1 //===-- PPCISelDAGToDAG.cpp - PPC --pattern matching inst selector --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a pattern matching instruction selector for PowerPC,
11 // converting from a legalized dag to a PPC dag.
13 //===----------------------------------------------------------------------===//
15 #include "MCTargetDesc/PPCMCTargetDesc.h"
16 #include "MCTargetDesc/PPCPredicates.h"
18 #include "PPCISelLowering.h"
19 #include "PPCMachineFunctionInfo.h"
20 #include "PPCSubtarget.h"
21 #include "PPCTargetMachine.h"
22 #include "llvm/ADT/APInt.h"
23 #include "llvm/ADT/DenseMap.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/SmallPtrSet.h"
26 #include "llvm/ADT/SmallVector.h"
27 #include "llvm/ADT/Statistic.h"
28 #include "llvm/Analysis/BranchProbabilityInfo.h"
29 #include "llvm/CodeGen/FunctionLoweringInfo.h"
30 #include "llvm/CodeGen/ISDOpcodes.h"
31 #include "llvm/CodeGen/MachineBasicBlock.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/MachineValueType.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/CodeGen/SelectionDAGISel.h"
38 #include "llvm/CodeGen/SelectionDAGNodes.h"
39 #include "llvm/CodeGen/ValueTypes.h"
40 #include "llvm/IR/BasicBlock.h"
41 #include "llvm/IR/DebugLoc.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/GlobalValue.h"
44 #include "llvm/IR/InlineAsm.h"
45 #include "llvm/IR/InstrTypes.h"
46 #include "llvm/IR/Module.h"
47 #include "llvm/Support/Casting.h"
48 #include "llvm/Support/CodeGen.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/KnownBits.h"
54 #include "llvm/Support/MathExtras.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Target/TargetInstrInfo.h"
57 #include "llvm/Target/TargetRegisterInfo.h"
70 #define DEBUG_TYPE "ppc-codegen"
72 STATISTIC(NumSextSetcc,
73 "Number of (sext(setcc)) nodes expanded into GPR sequence.");
74 STATISTIC(NumZextSetcc,
75 "Number of (zext(setcc)) nodes expanded into GPR sequence.");
76 STATISTIC(SignExtensionsAdded,
77 "Number of sign extensions for compare inputs added.");
78 STATISTIC(ZeroExtensionsAdded,
79 "Number of zero extensions for compare inputs added.");
80 STATISTIC(NumLogicOpsOnComparison,
81 "Number of logical ops on i1 values calculated in GPR.");
82 STATISTIC(OmittedForNonExtendUses,
83 "Number of compares not eliminated as they have non-extending uses.");
85 // FIXME: Remove this once the bug has been fixed!
86 cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug",
87 cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden);
90 UseBitPermRewriter("ppc-use-bit-perm-rewriter", cl::init(true),
91 cl::desc("use aggressive ppc isel for bit permutations"),
93 static cl::opt<bool> BPermRewriterNoMasking(
94 "ppc-bit-perm-rewriter-stress-rotates",
95 cl::desc("stress rotate selection in aggressive ppc isel for "
99 static cl::opt<bool> EnableBranchHint(
100 "ppc-use-branch-hint", cl::init(true),
101 cl::desc("Enable static hinting of branches on ppc"),
106 //===--------------------------------------------------------------------===//
107 /// PPCDAGToDAGISel - PPC specific code to select PPC machine
108 /// instructions for SelectionDAG operations.
110 class PPCDAGToDAGISel : public SelectionDAGISel {
111 const PPCTargetMachine &TM;
112 const PPCSubtarget *PPCSubTarget;
113 const PPCTargetLowering *PPCLowering;
114 unsigned GlobalBaseReg;
117 explicit PPCDAGToDAGISel(PPCTargetMachine &tm)
118 : SelectionDAGISel(tm), TM(tm) {}
120 bool runOnMachineFunction(MachineFunction &MF) override {
121 // Make sure we re-emit a set of the global base reg if necessary
123 PPCSubTarget = &MF.getSubtarget<PPCSubtarget>();
124 PPCLowering = PPCSubTarget->getTargetLowering();
125 SelectionDAGISel::runOnMachineFunction(MF);
127 if (!PPCSubTarget->isSVR4ABI())
128 InsertVRSaveCode(MF);
133 void PreprocessISelDAG() override;
134 void PostprocessISelDAG() override;
136 /// getI32Imm - Return a target constant with the specified value, of type
138 inline SDValue getI32Imm(unsigned Imm, const SDLoc &dl) {
139 return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
142 /// getI64Imm - Return a target constant with the specified value, of type
144 inline SDValue getI64Imm(uint64_t Imm, const SDLoc &dl) {
145 return CurDAG->getTargetConstant(Imm, dl, MVT::i64);
148 /// getSmallIPtrImm - Return a target constant of pointer type.
149 inline SDValue getSmallIPtrImm(unsigned Imm, const SDLoc &dl) {
150 return CurDAG->getTargetConstant(
151 Imm, dl, PPCLowering->getPointerTy(CurDAG->getDataLayout()));
154 /// isRotateAndMask - Returns true if Mask and Shift can be folded into a
155 /// rotate and mask opcode and mask operation.
156 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask,
157 unsigned &SH, unsigned &MB, unsigned &ME);
159 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
160 /// base register. Return the virtual register that holds this value.
161 SDNode *getGlobalBaseReg();
163 void selectFrameIndex(SDNode *SN, SDNode *N, unsigned Offset = 0);
165 // Select - Convert the specified operand from a target-independent to a
166 // target-specific node if it hasn't already been changed.
167 void Select(SDNode *N) override;
169 bool tryBitfieldInsert(SDNode *N);
170 bool tryBitPermutation(SDNode *N);
172 /// SelectCC - Select a comparison of the specified values with the
173 /// specified condition code, returning the CR# of the expression.
174 SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC,
177 /// SelectAddrImm - Returns true if the address N can be represented by
178 /// a base register plus a signed 16-bit displacement [r+imm].
179 bool SelectAddrImm(SDValue N, SDValue &Disp,
181 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, false);
184 /// SelectAddrImmOffs - Return true if the operand is valid for a preinc
185 /// immediate field. Note that the operand at this point is already the
186 /// result of a prior SelectAddressRegImm call.
187 bool SelectAddrImmOffs(SDValue N, SDValue &Out) const {
188 if (N.getOpcode() == ISD::TargetConstant ||
189 N.getOpcode() == ISD::TargetGlobalAddress) {
197 /// SelectAddrIdx - Given the specified addressed, check to see if it can be
198 /// represented as an indexed [r+r] operation. Returns false if it can
199 /// be represented by [r+imm], which are preferred.
200 bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) {
201 return PPCLowering->SelectAddressRegReg(N, Base, Index, *CurDAG);
204 /// SelectAddrIdxOnly - Given the specified addressed, force it to be
205 /// represented as an indexed [r+r] operation.
206 bool SelectAddrIdxOnly(SDValue N, SDValue &Base, SDValue &Index) {
207 return PPCLowering->SelectAddressRegRegOnly(N, Base, Index, *CurDAG);
210 /// SelectAddrImmX4 - Returns true if the address N can be represented by
211 /// a base register plus a signed 16-bit displacement that is a multiple of 4.
212 /// Suitable for use by STD and friends.
213 bool SelectAddrImmX4(SDValue N, SDValue &Disp, SDValue &Base) {
214 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, true);
217 // Select an address into a single register.
218 bool SelectAddr(SDValue N, SDValue &Base) {
223 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
224 /// inline asm expressions. It is always correct to compute the value into
225 /// a register. The case of adding a (possibly relocatable) constant to a
226 /// register can be improved, but it is wrong to substitute Reg+Reg for
227 /// Reg in an asm, because the load or store opcode would have to change.
228 bool SelectInlineAsmMemoryOperand(const SDValue &Op,
229 unsigned ConstraintID,
230 std::vector<SDValue> &OutOps) override {
231 switch(ConstraintID) {
233 errs() << "ConstraintID: " << ConstraintID << "\n";
234 llvm_unreachable("Unexpected asm memory constraint");
235 case InlineAsm::Constraint_es:
236 case InlineAsm::Constraint_i:
237 case InlineAsm::Constraint_m:
238 case InlineAsm::Constraint_o:
239 case InlineAsm::Constraint_Q:
240 case InlineAsm::Constraint_Z:
241 case InlineAsm::Constraint_Zy:
242 // We need to make sure that this one operand does not end up in r0
243 // (because we might end up lowering this as 0(%op)).
244 const TargetRegisterInfo *TRI = PPCSubTarget->getRegisterInfo();
245 const TargetRegisterClass *TRC = TRI->getPointerRegClass(*MF, /*Kind=*/1);
247 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32);
249 SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
250 dl, Op.getValueType(),
253 OutOps.push_back(NewOp);
259 void InsertVRSaveCode(MachineFunction &MF);
261 StringRef getPassName() const override {
262 return "PowerPC DAG->DAG Pattern Instruction Selection";
265 // Include the pieces autogenerated from the target description.
266 #include "PPCGenDAGISel.inc"
269 // Conversion type for interpreting results of a 32-bit instruction as
270 // a 64-bit value or vice versa.
271 enum ExtOrTruncConversion { Ext, Trunc };
273 // Modifiers to guide how an ISD::SETCC node's result is to be computed
275 // ZExtOrig - use the original condition code, zero-extend value
276 // ZExtInvert - invert the condition code, zero-extend value
277 // SExtOrig - use the original condition code, sign-extend value
278 // SExtInvert - invert the condition code, sign-extend value
279 enum SetccInGPROpts { ZExtOrig, ZExtInvert, SExtOrig, SExtInvert };
281 bool trySETCC(SDNode *N);
282 bool tryEXTEND(SDNode *N);
283 bool tryLogicOpOfCompares(SDNode *N);
284 SDValue computeLogicOpInGPR(SDValue LogicOp);
285 SDValue signExtendInputIfNeeded(SDValue Input);
286 SDValue zeroExtendInputIfNeeded(SDValue Input);
287 SDValue addExtOrTrunc(SDValue NatWidthRes, ExtOrTruncConversion Conv);
288 SDValue get32BitZExtCompare(SDValue LHS, SDValue RHS, ISD::CondCode CC,
289 int64_t RHSValue, SDLoc dl);
290 SDValue get32BitSExtCompare(SDValue LHS, SDValue RHS, ISD::CondCode CC,
291 int64_t RHSValue, SDLoc dl);
292 SDValue get64BitZExtCompare(SDValue LHS, SDValue RHS, ISD::CondCode CC,
293 int64_t RHSValue, SDLoc dl);
294 SDValue get64BitSExtCompare(SDValue LHS, SDValue RHS, ISD::CondCode CC,
295 int64_t RHSValue, SDLoc dl);
296 SDValue getSETCCInGPR(SDValue Compare, SetccInGPROpts ConvOpts);
298 void PeepholePPC64();
299 void PeepholePPC64ZExt();
300 void PeepholeCROps();
302 SDValue combineToCMPB(SDNode *N);
303 void foldBoolExts(SDValue &Res, SDNode *&N);
305 bool AllUsersSelectZero(SDNode *N);
306 void SwapAllSelectUsers(SDNode *N);
308 void transferMemOperands(SDNode *N, SDNode *Result);
311 } // end anonymous namespace
313 /// InsertVRSaveCode - Once the entire function has been instruction selected,
314 /// all virtual registers are created and all machine instructions are built,
315 /// check to see if we need to save/restore VRSAVE. If so, do it.
316 void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) {
317 // Check to see if this function uses vector registers, which means we have to
318 // save and restore the VRSAVE register and update it with the regs we use.
320 // In this case, there will be virtual registers of vector type created
321 // by the scheduler. Detect them now.
322 bool HasVectorVReg = false;
323 for (unsigned i = 0, e = RegInfo->getNumVirtRegs(); i != e; ++i) {
324 unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
325 if (RegInfo->getRegClass(Reg) == &PPC::VRRCRegClass) {
326 HasVectorVReg = true;
330 if (!HasVectorVReg) return; // nothing to do.
332 // If we have a vector register, we want to emit code into the entry and exit
333 // blocks to save and restore the VRSAVE register. We do this here (instead
334 // of marking all vector instructions as clobbering VRSAVE) for two reasons:
336 // 1. This (trivially) reduces the load on the register allocator, by not
337 // having to represent the live range of the VRSAVE register.
338 // 2. This (more significantly) allows us to create a temporary virtual
339 // register to hold the saved VRSAVE value, allowing this temporary to be
340 // register allocated, instead of forcing it to be spilled to the stack.
342 // Create two vregs - one to hold the VRSAVE register that is live-in to the
343 // function and one for the value after having bits or'd into it.
344 unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
345 unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
347 const TargetInstrInfo &TII = *PPCSubTarget->getInstrInfo();
348 MachineBasicBlock &EntryBB = *Fn.begin();
350 // Emit the following code into the entry block:
351 // InVRSAVE = MFVRSAVE
352 // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE
353 // MTVRSAVE UpdatedVRSAVE
354 MachineBasicBlock::iterator IP = EntryBB.begin(); // Insert Point
355 BuildMI(EntryBB, IP, dl, TII.get(PPC::MFVRSAVE), InVRSAVE);
356 BuildMI(EntryBB, IP, dl, TII.get(PPC::UPDATE_VRSAVE),
357 UpdatedVRSAVE).addReg(InVRSAVE);
358 BuildMI(EntryBB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(UpdatedVRSAVE);
360 // Find all return blocks, outputting a restore in each epilog.
361 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
362 if (BB->isReturnBlock()) {
363 IP = BB->end(); --IP;
365 // Skip over all terminator instructions, which are part of the return
367 MachineBasicBlock::iterator I2 = IP;
368 while (I2 != BB->begin() && (--I2)->isTerminator())
371 // Emit: MTVRSAVE InVRSave
372 BuildMI(*BB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(InVRSAVE);
377 /// getGlobalBaseReg - Output the instructions required to put the
378 /// base address to use for accessing globals into a register.
380 SDNode *PPCDAGToDAGISel::getGlobalBaseReg() {
381 if (!GlobalBaseReg) {
382 const TargetInstrInfo &TII = *PPCSubTarget->getInstrInfo();
383 // Insert the set of GlobalBaseReg into the first MBB of the function
384 MachineBasicBlock &FirstMBB = MF->front();
385 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
386 const Module *M = MF->getFunction()->getParent();
389 if (PPCLowering->getPointerTy(CurDAG->getDataLayout()) == MVT::i32) {
390 if (PPCSubTarget->isTargetELF()) {
391 GlobalBaseReg = PPC::R30;
392 if (M->getPICLevel() == PICLevel::SmallPIC) {
393 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MoveGOTtoLR));
394 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
395 MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true);
397 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
398 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
399 unsigned TempReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
400 BuildMI(FirstMBB, MBBI, dl,
401 TII.get(PPC::UpdateGBR), GlobalBaseReg)
402 .addReg(TempReg, RegState::Define).addReg(GlobalBaseReg);
403 MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true);
407 RegInfo->createVirtualRegister(&PPC::GPRC_and_GPRC_NOR0RegClass);
408 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
409 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
412 GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RC_and_G8RC_NOX0RegClass);
413 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8));
414 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg);
417 return CurDAG->getRegister(GlobalBaseReg,
418 PPCLowering->getPointerTy(CurDAG->getDataLayout()))
422 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
423 /// or 64-bit immediate, and if the value can be accurately represented as a
424 /// sign extension from a 16-bit value. If so, this returns true and the
426 static bool isIntS16Immediate(SDNode *N, short &Imm) {
427 if (N->getOpcode() != ISD::Constant)
430 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
431 if (N->getValueType(0) == MVT::i32)
432 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
434 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
437 static bool isIntS16Immediate(SDValue Op, short &Imm) {
438 return isIntS16Immediate(Op.getNode(), Imm);
441 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
442 /// operand. If so Imm will receive the 32-bit value.
443 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
444 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
445 Imm = cast<ConstantSDNode>(N)->getZExtValue();
451 /// isInt64Immediate - This method tests to see if the node is a 64-bit constant
452 /// operand. If so Imm will receive the 64-bit value.
453 static bool isInt64Immediate(SDNode *N, uint64_t &Imm) {
454 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i64) {
455 Imm = cast<ConstantSDNode>(N)->getZExtValue();
461 // isInt32Immediate - This method tests to see if a constant operand.
462 // If so Imm will receive the 32 bit value.
463 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
464 return isInt32Immediate(N.getNode(), Imm);
467 static unsigned getBranchHint(unsigned PCC, FunctionLoweringInfo *FuncInfo,
468 const SDValue &DestMBB) {
469 assert(isa<BasicBlockSDNode>(DestMBB));
471 if (!FuncInfo->BPI) return PPC::BR_NO_HINT;
473 const BasicBlock *BB = FuncInfo->MBB->getBasicBlock();
474 const TerminatorInst *BBTerm = BB->getTerminator();
476 if (BBTerm->getNumSuccessors() != 2) return PPC::BR_NO_HINT;
478 const BasicBlock *TBB = BBTerm->getSuccessor(0);
479 const BasicBlock *FBB = BBTerm->getSuccessor(1);
481 auto TProb = FuncInfo->BPI->getEdgeProbability(BB, TBB);
482 auto FProb = FuncInfo->BPI->getEdgeProbability(BB, FBB);
484 // We only want to handle cases which are easy to predict at static time, e.g.
485 // C++ throw statement, that is very likely not taken, or calling never
486 // returned function, e.g. stdlib exit(). So we set Threshold to filter
489 // Below is LLVM branch weight table, we only want to handle case 1, 2
491 // Case Taken:Nontaken Example
492 // 1. Unreachable 1048575:1 C++ throw, stdlib exit(),
493 // 2. Invoke-terminating 1:1048575
494 // 3. Coldblock 4:64 __builtin_expect
495 // 4. Loop Branch 124:4 For loop
496 // 5. PH/ZH/FPH 20:12
497 const uint32_t Threshold = 10000;
499 if (std::max(TProb, FProb) / Threshold < std::min(TProb, FProb))
500 return PPC::BR_NO_HINT;
502 DEBUG(dbgs() << "Use branch hint for '" << FuncInfo->Fn->getName() << "::"
503 << BB->getName() << "'\n"
504 << " -> " << TBB->getName() << ": " << TProb << "\n"
505 << " -> " << FBB->getName() << ": " << FProb << "\n");
507 const BasicBlockSDNode *BBDN = cast<BasicBlockSDNode>(DestMBB);
509 // If Dest BasicBlock is False-BasicBlock (FBB), swap branch probabilities,
510 // because we want 'TProb' stands for 'branch probability' to Dest BasicBlock
511 if (BBDN->getBasicBlock()->getBasicBlock() != TBB)
512 std::swap(TProb, FProb);
514 return (TProb > FProb) ? PPC::BR_TAKEN_HINT : PPC::BR_NONTAKEN_HINT;
517 // isOpcWithIntImmediate - This method tests to see if the node is a specific
518 // opcode and that it has a immediate integer right operand.
519 // If so Imm will receive the 32 bit value.
520 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
521 return N->getOpcode() == Opc
522 && isInt32Immediate(N->getOperand(1).getNode(), Imm);
525 void PPCDAGToDAGISel::selectFrameIndex(SDNode *SN, SDNode *N, unsigned Offset) {
527 int FI = cast<FrameIndexSDNode>(N)->getIndex();
528 SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0));
529 unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8;
531 CurDAG->SelectNodeTo(SN, Opc, N->getValueType(0), TFI,
532 getSmallIPtrImm(Offset, dl));
534 ReplaceNode(SN, CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI,
535 getSmallIPtrImm(Offset, dl)));
538 bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask,
539 bool isShiftMask, unsigned &SH,
540 unsigned &MB, unsigned &ME) {
541 // Don't even go down this path for i64, since different logic will be
542 // necessary for rldicl/rldicr/rldimi.
543 if (N->getValueType(0) != MVT::i32)
547 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
548 unsigned Opcode = N->getOpcode();
549 if (N->getNumOperands() != 2 ||
550 !isInt32Immediate(N->getOperand(1).getNode(), Shift) || (Shift > 31))
553 if (Opcode == ISD::SHL) {
554 // apply shift left to mask if it comes first
555 if (isShiftMask) Mask = Mask << Shift;
556 // determine which bits are made indeterminant by shift
557 Indeterminant = ~(0xFFFFFFFFu << Shift);
558 } else if (Opcode == ISD::SRL) {
559 // apply shift right to mask if it comes first
560 if (isShiftMask) Mask = Mask >> Shift;
561 // determine which bits are made indeterminant by shift
562 Indeterminant = ~(0xFFFFFFFFu >> Shift);
563 // adjust for the left rotate
565 } else if (Opcode == ISD::ROTL) {
571 // if the mask doesn't intersect any Indeterminant bits
572 if (Mask && !(Mask & Indeterminant)) {
574 // make sure the mask is still a mask (wrap arounds may not be)
575 return isRunOfOnes(Mask, MB, ME);
580 /// Turn an or of two masked values into the rotate left word immediate then
581 /// mask insert (rlwimi) instruction.
582 bool PPCDAGToDAGISel::tryBitfieldInsert(SDNode *N) {
583 SDValue Op0 = N->getOperand(0);
584 SDValue Op1 = N->getOperand(1);
587 KnownBits LKnown, RKnown;
588 CurDAG->computeKnownBits(Op0, LKnown);
589 CurDAG->computeKnownBits(Op1, RKnown);
591 unsigned TargetMask = LKnown.Zero.getZExtValue();
592 unsigned InsertMask = RKnown.Zero.getZExtValue();
594 if ((TargetMask | InsertMask) == 0xFFFFFFFF) {
595 unsigned Op0Opc = Op0.getOpcode();
596 unsigned Op1Opc = Op1.getOpcode();
597 unsigned Value, SH = 0;
598 TargetMask = ~TargetMask;
599 InsertMask = ~InsertMask;
601 // If the LHS has a foldable shift and the RHS does not, then swap it to the
602 // RHS so that we can fold the shift into the insert.
603 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
604 if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
605 Op0.getOperand(0).getOpcode() == ISD::SRL) {
606 if (Op1.getOperand(0).getOpcode() != ISD::SHL &&
607 Op1.getOperand(0).getOpcode() != ISD::SRL) {
609 std::swap(Op0Opc, Op1Opc);
610 std::swap(TargetMask, InsertMask);
613 } else if (Op0Opc == ISD::SHL || Op0Opc == ISD::SRL) {
614 if (Op1Opc == ISD::AND && Op1.getOperand(0).getOpcode() != ISD::SHL &&
615 Op1.getOperand(0).getOpcode() != ISD::SRL) {
617 std::swap(Op0Opc, Op1Opc);
618 std::swap(TargetMask, InsertMask);
623 if (isRunOfOnes(InsertMask, MB, ME)) {
626 if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) &&
627 isInt32Immediate(Op1.getOperand(1), Value)) {
628 Op1 = Op1.getOperand(0);
629 SH = (Op1Opc == ISD::SHL) ? Value : 32 - Value;
631 if (Op1Opc == ISD::AND) {
632 // The AND mask might not be a constant, and we need to make sure that
633 // if we're going to fold the masking with the insert, all bits not
634 // know to be zero in the mask are known to be one.
636 CurDAG->computeKnownBits(Op1.getOperand(1), MKnown);
637 bool CanFoldMask = InsertMask == MKnown.One.getZExtValue();
639 unsigned SHOpc = Op1.getOperand(0).getOpcode();
640 if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) && CanFoldMask &&
641 isInt32Immediate(Op1.getOperand(0).getOperand(1), Value)) {
642 // Note that Value must be in range here (less than 32) because
643 // otherwise there would not be any bits set in InsertMask.
644 Op1 = Op1.getOperand(0).getOperand(0);
645 SH = (SHOpc == ISD::SHL) ? Value : 32 - Value;
650 SDValue Ops[] = { Op0, Op1, getI32Imm(SH, dl), getI32Imm(MB, dl),
652 ReplaceNode(N, CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops));
659 // Predict the number of instructions that would be generated by calling
661 static unsigned getInt64CountDirect(int64_t Imm) {
662 // Assume no remaining bits.
663 unsigned Remainder = 0;
664 // Assume no shift required.
667 // If it can't be represented as a 32 bit value.
668 if (!isInt<32>(Imm)) {
669 Shift = countTrailingZeros<uint64_t>(Imm);
670 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
672 // If the shifted value fits 32 bits.
673 if (isInt<32>(ImmSh)) {
674 // Go with the shifted value.
677 // Still stuck with a 64 bit value.
684 // Intermediate operand.
687 // Handle first 32 bits.
688 unsigned Lo = Imm & 0xFFFF;
691 if (isInt<16>(Imm)) {
695 // Handle the Hi bits and Lo bits.
702 // If no shift, we're done.
703 if (!Shift) return Result;
705 // If Hi word == Lo word,
706 // we can use rldimi to insert the Lo word into Hi word.
707 if ((unsigned)(Imm & 0xFFFFFFFF) == Remainder) {
712 // Shift for next step if the upper 32-bits were not zero.
716 // Add in the last bits as required.
717 if ((Remainder >> 16) & 0xFFFF)
719 if (Remainder & 0xFFFF)
725 static uint64_t Rot64(uint64_t Imm, unsigned R) {
726 return (Imm << R) | (Imm >> (64 - R));
729 static unsigned getInt64Count(int64_t Imm) {
730 unsigned Count = getInt64CountDirect(Imm);
734 for (unsigned r = 1; r < 63; ++r) {
735 uint64_t RImm = Rot64(Imm, r);
736 unsigned RCount = getInt64CountDirect(RImm) + 1;
737 Count = std::min(Count, RCount);
739 // See comments in getInt64 for an explanation of the logic below.
740 unsigned LS = findLastSet(RImm);
744 uint64_t OnesMask = -(int64_t) (UINT64_C(1) << (LS+1));
745 uint64_t RImmWithOnes = RImm | OnesMask;
747 RCount = getInt64CountDirect(RImmWithOnes) + 1;
748 Count = std::min(Count, RCount);
754 // Select a 64-bit constant. For cost-modeling purposes, getInt64Count
755 // (above) needs to be kept in sync with this function.
756 static SDNode *getInt64Direct(SelectionDAG *CurDAG, const SDLoc &dl,
758 // Assume no remaining bits.
759 unsigned Remainder = 0;
760 // Assume no shift required.
763 // If it can't be represented as a 32 bit value.
764 if (!isInt<32>(Imm)) {
765 Shift = countTrailingZeros<uint64_t>(Imm);
766 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
768 // If the shifted value fits 32 bits.
769 if (isInt<32>(ImmSh)) {
770 // Go with the shifted value.
773 // Still stuck with a 64 bit value.
780 // Intermediate operand.
783 // Handle first 32 bits.
784 unsigned Lo = Imm & 0xFFFF;
785 unsigned Hi = (Imm >> 16) & 0xFFFF;
787 auto getI32Imm = [CurDAG, dl](unsigned Imm) {
788 return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
792 if (isInt<16>(Imm)) {
794 Result = CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(Lo));
796 // Handle the Hi bits.
797 unsigned OpC = Hi ? PPC::LIS8 : PPC::LI8;
798 Result = CurDAG->getMachineNode(OpC, dl, MVT::i64, getI32Imm(Hi));
800 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
801 SDValue(Result, 0), getI32Imm(Lo));
804 Result = CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(Hi));
807 // If no shift, we're done.
808 if (!Shift) return Result;
810 // If Hi word == Lo word,
811 // we can use rldimi to insert the Lo word into Hi word.
812 if ((unsigned)(Imm & 0xFFFFFFFF) == Remainder) {
814 { SDValue(Result, 0), SDValue(Result, 0), getI32Imm(Shift), getI32Imm(0)};
815 return CurDAG->getMachineNode(PPC::RLDIMI, dl, MVT::i64, Ops);
818 // Shift for next step if the upper 32-bits were not zero.
820 Result = CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64,
823 getI32Imm(63 - Shift));
826 // Add in the last bits as required.
827 if ((Hi = (Remainder >> 16) & 0xFFFF)) {
828 Result = CurDAG->getMachineNode(PPC::ORIS8, dl, MVT::i64,
829 SDValue(Result, 0), getI32Imm(Hi));
831 if ((Lo = Remainder & 0xFFFF)) {
832 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
833 SDValue(Result, 0), getI32Imm(Lo));
839 static SDNode *getInt64(SelectionDAG *CurDAG, const SDLoc &dl, int64_t Imm) {
840 unsigned Count = getInt64CountDirect(Imm);
842 return getInt64Direct(CurDAG, dl, Imm);
849 for (unsigned r = 1; r < 63; ++r) {
850 uint64_t RImm = Rot64(Imm, r);
851 unsigned RCount = getInt64CountDirect(RImm) + 1;
852 if (RCount < Count) {
859 // If the immediate to generate has many trailing zeros, it might be
860 // worthwhile to generate a rotated value with too many leading ones
861 // (because that's free with li/lis's sign-extension semantics), and then
862 // mask them off after rotation.
864 unsigned LS = findLastSet(RImm);
865 // We're adding (63-LS) higher-order ones, and we expect to mask them off
866 // after performing the inverse rotation by (64-r). So we need that:
867 // 63-LS == 64-r => LS == r-1
871 uint64_t OnesMask = -(int64_t) (UINT64_C(1) << (LS+1));
872 uint64_t RImmWithOnes = RImm | OnesMask;
874 RCount = getInt64CountDirect(RImmWithOnes) + 1;
875 if (RCount < Count) {
878 MatImm = RImmWithOnes;
884 return getInt64Direct(CurDAG, dl, Imm);
886 auto getI32Imm = [CurDAG, dl](unsigned Imm) {
887 return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
890 SDValue Val = SDValue(getInt64Direct(CurDAG, dl, MatImm), 0);
891 return CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, Val,
892 getI32Imm(64 - RMin), getI32Imm(MaskEnd));
895 // Select a 64-bit constant.
896 static SDNode *getInt64(SelectionDAG *CurDAG, SDNode *N) {
900 int64_t Imm = cast<ConstantSDNode>(N)->getZExtValue();
901 return getInt64(CurDAG, dl, Imm);
906 class BitPermutationSelector {
910 // The bit number in the value, using a convention where bit 0 is the
919 ValueBit(SDValue V, unsigned I, Kind K = Variable)
920 : V(V), Idx(I), K(K) {}
921 ValueBit(Kind K = Variable)
922 : V(SDValue(nullptr, 0)), Idx(UINT32_MAX), K(K) {}
924 bool isZero() const {
925 return K == ConstZero;
928 bool hasValue() const {
929 return K == Variable;
932 SDValue getValue() const {
933 assert(hasValue() && "Cannot get the value of a constant bit");
937 unsigned getValueBitIndex() const {
938 assert(hasValue() && "Cannot get the value bit index of a constant bit");
943 // A bit group has the same underlying value and the same rotate factor.
947 unsigned StartIdx, EndIdx;
949 // This rotation amount assumes that the lower 32 bits of the quantity are
950 // replicated in the high 32 bits by the rotation operator (which is done
951 // by rlwinm and friends in 64-bit mode).
953 // Did converting to Repl32 == true change the rotation factor? If it did,
954 // it decreased it by 32.
956 // Was this group coalesced after setting Repl32 to true?
957 bool Repl32Coalesced;
959 BitGroup(SDValue V, unsigned R, unsigned S, unsigned E)
960 : V(V), RLAmt(R), StartIdx(S), EndIdx(E), Repl32(false), Repl32CR(false),
961 Repl32Coalesced(false) {
962 DEBUG(dbgs() << "\tbit group for " << V.getNode() << " RLAmt = " << R <<
963 " [" << S << ", " << E << "]\n");
967 // Information on each (Value, RLAmt) pair (like the number of groups
968 // associated with each) used to choose the lowering method.
969 struct ValueRotInfo {
971 unsigned RLAmt = std::numeric_limits<unsigned>::max();
972 unsigned NumGroups = 0;
973 unsigned FirstGroupStartIdx = std::numeric_limits<unsigned>::max();
976 ValueRotInfo() = default;
978 // For sorting (in reverse order) by NumGroups, and then by
979 // FirstGroupStartIdx.
980 bool operator < (const ValueRotInfo &Other) const {
981 // We need to sort so that the non-Repl32 come first because, when we're
982 // doing masking, the Repl32 bit groups might be subsumed into the 64-bit
983 // masking operation.
984 if (Repl32 < Other.Repl32)
986 else if (Repl32 > Other.Repl32)
988 else if (NumGroups > Other.NumGroups)
990 else if (NumGroups < Other.NumGroups)
992 else if (FirstGroupStartIdx < Other.FirstGroupStartIdx)
998 using ValueBitsMemoizedValue = std::pair<bool, SmallVector<ValueBit, 64>>;
999 using ValueBitsMemoizer =
1000 DenseMap<SDValue, std::unique_ptr<ValueBitsMemoizedValue>>;
1001 ValueBitsMemoizer Memoizer;
1003 // Return a pair of bool and a SmallVector pointer to a memoization entry.
1004 // The bool is true if something interesting was deduced, otherwise if we're
1005 // providing only a generic representation of V (or something else likewise
1006 // uninteresting for instruction selection) through the SmallVector.
1007 std::pair<bool, SmallVector<ValueBit, 64> *> getValueBits(SDValue V,
1009 auto &ValueEntry = Memoizer[V];
1011 return std::make_pair(ValueEntry->first, &ValueEntry->second);
1012 ValueEntry.reset(new ValueBitsMemoizedValue());
1013 bool &Interesting = ValueEntry->first;
1014 SmallVector<ValueBit, 64> &Bits = ValueEntry->second;
1015 Bits.resize(NumBits);
1017 switch (V.getOpcode()) {
1020 if (isa<ConstantSDNode>(V.getOperand(1))) {
1021 unsigned RotAmt = V.getConstantOperandVal(1);
1023 const auto &LHSBits = *getValueBits(V.getOperand(0), NumBits).second;
1025 for (unsigned i = 0; i < NumBits; ++i)
1026 Bits[i] = LHSBits[i < RotAmt ? i + (NumBits - RotAmt) : i - RotAmt];
1028 return std::make_pair(Interesting = true, &Bits);
1032 if (isa<ConstantSDNode>(V.getOperand(1))) {
1033 unsigned ShiftAmt = V.getConstantOperandVal(1);
1035 const auto &LHSBits = *getValueBits(V.getOperand(0), NumBits).second;
1037 for (unsigned i = ShiftAmt; i < NumBits; ++i)
1038 Bits[i] = LHSBits[i - ShiftAmt];
1040 for (unsigned i = 0; i < ShiftAmt; ++i)
1041 Bits[i] = ValueBit(ValueBit::ConstZero);
1043 return std::make_pair(Interesting = true, &Bits);
1047 if (isa<ConstantSDNode>(V.getOperand(1))) {
1048 unsigned ShiftAmt = V.getConstantOperandVal(1);
1050 const auto &LHSBits = *getValueBits(V.getOperand(0), NumBits).second;
1052 for (unsigned i = 0; i < NumBits - ShiftAmt; ++i)
1053 Bits[i] = LHSBits[i + ShiftAmt];
1055 for (unsigned i = NumBits - ShiftAmt; i < NumBits; ++i)
1056 Bits[i] = ValueBit(ValueBit::ConstZero);
1058 return std::make_pair(Interesting = true, &Bits);
1062 if (isa<ConstantSDNode>(V.getOperand(1))) {
1063 uint64_t Mask = V.getConstantOperandVal(1);
1065 const SmallVector<ValueBit, 64> *LHSBits;
1066 // Mark this as interesting, only if the LHS was also interesting. This
1067 // prevents the overall procedure from matching a single immediate 'and'
1068 // (which is non-optimal because such an and might be folded with other
1069 // things if we don't select it here).
1070 std::tie(Interesting, LHSBits) = getValueBits(V.getOperand(0), NumBits);
1072 for (unsigned i = 0; i < NumBits; ++i)
1073 if (((Mask >> i) & 1) == 1)
1074 Bits[i] = (*LHSBits)[i];
1076 Bits[i] = ValueBit(ValueBit::ConstZero);
1078 return std::make_pair(Interesting, &Bits);
1082 const auto &LHSBits = *getValueBits(V.getOperand(0), NumBits).second;
1083 const auto &RHSBits = *getValueBits(V.getOperand(1), NumBits).second;
1085 bool AllDisjoint = true;
1086 for (unsigned i = 0; i < NumBits; ++i)
1087 if (LHSBits[i].isZero())
1088 Bits[i] = RHSBits[i];
1089 else if (RHSBits[i].isZero())
1090 Bits[i] = LHSBits[i];
1092 AllDisjoint = false;
1099 return std::make_pair(Interesting = true, &Bits);
1103 for (unsigned i = 0; i < NumBits; ++i)
1104 Bits[i] = ValueBit(V, i);
1106 return std::make_pair(Interesting = false, &Bits);
1109 // For each value (except the constant ones), compute the left-rotate amount
1110 // to get it from its original to final position.
1111 void computeRotationAmounts() {
1113 RLAmt.resize(Bits.size());
1114 for (unsigned i = 0; i < Bits.size(); ++i)
1115 if (Bits[i].hasValue()) {
1116 unsigned VBI = Bits[i].getValueBitIndex();
1120 RLAmt[i] = Bits.size() - (VBI - i);
1121 } else if (Bits[i].isZero()) {
1123 RLAmt[i] = UINT32_MAX;
1125 llvm_unreachable("Unknown value bit type");
1129 // Collect groups of consecutive bits with the same underlying value and
1130 // rotation factor. If we're doing late masking, we ignore zeros, otherwise
1131 // they break up groups.
1132 void collectBitGroups(bool LateMask) {
1135 unsigned LastRLAmt = RLAmt[0];
1136 SDValue LastValue = Bits[0].hasValue() ? Bits[0].getValue() : SDValue();
1137 unsigned LastGroupStartIdx = 0;
1138 for (unsigned i = 1; i < Bits.size(); ++i) {
1139 unsigned ThisRLAmt = RLAmt[i];
1140 SDValue ThisValue = Bits[i].hasValue() ? Bits[i].getValue() : SDValue();
1141 if (LateMask && !ThisValue) {
1142 ThisValue = LastValue;
1143 ThisRLAmt = LastRLAmt;
1144 // If we're doing late masking, then the first bit group always starts
1145 // at zero (even if the first bits were zero).
1146 if (BitGroups.empty())
1147 LastGroupStartIdx = 0;
1150 // If this bit has the same underlying value and the same rotate factor as
1151 // the last one, then they're part of the same group.
1152 if (ThisRLAmt == LastRLAmt && ThisValue == LastValue)
1155 if (LastValue.getNode())
1156 BitGroups.push_back(BitGroup(LastValue, LastRLAmt, LastGroupStartIdx,
1158 LastRLAmt = ThisRLAmt;
1159 LastValue = ThisValue;
1160 LastGroupStartIdx = i;
1162 if (LastValue.getNode())
1163 BitGroups.push_back(BitGroup(LastValue, LastRLAmt, LastGroupStartIdx,
1166 if (BitGroups.empty())
1169 // We might be able to combine the first and last groups.
1170 if (BitGroups.size() > 1) {
1171 // If the first and last groups are the same, then remove the first group
1172 // in favor of the last group, making the ending index of the last group
1173 // equal to the ending index of the to-be-removed first group.
1174 if (BitGroups[0].StartIdx == 0 &&
1175 BitGroups[BitGroups.size()-1].EndIdx == Bits.size()-1 &&
1176 BitGroups[0].V == BitGroups[BitGroups.size()-1].V &&
1177 BitGroups[0].RLAmt == BitGroups[BitGroups.size()-1].RLAmt) {
1178 DEBUG(dbgs() << "\tcombining final bit group with initial one\n");
1179 BitGroups[BitGroups.size()-1].EndIdx = BitGroups[0].EndIdx;
1180 BitGroups.erase(BitGroups.begin());
1185 // Take all (SDValue, RLAmt) pairs and sort them by the number of groups
1186 // associated with each. If there is a degeneracy, pick the one that occurs
1187 // first (in the final value).
1188 void collectValueRotInfo() {
1191 for (auto &BG : BitGroups) {
1192 unsigned RLAmtKey = BG.RLAmt + (BG.Repl32 ? 64 : 0);
1193 ValueRotInfo &VRI = ValueRots[std::make_pair(BG.V, RLAmtKey)];
1195 VRI.RLAmt = BG.RLAmt;
1196 VRI.Repl32 = BG.Repl32;
1198 VRI.FirstGroupStartIdx = std::min(VRI.FirstGroupStartIdx, BG.StartIdx);
1201 // Now that we've collected the various ValueRotInfo instances, we need to
1203 ValueRotsVec.clear();
1204 for (auto &I : ValueRots) {
1205 ValueRotsVec.push_back(I.second);
1207 std::sort(ValueRotsVec.begin(), ValueRotsVec.end());
1210 // In 64-bit mode, rlwinm and friends have a rotation operator that
1211 // replicates the low-order 32 bits into the high-order 32-bits. The mask
1212 // indices of these instructions can only be in the lower 32 bits, so they
1213 // can only represent some 64-bit bit groups. However, when they can be used,
1214 // the 32-bit replication can be used to represent, as a single bit group,
1215 // otherwise separate bit groups. We'll convert to replicated-32-bit bit
1216 // groups when possible. Returns true if any of the bit groups were
1218 void assignRepl32BitGroups() {
1219 // If we have bits like this:
1221 // Indices: 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
1222 // V bits: ... 7 6 5 4 3 2 1 0 31 30 29 28 27 26 25 24
1223 // Groups: | RLAmt = 8 | RLAmt = 40 |
1225 // But, making use of a 32-bit operation that replicates the low-order 32
1226 // bits into the high-order 32 bits, this can be one bit group with a RLAmt
1229 auto IsAllLow32 = [this](BitGroup & BG) {
1230 if (BG.StartIdx <= BG.EndIdx) {
1231 for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i) {
1232 if (!Bits[i].hasValue())
1234 if (Bits[i].getValueBitIndex() >= 32)
1238 for (unsigned i = BG.StartIdx; i < Bits.size(); ++i) {
1239 if (!Bits[i].hasValue())
1241 if (Bits[i].getValueBitIndex() >= 32)
1244 for (unsigned i = 0; i <= BG.EndIdx; ++i) {
1245 if (!Bits[i].hasValue())
1247 if (Bits[i].getValueBitIndex() >= 32)
1255 for (auto &BG : BitGroups) {
1256 if (BG.StartIdx < 32 && BG.EndIdx < 32) {
1257 if (IsAllLow32(BG)) {
1258 if (BG.RLAmt >= 32) {
1265 DEBUG(dbgs() << "\t32-bit replicated bit group for " <<
1266 BG.V.getNode() << " RLAmt = " << BG.RLAmt <<
1267 " [" << BG.StartIdx << ", " << BG.EndIdx << "]\n");
1272 // Now walk through the bit groups, consolidating where possible.
1273 for (auto I = BitGroups.begin(); I != BitGroups.end();) {
1274 // We might want to remove this bit group by merging it with the previous
1275 // group (which might be the ending group).
1276 auto IP = (I == BitGroups.begin()) ?
1277 std::prev(BitGroups.end()) : std::prev(I);
1278 if (I->Repl32 && IP->Repl32 && I->V == IP->V && I->RLAmt == IP->RLAmt &&
1279 I->StartIdx == (IP->EndIdx + 1) % 64 && I != IP) {
1281 DEBUG(dbgs() << "\tcombining 32-bit replicated bit group for " <<
1282 I->V.getNode() << " RLAmt = " << I->RLAmt <<
1283 " [" << I->StartIdx << ", " << I->EndIdx <<
1284 "] with group with range [" <<
1285 IP->StartIdx << ", " << IP->EndIdx << "]\n");
1287 IP->EndIdx = I->EndIdx;
1288 IP->Repl32CR = IP->Repl32CR || I->Repl32CR;
1289 IP->Repl32Coalesced = true;
1290 I = BitGroups.erase(I);
1293 // There is a special case worth handling: If there is a single group
1294 // covering the entire upper 32 bits, and it can be merged with both
1295 // the next and previous groups (which might be the same group), then
1296 // do so. If it is the same group (so there will be only one group in
1297 // total), then we need to reverse the order of the range so that it
1298 // covers the entire 64 bits.
1299 if (I->StartIdx == 32 && I->EndIdx == 63) {
1300 assert(std::next(I) == BitGroups.end() &&
1301 "bit group ends at index 63 but there is another?");
1302 auto IN = BitGroups.begin();
1304 if (IP->Repl32 && IN->Repl32 && I->V == IP->V && I->V == IN->V &&
1305 (I->RLAmt % 32) == IP->RLAmt && (I->RLAmt % 32) == IN->RLAmt &&
1306 IP->EndIdx == 31 && IN->StartIdx == 0 && I != IP &&
1309 DEBUG(dbgs() << "\tcombining bit group for " <<
1310 I->V.getNode() << " RLAmt = " << I->RLAmt <<
1311 " [" << I->StartIdx << ", " << I->EndIdx <<
1312 "] with 32-bit replicated groups with ranges [" <<
1313 IP->StartIdx << ", " << IP->EndIdx << "] and [" <<
1314 IN->StartIdx << ", " << IN->EndIdx << "]\n");
1317 // There is only one other group; change it to cover the whole
1318 // range (backward, so that it can still be Repl32 but cover the
1319 // whole 64-bit range).
1322 IP->Repl32CR = IP->Repl32CR || I->RLAmt >= 32;
1323 IP->Repl32Coalesced = true;
1324 I = BitGroups.erase(I);
1326 // There are two separate groups, one before this group and one
1327 // after us (at the beginning). We're going to remove this group,
1328 // but also the group at the very beginning.
1329 IP->EndIdx = IN->EndIdx;
1330 IP->Repl32CR = IP->Repl32CR || IN->Repl32CR || I->RLAmt >= 32;
1331 IP->Repl32Coalesced = true;
1332 I = BitGroups.erase(I);
1333 BitGroups.erase(BitGroups.begin());
1336 // This must be the last group in the vector (and we might have
1337 // just invalidated the iterator above), so break here.
1347 SDValue getI32Imm(unsigned Imm, const SDLoc &dl) {
1348 return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
1351 uint64_t getZerosMask() {
1353 for (unsigned i = 0; i < Bits.size(); ++i) {
1354 if (Bits[i].hasValue())
1356 Mask |= (UINT64_C(1) << i);
1362 // Depending on the number of groups for a particular value, it might be
1363 // better to rotate, mask explicitly (using andi/andis), and then or the
1364 // result. Select this part of the result first.
1365 void SelectAndParts32(const SDLoc &dl, SDValue &Res, unsigned *InstCnt) {
1366 if (BPermRewriterNoMasking)
1369 for (ValueRotInfo &VRI : ValueRotsVec) {
1371 for (unsigned i = 0; i < Bits.size(); ++i) {
1372 if (!Bits[i].hasValue() || Bits[i].getValue() != VRI.V)
1374 if (RLAmt[i] != VRI.RLAmt)
1379 // Compute the masks for andi/andis that would be necessary.
1380 unsigned ANDIMask = (Mask & UINT16_MAX), ANDISMask = Mask >> 16;
1381 assert((ANDIMask != 0 || ANDISMask != 0) &&
1382 "No set bits in mask for value bit groups");
1383 bool NeedsRotate = VRI.RLAmt != 0;
1385 // We're trying to minimize the number of instructions. If we have one
1386 // group, using one of andi/andis can break even. If we have three
1387 // groups, we can use both andi and andis and break even (to use both
1388 // andi and andis we also need to or the results together). We need four
1389 // groups if we also need to rotate. To use andi/andis we need to do more
1390 // than break even because rotate-and-mask instructions tend to be easier
1393 // FIXME: We've biased here against using andi/andis, which is right for
1394 // POWER cores, but not optimal everywhere. For example, on the A2,
1395 // andi/andis have single-cycle latency whereas the rotate-and-mask
1396 // instructions take two cycles, and it would be better to bias toward
1397 // andi/andis in break-even cases.
1399 unsigned NumAndInsts = (unsigned) NeedsRotate +
1400 (unsigned) (ANDIMask != 0) +
1401 (unsigned) (ANDISMask != 0) +
1402 (unsigned) (ANDIMask != 0 && ANDISMask != 0) +
1403 (unsigned) (bool) Res;
1405 DEBUG(dbgs() << "\t\trotation groups for " << VRI.V.getNode() <<
1406 " RL: " << VRI.RLAmt << ":" <<
1407 "\n\t\t\tisel using masking: " << NumAndInsts <<
1408 " using rotates: " << VRI.NumGroups << "\n");
1410 if (NumAndInsts >= VRI.NumGroups)
1413 DEBUG(dbgs() << "\t\t\t\tusing masking\n");
1415 if (InstCnt) *InstCnt += NumAndInsts;
1420 { VRI.V, getI32Imm(VRI.RLAmt, dl), getI32Imm(0, dl),
1421 getI32Imm(31, dl) };
1422 VRot = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32,
1428 SDValue ANDIVal, ANDISVal;
1430 ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32,
1431 VRot, getI32Imm(ANDIMask, dl)), 0);
1433 ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32,
1434 VRot, getI32Imm(ANDISMask, dl)), 0);
1438 TotalVal = ANDISVal;
1442 TotalVal = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32,
1443 ANDIVal, ANDISVal), 0);
1448 Res = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32,
1451 // Now, remove all groups with this underlying value and rotation
1453 eraseMatchingBitGroups([VRI](const BitGroup &BG) {
1454 return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt;
1459 // Instruction selection for the 32-bit case.
1460 SDNode *Select32(SDNode *N, bool LateMask, unsigned *InstCnt) {
1464 if (InstCnt) *InstCnt = 0;
1466 // Take care of cases that should use andi/andis first.
1467 SelectAndParts32(dl, Res, InstCnt);
1469 // If we've not yet selected a 'starting' instruction, and we have no zeros
1470 // to fill in, select the (Value, RLAmt) with the highest priority (largest
1471 // number of groups), and start with this rotated value.
1472 if ((!HasZeros || LateMask) && !Res) {
1473 ValueRotInfo &VRI = ValueRotsVec[0];
1475 if (InstCnt) *InstCnt += 1;
1477 { VRI.V, getI32Imm(VRI.RLAmt, dl), getI32Imm(0, dl),
1478 getI32Imm(31, dl) };
1479 Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops),
1485 // Now, remove all groups with this underlying value and rotation factor.
1486 eraseMatchingBitGroups([VRI](const BitGroup &BG) {
1487 return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt;
1491 if (InstCnt) *InstCnt += BitGroups.size();
1493 // Insert the other groups (one at a time).
1494 for (auto &BG : BitGroups) {
1497 { BG.V, getI32Imm(BG.RLAmt, dl),
1498 getI32Imm(Bits.size() - BG.EndIdx - 1, dl),
1499 getI32Imm(Bits.size() - BG.StartIdx - 1, dl) };
1500 Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
1503 { Res, BG.V, getI32Imm(BG.RLAmt, dl),
1504 getI32Imm(Bits.size() - BG.EndIdx - 1, dl),
1505 getI32Imm(Bits.size() - BG.StartIdx - 1, dl) };
1506 Res = SDValue(CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops), 0);
1511 unsigned Mask = (unsigned) getZerosMask();
1513 unsigned ANDIMask = (Mask & UINT16_MAX), ANDISMask = Mask >> 16;
1514 assert((ANDIMask != 0 || ANDISMask != 0) &&
1515 "No set bits in zeros mask?");
1517 if (InstCnt) *InstCnt += (unsigned) (ANDIMask != 0) +
1518 (unsigned) (ANDISMask != 0) +
1519 (unsigned) (ANDIMask != 0 && ANDISMask != 0);
1521 SDValue ANDIVal, ANDISVal;
1523 ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32,
1524 Res, getI32Imm(ANDIMask, dl)), 0);
1526 ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32,
1527 Res, getI32Imm(ANDISMask, dl)), 0);
1534 Res = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32,
1535 ANDIVal, ANDISVal), 0);
1538 return Res.getNode();
1541 unsigned SelectRotMask64Count(unsigned RLAmt, bool Repl32,
1542 unsigned MaskStart, unsigned MaskEnd,
1544 // In the notation used by the instructions, 'start' and 'end' are reversed
1545 // because bits are counted from high to low order.
1546 unsigned InstMaskStart = 64 - MaskEnd - 1,
1547 InstMaskEnd = 64 - MaskStart - 1;
1552 if ((!IsIns && (InstMaskEnd == 63 || InstMaskStart == 0)) ||
1553 InstMaskEnd == 63 - RLAmt)
1559 // For 64-bit values, not all combinations of rotates and masks are
1560 // available. Produce one if it is available.
1561 SDValue SelectRotMask64(SDValue V, const SDLoc &dl, unsigned RLAmt,
1562 bool Repl32, unsigned MaskStart, unsigned MaskEnd,
1563 unsigned *InstCnt = nullptr) {
1564 // In the notation used by the instructions, 'start' and 'end' are reversed
1565 // because bits are counted from high to low order.
1566 unsigned InstMaskStart = 64 - MaskEnd - 1,
1567 InstMaskEnd = 64 - MaskStart - 1;
1569 if (InstCnt) *InstCnt += 1;
1572 // This rotation amount assumes that the lower 32 bits of the quantity
1573 // are replicated in the high 32 bits by the rotation operator (which is
1574 // done by rlwinm and friends).
1575 assert(InstMaskStart >= 32 && "Mask cannot start out of range");
1576 assert(InstMaskEnd >= 32 && "Mask cannot end out of range");
1578 { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart - 32, dl),
1579 getI32Imm(InstMaskEnd - 32, dl) };
1580 return SDValue(CurDAG->getMachineNode(PPC::RLWINM8, dl, MVT::i64,
1584 if (InstMaskEnd == 63) {
1586 { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) };
1587 return SDValue(CurDAG->getMachineNode(PPC::RLDICL, dl, MVT::i64, Ops), 0);
1590 if (InstMaskStart == 0) {
1592 { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskEnd, dl) };
1593 return SDValue(CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, Ops), 0);
1596 if (InstMaskEnd == 63 - RLAmt) {
1598 { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) };
1599 return SDValue(CurDAG->getMachineNode(PPC::RLDIC, dl, MVT::i64, Ops), 0);
1602 // We cannot do this with a single instruction, so we'll use two. The
1603 // problem is that we're not free to choose both a rotation amount and mask
1604 // start and end independently. We can choose an arbitrary mask start and
1605 // end, but then the rotation amount is fixed. Rotation, however, can be
1606 // inverted, and so by applying an "inverse" rotation first, we can get the
1608 if (InstCnt) *InstCnt += 1;
1610 // The rotation mask for the second instruction must be MaskStart.
1611 unsigned RLAmt2 = MaskStart;
1612 // The first instruction must rotate V so that the overall rotation amount
1614 unsigned RLAmt1 = (64 + RLAmt - RLAmt2) % 64;
1616 V = SelectRotMask64(V, dl, RLAmt1, false, 0, 63);
1617 return SelectRotMask64(V, dl, RLAmt2, false, MaskStart, MaskEnd);
1620 // For 64-bit values, not all combinations of rotates and masks are
1621 // available. Produce a rotate-mask-and-insert if one is available.
1622 SDValue SelectRotMaskIns64(SDValue Base, SDValue V, const SDLoc &dl,
1623 unsigned RLAmt, bool Repl32, unsigned MaskStart,
1624 unsigned MaskEnd, unsigned *InstCnt = nullptr) {
1625 // In the notation used by the instructions, 'start' and 'end' are reversed
1626 // because bits are counted from high to low order.
1627 unsigned InstMaskStart = 64 - MaskEnd - 1,
1628 InstMaskEnd = 64 - MaskStart - 1;
1630 if (InstCnt) *InstCnt += 1;
1633 // This rotation amount assumes that the lower 32 bits of the quantity
1634 // are replicated in the high 32 bits by the rotation operator (which is
1635 // done by rlwinm and friends).
1636 assert(InstMaskStart >= 32 && "Mask cannot start out of range");
1637 assert(InstMaskEnd >= 32 && "Mask cannot end out of range");
1639 { Base, V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart - 32, dl),
1640 getI32Imm(InstMaskEnd - 32, dl) };
1641 return SDValue(CurDAG->getMachineNode(PPC::RLWIMI8, dl, MVT::i64,
1645 if (InstMaskEnd == 63 - RLAmt) {
1647 { Base, V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) };
1648 return SDValue(CurDAG->getMachineNode(PPC::RLDIMI, dl, MVT::i64, Ops), 0);
1651 // We cannot do this with a single instruction, so we'll use two. The
1652 // problem is that we're not free to choose both a rotation amount and mask
1653 // start and end independently. We can choose an arbitrary mask start and
1654 // end, but then the rotation amount is fixed. Rotation, however, can be
1655 // inverted, and so by applying an "inverse" rotation first, we can get the
1657 if (InstCnt) *InstCnt += 1;
1659 // The rotation mask for the second instruction must be MaskStart.
1660 unsigned RLAmt2 = MaskStart;
1661 // The first instruction must rotate V so that the overall rotation amount
1663 unsigned RLAmt1 = (64 + RLAmt - RLAmt2) % 64;
1665 V = SelectRotMask64(V, dl, RLAmt1, false, 0, 63);
1666 return SelectRotMaskIns64(Base, V, dl, RLAmt2, false, MaskStart, MaskEnd);
1669 void SelectAndParts64(const SDLoc &dl, SDValue &Res, unsigned *InstCnt) {
1670 if (BPermRewriterNoMasking)
1673 // The idea here is the same as in the 32-bit version, but with additional
1674 // complications from the fact that Repl32 might be true. Because we
1675 // aggressively convert bit groups to Repl32 form (which, for small
1676 // rotation factors, involves no other change), and then coalesce, it might
1677 // be the case that a single 64-bit masking operation could handle both
1678 // some Repl32 groups and some non-Repl32 groups. If converting to Repl32
1679 // form allowed coalescing, then we must use a 32-bit rotaton in order to
1680 // completely capture the new combined bit group.
1682 for (ValueRotInfo &VRI : ValueRotsVec) {
1685 // We need to add to the mask all bits from the associated bit groups.
1686 // If Repl32 is false, we need to add bits from bit groups that have
1687 // Repl32 true, but are trivially convertable to Repl32 false. Such a
1688 // group is trivially convertable if it overlaps only with the lower 32
1689 // bits, and the group has not been coalesced.
1690 auto MatchingBG = [VRI](const BitGroup &BG) {
1694 unsigned EffRLAmt = BG.RLAmt;
1695 if (!VRI.Repl32 && BG.Repl32) {
1696 if (BG.StartIdx < 32 && BG.EndIdx < 32 && BG.StartIdx <= BG.EndIdx &&
1697 !BG.Repl32Coalesced) {
1703 } else if (VRI.Repl32 != BG.Repl32) {
1707 return VRI.RLAmt == EffRLAmt;
1710 for (auto &BG : BitGroups) {
1711 if (!MatchingBG(BG))
1714 if (BG.StartIdx <= BG.EndIdx) {
1715 for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i)
1716 Mask |= (UINT64_C(1) << i);
1718 for (unsigned i = BG.StartIdx; i < Bits.size(); ++i)
1719 Mask |= (UINT64_C(1) << i);
1720 for (unsigned i = 0; i <= BG.EndIdx; ++i)
1721 Mask |= (UINT64_C(1) << i);
1725 // We can use the 32-bit andi/andis technique if the mask does not
1726 // require any higher-order bits. This can save an instruction compared
1727 // to always using the general 64-bit technique.
1728 bool Use32BitInsts = isUInt<32>(Mask);
1729 // Compute the masks for andi/andis that would be necessary.
1730 unsigned ANDIMask = (Mask & UINT16_MAX),
1731 ANDISMask = (Mask >> 16) & UINT16_MAX;
1733 bool NeedsRotate = VRI.RLAmt || (VRI.Repl32 && !isUInt<32>(Mask));
1735 unsigned NumAndInsts = (unsigned) NeedsRotate +
1736 (unsigned) (bool) Res;
1738 NumAndInsts += (unsigned) (ANDIMask != 0) + (unsigned) (ANDISMask != 0) +
1739 (unsigned) (ANDIMask != 0 && ANDISMask != 0);
1741 NumAndInsts += getInt64Count(Mask) + /* and */ 1;
1743 unsigned NumRLInsts = 0;
1744 bool FirstBG = true;
1745 bool MoreBG = false;
1746 for (auto &BG : BitGroups) {
1747 if (!MatchingBG(BG)) {
1752 SelectRotMask64Count(BG.RLAmt, BG.Repl32, BG.StartIdx, BG.EndIdx,
1757 DEBUG(dbgs() << "\t\trotation groups for " << VRI.V.getNode() <<
1758 " RL: " << VRI.RLAmt << (VRI.Repl32 ? " (32):" : ":") <<
1759 "\n\t\t\tisel using masking: " << NumAndInsts <<
1760 " using rotates: " << NumRLInsts << "\n");
1762 // When we'd use andi/andis, we bias toward using the rotates (andi only
1763 // has a record form, and is cracked on POWER cores). However, when using
1764 // general 64-bit constant formation, bias toward the constant form,
1765 // because that exposes more opportunities for CSE.
1766 if (NumAndInsts > NumRLInsts)
1768 // When merging multiple bit groups, instruction or is used.
1769 // But when rotate is used, rldimi can inert the rotated value into any
1770 // register, so instruction or can be avoided.
1771 if ((Use32BitInsts || MoreBG) && NumAndInsts == NumRLInsts)
1774 DEBUG(dbgs() << "\t\t\t\tusing masking\n");
1776 if (InstCnt) *InstCnt += NumAndInsts;
1779 // We actually need to generate a rotation if we have a non-zero rotation
1780 // factor or, in the Repl32 case, if we care about any of the
1781 // higher-order replicated bits. In the latter case, we generate a mask
1782 // backward so that it actually includes the entire 64 bits.
1783 if (VRI.RLAmt || (VRI.Repl32 && !isUInt<32>(Mask)))
1784 VRot = SelectRotMask64(VRI.V, dl, VRI.RLAmt, VRI.Repl32,
1785 VRI.Repl32 ? 31 : 0, VRI.Repl32 ? 30 : 63);
1790 if (Use32BitInsts) {
1791 assert((ANDIMask != 0 || ANDISMask != 0) &&
1792 "No set bits in mask when using 32-bit ands for 64-bit value");
1794 SDValue ANDIVal, ANDISVal;
1796 ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64,
1797 VRot, getI32Imm(ANDIMask, dl)), 0);
1799 ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64,
1800 VRot, getI32Imm(ANDISMask, dl)), 0);
1803 TotalVal = ANDISVal;
1807 TotalVal = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64,
1808 ANDIVal, ANDISVal), 0);
1810 TotalVal = SDValue(getInt64(CurDAG, dl, Mask), 0);
1812 SDValue(CurDAG->getMachineNode(PPC::AND8, dl, MVT::i64,
1813 VRot, TotalVal), 0);
1819 Res = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64,
1822 // Now, remove all groups with this underlying value and rotation
1824 eraseMatchingBitGroups(MatchingBG);
1828 // Instruction selection for the 64-bit case.
1829 SDNode *Select64(SDNode *N, bool LateMask, unsigned *InstCnt) {
1833 if (InstCnt) *InstCnt = 0;
1835 // Take care of cases that should use andi/andis first.
1836 SelectAndParts64(dl, Res, InstCnt);
1838 // If we've not yet selected a 'starting' instruction, and we have no zeros
1839 // to fill in, select the (Value, RLAmt) with the highest priority (largest
1840 // number of groups), and start with this rotated value.
1841 if ((!HasZeros || LateMask) && !Res) {
1842 // If we have both Repl32 groups and non-Repl32 groups, the non-Repl32
1843 // groups will come first, and so the VRI representing the largest number
1844 // of groups might not be first (it might be the first Repl32 groups).
1845 unsigned MaxGroupsIdx = 0;
1846 if (!ValueRotsVec[0].Repl32) {
1847 for (unsigned i = 0, ie = ValueRotsVec.size(); i < ie; ++i)
1848 if (ValueRotsVec[i].Repl32) {
1849 if (ValueRotsVec[i].NumGroups > ValueRotsVec[0].NumGroups)
1855 ValueRotInfo &VRI = ValueRotsVec[MaxGroupsIdx];
1856 bool NeedsRotate = false;
1859 } else if (VRI.Repl32) {
1860 for (auto &BG : BitGroups) {
1861 if (BG.V != VRI.V || BG.RLAmt != VRI.RLAmt ||
1862 BG.Repl32 != VRI.Repl32)
1865 // We don't need a rotate if the bit group is confined to the lower
1867 if (BG.StartIdx < 32 && BG.EndIdx < 32 && BG.StartIdx < BG.EndIdx)
1876 Res = SelectRotMask64(VRI.V, dl, VRI.RLAmt, VRI.Repl32,
1877 VRI.Repl32 ? 31 : 0, VRI.Repl32 ? 30 : 63,
1882 // Now, remove all groups with this underlying value and rotation factor.
1884 eraseMatchingBitGroups([VRI](const BitGroup &BG) {
1885 return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt &&
1886 BG.Repl32 == VRI.Repl32;
1890 // Because 64-bit rotates are more flexible than inserts, we might have a
1891 // preference regarding which one we do first (to save one instruction).
1893 for (auto I = BitGroups.begin(), IE = BitGroups.end(); I != IE; ++I) {
1894 if (SelectRotMask64Count(I->RLAmt, I->Repl32, I->StartIdx, I->EndIdx,
1896 SelectRotMask64Count(I->RLAmt, I->Repl32, I->StartIdx, I->EndIdx,
1898 if (I != BitGroups.begin()) {
1901 BitGroups.insert(BitGroups.begin(), BG);
1908 // Insert the other groups (one at a time).
1909 for (auto &BG : BitGroups) {
1911 Res = SelectRotMask64(BG.V, dl, BG.RLAmt, BG.Repl32, BG.StartIdx,
1912 BG.EndIdx, InstCnt);
1914 Res = SelectRotMaskIns64(Res, BG.V, dl, BG.RLAmt, BG.Repl32,
1915 BG.StartIdx, BG.EndIdx, InstCnt);
1919 uint64_t Mask = getZerosMask();
1921 // We can use the 32-bit andi/andis technique if the mask does not
1922 // require any higher-order bits. This can save an instruction compared
1923 // to always using the general 64-bit technique.
1924 bool Use32BitInsts = isUInt<32>(Mask);
1925 // Compute the masks for andi/andis that would be necessary.
1926 unsigned ANDIMask = (Mask & UINT16_MAX),
1927 ANDISMask = (Mask >> 16) & UINT16_MAX;
1929 if (Use32BitInsts) {
1930 assert((ANDIMask != 0 || ANDISMask != 0) &&
1931 "No set bits in mask when using 32-bit ands for 64-bit value");
1933 if (InstCnt) *InstCnt += (unsigned) (ANDIMask != 0) +
1934 (unsigned) (ANDISMask != 0) +
1935 (unsigned) (ANDIMask != 0 && ANDISMask != 0);
1937 SDValue ANDIVal, ANDISVal;
1939 ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64,
1940 Res, getI32Imm(ANDIMask, dl)), 0);
1942 ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64,
1943 Res, getI32Imm(ANDISMask, dl)), 0);
1950 Res = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64,
1951 ANDIVal, ANDISVal), 0);
1953 if (InstCnt) *InstCnt += getInt64Count(Mask) + /* and */ 1;
1955 SDValue MaskVal = SDValue(getInt64(CurDAG, dl, Mask), 0);
1957 SDValue(CurDAG->getMachineNode(PPC::AND8, dl, MVT::i64,
1962 return Res.getNode();
1965 SDNode *Select(SDNode *N, bool LateMask, unsigned *InstCnt = nullptr) {
1966 // Fill in BitGroups.
1967 collectBitGroups(LateMask);
1968 if (BitGroups.empty())
1971 // For 64-bit values, figure out when we can use 32-bit instructions.
1972 if (Bits.size() == 64)
1973 assignRepl32BitGroups();
1975 // Fill in ValueRotsVec.
1976 collectValueRotInfo();
1978 if (Bits.size() == 32) {
1979 return Select32(N, LateMask, InstCnt);
1981 assert(Bits.size() == 64 && "Not 64 bits here?");
1982 return Select64(N, LateMask, InstCnt);
1988 void eraseMatchingBitGroups(function_ref<bool(const BitGroup &)> F) {
1989 BitGroups.erase(remove_if(BitGroups, F), BitGroups.end());
1992 SmallVector<ValueBit, 64> Bits;
1995 SmallVector<unsigned, 64> RLAmt;
1997 SmallVector<BitGroup, 16> BitGroups;
1999 DenseMap<std::pair<SDValue, unsigned>, ValueRotInfo> ValueRots;
2000 SmallVector<ValueRotInfo, 16> ValueRotsVec;
2002 SelectionDAG *CurDAG;
2005 BitPermutationSelector(SelectionDAG *DAG)
2008 // Here we try to match complex bit permutations into a set of
2009 // rotate-and-shift/shift/and/or instructions, using a set of heuristics
2010 // known to produce optimial code for common cases (like i32 byte swapping).
2011 SDNode *Select(SDNode *N) {
2014 getValueBits(SDValue(N, 0), N->getValueType(0).getSizeInBits());
2017 Bits = std::move(*Result.second);
2019 DEBUG(dbgs() << "Considering bit-permutation-based instruction"
2020 " selection for: ");
2021 DEBUG(N->dump(CurDAG));
2023 // Fill it RLAmt and set HasZeros.
2024 computeRotationAmounts();
2027 return Select(N, false);
2029 // We currently have two techniques for handling results with zeros: early
2030 // masking (the default) and late masking. Late masking is sometimes more
2031 // efficient, but because the structure of the bit groups is different, it
2032 // is hard to tell without generating both and comparing the results. With
2033 // late masking, we ignore zeros in the resulting value when inserting each
2034 // set of bit groups, and then mask in the zeros at the end. With early
2035 // masking, we only insert the non-zero parts of the result at every step.
2037 unsigned InstCnt, InstCntLateMask;
2038 DEBUG(dbgs() << "\tEarly masking:\n");
2039 SDNode *RN = Select(N, false, &InstCnt);
2040 DEBUG(dbgs() << "\t\tisel would use " << InstCnt << " instructions\n");
2042 DEBUG(dbgs() << "\tLate masking:\n");
2043 SDNode *RNLM = Select(N, true, &InstCntLateMask);
2044 DEBUG(dbgs() << "\t\tisel would use " << InstCntLateMask <<
2047 if (InstCnt <= InstCntLateMask) {
2048 DEBUG(dbgs() << "\tUsing early-masking for isel\n");
2052 DEBUG(dbgs() << "\tUsing late-masking for isel\n");
2057 } // end anonymous namespace
2059 bool PPCDAGToDAGISel::tryBitPermutation(SDNode *N) {
2060 if (N->getValueType(0) != MVT::i32 &&
2061 N->getValueType(0) != MVT::i64)
2064 if (!UseBitPermRewriter)
2067 switch (N->getOpcode()) {
2074 BitPermutationSelector BPS(CurDAG);
2075 if (SDNode *New = BPS.Select(N)) {
2076 ReplaceNode(N, New);
2086 /// SelectCC - Select a comparison of the specified values with the specified
2087 /// condition code, returning the CR# of the expression.
2088 SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2090 // Always select the LHS.
2093 if (LHS.getValueType() == MVT::i32) {
2095 if (CC == ISD::SETEQ || CC == ISD::SETNE) {
2096 if (isInt32Immediate(RHS, Imm)) {
2097 // SETEQ/SETNE comparison with 16-bit immediate, fold it.
2098 if (isUInt<16>(Imm))
2099 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
2100 getI32Imm(Imm & 0xFFFF, dl)),
2102 // If this is a 16-bit signed immediate, fold it.
2103 if (isInt<16>((int)Imm))
2104 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
2105 getI32Imm(Imm & 0xFFFF, dl)),
2108 // For non-equality comparisons, the default code would materialize the
2109 // constant, then compare against it, like this:
2111 // ori r2, r2, 22136
2113 // Since we are just comparing for equality, we can emit this instead:
2114 // xoris r0,r3,0x1234
2115 // cmplwi cr0,r0,0x5678
2117 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS,
2118 getI32Imm(Imm >> 16, dl)), 0);
2119 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor,
2120 getI32Imm(Imm & 0xFFFF, dl)), 0);
2123 } else if (ISD::isUnsignedIntSetCC(CC)) {
2124 if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm))
2125 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
2126 getI32Imm(Imm & 0xFFFF, dl)), 0);
2130 if (isIntS16Immediate(RHS, SImm))
2131 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
2132 getI32Imm((int)SImm & 0xFFFF,
2137 } else if (LHS.getValueType() == MVT::i64) {
2139 if (CC == ISD::SETEQ || CC == ISD::SETNE) {
2140 if (isInt64Immediate(RHS.getNode(), Imm)) {
2141 // SETEQ/SETNE comparison with 16-bit immediate, fold it.
2142 if (isUInt<16>(Imm))
2143 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
2144 getI32Imm(Imm & 0xFFFF, dl)),
2146 // If this is a 16-bit signed immediate, fold it.
2148 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
2149 getI32Imm(Imm & 0xFFFF, dl)),
2152 // For non-equality comparisons, the default code would materialize the
2153 // constant, then compare against it, like this:
2155 // ori r2, r2, 22136
2157 // Since we are just comparing for equality, we can emit this instead:
2158 // xoris r0,r3,0x1234
2159 // cmpldi cr0,r0,0x5678
2161 if (isUInt<32>(Imm)) {
2162 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS,
2163 getI64Imm(Imm >> 16, dl)), 0);
2164 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor,
2165 getI64Imm(Imm & 0xFFFF, dl)),
2170 } else if (ISD::isUnsignedIntSetCC(CC)) {
2171 if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm))
2172 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
2173 getI64Imm(Imm & 0xFFFF, dl)), 0);
2177 if (isIntS16Immediate(RHS, SImm))
2178 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
2179 getI64Imm(SImm & 0xFFFF, dl)),
2183 } else if (LHS.getValueType() == MVT::f32) {
2186 assert(LHS.getValueType() == MVT::f64 && "Unknown vt!");
2187 Opc = PPCSubTarget->hasVSX() ? PPC::XSCMPUDP : PPC::FCMPUD;
2189 return SDValue(CurDAG->getMachineNode(Opc, dl, MVT::i32, LHS, RHS), 0);
2192 static PPC::Predicate getPredicateForSetCC(ISD::CondCode CC) {
2198 llvm_unreachable("Should be lowered by legalize!");
2199 default: llvm_unreachable("Unknown condition!");
2201 case ISD::SETEQ: return PPC::PRED_EQ;
2203 case ISD::SETNE: return PPC::PRED_NE;
2205 case ISD::SETLT: return PPC::PRED_LT;
2207 case ISD::SETLE: return PPC::PRED_LE;
2209 case ISD::SETGT: return PPC::PRED_GT;
2211 case ISD::SETGE: return PPC::PRED_GE;
2212 case ISD::SETO: return PPC::PRED_NU;
2213 case ISD::SETUO: return PPC::PRED_UN;
2214 // These two are invalid for floating point. Assume we have int.
2215 case ISD::SETULT: return PPC::PRED_LT;
2216 case ISD::SETUGT: return PPC::PRED_GT;
2220 /// getCRIdxForSetCC - Return the index of the condition register field
2221 /// associated with the SetCC condition, and whether or not the field is
2222 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
2223 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) {
2226 default: llvm_unreachable("Unknown condition!");
2228 case ISD::SETLT: return 0; // Bit #0 = SETOLT
2230 case ISD::SETGT: return 1; // Bit #1 = SETOGT
2232 case ISD::SETEQ: return 2; // Bit #2 = SETOEQ
2233 case ISD::SETUO: return 3; // Bit #3 = SETUO
2235 case ISD::SETGE: Invert = true; return 0; // !Bit #0 = SETUGE
2237 case ISD::SETLE: Invert = true; return 1; // !Bit #1 = SETULE
2239 case ISD::SETNE: Invert = true; return 2; // !Bit #2 = SETUNE
2240 case ISD::SETO: Invert = true; return 3; // !Bit #3 = SETO
2245 llvm_unreachable("Invalid branch code: should be expanded by legalize");
2246 // These are invalid for floating point. Assume integer.
2247 case ISD::SETULT: return 0;
2248 case ISD::SETUGT: return 1;
2252 // getVCmpInst: return the vector compare instruction for the specified
2253 // vector type and condition code. Since this is for altivec specific code,
2254 // only support the altivec types (v16i8, v8i16, v4i32, v2i64, and v4f32).
2255 static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC,
2256 bool HasVSX, bool &Swap, bool &Negate) {
2260 if (VecVT.isFloatingPoint()) {
2261 /* Handle some cases by swapping input operands. */
2263 case ISD::SETLE: CC = ISD::SETGE; Swap = true; break;
2264 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break;
2265 case ISD::SETOLE: CC = ISD::SETOGE; Swap = true; break;
2266 case ISD::SETOLT: CC = ISD::SETOGT; Swap = true; break;
2267 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break;
2268 case ISD::SETUGT: CC = ISD::SETULT; Swap = true; break;
2271 /* Handle some cases by negating the result. */
2273 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break;
2274 case ISD::SETUNE: CC = ISD::SETOEQ; Negate = true; break;
2275 case ISD::SETULE: CC = ISD::SETOGT; Negate = true; break;
2276 case ISD::SETULT: CC = ISD::SETOGE; Negate = true; break;
2279 /* We have instructions implementing the remaining cases. */
2283 if (VecVT == MVT::v4f32)
2284 return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP;
2285 else if (VecVT == MVT::v2f64)
2286 return PPC::XVCMPEQDP;
2290 if (VecVT == MVT::v4f32)
2291 return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP;
2292 else if (VecVT == MVT::v2f64)
2293 return PPC::XVCMPGTDP;
2297 if (VecVT == MVT::v4f32)
2298 return HasVSX ? PPC::XVCMPGESP : PPC::VCMPGEFP;
2299 else if (VecVT == MVT::v2f64)
2300 return PPC::XVCMPGEDP;
2305 llvm_unreachable("Invalid floating-point vector compare condition");
2307 /* Handle some cases by swapping input operands. */
2309 case ISD::SETGE: CC = ISD::SETLE; Swap = true; break;
2310 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break;
2311 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break;
2312 case ISD::SETULT: CC = ISD::SETUGT; Swap = true; break;
2315 /* Handle some cases by negating the result. */
2317 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break;
2318 case ISD::SETUNE: CC = ISD::SETUEQ; Negate = true; break;
2319 case ISD::SETLE: CC = ISD::SETGT; Negate = true; break;
2320 case ISD::SETULE: CC = ISD::SETUGT; Negate = true; break;
2323 /* We have instructions implementing the remaining cases. */
2327 if (VecVT == MVT::v16i8)
2328 return PPC::VCMPEQUB;
2329 else if (VecVT == MVT::v8i16)
2330 return PPC::VCMPEQUH;
2331 else if (VecVT == MVT::v4i32)
2332 return PPC::VCMPEQUW;
2333 else if (VecVT == MVT::v2i64)
2334 return PPC::VCMPEQUD;
2337 if (VecVT == MVT::v16i8)
2338 return PPC::VCMPGTSB;
2339 else if (VecVT == MVT::v8i16)
2340 return PPC::VCMPGTSH;
2341 else if (VecVT == MVT::v4i32)
2342 return PPC::VCMPGTSW;
2343 else if (VecVT == MVT::v2i64)
2344 return PPC::VCMPGTSD;
2347 if (VecVT == MVT::v16i8)
2348 return PPC::VCMPGTUB;
2349 else if (VecVT == MVT::v8i16)
2350 return PPC::VCMPGTUH;
2351 else if (VecVT == MVT::v4i32)
2352 return PPC::VCMPGTUW;
2353 else if (VecVT == MVT::v2i64)
2354 return PPC::VCMPGTUD;
2359 llvm_unreachable("Invalid integer vector compare condition");
2363 bool PPCDAGToDAGISel::trySETCC(SDNode *N) {
2366 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
2368 CurDAG->getTargetLoweringInfo().getPointerTy(CurDAG->getDataLayout());
2369 bool isPPC64 = (PtrVT == MVT::i64);
2371 if (!PPCSubTarget->useCRBits() &&
2372 isInt32Immediate(N->getOperand(1), Imm)) {
2373 // We can codegen setcc op, imm very efficiently compared to a brcond.
2374 // Check for those cases here.
2377 SDValue Op = N->getOperand(0);
2381 Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0);
2382 SDValue Ops[] = { Op, getI32Imm(27, dl), getI32Imm(5, dl),
2383 getI32Imm(31, dl) };
2384 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2390 SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
2391 Op, getI32Imm(~0U, dl)), 0);
2392 CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op, AD.getValue(1));
2396 SDValue Ops[] = { Op, getI32Imm(1, dl), getI32Imm(31, dl),
2397 getI32Imm(31, dl) };
2398 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2403 SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0);
2404 T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0);
2405 SDValue Ops[] = { T, getI32Imm(1, dl), getI32Imm(31, dl),
2406 getI32Imm(31, dl) };
2407 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2411 } else if (Imm == ~0U) { // setcc op, -1
2412 SDValue Op = N->getOperand(0);
2417 Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
2418 Op, getI32Imm(1, dl)), 0);
2419 CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
2420 SDValue(CurDAG->getMachineNode(PPC::LI, dl,
2423 0), Op.getValue(1));
2427 Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0);
2428 SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
2429 Op, getI32Imm(~0U, dl));
2430 CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0), Op,
2435 SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op,
2436 getI32Imm(1, dl)), 0);
2437 SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD,
2439 SDValue Ops[] = { AN, getI32Imm(1, dl), getI32Imm(31, dl),
2440 getI32Imm(31, dl) };
2441 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2445 SDValue Ops[] = { Op, getI32Imm(1, dl), getI32Imm(31, dl),
2446 getI32Imm(31, dl) };
2447 Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
2448 CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, getI32Imm(1, dl));
2455 SDValue LHS = N->getOperand(0);
2456 SDValue RHS = N->getOperand(1);
2458 // Altivec Vector compare instructions do not set any CR register by default and
2459 // vector compare operations return the same type as the operands.
2460 if (LHS.getValueType().isVector()) {
2461 if (PPCSubTarget->hasQPX())
2464 EVT VecVT = LHS.getValueType();
2466 unsigned int VCmpInst = getVCmpInst(VecVT.getSimpleVT(), CC,
2467 PPCSubTarget->hasVSX(), Swap, Negate);
2469 std::swap(LHS, RHS);
2471 EVT ResVT = VecVT.changeVectorElementTypeToInteger();
2473 SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, ResVT, LHS, RHS), 0);
2474 CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLNOR : PPC::VNOR,
2479 CurDAG->SelectNodeTo(N, VCmpInst, ResVT, LHS, RHS);
2483 if (PPCSubTarget->useCRBits())
2487 unsigned Idx = getCRIdxForSetCC(CC, Inv);
2488 SDValue CCReg = SelectCC(LHS, RHS, CC, dl);
2491 // Force the ccreg into CR7.
2492 SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
2494 SDValue InFlag(nullptr, 0); // Null incoming flag value.
2495 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg,
2496 InFlag).getValue(1);
2498 IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg,
2501 SDValue Ops[] = { IntCR, getI32Imm((32 - (3 - Idx)) & 31, dl),
2502 getI32Imm(31, dl), getI32Imm(31, dl) };
2504 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2508 // Get the specified bit.
2510 SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
2511 CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1, dl));
2515 // Is this opcode a bitwise logical operation?
2516 static bool isLogicOp(unsigned Opc) {
2517 return Opc == ISD::AND || Opc == ISD::OR || Opc == ISD::XOR;
2520 /// If this node is a sign/zero extension of an integer comparison,
2521 /// it can usually be computed in GPR's rather than using comparison
2522 /// instructions and ISEL. We only do this on 64-bit targets for now
2523 /// as the code is specialized for 64-bit (it uses 64-bit instructions
2524 /// and assumes 64-bit registers).
2525 bool PPCDAGToDAGISel::tryEXTEND(SDNode *N) {
2526 if (TM.getOptLevel() == CodeGenOpt::None || !TM.isPPC64())
2528 assert((N->getOpcode() == ISD::ZERO_EXTEND ||
2529 N->getOpcode() == ISD::SIGN_EXTEND) &&
2530 "Expecting a zero/sign extend node!");
2533 // If we are zero-extending the result of a logical operation on i1
2534 // values, we can keep the values in GPRs.
2535 if (isLogicOp(N->getOperand(0).getOpcode()) &&
2536 N->getOperand(0).getValueType() == MVT::i1 &&
2537 N->getOpcode() == ISD::ZERO_EXTEND)
2538 WideRes = computeLogicOpInGPR(N->getOperand(0));
2539 else if (N->getOperand(0).getOpcode() != ISD::SETCC)
2543 getSETCCInGPR(N->getOperand(0),
2544 N->getOpcode() == ISD::SIGN_EXTEND ?
2545 SetccInGPROpts::SExtOrig : SetccInGPROpts::ZExtOrig);
2551 bool Inputs32Bit = N->getOperand(0).getOperand(0).getValueType() == MVT::i32;
2552 bool Output32Bit = N->getValueType(0) == MVT::i32;
2554 NumSextSetcc += N->getOpcode() == ISD::SIGN_EXTEND ? 1 : 0;
2555 NumZextSetcc += N->getOpcode() == ISD::SIGN_EXTEND ? 0 : 1;
2557 SDValue ConvOp = WideRes;
2558 if (Inputs32Bit != Output32Bit)
2559 ConvOp = addExtOrTrunc(WideRes, Inputs32Bit ? ExtOrTruncConversion::Ext :
2560 ExtOrTruncConversion::Trunc);
2561 ReplaceNode(N, ConvOp.getNode());
2566 // Lower a logical operation on i1 values into a GPR sequence if possible.
2567 // The result can be kept in a GPR if requested.
2568 // Three types of inputs can be handled:
2571 // - Logical operation (AND/OR/XOR)
2572 // There is also a special case that is handled (namely a complement operation
2573 // achieved with xor %a, -1).
2574 SDValue PPCDAGToDAGISel::computeLogicOpInGPR(SDValue LogicOp) {
2575 assert(isLogicOp(LogicOp.getOpcode()) &&
2576 "Can only handle logic operations here.");
2577 assert(LogicOp.getValueType() == MVT::i1 &&
2578 "Can only handle logic operations on i1 values here.");
2582 // Special case: xor %a, -1
2583 bool IsBitwiseNegation = isBitwiseNot(LogicOp);
2585 // Produces a GPR sequence for each operand of the binary logic operation.
2586 // For SETCC, it produces the respective comparison, for TRUNCATE it truncates
2587 // the value in a GPR and for logic operations, it will recursively produce
2588 // a GPR sequence for the operation.
2589 auto getLogicOperand = [&] (SDValue Operand) -> SDValue {
2590 unsigned OperandOpcode = Operand.getOpcode();
2591 if (OperandOpcode == ISD::SETCC)
2592 return getSETCCInGPR(Operand, SetccInGPROpts::ZExtOrig);
2593 else if (OperandOpcode == ISD::TRUNCATE) {
2594 SDValue InputOp = Operand.getOperand(0);
2595 EVT InVT = InputOp.getValueType();
2597 SDValue(CurDAG->getMachineNode(InVT == MVT::i32 ? PPC::RLDICL_32 :
2598 PPC::RLDICL, dl, InVT, InputOp,
2599 getI64Imm(0, dl), getI64Imm(63, dl)), 0);
2600 } else if (isLogicOp(OperandOpcode))
2601 return computeLogicOpInGPR(Operand);
2604 LHS = getLogicOperand(LogicOp.getOperand(0));
2605 RHS = getLogicOperand(LogicOp.getOperand(1));
2607 // If a GPR sequence can't be produced for the LHS we can't proceed.
2608 // Not producing a GPR sequence for the RHS is only a problem if this isn't
2609 // a bitwise negation operation.
2610 if (!LHS || (!RHS && !IsBitwiseNegation))
2613 NumLogicOpsOnComparison++;
2615 // We will use the inputs as 64-bit values.
2616 if (LHS.getValueType() == MVT::i32)
2617 LHS = addExtOrTrunc(LHS, ExtOrTruncConversion::Ext);
2618 if (!IsBitwiseNegation && RHS.getValueType() == MVT::i32)
2619 RHS = addExtOrTrunc(RHS, ExtOrTruncConversion::Ext);
2622 switch (LogicOp.getOpcode()) {
2623 default: llvm_unreachable("Unknown logic operation.");
2624 case ISD::AND: NewOpc = PPC::AND8; break;
2625 case ISD::OR: NewOpc = PPC::OR8; break;
2626 case ISD::XOR: NewOpc = PPC::XOR8; break;
2629 if (IsBitwiseNegation) {
2630 RHS = getI64Imm(1, dl);
2631 NewOpc = PPC::XORI8;
2634 return SDValue(CurDAG->getMachineNode(NewOpc, dl, MVT::i64, LHS, RHS), 0);
2638 /// Try performing logical operations on results of comparisons in GPRs.
2639 /// It is typically preferred from a performance perspective over performing
2640 /// the operations on individual bits in the CR. We only do this on 64-bit
2641 /// targets for now as the code is specialized for 64-bit (it uses 64-bit
2642 /// instructions and assumes 64-bit registers).
2643 bool PPCDAGToDAGISel::tryLogicOpOfCompares(SDNode *N) {
2644 if (TM.getOptLevel() == CodeGenOpt::None || !TM.isPPC64())
2646 if (N->getValueType(0) != MVT::i1)
2648 assert(isLogicOp(N->getOpcode()) &&
2649 "Expected a logic operation on setcc results.");
2650 SDValue LoweredLogical = computeLogicOpInGPR(SDValue(N, 0));
2651 if (!LoweredLogical)
2655 bool IsBitwiseNegate = LoweredLogical.getMachineOpcode() == PPC::XORI8;
2656 unsigned SubRegToExtract = IsBitwiseNegate ? PPC::sub_eq : PPC::sub_gt;
2657 SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32);
2658 SDValue LHS = LoweredLogical.getOperand(0);
2659 SDValue RHS = LoweredLogical.getOperand(1);
2661 SDValue OpToConvToRecForm;
2663 // Look through any 32-bit to 64-bit implicit extend nodes to find the opcode
2664 // that is input to the XORI.
2665 if (IsBitwiseNegate &&
2666 LoweredLogical.getOperand(0).getMachineOpcode() == PPC::INSERT_SUBREG)
2667 OpToConvToRecForm = LoweredLogical.getOperand(0).getOperand(1);
2668 else if (IsBitwiseNegate)
2669 // If the input to the XORI isn't an extension, that's what we're after.
2670 OpToConvToRecForm = LoweredLogical.getOperand(0);
2672 // If this is not an XORI, it is a reg-reg logical op and we can convert it
2674 OpToConvToRecForm = LoweredLogical;
2676 // Get the record-form version of the node we're looking to use to get the
2678 uint16_t NonRecOpc = OpToConvToRecForm.getMachineOpcode();
2679 int NewOpc = PPCInstrInfo::getRecordFormOpcode(NonRecOpc);
2681 // Convert the right node to record-form. This is either the logical we're
2682 // looking at or it is the input node to the negation (if we're looking at
2683 // a bitwise negation).
2684 if (NewOpc != -1 && IsBitwiseNegate) {
2685 // The input to the XORI has a record-form. Use it.
2686 assert(LoweredLogical.getConstantOperandVal(1) == 1 &&
2687 "Expected a PPC::XORI8 only for bitwise negation.");
2688 // Emit the record-form instruction.
2689 std::vector<SDValue> Ops;
2690 for (int i = 0, e = OpToConvToRecForm.getNumOperands(); i < e; i++)
2691 Ops.push_back(OpToConvToRecForm.getOperand(i));
2694 SDValue(CurDAG->getMachineNode(NewOpc, dl,
2695 OpToConvToRecForm.getValueType(),
2696 MVT::Glue, Ops), 0);
2698 assert((NewOpc != -1 || !IsBitwiseNegate) &&
2699 "No record form available for AND8/OR8/XOR8?");
2701 SDValue(CurDAG->getMachineNode(NewOpc == -1 ? PPC::ANDIo8 : NewOpc, dl,
2702 MVT::i64, MVT::Glue, LHS, RHS), 0);
2705 // Select this node to a single bit from CR0 set by the record-form node
2706 // just created. For bitwise negation, use the EQ bit which is the equivalent
2707 // of negating the result (i.e. it is a bit set when the result of the
2708 // operation is zero).
2710 CurDAG->getTargetConstant(SubRegToExtract, dl, MVT::i32);
2712 SDValue(CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl,
2713 MVT::i1, CR0Reg, SRIdxVal,
2714 WideOp.getValue(1)), 0);
2715 ReplaceNode(N, CRBit.getNode());
2719 /// If the value isn't guaranteed to be sign-extended to 64-bits, extend it.
2720 /// Useful when emitting comparison code for 32-bit values without using
2721 /// the compare instruction (which only considers the lower 32-bits).
2722 SDValue PPCDAGToDAGISel::signExtendInputIfNeeded(SDValue Input) {
2723 assert(Input.getValueType() == MVT::i32 &&
2724 "Can only sign-extend 32-bit values here.");
2725 unsigned Opc = Input.getOpcode();
2727 // The value was sign extended and then truncated to 32-bits. No need to
2728 // sign extend it again.
2729 if (Opc == ISD::TRUNCATE &&
2730 (Input.getOperand(0).getOpcode() == ISD::AssertSext ||
2731 Input.getOperand(0).getOpcode() == ISD::SIGN_EXTEND))
2734 LoadSDNode *InputLoad = dyn_cast<LoadSDNode>(Input);
2735 // The input is a sign-extending load. No reason to sign-extend.
2736 if (InputLoad && InputLoad->getExtensionType() == ISD::SEXTLOAD)
2739 ConstantSDNode *InputConst = dyn_cast<ConstantSDNode>(Input);
2740 // We don't sign-extend constants and already sign-extended values.
2741 if (InputConst || Opc == ISD::AssertSext || Opc == ISD::SIGN_EXTEND_INREG ||
2742 Opc == ISD::SIGN_EXTEND)
2746 SignExtensionsAdded++;
2747 return SDValue(CurDAG->getMachineNode(PPC::EXTSW_32, dl, MVT::i32, Input), 0);
2750 /// If the value isn't guaranteed to be zero-extended to 64-bits, extend it.
2751 /// Useful when emitting comparison code for 32-bit values without using
2752 /// the compare instruction (which only considers the lower 32-bits).
2753 SDValue PPCDAGToDAGISel::zeroExtendInputIfNeeded(SDValue Input) {
2754 assert(Input.getValueType() == MVT::i32 &&
2755 "Can only zero-extend 32-bit values here.");
2756 LoadSDNode *InputLoad = dyn_cast<LoadSDNode>(Input);
2757 unsigned Opc = Input.getOpcode();
2759 // No need to zero-extend loaded values (unless they're loaded with
2760 // a sign-extending load).
2761 if (InputLoad && InputLoad->getExtensionType() != ISD::SEXTLOAD)
2764 ConstantSDNode *InputConst = dyn_cast<ConstantSDNode>(Input);
2765 bool InputZExtConst = InputConst && InputConst->getSExtValue() >= 0;
2766 // An ISD::TRUNCATE will be lowered to an EXTRACT_SUBREG so we have
2767 // to conservatively actually clear the high bits. We also don't need to
2768 // zero-extend constants or values that are already zero-extended.
2769 if (InputZExtConst || Opc == ISD::AssertZext || Opc == ISD::ZERO_EXTEND)
2773 ZeroExtensionsAdded++;
2774 return SDValue(CurDAG->getMachineNode(PPC::RLDICL_32, dl, MVT::i32, Input,
2775 getI64Imm(0, dl), getI64Imm(32, dl)),
2779 // Handle a 32-bit value in a 64-bit register and vice-versa. These are of
2780 // course not actual zero/sign extensions that will generate machine code,
2781 // they're just a way to reinterpret a 32 bit value in a register as a
2782 // 64 bit value and vice-versa.
2783 SDValue PPCDAGToDAGISel::addExtOrTrunc(SDValue NatWidthRes,
2784 ExtOrTruncConversion Conv) {
2785 SDLoc dl(NatWidthRes);
2787 // For reinterpreting 32-bit values as 64 bit values, we generate
2788 // INSERT_SUBREG IMPLICIT_DEF:i64, <input>, TargetConstant:i32<1>
2789 if (Conv == ExtOrTruncConversion::Ext) {
2790 SDValue ImDef(CurDAG->getMachineNode(PPC::IMPLICIT_DEF, dl, MVT::i64), 0);
2792 CurDAG->getTargetConstant(PPC::sub_32, dl, MVT::i32);
2793 return SDValue(CurDAG->getMachineNode(PPC::INSERT_SUBREG, dl, MVT::i64,
2794 ImDef, NatWidthRes, SubRegIdx), 0);
2797 assert(Conv == ExtOrTruncConversion::Trunc &&
2798 "Unknown convertion between 32 and 64 bit values.");
2799 // For reinterpreting 64-bit values as 32-bit values, we just need to
2800 // EXTRACT_SUBREG (i.e. extract the low word).
2802 CurDAG->getTargetConstant(PPC::sub_32, dl, MVT::i32);
2803 return SDValue(CurDAG->getMachineNode(PPC::EXTRACT_SUBREG, dl, MVT::i32,
2804 NatWidthRes, SubRegIdx), 0);
2807 /// Produces a zero-extended result of comparing two 32-bit values according to
2808 /// the passed condition code.
2809 SDValue PPCDAGToDAGISel::get32BitZExtCompare(SDValue LHS, SDValue RHS,
2811 int64_t RHSValue, SDLoc dl) {
2812 bool IsRHSZero = RHSValue == 0;
2814 default: return SDValue();
2816 // (zext (setcc %a, %b, seteq)) -> (lshr (cntlzw (xor %a, %b)), 5)
2817 // (zext (setcc %a, 0, seteq)) -> (lshr (cntlzw %a), 5)
2818 SDValue Xor = IsRHSZero ? LHS :
2819 SDValue(CurDAG->getMachineNode(PPC::XOR, dl, MVT::i32, LHS, RHS), 0);
2821 SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Xor), 0);
2822 SDValue ShiftOps[] = { Clz, getI32Imm(27, dl), getI32Imm(5, dl),
2823 getI32Imm(31, dl) };
2824 return SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32,
2828 // (zext (setcc %a, %b, setne)) -> (xor (lshr (cntlzw (xor %a, %b)), 5), 1)
2829 // (zext (setcc %a, 0, setne)) -> (xor (lshr (cntlzw %a), 5), 1)
2830 SDValue Xor = IsRHSZero ? LHS :
2831 SDValue(CurDAG->getMachineNode(PPC::XOR, dl, MVT::i32, LHS, RHS), 0);
2833 SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Xor), 0);
2834 SDValue ShiftOps[] = { Clz, getI32Imm(27, dl), getI32Imm(5, dl),
2835 getI32Imm(31, dl) };
2837 SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, ShiftOps), 0);
2838 return SDValue(CurDAG->getMachineNode(PPC::XORI, dl, MVT::i32, Shift,
2839 getI32Imm(1, dl)), 0);
2844 /// Produces a sign-extended result of comparing two 32-bit values according to
2845 /// the passed condition code.
2846 SDValue PPCDAGToDAGISel::get32BitSExtCompare(SDValue LHS, SDValue RHS,
2848 int64_t RHSValue, SDLoc dl) {
2849 bool IsRHSZero = RHSValue == 0;
2851 default: return SDValue();
2853 // (sext (setcc %a, %b, seteq)) ->
2854 // (ashr (shl (ctlz (xor %a, %b)), 58), 63)
2855 // (sext (setcc %a, 0, seteq)) ->
2856 // (ashr (shl (ctlz %a), 58), 63)
2857 SDValue CountInput = IsRHSZero ? LHS :
2858 SDValue(CurDAG->getMachineNode(PPC::XOR, dl, MVT::i32, LHS, RHS), 0);
2860 SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, CountInput), 0);
2861 SDValue SHLOps[] = { Cntlzw, getI32Imm(58, dl), getI32Imm(0, dl) };
2863 SDValue(CurDAG->getMachineNode(PPC::RLDICR_32, dl, MVT::i32, SHLOps), 0);
2864 return SDValue(CurDAG->getMachineNode(PPC::SRADI_32, dl, MVT::i32, Sldi,
2865 getI32Imm(63, dl)), 0);
2868 // Bitwise xor the operands, count leading zeros, shift right by 5 bits and
2869 // flip the bit, finally take 2's complement.
2870 // (sext (setcc %a, %b, setne)) ->
2871 // (neg (xor (lshr (ctlz (xor %a, %b)), 5), 1))
2872 // Same as above, but the first xor is not needed.
2873 // (sext (setcc %a, 0, setne)) ->
2874 // (neg (xor (lshr (ctlz %a), 5), 1))
2875 SDValue Xor = IsRHSZero ? LHS :
2876 SDValue(CurDAG->getMachineNode(PPC::XOR, dl, MVT::i32, LHS, RHS), 0);
2878 SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Xor), 0);
2879 SDValue ShiftOps[] =
2880 { Clz, getI32Imm(27, dl), getI32Imm(5, dl), getI32Imm(31, dl) };
2882 SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, ShiftOps), 0);
2884 SDValue(CurDAG->getMachineNode(PPC::XORI, dl, MVT::i32, Shift,
2885 getI32Imm(1, dl)), 0);
2886 return SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Xori), 0);
2891 /// Produces a zero-extended result of comparing two 64-bit values according to
2892 /// the passed condition code.
2893 SDValue PPCDAGToDAGISel::get64BitZExtCompare(SDValue LHS, SDValue RHS,
2895 int64_t RHSValue, SDLoc dl) {
2896 bool IsRHSZero = RHSValue == 0;
2898 default: return SDValue();
2900 // (zext (setcc %a, %b, seteq)) -> (lshr (ctlz (xor %a, %b)), 6)
2901 // (zext (setcc %a, 0, seteq)) -> (lshr (ctlz %a), 6)
2902 SDValue Xor = IsRHSZero ? LHS :
2903 SDValue(CurDAG->getMachineNode(PPC::XOR8, dl, MVT::i64, LHS, RHS), 0);
2905 SDValue(CurDAG->getMachineNode(PPC::CNTLZD, dl, MVT::i64, Xor), 0);
2906 return SDValue(CurDAG->getMachineNode(PPC::RLDICL, dl, MVT::i64, Clz,
2907 getI64Imm(58, dl), getI64Imm(63, dl)),
2911 // {addc.reg, addc.CA} = (addcarry (xor %a, %b), -1)
2912 // (zext (setcc %a, %b, setne)) -> (sube addc.reg, addc.reg, addc.CA)
2913 // {addcz.reg, addcz.CA} = (addcarry %a, -1)
2914 // (zext (setcc %a, 0, setne)) -> (sube addcz.reg, addcz.reg, addcz.CA)
2915 SDValue Xor = IsRHSZero ? LHS :
2916 SDValue(CurDAG->getMachineNode(PPC::XOR8, dl, MVT::i64, LHS, RHS), 0);
2918 SDValue(CurDAG->getMachineNode(PPC::ADDIC8, dl, MVT::i64, MVT::Glue,
2919 Xor, getI32Imm(~0U, dl)), 0);
2920 return SDValue(CurDAG->getMachineNode(PPC::SUBFE8, dl, MVT::i64, AC,
2921 Xor, AC.getValue(1)), 0);
2926 /// Produces a sign-extended result of comparing two 64-bit values according to
2927 /// the passed condition code.
2928 SDValue PPCDAGToDAGISel::get64BitSExtCompare(SDValue LHS, SDValue RHS,
2930 int64_t RHSValue, SDLoc dl) {
2931 bool IsRHSZero = RHSValue == 0;
2933 default: return SDValue();
2935 // {addc.reg, addc.CA} = (addcarry (xor %a, %b), -1)
2936 // (sext (setcc %a, %b, seteq)) -> (sube addc.reg, addc.reg, addc.CA)
2937 // {addcz.reg, addcz.CA} = (addcarry %a, -1)
2938 // (sext (setcc %a, 0, seteq)) -> (sube addcz.reg, addcz.reg, addcz.CA)
2939 SDValue AddInput = IsRHSZero ? LHS :
2940 SDValue(CurDAG->getMachineNode(PPC::XOR8, dl, MVT::i64, LHS, RHS), 0);
2942 SDValue(CurDAG->getMachineNode(PPC::ADDIC8, dl, MVT::i64, MVT::Glue,
2943 AddInput, getI32Imm(~0U, dl)), 0);
2944 return SDValue(CurDAG->getMachineNode(PPC::SUBFE8, dl, MVT::i64, Addic,
2945 Addic, Addic.getValue(1)), 0);
2948 // {subfc.reg, subfc.CA} = (subcarry 0, (xor %a, %b))
2949 // (sext (setcc %a, %b, setne)) -> (sube subfc.reg, subfc.reg, subfc.CA)
2950 // {subfcz.reg, subfcz.CA} = (subcarry 0, %a)
2951 // (sext (setcc %a, 0, setne)) -> (sube subfcz.reg, subfcz.reg, subfcz.CA)
2952 SDValue Xor = IsRHSZero ? LHS :
2953 SDValue(CurDAG->getMachineNode(PPC::XOR8, dl, MVT::i64, LHS, RHS), 0);
2955 SDValue(CurDAG->getMachineNode(PPC::SUBFIC8, dl, MVT::i64, MVT::Glue,
2956 Xor, getI32Imm(0, dl)), 0);
2957 return SDValue(CurDAG->getMachineNode(PPC::SUBFE8, dl, MVT::i64, SC,
2958 SC, SC.getValue(1)), 0);
2963 /// Does this SDValue have any uses for which keeping the value in a GPR is
2964 /// appropriate. This is meant to be used on values that have type i1 since
2965 /// it is somewhat meaningless to ask if values of other types can be kept in
2967 static bool allUsesExtend(SDValue Compare, SelectionDAG *CurDAG) {
2968 assert(Compare.getOpcode() == ISD::SETCC &&
2969 "An ISD::SETCC node required here.");
2971 // For values that have a single use, the caller should obviously already have
2972 // checked if that use is an extending use. We check the other uses here.
2973 if (Compare.hasOneUse())
2975 // We want the value in a GPR if it is being extended, used for a select, or
2976 // used in logical operations.
2977 for (auto CompareUse : Compare.getNode()->uses())
2978 if (CompareUse->getOpcode() != ISD::SIGN_EXTEND &&
2979 CompareUse->getOpcode() != ISD::ZERO_EXTEND &&
2980 CompareUse->getOpcode() != ISD::SELECT &&
2981 !isLogicOp(CompareUse->getOpcode())) {
2982 OmittedForNonExtendUses++;
2988 /// Returns an equivalent of a SETCC node but with the result the same width as
2989 /// the inputs. This can nalso be used for SELECT_CC if either the true or false
2990 /// values is a power of two while the other is zero.
2991 SDValue PPCDAGToDAGISel::getSETCCInGPR(SDValue Compare,
2992 SetccInGPROpts ConvOpts) {
2993 assert((Compare.getOpcode() == ISD::SETCC ||
2994 Compare.getOpcode() == ISD::SELECT_CC) &&
2995 "An ISD::SETCC node required here.");
2997 // Don't convert this comparison to a GPR sequence because there are uses
2998 // of the i1 result (i.e. uses that require the result in the CR).
2999 if ((Compare.getOpcode() == ISD::SETCC) && !allUsesExtend(Compare, CurDAG))
3002 SDValue LHS = Compare.getOperand(0);
3003 SDValue RHS = Compare.getOperand(1);
3005 // The condition code is operand 2 for SETCC and operand 4 for SELECT_CC.
3006 int CCOpNum = Compare.getOpcode() == ISD::SELECT_CC ? 4 : 2;
3008 cast<CondCodeSDNode>(Compare.getOperand(CCOpNum))->get();
3009 EVT InputVT = LHS.getValueType();
3010 if (InputVT != MVT::i32 && InputVT != MVT::i64)
3013 if (ConvOpts == SetccInGPROpts::ZExtInvert ||
3014 ConvOpts == SetccInGPROpts::SExtInvert)
3015 CC = ISD::getSetCCInverse(CC, true);
3017 bool Inputs32Bit = InputVT == MVT::i32;
3018 if (ISD::isSignedIntSetCC(CC) && Inputs32Bit) {
3019 LHS = signExtendInputIfNeeded(LHS);
3020 RHS = signExtendInputIfNeeded(RHS);
3021 } else if (ISD::isUnsignedIntSetCC(CC) && Inputs32Bit) {
3022 LHS = zeroExtendInputIfNeeded(LHS);
3023 RHS = zeroExtendInputIfNeeded(RHS);
3027 ConstantSDNode *RHSConst = dyn_cast<ConstantSDNode>(RHS);
3028 int64_t RHSValue = RHSConst ? RHSConst->getSExtValue() : INT64_MAX;
3029 bool IsSext = ConvOpts == SetccInGPROpts::SExtOrig ||
3030 ConvOpts == SetccInGPROpts::SExtInvert;
3032 if (IsSext && Inputs32Bit)
3033 return get32BitSExtCompare(LHS, RHS, CC, RHSValue, dl);
3034 else if (Inputs32Bit)
3035 return get32BitZExtCompare(LHS, RHS, CC, RHSValue, dl);
3037 return get64BitSExtCompare(LHS, RHS, CC, RHSValue, dl);
3038 return get64BitZExtCompare(LHS, RHS, CC, RHSValue, dl);
3041 void PPCDAGToDAGISel::transferMemOperands(SDNode *N, SDNode *Result) {
3042 // Transfer memoperands.
3043 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
3044 MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
3045 cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1);
3048 // Select - Convert the specified operand from a target-independent to a
3049 // target-specific node if it hasn't already been changed.
3050 void PPCDAGToDAGISel::Select(SDNode *N) {
3052 if (N->isMachineOpcode()) {
3054 return; // Already selected.
3057 // In case any misguided DAG-level optimizations form an ADD with a
3058 // TargetConstant operand, crash here instead of miscompiling (by selecting
3059 // an r+r add instead of some kind of r+i add).
3060 if (N->getOpcode() == ISD::ADD &&
3061 N->getOperand(1).getOpcode() == ISD::TargetConstant)
3062 llvm_unreachable("Invalid ADD with TargetConstant operand");
3064 // Try matching complex bit permutations before doing anything else.
3065 if (tryBitPermutation(N))
3068 switch (N->getOpcode()) {
3072 if (N->getValueType(0) == MVT::i64) {
3073 ReplaceNode(N, getInt64(CurDAG, N));
3078 case ISD::ZERO_EXTEND:
3079 case ISD::SIGN_EXTEND:
3089 case PPCISD::GlobalBaseReg:
3090 ReplaceNode(N, getGlobalBaseReg());
3093 case ISD::FrameIndex:
3094 selectFrameIndex(N, N);
3097 case PPCISD::MFOCRF: {
3098 SDValue InFlag = N->getOperand(1);
3099 ReplaceNode(N, CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32,
3100 N->getOperand(0), InFlag));
3104 case PPCISD::READ_TIME_BASE:
3105 ReplaceNode(N, CurDAG->getMachineNode(PPC::ReadTB, dl, MVT::i32, MVT::i32,
3106 MVT::Other, N->getOperand(0)));
3109 case PPCISD::SRA_ADDZE: {
3110 SDValue N0 = N->getOperand(0);
3112 CurDAG->getTargetConstant(*cast<ConstantSDNode>(N->getOperand(1))->
3113 getConstantIntValue(), dl,
3114 N->getValueType(0));
3115 if (N->getValueType(0) == MVT::i64) {
3117 CurDAG->getMachineNode(PPC::SRADI, dl, MVT::i64, MVT::Glue,
3119 CurDAG->SelectNodeTo(N, PPC::ADDZE8, MVT::i64, SDValue(Op, 0),
3123 assert(N->getValueType(0) == MVT::i32 &&
3124 "Expecting i64 or i32 in PPCISD::SRA_ADDZE");
3126 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
3128 CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, SDValue(Op, 0),
3135 // Handle preincrement loads.
3136 LoadSDNode *LD = cast<LoadSDNode>(N);
3137 EVT LoadedVT = LD->getMemoryVT();
3139 // Normal loads are handled by code generated from the .td file.
3140 if (LD->getAddressingMode() != ISD::PRE_INC)
3143 SDValue Offset = LD->getOffset();
3144 if (Offset.getOpcode() == ISD::TargetConstant ||
3145 Offset.getOpcode() == ISD::TargetGlobalAddress) {
3148 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
3149 if (LD->getValueType(0) != MVT::i64) {
3150 // Handle PPC32 integer and normal FP loads.
3151 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
3152 switch (LoadedVT.getSimpleVT().SimpleTy) {
3153 default: llvm_unreachable("Invalid PPC load type!");
3154 case MVT::f64: Opcode = PPC::LFDU; break;
3155 case MVT::f32: Opcode = PPC::LFSU; break;
3156 case MVT::i32: Opcode = PPC::LWZU; break;
3157 case MVT::i16: Opcode = isSExt ? PPC::LHAU : PPC::LHZU; break;
3159 case MVT::i8: Opcode = PPC::LBZU; break;
3162 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
3163 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
3164 switch (LoadedVT.getSimpleVT().SimpleTy) {
3165 default: llvm_unreachable("Invalid PPC load type!");
3166 case MVT::i64: Opcode = PPC::LDU; break;
3167 case MVT::i32: Opcode = PPC::LWZU8; break;
3168 case MVT::i16: Opcode = isSExt ? PPC::LHAU8 : PPC::LHZU8; break;
3170 case MVT::i8: Opcode = PPC::LBZU8; break;
3174 SDValue Chain = LD->getChain();
3175 SDValue Base = LD->getBasePtr();
3176 SDValue Ops[] = { Offset, Base, Chain };
3177 SDNode *MN = CurDAG->getMachineNode(
3178 Opcode, dl, LD->getValueType(0),
3179 PPCLowering->getPointerTy(CurDAG->getDataLayout()), MVT::Other, Ops);
3180 transferMemOperands(N, MN);
3185 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
3186 if (LD->getValueType(0) != MVT::i64) {
3187 // Handle PPC32 integer and normal FP loads.
3188 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
3189 switch (LoadedVT.getSimpleVT().SimpleTy) {
3190 default: llvm_unreachable("Invalid PPC load type!");
3191 case MVT::v4f64: Opcode = PPC::QVLFDUX; break; // QPX
3192 case MVT::v4f32: Opcode = PPC::QVLFSUX; break; // QPX
3193 case MVT::f64: Opcode = PPC::LFDUX; break;
3194 case MVT::f32: Opcode = PPC::LFSUX; break;
3195 case MVT::i32: Opcode = PPC::LWZUX; break;
3196 case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break;
3198 case MVT::i8: Opcode = PPC::LBZUX; break;
3201 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
3202 assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) &&
3203 "Invalid sext update load");
3204 switch (LoadedVT.getSimpleVT().SimpleTy) {
3205 default: llvm_unreachable("Invalid PPC load type!");
3206 case MVT::i64: Opcode = PPC::LDUX; break;
3207 case MVT::i32: Opcode = isSExt ? PPC::LWAUX : PPC::LWZUX8; break;
3208 case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break;
3210 case MVT::i8: Opcode = PPC::LBZUX8; break;
3214 SDValue Chain = LD->getChain();
3215 SDValue Base = LD->getBasePtr();
3216 SDValue Ops[] = { Base, Offset, Chain };
3217 SDNode *MN = CurDAG->getMachineNode(
3218 Opcode, dl, LD->getValueType(0),
3219 PPCLowering->getPointerTy(CurDAG->getDataLayout()), MVT::Other, Ops);
3220 transferMemOperands(N, MN);
3227 if (tryLogicOpOfCompares(N))
3230 unsigned Imm, Imm2, SH, MB, ME;
3233 // If this is an and of a value rotated between 0 and 31 bits and then and'd
3234 // with a mask, emit rlwinm
3235 if (isInt32Immediate(N->getOperand(1), Imm) &&
3236 isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) {
3237 SDValue Val = N->getOperand(0).getOperand(0);
3238 SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl),
3239 getI32Imm(ME, dl) };
3240 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
3243 // If this is just a masked value where the input is not handled above, and
3244 // is not a rotate-left (handled by a pattern in the .td file), emit rlwinm
3245 if (isInt32Immediate(N->getOperand(1), Imm) &&
3246 isRunOfOnes(Imm, MB, ME) &&
3247 N->getOperand(0).getOpcode() != ISD::ROTL) {
3248 SDValue Val = N->getOperand(0);
3249 SDValue Ops[] = { Val, getI32Imm(0, dl), getI32Imm(MB, dl),
3250 getI32Imm(ME, dl) };
3251 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
3254 // If this is a 64-bit zero-extension mask, emit rldicl.
3255 if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) &&
3257 SDValue Val = N->getOperand(0);
3258 MB = 64 - countTrailingOnes(Imm64);
3261 if (Val.getOpcode() == ISD::ANY_EXTEND) {
3262 auto Op0 = Val.getOperand(0);
3263 if ( Op0.getOpcode() == ISD::SRL &&
3264 isInt32Immediate(Op0.getOperand(1).getNode(), Imm) && Imm <= MB) {
3266 auto ResultType = Val.getNode()->getValueType(0);
3267 auto ImDef = CurDAG->getMachineNode(PPC::IMPLICIT_DEF, dl,
3269 SDValue IDVal (ImDef, 0);
3271 Val = SDValue(CurDAG->getMachineNode(PPC::INSERT_SUBREG, dl,
3272 ResultType, IDVal, Op0.getOperand(0),
3273 getI32Imm(1, dl)), 0);
3278 // If the operand is a logical right shift, we can fold it into this
3279 // instruction: rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb)
3280 // for n <= mb. The right shift is really a left rotate followed by a
3281 // mask, and this mask is a more-restrictive sub-mask of the mask implied
3283 if (Val.getOpcode() == ISD::SRL &&
3284 isInt32Immediate(Val.getOperand(1).getNode(), Imm) && Imm <= MB) {
3285 assert(Imm < 64 && "Illegal shift amount");
3286 Val = Val.getOperand(0);
3290 SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl) };
3291 CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops);
3294 // If this is a negated 64-bit zero-extension mask,
3295 // i.e. the immediate is a sequence of ones from most significant side
3296 // and all zero for reminder, we should use rldicr.
3297 if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) &&
3298 isMask_64(~Imm64)) {
3299 SDValue Val = N->getOperand(0);
3300 MB = 63 - countTrailingOnes(~Imm64);
3302 SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl) };
3303 CurDAG->SelectNodeTo(N, PPC::RLDICR, MVT::i64, Ops);
3307 // AND X, 0 -> 0, not "rlwinm 32".
3308 if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) {
3309 ReplaceUses(SDValue(N, 0), N->getOperand(1));
3312 // ISD::OR doesn't get all the bitfield insertion fun.
3313 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) might be a
3315 if (isInt32Immediate(N->getOperand(1), Imm) &&
3316 N->getOperand(0).getOpcode() == ISD::OR &&
3317 isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) {
3318 // The idea here is to check whether this is equivalent to:
3319 // (c1 & m) | (x & ~m)
3320 // where m is a run-of-ones mask. The logic here is that, for each bit in
3322 // - if both are 1, then the output will be 1.
3323 // - if both are 0, then the output will be 0.
3324 // - if the bit in c1 is 0, and the bit in c2 is 1, then the output will
3326 // - if the bit in c1 is 1, and the bit in c2 is 0, then the output will
3328 // If that last condition is never the case, then we can form m from the
3329 // bits that are the same between c1 and c2.
3331 if (isRunOfOnes(~(Imm^Imm2), MB, ME) && !(~Imm & Imm2)) {
3332 SDValue Ops[] = { N->getOperand(0).getOperand(0),
3333 N->getOperand(0).getOperand(1),
3334 getI32Imm(0, dl), getI32Imm(MB, dl),
3335 getI32Imm(ME, dl) };
3336 ReplaceNode(N, CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops));
3341 // Other cases are autogenerated.
3345 if (N->getValueType(0) == MVT::i32)
3346 if (tryBitfieldInsert(N))
3349 if (tryLogicOpOfCompares(N))
3353 if (N->getOperand(0)->getOpcode() == ISD::FrameIndex &&
3354 isIntS16Immediate(N->getOperand(1), Imm)) {
3356 CurDAG->computeKnownBits(N->getOperand(0), LHSKnown);
3358 // If this is equivalent to an add, then we can fold it with the
3359 // FrameIndex calculation.
3360 if ((LHSKnown.Zero.getZExtValue()|~(uint64_t)Imm) == ~0ULL) {
3361 selectFrameIndex(N, N->getOperand(0).getNode(), (int)Imm);
3366 // Other cases are autogenerated.
3370 if (tryLogicOpOfCompares(N))
3376 if (N->getOperand(0)->getOpcode() == ISD::FrameIndex &&
3377 isIntS16Immediate(N->getOperand(1), Imm)) {
3378 selectFrameIndex(N, N->getOperand(0).getNode(), (int)Imm);
3385 unsigned Imm, SH, MB, ME;
3386 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
3387 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
3388 SDValue Ops[] = { N->getOperand(0).getOperand(0),
3389 getI32Imm(SH, dl), getI32Imm(MB, dl),
3390 getI32Imm(ME, dl) };
3391 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
3395 // Other cases are autogenerated.
3399 unsigned Imm, SH, MB, ME;
3400 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
3401 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
3402 SDValue Ops[] = { N->getOperand(0).getOperand(0),
3403 getI32Imm(SH, dl), getI32Imm(MB, dl),
3404 getI32Imm(ME, dl) };
3405 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
3409 // Other cases are autogenerated.
3412 // FIXME: Remove this once the ANDI glue bug is fixed:
3413 case PPCISD::ANDIo_1_EQ_BIT:
3414 case PPCISD::ANDIo_1_GT_BIT: {
3418 EVT InVT = N->getOperand(0).getValueType();
3419 assert((InVT == MVT::i64 || InVT == MVT::i32) &&
3420 "Invalid input type for ANDIo_1_EQ_BIT");
3422 unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDIo8 : PPC::ANDIo;
3423 SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue,
3425 CurDAG->getTargetConstant(1, dl, InVT)),
3427 SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32);
3429 CurDAG->getTargetConstant(N->getOpcode() == PPCISD::ANDIo_1_EQ_BIT ?
3430 PPC::sub_eq : PPC::sub_gt, dl, MVT::i32);
3432 CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1, CR0Reg,
3433 SRIdxVal, SDValue(AndI.getNode(), 1) /* glue */);
3436 case ISD::SELECT_CC: {
3437 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
3439 CurDAG->getTargetLoweringInfo().getPointerTy(CurDAG->getDataLayout());
3440 bool isPPC64 = (PtrVT == MVT::i64);
3442 // If this is a select of i1 operands, we'll pattern match it.
3443 if (PPCSubTarget->useCRBits() &&
3444 N->getOperand(0).getValueType() == MVT::i1)
3447 // Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
3449 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
3450 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
3451 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
3452 if (N1C->isNullValue() && N3C->isNullValue() &&
3453 N2C->getZExtValue() == 1ULL && CC == ISD::SETNE &&
3454 // FIXME: Implement this optzn for PPC64.
3455 N->getValueType(0) == MVT::i32) {
3457 CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
3458 N->getOperand(0), getI32Imm(~0U, dl));
3459 CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(Tmp, 0),
3460 N->getOperand(0), SDValue(Tmp, 1));
3464 SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl);
3466 if (N->getValueType(0) == MVT::i1) {
3467 // An i1 select is: (c & t) | (!c & f).
3469 unsigned Idx = getCRIdxForSetCC(CC, Inv);
3473 default: llvm_unreachable("Invalid CC index");
3474 case 0: SRI = PPC::sub_lt; break;
3475 case 1: SRI = PPC::sub_gt; break;
3476 case 2: SRI = PPC::sub_eq; break;
3477 case 3: SRI = PPC::sub_un; break;
3480 SDValue CCBit = CurDAG->getTargetExtractSubreg(SRI, dl, MVT::i1, CCReg);
3482 SDValue NotCCBit(CurDAG->getMachineNode(PPC::CRNOR, dl, MVT::i1,
3484 SDValue C = Inv ? NotCCBit : CCBit,
3485 NotC = Inv ? CCBit : NotCCBit;
3487 SDValue CAndT(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1,
3488 C, N->getOperand(2)), 0);
3489 SDValue NotCAndF(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1,
3490 NotC, N->getOperand(3)), 0);
3492 CurDAG->SelectNodeTo(N, PPC::CROR, MVT::i1, CAndT, NotCAndF);
3496 unsigned BROpc = getPredicateForSetCC(CC);
3498 unsigned SelectCCOp;
3499 if (N->getValueType(0) == MVT::i32)
3500 SelectCCOp = PPC::SELECT_CC_I4;
3501 else if (N->getValueType(0) == MVT::i64)
3502 SelectCCOp = PPC::SELECT_CC_I8;
3503 else if (N->getValueType(0) == MVT::f32)
3504 if (PPCSubTarget->hasP8Vector())
3505 SelectCCOp = PPC::SELECT_CC_VSSRC;
3507 SelectCCOp = PPC::SELECT_CC_F4;
3508 else if (N->getValueType(0) == MVT::f64)
3509 if (PPCSubTarget->hasVSX())
3510 SelectCCOp = PPC::SELECT_CC_VSFRC;
3512 SelectCCOp = PPC::SELECT_CC_F8;
3513 else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4f64)
3514 SelectCCOp = PPC::SELECT_CC_QFRC;
3515 else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4f32)
3516 SelectCCOp = PPC::SELECT_CC_QSRC;
3517 else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4i1)
3518 SelectCCOp = PPC::SELECT_CC_QBRC;
3519 else if (N->getValueType(0) == MVT::v2f64 ||
3520 N->getValueType(0) == MVT::v2i64)
3521 SelectCCOp = PPC::SELECT_CC_VSRC;
3523 SelectCCOp = PPC::SELECT_CC_VRRC;
3525 SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3),
3526 getI32Imm(BROpc, dl) };
3527 CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops);
3531 if (PPCSubTarget->hasVSX()) {
3532 SDValue Ops[] = { N->getOperand(2), N->getOperand(1), N->getOperand(0) };
3533 CurDAG->SelectNodeTo(N, PPC::XXSEL, N->getValueType(0), Ops);
3538 case ISD::VECTOR_SHUFFLE:
3539 if (PPCSubTarget->hasVSX() && (N->getValueType(0) == MVT::v2f64 ||
3540 N->getValueType(0) == MVT::v2i64)) {
3541 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
3543 SDValue Op1 = N->getOperand(SVN->getMaskElt(0) < 2 ? 0 : 1),
3544 Op2 = N->getOperand(SVN->getMaskElt(1) < 2 ? 0 : 1);
3547 for (int i = 0; i < 2; ++i)
3548 if (SVN->getMaskElt(i) <= 0 || SVN->getMaskElt(i) == 2)
3553 if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 &&
3554 Op1.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3555 isa<LoadSDNode>(Op1.getOperand(0))) {
3556 LoadSDNode *LD = cast<LoadSDNode>(Op1.getOperand(0));
3557 SDValue Base, Offset;
3559 if (LD->isUnindexed() && LD->hasOneUse() && Op1.hasOneUse() &&
3560 (LD->getMemoryVT() == MVT::f64 ||
3561 LD->getMemoryVT() == MVT::i64) &&
3562 SelectAddrIdxOnly(LD->getBasePtr(), Base, Offset)) {
3563 SDValue Chain = LD->getChain();
3564 SDValue Ops[] = { Base, Offset, Chain };
3565 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
3566 MemOp[0] = LD->getMemOperand();
3567 SDNode *NewN = CurDAG->SelectNodeTo(N, PPC::LXVDSX,
3568 N->getValueType(0), Ops);
3569 cast<MachineSDNode>(NewN)->setMemRefs(MemOp, MemOp + 1);
3574 // For little endian, we must swap the input operands and adjust
3575 // the mask elements (reverse and invert them).
3576 if (PPCSubTarget->isLittleEndian()) {
3577 std::swap(Op1, Op2);
3578 unsigned tmp = DM[0];
3583 SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), dl,
3585 SDValue Ops[] = { Op1, Op2, DMV };
3586 CurDAG->SelectNodeTo(N, PPC::XXPERMDI, N->getValueType(0), Ops);
3593 bool IsPPC64 = PPCSubTarget->isPPC64();
3594 SDValue Ops[] = { N->getOperand(1), N->getOperand(0) };
3595 CurDAG->SelectNodeTo(N, N->getOpcode() == PPCISD::BDNZ
3596 ? (IsPPC64 ? PPC::BDNZ8 : PPC::BDNZ)
3597 : (IsPPC64 ? PPC::BDZ8 : PPC::BDZ),
3601 case PPCISD::COND_BRANCH: {
3602 // Op #0 is the Chain.
3603 // Op #1 is the PPC::PRED_* number.
3605 // Op #3 is the Dest MBB
3606 // Op #4 is the Flag.
3607 // Prevent PPC::PRED_* from being selected into LI.
3608 unsigned PCC = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
3609 if (EnableBranchHint)
3610 PCC |= getBranchHint(PCC, FuncInfo, N->getOperand(3));
3612 SDValue Pred = getI32Imm(PCC, dl);
3613 SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3),
3614 N->getOperand(0), N->getOperand(4) };
3615 CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
3619 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
3620 unsigned PCC = getPredicateForSetCC(CC);
3622 if (N->getOperand(2).getValueType() == MVT::i1) {
3626 default: llvm_unreachable("Unexpected Boolean-operand predicate");
3627 case PPC::PRED_LT: Opc = PPC::CRANDC; Swap = true; break;
3628 case PPC::PRED_LE: Opc = PPC::CRORC; Swap = true; break;
3629 case PPC::PRED_EQ: Opc = PPC::CREQV; Swap = false; break;
3630 case PPC::PRED_GE: Opc = PPC::CRORC; Swap = false; break;
3631 case PPC::PRED_GT: Opc = PPC::CRANDC; Swap = false; break;
3632 case PPC::PRED_NE: Opc = PPC::CRXOR; Swap = false; break;
3635 SDValue BitComp(CurDAG->getMachineNode(Opc, dl, MVT::i1,
3636 N->getOperand(Swap ? 3 : 2),
3637 N->getOperand(Swap ? 2 : 3)), 0);
3638 CurDAG->SelectNodeTo(N, PPC::BC, MVT::Other, BitComp, N->getOperand(4),
3643 if (EnableBranchHint)
3644 PCC |= getBranchHint(PCC, FuncInfo, N->getOperand(4));
3646 SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl);
3647 SDValue Ops[] = { getI32Imm(PCC, dl), CondCode,
3648 N->getOperand(4), N->getOperand(0) };
3649 CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
3653 // FIXME: Should custom lower this.
3654 SDValue Chain = N->getOperand(0);
3655 SDValue Target = N->getOperand(1);
3656 unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8;
3657 unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8;
3658 Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Target,
3660 CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain);
3663 case PPCISD::TOC_ENTRY: {
3664 assert ((PPCSubTarget->isPPC64() || PPCSubTarget->isSVR4ABI()) &&
3665 "Only supported for 64-bit ABI and 32-bit SVR4");
3666 if (PPCSubTarget->isSVR4ABI() && !PPCSubTarget->isPPC64()) {
3667 SDValue GA = N->getOperand(0);
3668 SDNode *MN = CurDAG->getMachineNode(PPC::LWZtoc, dl, MVT::i32, GA,
3670 transferMemOperands(N, MN);
3675 // For medium and large code model, we generate two instructions as
3676 // described below. Otherwise we allow SelectCodeCommon to handle this,
3677 // selecting one of LDtoc, LDtocJTI, LDtocCPT, and LDtocBA.
3678 CodeModel::Model CModel = TM.getCodeModel();
3679 if (CModel != CodeModel::Medium && CModel != CodeModel::Large)
3682 // The first source operand is a TargetGlobalAddress or a TargetJumpTable.
3683 // If it must be toc-referenced according to PPCSubTarget, we generate:
3684 // LDtocL(<ga:@sym>, ADDIStocHA(%X2, <ga:@sym>))
3685 // Otherwise we generate:
3686 // ADDItocL(ADDIStocHA(%X2, <ga:@sym>), <ga:@sym>)
3687 SDValue GA = N->getOperand(0);
3688 SDValue TOCbase = N->getOperand(1);
3689 SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64,
3692 if (isa<JumpTableSDNode>(GA) || isa<BlockAddressSDNode>(GA) ||
3693 CModel == CodeModel::Large) {
3694 SDNode *MN = CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
3696 transferMemOperands(N, MN);
3701 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) {
3702 const GlobalValue *GV = G->getGlobal();
3703 unsigned char GVFlags = PPCSubTarget->classifyGlobalReference(GV);
3704 if (GVFlags & PPCII::MO_NLP_FLAG) {
3705 SDNode *MN = CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
3707 transferMemOperands(N, MN);
3713 ReplaceNode(N, CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64,
3714 SDValue(Tmp, 0), GA));
3717 case PPCISD::PPC32_PICGOT:
3718 // Generate a PIC-safe GOT reference.
3719 assert(!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI() &&
3720 "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4");
3721 CurDAG->SelectNodeTo(N, PPC::PPC32PICGOT,
3722 PPCLowering->getPointerTy(CurDAG->getDataLayout()),
3726 case PPCISD::VADD_SPLAT: {
3727 // This expands into one of three sequences, depending on whether
3728 // the first operand is odd or even, positive or negative.
3729 assert(isa<ConstantSDNode>(N->getOperand(0)) &&
3730 isa<ConstantSDNode>(N->getOperand(1)) &&
3731 "Invalid operand on VADD_SPLAT!");
3733 int Elt = N->getConstantOperandVal(0);
3734 int EltSize = N->getConstantOperandVal(1);
3735 unsigned Opc1, Opc2, Opc3;
3739 Opc1 = PPC::VSPLTISB;
3740 Opc2 = PPC::VADDUBM;
3741 Opc3 = PPC::VSUBUBM;
3743 } else if (EltSize == 2) {
3744 Opc1 = PPC::VSPLTISH;
3745 Opc2 = PPC::VADDUHM;
3746 Opc3 = PPC::VSUBUHM;
3749 assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!");
3750 Opc1 = PPC::VSPLTISW;
3751 Opc2 = PPC::VADDUWM;
3752 Opc3 = PPC::VSUBUWM;
3756 if ((Elt & 1) == 0) {
3757 // Elt is even, in the range [-32,-18] + [16,30].
3759 // Convert: VADD_SPLAT elt, size
3760 // Into: tmp = VSPLTIS[BHW] elt
3761 // VADDU[BHW]M tmp, tmp
3762 // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4
3763 SDValue EltVal = getI32Imm(Elt >> 1, dl);
3764 SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
3765 SDValue TmpVal = SDValue(Tmp, 0);
3766 ReplaceNode(N, CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal));
3768 } else if (Elt > 0) {
3769 // Elt is odd and positive, in the range [17,31].
3771 // Convert: VADD_SPLAT elt, size
3772 // Into: tmp1 = VSPLTIS[BHW] elt-16
3773 // tmp2 = VSPLTIS[BHW] -16
3774 // VSUBU[BHW]M tmp1, tmp2
3775 SDValue EltVal = getI32Imm(Elt - 16, dl);
3776 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
3777 EltVal = getI32Imm(-16, dl);
3778 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
3779 ReplaceNode(N, CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0),
3783 // Elt is odd and negative, in the range [-31,-17].
3785 // Convert: VADD_SPLAT elt, size
3786 // Into: tmp1 = VSPLTIS[BHW] elt+16
3787 // tmp2 = VSPLTIS[BHW] -16
3788 // VADDU[BHW]M tmp1, tmp2
3789 SDValue EltVal = getI32Imm(Elt + 16, dl);
3790 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
3791 EltVal = getI32Imm(-16, dl);
3792 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
3793 ReplaceNode(N, CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0),
3803 // If the target supports the cmpb instruction, do the idiom recognition here.
3804 // We don't do this as a DAG combine because we don't want to do it as nodes
3805 // are being combined (because we might miss part of the eventual idiom). We
3806 // don't want to do it during instruction selection because we want to reuse
3807 // the logic for lowering the masking operations already part of the
3808 // instruction selector.
3809 SDValue PPCDAGToDAGISel::combineToCMPB(SDNode *N) {
3812 assert(N->getOpcode() == ISD::OR &&
3813 "Only OR nodes are supported for CMPB");
3816 if (!PPCSubTarget->hasCMPB())
3819 if (N->getValueType(0) != MVT::i32 &&
3820 N->getValueType(0) != MVT::i64)
3823 EVT VT = N->getValueType(0);
3826 bool BytesFound[8] = {false, false, false, false, false, false, false, false};
3827 uint64_t Mask = 0, Alt = 0;
3829 auto IsByteSelectCC = [this](SDValue O, unsigned &b,
3830 uint64_t &Mask, uint64_t &Alt,
3831 SDValue &LHS, SDValue &RHS) {
3832 if (O.getOpcode() != ISD::SELECT_CC)
3834 ISD::CondCode CC = cast<CondCodeSDNode>(O.getOperand(4))->get();
3836 if (!isa<ConstantSDNode>(O.getOperand(2)) ||
3837 !isa<ConstantSDNode>(O.getOperand(3)))
3840 uint64_t PM = O.getConstantOperandVal(2);
3841 uint64_t PAlt = O.getConstantOperandVal(3);
3842 for (b = 0; b < 8; ++b) {
3843 uint64_t Mask = UINT64_C(0xFF) << (8*b);
3844 if (PM && (PM & Mask) == PM && (PAlt & Mask) == PAlt)
3853 if (!isa<ConstantSDNode>(O.getOperand(1)) ||
3854 O.getConstantOperandVal(1) != 0) {
3855 SDValue Op0 = O.getOperand(0), Op1 = O.getOperand(1);
3856 if (Op0.getOpcode() == ISD::TRUNCATE)
3857 Op0 = Op0.getOperand(0);
3858 if (Op1.getOpcode() == ISD::TRUNCATE)
3859 Op1 = Op1.getOperand(0);
3861 if (Op0.getOpcode() == ISD::SRL && Op1.getOpcode() == ISD::SRL &&
3862 Op0.getOperand(1) == Op1.getOperand(1) && CC == ISD::SETEQ &&
3863 isa<ConstantSDNode>(Op0.getOperand(1))) {
3865 unsigned Bits = Op0.getValueSizeInBits();
3868 if (Op0.getConstantOperandVal(1) != Bits-8)
3871 LHS = Op0.getOperand(0);
3872 RHS = Op1.getOperand(0);
3876 // When we have small integers (i16 to be specific), the form present
3877 // post-legalization uses SETULT in the SELECT_CC for the
3878 // higher-order byte, depending on the fact that the
3879 // even-higher-order bytes are known to all be zero, for example:
3880 // select_cc (xor $lhs, $rhs), 256, 65280, 0, setult
3881 // (so when the second byte is the same, because all higher-order
3882 // bits from bytes 3 and 4 are known to be zero, the result of the
3883 // xor can be at most 255)
3884 if (Op0.getOpcode() == ISD::XOR && CC == ISD::SETULT &&
3885 isa<ConstantSDNode>(O.getOperand(1))) {
3887 uint64_t ULim = O.getConstantOperandVal(1);
3888 if (ULim != (UINT64_C(1) << b*8))
3891 // Now we need to make sure that the upper bytes are known to be
3893 unsigned Bits = Op0.getValueSizeInBits();
3894 if (!CurDAG->MaskedValueIsZero(
3895 Op0, APInt::getHighBitsSet(Bits, Bits - (b + 1) * 8)))
3898 LHS = Op0.getOperand(0);
3899 RHS = Op0.getOperand(1);
3906 if (CC != ISD::SETEQ)
3909 SDValue Op = O.getOperand(0);
3910 if (Op.getOpcode() == ISD::AND) {
3911 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3913 if (Op.getConstantOperandVal(1) != (UINT64_C(0xFF) << (8*b)))
3916 SDValue XOR = Op.getOperand(0);
3917 if (XOR.getOpcode() == ISD::TRUNCATE)
3918 XOR = XOR.getOperand(0);
3919 if (XOR.getOpcode() != ISD::XOR)
3922 LHS = XOR.getOperand(0);
3923 RHS = XOR.getOperand(1);
3925 } else if (Op.getOpcode() == ISD::SRL) {
3926 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3928 unsigned Bits = Op.getValueSizeInBits();
3931 if (Op.getConstantOperandVal(1) != Bits-8)
3934 SDValue XOR = Op.getOperand(0);
3935 if (XOR.getOpcode() == ISD::TRUNCATE)
3936 XOR = XOR.getOperand(0);
3937 if (XOR.getOpcode() != ISD::XOR)
3940 LHS = XOR.getOperand(0);
3941 RHS = XOR.getOperand(1);
3948 SmallVector<SDValue, 8> Queue(1, SDValue(N, 0));
3949 while (!Queue.empty()) {
3950 SDValue V = Queue.pop_back_val();
3952 for (const SDValue &O : V.getNode()->ops()) {
3954 uint64_t M = 0, A = 0;
3956 if (O.getOpcode() == ISD::OR) {
3958 } else if (IsByteSelectCC(O, b, M, A, OLHS, ORHS)) {
3962 BytesFound[b] = true;
3965 } else if ((LHS == ORHS && RHS == OLHS) ||
3966 (RHS == ORHS && LHS == OLHS)) {
3967 BytesFound[b] = true;
3979 unsigned LastB = 0, BCnt = 0;
3980 for (unsigned i = 0; i < 8; ++i)
3981 if (BytesFound[LastB]) {
3986 if (!LastB || BCnt < 2)
3989 // Because we'll be zero-extending the output anyway if don't have a specific
3990 // value for each input byte (via the Mask), we can 'anyext' the inputs.
3991 if (LHS.getValueType() != VT) {
3992 LHS = CurDAG->getAnyExtOrTrunc(LHS, dl, VT);
3993 RHS = CurDAG->getAnyExtOrTrunc(RHS, dl, VT);
3996 Res = CurDAG->getNode(PPCISD::CMPB, dl, VT, LHS, RHS);
3998 bool NonTrivialMask = ((int64_t) Mask) != INT64_C(-1);
3999 if (NonTrivialMask && !Alt) {
4000 // Res = Mask & CMPB
4001 Res = CurDAG->getNode(ISD::AND, dl, VT, Res,
4002 CurDAG->getConstant(Mask, dl, VT));
4004 // Res = (CMPB & Mask) | (~CMPB & Alt)
4005 // Which, as suggested here:
4006 // https://graphics.stanford.edu/~seander/bithacks.html#MaskedMerge
4007 // can be written as:
4008 // Res = Alt ^ ((Alt ^ Mask) & CMPB)
4009 // useful because the (Alt ^ Mask) can be pre-computed.
4010 Res = CurDAG->getNode(ISD::AND, dl, VT, Res,
4011 CurDAG->getConstant(Mask ^ Alt, dl, VT));
4012 Res = CurDAG->getNode(ISD::XOR, dl, VT, Res,
4013 CurDAG->getConstant(Alt, dl, VT));
4019 // When CR bit registers are enabled, an extension of an i1 variable to a i32
4020 // or i64 value is lowered in terms of a SELECT_I[48] operation, and thus
4021 // involves constant materialization of a 0 or a 1 or both. If the result of
4022 // the extension is then operated upon by some operator that can be constant
4023 // folded with a constant 0 or 1, and that constant can be materialized using
4024 // only one instruction (like a zero or one), then we should fold in those
4025 // operations with the select.
4026 void PPCDAGToDAGISel::foldBoolExts(SDValue &Res, SDNode *&N) {
4027 if (!PPCSubTarget->useCRBits())
4030 if (N->getOpcode() != ISD::ZERO_EXTEND &&
4031 N->getOpcode() != ISD::SIGN_EXTEND &&
4032 N->getOpcode() != ISD::ANY_EXTEND)
4035 if (N->getOperand(0).getValueType() != MVT::i1)
4038 if (!N->hasOneUse())
4042 EVT VT = N->getValueType(0);
4043 SDValue Cond = N->getOperand(0);
4045 CurDAG->getConstant(N->getOpcode() == ISD::SIGN_EXTEND ? -1 : 1, dl, VT);
4046 SDValue ConstFalse = CurDAG->getConstant(0, dl, VT);
4049 SDNode *User = *N->use_begin();
4050 if (User->getNumOperands() != 2)
4053 auto TryFold = [this, N, User, dl](SDValue Val) {
4054 SDValue UserO0 = User->getOperand(0), UserO1 = User->getOperand(1);
4055 SDValue O0 = UserO0.getNode() == N ? Val : UserO0;
4056 SDValue O1 = UserO1.getNode() == N ? Val : UserO1;
4058 return CurDAG->FoldConstantArithmetic(User->getOpcode(), dl,
4059 User->getValueType(0),
4060 O0.getNode(), O1.getNode());
4063 SDValue TrueRes = TryFold(ConstTrue);
4066 SDValue FalseRes = TryFold(ConstFalse);
4070 // For us to materialize these using one instruction, we must be able to
4071 // represent them as signed 16-bit integers.
4072 uint64_t True = cast<ConstantSDNode>(TrueRes)->getZExtValue(),
4073 False = cast<ConstantSDNode>(FalseRes)->getZExtValue();
4074 if (!isInt<16>(True) || !isInt<16>(False))
4077 // We can replace User with a new SELECT node, and try again to see if we
4078 // can fold the select with its user.
4079 Res = CurDAG->getSelect(dl, User->getValueType(0), Cond, TrueRes, FalseRes);
4081 ConstTrue = TrueRes;
4082 ConstFalse = FalseRes;
4083 } while (N->hasOneUse());
4086 void PPCDAGToDAGISel::PreprocessISelDAG() {
4087 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
4090 bool MadeChange = false;
4091 while (Position != CurDAG->allnodes_begin()) {
4092 SDNode *N = &*--Position;
4097 switch (N->getOpcode()) {
4100 Res = combineToCMPB(N);
4105 foldBoolExts(Res, N);
4108 DEBUG(dbgs() << "PPC DAG preprocessing replacing:\nOld: ");
4109 DEBUG(N->dump(CurDAG));
4110 DEBUG(dbgs() << "\nNew: ");
4111 DEBUG(Res.getNode()->dump(CurDAG));
4112 DEBUG(dbgs() << "\n");
4114 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Res);
4120 CurDAG->RemoveDeadNodes();
4123 /// PostprocessISelDAG - Perform some late peephole optimizations
4124 /// on the DAG representation.
4125 void PPCDAGToDAGISel::PostprocessISelDAG() {
4126 // Skip peepholes at -O0.
4127 if (TM.getOptLevel() == CodeGenOpt::None)
4132 PeepholePPC64ZExt();
4135 // Check if all users of this node will become isel where the second operand
4136 // is the constant zero. If this is so, and if we can negate the condition,
4137 // then we can flip the true and false operands. This will allow the zero to
4138 // be folded with the isel so that we don't need to materialize a register
4140 bool PPCDAGToDAGISel::AllUsersSelectZero(SDNode *N) {
4141 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4144 if (!User->isMachineOpcode())
4146 if (User->getMachineOpcode() != PPC::SELECT_I4 &&
4147 User->getMachineOpcode() != PPC::SELECT_I8)
4150 SDNode *Op2 = User->getOperand(2).getNode();
4151 if (!Op2->isMachineOpcode())
4154 if (Op2->getMachineOpcode() != PPC::LI &&
4155 Op2->getMachineOpcode() != PPC::LI8)
4158 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op2->getOperand(0));
4162 if (!C->isNullValue())
4169 void PPCDAGToDAGISel::SwapAllSelectUsers(SDNode *N) {
4170 SmallVector<SDNode *, 4> ToReplace;
4171 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4174 assert((User->getMachineOpcode() == PPC::SELECT_I4 ||
4175 User->getMachineOpcode() == PPC::SELECT_I8) &&
4176 "Must have all select users");
4177 ToReplace.push_back(User);
4180 for (SmallVector<SDNode *, 4>::iterator UI = ToReplace.begin(),
4181 UE = ToReplace.end(); UI != UE; ++UI) {
4184 CurDAG->getMachineNode(User->getMachineOpcode(), SDLoc(User),
4185 User->getValueType(0), User->getOperand(0),
4186 User->getOperand(2),
4187 User->getOperand(1));
4189 DEBUG(dbgs() << "CR Peephole replacing:\nOld: ");
4190 DEBUG(User->dump(CurDAG));
4191 DEBUG(dbgs() << "\nNew: ");
4192 DEBUG(ResNode->dump(CurDAG));
4193 DEBUG(dbgs() << "\n");
4195 ReplaceUses(User, ResNode);
4199 void PPCDAGToDAGISel::PeepholeCROps() {
4203 for (SDNode &Node : CurDAG->allnodes()) {
4204 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(&Node);
4205 if (!MachineNode || MachineNode->use_empty())
4207 SDNode *ResNode = MachineNode;
4209 bool Op1Set = false, Op1Unset = false,
4211 Op2Set = false, Op2Unset = false,
4214 unsigned Opcode = MachineNode->getMachineOpcode();
4225 SDValue Op = MachineNode->getOperand(1);
4226 if (Op.isMachineOpcode()) {
4227 if (Op.getMachineOpcode() == PPC::CRSET)
4229 else if (Op.getMachineOpcode() == PPC::CRUNSET)
4231 else if (Op.getMachineOpcode() == PPC::CRNOR &&
4232 Op.getOperand(0) == Op.getOperand(1))
4239 case PPC::SELECT_I4:
4240 case PPC::SELECT_I8:
4241 case PPC::SELECT_F4:
4242 case PPC::SELECT_F8:
4243 case PPC::SELECT_QFRC:
4244 case PPC::SELECT_QSRC:
4245 case PPC::SELECT_QBRC:
4246 case PPC::SELECT_VRRC:
4247 case PPC::SELECT_VSFRC:
4248 case PPC::SELECT_VSSRC:
4249 case PPC::SELECT_VSRC: {
4250 SDValue Op = MachineNode->getOperand(0);
4251 if (Op.isMachineOpcode()) {
4252 if (Op.getMachineOpcode() == PPC::CRSET)
4254 else if (Op.getMachineOpcode() == PPC::CRUNSET)
4256 else if (Op.getMachineOpcode() == PPC::CRNOR &&
4257 Op.getOperand(0) == Op.getOperand(1))
4264 bool SelectSwap = false;
4268 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
4270 ResNode = MachineNode->getOperand(0).getNode();
4273 ResNode = MachineNode->getOperand(1).getNode();
4276 ResNode = MachineNode->getOperand(0).getNode();
4277 else if (Op1Unset || Op2Unset)
4278 // x & 0 = 0 & y = 0
4279 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
4282 // ~x & y = andc(y, x)
4283 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
4284 MVT::i1, MachineNode->getOperand(1),
4285 MachineNode->getOperand(0).
4288 // x & ~y = andc(x, y)
4289 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
4290 MVT::i1, MachineNode->getOperand(0),
4291 MachineNode->getOperand(1).
4293 else if (AllUsersSelectZero(MachineNode)) {
4294 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode),
4295 MVT::i1, MachineNode->getOperand(0),
4296 MachineNode->getOperand(1));
4301 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
4302 // nand(x, x) -> nor(x, x)
4303 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4304 MVT::i1, MachineNode->getOperand(0),
4305 MachineNode->getOperand(0));
4307 // nand(1, y) -> nor(y, y)
4308 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4309 MVT::i1, MachineNode->getOperand(1),
4310 MachineNode->getOperand(1));
4312 // nand(x, 1) -> nor(x, x)
4313 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4314 MVT::i1, MachineNode->getOperand(0),
4315 MachineNode->getOperand(0));
4316 else if (Op1Unset || Op2Unset)
4317 // nand(x, 0) = nand(0, y) = 1
4318 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
4321 // nand(~x, y) = ~(~x & y) = x | ~y = orc(x, y)
4322 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
4323 MVT::i1, MachineNode->getOperand(0).
4325 MachineNode->getOperand(1));
4327 // nand(x, ~y) = ~x | y = orc(y, x)
4328 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
4329 MVT::i1, MachineNode->getOperand(1).
4331 MachineNode->getOperand(0));
4332 else if (AllUsersSelectZero(MachineNode)) {
4333 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode),
4334 MVT::i1, MachineNode->getOperand(0),
4335 MachineNode->getOperand(1));
4340 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
4342 ResNode = MachineNode->getOperand(0).getNode();
4343 else if (Op1Set || Op2Set)
4344 // x | 1 = 1 | y = 1
4345 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
4349 ResNode = MachineNode->getOperand(1).getNode();
4352 ResNode = MachineNode->getOperand(0).getNode();
4354 // ~x | y = orc(y, x)
4355 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
4356 MVT::i1, MachineNode->getOperand(1),
4357 MachineNode->getOperand(0).
4360 // x | ~y = orc(x, y)
4361 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
4362 MVT::i1, MachineNode->getOperand(0),
4363 MachineNode->getOperand(1).
4365 else if (AllUsersSelectZero(MachineNode)) {
4366 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4367 MVT::i1, MachineNode->getOperand(0),
4368 MachineNode->getOperand(1));
4373 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
4375 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
4378 // xor(1, y) -> nor(y, y)
4379 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4380 MVT::i1, MachineNode->getOperand(1),
4381 MachineNode->getOperand(1));
4383 // xor(x, 1) -> nor(x, x)
4384 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4385 MVT::i1, MachineNode->getOperand(0),
4386 MachineNode->getOperand(0));
4389 ResNode = MachineNode->getOperand(1).getNode();
4392 ResNode = MachineNode->getOperand(0).getNode();
4394 // xor(~x, y) = eqv(x, y)
4395 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
4396 MVT::i1, MachineNode->getOperand(0).
4398 MachineNode->getOperand(1));
4400 // xor(x, ~y) = eqv(x, y)
4401 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
4402 MVT::i1, MachineNode->getOperand(0),
4403 MachineNode->getOperand(1).
4405 else if (AllUsersSelectZero(MachineNode)) {
4406 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
4407 MVT::i1, MachineNode->getOperand(0),
4408 MachineNode->getOperand(1));
4413 if (Op1Set || Op2Set)
4415 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
4418 // nor(0, y) = ~y -> nor(y, y)
4419 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4420 MVT::i1, MachineNode->getOperand(1),
4421 MachineNode->getOperand(1));
4424 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4425 MVT::i1, MachineNode->getOperand(0),
4426 MachineNode->getOperand(0));
4428 // nor(~x, y) = andc(x, y)
4429 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
4430 MVT::i1, MachineNode->getOperand(0).
4432 MachineNode->getOperand(1));
4434 // nor(x, ~y) = andc(y, x)
4435 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
4436 MVT::i1, MachineNode->getOperand(1).
4438 MachineNode->getOperand(0));
4439 else if (AllUsersSelectZero(MachineNode)) {
4440 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode),
4441 MVT::i1, MachineNode->getOperand(0),
4442 MachineNode->getOperand(1));
4447 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
4449 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
4453 ResNode = MachineNode->getOperand(1).getNode();
4456 ResNode = MachineNode->getOperand(0).getNode();
4458 // eqv(0, y) = ~y -> nor(y, y)
4459 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4460 MVT::i1, MachineNode->getOperand(1),
4461 MachineNode->getOperand(1));
4464 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4465 MVT::i1, MachineNode->getOperand(0),
4466 MachineNode->getOperand(0));
4468 // eqv(~x, y) = xor(x, y)
4469 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
4470 MVT::i1, MachineNode->getOperand(0).
4472 MachineNode->getOperand(1));
4474 // eqv(x, ~y) = xor(x, y)
4475 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
4476 MVT::i1, MachineNode->getOperand(0),
4477 MachineNode->getOperand(1).
4479 else if (AllUsersSelectZero(MachineNode)) {
4480 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
4481 MVT::i1, MachineNode->getOperand(0),
4482 MachineNode->getOperand(1));
4487 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
4489 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
4493 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4494 MVT::i1, MachineNode->getOperand(1),
4495 MachineNode->getOperand(1));
4496 else if (Op1Unset || Op2Set)
4497 // andc(0, y) = andc(x, 1) = 0
4498 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
4502 ResNode = MachineNode->getOperand(0).getNode();
4504 // andc(~x, y) = ~(x | y) = nor(x, y)
4505 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4506 MVT::i1, MachineNode->getOperand(0).
4508 MachineNode->getOperand(1));
4510 // andc(x, ~y) = x & y
4511 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode),
4512 MVT::i1, MachineNode->getOperand(0),
4513 MachineNode->getOperand(1).
4515 else if (AllUsersSelectZero(MachineNode)) {
4516 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
4517 MVT::i1, MachineNode->getOperand(1),
4518 MachineNode->getOperand(0));
4523 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
4525 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
4527 else if (Op1Set || Op2Unset)
4528 // orc(1, y) = orc(x, 0) = 1
4529 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
4533 ResNode = MachineNode->getOperand(0).getNode();
4536 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
4537 MVT::i1, MachineNode->getOperand(1),
4538 MachineNode->getOperand(1));
4540 // orc(~x, y) = ~(x & y) = nand(x, y)
4541 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode),
4542 MVT::i1, MachineNode->getOperand(0).
4544 MachineNode->getOperand(1));
4546 // orc(x, ~y) = x | y
4547 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode),
4548 MVT::i1, MachineNode->getOperand(0),
4549 MachineNode->getOperand(1).
4551 else if (AllUsersSelectZero(MachineNode)) {
4552 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
4553 MVT::i1, MachineNode->getOperand(1),
4554 MachineNode->getOperand(0));
4558 case PPC::SELECT_I4:
4559 case PPC::SELECT_I8:
4560 case PPC::SELECT_F4:
4561 case PPC::SELECT_F8:
4562 case PPC::SELECT_QFRC:
4563 case PPC::SELECT_QSRC:
4564 case PPC::SELECT_QBRC:
4565 case PPC::SELECT_VRRC:
4566 case PPC::SELECT_VSFRC:
4567 case PPC::SELECT_VSSRC:
4568 case PPC::SELECT_VSRC:
4570 ResNode = MachineNode->getOperand(1).getNode();
4572 ResNode = MachineNode->getOperand(2).getNode();
4574 ResNode = CurDAG->getMachineNode(MachineNode->getMachineOpcode(),
4576 MachineNode->getValueType(0),
4577 MachineNode->getOperand(0).
4579 MachineNode->getOperand(2),
4580 MachineNode->getOperand(1));
4585 ResNode = CurDAG->getMachineNode(Opcode == PPC::BC ? PPC::BCn :
4589 MachineNode->getOperand(0).
4591 MachineNode->getOperand(1),
4592 MachineNode->getOperand(2));
4593 // FIXME: Handle Op1Set, Op1Unset here too.
4597 // If we're inverting this node because it is used only by selects that
4598 // we'd like to swap, then swap the selects before the node replacement.
4600 SwapAllSelectUsers(MachineNode);
4602 if (ResNode != MachineNode) {
4603 DEBUG(dbgs() << "CR Peephole replacing:\nOld: ");
4604 DEBUG(MachineNode->dump(CurDAG));
4605 DEBUG(dbgs() << "\nNew: ");
4606 DEBUG(ResNode->dump(CurDAG));
4607 DEBUG(dbgs() << "\n");
4609 ReplaceUses(MachineNode, ResNode);
4614 CurDAG->RemoveDeadNodes();
4615 } while (IsModified);
4618 // Gather the set of 32-bit operations that are known to have their
4619 // higher-order 32 bits zero, where ToPromote contains all such operations.
4620 static bool PeepholePPC64ZExtGather(SDValue Op32,
4621 SmallPtrSetImpl<SDNode *> &ToPromote) {
4622 if (!Op32.isMachineOpcode())
4625 // First, check for the "frontier" instructions (those that will clear the
4626 // higher-order 32 bits.
4628 // For RLWINM and RLWNM, we need to make sure that the mask does not wrap
4629 // around. If it does not, then these instructions will clear the
4630 // higher-order bits.
4631 if ((Op32.getMachineOpcode() == PPC::RLWINM ||
4632 Op32.getMachineOpcode() == PPC::RLWNM) &&
4633 Op32.getConstantOperandVal(2) <= Op32.getConstantOperandVal(3)) {
4634 ToPromote.insert(Op32.getNode());
4638 // SLW and SRW always clear the higher-order bits.
4639 if (Op32.getMachineOpcode() == PPC::SLW ||
4640 Op32.getMachineOpcode() == PPC::SRW) {
4641 ToPromote.insert(Op32.getNode());
4645 // For LI and LIS, we need the immediate to be positive (so that it is not
4647 if (Op32.getMachineOpcode() == PPC::LI ||
4648 Op32.getMachineOpcode() == PPC::LIS) {
4649 if (!isUInt<15>(Op32.getConstantOperandVal(0)))
4652 ToPromote.insert(Op32.getNode());
4656 // LHBRX and LWBRX always clear the higher-order bits.
4657 if (Op32.getMachineOpcode() == PPC::LHBRX ||
4658 Op32.getMachineOpcode() == PPC::LWBRX) {
4659 ToPromote.insert(Op32.getNode());
4663 // CNT[LT]ZW always produce a 64-bit value in [0,32], and so is zero extended.
4664 if (Op32.getMachineOpcode() == PPC::CNTLZW ||
4665 Op32.getMachineOpcode() == PPC::CNTTZW) {
4666 ToPromote.insert(Op32.getNode());
4670 // Next, check for those instructions we can look through.
4672 // Assuming the mask does not wrap around, then the higher-order bits are
4673 // taken directly from the first operand.
4674 if (Op32.getMachineOpcode() == PPC::RLWIMI &&
4675 Op32.getConstantOperandVal(3) <= Op32.getConstantOperandVal(4)) {
4676 SmallPtrSet<SDNode *, 16> ToPromote1;
4677 if (!PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1))
4680 ToPromote.insert(Op32.getNode());
4681 ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
4685 // For OR, the higher-order bits are zero if that is true for both operands.
4686 // For SELECT_I4, the same is true (but the relevant operand numbers are
4688 if (Op32.getMachineOpcode() == PPC::OR ||
4689 Op32.getMachineOpcode() == PPC::SELECT_I4) {
4690 unsigned B = Op32.getMachineOpcode() == PPC::SELECT_I4 ? 1 : 0;
4691 SmallPtrSet<SDNode *, 16> ToPromote1;
4692 if (!PeepholePPC64ZExtGather(Op32.getOperand(B+0), ToPromote1))
4694 if (!PeepholePPC64ZExtGather(Op32.getOperand(B+1), ToPromote1))
4697 ToPromote.insert(Op32.getNode());
4698 ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
4702 // For ORI and ORIS, we need the higher-order bits of the first operand to be
4703 // zero, and also for the constant to be positive (so that it is not sign
4705 if (Op32.getMachineOpcode() == PPC::ORI ||
4706 Op32.getMachineOpcode() == PPC::ORIS) {
4707 SmallPtrSet<SDNode *, 16> ToPromote1;
4708 if (!PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1))
4710 if (!isUInt<15>(Op32.getConstantOperandVal(1)))
4713 ToPromote.insert(Op32.getNode());
4714 ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
4718 // The higher-order bits of AND are zero if that is true for at least one of
4720 if (Op32.getMachineOpcode() == PPC::AND) {
4721 SmallPtrSet<SDNode *, 16> ToPromote1, ToPromote2;
4723 PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1);
4725 PeepholePPC64ZExtGather(Op32.getOperand(1), ToPromote2);
4726 if (!Op0OK && !Op1OK)
4729 ToPromote.insert(Op32.getNode());
4732 ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
4735 ToPromote.insert(ToPromote2.begin(), ToPromote2.end());
4740 // For ANDI and ANDIS, the higher-order bits are zero if either that is true
4741 // of the first operand, or if the second operand is positive (so that it is
4742 // not sign extended).
4743 if (Op32.getMachineOpcode() == PPC::ANDIo ||
4744 Op32.getMachineOpcode() == PPC::ANDISo) {
4745 SmallPtrSet<SDNode *, 16> ToPromote1;
4747 PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1);
4748 bool Op1OK = isUInt<15>(Op32.getConstantOperandVal(1));
4749 if (!Op0OK && !Op1OK)
4752 ToPromote.insert(Op32.getNode());
4755 ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
4763 void PPCDAGToDAGISel::PeepholePPC64ZExt() {
4764 if (!PPCSubTarget->isPPC64())
4767 // When we zero-extend from i32 to i64, we use a pattern like this:
4768 // def : Pat<(i64 (zext i32:$in)),
4769 // (RLDICL (INSERT_SUBREG (i64 (IMPLICIT_DEF)), $in, sub_32),
4771 // There are several 32-bit shift/rotate instructions, however, that will
4772 // clear the higher-order bits of their output, rendering the RLDICL
4773 // unnecessary. When that happens, we remove it here, and redefine the
4774 // relevant 32-bit operation to be a 64-bit operation.
4776 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
4779 bool MadeChange = false;
4780 while (Position != CurDAG->allnodes_begin()) {
4781 SDNode *N = &*--Position;
4782 // Skip dead nodes and any non-machine opcodes.
4783 if (N->use_empty() || !N->isMachineOpcode())
4786 if (N->getMachineOpcode() != PPC::RLDICL)
4789 if (N->getConstantOperandVal(1) != 0 ||
4790 N->getConstantOperandVal(2) != 32)
4793 SDValue ISR = N->getOperand(0);
4794 if (!ISR.isMachineOpcode() ||
4795 ISR.getMachineOpcode() != TargetOpcode::INSERT_SUBREG)
4798 if (!ISR.hasOneUse())
4801 if (ISR.getConstantOperandVal(2) != PPC::sub_32)
4804 SDValue IDef = ISR.getOperand(0);
4805 if (!IDef.isMachineOpcode() ||
4806 IDef.getMachineOpcode() != TargetOpcode::IMPLICIT_DEF)
4809 // We now know that we're looking at a canonical i32 -> i64 zext. See if we
4810 // can get rid of it.
4812 SDValue Op32 = ISR->getOperand(1);
4813 if (!Op32.isMachineOpcode())
4816 // There are some 32-bit instructions that always clear the high-order 32
4817 // bits, there are also some instructions (like AND) that we can look
4819 SmallPtrSet<SDNode *, 16> ToPromote;
4820 if (!PeepholePPC64ZExtGather(Op32, ToPromote))
4823 // If the ToPromote set contains nodes that have uses outside of the set
4824 // (except for the original INSERT_SUBREG), then abort the transformation.
4825 bool OutsideUse = false;
4826 for (SDNode *PN : ToPromote) {
4827 for (SDNode *UN : PN->uses()) {
4828 if (!ToPromote.count(UN) && UN != ISR.getNode()) {
4842 // We now know that this zero extension can be removed by promoting to
4843 // nodes in ToPromote to 64-bit operations, where for operations in the
4844 // frontier of the set, we need to insert INSERT_SUBREGs for their
4846 for (SDNode *PN : ToPromote) {
4848 switch (PN->getMachineOpcode()) {
4850 llvm_unreachable("Don't know the 64-bit variant of this instruction");
4851 case PPC::RLWINM: NewOpcode = PPC::RLWINM8; break;
4852 case PPC::RLWNM: NewOpcode = PPC::RLWNM8; break;
4853 case PPC::SLW: NewOpcode = PPC::SLW8; break;
4854 case PPC::SRW: NewOpcode = PPC::SRW8; break;
4855 case PPC::LI: NewOpcode = PPC::LI8; break;
4856 case PPC::LIS: NewOpcode = PPC::LIS8; break;
4857 case PPC::LHBRX: NewOpcode = PPC::LHBRX8; break;
4858 case PPC::LWBRX: NewOpcode = PPC::LWBRX8; break;
4859 case PPC::CNTLZW: NewOpcode = PPC::CNTLZW8; break;
4860 case PPC::CNTTZW: NewOpcode = PPC::CNTTZW8; break;
4861 case PPC::RLWIMI: NewOpcode = PPC::RLWIMI8; break;
4862 case PPC::OR: NewOpcode = PPC::OR8; break;
4863 case PPC::SELECT_I4: NewOpcode = PPC::SELECT_I8; break;
4864 case PPC::ORI: NewOpcode = PPC::ORI8; break;
4865 case PPC::ORIS: NewOpcode = PPC::ORIS8; break;
4866 case PPC::AND: NewOpcode = PPC::AND8; break;
4867 case PPC::ANDIo: NewOpcode = PPC::ANDIo8; break;
4868 case PPC::ANDISo: NewOpcode = PPC::ANDISo8; break;
4871 // Note: During the replacement process, the nodes will be in an
4872 // inconsistent state (some instructions will have operands with values
4873 // of the wrong type). Once done, however, everything should be right
4876 SmallVector<SDValue, 4> Ops;
4877 for (const SDValue &V : PN->ops()) {
4878 if (!ToPromote.count(V.getNode()) && V.getValueType() == MVT::i32 &&
4879 !isa<ConstantSDNode>(V)) {
4880 SDValue ReplOpOps[] = { ISR.getOperand(0), V, ISR.getOperand(2) };
4882 CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, SDLoc(V),
4883 ISR.getNode()->getVTList(), ReplOpOps);
4884 Ops.push_back(SDValue(ReplOp, 0));
4890 // Because all to-be-promoted nodes only have users that are other
4891 // promoted nodes (or the original INSERT_SUBREG), we can safely replace
4892 // the i32 result value type with i64.
4894 SmallVector<EVT, 2> NewVTs;
4895 SDVTList VTs = PN->getVTList();
4896 for (unsigned i = 0, ie = VTs.NumVTs; i != ie; ++i)
4897 if (VTs.VTs[i] == MVT::i32)
4898 NewVTs.push_back(MVT::i64);
4900 NewVTs.push_back(VTs.VTs[i]);
4902 DEBUG(dbgs() << "PPC64 ZExt Peephole morphing:\nOld: ");
4903 DEBUG(PN->dump(CurDAG));
4905 CurDAG->SelectNodeTo(PN, NewOpcode, CurDAG->getVTList(NewVTs), Ops);
4907 DEBUG(dbgs() << "\nNew: ");
4908 DEBUG(PN->dump(CurDAG));
4909 DEBUG(dbgs() << "\n");
4912 // Now we replace the original zero extend and its associated INSERT_SUBREG
4913 // with the value feeding the INSERT_SUBREG (which has now been promoted to
4916 DEBUG(dbgs() << "PPC64 ZExt Peephole replacing:\nOld: ");
4917 DEBUG(N->dump(CurDAG));
4918 DEBUG(dbgs() << "\nNew: ");
4919 DEBUG(Op32.getNode()->dump(CurDAG));
4920 DEBUG(dbgs() << "\n");
4922 ReplaceUses(N, Op32.getNode());
4926 CurDAG->RemoveDeadNodes();
4929 void PPCDAGToDAGISel::PeepholePPC64() {
4930 // These optimizations are currently supported only for 64-bit SVR4.
4931 if (PPCSubTarget->isDarwin() || !PPCSubTarget->isPPC64())
4934 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
4937 while (Position != CurDAG->allnodes_begin()) {
4938 SDNode *N = &*--Position;
4939 // Skip dead nodes and any non-machine opcodes.
4940 if (N->use_empty() || !N->isMachineOpcode())
4944 unsigned StorageOpcode = N->getMachineOpcode();
4946 switch (StorageOpcode) {
4977 // If this is a load or store with a zero offset, or within the alignment,
4978 // we may be able to fold an add-immediate into the memory operation.
4979 // The check against alignment is below, as it can't occur until we check
4980 // the arguments to N
4981 if (!isa<ConstantSDNode>(N->getOperand(FirstOp)))
4984 SDValue Base = N->getOperand(FirstOp + 1);
4985 if (!Base.isMachineOpcode())
4989 bool ReplaceFlags = true;
4991 // When the feeding operation is an add-immediate of some sort,
4992 // determine whether we need to add relocation information to the
4993 // target flags on the immediate operand when we fold it into the
4994 // load instruction.
4996 // For something like ADDItocL, the relocation information is
4997 // inferred from the opcode; when we process it in the AsmPrinter,
4998 // we add the necessary relocation there. A load, though, can receive
4999 // relocation from various flavors of ADDIxxx, so we need to carry
5000 // the relocation information in the target flags.
5001 switch (Base.getMachineOpcode()) {
5006 // In some cases (such as TLS) the relocation information
5007 // is already in place on the operand, so copying the operand
5009 ReplaceFlags = false;
5010 // For these cases, the immediate may not be divisible by 4, in
5011 // which case the fold is illegal for DS-form instructions. (The
5012 // other cases provide aligned addresses and are always safe.)
5013 if ((StorageOpcode == PPC::LWA ||
5014 StorageOpcode == PPC::LD ||
5015 StorageOpcode == PPC::STD) &&
5016 (!isa<ConstantSDNode>(Base.getOperand(1)) ||
5017 Base.getConstantOperandVal(1) % 4 != 0))
5020 case PPC::ADDIdtprelL:
5021 Flags = PPCII::MO_DTPREL_LO;
5023 case PPC::ADDItlsldL:
5024 Flags = PPCII::MO_TLSLD_LO;
5027 Flags = PPCII::MO_TOC_LO;
5031 SDValue ImmOpnd = Base.getOperand(1);
5033 // On PPC64, the TOC base pointer is guaranteed by the ABI only to have
5034 // 8-byte alignment, and so we can only use offsets less than 8 (otherwise,
5035 // we might have needed different @ha relocation values for the offset
5037 int MaxDisplacement = 7;
5038 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) {
5039 const GlobalValue *GV = GA->getGlobal();
5040 MaxDisplacement = std::min((int) GV->getAlignment() - 1, MaxDisplacement);
5043 bool UpdateHBase = false;
5044 SDValue HBase = Base.getOperand(0);
5046 int Offset = N->getConstantOperandVal(FirstOp);
5048 if (Offset < 0 || Offset > MaxDisplacement) {
5049 // If we have a addi(toc@l)/addis(toc@ha) pair, and the addis has only
5050 // one use, then we can do this for any offset, we just need to also
5051 // update the offset (i.e. the symbol addend) on the addis also.
5052 if (Base.getMachineOpcode() != PPC::ADDItocL)
5055 if (!HBase.isMachineOpcode() ||
5056 HBase.getMachineOpcode() != PPC::ADDIStocHA)
5059 if (!Base.hasOneUse() || !HBase.hasOneUse())
5062 SDValue HImmOpnd = HBase.getOperand(1);
5063 if (HImmOpnd != ImmOpnd)
5069 // If we're directly folding the addend from an addi instruction, then:
5070 // 1. In general, the offset on the memory access must be zero.
5071 // 2. If the addend is a constant, then it can be combined with a
5072 // non-zero offset, but only if the result meets the encoding
5074 if (auto *C = dyn_cast<ConstantSDNode>(ImmOpnd)) {
5075 Offset += C->getSExtValue();
5077 if ((StorageOpcode == PPC::LWA || StorageOpcode == PPC::LD ||
5078 StorageOpcode == PPC::STD) && (Offset % 4) != 0)
5081 if (!isInt<16>(Offset))
5084 ImmOpnd = CurDAG->getTargetConstant(Offset, SDLoc(ImmOpnd),
5085 ImmOpnd.getValueType());
5086 } else if (Offset != 0) {
5091 // We found an opportunity. Reverse the operands from the add
5092 // immediate and substitute them into the load or store. If
5093 // needed, update the target flags for the immediate operand to
5094 // reflect the necessary relocation information.
5095 DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase: ");
5096 DEBUG(Base->dump(CurDAG));
5097 DEBUG(dbgs() << "\nN: ");
5098 DEBUG(N->dump(CurDAG));
5099 DEBUG(dbgs() << "\n");
5101 // If the relocation information isn't already present on the
5102 // immediate operand, add it now.
5104 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) {
5106 const GlobalValue *GV = GA->getGlobal();
5107 // We can't perform this optimization for data whose alignment
5108 // is insufficient for the instruction encoding.
5109 if (GV->getAlignment() < 4 &&
5110 (StorageOpcode == PPC::LD || StorageOpcode == PPC::STD ||
5111 StorageOpcode == PPC::LWA || (Offset % 4) != 0)) {
5112 DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n");
5115 ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, Offset, Flags);
5116 } else if (ConstantPoolSDNode *CP =
5117 dyn_cast<ConstantPoolSDNode>(ImmOpnd)) {
5118 const Constant *C = CP->getConstVal();
5119 ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64,
5125 if (FirstOp == 1) // Store
5126 (void)CurDAG->UpdateNodeOperands(N, N->getOperand(0), ImmOpnd,
5127 Base.getOperand(0), N->getOperand(3));
5129 (void)CurDAG->UpdateNodeOperands(N, ImmOpnd, Base.getOperand(0),
5133 (void)CurDAG->UpdateNodeOperands(HBase.getNode(), HBase.getOperand(0),
5136 // The add-immediate may now be dead, in which case remove it.
5137 if (Base.getNode()->use_empty())
5138 CurDAG->RemoveDeadNode(Base.getNode());
5142 /// createPPCISelDag - This pass converts a legalized DAG into a
5143 /// PowerPC-specific DAG, ready for instruction scheduling.
5145 FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM) {
5146 return new PPCDAGToDAGISel(TM);