the operands to swap are actually the second and third.
457 int AltOpc = PPC::getAltVSXFMAOpcode(MI.getOpcode());
459 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
461 // The commutable operand indices are 2 and 3. Return them in SrcOpIdx1
463 return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 2, 3);
466 void PPCInstrInfo::insertNoop(MachineBasicBlock &MBB,
467 MachineBasicBlock::iterator MI) const {
468 // This function is used for scheduling, and the nop wanted here is the type
469 // that terminates dispatch groups on the POWER cores.
470 unsigned Directive = Subtarget.getDarwinDirective();
473 default: Opcode = PPC::NOP; break;
474 case PPC::DIR_PWR6: Opcode = PPC::NOP_GT_PWR6; break;
475 case PPC::DIR_PWR7: Opcode = PPC::NOP_GT_PWR7; break;
476 case PPC::DIR_PWR8: Opcode = PPC::NOP_GT_PWR7; break; /* FIXME: Update when P8 InstrScheduling model is ready */
477 // FIXME: Update when POWER9 scheduling model is ready.
478 case PPC::DIR_PWR9: Opcode = PPC::NOP_GT_PWR7; break;
482 BuildMI(MBB, MI, DL, get(Opcode));
485 /// Return the noop instruction to use for a noop.
486 void PPCInstrInfo::getNoop(MCInst &NopInst) const {
487 NopInst.setOpcode(PPC::NOP);
491 // Note: If the condition register is set to CTR or CTR8 then this is a
492 // BDNZ (imm == 1) or BDZ (imm == 0) branch.
493 bool PPCInstrInfo::analyzeBranch(MachineBasicBlock &MBB,
494 MachineBasicBlock *&TBB,
495 MachineBasicBlock *&FBB,
496 SmallVectorImpl<MachineOperand> &Cond,
497 bool AllowModify) const {
498 bool isPPC64 = Subtarget.isPPC64();
500 // If the block has no terminators, it just falls into the block after it.
501 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
505 if (!isUnpredicatedTerminator(*I))
509 // If the BB ends with an unconditional branch to the fallthrough BB,
510 // we eliminate the branch instruction.
511 if (I->getOpcode() == PPC::B &&
512 MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
513 I->eraseFromParent();
515 // We update iterator after deleting the last branch.
516 I = MBB.getLastNonDebugInstr();
517 if (I == MBB.end() || !isUnpredicatedTerminator(*I))
522 // Get the last instruction in the block.
523 MachineInstr &LastInst = *I;
525 // If there is only one terminator instruction, process it.
526 if (I == MBB.begin() || !isUnpredicatedTerminator(*--I)) {
527 if (LastInst.getOpcode() == PPC::B) {
528 if (!LastInst.getOperand(0).isMBB())
530 TBB = LastInst.getOperand(0).getMBB();
532 } else if (LastInst.getOpcode() == PPC::BCC) {
533 if (!LastInst.getOperand(2).isMBB())
535 // Block ends with fall-through condbranch.
536 TBB = LastInst.getOperand(2).getMBB();
537 Cond.push_back(LastInst.getOperand(0));
538 Cond.push_back(LastInst.getOperand(1));
540 } else if (LastInst.getOpcode() == PPC::BC) {
541 if (!LastInst.getOperand(1).isMBB())
543 // Block ends with fall-through condbranch.
544 TBB = LastInst.getOperand(1).getMBB();
545 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET));
546 Cond.push_back(LastInst.getOperand(0));
548 } else if (LastInst.getOpcode() == PPC::BCn) {
549 if (!LastInst.getOperand(1).isMBB())
551 // Block ends with fall-through condbranch.
552 TBB = LastInst.getOperand(1).getMBB();
553 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET));
554 Cond.push_back(LastInst.getOperand(0));
556 } else if (LastInst.getOpcode() == PPC::BDNZ8 ||
557 LastInst.getOpcode() == PPC::BDNZ) {
558 if (!LastInst.getOperand(0).isMBB())
560 if (DisableCTRLoopAnal)
562 TBB = LastInst.getOperand(0).getMBB();
563 Cond.push_back(MachineOperand::CreateImm(1));
564 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR,
567 } else if (LastInst.getOpcode() == PPC::BDZ8 ||
568 LastInst.getOpcode() == PPC::BDZ) {
569 if (!LastInst.getOperand(0).isMBB())
571 if (DisableCTRLoopAnal)
573 TBB = LastInst.getOperand(0).getMBB();
574 Cond.push_back(MachineOperand::CreateImm(0));
575 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR,
580 // Otherwise, don't know what this is.
584 // Get the instruction before it if it's a terminator.
585 MachineInstr &SecondLastInst = *I;
587 // If there are three terminators, we don't know what sort of block this is.
588 if (I != MBB.begin() && isUnpredicatedTerminator(*--I))
591 // If the block ends with PPC::B and PPC:BCC, handle it.
592 if (SecondLastInst.getOpcode() == PPC::BCC &&
593 LastInst.getOpcode() == PPC::B) {
594 if (!SecondLastInst.getOperand(2).isMBB() ||
595 !LastInst.getOperand(0).isMBB())
597 TBB = SecondLastInst.getOperand(2).getMBB();
598 Cond.push_back(SecondLastInst.getOperand(0));
599 Cond.push_back(SecondLastInst.getOperand(1));
600 FBB = LastInst.getOperand(0).getMBB();
602 } else if (SecondLastInst.getOpcode() == PPC::BC &&
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_SET));
609 Cond.push_back(SecondLastInst.getOperand(0));
610 FBB = LastInst.getOperand(0).getMBB();
612 } else if (SecondLastInst.getOpcode() == PPC::BCn &&
613 LastInst.getOpcode() == PPC::B) {
614 if (!SecondLastInst.getOperand(1).isMBB() ||
615 !LastInst.getOperand(0).isMBB())
617 TBB = SecondLastInst.getOperand(1).getMBB();
618 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET));
619 Cond.push_back(SecondLastInst.getOperand(0));
620 FBB = LastInst.getOperand(0).getMBB();
622 } else if ((SecondLastInst.getOpcode() == PPC::BDNZ8 ||
623 SecondLastInst.getOpcode() == PPC::BDNZ) &&
624 LastInst.getOpcode() == PPC::B) {
625 if (!SecondLastInst.getOperand(0).isMBB() ||
626 !LastInst.getOperand(0).isMBB())
628 if (DisableCTRLoopAnal)
630 TBB = SecondLastInst.getOperand(0).getMBB();
631 Cond.push_back(MachineOperand::CreateImm(1));
632 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR,
634 FBB = LastInst.getOperand(0).getMBB();
636 } else if ((SecondLastInst.getOpcode() == PPC::BDZ8 ||
637 SecondLastInst.getOpcode() == PPC::BDZ) &&
638 LastInst.getOpcode() == PPC::B) {
639 if (!SecondLastInst.getOperand(0).isMBB() ||
640 !LastInst.getOperand(0).isMBB())
642 if (DisableCTRLoopAnal)
644 TBB = SecondLastInst.getOperand(0).getMBB();
645 Cond.push_back(MachineOperand::CreateImm(0));
646 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR,
648 FBB = LastInst.getOperand(0).getMBB();
652 // If the block ends with two PPC:Bs, handle it. The second one is not
653 // executed, so remove it.
654 if (SecondLastInst.getOpcode() == PPC::B && LastInst.getOpcode() == PPC::B) {
655 if (!SecondLastInst.getOperand(0).isMBB())
657 TBB = SecondLastInst.getOperand(0).getMBB();
660 I->eraseFromParent();
664 // Otherwise, can't handle this.
668 unsigned PPCInstrInfo::removeBranch(MachineBasicBlock &MBB,
669 int *BytesRemoved) const {
670 assert(!BytesRemoved && "code size not handled");
672 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
676 if (I->getOpcode() != PPC::B && I->getOpcode() != PPC::BCC &&
677 I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn &&
678 I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ &&
679 I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ)
682 // Remove the branch.
683 I->eraseFromParent();
687 if (I == MBB.begin()) return 1;
689 if (I->getOpcode() != PPC::BCC &&
690 I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn &&
691 I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ &&
692 I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ)
695 // Remove the branch.
696 I->eraseFromParent();
700 unsigned PPCInstrInfo::insertBranch(MachineBasicBlock &MBB,
701 MachineBasicBlock *TBB,
702 MachineBasicBlock *FBB,
703 ArrayRef<MachineOperand> Cond,
705 int *BytesAdded) const {
706 // Shouldn't be a fall through.
707 assert(TBB && "insertBranch must not be told to insert a fallthrough");
708 assert((Cond.size() == 2 || Cond.size() == 0) &&
709 "PPC branch conditions have two components!");
710 assert(!BytesAdded && "code size not handled");
712 bool isPPC64 = Subtarget.isPPC64();
716 if (Cond.empty()) // Unconditional branch
717 BuildMI(&MBB, DL, get(PPC::B)).addMBB(TBB);
718 else if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8)
719 BuildMI(&MBB, DL, get(Cond[0].getImm() ?
720 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
721 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB);
722 else if (Cond[0].getImm() == PPC::PRED_BIT_SET)
723 BuildMI(&MBB, DL, get(PPC::BC)).add(Cond[1]).addMBB(TBB);
724 else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET)
725 BuildMI(&MBB, DL, get(PPC::BCn)).add(Cond[1]).addMBB(TBB);
726 else // Conditional branch
727 BuildMI(&MBB, DL, get(PPC::BCC))
728 .addImm(Cond[0].getImm())
734 // Two-way Conditional Branch.
735 if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8)
736 BuildMI(&MBB, DL, get(Cond[0].getImm() ?
737 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
738 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB);
739 else if (Cond[0].getImm() == PPC::PRED_BIT_SET)
740 BuildMI(&MBB, DL, get(PPC::BC)).add(Cond[1]).addMBB(TBB);
741 else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET)
742 BuildMI(&MBB, DL, get(PPC::BCn)).add(Cond[1]).addMBB(TBB);
744 BuildMI(&MBB, DL, get(PPC::BCC))
745 .addImm(Cond[0].getImm())
748 BuildMI(&MBB, DL, get(PPC::B)).addMBB(FBB);
753 bool PPCInstrInfo::canInsertSelect(const MachineBasicBlock &MBB,
754 ArrayRef<MachineOperand> Cond,
755 unsigned TrueReg, unsigned FalseReg,
756 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
757 if (Cond.size() != 2)
760 // If this is really a bdnz-like condition, then it cannot be turned into a
762 if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8)
765 // Check register classes.
766 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
767 const TargetRegisterClass *RC =
768 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
772 // isel is for regular integer GPRs only.
773 if (!PPC::GPRCRegClass.hasSubClassEq(RC) &&
774 !PPC::GPRC_NOR0RegClass.hasSubClassEq(RC) &&
775 !PPC::G8RCRegClass.hasSubClassEq(RC) &&
776 !PPC::G8RC_NOX0RegClass.hasSubClassEq(RC))
779 // FIXME: These numbers are for the A2, how well they work for other cores is
780 // an open question. On the A2, the isel instruction has a 2-cycle latency
781 // but single-cycle throughput. These numbers are used in combination with
782 // the MispredictPenalty setting from the active SchedMachineModel.
790 void PPCInstrInfo::insertSelect(MachineBasicBlock &MBB,
791 MachineBasicBlock::iterator MI,
792 const DebugLoc &dl, unsigned DestReg,
793 ArrayRef<MachineOperand> Cond, unsigned TrueReg,
794 unsigned FalseReg) const {
795 assert(Cond.size() == 2 &&
796 "PPC branch conditions have two components!");
798 // Get the register classes.
799 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
800 const TargetRegisterClass *RC =
801 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
802 assert(RC && "TrueReg and FalseReg must have overlapping register classes");
804 bool Is64Bit = PPC::G8RCRegClass.hasSubClassEq(RC) ||
805 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC);
807 PPC::GPRCRegClass.hasSubClassEq(RC) ||
808 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) &&
809 "isel is for regular integer GPRs only");
811 unsigned OpCode = Is64Bit ? PPC::ISEL8 : PPC::ISEL;
812 auto SelectPred = static_cast<PPC::Predicate>(Cond[0].getImm());
815 bool SwapOps = false;
816 switch (SelectPred) {
818 case PPC::PRED_EQ_MINUS:
819 case PPC::PRED_EQ_PLUS:
820 SubIdx = PPC::sub_eq; SwapOps = false; break;
822 case PPC::PRED_NE_MINUS:
823 case PPC::PRED_NE_PLUS:
824 SubIdx = PPC::sub_eq; SwapOps = true; break;
826 case PPC::PRED_LT_MINUS:
827 case PPC::PRED_LT_PLUS:
828 SubIdx = PPC::sub_lt; SwapOps = false; break;
830 case PPC::PRED_GE_MINUS:
831 case PPC::PRED_GE_PLUS:
832 SubIdx = PPC::sub_lt; SwapOps = true; break;
834 case PPC::PRED_GT_MINUS:
835 case PPC::PRED_GT_PLUS:
836 SubIdx = PPC::sub_gt; SwapOps = false; break;
838 case PPC::PRED_LE_MINUS:
839 case PPC::PRED_LE_PLUS:
840 SubIdx = PPC::sub_gt; SwapOps = true; break;
842 case PPC::PRED_UN_MINUS:
843 case PPC::PRED_UN_PLUS:
844 SubIdx = PPC::sub_un; SwapOps = false; break;
846 case PPC::PRED_NU_MINUS:
847 case PPC::PRED_NU_PLUS:
848 SubIdx = PPC::sub_un; SwapOps = true; break;
849 case PPC::PRED_BIT_SET: SubIdx = 0; SwapOps = false; break;
850 case PPC::PRED_BIT_UNSET: SubIdx = 0; SwapOps = true; break;
853 unsigned FirstReg = SwapOps ? FalseReg : TrueReg,
854 SecondReg = SwapOps ? TrueReg : FalseReg;
856 // The first input register of isel cannot be r0. If it is a member
857 // of a register class that can be r0, then copy it first (the
858 // register allocator should eliminate the copy).
859 if (MRI.getRegClass(FirstReg)->contains(PPC::R0) ||
860 MRI.getRegClass(FirstReg)->contains(PPC::X0)) {
861 const TargetRegisterClass *FirstRC =
862 MRI.getRegClass(FirstReg)->contains(PPC::X0) ?
863 &PPC::G8RC_NOX0RegClass : &PPC::GPRC_NOR0RegClass;
864 unsigned OldFirstReg = FirstReg;
865 FirstReg = MRI.createVirtualRegister(FirstRC);
866 BuildMI(MBB, MI, dl, get(TargetOpcode::COPY), FirstReg)
867 .addReg(OldFirstReg);
870 BuildMI(MBB, MI, dl, get(OpCode), DestReg)
871 .addReg(FirstReg).addReg(SecondReg)
872 .addReg(Cond[1].getReg(), 0, SubIdx);
875 static unsigned getCRBitValue(unsigned CRBit) {
877 if (CRBit == PPC::CR0LT || CRBit == PPC::CR1LT ||
878 CRBit == PPC::CR2LT || CRBit == PPC::CR3LT ||
879 CRBit == PPC::CR4LT || CRBit == PPC::CR5LT ||
880 CRBit == PPC::CR6LT || CRBit == PPC::CR7LT)
882 if (CRBit == PPC::CR0GT || CRBit == PPC::CR1GT ||
883 CRBit == PPC::CR2GT || CRBit == PPC::CR3GT ||
884 CRBit == PPC::CR4GT || CRBit == PPC::CR5GT ||
885 CRBit == PPC::CR6GT || CRBit == PPC::CR7GT)
887 if (CRBit == PPC::CR0EQ || CRBit == PPC::CR1EQ ||
888 CRBit == PPC::CR2EQ || CRBit == PPC::CR3EQ ||
889 CRBit == PPC::CR4EQ || CRBit == PPC::CR5EQ ||
890 CRBit == PPC::CR6EQ || CRBit == PPC::CR7EQ)
892 if (CRBit == PPC::CR0UN || CRBit == PPC::CR1UN ||
893 CRBit == PPC::CR2UN || CRBit == PPC::CR3UN ||
894 CRBit == PPC::CR4UN || CRBit == PPC::CR5UN ||
895 CRBit == PPC::CR6UN || CRBit == PPC::CR7UN)
898 assert(Ret != 4 && "Invalid CR bit register");
902 void PPCInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
903 MachineBasicBlock::iterator I,
904 const DebugLoc &DL, unsigned DestReg,
905 unsigned SrcReg, bool KillSrc) const {
906 // We can end up with self copies and similar things as a result of VSX copy
907 // legalization. Promote them here.
908 const TargetRegisterInfo *TRI = &getRegisterInfo();
909 if (PPC::F8RCRegClass.contains(DestReg) &&
910 PPC::VSRCRegClass.contains(SrcReg)) {
912 TRI->getMatchingSuperReg(DestReg, PPC::sub_64, &PPC::VSRCRegClass);
914 if (VSXSelfCopyCrash && SrcReg == SuperReg)
915 llvm_unreachable("nop VSX copy");
918 } else if (PPC::F8RCRegClass.contains(SrcReg) &&
919 PPC::VSRCRegClass.contains(DestReg)) {
921 TRI->getMatchingSuperReg(SrcReg, PPC::sub_64, &PPC::VSRCRegClass);
923 if (VSXSelfCopyCrash && DestReg == SuperReg)
924 llvm_unreachable("nop VSX copy");
929 // Different class register copy
930 if (PPC::CRBITRCRegClass.contains(SrcReg) &&
931 PPC::GPRCRegClass.contains(DestReg)) {
932 unsigned CRReg = getCRFromCRBit(SrcReg);
933 BuildMI(MBB, I, DL, get(PPC::MFOCRF), DestReg).addReg(CRReg);
934 getKillRegState(KillSrc);
935 // Rotate the CR bit in the CR fields to be the least significant bit and
936 // then mask with 0x1 (MB = ME = 31).
937 BuildMI(MBB, I, DL, get(PPC::RLWINM), DestReg)
938 .addReg(DestReg, RegState::Kill)
939 .addImm(TRI->getEncodingValue(CRReg) * 4 + (4 - getCRBitValue(SrcReg)))
943 } else if (PPC::CRRCRegClass.contains(SrcReg) &&
944 PPC::G8RCRegClass.contains(DestReg)) {
945 BuildMI(MBB, I, DL, get(PPC::MFOCRF8), DestReg).addReg(SrcReg);
946 getKillRegState(KillSrc);
948 } else if (PPC::CRRCRegClass.contains(SrcReg) &&
949 PPC::GPRCRegClass.contains(DestReg)) {
950 BuildMI(MBB, I, DL, get(PPC::MFOCRF), DestReg).addReg(SrcReg);
951 getKillRegState(KillSrc);
953 } else if (PPC::G8RCRegClass.contains(SrcReg) &&
954 PPC::VSFRCRegClass.contains(DestReg)) {
955 assert(Subtarget.hasDirectMove() &&
956 "Subtarget doesn't support directmove, don't know how to copy.");
957 BuildMI(MBB, I, DL, get(PPC::MTVSRD), DestReg).addReg(SrcReg);
959 getKillRegState(KillSrc);
961 } else if (PPC::VSFRCRegClass.contains(SrcReg) &&
962 PPC::G8RCRegClass.contains(DestReg)) {
963 assert(Subtarget.hasDirectMove() &&
964 "Subtarget doesn't support directmove, don't know how to copy.");
965 BuildMI(MBB, I, DL, get(PPC::MFVSRD), DestReg).addReg(SrcReg);
966 getKillRegState(KillSrc);
968 } else if (PPC::SPERCRegClass.contains(SrcReg) &&
969 PPC::SPE4RCRegClass.contains(DestReg)) {
970 BuildMI(MBB, I, DL, get(PPC::EFSCFD), DestReg).addReg(SrcReg);
971 getKillRegState(KillSrc);
973 } else if (PPC::SPE4RCRegClass.contains(SrcReg) &&
974 PPC::SPERCRegClass.contains(DestReg)) {
975 BuildMI(MBB, I, DL, get(PPC::EFDCFS), DestReg).addReg(SrcReg);
976 getKillRegState(KillSrc);
981 if (PPC::GPRCRegClass.contains(DestReg, SrcReg))
983 else if (PPC::G8RCRegClass.contains(DestReg, SrcReg))
985 else if (PPC::F4RCRegClass.contains(DestReg, SrcReg))
987 else if (PPC::CRRCRegClass.contains(DestReg, SrcReg))
989 else if (PPC::VRRCRegClass.contains(DestReg, SrcReg))
991 else if (PPC::VSRCRegClass.contains(DestReg, SrcReg))
992 // There are two different ways this can be done:
993 // 1. xxlor : This has lower latency (on the P7), 2 cycles, but can only
994 // issue in VSU pipeline 0.
995 // 2. xmovdp/xmovsp: This has higher latency (on the P7), 6 cycles, but
996 // can go to either pipeline.
997 // We'll always use xxlor here, because in practically all cases where
998 // copies are generated, they are close enough to some use that the
999 // lower-latency form is preferable.
1001 else if (PPC::VSFRCRegClass.contains(DestReg, SrcReg) ||
1002 PPC::VSSRCRegClass.contains(DestReg, SrcReg))
1003 Opc = (Subtarget.hasP9Vector()) ? PPC::XSCPSGNDP : PPC::XXLORf;
1004 else if (PPC::QFRCRegClass.contains(DestReg, SrcReg))
1006 else if (PPC::QSRCRegClass.contains(DestReg, SrcReg))
1008 else if (PPC::QBRCRegClass.contains(DestReg, SrcReg))
1010 else if (PPC::CRBITRCRegClass.contains(DestReg, SrcReg))
1012 else if (PPC::SPE4RCRegClass.contains(DestReg, SrcReg))
1014 else if (PPC::SPERCRegClass.contains(DestReg, SrcReg))
1017 llvm_unreachable("Impossible reg-to-reg copy");
1019 const MCInstrDesc &MCID = get(Opc);
1020 if (MCID.getNumOperands() == 3)
1021 BuildMI(MBB, I, DL, MCID, DestReg)
1022 .addReg(SrcReg).addReg(SrcReg, getKillRegState(KillSrc));
1024 BuildMI(MBB, I, DL, MCID, DestReg).addReg(SrcReg, getKillRegState(KillSrc));
1027 unsigned PPCInstrInfo::getStoreOpcodeForSpill(unsigned Reg,
1028 const TargetRegisterClass *RC)
1030 const unsigned *OpcodesForSpill = getStoreOpcodesForSpillArray();
1031 int OpcodeIndex = 0;
1033 if (RC != nullptr) {
1034 if (PPC::GPRCRegClass.hasSubClassEq(RC) ||
1035 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) {
1036 OpcodeIndex = SOK_Int4Spill;
1037 } else if (PPC::G8RCRegClass.hasSubClassEq(RC) ||
1038 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) {
1039 OpcodeIndex = SOK_Int8Spill;
1040 } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) {
1041 OpcodeIndex = SOK_Float8Spill;
1042 } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) {
1043 OpcodeIndex = SOK_Float4Spill;
1044 } else if (PPC::SPERCRegClass.hasSubClassEq(RC)) {
1045 OpcodeIndex = SOK_SPESpill;
1046 } else if (PPC::SPE4RCRegClass.hasSubClassEq(RC)) {
1047 OpcodeIndex = SOK_SPE4Spill;
1048 } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) {
1049 OpcodeIndex = SOK_CRSpill;
1050 } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) {
1051 OpcodeIndex = SOK_CRBitSpill;
1052 } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) {
1053 OpcodeIndex = SOK_VRVectorSpill;
1054 } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) {
1055 OpcodeIndex = SOK_VSXVectorSpill;
1056 } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) {
1057 OpcodeIndex = SOK_VectorFloat8Spill;
1058 } else if (PPC::VSSRCRegClass.hasSubClassEq(RC)) {
1059 OpcodeIndex = SOK_VectorFloat4Spill;
1060 } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) {
1061 OpcodeIndex = SOK_VRSaveSpill;
1062 } else if (PPC::QFRCRegClass.hasSubClassEq(RC)) {
1063 OpcodeIndex = SOK_QuadFloat8Spill;
1064 } else if (PPC::QSRCRegClass.hasSubClassEq(RC)) {
1065 OpcodeIndex = SOK_QuadFloat4Spill;
1066 } else if (PPC::QBRCRegClass.hasSubClassEq(RC)) {
1067 OpcodeIndex = SOK_QuadBitSpill;
1068 } else if (PPC::SPILLTOVSRRCRegClass.hasSubClassEq(RC)) {
1069 OpcodeIndex = SOK_SpillToVSR;
1071 llvm_unreachable("Unknown regclass!");
1074 if (PPC::GPRCRegClass.contains(Reg) ||
1075 PPC::GPRC_NOR0RegClass.contains(Reg)) {
1076 OpcodeIndex = SOK_Int4Spill;
1077 } else if (PPC::G8RCRegClass.contains(Reg) ||
1078 PPC::G8RC_NOX0RegClass.contains(Reg)) {
1079 OpcodeIndex = SOK_Int8Spill;
1080 } else if (PPC::F8RCRegClass.contains(Reg)) {
1081 OpcodeIndex = SOK_Float8Spill;
1082 } else if (PPC::F4RCRegClass.contains(Reg)) {
1083 OpcodeIndex = SOK_Float4Spill;
1084 } else if (PPC::SPERCRegClass.contains(Reg)) {
1085 OpcodeIndex = SOK_SPESpill;
1086 } else if (PPC::SPE4RCRegClass.contains(Reg)) {
1087 OpcodeIndex = SOK_SPE4Spill;
1088 } else if (PPC::CRRCRegClass.contains(Reg)) {
1089 OpcodeIndex = SOK_CRSpill;
1090 } else if (PPC::CRBITRCRegClass.contains(Reg)) {
1091 OpcodeIndex = SOK_CRBitSpill;
1092 } else if (PPC::VRRCRegClass.contains(Reg)) {
1093 OpcodeIndex = SOK_VRVectorSpill;
1094 } else if (PPC::VSRCRegClass.contains(Reg)) {
1095 OpcodeIndex = SOK_VSXVectorSpill;
1096 } else if (PPC::VSFRCRegClass.contains(Reg)) {
1097 OpcodeIndex = SOK_VectorFloat8Spill;
1098 } else if (PPC::VSSRCRegClass.contains(Reg)) {
1099 OpcodeIndex = SOK_VectorFloat4Spill;
1100 } else if (PPC::VRSAVERCRegClass.contains(Reg)) {
1101 OpcodeIndex = SOK_VRSaveSpill;
1102 } else if (PPC::QFRCRegClass.contains(Reg)) {
1103 OpcodeIndex = SOK_QuadFloat8Spill;
1104 } else if (PPC::QSRCRegClass.contains(Reg)) {
1105 OpcodeIndex = SOK_QuadFloat4Spill;
1106 } else if (PPC::QBRCRegClass.contains(Reg)) {
1107 OpcodeIndex = SOK_QuadBitSpill;
1108 } else if (PPC::SPILLTOVSRRCRegClass.contains(Reg)) {
1109 OpcodeIndex = SOK_SpillToVSR;
1111 llvm_unreachable("Unknown regclass!");
1114 return OpcodesForSpill[OpcodeIndex];
1118 PPCInstrInfo::getLoadOpcodeForSpill(unsigned Reg,
1119 const TargetRegisterClass *RC) const {
1120 const unsigned *OpcodesForSpill = getLoadOpcodesForSpillArray();
1121 int OpcodeIndex = 0;
1123 if (RC != nullptr) {
1124 if (PPC::GPRCRegClass.hasSubClassEq(RC) ||
1125 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) {
1126 OpcodeIndex = SOK_Int4Spill;
1127 } else if (PPC::G8RCRegClass.hasSubClassEq(RC) ||
1128 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) {
1129 OpcodeIndex = SOK_Int8Spill;
1130 } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) {
1131 OpcodeIndex = SOK_Float8Spill;
1132 } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) {
1133 OpcodeIndex = SOK_Float4Spill;
1134 } else if (PPC::SPERCRegClass.hasSubClassEq(RC)) {
1135 OpcodeIndex = SOK_SPESpill;
1136 } else if (PPC::SPE4RCRegClass.hasSubClassEq(RC)) {
1137 OpcodeIndex = SOK_SPE4Spill;
1138 } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) {
1139 OpcodeIndex = SOK_CRSpill;
1140 } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) {
1141 OpcodeIndex = SOK_CRBitSpill;
1142 } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) {
1143 OpcodeIndex = SOK_VRVectorSpill;
1144 } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) {
1145 OpcodeIndex = SOK_VSXVectorSpill;
1146 } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) {
1147 OpcodeIndex = SOK_VectorFloat8Spill;
1148 } else if (PPC::VSSRCRegClass.hasSubClassEq(RC)) {
1149 OpcodeIndex = SOK_VectorFloat4Spill;
1150 } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) {
1151 OpcodeIndex = SOK_VRSaveSpill;
1152 } else if (PPC::QFRCRegClass.hasSubClassEq(RC)) {
1153 OpcodeIndex = SOK_QuadFloat8Spill;
1154 } else if (PPC::QSRCRegClass.hasSubClassEq(RC)) {
1155 OpcodeIndex = SOK_QuadFloat4Spill;
1156 } else if (PPC::QBRCRegClass.hasSubClassEq(RC)) {
1157 OpcodeIndex = SOK_QuadBitSpill;
1158 } else if (PPC::SPILLTOVSRRCRegClass.hasSubClassEq(RC)) {
1159 OpcodeIndex = SOK_SpillToVSR;
1161 llvm_unreachable("Unknown regclass!");
1164 if (PPC::GPRCRegClass.contains(Reg) ||
1165 PPC::GPRC_NOR0RegClass.contains(Reg)) {
1166 OpcodeIndex = SOK_Int4Spill;
1167 } else if (PPC::G8RCRegClass.contains(Reg) ||
1168 PPC::G8RC_NOX0RegClass.contains(Reg)) {
1169 OpcodeIndex = SOK_Int8Spill;
1170 } else if (PPC::F8RCRegClass.contains(Reg)) {
1171 OpcodeIndex = SOK_Float8Spill;
1172 } else if (PPC::F4RCRegClass.contains(Reg)) {
1173 OpcodeIndex = SOK_Float4Spill;
1174 } else if (PPC::SPERCRegClass.contains(Reg)) {
1175 OpcodeIndex = SOK_SPESpill;
1176 } else if (PPC::SPE4RCRegClass.contains(Reg)) {
1177 OpcodeIndex = SOK_SPE4Spill;
1178 } else if (PPC::CRRCRegClass.contains(Reg)) {
1179 OpcodeIndex = SOK_CRSpill;
1180 } else if (PPC::CRBITRCRegClass.contains(Reg)) {
1181 OpcodeIndex = SOK_CRBitSpill;
1182 } else if (PPC::VRRCRegClass.contains(Reg)) {
1183 OpcodeIndex = SOK_VRVectorSpill;
1184 } else if (PPC::VSRCRegClass.contains(Reg)) {
1185 OpcodeIndex = SOK_VSXVectorSpill;
1186 } else if (PPC::VSFRCRegClass.contains(Reg)) {
1187 OpcodeIndex = SOK_VectorFloat8Spill;
1188 } else if (PPC::VSSRCRegClass.contains(Reg)) {
1189 OpcodeIndex = SOK_VectorFloat4Spill;
1190 } else if (PPC::VRSAVERCRegClass.contains(Reg)) {
1191 OpcodeIndex = SOK_VRSaveSpill;
1192 } else if (PPC::QFRCRegClass.contains(Reg)) {
1193 OpcodeIndex = SOK_QuadFloat8Spill;
1194 } else if (PPC::QSRCRegClass.contains(Reg)) {
1195 OpcodeIndex = SOK_QuadFloat4Spill;
1196 } else if (PPC::QBRCRegClass.contains(Reg)) {
1197 OpcodeIndex = SOK_QuadBitSpill;
1198 } else if (PPC::SPILLTOVSRRCRegClass.contains(Reg)) {
1199 OpcodeIndex = SOK_SpillToVSR;
1201 llvm_unreachable("Unknown regclass!");
1204 return OpcodesForSpill[OpcodeIndex];
1207 void PPCInstrInfo::StoreRegToStackSlot(
1208 MachineFunction &MF, unsigned SrcReg, bool isKill, int FrameIdx,
1209 const TargetRegisterClass *RC,
1210 SmallVectorImpl<MachineInstr *> &NewMIs) const {
1211 unsigned Opcode = getStoreOpcodeForSpill(PPC::NoRegister, RC);
1214 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1215 FuncInfo->setHasSpills();
1217 NewMIs.push_back(addFrameReference(
1218 BuildMI(MF, DL, get(Opcode)).addReg(SrcReg, getKillRegState(isKill)),
1221 if (PPC::CRRCRegClass.hasSubClassEq(RC) ||
1222 PPC::CRBITRCRegClass.hasSubClassEq(RC))
1223 FuncInfo->setSpillsCR();
1225 if (PPC::VRSAVERCRegClass.hasSubClassEq(RC))
1226 FuncInfo->setSpillsVRSAVE();
1228 if (isXFormMemOp(Opcode))
1229 FuncInfo->setHasNonRISpills();
1232 void PPCInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
1233 MachineBasicBlock::iterator MI,
1234 unsigned SrcReg, bool isKill,
1236 const TargetRegisterClass *RC,
1237 const TargetRegisterInfo *TRI) const {
1238 MachineFunction &MF = *MBB.getParent();
1239 SmallVector<MachineInstr *, 4> NewMIs;
1241 // We need to avoid a situation in which the value from a VRRC register is
1242 // spilled using an Altivec instruction and reloaded into a VSRC register
1243 // using a VSX instruction. The issue with this is that the VSX
1244 // load/store instructions swap the doublewords in the vector and the Altivec
1245 // ones don't. The register classes on the spill/reload may be different if
1246 // the register is defined using an Altivec instruction and is then used by a
1250 StoreRegToStackSlot(MF, SrcReg, isKill, FrameIdx, RC, NewMIs);
1252 for (unsigned i = 0, e = NewMIs.size(); i != e; ++i)
1253 MBB.insert(MI, NewMIs[i]);
1255 const MachineFrameInfo &MFI = MF.getFrameInfo();
1256 MachineMemOperand *MMO = MF.getMachineMemOperand(
1257 MachinePointerInfo::getFixedStack(MF, FrameIdx),
1258 MachineMemOperand::MOStore, MFI.getObjectSize(FrameIdx),
1259 MFI.getObjectAlignment(FrameIdx));
1260 NewMIs.back()->addMemOperand(MF, MMO);
1263 void PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, const DebugLoc &DL,
1264 unsigned DestReg, int FrameIdx,
1265 const TargetRegisterClass *RC,
1266 SmallVectorImpl<MachineInstr *> &NewMIs)
1268 unsigned Opcode = getLoadOpcodeForSpill(PPC::NoRegister, RC);
1269 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(Opcode), DestReg),
1271 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1273 if (PPC::CRRCRegClass.hasSubClassEq(RC) ||
1274 PPC::CRBITRCRegClass.hasSubClassEq(RC))
1275 FuncInfo->setSpillsCR();
1277 if (PPC::VRSAVERCRegClass.hasSubClassEq(RC))
1278 FuncInfo->setSpillsVRSAVE();
1280 if (isXFormMemOp(Opcode))
1281 FuncInfo->setHasNonRISpills();
1285 PPCInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1286 MachineBasicBlock::iterator MI,
1287 unsigned DestReg, int FrameIdx,
1288 const TargetRegisterClass *RC,
1289 const TargetRegisterInfo *TRI) const {
1290 MachineFunction &MF = *MBB.getParent();
1291 SmallVector<MachineInstr*, 4> NewMIs;
1293 if (MI != MBB.end()) DL = MI->getDebugLoc();
1295 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1296 FuncInfo->setHasSpills();
1298 // We need to avoid a situation in which the value from a VRRC register is
1299 // spilled using an Altivec instruction and reloaded into a VSRC register
1300 // using a VSX instruction. The issue with this is that the VSX
1301 // load/store instructions swap the doublewords in the vector and the Altivec
1302 // ones don't. The register classes on the spill/reload may be different if
1303 // the register is defined using an Altivec instruction and is then used by a
1305 if (Subtarget.hasVSX() && RC == &PPC::VRRCRegClass)
1306 RC = &PPC::VSRCRegClass;
1308 LoadRegFromStackSlot(MF, DL, DestReg, FrameIdx, RC, NewMIs);
1310 for (unsigned i = 0, e = NewMIs.size(); i != e; ++i)
1311 MBB.insert(MI, NewMIs[i]);
1313 const MachineFrameInfo &MFI = MF.getFrameInfo();
1314 MachineMemOperand *MMO = MF.getMachineMemOperand(
1315 MachinePointerInfo::getFixedStack(MF, FrameIdx),
1316 MachineMemOperand::MOLoad, MFI.getObjectSize(FrameIdx),
1317 MFI.getObjectAlignment(FrameIdx));
1318 NewMIs.back()->addMemOperand(MF, MMO);
1322 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
1323 assert(Cond.size() == 2 && "Invalid PPC branch opcode!");
1324 if (Cond[1].getReg() == PPC::CTR8 || Cond[1].getReg() == PPC::CTR)
1325 Cond[0].setImm(Cond[0].getImm() == 0 ? 1 : 0);
1327 // Leave the CR# the same, but invert the condition.
1328 Cond[0].setImm(PPC::InvertPredicate((PPC::Predicate)Cond[0].getImm()));
1332 bool PPCInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
1333 unsigned Reg, MachineRegisterInfo *MRI) const {
1334 // For some instructions, it is legal to fold ZERO into the RA register field.
1335 // A zero immediate should always be loaded with a single li.
1336 unsigned DefOpc = DefMI.getOpcode();
1337 if (DefOpc != PPC::LI && DefOpc != PPC::LI8)
1339 if (!DefMI.getOperand(1).isImm())
1341 if (DefMI.getOperand(1).getImm() != 0)
1344 // Note that we cannot here invert the arguments of an isel in order to fold
1345 // a ZERO into what is presented as the second argument. All we have here
1346 // is the condition bit, and that might come from a CR-logical bit operation.
1348 const MCInstrDesc &UseMCID = UseMI.getDesc();
1350 // Only fold into real machine instructions.
1351 if (UseMCID.isPseudo())
1355 for (UseIdx = 0; UseIdx < UseMI.getNumOperands(); ++UseIdx)
1356 if (UseMI.getOperand(UseIdx).isReg() &&
1357 UseMI.getOperand(UseIdx).getReg() == Reg)
1360 assert(UseIdx < UseMI.getNumOperands() && "Cannot find Reg in UseMI");
1361 assert(UseIdx < UseMCID.getNumOperands() && "No operand description for Reg");
1363 const MCOperandInfo *UseInfo = &UseMCID.OpInfo[UseIdx];
1365 // We can fold the zero if this register requires a GPRC_NOR0/G8RC_NOX0
1366 // register (which might also be specified as a pointer class kind).
1367 if (UseInfo->isLookupPtrRegClass()) {
1368 if (UseInfo->RegClass /* Kind */ != 1)
1371 if (UseInfo->RegClass != PPC::GPRC_NOR0RegClassID &&
1372 UseInfo->RegClass != PPC::G8RC_NOX0RegClassID)
1376 // Make sure this is not tied to an output register (or otherwise
1377 // constrained). This is true for ST?UX registers, for example, which
1378 // are tied to their output registers.
1379 if (UseInfo->Constraints != 0)
1383 if (UseInfo->isLookupPtrRegClass()) {
1384 bool isPPC64 = Subtarget.isPPC64();
1385 ZeroReg = isPPC64 ? PPC::ZERO8 : PPC::ZERO;
1387 ZeroReg = UseInfo->RegClass == PPC::G8RC_NOX0RegClassID ?
1388 PPC::ZERO8 : PPC::ZERO;
1391 bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
1392 UseMI.getOperand(UseIdx).setReg(ZeroReg);
1395 DefMI.eraseFromParent();
1400 static bool MBBDefinesCTR(MachineBasicBlock &MBB) {
1401 for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end();
1403 if (I->definesRegister(PPC::CTR) || I->definesRegister(PPC::CTR8))
1408 // We should make sure that, if we're going to predicate both sides of a
1409 // condition (a diamond), that both sides don't define the counter register. We
1410 // can predicate counter-decrement-based branches, but while that predicates
1411 // the branching, it does not predicate the counter decrement. If we tried to
1412 // merge the triangle into one predicated block, we'd decrement the counter
1414 bool PPCInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
1415 unsigned NumT, unsigned ExtraT,
1416 MachineBasicBlock &FMBB,
1417 unsigned NumF, unsigned ExtraF,
1418 BranchProbability Probability) const {
1419 return !(MBBDefinesCTR(TMBB) && MBBDefinesCTR(FMBB));
1423 bool PPCInstrInfo::isPredicated(const MachineInstr &MI) const {
1424 // The predicated branches are identified by their type, not really by the
1425 // explicit presence of a predicate. Furthermore, some of them can be
1426 // predicated more than once. Because if conversion won't try to predicate
1427 // any instruction which already claims to be predicated (by returning true
1428 // here), always return false. In doing so, we let isPredicable() be the
1429 // final word on whether not the instruction can be (further) predicated.
1434 bool PPCInstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const {
1435 if (!MI.isTerminator())
1438 // Conditional branch is a special case.
1439 if (MI.isBranch() && !MI.isBarrier())
1442 return !isPredicated(MI);
1445 bool PPCInstrInfo::PredicateInstruction(MachineInstr &MI,
1446 ArrayRef<MachineOperand> Pred) const {
1447 unsigned OpC = MI.getOpcode();
1448 if (OpC == PPC::BLR || OpC == PPC::BLR8) {
1449 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) {
1450 bool isPPC64 = Subtarget.isPPC64();
1451 MI.setDesc(get(Pred[0].getImm() ? (isPPC64 ? PPC::BDNZLR8 : PPC::BDNZLR)
1452 : (isPPC64 ? PPC::BDZLR8 : PPC::BDZLR)));
1453 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
1454 MI.setDesc(get(PPC::BCLR));
1455 MachineInstrBuilder(*MI.getParent()->getParent(), MI).add(Pred[1]);
1456 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
1457 MI.setDesc(get(PPC::BCLRn));
1458 MachineInstrBuilder(*MI.getParent()->getParent(), MI).add(Pred[1]);
1460 MI.setDesc(get(PPC::BCCLR));
1461 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1462 .addImm(Pred[0].getImm())
1467 } else if (OpC == PPC::B) {
1468 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) {
1469 bool isPPC64 = Subtarget.isPPC64();
1470 MI.setDesc(get(Pred[0].getImm() ? (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ)
1471 : (isPPC64 ? PPC::BDZ8 : PPC::BDZ)));
1472 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
1473 MachineBasicBlock *MBB = MI.getOperand(0).getMBB();
1474 MI.RemoveOperand(0);
1476 MI.setDesc(get(PPC::BC));
1477 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1480 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
1481 MachineBasicBlock *MBB = MI.getOperand(0).getMBB();
1482 MI.RemoveOperand(0);
1484 MI.setDesc(get(PPC::BCn));
1485 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1489 MachineBasicBlock *MBB = MI.getOperand(0).getMBB();
1490 MI.RemoveOperand(0);
1492 MI.setDesc(get(PPC::BCC));
1493 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1494 .addImm(Pred[0].getImm())
1500 } else if (OpC == PPC::BCTR || OpC == PPC::BCTR8 || OpC == PPC::BCTRL ||
1501 OpC == PPC::BCTRL8) {
1502 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR)
1503 llvm_unreachable("Cannot predicate bctr[l] on the ctr register");
1505 bool setLR = OpC == PPC::BCTRL || OpC == PPC::BCTRL8;
1506 bool isPPC64 = Subtarget.isPPC64();
1508 if (Pred[0].getImm() == PPC::PRED_BIT_SET) {
1509 MI.setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8 : PPC::BCCTR8)
1510 : (setLR ? PPC::BCCTRL : PPC::BCCTR)));
1511 MachineInstrBuilder(*MI.getParent()->getParent(), MI).add(Pred[1]);
1513 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) {
1514 MI.setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8n : PPC::BCCTR8n)
1515 : (setLR ? PPC::BCCTRLn : PPC::BCCTRn)));
1516 MachineInstrBuilder(*MI.getParent()->getParent(), MI).add(Pred[1]);
1520 MI.setDesc(get(isPPC64 ? (setLR ? PPC::BCCCTRL8 : PPC::BCCCTR8)
1521 : (setLR ? PPC::BCCCTRL : PPC::BCCCTR)));
1522 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
1523 .addImm(Pred[0].getImm())
1531 bool PPCInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
1532 ArrayRef<MachineOperand> Pred2) const {
1533 assert(Pred1.size() == 2 && "Invalid PPC first predicate");
1534 assert(Pred2.size() == 2 && "Invalid PPC second predicate");
1536 if (Pred1[1].getReg() == PPC::CTR8 || Pred1[1].getReg() == PPC::CTR)
1538 if (Pred2[1].getReg() == PPC::CTR8 || Pred2[1].getReg() == PPC::CTR)
1541 // P1 can only subsume P2 if they test the same condition register.
1542 if (Pred1[1].getReg() != Pred2[1].getReg())
1545 PPC::Predicate P1 = (PPC::Predicate) Pred1[0].getImm();
1546 PPC::Predicate P2 = (PPC::Predicate) Pred2[0].getImm();
1551 // Does P1 subsume P2, e.g. GE subsumes GT.
1552 if (P1 == PPC::PRED_LE &&
1553 (P2 == PPC::PRED_LT || P2 == PPC::PRED_EQ))
1555 if (P1 == PPC::PRED_GE &&
1556 (P2 == PPC::PRED_GT || P2 == PPC::PRED_EQ))
1562 bool PPCInstrInfo::DefinesPredicate(MachineInstr &MI,
1563 std::vector<MachineOperand> &Pred) const {
1564 // Note: At the present time, the contents of Pred from this function is
1565 // unused by IfConversion. This implementation follows ARM by pushing the
1566 // CR-defining operand. Because the 'DZ' and 'DNZ' count as types of
1567 // predicate, instructions defining CTR or CTR8 are also included as
1568 // predicate-defining instructions.
1570 const TargetRegisterClass *RCs[] =
1571 { &PPC::CRRCRegClass, &PPC::CRBITRCRegClass,
1572 &PPC::CTRRCRegClass, &PPC::CTRRC8RegClass };
1575 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1576 const MachineOperand &MO = MI.getOperand(i);
1577 for (unsigned c = 0; c < array_lengthof(RCs) && !Found; ++c) {
1578 const TargetRegisterClass *RC = RCs[c];
1580 if (MO.isDef() && RC->contains(MO.getReg())) {
1584 } else if (MO.isRegMask()) {
1585 for (TargetRegisterClass::iterator I = RC->begin(),
1586 IE = RC->end(); I != IE; ++I)
1587 if (MO.clobbersPhysReg(*I)) {
1598 bool PPCInstrInfo::isPredicable(const MachineInstr &MI) const {
1599 unsigned OpC = MI.getOpcode();
1614 bool PPCInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
1615 unsigned &SrcReg2, int &Mask,
1617 unsigned Opc = MI.getOpcode();
1620 default: return false;
1625 SrcReg = MI.getOperand(1).getReg();
1627 Value = MI.getOperand(2).getImm();
1636 SrcReg = MI.getOperand(1).getReg();
1637 SrcReg2 = MI.getOperand(2).getReg();
1644 bool PPCInstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
1645 unsigned SrcReg2, int Mask, int Value,
1646 const MachineRegisterInfo *MRI) const {
1650 int OpC = CmpInstr.getOpcode();
1651 unsigned CRReg = CmpInstr.getOperand(0).getReg();
1653 // FP record forms set CR1 based on the exception status bits, not a
1654 // comparison with zero.
1655 if (OpC == PPC::FCMPUS || OpC == PPC::FCMPUD)
1658 const TargetRegisterInfo *TRI = &getRegisterInfo();
1659 // The record forms set the condition register based on a signed comparison
1660 // with zero (so says the ISA manual). This is not as straightforward as it
1661 // seems, however, because this is always a 64-bit comparison on PPC64, even
1662 // for instructions that are 32-bit in nature (like slw for example).
1663 // So, on PPC32, for unsigned comparisons, we can use the record forms only
1664 // for equality checks (as those don't depend on the sign). On PPC64,
1665 // we are restricted to equality for unsigned 64-bit comparisons and for
1666 // signed 32-bit comparisons the applicability is more restricted.
1667 bool isPPC64 = Subtarget.isPPC64();
1668 bool is32BitSignedCompare = OpC == PPC::CMPWI || OpC == PPC::CMPW;
1669 bool is32BitUnsignedCompare = OpC == PPC::CMPLWI || OpC == PPC::CMPLW;
1670 bool is64BitUnsignedCompare = OpC == PPC::CMPLDI || OpC == PPC::CMPLD;
1672 // Look through copies unless that gets us to a physical register.
1673 unsigned ActualSrc = TRI->lookThruCopyLike(SrcReg, MRI);
1674 if (TargetRegisterInfo::isVirtualRegister(ActualSrc))
1677 // Get the unique definition of SrcReg.
1678 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
1679 if (!MI) return false;
1681 bool equalityOnly = false;
1684 if (is32BitSignedCompare) {
1685 // We can perform this optimization only if MI is sign-extending.
1686 if (isSignExtended(*MI))
1690 } else if (is32BitUnsignedCompare) {
1691 // We can perform this optimization, equality only, if MI is
1693 if (isZeroExtended(*MI)) {
1695 equalityOnly = true;
1699 equalityOnly = is64BitUnsignedCompare;
1701 equalityOnly = is32BitUnsignedCompare;
1704 // We need to check the uses of the condition register in order to reject
1705 // non-equality comparisons.
1706 for (MachineRegisterInfo::use_instr_iterator
1707 I = MRI->use_instr_begin(CRReg), IE = MRI->use_instr_end();
1709 MachineInstr *UseMI = &*I;
1710 if (UseMI->getOpcode() == PPC::BCC) {
1711 PPC::Predicate Pred = (PPC::Predicate)UseMI->getOperand(0).getImm();
1712 unsigned PredCond = PPC::getPredicateCondition(Pred);
1713 // We ignore hint bits when checking for non-equality comparisons.
1714 if (PredCond != PPC::PRED_EQ && PredCond != PPC::PRED_NE)
1716 } else if (UseMI->getOpcode() == PPC::ISEL ||
1717 UseMI->getOpcode() == PPC::ISEL8) {
1718 unsigned SubIdx = UseMI->getOperand(3).getSubReg();
1719 if (SubIdx != PPC::sub_eq)
1726 MachineBasicBlock::iterator I = CmpInstr;
1728 // Scan forward to find the first use of the compare.
1729 for (MachineBasicBlock::iterator EL = CmpInstr.getParent()->end(); I != EL;
1731 bool FoundUse = false;
1732 for (MachineRegisterInfo::use_instr_iterator
1733 J = MRI->use_instr_begin(CRReg), JE = MRI->use_instr_end();
1744 SmallVector<std::pair<MachineOperand*, PPC::Predicate>, 4> PredsToUpdate;
1745 SmallVector<std::pair<MachineOperand*, unsigned>, 4> SubRegsToUpdate;
1747 // There are two possible candidates which can be changed to set CR[01].
1748 // One is MI, the other is a SUB instruction.
1749 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
1750 MachineInstr *Sub = nullptr;
1752 // MI is not a candidate for CMPrr.
1754 // FIXME: Conservatively refuse to convert an instruction which isn't in the
1755 // same BB as the comparison. This is to allow the check below to avoid calls
1756 // (and other explicit clobbers); instead we should really check for these
1757 // more explicitly (in at least a few predecessors).
1758 else if (MI->getParent() != CmpInstr.getParent())
1760 else if (Value != 0) {
1761 // The record-form instructions set CR bit based on signed comparison
1762 // against 0. We try to convert a compare against 1 or -1 into a compare
1763 // against 0 to exploit record-form instructions. For example, we change
1764 // the condition "greater than -1" into "greater than or equal to 0"
1765 // and "less than 1" into "less than or equal to 0".
1767 // Since we optimize comparison based on a specific branch condition,
1768 // we don't optimize if condition code is used by more than once.
1769 if (equalityOnly || !MRI->hasOneUse(CRReg))
1772 MachineInstr *UseMI = &*MRI->use_instr_begin(CRReg);
1773 if (UseMI->getOpcode() != PPC::BCC)
1776 PPC::Predicate Pred = (PPC::Predicate)UseMI->getOperand(0).getImm();
1777 unsigned PredCond = PPC::getPredicateCondition(Pred);
1778 unsigned PredHint = PPC::getPredicateHint(Pred);
1779 int16_t Immed = (int16_t)Value;
1781 // When modifying the condition in the predicate, we propagate hint bits
1782 // from the original predicate to the new one.
1783 if (Immed == -1 && PredCond == PPC::PRED_GT)
1784 // We convert "greater than -1" into "greater than or equal to 0",
1785 // since we are assuming signed comparison by !equalityOnly
1786 Pred = PPC::getPredicate(PPC::PRED_GE, PredHint);
1787 else if (Immed == -1 && PredCond == PPC::PRED_LE)
1788 // We convert "less than or equal to -1" into "less than 0".
1789 Pred = PPC::getPredicate(PPC::PRED_LT, PredHint);
1790 else if (Immed == 1 && PredCond == PPC::PRED_LT)
1791 // We convert "less than 1" into "less than or equal to 0".
1792 Pred = PPC::getPredicate(PPC::PRED_LE, PredHint);
1793 else if (Immed == 1 && PredCond == PPC::PRED_GE)
1794 // We convert "greater than or equal to 1" into "greater than 0".
1795 Pred = PPC::getPredicate(PPC::PRED_GT, PredHint);
1799 PredsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(0)), Pred));
1805 // Get ready to iterate backward from CmpInstr.
1806 MachineBasicBlock::iterator E = MI, B = CmpInstr.getParent()->begin();
1808 for (; I != E && !noSub; --I) {
1809 const MachineInstr &Instr = *I;
1810 unsigned IOpC = Instr.getOpcode();
1812 if (&*I != &CmpInstr && (Instr.modifiesRegister(PPC::CR0, TRI) ||
1813 Instr.readsRegister(PPC::CR0, TRI)))
1814 // This instruction modifies or uses the record condition register after
1815 // the one we want to change. While we could do this transformation, it
1816 // would likely not be profitable. This transformation removes one
1817 // instruction, and so even forcing RA to generate one move probably
1818 // makes it unprofitable.
1821 // Check whether CmpInstr can be made redundant by the current instruction.
1822 if ((OpC == PPC::CMPW || OpC == PPC::CMPLW ||
1823 OpC == PPC::CMPD || OpC == PPC::CMPLD) &&
1824 (IOpC == PPC::SUBF || IOpC == PPC::SUBF8) &&
1825 ((Instr.getOperand(1).getReg() == SrcReg &&
1826 Instr.getOperand(2).getReg() == SrcReg2) ||
1827 (Instr.getOperand(1).getReg() == SrcReg2 &&
1828 Instr.getOperand(2).getReg() == SrcReg))) {
1834 // The 'and' is below the comparison instruction.
1838 // Return false if no candidates exist.
1842 // The single candidate is called MI.
1846 int MIOpC = MI->getOpcode();
1847 if (MIOpC == PPC::ANDIo || MIOpC == PPC::ANDIo8 ||
1848 MIOpC == PPC::ANDISo || MIOpC == PPC::ANDISo8)
1851 NewOpC = PPC::getRecordFormOpcode(MIOpC);
1852 if (NewOpC == -1 && PPC::getNonRecordFormOpcode(MIOpC) != -1)
1856 // FIXME: On the non-embedded POWER architectures, only some of the record
1857 // forms are fast, and we should use only the fast ones.
1859 // The defining instruction has a record form (or is already a record
1860 // form). It is possible, however, that we'll need to reverse the condition
1861 // code of the users.
1865 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based on CMP
1866 // needs to be updated to be based on SUB. Push the condition code
1867 // operands to OperandsToUpdate. If it is safe to remove CmpInstr, the
1868 // condition code of these operands will be modified.
1869 // Here, Value == 0 means we haven't converted comparison against 1 or -1 to
1870 // comparison against 0, which may modify predicate.
1871 bool ShouldSwap = false;
1872 if (Sub && Value == 0) {
1873 ShouldSwap = SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
1874 Sub->getOperand(2).getReg() == SrcReg;
1876 // The operands to subf are the opposite of sub, so only in the fixed-point
1877 // case, invert the order.
1878 ShouldSwap = !ShouldSwap;
1882 for (MachineRegisterInfo::use_instr_iterator
1883 I = MRI->use_instr_begin(CRReg), IE = MRI->use_instr_end();
1885 MachineInstr *UseMI = &*I;
1886 if (UseMI->getOpcode() == PPC::BCC) {
1887 PPC::Predicate Pred = (PPC::Predicate) UseMI->getOperand(0).getImm();
1888 unsigned PredCond = PPC::getPredicateCondition(Pred);
1889 assert((!equalityOnly ||
1890 PredCond == PPC::PRED_EQ || PredCond == PPC::PRED_NE) &&
1891 "Invalid predicate for equality-only optimization");
1892 (void)PredCond; // To suppress warning in release build.
1893 PredsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(0)),
1894 PPC::getSwappedPredicate(Pred)));
1895 } else if (UseMI->getOpcode() == PPC::ISEL ||
1896 UseMI->getOpcode() == PPC::ISEL8) {
1897 unsigned NewSubReg = UseMI->getOperand(3).getSubReg();
1898 assert((!equalityOnly || NewSubReg == PPC::sub_eq) &&
1899 "Invalid CR bit for equality-only optimization");
1901 if (NewSubReg == PPC::sub_lt)
1902 NewSubReg = PPC::sub_gt;
1903 else if (NewSubReg == PPC::sub_gt)
1904 NewSubReg = PPC::sub_lt;
1906 SubRegsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(3)),
1908 } else // We need to abort on a user we don't understand.
1911 assert(!(Value != 0 && ShouldSwap) &&
1912 "Non-zero immediate support and ShouldSwap"
1913 "may conflict in updating predicate");
1915 // Create a new virtual register to hold the value of the CR set by the
1916 // record-form instruction. If the instruction was not previously in
1917 // record form, then set the kill flag on the CR.
1918 CmpInstr.eraseFromParent();
1920 MachineBasicBlock::iterator MII = MI;
1921 BuildMI(*MI->getParent(), std::next(MII), MI->getDebugLoc(),
1922 get(TargetOpcode::COPY), CRReg)
1923 .addReg(PPC::CR0, MIOpC != NewOpC ? RegState::Kill : 0);
1925 // Even if CR0 register were dead before, it is alive now since the
1926 // instruction we just built uses it.
1927 MI->clearRegisterDeads(PPC::CR0);
1929 if (MIOpC != NewOpC) {
1930 // We need to be careful here: we're replacing one instruction with
1931 // another, and we need to make sure that we get all of the right
1932 // implicit uses and defs. On the other hand, the caller may be holding
1933 // an iterator to this instruction, and so we can't delete it (this is
1934 // specifically the case if this is the instruction directly after the
1937 // Rotates are expensive instructions. If we're emitting a record-form
1938 // rotate that can just be an andi/andis, we should just emit that.
1939 if (MIOpC == PPC::RLWINM || MIOpC == PPC::RLWINM8) {
1940 unsigned GPRRes = MI->getOperand(0).getReg();
1941 int64_t SH = MI->getOperand(2).getImm();
1942 int64_t MB = MI->getOperand(3).getImm();
1943 int64_t ME = MI->getOperand(4).getImm();
1944 // We can only do this if both the start and end of the mask are in the
1946 bool MBInLoHWord = MB >= 16;
1947 bool MEInLoHWord = ME >= 16;
1948 uint64_t Mask = ~0LLU;
1950 if (MB <= ME && MBInLoHWord == MEInLoHWord && SH == 0) {
1951 Mask = ((1LLU << (32 - MB)) - 1) & ~((1LLU << (31 - ME)) - 1);
1952 // The mask value needs to shift right 16 if we're emitting andis.
1953 Mask >>= MBInLoHWord ? 0 : 16;
1954 NewOpC = MIOpC == PPC::RLWINM ?
1955 (MBInLoHWord ? PPC::ANDIo : PPC::ANDISo) :
1956 (MBInLoHWord ? PPC::ANDIo8 :PPC::ANDISo8);
1957 } else if (MRI->use_empty(GPRRes) && (ME == 31) &&
1958 (ME - MB + 1 == SH) && (MB >= 16)) {
1959 // If we are rotating by the exact number of bits as are in the mask
1960 // and the mask is in the least significant bits of the register,
1961 // that's just an andis. (as long as the GPR result has no uses).
1962 Mask = ((1LLU << 32) - 1) & ~((1LLU << (32 - SH)) - 1);
1964 NewOpC = MIOpC == PPC::RLWINM ? PPC::ANDISo :PPC::ANDISo8;
1966 // If we've set the mask, we can transform.
1967 if (Mask != ~0LLU) {
1968 MI->RemoveOperand(4);
1969 MI->RemoveOperand(3);
1970 MI->getOperand(2).setImm(Mask);
1971 NumRcRotatesConvertedToRcAnd++;
1973 } else if (MIOpC == PPC::RLDICL && MI->getOperand(2).getImm() == 0) {
1974 int64_t MB = MI->getOperand(3).getImm();
1976 uint64_t Mask = (1LLU << (63 - MB + 1)) - 1;
1977 NewOpC = PPC::ANDIo8;
1978 MI->RemoveOperand(3);
1979 MI->getOperand(2).setImm(Mask);
1980 NumRcRotatesConvertedToRcAnd++;
1984 const MCInstrDesc &NewDesc = get(NewOpC);
1985 MI->setDesc(NewDesc);
1987 if (NewDesc.ImplicitDefs)
1988 for (const MCPhysReg *ImpDefs = NewDesc.getImplicitDefs();
1989 *ImpDefs; ++ImpDefs)
1990 if (!MI->definesRegister(*ImpDefs))
1991 MI->addOperand(*MI->getParent()->getParent(),
1992 MachineOperand::CreateReg(*ImpDefs, true, true));
1993 if (NewDesc.ImplicitUses)
1994 for (const MCPhysReg *ImpUses = NewDesc.getImplicitUses();
1995 *ImpUses; ++ImpUses)
1996 if (!MI->readsRegister(*ImpUses))
1997 MI->addOperand(*MI->getParent()->getParent(),
1998 MachineOperand::CreateReg(*ImpUses, false, true));
2000 assert(MI->definesRegister(PPC::CR0) &&
2001 "Record-form instruction does not define cr0?");
2003 // Modify the condition code of operands in OperandsToUpdate.
2004 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to
2005 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
2006 for (unsigned i = 0, e = PredsToUpdate.size(); i < e; i++)
2007 PredsToUpdate[i].first->setImm(PredsToUpdate[i].second);
2009 for (unsigned i = 0, e = SubRegsToUpdate.size(); i < e; i++)
2010 SubRegsToUpdate[i].first->setSubReg(SubRegsToUpdate[i].second);
2015 /// GetInstSize - Return the number of bytes of code the specified
2016 /// instruction may be. This returns the maximum number of bytes.
2018 unsigned PPCInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
2019 unsigned Opcode = MI.getOpcode();
2021 if (Opcode == PPC::INLINEASM || Opcode == PPC::INLINEASM_BR) {
2022 const MachineFunction *MF = MI.getParent()->getParent();
2023 const char *AsmStr = MI.getOperand(0).getSymbolName();
2024 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
2025 } else if (Opcode == TargetOpcode::STACKMAP) {
2026 StackMapOpers Opers(&MI);
2027 return Opers.getNumPatchBytes();
2028 } else if (Opcode == TargetOpcode::PATCHPOINT) {
2029 PatchPointOpers Opers(&MI);
2030 return Opers.getNumPatchBytes();
2032 return get(Opcode).getSize();
2036 std::pair<unsigned, unsigned>
2037 PPCInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
2038 const unsigned Mask = PPCII::MO_ACCESS_MASK;
2039 return std::make_pair(TF & Mask, TF & ~Mask);
2042 ArrayRef<std::pair<unsigned, const char *>>
2043 PPCInstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
2044 using namespace PPCII;
2045 static const std::pair<unsigned, const char *> TargetFlags[] = {
2048 {MO_TPREL_LO, "ppc-tprel-lo"},
2049 {MO_TPREL_HA, "ppc-tprel-ha"},
2050 {MO_DTPREL_LO, "ppc-dtprel-lo"},
2051 {MO_TLSLD_LO, "ppc-tlsld-lo"},
2052 {MO_TOC_LO, "ppc-toc-lo"},
2053 {MO_TLS, "ppc-tls"}};
2054 return makeArrayRef(TargetFlags);
2057 ArrayRef<std::pair<unsigned, const char *>>
2058 PPCInstrInfo::getSerializableBitmaskMachineOperandTargetFlags() const {
2059 using namespace PPCII;
2060 static const std::pair<unsigned, const char *> TargetFlags[] = {
2061 {MO_PLT, "ppc-plt"},
2062 {MO_PIC_FLAG, "ppc-pic"},
2063 {MO_NLP_FLAG, "ppc-nlp"},
2064 {MO_NLP_HIDDEN_FLAG, "ppc-nlp-hidden"}};
2065 return makeArrayRef(TargetFlags);
2068 // Expand VSX Memory Pseudo instruction to either a VSX or a FP instruction.
2069 // The VSX versions have the advantage of a full 64-register target whereas
2070 // the FP ones have the advantage of lower latency and higher throughput. So
2071 // what we are after is using the faster instructions in low register pressure
2072 // situations and using the larger register file in high register pressure
2074 bool PPCInstrInfo::expandVSXMemPseudo(MachineInstr &MI) const {
2075 unsigned UpperOpcode, LowerOpcode;
2076 switch (MI.getOpcode()) {
2077 case PPC::DFLOADf32:
2078 UpperOpcode = PPC::LXSSP;
2079 LowerOpcode = PPC::LFS;
2081 case PPC::DFLOADf64:
2082 UpperOpcode = PPC::LXSD;
2083 LowerOpcode = PPC::LFD;
2085 case PPC::DFSTOREf32:
2086 UpperOpcode = PPC::STXSSP;
2087 LowerOpcode = PPC::STFS;
2089 case PPC::DFSTOREf64:
2090 UpperOpcode = PPC::STXSD;
2091 LowerOpcode = PPC::STFD;
2093 case PPC::XFLOADf32:
2094 UpperOpcode = PPC::LXSSPX;
2095 LowerOpcode = PPC::LFSX;
2097 case PPC::XFLOADf64:
2098 UpperOpcode = PPC::LXSDX;
2099 LowerOpcode = PPC::LFDX;
2101 case PPC::XFSTOREf32:
2102 UpperOpcode = PPC::STXSSPX;
2103 LowerOpcode = PPC::STFSX;
2105 case PPC::XFSTOREf64:
2106 UpperOpcode = PPC::STXSDX;
2107 LowerOpcode = PPC::STFDX;
2110 UpperOpcode = PPC::LXSIWAX;
2111 LowerOpcode = PPC::LFIWAX;
2114 UpperOpcode = PPC::LXSIWZX;
2115 LowerOpcode = PPC::LFIWZX;
2118 UpperOpcode = PPC::STXSIWX;
2119 LowerOpcode = PPC::STFIWX;
2122 llvm_unreachable("Unknown Operation!");
2125 unsigned TargetReg = MI.getOperand(0).getReg();
2127 if ((TargetReg >= PPC::F0 && TargetReg <= PPC::F31) ||
2128 (TargetReg >= PPC::VSL0 && TargetReg <= PPC::VSL31))
2129 Opcode = LowerOpcode;
2131 Opcode = UpperOpcode;
2132 MI.setDesc(get(Opcode));
2136 static bool isAnImmediateOperand(const MachineOperand &MO) {
2137 return MO.isCPI() || MO.isGlobal() || MO.isImm();
2140 bool PPCInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
2141 auto &MBB = *MI.getParent();
2142 auto DL = MI.getDebugLoc();
2144 switch (MI.getOpcode()) {
2145 case TargetOpcode::LOAD_STACK_GUARD: {
2146 assert(Subtarget.isTargetLinux() &&
2147 "Only Linux target is expected to contain LOAD_STACK_GUARD");
2148 const int64_t Offset = Subtarget.isPPC64() ? -0x7010 : -0x7008;
2149 const unsigned Reg = Subtarget.isPPC64() ? PPC::X13 : PPC::R2;
2150 MI.setDesc(get(Subtarget.isPPC64() ? PPC::LD : PPC::LWZ));
2151 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
2156 case PPC::DFLOADf32:
2157 case PPC::DFLOADf64:
2158 case PPC::DFSTOREf32:
2159 case PPC::DFSTOREf64: {
2160 assert(Subtarget.hasP9Vector() &&
2161 "Invalid D-Form Pseudo-ops on Pre-P9 target.");
2162 assert(MI.getOperand(2).isReg() &&
2163 isAnImmediateOperand(MI.getOperand(1)) &&
2164 "D-form op must have register and immediate operands");
2165 return expandVSXMemPseudo(MI);
2167 case PPC::XFLOADf32:
2168 case PPC::XFSTOREf32:
2172 assert(Subtarget.hasP8Vector() &&
2173 "Invalid X-Form Pseudo-ops on Pre-P8 target.");
2174 assert(MI.getOperand(2).isReg() && MI.getOperand(1).isReg() &&
2175 "X-form op must have register and register operands");
2176 return expandVSXMemPseudo(MI);
2178 case PPC::XFLOADf64:
2179 case PPC::XFSTOREf64: {
2180 assert(Subtarget.hasVSX() &&
2181 "Invalid X-Form Pseudo-ops on target that has no VSX.");
2182 assert(MI.getOperand(2).isReg() && MI.getOperand(1).isReg() &&
2183 "X-form op must have register and register operands");
2184 return expandVSXMemPseudo(MI);
2186 case PPC::SPILLTOVSR_LD: {
2187 unsigned TargetReg = MI.getOperand(0).getReg();
2188 if (PPC::VSFRCRegClass.contains(TargetReg)) {
2189 MI.setDesc(get(PPC::DFLOADf64));
2190 return expandPostRAPseudo(MI);
2193 MI.setDesc(get(PPC::LD));
2196 case PPC::SPILLTOVSR_ST: {
2197 unsigned SrcReg = MI.getOperand(0).getReg();
2198 if (PPC::VSFRCRegClass.contains(SrcReg)) {
2199 NumStoreSPILLVSRRCAsVec++;
2200 MI.setDesc(get(PPC::DFSTOREf64));
2201 return expandPostRAPseudo(MI);
2203 NumStoreSPILLVSRRCAsGpr++;
2204 MI.setDesc(get(PPC::STD));
2208 case PPC::SPILLTOVSR_LDX: {
2209 unsigned TargetReg = MI.getOperand(0).getReg();
2210 if (PPC::VSFRCRegClass.contains(TargetReg))
2211 MI.setDesc(get(PPC::LXSDX));
2213 MI.setDesc(get(PPC::LDX));
2216 case PPC::SPILLTOVSR_STX: {
2217 unsigned SrcReg = MI.getOperand(0).getReg();
2218 if (PPC::VSFRCRegClass.contains(SrcReg)) {
2219 NumStoreSPILLVSRRCAsVec++;
2220 MI.setDesc(get(PPC::STXSDX));
2222 NumStoreSPILLVSRRCAsGpr++;
2223 MI.setDesc(get(PPC::STDX));
2228 case PPC::CFENCE8: {
2229 auto Val = MI.getOperand(0).getReg();
2230 BuildMI(MBB, MI, DL, get(PPC::CMPD), PPC::CR7).addReg(Val).addReg(Val);
2231 BuildMI(MBB, MI, DL, get(PPC::CTRL_DEP))
2232 .addImm(PPC::PRED_NE_MINUS)
2235 MI.setDesc(get(PPC::ISYNC));
2236 MI.RemoveOperand(0);
2243 // Essentially a compile-time implementation of a compare->isel sequence.
2244 // It takes two constants to compare, along with the true/false registers
2245 // and the comparison type (as a subreg to a CR field) and returns one
2246 // of the true/false registers, depending on the comparison results.
2247 static unsigned selectReg(int64_t Imm1, int64_t Imm2, unsigned CompareOpc,
2248 unsigned TrueReg, unsigned FalseReg,
2249 unsigned CRSubReg) {
2250 // Signed comparisons. The immediates are assumed to be sign-extended.
2251 if (CompareOpc == PPC::CMPWI || CompareOpc == PPC::CMPDI) {
2253 default: llvm_unreachable("Unknown integer comparison type.");
2255 return Imm1 < Imm2 ? TrueReg : FalseReg;
2257 return Imm1 > Imm2 ? TrueReg : FalseReg;
2259 return Imm1 == Imm2 ? TrueReg : FalseReg;
2262 // Unsigned comparisons.
2263 else if (CompareOpc == PPC::CMPLWI || CompareOpc == PPC::CMPLDI) {
2265 default: llvm_unreachable("Unknown integer comparison type.");
2267 return (uint64_t)Imm1 < (uint64_t)Imm2 ? TrueReg : FalseReg;
2269 return (uint64_t)Imm1 > (uint64_t)Imm2 ? TrueReg : FalseReg;
2271 return Imm1 == Imm2 ? TrueReg : FalseReg;
2274 return PPC::NoRegister;
2277 void PPCInstrInfo::replaceInstrOperandWithImm(MachineInstr &MI,
2279 int64_t Imm) const {
2280 assert(MI.getOperand(OpNo).isReg() && "Operand must be a REG");
2281 // Replace the REG with the Immediate.
2282 unsigned InUseReg = MI.getOperand(OpNo).getReg();
2283 MI.getOperand(OpNo).ChangeToImmediate(Imm);
2285 if (empty(MI.implicit_operands()))
2288 // We need to make sure that the MI didn't have any implicit use
2289 // of this REG any more.
2290 const TargetRegisterInfo *TRI = &getRegisterInfo();
2291 int UseOpIdx = MI.findRegisterUseOperandIdx(InUseReg, false, TRI);
2292 if (UseOpIdx >= 0) {
2293 MachineOperand &MO = MI.getOperand(UseOpIdx);
2294 if (MO.isImplicit())
2295 // The operands must always be in the following order:
2296 // - explicit reg defs,
2297 // - other explicit operands (reg uses, immediates, etc.),
2298 // - implicit reg defs
2299 // - implicit reg uses
2300 // Therefore, removing the implicit operand won't change the explicit
2302 MI.RemoveOperand(UseOpIdx);
2306 // Replace an instruction with one that materializes a constant (and sets
2307 // CR0 if the original instruction was a record-form instruction).
2308 void PPCInstrInfo::replaceInstrWithLI(MachineInstr &MI,
2309 const LoadImmediateInfo &LII) const {
2310 // Remove existing operands.
2311 int OperandToKeep = LII.SetCR ? 1 : 0;
2312 for (int i = MI.getNumOperands() - 1; i > OperandToKeep; i--)
2313 MI.RemoveOperand(i);
2315 // Replace the instruction.
2317 MI.setDesc(get(LII.Is64Bit ? PPC::ANDIo8 : PPC::ANDIo));
2318 // Set the immediate.
2319 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
2320 .addImm(LII.Imm).addReg(PPC::CR0, RegState::ImplicitDefine);
2324 MI.setDesc(get(LII.Is64Bit ? PPC::LI8 : PPC::LI));
2326 // Set the immediate.
2327 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
2331 MachineInstr *PPCInstrInfo::getForwardingDefMI(
2333 unsigned &OpNoForForwarding,
2334 bool &SeenIntermediateUse) const {
2335 OpNoForForwarding = ~0U;
2336 MachineInstr *DefMI = nullptr;
2337 MachineRegisterInfo *MRI = &MI.getParent()->getParent()->getRegInfo();
2338 const TargetRegisterInfo *TRI = &getRegisterInfo();
2339 // If we're in SSA, get the defs through the MRI. Otherwise, only look
2340 // within the basic block to see if the register is defined using an LI/LI8.
2342 for (int i = 1, e = MI.getNumOperands(); i < e; i++) {
2343 if (!MI.getOperand(i).isReg())
2345 unsigned Reg = MI.getOperand(i).getReg();
2346 if (!TargetRegisterInfo::isVirtualRegister(Reg))
2348 unsigned TrueReg = TRI->lookThruCopyLike(Reg, MRI);
2349 if (TargetRegisterInfo::isVirtualRegister(TrueReg)) {
2350 DefMI = MRI->getVRegDef(TrueReg);
2351 if (DefMI->getOpcode() == PPC::LI || DefMI->getOpcode() == PPC::LI8) {
2352 OpNoForForwarding = i;
2358 // Looking back through the definition for each operand could be expensive,
2359 // so exit early if this isn't an instruction that either has an immediate
2360 // form or is already an immediate form that we can handle.
2362 unsigned Opc = MI.getOpcode();
2363 bool ConvertibleImmForm =
2364 Opc == PPC::CMPWI || Opc == PPC::CMPLWI ||
2365 Opc == PPC::CMPDI || Opc == PPC::CMPLDI ||
2366 Opc == PPC::ADDI || Opc == PPC::ADDI8 ||
2367 Opc == PPC::ORI || Opc == PPC::ORI8 ||
2368 Opc == PPC::XORI || Opc == PPC::XORI8 ||
2369 Opc == PPC::RLDICL || Opc == PPC::RLDICLo ||
2370 Opc == PPC::RLDICL_32 || Opc == PPC::RLDICL_32_64 ||
2371 Opc == PPC::RLWINM || Opc == PPC::RLWINMo ||
2372 Opc == PPC::RLWINM8 || Opc == PPC::RLWINM8o;
2373 if (!instrHasImmForm(MI, III, true) && !ConvertibleImmForm)
2376 // Don't convert or %X, %Y, %Y since that's just a register move.
2377 if ((Opc == PPC::OR || Opc == PPC::OR8) &&
2378 MI.getOperand(1).getReg() == MI.getOperand(2).getReg())
2380 for (int i = 1, e = MI.getNumOperands(); i < e; i++) {
2381 MachineOperand &MO = MI.getOperand(i);
2382 SeenIntermediateUse = false;
2383 if (MO.isReg() && MO.isUse() && !MO.isImplicit()) {
2384 MachineBasicBlock::reverse_iterator E = MI.getParent()->rend(), It = MI;
2386 unsigned Reg = MI.getOperand(i).getReg();
2388 // Is this register defined by some form of add-immediate (including
2389 // load-immediate) within this basic block?
2390 for ( ; It != E; ++It) {
2391 if (It->modifiesRegister(Reg, &getRegisterInfo())) {
2392 switch (It->getOpcode()) {
2399 OpNoForForwarding = i;
2403 } else if (It->readsRegister(Reg, &getRegisterInfo()))
2404 // If we see another use of this reg between the def and the MI,
2405 // we want to flat it so the def isn't deleted.
2406 SeenIntermediateUse = true;
2411 return OpNoForForwarding == ~0U ? nullptr : DefMI;
2414 const unsigned *PPCInstrInfo::getStoreOpcodesForSpillArray() const {
2415 static const unsigned OpcodesForSpill[2][SOK_LastOpcodeSpill] = {
2417 {PPC::STW, PPC::STD, PPC::STFD, PPC::STFS, PPC::SPILL_CR,
2418 PPC::SPILL_CRBIT, PPC::STVX, PPC::STXVD2X, PPC::STXSDX, PPC::STXSSPX,
2419 PPC::SPILL_VRSAVE, PPC::QVSTFDX, PPC::QVSTFSXs, PPC::QVSTFDXb,
2420 PPC::SPILLTOVSR_ST, PPC::EVSTDD, PPC::SPESTW},
2422 {PPC::STW, PPC::STD, PPC::STFD, PPC::STFS, PPC::SPILL_CR,
2423 PPC::SPILL_CRBIT, PPC::STVX, PPC::STXV, PPC::DFSTOREf64, PPC::DFSTOREf32,
2424 PPC::SPILL_VRSAVE, PPC::QVSTFDX, PPC::QVSTFSXs, PPC::QVSTFDXb,
2425 PPC::SPILLTOVSR_ST}};
2427 return OpcodesForSpill[(Subtarget.hasP9Vector()) ? 1 : 0];
2430 const unsigned *PPCInstrInfo::getLoadOpcodesForSpillArray() const {
2431 static const unsigned OpcodesForSpill[2][SOK_LastOpcodeSpill] = {
2433 {PPC::LWZ, PPC::LD, PPC::LFD, PPC::LFS, PPC::RESTORE_CR,
2434 PPC::RESTORE_CRBIT, PPC::LVX, PPC::LXVD2X, PPC::LXSDX, PPC::LXSSPX,
2435 PPC::RESTORE_VRSAVE, PPC::QVLFDX, PPC::QVLFSXs, PPC::QVLFDXb,
2436 PPC::SPILLTOVSR_LD, PPC::EVLDD, PPC::SPELWZ},
2438 {PPC::LWZ, PPC::LD, PPC::LFD, PPC::LFS, PPC::RESTORE_CR,
2439 PPC::RESTORE_CRBIT, PPC::LVX, PPC::LXV, PPC::DFLOADf64, PPC::DFLOADf32,
2440 PPC::RESTORE_VRSAVE, PPC::QVLFDX, PPC::QVLFSXs, PPC::QVLFDXb,
2441 PPC::SPILLTOVSR_LD}};
2443 return OpcodesForSpill[(Subtarget.hasP9Vector()) ? 1 : 0];
2446 void PPCInstrInfo::fixupIsDeadOrKill(MachineInstr &StartMI, MachineInstr &EndMI,
2447 unsigned RegNo) const {
2448 const MachineRegisterInfo &MRI =
2449 StartMI.getParent()->getParent()->getRegInfo();
2453 // Instructions between [StartMI, EndMI] should be in same basic block.
2454 assert((StartMI.getParent() == EndMI.getParent()) &&
2455 "Instructions are not in same basic block");
2457 bool IsKillSet = false;
2459 auto clearOperandKillInfo = [=] (MachineInstr &MI, unsigned Index) {
2460 MachineOperand &MO = MI.getOperand(Index);
2461 if (MO.isReg() && MO.isUse() && MO.isKill() &&
2462 getRegisterInfo().regsOverlap(MO.getReg(), RegNo))
2463 MO.setIsKill(false);
2466 // Set killed flag for EndMI.
2467 // No need to do anything if EndMI defines RegNo.
2469 EndMI.findRegisterUseOperandIdx(RegNo, false, &getRegisterInfo());
2470 if (UseIndex != -1) {
2471 EndMI.getOperand(UseIndex).setIsKill(true);
2473 // Clear killed flag for other EndMI operands related to RegNo. In some
2474 // upexpected cases, killed may be set multiple times for same register
2475 // operand in same MI.
2476 for (int i = 0, e = EndMI.getNumOperands(); i != e; ++i)
2478 clearOperandKillInfo(EndMI, i);
2481 // Walking the inst in reverse order (EndMI -> StartMI].
2482 MachineBasicBlock::reverse_iterator It = EndMI;
2483 MachineBasicBlock::reverse_iterator E = EndMI.getParent()->rend();
2484 // EndMI has been handled above, skip it here.
2486 MachineOperand *MO = nullptr;
2487 for (; It != E; ++It) {
2488 // Skip insturctions which could not be a def/use of RegNo.
2489 if (It->isDebugInstr() || It->isPosition())
2492 // Clear killed flag for all It operands related to RegNo. In some
2493 // upexpected cases, killed may be set multiple times for same register
2494 // operand in same MI.
2495 for (int i = 0, e = It->getNumOperands(); i != e; ++i)
2496 clearOperandKillInfo(*It, i);
2498 // If killed is not set, set killed for its last use or set dead for its def
2501 if ((MO = It->findRegisterUseOperand(RegNo, false, &getRegisterInfo()))) {
2502 // Use found, set it killed.
2504 MO->setIsKill(true);
2506 } else if ((MO = It->findRegisterDefOperand(RegNo, false, true,
2507 &getRegisterInfo()))) {
2508 // No use found, set dead for its def.
2509 assert(&*It == &StartMI && "No new def between StartMI and EndMI.");
2510 MO->setIsDead(true);
2515 if ((&*It) == &StartMI)
2518 // Ensure RegMo liveness is killed after EndMI.
2519 assert((IsKillSet || (MO && MO->isDead())) &&
2520 "RegNo should be killed or dead");
2523 // If this instruction has an immediate form and one of its operands is a
2524 // result of a load-immediate or an add-immediate, convert it to
2525 // the immediate form if the constant is in range.
2526 bool PPCInstrInfo::convertToImmediateForm(MachineInstr &MI,
2527 MachineInstr **KilledDef) const {
2528 MachineFunction *MF = MI.getParent()->getParent();
2529 MachineRegisterInfo *MRI = &MF->getRegInfo();
2530 bool PostRA = !MRI->isSSA();
2531 bool SeenIntermediateUse = true;
2532 unsigned ForwardingOperand = ~0U;
2533 MachineInstr *DefMI = getForwardingDefMI(MI, ForwardingOperand,
2534 SeenIntermediateUse);
2537 assert(ForwardingOperand < MI.getNumOperands() &&
2538 "The forwarding operand needs to be valid at this point");
2539 bool IsForwardingOperandKilled = MI.getOperand(ForwardingOperand).isKill();
2540 bool KillFwdDefMI = !SeenIntermediateUse && IsForwardingOperandKilled;
2541 unsigned ForwardingOperandReg = MI.getOperand(ForwardingOperand).getReg();
2542 if (KilledDef && KillFwdDefMI)
2546 bool HasImmForm = instrHasImmForm(MI, III, PostRA);
2547 // If this is a reg+reg instruction that has a reg+imm form,
2548 // and one of the operands is produced by an add-immediate,
2549 // try to convert it.
2551 transformToImmFormFedByAdd(MI, III, ForwardingOperand, *DefMI,
2555 if ((DefMI->getOpcode() != PPC::LI && DefMI->getOpcode() != PPC::LI8) ||
2556 !DefMI->getOperand(1).isImm())
2559 int64_t Immediate = DefMI->getOperand(1).getImm();
2560 // Sign-extend to 64-bits.
2561 int64_t SExtImm = ((uint64_t)Immediate & ~0x7FFFuLL) != 0 ?
2562 (Immediate | 0xFFFFFFFFFFFF0000) : Immediate;
2564 // If this is a reg+reg instruction that has a reg+imm form,
2565 // and one of the operands is produced by LI, convert it now.
2567 return transformToImmFormFedByLI(MI, III, ForwardingOperand, *DefMI, SExtImm);
2569 bool ReplaceWithLI = false;
2570 bool Is64BitLI = false;
2573 unsigned Opc = MI.getOpcode();
2575 default: return false;
2577 // FIXME: Any branches conditional on such a comparison can be made
2578 // unconditional. At this time, this happens too infrequently to be worth
2579 // the implementation effort, but if that ever changes, we could convert
2580 // such a pattern here.
2585 // Doing this post-RA would require dataflow analysis to reliably find uses
2586 // of the CR register set by the compare.
2587 // No need to fixup killed/dead flag since this transformation is only valid
2591 // If a compare-immediate is fed by an immediate and is itself an input of
2592 // an ISEL (the most common case) into a COPY of the correct register.
2593 bool Changed = false;
2594 unsigned DefReg = MI.getOperand(0).getReg();
2595 int64_t Comparand = MI.getOperand(2).getImm();
2596 int64_t SExtComparand = ((uint64_t)Comparand & ~0x7FFFuLL) != 0 ?
2597 (Comparand | 0xFFFFFFFFFFFF0000) : Comparand;
2599 for (auto &CompareUseMI : MRI->use_instructions(DefReg)) {
2600 unsigned UseOpc = CompareUseMI.getOpcode();
2601 if (UseOpc != PPC::ISEL && UseOpc != PPC::ISEL8)
2603 unsigned CRSubReg = CompareUseMI.getOperand(3).getSubReg();
2604 unsigned TrueReg = CompareUseMI.getOperand(1).getReg();
2605 unsigned FalseReg = CompareUseMI.getOperand(2).getReg();
2606 unsigned RegToCopy = selectReg(SExtImm, SExtComparand, Opc, TrueReg,
2607 FalseReg, CRSubReg);
2608 if (RegToCopy == PPC::NoRegister)
2610 // Can't use PPC::COPY to copy PPC::ZERO[8]. Convert it to LI[8] 0.
2611 if (RegToCopy == PPC::ZERO || RegToCopy == PPC::ZERO8) {
2612 CompareUseMI.setDesc(get(UseOpc == PPC::ISEL8 ? PPC::LI8 : PPC::LI));
2613 replaceInstrOperandWithImm(CompareUseMI, 1, 0);
2614 CompareUseMI.RemoveOperand(3);
2615 CompareUseMI.RemoveOperand(2);
2619 dbgs() << "Found LI -> CMPI -> ISEL, replacing with a copy.\n");
2620 LLVM_DEBUG(DefMI->dump(); MI.dump(); CompareUseMI.dump());
2621 LLVM_DEBUG(dbgs() << "Is converted to:\n");
2622 // Convert to copy and remove unneeded operands.
2623 CompareUseMI.setDesc(get(PPC::COPY));
2624 CompareUseMI.RemoveOperand(3);
2625 CompareUseMI.RemoveOperand(RegToCopy == TrueReg ? 2 : 1);
2626 CmpIselsConverted++;
2628 LLVM_DEBUG(CompareUseMI.dump());
2632 // This may end up incremented multiple times since this function is called
2633 // during a fixed-point transformation, but it is only meant to indicate the
2634 // presence of this opportunity.
2635 MissedConvertibleImmediateInstrs++;
2639 // Immediate forms - may simply be convertable to an LI.
2642 // Does the sum fit in a 16-bit signed field?
2643 int64_t Addend = MI.getOperand(2).getImm();
2644 if (isInt<16>(Addend + SExtImm)) {
2645 ReplaceWithLI = true;
2646 Is64BitLI = Opc == PPC::ADDI8;
2647 NewImm = Addend + SExtImm;
2654 case PPC::RLDICL_32:
2655 case PPC::RLDICL_32_64: {
2656 // Use APInt's rotate function.
2657 int64_t SH = MI.getOperand(2).getImm();
2658 int64_t MB = MI.getOperand(3).getImm();
2659 APInt InVal((Opc == PPC::RLDICL || Opc == PPC::RLDICLo) ?
2660 64 : 32, SExtImm, true);
2661 InVal = InVal.rotl(SH);
2662 uint64_t Mask = (1LLU << (63 - MB + 1)) - 1;
2664 // Can't replace negative values with an LI as that will sign-extend
2665 // and not clear the left bits. If we're setting the CR bit, we will use
2666 // ANDIo which won't sign extend, so that's safe.
2667 if (isUInt<15>(InVal.getSExtValue()) ||
2668 (Opc == PPC::RLDICLo && isUInt<16>(InVal.getSExtValue()))) {
2669 ReplaceWithLI = true;
2670 Is64BitLI = Opc != PPC::RLDICL_32;
2671 NewImm = InVal.getSExtValue();
2672 SetCR = Opc == PPC::RLDICLo;
2680 case PPC::RLWINM8o: {
2681 int64_t SH = MI.getOperand(2).getImm();
2682 int64_t MB = MI.getOperand(3).getImm();
2683 int64_t ME = MI.getOperand(4).getImm();
2684 APInt InVal(32, SExtImm, true);
2685 InVal = InVal.rotl(SH);
2686 // Set the bits ( MB + 32 ) to ( ME + 32 ).
2687 uint64_t Mask = ((1LLU << (32 - MB)) - 1) & ~((1LLU << (31 - ME)) - 1);
2689 // Can't replace negative values with an LI as that will sign-extend
2690 // and not clear the left bits. If we're setting the CR bit, we will use
2691 // ANDIo which won't sign extend, so that's safe.
2692 bool ValueFits = isUInt<15>(InVal.getSExtValue());
2693 ValueFits |= ((Opc == PPC::RLWINMo || Opc == PPC::RLWINM8o) &&
2694 isUInt<16>(InVal.getSExtValue()));
2696 ReplaceWithLI = true;
2697 Is64BitLI = Opc == PPC::RLWINM8 || Opc == PPC::RLWINM8o;
2698 NewImm = InVal.getSExtValue();
2699 SetCR = Opc == PPC::RLWINMo || Opc == PPC::RLWINM8o;
2708 int64_t LogicalImm = MI.getOperand(2).getImm();
2710 if (Opc == PPC::ORI || Opc == PPC::ORI8)
2711 Result = LogicalImm | SExtImm;
2713 Result = LogicalImm ^ SExtImm;
2714 if (isInt<16>(Result)) {
2715 ReplaceWithLI = true;
2716 Is64BitLI = Opc == PPC::ORI8 || Opc == PPC::XORI8;
2724 if (ReplaceWithLI) {
2725 // We need to be careful with CR-setting instructions we're replacing.
2727 // We don't know anything about uses when we're out of SSA, so only
2728 // replace if the new immediate will be reproduced.
2729 bool ImmChanged = (SExtImm & NewImm) != NewImm;
2730 if (PostRA && ImmChanged)
2734 // If the defining load-immediate has no other uses, we can just replace
2735 // the immediate with the new immediate.
2736 if (MRI->hasOneUse(DefMI->getOperand(0).getReg()))
2737 DefMI->getOperand(1).setImm(NewImm);
2739 // If we're not using the GPR result of the CR-setting instruction, we
2740 // just need to and with zero/non-zero depending on the new immediate.
2741 else if (MRI->use_empty(MI.getOperand(0).getReg())) {
2743 assert(Immediate && "Transformation converted zero to non-zero?");
2747 else if (ImmChanged)
2752 LLVM_DEBUG(dbgs() << "Replacing instruction:\n");
2753 LLVM_DEBUG(MI.dump());
2754 LLVM_DEBUG(dbgs() << "Fed by:\n");
2755 LLVM_DEBUG(DefMI->dump());
2756 LoadImmediateInfo LII;
2758 LII.Is64Bit = Is64BitLI;
2760 // If we're setting the CR, the original load-immediate must be kept (as an
2761 // operand to ANDIo/ANDI8o).
2762 if (KilledDef && SetCR)
2763 *KilledDef = nullptr;
2764 replaceInstrWithLI(MI, LII);
2766 // Fixup killed/dead flag after transformation.
2768 // ForwardingOperandReg = LI imm1
2769 // y = op2 imm2, ForwardingOperandReg(killed)
2770 if (IsForwardingOperandKilled)
2771 fixupIsDeadOrKill(*DefMI, MI, ForwardingOperandReg);
2773 LLVM_DEBUG(dbgs() << "With:\n");
2774 LLVM_DEBUG(MI.dump());
2780 bool PPCInstrInfo::instrHasImmForm(const MachineInstr &MI,
2781 ImmInstrInfo &III, bool PostRA) const {
2782 unsigned Opc = MI.getOpcode();
2783 // The vast majority of the instructions would need their operand 2 replaced
2784 // with an immediate when switching to the reg+imm form. A marked exception
2785 // are the update form loads/stores for which a constant operand 2 would need
2786 // to turn into a displacement and move operand 1 to the operand 2 position.
2788 III.OpNoForForwarding = 2;
2790 III.ImmMustBeMultipleOf = 1;
2791 III.TruncateImmTo = 0;
2792 III.IsSummingOperands = false;
2794 default: return false;
2797 III.SignedImm = true;
2798 III.ZeroIsSpecialOrig = 0;
2799 III.ZeroIsSpecialNew = 1;
2800 III.IsCommutative = true;
2801 III.IsSummingOperands = true;
2802 III.ImmOpcode = Opc == PPC::ADD4 ? PPC::ADDI : PPC::ADDI8;
2806 III.SignedImm = true;
2807 III.ZeroIsSpecialOrig = 0;
2808 III.ZeroIsSpecialNew = 0;
2809 III.IsCommutative = true;
2810 III.IsSummingOperands = true;
2811 III.ImmOpcode = Opc == PPC::ADDC ? PPC::ADDIC : PPC::ADDIC8;
2814 III.SignedImm = true;
2815 III.ZeroIsSpecialOrig = 0;
2816 III.ZeroIsSpecialNew = 0;
2817 III.IsCommutative = true;
2818 III.IsSummingOperands = true;
2819 III.ImmOpcode = PPC::ADDICo;
2823 III.SignedImm = true;
2824 III.ZeroIsSpecialOrig = 0;
2825 III.ZeroIsSpecialNew = 0;
2826 III.IsCommutative = false;
2827 III.ImmOpcode = Opc == PPC::SUBFC ? PPC::SUBFIC : PPC::SUBFIC8;
2831 III.SignedImm = true;
2832 III.ZeroIsSpecialOrig = 0;
2833 III.ZeroIsSpecialNew = 0;
2834 III.IsCommutative = false;
2835 III.ImmOpcode = Opc == PPC::CMPW ? PPC::CMPWI : PPC::CMPDI;
2839 III.SignedImm = false;
2840 III.ZeroIsSpecialOrig = 0;
2841 III.ZeroIsSpecialNew = 0;
2842 III.IsCommutative = false;
2843 III.ImmOpcode = Opc == PPC::CMPLW ? PPC::CMPLWI : PPC::CMPLDI;
2851 III.SignedImm = false;
2852 III.ZeroIsSpecialOrig = 0;
2853 III.ZeroIsSpecialNew = 0;
2854 III.IsCommutative = true;
2856 default: llvm_unreachable("Unknown opcode");
2857 case PPC::ANDo: III.ImmOpcode = PPC::ANDIo; break;
2858 case PPC::AND8o: III.ImmOpcode = PPC::ANDIo8; break;
2859 case PPC::OR: III.ImmOpcode = PPC::ORI; break;
2860 case PPC::OR8: III.ImmOpcode = PPC::ORI8; break;
2861 case PPC::XOR: III.ImmOpcode = PPC::XORI; break;
2862 case PPC::XOR8: III.ImmOpcode = PPC::XORI8; break;
2879 III.SignedImm = false;
2880 III.ZeroIsSpecialOrig = 0;
2881 III.ZeroIsSpecialNew = 0;
2882 III.IsCommutative = false;
2883 // This isn't actually true, but the instructions ignore any of the
2884 // upper bits, so any immediate loaded with an LI is acceptable.
2885 // This does not apply to shift right algebraic because a value
2886 // out of range will produce a -1/0.
2888 if (Opc == PPC::RLWNM || Opc == PPC::RLWNM8 ||
2889 Opc == PPC::RLWNMo || Opc == PPC::RLWNM8o)
2890 III.TruncateImmTo = 5;
2892 III.TruncateImmTo = 6;
2894 default: llvm_unreachable("Unknown opcode");
2895 case PPC::RLWNM: III.ImmOpcode = PPC::RLWINM; break;
2896 case PPC::RLWNM8: III.ImmOpcode = PPC::RLWINM8; break;
2897 case PPC::RLWNMo: III.ImmOpcode = PPC::RLWINMo; break;
2898 case PPC::RLWNM8o: III.ImmOpcode = PPC::RLWINM8o; break;
2899 case PPC::SLW: III.ImmOpcode = PPC::RLWINM; break;
2900 case PPC::SLW8: III.ImmOpcode = PPC::RLWINM8; break;
2901 case PPC::SLWo: III.ImmOpcode = PPC::RLWINMo; break;
2902 case PPC::SLW8o: III.ImmOpcode = PPC::RLWINM8o; break;
2903 case PPC::SRW: III.ImmOpcode = PPC::RLWINM; break;
2904 case PPC::SRW8: III.ImmOpcode = PPC::RLWINM8; break;
2905 case PPC::SRWo: III.ImmOpcode = PPC::RLWINMo; break;
2906 case PPC::SRW8o: III.ImmOpcode = PPC::RLWINM8o; break;
2909 III.TruncateImmTo = 0;
2910 III.ImmOpcode = PPC::SRAWI;
2914 III.TruncateImmTo = 0;
2915 III.ImmOpcode = PPC::SRAWIo;
2929 III.SignedImm = false;
2930 III.ZeroIsSpecialOrig = 0;
2931 III.ZeroIsSpecialNew = 0;
2932 III.IsCommutative = false;
2933 // This isn't actually true, but the instructions ignore any of the
2934 // upper bits, so any immediate loaded with an LI is acceptable.
2935 // This does not apply to shift right algebraic because a value
2936 // out of range will produce a -1/0.
2938 if (Opc == PPC::RLDCL || Opc == PPC::RLDCLo ||
2939 Opc == PPC::RLDCR || Opc == PPC::RLDCRo)
2940 III.TruncateImmTo = 6;
2942 III.TruncateImmTo = 7;
2944 default: llvm_unreachable("Unknown opcode");
2945 case PPC::RLDCL: III.ImmOpcode = PPC::RLDICL; break;
2946 case PPC::RLDCLo: III.ImmOpcode = PPC::RLDICLo; break;
2947 case PPC::RLDCR: III.ImmOpcode = PPC::RLDICR; break;
2948 case PPC::RLDCRo: III.ImmOpcode = PPC::RLDICRo; break;
2949 case PPC::SLD: III.ImmOpcode = PPC::RLDICR; break;
2950 case PPC::SLDo: III.ImmOpcode = PPC::RLDICRo; break;
2951 case PPC::SRD: III.ImmOpcode = PPC::RLDICL; break;
2952 case PPC::SRDo: III.ImmOpcode = PPC::RLDICLo; break;
2955 III.TruncateImmTo = 0;
2956 III.ImmOpcode = PPC::SRADI;
2960 III.TruncateImmTo = 0;
2961 III.ImmOpcode = PPC::SRADIo;
2965 // Loads and stores:
2987 III.SignedImm = true;
2988 III.ZeroIsSpecialOrig = 1;
2989 III.ZeroIsSpecialNew = 2;
2990 III.IsCommutative = true;
2991 III.IsSummingOperands = true;
2993 III.OpNoForForwarding = 2;
2995 default: llvm_unreachable("Unknown opcode");
2996 case PPC::LBZX: III.ImmOpcode = PPC::LBZ; break;
2997 case PPC::LBZX8: III.ImmOpcode = PPC::LBZ8; break;
2998 case PPC::LHZX: III.ImmOpcode = PPC::LHZ; break;
2999 case PPC::LHZX8: III.ImmOpcode = PPC::LHZ8; break;
3000 case PPC::LHAX: III.ImmOpcode = PPC::LHA; break;
3001 case PPC::LHAX8: III.ImmOpcode = PPC::LHA8; break;
3002 case PPC::LWZX: III.ImmOpcode = PPC::LWZ; break;
3003 case PPC::LWZX8: III.ImmOpcode = PPC::LWZ8; break;
3005 III.ImmOpcode = PPC::LWA;
3006 III.ImmMustBeMultipleOf = 4;
3008 case PPC::LDX: III.ImmOpcode = PPC::LD; III.ImmMustBeMultipleOf = 4; break;
3009 case PPC::LFSX: III.ImmOpcode = PPC::LFS; break;
3010 case PPC::LFDX: III.ImmOpcode = PPC::LFD; break;
3011 case PPC::STBX: III.ImmOpcode = PPC::STB; break;
3012 case PPC::STBX8: III.ImmOpcode = PPC::STB8; break;
3013 case PPC::STHX: III.ImmOpcode = PPC::STH; break;
3014 case PPC::STHX8: III.ImmOpcode = PPC::STH8; break;
3015 case PPC::STWX: III.ImmOpcode = PPC::STW; break;
3016 case PPC::STWX8: III.ImmOpcode = PPC::STW8; break;
3018 III.ImmOpcode = PPC::STD;
3019 III.ImmMustBeMultipleOf = 4;
3021 case PPC::STFSX: III.ImmOpcode = PPC::STFS; break;
3022 case PPC::STFDX: III.ImmOpcode = PPC::STFD; break;
3045 III.SignedImm = true;
3046 III.ZeroIsSpecialOrig = 2;
3047 III.ZeroIsSpecialNew = 3;
3048 III.IsCommutative = false;
3049 III.IsSummingOperands = true;
3051 III.OpNoForForwarding = 3;
3053 default: llvm_unreachable("Unknown opcode");
3054 case PPC::LBZUX: III.ImmOpcode = PPC::LBZU; break;
3055 case PPC::LBZUX8: III.ImmOpcode = PPC::LBZU8; break;
3056 case PPC::LHZUX: III.ImmOpcode = PPC::LHZU; break;
3057 case PPC::LHZUX8: III.ImmOpcode = PPC::LHZU8; break;
3058 case PPC::LHAUX: III.ImmOpcode = PPC::LHAU; break;
3059 case PPC::LHAUX8: III.ImmOpcode = PPC::LHAU8; break;
3060 case PPC::LWZUX: III.ImmOpcode = PPC::LWZU; break;
3061 case PPC::LWZUX8: III.ImmOpcode = PPC::LWZU8; break;
3063 III.ImmOpcode = PPC::LDU;
3064 III.ImmMustBeMultipleOf = 4;
3066 case PPC::LFSUX: III.ImmOpcode = PPC::LFSU; break;
3067 case PPC::LFDUX: III.ImmOpcode = PPC::LFDU; break;
3068 case PPC::STBUX: III.ImmOpcode = PPC::STBU; break;
3069 case PPC::STBUX8: III.ImmOpcode = PPC::STBU8; break;
3070 case PPC::STHUX: III.ImmOpcode = PPC::STHU; break;
3071 case PPC::STHUX8: III.ImmOpcode = PPC::STHU8; break;
3072 case PPC::STWUX: III.ImmOpcode = PPC::STWU; break;
3073 case PPC::STWUX8: III.ImmOpcode = PPC::STWU8; break;
3075 III.ImmOpcode = PPC::STDU;
3076 III.ImmMustBeMultipleOf = 4;
3078 case PPC::STFSUX: III.ImmOpcode = PPC::STFSU; break;
3079 case PPC::STFDUX: III.ImmOpcode = PPC::STFDU; break;
3082 // Power9 and up only. For some of these, the X-Form version has access to all
3083 // 64 VSR's whereas the D-Form only has access to the VR's. We replace those
3084 // with pseudo-ops pre-ra and for post-ra, we check that the register loaded
3085 // into or stored from is one of the VR registers.
3092 case PPC::XFLOADf32:
3093 case PPC::XFLOADf64:
3094 case PPC::XFSTOREf32:
3095 case PPC::XFSTOREf64:
3096 if (!Subtarget.hasP9Vector())
3098 III.SignedImm = true;
3099 III.ZeroIsSpecialOrig = 1;
3100 III.ZeroIsSpecialNew = 2;
3101 III.IsCommutative = true;
3102 III.IsSummingOperands = true;
3104 III.OpNoForForwarding = 2;
3105 III.ImmMustBeMultipleOf = 4;
3107 default: llvm_unreachable("Unknown opcode");
3109 III.ImmOpcode = PPC::LXV;
3110 III.ImmMustBeMultipleOf = 16;
3114 if (isVFRegister(MI.getOperand(0).getReg()))
3115 III.ImmOpcode = PPC::LXSSP;
3117 III.ImmOpcode = PPC::LFS;
3118 III.ImmMustBeMultipleOf = 1;
3123 case PPC::XFLOADf32:
3124 III.ImmOpcode = PPC::DFLOADf32;
3128 if (isVFRegister(MI.getOperand(0).getReg()))
3129 III.ImmOpcode = PPC::LXSD;
3131 III.ImmOpcode = PPC::LFD;
3132 III.ImmMustBeMultipleOf = 1;
3137 case PPC::XFLOADf64:
3138 III.ImmOpcode = PPC::DFLOADf64;
3141 III.ImmOpcode = PPC::STXV;
3142 III.ImmMustBeMultipleOf = 16;
3146 if (isVFRegister(MI.getOperand(0).getReg()))
3147 III.ImmOpcode = PPC::STXSSP;
3149 III.ImmOpcode = PPC::STFS;
3150 III.ImmMustBeMultipleOf = 1;
3155 case PPC::XFSTOREf32:
3156 III.ImmOpcode = PPC::DFSTOREf32;
3160 if (isVFRegister(MI.getOperand(0).getReg()))
3161 III.ImmOpcode = PPC::STXSD;
3163 III.ImmOpcode = PPC::STFD;
3164 III.ImmMustBeMultipleOf = 1;
3169 case PPC::XFSTOREf64:
3170 III.ImmOpcode = PPC::DFSTOREf64;
3178 // Utility function for swaping two arbitrary operands of an instruction.
3179 static void swapMIOperands(MachineInstr &MI, unsigned Op1, unsigned Op2) {
3180 assert(Op1 != Op2 && "Cannot swap operand with itself.");
3182 unsigned MaxOp = std::max(Op1, Op2);
3183 unsigned MinOp = std::min(Op1, Op2);
3184 MachineOperand MOp1 = MI.getOperand(MinOp);
3185 MachineOperand MOp2 = MI.getOperand(MaxOp);
3186 MI.RemoveOperand(std::max(Op1, Op2));
3187 MI.RemoveOperand(std::min(Op1, Op2));
3189 // If the operands we are swapping are the two at the end (the common case)
3190 // we can just remove both and add them in the opposite order.
3191 if (MaxOp - MinOp == 1 && MI.getNumOperands() == MinOp) {
3192 MI.addOperand(MOp2);
3193 MI.addOperand(MOp1);
3195 // Store all operands in a temporary vector, remove them and re-add in the
3197 SmallVector<MachineOperand, 2> MOps;
3198 unsigned TotalOps = MI.getNumOperands() + 2; // We've already removed 2 ops.
3199 for (unsigned i = MI.getNumOperands() - 1; i >= MinOp; i--) {
3200 MOps.push_back(MI.getOperand(i));
3201 MI.RemoveOperand(i);
3203 // MOp2 needs to be added next.
3204 MI.addOperand(MOp2);
3205 // Now add the rest.
3206 for (unsigned i = MI.getNumOperands(); i < TotalOps; i++) {
3208 MI.addOperand(MOp1);
3210 MI.addOperand(MOps.back());
3217 // Check if the 'MI' that has the index OpNoForForwarding
3218 // meets the requirement described in the ImmInstrInfo.
3219 bool PPCInstrInfo::isUseMIElgibleForForwarding(MachineInstr &MI,
3220 const ImmInstrInfo &III,
3221 unsigned OpNoForForwarding
3223 // As the algorithm of checking for PPC::ZERO/PPC::ZERO8
3224 // would not work pre-RA, we can only do the check post RA.
3225 MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
3229 // Cannot do the transform if MI isn't summing the operands.
3230 if (!III.IsSummingOperands)
3233 // The instruction we are trying to replace must have the ZeroIsSpecialOrig set.
3234 if (!III.ZeroIsSpecialOrig)
3237 // We cannot do the transform if the operand we are trying to replace
3238 // isn't the same as the operand the instruction allows.
3239 if (OpNoForForwarding != III.OpNoForForwarding)
3242 // Check if the instruction we are trying to transform really has
3243 // the special zero register as its operand.
3244 if (MI.getOperand(III.ZeroIsSpecialOrig).getReg() != PPC::ZERO &&
3245 MI.getOperand(III.ZeroIsSpecialOrig).getReg() != PPC::ZERO8)
3248 // This machine instruction is convertible if it is,
3249 // 1. summing the operands.
3250 // 2. one of the operands is special zero register.
3251 // 3. the operand we are trying to replace is allowed by the MI.
3255 // Check if the DefMI is the add inst and set the ImmMO and RegMO
3257 bool PPCInstrInfo::isDefMIElgibleForForwarding(MachineInstr &DefMI,
3258 const ImmInstrInfo &III,
3259 MachineOperand *&ImmMO,
3260 MachineOperand *&RegMO) const {
3261 unsigned Opc = DefMI.getOpcode();
3262 if (Opc != PPC::ADDItocL && Opc != PPC::ADDI && Opc != PPC::ADDI8)
3265 assert(DefMI.getNumOperands() >= 3 &&
3266 "Add inst must have at least three operands");
3267 RegMO = &DefMI.getOperand(1);
3268 ImmMO = &DefMI.getOperand(2);
3270 // This DefMI is elgible for forwarding if it is:
3272 // 2. one of the operands is Imm/CPI/Global.
3273 return isAnImmediateOperand(*ImmMO);
3276 bool PPCInstrInfo::isRegElgibleForForwarding(
3277 const MachineOperand &RegMO, const MachineInstr &DefMI,
3278 const MachineInstr &MI, bool KillDefMI,
3279 bool &IsFwdFeederRegKilled) const {
3282 // z = lfdx 0, x -> z = lfd imm(y)
3283 // The Reg "y" can be forwarded to the MI(z) only when there is no DEF
3284 // of "y" between the DEF of "x" and "z".
3285 // The query is only valid post RA.
3286 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
3290 unsigned Reg = RegMO.getReg();
3292 // Walking the inst in reverse(MI-->DefMI) to get the last DEF of the Reg.
3293 MachineBasicBlock::const_reverse_iterator It = MI;
3294 MachineBasicBlock::const_reverse_iterator E = MI.getParent()->rend();
3296 for (; It != E; ++It) {
3297 if (It->modifiesRegister(Reg, &getRegisterInfo()) && (&*It) != &DefMI)
3299 else if (It->killsRegister(Reg, &getRegisterInfo()) && (&*It) != &DefMI)
3300 IsFwdFeederRegKilled = true;
3301 // Made it to DefMI without encountering a clobber.
3302 if ((&*It) == &DefMI)
3305 assert((&*It) == &DefMI && "DefMI is missing");
3307 // If DefMI also defines the register to be forwarded, we can only forward it
3308 // if DefMI is being erased.
3309 if (DefMI.modifiesRegister(Reg, &getRegisterInfo()))
3315 bool PPCInstrInfo::isImmElgibleForForwarding(const MachineOperand &ImmMO,
3316 const MachineInstr &DefMI,
3317 const ImmInstrInfo &III,
3318 int64_t &Imm) const {
3319 assert(isAnImmediateOperand(ImmMO) && "ImmMO is NOT an immediate");
3320 if (DefMI.getOpcode() == PPC::ADDItocL) {
3321 // The operand for ADDItocL is CPI, which isn't imm at compiling time,
3322 // However, we know that, it is 16-bit width, and has the alignment of 4.
3323 // Check if the instruction met the requirement.
3324 if (III.ImmMustBeMultipleOf > 4 ||
3325 III.TruncateImmTo || III.ImmWidth != 16)
3328 // Going from XForm to DForm loads means that the displacement needs to be
3329 // not just an immediate but also a multiple of 4, or 16 depending on the
3330 // load. A DForm load cannot be represented if it is a multiple of say 2.
3331 // XForm loads do not have this restriction.
3332 if (ImmMO.isGlobal() &&
3333 ImmMO.getGlobal()->getAlignment() < III.ImmMustBeMultipleOf)
3339 if (ImmMO.isImm()) {
3340 // It is Imm, we need to check if the Imm fit the range.
3341 int64_t Immediate = ImmMO.getImm();
3342 // Sign-extend to 64-bits.
3343 Imm = ((uint64_t)Immediate & ~0x7FFFuLL) != 0 ?
3344 (Immediate | 0xFFFFFFFFFFFF0000) : Immediate;
3346 if (Imm % III.ImmMustBeMultipleOf)
3348 if (III.TruncateImmTo)
3349 Imm &= ((1 << III.TruncateImmTo) - 1);
3350 if (III.SignedImm) {
3351 APInt ActualValue(64, Imm, true);
3352 if (!ActualValue.isSignedIntN(III.ImmWidth))
3355 uint64_t UnsignedMax = (1 << III.ImmWidth) - 1;
3356 if ((uint64_t)Imm > UnsignedMax)
3363 // This ImmMO is forwarded if it meets the requriement describle
3368 // If an X-Form instruction is fed by an add-immediate and one of its operands
3369 // is the literal zero, attempt to forward the source of the add-immediate to
3370 // the corresponding D-Form instruction with the displacement coming from
3371 // the immediate being added.
3372 bool PPCInstrInfo::transformToImmFormFedByAdd(
3373 MachineInstr &MI, const ImmInstrInfo &III, unsigned OpNoForForwarding,
3374 MachineInstr &DefMI, bool KillDefMI) const {
3377 // x = addi reg, imm <----- DefMI
3378 // y = op 0 , x <----- MI
3380 // OpNoForForwarding
3381 // Check if the MI meet the requirement described in the III.
3382 if (!isUseMIElgibleForForwarding(MI, III, OpNoForForwarding))
3385 // Check if the DefMI meet the requirement
3386 // described in the III. If yes, set the ImmMO and RegMO accordingly.
3387 MachineOperand *ImmMO = nullptr;
3388 MachineOperand *RegMO = nullptr;
3389 if (!isDefMIElgibleForForwarding(DefMI, III, ImmMO, RegMO))
3391 assert(ImmMO && RegMO && "Imm and Reg operand must have been set");
3393 // As we get the Imm operand now, we need to check if the ImmMO meet
3394 // the requirement described in the III. If yes set the Imm.
3396 if (!isImmElgibleForForwarding(*ImmMO, DefMI, III, Imm))
3399 bool IsFwdFeederRegKilled = false;
3400 // Check if the RegMO can be forwarded to MI.
3401 if (!isRegElgibleForForwarding(*RegMO, DefMI, MI, KillDefMI,
3402 IsFwdFeederRegKilled))
3405 // Get killed info in case fixup needed after transformation.
3406 unsigned ForwardKilledOperandReg = ~0U;
3407 MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
3408 bool PostRA = !MRI.isSSA();
3409 if (PostRA && MI.getOperand(OpNoForForwarding).isKill())
3410 ForwardKilledOperandReg = MI.getOperand(OpNoForForwarding).getReg();
3412 // We know that, the MI and DefMI both meet the pattern, and
3413 // the Imm also meet the requirement with the new Imm-form.
3414 // It is safe to do the transformation now.
3415 LLVM_DEBUG(dbgs() << "Replacing instruction:\n");
3416 LLVM_DEBUG(MI.dump());
3417 LLVM_DEBUG(dbgs() << "Fed by:\n");
3418 LLVM_DEBUG(DefMI.dump());
3420 // Update the base reg first.
3421 MI.getOperand(III.OpNoForForwarding).ChangeToRegister(RegMO->getReg(),
3425 // Then, update the imm.
3426 if (ImmMO->isImm()) {
3427 // If the ImmMO is Imm, change the operand that has ZERO to that Imm
3429 replaceInstrOperandWithImm(MI, III.ZeroIsSpecialOrig, Imm);
3432 // Otherwise, it is Constant Pool Index(CPI) or Global,
3433 // which is relocation in fact. We need to replace the special zero
3434 // register with ImmMO.
3435 // Before that, we need to fixup the target flags for imm.
3436 // For some reason, we miss to set the flag for the ImmMO if it is CPI.
3437 if (DefMI.getOpcode() == PPC::ADDItocL)
3438 ImmMO->setTargetFlags(PPCII::MO_TOC_LO);
3440 // MI didn't have the interface such as MI.setOperand(i) though
3441 // it has MI.getOperand(i). To repalce the ZERO MachineOperand with
3442 // ImmMO, we need to remove ZERO operand and all the operands behind it,
3443 // and, add the ImmMO, then, move back all the operands behind ZERO.
3444 SmallVector<MachineOperand, 2> MOps;
3445 for (unsigned i = MI.getNumOperands() - 1; i >= III.ZeroIsSpecialOrig; i--) {
3446 MOps.push_back(MI.getOperand(i));
3447 MI.RemoveOperand(i);
3450 // Remove the last MO in the list, which is ZERO operand in fact.
3452 // Add the imm operand.
3453 MI.addOperand(*ImmMO);
3454 // Now add the rest back.
3455 for (auto &MO : MOps)
3459 // Update the opcode.
3460 MI.setDesc(get(III.ImmOpcode));
3462 // Fix up killed/dead flag after transformation.
3464 // x = ADD KilledFwdFeederReg, imm
3465 // n = opn KilledFwdFeederReg(killed), regn
3468 // x = ADD reg(killed), imm
3470 if (IsFwdFeederRegKilled || RegMO->isKill())
3471 fixupIsDeadOrKill(DefMI, MI, RegMO->getReg());
3473 // ForwardKilledOperandReg = ADD reg, imm
3474 // y = XOP 0, ForwardKilledOperandReg(killed)
3475 if (ForwardKilledOperandReg != ~0U)
3476 fixupIsDeadOrKill(DefMI, MI, ForwardKilledOperandReg);
3478 LLVM_DEBUG(dbgs() << "With:\n");
3479 LLVM_DEBUG(MI.dump());
3484 bool PPCInstrInfo::transformToImmFormFedByLI(MachineInstr &MI,
3485 const ImmInstrInfo &III,
3486 unsigned ConstantOpNo,
3487 MachineInstr &DefMI,
3488 int64_t Imm) const {
3489 MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
3490 bool PostRA = !MRI.isSSA();
3491 // Exit early if we can't convert this.
3492 if ((ConstantOpNo != III.OpNoForForwarding) && !III.IsCommutative)
3494 if (Imm % III.ImmMustBeMultipleOf)
3496 if (III.TruncateImmTo)
3497 Imm &= ((1 << III.TruncateImmTo) - 1);
3498 if (III.SignedImm) {
3499 APInt ActualValue(64, Imm, true);
3500 if (!ActualValue.isSignedIntN(III.ImmWidth))
3503 uint64_t UnsignedMax = (1 << III.ImmWidth) - 1;
3504 if ((uint64_t)Imm > UnsignedMax)
3508 // If we're post-RA, the instructions don't agree on whether register zero is
3509 // special, we can transform this as long as the register operand that will
3510 // end up in the location where zero is special isn't R0.
3511 if (PostRA && III.ZeroIsSpecialOrig != III.ZeroIsSpecialNew) {
3512 unsigned PosForOrigZero = III.ZeroIsSpecialOrig ? III.ZeroIsSpecialOrig :
3513 III.ZeroIsSpecialNew + 1;
3514 unsigned OrigZeroReg = MI.getOperand(PosForOrigZero).getReg();
3515 unsigned NewZeroReg = MI.getOperand(III.ZeroIsSpecialNew).getReg();
3516 // If R0 is in the operand where zero is special for the new instruction,
3517 // it is unsafe to transform if the constant operand isn't that operand.
3518 if ((NewZeroReg == PPC::R0 || NewZeroReg == PPC::X0) &&
3519 ConstantOpNo != III.ZeroIsSpecialNew)
3521 if ((OrigZeroReg == PPC::R0 || OrigZeroReg == PPC::X0) &&
3522 ConstantOpNo != PosForOrigZero)
3526 // Get killed info in case fixup needed after transformation.
3527 unsigned ForwardKilledOperandReg = ~0U;
3528 if (PostRA && MI.getOperand(ConstantOpNo).isKill())
3529 ForwardKilledOperandReg = MI.getOperand(ConstantOpNo).getReg();
3531 unsigned Opc = MI.getOpcode();
3532 bool SpecialShift32 = Opc == PPC::SLW || Opc == PPC::SLWo ||
3533 Opc == PPC::SRW || Opc == PPC::SRWo ||
3534 Opc == PPC::SLW8 || Opc == PPC::SLW8o ||
3535 Opc == PPC::SRW8 || Opc == PPC::SRW8o;
3536 bool SpecialShift64 =
3537 Opc == PPC::SLD || Opc == PPC::SLDo || Opc == PPC::SRD || Opc == PPC::SRDo;
3538 bool SetCR = Opc == PPC::SLWo || Opc == PPC::SRWo ||
3539 Opc == PPC::SLDo || Opc == PPC::SRDo;
3541 Opc == PPC::SRW || Opc == PPC::SRWo || Opc == PPC::SRD || Opc == PPC::SRDo;
3543 MI.setDesc(get(III.ImmOpcode));
3544 if (ConstantOpNo == III.OpNoForForwarding) {
3545 // Converting shifts to immediate form is a bit tricky since they may do
3546 // one of three things:
3547 // 1. If the shift amount is between OpSize and 2*OpSize, the result is zero
3548 // 2. If the shift amount is zero, the result is unchanged (save for maybe
3550 // 3. If the shift amount is in [1, OpSize), it's just a shift
3551 if (SpecialShift32 || SpecialShift64) {
3552 LoadImmediateInfo LII;
3555 LII.Is64Bit = SpecialShift64;
3556 uint64_t ShAmt = Imm & (SpecialShift32 ? 0x1F : 0x3F);
3557 if (Imm & (SpecialShift32 ? 0x20 : 0x40))
3558 replaceInstrWithLI(MI, LII);
3559 // Shifts by zero don't change the value. If we don't need to set CR0,
3560 // just convert this to a COPY. Can't do this post-RA since we've already
3561 // cleaned up the copies.
3562 else if (!SetCR && ShAmt == 0 && !PostRA) {
3563 MI.RemoveOperand(2);
3564 MI.setDesc(get(PPC::COPY));
3566 // The 32 bit and 64 bit instructions are quite different.
3567 if (SpecialShift32) {
3568 // Left shifts use (N, 0, 31-N), right shifts use (32-N, N, 31).
3569 uint64_t SH = RightShift ? 32 - ShAmt : ShAmt;
3570 uint64_t MB = RightShift ? ShAmt : 0;
3571 uint64_t ME = RightShift ? 31 : 31 - ShAmt;
3572 replaceInstrOperandWithImm(MI, III.OpNoForForwarding, SH);
3573 MachineInstrBuilder(*MI.getParent()->getParent(), MI).addImm(MB)
3576 // Left shifts use (N, 63-N), right shifts use (64-N, N).
3577 uint64_t SH = RightShift ? 64 - ShAmt : ShAmt;
3578 uint64_t ME = RightShift ? ShAmt : 63 - ShAmt;
3579 replaceInstrOperandWithImm(MI, III.OpNoForForwarding, SH);
3580 MachineInstrBuilder(*MI.getParent()->getParent(), MI).addImm(ME);
3584 replaceInstrOperandWithImm(MI, ConstantOpNo, Imm);
3586 // Convert commutative instructions (switch the operands and convert the
3587 // desired one to an immediate.
3588 else if (III.IsCommutative) {
3589 replaceInstrOperandWithImm(MI, ConstantOpNo, Imm);
3590 swapMIOperands(MI, ConstantOpNo, III.OpNoForForwarding);
3592 llvm_unreachable("Should have exited early!");
3594 // For instructions for which the constant register replaces a different
3595 // operand than where the immediate goes, we need to swap them.
3596 if (III.OpNoForForwarding != III.ImmOpNo)
3597 swapMIOperands(MI, III.OpNoForForwarding, III.ImmOpNo);
3599 // If the special R0/X0 register index are different for original instruction
3600 // and new instruction, we need to fix up the register class in new
3602 if (!PostRA && III.ZeroIsSpecialOrig != III.ZeroIsSpecialNew) {
3603 if (III.ZeroIsSpecialNew) {
3604 // If operand at III.ZeroIsSpecialNew is physical reg(eg: ZERO/ZERO8), no
3605 // need to fix up register class.
3606 unsigned RegToModify = MI.getOperand(III.ZeroIsSpecialNew).getReg();
3607 if (TargetRegisterInfo::isVirtualRegister(RegToModify)) {
3608 const TargetRegisterClass *NewRC =
3609 MRI.getRegClass(RegToModify)->hasSuperClassEq(&PPC::GPRCRegClass) ?
3610 &PPC::GPRC_and_GPRC_NOR0RegClass : &PPC::G8RC_and_G8RC_NOX0RegClass;
3611 MRI.setRegClass(RegToModify, NewRC);
3616 // Fix up killed/dead flag after transformation.
3618 // ForwardKilledOperandReg = LI imm
3619 // y = XOP reg, ForwardKilledOperandReg(killed)
3620 if (ForwardKilledOperandReg != ~0U)
3621 fixupIsDeadOrKill(DefMI, MI, ForwardKilledOperandReg);
3625 const TargetRegisterClass *
3626 PPCInstrInfo::updatedRC(const TargetRegisterClass *RC) const {
3627 if (Subtarget.hasVSX() && RC == &PPC::VRRCRegClass)
3628 return &PPC::VSRCRegClass;
3632 int PPCInstrInfo::getRecordFormOpcode(unsigned Opcode) {
3633 return PPC::getRecordFormOpcode(Opcode);
3636 // This function returns true if the machine instruction
3637 // always outputs a value by sign-extending a 32 bit value,
3638 // i.e. 0 to 31-th bits are same as 32-th bit.
3639 static bool isSignExtendingOp(const MachineInstr &MI) {
3640 int Opcode = MI.getOpcode();
3641 if (Opcode == PPC::LI || Opcode == PPC::LI8 ||
3642 Opcode == PPC::LIS || Opcode == PPC::LIS8 ||
3643 Opcode == PPC::SRAW || Opcode == PPC::SRAWo ||
3644 Opcode == PPC::SRAWI || Opcode == PPC::SRAWIo ||
3645 Opcode == PPC::LWA || Opcode == PPC::LWAX ||
3646 Opcode == PPC::LWA_32 || Opcode == PPC::LWAX_32 ||
3647 Opcode == PPC::LHA || Opcode == PPC::LHAX ||
3648 Opcode == PPC::LHA8 || Opcode == PPC::LHAX8 ||
3649 Opcode == PPC::LBZ || Opcode == PPC::LBZX ||
3650 Opcode == PPC::LBZ8 || Opcode == PPC::LBZX8 ||
3651 Opcode == PPC::LBZU || Opcode == PPC::LBZUX ||
3652 Opcode == PPC::LBZU8 || Opcode == PPC::LBZUX8 ||
3653 Opcode == PPC::LHZ || Opcode == PPC::LHZX ||
3654 Opcode == PPC::LHZ8 || Opcode == PPC::LHZX8 ||
3655 Opcode == PPC::LHZU || Opcode == PPC::LHZUX ||
3656 Opcode == PPC::LHZU8 || Opcode == PPC::LHZUX8 ||
3657 Opcode == PPC::EXTSB || Opcode == PPC::EXTSBo ||
3658 Opcode == PPC::EXTSH || Opcode == PPC::EXTSHo ||
3659 Opcode == PPC::EXTSB8 || Opcode == PPC::EXTSH8 ||
3660 Opcode == PPC::EXTSW || Opcode == PPC::EXTSWo ||
3661 Opcode == PPC::SETB || Opcode == PPC::SETB8 ||
3662 Opcode == PPC::EXTSH8_32_64 || Opcode == PPC::EXTSW_32_64 ||
3663 Opcode == PPC::EXTSB8_32_64)
3666 if (Opcode == PPC::RLDICL && MI.getOperand(3).getImm() >= 33)
3669 if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINMo ||
3670 Opcode == PPC::RLWNM || Opcode == PPC::RLWNMo) &&
3671 MI.getOperand(3).getImm() > 0 &&
3672 MI.getOperand(3).getImm() <= MI.getOperand(4).getImm())
3678 // This function returns true if the machine instruction
3679 // always outputs zeros in higher 32 bits.
3680 static bool isZeroExtendingOp(const MachineInstr &MI) {
3681 int Opcode = MI.getOpcode();
3682 // The 16-bit immediate is sign-extended in li/lis.
3683 // If the most significant bit is zero, all higher bits are zero.
3684 if (Opcode == PPC::LI || Opcode == PPC::LI8 ||
3685 Opcode == PPC::LIS || Opcode == PPC::LIS8) {
3686 int64_t Imm = MI.getOperand(1).getImm();
3687 if (((uint64_t)Imm & ~0x7FFFuLL) == 0)
3691 // We have some variations of rotate-and-mask instructions
3692 // that clear higher 32-bits.
3693 if ((Opcode == PPC::RLDICL || Opcode == PPC::RLDICLo ||
3694 Opcode == PPC::RLDCL || Opcode == PPC::RLDCLo ||
3695 Opcode == PPC::RLDICL_32_64) &&
3696 MI.getOperand(3).getImm() >= 32)
3699 if ((Opcode == PPC::RLDIC || Opcode == PPC::RLDICo) &&
3700 MI.getOperand(3).getImm() >= 32 &&
3701 MI.getOperand(3).getImm() <= 63 - MI.getOperand(2).getImm())
3704 if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINMo ||
3705 Opcode == PPC::RLWNM || Opcode == PPC::RLWNMo ||
3706 Opcode == PPC::RLWINM8 || Opcode == PPC::RLWNM8) &&
3707 MI.getOperand(3).getImm() <= MI.getOperand(4).getImm())
3710 // There are other instructions that clear higher 32-bits.
3711 if (Opcode == PPC::CNTLZW || Opcode == PPC::CNTLZWo ||
3712 Opcode == PPC::CNTTZW || Opcode == PPC::CNTTZWo ||
3713 Opcode == PPC::CNTLZW8 || Opcode == PPC::CNTTZW8 ||
3714 Opcode == PPC::CNTLZD || Opcode == PPC::CNTLZDo ||
3715 Opcode == PPC::CNTTZD || Opcode == PPC::CNTTZDo ||
3716 Opcode == PPC::POPCNTD || Opcode == PPC::POPCNTW ||
3717 Opcode == PPC::SLW || Opcode == PPC::SLWo ||
3718 Opcode == PPC::SRW || Opcode == PPC::SRWo ||
3719 Opcode == PPC::SLW8 || Opcode == PPC::SRW8 ||
3720 Opcode == PPC::SLWI || Opcode == PPC::SLWIo ||
3721 Opcode == PPC::SRWI || Opcode == PPC::SRWIo ||
3722 Opcode == PPC::LWZ || Opcode == PPC::LWZX ||
3723 Opcode == PPC::LWZU || Opcode == PPC::LWZUX ||
3724 Opcode == PPC::LWBRX || Opcode == PPC::LHBRX ||
3725 Opcode == PPC::LHZ || Opcode == PPC::LHZX ||
3726 Opcode == PPC::LHZU || Opcode == PPC::LHZUX ||
3727 Opcode == PPC::LBZ || Opcode == PPC::LBZX ||
3728 Opcode == PPC::LBZU || Opcode == PPC::LBZUX ||
3729 Opcode == PPC::LWZ8 || Opcode == PPC::LWZX8 ||
3730 Opcode == PPC::LWZU8 || Opcode == PPC::LWZUX8 ||
3731 Opcode == PPC::LWBRX8 || Opcode == PPC::LHBRX8 ||
3732 Opcode == PPC::LHZ8 || Opcode == PPC::LHZX8 ||
3733 Opcode == PPC::LHZU8 || Opcode == PPC::LHZUX8 ||
3734 Opcode == PPC::LBZ8 || Opcode == PPC::LBZX8 ||
3735 Opcode == PPC::LBZU8 || Opcode == PPC::LBZUX8 ||
3736 Opcode == PPC::ANDIo || Opcode == PPC::ANDISo ||
3737 Opcode == PPC::ROTRWI || Opcode == PPC::ROTRWIo ||
3738 Opcode == PPC::EXTLWI || Opcode == PPC::EXTLWIo ||
3739 Opcode == PPC::MFVSRWZ)
3745 // This function returns true if the input MachineInstr is a TOC save
3747 bool PPCInstrInfo::isTOCSaveMI(const MachineInstr &MI) const {
3748 if (!MI.getOperand(1).isImm() || !MI.getOperand(2).isReg())
3750 unsigned TOCSaveOffset = Subtarget.getFrameLowering()->getTOCSaveOffset();
3751 unsigned StackOffset = MI.getOperand(1).getImm();
3752 unsigned StackReg = MI.getOperand(2).getReg();
3753 if (StackReg == PPC::X1 && StackOffset == TOCSaveOffset)
3759 // We limit the max depth to track incoming values of PHIs or binary ops
3760 // (e.g. AND) to avoid excessive cost.
3761 const unsigned MAX_DEPTH = 1;
3764 PPCInstrInfo::isSignOrZeroExtended(const MachineInstr &MI, bool SignExt,
3765 const unsigned Depth) const {
3766 const MachineFunction *MF = MI.getParent()->getParent();
3767 const MachineRegisterInfo *MRI = &MF->getRegInfo();
3769 // If we know this instruction returns sign- or zero-extended result,
3771 if (SignExt ? isSignExtendingOp(MI):
3772 isZeroExtendingOp(MI))
3775 switch (MI.getOpcode()) {
3777 unsigned SrcReg = MI.getOperand(1).getReg();
3779 // In both ELFv1 and v2 ABI, method parameters and the return value
3780 // are sign- or zero-extended.
3781 if (MF->getSubtarget<PPCSubtarget>().isSVR4ABI()) {
3782 const PPCFunctionInfo *FuncInfo = MF->getInfo<PPCFunctionInfo>();
3783 // We check the ZExt/SExt flags for a method parameter.
3784 if (MI.getParent()->getBasicBlock() ==
3785 &MF->getFunction().getEntryBlock()) {
3786 unsigned VReg = MI.getOperand(0).getReg();
3787 if (MF->getRegInfo().isLiveIn(VReg))
3788 return SignExt ? FuncInfo->isLiveInSExt(VReg) :
3789 FuncInfo->isLiveInZExt(VReg);
3792 // For a method return value, we check the ZExt/SExt flags in attribute.
3793 // We assume the following code sequence for method call.
3794 // ADJCALLSTACKDOWN 32, implicit dead %r1, implicit %r1
3795 // BL8_NOP @func,...
3796 // ADJCALLSTACKUP 32, 0, implicit dead %r1, implicit %r1
3797 // %5 = COPY %x3; G8RC:%5
3798 if (SrcReg == PPC::X3) {
3799 const MachineBasicBlock *MBB = MI.getParent();
3800 MachineBasicBlock::const_instr_iterator II =
3801 MachineBasicBlock::const_instr_iterator(&MI);
3802 if (II != MBB->instr_begin() &&
3803 (--II)->getOpcode() == PPC::ADJCALLSTACKUP) {
3804 const MachineInstr &CallMI = *(--II);
3805 if (CallMI.isCall() && CallMI.getOperand(0).isGlobal()) {
3806 const Function *CalleeFn =
3807 dyn_cast<Function>(CallMI.getOperand(0).getGlobal());
3810 const IntegerType *IntTy =
3811 dyn_cast<IntegerType>(CalleeFn->getReturnType());
3812 const AttributeSet &Attrs =
3813 CalleeFn->getAttributes().getRetAttributes();
3814 if (IntTy && IntTy->getBitWidth() <= 32)
3815 return Attrs.hasAttribute(SignExt ? Attribute::SExt :
3822 // If this is a copy from another register, we recursively check source.
3823 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
3825 const MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
3827 return isSignOrZeroExtended(*SrcMI, SignExt, Depth);
3844 // logical operation with 16-bit immediate does not change the upper bits.
3845 // So, we track the operand register as we do for register copy.
3846 unsigned SrcReg = MI.getOperand(1).getReg();
3847 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
3849 const MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
3851 return isSignOrZeroExtended(*SrcMI, SignExt, Depth);
3856 // If all incoming values are sign-/zero-extended,
3857 // the output of OR, ISEL or PHI is also sign-/zero-extended.
3862 if (Depth >= MAX_DEPTH)
3865 // The input registers for PHI are operand 1, 3, ...
3866 // The input registers for others are operand 1 and 2.
3867 unsigned E = 3, D = 1;
3868 if (MI.getOpcode() == PPC::PHI) {
3869 E = MI.getNumOperands();
3873 for (unsigned I = 1; I != E; I += D) {
3874 if (MI.getOperand(I).isReg()) {
3875 unsigned SrcReg = MI.getOperand(I).getReg();
3876 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
3878 const MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
3879 if (SrcMI == NULL || !isSignOrZeroExtended(*SrcMI, SignExt, Depth+1))
3888 // If at least one of the incoming values of an AND is zero extended
3889 // then the output is also zero-extended. If both of the incoming values
3890 // are sign-extended then the output is also sign extended.
3893 if (Depth >= MAX_DEPTH)
3896 assert(MI.getOperand(1).isReg() && MI.getOperand(2).isReg());
3898 unsigned SrcReg1 = MI.getOperand(1).getReg();
3899 unsigned SrcReg2 = MI.getOperand(2).getReg();
3901 if (!TargetRegisterInfo::isVirtualRegister(SrcReg1) ||
3902 !TargetRegisterInfo::isVirtualRegister(SrcReg2))
3905 const MachineInstr *MISrc1 = MRI->getVRegDef(SrcReg1);
3906 const MachineInstr *MISrc2 = MRI->getVRegDef(SrcReg2);
3907 if (!MISrc1 || !MISrc2)
3911 return isSignOrZeroExtended(*MISrc1, SignExt, Depth+1) &&
3912 isSignOrZeroExtended(*MISrc2, SignExt, Depth+1);
3914 return isSignOrZeroExtended(*MISrc1, SignExt, Depth+1) ||
3915 isSignOrZeroExtended(*MISrc2, SignExt, Depth+1);
3924 bool PPCInstrInfo::isBDNZ(unsigned Opcode) const {
3925 return (Opcode == (Subtarget.isPPC64() ? PPC::BDNZ8 : PPC::BDNZ));
3928 bool PPCInstrInfo::analyzeLoop(MachineLoop &L, MachineInstr *&IndVarInst,
3929 MachineInstr *&CmpInst) const {
3930 MachineBasicBlock *LoopEnd = L.getBottomBlock();
3931 MachineBasicBlock::iterator I = LoopEnd->getFirstTerminator();
3932 // We really "analyze" only CTR loops right now.
3933 if (I != LoopEnd->end() && isBDNZ(I->getOpcode())) {
3934 IndVarInst = nullptr;
3942 PPCInstrInfo::findLoopInstr(MachineBasicBlock &PreHeader) const {
3944 unsigned LOOPi = (Subtarget.isPPC64() ? PPC::MTCTR8loop : PPC::MTCTRloop);
3946 // The loop set-up instruction should be in preheader
3947 for (auto &I : PreHeader.instrs())
3948 if (I.getOpcode() == LOOPi)
3953 unsigned PPCInstrInfo::reduceLoopCount(
3954 MachineBasicBlock &MBB, MachineBasicBlock &PreHeader, MachineInstr *IndVar,
3955 MachineInstr &Cmp, SmallVectorImpl<MachineOperand> &Cond,
3956 SmallVectorImpl<MachineInstr *> &PrevInsts, unsigned Iter,
3957 unsigned MaxIter) const {
3958 // We expect a hardware loop currently. This means that IndVar is set
3959 // to null, and the compare is the ENDLOOP instruction.
3960 assert((!IndVar) && isBDNZ(Cmp.getOpcode()) && "Expecting a CTR loop");
3961 MachineFunction *MF = MBB.getParent();
3962 DebugLoc DL = Cmp.getDebugLoc();
3963 MachineInstr *Loop = findLoopInstr(PreHeader);
3966 unsigned LoopCountReg = Loop->getOperand(0).getReg();
3967 MachineRegisterInfo &MRI = MF->getRegInfo();
3968 MachineInstr *LoopCount = MRI.getUniqueVRegDef(LoopCountReg);
3972 // If the loop trip count is a compile-time value, then just change the
3974 if (LoopCount->getOpcode() == PPC::LI8 || LoopCount->getOpcode() == PPC::LI) {
3975 int64_t Offset = LoopCount->getOperand(1).getImm();
3977 LoopCount->eraseFromParent();
3978 Loop->eraseFromParent();
3981 LoopCount->getOperand(1).setImm(Offset - 1);
3985 // The loop trip count is a run-time value.
3986 // We need to subtract one from the trip count,
3987 // and insert branch later to check if we're done with the loop.
3989 // Since BDZ/BDZ8 that we will insert will also decrease the ctr by 1,
3990 // so we don't need to generate any thing here.
3991 Cond.push_back(MachineOperand::CreateImm(0));
3992 Cond.push_back(MachineOperand::CreateReg(
3993 Subtarget.isPPC64() ? PPC::CTR8 : PPC::CTR, true));
3994 return LoopCountReg;
3997 // Return true if get the base operand, byte offset of an instruction and the
3998 // memory width. Width is the size of memory that is being loaded/stored.
3999 bool PPCInstrInfo::getMemOperandWithOffsetWidth(
4000 const MachineInstr &LdSt,
4001 const MachineOperand *&BaseReg,
4004 const TargetRegisterInfo *TRI) const {
4005 assert(LdSt.mayLoadOrStore() && "Expected a memory operation.");
4007 // Handle only loads/stores with base register followed by immediate offset.
4008 if (LdSt.getNumExplicitOperands() != 3)
4010 if (!LdSt.getOperand(1).isImm() || !LdSt.getOperand(2).isReg())
4013 if (!LdSt.hasOneMemOperand())
4016 Width = (*LdSt.memoperands_begin())->getSize();
4017 Offset = LdSt.getOperand(1).getImm();
4018 BaseReg = &LdSt.getOperand(2);
4022 bool PPCInstrInfo::areMemAccessesTriviallyDisjoint(
4023 const MachineInstr &MIa, const MachineInstr &MIb,
4024 AliasAnalysis * /*AA*/) const {
4025 assert(MIa.mayLoadOrStore() && "MIa must be a load or store.");
4026 assert(MIb.mayLoadOrStore() && "MIb must be a load or store.");
4028 if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects() ||
4029 MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef())
4032 // Retrieve the base register, offset from the base register and width. Width
4033 // is the size of memory that is being loaded/stored (e.g. 1, 2, 4). If
4034 // base registers are identical, and the offset of a lower memory access +
4035 // the width doesn't overlap the offset of a higher memory access,
4036 // then the memory accesses are different.
4037 const TargetRegisterInfo *TRI = &getRegisterInfo();
4038 const MachineOperand *BaseOpA = nullptr, *BaseOpB = nullptr;
4039 int64_t OffsetA = 0, OffsetB = 0;
4040 unsigned int WidthA = 0, WidthB = 0;
4041 if (getMemOperandWithOffsetWidth(MIa, BaseOpA, OffsetA, WidthA, TRI) &&
4042 getMemOperandWithOffsetWidth(MIb, BaseOpB, OffsetB, WidthB, TRI)) {
4043 if (BaseOpA->isIdenticalTo(*BaseOpB)) {
4044 int LowOffset = std::min(OffsetA, OffsetB);
4045 int HighOffset = std::max(OffsetA, OffsetB);
4046 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
4047 if (LowOffset + LowWidth <= HighOffset)