1 //===-- PPCFastISel.cpp - PowerPC FastISel implementation -----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the PowerPC-specific support for the FastISel class. Some
11 // of the target-specific code is generated by tablegen in the file
12 // PPCGenFastISel.inc, which is #included here.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "ppcfastisel"
18 #include "MCTargetDesc/PPCPredicates.h"
19 #include "PPCISelLowering.h"
20 #include "PPCSubtarget.h"
21 #include "PPCTargetMachine.h"
22 #include "llvm/ADT/Optional.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/FastISel.h"
25 #include "llvm/CodeGen/FunctionLoweringInfo.h"
26 #include "llvm/CodeGen/MachineConstantPool.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/IR/CallingConv.h"
31 #include "llvm/IR/GetElementPtrTypeIterator.h"
32 #include "llvm/IR/GlobalAlias.h"
33 #include "llvm/IR/GlobalVariable.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/Operator.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Target/TargetLowering.h"
38 #include "llvm/Target/TargetMachine.h"
40 //===----------------------------------------------------------------------===//
43 // FastLowerArguments: Handle simple cases.
44 // PPCMaterializeGV: Handle TLS.
45 // SelectCall: Handle function pointers.
46 // SelectCall: Handle multi-register return values.
47 // SelectCall: Optimize away nops for local calls.
48 // processCallArgs: Handle bit-converted arguments.
49 // finishCall: Handle multi-register return values.
50 // PPCComputeAddress: Handle parameter references as FrameIndex's.
51 // PPCEmitCmp: Handle immediate as operand 1.
52 // SelectCall: Handle small byval arguments.
53 // SelectIntrinsicCall: Implement.
54 // SelectSelect: Implement.
55 // Consider factoring isTypeLegal into the base class.
56 // Implement switches and jump tables.
58 //===----------------------------------------------------------------------===//
63 typedef struct Address {
76 // Innocuous defaults for our address.
78 : BaseType(RegBase), Offset(0) {
83 class PPCFastISel final : public FastISel {
85 const TargetMachine &TM;
86 const TargetInstrInfo &TII;
87 const TargetLowering &TLI;
88 const PPCSubtarget &PPCSubTarget;
92 explicit PPCFastISel(FunctionLoweringInfo &FuncInfo,
93 const TargetLibraryInfo *LibInfo)
94 : FastISel(FuncInfo, LibInfo),
95 TM(FuncInfo.MF->getTarget()),
96 TII(*TM.getInstrInfo()),
97 TLI(*TM.getTargetLowering()),
99 *((static_cast<const PPCTargetMachine *>(&TM))->getSubtargetImpl())
101 Context(&FuncInfo.Fn->getContext()) { }
103 // Backend specific FastISel code.
105 virtual bool TargetSelectInstruction(const Instruction *I);
106 virtual unsigned TargetMaterializeConstant(const Constant *C);
107 virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
108 virtual bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
110 virtual bool FastLowerArguments();
111 virtual unsigned FastEmit_i(MVT Ty, MVT RetTy, unsigned Opc, uint64_t Imm);
112 virtual unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
113 const TargetRegisterClass *RC,
114 unsigned Op0, bool Op0IsKill,
116 virtual unsigned FastEmitInst_r(unsigned MachineInstOpcode,
117 const TargetRegisterClass *RC,
118 unsigned Op0, bool Op0IsKill);
119 virtual unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
120 const TargetRegisterClass *RC,
121 unsigned Op0, bool Op0IsKill,
122 unsigned Op1, bool Op1IsKill);
124 // Instruction selection routines.
126 bool SelectLoad(const Instruction *I);
127 bool SelectStore(const Instruction *I);
128 bool SelectBranch(const Instruction *I);
129 bool SelectIndirectBr(const Instruction *I);
130 bool SelectCmp(const Instruction *I);
131 bool SelectFPExt(const Instruction *I);
132 bool SelectFPTrunc(const Instruction *I);
133 bool SelectIToFP(const Instruction *I, bool IsSigned);
134 bool SelectFPToI(const Instruction *I, bool IsSigned);
135 bool SelectBinaryIntOp(const Instruction *I, unsigned ISDOpcode);
136 bool SelectCall(const Instruction *I);
137 bool SelectRet(const Instruction *I);
138 bool SelectTrunc(const Instruction *I);
139 bool SelectIntExt(const Instruction *I);
143 bool isTypeLegal(Type *Ty, MVT &VT);
144 bool isLoadTypeLegal(Type *Ty, MVT &VT);
145 bool PPCEmitCmp(const Value *Src1Value, const Value *Src2Value,
146 bool isZExt, unsigned DestReg);
147 bool PPCEmitLoad(MVT VT, unsigned &ResultReg, Address &Addr,
148 const TargetRegisterClass *RC, bool IsZExt = true,
149 unsigned FP64LoadOpc = PPC::LFD);
150 bool PPCEmitStore(MVT VT, unsigned SrcReg, Address &Addr);
151 bool PPCComputeAddress(const Value *Obj, Address &Addr);
152 void PPCSimplifyAddress(Address &Addr, MVT VT, bool &UseOffset,
154 bool PPCEmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
155 unsigned DestReg, bool IsZExt);
156 unsigned PPCMaterializeFP(const ConstantFP *CFP, MVT VT);
157 unsigned PPCMaterializeGV(const GlobalValue *GV, MVT VT);
158 unsigned PPCMaterializeInt(const Constant *C, MVT VT);
159 unsigned PPCMaterialize32BitInt(int64_t Imm,
160 const TargetRegisterClass *RC);
161 unsigned PPCMaterialize64BitInt(int64_t Imm,
162 const TargetRegisterClass *RC);
163 unsigned PPCMoveToIntReg(const Instruction *I, MVT VT,
164 unsigned SrcReg, bool IsSigned);
165 unsigned PPCMoveToFPReg(MVT VT, unsigned SrcReg, bool IsSigned);
167 // Call handling routines.
169 bool processCallArgs(SmallVectorImpl<Value*> &Args,
170 SmallVectorImpl<unsigned> &ArgRegs,
171 SmallVectorImpl<MVT> &ArgVTs,
172 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
173 SmallVectorImpl<unsigned> &RegArgs,
177 void finishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
178 const Instruction *I, CallingConv::ID CC,
179 unsigned &NumBytes, bool IsVarArg);
180 CCAssignFn *usePPC32CCs(unsigned Flag);
183 #include "PPCGenFastISel.inc"
187 } // end anonymous namespace
189 #include "PPCGenCallingConv.inc"
191 // Function whose sole purpose is to kill compiler warnings
192 // stemming from unused functions included from PPCGenCallingConv.inc.
193 CCAssignFn *PPCFastISel::usePPC32CCs(unsigned Flag) {
195 return CC_PPC32_SVR4;
197 return CC_PPC32_SVR4_ByVal;
199 return CC_PPC32_SVR4_VarArg;
204 static Optional<PPC::Predicate> getComparePred(CmpInst::Predicate Pred) {
206 // These are not representable with any single compare.
207 case CmpInst::FCMP_FALSE:
208 case CmpInst::FCMP_UEQ:
209 case CmpInst::FCMP_UGT:
210 case CmpInst::FCMP_UGE:
211 case CmpInst::FCMP_ULT:
212 case CmpInst::FCMP_ULE:
213 case CmpInst::FCMP_UNE:
214 case CmpInst::FCMP_TRUE:
216 return Optional<PPC::Predicate>();
218 case CmpInst::FCMP_OEQ:
219 case CmpInst::ICMP_EQ:
222 case CmpInst::FCMP_OGT:
223 case CmpInst::ICMP_UGT:
224 case CmpInst::ICMP_SGT:
227 case CmpInst::FCMP_OGE:
228 case CmpInst::ICMP_UGE:
229 case CmpInst::ICMP_SGE:
232 case CmpInst::FCMP_OLT:
233 case CmpInst::ICMP_ULT:
234 case CmpInst::ICMP_SLT:
237 case CmpInst::FCMP_OLE:
238 case CmpInst::ICMP_ULE:
239 case CmpInst::ICMP_SLE:
242 case CmpInst::FCMP_ONE:
243 case CmpInst::ICMP_NE:
246 case CmpInst::FCMP_ORD:
249 case CmpInst::FCMP_UNO:
254 // Determine whether the type Ty is simple enough to be handled by
255 // fast-isel, and return its equivalent machine type in VT.
256 // FIXME: Copied directly from ARM -- factor into base class?
257 bool PPCFastISel::isTypeLegal(Type *Ty, MVT &VT) {
258 EVT Evt = TLI.getValueType(Ty, true);
260 // Only handle simple types.
261 if (Evt == MVT::Other || !Evt.isSimple()) return false;
262 VT = Evt.getSimpleVT();
264 // Handle all legal types, i.e. a register that will directly hold this
266 return TLI.isTypeLegal(VT);
269 // Determine whether the type Ty is simple enough to be handled by
270 // fast-isel as a load target, and return its equivalent machine type in VT.
271 bool PPCFastISel::isLoadTypeLegal(Type *Ty, MVT &VT) {
272 if (isTypeLegal(Ty, VT)) return true;
274 // If this is a type than can be sign or zero-extended to a basic operation
275 // go ahead and accept it now.
276 if (VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) {
283 // Given a value Obj, create an Address object Addr that represents its
284 // address. Return false if we can't handle it.
285 bool PPCFastISel::PPCComputeAddress(const Value *Obj, Address &Addr) {
286 const User *U = NULL;
287 unsigned Opcode = Instruction::UserOp1;
288 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
289 // Don't walk into other basic blocks unless the object is an alloca from
290 // another block, otherwise it may not have a virtual register assigned.
291 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
292 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
293 Opcode = I->getOpcode();
296 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
297 Opcode = C->getOpcode();
304 case Instruction::BitCast:
305 // Look through bitcasts.
306 return PPCComputeAddress(U->getOperand(0), Addr);
307 case Instruction::IntToPtr:
308 // Look past no-op inttoptrs.
309 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
310 return PPCComputeAddress(U->getOperand(0), Addr);
312 case Instruction::PtrToInt:
313 // Look past no-op ptrtoints.
314 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
315 return PPCComputeAddress(U->getOperand(0), Addr);
317 case Instruction::GetElementPtr: {
318 Address SavedAddr = Addr;
319 long TmpOffset = Addr.Offset;
321 // Iterate through the GEP folding the constants into offsets where
323 gep_type_iterator GTI = gep_type_begin(U);
324 for (User::const_op_iterator II = U->op_begin() + 1, IE = U->op_end();
325 II != IE; ++II, ++GTI) {
326 const Value *Op = *II;
327 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
328 const StructLayout *SL = DL.getStructLayout(STy);
329 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
330 TmpOffset += SL->getElementOffset(Idx);
332 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
334 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
335 // Constant-offset addressing.
336 TmpOffset += CI->getSExtValue() * S;
339 if (canFoldAddIntoGEP(U, Op)) {
340 // A compatible add with a constant operand. Fold the constant.
342 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
343 TmpOffset += CI->getSExtValue() * S;
344 // Iterate on the other operand.
345 Op = cast<AddOperator>(Op)->getOperand(0);
349 goto unsupported_gep;
354 // Try to grab the base operand now.
355 Addr.Offset = TmpOffset;
356 if (PPCComputeAddress(U->getOperand(0), Addr)) return true;
358 // We failed, restore everything and try the other options.
364 case Instruction::Alloca: {
365 const AllocaInst *AI = cast<AllocaInst>(Obj);
366 DenseMap<const AllocaInst*, int>::iterator SI =
367 FuncInfo.StaticAllocaMap.find(AI);
368 if (SI != FuncInfo.StaticAllocaMap.end()) {
369 Addr.BaseType = Address::FrameIndexBase;
370 Addr.Base.FI = SI->second;
377 // FIXME: References to parameters fall through to the behavior
378 // below. They should be able to reference a frame index since
379 // they are stored to the stack, so we can get "ld rx, offset(r1)"
380 // instead of "addi ry, r1, offset / ld rx, 0(ry)". Obj will
381 // just contain the parameter. Try to handle this with a FI.
383 // Try to get this in a register if nothing else has worked.
384 if (Addr.Base.Reg == 0)
385 Addr.Base.Reg = getRegForValue(Obj);
387 // Prevent assignment of base register to X0, which is inappropriate
388 // for loads and stores alike.
389 if (Addr.Base.Reg != 0)
390 MRI.setRegClass(Addr.Base.Reg, &PPC::G8RC_and_G8RC_NOX0RegClass);
392 return Addr.Base.Reg != 0;
395 // Fix up some addresses that can't be used directly. For example, if
396 // an offset won't fit in an instruction field, we may need to move it
397 // into an index register.
398 void PPCFastISel::PPCSimplifyAddress(Address &Addr, MVT VT, bool &UseOffset,
399 unsigned &IndexReg) {
401 // Check whether the offset fits in the instruction field.
402 if (!isInt<16>(Addr.Offset))
405 // If this is a stack pointer and the offset needs to be simplified then
406 // put the alloca address into a register, set the base type back to
407 // register and continue. This should almost never happen.
408 if (!UseOffset && Addr.BaseType == Address::FrameIndexBase) {
409 unsigned ResultReg = createResultReg(&PPC::G8RC_and_G8RC_NOX0RegClass);
410 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::ADDI8),
411 ResultReg).addFrameIndex(Addr.Base.FI).addImm(0);
412 Addr.Base.Reg = ResultReg;
413 Addr.BaseType = Address::RegBase;
417 IntegerType *OffsetTy = ((VT == MVT::i32) ? Type::getInt32Ty(*Context)
418 : Type::getInt64Ty(*Context));
419 const ConstantInt *Offset =
420 ConstantInt::getSigned(OffsetTy, (int64_t)(Addr.Offset));
421 IndexReg = PPCMaterializeInt(Offset, MVT::i64);
422 assert(IndexReg && "Unexpected error in PPCMaterializeInt!");
426 // Emit a load instruction if possible, returning true if we succeeded,
427 // otherwise false. See commentary below for how the register class of
428 // the load is determined.
429 bool PPCFastISel::PPCEmitLoad(MVT VT, unsigned &ResultReg, Address &Addr,
430 const TargetRegisterClass *RC,
431 bool IsZExt, unsigned FP64LoadOpc) {
433 bool UseOffset = true;
435 // If ResultReg is given, it determines the register class of the load.
436 // Otherwise, RC is the register class to use. If the result of the
437 // load isn't anticipated in this block, both may be zero, in which
438 // case we must make a conservative guess. In particular, don't assign
439 // R0 or X0 to the result register, as the result may be used in a load,
440 // store, add-immediate, or isel that won't permit this. (Though
441 // perhaps the spill and reload of live-exit values would handle this?)
442 const TargetRegisterClass *UseRC =
443 (ResultReg ? MRI.getRegClass(ResultReg) :
445 (VT == MVT::f64 ? &PPC::F8RCRegClass :
446 (VT == MVT::f32 ? &PPC::F4RCRegClass :
447 (VT == MVT::i64 ? &PPC::G8RC_and_G8RC_NOX0RegClass :
448 &PPC::GPRC_and_GPRC_NOR0RegClass)))));
450 bool Is32BitInt = UseRC->hasSuperClassEq(&PPC::GPRCRegClass);
452 switch (VT.SimpleTy) {
453 default: // e.g., vector types not handled
456 Opc = Is32BitInt ? PPC::LBZ : PPC::LBZ8;
460 (Is32BitInt ? PPC::LHZ : PPC::LHZ8) :
461 (Is32BitInt ? PPC::LHA : PPC::LHA8));
465 (Is32BitInt ? PPC::LWZ : PPC::LWZ8) :
466 (Is32BitInt ? PPC::LWA_32 : PPC::LWA));
467 if ((Opc == PPC::LWA || Opc == PPC::LWA_32) && ((Addr.Offset & 3) != 0))
472 assert(UseRC->hasSuperClassEq(&PPC::G8RCRegClass) &&
473 "64-bit load with 32-bit target??");
474 UseOffset = ((Addr.Offset & 3) == 0);
484 // If necessary, materialize the offset into a register and use
485 // the indexed form. Also handle stack pointers with special needs.
486 unsigned IndexReg = 0;
487 PPCSimplifyAddress(Addr, VT, UseOffset, IndexReg);
489 ResultReg = createResultReg(UseRC);
491 // Note: If we still have a frame index here, we know the offset is
492 // in range, as otherwise PPCSimplifyAddress would have converted it
494 if (Addr.BaseType == Address::FrameIndexBase) {
496 MachineMemOperand *MMO =
497 FuncInfo.MF->getMachineMemOperand(
498 MachinePointerInfo::getFixedStack(Addr.Base.FI, Addr.Offset),
499 MachineMemOperand::MOLoad, MFI.getObjectSize(Addr.Base.FI),
500 MFI.getObjectAlignment(Addr.Base.FI));
502 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
503 .addImm(Addr.Offset).addFrameIndex(Addr.Base.FI).addMemOperand(MMO);
505 // Base reg with offset in range.
506 } else if (UseOffset) {
508 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
509 .addImm(Addr.Offset).addReg(Addr.Base.Reg);
513 // Get the RR opcode corresponding to the RI one. FIXME: It would be
514 // preferable to use the ImmToIdxMap from PPCRegisterInfo.cpp, but it
515 // is hard to get at.
517 default: llvm_unreachable("Unexpected opcode!");
518 case PPC::LBZ: Opc = PPC::LBZX; break;
519 case PPC::LBZ8: Opc = PPC::LBZX8; break;
520 case PPC::LHZ: Opc = PPC::LHZX; break;
521 case PPC::LHZ8: Opc = PPC::LHZX8; break;
522 case PPC::LHA: Opc = PPC::LHAX; break;
523 case PPC::LHA8: Opc = PPC::LHAX8; break;
524 case PPC::LWZ: Opc = PPC::LWZX; break;
525 case PPC::LWZ8: Opc = PPC::LWZX8; break;
526 case PPC::LWA: Opc = PPC::LWAX; break;
527 case PPC::LWA_32: Opc = PPC::LWAX_32; break;
528 case PPC::LD: Opc = PPC::LDX; break;
529 case PPC::LFS: Opc = PPC::LFSX; break;
530 case PPC::LFD: Opc = PPC::LFDX; break;
532 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
533 .addReg(Addr.Base.Reg).addReg(IndexReg);
539 // Attempt to fast-select a load instruction.
540 bool PPCFastISel::SelectLoad(const Instruction *I) {
541 // FIXME: No atomic loads are supported.
542 if (cast<LoadInst>(I)->isAtomic())
545 // Verify we have a legal type before going any further.
547 if (!isLoadTypeLegal(I->getType(), VT))
550 // See if we can handle this address.
552 if (!PPCComputeAddress(I->getOperand(0), Addr))
555 // Look at the currently assigned register for this instruction
556 // to determine the required register class. This is necessary
557 // to constrain RA from using R0/X0 when this is not legal.
558 unsigned AssignedReg = FuncInfo.ValueMap[I];
559 const TargetRegisterClass *RC =
560 AssignedReg ? MRI.getRegClass(AssignedReg) : 0;
562 unsigned ResultReg = 0;
563 if (!PPCEmitLoad(VT, ResultReg, Addr, RC))
565 UpdateValueMap(I, ResultReg);
569 // Emit a store instruction to store SrcReg at Addr.
570 bool PPCFastISel::PPCEmitStore(MVT VT, unsigned SrcReg, Address &Addr) {
571 assert(SrcReg && "Nothing to store!");
573 bool UseOffset = true;
575 const TargetRegisterClass *RC = MRI.getRegClass(SrcReg);
576 bool Is32BitInt = RC->hasSuperClassEq(&PPC::GPRCRegClass);
578 switch (VT.SimpleTy) {
579 default: // e.g., vector types not handled
582 Opc = Is32BitInt ? PPC::STB : PPC::STB8;
585 Opc = Is32BitInt ? PPC::STH : PPC::STH8;
588 assert(Is32BitInt && "Not GPRC for i32??");
593 UseOffset = ((Addr.Offset & 3) == 0);
603 // If necessary, materialize the offset into a register and use
604 // the indexed form. Also handle stack pointers with special needs.
605 unsigned IndexReg = 0;
606 PPCSimplifyAddress(Addr, VT, UseOffset, IndexReg);
608 // Note: If we still have a frame index here, we know the offset is
609 // in range, as otherwise PPCSimplifyAddress would have converted it
611 if (Addr.BaseType == Address::FrameIndexBase) {
612 MachineMemOperand *MMO =
613 FuncInfo.MF->getMachineMemOperand(
614 MachinePointerInfo::getFixedStack(Addr.Base.FI, Addr.Offset),
615 MachineMemOperand::MOStore, MFI.getObjectSize(Addr.Base.FI),
616 MFI.getObjectAlignment(Addr.Base.FI));
618 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
621 .addFrameIndex(Addr.Base.FI)
624 // Base reg with offset in range.
625 } else if (UseOffset)
626 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
627 .addReg(SrcReg).addImm(Addr.Offset).addReg(Addr.Base.Reg);
631 // Get the RR opcode corresponding to the RI one. FIXME: It would be
632 // preferable to use the ImmToIdxMap from PPCRegisterInfo.cpp, but it
633 // is hard to get at.
635 default: llvm_unreachable("Unexpected opcode!");
636 case PPC::STB: Opc = PPC::STBX; break;
637 case PPC::STH : Opc = PPC::STHX; break;
638 case PPC::STW : Opc = PPC::STWX; break;
639 case PPC::STB8: Opc = PPC::STBX8; break;
640 case PPC::STH8: Opc = PPC::STHX8; break;
641 case PPC::STW8: Opc = PPC::STWX8; break;
642 case PPC::STD: Opc = PPC::STDX; break;
643 case PPC::STFS: Opc = PPC::STFSX; break;
644 case PPC::STFD: Opc = PPC::STFDX; break;
646 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
647 .addReg(SrcReg).addReg(Addr.Base.Reg).addReg(IndexReg);
653 // Attempt to fast-select a store instruction.
654 bool PPCFastISel::SelectStore(const Instruction *I) {
655 Value *Op0 = I->getOperand(0);
658 // FIXME: No atomics loads are supported.
659 if (cast<StoreInst>(I)->isAtomic())
662 // Verify we have a legal type before going any further.
664 if (!isLoadTypeLegal(Op0->getType(), VT))
667 // Get the value to be stored into a register.
668 SrcReg = getRegForValue(Op0);
672 // See if we can handle this address.
674 if (!PPCComputeAddress(I->getOperand(1), Addr))
677 if (!PPCEmitStore(VT, SrcReg, Addr))
683 // Attempt to fast-select a branch instruction.
684 bool PPCFastISel::SelectBranch(const Instruction *I) {
685 const BranchInst *BI = cast<BranchInst>(I);
686 MachineBasicBlock *BrBB = FuncInfo.MBB;
687 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
688 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
690 // For now, just try the simplest case where it's fed by a compare.
691 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
692 Optional<PPC::Predicate> OptPPCPred = getComparePred(CI->getPredicate());
696 PPC::Predicate PPCPred = OptPPCPred.getValue();
698 // Take advantage of fall-through opportunities.
699 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
701 PPCPred = PPC::InvertPredicate(PPCPred);
704 unsigned CondReg = createResultReg(&PPC::CRRCRegClass);
706 if (!PPCEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned(),
710 BuildMI(*BrBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::BCC))
711 .addImm(PPCPred).addReg(CondReg).addMBB(TBB);
712 FastEmitBranch(FBB, DbgLoc);
713 FuncInfo.MBB->addSuccessor(TBB);
716 } else if (const ConstantInt *CI =
717 dyn_cast<ConstantInt>(BI->getCondition())) {
718 uint64_t Imm = CI->getZExtValue();
719 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
720 FastEmitBranch(Target, DbgLoc);
724 // FIXME: ARM looks for a case where the block containing the compare
725 // has been split from the block containing the branch. If this happens,
726 // there is a vreg available containing the result of the compare. I'm
727 // not sure we can do much, as we've lost the predicate information with
728 // the compare instruction -- we have a 4-bit CR but don't know which bit
733 // Attempt to emit a compare of the two source values. Signed and unsigned
734 // comparisons are supported. Return false if we can't handle it.
735 bool PPCFastISel::PPCEmitCmp(const Value *SrcValue1, const Value *SrcValue2,
736 bool IsZExt, unsigned DestReg) {
737 Type *Ty = SrcValue1->getType();
738 EVT SrcEVT = TLI.getValueType(Ty, true);
739 if (!SrcEVT.isSimple())
741 MVT SrcVT = SrcEVT.getSimpleVT();
743 if (SrcVT == MVT::i1 && PPCSubTarget.useCRBits())
746 // See if operand 2 is an immediate encodeable in the compare.
747 // FIXME: Operands are not in canonical order at -O0, so an immediate
748 // operand in position 1 is a lost opportunity for now. We are
749 // similar to ARM in this regard.
753 // Only 16-bit integer constants can be represented in compares for
754 // PowerPC. Others will be materialized into a register.
755 if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(SrcValue2)) {
756 if (SrcVT == MVT::i64 || SrcVT == MVT::i32 || SrcVT == MVT::i16 ||
757 SrcVT == MVT::i8 || SrcVT == MVT::i1) {
758 const APInt &CIVal = ConstInt->getValue();
759 Imm = (IsZExt) ? (long)CIVal.getZExtValue() : (long)CIVal.getSExtValue();
760 if ((IsZExt && isUInt<16>(Imm)) || (!IsZExt && isInt<16>(Imm)))
766 bool NeedsExt = false;
767 switch (SrcVT.SimpleTy) {
768 default: return false;
770 CmpOpc = PPC::FCMPUS;
773 CmpOpc = PPC::FCMPUD;
779 // Intentional fall-through.
782 CmpOpc = IsZExt ? PPC::CMPLW : PPC::CMPW;
784 CmpOpc = IsZExt ? PPC::CMPLWI : PPC::CMPWI;
788 CmpOpc = IsZExt ? PPC::CMPLD : PPC::CMPD;
790 CmpOpc = IsZExt ? PPC::CMPLDI : PPC::CMPDI;
794 unsigned SrcReg1 = getRegForValue(SrcValue1);
798 unsigned SrcReg2 = 0;
800 SrcReg2 = getRegForValue(SrcValue2);
806 unsigned ExtReg = createResultReg(&PPC::GPRCRegClass);
807 if (!PPCEmitIntExt(SrcVT, SrcReg1, MVT::i32, ExtReg, IsZExt))
812 unsigned ExtReg = createResultReg(&PPC::GPRCRegClass);
813 if (!PPCEmitIntExt(SrcVT, SrcReg2, MVT::i32, ExtReg, IsZExt))
820 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc), DestReg)
821 .addReg(SrcReg1).addReg(SrcReg2);
823 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc), DestReg)
824 .addReg(SrcReg1).addImm(Imm);
829 // Attempt to fast-select a floating-point extend instruction.
830 bool PPCFastISel::SelectFPExt(const Instruction *I) {
831 Value *Src = I->getOperand(0);
832 EVT SrcVT = TLI.getValueType(Src->getType(), true);
833 EVT DestVT = TLI.getValueType(I->getType(), true);
835 if (SrcVT != MVT::f32 || DestVT != MVT::f64)
838 unsigned SrcReg = getRegForValue(Src);
842 // No code is generated for a FP extend.
843 UpdateValueMap(I, SrcReg);
847 // Attempt to fast-select a floating-point truncate instruction.
848 bool PPCFastISel::SelectFPTrunc(const Instruction *I) {
849 Value *Src = I->getOperand(0);
850 EVT SrcVT = TLI.getValueType(Src->getType(), true);
851 EVT DestVT = TLI.getValueType(I->getType(), true);
853 if (SrcVT != MVT::f64 || DestVT != MVT::f32)
856 unsigned SrcReg = getRegForValue(Src);
860 // Round the result to single precision.
861 unsigned DestReg = createResultReg(&PPC::F4RCRegClass);
862 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::FRSP), DestReg)
865 UpdateValueMap(I, DestReg);
869 // Move an i32 or i64 value in a GPR to an f64 value in an FPR.
870 // FIXME: When direct register moves are implemented (see PowerISA 2.08),
871 // those should be used instead of moving via a stack slot when the
872 // subtarget permits.
873 // FIXME: The code here is sloppy for the 4-byte case. Can use a 4-byte
874 // stack slot and 4-byte store/load sequence. Or just sext the 4-byte
875 // case to 8 bytes which produces tighter code but wastes stack space.
876 unsigned PPCFastISel::PPCMoveToFPReg(MVT SrcVT, unsigned SrcReg,
879 // If necessary, extend 32-bit int to 64-bit.
880 if (SrcVT == MVT::i32) {
881 unsigned TmpReg = createResultReg(&PPC::G8RCRegClass);
882 if (!PPCEmitIntExt(MVT::i32, SrcReg, MVT::i64, TmpReg, !IsSigned))
887 // Get a stack slot 8 bytes wide, aligned on an 8-byte boundary.
889 Addr.BaseType = Address::FrameIndexBase;
890 Addr.Base.FI = MFI.CreateStackObject(8, 8, false);
892 // Store the value from the GPR.
893 if (!PPCEmitStore(MVT::i64, SrcReg, Addr))
896 // Load the integer value into an FPR. The kind of load used depends
897 // on a number of conditions.
898 unsigned LoadOpc = PPC::LFD;
900 if (SrcVT == MVT::i32) {
902 LoadOpc = PPC::LFIWZX;
904 } else if (PPCSubTarget.hasLFIWAX()) {
905 LoadOpc = PPC::LFIWAX;
910 const TargetRegisterClass *RC = &PPC::F8RCRegClass;
911 unsigned ResultReg = 0;
912 if (!PPCEmitLoad(MVT::f64, ResultReg, Addr, RC, !IsSigned, LoadOpc))
918 // Attempt to fast-select an integer-to-floating-point conversion.
919 bool PPCFastISel::SelectIToFP(const Instruction *I, bool IsSigned) {
921 Type *DstTy = I->getType();
922 if (!isTypeLegal(DstTy, DstVT))
925 if (DstVT != MVT::f32 && DstVT != MVT::f64)
928 Value *Src = I->getOperand(0);
929 EVT SrcEVT = TLI.getValueType(Src->getType(), true);
930 if (!SrcEVT.isSimple())
933 MVT SrcVT = SrcEVT.getSimpleVT();
935 if (SrcVT != MVT::i8 && SrcVT != MVT::i16 &&
936 SrcVT != MVT::i32 && SrcVT != MVT::i64)
939 unsigned SrcReg = getRegForValue(Src);
943 // We can only lower an unsigned convert if we have the newer
944 // floating-point conversion operations.
945 if (!IsSigned && !PPCSubTarget.hasFPCVT())
948 // FIXME: For now we require the newer floating-point conversion operations
949 // (which are present only on P7 and A2 server models) when converting
950 // to single-precision float. Otherwise we have to generate a lot of
951 // fiddly code to avoid double rounding. If necessary, the fiddly code
952 // can be found in PPCTargetLowering::LowerINT_TO_FP().
953 if (DstVT == MVT::f32 && !PPCSubTarget.hasFPCVT())
956 // Extend the input if necessary.
957 if (SrcVT == MVT::i8 || SrcVT == MVT::i16) {
958 unsigned TmpReg = createResultReg(&PPC::G8RCRegClass);
959 if (!PPCEmitIntExt(SrcVT, SrcReg, MVT::i64, TmpReg, !IsSigned))
965 // Move the integer value to an FPR.
966 unsigned FPReg = PPCMoveToFPReg(SrcVT, SrcReg, IsSigned);
970 // Determine the opcode for the conversion.
971 const TargetRegisterClass *RC = &PPC::F8RCRegClass;
972 unsigned DestReg = createResultReg(RC);
975 if (DstVT == MVT::f32)
976 Opc = IsSigned ? PPC::FCFIDS : PPC::FCFIDUS;
978 Opc = IsSigned ? PPC::FCFID : PPC::FCFIDU;
980 // Generate the convert.
981 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), DestReg)
984 UpdateValueMap(I, DestReg);
988 // Move the floating-point value in SrcReg into an integer destination
989 // register, and return the register (or zero if we can't handle it).
990 // FIXME: When direct register moves are implemented (see PowerISA 2.08),
991 // those should be used instead of moving via a stack slot when the
992 // subtarget permits.
993 unsigned PPCFastISel::PPCMoveToIntReg(const Instruction *I, MVT VT,
994 unsigned SrcReg, bool IsSigned) {
995 // Get a stack slot 8 bytes wide, aligned on an 8-byte boundary.
996 // Note that if have STFIWX available, we could use a 4-byte stack
997 // slot for i32, but this being fast-isel we'll just go with the
998 // easiest code gen possible.
1000 Addr.BaseType = Address::FrameIndexBase;
1001 Addr.Base.FI = MFI.CreateStackObject(8, 8, false);
1003 // Store the value from the FPR.
1004 if (!PPCEmitStore(MVT::f64, SrcReg, Addr))
1007 // Reload it into a GPR. If we want an i32, modify the address
1008 // to have a 4-byte offset so we load from the right place.
1012 // Look at the currently assigned register for this instruction
1013 // to determine the required register class.
1014 unsigned AssignedReg = FuncInfo.ValueMap[I];
1015 const TargetRegisterClass *RC =
1016 AssignedReg ? MRI.getRegClass(AssignedReg) : 0;
1018 unsigned ResultReg = 0;
1019 if (!PPCEmitLoad(VT, ResultReg, Addr, RC, !IsSigned))
1025 // Attempt to fast-select a floating-point-to-integer conversion.
1026 bool PPCFastISel::SelectFPToI(const Instruction *I, bool IsSigned) {
1028 Type *DstTy = I->getType();
1029 if (!isTypeLegal(DstTy, DstVT))
1032 if (DstVT != MVT::i32 && DstVT != MVT::i64)
1035 Value *Src = I->getOperand(0);
1036 Type *SrcTy = Src->getType();
1037 if (!isTypeLegal(SrcTy, SrcVT))
1040 if (SrcVT != MVT::f32 && SrcVT != MVT::f64)
1043 unsigned SrcReg = getRegForValue(Src);
1047 // Convert f32 to f64 if necessary. This is just a meaningless copy
1048 // to get the register class right. COPY_TO_REGCLASS is needed since
1049 // a COPY from F4RC to F8RC is converted to a F4RC-F4RC copy downstream.
1050 const TargetRegisterClass *InRC = MRI.getRegClass(SrcReg);
1051 if (InRC == &PPC::F4RCRegClass) {
1052 unsigned TmpReg = createResultReg(&PPC::F8RCRegClass);
1053 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1054 TII.get(TargetOpcode::COPY_TO_REGCLASS), TmpReg)
1055 .addReg(SrcReg).addImm(PPC::F8RCRegClassID);
1059 // Determine the opcode for the conversion, which takes place
1060 // entirely within FPRs.
1061 unsigned DestReg = createResultReg(&PPC::F8RCRegClass);
1064 if (DstVT == MVT::i32)
1068 Opc = PPCSubTarget.hasFPCVT() ? PPC::FCTIWUZ : PPC::FCTIDZ;
1070 Opc = IsSigned ? PPC::FCTIDZ : PPC::FCTIDUZ;
1072 // Generate the convert.
1073 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), DestReg)
1076 // Now move the integer value from a float register to an integer register.
1077 unsigned IntReg = PPCMoveToIntReg(I, DstVT, DestReg, IsSigned);
1081 UpdateValueMap(I, IntReg);
1085 // Attempt to fast-select a binary integer operation that isn't already
1086 // handled automatically.
1087 bool PPCFastISel::SelectBinaryIntOp(const Instruction *I, unsigned ISDOpcode) {
1088 EVT DestVT = TLI.getValueType(I->getType(), true);
1090 // We can get here in the case when we have a binary operation on a non-legal
1091 // type and the target independent selector doesn't know how to handle it.
1092 if (DestVT != MVT::i16 && DestVT != MVT::i8)
1095 // Look at the currently assigned register for this instruction
1096 // to determine the required register class. If there is no register,
1097 // make a conservative choice (don't assign R0).
1098 unsigned AssignedReg = FuncInfo.ValueMap[I];
1099 const TargetRegisterClass *RC =
1100 (AssignedReg ? MRI.getRegClass(AssignedReg) :
1101 &PPC::GPRC_and_GPRC_NOR0RegClass);
1102 bool IsGPRC = RC->hasSuperClassEq(&PPC::GPRCRegClass);
1105 switch (ISDOpcode) {
1106 default: return false;
1108 Opc = IsGPRC ? PPC::ADD4 : PPC::ADD8;
1111 Opc = IsGPRC ? PPC::OR : PPC::OR8;
1114 Opc = IsGPRC ? PPC::SUBF : PPC::SUBF8;
1118 unsigned ResultReg = createResultReg(RC ? RC : &PPC::G8RCRegClass);
1119 unsigned SrcReg1 = getRegForValue(I->getOperand(0));
1120 if (SrcReg1 == 0) return false;
1122 // Handle case of small immediate operand.
1123 if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(I->getOperand(1))) {
1124 const APInt &CIVal = ConstInt->getValue();
1125 int Imm = (int)CIVal.getSExtValue();
1127 if (isInt<16>(Imm)) {
1130 llvm_unreachable("Missing case!");
1133 MRI.setRegClass(SrcReg1, &PPC::GPRC_and_GPRC_NOR0RegClass);
1137 MRI.setRegClass(SrcReg1, &PPC::G8RC_and_G8RC_NOX0RegClass);
1150 MRI.setRegClass(SrcReg1, &PPC::GPRC_and_GPRC_NOR0RegClass);
1159 MRI.setRegClass(SrcReg1, &PPC::G8RC_and_G8RC_NOX0RegClass);
1166 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc),
1170 UpdateValueMap(I, ResultReg);
1177 unsigned SrcReg2 = getRegForValue(I->getOperand(1));
1178 if (SrcReg2 == 0) return false;
1180 // Reverse operands for subtract-from.
1181 if (ISDOpcode == ISD::SUB)
1182 std::swap(SrcReg1, SrcReg2);
1184 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
1185 .addReg(SrcReg1).addReg(SrcReg2);
1186 UpdateValueMap(I, ResultReg);
1190 // Handle arguments to a call that we're attempting to fast-select.
1191 // Return false if the arguments are too complex for us at the moment.
1192 bool PPCFastISel::processCallArgs(SmallVectorImpl<Value*> &Args,
1193 SmallVectorImpl<unsigned> &ArgRegs,
1194 SmallVectorImpl<MVT> &ArgVTs,
1195 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1196 SmallVectorImpl<unsigned> &RegArgs,
1200 SmallVector<CCValAssign, 16> ArgLocs;
1201 CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, TM, ArgLocs, *Context);
1202 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CC_PPC64_ELF_FIS);
1204 // Bail out if we can't handle any of the arguments.
1205 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
1206 CCValAssign &VA = ArgLocs[I];
1207 MVT ArgVT = ArgVTs[VA.getValNo()];
1209 // Skip vector arguments for now, as well as long double and
1210 // uint128_t, and anything that isn't passed in a register.
1211 if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64 || ArgVT == MVT::i1 ||
1212 !VA.isRegLoc() || VA.needsCustom())
1215 // Skip bit-converted arguments for now.
1216 if (VA.getLocInfo() == CCValAssign::BCvt)
1220 // Get a count of how many bytes are to be pushed onto the stack.
1221 NumBytes = CCInfo.getNextStackOffset();
1223 // Issue CALLSEQ_START.
1224 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1225 TII.get(TII.getCallFrameSetupOpcode()))
1228 // Prepare to assign register arguments. Every argument uses up a
1229 // GPR protocol register even if it's passed in a floating-point
1231 unsigned NextGPR = PPC::X3;
1232 unsigned NextFPR = PPC::F1;
1234 // Process arguments.
1235 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
1236 CCValAssign &VA = ArgLocs[I];
1237 unsigned Arg = ArgRegs[VA.getValNo()];
1238 MVT ArgVT = ArgVTs[VA.getValNo()];
1240 // Handle argument promotion and bitcasts.
1241 switch (VA.getLocInfo()) {
1243 llvm_unreachable("Unknown loc info!");
1244 case CCValAssign::Full:
1246 case CCValAssign::SExt: {
1247 MVT DestVT = VA.getLocVT();
1248 const TargetRegisterClass *RC =
1249 (DestVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass;
1250 unsigned TmpReg = createResultReg(RC);
1251 if (!PPCEmitIntExt(ArgVT, Arg, DestVT, TmpReg, /*IsZExt*/false))
1252 llvm_unreachable("Failed to emit a sext!");
1257 case CCValAssign::AExt:
1258 case CCValAssign::ZExt: {
1259 MVT DestVT = VA.getLocVT();
1260 const TargetRegisterClass *RC =
1261 (DestVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass;
1262 unsigned TmpReg = createResultReg(RC);
1263 if (!PPCEmitIntExt(ArgVT, Arg, DestVT, TmpReg, /*IsZExt*/true))
1264 llvm_unreachable("Failed to emit a zext!");
1269 case CCValAssign::BCvt: {
1270 // FIXME: Not yet handled.
1271 llvm_unreachable("Should have bailed before getting here!");
1276 // Copy this argument to the appropriate register.
1278 if (ArgVT == MVT::f32 || ArgVT == MVT::f64) {
1284 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1285 TII.get(TargetOpcode::COPY), ArgReg).addReg(Arg);
1286 RegArgs.push_back(ArgReg);
1292 // For a call that we've determined we can fast-select, finish the
1293 // call sequence and generate a copy to obtain the return value (if any).
1294 void PPCFastISel::finishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1295 const Instruction *I, CallingConv::ID CC,
1296 unsigned &NumBytes, bool IsVarArg) {
1297 // Issue CallSEQ_END.
1298 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1299 TII.get(TII.getCallFrameDestroyOpcode()))
1300 .addImm(NumBytes).addImm(0);
1302 // Next, generate a copy to obtain the return value.
1303 // FIXME: No multi-register return values yet, though I don't foresee
1304 // any real difficulties there.
1305 if (RetVT != MVT::isVoid) {
1306 SmallVector<CCValAssign, 16> RVLocs;
1307 CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, TM, RVLocs, *Context);
1308 CCInfo.AnalyzeCallResult(RetVT, RetCC_PPC64_ELF_FIS);
1309 CCValAssign &VA = RVLocs[0];
1310 assert(RVLocs.size() == 1 && "No support for multi-reg return values!");
1311 assert(VA.isRegLoc() && "Can only return in registers!");
1313 MVT DestVT = VA.getValVT();
1314 MVT CopyVT = DestVT;
1316 // Ints smaller than a register still arrive in a full 64-bit
1317 // register, so make sure we recognize this.
1318 if (RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32)
1321 unsigned SourcePhysReg = VA.getLocReg();
1322 unsigned ResultReg = 0;
1324 if (RetVT == CopyVT) {
1325 const TargetRegisterClass *CpyRC = TLI.getRegClassFor(CopyVT);
1326 ResultReg = createResultReg(CpyRC);
1328 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1329 TII.get(TargetOpcode::COPY), ResultReg)
1330 .addReg(SourcePhysReg);
1332 // If necessary, round the floating result to single precision.
1333 } else if (CopyVT == MVT::f64) {
1334 ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
1335 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::FRSP),
1336 ResultReg).addReg(SourcePhysReg);
1338 // If only the low half of a general register is needed, generate
1339 // a GPRC copy instead of a G8RC copy. (EXTRACT_SUBREG can't be
1340 // used along the fast-isel path (not lowered), and downstream logic
1341 // also doesn't like a direct subreg copy on a physical reg.)
1342 } else if (RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32) {
1343 ResultReg = createResultReg(&PPC::GPRCRegClass);
1344 // Convert physical register from G8RC to GPRC.
1345 SourcePhysReg -= PPC::X0 - PPC::R0;
1346 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1347 TII.get(TargetOpcode::COPY), ResultReg)
1348 .addReg(SourcePhysReg);
1351 assert(ResultReg && "ResultReg unset!");
1352 UsedRegs.push_back(SourcePhysReg);
1353 UpdateValueMap(I, ResultReg);
1357 // Attempt to fast-select a call instruction.
1358 bool PPCFastISel::SelectCall(const Instruction *I) {
1359 const CallInst *CI = cast<CallInst>(I);
1360 const Value *Callee = CI->getCalledValue();
1362 // Can't handle inline asm.
1363 if (isa<InlineAsm>(Callee))
1366 // Allow SelectionDAG isel to handle tail calls.
1367 if (CI->isTailCall())
1370 // Obtain calling convention.
1371 ImmutableCallSite CS(CI);
1372 CallingConv::ID CC = CS.getCallingConv();
1374 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1375 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1376 bool IsVarArg = FTy->isVarArg();
1378 // Not ready for varargs yet.
1382 // Handle simple calls for now, with legal return types and
1383 // those that can be extended.
1384 Type *RetTy = I->getType();
1386 if (RetTy->isVoidTy())
1387 RetVT = MVT::isVoid;
1388 else if (!isTypeLegal(RetTy, RetVT) && RetVT != MVT::i16 &&
1392 // FIXME: No multi-register return values yet.
1393 if (RetVT != MVT::isVoid && RetVT != MVT::i8 && RetVT != MVT::i16 &&
1394 RetVT != MVT::i32 && RetVT != MVT::i64 && RetVT != MVT::f32 &&
1395 RetVT != MVT::f64) {
1396 SmallVector<CCValAssign, 16> RVLocs;
1397 CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, TM, RVLocs, *Context);
1398 CCInfo.AnalyzeCallResult(RetVT, RetCC_PPC64_ELF_FIS);
1399 if (RVLocs.size() > 1)
1403 // Bail early if more than 8 arguments, as we only currently
1404 // handle arguments passed in registers.
1405 unsigned NumArgs = CS.arg_size();
1409 // Set up the argument vectors.
1410 SmallVector<Value*, 8> Args;
1411 SmallVector<unsigned, 8> ArgRegs;
1412 SmallVector<MVT, 8> ArgVTs;
1413 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1415 Args.reserve(NumArgs);
1416 ArgRegs.reserve(NumArgs);
1417 ArgVTs.reserve(NumArgs);
1418 ArgFlags.reserve(NumArgs);
1420 for (ImmutableCallSite::arg_iterator II = CS.arg_begin(), IE = CS.arg_end();
1422 // FIXME: ARM does something for intrinsic calls here, check into that.
1424 unsigned AttrIdx = II - CS.arg_begin() + 1;
1426 // Only handle easy calls for now. It would be reasonably easy
1427 // to handle <= 8-byte structures passed ByVal in registers, but we
1428 // have to ensure they are right-justified in the register.
1429 if (CS.paramHasAttr(AttrIdx, Attribute::InReg) ||
1430 CS.paramHasAttr(AttrIdx, Attribute::StructRet) ||
1431 CS.paramHasAttr(AttrIdx, Attribute::Nest) ||
1432 CS.paramHasAttr(AttrIdx, Attribute::ByVal))
1435 ISD::ArgFlagsTy Flags;
1436 if (CS.paramHasAttr(AttrIdx, Attribute::SExt))
1438 if (CS.paramHasAttr(AttrIdx, Attribute::ZExt))
1441 Type *ArgTy = (*II)->getType();
1443 if (!isTypeLegal(ArgTy, ArgVT) && ArgVT != MVT::i16 && ArgVT != MVT::i8)
1446 if (ArgVT.isVector())
1449 unsigned Arg = getRegForValue(*II);
1453 unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
1454 Flags.setOrigAlign(OriginalAlignment);
1456 Args.push_back(*II);
1457 ArgRegs.push_back(Arg);
1458 ArgVTs.push_back(ArgVT);
1459 ArgFlags.push_back(Flags);
1462 // Process the arguments.
1463 SmallVector<unsigned, 8> RegArgs;
1466 if (!processCallArgs(Args, ArgRegs, ArgVTs, ArgFlags,
1467 RegArgs, CC, NumBytes, IsVarArg))
1470 // FIXME: No handling for function pointers yet. This requires
1471 // implementing the function descriptor (OPD) setup.
1472 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
1476 // Build direct call with NOP for TOC restore.
1477 // FIXME: We can and should optimize away the NOP for local calls.
1478 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1479 TII.get(PPC::BL8_NOP));
1481 MIB.addGlobalAddress(GV);
1483 // Add implicit physical register uses to the call.
1484 for (unsigned II = 0, IE = RegArgs.size(); II != IE; ++II)
1485 MIB.addReg(RegArgs[II], RegState::Implicit);
1487 // Add a register mask with the call-preserved registers. Proper
1488 // defs for return values will be added by setPhysRegsDeadExcept().
1489 MIB.addRegMask(TRI.getCallPreservedMask(CC));
1491 // Finish off the call including any return values.
1492 SmallVector<unsigned, 4> UsedRegs;
1493 finishCall(RetVT, UsedRegs, I, CC, NumBytes, IsVarArg);
1495 // Set all unused physregs defs as dead.
1496 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1501 // Attempt to fast-select a return instruction.
1502 bool PPCFastISel::SelectRet(const Instruction *I) {
1504 if (!FuncInfo.CanLowerReturn)
1507 const ReturnInst *Ret = cast<ReturnInst>(I);
1508 const Function &F = *I->getParent()->getParent();
1510 // Build a list of return value registers.
1511 SmallVector<unsigned, 4> RetRegs;
1512 CallingConv::ID CC = F.getCallingConv();
1514 if (Ret->getNumOperands() > 0) {
1515 SmallVector<ISD::OutputArg, 4> Outs;
1516 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
1518 // Analyze operands of the call, assigning locations to each operand.
1519 SmallVector<CCValAssign, 16> ValLocs;
1520 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs, *Context);
1521 CCInfo.AnalyzeReturn(Outs, RetCC_PPC64_ELF_FIS);
1522 const Value *RV = Ret->getOperand(0);
1524 // FIXME: Only one output register for now.
1525 if (ValLocs.size() > 1)
1528 // Special case for returning a constant integer of any size.
1529 // Materialize the constant as an i64 and copy it to the return
1530 // register. This avoids an unnecessary extend or truncate.
1531 if (isa<ConstantInt>(*RV)) {
1532 const Constant *C = cast<Constant>(RV);
1533 unsigned SrcReg = PPCMaterializeInt(C, MVT::i64);
1534 unsigned RetReg = ValLocs[0].getLocReg();
1535 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1536 TII.get(TargetOpcode::COPY), RetReg).addReg(SrcReg);
1537 RetRegs.push_back(RetReg);
1540 unsigned Reg = getRegForValue(RV);
1545 // Copy the result values into the output registers.
1546 for (unsigned i = 0; i < ValLocs.size(); ++i) {
1548 CCValAssign &VA = ValLocs[i];
1549 assert(VA.isRegLoc() && "Can only return in registers!");
1550 RetRegs.push_back(VA.getLocReg());
1551 unsigned SrcReg = Reg + VA.getValNo();
1553 EVT RVEVT = TLI.getValueType(RV->getType());
1554 if (!RVEVT.isSimple())
1556 MVT RVVT = RVEVT.getSimpleVT();
1557 MVT DestVT = VA.getLocVT();
1559 if (RVVT != DestVT && RVVT != MVT::i8 &&
1560 RVVT != MVT::i16 && RVVT != MVT::i32)
1563 if (RVVT != DestVT) {
1564 switch (VA.getLocInfo()) {
1566 llvm_unreachable("Unknown loc info!");
1567 case CCValAssign::Full:
1568 llvm_unreachable("Full value assign but types don't match?");
1569 case CCValAssign::AExt:
1570 case CCValAssign::ZExt: {
1571 const TargetRegisterClass *RC =
1572 (DestVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass;
1573 unsigned TmpReg = createResultReg(RC);
1574 if (!PPCEmitIntExt(RVVT, SrcReg, DestVT, TmpReg, true))
1579 case CCValAssign::SExt: {
1580 const TargetRegisterClass *RC =
1581 (DestVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass;
1582 unsigned TmpReg = createResultReg(RC);
1583 if (!PPCEmitIntExt(RVVT, SrcReg, DestVT, TmpReg, false))
1591 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1592 TII.get(TargetOpcode::COPY), RetRegs[i])
1598 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1601 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
1602 MIB.addReg(RetRegs[i], RegState::Implicit);
1607 // Attempt to emit an integer extend of SrcReg into DestReg. Both
1608 // signed and zero extensions are supported. Return false if we
1610 bool PPCFastISel::PPCEmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
1611 unsigned DestReg, bool IsZExt) {
1612 if (DestVT != MVT::i32 && DestVT != MVT::i64)
1614 if (SrcVT != MVT::i8 && SrcVT != MVT::i16 && SrcVT != MVT::i32)
1617 // Signed extensions use EXTSB, EXTSH, EXTSW.
1620 if (SrcVT == MVT::i8)
1621 Opc = (DestVT == MVT::i32) ? PPC::EXTSB : PPC::EXTSB8_32_64;
1622 else if (SrcVT == MVT::i16)
1623 Opc = (DestVT == MVT::i32) ? PPC::EXTSH : PPC::EXTSH8_32_64;
1625 assert(DestVT == MVT::i64 && "Signed extend from i32 to i32??");
1626 Opc = PPC::EXTSW_32_64;
1628 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), DestReg)
1631 // Unsigned 32-bit extensions use RLWINM.
1632 } else if (DestVT == MVT::i32) {
1634 if (SrcVT == MVT::i8)
1637 assert(SrcVT == MVT::i16 && "Unsigned extend from i32 to i32??");
1640 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::RLWINM),
1642 .addReg(SrcReg).addImm(/*SH=*/0).addImm(MB).addImm(/*ME=*/31);
1644 // Unsigned 64-bit extensions use RLDICL (with a 32-bit source).
1647 if (SrcVT == MVT::i8)
1649 else if (SrcVT == MVT::i16)
1653 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1654 TII.get(PPC::RLDICL_32_64), DestReg)
1655 .addReg(SrcReg).addImm(/*SH=*/0).addImm(MB);
1661 // Attempt to fast-select an indirect branch instruction.
1662 bool PPCFastISel::SelectIndirectBr(const Instruction *I) {
1663 unsigned AddrReg = getRegForValue(I->getOperand(0));
1667 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::MTCTR8))
1669 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::BCTR8));
1671 const IndirectBrInst *IB = cast<IndirectBrInst>(I);
1672 for (unsigned i = 0, e = IB->getNumSuccessors(); i != e; ++i)
1673 FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[IB->getSuccessor(i)]);
1678 // Attempt to fast-select an integer truncate instruction.
1679 bool PPCFastISel::SelectTrunc(const Instruction *I) {
1680 Value *Src = I->getOperand(0);
1681 EVT SrcVT = TLI.getValueType(Src->getType(), true);
1682 EVT DestVT = TLI.getValueType(I->getType(), true);
1684 if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16)
1687 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8)
1690 unsigned SrcReg = getRegForValue(Src);
1694 // The only interesting case is when we need to switch register classes.
1695 if (SrcVT == MVT::i64) {
1696 unsigned ResultReg = createResultReg(&PPC::GPRCRegClass);
1697 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1698 TII.get(TargetOpcode::COPY),
1699 ResultReg).addReg(SrcReg, 0, PPC::sub_32);
1703 UpdateValueMap(I, SrcReg);
1707 // Attempt to fast-select an integer extend instruction.
1708 bool PPCFastISel::SelectIntExt(const Instruction *I) {
1709 Type *DestTy = I->getType();
1710 Value *Src = I->getOperand(0);
1711 Type *SrcTy = Src->getType();
1713 bool IsZExt = isa<ZExtInst>(I);
1714 unsigned SrcReg = getRegForValue(Src);
1715 if (!SrcReg) return false;
1717 EVT SrcEVT, DestEVT;
1718 SrcEVT = TLI.getValueType(SrcTy, true);
1719 DestEVT = TLI.getValueType(DestTy, true);
1720 if (!SrcEVT.isSimple())
1722 if (!DestEVT.isSimple())
1725 MVT SrcVT = SrcEVT.getSimpleVT();
1726 MVT DestVT = DestEVT.getSimpleVT();
1728 // If we know the register class needed for the result of this
1729 // instruction, use it. Otherwise pick the register class of the
1730 // correct size that does not contain X0/R0, since we don't know
1731 // whether downstream uses permit that assignment.
1732 unsigned AssignedReg = FuncInfo.ValueMap[I];
1733 const TargetRegisterClass *RC =
1734 (AssignedReg ? MRI.getRegClass(AssignedReg) :
1735 (DestVT == MVT::i64 ? &PPC::G8RC_and_G8RC_NOX0RegClass :
1736 &PPC::GPRC_and_GPRC_NOR0RegClass));
1737 unsigned ResultReg = createResultReg(RC);
1739 if (!PPCEmitIntExt(SrcVT, SrcReg, DestVT, ResultReg, IsZExt))
1742 UpdateValueMap(I, ResultReg);
1746 // Attempt to fast-select an instruction that wasn't handled by
1747 // the table-generated machinery.
1748 bool PPCFastISel::TargetSelectInstruction(const Instruction *I) {
1750 switch (I->getOpcode()) {
1751 case Instruction::Load:
1752 return SelectLoad(I);
1753 case Instruction::Store:
1754 return SelectStore(I);
1755 case Instruction::Br:
1756 return SelectBranch(I);
1757 case Instruction::IndirectBr:
1758 return SelectIndirectBr(I);
1759 case Instruction::FPExt:
1760 return SelectFPExt(I);
1761 case Instruction::FPTrunc:
1762 return SelectFPTrunc(I);
1763 case Instruction::SIToFP:
1764 return SelectIToFP(I, /*IsSigned*/ true);
1765 case Instruction::UIToFP:
1766 return SelectIToFP(I, /*IsSigned*/ false);
1767 case Instruction::FPToSI:
1768 return SelectFPToI(I, /*IsSigned*/ true);
1769 case Instruction::FPToUI:
1770 return SelectFPToI(I, /*IsSigned*/ false);
1771 case Instruction::Add:
1772 return SelectBinaryIntOp(I, ISD::ADD);
1773 case Instruction::Or:
1774 return SelectBinaryIntOp(I, ISD::OR);
1775 case Instruction::Sub:
1776 return SelectBinaryIntOp(I, ISD::SUB);
1777 case Instruction::Call:
1778 if (dyn_cast<IntrinsicInst>(I))
1780 return SelectCall(I);
1781 case Instruction::Ret:
1782 return SelectRet(I);
1783 case Instruction::Trunc:
1784 return SelectTrunc(I);
1785 case Instruction::ZExt:
1786 case Instruction::SExt:
1787 return SelectIntExt(I);
1788 // Here add other flavors of Instruction::XXX that automated
1789 // cases don't catch. For example, switches are terminators
1790 // that aren't yet handled.
1797 // Materialize a floating-point constant into a register, and return
1798 // the register number (or zero if we failed to handle it).
1799 unsigned PPCFastISel::PPCMaterializeFP(const ConstantFP *CFP, MVT VT) {
1800 // No plans to handle long double here.
1801 if (VT != MVT::f32 && VT != MVT::f64)
1804 // All FP constants are loaded from the constant pool.
1805 unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
1806 assert(Align > 0 && "Unexpectedly missing alignment information!");
1807 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
1808 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
1809 CodeModel::Model CModel = TM.getCodeModel();
1811 MachineMemOperand *MMO =
1812 FuncInfo.MF->getMachineMemOperand(
1813 MachinePointerInfo::getConstantPool(), MachineMemOperand::MOLoad,
1814 (VT == MVT::f32) ? 4 : 8, Align);
1816 unsigned Opc = (VT == MVT::f32) ? PPC::LFS : PPC::LFD;
1817 unsigned TmpReg = createResultReg(&PPC::G8RC_and_G8RC_NOX0RegClass);
1819 // For small code model, generate a LF[SD](0, LDtocCPT(Idx, X2)).
1820 if (CModel == CodeModel::Small || CModel == CodeModel::JITDefault) {
1821 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::LDtocCPT),
1823 .addConstantPoolIndex(Idx).addReg(PPC::X2);
1824 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), DestReg)
1825 .addImm(0).addReg(TmpReg).addMemOperand(MMO);
1827 // Otherwise we generate LF[SD](Idx[lo], ADDIStocHA(X2, Idx)).
1828 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::ADDIStocHA),
1829 TmpReg).addReg(PPC::X2).addConstantPoolIndex(Idx);
1830 // But for large code model, we must generate a LDtocL followed
1832 if (CModel == CodeModel::Large) {
1833 unsigned TmpReg2 = createResultReg(&PPC::G8RC_and_G8RC_NOX0RegClass);
1834 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::LDtocL),
1835 TmpReg2).addConstantPoolIndex(Idx).addReg(TmpReg);
1836 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), DestReg)
1837 .addImm(0).addReg(TmpReg2);
1839 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), DestReg)
1840 .addConstantPoolIndex(Idx, 0, PPCII::MO_TOC_LO)
1842 .addMemOperand(MMO);
1848 // Materialize the address of a global value into a register, and return
1849 // the register number (or zero if we failed to handle it).
1850 unsigned PPCFastISel::PPCMaterializeGV(const GlobalValue *GV, MVT VT) {
1851 assert(VT == MVT::i64 && "Non-address!");
1852 const TargetRegisterClass *RC = &PPC::G8RC_and_G8RC_NOX0RegClass;
1853 unsigned DestReg = createResultReg(RC);
1855 // Global values may be plain old object addresses, TLS object
1856 // addresses, constant pool entries, or jump tables. How we generate
1857 // code for these may depend on small, medium, or large code model.
1858 CodeModel::Model CModel = TM.getCodeModel();
1860 // FIXME: Jump tables are not yet required because fast-isel doesn't
1861 // handle switches; if that changes, we need them as well. For now,
1862 // what follows assumes everything's a generic (or TLS) global address.
1863 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
1865 // If GV is an alias, use the aliasee for determining thread-locality.
1866 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
1867 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false));
1870 // FIXME: We don't yet handle the complexity of TLS.
1871 bool IsTLS = GVar && GVar->isThreadLocal();
1875 // For small code model, generate a simple TOC load.
1876 if (CModel == CodeModel::Small || CModel == CodeModel::JITDefault)
1877 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::LDtoc),
1879 .addGlobalAddress(GV)
1882 // If the address is an externally defined symbol, a symbol with
1883 // common or externally available linkage, a function address, or a
1884 // jump table address (not yet needed), or if we are generating code
1885 // for large code model, we generate:
1886 // LDtocL(GV, ADDIStocHA(%X2, GV))
1887 // Otherwise we generate:
1888 // ADDItocL(ADDIStocHA(%X2, GV), GV)
1889 // Either way, start with the ADDIStocHA:
1890 unsigned HighPartReg = createResultReg(RC);
1891 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::ADDIStocHA),
1892 HighPartReg).addReg(PPC::X2).addGlobalAddress(GV);
1894 // !GVar implies a function address. An external variable is one
1895 // without an initializer.
1896 // If/when switches are implemented, jump tables should be handled
1897 // on the "if" path here.
1898 if (CModel == CodeModel::Large || !GVar || !GVar->hasInitializer() ||
1899 GVar->hasCommonLinkage() || GVar->hasAvailableExternallyLinkage())
1900 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::LDtocL),
1901 DestReg).addGlobalAddress(GV).addReg(HighPartReg);
1903 // Otherwise generate the ADDItocL.
1904 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::ADDItocL),
1905 DestReg).addReg(HighPartReg).addGlobalAddress(GV);
1911 // Materialize a 32-bit integer constant into a register, and return
1912 // the register number (or zero if we failed to handle it).
1913 unsigned PPCFastISel::PPCMaterialize32BitInt(int64_t Imm,
1914 const TargetRegisterClass *RC) {
1915 unsigned Lo = Imm & 0xFFFF;
1916 unsigned Hi = (Imm >> 16) & 0xFFFF;
1918 unsigned ResultReg = createResultReg(RC);
1919 bool IsGPRC = RC->hasSuperClassEq(&PPC::GPRCRegClass);
1922 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1923 TII.get(IsGPRC ? PPC::LI : PPC::LI8), ResultReg)
1926 // Both Lo and Hi have nonzero bits.
1927 unsigned TmpReg = createResultReg(RC);
1928 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1929 TII.get(IsGPRC ? PPC::LIS : PPC::LIS8), TmpReg)
1931 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1932 TII.get(IsGPRC ? PPC::ORI : PPC::ORI8), ResultReg)
1933 .addReg(TmpReg).addImm(Lo);
1936 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1937 TII.get(IsGPRC ? PPC::LIS : PPC::LIS8), ResultReg)
1943 // Materialize a 64-bit integer constant into a register, and return
1944 // the register number (or zero if we failed to handle it).
1945 unsigned PPCFastISel::PPCMaterialize64BitInt(int64_t Imm,
1946 const TargetRegisterClass *RC) {
1947 unsigned Remainder = 0;
1950 // If the value doesn't fit in 32 bits, see if we can shift it
1951 // so that it fits in 32 bits.
1952 if (!isInt<32>(Imm)) {
1953 Shift = countTrailingZeros<uint64_t>(Imm);
1954 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
1956 if (isInt<32>(ImmSh))
1965 // Handle the high-order 32 bits (if shifted) or the whole 32 bits
1966 // (if not shifted).
1967 unsigned TmpReg1 = PPCMaterialize32BitInt(Imm, RC);
1971 // If upper 32 bits were not zero, we've built them and need to shift
1975 TmpReg2 = createResultReg(RC);
1976 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::RLDICR),
1977 TmpReg2).addReg(TmpReg1).addImm(Shift).addImm(63 - Shift);
1981 unsigned TmpReg3, Hi, Lo;
1982 if ((Hi = (Remainder >> 16) & 0xFFFF)) {
1983 TmpReg3 = createResultReg(RC);
1984 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::ORIS8),
1985 TmpReg3).addReg(TmpReg2).addImm(Hi);
1989 if ((Lo = Remainder & 0xFFFF)) {
1990 unsigned ResultReg = createResultReg(RC);
1991 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::ORI8),
1992 ResultReg).addReg(TmpReg3).addImm(Lo);
2000 // Materialize an integer constant into a register, and return
2001 // the register number (or zero if we failed to handle it).
2002 unsigned PPCFastISel::PPCMaterializeInt(const Constant *C, MVT VT) {
2003 // If we're using CR bit registers for i1 values, handle that as a special
2005 if (VT == MVT::i1 && PPCSubTarget.useCRBits()) {
2006 const ConstantInt *CI = cast<ConstantInt>(C);
2007 unsigned ImmReg = createResultReg(&PPC::CRBITRCRegClass);
2008 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2009 TII.get(CI->isZero() ? PPC::CRUNSET : PPC::CRSET), ImmReg);
2013 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16 &&
2014 VT != MVT::i8 && VT != MVT::i1)
2017 const TargetRegisterClass *RC = ((VT == MVT::i64) ? &PPC::G8RCRegClass :
2018 &PPC::GPRCRegClass);
2020 // If the constant is in range, use a load-immediate.
2021 const ConstantInt *CI = cast<ConstantInt>(C);
2022 if (isInt<16>(CI->getSExtValue())) {
2023 unsigned Opc = (VT == MVT::i64) ? PPC::LI8 : PPC::LI;
2024 unsigned ImmReg = createResultReg(RC);
2025 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ImmReg)
2026 .addImm(CI->getSExtValue());
2030 // Construct the constant piecewise.
2031 int64_t Imm = CI->getZExtValue();
2034 return PPCMaterialize64BitInt(Imm, RC);
2035 else if (VT == MVT::i32)
2036 return PPCMaterialize32BitInt(Imm, RC);
2041 // Materialize a constant into a register, and return the register
2042 // number (or zero if we failed to handle it).
2043 unsigned PPCFastISel::TargetMaterializeConstant(const Constant *C) {
2044 EVT CEVT = TLI.getValueType(C->getType(), true);
2046 // Only handle simple types.
2047 if (!CEVT.isSimple()) return 0;
2048 MVT VT = CEVT.getSimpleVT();
2050 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
2051 return PPCMaterializeFP(CFP, VT);
2052 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
2053 return PPCMaterializeGV(GV, VT);
2054 else if (isa<ConstantInt>(C))
2055 return PPCMaterializeInt(C, VT);
2060 // Materialize the address created by an alloca into a register, and
2061 // return the register number (or zero if we failed to handle it).
2062 unsigned PPCFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
2063 // Don't handle dynamic allocas.
2064 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
2067 if (!isLoadTypeLegal(AI->getType(), VT)) return 0;
2069 DenseMap<const AllocaInst*, int>::iterator SI =
2070 FuncInfo.StaticAllocaMap.find(AI);
2072 if (SI != FuncInfo.StaticAllocaMap.end()) {
2073 unsigned ResultReg = createResultReg(&PPC::G8RC_and_G8RC_NOX0RegClass);
2074 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::ADDI8),
2075 ResultReg).addFrameIndex(SI->second).addImm(0);
2082 // Fold loads into extends when possible.
2083 // FIXME: We can have multiple redundant extend/trunc instructions
2084 // following a load. The folding only picks up one. Extend this
2085 // to check subsequent instructions for the same pattern and remove
2086 // them. Thus ResultReg should be the def reg for the last redundant
2087 // instruction in a chain, and all intervening instructions can be
2088 // removed from parent. Change test/CodeGen/PowerPC/fast-isel-fold.ll
2089 // to add ELF64-NOT: rldicl to the appropriate tests when this works.
2090 bool PPCFastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
2091 const LoadInst *LI) {
2092 // Verify we have a legal type before going any further.
2094 if (!isLoadTypeLegal(LI->getType(), VT))
2097 // Combine load followed by zero- or sign-extend.
2098 bool IsZExt = false;
2099 switch(MI->getOpcode()) {
2104 case PPC::RLDICL_32_64: {
2106 unsigned MB = MI->getOperand(3).getImm();
2107 if ((VT == MVT::i8 && MB <= 56) ||
2108 (VT == MVT::i16 && MB <= 48) ||
2109 (VT == MVT::i32 && MB <= 32))
2115 case PPC::RLWINM8: {
2117 unsigned MB = MI->getOperand(3).getImm();
2118 if ((VT == MVT::i8 && MB <= 24) ||
2119 (VT == MVT::i16 && MB <= 16))
2126 case PPC::EXTSB8_32_64:
2127 /* There is no sign-extending load-byte instruction. */
2132 case PPC::EXTSH8_32_64: {
2133 if (VT != MVT::i16 && VT != MVT::i8)
2139 case PPC::EXTSW_32_64: {
2140 if (VT != MVT::i32 && VT != MVT::i16 && VT != MVT::i8)
2146 // See if we can handle this address.
2148 if (!PPCComputeAddress(LI->getOperand(0), Addr))
2151 unsigned ResultReg = MI->getOperand(0).getReg();
2153 if (!PPCEmitLoad(VT, ResultReg, Addr, 0, IsZExt))
2156 MI->eraseFromParent();
2160 // Attempt to lower call arguments in a faster way than done by
2161 // the selection DAG code.
2162 bool PPCFastISel::FastLowerArguments() {
2163 // Defer to normal argument lowering for now. It's reasonably
2164 // efficient. Consider doing something like ARM to handle the
2165 // case where all args fit in registers, no varargs, no float
2170 // Handle materializing integer constants into a register. This is not
2171 // automatically generated for PowerPC, so must be explicitly created here.
2172 unsigned PPCFastISel::FastEmit_i(MVT Ty, MVT VT, unsigned Opc, uint64_t Imm) {
2174 if (Opc != ISD::Constant)
2177 // If we're using CR bit registers for i1 values, handle that as a special
2179 if (VT == MVT::i1 && PPCSubTarget.useCRBits()) {
2180 unsigned ImmReg = createResultReg(&PPC::CRBITRCRegClass);
2181 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2182 TII.get(Imm == 0 ? PPC::CRUNSET : PPC::CRSET), ImmReg);
2186 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16 &&
2187 VT != MVT::i8 && VT != MVT::i1)
2190 const TargetRegisterClass *RC = ((VT == MVT::i64) ? &PPC::G8RCRegClass :
2191 &PPC::GPRCRegClass);
2193 return PPCMaterialize64BitInt(Imm, RC);
2195 return PPCMaterialize32BitInt(Imm, RC);
2198 // Override for ADDI and ADDI8 to set the correct register class
2199 // on RHS operand 0. The automatic infrastructure naively assumes
2200 // GPRC for i32 and G8RC for i64; the concept of "no R0" is lost
2201 // for these cases. At the moment, none of the other automatically
2202 // generated RI instructions require special treatment. However, once
2203 // SelectSelect is implemented, "isel" requires similar handling.
2205 // Also be conservative about the output register class. Avoid
2206 // assigning R0 or X0 to the output register for GPRC and G8RC
2207 // register classes, as any such result could be used in ADDI, etc.,
2208 // where those regs have another meaning.
2209 unsigned PPCFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
2210 const TargetRegisterClass *RC,
2211 unsigned Op0, bool Op0IsKill,
2213 if (MachineInstOpcode == PPC::ADDI)
2214 MRI.setRegClass(Op0, &PPC::GPRC_and_GPRC_NOR0RegClass);
2215 else if (MachineInstOpcode == PPC::ADDI8)
2216 MRI.setRegClass(Op0, &PPC::G8RC_and_G8RC_NOX0RegClass);
2218 const TargetRegisterClass *UseRC =
2219 (RC == &PPC::GPRCRegClass ? &PPC::GPRC_and_GPRC_NOR0RegClass :
2220 (RC == &PPC::G8RCRegClass ? &PPC::G8RC_and_G8RC_NOX0RegClass : RC));
2222 return FastISel::FastEmitInst_ri(MachineInstOpcode, UseRC,
2223 Op0, Op0IsKill, Imm);
2226 // Override for instructions with one register operand to avoid use of
2227 // R0/X0. The automatic infrastructure isn't aware of the context so
2228 // we must be conservative.
2229 unsigned PPCFastISel::FastEmitInst_r(unsigned MachineInstOpcode,
2230 const TargetRegisterClass* RC,
2231 unsigned Op0, bool Op0IsKill) {
2232 const TargetRegisterClass *UseRC =
2233 (RC == &PPC::GPRCRegClass ? &PPC::GPRC_and_GPRC_NOR0RegClass :
2234 (RC == &PPC::G8RCRegClass ? &PPC::G8RC_and_G8RC_NOX0RegClass : RC));
2236 return FastISel::FastEmitInst_r(MachineInstOpcode, UseRC, Op0, Op0IsKill);
2239 // Override for instructions with two register operands to avoid use
2240 // of R0/X0. The automatic infrastructure isn't aware of the context
2241 // so we must be conservative.
2242 unsigned PPCFastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
2243 const TargetRegisterClass* RC,
2244 unsigned Op0, bool Op0IsKill,
2245 unsigned Op1, bool Op1IsKill) {
2246 const TargetRegisterClass *UseRC =
2247 (RC == &PPC::GPRCRegClass ? &PPC::GPRC_and_GPRC_NOR0RegClass :
2248 (RC == &PPC::G8RCRegClass ? &PPC::G8RC_and_G8RC_NOX0RegClass : RC));
2250 return FastISel::FastEmitInst_rr(MachineInstOpcode, UseRC, Op0, Op0IsKill,
2255 // Create the fast instruction selector for PowerPC64 ELF.
2256 FastISel *PPC::createFastISel(FunctionLoweringInfo &FuncInfo,
2257 const TargetLibraryInfo *LibInfo) {
2258 const TargetMachine &TM = FuncInfo.MF->getTarget();
2260 // Only available on 64-bit ELF for now.
2261 const PPCSubtarget *Subtarget = &TM.getSubtarget<PPCSubtarget>();
2262 if (Subtarget->isPPC64() && Subtarget->isSVR4ABI())
2263 return new PPCFastISel(FuncInfo, LibInfo);