1 //===-- PPCISelDAGToDAG.cpp - PPC --pattern matching inst selector --------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines a pattern matching instruction selector for PowerPC,
11 // converting from a legalized dag to a PPC dag.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "PPC.h"
16 #include "MCTargetDesc/PPCPredicates.h"
17 #include "PPCMachineFunctionInfo.h"
18 #include "PPCTargetMachine.h"
19 #include "llvm/Analysis/BranchProbabilityInfo.h"
20 #include "llvm/CodeGen/FunctionLoweringInfo.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/SelectionDAG.h"
25 #include "llvm/CodeGen/SelectionDAGISel.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GlobalAlias.h"
29 #include "llvm/IR/GlobalValue.h"
30 #include "llvm/IR/GlobalVariable.h"
31 #include "llvm/IR/Intrinsics.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include "llvm/Target/TargetOptions.h"
39 using namespace llvm;
40 
41 #define DEBUG_TYPE "ppc-codegen"
42 
43 // FIXME: Remove this once the bug has been fixed!
44 cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug",
45 cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden);
46 
47 static cl::opt<bool>
48     UseBitPermRewriter("ppc-use-bit-perm-rewriter", cl::init(true),
49                        cl::desc("use aggressive ppc isel for bit permutations"),
50                        cl::Hidden);
51 static cl::opt<bool> BPermRewriterNoMasking(
52     "ppc-bit-perm-rewriter-stress-rotates",
53     cl::desc("stress rotate selection in aggressive ppc isel for "
54              "bit permutations"),
55     cl::Hidden);
56 
57 static cl::opt<bool> EnableBranchHint(
58   "ppc-use-branch-hint", cl::init(true),
59     cl::desc("Enable static hinting of branches on ppc"),
60     cl::Hidden);
61 
62 namespace llvm {
63   void initializePPCDAGToDAGISelPass(PassRegistry&);
64 }
65 
66 namespace {
67   //===--------------------------------------------------------------------===//
68   /// PPCDAGToDAGISel - PPC specific code to select PPC machine
69   /// instructions for SelectionDAG operations.
70   ///
71   class PPCDAGToDAGISel : public SelectionDAGISel {
72     const PPCTargetMachine &TM;
73     const PPCSubtarget *PPCSubTarget;
74     const PPCTargetLowering *PPCLowering;
75     unsigned GlobalBaseReg;
76   public:
PPCDAGToDAGISel(PPCTargetMachine & tm)77     explicit PPCDAGToDAGISel(PPCTargetMachine &tm)
78         : SelectionDAGISel(tm), TM(tm) {
79       initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry());
80     }
81 
runOnMachineFunction(MachineFunction & MF)82     bool runOnMachineFunction(MachineFunction &MF) override {
83       // Make sure we re-emit a set of the global base reg if necessary
84       GlobalBaseReg = 0;
85       PPCSubTarget = &MF.getSubtarget<PPCSubtarget>();
86       PPCLowering = PPCSubTarget->getTargetLowering();
87       SelectionDAGISel::runOnMachineFunction(MF);
88 
89       if (!PPCSubTarget->isSVR4ABI())
90         InsertVRSaveCode(MF);
91 
92       return true;
93     }
94 
95     void PreprocessISelDAG() override;
96     void PostprocessISelDAG() override;
97 
98     /// getI32Imm - Return a target constant with the specified value, of type
99     /// i32.
getI32Imm(unsigned Imm,SDLoc dl)100     inline SDValue getI32Imm(unsigned Imm, SDLoc dl) {
101       return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
102     }
103 
104     /// getI64Imm - Return a target constant with the specified value, of type
105     /// i64.
getI64Imm(uint64_t Imm,SDLoc dl)106     inline SDValue getI64Imm(uint64_t Imm, SDLoc dl) {
107       return CurDAG->getTargetConstant(Imm, dl, MVT::i64);
108     }
109 
110     /// getSmallIPtrImm - Return a target constant of pointer type.
getSmallIPtrImm(unsigned Imm,SDLoc dl)111     inline SDValue getSmallIPtrImm(unsigned Imm, SDLoc dl) {
112       return CurDAG->getTargetConstant(
113           Imm, dl, PPCLowering->getPointerTy(CurDAG->getDataLayout()));
114     }
115 
116     /// isRotateAndMask - Returns true if Mask and Shift can be folded into a
117     /// rotate and mask opcode and mask operation.
118     static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask,
119                                 unsigned &SH, unsigned &MB, unsigned &ME);
120 
121     /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
122     /// base register.  Return the virtual register that holds this value.
123     SDNode *getGlobalBaseReg();
124 
125     SDNode *getFrameIndex(SDNode *SN, SDNode *N, unsigned Offset = 0);
126 
127     // Select - Convert the specified operand from a target-independent to a
128     // target-specific node if it hasn't already been changed.
129     SDNode *Select(SDNode *N) override;
130 
131     SDNode *SelectBitfieldInsert(SDNode *N);
132     SDNode *SelectBitPermutation(SDNode *N);
133 
134     /// SelectCC - Select a comparison of the specified values with the
135     /// specified condition code, returning the CR# of the expression.
136     SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, SDLoc dl);
137 
138     /// SelectAddrImm - Returns true if the address N can be represented by
139     /// a base register plus a signed 16-bit displacement [r+imm].
SelectAddrImm(SDValue N,SDValue & Disp,SDValue & Base)140     bool SelectAddrImm(SDValue N, SDValue &Disp,
141                        SDValue &Base) {
142       return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, false);
143     }
144 
145     /// SelectAddrImmOffs - Return true if the operand is valid for a preinc
146     /// immediate field.  Note that the operand at this point is already the
147     /// result of a prior SelectAddressRegImm call.
SelectAddrImmOffs(SDValue N,SDValue & Out) const148     bool SelectAddrImmOffs(SDValue N, SDValue &Out) const {
149       if (N.getOpcode() == ISD::TargetConstant ||
150           N.getOpcode() == ISD::TargetGlobalAddress) {
151         Out = N;
152         return true;
153       }
154 
155       return false;
156     }
157 
158     /// SelectAddrIdx - Given the specified addressed, check to see if it can be
159     /// represented as an indexed [r+r] operation.  Returns false if it can
160     /// be represented by [r+imm], which are preferred.
SelectAddrIdx(SDValue N,SDValue & Base,SDValue & Index)161     bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) {
162       return PPCLowering->SelectAddressRegReg(N, Base, Index, *CurDAG);
163     }
164 
165     /// SelectAddrIdxOnly - Given the specified addressed, force it to be
166     /// represented as an indexed [r+r] operation.
SelectAddrIdxOnly(SDValue N,SDValue & Base,SDValue & Index)167     bool SelectAddrIdxOnly(SDValue N, SDValue &Base, SDValue &Index) {
168       return PPCLowering->SelectAddressRegRegOnly(N, Base, Index, *CurDAG);
169     }
170 
171     /// SelectAddrImmX4 - Returns true if the address N can be represented by
172     /// a base register plus a signed 16-bit displacement that is a multiple of 4.
173     /// Suitable for use by STD and friends.
SelectAddrImmX4(SDValue N,SDValue & Disp,SDValue & Base)174     bool SelectAddrImmX4(SDValue N, SDValue &Disp, SDValue &Base) {
175       return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, true);
176     }
177 
178     // Select an address into a single register.
SelectAddr(SDValue N,SDValue & Base)179     bool SelectAddr(SDValue N, SDValue &Base) {
180       Base = N;
181       return true;
182     }
183 
184     /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
185     /// inline asm expressions.  It is always correct to compute the value into
186     /// a register.  The case of adding a (possibly relocatable) constant to a
187     /// register can be improved, but it is wrong to substitute Reg+Reg for
188     /// Reg in an asm, because the load or store opcode would have to change.
SelectInlineAsmMemoryOperand(const SDValue & Op,unsigned ConstraintID,std::vector<SDValue> & OutOps)189     bool SelectInlineAsmMemoryOperand(const SDValue &Op,
190                                       unsigned ConstraintID,
191                                       std::vector<SDValue> &OutOps) override {
192 
193       switch(ConstraintID) {
194       default:
195         errs() << "ConstraintID: " << ConstraintID << "\n";
196         llvm_unreachable("Unexpected asm memory constraint");
197       case InlineAsm::Constraint_es:
198       case InlineAsm::Constraint_i:
199       case InlineAsm::Constraint_m:
200       case InlineAsm::Constraint_o:
201       case InlineAsm::Constraint_Q:
202       case InlineAsm::Constraint_Z:
203       case InlineAsm::Constraint_Zy:
204         // We need to make sure that this one operand does not end up in r0
205         // (because we might end up lowering this as 0(%op)).
206         const TargetRegisterInfo *TRI = PPCSubTarget->getRegisterInfo();
207         const TargetRegisterClass *TRC = TRI->getPointerRegClass(*MF, /*Kind=*/1);
208         SDLoc dl(Op);
209         SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32);
210         SDValue NewOp =
211           SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
212                                          dl, Op.getValueType(),
213                                          Op, RC), 0);
214 
215         OutOps.push_back(NewOp);
216         return false;
217       }
218       return true;
219     }
220 
221     void InsertVRSaveCode(MachineFunction &MF);
222 
getPassName() const223     const char *getPassName() const override {
224       return "PowerPC DAG->DAG Pattern Instruction Selection";
225     }
226 
227 // Include the pieces autogenerated from the target description.
228 #include "PPCGenDAGISel.inc"
229 
230 private:
231     SDNode *SelectSETCC(SDNode *N);
232 
233     void PeepholePPC64();
234     void PeepholePPC64ZExt();
235     void PeepholeCROps();
236 
237     SDValue combineToCMPB(SDNode *N);
238     void foldBoolExts(SDValue &Res, SDNode *&N);
239 
240     bool AllUsersSelectZero(SDNode *N);
241     void SwapAllSelectUsers(SDNode *N);
242 
243     SDNode *transferMemOperands(SDNode *N, SDNode *Result);
244   };
245 }
246 
247 /// InsertVRSaveCode - Once the entire function has been instruction selected,
248 /// all virtual registers are created and all machine instructions are built,
249 /// check to see if we need to save/restore VRSAVE.  If so, do it.
InsertVRSaveCode(MachineFunction & Fn)250 void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) {
251   // Check to see if this function uses vector registers, which means we have to
252   // save and restore the VRSAVE register and update it with the regs we use.
253   //
254   // In this case, there will be virtual registers of vector type created
255   // by the scheduler.  Detect them now.
256   bool HasVectorVReg = false;
257   for (unsigned i = 0, e = RegInfo->getNumVirtRegs(); i != e; ++i) {
258     unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
259     if (RegInfo->getRegClass(Reg) == &PPC::VRRCRegClass) {
260       HasVectorVReg = true;
261       break;
262     }
263   }
264   if (!HasVectorVReg) return;  // nothing to do.
265 
266   // If we have a vector register, we want to emit code into the entry and exit
267   // blocks to save and restore the VRSAVE register.  We do this here (instead
268   // of marking all vector instructions as clobbering VRSAVE) for two reasons:
269   //
270   // 1. This (trivially) reduces the load on the register allocator, by not
271   //    having to represent the live range of the VRSAVE register.
272   // 2. This (more significantly) allows us to create a temporary virtual
273   //    register to hold the saved VRSAVE value, allowing this temporary to be
274   //    register allocated, instead of forcing it to be spilled to the stack.
275 
276   // Create two vregs - one to hold the VRSAVE register that is live-in to the
277   // function and one for the value after having bits or'd into it.
278   unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
279   unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
280 
281   const TargetInstrInfo &TII = *PPCSubTarget->getInstrInfo();
282   MachineBasicBlock &EntryBB = *Fn.begin();
283   DebugLoc dl;
284   // Emit the following code into the entry block:
285   // InVRSAVE = MFVRSAVE
286   // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE
287   // MTVRSAVE UpdatedVRSAVE
288   MachineBasicBlock::iterator IP = EntryBB.begin();  // Insert Point
289   BuildMI(EntryBB, IP, dl, TII.get(PPC::MFVRSAVE), InVRSAVE);
290   BuildMI(EntryBB, IP, dl, TII.get(PPC::UPDATE_VRSAVE),
291           UpdatedVRSAVE).addReg(InVRSAVE);
292   BuildMI(EntryBB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(UpdatedVRSAVE);
293 
294   // Find all return blocks, outputting a restore in each epilog.
295   for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
296     if (BB->isReturnBlock()) {
297       IP = BB->end(); --IP;
298 
299       // Skip over all terminator instructions, which are part of the return
300       // sequence.
301       MachineBasicBlock::iterator I2 = IP;
302       while (I2 != BB->begin() && (--I2)->isTerminator())
303         IP = I2;
304 
305       // Emit: MTVRSAVE InVRSave
306       BuildMI(*BB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(InVRSAVE);
307     }
308   }
309 }
310 
311 
312 /// getGlobalBaseReg - Output the instructions required to put the
313 /// base address to use for accessing globals into a register.
314 ///
getGlobalBaseReg()315 SDNode *PPCDAGToDAGISel::getGlobalBaseReg() {
316   if (!GlobalBaseReg) {
317     const TargetInstrInfo &TII = *PPCSubTarget->getInstrInfo();
318     // Insert the set of GlobalBaseReg into the first MBB of the function
319     MachineBasicBlock &FirstMBB = MF->front();
320     MachineBasicBlock::iterator MBBI = FirstMBB.begin();
321     const Module *M = MF->getFunction()->getParent();
322     DebugLoc dl;
323 
324     if (PPCLowering->getPointerTy(CurDAG->getDataLayout()) == MVT::i32) {
325       if (PPCSubTarget->isTargetELF()) {
326         GlobalBaseReg = PPC::R30;
327         if (M->getPICLevel() == PICLevel::Small) {
328           BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MoveGOTtoLR));
329           BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
330           MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true);
331         } else {
332           BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
333           BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
334           unsigned TempReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
335           BuildMI(FirstMBB, MBBI, dl,
336                   TII.get(PPC::UpdateGBR), GlobalBaseReg)
337                   .addReg(TempReg, RegState::Define).addReg(GlobalBaseReg);
338           MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true);
339         }
340       } else {
341         GlobalBaseReg =
342           RegInfo->createVirtualRegister(&PPC::GPRC_NOR0RegClass);
343         BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
344         BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
345       }
346     } else {
347       GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RC_NOX0RegClass);
348       BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8));
349       BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg);
350     }
351   }
352   return CurDAG->getRegister(GlobalBaseReg,
353                              PPCLowering->getPointerTy(CurDAG->getDataLayout()))
354       .getNode();
355 }
356 
357 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
358 /// or 64-bit immediate, and if the value can be accurately represented as a
359 /// sign extension from a 16-bit value.  If so, this returns true and the
360 /// immediate.
isIntS16Immediate(SDNode * N,short & Imm)361 static bool isIntS16Immediate(SDNode *N, short &Imm) {
362   if (N->getOpcode() != ISD::Constant)
363     return false;
364 
365   Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
366   if (N->getValueType(0) == MVT::i32)
367     return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
368   else
369     return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
370 }
371 
isIntS16Immediate(SDValue Op,short & Imm)372 static bool isIntS16Immediate(SDValue Op, short &Imm) {
373   return isIntS16Immediate(Op.getNode(), Imm);
374 }
375 
376 
377 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
378 /// operand. If so Imm will receive the 32-bit value.
isInt32Immediate(SDNode * N,unsigned & Imm)379 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
380   if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
381     Imm = cast<ConstantSDNode>(N)->getZExtValue();
382     return true;
383   }
384   return false;
385 }
386 
387 /// isInt64Immediate - This method tests to see if the node is a 64-bit constant
388 /// operand.  If so Imm will receive the 64-bit value.
isInt64Immediate(SDNode * N,uint64_t & Imm)389 static bool isInt64Immediate(SDNode *N, uint64_t &Imm) {
390   if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i64) {
391     Imm = cast<ConstantSDNode>(N)->getZExtValue();
392     return true;
393   }
394   return false;
395 }
396 
397 // isInt32Immediate - This method tests to see if a constant operand.
398 // If so Imm will receive the 32 bit value.
isInt32Immediate(SDValue N,unsigned & Imm)399 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
400   return isInt32Immediate(N.getNode(), Imm);
401 }
402 
getBranchHint(unsigned PCC,FunctionLoweringInfo * FuncInfo,const SDValue & DestMBB)403 static unsigned getBranchHint(unsigned PCC, FunctionLoweringInfo *FuncInfo,
404                               const SDValue &DestMBB) {
405   assert(isa<BasicBlockSDNode>(DestMBB));
406 
407   if (!FuncInfo->BPI) return PPC::BR_NO_HINT;
408 
409   const BasicBlock *BB = FuncInfo->MBB->getBasicBlock();
410   const TerminatorInst *BBTerm = BB->getTerminator();
411 
412   if (BBTerm->getNumSuccessors() != 2) return PPC::BR_NO_HINT;
413 
414   const BasicBlock *TBB = BBTerm->getSuccessor(0);
415   const BasicBlock *FBB = BBTerm->getSuccessor(1);
416 
417   uint32_t TWeight = FuncInfo->BPI->getEdgeWeight(BB, TBB);
418   uint32_t FWeight = FuncInfo->BPI->getEdgeWeight(BB, FBB);
419 
420   // We only want to handle cases which are easy to predict at static time, e.g.
421   // C++ throw statement, that is very likely not taken, or calling never
422   // returned function, e.g. stdlib exit(). So we set Threshold to filter
423   // unwanted cases.
424   //
425   // Below is LLVM branch weight table, we only want to handle case 1, 2
426   //
427   // Case                  Taken:Nontaken  Example
428   // 1. Unreachable        1048575:1       C++ throw, stdlib exit(),
429   // 2. Invoke-terminating 1:1048575
430   // 3. Coldblock          4:64            __builtin_expect
431   // 4. Loop Branch        124:4           For loop
432   // 5. PH/ZH/FPH          20:12
433   const uint32_t Threshold = 10000;
434 
435   // Minimal weight should be at least 1
436   if (std::max(TWeight, FWeight) /
437       std::max(1u, std::min(TWeight, FWeight)) < Threshold)
438     return PPC::BR_NO_HINT;
439 
440   DEBUG(dbgs() << "Use branch hint for '" << FuncInfo->Fn->getName() << "::"
441                << BB->getName() << "'\n"
442                << " -> " << TBB->getName() << ": " << TWeight << "\n"
443                << " -> " << FBB->getName() << ": " << FWeight << "\n");
444 
445   const BasicBlockSDNode *BBDN = cast<BasicBlockSDNode>(DestMBB);
446 
447   // If Dest BasicBlock is False-BasicBlock (FBB), swap branch weight,
448   // because we want 'TWeight' stands for 'branch weight' to Dest BasicBlock
449   if (BBDN->getBasicBlock()->getBasicBlock() != TBB)
450     std::swap(TWeight, FWeight);
451 
452   return (TWeight > FWeight) ? PPC::BR_TAKEN_HINT : PPC::BR_NONTAKEN_HINT;
453 }
454 
455 // isOpcWithIntImmediate - This method tests to see if the node is a specific
456 // opcode and that it has a immediate integer right operand.
457 // If so Imm will receive the 32 bit value.
isOpcWithIntImmediate(SDNode * N,unsigned Opc,unsigned & Imm)458 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
459   return N->getOpcode() == Opc
460          && isInt32Immediate(N->getOperand(1).getNode(), Imm);
461 }
462 
getFrameIndex(SDNode * SN,SDNode * N,unsigned Offset)463 SDNode *PPCDAGToDAGISel::getFrameIndex(SDNode *SN, SDNode *N, unsigned Offset) {
464   SDLoc dl(SN);
465   int FI = cast<FrameIndexSDNode>(N)->getIndex();
466   SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0));
467   unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8;
468   if (SN->hasOneUse())
469     return CurDAG->SelectNodeTo(SN, Opc, N->getValueType(0), TFI,
470                                 getSmallIPtrImm(Offset, dl));
471   return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI,
472                                 getSmallIPtrImm(Offset, dl));
473 }
474 
isRotateAndMask(SDNode * N,unsigned Mask,bool isShiftMask,unsigned & SH,unsigned & MB,unsigned & ME)475 bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask,
476                                       bool isShiftMask, unsigned &SH,
477                                       unsigned &MB, unsigned &ME) {
478   // Don't even go down this path for i64, since different logic will be
479   // necessary for rldicl/rldicr/rldimi.
480   if (N->getValueType(0) != MVT::i32)
481     return false;
482 
483   unsigned Shift  = 32;
484   unsigned Indeterminant = ~0;  // bit mask marking indeterminant results
485   unsigned Opcode = N->getOpcode();
486   if (N->getNumOperands() != 2 ||
487       !isInt32Immediate(N->getOperand(1).getNode(), Shift) || (Shift > 31))
488     return false;
489 
490   if (Opcode == ISD::SHL) {
491     // apply shift left to mask if it comes first
492     if (isShiftMask) Mask = Mask << Shift;
493     // determine which bits are made indeterminant by shift
494     Indeterminant = ~(0xFFFFFFFFu << Shift);
495   } else if (Opcode == ISD::SRL) {
496     // apply shift right to mask if it comes first
497     if (isShiftMask) Mask = Mask >> Shift;
498     // determine which bits are made indeterminant by shift
499     Indeterminant = ~(0xFFFFFFFFu >> Shift);
500     // adjust for the left rotate
501     Shift = 32 - Shift;
502   } else if (Opcode == ISD::ROTL) {
503     Indeterminant = 0;
504   } else {
505     return false;
506   }
507 
508   // if the mask doesn't intersect any Indeterminant bits
509   if (Mask && !(Mask & Indeterminant)) {
510     SH = Shift & 31;
511     // make sure the mask is still a mask (wrap arounds may not be)
512     return isRunOfOnes(Mask, MB, ME);
513   }
514   return false;
515 }
516 
517 /// SelectBitfieldInsert - turn an or of two masked values into
518 /// the rotate left word immediate then mask insert (rlwimi) instruction.
SelectBitfieldInsert(SDNode * N)519 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
520   SDValue Op0 = N->getOperand(0);
521   SDValue Op1 = N->getOperand(1);
522   SDLoc dl(N);
523 
524   APInt LKZ, LKO, RKZ, RKO;
525   CurDAG->computeKnownBits(Op0, LKZ, LKO);
526   CurDAG->computeKnownBits(Op1, RKZ, RKO);
527 
528   unsigned TargetMask = LKZ.getZExtValue();
529   unsigned InsertMask = RKZ.getZExtValue();
530 
531   if ((TargetMask | InsertMask) == 0xFFFFFFFF) {
532     unsigned Op0Opc = Op0.getOpcode();
533     unsigned Op1Opc = Op1.getOpcode();
534     unsigned Value, SH = 0;
535     TargetMask = ~TargetMask;
536     InsertMask = ~InsertMask;
537 
538     // If the LHS has a foldable shift and the RHS does not, then swap it to the
539     // RHS so that we can fold the shift into the insert.
540     if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
541       if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
542           Op0.getOperand(0).getOpcode() == ISD::SRL) {
543         if (Op1.getOperand(0).getOpcode() != ISD::SHL &&
544             Op1.getOperand(0).getOpcode() != ISD::SRL) {
545           std::swap(Op0, Op1);
546           std::swap(Op0Opc, Op1Opc);
547           std::swap(TargetMask, InsertMask);
548         }
549       }
550     } else if (Op0Opc == ISD::SHL || Op0Opc == ISD::SRL) {
551       if (Op1Opc == ISD::AND && Op1.getOperand(0).getOpcode() != ISD::SHL &&
552           Op1.getOperand(0).getOpcode() != ISD::SRL) {
553         std::swap(Op0, Op1);
554         std::swap(Op0Opc, Op1Opc);
555         std::swap(TargetMask, InsertMask);
556       }
557     }
558 
559     unsigned MB, ME;
560     if (isRunOfOnes(InsertMask, MB, ME)) {
561       SDValue Tmp1, Tmp2;
562 
563       if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) &&
564           isInt32Immediate(Op1.getOperand(1), Value)) {
565         Op1 = Op1.getOperand(0);
566         SH  = (Op1Opc == ISD::SHL) ? Value : 32 - Value;
567       }
568       if (Op1Opc == ISD::AND) {
569        // The AND mask might not be a constant, and we need to make sure that
570        // if we're going to fold the masking with the insert, all bits not
571        // know to be zero in the mask are known to be one.
572         APInt MKZ, MKO;
573         CurDAG->computeKnownBits(Op1.getOperand(1), MKZ, MKO);
574         bool CanFoldMask = InsertMask == MKO.getZExtValue();
575 
576         unsigned SHOpc = Op1.getOperand(0).getOpcode();
577         if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) && CanFoldMask &&
578             isInt32Immediate(Op1.getOperand(0).getOperand(1), Value)) {
579           // Note that Value must be in range here (less than 32) because
580           // otherwise there would not be any bits set in InsertMask.
581           Op1 = Op1.getOperand(0).getOperand(0);
582           SH  = (SHOpc == ISD::SHL) ? Value : 32 - Value;
583         }
584       }
585 
586       SH &= 31;
587       SDValue Ops[] = { Op0, Op1, getI32Imm(SH, dl), getI32Imm(MB, dl),
588                           getI32Imm(ME, dl) };
589       return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
590     }
591   }
592   return nullptr;
593 }
594 
595 // Predict the number of instructions that would be generated by calling
596 // SelectInt64(N).
SelectInt64CountDirect(int64_t Imm)597 static unsigned SelectInt64CountDirect(int64_t Imm) {
598   // Assume no remaining bits.
599   unsigned Remainder = 0;
600   // Assume no shift required.
601   unsigned Shift = 0;
602 
603   // If it can't be represented as a 32 bit value.
604   if (!isInt<32>(Imm)) {
605     Shift = countTrailingZeros<uint64_t>(Imm);
606     int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
607 
608     // If the shifted value fits 32 bits.
609     if (isInt<32>(ImmSh)) {
610       // Go with the shifted value.
611       Imm = ImmSh;
612     } else {
613       // Still stuck with a 64 bit value.
614       Remainder = Imm;
615       Shift = 32;
616       Imm >>= 32;
617     }
618   }
619 
620   // Intermediate operand.
621   unsigned Result = 0;
622 
623   // Handle first 32 bits.
624   unsigned Lo = Imm & 0xFFFF;
625 
626   // Simple value.
627   if (isInt<16>(Imm)) {
628     // Just the Lo bits.
629     ++Result;
630   } else if (Lo) {
631     // Handle the Hi bits and Lo bits.
632     Result += 2;
633   } else {
634     // Just the Hi bits.
635     ++Result;
636   }
637 
638   // If no shift, we're done.
639   if (!Shift) return Result;
640 
641   // Shift for next step if the upper 32-bits were not zero.
642   if (Imm)
643     ++Result;
644 
645   // Add in the last bits as required.
646   if ((Remainder >> 16) & 0xFFFF)
647     ++Result;
648   if (Remainder & 0xFFFF)
649     ++Result;
650 
651   return Result;
652 }
653 
Rot64(uint64_t Imm,unsigned R)654 static uint64_t Rot64(uint64_t Imm, unsigned R) {
655   return (Imm << R) | (Imm >> (64 - R));
656 }
657 
SelectInt64Count(int64_t Imm)658 static unsigned SelectInt64Count(int64_t Imm) {
659   unsigned Count = SelectInt64CountDirect(Imm);
660   if (Count == 1)
661     return Count;
662 
663   for (unsigned r = 1; r < 63; ++r) {
664     uint64_t RImm = Rot64(Imm, r);
665     unsigned RCount = SelectInt64CountDirect(RImm) + 1;
666     Count = std::min(Count, RCount);
667 
668     // See comments in SelectInt64 for an explanation of the logic below.
669     unsigned LS = findLastSet(RImm);
670     if (LS != r-1)
671       continue;
672 
673     uint64_t OnesMask = -(int64_t) (UINT64_C(1) << (LS+1));
674     uint64_t RImmWithOnes = RImm | OnesMask;
675 
676     RCount = SelectInt64CountDirect(RImmWithOnes) + 1;
677     Count = std::min(Count, RCount);
678   }
679 
680   return Count;
681 }
682 
683 // Select a 64-bit constant. For cost-modeling purposes, SelectInt64Count
684 // (above) needs to be kept in sync with this function.
SelectInt64Direct(SelectionDAG * CurDAG,SDLoc dl,int64_t Imm)685 static SDNode *SelectInt64Direct(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) {
686   // Assume no remaining bits.
687   unsigned Remainder = 0;
688   // Assume no shift required.
689   unsigned Shift = 0;
690 
691   // If it can't be represented as a 32 bit value.
692   if (!isInt<32>(Imm)) {
693     Shift = countTrailingZeros<uint64_t>(Imm);
694     int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
695 
696     // If the shifted value fits 32 bits.
697     if (isInt<32>(ImmSh)) {
698       // Go with the shifted value.
699       Imm = ImmSh;
700     } else {
701       // Still stuck with a 64 bit value.
702       Remainder = Imm;
703       Shift = 32;
704       Imm >>= 32;
705     }
706   }
707 
708   // Intermediate operand.
709   SDNode *Result;
710 
711   // Handle first 32 bits.
712   unsigned Lo = Imm & 0xFFFF;
713   unsigned Hi = (Imm >> 16) & 0xFFFF;
714 
715   auto getI32Imm = [CurDAG, dl](unsigned Imm) {
716       return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
717   };
718 
719   // Simple value.
720   if (isInt<16>(Imm)) {
721     // Just the Lo bits.
722     Result = CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(Lo));
723   } else if (Lo) {
724     // Handle the Hi bits.
725     unsigned OpC = Hi ? PPC::LIS8 : PPC::LI8;
726     Result = CurDAG->getMachineNode(OpC, dl, MVT::i64, getI32Imm(Hi));
727     // And Lo bits.
728     Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
729                                     SDValue(Result, 0), getI32Imm(Lo));
730   } else {
731     // Just the Hi bits.
732     Result = CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(Hi));
733   }
734 
735   // If no shift, we're done.
736   if (!Shift) return Result;
737 
738   // Shift for next step if the upper 32-bits were not zero.
739   if (Imm) {
740     Result = CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64,
741                                     SDValue(Result, 0),
742                                     getI32Imm(Shift),
743                                     getI32Imm(63 - Shift));
744   }
745 
746   // Add in the last bits as required.
747   if ((Hi = (Remainder >> 16) & 0xFFFF)) {
748     Result = CurDAG->getMachineNode(PPC::ORIS8, dl, MVT::i64,
749                                     SDValue(Result, 0), getI32Imm(Hi));
750   }
751   if ((Lo = Remainder & 0xFFFF)) {
752     Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
753                                     SDValue(Result, 0), getI32Imm(Lo));
754   }
755 
756   return Result;
757 }
758 
SelectInt64(SelectionDAG * CurDAG,SDLoc dl,int64_t Imm)759 static SDNode *SelectInt64(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) {
760   unsigned Count = SelectInt64CountDirect(Imm);
761   if (Count == 1)
762     return SelectInt64Direct(CurDAG, dl, Imm);
763 
764   unsigned RMin = 0;
765 
766   int64_t MatImm;
767   unsigned MaskEnd;
768 
769   for (unsigned r = 1; r < 63; ++r) {
770     uint64_t RImm = Rot64(Imm, r);
771     unsigned RCount = SelectInt64CountDirect(RImm) + 1;
772     if (RCount < Count) {
773       Count = RCount;
774       RMin = r;
775       MatImm = RImm;
776       MaskEnd = 63;
777     }
778 
779     // If the immediate to generate has many trailing zeros, it might be
780     // worthwhile to generate a rotated value with too many leading ones
781     // (because that's free with li/lis's sign-extension semantics), and then
782     // mask them off after rotation.
783 
784     unsigned LS = findLastSet(RImm);
785     // We're adding (63-LS) higher-order ones, and we expect to mask them off
786     // after performing the inverse rotation by (64-r). So we need that:
787     //   63-LS == 64-r => LS == r-1
788     if (LS != r-1)
789       continue;
790 
791     uint64_t OnesMask = -(int64_t) (UINT64_C(1) << (LS+1));
792     uint64_t RImmWithOnes = RImm | OnesMask;
793 
794     RCount = SelectInt64CountDirect(RImmWithOnes) + 1;
795     if (RCount < Count) {
796       Count = RCount;
797       RMin = r;
798       MatImm = RImmWithOnes;
799       MaskEnd = LS;
800     }
801   }
802 
803   if (!RMin)
804     return SelectInt64Direct(CurDAG, dl, Imm);
805 
806   auto getI32Imm = [CurDAG, dl](unsigned Imm) {
807       return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
808   };
809 
810   SDValue Val = SDValue(SelectInt64Direct(CurDAG, dl, MatImm), 0);
811   return CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, Val,
812                                 getI32Imm(64 - RMin), getI32Imm(MaskEnd));
813 }
814 
815 // Select a 64-bit constant.
SelectInt64(SelectionDAG * CurDAG,SDNode * N)816 static SDNode *SelectInt64(SelectionDAG *CurDAG, SDNode *N) {
817   SDLoc dl(N);
818 
819   // Get 64 bit value.
820   int64_t Imm = cast<ConstantSDNode>(N)->getZExtValue();
821   return SelectInt64(CurDAG, dl, Imm);
822 }
823 
824 namespace {
825 class BitPermutationSelector {
826   struct ValueBit {
827     SDValue V;
828 
829     // The bit number in the value, using a convention where bit 0 is the
830     // lowest-order bit.
831     unsigned Idx;
832 
833     enum Kind {
834       ConstZero,
835       Variable
836     } K;
837 
ValueBit__anon7bd595e20411::BitPermutationSelector::ValueBit838     ValueBit(SDValue V, unsigned I, Kind K = Variable)
839       : V(V), Idx(I), K(K) {}
ValueBit__anon7bd595e20411::BitPermutationSelector::ValueBit840     ValueBit(Kind K = Variable)
841       : V(SDValue(nullptr, 0)), Idx(UINT32_MAX), K(K) {}
842 
isZero__anon7bd595e20411::BitPermutationSelector::ValueBit843     bool isZero() const {
844       return K == ConstZero;
845     }
846 
hasValue__anon7bd595e20411::BitPermutationSelector::ValueBit847     bool hasValue() const {
848       return K == Variable;
849     }
850 
getValue__anon7bd595e20411::BitPermutationSelector::ValueBit851     SDValue getValue() const {
852       assert(hasValue() && "Cannot get the value of a constant bit");
853       return V;
854     }
855 
getValueBitIndex__anon7bd595e20411::BitPermutationSelector::ValueBit856     unsigned getValueBitIndex() const {
857       assert(hasValue() && "Cannot get the value bit index of a constant bit");
858       return Idx;
859     }
860   };
861 
862   // A bit group has the same underlying value and the same rotate factor.
863   struct BitGroup {
864     SDValue V;
865     unsigned RLAmt;
866     unsigned StartIdx, EndIdx;
867 
868     // This rotation amount assumes that the lower 32 bits of the quantity are
869     // replicated in the high 32 bits by the rotation operator (which is done
870     // by rlwinm and friends in 64-bit mode).
871     bool Repl32;
872     // Did converting to Repl32 == true change the rotation factor? If it did,
873     // it decreased it by 32.
874     bool Repl32CR;
875     // Was this group coalesced after setting Repl32 to true?
876     bool Repl32Coalesced;
877 
BitGroup__anon7bd595e20411::BitPermutationSelector::BitGroup878     BitGroup(SDValue V, unsigned R, unsigned S, unsigned E)
879       : V(V), RLAmt(R), StartIdx(S), EndIdx(E), Repl32(false), Repl32CR(false),
880         Repl32Coalesced(false) {
881       DEBUG(dbgs() << "\tbit group for " << V.getNode() << " RLAmt = " << R <<
882                       " [" << S << ", " << E << "]\n");
883     }
884   };
885 
886   // Information on each (Value, RLAmt) pair (like the number of groups
887   // associated with each) used to choose the lowering method.
888   struct ValueRotInfo {
889     SDValue V;
890     unsigned RLAmt;
891     unsigned NumGroups;
892     unsigned FirstGroupStartIdx;
893     bool Repl32;
894 
ValueRotInfo__anon7bd595e20411::BitPermutationSelector::ValueRotInfo895     ValueRotInfo()
896       : RLAmt(UINT32_MAX), NumGroups(0), FirstGroupStartIdx(UINT32_MAX),
897         Repl32(false) {}
898 
899     // For sorting (in reverse order) by NumGroups, and then by
900     // FirstGroupStartIdx.
operator <__anon7bd595e20411::BitPermutationSelector::ValueRotInfo901     bool operator < (const ValueRotInfo &Other) const {
902       // We need to sort so that the non-Repl32 come first because, when we're
903       // doing masking, the Repl32 bit groups might be subsumed into the 64-bit
904       // masking operation.
905       if (Repl32 < Other.Repl32)
906         return true;
907       else if (Repl32 > Other.Repl32)
908         return false;
909       else if (NumGroups > Other.NumGroups)
910         return true;
911       else if (NumGroups < Other.NumGroups)
912         return false;
913       else if (FirstGroupStartIdx < Other.FirstGroupStartIdx)
914         return true;
915       return false;
916     }
917   };
918 
919   // Return true if something interesting was deduced, return false if we're
920   // providing only a generic representation of V (or something else likewise
921   // uninteresting for instruction selection).
getValueBits(SDValue V,SmallVector<ValueBit,64> & Bits)922   bool getValueBits(SDValue V, SmallVector<ValueBit, 64> &Bits) {
923     switch (V.getOpcode()) {
924     default: break;
925     case ISD::ROTL:
926       if (isa<ConstantSDNode>(V.getOperand(1))) {
927         unsigned RotAmt = V.getConstantOperandVal(1);
928 
929         SmallVector<ValueBit, 64> LHSBits(Bits.size());
930         getValueBits(V.getOperand(0), LHSBits);
931 
932         for (unsigned i = 0; i < Bits.size(); ++i)
933           Bits[i] = LHSBits[i < RotAmt ? i + (Bits.size() - RotAmt) : i - RotAmt];
934 
935         return true;
936       }
937       break;
938     case ISD::SHL:
939       if (isa<ConstantSDNode>(V.getOperand(1))) {
940         unsigned ShiftAmt = V.getConstantOperandVal(1);
941 
942         SmallVector<ValueBit, 64> LHSBits(Bits.size());
943         getValueBits(V.getOperand(0), LHSBits);
944 
945         for (unsigned i = ShiftAmt; i < Bits.size(); ++i)
946           Bits[i] = LHSBits[i - ShiftAmt];
947 
948         for (unsigned i = 0; i < ShiftAmt; ++i)
949           Bits[i] = ValueBit(ValueBit::ConstZero);
950 
951         return true;
952       }
953       break;
954     case ISD::SRL:
955       if (isa<ConstantSDNode>(V.getOperand(1))) {
956         unsigned ShiftAmt = V.getConstantOperandVal(1);
957 
958         SmallVector<ValueBit, 64> LHSBits(Bits.size());
959         getValueBits(V.getOperand(0), LHSBits);
960 
961         for (unsigned i = 0; i < Bits.size() - ShiftAmt; ++i)
962           Bits[i] = LHSBits[i + ShiftAmt];
963 
964         for (unsigned i = Bits.size() - ShiftAmt; i < Bits.size(); ++i)
965           Bits[i] = ValueBit(ValueBit::ConstZero);
966 
967         return true;
968       }
969       break;
970     case ISD::AND:
971       if (isa<ConstantSDNode>(V.getOperand(1))) {
972         uint64_t Mask = V.getConstantOperandVal(1);
973 
974         SmallVector<ValueBit, 64> LHSBits(Bits.size());
975         bool LHSTrivial = getValueBits(V.getOperand(0), LHSBits);
976 
977         for (unsigned i = 0; i < Bits.size(); ++i)
978           if (((Mask >> i) & 1) == 1)
979             Bits[i] = LHSBits[i];
980           else
981             Bits[i] = ValueBit(ValueBit::ConstZero);
982 
983         // Mark this as interesting, only if the LHS was also interesting. This
984         // prevents the overall procedure from matching a single immediate 'and'
985         // (which is non-optimal because such an and might be folded with other
986         // things if we don't select it here).
987         return LHSTrivial;
988       }
989       break;
990     case ISD::OR: {
991       SmallVector<ValueBit, 64> LHSBits(Bits.size()), RHSBits(Bits.size());
992       getValueBits(V.getOperand(0), LHSBits);
993       getValueBits(V.getOperand(1), RHSBits);
994 
995       bool AllDisjoint = true;
996       for (unsigned i = 0; i < Bits.size(); ++i)
997         if (LHSBits[i].isZero())
998           Bits[i] = RHSBits[i];
999         else if (RHSBits[i].isZero())
1000           Bits[i] = LHSBits[i];
1001         else {
1002           AllDisjoint = false;
1003           break;
1004         }
1005 
1006       if (!AllDisjoint)
1007         break;
1008 
1009       return true;
1010     }
1011     }
1012 
1013     for (unsigned i = 0; i < Bits.size(); ++i)
1014       Bits[i] = ValueBit(V, i);
1015 
1016     return false;
1017   }
1018 
1019   // For each value (except the constant ones), compute the left-rotate amount
1020   // to get it from its original to final position.
computeRotationAmounts()1021   void computeRotationAmounts() {
1022     HasZeros = false;
1023     RLAmt.resize(Bits.size());
1024     for (unsigned i = 0; i < Bits.size(); ++i)
1025       if (Bits[i].hasValue()) {
1026         unsigned VBI = Bits[i].getValueBitIndex();
1027         if (i >= VBI)
1028           RLAmt[i] = i - VBI;
1029         else
1030           RLAmt[i] = Bits.size() - (VBI - i);
1031       } else if (Bits[i].isZero()) {
1032         HasZeros = true;
1033         RLAmt[i] = UINT32_MAX;
1034       } else {
1035         llvm_unreachable("Unknown value bit type");
1036       }
1037   }
1038 
1039   // Collect groups of consecutive bits with the same underlying value and
1040   // rotation factor. If we're doing late masking, we ignore zeros, otherwise
1041   // they break up groups.
collectBitGroups(bool LateMask)1042   void collectBitGroups(bool LateMask) {
1043     BitGroups.clear();
1044 
1045     unsigned LastRLAmt = RLAmt[0];
1046     SDValue LastValue = Bits[0].hasValue() ? Bits[0].getValue() : SDValue();
1047     unsigned LastGroupStartIdx = 0;
1048     for (unsigned i = 1; i < Bits.size(); ++i) {
1049       unsigned ThisRLAmt = RLAmt[i];
1050       SDValue ThisValue = Bits[i].hasValue() ? Bits[i].getValue() : SDValue();
1051       if (LateMask && !ThisValue) {
1052         ThisValue = LastValue;
1053         ThisRLAmt = LastRLAmt;
1054         // If we're doing late masking, then the first bit group always starts
1055         // at zero (even if the first bits were zero).
1056         if (BitGroups.empty())
1057           LastGroupStartIdx = 0;
1058       }
1059 
1060       // If this bit has the same underlying value and the same rotate factor as
1061       // the last one, then they're part of the same group.
1062       if (ThisRLAmt == LastRLAmt && ThisValue == LastValue)
1063         continue;
1064 
1065       if (LastValue.getNode())
1066         BitGroups.push_back(BitGroup(LastValue, LastRLAmt, LastGroupStartIdx,
1067                                      i-1));
1068       LastRLAmt = ThisRLAmt;
1069       LastValue = ThisValue;
1070       LastGroupStartIdx = i;
1071     }
1072     if (LastValue.getNode())
1073       BitGroups.push_back(BitGroup(LastValue, LastRLAmt, LastGroupStartIdx,
1074                                    Bits.size()-1));
1075 
1076     if (BitGroups.empty())
1077       return;
1078 
1079     // We might be able to combine the first and last groups.
1080     if (BitGroups.size() > 1) {
1081       // If the first and last groups are the same, then remove the first group
1082       // in favor of the last group, making the ending index of the last group
1083       // equal to the ending index of the to-be-removed first group.
1084       if (BitGroups[0].StartIdx == 0 &&
1085           BitGroups[BitGroups.size()-1].EndIdx == Bits.size()-1 &&
1086           BitGroups[0].V == BitGroups[BitGroups.size()-1].V &&
1087           BitGroups[0].RLAmt == BitGroups[BitGroups.size()-1].RLAmt) {
1088         DEBUG(dbgs() << "\tcombining final bit group with initial one\n");
1089         BitGroups[BitGroups.size()-1].EndIdx = BitGroups[0].EndIdx;
1090         BitGroups.erase(BitGroups.begin());
1091       }
1092     }
1093   }
1094 
1095   // Take all (SDValue, RLAmt) pairs and sort them by the number of groups
1096   // associated with each. If there is a degeneracy, pick the one that occurs
1097   // first (in the final value).
collectValueRotInfo()1098   void collectValueRotInfo() {
1099     ValueRots.clear();
1100 
1101     for (auto &BG : BitGroups) {
1102       unsigned RLAmtKey = BG.RLAmt + (BG.Repl32 ? 64 : 0);
1103       ValueRotInfo &VRI = ValueRots[std::make_pair(BG.V, RLAmtKey)];
1104       VRI.V = BG.V;
1105       VRI.RLAmt = BG.RLAmt;
1106       VRI.Repl32 = BG.Repl32;
1107       VRI.NumGroups += 1;
1108       VRI.FirstGroupStartIdx = std::min(VRI.FirstGroupStartIdx, BG.StartIdx);
1109     }
1110 
1111     // Now that we've collected the various ValueRotInfo instances, we need to
1112     // sort them.
1113     ValueRotsVec.clear();
1114     for (auto &I : ValueRots) {
1115       ValueRotsVec.push_back(I.second);
1116     }
1117     std::sort(ValueRotsVec.begin(), ValueRotsVec.end());
1118   }
1119 
1120   // In 64-bit mode, rlwinm and friends have a rotation operator that
1121   // replicates the low-order 32 bits into the high-order 32-bits. The mask
1122   // indices of these instructions can only be in the lower 32 bits, so they
1123   // can only represent some 64-bit bit groups. However, when they can be used,
1124   // the 32-bit replication can be used to represent, as a single bit group,
1125   // otherwise separate bit groups. We'll convert to replicated-32-bit bit
1126   // groups when possible. Returns true if any of the bit groups were
1127   // converted.
assignRepl32BitGroups()1128   void assignRepl32BitGroups() {
1129     // If we have bits like this:
1130     //
1131     // Indices:    15 14 13 12 11 10 9 8  7  6  5  4  3  2  1  0
1132     // V bits: ... 7  6  5  4  3  2  1 0 31 30 29 28 27 26 25 24
1133     // Groups:    |      RLAmt = 8      |      RLAmt = 40       |
1134     //
1135     // But, making use of a 32-bit operation that replicates the low-order 32
1136     // bits into the high-order 32 bits, this can be one bit group with a RLAmt
1137     // of 8.
1138 
1139     auto IsAllLow32 = [this](BitGroup & BG) {
1140       if (BG.StartIdx <= BG.EndIdx) {
1141         for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i) {
1142           if (!Bits[i].hasValue())
1143             continue;
1144           if (Bits[i].getValueBitIndex() >= 32)
1145             return false;
1146         }
1147       } else {
1148         for (unsigned i = BG.StartIdx; i < Bits.size(); ++i) {
1149           if (!Bits[i].hasValue())
1150             continue;
1151           if (Bits[i].getValueBitIndex() >= 32)
1152             return false;
1153         }
1154         for (unsigned i = 0; i <= BG.EndIdx; ++i) {
1155           if (!Bits[i].hasValue())
1156             continue;
1157           if (Bits[i].getValueBitIndex() >= 32)
1158             return false;
1159         }
1160       }
1161 
1162       return true;
1163     };
1164 
1165     for (auto &BG : BitGroups) {
1166       if (BG.StartIdx < 32 && BG.EndIdx < 32) {
1167         if (IsAllLow32(BG)) {
1168           if (BG.RLAmt >= 32) {
1169             BG.RLAmt -= 32;
1170             BG.Repl32CR = true;
1171           }
1172 
1173           BG.Repl32 = true;
1174 
1175           DEBUG(dbgs() << "\t32-bit replicated bit group for " <<
1176                           BG.V.getNode() << " RLAmt = " << BG.RLAmt <<
1177                           " [" << BG.StartIdx << ", " << BG.EndIdx << "]\n");
1178         }
1179       }
1180     }
1181 
1182     // Now walk through the bit groups, consolidating where possible.
1183     for (auto I = BitGroups.begin(); I != BitGroups.end();) {
1184       // We might want to remove this bit group by merging it with the previous
1185       // group (which might be the ending group).
1186       auto IP = (I == BitGroups.begin()) ?
1187                 std::prev(BitGroups.end()) : std::prev(I);
1188       if (I->Repl32 && IP->Repl32 && I->V == IP->V && I->RLAmt == IP->RLAmt &&
1189           I->StartIdx == (IP->EndIdx + 1) % 64 && I != IP) {
1190 
1191         DEBUG(dbgs() << "\tcombining 32-bit replicated bit group for " <<
1192                         I->V.getNode() << " RLAmt = " << I->RLAmt <<
1193                         " [" << I->StartIdx << ", " << I->EndIdx <<
1194                         "] with group with range [" <<
1195                         IP->StartIdx << ", " << IP->EndIdx << "]\n");
1196 
1197         IP->EndIdx = I->EndIdx;
1198         IP->Repl32CR = IP->Repl32CR || I->Repl32CR;
1199         IP->Repl32Coalesced = true;
1200         I = BitGroups.erase(I);
1201         continue;
1202       } else {
1203         // There is a special case worth handling: If there is a single group
1204         // covering the entire upper 32 bits, and it can be merged with both
1205         // the next and previous groups (which might be the same group), then
1206         // do so. If it is the same group (so there will be only one group in
1207         // total), then we need to reverse the order of the range so that it
1208         // covers the entire 64 bits.
1209         if (I->StartIdx == 32 && I->EndIdx == 63) {
1210           assert(std::next(I) == BitGroups.end() &&
1211                  "bit group ends at index 63 but there is another?");
1212           auto IN = BitGroups.begin();
1213 
1214           if (IP->Repl32 && IN->Repl32 && I->V == IP->V && I->V == IN->V &&
1215               (I->RLAmt % 32) == IP->RLAmt && (I->RLAmt % 32) == IN->RLAmt &&
1216               IP->EndIdx == 31 && IN->StartIdx == 0 && I != IP &&
1217               IsAllLow32(*I)) {
1218 
1219             DEBUG(dbgs() << "\tcombining bit group for " <<
1220                             I->V.getNode() << " RLAmt = " << I->RLAmt <<
1221                             " [" << I->StartIdx << ", " << I->EndIdx <<
1222                             "] with 32-bit replicated groups with ranges [" <<
1223                             IP->StartIdx << ", " << IP->EndIdx << "] and [" <<
1224                             IN->StartIdx << ", " << IN->EndIdx << "]\n");
1225 
1226             if (IP == IN) {
1227               // There is only one other group; change it to cover the whole
1228               // range (backward, so that it can still be Repl32 but cover the
1229               // whole 64-bit range).
1230               IP->StartIdx = 31;
1231               IP->EndIdx = 30;
1232               IP->Repl32CR = IP->Repl32CR || I->RLAmt >= 32;
1233               IP->Repl32Coalesced = true;
1234               I = BitGroups.erase(I);
1235             } else {
1236               // There are two separate groups, one before this group and one
1237               // after us (at the beginning). We're going to remove this group,
1238               // but also the group at the very beginning.
1239               IP->EndIdx = IN->EndIdx;
1240               IP->Repl32CR = IP->Repl32CR || IN->Repl32CR || I->RLAmt >= 32;
1241               IP->Repl32Coalesced = true;
1242               I = BitGroups.erase(I);
1243               BitGroups.erase(BitGroups.begin());
1244             }
1245 
1246             // This must be the last group in the vector (and we might have
1247             // just invalidated the iterator above), so break here.
1248             break;
1249           }
1250         }
1251       }
1252 
1253       ++I;
1254     }
1255   }
1256 
getI32Imm(unsigned Imm,SDLoc dl)1257   SDValue getI32Imm(unsigned Imm, SDLoc dl) {
1258     return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
1259   }
1260 
getZerosMask()1261   uint64_t getZerosMask() {
1262     uint64_t Mask = 0;
1263     for (unsigned i = 0; i < Bits.size(); ++i) {
1264       if (Bits[i].hasValue())
1265         continue;
1266       Mask |= (UINT64_C(1) << i);
1267     }
1268 
1269     return ~Mask;
1270   }
1271 
1272   // Depending on the number of groups for a particular value, it might be
1273   // better to rotate, mask explicitly (using andi/andis), and then or the
1274   // result. Select this part of the result first.
SelectAndParts32(SDLoc dl,SDValue & Res,unsigned * InstCnt)1275   void SelectAndParts32(SDLoc dl, SDValue &Res, unsigned *InstCnt) {
1276     if (BPermRewriterNoMasking)
1277       return;
1278 
1279     for (ValueRotInfo &VRI : ValueRotsVec) {
1280       unsigned Mask = 0;
1281       for (unsigned i = 0; i < Bits.size(); ++i) {
1282         if (!Bits[i].hasValue() || Bits[i].getValue() != VRI.V)
1283           continue;
1284         if (RLAmt[i] != VRI.RLAmt)
1285           continue;
1286         Mask |= (1u << i);
1287       }
1288 
1289       // Compute the masks for andi/andis that would be necessary.
1290       unsigned ANDIMask = (Mask & UINT16_MAX), ANDISMask = Mask >> 16;
1291       assert((ANDIMask != 0 || ANDISMask != 0) &&
1292              "No set bits in mask for value bit groups");
1293       bool NeedsRotate = VRI.RLAmt != 0;
1294 
1295       // We're trying to minimize the number of instructions. If we have one
1296       // group, using one of andi/andis can break even.  If we have three
1297       // groups, we can use both andi and andis and break even (to use both
1298       // andi and andis we also need to or the results together). We need four
1299       // groups if we also need to rotate. To use andi/andis we need to do more
1300       // than break even because rotate-and-mask instructions tend to be easier
1301       // to schedule.
1302 
1303       // FIXME: We've biased here against using andi/andis, which is right for
1304       // POWER cores, but not optimal everywhere. For example, on the A2,
1305       // andi/andis have single-cycle latency whereas the rotate-and-mask
1306       // instructions take two cycles, and it would be better to bias toward
1307       // andi/andis in break-even cases.
1308 
1309       unsigned NumAndInsts = (unsigned) NeedsRotate +
1310                              (unsigned) (ANDIMask != 0) +
1311                              (unsigned) (ANDISMask != 0) +
1312                              (unsigned) (ANDIMask != 0 && ANDISMask != 0) +
1313                              (unsigned) (bool) Res;
1314 
1315       DEBUG(dbgs() << "\t\trotation groups for " << VRI.V.getNode() <<
1316                       " RL: " << VRI.RLAmt << ":" <<
1317                       "\n\t\t\tisel using masking: " << NumAndInsts <<
1318                       " using rotates: " << VRI.NumGroups << "\n");
1319 
1320       if (NumAndInsts >= VRI.NumGroups)
1321         continue;
1322 
1323       DEBUG(dbgs() << "\t\t\t\tusing masking\n");
1324 
1325       if (InstCnt) *InstCnt += NumAndInsts;
1326 
1327       SDValue VRot;
1328       if (VRI.RLAmt) {
1329         SDValue Ops[] =
1330           { VRI.V, getI32Imm(VRI.RLAmt, dl), getI32Imm(0, dl),
1331             getI32Imm(31, dl) };
1332         VRot = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32,
1333                                               Ops), 0);
1334       } else {
1335         VRot = VRI.V;
1336       }
1337 
1338       SDValue ANDIVal, ANDISVal;
1339       if (ANDIMask != 0)
1340         ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32,
1341                             VRot, getI32Imm(ANDIMask, dl)), 0);
1342       if (ANDISMask != 0)
1343         ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32,
1344                              VRot, getI32Imm(ANDISMask, dl)), 0);
1345 
1346       SDValue TotalVal;
1347       if (!ANDIVal)
1348         TotalVal = ANDISVal;
1349       else if (!ANDISVal)
1350         TotalVal = ANDIVal;
1351       else
1352         TotalVal = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32,
1353                              ANDIVal, ANDISVal), 0);
1354 
1355       if (!Res)
1356         Res = TotalVal;
1357       else
1358         Res = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32,
1359                         Res, TotalVal), 0);
1360 
1361       // Now, remove all groups with this underlying value and rotation
1362       // factor.
1363       eraseMatchingBitGroups([VRI](const BitGroup &BG) {
1364         return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt;
1365       });
1366     }
1367   }
1368 
1369   // Instruction selection for the 32-bit case.
Select32(SDNode * N,bool LateMask,unsigned * InstCnt)1370   SDNode *Select32(SDNode *N, bool LateMask, unsigned *InstCnt) {
1371     SDLoc dl(N);
1372     SDValue Res;
1373 
1374     if (InstCnt) *InstCnt = 0;
1375 
1376     // Take care of cases that should use andi/andis first.
1377     SelectAndParts32(dl, Res, InstCnt);
1378 
1379     // If we've not yet selected a 'starting' instruction, and we have no zeros
1380     // to fill in, select the (Value, RLAmt) with the highest priority (largest
1381     // number of groups), and start with this rotated value.
1382     if ((!HasZeros || LateMask) && !Res) {
1383       ValueRotInfo &VRI = ValueRotsVec[0];
1384       if (VRI.RLAmt) {
1385         if (InstCnt) *InstCnt += 1;
1386         SDValue Ops[] =
1387           { VRI.V, getI32Imm(VRI.RLAmt, dl), getI32Imm(0, dl),
1388             getI32Imm(31, dl) };
1389         Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops),
1390                       0);
1391       } else {
1392         Res = VRI.V;
1393       }
1394 
1395       // Now, remove all groups with this underlying value and rotation factor.
1396       eraseMatchingBitGroups([VRI](const BitGroup &BG) {
1397         return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt;
1398       });
1399     }
1400 
1401     if (InstCnt) *InstCnt += BitGroups.size();
1402 
1403     // Insert the other groups (one at a time).
1404     for (auto &BG : BitGroups) {
1405       if (!Res) {
1406         SDValue Ops[] =
1407           { BG.V, getI32Imm(BG.RLAmt, dl),
1408             getI32Imm(Bits.size() - BG.EndIdx - 1, dl),
1409             getI32Imm(Bits.size() - BG.StartIdx - 1, dl) };
1410         Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
1411       } else {
1412         SDValue Ops[] =
1413           { Res, BG.V, getI32Imm(BG.RLAmt, dl),
1414               getI32Imm(Bits.size() - BG.EndIdx - 1, dl),
1415             getI32Imm(Bits.size() - BG.StartIdx - 1, dl) };
1416         Res = SDValue(CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops), 0);
1417       }
1418     }
1419 
1420     if (LateMask) {
1421       unsigned Mask = (unsigned) getZerosMask();
1422 
1423       unsigned ANDIMask = (Mask & UINT16_MAX), ANDISMask = Mask >> 16;
1424       assert((ANDIMask != 0 || ANDISMask != 0) &&
1425              "No set bits in zeros mask?");
1426 
1427       if (InstCnt) *InstCnt += (unsigned) (ANDIMask != 0) +
1428                                (unsigned) (ANDISMask != 0) +
1429                                (unsigned) (ANDIMask != 0 && ANDISMask != 0);
1430 
1431       SDValue ANDIVal, ANDISVal;
1432       if (ANDIMask != 0)
1433         ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32,
1434                             Res, getI32Imm(ANDIMask, dl)), 0);
1435       if (ANDISMask != 0)
1436         ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32,
1437                              Res, getI32Imm(ANDISMask, dl)), 0);
1438 
1439       if (!ANDIVal)
1440         Res = ANDISVal;
1441       else if (!ANDISVal)
1442         Res = ANDIVal;
1443       else
1444         Res = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32,
1445                         ANDIVal, ANDISVal), 0);
1446     }
1447 
1448     return Res.getNode();
1449   }
1450 
SelectRotMask64Count(unsigned RLAmt,bool Repl32,unsigned MaskStart,unsigned MaskEnd,bool IsIns)1451   unsigned SelectRotMask64Count(unsigned RLAmt, bool Repl32,
1452                                 unsigned MaskStart, unsigned MaskEnd,
1453                                 bool IsIns) {
1454     // In the notation used by the instructions, 'start' and 'end' are reversed
1455     // because bits are counted from high to low order.
1456     unsigned InstMaskStart = 64 - MaskEnd - 1,
1457              InstMaskEnd   = 64 - MaskStart - 1;
1458 
1459     if (Repl32)
1460       return 1;
1461 
1462     if ((!IsIns && (InstMaskEnd == 63 || InstMaskStart == 0)) ||
1463         InstMaskEnd == 63 - RLAmt)
1464       return 1;
1465 
1466     return 2;
1467   }
1468 
1469   // For 64-bit values, not all combinations of rotates and masks are
1470   // available. Produce one if it is available.
SelectRotMask64(SDValue V,SDLoc dl,unsigned RLAmt,bool Repl32,unsigned MaskStart,unsigned MaskEnd,unsigned * InstCnt=nullptr)1471   SDValue SelectRotMask64(SDValue V, SDLoc dl, unsigned RLAmt, bool Repl32,
1472                           unsigned MaskStart, unsigned MaskEnd,
1473                           unsigned *InstCnt = nullptr) {
1474     // In the notation used by the instructions, 'start' and 'end' are reversed
1475     // because bits are counted from high to low order.
1476     unsigned InstMaskStart = 64 - MaskEnd - 1,
1477              InstMaskEnd   = 64 - MaskStart - 1;
1478 
1479     if (InstCnt) *InstCnt += 1;
1480 
1481     if (Repl32) {
1482       // This rotation amount assumes that the lower 32 bits of the quantity
1483       // are replicated in the high 32 bits by the rotation operator (which is
1484       // done by rlwinm and friends).
1485       assert(InstMaskStart >= 32 && "Mask cannot start out of range");
1486       assert(InstMaskEnd   >= 32 && "Mask cannot end out of range");
1487       SDValue Ops[] =
1488         { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart - 32, dl),
1489           getI32Imm(InstMaskEnd - 32, dl) };
1490       return SDValue(CurDAG->getMachineNode(PPC::RLWINM8, dl, MVT::i64,
1491                                             Ops), 0);
1492     }
1493 
1494     if (InstMaskEnd == 63) {
1495       SDValue Ops[] =
1496         { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) };
1497       return SDValue(CurDAG->getMachineNode(PPC::RLDICL, dl, MVT::i64, Ops), 0);
1498     }
1499 
1500     if (InstMaskStart == 0) {
1501       SDValue Ops[] =
1502         { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskEnd, dl) };
1503       return SDValue(CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, Ops), 0);
1504     }
1505 
1506     if (InstMaskEnd == 63 - RLAmt) {
1507       SDValue Ops[] =
1508         { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) };
1509       return SDValue(CurDAG->getMachineNode(PPC::RLDIC, dl, MVT::i64, Ops), 0);
1510     }
1511 
1512     // We cannot do this with a single instruction, so we'll use two. The
1513     // problem is that we're not free to choose both a rotation amount and mask
1514     // start and end independently. We can choose an arbitrary mask start and
1515     // end, but then the rotation amount is fixed. Rotation, however, can be
1516     // inverted, and so by applying an "inverse" rotation first, we can get the
1517     // desired result.
1518     if (InstCnt) *InstCnt += 1;
1519 
1520     // The rotation mask for the second instruction must be MaskStart.
1521     unsigned RLAmt2 = MaskStart;
1522     // The first instruction must rotate V so that the overall rotation amount
1523     // is RLAmt.
1524     unsigned RLAmt1 = (64 + RLAmt - RLAmt2) % 64;
1525     if (RLAmt1)
1526       V = SelectRotMask64(V, dl, RLAmt1, false, 0, 63);
1527     return SelectRotMask64(V, dl, RLAmt2, false, MaskStart, MaskEnd);
1528   }
1529 
1530   // For 64-bit values, not all combinations of rotates and masks are
1531   // available. Produce a rotate-mask-and-insert if one is available.
SelectRotMaskIns64(SDValue Base,SDValue V,SDLoc dl,unsigned RLAmt,bool Repl32,unsigned MaskStart,unsigned MaskEnd,unsigned * InstCnt=nullptr)1532   SDValue SelectRotMaskIns64(SDValue Base, SDValue V, SDLoc dl, unsigned RLAmt,
1533                              bool Repl32, unsigned MaskStart,
1534                              unsigned MaskEnd, unsigned *InstCnt = nullptr) {
1535     // In the notation used by the instructions, 'start' and 'end' are reversed
1536     // because bits are counted from high to low order.
1537     unsigned InstMaskStart = 64 - MaskEnd - 1,
1538              InstMaskEnd   = 64 - MaskStart - 1;
1539 
1540     if (InstCnt) *InstCnt += 1;
1541 
1542     if (Repl32) {
1543       // This rotation amount assumes that the lower 32 bits of the quantity
1544       // are replicated in the high 32 bits by the rotation operator (which is
1545       // done by rlwinm and friends).
1546       assert(InstMaskStart >= 32 && "Mask cannot start out of range");
1547       assert(InstMaskEnd   >= 32 && "Mask cannot end out of range");
1548       SDValue Ops[] =
1549         { Base, V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart - 32, dl),
1550           getI32Imm(InstMaskEnd - 32, dl) };
1551       return SDValue(CurDAG->getMachineNode(PPC::RLWIMI8, dl, MVT::i64,
1552                                             Ops), 0);
1553     }
1554 
1555     if (InstMaskEnd == 63 - RLAmt) {
1556       SDValue Ops[] =
1557         { Base, V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) };
1558       return SDValue(CurDAG->getMachineNode(PPC::RLDIMI, dl, MVT::i64, Ops), 0);
1559     }
1560 
1561     // We cannot do this with a single instruction, so we'll use two. The
1562     // problem is that we're not free to choose both a rotation amount and mask
1563     // start and end independently. We can choose an arbitrary mask start and
1564     // end, but then the rotation amount is fixed. Rotation, however, can be
1565     // inverted, and so by applying an "inverse" rotation first, we can get the
1566     // desired result.
1567     if (InstCnt) *InstCnt += 1;
1568 
1569     // The rotation mask for the second instruction must be MaskStart.
1570     unsigned RLAmt2 = MaskStart;
1571     // The first instruction must rotate V so that the overall rotation amount
1572     // is RLAmt.
1573     unsigned RLAmt1 = (64 + RLAmt - RLAmt2) % 64;
1574     if (RLAmt1)
1575       V = SelectRotMask64(V, dl, RLAmt1, false, 0, 63);
1576     return SelectRotMaskIns64(Base, V, dl, RLAmt2, false, MaskStart, MaskEnd);
1577   }
1578 
SelectAndParts64(SDLoc dl,SDValue & Res,unsigned * InstCnt)1579   void SelectAndParts64(SDLoc dl, SDValue &Res, unsigned *InstCnt) {
1580     if (BPermRewriterNoMasking)
1581       return;
1582 
1583     // The idea here is the same as in the 32-bit version, but with additional
1584     // complications from the fact that Repl32 might be true. Because we
1585     // aggressively convert bit groups to Repl32 form (which, for small
1586     // rotation factors, involves no other change), and then coalesce, it might
1587     // be the case that a single 64-bit masking operation could handle both
1588     // some Repl32 groups and some non-Repl32 groups. If converting to Repl32
1589     // form allowed coalescing, then we must use a 32-bit rotaton in order to
1590     // completely capture the new combined bit group.
1591 
1592     for (ValueRotInfo &VRI : ValueRotsVec) {
1593       uint64_t Mask = 0;
1594 
1595       // We need to add to the mask all bits from the associated bit groups.
1596       // If Repl32 is false, we need to add bits from bit groups that have
1597       // Repl32 true, but are trivially convertable to Repl32 false. Such a
1598       // group is trivially convertable if it overlaps only with the lower 32
1599       // bits, and the group has not been coalesced.
1600       auto MatchingBG = [VRI](const BitGroup &BG) {
1601         if (VRI.V != BG.V)
1602           return false;
1603 
1604         unsigned EffRLAmt = BG.RLAmt;
1605         if (!VRI.Repl32 && BG.Repl32) {
1606           if (BG.StartIdx < 32 && BG.EndIdx < 32 && BG.StartIdx <= BG.EndIdx &&
1607               !BG.Repl32Coalesced) {
1608             if (BG.Repl32CR)
1609               EffRLAmt += 32;
1610           } else {
1611             return false;
1612           }
1613         } else if (VRI.Repl32 != BG.Repl32) {
1614           return false;
1615         }
1616 
1617         if (VRI.RLAmt != EffRLAmt)
1618           return false;
1619 
1620         return true;
1621       };
1622 
1623       for (auto &BG : BitGroups) {
1624         if (!MatchingBG(BG))
1625           continue;
1626 
1627         if (BG.StartIdx <= BG.EndIdx) {
1628           for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i)
1629             Mask |= (UINT64_C(1) << i);
1630         } else {
1631           for (unsigned i = BG.StartIdx; i < Bits.size(); ++i)
1632             Mask |= (UINT64_C(1) << i);
1633           for (unsigned i = 0; i <= BG.EndIdx; ++i)
1634             Mask |= (UINT64_C(1) << i);
1635         }
1636       }
1637 
1638       // We can use the 32-bit andi/andis technique if the mask does not
1639       // require any higher-order bits. This can save an instruction compared
1640       // to always using the general 64-bit technique.
1641       bool Use32BitInsts = isUInt<32>(Mask);
1642       // Compute the masks for andi/andis that would be necessary.
1643       unsigned ANDIMask = (Mask & UINT16_MAX),
1644                ANDISMask = (Mask >> 16) & UINT16_MAX;
1645 
1646       bool NeedsRotate = VRI.RLAmt || (VRI.Repl32 && !isUInt<32>(Mask));
1647 
1648       unsigned NumAndInsts = (unsigned) NeedsRotate +
1649                              (unsigned) (bool) Res;
1650       if (Use32BitInsts)
1651         NumAndInsts += (unsigned) (ANDIMask != 0) + (unsigned) (ANDISMask != 0) +
1652                        (unsigned) (ANDIMask != 0 && ANDISMask != 0);
1653       else
1654         NumAndInsts += SelectInt64Count(Mask) + /* and */ 1;
1655 
1656       unsigned NumRLInsts = 0;
1657       bool FirstBG = true;
1658       for (auto &BG : BitGroups) {
1659         if (!MatchingBG(BG))
1660           continue;
1661         NumRLInsts +=
1662           SelectRotMask64Count(BG.RLAmt, BG.Repl32, BG.StartIdx, BG.EndIdx,
1663                                !FirstBG);
1664         FirstBG = false;
1665       }
1666 
1667       DEBUG(dbgs() << "\t\trotation groups for " << VRI.V.getNode() <<
1668                       " RL: " << VRI.RLAmt << (VRI.Repl32 ? " (32):" : ":") <<
1669                       "\n\t\t\tisel using masking: " << NumAndInsts <<
1670                       " using rotates: " << NumRLInsts << "\n");
1671 
1672       // When we'd use andi/andis, we bias toward using the rotates (andi only
1673       // has a record form, and is cracked on POWER cores). However, when using
1674       // general 64-bit constant formation, bias toward the constant form,
1675       // because that exposes more opportunities for CSE.
1676       if (NumAndInsts > NumRLInsts)
1677         continue;
1678       if (Use32BitInsts && NumAndInsts == NumRLInsts)
1679         continue;
1680 
1681       DEBUG(dbgs() << "\t\t\t\tusing masking\n");
1682 
1683       if (InstCnt) *InstCnt += NumAndInsts;
1684 
1685       SDValue VRot;
1686       // We actually need to generate a rotation if we have a non-zero rotation
1687       // factor or, in the Repl32 case, if we care about any of the
1688       // higher-order replicated bits. In the latter case, we generate a mask
1689       // backward so that it actually includes the entire 64 bits.
1690       if (VRI.RLAmt || (VRI.Repl32 && !isUInt<32>(Mask)))
1691         VRot = SelectRotMask64(VRI.V, dl, VRI.RLAmt, VRI.Repl32,
1692                                VRI.Repl32 ? 31 : 0, VRI.Repl32 ? 30 : 63);
1693       else
1694         VRot = VRI.V;
1695 
1696       SDValue TotalVal;
1697       if (Use32BitInsts) {
1698         assert((ANDIMask != 0 || ANDISMask != 0) &&
1699                "No set bits in mask when using 32-bit ands for 64-bit value");
1700 
1701         SDValue ANDIVal, ANDISVal;
1702         if (ANDIMask != 0)
1703           ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64,
1704                               VRot, getI32Imm(ANDIMask, dl)), 0);
1705         if (ANDISMask != 0)
1706           ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64,
1707                                VRot, getI32Imm(ANDISMask, dl)), 0);
1708 
1709         if (!ANDIVal)
1710           TotalVal = ANDISVal;
1711         else if (!ANDISVal)
1712           TotalVal = ANDIVal;
1713         else
1714           TotalVal = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64,
1715                                ANDIVal, ANDISVal), 0);
1716       } else {
1717         TotalVal = SDValue(SelectInt64(CurDAG, dl, Mask), 0);
1718         TotalVal =
1719           SDValue(CurDAG->getMachineNode(PPC::AND8, dl, MVT::i64,
1720                                          VRot, TotalVal), 0);
1721      }
1722 
1723       if (!Res)
1724         Res = TotalVal;
1725       else
1726         Res = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64,
1727                                              Res, TotalVal), 0);
1728 
1729       // Now, remove all groups with this underlying value and rotation
1730       // factor.
1731       eraseMatchingBitGroups(MatchingBG);
1732     }
1733   }
1734 
1735   // Instruction selection for the 64-bit case.
Select64(SDNode * N,bool LateMask,unsigned * InstCnt)1736   SDNode *Select64(SDNode *N, bool LateMask, unsigned *InstCnt) {
1737     SDLoc dl(N);
1738     SDValue Res;
1739 
1740     if (InstCnt) *InstCnt = 0;
1741 
1742     // Take care of cases that should use andi/andis first.
1743     SelectAndParts64(dl, Res, InstCnt);
1744 
1745     // If we've not yet selected a 'starting' instruction, and we have no zeros
1746     // to fill in, select the (Value, RLAmt) with the highest priority (largest
1747     // number of groups), and start with this rotated value.
1748     if ((!HasZeros || LateMask) && !Res) {
1749       // If we have both Repl32 groups and non-Repl32 groups, the non-Repl32
1750       // groups will come first, and so the VRI representing the largest number
1751       // of groups might not be first (it might be the first Repl32 groups).
1752       unsigned MaxGroupsIdx = 0;
1753       if (!ValueRotsVec[0].Repl32) {
1754         for (unsigned i = 0, ie = ValueRotsVec.size(); i < ie; ++i)
1755           if (ValueRotsVec[i].Repl32) {
1756             if (ValueRotsVec[i].NumGroups > ValueRotsVec[0].NumGroups)
1757               MaxGroupsIdx = i;
1758             break;
1759           }
1760       }
1761 
1762       ValueRotInfo &VRI = ValueRotsVec[MaxGroupsIdx];
1763       bool NeedsRotate = false;
1764       if (VRI.RLAmt) {
1765         NeedsRotate = true;
1766       } else if (VRI.Repl32) {
1767         for (auto &BG : BitGroups) {
1768           if (BG.V != VRI.V || BG.RLAmt != VRI.RLAmt ||
1769               BG.Repl32 != VRI.Repl32)
1770             continue;
1771 
1772           // We don't need a rotate if the bit group is confined to the lower
1773           // 32 bits.
1774           if (BG.StartIdx < 32 && BG.EndIdx < 32 && BG.StartIdx < BG.EndIdx)
1775             continue;
1776 
1777           NeedsRotate = true;
1778           break;
1779         }
1780       }
1781 
1782       if (NeedsRotate)
1783         Res = SelectRotMask64(VRI.V, dl, VRI.RLAmt, VRI.Repl32,
1784                               VRI.Repl32 ? 31 : 0, VRI.Repl32 ? 30 : 63,
1785                               InstCnt);
1786       else
1787         Res = VRI.V;
1788 
1789       // Now, remove all groups with this underlying value and rotation factor.
1790       if (Res)
1791         eraseMatchingBitGroups([VRI](const BitGroup &BG) {
1792           return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt &&
1793                  BG.Repl32 == VRI.Repl32;
1794         });
1795     }
1796 
1797     // Because 64-bit rotates are more flexible than inserts, we might have a
1798     // preference regarding which one we do first (to save one instruction).
1799     if (!Res)
1800       for (auto I = BitGroups.begin(), IE = BitGroups.end(); I != IE; ++I) {
1801         if (SelectRotMask64Count(I->RLAmt, I->Repl32, I->StartIdx, I->EndIdx,
1802                                 false) <
1803             SelectRotMask64Count(I->RLAmt, I->Repl32, I->StartIdx, I->EndIdx,
1804                                 true)) {
1805           if (I != BitGroups.begin()) {
1806             BitGroup BG = *I;
1807             BitGroups.erase(I);
1808             BitGroups.insert(BitGroups.begin(), BG);
1809           }
1810 
1811           break;
1812         }
1813       }
1814 
1815     // Insert the other groups (one at a time).
1816     for (auto &BG : BitGroups) {
1817       if (!Res)
1818         Res = SelectRotMask64(BG.V, dl, BG.RLAmt, BG.Repl32, BG.StartIdx,
1819                               BG.EndIdx, InstCnt);
1820       else
1821         Res = SelectRotMaskIns64(Res, BG.V, dl, BG.RLAmt, BG.Repl32,
1822                                  BG.StartIdx, BG.EndIdx, InstCnt);
1823     }
1824 
1825     if (LateMask) {
1826       uint64_t Mask = getZerosMask();
1827 
1828       // We can use the 32-bit andi/andis technique if the mask does not
1829       // require any higher-order bits. This can save an instruction compared
1830       // to always using the general 64-bit technique.
1831       bool Use32BitInsts = isUInt<32>(Mask);
1832       // Compute the masks for andi/andis that would be necessary.
1833       unsigned ANDIMask = (Mask & UINT16_MAX),
1834                ANDISMask = (Mask >> 16) & UINT16_MAX;
1835 
1836       if (Use32BitInsts) {
1837         assert((ANDIMask != 0 || ANDISMask != 0) &&
1838                "No set bits in mask when using 32-bit ands for 64-bit value");
1839 
1840         if (InstCnt) *InstCnt += (unsigned) (ANDIMask != 0) +
1841                                  (unsigned) (ANDISMask != 0) +
1842                                  (unsigned) (ANDIMask != 0 && ANDISMask != 0);
1843 
1844         SDValue ANDIVal, ANDISVal;
1845         if (ANDIMask != 0)
1846           ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64,
1847                               Res, getI32Imm(ANDIMask, dl)), 0);
1848         if (ANDISMask != 0)
1849           ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64,
1850                                Res, getI32Imm(ANDISMask, dl)), 0);
1851 
1852         if (!ANDIVal)
1853           Res = ANDISVal;
1854         else if (!ANDISVal)
1855           Res = ANDIVal;
1856         else
1857           Res = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64,
1858                           ANDIVal, ANDISVal), 0);
1859       } else {
1860         if (InstCnt) *InstCnt += SelectInt64Count(Mask) + /* and */ 1;
1861 
1862         SDValue MaskVal = SDValue(SelectInt64(CurDAG, dl, Mask), 0);
1863         Res =
1864           SDValue(CurDAG->getMachineNode(PPC::AND8, dl, MVT::i64,
1865                                          Res, MaskVal), 0);
1866       }
1867     }
1868 
1869     return Res.getNode();
1870   }
1871 
Select(SDNode * N,bool LateMask,unsigned * InstCnt=nullptr)1872   SDNode *Select(SDNode *N, bool LateMask, unsigned *InstCnt = nullptr) {
1873     // Fill in BitGroups.
1874     collectBitGroups(LateMask);
1875     if (BitGroups.empty())
1876       return nullptr;
1877 
1878     // For 64-bit values, figure out when we can use 32-bit instructions.
1879     if (Bits.size() == 64)
1880       assignRepl32BitGroups();
1881 
1882     // Fill in ValueRotsVec.
1883     collectValueRotInfo();
1884 
1885     if (Bits.size() == 32) {
1886       return Select32(N, LateMask, InstCnt);
1887     } else {
1888       assert(Bits.size() == 64 && "Not 64 bits here?");
1889       return Select64(N, LateMask, InstCnt);
1890     }
1891 
1892     return nullptr;
1893   }
1894 
eraseMatchingBitGroups(function_ref<bool (const BitGroup &)> F)1895   void eraseMatchingBitGroups(function_ref<bool(const BitGroup &)> F) {
1896     BitGroups.erase(std::remove_if(BitGroups.begin(), BitGroups.end(), F),
1897                     BitGroups.end());
1898   }
1899 
1900   SmallVector<ValueBit, 64> Bits;
1901 
1902   bool HasZeros;
1903   SmallVector<unsigned, 64> RLAmt;
1904 
1905   SmallVector<BitGroup, 16> BitGroups;
1906 
1907   DenseMap<std::pair<SDValue, unsigned>, ValueRotInfo> ValueRots;
1908   SmallVector<ValueRotInfo, 16> ValueRotsVec;
1909 
1910   SelectionDAG *CurDAG;
1911 
1912 public:
BitPermutationSelector(SelectionDAG * DAG)1913   BitPermutationSelector(SelectionDAG *DAG)
1914     : CurDAG(DAG) {}
1915 
1916   // Here we try to match complex bit permutations into a set of
1917   // rotate-and-shift/shift/and/or instructions, using a set of heuristics
1918   // known to produce optimial code for common cases (like i32 byte swapping).
Select(SDNode * N)1919   SDNode *Select(SDNode *N) {
1920     Bits.resize(N->getValueType(0).getSizeInBits());
1921     if (!getValueBits(SDValue(N, 0), Bits))
1922       return nullptr;
1923 
1924     DEBUG(dbgs() << "Considering bit-permutation-based instruction"
1925                     " selection for:    ");
1926     DEBUG(N->dump(CurDAG));
1927 
1928     // Fill it RLAmt and set HasZeros.
1929     computeRotationAmounts();
1930 
1931     if (!HasZeros)
1932       return Select(N, false);
1933 
1934     // We currently have two techniques for handling results with zeros: early
1935     // masking (the default) and late masking. Late masking is sometimes more
1936     // efficient, but because the structure of the bit groups is different, it
1937     // is hard to tell without generating both and comparing the results. With
1938     // late masking, we ignore zeros in the resulting value when inserting each
1939     // set of bit groups, and then mask in the zeros at the end. With early
1940     // masking, we only insert the non-zero parts of the result at every step.
1941 
1942     unsigned InstCnt, InstCntLateMask;
1943     DEBUG(dbgs() << "\tEarly masking:\n");
1944     SDNode *RN = Select(N, false, &InstCnt);
1945     DEBUG(dbgs() << "\t\tisel would use " << InstCnt << " instructions\n");
1946 
1947     DEBUG(dbgs() << "\tLate masking:\n");
1948     SDNode *RNLM = Select(N, true, &InstCntLateMask);
1949     DEBUG(dbgs() << "\t\tisel would use " << InstCntLateMask <<
1950                     " instructions\n");
1951 
1952     if (InstCnt <= InstCntLateMask) {
1953       DEBUG(dbgs() << "\tUsing early-masking for isel\n");
1954       return RN;
1955     }
1956 
1957     DEBUG(dbgs() << "\tUsing late-masking for isel\n");
1958     return RNLM;
1959   }
1960 };
1961 } // anonymous namespace
1962 
SelectBitPermutation(SDNode * N)1963 SDNode *PPCDAGToDAGISel::SelectBitPermutation(SDNode *N) {
1964   if (N->getValueType(0) != MVT::i32 &&
1965       N->getValueType(0) != MVT::i64)
1966     return nullptr;
1967 
1968   if (!UseBitPermRewriter)
1969     return nullptr;
1970 
1971   switch (N->getOpcode()) {
1972   default: break;
1973   case ISD::ROTL:
1974   case ISD::SHL:
1975   case ISD::SRL:
1976   case ISD::AND:
1977   case ISD::OR: {
1978     BitPermutationSelector BPS(CurDAG);
1979     return BPS.Select(N);
1980   }
1981   }
1982 
1983   return nullptr;
1984 }
1985 
1986 /// SelectCC - Select a comparison of the specified values with the specified
1987 /// condition code, returning the CR# of the expression.
SelectCC(SDValue LHS,SDValue RHS,ISD::CondCode CC,SDLoc dl)1988 SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS,
1989                                     ISD::CondCode CC, SDLoc dl) {
1990   // Always select the LHS.
1991   unsigned Opc;
1992 
1993   if (LHS.getValueType() == MVT::i32) {
1994     unsigned Imm;
1995     if (CC == ISD::SETEQ || CC == ISD::SETNE) {
1996       if (isInt32Immediate(RHS, Imm)) {
1997         // SETEQ/SETNE comparison with 16-bit immediate, fold it.
1998         if (isUInt<16>(Imm))
1999           return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
2000                                                 getI32Imm(Imm & 0xFFFF, dl)),
2001                          0);
2002         // If this is a 16-bit signed immediate, fold it.
2003         if (isInt<16>((int)Imm))
2004           return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
2005                                                 getI32Imm(Imm & 0xFFFF, dl)),
2006                          0);
2007 
2008         // For non-equality comparisons, the default code would materialize the
2009         // constant, then compare against it, like this:
2010         //   lis r2, 4660
2011         //   ori r2, r2, 22136
2012         //   cmpw cr0, r3, r2
2013         // Since we are just comparing for equality, we can emit this instead:
2014         //   xoris r0,r3,0x1234
2015         //   cmplwi cr0,r0,0x5678
2016         //   beq cr0,L6
2017         SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS,
2018                                            getI32Imm(Imm >> 16, dl)), 0);
2019         return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor,
2020                                               getI32Imm(Imm & 0xFFFF, dl)), 0);
2021       }
2022       Opc = PPC::CMPLW;
2023     } else if (ISD::isUnsignedIntSetCC(CC)) {
2024       if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm))
2025         return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
2026                                               getI32Imm(Imm & 0xFFFF, dl)), 0);
2027       Opc = PPC::CMPLW;
2028     } else {
2029       short SImm;
2030       if (isIntS16Immediate(RHS, SImm))
2031         return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
2032                                               getI32Imm((int)SImm & 0xFFFF,
2033                                                         dl)),
2034                          0);
2035       Opc = PPC::CMPW;
2036     }
2037   } else if (LHS.getValueType() == MVT::i64) {
2038     uint64_t Imm;
2039     if (CC == ISD::SETEQ || CC == ISD::SETNE) {
2040       if (isInt64Immediate(RHS.getNode(), Imm)) {
2041         // SETEQ/SETNE comparison with 16-bit immediate, fold it.
2042         if (isUInt<16>(Imm))
2043           return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
2044                                                 getI32Imm(Imm & 0xFFFF, dl)),
2045                          0);
2046         // If this is a 16-bit signed immediate, fold it.
2047         if (isInt<16>(Imm))
2048           return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
2049                                                 getI32Imm(Imm & 0xFFFF, dl)),
2050                          0);
2051 
2052         // For non-equality comparisons, the default code would materialize the
2053         // constant, then compare against it, like this:
2054         //   lis r2, 4660
2055         //   ori r2, r2, 22136
2056         //   cmpd cr0, r3, r2
2057         // Since we are just comparing for equality, we can emit this instead:
2058         //   xoris r0,r3,0x1234
2059         //   cmpldi cr0,r0,0x5678
2060         //   beq cr0,L6
2061         if (isUInt<32>(Imm)) {
2062           SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS,
2063                                              getI64Imm(Imm >> 16, dl)), 0);
2064           return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor,
2065                                                 getI64Imm(Imm & 0xFFFF, dl)),
2066                          0);
2067         }
2068       }
2069       Opc = PPC::CMPLD;
2070     } else if (ISD::isUnsignedIntSetCC(CC)) {
2071       if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm))
2072         return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
2073                                               getI64Imm(Imm & 0xFFFF, dl)), 0);
2074       Opc = PPC::CMPLD;
2075     } else {
2076       short SImm;
2077       if (isIntS16Immediate(RHS, SImm))
2078         return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
2079                                               getI64Imm(SImm & 0xFFFF, dl)),
2080                          0);
2081       Opc = PPC::CMPD;
2082     }
2083   } else if (LHS.getValueType() == MVT::f32) {
2084     Opc = PPC::FCMPUS;
2085   } else {
2086     assert(LHS.getValueType() == MVT::f64 && "Unknown vt!");
2087     Opc = PPCSubTarget->hasVSX() ? PPC::XSCMPUDP : PPC::FCMPUD;
2088   }
2089   return SDValue(CurDAG->getMachineNode(Opc, dl, MVT::i32, LHS, RHS), 0);
2090 }
2091 
getPredicateForSetCC(ISD::CondCode CC)2092 static PPC::Predicate getPredicateForSetCC(ISD::CondCode CC) {
2093   switch (CC) {
2094   case ISD::SETUEQ:
2095   case ISD::SETONE:
2096   case ISD::SETOLE:
2097   case ISD::SETOGE:
2098     llvm_unreachable("Should be lowered by legalize!");
2099   default: llvm_unreachable("Unknown condition!");
2100   case ISD::SETOEQ:
2101   case ISD::SETEQ:  return PPC::PRED_EQ;
2102   case ISD::SETUNE:
2103   case ISD::SETNE:  return PPC::PRED_NE;
2104   case ISD::SETOLT:
2105   case ISD::SETLT:  return PPC::PRED_LT;
2106   case ISD::SETULE:
2107   case ISD::SETLE:  return PPC::PRED_LE;
2108   case ISD::SETOGT:
2109   case ISD::SETGT:  return PPC::PRED_GT;
2110   case ISD::SETUGE:
2111   case ISD::SETGE:  return PPC::PRED_GE;
2112   case ISD::SETO:   return PPC::PRED_NU;
2113   case ISD::SETUO:  return PPC::PRED_UN;
2114     // These two are invalid for floating point.  Assume we have int.
2115   case ISD::SETULT: return PPC::PRED_LT;
2116   case ISD::SETUGT: return PPC::PRED_GT;
2117   }
2118 }
2119 
2120 /// getCRIdxForSetCC - Return the index of the condition register field
2121 /// associated with the SetCC condition, and whether or not the field is
2122 /// treated as inverted.  That is, lt = 0; ge = 0 inverted.
getCRIdxForSetCC(ISD::CondCode CC,bool & Invert)2123 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) {
2124   Invert = false;
2125   switch (CC) {
2126   default: llvm_unreachable("Unknown condition!");
2127   case ISD::SETOLT:
2128   case ISD::SETLT:  return 0;                  // Bit #0 = SETOLT
2129   case ISD::SETOGT:
2130   case ISD::SETGT:  return 1;                  // Bit #1 = SETOGT
2131   case ISD::SETOEQ:
2132   case ISD::SETEQ:  return 2;                  // Bit #2 = SETOEQ
2133   case ISD::SETUO:  return 3;                  // Bit #3 = SETUO
2134   case ISD::SETUGE:
2135   case ISD::SETGE:  Invert = true; return 0;   // !Bit #0 = SETUGE
2136   case ISD::SETULE:
2137   case ISD::SETLE:  Invert = true; return 1;   // !Bit #1 = SETULE
2138   case ISD::SETUNE:
2139   case ISD::SETNE:  Invert = true; return 2;   // !Bit #2 = SETUNE
2140   case ISD::SETO:   Invert = true; return 3;   // !Bit #3 = SETO
2141   case ISD::SETUEQ:
2142   case ISD::SETOGE:
2143   case ISD::SETOLE:
2144   case ISD::SETONE:
2145     llvm_unreachable("Invalid branch code: should be expanded by legalize");
2146   // These are invalid for floating point.  Assume integer.
2147   case ISD::SETULT: return 0;
2148   case ISD::SETUGT: return 1;
2149   }
2150 }
2151 
2152 // getVCmpInst: return the vector compare instruction for the specified
2153 // vector type and condition code. Since this is for altivec specific code,
2154 // only support the altivec types (v16i8, v8i16, v4i32, v2i64, and v4f32).
getVCmpInst(MVT VecVT,ISD::CondCode CC,bool HasVSX,bool & Swap,bool & Negate)2155 static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC,
2156                                 bool HasVSX, bool &Swap, bool &Negate) {
2157   Swap = false;
2158   Negate = false;
2159 
2160   if (VecVT.isFloatingPoint()) {
2161     /* Handle some cases by swapping input operands.  */
2162     switch (CC) {
2163       case ISD::SETLE: CC = ISD::SETGE; Swap = true; break;
2164       case ISD::SETLT: CC = ISD::SETGT; Swap = true; break;
2165       case ISD::SETOLE: CC = ISD::SETOGE; Swap = true; break;
2166       case ISD::SETOLT: CC = ISD::SETOGT; Swap = true; break;
2167       case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break;
2168       case ISD::SETUGT: CC = ISD::SETULT; Swap = true; break;
2169       default: break;
2170     }
2171     /* Handle some cases by negating the result.  */
2172     switch (CC) {
2173       case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break;
2174       case ISD::SETUNE: CC = ISD::SETOEQ; Negate = true; break;
2175       case ISD::SETULE: CC = ISD::SETOGT; Negate = true; break;
2176       case ISD::SETULT: CC = ISD::SETOGE; Negate = true; break;
2177       default: break;
2178     }
2179     /* We have instructions implementing the remaining cases.  */
2180     switch (CC) {
2181       case ISD::SETEQ:
2182       case ISD::SETOEQ:
2183         if (VecVT == MVT::v4f32)
2184           return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP;
2185         else if (VecVT == MVT::v2f64)
2186           return PPC::XVCMPEQDP;
2187         break;
2188       case ISD::SETGT:
2189       case ISD::SETOGT:
2190         if (VecVT == MVT::v4f32)
2191           return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP;
2192         else if (VecVT == MVT::v2f64)
2193           return PPC::XVCMPGTDP;
2194         break;
2195       case ISD::SETGE:
2196       case ISD::SETOGE:
2197         if (VecVT == MVT::v4f32)
2198           return HasVSX ? PPC::XVCMPGESP : PPC::VCMPGEFP;
2199         else if (VecVT == MVT::v2f64)
2200           return PPC::XVCMPGEDP;
2201         break;
2202       default:
2203         break;
2204     }
2205     llvm_unreachable("Invalid floating-point vector compare condition");
2206   } else {
2207     /* Handle some cases by swapping input operands.  */
2208     switch (CC) {
2209       case ISD::SETGE: CC = ISD::SETLE; Swap = true; break;
2210       case ISD::SETLT: CC = ISD::SETGT; Swap = true; break;
2211       case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break;
2212       case ISD::SETULT: CC = ISD::SETUGT; Swap = true; break;
2213       default: break;
2214     }
2215     /* Handle some cases by negating the result.  */
2216     switch (CC) {
2217       case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break;
2218       case ISD::SETUNE: CC = ISD::SETUEQ; Negate = true; break;
2219       case ISD::SETLE: CC = ISD::SETGT; Negate = true; break;
2220       case ISD::SETULE: CC = ISD::SETUGT; Negate = true; break;
2221       default: break;
2222     }
2223     /* We have instructions implementing the remaining cases.  */
2224     switch (CC) {
2225       case ISD::SETEQ:
2226       case ISD::SETUEQ:
2227         if (VecVT == MVT::v16i8)
2228           return PPC::VCMPEQUB;
2229         else if (VecVT == MVT::v8i16)
2230           return PPC::VCMPEQUH;
2231         else if (VecVT == MVT::v4i32)
2232           return PPC::VCMPEQUW;
2233         else if (VecVT == MVT::v2i64)
2234           return PPC::VCMPEQUD;
2235         break;
2236       case ISD::SETGT:
2237         if (VecVT == MVT::v16i8)
2238           return PPC::VCMPGTSB;
2239         else if (VecVT == MVT::v8i16)
2240           return PPC::VCMPGTSH;
2241         else if (VecVT == MVT::v4i32)
2242           return PPC::VCMPGTSW;
2243         else if (VecVT == MVT::v2i64)
2244           return PPC::VCMPGTSD;
2245         break;
2246       case ISD::SETUGT:
2247         if (VecVT == MVT::v16i8)
2248           return PPC::VCMPGTUB;
2249         else if (VecVT == MVT::v8i16)
2250           return PPC::VCMPGTUH;
2251         else if (VecVT == MVT::v4i32)
2252           return PPC::VCMPGTUW;
2253         else if (VecVT == MVT::v2i64)
2254           return PPC::VCMPGTUD;
2255         break;
2256       default:
2257         break;
2258     }
2259     llvm_unreachable("Invalid integer vector compare condition");
2260   }
2261 }
2262 
SelectSETCC(SDNode * N)2263 SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) {
2264   SDLoc dl(N);
2265   unsigned Imm;
2266   ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
2267   EVT PtrVT =
2268       CurDAG->getTargetLoweringInfo().getPointerTy(CurDAG->getDataLayout());
2269   bool isPPC64 = (PtrVT == MVT::i64);
2270 
2271   if (!PPCSubTarget->useCRBits() &&
2272       isInt32Immediate(N->getOperand(1), Imm)) {
2273     // We can codegen setcc op, imm very efficiently compared to a brcond.
2274     // Check for those cases here.
2275     // setcc op, 0
2276     if (Imm == 0) {
2277       SDValue Op = N->getOperand(0);
2278       switch (CC) {
2279       default: break;
2280       case ISD::SETEQ: {
2281         Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0);
2282         SDValue Ops[] = { Op, getI32Imm(27, dl), getI32Imm(5, dl),
2283                           getI32Imm(31, dl) };
2284         return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2285       }
2286       case ISD::SETNE: {
2287         if (isPPC64) break;
2288         SDValue AD =
2289           SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
2290                                          Op, getI32Imm(~0U, dl)), 0);
2291         return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
2292                                     AD.getValue(1));
2293       }
2294       case ISD::SETLT: {
2295         SDValue Ops[] = { Op, getI32Imm(1, dl), getI32Imm(31, dl),
2296                           getI32Imm(31, dl) };
2297         return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2298       }
2299       case ISD::SETGT: {
2300         SDValue T =
2301           SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0);
2302         T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0);
2303         SDValue Ops[] = { T, getI32Imm(1, dl), getI32Imm(31, dl),
2304                           getI32Imm(31, dl) };
2305         return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2306       }
2307       }
2308     } else if (Imm == ~0U) {        // setcc op, -1
2309       SDValue Op = N->getOperand(0);
2310       switch (CC) {
2311       default: break;
2312       case ISD::SETEQ:
2313         if (isPPC64) break;
2314         Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
2315                                             Op, getI32Imm(1, dl)), 0);
2316         return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
2317                               SDValue(CurDAG->getMachineNode(PPC::LI, dl,
2318                                                              MVT::i32,
2319                                                              getI32Imm(0, dl)),
2320                                       0), Op.getValue(1));
2321       case ISD::SETNE: {
2322         if (isPPC64) break;
2323         Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0);
2324         SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
2325                                             Op, getI32Imm(~0U, dl));
2326         return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0),
2327                                     Op, SDValue(AD, 1));
2328       }
2329       case ISD::SETLT: {
2330         SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op,
2331                                                     getI32Imm(1, dl)), 0);
2332         SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD,
2333                                                     Op), 0);
2334         SDValue Ops[] = { AN, getI32Imm(1, dl), getI32Imm(31, dl),
2335                           getI32Imm(31, dl) };
2336         return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2337       }
2338       case ISD::SETGT: {
2339         SDValue Ops[] = { Op, getI32Imm(1, dl), getI32Imm(31, dl),
2340                           getI32Imm(31, dl) };
2341         Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
2342         return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op,
2343                                     getI32Imm(1, dl));
2344       }
2345       }
2346     }
2347   }
2348 
2349   SDValue LHS = N->getOperand(0);
2350   SDValue RHS = N->getOperand(1);
2351 
2352   // Altivec Vector compare instructions do not set any CR register by default and
2353   // vector compare operations return the same type as the operands.
2354   if (LHS.getValueType().isVector()) {
2355     if (PPCSubTarget->hasQPX())
2356       return nullptr;
2357 
2358     EVT VecVT = LHS.getValueType();
2359     bool Swap, Negate;
2360     unsigned int VCmpInst = getVCmpInst(VecVT.getSimpleVT(), CC,
2361                                         PPCSubTarget->hasVSX(), Swap, Negate);
2362     if (Swap)
2363       std::swap(LHS, RHS);
2364 
2365     EVT ResVT = VecVT.changeVectorElementTypeToInteger();
2366     if (Negate) {
2367       SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, ResVT, LHS, RHS), 0);
2368       return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLNOR :
2369                                                               PPC::VNOR,
2370                                   ResVT, VCmp, VCmp);
2371     }
2372 
2373     return CurDAG->SelectNodeTo(N, VCmpInst, ResVT, LHS, RHS);
2374   }
2375 
2376   if (PPCSubTarget->useCRBits())
2377     return nullptr;
2378 
2379   bool Inv;
2380   unsigned Idx = getCRIdxForSetCC(CC, Inv);
2381   SDValue CCReg = SelectCC(LHS, RHS, CC, dl);
2382   SDValue IntCR;
2383 
2384   // Force the ccreg into CR7.
2385   SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
2386 
2387   SDValue InFlag(nullptr, 0);  // Null incoming flag value.
2388   CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg,
2389                                InFlag).getValue(1);
2390 
2391   IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg,
2392                                          CCReg), 0);
2393 
2394   SDValue Ops[] = { IntCR, getI32Imm((32 - (3 - Idx)) & 31, dl),
2395                       getI32Imm(31, dl), getI32Imm(31, dl) };
2396   if (!Inv)
2397     return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2398 
2399   // Get the specified bit.
2400   SDValue Tmp =
2401     SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
2402   return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1, dl));
2403 }
2404 
transferMemOperands(SDNode * N,SDNode * Result)2405 SDNode *PPCDAGToDAGISel::transferMemOperands(SDNode *N, SDNode *Result) {
2406   // Transfer memoperands.
2407   MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2408   MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
2409   cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1);
2410   return Result;
2411 }
2412 
2413 
2414 // Select - Convert the specified operand from a target-independent to a
2415 // target-specific node if it hasn't already been changed.
Select(SDNode * N)2416 SDNode *PPCDAGToDAGISel::Select(SDNode *N) {
2417   SDLoc dl(N);
2418   if (N->isMachineOpcode()) {
2419     N->setNodeId(-1);
2420     return nullptr;   // Already selected.
2421   }
2422 
2423   // In case any misguided DAG-level optimizations form an ADD with a
2424   // TargetConstant operand, crash here instead of miscompiling (by selecting
2425   // an r+r add instead of some kind of r+i add).
2426   if (N->getOpcode() == ISD::ADD &&
2427       N->getOperand(1).getOpcode() == ISD::TargetConstant)
2428     llvm_unreachable("Invalid ADD with TargetConstant operand");
2429 
2430   // Try matching complex bit permutations before doing anything else.
2431   if (SDNode *NN = SelectBitPermutation(N))
2432     return NN;
2433 
2434   switch (N->getOpcode()) {
2435   default: break;
2436 
2437   case ISD::Constant: {
2438     if (N->getValueType(0) == MVT::i64)
2439       return SelectInt64(CurDAG, N);
2440     break;
2441   }
2442 
2443   case ISD::SETCC: {
2444     SDNode *SN = SelectSETCC(N);
2445     if (SN)
2446       return SN;
2447     break;
2448   }
2449   case PPCISD::GlobalBaseReg:
2450     return getGlobalBaseReg();
2451 
2452   case ISD::FrameIndex:
2453     return getFrameIndex(N, N);
2454 
2455   case PPCISD::MFOCRF: {
2456     SDValue InFlag = N->getOperand(1);
2457     return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32,
2458                                   N->getOperand(0), InFlag);
2459   }
2460 
2461   case PPCISD::READ_TIME_BASE: {
2462     return CurDAG->getMachineNode(PPC::ReadTB, dl, MVT::i32, MVT::i32,
2463                                   MVT::Other, N->getOperand(0));
2464   }
2465 
2466   case PPCISD::SRA_ADDZE: {
2467     SDValue N0 = N->getOperand(0);
2468     SDValue ShiftAmt =
2469       CurDAG->getTargetConstant(*cast<ConstantSDNode>(N->getOperand(1))->
2470                                   getConstantIntValue(), dl,
2471                                   N->getValueType(0));
2472     if (N->getValueType(0) == MVT::i64) {
2473       SDNode *Op =
2474         CurDAG->getMachineNode(PPC::SRADI, dl, MVT::i64, MVT::Glue,
2475                                N0, ShiftAmt);
2476       return CurDAG->SelectNodeTo(N, PPC::ADDZE8, MVT::i64,
2477                                   SDValue(Op, 0), SDValue(Op, 1));
2478     } else {
2479       assert(N->getValueType(0) == MVT::i32 &&
2480              "Expecting i64 or i32 in PPCISD::SRA_ADDZE");
2481       SDNode *Op =
2482         CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
2483                                N0, ShiftAmt);
2484       return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
2485                                   SDValue(Op, 0), SDValue(Op, 1));
2486     }
2487   }
2488 
2489   case ISD::LOAD: {
2490     // Handle preincrement loads.
2491     LoadSDNode *LD = cast<LoadSDNode>(N);
2492     EVT LoadedVT = LD->getMemoryVT();
2493 
2494     // Normal loads are handled by code generated from the .td file.
2495     if (LD->getAddressingMode() != ISD::PRE_INC)
2496       break;
2497 
2498     SDValue Offset = LD->getOffset();
2499     if (Offset.getOpcode() == ISD::TargetConstant ||
2500         Offset.getOpcode() == ISD::TargetGlobalAddress) {
2501 
2502       unsigned Opcode;
2503       bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
2504       if (LD->getValueType(0) != MVT::i64) {
2505         // Handle PPC32 integer and normal FP loads.
2506         assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
2507         switch (LoadedVT.getSimpleVT().SimpleTy) {
2508           default: llvm_unreachable("Invalid PPC load type!");
2509           case MVT::f64: Opcode = PPC::LFDU; break;
2510           case MVT::f32: Opcode = PPC::LFSU; break;
2511           case MVT::i32: Opcode = PPC::LWZU; break;
2512           case MVT::i16: Opcode = isSExt ? PPC::LHAU : PPC::LHZU; break;
2513           case MVT::i1:
2514           case MVT::i8:  Opcode = PPC::LBZU; break;
2515         }
2516       } else {
2517         assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
2518         assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
2519         switch (LoadedVT.getSimpleVT().SimpleTy) {
2520           default: llvm_unreachable("Invalid PPC load type!");
2521           case MVT::i64: Opcode = PPC::LDU; break;
2522           case MVT::i32: Opcode = PPC::LWZU8; break;
2523           case MVT::i16: Opcode = isSExt ? PPC::LHAU8 : PPC::LHZU8; break;
2524           case MVT::i1:
2525           case MVT::i8:  Opcode = PPC::LBZU8; break;
2526         }
2527       }
2528 
2529       SDValue Chain = LD->getChain();
2530       SDValue Base = LD->getBasePtr();
2531       SDValue Ops[] = { Offset, Base, Chain };
2532       return transferMemOperands(
2533           N, CurDAG->getMachineNode(
2534                  Opcode, dl, LD->getValueType(0),
2535                  PPCLowering->getPointerTy(CurDAG->getDataLayout()), MVT::Other,
2536                  Ops));
2537     } else {
2538       unsigned Opcode;
2539       bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
2540       if (LD->getValueType(0) != MVT::i64) {
2541         // Handle PPC32 integer and normal FP loads.
2542         assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
2543         switch (LoadedVT.getSimpleVT().SimpleTy) {
2544           default: llvm_unreachable("Invalid PPC load type!");
2545           case MVT::v4f64: Opcode = PPC::QVLFDUX; break; // QPX
2546           case MVT::v4f32: Opcode = PPC::QVLFSUX; break; // QPX
2547           case MVT::f64: Opcode = PPC::LFDUX; break;
2548           case MVT::f32: Opcode = PPC::LFSUX; break;
2549           case MVT::i32: Opcode = PPC::LWZUX; break;
2550           case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break;
2551           case MVT::i1:
2552           case MVT::i8:  Opcode = PPC::LBZUX; break;
2553         }
2554       } else {
2555         assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
2556         assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) &&
2557                "Invalid sext update load");
2558         switch (LoadedVT.getSimpleVT().SimpleTy) {
2559           default: llvm_unreachable("Invalid PPC load type!");
2560           case MVT::i64: Opcode = PPC::LDUX; break;
2561           case MVT::i32: Opcode = isSExt ? PPC::LWAUX  : PPC::LWZUX8; break;
2562           case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break;
2563           case MVT::i1:
2564           case MVT::i8:  Opcode = PPC::LBZUX8; break;
2565         }
2566       }
2567 
2568       SDValue Chain = LD->getChain();
2569       SDValue Base = LD->getBasePtr();
2570       SDValue Ops[] = { Base, Offset, Chain };
2571       return transferMemOperands(
2572           N, CurDAG->getMachineNode(
2573                  Opcode, dl, LD->getValueType(0),
2574                  PPCLowering->getPointerTy(CurDAG->getDataLayout()), MVT::Other,
2575                  Ops));
2576     }
2577   }
2578 
2579   case ISD::AND: {
2580     unsigned Imm, Imm2, SH, MB, ME;
2581     uint64_t Imm64;
2582 
2583     // If this is an and of a value rotated between 0 and 31 bits and then and'd
2584     // with a mask, emit rlwinm
2585     if (isInt32Immediate(N->getOperand(1), Imm) &&
2586         isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) {
2587       SDValue Val = N->getOperand(0).getOperand(0);
2588       SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl),
2589                         getI32Imm(ME, dl) };
2590       return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2591     }
2592     // If this is just a masked value where the input is not handled above, and
2593     // is not a rotate-left (handled by a pattern in the .td file), emit rlwinm
2594     if (isInt32Immediate(N->getOperand(1), Imm) &&
2595         isRunOfOnes(Imm, MB, ME) &&
2596         N->getOperand(0).getOpcode() != ISD::ROTL) {
2597       SDValue Val = N->getOperand(0);
2598       SDValue Ops[] = { Val, getI32Imm(0, dl), getI32Imm(MB, dl),
2599                         getI32Imm(ME, dl) };
2600       return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2601     }
2602     // If this is a 64-bit zero-extension mask, emit rldicl.
2603     if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) &&
2604         isMask_64(Imm64)) {
2605       SDValue Val = N->getOperand(0);
2606       MB = 64 - countTrailingOnes(Imm64);
2607       SH = 0;
2608 
2609       // If the operand is a logical right shift, we can fold it into this
2610       // instruction: rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb)
2611       // for n <= mb. The right shift is really a left rotate followed by a
2612       // mask, and this mask is a more-restrictive sub-mask of the mask implied
2613       // by the shift.
2614       if (Val.getOpcode() == ISD::SRL &&
2615           isInt32Immediate(Val.getOperand(1).getNode(), Imm) && Imm <= MB) {
2616         assert(Imm < 64 && "Illegal shift amount");
2617         Val = Val.getOperand(0);
2618         SH = 64 - Imm;
2619       }
2620 
2621       SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl) };
2622       return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops);
2623     }
2624     // AND X, 0 -> 0, not "rlwinm 32".
2625     if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) {
2626       ReplaceUses(SDValue(N, 0), N->getOperand(1));
2627       return nullptr;
2628     }
2629     // ISD::OR doesn't get all the bitfield insertion fun.
2630     // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) might be a
2631     // bitfield insert.
2632     if (isInt32Immediate(N->getOperand(1), Imm) &&
2633         N->getOperand(0).getOpcode() == ISD::OR &&
2634         isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) {
2635       // The idea here is to check whether this is equivalent to:
2636       //   (c1 & m) | (x & ~m)
2637       // where m is a run-of-ones mask. The logic here is that, for each bit in
2638       // c1 and c2:
2639       //  - if both are 1, then the output will be 1.
2640       //  - if both are 0, then the output will be 0.
2641       //  - if the bit in c1 is 0, and the bit in c2 is 1, then the output will
2642       //    come from x.
2643       //  - if the bit in c1 is 1, and the bit in c2 is 0, then the output will
2644       //    be 0.
2645       //  If that last condition is never the case, then we can form m from the
2646       //  bits that are the same between c1 and c2.
2647       unsigned MB, ME;
2648       if (isRunOfOnes(~(Imm^Imm2), MB, ME) && !(~Imm & Imm2)) {
2649         SDValue Ops[] = { N->getOperand(0).getOperand(0),
2650                             N->getOperand(0).getOperand(1),
2651                             getI32Imm(0, dl), getI32Imm(MB, dl),
2652                             getI32Imm(ME, dl) };
2653         return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
2654       }
2655     }
2656 
2657     // Other cases are autogenerated.
2658     break;
2659   }
2660   case ISD::OR: {
2661     if (N->getValueType(0) == MVT::i32)
2662       if (SDNode *I = SelectBitfieldInsert(N))
2663         return I;
2664 
2665     short Imm;
2666     if (N->getOperand(0)->getOpcode() == ISD::FrameIndex &&
2667         isIntS16Immediate(N->getOperand(1), Imm)) {
2668       APInt LHSKnownZero, LHSKnownOne;
2669       CurDAG->computeKnownBits(N->getOperand(0), LHSKnownZero, LHSKnownOne);
2670 
2671       // If this is equivalent to an add, then we can fold it with the
2672       // FrameIndex calculation.
2673       if ((LHSKnownZero.getZExtValue()|~(uint64_t)Imm) == ~0ULL)
2674         return getFrameIndex(N, N->getOperand(0).getNode(), (int)Imm);
2675     }
2676 
2677     // Other cases are autogenerated.
2678     break;
2679   }
2680   case ISD::ADD: {
2681     short Imm;
2682     if (N->getOperand(0)->getOpcode() == ISD::FrameIndex &&
2683         isIntS16Immediate(N->getOperand(1), Imm))
2684       return getFrameIndex(N, N->getOperand(0).getNode(), (int)Imm);
2685 
2686     break;
2687   }
2688   case ISD::SHL: {
2689     unsigned Imm, SH, MB, ME;
2690     if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
2691         isRotateAndMask(N, Imm, true, SH, MB, ME)) {
2692       SDValue Ops[] = { N->getOperand(0).getOperand(0),
2693                           getI32Imm(SH, dl), getI32Imm(MB, dl),
2694                           getI32Imm(ME, dl) };
2695       return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2696     }
2697 
2698     // Other cases are autogenerated.
2699     break;
2700   }
2701   case ISD::SRL: {
2702     unsigned Imm, SH, MB, ME;
2703     if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
2704         isRotateAndMask(N, Imm, true, SH, MB, ME)) {
2705       SDValue Ops[] = { N->getOperand(0).getOperand(0),
2706                           getI32Imm(SH, dl), getI32Imm(MB, dl),
2707                           getI32Imm(ME, dl) };
2708       return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2709     }
2710 
2711     // Other cases are autogenerated.
2712     break;
2713   }
2714   // FIXME: Remove this once the ANDI glue bug is fixed:
2715   case PPCISD::ANDIo_1_EQ_BIT:
2716   case PPCISD::ANDIo_1_GT_BIT: {
2717     if (!ANDIGlueBug)
2718       break;
2719 
2720     EVT InVT = N->getOperand(0).getValueType();
2721     assert((InVT == MVT::i64 || InVT == MVT::i32) &&
2722            "Invalid input type for ANDIo_1_EQ_BIT");
2723 
2724     unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDIo8 : PPC::ANDIo;
2725     SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue,
2726                                         N->getOperand(0),
2727                                         CurDAG->getTargetConstant(1, dl, InVT)),
2728                  0);
2729     SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32);
2730     SDValue SRIdxVal =
2731       CurDAG->getTargetConstant(N->getOpcode() == PPCISD::ANDIo_1_EQ_BIT ?
2732                                 PPC::sub_eq : PPC::sub_gt, dl, MVT::i32);
2733 
2734     return CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1,
2735                                 CR0Reg, SRIdxVal,
2736                                 SDValue(AndI.getNode(), 1) /* glue */);
2737   }
2738   case ISD::SELECT_CC: {
2739     ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
2740     EVT PtrVT =
2741         CurDAG->getTargetLoweringInfo().getPointerTy(CurDAG->getDataLayout());
2742     bool isPPC64 = (PtrVT == MVT::i64);
2743 
2744     // If this is a select of i1 operands, we'll pattern match it.
2745     if (PPCSubTarget->useCRBits() &&
2746         N->getOperand(0).getValueType() == MVT::i1)
2747       break;
2748 
2749     // Handle the setcc cases here.  select_cc lhs, 0, 1, 0, cc
2750     if (!isPPC64)
2751       if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2752         if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
2753           if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
2754             if (N1C->isNullValue() && N3C->isNullValue() &&
2755                 N2C->getZExtValue() == 1ULL && CC == ISD::SETNE &&
2756                 // FIXME: Implement this optzn for PPC64.
2757                 N->getValueType(0) == MVT::i32) {
2758               SDNode *Tmp =
2759                 CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
2760                                        N->getOperand(0), getI32Imm(~0U, dl));
2761               return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32,
2762                                           SDValue(Tmp, 0), N->getOperand(0),
2763                                           SDValue(Tmp, 1));
2764             }
2765 
2766     SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl);
2767 
2768     if (N->getValueType(0) == MVT::i1) {
2769       // An i1 select is: (c & t) | (!c & f).
2770       bool Inv;
2771       unsigned Idx = getCRIdxForSetCC(CC, Inv);
2772 
2773       unsigned SRI;
2774       switch (Idx) {
2775       default: llvm_unreachable("Invalid CC index");
2776       case 0: SRI = PPC::sub_lt; break;
2777       case 1: SRI = PPC::sub_gt; break;
2778       case 2: SRI = PPC::sub_eq; break;
2779       case 3: SRI = PPC::sub_un; break;
2780       }
2781 
2782       SDValue CCBit = CurDAG->getTargetExtractSubreg(SRI, dl, MVT::i1, CCReg);
2783 
2784       SDValue NotCCBit(CurDAG->getMachineNode(PPC::CRNOR, dl, MVT::i1,
2785                                               CCBit, CCBit), 0);
2786       SDValue C =    Inv ? NotCCBit : CCBit,
2787               NotC = Inv ? CCBit    : NotCCBit;
2788 
2789       SDValue CAndT(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1,
2790                                            C, N->getOperand(2)), 0);
2791       SDValue NotCAndF(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1,
2792                                               NotC, N->getOperand(3)), 0);
2793 
2794       return CurDAG->SelectNodeTo(N, PPC::CROR, MVT::i1, CAndT, NotCAndF);
2795     }
2796 
2797     unsigned BROpc = getPredicateForSetCC(CC);
2798 
2799     unsigned SelectCCOp;
2800     if (N->getValueType(0) == MVT::i32)
2801       SelectCCOp = PPC::SELECT_CC_I4;
2802     else if (N->getValueType(0) == MVT::i64)
2803       SelectCCOp = PPC::SELECT_CC_I8;
2804     else if (N->getValueType(0) == MVT::f32)
2805       if (PPCSubTarget->hasP8Vector())
2806         SelectCCOp = PPC::SELECT_CC_VSSRC;
2807       else
2808         SelectCCOp = PPC::SELECT_CC_F4;
2809     else if (N->getValueType(0) == MVT::f64)
2810       if (PPCSubTarget->hasVSX())
2811         SelectCCOp = PPC::SELECT_CC_VSFRC;
2812       else
2813         SelectCCOp = PPC::SELECT_CC_F8;
2814     else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4f64)
2815       SelectCCOp = PPC::SELECT_CC_QFRC;
2816     else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4f32)
2817       SelectCCOp = PPC::SELECT_CC_QSRC;
2818     else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4i1)
2819       SelectCCOp = PPC::SELECT_CC_QBRC;
2820     else if (N->getValueType(0) == MVT::v2f64 ||
2821              N->getValueType(0) == MVT::v2i64)
2822       SelectCCOp = PPC::SELECT_CC_VSRC;
2823     else
2824       SelectCCOp = PPC::SELECT_CC_VRRC;
2825 
2826     SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3),
2827                         getI32Imm(BROpc, dl) };
2828     return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops);
2829   }
2830   case ISD::VSELECT:
2831     if (PPCSubTarget->hasVSX()) {
2832       SDValue Ops[] = { N->getOperand(2), N->getOperand(1), N->getOperand(0) };
2833       return CurDAG->SelectNodeTo(N, PPC::XXSEL, N->getValueType(0), Ops);
2834     }
2835 
2836     break;
2837   case ISD::VECTOR_SHUFFLE:
2838     if (PPCSubTarget->hasVSX() && (N->getValueType(0) == MVT::v2f64 ||
2839                                   N->getValueType(0) == MVT::v2i64)) {
2840       ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
2841 
2842       SDValue Op1 = N->getOperand(SVN->getMaskElt(0) < 2 ? 0 : 1),
2843               Op2 = N->getOperand(SVN->getMaskElt(1) < 2 ? 0 : 1);
2844       unsigned DM[2];
2845 
2846       for (int i = 0; i < 2; ++i)
2847         if (SVN->getMaskElt(i) <= 0 || SVN->getMaskElt(i) == 2)
2848           DM[i] = 0;
2849         else
2850           DM[i] = 1;
2851 
2852       if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 &&
2853           Op1.getOpcode() == ISD::SCALAR_TO_VECTOR &&
2854           isa<LoadSDNode>(Op1.getOperand(0))) {
2855         LoadSDNode *LD = cast<LoadSDNode>(Op1.getOperand(0));
2856         SDValue Base, Offset;
2857 
2858         if (LD->isUnindexed() && LD->hasOneUse() && Op1.hasOneUse() &&
2859             (LD->getMemoryVT() == MVT::f64 ||
2860              LD->getMemoryVT() == MVT::i64) &&
2861             SelectAddrIdxOnly(LD->getBasePtr(), Base, Offset)) {
2862           SDValue Chain = LD->getChain();
2863           SDValue Ops[] = { Base, Offset, Chain };
2864           return CurDAG->SelectNodeTo(N, PPC::LXVDSX,
2865                                       N->getValueType(0), Ops);
2866         }
2867       }
2868 
2869       // For little endian, we must swap the input operands and adjust
2870       // the mask elements (reverse and invert them).
2871       if (PPCSubTarget->isLittleEndian()) {
2872         std::swap(Op1, Op2);
2873         unsigned tmp = DM[0];
2874         DM[0] = 1 - DM[1];
2875         DM[1] = 1 - tmp;
2876       }
2877 
2878       SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), dl,
2879                                               MVT::i32);
2880       SDValue Ops[] = { Op1, Op2, DMV };
2881       return CurDAG->SelectNodeTo(N, PPC::XXPERMDI, N->getValueType(0), Ops);
2882     }
2883 
2884     break;
2885   case PPCISD::BDNZ:
2886   case PPCISD::BDZ: {
2887     bool IsPPC64 = PPCSubTarget->isPPC64();
2888     SDValue Ops[] = { N->getOperand(1), N->getOperand(0) };
2889     return CurDAG->SelectNodeTo(N, N->getOpcode() == PPCISD::BDNZ ?
2890                                    (IsPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
2891                                    (IsPPC64 ? PPC::BDZ8 : PPC::BDZ),
2892                                 MVT::Other, Ops);
2893   }
2894   case PPCISD::COND_BRANCH: {
2895     // Op #0 is the Chain.
2896     // Op #1 is the PPC::PRED_* number.
2897     // Op #2 is the CR#
2898     // Op #3 is the Dest MBB
2899     // Op #4 is the Flag.
2900     // Prevent PPC::PRED_* from being selected into LI.
2901     unsigned PCC = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2902     if (EnableBranchHint)
2903       PCC |= getBranchHint(PCC, FuncInfo, N->getOperand(3));
2904 
2905     SDValue Pred = getI32Imm(PCC, dl);
2906     SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3),
2907       N->getOperand(0), N->getOperand(4) };
2908     return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
2909   }
2910   case ISD::BR_CC: {
2911     ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
2912     unsigned PCC = getPredicateForSetCC(CC);
2913 
2914     if (N->getOperand(2).getValueType() == MVT::i1) {
2915       unsigned Opc;
2916       bool Swap;
2917       switch (PCC) {
2918       default: llvm_unreachable("Unexpected Boolean-operand predicate");
2919       case PPC::PRED_LT: Opc = PPC::CRANDC; Swap = true;  break;
2920       case PPC::PRED_LE: Opc = PPC::CRORC;  Swap = true;  break;
2921       case PPC::PRED_EQ: Opc = PPC::CREQV;  Swap = false; break;
2922       case PPC::PRED_GE: Opc = PPC::CRORC;  Swap = false; break;
2923       case PPC::PRED_GT: Opc = PPC::CRANDC; Swap = false; break;
2924       case PPC::PRED_NE: Opc = PPC::CRXOR;  Swap = false; break;
2925       }
2926 
2927       SDValue BitComp(CurDAG->getMachineNode(Opc, dl, MVT::i1,
2928                                              N->getOperand(Swap ? 3 : 2),
2929                                              N->getOperand(Swap ? 2 : 3)), 0);
2930       return CurDAG->SelectNodeTo(N, PPC::BC, MVT::Other,
2931                                   BitComp, N->getOperand(4), N->getOperand(0));
2932     }
2933 
2934     if (EnableBranchHint)
2935       PCC |= getBranchHint(PCC, FuncInfo, N->getOperand(4));
2936 
2937     SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl);
2938     SDValue Ops[] = { getI32Imm(PCC, dl), CondCode,
2939                         N->getOperand(4), N->getOperand(0) };
2940     return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
2941   }
2942   case ISD::BRIND: {
2943     // FIXME: Should custom lower this.
2944     SDValue Chain = N->getOperand(0);
2945     SDValue Target = N->getOperand(1);
2946     unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8;
2947     unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8;
2948     Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Target,
2949                                            Chain), 0);
2950     return CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain);
2951   }
2952   case PPCISD::TOC_ENTRY: {
2953     assert ((PPCSubTarget->isPPC64() || PPCSubTarget->isSVR4ABI()) &&
2954             "Only supported for 64-bit ABI and 32-bit SVR4");
2955     if (PPCSubTarget->isSVR4ABI() && !PPCSubTarget->isPPC64()) {
2956       SDValue GA = N->getOperand(0);
2957       return transferMemOperands(N, CurDAG->getMachineNode(PPC::LWZtoc, dl,
2958                                       MVT::i32, GA, N->getOperand(1)));
2959     }
2960 
2961     // For medium and large code model, we generate two instructions as
2962     // described below.  Otherwise we allow SelectCodeCommon to handle this,
2963     // selecting one of LDtoc, LDtocJTI, LDtocCPT, and LDtocBA.
2964     CodeModel::Model CModel = TM.getCodeModel();
2965     if (CModel != CodeModel::Medium && CModel != CodeModel::Large)
2966       break;
2967 
2968     // The first source operand is a TargetGlobalAddress or a TargetJumpTable.
2969     // If it must be toc-referenced according to PPCSubTarget, we generate:
2970     //   LDtocL(<ga:@sym>, ADDIStocHA(%X2, <ga:@sym>))
2971     // Otherwise we generate:
2972     //   ADDItocL(ADDIStocHA(%X2, <ga:@sym>), <ga:@sym>)
2973     SDValue GA = N->getOperand(0);
2974     SDValue TOCbase = N->getOperand(1);
2975     SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64,
2976                                          TOCbase, GA);
2977 
2978     if (isa<JumpTableSDNode>(GA) || isa<BlockAddressSDNode>(GA) ||
2979         CModel == CodeModel::Large)
2980       return transferMemOperands(N, CurDAG->getMachineNode(PPC::LDtocL, dl,
2981                                       MVT::i64, GA, SDValue(Tmp, 0)));
2982 
2983     if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) {
2984       const GlobalValue *GV = G->getGlobal();
2985       unsigned char GVFlags = PPCSubTarget->classifyGlobalReference(GV);
2986       if (GVFlags & PPCII::MO_NLP_FLAG) {
2987         return transferMemOperands(N, CurDAG->getMachineNode(PPC::LDtocL, dl,
2988                                         MVT::i64, GA, SDValue(Tmp, 0)));
2989       }
2990     }
2991 
2992     return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64,
2993                                   SDValue(Tmp, 0), GA);
2994   }
2995   case PPCISD::PPC32_PICGOT: {
2996     // Generate a PIC-safe GOT reference.
2997     assert(!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI() &&
2998       "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4");
2999     return CurDAG->SelectNodeTo(
3000         N, PPC::PPC32PICGOT, PPCLowering->getPointerTy(CurDAG->getDataLayout()),
3001         MVT::i32);
3002   }
3003   case PPCISD::VADD_SPLAT: {
3004     // This expands into one of three sequences, depending on whether
3005     // the first operand is odd or even, positive or negative.
3006     assert(isa<ConstantSDNode>(N->getOperand(0)) &&
3007            isa<ConstantSDNode>(N->getOperand(1)) &&
3008            "Invalid operand on VADD_SPLAT!");
3009 
3010     int Elt     = N->getConstantOperandVal(0);
3011     int EltSize = N->getConstantOperandVal(1);
3012     unsigned Opc1, Opc2, Opc3;
3013     EVT VT;
3014 
3015     if (EltSize == 1) {
3016       Opc1 = PPC::VSPLTISB;
3017       Opc2 = PPC::VADDUBM;
3018       Opc3 = PPC::VSUBUBM;
3019       VT = MVT::v16i8;
3020     } else if (EltSize == 2) {
3021       Opc1 = PPC::VSPLTISH;
3022       Opc2 = PPC::VADDUHM;
3023       Opc3 = PPC::VSUBUHM;
3024       VT = MVT::v8i16;
3025     } else {
3026       assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!");
3027       Opc1 = PPC::VSPLTISW;
3028       Opc2 = PPC::VADDUWM;
3029       Opc3 = PPC::VSUBUWM;
3030       VT = MVT::v4i32;
3031     }
3032 
3033     if ((Elt & 1) == 0) {
3034       // Elt is even, in the range [-32,-18] + [16,30].
3035       //
3036       // Convert: VADD_SPLAT elt, size
3037       // Into:    tmp = VSPLTIS[BHW] elt
3038       //          VADDU[BHW]M tmp, tmp
3039       // Where:   [BHW] = B for size = 1, H for size = 2, W for size = 4
3040       SDValue EltVal = getI32Imm(Elt >> 1, dl);
3041       SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
3042       SDValue TmpVal = SDValue(Tmp, 0);
3043       return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal);
3044 
3045     } else if (Elt > 0) {
3046       // Elt is odd and positive, in the range [17,31].
3047       //
3048       // Convert: VADD_SPLAT elt, size
3049       // Into:    tmp1 = VSPLTIS[BHW] elt-16
3050       //          tmp2 = VSPLTIS[BHW] -16
3051       //          VSUBU[BHW]M tmp1, tmp2
3052       SDValue EltVal = getI32Imm(Elt - 16, dl);
3053       SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
3054       EltVal = getI32Imm(-16, dl);
3055       SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
3056       return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0),
3057                                     SDValue(Tmp2, 0));
3058 
3059     } else {
3060       // Elt is odd and negative, in the range [-31,-17].
3061       //
3062       // Convert: VADD_SPLAT elt, size
3063       // Into:    tmp1 = VSPLTIS[BHW] elt+16
3064       //          tmp2 = VSPLTIS[BHW] -16
3065       //          VADDU[BHW]M tmp1, tmp2
3066       SDValue EltVal = getI32Imm(Elt + 16, dl);
3067       SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
3068       EltVal = getI32Imm(-16, dl);
3069       SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
3070       return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0),
3071                                     SDValue(Tmp2, 0));
3072     }
3073   }
3074   }
3075 
3076   return SelectCode(N);
3077 }
3078 
3079 // If the target supports the cmpb instruction, do the idiom recognition here.
3080 // We don't do this as a DAG combine because we don't want to do it as nodes
3081 // are being combined (because we might miss part of the eventual idiom). We
3082 // don't want to do it during instruction selection because we want to reuse
3083 // the logic for lowering the masking operations already part of the
3084 // instruction selector.
combineToCMPB(SDNode * N)3085 SDValue PPCDAGToDAGISel::combineToCMPB(SDNode *N) {
3086   SDLoc dl(N);
3087 
3088   assert(N->getOpcode() == ISD::OR &&
3089          "Only OR nodes are supported for CMPB");
3090 
3091   SDValue Res;
3092   if (!PPCSubTarget->hasCMPB())
3093     return Res;
3094 
3095   if (N->getValueType(0) != MVT::i32 &&
3096       N->getValueType(0) != MVT::i64)
3097     return Res;
3098 
3099   EVT VT = N->getValueType(0);
3100 
3101   SDValue RHS, LHS;
3102   bool BytesFound[8] = { 0, 0, 0, 0, 0, 0, 0, 0 };
3103   uint64_t Mask = 0, Alt = 0;
3104 
3105   auto IsByteSelectCC = [this](SDValue O, unsigned &b,
3106                                uint64_t &Mask, uint64_t &Alt,
3107                                SDValue &LHS, SDValue &RHS) {
3108     if (O.getOpcode() != ISD::SELECT_CC)
3109       return false;
3110     ISD::CondCode CC = cast<CondCodeSDNode>(O.getOperand(4))->get();
3111 
3112     if (!isa<ConstantSDNode>(O.getOperand(2)) ||
3113         !isa<ConstantSDNode>(O.getOperand(3)))
3114       return false;
3115 
3116     uint64_t PM = O.getConstantOperandVal(2);
3117     uint64_t PAlt = O.getConstantOperandVal(3);
3118     for (b = 0; b < 8; ++b) {
3119       uint64_t Mask = UINT64_C(0xFF) << (8*b);
3120       if (PM && (PM & Mask) == PM && (PAlt & Mask) == PAlt)
3121         break;
3122     }
3123 
3124     if (b == 8)
3125       return false;
3126     Mask |= PM;
3127     Alt  |= PAlt;
3128 
3129     if (!isa<ConstantSDNode>(O.getOperand(1)) ||
3130         O.getConstantOperandVal(1) != 0) {
3131       SDValue Op0 = O.getOperand(0), Op1 = O.getOperand(1);
3132       if (Op0.getOpcode() == ISD::TRUNCATE)
3133         Op0 = Op0.getOperand(0);
3134       if (Op1.getOpcode() == ISD::TRUNCATE)
3135         Op1 = Op1.getOperand(0);
3136 
3137       if (Op0.getOpcode() == ISD::SRL && Op1.getOpcode() == ISD::SRL &&
3138           Op0.getOperand(1) == Op1.getOperand(1) && CC == ISD::SETEQ &&
3139           isa<ConstantSDNode>(Op0.getOperand(1))) {
3140 
3141         unsigned Bits = Op0.getValueType().getSizeInBits();
3142         if (b != Bits/8-1)
3143           return false;
3144         if (Op0.getConstantOperandVal(1) != Bits-8)
3145           return false;
3146 
3147         LHS = Op0.getOperand(0);
3148         RHS = Op1.getOperand(0);
3149         return true;
3150       }
3151 
3152       // When we have small integers (i16 to be specific), the form present
3153       // post-legalization uses SETULT in the SELECT_CC for the
3154       // higher-order byte, depending on the fact that the
3155       // even-higher-order bytes are known to all be zero, for example:
3156       //   select_cc (xor $lhs, $rhs), 256, 65280, 0, setult
3157       // (so when the second byte is the same, because all higher-order
3158       // bits from bytes 3 and 4 are known to be zero, the result of the
3159       // xor can be at most 255)
3160       if (Op0.getOpcode() == ISD::XOR && CC == ISD::SETULT &&
3161           isa<ConstantSDNode>(O.getOperand(1))) {
3162 
3163         uint64_t ULim = O.getConstantOperandVal(1);
3164         if (ULim != (UINT64_C(1) << b*8))
3165           return false;
3166 
3167         // Now we need to make sure that the upper bytes are known to be
3168         // zero.
3169         unsigned Bits = Op0.getValueType().getSizeInBits();
3170         if (!CurDAG->MaskedValueIsZero(Op0,
3171               APInt::getHighBitsSet(Bits, Bits - (b+1)*8)))
3172           return false;
3173 
3174         LHS = Op0.getOperand(0);
3175         RHS = Op0.getOperand(1);
3176         return true;
3177       }
3178 
3179       return false;
3180     }
3181 
3182     if (CC != ISD::SETEQ)
3183       return false;
3184 
3185     SDValue Op = O.getOperand(0);
3186     if (Op.getOpcode() == ISD::AND) {
3187       if (!isa<ConstantSDNode>(Op.getOperand(1)))
3188         return false;
3189       if (Op.getConstantOperandVal(1) != (UINT64_C(0xFF) << (8*b)))
3190         return false;
3191 
3192       SDValue XOR = Op.getOperand(0);
3193       if (XOR.getOpcode() == ISD::TRUNCATE)
3194         XOR = XOR.getOperand(0);
3195       if (XOR.getOpcode() != ISD::XOR)
3196         return false;
3197 
3198       LHS = XOR.getOperand(0);
3199       RHS = XOR.getOperand(1);
3200       return true;
3201     } else if (Op.getOpcode() == ISD::SRL) {
3202       if (!isa<ConstantSDNode>(Op.getOperand(1)))
3203         return false;
3204       unsigned Bits = Op.getValueType().getSizeInBits();
3205       if (b != Bits/8-1)
3206         return false;
3207       if (Op.getConstantOperandVal(1) != Bits-8)
3208         return false;
3209 
3210       SDValue XOR = Op.getOperand(0);
3211       if (XOR.getOpcode() == ISD::TRUNCATE)
3212         XOR = XOR.getOperand(0);
3213       if (XOR.getOpcode() != ISD::XOR)
3214         return false;
3215 
3216       LHS = XOR.getOperand(0);
3217       RHS = XOR.getOperand(1);
3218       return true;
3219     }
3220 
3221     return false;
3222   };
3223 
3224   SmallVector<SDValue, 8> Queue(1, SDValue(N, 0));
3225   while (!Queue.empty()) {
3226     SDValue V = Queue.pop_back_val();
3227 
3228     for (const SDValue &O : V.getNode()->ops()) {
3229       unsigned b;
3230       uint64_t M = 0, A = 0;
3231       SDValue OLHS, ORHS;
3232       if (O.getOpcode() == ISD::OR) {
3233         Queue.push_back(O);
3234       } else if (IsByteSelectCC(O, b, M, A, OLHS, ORHS)) {
3235         if (!LHS) {
3236           LHS = OLHS;
3237           RHS = ORHS;
3238           BytesFound[b] = true;
3239           Mask |= M;
3240           Alt  |= A;
3241         } else if ((LHS == ORHS && RHS == OLHS) ||
3242                    (RHS == ORHS && LHS == OLHS)) {
3243           BytesFound[b] = true;
3244           Mask |= M;
3245           Alt  |= A;
3246         } else {
3247           return Res;
3248         }
3249       } else {
3250         return Res;
3251       }
3252     }
3253   }
3254 
3255   unsigned LastB = 0, BCnt = 0;
3256   for (unsigned i = 0; i < 8; ++i)
3257     if (BytesFound[LastB]) {
3258       ++BCnt;
3259       LastB = i;
3260     }
3261 
3262   if (!LastB || BCnt < 2)
3263     return Res;
3264 
3265   // Because we'll be zero-extending the output anyway if don't have a specific
3266   // value for each input byte (via the Mask), we can 'anyext' the inputs.
3267   if (LHS.getValueType() != VT) {
3268     LHS = CurDAG->getAnyExtOrTrunc(LHS, dl, VT);
3269     RHS = CurDAG->getAnyExtOrTrunc(RHS, dl, VT);
3270   }
3271 
3272   Res = CurDAG->getNode(PPCISD::CMPB, dl, VT, LHS, RHS);
3273 
3274   bool NonTrivialMask = ((int64_t) Mask) != INT64_C(-1);
3275   if (NonTrivialMask && !Alt) {
3276     // Res = Mask & CMPB
3277     Res = CurDAG->getNode(ISD::AND, dl, VT, Res,
3278                           CurDAG->getConstant(Mask, dl, VT));
3279   } else if (Alt) {
3280     // Res = (CMPB & Mask) | (~CMPB & Alt)
3281     // Which, as suggested here:
3282     //   https://graphics.stanford.edu/~seander/bithacks.html#MaskedMerge
3283     // can be written as:
3284     // Res = Alt ^ ((Alt ^ Mask) & CMPB)
3285     // useful because the (Alt ^ Mask) can be pre-computed.
3286     Res = CurDAG->getNode(ISD::AND, dl, VT, Res,
3287                           CurDAG->getConstant(Mask ^ Alt, dl, VT));
3288     Res = CurDAG->getNode(ISD::XOR, dl, VT, Res,
3289                           CurDAG->getConstant(Alt, dl, VT));
3290   }
3291 
3292   return Res;
3293 }
3294 
3295 // When CR bit registers are enabled, an extension of an i1 variable to a i32
3296 // or i64 value is lowered in terms of a SELECT_I[48] operation, and thus
3297 // involves constant materialization of a 0 or a 1 or both. If the result of
3298 // the extension is then operated upon by some operator that can be constant
3299 // folded with a constant 0 or 1, and that constant can be materialized using
3300 // only one instruction (like a zero or one), then we should fold in those
3301 // operations with the select.
foldBoolExts(SDValue & Res,SDNode * & N)3302 void PPCDAGToDAGISel::foldBoolExts(SDValue &Res, SDNode *&N) {
3303   if (!PPCSubTarget->useCRBits())
3304     return;
3305 
3306   if (N->getOpcode() != ISD::ZERO_EXTEND &&
3307       N->getOpcode() != ISD::SIGN_EXTEND &&
3308       N->getOpcode() != ISD::ANY_EXTEND)
3309     return;
3310 
3311   if (N->getOperand(0).getValueType() != MVT::i1)
3312     return;
3313 
3314   if (!N->hasOneUse())
3315     return;
3316 
3317   SDLoc dl(N);
3318   EVT VT = N->getValueType(0);
3319   SDValue Cond = N->getOperand(0);
3320   SDValue ConstTrue =
3321     CurDAG->getConstant(N->getOpcode() == ISD::SIGN_EXTEND ? -1 : 1, dl, VT);
3322   SDValue ConstFalse = CurDAG->getConstant(0, dl, VT);
3323 
3324   do {
3325     SDNode *User = *N->use_begin();
3326     if (User->getNumOperands() != 2)
3327       break;
3328 
3329     auto TryFold = [this, N, User, dl](SDValue Val) {
3330       SDValue UserO0 = User->getOperand(0), UserO1 = User->getOperand(1);
3331       SDValue O0 = UserO0.getNode() == N ? Val : UserO0;
3332       SDValue O1 = UserO1.getNode() == N ? Val : UserO1;
3333 
3334       return CurDAG->FoldConstantArithmetic(User->getOpcode(), dl,
3335                                             User->getValueType(0),
3336                                             O0.getNode(), O1.getNode());
3337     };
3338 
3339     SDValue TrueRes = TryFold(ConstTrue);
3340     if (!TrueRes)
3341       break;
3342     SDValue FalseRes = TryFold(ConstFalse);
3343     if (!FalseRes)
3344       break;
3345 
3346     // For us to materialize these using one instruction, we must be able to
3347     // represent them as signed 16-bit integers.
3348     uint64_t True  = cast<ConstantSDNode>(TrueRes)->getZExtValue(),
3349              False = cast<ConstantSDNode>(FalseRes)->getZExtValue();
3350     if (!isInt<16>(True) || !isInt<16>(False))
3351       break;
3352 
3353     // We can replace User with a new SELECT node, and try again to see if we
3354     // can fold the select with its user.
3355     Res = CurDAG->getSelect(dl, User->getValueType(0), Cond, TrueRes, FalseRes);
3356     N = User;
3357     ConstTrue = TrueRes;
3358     ConstFalse = FalseRes;
3359   } while (N->hasOneUse());
3360 }
3361 
PreprocessISelDAG()3362 void PPCDAGToDAGISel::PreprocessISelDAG() {
3363   SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
3364   ++Position;
3365 
3366   bool MadeChange = false;
3367   while (Position != CurDAG->allnodes_begin()) {
3368     SDNode *N = &*--Position;
3369     if (N->use_empty())
3370       continue;
3371 
3372     SDValue Res;
3373     switch (N->getOpcode()) {
3374     default: break;
3375     case ISD::OR:
3376       Res = combineToCMPB(N);
3377       break;
3378     }
3379 
3380     if (!Res)
3381       foldBoolExts(Res, N);
3382 
3383     if (Res) {
3384       DEBUG(dbgs() << "PPC DAG preprocessing replacing:\nOld:    ");
3385       DEBUG(N->dump(CurDAG));
3386       DEBUG(dbgs() << "\nNew: ");
3387       DEBUG(Res.getNode()->dump(CurDAG));
3388       DEBUG(dbgs() << "\n");
3389 
3390       CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Res);
3391       MadeChange = true;
3392     }
3393   }
3394 
3395   if (MadeChange)
3396     CurDAG->RemoveDeadNodes();
3397 }
3398 
3399 /// PostprocessISelDAG - Perform some late peephole optimizations
3400 /// on the DAG representation.
PostprocessISelDAG()3401 void PPCDAGToDAGISel::PostprocessISelDAG() {
3402 
3403   // Skip peepholes at -O0.
3404   if (TM.getOptLevel() == CodeGenOpt::None)
3405     return;
3406 
3407   PeepholePPC64();
3408   PeepholeCROps();
3409   PeepholePPC64ZExt();
3410 }
3411 
3412 // Check if all users of this node will become isel where the second operand
3413 // is the constant zero. If this is so, and if we can negate the condition,
3414 // then we can flip the true and false operands. This will allow the zero to
3415 // be folded with the isel so that we don't need to materialize a register
3416 // containing zero.
AllUsersSelectZero(SDNode * N)3417 bool PPCDAGToDAGISel::AllUsersSelectZero(SDNode *N) {
3418   // If we're not using isel, then this does not matter.
3419   if (!PPCSubTarget->hasISEL())
3420     return false;
3421 
3422   for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
3423        UI != UE; ++UI) {
3424     SDNode *User = *UI;
3425     if (!User->isMachineOpcode())
3426       return false;
3427     if (User->getMachineOpcode() != PPC::SELECT_I4 &&
3428         User->getMachineOpcode() != PPC::SELECT_I8)
3429       return false;
3430 
3431     SDNode *Op2 = User->getOperand(2).getNode();
3432     if (!Op2->isMachineOpcode())
3433       return false;
3434 
3435     if (Op2->getMachineOpcode() != PPC::LI &&
3436         Op2->getMachineOpcode() != PPC::LI8)
3437       return false;
3438 
3439     ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op2->getOperand(0));
3440     if (!C)
3441       return false;
3442 
3443     if (!C->isNullValue())
3444       return false;
3445   }
3446 
3447   return true;
3448 }
3449 
SwapAllSelectUsers(SDNode * N)3450 void PPCDAGToDAGISel::SwapAllSelectUsers(SDNode *N) {
3451   SmallVector<SDNode *, 4> ToReplace;
3452   for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
3453        UI != UE; ++UI) {
3454     SDNode *User = *UI;
3455     assert((User->getMachineOpcode() == PPC::SELECT_I4 ||
3456             User->getMachineOpcode() == PPC::SELECT_I8) &&
3457            "Must have all select users");
3458     ToReplace.push_back(User);
3459   }
3460 
3461   for (SmallVector<SDNode *, 4>::iterator UI = ToReplace.begin(),
3462        UE = ToReplace.end(); UI != UE; ++UI) {
3463     SDNode *User = *UI;
3464     SDNode *ResNode =
3465       CurDAG->getMachineNode(User->getMachineOpcode(), SDLoc(User),
3466                              User->getValueType(0), User->getOperand(0),
3467                              User->getOperand(2),
3468                              User->getOperand(1));
3469 
3470       DEBUG(dbgs() << "CR Peephole replacing:\nOld:    ");
3471       DEBUG(User->dump(CurDAG));
3472       DEBUG(dbgs() << "\nNew: ");
3473       DEBUG(ResNode->dump(CurDAG));
3474       DEBUG(dbgs() << "\n");
3475 
3476       ReplaceUses(User, ResNode);
3477   }
3478 }
3479 
PeepholeCROps()3480 void PPCDAGToDAGISel::PeepholeCROps() {
3481   bool IsModified;
3482   do {
3483     IsModified = false;
3484     for (SDNode &Node : CurDAG->allnodes()) {
3485       MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(&Node);
3486       if (!MachineNode || MachineNode->use_empty())
3487         continue;
3488       SDNode *ResNode = MachineNode;
3489 
3490       bool Op1Set   = false, Op1Unset = false,
3491            Op1Not   = false,
3492            Op2Set   = false, Op2Unset = false,
3493            Op2Not   = false;
3494 
3495       unsigned Opcode = MachineNode->getMachineOpcode();
3496       switch (Opcode) {
3497       default: break;
3498       case PPC::CRAND:
3499       case PPC::CRNAND:
3500       case PPC::CROR:
3501       case PPC::CRXOR:
3502       case PPC::CRNOR:
3503       case PPC::CREQV:
3504       case PPC::CRANDC:
3505       case PPC::CRORC: {
3506         SDValue Op = MachineNode->getOperand(1);
3507         if (Op.isMachineOpcode()) {
3508           if (Op.getMachineOpcode() == PPC::CRSET)
3509             Op2Set = true;
3510           else if (Op.getMachineOpcode() == PPC::CRUNSET)
3511             Op2Unset = true;
3512           else if (Op.getMachineOpcode() == PPC::CRNOR &&
3513                    Op.getOperand(0) == Op.getOperand(1))
3514             Op2Not = true;
3515         }
3516         }  // fallthrough
3517       case PPC::BC:
3518       case PPC::BCn:
3519       case PPC::SELECT_I4:
3520       case PPC::SELECT_I8:
3521       case PPC::SELECT_F4:
3522       case PPC::SELECT_F8:
3523       case PPC::SELECT_QFRC:
3524       case PPC::SELECT_QSRC:
3525       case PPC::SELECT_QBRC:
3526       case PPC::SELECT_VRRC:
3527       case PPC::SELECT_VSFRC:
3528       case PPC::SELECT_VSSRC:
3529       case PPC::SELECT_VSRC: {
3530         SDValue Op = MachineNode->getOperand(0);
3531         if (Op.isMachineOpcode()) {
3532           if (Op.getMachineOpcode() == PPC::CRSET)
3533             Op1Set = true;
3534           else if (Op.getMachineOpcode() == PPC::CRUNSET)
3535             Op1Unset = true;
3536           else if (Op.getMachineOpcode() == PPC::CRNOR &&
3537                    Op.getOperand(0) == Op.getOperand(1))
3538             Op1Not = true;
3539         }
3540         }
3541         break;
3542       }
3543 
3544       bool SelectSwap = false;
3545       switch (Opcode) {
3546       default: break;
3547       case PPC::CRAND:
3548         if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3549           // x & x = x
3550           ResNode = MachineNode->getOperand(0).getNode();
3551         else if (Op1Set)
3552           // 1 & y = y
3553           ResNode = MachineNode->getOperand(1).getNode();
3554         else if (Op2Set)
3555           // x & 1 = x
3556           ResNode = MachineNode->getOperand(0).getNode();
3557         else if (Op1Unset || Op2Unset)
3558           // x & 0 = 0 & y = 0
3559           ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
3560                                            MVT::i1);
3561         else if (Op1Not)
3562           // ~x & y = andc(y, x)
3563           ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
3564                                            MVT::i1, MachineNode->getOperand(1),
3565                                            MachineNode->getOperand(0).
3566                                              getOperand(0));
3567         else if (Op2Not)
3568           // x & ~y = andc(x, y)
3569           ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
3570                                            MVT::i1, MachineNode->getOperand(0),
3571                                            MachineNode->getOperand(1).
3572                                              getOperand(0));
3573         else if (AllUsersSelectZero(MachineNode))
3574           ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode),
3575                                            MVT::i1, MachineNode->getOperand(0),
3576                                            MachineNode->getOperand(1)),
3577           SelectSwap = true;
3578         break;
3579       case PPC::CRNAND:
3580         if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3581           // nand(x, x) -> nor(x, x)
3582           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3583                                            MVT::i1, MachineNode->getOperand(0),
3584                                            MachineNode->getOperand(0));
3585         else if (Op1Set)
3586           // nand(1, y) -> nor(y, y)
3587           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3588                                            MVT::i1, MachineNode->getOperand(1),
3589                                            MachineNode->getOperand(1));
3590         else if (Op2Set)
3591           // nand(x, 1) -> nor(x, x)
3592           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3593                                            MVT::i1, MachineNode->getOperand(0),
3594                                            MachineNode->getOperand(0));
3595         else if (Op1Unset || Op2Unset)
3596           // nand(x, 0) = nand(0, y) = 1
3597           ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
3598                                            MVT::i1);
3599         else if (Op1Not)
3600           // nand(~x, y) = ~(~x & y) = x | ~y = orc(x, y)
3601           ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
3602                                            MVT::i1, MachineNode->getOperand(0).
3603                                                       getOperand(0),
3604                                            MachineNode->getOperand(1));
3605         else if (Op2Not)
3606           // nand(x, ~y) = ~x | y = orc(y, x)
3607           ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
3608                                            MVT::i1, MachineNode->getOperand(1).
3609                                                       getOperand(0),
3610                                            MachineNode->getOperand(0));
3611         else if (AllUsersSelectZero(MachineNode))
3612           ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode),
3613                                            MVT::i1, MachineNode->getOperand(0),
3614                                            MachineNode->getOperand(1)),
3615           SelectSwap = true;
3616         break;
3617       case PPC::CROR:
3618         if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3619           // x | x = x
3620           ResNode = MachineNode->getOperand(0).getNode();
3621         else if (Op1Set || Op2Set)
3622           // x | 1 = 1 | y = 1
3623           ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
3624                                            MVT::i1);
3625         else if (Op1Unset)
3626           // 0 | y = y
3627           ResNode = MachineNode->getOperand(1).getNode();
3628         else if (Op2Unset)
3629           // x | 0 = x
3630           ResNode = MachineNode->getOperand(0).getNode();
3631         else if (Op1Not)
3632           // ~x | y = orc(y, x)
3633           ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
3634                                            MVT::i1, MachineNode->getOperand(1),
3635                                            MachineNode->getOperand(0).
3636                                              getOperand(0));
3637         else if (Op2Not)
3638           // x | ~y = orc(x, y)
3639           ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
3640                                            MVT::i1, MachineNode->getOperand(0),
3641                                            MachineNode->getOperand(1).
3642                                              getOperand(0));
3643         else if (AllUsersSelectZero(MachineNode))
3644           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3645                                            MVT::i1, MachineNode->getOperand(0),
3646                                            MachineNode->getOperand(1)),
3647           SelectSwap = true;
3648         break;
3649       case PPC::CRXOR:
3650         if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3651           // xor(x, x) = 0
3652           ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
3653                                            MVT::i1);
3654         else if (Op1Set)
3655           // xor(1, y) -> nor(y, y)
3656           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3657                                            MVT::i1, MachineNode->getOperand(1),
3658                                            MachineNode->getOperand(1));
3659         else if (Op2Set)
3660           // xor(x, 1) -> nor(x, x)
3661           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3662                                            MVT::i1, MachineNode->getOperand(0),
3663                                            MachineNode->getOperand(0));
3664         else if (Op1Unset)
3665           // xor(0, y) = y
3666           ResNode = MachineNode->getOperand(1).getNode();
3667         else if (Op2Unset)
3668           // xor(x, 0) = x
3669           ResNode = MachineNode->getOperand(0).getNode();
3670         else if (Op1Not)
3671           // xor(~x, y) = eqv(x, y)
3672           ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
3673                                            MVT::i1, MachineNode->getOperand(0).
3674                                                       getOperand(0),
3675                                            MachineNode->getOperand(1));
3676         else if (Op2Not)
3677           // xor(x, ~y) = eqv(x, y)
3678           ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
3679                                            MVT::i1, MachineNode->getOperand(0),
3680                                            MachineNode->getOperand(1).
3681                                              getOperand(0));
3682         else if (AllUsersSelectZero(MachineNode))
3683           ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
3684                                            MVT::i1, MachineNode->getOperand(0),
3685                                            MachineNode->getOperand(1)),
3686           SelectSwap = true;
3687         break;
3688       case PPC::CRNOR:
3689         if (Op1Set || Op2Set)
3690           // nor(1, y) -> 0
3691           ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
3692                                            MVT::i1);
3693         else if (Op1Unset)
3694           // nor(0, y) = ~y -> nor(y, y)
3695           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3696                                            MVT::i1, MachineNode->getOperand(1),
3697                                            MachineNode->getOperand(1));
3698         else if (Op2Unset)
3699           // nor(x, 0) = ~x
3700           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3701                                            MVT::i1, MachineNode->getOperand(0),
3702                                            MachineNode->getOperand(0));
3703         else if (Op1Not)
3704           // nor(~x, y) = andc(x, y)
3705           ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
3706                                            MVT::i1, MachineNode->getOperand(0).
3707                                                       getOperand(0),
3708                                            MachineNode->getOperand(1));
3709         else if (Op2Not)
3710           // nor(x, ~y) = andc(y, x)
3711           ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
3712                                            MVT::i1, MachineNode->getOperand(1).
3713                                                       getOperand(0),
3714                                            MachineNode->getOperand(0));
3715         else if (AllUsersSelectZero(MachineNode))
3716           ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode),
3717                                            MVT::i1, MachineNode->getOperand(0),
3718                                            MachineNode->getOperand(1)),
3719           SelectSwap = true;
3720         break;
3721       case PPC::CREQV:
3722         if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3723           // eqv(x, x) = 1
3724           ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
3725                                            MVT::i1);
3726         else if (Op1Set)
3727           // eqv(1, y) = y
3728           ResNode = MachineNode->getOperand(1).getNode();
3729         else if (Op2Set)
3730           // eqv(x, 1) = x
3731           ResNode = MachineNode->getOperand(0).getNode();
3732         else if (Op1Unset)
3733           // eqv(0, y) = ~y -> nor(y, y)
3734           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3735                                            MVT::i1, MachineNode->getOperand(1),
3736                                            MachineNode->getOperand(1));
3737         else if (Op2Unset)
3738           // eqv(x, 0) = ~x
3739           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3740                                            MVT::i1, MachineNode->getOperand(0),
3741                                            MachineNode->getOperand(0));
3742         else if (Op1Not)
3743           // eqv(~x, y) = xor(x, y)
3744           ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
3745                                            MVT::i1, MachineNode->getOperand(0).
3746                                                       getOperand(0),
3747                                            MachineNode->getOperand(1));
3748         else if (Op2Not)
3749           // eqv(x, ~y) = xor(x, y)
3750           ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
3751                                            MVT::i1, MachineNode->getOperand(0),
3752                                            MachineNode->getOperand(1).
3753                                              getOperand(0));
3754         else if (AllUsersSelectZero(MachineNode))
3755           ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
3756                                            MVT::i1, MachineNode->getOperand(0),
3757                                            MachineNode->getOperand(1)),
3758           SelectSwap = true;
3759         break;
3760       case PPC::CRANDC:
3761         if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3762           // andc(x, x) = 0
3763           ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
3764                                            MVT::i1);
3765         else if (Op1Set)
3766           // andc(1, y) = ~y
3767           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3768                                            MVT::i1, MachineNode->getOperand(1),
3769                                            MachineNode->getOperand(1));
3770         else if (Op1Unset || Op2Set)
3771           // andc(0, y) = andc(x, 1) = 0
3772           ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
3773                                            MVT::i1);
3774         else if (Op2Unset)
3775           // andc(x, 0) = x
3776           ResNode = MachineNode->getOperand(0).getNode();
3777         else if (Op1Not)
3778           // andc(~x, y) = ~(x | y) = nor(x, y)
3779           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3780                                            MVT::i1, MachineNode->getOperand(0).
3781                                                       getOperand(0),
3782                                            MachineNode->getOperand(1));
3783         else if (Op2Not)
3784           // andc(x, ~y) = x & y
3785           ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode),
3786                                            MVT::i1, MachineNode->getOperand(0),
3787                                            MachineNode->getOperand(1).
3788                                              getOperand(0));
3789         else if (AllUsersSelectZero(MachineNode))
3790           ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
3791                                            MVT::i1, MachineNode->getOperand(1),
3792                                            MachineNode->getOperand(0)),
3793           SelectSwap = true;
3794         break;
3795       case PPC::CRORC:
3796         if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3797           // orc(x, x) = 1
3798           ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
3799                                            MVT::i1);
3800         else if (Op1Set || Op2Unset)
3801           // orc(1, y) = orc(x, 0) = 1
3802           ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
3803                                            MVT::i1);
3804         else if (Op2Set)
3805           // orc(x, 1) = x
3806           ResNode = MachineNode->getOperand(0).getNode();
3807         else if (Op1Unset)
3808           // orc(0, y) = ~y
3809           ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3810                                            MVT::i1, MachineNode->getOperand(1),
3811                                            MachineNode->getOperand(1));
3812         else if (Op1Not)
3813           // orc(~x, y) = ~(x & y) = nand(x, y)
3814           ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode),
3815                                            MVT::i1, MachineNode->getOperand(0).
3816                                                       getOperand(0),
3817                                            MachineNode->getOperand(1));
3818         else if (Op2Not)
3819           // orc(x, ~y) = x | y
3820           ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode),
3821                                            MVT::i1, MachineNode->getOperand(0),
3822                                            MachineNode->getOperand(1).
3823                                              getOperand(0));
3824         else if (AllUsersSelectZero(MachineNode))
3825           ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
3826                                            MVT::i1, MachineNode->getOperand(1),
3827                                            MachineNode->getOperand(0)),
3828           SelectSwap = true;
3829         break;
3830       case PPC::SELECT_I4:
3831       case PPC::SELECT_I8:
3832       case PPC::SELECT_F4:
3833       case PPC::SELECT_F8:
3834       case PPC::SELECT_QFRC:
3835       case PPC::SELECT_QSRC:
3836       case PPC::SELECT_QBRC:
3837       case PPC::SELECT_VRRC:
3838       case PPC::SELECT_VSFRC:
3839       case PPC::SELECT_VSSRC:
3840       case PPC::SELECT_VSRC:
3841         if (Op1Set)
3842           ResNode = MachineNode->getOperand(1).getNode();
3843         else if (Op1Unset)
3844           ResNode = MachineNode->getOperand(2).getNode();
3845         else if (Op1Not)
3846           ResNode = CurDAG->getMachineNode(MachineNode->getMachineOpcode(),
3847                                            SDLoc(MachineNode),
3848                                            MachineNode->getValueType(0),
3849                                            MachineNode->getOperand(0).
3850                                              getOperand(0),
3851                                            MachineNode->getOperand(2),
3852                                            MachineNode->getOperand(1));
3853         break;
3854       case PPC::BC:
3855       case PPC::BCn:
3856         if (Op1Not)
3857           ResNode = CurDAG->getMachineNode(Opcode == PPC::BC ? PPC::BCn :
3858                                                                PPC::BC,
3859                                            SDLoc(MachineNode),
3860                                            MVT::Other,
3861                                            MachineNode->getOperand(0).
3862                                              getOperand(0),
3863                                            MachineNode->getOperand(1),
3864                                            MachineNode->getOperand(2));
3865         // FIXME: Handle Op1Set, Op1Unset here too.
3866         break;
3867       }
3868 
3869       // If we're inverting this node because it is used only by selects that
3870       // we'd like to swap, then swap the selects before the node replacement.
3871       if (SelectSwap)
3872         SwapAllSelectUsers(MachineNode);
3873 
3874       if (ResNode != MachineNode) {
3875         DEBUG(dbgs() << "CR Peephole replacing:\nOld:    ");
3876         DEBUG(MachineNode->dump(CurDAG));
3877         DEBUG(dbgs() << "\nNew: ");
3878         DEBUG(ResNode->dump(CurDAG));
3879         DEBUG(dbgs() << "\n");
3880 
3881         ReplaceUses(MachineNode, ResNode);
3882         IsModified = true;
3883       }
3884     }
3885     if (IsModified)
3886       CurDAG->RemoveDeadNodes();
3887   } while (IsModified);
3888 }
3889 
3890 // Gather the set of 32-bit operations that are known to have their
3891 // higher-order 32 bits zero, where ToPromote contains all such operations.
PeepholePPC64ZExtGather(SDValue Op32,SmallPtrSetImpl<SDNode * > & ToPromote)3892 static bool PeepholePPC64ZExtGather(SDValue Op32,
3893                                     SmallPtrSetImpl<SDNode *> &ToPromote) {
3894   if (!Op32.isMachineOpcode())
3895     return false;
3896 
3897   // First, check for the "frontier" instructions (those that will clear the
3898   // higher-order 32 bits.
3899 
3900   // For RLWINM and RLWNM, we need to make sure that the mask does not wrap
3901   // around. If it does not, then these instructions will clear the
3902   // higher-order bits.
3903   if ((Op32.getMachineOpcode() == PPC::RLWINM ||
3904        Op32.getMachineOpcode() == PPC::RLWNM) &&
3905       Op32.getConstantOperandVal(2) <= Op32.getConstantOperandVal(3)) {
3906     ToPromote.insert(Op32.getNode());
3907     return true;
3908   }
3909 
3910   // SLW and SRW always clear the higher-order bits.
3911   if (Op32.getMachineOpcode() == PPC::SLW ||
3912       Op32.getMachineOpcode() == PPC::SRW) {
3913     ToPromote.insert(Op32.getNode());
3914     return true;
3915   }
3916 
3917   // For LI and LIS, we need the immediate to be positive (so that it is not
3918   // sign extended).
3919   if (Op32.getMachineOpcode() == PPC::LI ||
3920       Op32.getMachineOpcode() == PPC::LIS) {
3921     if (!isUInt<15>(Op32.getConstantOperandVal(0)))
3922       return false;
3923 
3924     ToPromote.insert(Op32.getNode());
3925     return true;
3926   }
3927 
3928   // LHBRX and LWBRX always clear the higher-order bits.
3929   if (Op32.getMachineOpcode() == PPC::LHBRX ||
3930       Op32.getMachineOpcode() == PPC::LWBRX) {
3931     ToPromote.insert(Op32.getNode());
3932     return true;
3933   }
3934 
3935   // CNTLZW always produces a 64-bit value in [0,32], and so is zero extended.
3936   if (Op32.getMachineOpcode() == PPC::CNTLZW) {
3937     ToPromote.insert(Op32.getNode());
3938     return true;
3939   }
3940 
3941   // Next, check for those instructions we can look through.
3942 
3943   // Assuming the mask does not wrap around, then the higher-order bits are
3944   // taken directly from the first operand.
3945   if (Op32.getMachineOpcode() == PPC::RLWIMI &&
3946       Op32.getConstantOperandVal(3) <= Op32.getConstantOperandVal(4)) {
3947     SmallPtrSet<SDNode *, 16> ToPromote1;
3948     if (!PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1))
3949       return false;
3950 
3951     ToPromote.insert(Op32.getNode());
3952     ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
3953     return true;
3954   }
3955 
3956   // For OR, the higher-order bits are zero if that is true for both operands.
3957   // For SELECT_I4, the same is true (but the relevant operand numbers are
3958   // shifted by 1).
3959   if (Op32.getMachineOpcode() == PPC::OR ||
3960       Op32.getMachineOpcode() == PPC::SELECT_I4) {
3961     unsigned B = Op32.getMachineOpcode() == PPC::SELECT_I4 ? 1 : 0;
3962     SmallPtrSet<SDNode *, 16> ToPromote1;
3963     if (!PeepholePPC64ZExtGather(Op32.getOperand(B+0), ToPromote1))
3964       return false;
3965     if (!PeepholePPC64ZExtGather(Op32.getOperand(B+1), ToPromote1))
3966       return false;
3967 
3968     ToPromote.insert(Op32.getNode());
3969     ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
3970     return true;
3971   }
3972 
3973   // For ORI and ORIS, we need the higher-order bits of the first operand to be
3974   // zero, and also for the constant to be positive (so that it is not sign
3975   // extended).
3976   if (Op32.getMachineOpcode() == PPC::ORI ||
3977       Op32.getMachineOpcode() == PPC::ORIS) {
3978     SmallPtrSet<SDNode *, 16> ToPromote1;
3979     if (!PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1))
3980       return false;
3981     if (!isUInt<15>(Op32.getConstantOperandVal(1)))
3982       return false;
3983 
3984     ToPromote.insert(Op32.getNode());
3985     ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
3986     return true;
3987   }
3988 
3989   // The higher-order bits of AND are zero if that is true for at least one of
3990   // the operands.
3991   if (Op32.getMachineOpcode() == PPC::AND) {
3992     SmallPtrSet<SDNode *, 16> ToPromote1, ToPromote2;
3993     bool Op0OK =
3994       PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1);
3995     bool Op1OK =
3996       PeepholePPC64ZExtGather(Op32.getOperand(1), ToPromote2);
3997     if (!Op0OK && !Op1OK)
3998       return false;
3999 
4000     ToPromote.insert(Op32.getNode());
4001 
4002     if (Op0OK)
4003       ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
4004 
4005     if (Op1OK)
4006       ToPromote.insert(ToPromote2.begin(), ToPromote2.end());
4007 
4008     return true;
4009   }
4010 
4011   // For ANDI and ANDIS, the higher-order bits are zero if either that is true
4012   // of the first operand, or if the second operand is positive (so that it is
4013   // not sign extended).
4014   if (Op32.getMachineOpcode() == PPC::ANDIo ||
4015       Op32.getMachineOpcode() == PPC::ANDISo) {
4016     SmallPtrSet<SDNode *, 16> ToPromote1;
4017     bool Op0OK =
4018       PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1);
4019     bool Op1OK = isUInt<15>(Op32.getConstantOperandVal(1));
4020     if (!Op0OK && !Op1OK)
4021       return false;
4022 
4023     ToPromote.insert(Op32.getNode());
4024 
4025     if (Op0OK)
4026       ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
4027 
4028     return true;
4029   }
4030 
4031   return false;
4032 }
4033 
PeepholePPC64ZExt()4034 void PPCDAGToDAGISel::PeepholePPC64ZExt() {
4035   if (!PPCSubTarget->isPPC64())
4036     return;
4037 
4038   // When we zero-extend from i32 to i64, we use a pattern like this:
4039   // def : Pat<(i64 (zext i32:$in)),
4040   //           (RLDICL (INSERT_SUBREG (i64 (IMPLICIT_DEF)), $in, sub_32),
4041   //                   0, 32)>;
4042   // There are several 32-bit shift/rotate instructions, however, that will
4043   // clear the higher-order bits of their output, rendering the RLDICL
4044   // unnecessary. When that happens, we remove it here, and redefine the
4045   // relevant 32-bit operation to be a 64-bit operation.
4046 
4047   SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
4048   ++Position;
4049 
4050   bool MadeChange = false;
4051   while (Position != CurDAG->allnodes_begin()) {
4052     SDNode *N = &*--Position;
4053     // Skip dead nodes and any non-machine opcodes.
4054     if (N->use_empty() || !N->isMachineOpcode())
4055       continue;
4056 
4057     if (N->getMachineOpcode() != PPC::RLDICL)
4058       continue;
4059 
4060     if (N->getConstantOperandVal(1) != 0 ||
4061         N->getConstantOperandVal(2) != 32)
4062       continue;
4063 
4064     SDValue ISR = N->getOperand(0);
4065     if (!ISR.isMachineOpcode() ||
4066         ISR.getMachineOpcode() != TargetOpcode::INSERT_SUBREG)
4067       continue;
4068 
4069     if (!ISR.hasOneUse())
4070       continue;
4071 
4072     if (ISR.getConstantOperandVal(2) != PPC::sub_32)
4073       continue;
4074 
4075     SDValue IDef = ISR.getOperand(0);
4076     if (!IDef.isMachineOpcode() ||
4077         IDef.getMachineOpcode() != TargetOpcode::IMPLICIT_DEF)
4078       continue;
4079 
4080     // We now know that we're looking at a canonical i32 -> i64 zext. See if we
4081     // can get rid of it.
4082 
4083     SDValue Op32 = ISR->getOperand(1);
4084     if (!Op32.isMachineOpcode())
4085       continue;
4086 
4087     // There are some 32-bit instructions that always clear the high-order 32
4088     // bits, there are also some instructions (like AND) that we can look
4089     // through.
4090     SmallPtrSet<SDNode *, 16> ToPromote;
4091     if (!PeepholePPC64ZExtGather(Op32, ToPromote))
4092       continue;
4093 
4094     // If the ToPromote set contains nodes that have uses outside of the set
4095     // (except for the original INSERT_SUBREG), then abort the transformation.
4096     bool OutsideUse = false;
4097     for (SDNode *PN : ToPromote) {
4098       for (SDNode *UN : PN->uses()) {
4099         if (!ToPromote.count(UN) && UN != ISR.getNode()) {
4100           OutsideUse = true;
4101           break;
4102         }
4103       }
4104 
4105       if (OutsideUse)
4106         break;
4107     }
4108     if (OutsideUse)
4109       continue;
4110 
4111     MadeChange = true;
4112 
4113     // We now know that this zero extension can be removed by promoting to
4114     // nodes in ToPromote to 64-bit operations, where for operations in the
4115     // frontier of the set, we need to insert INSERT_SUBREGs for their
4116     // operands.
4117     for (SDNode *PN : ToPromote) {
4118       unsigned NewOpcode;
4119       switch (PN->getMachineOpcode()) {
4120       default:
4121         llvm_unreachable("Don't know the 64-bit variant of this instruction");
4122       case PPC::RLWINM:    NewOpcode = PPC::RLWINM8; break;
4123       case PPC::RLWNM:     NewOpcode = PPC::RLWNM8; break;
4124       case PPC::SLW:       NewOpcode = PPC::SLW8; break;
4125       case PPC::SRW:       NewOpcode = PPC::SRW8; break;
4126       case PPC::LI:        NewOpcode = PPC::LI8; break;
4127       case PPC::LIS:       NewOpcode = PPC::LIS8; break;
4128       case PPC::LHBRX:     NewOpcode = PPC::LHBRX8; break;
4129       case PPC::LWBRX:     NewOpcode = PPC::LWBRX8; break;
4130       case PPC::CNTLZW:    NewOpcode = PPC::CNTLZW8; break;
4131       case PPC::RLWIMI:    NewOpcode = PPC::RLWIMI8; break;
4132       case PPC::OR:        NewOpcode = PPC::OR8; break;
4133       case PPC::SELECT_I4: NewOpcode = PPC::SELECT_I8; break;
4134       case PPC::ORI:       NewOpcode = PPC::ORI8; break;
4135       case PPC::ORIS:      NewOpcode = PPC::ORIS8; break;
4136       case PPC::AND:       NewOpcode = PPC::AND8; break;
4137       case PPC::ANDIo:     NewOpcode = PPC::ANDIo8; break;
4138       case PPC::ANDISo:    NewOpcode = PPC::ANDISo8; break;
4139       }
4140 
4141       // Note: During the replacement process, the nodes will be in an
4142       // inconsistent state (some instructions will have operands with values
4143       // of the wrong type). Once done, however, everything should be right
4144       // again.
4145 
4146       SmallVector<SDValue, 4> Ops;
4147       for (const SDValue &V : PN->ops()) {
4148         if (!ToPromote.count(V.getNode()) && V.getValueType() == MVT::i32 &&
4149             !isa<ConstantSDNode>(V)) {
4150           SDValue ReplOpOps[] = { ISR.getOperand(0), V, ISR.getOperand(2) };
4151           SDNode *ReplOp =
4152             CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, SDLoc(V),
4153                                    ISR.getNode()->getVTList(), ReplOpOps);
4154           Ops.push_back(SDValue(ReplOp, 0));
4155         } else {
4156           Ops.push_back(V);
4157         }
4158       }
4159 
4160       // Because all to-be-promoted nodes only have users that are other
4161       // promoted nodes (or the original INSERT_SUBREG), we can safely replace
4162       // the i32 result value type with i64.
4163 
4164       SmallVector<EVT, 2> NewVTs;
4165       SDVTList VTs = PN->getVTList();
4166       for (unsigned i = 0, ie = VTs.NumVTs; i != ie; ++i)
4167         if (VTs.VTs[i] == MVT::i32)
4168           NewVTs.push_back(MVT::i64);
4169         else
4170           NewVTs.push_back(VTs.VTs[i]);
4171 
4172       DEBUG(dbgs() << "PPC64 ZExt Peephole morphing:\nOld:    ");
4173       DEBUG(PN->dump(CurDAG));
4174 
4175       CurDAG->SelectNodeTo(PN, NewOpcode, CurDAG->getVTList(NewVTs), Ops);
4176 
4177       DEBUG(dbgs() << "\nNew: ");
4178       DEBUG(PN->dump(CurDAG));
4179       DEBUG(dbgs() << "\n");
4180     }
4181 
4182     // Now we replace the original zero extend and its associated INSERT_SUBREG
4183     // with the value feeding the INSERT_SUBREG (which has now been promoted to
4184     // return an i64).
4185 
4186     DEBUG(dbgs() << "PPC64 ZExt Peephole replacing:\nOld:    ");
4187     DEBUG(N->dump(CurDAG));
4188     DEBUG(dbgs() << "\nNew: ");
4189     DEBUG(Op32.getNode()->dump(CurDAG));
4190     DEBUG(dbgs() << "\n");
4191 
4192     ReplaceUses(N, Op32.getNode());
4193   }
4194 
4195   if (MadeChange)
4196     CurDAG->RemoveDeadNodes();
4197 }
4198 
PeepholePPC64()4199 void PPCDAGToDAGISel::PeepholePPC64() {
4200   // These optimizations are currently supported only for 64-bit SVR4.
4201   if (PPCSubTarget->isDarwin() || !PPCSubTarget->isPPC64())
4202     return;
4203 
4204   SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
4205   ++Position;
4206 
4207   while (Position != CurDAG->allnodes_begin()) {
4208     SDNode *N = &*--Position;
4209     // Skip dead nodes and any non-machine opcodes.
4210     if (N->use_empty() || !N->isMachineOpcode())
4211       continue;
4212 
4213     unsigned FirstOp;
4214     unsigned StorageOpcode = N->getMachineOpcode();
4215 
4216     switch (StorageOpcode) {
4217     default: continue;
4218 
4219     case PPC::LBZ:
4220     case PPC::LBZ8:
4221     case PPC::LD:
4222     case PPC::LFD:
4223     case PPC::LFS:
4224     case PPC::LHA:
4225     case PPC::LHA8:
4226     case PPC::LHZ:
4227     case PPC::LHZ8:
4228     case PPC::LWA:
4229     case PPC::LWZ:
4230     case PPC::LWZ8:
4231       FirstOp = 0;
4232       break;
4233 
4234     case PPC::STB:
4235     case PPC::STB8:
4236     case PPC::STD:
4237     case PPC::STFD:
4238     case PPC::STFS:
4239     case PPC::STH:
4240     case PPC::STH8:
4241     case PPC::STW:
4242     case PPC::STW8:
4243       FirstOp = 1;
4244       break;
4245     }
4246 
4247     // If this is a load or store with a zero offset, or within the alignment,
4248     // we may be able to fold an add-immediate into the memory operation.
4249     // The check against alignment is below, as it can't occur until we check
4250     // the arguments to N
4251     if (!isa<ConstantSDNode>(N->getOperand(FirstOp)))
4252       continue;
4253 
4254     SDValue Base = N->getOperand(FirstOp + 1);
4255     if (!Base.isMachineOpcode())
4256       continue;
4257 
4258     // On targets with fusion, we don't want this to fire and remove a fusion
4259     // opportunity, unless a) it results in another fusion opportunity or
4260     // b) optimizing for size.
4261     if (PPCSubTarget->hasFusion() &&
4262         (!MF->getFunction()->optForSize() && !Base.hasOneUse()))
4263       continue;
4264 
4265     unsigned Flags = 0;
4266     bool ReplaceFlags = true;
4267 
4268     // When the feeding operation is an add-immediate of some sort,
4269     // determine whether we need to add relocation information to the
4270     // target flags on the immediate operand when we fold it into the
4271     // load instruction.
4272     //
4273     // For something like ADDItocL, the relocation information is
4274     // inferred from the opcode; when we process it in the AsmPrinter,
4275     // we add the necessary relocation there.  A load, though, can receive
4276     // relocation from various flavors of ADDIxxx, so we need to carry
4277     // the relocation information in the target flags.
4278     switch (Base.getMachineOpcode()) {
4279     default: continue;
4280 
4281     case PPC::ADDI8:
4282     case PPC::ADDI:
4283       // In some cases (such as TLS) the relocation information
4284       // is already in place on the operand, so copying the operand
4285       // is sufficient.
4286       ReplaceFlags = false;
4287       // For these cases, the immediate may not be divisible by 4, in
4288       // which case the fold is illegal for DS-form instructions.  (The
4289       // other cases provide aligned addresses and are always safe.)
4290       if ((StorageOpcode == PPC::LWA ||
4291            StorageOpcode == PPC::LD  ||
4292            StorageOpcode == PPC::STD) &&
4293           (!isa<ConstantSDNode>(Base.getOperand(1)) ||
4294            Base.getConstantOperandVal(1) % 4 != 0))
4295         continue;
4296       break;
4297     case PPC::ADDIdtprelL:
4298       Flags = PPCII::MO_DTPREL_LO;
4299       break;
4300     case PPC::ADDItlsldL:
4301       Flags = PPCII::MO_TLSLD_LO;
4302       break;
4303     case PPC::ADDItocL:
4304       Flags = PPCII::MO_TOC_LO;
4305       break;
4306     }
4307 
4308     SDValue ImmOpnd = Base.getOperand(1);
4309     int MaxDisplacement = 0;
4310     if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) {
4311       const GlobalValue *GV = GA->getGlobal();
4312       MaxDisplacement = GV->getAlignment() - 1;
4313     }
4314 
4315     int Offset = N->getConstantOperandVal(FirstOp);
4316     if (Offset < 0 || Offset > MaxDisplacement)
4317       continue;
4318 
4319     // We found an opportunity.  Reverse the operands from the add
4320     // immediate and substitute them into the load or store.  If
4321     // needed, update the target flags for the immediate operand to
4322     // reflect the necessary relocation information.
4323     DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase:    ");
4324     DEBUG(Base->dump(CurDAG));
4325     DEBUG(dbgs() << "\nN: ");
4326     DEBUG(N->dump(CurDAG));
4327     DEBUG(dbgs() << "\n");
4328 
4329     // If the relocation information isn't already present on the
4330     // immediate operand, add it now.
4331     if (ReplaceFlags) {
4332       if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) {
4333         SDLoc dl(GA);
4334         const GlobalValue *GV = GA->getGlobal();
4335         // We can't perform this optimization for data whose alignment
4336         // is insufficient for the instruction encoding.
4337         if (GV->getAlignment() < 4 &&
4338             (StorageOpcode == PPC::LD || StorageOpcode == PPC::STD ||
4339              StorageOpcode == PPC::LWA || (Offset % 4) != 0)) {
4340           DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n");
4341           continue;
4342         }
4343         ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, Offset, Flags);
4344       } else if (ConstantPoolSDNode *CP =
4345                  dyn_cast<ConstantPoolSDNode>(ImmOpnd)) {
4346         const Constant *C = CP->getConstVal();
4347         ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64,
4348                                                 CP->getAlignment(),
4349                                                 Offset, Flags);
4350       }
4351     }
4352 
4353     if (FirstOp == 1) // Store
4354       (void)CurDAG->UpdateNodeOperands(N, N->getOperand(0), ImmOpnd,
4355                                        Base.getOperand(0), N->getOperand(3));
4356     else // Load
4357       (void)CurDAG->UpdateNodeOperands(N, ImmOpnd, Base.getOperand(0),
4358                                        N->getOperand(2));
4359 
4360     // The add-immediate may now be dead, in which case remove it.
4361     if (Base.getNode()->use_empty())
4362       CurDAG->RemoveDeadNode(Base.getNode());
4363   }
4364 }
4365 
4366 
4367 /// createPPCISelDag - This pass converts a legalized DAG into a
4368 /// PowerPC-specific DAG, ready for instruction scheduling.
4369 ///
createPPCISelDag(PPCTargetMachine & TM)4370 FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM) {
4371   return new PPCDAGToDAGISel(TM);
4372 }
4373 
initializePassOnce(PassRegistry & Registry)4374 static void initializePassOnce(PassRegistry &Registry) {
4375   const char *Name = "PowerPC DAG->DAG Pattern Instruction Selection";
4376   PassInfo *PI = new PassInfo(Name, "ppc-codegen", &SelectionDAGISel::ID,
4377                               nullptr, false, false);
4378   Registry.registerPass(*PI, true);
4379 }
4380 
initializePPCDAGToDAGISelPass(PassRegistry & Registry)4381 void llvm::initializePPCDAGToDAGISelPass(PassRegistry &Registry) {
4382   CALL_ONCE_INITIALIZATION(initializePassOnce);
4383 }
4384 
4385