1 //===-- PPC64ISelPattern.cpp - A pattern matching inst selector for PPC64 -===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Nate Begeman and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a pattern matching instruction selector for 64 bit PowerPC.
12 //===----------------------------------------------------------------------===//
15 #include "PowerPCInstrBuilder.h"
16 #include "PowerPCInstrInfo.h"
17 #include "PPC64RegisterInfo.h"
18 #include "llvm/Constants.h" // FIXME: REMOVE
19 #include "llvm/Function.h"
20 #include "llvm/CodeGen/MachineConstantPool.h" // FIXME: REMOVE
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/SelectionDAG.h"
24 #include "llvm/CodeGen/SelectionDAGISel.h"
25 #include "llvm/CodeGen/SSARegMap.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Target/TargetLowering.h"
28 #include "llvm/Target/TargetOptions.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/MathExtras.h"
31 #include "llvm/ADT/Statistic.h"
36 //===----------------------------------------------------------------------===//
37 // PPC32TargetLowering - PPC32 Implementation of the TargetLowering interface
39 class PPC64TargetLowering : public TargetLowering {
40 int VarArgsFrameIndex; // FrameIndex for start of varargs area.
41 int ReturnAddrIndex; // FrameIndex for return slot.
43 PPC64TargetLowering(TargetMachine &TM) : TargetLowering(TM) {
44 // Set up the register classes.
45 addRegisterClass(MVT::i64, PPC64::GPRCRegisterClass);
46 addRegisterClass(MVT::f32, PPC64::FPRCRegisterClass);
47 addRegisterClass(MVT::f64, PPC64::FPRCRegisterClass);
49 // PowerPC has no intrinsics for these particular operations
50 setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
51 setOperationAction(ISD::MEMSET, MVT::Other, Expand);
52 setOperationAction(ISD::MEMCPY, MVT::Other, Expand);
54 // PPC 64 has i16 and i32 but no i8 (or i1) SEXTLOAD
55 setOperationAction(ISD::SEXTLOAD, MVT::i1, Expand);
56 setOperationAction(ISD::SEXTLOAD, MVT::i8, Expand);
58 // PowerPC has no SREM/UREM instructions
59 setOperationAction(ISD::SREM, MVT::i64, Expand);
60 setOperationAction(ISD::UREM, MVT::i64, Expand);
62 setShiftAmountFlavor(Extend); // shl X, 32 == 0
63 addLegalFPImmediate(+0.0); // Necessary for FSEL
64 addLegalFPImmediate(-0.0); //
66 computeRegisterProperties();
69 /// LowerArguments - This hook must be implemented to indicate how we should
70 /// lower the arguments for the specified function, into the specified DAG.
71 virtual std::vector<SDOperand>
72 LowerArguments(Function &F, SelectionDAG &DAG);
74 /// LowerCallTo - This hook lowers an abstract call to a function into an
76 virtual std::pair<SDOperand, SDOperand>
77 LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg,
78 SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG);
80 virtual std::pair<SDOperand, SDOperand>
81 LowerVAStart(SDOperand Chain, SelectionDAG &DAG);
83 virtual std::pair<SDOperand,SDOperand>
84 LowerVAArgNext(bool isVANext, SDOperand Chain, SDOperand VAList,
85 const Type *ArgTy, SelectionDAG &DAG);
87 virtual std::pair<SDOperand, SDOperand>
88 LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
94 std::vector<SDOperand>
95 PPC64TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
97 // add beautiful description of PPC stack frame format, or at least some docs
99 MachineFunction &MF = DAG.getMachineFunction();
100 MachineFrameInfo *MFI = MF.getFrameInfo();
101 MachineBasicBlock& BB = MF.front();
102 std::vector<SDOperand> ArgValues;
104 // Due to the rather complicated nature of the PowerPC ABI, rather than a
105 // fixed size array of physical args, for the sake of simplicity let the STL
106 // handle tracking them for us.
107 std::vector<unsigned> argVR, argPR, argOp;
108 unsigned ArgOffset = 48;
109 unsigned GPR_remaining = 8;
110 unsigned FPR_remaining = 13;
111 unsigned GPR_idx = 0, FPR_idx = 0;
112 static const unsigned GPR[] = {
113 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
114 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
116 static const unsigned FPR[] = {
117 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
118 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
121 // Add DAG nodes to load the arguments... On entry to a function on PPC,
122 // the arguments start at offset 48, although they are likely to be passed
124 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
125 SDOperand newroot, argt;
126 bool needsLoad = false;
127 MVT::ValueType ObjectVT = getValueType(I->getType());
130 default: assert(0 && "Unhandled argument type!");
136 if (GPR_remaining > 0) {
137 BuildMI(&BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]);
138 argt = newroot = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32,
140 if (ObjectVT != MVT::i64)
141 argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, newroot);
148 if (FPR_remaining > 0) {
149 BuildMI(&BB, PPC::IMPLICIT_DEF, 0, FPR[FPR_idx]);
150 argt = newroot = DAG.getCopyFromReg(FPR[FPR_idx], ObjectVT,
160 // We need to load the argument to a virtual register if we determined above
161 // that we ran out of physical registers of the appropriate type
163 unsigned SubregOffset = 0;
165 default: assert(0 && "Unhandled argument type!");
167 case MVT::i8: SubregOffset = 7; break;
168 case MVT::i16: SubregOffset = 6; break;
170 case MVT::f32: SubregOffset = 4; break;
172 case MVT::f64: SubregOffset = 0; break;
174 int FI = MFI->CreateFixedObject(8, ArgOffset);
175 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i64);
176 FIN = DAG.getNode(ISD::ADD, MVT::i64, FIN,
177 DAG.getConstant(SubregOffset, MVT::i64));
178 argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN);
181 // Every 4 bytes of argument space consumes one of the GPRs available for
183 if (GPR_remaining > 0) {
189 DAG.setRoot(newroot.getValue(1));
190 ArgValues.push_back(argt);
193 // If the function takes variable number of arguments, make a frame index for
194 // the start of the first vararg value... for expansion of llvm.va_start.
196 VarArgsFrameIndex = MFI->CreateFixedObject(8, ArgOffset);
197 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i64);
198 // If this function is vararg, store any remaining integer argument regs
199 // to their spots on the stack so that they may be loaded by deferencing the
200 // result of va_next.
201 std::vector<SDOperand> MemOps;
202 for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) {
203 BuildMI(&BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]);
204 SDOperand Val = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i64, DAG.getRoot());
205 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
207 MemOps.push_back(Store);
208 // Increment the address by eight for the next argument to store
209 SDOperand PtrOff = DAG.getConstant(8, getPointerTy());
210 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff);
212 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps));
218 std::pair<SDOperand, SDOperand>
219 PPC64TargetLowering::LowerCallTo(SDOperand Chain,
220 const Type *RetTy, bool isVarArg,
221 SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG) {
222 // args_to_use will accumulate outgoing args for the ISD::CALL case in
223 // SelectExpr to use to put the arguments in the appropriate registers.
224 std::vector<SDOperand> args_to_use;
226 // Count how many bytes are to be pushed on the stack, including the linkage
227 // area, and parameter passing area.
228 unsigned NumBytes = 48;
231 Chain = DAG.getNode(ISD::ADJCALLSTACKDOWN, MVT::Other, Chain,
232 DAG.getConstant(NumBytes, getPointerTy()));
234 NumBytes = 8 * Args.size(); // All arguments are rounded up to 8 bytes
236 // Just to be safe, we'll always reserve the full 48 bytes of linkage area
237 // plus 64 bytes of argument space in case any called code gets funky on us.
238 if (NumBytes < 112) NumBytes = 112;
240 // Adjust the stack pointer for the new arguments...
241 // These operations are automatically eliminated by the prolog/epilog pass
242 Chain = DAG.getNode(ISD::ADJCALLSTACKDOWN, MVT::Other, Chain,
243 DAG.getConstant(NumBytes, getPointerTy()));
245 // Set up a copy of the stack pointer for use loading and storing any
246 // arguments that may not fit in the registers available for argument
248 SDOperand StackPtr = DAG.getCopyFromReg(PPC::R1, MVT::i32,
251 // Figure out which arguments are going to go in registers, and which in
252 // memory. Also, if this is a vararg function, floating point operations
253 // must be stored to our stack, and loaded into integer regs as well, if
254 // any integer regs are available for argument passing.
255 unsigned ArgOffset = 48;
256 unsigned GPR_remaining = 8;
257 unsigned FPR_remaining = 13;
259 std::vector<SDOperand> MemOps;
260 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
261 // PtrOff will be used to store the current argument to the stack if a
262 // register cannot be found for it.
263 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
264 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
265 MVT::ValueType ArgVT = getValueType(Args[i].second);
268 default: assert(0 && "Unexpected ValueType for argument!");
273 // Promote the integer to 64 bits. If the input type is signed use a
274 // sign extend, otherwise use a zero extend.
275 if (Args[i].second->isSigned())
276 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i64, Args[i].first);
278 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i64, Args[i].first);
281 if (GPR_remaining > 0) {
282 args_to_use.push_back(Args[i].first);
285 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
286 Args[i].first, PtrOff));
292 if (FPR_remaining > 0) {
293 args_to_use.push_back(Args[i].first);
296 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
297 Args[i].first, PtrOff);
298 MemOps.push_back(Store);
299 // Float varargs are always shadowed in available integer registers
300 if (GPR_remaining > 0) {
301 SDOperand Load = DAG.getLoad(MVT::i64, Store, PtrOff);
302 MemOps.push_back(Load);
303 args_to_use.push_back(Load);
307 // If we have any FPRs remaining, we may also have GPRs remaining.
308 // Args passed in FPRs also consume an available GPR.
309 if (GPR_remaining > 0) {
310 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i64));
315 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
316 Args[i].first, PtrOff));
323 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
326 std::vector<MVT::ValueType> RetVals;
327 MVT::ValueType RetTyVT = getValueType(RetTy);
328 if (RetTyVT != MVT::isVoid)
329 RetVals.push_back(RetTyVT);
330 RetVals.push_back(MVT::Other);
332 SDOperand TheCall = SDOperand(DAG.getCall(RetVals,
333 Chain, Callee, args_to_use), 0);
334 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
335 Chain = DAG.getNode(ISD::ADJCALLSTACKUP, MVT::Other, Chain,
336 DAG.getConstant(NumBytes, getPointerTy()));
337 return std::make_pair(TheCall, Chain);
340 std::pair<SDOperand, SDOperand>
341 PPC64TargetLowering::LowerVAStart(SDOperand Chain, SelectionDAG &DAG) {
342 //vastart just returns the address of the VarArgsFrameIndex slot.
343 return std::make_pair(DAG.getFrameIndex(VarArgsFrameIndex, MVT::i64), Chain);
346 std::pair<SDOperand,SDOperand> PPC64TargetLowering::
347 LowerVAArgNext(bool isVANext, SDOperand Chain, SDOperand VAList,
348 const Type *ArgTy, SelectionDAG &DAG) {
349 MVT::ValueType ArgVT = getValueType(ArgTy);
352 Result = DAG.getLoad(ArgVT, DAG.getEntryNode(), VAList);
354 Result = DAG.getNode(ISD::ADD, VAList.getValueType(), VAList,
355 DAG.getConstant(8, VAList.getValueType()));
357 return std::make_pair(Result, Chain);
361 std::pair<SDOperand, SDOperand> PPC64TargetLowering::
362 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
364 assert(0 && "LowerFrameReturnAddress unimplemented");
369 Statistic<>NotLogic("ppc-codegen", "Number of inverted logical ops");
370 Statistic<>FusedFP("ppc-codegen", "Number of fused fp operations");
371 //===--------------------------------------------------------------------===//
372 /// ISel - PPC32 specific code to select PPC32 machine instructions for
373 /// SelectionDAG operations.
374 //===--------------------------------------------------------------------===//
375 class ISel : public SelectionDAGISel {
378 PPC64TargetLowering PPC64Lowering;
380 /// ExprMap - As shared expressions are codegen'd, we keep track of which
381 /// vreg the value is produced in, so we only emit one copy of each compiled
383 std::map<SDOperand, unsigned> ExprMap;
385 unsigned GlobalBaseReg;
386 bool GlobalBaseInitialized;
389 ISel(TargetMachine &TM) : SelectionDAGISel(PPC64Lowering), PPC64Lowering(TM)
392 /// runOnFunction - Override this function in order to reset our per-function
394 virtual bool runOnFunction(Function &Fn) {
395 // Make sure we re-emit a set of the global base reg if necessary
396 GlobalBaseInitialized = false;
397 return SelectionDAGISel::runOnFunction(Fn);
400 /// InstructionSelectBasicBlock - This callback is invoked by
401 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
402 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG) {
404 // Codegen the basic block.
405 Select(DAG.getRoot());
407 // Clear state used for selection.
411 unsigned getGlobalBaseReg();
412 unsigned getConstDouble(double floatVal, unsigned Result);
413 unsigned SelectSetCR0(SDOperand CC);
414 unsigned SelectExpr(SDOperand N);
415 unsigned SelectExprFP(SDOperand N, unsigned Result);
416 void Select(SDOperand N);
418 bool SelectAddr(SDOperand N, unsigned& Reg, int& offset);
419 void SelectBranchCC(SDOperand N);
422 /// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
423 /// returns zero when the input is not exactly a power of two.
424 static unsigned ExactLog2(unsigned Val) {
425 if (Val == 0 || (Val & (Val-1))) return 0;
434 /// getImmediateForOpcode - This method returns a value indicating whether
435 /// the ConstantSDNode N can be used as an immediate to Opcode. The return
436 /// values are either 0, 1 or 2. 0 indicates that either N is not a
437 /// ConstantSDNode, or is not suitable for use by that opcode. A return value
438 /// of 1 indicates that the constant may be used in normal immediate form. A
439 /// return value of 2 indicates that the constant may be used in shifted
440 /// immediate form. A return value of 3 indicates that log base 2 of the
441 /// constant may be used.
443 static unsigned getImmediateForOpcode(SDOperand N, unsigned Opcode,
444 unsigned& Imm, bool U = false) {
445 if (N.getOpcode() != ISD::Constant) return 0;
447 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
452 if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
453 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
458 if (v >= 0 && v <= 65535) { Imm = v & 0xFFFF; return 1; }
459 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
463 if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
466 if (U && (v >= 0 && v <= 65535)) { Imm = v & 0xFFFF; return 1; }
467 if (!U && (v <= 32767 && v >= -32768)) { Imm = v & 0xFFFF; return 1; }
470 if ((Imm = ExactLog2(v))) { return 3; }
476 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
477 /// to Condition. If the Condition is unordered or unsigned, the bool argument
478 /// U is set to true, otherwise it is set to false.
479 static unsigned getBCCForSetCC(unsigned Condition, bool& U) {
482 default: assert(0 && "Unknown condition!"); abort();
483 case ISD::SETEQ: return PPC::BEQ;
484 case ISD::SETNE: return PPC::BNE;
485 case ISD::SETULT: U = true;
486 case ISD::SETLT: return PPC::BLT;
487 case ISD::SETULE: U = true;
488 case ISD::SETLE: return PPC::BLE;
489 case ISD::SETUGT: U = true;
490 case ISD::SETGT: return PPC::BGT;
491 case ISD::SETUGE: U = true;
492 case ISD::SETGE: return PPC::BGE;
497 /// IndexedOpForOp - Return the indexed variant for each of the PowerPC load
498 /// and store immediate instructions.
499 static unsigned IndexedOpForOp(unsigned Opcode) {
501 default: assert(0 && "Unknown opcode!"); abort();
502 case PPC::LBZ: return PPC::LBZX; case PPC::STB: return PPC::STBX;
503 case PPC::LHZ: return PPC::LHZX; case PPC::STH: return PPC::STHX;
504 case PPC::LHA: return PPC::LHAX; case PPC::STW: return PPC::STWX;
505 case PPC::LWZ: return PPC::LWZX; case PPC::STD: return PPC::STDX;
506 case PPC::LD: return PPC::LDX; case PPC::STFS: return PPC::STFSX;
507 case PPC::LFS: return PPC::LFSX; case PPC::STFD: return PPC::STFDX;
508 case PPC::LFD: return PPC::LFDX;
514 /// getGlobalBaseReg - Output the instructions required to put the
515 /// base address to use for accessing globals into a register.
517 unsigned ISel::getGlobalBaseReg() {
518 if (!GlobalBaseInitialized) {
519 // Insert the set of GlobalBaseReg into the first MBB of the function
520 MachineBasicBlock &FirstMBB = BB->getParent()->front();
521 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
522 GlobalBaseReg = MakeReg(MVT::i64);
523 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
524 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg).addReg(PPC::LR);
525 GlobalBaseInitialized = true;
527 return GlobalBaseReg;
530 /// getConstDouble - Loads a floating point value into a register, via the
531 /// Constant Pool. Optionally takes a register in which to load the value.
532 unsigned ISel::getConstDouble(double doubleVal, unsigned Result=0) {
533 unsigned Tmp1 = MakeReg(MVT::i64);
534 if (0 == Result) Result = MakeReg(MVT::f64);
535 MachineConstantPool *CP = BB->getParent()->getConstantPool();
536 ConstantFP *CFP = ConstantFP::get(Type::DoubleTy, doubleVal);
537 unsigned CPI = CP->getConstantPoolIndex(CFP);
538 BuildMI(BB, PPC::LOADHiAddr, 2, Tmp1).addReg(getGlobalBaseReg())
539 .addConstantPoolIndex(CPI);
540 BuildMI(BB, PPC::LFD, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
544 unsigned ISel::SelectSetCR0(SDOperand CC) {
545 unsigned Opc, Tmp1, Tmp2;
546 static const unsigned CompareOpcodes[] =
547 { PPC::FCMPU, PPC::FCMPU, PPC::CMPW, PPC::CMPLW };
549 // If the first operand to the select is a SETCC node, then we can fold it
550 // into the branch that selects which value to return.
551 SetCCSDNode* SetCC = dyn_cast<SetCCSDNode>(CC.Val);
552 if (SetCC && CC.getOpcode() == ISD::SETCC) {
554 Opc = getBCCForSetCC(SetCC->getCondition(), U);
555 Tmp1 = SelectExpr(SetCC->getOperand(0));
557 // Pass the optional argument U to getImmediateForOpcode for SETCC,
558 // so that it knows whether the SETCC immediate range is signed or not.
559 if (1 == getImmediateForOpcode(SetCC->getOperand(1), ISD::SETCC,
562 BuildMI(BB, PPC::CMPLWI, 2, PPC::CR0).addReg(Tmp1).addImm(Tmp2);
564 BuildMI(BB, PPC::CMPWI, 2, PPC::CR0).addReg(Tmp1).addSImm(Tmp2);
566 bool IsInteger = MVT::isInteger(SetCC->getOperand(0).getValueType());
567 unsigned CompareOpc = CompareOpcodes[2 * IsInteger + U];
568 Tmp2 = SelectExpr(SetCC->getOperand(1));
569 BuildMI(BB, CompareOpc, 2, PPC::CR0).addReg(Tmp1).addReg(Tmp2);
572 Tmp1 = SelectExpr(CC);
573 BuildMI(BB, PPC::CMPLWI, 2, PPC::CR0).addReg(Tmp1).addImm(0);
579 /// Check to see if the load is a constant offset from a base register
580 bool ISel::SelectAddr(SDOperand N, unsigned& Reg, int& offset)
582 unsigned imm = 0, opcode = N.getOpcode();
583 if (N.getOpcode() == ISD::ADD) {
584 Reg = SelectExpr(N.getOperand(0));
585 if (1 == getImmediateForOpcode(N.getOperand(1), opcode, imm)) {
589 offset = SelectExpr(N.getOperand(1));
597 void ISel::SelectBranchCC(SDOperand N)
599 assert(N.getOpcode() == ISD::BRCOND && "Not a BranchCC???");
600 MachineBasicBlock *Dest =
601 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
603 // Get the MBB we will fall through to so that we can hand it off to the
604 // branch selection pass as an argument to the PPC::COND_BRANCH pseudo op.
605 //ilist<MachineBasicBlock>::iterator It = BB;
606 //MachineBasicBlock *Fallthrough = ++It;
608 Select(N.getOperand(0)); //chain
609 unsigned Opc = SelectSetCR0(N.getOperand(1));
610 // FIXME: Use this once we have something approximating two-way branches
611 // We cannot currently use this in case the ISel hands us something like
614 // since the fallthrough basic block for the conditional branch does not start
615 // with the unconditional branch (it is skipped over).
616 //BuildMI(BB, PPC::COND_BRANCH, 4).addReg(PPC::CR0).addImm(Opc)
617 // .addMBB(Dest).addMBB(Fallthrough);
618 BuildMI(BB, Opc, 2).addReg(PPC::CR0).addMBB(Dest);
622 unsigned ISel::SelectExprFP(SDOperand N, unsigned Result)
624 unsigned Tmp1, Tmp2, Tmp3;
626 SDNode *Node = N.Val;
627 MVT::ValueType DestType = N.getValueType();
628 unsigned opcode = N.getOpcode();
633 assert(0 && "Node not handled!\n");
636 // Attempt to generate FSEL. We can do this whenever we have an FP result,
637 // and an FP comparison in the SetCC node.
638 SetCCSDNode* SetCC = dyn_cast<SetCCSDNode>(N.getOperand(0).Val);
639 if (SetCC && N.getOperand(0).getOpcode() == ISD::SETCC &&
640 !MVT::isInteger(SetCC->getOperand(0).getValueType()) &&
641 SetCC->getCondition() != ISD::SETEQ &&
642 SetCC->getCondition() != ISD::SETNE) {
643 MVT::ValueType VT = SetCC->getOperand(0).getValueType();
644 Tmp1 = SelectExpr(SetCC->getOperand(0)); // Val to compare against
645 unsigned TV = SelectExpr(N.getOperand(1)); // Use if TRUE
646 unsigned FV = SelectExpr(N.getOperand(2)); // Use if FALSE
648 ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(SetCC->getOperand(1));
649 if (CN && (CN->isExactlyValue(-0.0) || CN->isExactlyValue(0.0))) {
650 switch(SetCC->getCondition()) {
651 default: assert(0 && "Invalid FSEL condition"); abort();
654 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp1).addReg(FV).addReg(TV);
658 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp1).addReg(TV).addReg(FV);
663 BuildMI(BB, PPC::FNEG, 1, Tmp2).addReg(Tmp1);
664 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp2).addReg(FV).addReg(TV);
670 BuildMI(BB, PPC::FNEG, 1, Tmp2).addReg(Tmp1);
671 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp2).addReg(TV).addReg(FV);
676 Opc = (MVT::f64 == VT) ? PPC::FSUB : PPC::FSUBS;
677 Tmp2 = SelectExpr(SetCC->getOperand(1));
679 switch(SetCC->getCondition()) {
680 default: assert(0 && "Invalid FSEL condition"); abort();
683 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
684 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
688 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
689 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
693 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
694 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
698 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
699 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
703 assert(0 && "Should never get here");
707 unsigned TrueValue = SelectExpr(N.getOperand(1)); //Use if TRUE
708 unsigned FalseValue = SelectExpr(N.getOperand(2)); //Use if FALSE
709 Opc = SelectSetCR0(N.getOperand(0));
711 // Create an iterator with which to insert the MBB for copying the false
712 // value and the MBB to hold the PHI instruction for this SetCC.
713 MachineBasicBlock *thisMBB = BB;
714 const BasicBlock *LLVM_BB = BB->getBasicBlock();
715 ilist<MachineBasicBlock>::iterator It = BB;
723 // fallthrough --> copy0MBB
724 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
725 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
726 BuildMI(BB, Opc, 2).addReg(PPC::CR0).addMBB(sinkMBB);
727 MachineFunction *F = BB->getParent();
728 F->getBasicBlockList().insert(It, copy0MBB);
729 F->getBasicBlockList().insert(It, sinkMBB);
730 // Update machine-CFG edges
731 BB->addSuccessor(copy0MBB);
732 BB->addSuccessor(sinkMBB);
736 // # fallthrough to sinkMBB
738 // Update machine-CFG edges
739 BB->addSuccessor(sinkMBB);
742 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
745 BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue)
746 .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB);
751 if (!NoExcessFPPrecision &&
752 ISD::ADD == N.getOperand(0).getOpcode() &&
753 N.getOperand(0).Val->hasOneUse() &&
754 ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() &&
755 N.getOperand(0).getOperand(0).Val->hasOneUse()) {
756 ++FusedFP; // Statistic
757 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0));
758 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1));
759 Tmp3 = SelectExpr(N.getOperand(0).getOperand(1));
760 Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
761 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
762 } else if (!NoExcessFPPrecision &&
763 ISD::SUB == N.getOperand(0).getOpcode() &&
764 N.getOperand(0).Val->hasOneUse() &&
765 ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() &&
766 N.getOperand(0).getOperand(0).Val->hasOneUse()) {
767 ++FusedFP; // Statistic
768 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0));
769 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1));
770 Tmp3 = SelectExpr(N.getOperand(0).getOperand(1));
771 Opc = DestType == MVT::f64 ? PPC::FNMSUB : PPC::FNMSUBS;
772 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
773 } else if (ISD::FABS == N.getOperand(0).getOpcode()) {
774 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
775 BuildMI(BB, PPC::FNABS, 1, Result).addReg(Tmp1);
777 Tmp1 = SelectExpr(N.getOperand(0));
778 BuildMI(BB, PPC::FNEG, 1, Result).addReg(Tmp1);
783 Tmp1 = SelectExpr(N.getOperand(0));
784 BuildMI(BB, PPC::FABS, 1, Result).addReg(Tmp1);
788 assert (DestType == MVT::f32 &&
789 N.getOperand(0).getValueType() == MVT::f64 &&
790 "only f64 to f32 conversion supported here");
791 Tmp1 = SelectExpr(N.getOperand(0));
792 BuildMI(BB, PPC::FRSP, 1, Result).addReg(Tmp1);
796 assert (DestType == MVT::f64 &&
797 N.getOperand(0).getValueType() == MVT::f32 &&
798 "only f32 to f64 conversion supported here");
799 Tmp1 = SelectExpr(N.getOperand(0));
800 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
803 case ISD::CopyFromReg:
805 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
806 Tmp1 = dyn_cast<RegSDNode>(Node)->getReg();
807 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
810 case ISD::ConstantFP: {
811 ConstantFPSDNode *CN = cast<ConstantFPSDNode>(N);
812 Result = getConstDouble(CN->getValue(), Result);
817 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
818 N.getOperand(0).Val->hasOneUse()) {
819 ++FusedFP; // Statistic
820 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
821 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
822 Tmp3 = SelectExpr(N.getOperand(1));
823 Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
824 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
827 Opc = DestType == MVT::f64 ? PPC::FADD : PPC::FADDS;
828 Tmp1 = SelectExpr(N.getOperand(0));
829 Tmp2 = SelectExpr(N.getOperand(1));
830 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
834 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
835 N.getOperand(0).Val->hasOneUse()) {
836 ++FusedFP; // Statistic
837 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
838 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
839 Tmp3 = SelectExpr(N.getOperand(1));
840 Opc = DestType == MVT::f64 ? PPC::FMSUB : PPC::FMSUBS;
841 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
844 Opc = DestType == MVT::f64 ? PPC::FSUB : PPC::FSUBS;
845 Tmp1 = SelectExpr(N.getOperand(0));
846 Tmp2 = SelectExpr(N.getOperand(1));
847 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
853 case ISD::MUL: Opc = DestType == MVT::f64 ? PPC::FMUL : PPC::FMULS; break;
854 case ISD::SDIV: Opc = DestType == MVT::f64 ? PPC::FDIV : PPC::FDIVS; break;
856 Tmp1 = SelectExpr(N.getOperand(0));
857 Tmp2 = SelectExpr(N.getOperand(1));
858 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
861 case ISD::UINT_TO_FP:
862 case ISD::SINT_TO_FP: {
863 bool IsUnsigned = (ISD::UINT_TO_FP == opcode);
864 Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
865 Tmp2 = MakeReg(MVT::f64); // temp reg to load the integer value into
866 Tmp3 = MakeReg(MVT::i64); // temp reg to hold the conversion constant
867 unsigned ConstF = MakeReg(MVT::f64); // temp reg to hold the fp constant
869 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
870 MachineConstantPool *CP = BB->getParent()->getConstantPool();
872 // FIXME: pull this FP constant generation stuff out into something like
873 // the simple ISel's getReg.
875 addFrameReference(BuildMI(BB, PPC::STD, 3).addReg(Tmp1), FrameIdx);
876 addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
877 BuildMI(BB, PPC::FCFID, 1, Result).addReg(Tmp2);
879 ConstantFP *CFP = ConstantFP::get(Type::DoubleTy, 0x1.000008p52);
880 unsigned CPI = CP->getConstantPoolIndex(CFP);
881 // Load constant fp value
882 unsigned Tmp4 = MakeReg(MVT::i32);
883 unsigned TmpL = MakeReg(MVT::i32);
884 BuildMI(BB, PPC::LOADHiAddr, 2, Tmp4).addReg(getGlobalBaseReg())
885 .addConstantPoolIndex(CPI);
886 BuildMI(BB, PPC::LFD, 2, ConstF).addConstantPoolIndex(CPI).addReg(Tmp4);
887 // Store the hi & low halves of the fp value, currently in int regs
888 BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330);
889 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx);
890 BuildMI(BB, PPC::XORIS, 2, TmpL).addReg(Tmp1).addImm(0x8000);
891 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(TmpL), FrameIdx, 4);
892 addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
893 // Generate the return value with a subtract
894 BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF);
899 assert(0 && "Should never get here");
903 unsigned ISel::SelectExpr(SDOperand N) {
905 unsigned Tmp1, Tmp2, Tmp3;
907 unsigned opcode = N.getOpcode();
909 SDNode *Node = N.Val;
910 MVT::ValueType DestType = N.getValueType();
912 unsigned &Reg = ExprMap[N];
915 switch (N.getOpcode()) {
917 Reg = Result = (N.getValueType() != MVT::Other) ?
918 MakeReg(N.getValueType()) : 1;
921 // If this is a call instruction, make sure to prepare ALL of the result
922 // values as well as the chain.
923 if (Node->getNumValues() == 1)
924 Reg = Result = 1; // Void call, just a chain.
926 Result = MakeReg(Node->getValueType(0));
927 ExprMap[N.getValue(0)] = Result;
928 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
929 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
930 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
935 if (ISD::CopyFromReg == opcode)
936 DestType = N.getValue(0).getValueType();
938 if (DestType == MVT::f64 || DestType == MVT::f32)
939 if (ISD::LOAD != opcode && ISD::EXTLOAD != opcode && ISD::UNDEF != opcode)
940 return SelectExprFP(N, Result);
945 assert(0 && "Node not handled!\n");
947 BuildMI(BB, PPC::IMPLICIT_DEF, 0, Result);
949 case ISD::DYNAMIC_STACKALLOC:
950 // Generate both result values. FIXME: Need a better commment here?
952 ExprMap[N.getValue(1)] = 1;
954 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
956 // FIXME: We are currently ignoring the requested alignment for handling
957 // greater than the stack alignment. This will need to be revisited at some
958 // point. Align = N.getOperand(2);
959 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
960 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
961 std::cerr << "Cannot allocate stack object with greater alignment than"
962 << " the stack alignment yet!";
965 Select(N.getOperand(0));
966 Tmp1 = SelectExpr(N.getOperand(1));
967 // Subtract size from stack pointer, thereby allocating some space.
968 BuildMI(BB, PPC::SUBF, 2, PPC::R1).addReg(Tmp1).addReg(PPC::R1);
969 // Put a pointer to the space into the result register by copying the SP
970 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R1).addReg(PPC::R1);
973 case ISD::ConstantPool:
974 Tmp1 = cast<ConstantPoolSDNode>(N)->getIndex();
975 Tmp2 = MakeReg(MVT::i64);
976 BuildMI(BB, PPC::LOADHiAddr, 2, Tmp2).addReg(getGlobalBaseReg())
977 .addConstantPoolIndex(Tmp1);
978 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp2).addConstantPoolIndex(Tmp1);
981 case ISD::FrameIndex:
982 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
983 addFrameReference(BuildMI(BB, PPC::ADDI, 2, Result), (int)Tmp1, 0, false);
986 case ISD::GlobalAddress: {
987 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
988 Tmp1 = MakeReg(MVT::i64);
989 BuildMI(BB, PPC::LOADHiAddr, 2, Tmp1).addReg(getGlobalBaseReg())
990 .addGlobalAddress(GV);
991 if (GV->hasWeakLinkage() || GV->isExternal()) {
992 BuildMI(BB, PPC::LD, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
994 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp1).addGlobalAddress(GV);
1002 case ISD::SEXTLOAD: {
1003 MVT::ValueType TypeBeingLoaded = (ISD::LOAD == opcode) ?
1004 Node->getValueType(0) : cast<MVTSDNode>(Node)->getExtraValueType();
1005 bool sext = (ISD::SEXTLOAD == opcode);
1007 // Make sure we generate both values.
1009 ExprMap[N.getValue(1)] = 1; // Generate the token
1011 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1013 SDOperand Chain = N.getOperand(0);
1014 SDOperand Address = N.getOperand(1);
1017 switch (TypeBeingLoaded) {
1018 default: Node->dump(); assert(0 && "Cannot load this type!");
1019 case MVT::i1: Opc = PPC::LBZ; break;
1020 case MVT::i8: Opc = PPC::LBZ; break;
1021 case MVT::i16: Opc = sext ? PPC::LHA : PPC::LHZ; break;
1022 case MVT::i32: Opc = sext ? PPC::LWA : PPC::LWZ; break;
1023 case MVT::i64: Opc = PPC::LD; break;
1024 case MVT::f32: Opc = PPC::LFS; break;
1025 case MVT::f64: Opc = PPC::LFD; break;
1028 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Address)) {
1029 Tmp1 = MakeReg(MVT::i64);
1030 int CPI = CP->getIndex();
1031 BuildMI(BB, PPC::LOADHiAddr, 2, Tmp1).addReg(getGlobalBaseReg())
1032 .addConstantPoolIndex(CPI);
1033 BuildMI(BB, Opc, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
1035 else if(Address.getOpcode() == ISD::FrameIndex) {
1036 Tmp1 = cast<FrameIndexSDNode>(Address)->getIndex();
1037 addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1);
1040 bool idx = SelectAddr(Address, Tmp1, offset);
1042 Opc = IndexedOpForOp(Opc);
1043 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(offset);
1045 BuildMI(BB, Opc, 2, Result).addSImm(offset).addReg(Tmp1);
1052 unsigned GPR_idx = 0, FPR_idx = 0;
1053 static const unsigned GPR[] = {
1054 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1055 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1057 static const unsigned FPR[] = {
1058 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1059 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1062 // Lower the chain for this call.
1063 Select(N.getOperand(0));
1064 ExprMap[N.getValue(Node->getNumValues()-1)] = 1;
1066 MachineInstr *CallMI;
1067 // Emit the correct call instruction based on the type of symbol called.
1068 if (GlobalAddressSDNode *GASD =
1069 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
1070 CallMI = BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(GASD->getGlobal(),
1072 } else if (ExternalSymbolSDNode *ESSDN =
1073 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
1074 CallMI = BuildMI(PPC::CALLpcrel, 1).addExternalSymbol(ESSDN->getSymbol(),
1077 Tmp1 = SelectExpr(N.getOperand(1));
1078 BuildMI(BB, PPC::OR, 2, PPC::R12).addReg(Tmp1).addReg(Tmp1);
1079 BuildMI(BB, PPC::MTCTR, 1).addReg(PPC::R12);
1080 CallMI = BuildMI(PPC::CALLindirect, 3).addImm(20).addImm(0)
1084 // Load the register args to virtual regs
1085 std::vector<unsigned> ArgVR;
1086 for(int i = 2, e = Node->getNumOperands(); i < e; ++i)
1087 ArgVR.push_back(SelectExpr(N.getOperand(i)));
1089 // Copy the virtual registers into the appropriate argument register
1090 for(int i = 0, e = ArgVR.size(); i < e; ++i) {
1091 switch(N.getOperand(i+2).getValueType()) {
1092 default: Node->dump(); assert(0 && "Unknown value type for call");
1098 assert(GPR_idx < 8 && "Too many int args");
1099 if (N.getOperand(i+2).getOpcode() != ISD::UNDEF) {
1100 BuildMI(BB, PPC::OR,2,GPR[GPR_idx]).addReg(ArgVR[i]).addReg(ArgVR[i]);
1101 CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use);
1107 assert(FPR_idx < 13 && "Too many fp args");
1108 BuildMI(BB, PPC::FMR, 1, FPR[FPR_idx]).addReg(ArgVR[i]);
1109 CallMI->addRegOperand(FPR[FPR_idx], MachineOperand::Use);
1115 // Put the call instruction in the correct place in the MachineBasicBlock
1116 BB->push_back(CallMI);
1118 switch (Node->getValueType(0)) {
1119 default: assert(0 && "Unknown value type for call result!");
1120 case MVT::Other: return 1;
1126 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R3).addReg(PPC::R3);
1130 BuildMI(BB, PPC::FMR, 1, Result).addReg(PPC::F1);
1133 return Result+N.ResNo;
1136 case ISD::SIGN_EXTEND:
1137 case ISD::SIGN_EXTEND_INREG:
1138 Tmp1 = SelectExpr(N.getOperand(0));
1139 switch(cast<MVTSDNode>(Node)->getExtraValueType()) {
1140 default: Node->dump(); assert(0 && "Unhandled SIGN_EXTEND type"); break;
1142 BuildMI(BB, PPC::EXTSW, 1, Result).addReg(Tmp1);
1145 BuildMI(BB, PPC::EXTSH, 1, Result).addReg(Tmp1);
1148 BuildMI(BB, PPC::EXTSB, 1, Result).addReg(Tmp1);
1151 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp1).addSImm(0);
1156 case ISD::ZERO_EXTEND_INREG:
1157 Tmp1 = SelectExpr(N.getOperand(0));
1158 switch(cast<MVTSDNode>(Node)->getExtraValueType()) {
1159 default: Node->dump(); assert(0 && "Unhandled ZERO_EXTEND type"); break;
1160 case MVT::i16: Tmp2 = 16; break;
1161 case MVT::i8: Tmp2 = 24; break;
1162 case MVT::i1: Tmp2 = 31; break;
1164 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(0).addImm(Tmp2)
1168 case ISD::CopyFromReg:
1170 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1171 Tmp1 = dyn_cast<RegSDNode>(Node)->getReg();
1172 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1176 Tmp1 = SelectExpr(N.getOperand(0));
1177 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1178 Tmp2 = CN->getValue() & 0x3F;
1179 BuildMI(BB, PPC::RLDICR, 3, Result).addReg(Tmp1).addImm(Tmp2)
1182 Tmp2 = SelectExpr(N.getOperand(1));
1183 BuildMI(BB, PPC::SLD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1188 Tmp1 = SelectExpr(N.getOperand(0));
1189 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1190 Tmp2 = CN->getValue() & 0x3F;
1191 BuildMI(BB, PPC::RLDICL, 3, Result).addReg(Tmp1).addImm(64-Tmp2)
1194 Tmp2 = SelectExpr(N.getOperand(1));
1195 BuildMI(BB, PPC::SRD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1200 Tmp1 = SelectExpr(N.getOperand(0));
1201 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1202 Tmp2 = CN->getValue() & 0x3F;
1203 BuildMI(BB, PPC::SRADI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1205 Tmp2 = SelectExpr(N.getOperand(1));
1206 BuildMI(BB, PPC::SRAD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1211 Tmp1 = SelectExpr(N.getOperand(0));
1212 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1213 default: assert(0 && "unhandled result code");
1214 case 0: // No immediate
1215 Tmp2 = SelectExpr(N.getOperand(1));
1216 BuildMI(BB, PPC::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1218 case 1: // Low immediate
1219 BuildMI(BB, PPC::ADDI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1221 case 2: // Shifted immediate
1222 BuildMI(BB, PPC::ADDIS, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1229 Tmp1 = SelectExpr(N.getOperand(0));
1230 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1231 default: assert(0 && "unhandled result code");
1232 case 0: // No immediate
1233 Tmp2 = SelectExpr(N.getOperand(1));
1235 case ISD::AND: Opc = PPC::AND; break;
1236 case ISD::OR: Opc = PPC::OR; break;
1238 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1240 case 1: // Low immediate
1242 case ISD::AND: Opc = PPC::ANDIo; break;
1243 case ISD::OR: Opc = PPC::ORI; break;
1245 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(Tmp2);
1247 case 2: // Shifted immediate
1249 case ISD::AND: Opc = PPC::ANDISo; break;
1250 case ISD::OR: Opc = PPC::ORIS; break;
1252 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(Tmp2);
1258 // Check for EQV: xor, (xor a, -1), b
1259 if (N.getOperand(0).getOpcode() == ISD::XOR &&
1260 N.getOperand(0).getOperand(1).getOpcode() == ISD::Constant &&
1261 cast<ConstantSDNode>(N.getOperand(0).getOperand(1))->isAllOnesValue()) {
1263 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1264 Tmp2 = SelectExpr(N.getOperand(1));
1265 BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1268 // Check for NOT, NOR, and NAND: xor (copy, or, and), -1
1269 if (N.getOperand(1).getOpcode() == ISD::Constant &&
1270 cast<ConstantSDNode>(N.getOperand(1))->isAllOnesValue()) {
1272 switch(N.getOperand(0).getOpcode()) {
1274 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1275 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1276 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1279 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1280 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1281 BuildMI(BB, PPC::NAND, 2, Result).addReg(Tmp1).addReg(Tmp2);
1284 Tmp1 = SelectExpr(N.getOperand(0));
1285 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1290 Tmp1 = SelectExpr(N.getOperand(0));
1291 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1292 default: assert(0 && "unhandled result code");
1293 case 0: // No immediate
1294 Tmp2 = SelectExpr(N.getOperand(1));
1295 BuildMI(BB, PPC::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1297 case 1: // Low immediate
1298 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1300 case 2: // Shifted immediate
1301 BuildMI(BB, PPC::XORIS, 2, Result).addReg(Tmp1).addImm(Tmp2);
1308 Tmp2 = SelectExpr(N.getOperand(1));
1309 if (1 == getImmediateForOpcode(N.getOperand(0), opcode, Tmp1))
1310 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp2).addSImm(Tmp1);
1312 Tmp1 = SelectExpr(N.getOperand(0));
1313 BuildMI(BB, PPC::SUBF, 2, Result).addReg(Tmp2).addReg(Tmp1);
1318 Tmp1 = SelectExpr(N.getOperand(0));
1319 if (1 == getImmediateForOpcode(N.getOperand(1), opcode, Tmp2))
1320 BuildMI(BB, PPC::MULLI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1322 Tmp2 = SelectExpr(N.getOperand(1));
1323 BuildMI(BB, PPC::MULLD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1329 if (3 == getImmediateForOpcode(N.getOperand(1), opcode, Tmp3)) {
1330 Tmp1 = MakeReg(MVT::i64);
1331 Tmp2 = SelectExpr(N.getOperand(0));
1332 BuildMI(BB, PPC::SRADI, 2, Tmp1).addReg(Tmp2).addImm(Tmp3);
1333 BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp1);
1336 Tmp1 = SelectExpr(N.getOperand(0));
1337 Tmp2 = SelectExpr(N.getOperand(1));
1338 Opc = (ISD::UDIV == opcode) ? PPC::DIVWU : PPC::DIVW;
1339 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1342 case ISD::FP_TO_UINT:
1343 case ISD::FP_TO_SINT: {
1344 Tmp1 = SelectExpr(N.getOperand(0));
1345 Tmp2 = MakeReg(MVT::f64);
1346 BuildMI(BB, PPC::FCTIDZ, 1, Tmp2).addReg(Tmp1);
1347 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
1348 addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(Tmp2), FrameIdx);
1349 addFrameReference(BuildMI(BB, PPC::LD, 2, Result), FrameIdx);
1354 if (SetCCSDNode *SetCC = dyn_cast<SetCCSDNode>(Node)) {
1355 Opc = SelectSetCR0(N);
1357 unsigned TrueValue = MakeReg(MVT::i32);
1358 BuildMI(BB, PPC::LI, 1, TrueValue).addSImm(1);
1359 unsigned FalseValue = MakeReg(MVT::i32);
1360 BuildMI(BB, PPC::LI, 1, FalseValue).addSImm(0);
1362 // Create an iterator with which to insert the MBB for copying the false
1363 // value and the MBB to hold the PHI instruction for this SetCC.
1364 MachineBasicBlock *thisMBB = BB;
1365 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1366 ilist<MachineBasicBlock>::iterator It = BB;
1371 // cmpTY cr0, r1, r2
1372 // %TrueValue = li 1
1374 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1375 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1376 BuildMI(BB, Opc, 2).addReg(PPC::CR0).addMBB(sinkMBB);
1377 MachineFunction *F = BB->getParent();
1378 F->getBasicBlockList().insert(It, copy0MBB);
1379 F->getBasicBlockList().insert(It, sinkMBB);
1380 // Update machine-CFG edges
1381 BB->addSuccessor(copy0MBB);
1382 BB->addSuccessor(sinkMBB);
1385 // %FalseValue = li 0
1388 // Update machine-CFG edges
1389 BB->addSuccessor(sinkMBB);
1392 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1395 BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue)
1396 .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB);
1399 assert(0 && "Is this legal?");
1403 unsigned TrueValue = SelectExpr(N.getOperand(1)); //Use if TRUE
1404 unsigned FalseValue = SelectExpr(N.getOperand(2)); //Use if FALSE
1405 Opc = SelectSetCR0(N.getOperand(0));
1407 // Create an iterator with which to insert the MBB for copying the false
1408 // value and the MBB to hold the PHI instruction for this SetCC.
1409 MachineBasicBlock *thisMBB = BB;
1410 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1411 ilist<MachineBasicBlock>::iterator It = BB;
1417 // cmpTY cr0, r1, r2
1419 // fallthrough --> copy0MBB
1420 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1421 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1422 BuildMI(BB, Opc, 2).addReg(PPC::CR0).addMBB(sinkMBB);
1423 MachineFunction *F = BB->getParent();
1424 F->getBasicBlockList().insert(It, copy0MBB);
1425 F->getBasicBlockList().insert(It, sinkMBB);
1426 // Update machine-CFG edges
1427 BB->addSuccessor(copy0MBB);
1428 BB->addSuccessor(sinkMBB);
1431 // %FalseValue = ...
1432 // # fallthrough to sinkMBB
1434 // Update machine-CFG edges
1435 BB->addSuccessor(sinkMBB);
1438 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1441 BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue)
1442 .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB);
1444 // FIXME: Select i64?
1449 switch (N.getValueType()) {
1450 default: assert(0 && "Cannot use constants of this type!");
1452 BuildMI(BB, PPC::LI, 1, Result)
1453 .addSImm(!cast<ConstantSDNode>(N)->isNullValue());
1457 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
1458 if (v < 32768 && v >= -32768) {
1459 BuildMI(BB, PPC::LI, 1, Result).addSImm(v);
1461 Tmp1 = MakeReg(MVT::i32);
1462 BuildMI(BB, PPC::LIS, 1, Tmp1).addSImm(v >> 16);
1463 BuildMI(BB, PPC::ORI, 2, Result).addReg(Tmp1).addImm(v & 0xFFFF);
1473 void ISel::Select(SDOperand N) {
1474 unsigned Tmp1, Tmp2, Opc;
1475 unsigned opcode = N.getOpcode();
1477 if (!ExprMap.insert(std::make_pair(N, 1)).second)
1478 return; // Already selected.
1480 SDNode *Node = N.Val;
1482 switch (Node->getOpcode()) {
1484 Node->dump(); std::cerr << "\n";
1485 assert(0 && "Node not handled yet!");
1486 case ISD::EntryToken: return; // Noop
1487 case ISD::TokenFactor:
1488 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1489 Select(Node->getOperand(i));
1491 case ISD::ADJCALLSTACKDOWN:
1492 case ISD::ADJCALLSTACKUP:
1493 Select(N.getOperand(0));
1494 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
1495 Opc = N.getOpcode() == ISD::ADJCALLSTACKDOWN ? PPC::ADJCALLSTACKDOWN :
1496 PPC::ADJCALLSTACKUP;
1497 BuildMI(BB, Opc, 1).addImm(Tmp1);
1500 MachineBasicBlock *Dest =
1501 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
1502 Select(N.getOperand(0));
1503 BuildMI(BB, PPC::B, 1).addMBB(Dest);
1509 case ISD::CopyToReg:
1510 Select(N.getOperand(0));
1511 Tmp1 = SelectExpr(N.getOperand(1));
1512 Tmp2 = cast<RegSDNode>(N)->getReg();
1515 if (N.getOperand(1).getValueType() == MVT::f64 ||
1516 N.getOperand(1).getValueType() == MVT::f32)
1517 BuildMI(BB, PPC::FMR, 1, Tmp2).addReg(Tmp1);
1519 BuildMI(BB, PPC::OR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
1522 case ISD::ImplicitDef:
1523 Select(N.getOperand(0));
1524 BuildMI(BB, PPC::IMPLICIT_DEF, 0, cast<RegSDNode>(N)->getReg());
1527 switch (N.getNumOperands()) {
1529 assert(0 && "Unknown return instruction!");
1531 assert(N.getOperand(1).getValueType() == MVT::i32 &&
1532 N.getOperand(2).getValueType() == MVT::i32 &&
1533 "Unknown two-register value!");
1534 Select(N.getOperand(0));
1535 Tmp1 = SelectExpr(N.getOperand(1));
1536 Tmp2 = SelectExpr(N.getOperand(2));
1537 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp2).addReg(Tmp2);
1538 BuildMI(BB, PPC::OR, 2, PPC::R4).addReg(Tmp1).addReg(Tmp1);
1541 Select(N.getOperand(0));
1542 Tmp1 = SelectExpr(N.getOperand(1));
1543 switch (N.getOperand(1).getValueType()) {
1545 assert(0 && "Unknown return type!");
1548 BuildMI(BB, PPC::FMR, 1, PPC::F1).addReg(Tmp1);
1551 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp1).addReg(Tmp1);
1555 Select(N.getOperand(0));
1558 BuildMI(BB, PPC::BLR, 0); // Just emit a 'ret' instruction
1560 case ISD::TRUNCSTORE:
1563 SDOperand Chain = N.getOperand(0);
1564 SDOperand Value = N.getOperand(1);
1565 SDOperand Address = N.getOperand(2);
1568 Tmp1 = SelectExpr(Value); //value
1570 if (opcode == ISD::STORE) {
1571 switch(Value.getValueType()) {
1572 default: assert(0 && "unknown Type in store");
1573 case MVT::i64: Opc = PPC::STD; break;
1574 case MVT::f64: Opc = PPC::STFD; break;
1575 case MVT::f32: Opc = PPC::STFS; break;
1577 } else { //ISD::TRUNCSTORE
1578 switch(cast<MVTSDNode>(Node)->getExtraValueType()) {
1579 default: assert(0 && "unknown Type in store");
1580 case MVT::i1: //FIXME: DAG does not promote this load
1581 case MVT::i8: Opc= PPC::STB; break;
1582 case MVT::i16: Opc = PPC::STH; break;
1583 case MVT::i32: Opc = PPC::STW; break;
1587 if(Address.getOpcode() == ISD::FrameIndex)
1589 Tmp2 = cast<FrameIndexSDNode>(Address)->getIndex();
1590 addFrameReference(BuildMI(BB, Opc, 3).addReg(Tmp1), (int)Tmp2);
1595 bool idx = SelectAddr(Address, Tmp2, offset);
1597 Opc = IndexedOpForOp(Opc);
1598 BuildMI(BB, Opc, 3).addReg(Tmp1).addReg(Tmp2).addReg(offset);
1600 BuildMI(BB, Opc, 3).addReg(Tmp1).addImm(offset).addReg(Tmp2);
1609 case ISD::CopyFromReg:
1611 case ISD::DYNAMIC_STACKALLOC:
1616 assert(0 && "Should not be reached!");
1620 /// createPPC32PatternInstructionSelector - This pass converts an LLVM function
1621 /// into a machine code representation using pattern matching and a machine
1622 /// description file.
1624 FunctionPass *llvm::createPPC64ISelPattern(TargetMachine &TM) {
1625 return new ISel(TM);