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 setShiftAmountFlavor(Extend); // shl X, 32 == 0
59 addLegalFPImmediate(+0.0); // Necessary for FSEL
60 addLegalFPImmediate(-0.0); //
62 computeRegisterProperties();
65 /// LowerArguments - This hook must be implemented to indicate how we should
66 /// lower the arguments for the specified function, into the specified DAG.
67 virtual std::vector<SDOperand>
68 LowerArguments(Function &F, SelectionDAG &DAG);
70 /// LowerCallTo - This hook lowers an abstract call to a function into an
72 virtual std::pair<SDOperand, SDOperand>
73 LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg,
74 SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG);
76 virtual std::pair<SDOperand, SDOperand>
77 LowerVAStart(SDOperand Chain, SelectionDAG &DAG);
79 virtual std::pair<SDOperand,SDOperand>
80 LowerVAArgNext(bool isVANext, SDOperand Chain, SDOperand VAList,
81 const Type *ArgTy, SelectionDAG &DAG);
83 virtual std::pair<SDOperand, SDOperand>
84 LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
90 std::vector<SDOperand>
91 PPC64TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
93 // add beautiful description of PPC stack frame format, or at least some docs
95 MachineFunction &MF = DAG.getMachineFunction();
96 MachineFrameInfo *MFI = MF.getFrameInfo();
97 MachineBasicBlock& BB = MF.front();
98 std::vector<SDOperand> ArgValues;
100 // Due to the rather complicated nature of the PowerPC ABI, rather than a
101 // fixed size array of physical args, for the sake of simplicity let the STL
102 // handle tracking them for us.
103 std::vector<unsigned> argVR, argPR, argOp;
104 unsigned ArgOffset = 48;
105 unsigned GPR_remaining = 8;
106 unsigned FPR_remaining = 13;
107 unsigned GPR_idx = 0, FPR_idx = 0;
108 static const unsigned GPR[] = {
109 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
110 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
112 static const unsigned FPR[] = {
113 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
114 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
117 // Add DAG nodes to load the arguments... On entry to a function on PPC,
118 // the arguments start at offset 48, although they are likely to be passed
120 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
121 SDOperand newroot, argt;
122 bool needsLoad = false;
123 MVT::ValueType ObjectVT = getValueType(I->getType());
126 default: assert(0 && "Unhandled argument type!");
132 if (GPR_remaining > 0) {
133 BuildMI(&BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]);
134 argt = newroot = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32,
136 if (ObjectVT != MVT::i64)
137 argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, newroot);
144 if (FPR_remaining > 0) {
145 BuildMI(&BB, PPC::IMPLICIT_DEF, 0, FPR[FPR_idx]);
146 argt = newroot = DAG.getCopyFromReg(FPR[FPR_idx], ObjectVT,
156 // We need to load the argument to a virtual register if we determined above
157 // that we ran out of physical registers of the appropriate type
159 unsigned SubregOffset = 0;
161 default: assert(0 && "Unhandled argument type!");
163 case MVT::i8: SubregOffset = 7; break;
164 case MVT::i16: SubregOffset = 6; break;
166 case MVT::f32: SubregOffset = 4; break;
168 case MVT::f64: SubregOffset = 0; break;
170 int FI = MFI->CreateFixedObject(8, ArgOffset);
171 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i64);
172 FIN = DAG.getNode(ISD::ADD, MVT::i64, FIN,
173 DAG.getConstant(SubregOffset, MVT::i64));
174 argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN);
177 // Every 4 bytes of argument space consumes one of the GPRs available for
179 if (GPR_remaining > 0) {
185 DAG.setRoot(newroot.getValue(1));
186 ArgValues.push_back(argt);
189 // If the function takes variable number of arguments, make a frame index for
190 // the start of the first vararg value... for expansion of llvm.va_start.
192 VarArgsFrameIndex = MFI->CreateFixedObject(8, ArgOffset);
193 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i64);
194 // If this function is vararg, store any remaining integer argument regs
195 // to their spots on the stack so that they may be loaded by deferencing the
196 // result of va_next.
197 std::vector<SDOperand> MemOps;
198 for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) {
199 BuildMI(&BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]);
200 SDOperand Val = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i64, DAG.getRoot());
201 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
203 MemOps.push_back(Store);
204 // Increment the address by eight for the next argument to store
205 SDOperand PtrOff = DAG.getConstant(8, getPointerTy());
206 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff);
208 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps));
214 std::pair<SDOperand, SDOperand>
215 PPC64TargetLowering::LowerCallTo(SDOperand Chain,
216 const Type *RetTy, bool isVarArg,
217 SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG) {
218 // args_to_use will accumulate outgoing args for the ISD::CALL case in
219 // SelectExpr to use to put the arguments in the appropriate registers.
220 std::vector<SDOperand> args_to_use;
222 // Count how many bytes are to be pushed on the stack, including the linkage
223 // area, and parameter passing area.
224 unsigned NumBytes = 48;
227 Chain = DAG.getNode(ISD::ADJCALLSTACKDOWN, MVT::Other, Chain,
228 DAG.getConstant(NumBytes, getPointerTy()));
230 NumBytes = 8 * Args.size(); // All arguments are rounded up to 8 bytes
232 // Just to be safe, we'll always reserve the full 48 bytes of linkage area
233 // plus 64 bytes of argument space in case any called code gets funky on us.
234 if (NumBytes < 112) NumBytes = 112;
236 // Adjust the stack pointer for the new arguments...
237 // These operations are automatically eliminated by the prolog/epilog pass
238 Chain = DAG.getNode(ISD::ADJCALLSTACKDOWN, MVT::Other, Chain,
239 DAG.getConstant(NumBytes, getPointerTy()));
241 // Set up a copy of the stack pointer for use loading and storing any
242 // arguments that may not fit in the registers available for argument
244 SDOperand StackPtr = DAG.getCopyFromReg(PPC::R1, MVT::i32,
247 // Figure out which arguments are going to go in registers, and which in
248 // memory. Also, if this is a vararg function, floating point operations
249 // must be stored to our stack, and loaded into integer regs as well, if
250 // any integer regs are available for argument passing.
251 unsigned ArgOffset = 48;
252 unsigned GPR_remaining = 8;
253 unsigned FPR_remaining = 13;
255 std::vector<SDOperand> MemOps;
256 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
257 // PtrOff will be used to store the current argument to the stack if a
258 // register cannot be found for it.
259 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
260 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
261 MVT::ValueType ArgVT = getValueType(Args[i].second);
264 default: assert(0 && "Unexpected ValueType for argument!");
269 // Promote the integer to 64 bits. If the input type is signed use a
270 // sign extend, otherwise use a zero extend.
271 if (Args[i].second->isSigned())
272 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i64, Args[i].first);
274 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i64, Args[i].first);
277 if (GPR_remaining > 0) {
278 args_to_use.push_back(Args[i].first);
281 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
282 Args[i].first, PtrOff));
288 if (FPR_remaining > 0) {
289 args_to_use.push_back(Args[i].first);
292 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
293 Args[i].first, PtrOff);
294 MemOps.push_back(Store);
295 // Float varargs are always shadowed in available integer registers
296 if (GPR_remaining > 0) {
297 SDOperand Load = DAG.getLoad(MVT::i64, Store, PtrOff);
298 MemOps.push_back(Load);
299 args_to_use.push_back(Load);
303 // If we have any FPRs remaining, we may also have GPRs remaining.
304 // Args passed in FPRs also consume an available GPR.
305 if (GPR_remaining > 0) {
306 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i64));
311 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
312 Args[i].first, PtrOff));
319 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
322 std::vector<MVT::ValueType> RetVals;
323 MVT::ValueType RetTyVT = getValueType(RetTy);
324 if (RetTyVT != MVT::isVoid)
325 RetVals.push_back(RetTyVT);
326 RetVals.push_back(MVT::Other);
328 SDOperand TheCall = SDOperand(DAG.getCall(RetVals,
329 Chain, Callee, args_to_use), 0);
330 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
331 Chain = DAG.getNode(ISD::ADJCALLSTACKUP, MVT::Other, Chain,
332 DAG.getConstant(NumBytes, getPointerTy()));
333 return std::make_pair(TheCall, Chain);
336 std::pair<SDOperand, SDOperand>
337 PPC64TargetLowering::LowerVAStart(SDOperand Chain, SelectionDAG &DAG) {
338 //vastart just returns the address of the VarArgsFrameIndex slot.
339 return std::make_pair(DAG.getFrameIndex(VarArgsFrameIndex, MVT::i64), Chain);
342 std::pair<SDOperand,SDOperand> PPC64TargetLowering::
343 LowerVAArgNext(bool isVANext, SDOperand Chain, SDOperand VAList,
344 const Type *ArgTy, SelectionDAG &DAG) {
345 MVT::ValueType ArgVT = getValueType(ArgTy);
348 Result = DAG.getLoad(ArgVT, DAG.getEntryNode(), VAList);
350 Result = DAG.getNode(ISD::ADD, VAList.getValueType(), VAList,
351 DAG.getConstant(8, VAList.getValueType()));
353 return std::make_pair(Result, Chain);
357 std::pair<SDOperand, SDOperand> PPC64TargetLowering::
358 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
360 assert(0 && "LowerFrameReturnAddress unimplemented");
365 Statistic<>NotLogic("ppc-codegen", "Number of inverted logical ops");
366 Statistic<>FusedFP("ppc-codegen", "Number of fused fp operations");
367 //===--------------------------------------------------------------------===//
368 /// ISel - PPC32 specific code to select PPC32 machine instructions for
369 /// SelectionDAG operations.
370 //===--------------------------------------------------------------------===//
371 class ISel : public SelectionDAGISel {
374 PPC64TargetLowering PPC64Lowering;
376 /// ExprMap - As shared expressions are codegen'd, we keep track of which
377 /// vreg the value is produced in, so we only emit one copy of each compiled
379 std::map<SDOperand, unsigned> ExprMap;
381 unsigned GlobalBaseReg;
382 bool GlobalBaseInitialized;
385 ISel(TargetMachine &TM) : SelectionDAGISel(PPC64Lowering), PPC64Lowering(TM)
388 /// runOnFunction - Override this function in order to reset our per-function
390 virtual bool runOnFunction(Function &Fn) {
391 // Make sure we re-emit a set of the global base reg if necessary
392 GlobalBaseInitialized = false;
393 return SelectionDAGISel::runOnFunction(Fn);
396 /// InstructionSelectBasicBlock - This callback is invoked by
397 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
398 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG) {
400 // Codegen the basic block.
401 Select(DAG.getRoot());
403 // Clear state used for selection.
407 unsigned getGlobalBaseReg();
408 unsigned getConstDouble(double floatVal, unsigned Result);
409 unsigned SelectSetCR0(SDOperand CC);
410 unsigned SelectExpr(SDOperand N);
411 unsigned SelectExprFP(SDOperand N, unsigned Result);
412 void Select(SDOperand N);
414 bool SelectAddr(SDOperand N, unsigned& Reg, int& offset);
415 void SelectBranchCC(SDOperand N);
418 /// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
419 /// returns zero when the input is not exactly a power of two.
420 static unsigned ExactLog2(unsigned Val) {
421 if (Val == 0 || (Val & (Val-1))) return 0;
430 /// getImmediateForOpcode - This method returns a value indicating whether
431 /// the ConstantSDNode N can be used as an immediate to Opcode. The return
432 /// values are either 0, 1 or 2. 0 indicates that either N is not a
433 /// ConstantSDNode, or is not suitable for use by that opcode. A return value
434 /// of 1 indicates that the constant may be used in normal immediate form. A
435 /// return value of 2 indicates that the constant may be used in shifted
436 /// immediate form. A return value of 3 indicates that log base 2 of the
437 /// constant may be used.
439 static unsigned getImmediateForOpcode(SDOperand N, unsigned Opcode,
440 unsigned& Imm, bool U = false) {
441 if (N.getOpcode() != ISD::Constant) return 0;
443 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
448 if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
449 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
454 if (v >= 0 && v <= 65535) { Imm = v & 0xFFFF; return 1; }
455 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
459 if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
462 if (U && (v >= 0 && v <= 65535)) { Imm = v & 0xFFFF; return 1; }
463 if (!U && (v <= 32767 && v >= -32768)) { Imm = v & 0xFFFF; return 1; }
466 if ((Imm = ExactLog2(v))) { return 3; }
472 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
473 /// to Condition. If the Condition is unordered or unsigned, the bool argument
474 /// U is set to true, otherwise it is set to false.
475 static unsigned getBCCForSetCC(unsigned Condition, bool& U) {
478 default: assert(0 && "Unknown condition!"); abort();
479 case ISD::SETEQ: return PPC::BEQ;
480 case ISD::SETNE: return PPC::BNE;
481 case ISD::SETULT: U = true;
482 case ISD::SETLT: return PPC::BLT;
483 case ISD::SETULE: U = true;
484 case ISD::SETLE: return PPC::BLE;
485 case ISD::SETUGT: U = true;
486 case ISD::SETGT: return PPC::BGT;
487 case ISD::SETUGE: U = true;
488 case ISD::SETGE: return PPC::BGE;
493 /// IndexedOpForOp - Return the indexed variant for each of the PowerPC load
494 /// and store immediate instructions.
495 static unsigned IndexedOpForOp(unsigned Opcode) {
497 default: assert(0 && "Unknown opcode!"); abort();
498 case PPC::LBZ: return PPC::LBZX; case PPC::STB: return PPC::STBX;
499 case PPC::LHZ: return PPC::LHZX; case PPC::STH: return PPC::STHX;
500 case PPC::LHA: return PPC::LHAX; case PPC::STW: return PPC::STWX;
501 case PPC::LWZ: return PPC::LWZX; case PPC::STD: return PPC::STDX;
502 case PPC::LD: return PPC::LDX; case PPC::STFS: return PPC::STFSX;
503 case PPC::LFS: return PPC::LFSX; case PPC::STFD: return PPC::STFDX;
504 case PPC::LFD: return PPC::LFDX;
510 /// getGlobalBaseReg - Output the instructions required to put the
511 /// base address to use for accessing globals into a register.
513 unsigned ISel::getGlobalBaseReg() {
514 if (!GlobalBaseInitialized) {
515 // Insert the set of GlobalBaseReg into the first MBB of the function
516 MachineBasicBlock &FirstMBB = BB->getParent()->front();
517 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
518 GlobalBaseReg = MakeReg(MVT::i64);
519 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
520 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg).addReg(PPC::LR);
521 GlobalBaseInitialized = true;
523 return GlobalBaseReg;
526 /// getConstDouble - Loads a floating point value into a register, via the
527 /// Constant Pool. Optionally takes a register in which to load the value.
528 unsigned ISel::getConstDouble(double doubleVal, unsigned Result=0) {
529 unsigned Tmp1 = MakeReg(MVT::i64);
530 if (0 == Result) Result = MakeReg(MVT::f64);
531 MachineConstantPool *CP = BB->getParent()->getConstantPool();
532 ConstantFP *CFP = ConstantFP::get(Type::DoubleTy, doubleVal);
533 unsigned CPI = CP->getConstantPoolIndex(CFP);
534 BuildMI(BB, PPC::LOADHiAddr, 2, Tmp1).addReg(getGlobalBaseReg())
535 .addConstantPoolIndex(CPI);
536 BuildMI(BB, PPC::LFD, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
540 unsigned ISel::SelectSetCR0(SDOperand CC) {
541 unsigned Opc, Tmp1, Tmp2;
542 static const unsigned CompareOpcodes[] =
543 { PPC::FCMPU, PPC::FCMPU, PPC::CMPW, PPC::CMPLW };
545 // If the first operand to the select is a SETCC node, then we can fold it
546 // into the branch that selects which value to return.
547 SetCCSDNode* SetCC = dyn_cast<SetCCSDNode>(CC.Val);
548 if (SetCC && CC.getOpcode() == ISD::SETCC) {
550 Opc = getBCCForSetCC(SetCC->getCondition(), U);
551 Tmp1 = SelectExpr(SetCC->getOperand(0));
553 // Pass the optional argument U to getImmediateForOpcode for SETCC,
554 // so that it knows whether the SETCC immediate range is signed or not.
555 if (1 == getImmediateForOpcode(SetCC->getOperand(1), ISD::SETCC,
558 BuildMI(BB, PPC::CMPLWI, 2, PPC::CR0).addReg(Tmp1).addImm(Tmp2);
560 BuildMI(BB, PPC::CMPWI, 2, PPC::CR0).addReg(Tmp1).addSImm(Tmp2);
562 bool IsInteger = MVT::isInteger(SetCC->getOperand(0).getValueType());
563 unsigned CompareOpc = CompareOpcodes[2 * IsInteger + U];
564 Tmp2 = SelectExpr(SetCC->getOperand(1));
565 BuildMI(BB, CompareOpc, 2, PPC::CR0).addReg(Tmp1).addReg(Tmp2);
568 Tmp1 = SelectExpr(CC);
569 BuildMI(BB, PPC::CMPLWI, 2, PPC::CR0).addReg(Tmp1).addImm(0);
575 /// Check to see if the load is a constant offset from a base register
576 bool ISel::SelectAddr(SDOperand N, unsigned& Reg, int& offset)
578 unsigned imm = 0, opcode = N.getOpcode();
579 if (N.getOpcode() == ISD::ADD) {
580 Reg = SelectExpr(N.getOperand(0));
581 if (1 == getImmediateForOpcode(N.getOperand(1), opcode, imm)) {
585 offset = SelectExpr(N.getOperand(1));
593 void ISel::SelectBranchCC(SDOperand N)
595 assert(N.getOpcode() == ISD::BRCOND && "Not a BranchCC???");
596 MachineBasicBlock *Dest =
597 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
599 // Get the MBB we will fall through to so that we can hand it off to the
600 // branch selection pass as an argument to the PPC::COND_BRANCH pseudo op.
601 //ilist<MachineBasicBlock>::iterator It = BB;
602 //MachineBasicBlock *Fallthrough = ++It;
604 Select(N.getOperand(0)); //chain
605 unsigned Opc = SelectSetCR0(N.getOperand(1));
606 // FIXME: Use this once we have something approximating two-way branches
607 // We cannot currently use this in case the ISel hands us something like
610 // since the fallthrough basic block for the conditional branch does not start
611 // with the unconditional branch (it is skipped over).
612 //BuildMI(BB, PPC::COND_BRANCH, 4).addReg(PPC::CR0).addImm(Opc)
613 // .addMBB(Dest).addMBB(Fallthrough);
614 BuildMI(BB, Opc, 2).addReg(PPC::CR0).addMBB(Dest);
618 unsigned ISel::SelectExprFP(SDOperand N, unsigned Result)
620 unsigned Tmp1, Tmp2, Tmp3;
622 SDNode *Node = N.Val;
623 MVT::ValueType DestType = N.getValueType();
624 unsigned opcode = N.getOpcode();
629 assert(0 && "Node not handled!\n");
632 // Attempt to generate FSEL. We can do this whenever we have an FP result,
633 // and an FP comparison in the SetCC node.
634 SetCCSDNode* SetCC = dyn_cast<SetCCSDNode>(N.getOperand(0).Val);
635 if (SetCC && N.getOperand(0).getOpcode() == ISD::SETCC &&
636 !MVT::isInteger(SetCC->getOperand(0).getValueType()) &&
637 SetCC->getCondition() != ISD::SETEQ &&
638 SetCC->getCondition() != ISD::SETNE) {
639 MVT::ValueType VT = SetCC->getOperand(0).getValueType();
640 Tmp1 = SelectExpr(SetCC->getOperand(0)); // Val to compare against
641 unsigned TV = SelectExpr(N.getOperand(1)); // Use if TRUE
642 unsigned FV = SelectExpr(N.getOperand(2)); // Use if FALSE
644 ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(SetCC->getOperand(1));
645 if (CN && (CN->isExactlyValue(-0.0) || CN->isExactlyValue(0.0))) {
646 switch(SetCC->getCondition()) {
647 default: assert(0 && "Invalid FSEL condition"); abort();
650 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp1).addReg(FV).addReg(TV);
654 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp1).addReg(TV).addReg(FV);
659 BuildMI(BB, PPC::FNEG, 1, Tmp2).addReg(Tmp1);
660 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp2).addReg(FV).addReg(TV);
666 BuildMI(BB, PPC::FNEG, 1, Tmp2).addReg(Tmp1);
667 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp2).addReg(TV).addReg(FV);
672 Opc = (MVT::f64 == VT) ? PPC::FSUB : PPC::FSUBS;
673 Tmp2 = SelectExpr(SetCC->getOperand(1));
675 switch(SetCC->getCondition()) {
676 default: assert(0 && "Invalid FSEL condition"); abort();
679 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
680 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
684 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
685 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
689 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
690 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
694 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
695 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
699 assert(0 && "Should never get here");
703 unsigned TrueValue = SelectExpr(N.getOperand(1)); //Use if TRUE
704 unsigned FalseValue = SelectExpr(N.getOperand(2)); //Use if FALSE
705 Opc = SelectSetCR0(N.getOperand(0));
707 // Create an iterator with which to insert the MBB for copying the false
708 // value and the MBB to hold the PHI instruction for this SetCC.
709 MachineBasicBlock *thisMBB = BB;
710 const BasicBlock *LLVM_BB = BB->getBasicBlock();
711 ilist<MachineBasicBlock>::iterator It = BB;
719 // fallthrough --> copy0MBB
720 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
721 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
722 BuildMI(BB, Opc, 2).addReg(PPC::CR0).addMBB(sinkMBB);
723 MachineFunction *F = BB->getParent();
724 F->getBasicBlockList().insert(It, copy0MBB);
725 F->getBasicBlockList().insert(It, sinkMBB);
726 // Update machine-CFG edges
727 BB->addSuccessor(copy0MBB);
728 BB->addSuccessor(sinkMBB);
732 // # fallthrough to sinkMBB
734 // Update machine-CFG edges
735 BB->addSuccessor(sinkMBB);
738 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
741 BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue)
742 .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB);
747 if (!NoExcessFPPrecision &&
748 ISD::ADD == N.getOperand(0).getOpcode() &&
749 N.getOperand(0).Val->hasOneUse() &&
750 ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() &&
751 N.getOperand(0).getOperand(0).Val->hasOneUse()) {
752 ++FusedFP; // Statistic
753 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0));
754 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1));
755 Tmp3 = SelectExpr(N.getOperand(0).getOperand(1));
756 Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
757 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
758 } else if (!NoExcessFPPrecision &&
759 ISD::SUB == N.getOperand(0).getOpcode() &&
760 N.getOperand(0).Val->hasOneUse() &&
761 ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() &&
762 N.getOperand(0).getOperand(0).Val->hasOneUse()) {
763 ++FusedFP; // Statistic
764 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0));
765 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1));
766 Tmp3 = SelectExpr(N.getOperand(0).getOperand(1));
767 Opc = DestType == MVT::f64 ? PPC::FNMSUB : PPC::FNMSUBS;
768 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
769 } else if (ISD::FABS == N.getOperand(0).getOpcode()) {
770 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
771 BuildMI(BB, PPC::FNABS, 1, Result).addReg(Tmp1);
773 Tmp1 = SelectExpr(N.getOperand(0));
774 BuildMI(BB, PPC::FNEG, 1, Result).addReg(Tmp1);
779 Tmp1 = SelectExpr(N.getOperand(0));
780 BuildMI(BB, PPC::FABS, 1, Result).addReg(Tmp1);
784 assert (DestType == MVT::f32 &&
785 N.getOperand(0).getValueType() == MVT::f64 &&
786 "only f64 to f32 conversion supported here");
787 Tmp1 = SelectExpr(N.getOperand(0));
788 BuildMI(BB, PPC::FRSP, 1, Result).addReg(Tmp1);
792 assert (DestType == MVT::f64 &&
793 N.getOperand(0).getValueType() == MVT::f32 &&
794 "only f32 to f64 conversion supported here");
795 Tmp1 = SelectExpr(N.getOperand(0));
796 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
799 case ISD::CopyFromReg:
801 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
802 Tmp1 = dyn_cast<RegSDNode>(Node)->getReg();
803 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
806 case ISD::ConstantFP: {
807 ConstantFPSDNode *CN = cast<ConstantFPSDNode>(N);
808 Result = getConstDouble(CN->getValue(), Result);
813 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
814 N.getOperand(0).Val->hasOneUse()) {
815 ++FusedFP; // Statistic
816 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
817 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
818 Tmp3 = SelectExpr(N.getOperand(1));
819 Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
820 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
823 Opc = DestType == MVT::f64 ? PPC::FADD : PPC::FADDS;
824 Tmp1 = SelectExpr(N.getOperand(0));
825 Tmp2 = SelectExpr(N.getOperand(1));
826 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
830 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
831 N.getOperand(0).Val->hasOneUse()) {
832 ++FusedFP; // Statistic
833 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
834 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
835 Tmp3 = SelectExpr(N.getOperand(1));
836 Opc = DestType == MVT::f64 ? PPC::FMSUB : PPC::FMSUBS;
837 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
840 Opc = DestType == MVT::f64 ? PPC::FSUB : PPC::FSUBS;
841 Tmp1 = SelectExpr(N.getOperand(0));
842 Tmp2 = SelectExpr(N.getOperand(1));
843 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
849 case ISD::MUL: Opc = DestType == MVT::f64 ? PPC::FMUL : PPC::FMULS; break;
850 case ISD::SDIV: Opc = DestType == MVT::f64 ? PPC::FDIV : PPC::FDIVS; break;
852 Tmp1 = SelectExpr(N.getOperand(0));
853 Tmp2 = SelectExpr(N.getOperand(1));
854 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
857 case ISD::UINT_TO_FP:
858 case ISD::SINT_TO_FP: {
859 bool IsUnsigned = (ISD::UINT_TO_FP == opcode);
860 Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
861 Tmp2 = MakeReg(MVT::f64); // temp reg to load the integer value into
862 Tmp3 = MakeReg(MVT::i64); // temp reg to hold the conversion constant
863 unsigned ConstF = MakeReg(MVT::f64); // temp reg to hold the fp constant
865 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
866 MachineConstantPool *CP = BB->getParent()->getConstantPool();
868 // FIXME: pull this FP constant generation stuff out into something like
869 // the simple ISel's getReg.
871 addFrameReference(BuildMI(BB, PPC::STD, 3).addReg(Tmp1), FrameIdx);
872 addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
873 BuildMI(BB, PPC::FCFID, 1, Result).addReg(Tmp2);
875 ConstantFP *CFP = ConstantFP::get(Type::DoubleTy, 0x1.000008p52);
876 unsigned CPI = CP->getConstantPoolIndex(CFP);
877 // Load constant fp value
878 unsigned Tmp4 = MakeReg(MVT::i32);
879 unsigned TmpL = MakeReg(MVT::i32);
880 BuildMI(BB, PPC::LOADHiAddr, 2, Tmp4).addReg(getGlobalBaseReg())
881 .addConstantPoolIndex(CPI);
882 BuildMI(BB, PPC::LFD, 2, ConstF).addConstantPoolIndex(CPI).addReg(Tmp4);
883 // Store the hi & low halves of the fp value, currently in int regs
884 BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330);
885 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx);
886 BuildMI(BB, PPC::XORIS, 2, TmpL).addReg(Tmp1).addImm(0x8000);
887 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(TmpL), FrameIdx, 4);
888 addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
889 // Generate the return value with a subtract
890 BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF);
895 assert(0 && "Should never get here");
899 unsigned ISel::SelectExpr(SDOperand N) {
901 unsigned Tmp1, Tmp2, Tmp3;
903 unsigned opcode = N.getOpcode();
905 SDNode *Node = N.Val;
906 MVT::ValueType DestType = N.getValueType();
908 unsigned &Reg = ExprMap[N];
911 switch (N.getOpcode()) {
913 Reg = Result = (N.getValueType() != MVT::Other) ?
914 MakeReg(N.getValueType()) : 1;
917 // If this is a call instruction, make sure to prepare ALL of the result
918 // values as well as the chain.
919 if (Node->getNumValues() == 1)
920 Reg = Result = 1; // Void call, just a chain.
922 Result = MakeReg(Node->getValueType(0));
923 ExprMap[N.getValue(0)] = Result;
924 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
925 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
926 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
931 if (ISD::CopyFromReg == opcode)
932 DestType = N.getValue(0).getValueType();
934 if (DestType == MVT::f64 || DestType == MVT::f32)
935 if (ISD::LOAD != opcode && ISD::EXTLOAD != opcode && ISD::UNDEF != opcode)
936 return SelectExprFP(N, Result);
941 assert(0 && "Node not handled!\n");
943 BuildMI(BB, PPC::IMPLICIT_DEF, 0, Result);
945 case ISD::DYNAMIC_STACKALLOC:
946 // Generate both result values. FIXME: Need a better commment here?
948 ExprMap[N.getValue(1)] = 1;
950 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
952 // FIXME: We are currently ignoring the requested alignment for handling
953 // greater than the stack alignment. This will need to be revisited at some
954 // point. Align = N.getOperand(2);
955 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
956 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
957 std::cerr << "Cannot allocate stack object with greater alignment than"
958 << " the stack alignment yet!";
961 Select(N.getOperand(0));
962 Tmp1 = SelectExpr(N.getOperand(1));
963 // Subtract size from stack pointer, thereby allocating some space.
964 BuildMI(BB, PPC::SUBF, 2, PPC::R1).addReg(Tmp1).addReg(PPC::R1);
965 // Put a pointer to the space into the result register by copying the SP
966 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R1).addReg(PPC::R1);
969 case ISD::ConstantPool:
970 Tmp1 = cast<ConstantPoolSDNode>(N)->getIndex();
971 Tmp2 = MakeReg(MVT::i64);
972 BuildMI(BB, PPC::LOADHiAddr, 2, Tmp2).addReg(getGlobalBaseReg())
973 .addConstantPoolIndex(Tmp1);
974 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp2).addConstantPoolIndex(Tmp1);
977 case ISD::FrameIndex:
978 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
979 addFrameReference(BuildMI(BB, PPC::ADDI, 2, Result), (int)Tmp1, 0, false);
982 case ISD::GlobalAddress: {
983 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
984 Tmp1 = MakeReg(MVT::i64);
985 BuildMI(BB, PPC::LOADHiAddr, 2, Tmp1).addReg(getGlobalBaseReg())
986 .addGlobalAddress(GV);
987 if (GV->hasWeakLinkage() || GV->isExternal()) {
988 BuildMI(BB, PPC::LD, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
990 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp1).addGlobalAddress(GV);
998 case ISD::SEXTLOAD: {
999 MVT::ValueType TypeBeingLoaded = (ISD::LOAD == opcode) ?
1000 Node->getValueType(0) : cast<MVTSDNode>(Node)->getExtraValueType();
1001 bool sext = (ISD::SEXTLOAD == opcode);
1003 // Make sure we generate both values.
1005 ExprMap[N.getValue(1)] = 1; // Generate the token
1007 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1009 SDOperand Chain = N.getOperand(0);
1010 SDOperand Address = N.getOperand(1);
1013 switch (TypeBeingLoaded) {
1014 default: Node->dump(); assert(0 && "Cannot load this type!");
1015 case MVT::i1: Opc = PPC::LBZ; break;
1016 case MVT::i8: Opc = PPC::LBZ; break;
1017 case MVT::i16: Opc = sext ? PPC::LHA : PPC::LHZ; break;
1018 case MVT::i32: Opc = sext ? PPC::LWA : PPC::LWZ; break;
1019 case MVT::i64: Opc = PPC::LD; break;
1020 case MVT::f32: Opc = PPC::LFS; break;
1021 case MVT::f64: Opc = PPC::LFD; break;
1024 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Address)) {
1025 Tmp1 = MakeReg(MVT::i64);
1026 int CPI = CP->getIndex();
1027 BuildMI(BB, PPC::LOADHiAddr, 2, Tmp1).addReg(getGlobalBaseReg())
1028 .addConstantPoolIndex(CPI);
1029 BuildMI(BB, Opc, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
1031 else if(Address.getOpcode() == ISD::FrameIndex) {
1032 Tmp1 = cast<FrameIndexSDNode>(Address)->getIndex();
1033 addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1);
1036 bool idx = SelectAddr(Address, Tmp1, offset);
1038 Opc = IndexedOpForOp(Opc);
1039 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(offset);
1041 BuildMI(BB, Opc, 2, Result).addSImm(offset).addReg(Tmp1);
1048 unsigned GPR_idx = 0, FPR_idx = 0;
1049 static const unsigned GPR[] = {
1050 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1051 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1053 static const unsigned FPR[] = {
1054 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1055 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1058 // Lower the chain for this call.
1059 Select(N.getOperand(0));
1060 ExprMap[N.getValue(Node->getNumValues()-1)] = 1;
1062 MachineInstr *CallMI;
1063 // Emit the correct call instruction based on the type of symbol called.
1064 if (GlobalAddressSDNode *GASD =
1065 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
1066 CallMI = BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(GASD->getGlobal(),
1068 } else if (ExternalSymbolSDNode *ESSDN =
1069 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
1070 CallMI = BuildMI(PPC::CALLpcrel, 1).addExternalSymbol(ESSDN->getSymbol(),
1073 Tmp1 = SelectExpr(N.getOperand(1));
1074 BuildMI(BB, PPC::OR, 2, PPC::R12).addReg(Tmp1).addReg(Tmp1);
1075 BuildMI(BB, PPC::MTCTR, 1).addReg(PPC::R12);
1076 CallMI = BuildMI(PPC::CALLindirect, 3).addImm(20).addImm(0)
1080 // Load the register args to virtual regs
1081 std::vector<unsigned> ArgVR;
1082 for(int i = 2, e = Node->getNumOperands(); i < e; ++i)
1083 ArgVR.push_back(SelectExpr(N.getOperand(i)));
1085 // Copy the virtual registers into the appropriate argument register
1086 for(int i = 0, e = ArgVR.size(); i < e; ++i) {
1087 switch(N.getOperand(i+2).getValueType()) {
1088 default: Node->dump(); assert(0 && "Unknown value type for call");
1094 assert(GPR_idx < 8 && "Too many int args");
1095 if (N.getOperand(i+2).getOpcode() != ISD::UNDEF) {
1096 BuildMI(BB, PPC::OR,2,GPR[GPR_idx]).addReg(ArgVR[i]).addReg(ArgVR[i]);
1097 CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use);
1103 assert(FPR_idx < 13 && "Too many fp args");
1104 BuildMI(BB, PPC::FMR, 1, FPR[FPR_idx]).addReg(ArgVR[i]);
1105 CallMI->addRegOperand(FPR[FPR_idx], MachineOperand::Use);
1111 // Put the call instruction in the correct place in the MachineBasicBlock
1112 BB->push_back(CallMI);
1114 switch (Node->getValueType(0)) {
1115 default: assert(0 && "Unknown value type for call result!");
1116 case MVT::Other: return 1;
1122 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R3).addReg(PPC::R3);
1126 BuildMI(BB, PPC::FMR, 1, Result).addReg(PPC::F1);
1129 return Result+N.ResNo;
1132 case ISD::SIGN_EXTEND:
1133 case ISD::SIGN_EXTEND_INREG:
1134 Tmp1 = SelectExpr(N.getOperand(0));
1135 switch(cast<MVTSDNode>(Node)->getExtraValueType()) {
1136 default: Node->dump(); assert(0 && "Unhandled SIGN_EXTEND type"); break;
1138 BuildMI(BB, PPC::EXTSW, 1, Result).addReg(Tmp1);
1141 BuildMI(BB, PPC::EXTSH, 1, Result).addReg(Tmp1);
1144 BuildMI(BB, PPC::EXTSB, 1, Result).addReg(Tmp1);
1147 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp1).addSImm(0);
1152 case ISD::ZERO_EXTEND_INREG:
1153 Tmp1 = SelectExpr(N.getOperand(0));
1154 switch(cast<MVTSDNode>(Node)->getExtraValueType()) {
1155 default: Node->dump(); assert(0 && "Unhandled ZERO_EXTEND type"); break;
1156 case MVT::i16: Tmp2 = 16; break;
1157 case MVT::i8: Tmp2 = 24; break;
1158 case MVT::i1: Tmp2 = 31; break;
1160 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(0).addImm(Tmp2)
1164 case ISD::CopyFromReg:
1166 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1167 Tmp1 = dyn_cast<RegSDNode>(Node)->getReg();
1168 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1172 Tmp1 = SelectExpr(N.getOperand(0));
1173 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1174 Tmp2 = CN->getValue() & 0x3F;
1175 BuildMI(BB, PPC::RLDICR, 3, Result).addReg(Tmp1).addImm(Tmp2)
1178 Tmp2 = SelectExpr(N.getOperand(1));
1179 BuildMI(BB, PPC::SLD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1184 Tmp1 = SelectExpr(N.getOperand(0));
1185 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1186 Tmp2 = CN->getValue() & 0x3F;
1187 BuildMI(BB, PPC::RLDICL, 3, Result).addReg(Tmp1).addImm(64-Tmp2)
1190 Tmp2 = SelectExpr(N.getOperand(1));
1191 BuildMI(BB, PPC::SRD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1196 Tmp1 = SelectExpr(N.getOperand(0));
1197 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1198 Tmp2 = CN->getValue() & 0x3F;
1199 BuildMI(BB, PPC::SRADI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1201 Tmp2 = SelectExpr(N.getOperand(1));
1202 BuildMI(BB, PPC::SRAD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1207 Tmp1 = SelectExpr(N.getOperand(0));
1208 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1209 default: assert(0 && "unhandled result code");
1210 case 0: // No immediate
1211 Tmp2 = SelectExpr(N.getOperand(1));
1212 BuildMI(BB, PPC::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1214 case 1: // Low immediate
1215 BuildMI(BB, PPC::ADDI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1217 case 2: // Shifted immediate
1218 BuildMI(BB, PPC::ADDIS, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1225 Tmp1 = SelectExpr(N.getOperand(0));
1226 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1227 default: assert(0 && "unhandled result code");
1228 case 0: // No immediate
1229 Tmp2 = SelectExpr(N.getOperand(1));
1231 case ISD::AND: Opc = PPC::AND; break;
1232 case ISD::OR: Opc = PPC::OR; break;
1234 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1236 case 1: // Low immediate
1238 case ISD::AND: Opc = PPC::ANDIo; break;
1239 case ISD::OR: Opc = PPC::ORI; break;
1241 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(Tmp2);
1243 case 2: // Shifted immediate
1245 case ISD::AND: Opc = PPC::ANDISo; break;
1246 case ISD::OR: Opc = PPC::ORIS; break;
1248 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(Tmp2);
1254 // Check for EQV: xor, (xor a, -1), b
1255 if (N.getOperand(0).getOpcode() == ISD::XOR &&
1256 N.getOperand(0).getOperand(1).getOpcode() == ISD::Constant &&
1257 cast<ConstantSDNode>(N.getOperand(0).getOperand(1))->isAllOnesValue()) {
1259 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1260 Tmp2 = SelectExpr(N.getOperand(1));
1261 BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1264 // Check for NOT, NOR, and NAND: xor (copy, or, and), -1
1265 if (N.getOperand(1).getOpcode() == ISD::Constant &&
1266 cast<ConstantSDNode>(N.getOperand(1))->isAllOnesValue()) {
1268 switch(N.getOperand(0).getOpcode()) {
1270 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1271 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1272 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1275 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1276 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1277 BuildMI(BB, PPC::NAND, 2, Result).addReg(Tmp1).addReg(Tmp2);
1280 Tmp1 = SelectExpr(N.getOperand(0));
1281 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1286 Tmp1 = SelectExpr(N.getOperand(0));
1287 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1288 default: assert(0 && "unhandled result code");
1289 case 0: // No immediate
1290 Tmp2 = SelectExpr(N.getOperand(1));
1291 BuildMI(BB, PPC::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1293 case 1: // Low immediate
1294 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1296 case 2: // Shifted immediate
1297 BuildMI(BB, PPC::XORIS, 2, Result).addReg(Tmp1).addImm(Tmp2);
1304 Tmp2 = SelectExpr(N.getOperand(1));
1305 if (1 == getImmediateForOpcode(N.getOperand(0), opcode, Tmp1))
1306 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp2).addSImm(Tmp1);
1308 Tmp1 = SelectExpr(N.getOperand(0));
1309 BuildMI(BB, PPC::SUBF, 2, Result).addReg(Tmp2).addReg(Tmp1);
1314 Tmp1 = SelectExpr(N.getOperand(0));
1315 if (1 == getImmediateForOpcode(N.getOperand(1), opcode, Tmp2))
1316 BuildMI(BB, PPC::MULLI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1318 Tmp2 = SelectExpr(N.getOperand(1));
1319 BuildMI(BB, PPC::MULLD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1325 if (3 == getImmediateForOpcode(N.getOperand(1), opcode, Tmp3)) {
1326 Tmp1 = MakeReg(MVT::i64);
1327 Tmp2 = SelectExpr(N.getOperand(0));
1328 BuildMI(BB, PPC::SRADI, 2, Tmp1).addReg(Tmp2).addImm(Tmp3);
1329 BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp1);
1332 Tmp1 = SelectExpr(N.getOperand(0));
1333 Tmp2 = SelectExpr(N.getOperand(1));
1334 Opc = (ISD::UDIV == opcode) ? PPC::DIVWU : PPC::DIVW;
1335 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1338 case ISD::FP_TO_UINT:
1339 case ISD::FP_TO_SINT: {
1340 Tmp1 = SelectExpr(N.getOperand(0));
1341 Tmp2 = MakeReg(MVT::f64);
1342 BuildMI(BB, PPC::FCTIDZ, 1, Tmp2).addReg(Tmp1);
1343 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
1344 addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(Tmp2), FrameIdx);
1345 addFrameReference(BuildMI(BB, PPC::LD, 2, Result), FrameIdx);
1350 if (SetCCSDNode *SetCC = dyn_cast<SetCCSDNode>(Node)) {
1351 Opc = SelectSetCR0(N);
1353 unsigned TrueValue = MakeReg(MVT::i32);
1354 BuildMI(BB, PPC::LI, 1, TrueValue).addSImm(1);
1355 unsigned FalseValue = MakeReg(MVT::i32);
1356 BuildMI(BB, PPC::LI, 1, FalseValue).addSImm(0);
1358 // Create an iterator with which to insert the MBB for copying the false
1359 // value and the MBB to hold the PHI instruction for this SetCC.
1360 MachineBasicBlock *thisMBB = BB;
1361 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1362 ilist<MachineBasicBlock>::iterator It = BB;
1367 // cmpTY cr0, r1, r2
1368 // %TrueValue = li 1
1370 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1371 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1372 BuildMI(BB, Opc, 2).addReg(PPC::CR0).addMBB(sinkMBB);
1373 MachineFunction *F = BB->getParent();
1374 F->getBasicBlockList().insert(It, copy0MBB);
1375 F->getBasicBlockList().insert(It, sinkMBB);
1376 // Update machine-CFG edges
1377 BB->addSuccessor(copy0MBB);
1378 BB->addSuccessor(sinkMBB);
1381 // %FalseValue = li 0
1384 // Update machine-CFG edges
1385 BB->addSuccessor(sinkMBB);
1388 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1391 BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue)
1392 .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB);
1395 assert(0 && "Is this legal?");
1399 unsigned TrueValue = SelectExpr(N.getOperand(1)); //Use if TRUE
1400 unsigned FalseValue = SelectExpr(N.getOperand(2)); //Use if FALSE
1401 Opc = SelectSetCR0(N.getOperand(0));
1403 // Create an iterator with which to insert the MBB for copying the false
1404 // value and the MBB to hold the PHI instruction for this SetCC.
1405 MachineBasicBlock *thisMBB = BB;
1406 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1407 ilist<MachineBasicBlock>::iterator It = BB;
1413 // cmpTY cr0, r1, r2
1415 // fallthrough --> copy0MBB
1416 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1417 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1418 BuildMI(BB, Opc, 2).addReg(PPC::CR0).addMBB(sinkMBB);
1419 MachineFunction *F = BB->getParent();
1420 F->getBasicBlockList().insert(It, copy0MBB);
1421 F->getBasicBlockList().insert(It, sinkMBB);
1422 // Update machine-CFG edges
1423 BB->addSuccessor(copy0MBB);
1424 BB->addSuccessor(sinkMBB);
1427 // %FalseValue = ...
1428 // # fallthrough to sinkMBB
1430 // Update machine-CFG edges
1431 BB->addSuccessor(sinkMBB);
1434 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1437 BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue)
1438 .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB);
1440 // FIXME: Select i64?
1445 switch (N.getValueType()) {
1446 default: assert(0 && "Cannot use constants of this type!");
1448 BuildMI(BB, PPC::LI, 1, Result)
1449 .addSImm(!cast<ConstantSDNode>(N)->isNullValue());
1453 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
1454 if (v < 32768 && v >= -32768) {
1455 BuildMI(BB, PPC::LI, 1, Result).addSImm(v);
1457 Tmp1 = MakeReg(MVT::i32);
1458 BuildMI(BB, PPC::LIS, 1, Tmp1).addSImm(v >> 16);
1459 BuildMI(BB, PPC::ORI, 2, Result).addReg(Tmp1).addImm(v & 0xFFFF);
1469 void ISel::Select(SDOperand N) {
1470 unsigned Tmp1, Tmp2, Opc;
1471 unsigned opcode = N.getOpcode();
1473 if (!ExprMap.insert(std::make_pair(N, 1)).second)
1474 return; // Already selected.
1476 SDNode *Node = N.Val;
1478 switch (Node->getOpcode()) {
1480 Node->dump(); std::cerr << "\n";
1481 assert(0 && "Node not handled yet!");
1482 case ISD::EntryToken: return; // Noop
1483 case ISD::TokenFactor:
1484 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1485 Select(Node->getOperand(i));
1487 case ISD::ADJCALLSTACKDOWN:
1488 case ISD::ADJCALLSTACKUP:
1489 Select(N.getOperand(0));
1490 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
1491 Opc = N.getOpcode() == ISD::ADJCALLSTACKDOWN ? PPC::ADJCALLSTACKDOWN :
1492 PPC::ADJCALLSTACKUP;
1493 BuildMI(BB, Opc, 1).addImm(Tmp1);
1496 MachineBasicBlock *Dest =
1497 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
1498 Select(N.getOperand(0));
1499 BuildMI(BB, PPC::B, 1).addMBB(Dest);
1505 case ISD::CopyToReg:
1506 Select(N.getOperand(0));
1507 Tmp1 = SelectExpr(N.getOperand(1));
1508 Tmp2 = cast<RegSDNode>(N)->getReg();
1511 if (N.getOperand(1).getValueType() == MVT::f64 ||
1512 N.getOperand(1).getValueType() == MVT::f32)
1513 BuildMI(BB, PPC::FMR, 1, Tmp2).addReg(Tmp1);
1515 BuildMI(BB, PPC::OR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
1518 case ISD::ImplicitDef:
1519 Select(N.getOperand(0));
1520 BuildMI(BB, PPC::IMPLICIT_DEF, 0, cast<RegSDNode>(N)->getReg());
1523 switch (N.getNumOperands()) {
1525 assert(0 && "Unknown return instruction!");
1527 assert(N.getOperand(1).getValueType() == MVT::i32 &&
1528 N.getOperand(2).getValueType() == MVT::i32 &&
1529 "Unknown two-register value!");
1530 Select(N.getOperand(0));
1531 Tmp1 = SelectExpr(N.getOperand(1));
1532 Tmp2 = SelectExpr(N.getOperand(2));
1533 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp2).addReg(Tmp2);
1534 BuildMI(BB, PPC::OR, 2, PPC::R4).addReg(Tmp1).addReg(Tmp1);
1537 Select(N.getOperand(0));
1538 Tmp1 = SelectExpr(N.getOperand(1));
1539 switch (N.getOperand(1).getValueType()) {
1541 assert(0 && "Unknown return type!");
1544 BuildMI(BB, PPC::FMR, 1, PPC::F1).addReg(Tmp1);
1547 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp1).addReg(Tmp1);
1551 Select(N.getOperand(0));
1554 BuildMI(BB, PPC::BLR, 0); // Just emit a 'ret' instruction
1556 case ISD::TRUNCSTORE:
1559 SDOperand Chain = N.getOperand(0);
1560 SDOperand Value = N.getOperand(1);
1561 SDOperand Address = N.getOperand(2);
1564 Tmp1 = SelectExpr(Value); //value
1566 if (opcode == ISD::STORE) {
1567 switch(Value.getValueType()) {
1568 default: assert(0 && "unknown Type in store");
1569 case MVT::i64: Opc = PPC::STD; break;
1570 case MVT::f64: Opc = PPC::STFD; break;
1571 case MVT::f32: Opc = PPC::STFS; break;
1573 } else { //ISD::TRUNCSTORE
1574 switch(cast<MVTSDNode>(Node)->getExtraValueType()) {
1575 default: assert(0 && "unknown Type in store");
1576 case MVT::i1: //FIXME: DAG does not promote this load
1577 case MVT::i8: Opc= PPC::STB; break;
1578 case MVT::i16: Opc = PPC::STH; break;
1579 case MVT::i32: Opc = PPC::STW; break;
1583 if(Address.getOpcode() == ISD::FrameIndex)
1585 Tmp2 = cast<FrameIndexSDNode>(Address)->getIndex();
1586 addFrameReference(BuildMI(BB, Opc, 3).addReg(Tmp1), (int)Tmp2);
1591 bool idx = SelectAddr(Address, Tmp2, offset);
1593 Opc = IndexedOpForOp(Opc);
1594 BuildMI(BB, Opc, 3).addReg(Tmp1).addReg(Tmp2).addReg(offset);
1596 BuildMI(BB, Opc, 3).addReg(Tmp1).addImm(offset).addReg(Tmp2);
1605 case ISD::CopyFromReg:
1607 case ISD::DYNAMIC_STACKALLOC:
1612 assert(0 && "Should not be reached!");
1616 /// createPPC32PatternInstructionSelector - This pass converts an LLVM function
1617 /// into a machine code representation using pattern matching and a machine
1618 /// description file.
1620 FunctionPass *llvm::createPPC64ISelPattern(TargetMachine &TM) {
1621 return new ISel(TM);