1 //===-- PPC32ISelPattern.cpp - A pattern matching inst selector for PPC32 -===//
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 32 bit PowerPC.
11 // Magic number generation for integer divide from the PowerPC Compiler Writer's
12 // Guide, section 3.2.3.5
14 //===----------------------------------------------------------------------===//
17 #include "PowerPCInstrBuilder.h"
18 #include "PowerPCInstrInfo.h"
19 #include "PPC32TargetMachine.h"
20 #include "llvm/Constants.h"
21 #include "llvm/Function.h"
22 #include "llvm/CodeGen/MachineConstantPool.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/SelectionDAG.h"
26 #include "llvm/CodeGen/SelectionDAGISel.h"
27 #include "llvm/CodeGen/SSARegMap.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Target/TargetLowering.h"
30 #include "llvm/Target/TargetOptions.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/ADT/Statistic.h"
39 //===----------------------------------------------------------------------===//
40 // PPC32TargetLowering - PPC32 Implementation of the TargetLowering interface
42 class PPC32TargetLowering : public TargetLowering {
43 int VarArgsFrameIndex; // FrameIndex for start of varargs area.
44 int ReturnAddrIndex; // FrameIndex for return slot.
46 PPC32TargetLowering(TargetMachine &TM) : TargetLowering(TM) {
47 // Fold away setcc operations if possible.
48 setSetCCIsExpensive();
50 // Set up the register classes.
51 addRegisterClass(MVT::i32, PPC32::GPRCRegisterClass);
52 addRegisterClass(MVT::f32, PPC32::FPRCRegisterClass);
53 addRegisterClass(MVT::f64, PPC32::FPRCRegisterClass);
55 // PowerPC has no intrinsics for these particular operations
56 setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
57 setOperationAction(ISD::MEMSET, MVT::Other, Expand);
58 setOperationAction(ISD::MEMCPY, MVT::Other, Expand);
60 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
61 setOperationAction(ISD::SEXTLOAD, MVT::i1, Expand);
62 setOperationAction(ISD::SEXTLOAD, MVT::i8, Expand);
64 // PowerPC has no SREM/UREM instructions
65 setOperationAction(ISD::SREM, MVT::i32, Expand);
66 setOperationAction(ISD::UREM, MVT::i32, Expand);
68 // We don't support sin/cos/sqrt/fmod
69 setOperationAction(ISD::FSIN , MVT::f64, Expand);
70 setOperationAction(ISD::FCOS , MVT::f64, Expand);
71 setOperationAction(ISD::SREM , MVT::f64, Expand);
72 setOperationAction(ISD::FSIN , MVT::f32, Expand);
73 setOperationAction(ISD::FCOS , MVT::f32, Expand);
74 setOperationAction(ISD::SREM , MVT::f32, Expand);
76 // If we're enabling GP optimizations, use hardware square root
77 if (!TM.getSubtarget<PPCSubtarget>().isGigaProcessor()) {
78 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
79 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
82 //PowerPC does not have CTPOP or CTTZ
83 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
84 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
86 setSetCCResultContents(ZeroOrOneSetCCResult);
87 addLegalFPImmediate(+0.0); // Necessary for FSEL
88 addLegalFPImmediate(-0.0); //
90 computeRegisterProperties();
93 /// LowerArguments - This hook must be implemented to indicate how we should
94 /// lower the arguments for the specified function, into the specified DAG.
95 virtual std::vector<SDOperand>
96 LowerArguments(Function &F, SelectionDAG &DAG);
98 /// LowerCallTo - This hook lowers an abstract call to a function into an
100 virtual std::pair<SDOperand, SDOperand>
101 LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, unsigned CC,
102 bool isTailCall, SDOperand Callee, ArgListTy &Args,
105 virtual SDOperand LowerVAStart(SDOperand Chain, SDOperand VAListP,
106 Value *VAListV, SelectionDAG &DAG);
108 virtual std::pair<SDOperand,SDOperand>
109 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
110 const Type *ArgTy, SelectionDAG &DAG);
112 virtual std::pair<SDOperand, SDOperand>
113 LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
119 std::vector<SDOperand>
120 PPC32TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
122 // add beautiful description of PPC stack frame format, or at least some docs
124 MachineFunction &MF = DAG.getMachineFunction();
125 MachineFrameInfo *MFI = MF.getFrameInfo();
126 MachineBasicBlock& BB = MF.front();
127 std::vector<SDOperand> ArgValues;
129 // Due to the rather complicated nature of the PowerPC ABI, rather than a
130 // fixed size array of physical args, for the sake of simplicity let the STL
131 // handle tracking them for us.
132 std::vector<unsigned> argVR, argPR, argOp;
133 unsigned ArgOffset = 24;
134 unsigned GPR_remaining = 8;
135 unsigned FPR_remaining = 13;
136 unsigned GPR_idx = 0, FPR_idx = 0;
137 static const unsigned GPR[] = {
138 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
139 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
141 static const unsigned FPR[] = {
142 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
143 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
146 // Add DAG nodes to load the arguments... On entry to a function on PPC,
147 // the arguments start at offset 24, although they are likely to be passed
149 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
150 SDOperand newroot, argt;
152 bool needsLoad = false;
153 bool ArgLive = !I->use_empty();
154 MVT::ValueType ObjectVT = getValueType(I->getType());
157 default: assert(0 && "Unhandled argument type!");
164 if (GPR_remaining > 0) {
165 MF.addLiveIn(GPR[GPR_idx]);
166 argt = newroot = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32,
168 if (ObjectVT != MVT::i32)
169 argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, newroot);
174 case MVT::i64: ObjSize = 8;
176 if (GPR_remaining > 0) {
177 SDOperand argHi, argLo;
178 MF.addLiveIn(GPR[GPR_idx]);
179 argHi = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot());
180 // If we have two or more remaining argument registers, then both halves
181 // of the i64 can be sourced from there. Otherwise, the lower half will
182 // have to come off the stack. This can happen when an i64 is preceded
183 // by 28 bytes of arguments.
184 if (GPR_remaining > 1) {
185 MF.addLiveIn(GPR[GPR_idx+1]);
186 argLo = DAG.getCopyFromReg(GPR[GPR_idx+1], MVT::i32, argHi);
188 int FI = MFI->CreateFixedObject(4, ArgOffset+4);
189 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
190 argLo = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
191 DAG.getSrcValue(NULL));
193 // Build the outgoing arg thingy
194 argt = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, argLo, argHi);
202 ObjSize = (ObjectVT == MVT::f64) ? 8 : 4;
204 if (FPR_remaining > 0) {
205 MF.addLiveIn(FPR[FPR_idx]);
206 argt = newroot = DAG.getCopyFromReg(FPR[FPR_idx], ObjectVT,
216 // We need to load the argument to a virtual register if we determined above
217 // that we ran out of physical registers of the appropriate type
219 unsigned SubregOffset = 0;
220 if (ObjectVT == MVT::i8 || ObjectVT == MVT::i1) SubregOffset = 3;
221 if (ObjectVT == MVT::i16) SubregOffset = 2;
222 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
223 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
224 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN,
225 DAG.getConstant(SubregOffset, MVT::i32));
226 argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
227 DAG.getSrcValue(NULL));
230 // Every 4 bytes of argument space consumes one of the GPRs available for
232 if (GPR_remaining > 0) {
233 unsigned delta = (GPR_remaining > 1 && ObjSize == 8) ? 2 : 1;
234 GPR_remaining -= delta;
237 ArgOffset += ObjSize;
239 DAG.setRoot(newroot.getValue(1));
241 ArgValues.push_back(argt);
244 // If the function takes variable number of arguments, make a frame index for
245 // the start of the first vararg value... for expansion of llvm.va_start.
247 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
248 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
249 // If this function is vararg, store any remaining integer argument regs
250 // to their spots on the stack so that they may be loaded by deferencing the
251 // result of va_next.
252 std::vector<SDOperand> MemOps;
253 for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) {
254 MF.addLiveIn(GPR[GPR_idx]);
255 SDOperand Val = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot());
256 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
257 Val, FIN, DAG.getSrcValue(NULL));
258 MemOps.push_back(Store);
259 // Increment the address by four for the next argument to store
260 SDOperand PtrOff = DAG.getConstant(4, getPointerTy());
261 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff);
263 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps));
266 // Finally, inform the code generator which regs we return values in.
267 switch (getValueType(F.getReturnType())) {
268 default: assert(0 && "Unknown type!");
269 case MVT::isVoid: break;
274 MF.addLiveOut(PPC::R3);
277 MF.addLiveOut(PPC::R3);
278 MF.addLiveOut(PPC::R4);
282 MF.addLiveOut(PPC::F1);
289 std::pair<SDOperand, SDOperand>
290 PPC32TargetLowering::LowerCallTo(SDOperand Chain,
291 const Type *RetTy, bool isVarArg,
292 unsigned CallingConv, bool isTailCall,
293 SDOperand Callee, ArgListTy &Args,
295 // args_to_use will accumulate outgoing args for the ISD::CALL case in
296 // SelectExpr to use to put the arguments in the appropriate registers.
297 std::vector<SDOperand> args_to_use;
299 // Count how many bytes are to be pushed on the stack, including the linkage
300 // area, and parameter passing area.
301 unsigned NumBytes = 24;
304 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
305 DAG.getConstant(NumBytes, getPointerTy()));
307 for (unsigned i = 0, e = Args.size(); i != e; ++i)
308 switch (getValueType(Args[i].second)) {
309 default: assert(0 && "Unknown value type!");
323 // Just to be safe, we'll always reserve the full 24 bytes of linkage area
324 // plus 32 bytes of argument space in case any called code gets funky on us.
325 // (Required by ABI to support var arg)
326 if (NumBytes < 56) NumBytes = 56;
328 // Adjust the stack pointer for the new arguments...
329 // These operations are automatically eliminated by the prolog/epilog pass
330 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
331 DAG.getConstant(NumBytes, getPointerTy()));
333 // Set up a copy of the stack pointer for use loading and storing any
334 // arguments that may not fit in the registers available for argument
336 SDOperand StackPtr = DAG.getCopyFromReg(PPC::R1, MVT::i32,
339 // Figure out which arguments are going to go in registers, and which in
340 // memory. Also, if this is a vararg function, floating point operations
341 // must be stored to our stack, and loaded into integer regs as well, if
342 // any integer regs are available for argument passing.
343 unsigned ArgOffset = 24;
344 unsigned GPR_remaining = 8;
345 unsigned FPR_remaining = 13;
347 std::vector<SDOperand> MemOps;
348 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
349 // PtrOff will be used to store the current argument to the stack if a
350 // register cannot be found for it.
351 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
352 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
353 MVT::ValueType ArgVT = getValueType(Args[i].second);
356 default: assert(0 && "Unexpected ValueType for argument!");
360 // Promote the integer to 32 bits. If the input type is signed use a
361 // sign extend, otherwise use a zero extend.
362 if (Args[i].second->isSigned())
363 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
365 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
368 if (GPR_remaining > 0) {
369 args_to_use.push_back(Args[i].first);
372 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
373 Args[i].first, PtrOff,
374 DAG.getSrcValue(NULL)));
379 // If we have one free GPR left, we can place the upper half of the i64
380 // in it, and store the other half to the stack. If we have two or more
381 // free GPRs, then we can pass both halves of the i64 in registers.
382 if (GPR_remaining > 0) {
383 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
384 Args[i].first, DAG.getConstant(1, MVT::i32));
385 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
386 Args[i].first, DAG.getConstant(0, MVT::i32));
387 args_to_use.push_back(Hi);
389 if (GPR_remaining > 0) {
390 args_to_use.push_back(Lo);
393 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
394 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
395 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
396 Lo, PtrOff, DAG.getSrcValue(NULL)));
399 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
400 Args[i].first, PtrOff,
401 DAG.getSrcValue(NULL)));
407 if (FPR_remaining > 0) {
408 args_to_use.push_back(Args[i].first);
411 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
412 Args[i].first, PtrOff,
413 DAG.getSrcValue(NULL));
414 MemOps.push_back(Store);
415 // Float varargs are always shadowed in available integer registers
416 if (GPR_remaining > 0) {
417 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
418 DAG.getSrcValue(NULL));
419 MemOps.push_back(Load);
420 args_to_use.push_back(Load);
423 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
424 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
425 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
426 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
427 DAG.getSrcValue(NULL));
428 MemOps.push_back(Load);
429 args_to_use.push_back(Load);
433 // If we have any FPRs remaining, we may also have GPRs remaining.
434 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
436 if (GPR_remaining > 0) {
437 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
440 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
441 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
446 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
447 Args[i].first, PtrOff,
448 DAG.getSrcValue(NULL)));
450 ArgOffset += (ArgVT == MVT::f32) ? 4 : 8;
455 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
458 std::vector<MVT::ValueType> RetVals;
459 MVT::ValueType RetTyVT = getValueType(RetTy);
460 if (RetTyVT != MVT::isVoid)
461 RetVals.push_back(RetTyVT);
462 RetVals.push_back(MVT::Other);
464 SDOperand TheCall = SDOperand(DAG.getCall(RetVals,
465 Chain, Callee, args_to_use), 0);
466 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
467 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
468 DAG.getConstant(NumBytes, getPointerTy()));
469 return std::make_pair(TheCall, Chain);
472 SDOperand PPC32TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
473 Value *VAListV, SelectionDAG &DAG) {
474 // vastart just stores the address of the VarArgsFrameIndex slot into the
475 // memory location argument.
476 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
477 return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR, VAListP,
478 DAG.getSrcValue(VAListV));
481 std::pair<SDOperand,SDOperand>
482 PPC32TargetLowering::LowerVAArg(SDOperand Chain,
483 SDOperand VAListP, Value *VAListV,
484 const Type *ArgTy, SelectionDAG &DAG) {
485 MVT::ValueType ArgVT = getValueType(ArgTy);
488 DAG.getLoad(MVT::i32, Chain, VAListP, DAG.getSrcValue(VAListV));
489 SDOperand Result = DAG.getLoad(ArgVT, Chain, VAList, DAG.getSrcValue(NULL));
491 if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
494 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
495 "Other types should have been promoted for varargs!");
498 VAList = DAG.getNode(ISD::ADD, VAList.getValueType(), VAList,
499 DAG.getConstant(Amt, VAList.getValueType()));
500 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
501 VAList, VAListP, DAG.getSrcValue(VAListV));
502 return std::make_pair(Result, Chain);
506 std::pair<SDOperand, SDOperand> PPC32TargetLowering::
507 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
509 assert(0 && "LowerFrameReturnAddress unimplemented");
514 Statistic<>Recorded("ppc-codegen", "Number of recording ops emitted");
515 Statistic<>FusedFP("ppc-codegen", "Number of fused fp operations");
516 Statistic<>FrameOff("ppc-codegen", "Number of frame idx offsets collapsed");
518 //===--------------------------------------------------------------------===//
519 /// ISel - PPC32 specific code to select PPC32 machine instructions for
520 /// SelectionDAG operations.
521 //===--------------------------------------------------------------------===//
522 class ISel : public SelectionDAGISel {
523 PPC32TargetLowering PPC32Lowering;
524 SelectionDAG *ISelDAG; // Hack to support us having a dag->dag transform
525 // for sdiv and udiv until it is put into the future
528 /// ExprMap - As shared expressions are codegen'd, we keep track of which
529 /// vreg the value is produced in, so we only emit one copy of each compiled
531 std::map<SDOperand, unsigned> ExprMap;
533 unsigned GlobalBaseReg;
534 bool GlobalBaseInitialized;
537 ISel(TargetMachine &TM) : SelectionDAGISel(PPC32Lowering), PPC32Lowering(TM),
540 /// runOnFunction - Override this function in order to reset our per-function
542 virtual bool runOnFunction(Function &Fn) {
543 // Make sure we re-emit a set of the global base reg if necessary
544 GlobalBaseInitialized = false;
545 return SelectionDAGISel::runOnFunction(Fn);
548 /// InstructionSelectBasicBlock - This callback is invoked by
549 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
550 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG) {
552 // Codegen the basic block.
554 Select(DAG.getRoot());
556 // Clear state used for selection.
561 // dag -> dag expanders for integer divide by constant
562 SDOperand BuildSDIVSequence(SDOperand N);
563 SDOperand BuildUDIVSequence(SDOperand N);
565 unsigned getGlobalBaseReg();
566 unsigned getConstDouble(double floatVal, unsigned Result);
567 void MoveCRtoGPR(unsigned CCReg, bool Inv, unsigned Idx, unsigned Result);
568 bool SelectBitfieldInsert(SDOperand OR, unsigned Result);
569 unsigned FoldIfWideZeroExtend(SDOperand N);
570 unsigned SelectCC(SDOperand CC, unsigned &Opc, bool &Inv, unsigned &Idx);
571 unsigned SelectCCExpr(SDOperand N, unsigned& Opc, bool &Inv, unsigned &Idx);
572 unsigned SelectExpr(SDOperand N, bool Recording=false);
573 void Select(SDOperand N);
575 unsigned SelectAddr(SDOperand N, unsigned& Reg, int& offset);
576 void SelectBranchCC(SDOperand N);
578 virtual const char *getPassName() const {
579 return "PowerPC Pattern Instruction Selection";
583 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
584 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
585 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
586 // not, since all 1s are not contiguous.
587 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
588 if (isShiftedMask_32(Val)) {
589 // look for the first non-zero bit
590 MB = CountLeadingZeros_32(Val);
591 // look for the first zero bit after the run of ones
592 ME = CountLeadingZeros_32((Val - 1) ^ Val);
594 } else if (isShiftedMask_32(Val = ~Val)) { // invert mask
595 // effectively look for the first zero bit
596 ME = CountLeadingZeros_32(Val) - 1;
597 // effectively look for the first one bit after the run of zeros
598 MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
605 // isRotateAndMask - Returns true if Mask and Shift can be folded in to a rotate
606 // and mask opcode and mask operation.
607 static bool isRotateAndMask(unsigned Opcode, unsigned Shift, unsigned Mask,
609 unsigned &SH, unsigned &MB, unsigned &ME) {
610 if (Shift > 31) return false;
611 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
613 if (Opcode == ISD::SHL) { // shift left
614 // apply shift to mask if it comes first
615 if (IsShiftMask) Mask = Mask << Shift;
616 // determine which bits are made indeterminant by shift
617 Indeterminant = ~(0xFFFFFFFFu << Shift);
618 } else if (Opcode == ISD::SRA || Opcode == ISD::SRL) { // shift rights
619 // apply shift to mask if it comes first
620 if (IsShiftMask) Mask = Mask >> Shift;
621 // determine which bits are made indeterminant by shift
622 Indeterminant = ~(0xFFFFFFFFu >> Shift);
623 // adjust for the left rotate
627 // if the mask doesn't intersect any Indeterminant bits
628 if (!(Mask & Indeterminant)) {
630 // make sure the mask is still a mask (wrap arounds may not be)
631 return isRunOfOnes(Mask, MB, ME);
638 // isImmediate - This method tests to see if a constant operand.
639 // If so Imm will receive the 32 bit value.
640 static bool isImmediate(SDOperand N, unsigned& Imm) {
642 if (N.getOpcode() == ISD::Constant) {
644 Imm = (unsigned)cast<ConstantSDNode>(N)->getSignExtended();
652 // isOprShiftImm - Returns true if the specified operand is a shift opcode with
653 // a immediate shift count less than 32.
654 static bool isOprShiftImm(SDOperand N, unsigned& Opc, unsigned& SH) {
656 return (Opc == ISD::SHL || Opc == ISD::SRL || Opc == ISD::SRA) &&
657 isImmediate(N.getOperand(1), SH) && SH < 32;
660 // isOprNot - Returns true if the specified operand is an xor with immediate -1.
661 static bool isOprNot(SDOperand N) {
663 return N.getOpcode() == ISD::XOR &&
664 isImmediate(N.getOperand(1), Imm) && (signed)Imm == -1;
667 // Immediate constant composers.
668 // Lo16 - grabs the lo 16 bits from a 32 bit constant.
669 // Hi16 - grabs the hi 16 bits from a 32 bit constant.
670 // HA16 - computes the hi bits required if the lo bits are add/subtracted in
672 static unsigned Lo16(unsigned x) { return x & 0x0000FFFF; }
673 static unsigned Hi16(unsigned x) { return Lo16(x >> 16); }
674 static unsigned HA16(unsigned x) { return Hi16((signed)x - (signed short)x); }
676 /// getImmediateForOpcode - This method returns a value indicating whether
677 /// the ConstantSDNode N can be used as an immediate to Opcode. The return
678 /// values are either 0, 1 or 2. 0 indicates that either N is not a
679 /// ConstantSDNode, or is not suitable for use by that opcode.
680 /// Return value codes for turning into an enum someday:
681 /// 1: constant may be used in normal immediate form.
682 /// 2: constant may be used in shifted immediate form.
683 /// 3: log base 2 of the constant may be used.
684 /// 4: constant is suitable for integer division conversion
685 /// 5: constant is a bitfield mask
687 static unsigned getImmediateForOpcode(SDOperand N, unsigned Opcode,
688 unsigned& Imm, bool U = false) {
689 if (N.getOpcode() != ISD::Constant) return 0;
691 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
696 if (isInt16(v)) { Imm = v & 0xFFFF; return 1; }
697 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
701 if (isRunOfOnes(v, MB, ME)) { Imm = MB << 16 | ME & 0xFFFF; return 5; }
702 if (isUInt16(v)) { Imm = v & 0xFFFF; return 1; }
703 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
708 if (isUInt16(v)) { Imm = v & 0xFFFF; return 1; }
709 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
712 if (isInt16(v)) { Imm = v & 0xFFFF; return 1; }
715 // handle subtract-from separately from subtract, since subi is really addi
716 if (U && isInt16(v)) { Imm = v & 0xFFFF; return 1; }
717 if (!U && isInt16(-v)) { Imm = (-v) & 0xFFFF; return 1; }
720 if (U && isUInt16(v)) { Imm = v & 0xFFFF; return 1; }
721 if (!U && isInt16(v)) { Imm = v & 0xFFFF; return 1; }
724 if (isPowerOf2_32(v)) { Imm = Log2_32(v); return 3; }
725 if (isPowerOf2_32(-v)) { Imm = Log2_32(-v); return 3; }
726 if (v <= -2 || v >= 2) { return 4; }
729 if (v > 1) { return 4; }
735 /// NodeHasRecordingVariant - If SelectExpr can always produce code for
736 /// NodeOpcode that also sets CR0 as a side effect, return true. Otherwise,
738 static bool NodeHasRecordingVariant(unsigned NodeOpcode) {
740 default: return false;
747 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
748 /// to Condition. If the Condition is unordered or unsigned, the bool argument
749 /// U is set to true, otherwise it is set to false.
750 static unsigned getBCCForSetCC(unsigned Condition, bool& U) {
753 default: assert(0 && "Unknown condition!"); abort();
754 case ISD::SETEQ: return PPC::BEQ;
755 case ISD::SETNE: return PPC::BNE;
756 case ISD::SETULT: U = true;
757 case ISD::SETLT: return PPC::BLT;
758 case ISD::SETULE: U = true;
759 case ISD::SETLE: return PPC::BLE;
760 case ISD::SETUGT: U = true;
761 case ISD::SETGT: return PPC::BGT;
762 case ISD::SETUGE: U = true;
763 case ISD::SETGE: return PPC::BGE;
768 /// getCROpForOp - Return the condition register opcode (or inverted opcode)
769 /// associated with the SelectionDAG opcode.
770 static unsigned getCROpForSetCC(unsigned Opcode, bool Inv1, bool Inv2) {
772 default: assert(0 && "Unknown opcode!"); abort();
774 if (Inv1 && Inv2) return PPC::CRNOR; // De Morgan's Law
775 if (!Inv1 && !Inv2) return PPC::CRAND;
776 if (Inv1 ^ Inv2) return PPC::CRANDC;
778 if (Inv1 && Inv2) return PPC::CRNAND; // De Morgan's Law
779 if (!Inv1 && !Inv2) return PPC::CROR;
780 if (Inv1 ^ Inv2) return PPC::CRORC;
785 /// getCRIdxForSetCC - Return the index of the condition register field
786 /// associated with the SetCC condition, and whether or not the field is
787 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
788 static unsigned getCRIdxForSetCC(unsigned Condition, bool& Inv) {
790 default: assert(0 && "Unknown condition!"); abort();
792 case ISD::SETLT: Inv = false; return 0;
794 case ISD::SETGE: Inv = true; return 0;
796 case ISD::SETGT: Inv = false; return 1;
798 case ISD::SETLE: Inv = true; return 1;
799 case ISD::SETEQ: Inv = false; return 2;
800 case ISD::SETNE: Inv = true; return 2;
805 /// IndexedOpForOp - Return the indexed variant for each of the PowerPC load
806 /// and store immediate instructions.
807 static unsigned IndexedOpForOp(unsigned Opcode) {
809 default: assert(0 && "Unknown opcode!"); abort();
810 case PPC::LBZ: return PPC::LBZX; case PPC::STB: return PPC::STBX;
811 case PPC::LHZ: return PPC::LHZX; case PPC::STH: return PPC::STHX;
812 case PPC::LHA: return PPC::LHAX; case PPC::STW: return PPC::STWX;
813 case PPC::LWZ: return PPC::LWZX; case PPC::STFS: return PPC::STFSX;
814 case PPC::LFS: return PPC::LFSX; case PPC::STFD: return PPC::STFDX;
815 case PPC::LFD: return PPC::LFDX;
820 // Structure used to return the necessary information to codegen an SDIV as
823 int m; // magic number
824 int s; // shift amount
828 unsigned int m; // magic number
829 int a; // add indicator
830 int s; // shift amount
833 /// magic - calculate the magic numbers required to codegen an integer sdiv as
834 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
836 static struct ms magic(int d) {
838 unsigned int ad, anc, delta, q1, r1, q2, r2, t;
839 const unsigned int two31 = 0x80000000U;
843 t = two31 + ((unsigned int)d >> 31);
844 anc = t - 1 - t%ad; // absolute value of nc
845 p = 31; // initialize p
846 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
847 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
848 q2 = two31/ad; // initialize q2 = 2p/abs(d)
849 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
852 q1 = 2*q1; // update q1 = 2p/abs(nc)
853 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
854 if (r1 >= anc) { // must be unsigned comparison
858 q2 = 2*q2; // update q2 = 2p/abs(d)
859 r2 = 2*r2; // update r2 = rem(2p/abs(d))
860 if (r2 >= ad) { // must be unsigned comparison
865 } while (q1 < delta || (q1 == delta && r1 == 0));
868 if (d < 0) mag.m = -mag.m; // resulting magic number
869 mag.s = p - 32; // resulting shift
873 /// magicu - calculate the magic numbers required to codegen an integer udiv as
874 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
875 static struct mu magicu(unsigned d)
878 unsigned int nc, delta, q1, r1, q2, r2;
880 magu.a = 0; // initialize "add" indicator
882 p = 31; // initialize p
883 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
884 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
885 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
886 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
889 if (r1 >= nc - r1 ) {
890 q1 = 2*q1 + 1; // update q1
891 r1 = 2*r1 - nc; // update r1
894 q1 = 2*q1; // update q1
895 r1 = 2*r1; // update r1
897 if (r2 + 1 >= d - r2) {
898 if (q2 >= 0x7FFFFFFF) magu.a = 1;
899 q2 = 2*q2 + 1; // update q2
900 r2 = 2*r2 + 1 - d; // update r2
903 if (q2 >= 0x80000000) magu.a = 1;
904 q2 = 2*q2; // update q2
905 r2 = 2*r2 + 1; // update r2
908 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
909 magu.m = q2 + 1; // resulting magic number
910 magu.s = p - 32; // resulting shift
915 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
916 /// return a DAG expression to select that will generate the same value by
917 /// multiplying by a magic number. See:
918 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
919 SDOperand ISel::BuildSDIVSequence(SDOperand N) {
920 int d = (int)cast<ConstantSDNode>(N.getOperand(1))->getSignExtended();
921 ms magics = magic(d);
922 // Multiply the numerator (operand 0) by the magic value
923 SDOperand Q = ISelDAG->getNode(ISD::MULHS, MVT::i32, N.getOperand(0),
924 ISelDAG->getConstant(magics.m, MVT::i32));
925 // If d > 0 and m < 0, add the numerator
926 if (d > 0 && magics.m < 0)
927 Q = ISelDAG->getNode(ISD::ADD, MVT::i32, Q, N.getOperand(0));
928 // If d < 0 and m > 0, subtract the numerator.
929 if (d < 0 && magics.m > 0)
930 Q = ISelDAG->getNode(ISD::SUB, MVT::i32, Q, N.getOperand(0));
931 // Shift right algebraic if shift value is nonzero
933 Q = ISelDAG->getNode(ISD::SRA, MVT::i32, Q,
934 ISelDAG->getConstant(magics.s, MVT::i32));
935 // Extract the sign bit and add it to the quotient
937 ISelDAG->getNode(ISD::SRL, MVT::i32, Q, ISelDAG->getConstant(31, MVT::i32));
938 return ISelDAG->getNode(ISD::ADD, MVT::i32, Q, T);
941 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
942 /// return a DAG expression to select that will generate the same value by
943 /// multiplying by a magic number. See:
944 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
945 SDOperand ISel::BuildUDIVSequence(SDOperand N) {
947 (unsigned)cast<ConstantSDNode>(N.getOperand(1))->getSignExtended();
948 mu magics = magicu(d);
949 // Multiply the numerator (operand 0) by the magic value
950 SDOperand Q = ISelDAG->getNode(ISD::MULHU, MVT::i32, N.getOperand(0),
951 ISelDAG->getConstant(magics.m, MVT::i32));
953 Q = ISelDAG->getNode(ISD::SRL, MVT::i32, Q,
954 ISelDAG->getConstant(magics.s, MVT::i32));
956 SDOperand NPQ = ISelDAG->getNode(ISD::SUB, MVT::i32, N.getOperand(0), Q);
957 NPQ = ISelDAG->getNode(ISD::SRL, MVT::i32, NPQ,
958 ISelDAG->getConstant(1, MVT::i32));
959 NPQ = ISelDAG->getNode(ISD::ADD, MVT::i32, NPQ, Q);
960 Q = ISelDAG->getNode(ISD::SRL, MVT::i32, NPQ,
961 ISelDAG->getConstant(magics.s-1, MVT::i32));
966 /// getGlobalBaseReg - Output the instructions required to put the
967 /// base address to use for accessing globals into a register.
969 unsigned ISel::getGlobalBaseReg() {
970 if (!GlobalBaseInitialized) {
971 // Insert the set of GlobalBaseReg into the first MBB of the function
972 MachineBasicBlock &FirstMBB = BB->getParent()->front();
973 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
974 GlobalBaseReg = MakeReg(MVT::i32);
975 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
976 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg).addReg(PPC::LR);
977 GlobalBaseInitialized = true;
979 return GlobalBaseReg;
982 /// getConstDouble - Loads a floating point value into a register, via the
983 /// Constant Pool. Optionally takes a register in which to load the value.
984 unsigned ISel::getConstDouble(double doubleVal, unsigned Result=0) {
985 unsigned Tmp1 = MakeReg(MVT::i32);
986 if (0 == Result) Result = MakeReg(MVT::f64);
987 MachineConstantPool *CP = BB->getParent()->getConstantPool();
988 ConstantFP *CFP = ConstantFP::get(Type::DoubleTy, doubleVal);
989 unsigned CPI = CP->getConstantPoolIndex(CFP);
991 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
992 .addConstantPoolIndex(CPI);
994 BuildMI(BB, PPC::LIS, 1, Tmp1).addConstantPoolIndex(CPI);
995 BuildMI(BB, PPC::LFD, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
999 /// MoveCRtoGPR - Move CCReg[Idx] to the least significant bit of Result. If
1000 /// Inv is true, then invert the result.
1001 void ISel::MoveCRtoGPR(unsigned CCReg, bool Inv, unsigned Idx, unsigned Result){
1002 unsigned IntCR = MakeReg(MVT::i32);
1003 BuildMI(BB, PPC::MCRF, 1, PPC::CR7).addReg(CCReg);
1005 TLI.getTargetMachine().getSubtarget<PPCSubtarget>().isGigaProcessor();
1006 BuildMI(BB, GPOpt ? PPC::MFOCRF : PPC::MFCR, 1, IntCR).addReg(PPC::CR7);
1008 unsigned Tmp1 = MakeReg(MVT::i32);
1009 BuildMI(BB, PPC::RLWINM, 4, Tmp1).addReg(IntCR).addImm(32-(3-Idx))
1010 .addImm(31).addImm(31);
1011 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp1).addImm(1);
1013 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(IntCR).addImm(32-(3-Idx))
1014 .addImm(31).addImm(31);
1018 /// SelectBitfieldInsert - turn an or of two masked values into
1019 /// the rotate left word immediate then mask insert (rlwimi) instruction.
1020 /// Returns true on success, false if the caller still needs to select OR.
1022 /// Patterns matched:
1023 /// 1. or shl, and 5. or and, and
1024 /// 2. or and, shl 6. or shl, shr
1025 /// 3. or shr, and 7. or shr, shl
1027 bool ISel::SelectBitfieldInsert(SDOperand OR, unsigned Result) {
1028 bool IsRotate = false;
1029 unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, Amount = 0;
1031 SDOperand Op0 = OR.getOperand(0);
1032 SDOperand Op1 = OR.getOperand(1);
1034 unsigned Op0Opc = Op0.getOpcode();
1035 unsigned Op1Opc = Op1.getOpcode();
1037 // Verify that we have the correct opcodes
1038 if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc)
1040 if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc)
1043 // Generate Mask value for Target
1044 if (ConstantSDNode *CN =
1045 dyn_cast<ConstantSDNode>(Op0.getOperand(1).Val)) {
1047 case ISD::SHL: TgtMask <<= (unsigned)CN->getValue(); break;
1048 case ISD::SRL: TgtMask >>= (unsigned)CN->getValue(); break;
1049 case ISD::AND: TgtMask &= (unsigned)CN->getValue(); break;
1055 // Generate Mask value for Insert
1056 if (ConstantSDNode *CN =
1057 dyn_cast<ConstantSDNode>(Op1.getOperand(1).Val)) {
1060 Amount = CN->getValue();
1062 if (Op0Opc == ISD::SRL) IsRotate = true;
1065 Amount = CN->getValue();
1068 if (Op0Opc == ISD::SHL) IsRotate = true;
1071 InsMask &= (unsigned)CN->getValue();
1080 // If both of the inputs are ANDs and one of them has a logical shift by
1081 // constant as its input, make that the inserted value so that we can combine
1082 // the shift into the rotate part of the rlwimi instruction
1083 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
1084 if (Op1.getOperand(0).getOpcode() == ISD::SHL ||
1085 Op1.getOperand(0).getOpcode() == ISD::SRL) {
1086 if (ConstantSDNode *CN =
1087 dyn_cast<ConstantSDNode>(Op1.getOperand(0).getOperand(1).Val)) {
1088 Amount = Op1.getOperand(0).getOpcode() == ISD::SHL ?
1089 CN->getValue() : 32 - CN->getValue();
1090 Tmp3 = SelectExpr(Op1.getOperand(0).getOperand(0));
1092 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
1093 Op0.getOperand(0).getOpcode() == ISD::SRL) {
1094 if (ConstantSDNode *CN =
1095 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(1).Val)) {
1096 std::swap(Op0, Op1);
1097 std::swap(TgtMask, InsMask);
1098 Amount = Op1.getOperand(0).getOpcode() == ISD::SHL ?
1099 CN->getValue() : 32 - CN->getValue();
1100 Tmp3 = SelectExpr(Op1.getOperand(0).getOperand(0));
1105 // Verify that the Target mask and Insert mask together form a full word mask
1106 // and that the Insert mask is a run of set bits (which implies both are runs
1107 // of set bits). Given that, Select the arguments and generate the rlwimi
1110 if (((TgtMask & InsMask) == 0) && isRunOfOnes(InsMask, MB, ME)) {
1111 unsigned Tmp1, Tmp2;
1112 bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF;
1113 // Check for rotlwi / rotrwi here, a special case of bitfield insert
1114 // where both bitfield halves are sourced from the same value.
1115 if (IsRotate && fullMask &&
1116 OR.getOperand(0).getOperand(0) == OR.getOperand(1).getOperand(0)) {
1117 Tmp1 = SelectExpr(OR.getOperand(0).getOperand(0));
1118 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(Amount)
1119 .addImm(0).addImm(31);
1122 if (Op0Opc == ISD::AND && fullMask)
1123 Tmp1 = SelectExpr(Op0.getOperand(0));
1125 Tmp1 = SelectExpr(Op0);
1126 Tmp2 = Tmp3 ? Tmp3 : SelectExpr(Op1.getOperand(0));
1127 BuildMI(BB, PPC::RLWIMI, 5, Result).addReg(Tmp1).addReg(Tmp2)
1128 .addImm(Amount).addImm(MB).addImm(ME);
1134 /// FoldIfWideZeroExtend - 32 bit PowerPC implicit masks shift amounts to the
1135 /// low six bits. If the shift amount is an ISD::AND node with a mask that is
1136 /// wider than the implicit mask, then we can get rid of the AND and let the
1137 /// shift do the mask.
1138 unsigned ISel::FoldIfWideZeroExtend(SDOperand N) {
1140 if (N.getOpcode() == ISD::AND &&
1141 isImmediate(N.getOperand(1), C) && isRunOfOnes(C, MB, ME) &&
1142 MB <= 26 && ME == 31)
1143 return SelectExpr(N.getOperand(0));
1145 return SelectExpr(N);
1148 unsigned ISel::SelectCC(SDOperand CC, unsigned& Opc, bool &Inv, unsigned& Idx) {
1149 unsigned Result, Tmp1, Tmp2;
1150 bool AlreadySelected = false;
1151 static const unsigned CompareOpcodes[] =
1152 { PPC::FCMPU, PPC::FCMPU, PPC::CMPW, PPC::CMPLW };
1154 // Allocate a condition register for this expression
1155 Result = RegMap->createVirtualRegister(PPC32::CRRCRegisterClass);
1157 // If the first operand to the select is a SETCC node, then we can fold it
1158 // into the branch that selects which value to return.
1159 if (SetCCSDNode* SetCC = dyn_cast<SetCCSDNode>(CC.Val)) {
1161 Opc = getBCCForSetCC(SetCC->getCondition(), U);
1162 Idx = getCRIdxForSetCC(SetCC->getCondition(), Inv);
1164 // Use U to determine whether the SETCC immediate range is signed or not.
1165 if (isImmediate(SetCC->getOperand(1), Tmp2) &&
1166 ((U && isUInt16(Tmp2)) || (!U && isInt16(Tmp2)))) {
1168 // For comparisons against zero, we can implicity set CR0 if a recording
1169 // variant (e.g. 'or.' instead of 'or') of the instruction that defines
1170 // operand zero of the SetCC node is available.
1172 NodeHasRecordingVariant(SetCC->getOperand(0).getOpcode()) &&
1173 SetCC->getOperand(0).Val->hasOneUse()) {
1174 RecordSuccess = false;
1175 Tmp1 = SelectExpr(SetCC->getOperand(0), true);
1176 if (RecordSuccess) {
1178 BuildMI(BB, PPC::MCRF, 1, Result).addReg(PPC::CR0);
1181 AlreadySelected = true;
1183 // If we could not implicitly set CR0, then emit a compare immediate
1185 if (!AlreadySelected) Tmp1 = SelectExpr(SetCC->getOperand(0));
1187 BuildMI(BB, PPC::CMPLWI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1189 BuildMI(BB, PPC::CMPWI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1191 bool IsInteger = MVT::isInteger(SetCC->getOperand(0).getValueType());
1192 unsigned CompareOpc = CompareOpcodes[2 * IsInteger + U];
1193 Tmp1 = SelectExpr(SetCC->getOperand(0));
1194 Tmp2 = SelectExpr(SetCC->getOperand(1));
1195 BuildMI(BB, CompareOpc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1198 // If this isn't a SetCC, then select the value and compare it against zero,
1199 // treating it as if it were a boolean.
1201 Idx = getCRIdxForSetCC(ISD::SETNE, Inv);
1202 Tmp1 = SelectExpr(CC);
1203 BuildMI(BB, PPC::CMPLWI, 2, Result).addReg(Tmp1).addImm(0);
1208 unsigned ISel::SelectCCExpr(SDOperand N, unsigned& Opc, bool &Inv,
1211 unsigned Idx0, Idx1, CROpc, Opc1, Tmp1, Tmp2;
1213 // Allocate a condition register for this expression
1214 unsigned Result = RegMap->createVirtualRegister(PPC32::CRRCRegisterClass);
1216 // Check for the operations we support:
1217 switch(N.getOpcode()) {
1220 Idx = getCRIdxForSetCC(ISD::SETNE, Inv);
1221 Tmp1 = SelectExpr(N);
1222 BuildMI(BB, PPC::CMPLWI, 2, Result).addReg(Tmp1).addImm(0);
1226 Tmp1 = SelectCCExpr(N.getOperand(0), Opc, Inv0, Idx0);
1227 Tmp2 = SelectCCExpr(N.getOperand(1), Opc1, Inv1, Idx1);
1228 CROpc = getCROpForSetCC(N.getOpcode(), Inv0, Inv1);
1229 if (Inv0 && !Inv1) {
1230 std::swap(Tmp1, Tmp2);
1231 std::swap(Idx0, Idx1);
1234 if (Inv0 && Inv1) Opc = PPC32InstrInfo::invertPPCBranchOpcode(Opc);
1235 BuildMI(BB, CROpc, 5, Result).addImm(Idx0).addReg(Tmp1).addImm(Idx0)
1236 .addReg(Tmp2).addImm(Idx1);
1241 Tmp1 = SelectCC(N, Opc, Inv, Idx);
1248 /// Check to see if the load is a constant offset from a base register
1249 unsigned ISel::SelectAddr(SDOperand N, unsigned& Reg, int& offset)
1251 unsigned imm = 0, opcode = N.getOpcode();
1252 if (N.getOpcode() == ISD::ADD) {
1253 bool isFrame = N.getOperand(0).getOpcode() == ISD::FrameIndex;
1254 if (isImmediate(N.getOperand(1), imm) && isInt16(imm)) {
1258 Reg = cast<FrameIndexSDNode>(N.getOperand(0))->getIndex();
1261 Reg = SelectExpr(N.getOperand(0));
1265 Reg = SelectExpr(N.getOperand(0));
1266 offset = SelectExpr(N.getOperand(1));
1270 Reg = SelectExpr(N);
1275 void ISel::SelectBranchCC(SDOperand N)
1277 MachineBasicBlock *Dest =
1278 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
1281 unsigned Opc, CCReg, Idx;
1282 Select(N.getOperand(0)); //chain
1283 CCReg = SelectCC(N.getOperand(1), Opc, Inv, Idx);
1285 // Iterate to the next basic block
1286 ilist<MachineBasicBlock>::iterator It = BB;
1289 // If this is a two way branch, then grab the fallthrough basic block argument
1290 // and build a PowerPC branch pseudo-op, suitable for long branch conversion
1291 // if necessary by the branch selection pass. Otherwise, emit a standard
1292 // conditional branch.
1293 if (N.getOpcode() == ISD::BRCONDTWOWAY) {
1294 MachineBasicBlock *Fallthrough =
1295 cast<BasicBlockSDNode>(N.getOperand(3))->getBasicBlock();
1297 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
1298 .addMBB(Dest).addMBB(Fallthrough);
1299 if (Fallthrough != It)
1300 BuildMI(BB, PPC::B, 1).addMBB(Fallthrough);
1302 if (Fallthrough != It) {
1303 Opc = PPC32InstrInfo::invertPPCBranchOpcode(Opc);
1304 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
1305 .addMBB(Fallthrough).addMBB(Dest);
1309 // If the fallthrough path is off the end of the function, which would be
1310 // undefined behavior, set it to be the same as the current block because
1311 // we have nothing better to set it to, and leaving it alone will cause the
1312 // PowerPC Branch Selection pass to crash.
1313 if (It == BB->getParent()->end()) It = Dest;
1314 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
1315 .addMBB(Dest).addMBB(It);
1320 unsigned ISel::SelectExpr(SDOperand N, bool Recording) {
1322 unsigned Tmp1, Tmp2, Tmp3;
1324 unsigned opcode = N.getOpcode();
1326 SDNode *Node = N.Val;
1327 MVT::ValueType DestType = N.getValueType();
1329 if (Node->getOpcode() == ISD::CopyFromReg &&
1330 (MRegisterInfo::isVirtualRegister(cast<RegSDNode>(Node)->getReg()) ||
1331 cast<RegSDNode>(Node)->getReg() == PPC::R1))
1332 // Just use the specified register as our input.
1333 return cast<RegSDNode>(Node)->getReg();
1335 unsigned &Reg = ExprMap[N];
1336 if (Reg) return Reg;
1338 switch (N.getOpcode()) {
1340 Reg = Result = (N.getValueType() != MVT::Other) ?
1341 MakeReg(N.getValueType()) : 1;
1345 // If this is a call instruction, make sure to prepare ALL of the result
1346 // values as well as the chain.
1347 if (Node->getNumValues() == 1)
1348 Reg = Result = 1; // Void call, just a chain.
1350 Result = MakeReg(Node->getValueType(0));
1351 ExprMap[N.getValue(0)] = Result;
1352 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
1353 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
1354 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
1357 case ISD::ADD_PARTS:
1358 case ISD::SUB_PARTS:
1359 case ISD::SHL_PARTS:
1360 case ISD::SRL_PARTS:
1361 case ISD::SRA_PARTS:
1362 Result = MakeReg(Node->getValueType(0));
1363 ExprMap[N.getValue(0)] = Result;
1364 for (unsigned i = 1, e = N.Val->getNumValues(); i != e; ++i)
1365 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
1372 assert(0 && "Node not handled!\n");
1374 BuildMI(BB, PPC::IMPLICIT_DEF, 0, Result);
1376 case ISD::DYNAMIC_STACKALLOC:
1377 // Generate both result values. FIXME: Need a better commment here?
1379 ExprMap[N.getValue(1)] = 1;
1381 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1383 // FIXME: We are currently ignoring the requested alignment for handling
1384 // greater than the stack alignment. This will need to be revisited at some
1385 // point. Align = N.getOperand(2);
1386 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
1387 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
1388 std::cerr << "Cannot allocate stack object with greater alignment than"
1389 << " the stack alignment yet!";
1392 Select(N.getOperand(0));
1393 Tmp1 = SelectExpr(N.getOperand(1));
1394 // Subtract size from stack pointer, thereby allocating some space.
1395 BuildMI(BB, PPC::SUBF, 2, PPC::R1).addReg(Tmp1).addReg(PPC::R1);
1396 // Put a pointer to the space into the result register by copying the SP
1397 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R1).addReg(PPC::R1);
1400 case ISD::ConstantPool:
1401 Tmp1 = cast<ConstantPoolSDNode>(N)->getIndex();
1402 Tmp2 = MakeReg(MVT::i32);
1404 BuildMI(BB, PPC::ADDIS, 2, Tmp2).addReg(getGlobalBaseReg())
1405 .addConstantPoolIndex(Tmp1);
1407 BuildMI(BB, PPC::LIS, 1, Tmp2).addConstantPoolIndex(Tmp1);
1408 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp2).addConstantPoolIndex(Tmp1);
1411 case ISD::FrameIndex:
1412 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
1413 addFrameReference(BuildMI(BB, PPC::ADDI, 2, Result), (int)Tmp1, 0, false);
1416 case ISD::GlobalAddress: {
1417 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
1418 Tmp1 = MakeReg(MVT::i32);
1420 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
1421 .addGlobalAddress(GV);
1423 BuildMI(BB, PPC::LIS, 1, Tmp1).addGlobalAddress(GV);
1424 if (GV->hasWeakLinkage() || GV->isExternal()) {
1425 BuildMI(BB, PPC::LWZ, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
1427 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp1).addGlobalAddress(GV);
1435 case ISD::SEXTLOAD: {
1436 MVT::ValueType TypeBeingLoaded = (ISD::LOAD == opcode) ?
1437 Node->getValueType(0) : cast<VTSDNode>(Node->getOperand(3))->getVT();
1438 bool sext = (ISD::SEXTLOAD == opcode);
1440 // Make sure we generate both values.
1442 ExprMap[N.getValue(1)] = 1; // Generate the token
1444 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1446 SDOperand Chain = N.getOperand(0);
1447 SDOperand Address = N.getOperand(1);
1450 switch (TypeBeingLoaded) {
1451 default: Node->dump(); assert(0 && "Cannot load this type!");
1452 case MVT::i1: Opc = PPC::LBZ; break;
1453 case MVT::i8: Opc = PPC::LBZ; break;
1454 case MVT::i16: Opc = sext ? PPC::LHA : PPC::LHZ; break;
1455 case MVT::i32: Opc = PPC::LWZ; break;
1456 case MVT::f32: Opc = PPC::LFS; break;
1457 case MVT::f64: Opc = PPC::LFD; break;
1460 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Address)) {
1461 Tmp1 = MakeReg(MVT::i32);
1462 int CPI = CP->getIndex();
1464 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
1465 .addConstantPoolIndex(CPI);
1467 BuildMI(BB, PPC::LIS, 1, Tmp1).addConstantPoolIndex(CPI);
1468 BuildMI(BB, Opc, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
1469 } else if (Address.getOpcode() == ISD::FrameIndex) {
1470 Tmp1 = cast<FrameIndexSDNode>(Address)->getIndex();
1471 addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1);
1472 } else if(GlobalAddressSDNode *GN = dyn_cast<GlobalAddressSDNode>(Address)){
1473 GlobalValue *GV = GN->getGlobal();
1474 Tmp1 = MakeReg(MVT::i32);
1476 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
1477 .addGlobalAddress(GV);
1479 BuildMI(BB, PPC::LIS, 1, Tmp1).addGlobalAddress(GV);
1480 if (GV->hasWeakLinkage() || GV->isExternal()) {
1481 Tmp2 = MakeReg(MVT::i32);
1482 BuildMI(BB, PPC::LWZ, 2, Tmp2).addGlobalAddress(GV).addReg(Tmp1);
1483 BuildMI(BB, Opc, 2, Result).addSImm(0).addReg(Tmp2);
1485 BuildMI(BB, Opc, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
1489 switch(SelectAddr(Address, Tmp1, offset)) {
1490 default: assert(0 && "Unhandled return value from SelectAddr");
1491 case 0: // imm offset, no frame, no index
1492 BuildMI(BB, Opc, 2, Result).addSImm(offset).addReg(Tmp1);
1494 case 1: // imm offset + frame index
1495 addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1, offset);
1497 case 2: // base+index addressing
1498 Opc = IndexedOpForOp(Opc);
1499 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(offset);
1508 unsigned GPR_idx = 0, FPR_idx = 0;
1509 static const unsigned GPR[] = {
1510 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1511 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1513 static const unsigned FPR[] = {
1514 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1515 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1518 // Lower the chain for this call.
1519 Select(N.getOperand(0));
1520 ExprMap[N.getValue(Node->getNumValues()-1)] = 1;
1522 MachineInstr *CallMI;
1523 // Emit the correct call instruction based on the type of symbol called.
1524 if (GlobalAddressSDNode *GASD =
1525 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
1526 CallMI = BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(GASD->getGlobal(),
1528 } else if (ExternalSymbolSDNode *ESSDN =
1529 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
1530 CallMI = BuildMI(PPC::CALLpcrel, 1).addExternalSymbol(ESSDN->getSymbol(),
1533 Tmp1 = SelectExpr(N.getOperand(1));
1534 BuildMI(BB, PPC::OR, 2, PPC::R12).addReg(Tmp1).addReg(Tmp1);
1535 BuildMI(BB, PPC::MTCTR, 1).addReg(PPC::R12);
1536 CallMI = BuildMI(PPC::CALLindirect, 3).addImm(20).addImm(0)
1540 // Load the register args to virtual regs
1541 std::vector<unsigned> ArgVR;
1542 for(int i = 2, e = Node->getNumOperands(); i < e; ++i)
1543 ArgVR.push_back(SelectExpr(N.getOperand(i)));
1545 // Copy the virtual registers into the appropriate argument register
1546 for(int i = 0, e = ArgVR.size(); i < e; ++i) {
1547 switch(N.getOperand(i+2).getValueType()) {
1548 default: Node->dump(); assert(0 && "Unknown value type for call");
1553 assert(GPR_idx < 8 && "Too many int args");
1554 if (N.getOperand(i+2).getOpcode() != ISD::UNDEF) {
1555 BuildMI(BB, PPC::OR,2,GPR[GPR_idx]).addReg(ArgVR[i]).addReg(ArgVR[i]);
1556 CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use);
1562 assert(FPR_idx < 13 && "Too many fp args");
1563 BuildMI(BB, PPC::FMR, 1, FPR[FPR_idx]).addReg(ArgVR[i]);
1564 CallMI->addRegOperand(FPR[FPR_idx], MachineOperand::Use);
1570 // Put the call instruction in the correct place in the MachineBasicBlock
1571 BB->push_back(CallMI);
1573 switch (Node->getValueType(0)) {
1574 default: assert(0 && "Unknown value type for call result!");
1575 case MVT::Other: return 1;
1580 if (Node->getValueType(1) == MVT::i32) {
1581 BuildMI(BB, PPC::OR, 2, Result+1).addReg(PPC::R3).addReg(PPC::R3);
1582 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R4).addReg(PPC::R4);
1584 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R3).addReg(PPC::R3);
1589 BuildMI(BB, PPC::FMR, 1, Result).addReg(PPC::F1);
1592 return Result+N.ResNo;
1595 case ISD::SIGN_EXTEND:
1596 case ISD::SIGN_EXTEND_INREG:
1597 Tmp1 = SelectExpr(N.getOperand(0));
1598 switch(cast<VTSDNode>(Node->getOperand(1))->getVT()) {
1599 default: Node->dump(); assert(0 && "Unhandled SIGN_EXTEND type"); break;
1601 BuildMI(BB, PPC::EXTSH, 1, Result).addReg(Tmp1);
1604 BuildMI(BB, PPC::EXTSB, 1, Result).addReg(Tmp1);
1607 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp1).addSImm(0);
1612 case ISD::CopyFromReg:
1613 DestType = N.getValue(0).getValueType();
1615 Result = ExprMap[N.getValue(0)] = MakeReg(DestType);
1616 Tmp1 = dyn_cast<RegSDNode>(Node)->getReg();
1617 if (MVT::isInteger(DestType))
1618 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1620 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
1624 Tmp1 = SelectExpr(N.getOperand(0));
1625 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1626 Tmp2 = CN->getValue() & 0x1F;
1627 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(Tmp2).addImm(0)
1630 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1631 BuildMI(BB, PPC::SLW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1636 Tmp1 = SelectExpr(N.getOperand(0));
1637 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1638 Tmp2 = CN->getValue() & 0x1F;
1639 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(32-Tmp2)
1640 .addImm(Tmp2).addImm(31);
1642 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1643 BuildMI(BB, PPC::SRW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1648 Tmp1 = SelectExpr(N.getOperand(0));
1649 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1650 Tmp2 = CN->getValue() & 0x1F;
1651 BuildMI(BB, PPC::SRAWI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1653 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1654 BuildMI(BB, PPC::SRAW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1659 Tmp1 = SelectExpr(N.getOperand(0));
1660 BuildMI(BB, PPC::CNTLZW, 1, Result).addReg(Tmp1);
1664 if (!MVT::isInteger(DestType)) {
1665 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
1666 N.getOperand(0).Val->hasOneUse()) {
1667 ++FusedFP; // Statistic
1668 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1669 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1670 Tmp3 = SelectExpr(N.getOperand(1));
1671 Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
1672 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1675 if (!NoExcessFPPrecision && N.getOperand(1).getOpcode() == ISD::MUL &&
1676 N.getOperand(1).Val->hasOneUse()) {
1677 ++FusedFP; // Statistic
1678 Tmp1 = SelectExpr(N.getOperand(1).getOperand(0));
1679 Tmp2 = SelectExpr(N.getOperand(1).getOperand(1));
1680 Tmp3 = SelectExpr(N.getOperand(0));
1681 Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
1682 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1685 Opc = DestType == MVT::f64 ? PPC::FADD : PPC::FADDS;
1686 Tmp1 = SelectExpr(N.getOperand(0));
1687 Tmp2 = SelectExpr(N.getOperand(1));
1688 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1691 Tmp1 = SelectExpr(N.getOperand(0));
1692 if (isImmediate(N.getOperand(1), Tmp2)) {
1696 unsigned Reg = MakeReg(MVT::i32);
1697 BuildMI(BB, PPC::ADDI, 2, Reg).addReg(Tmp1).addSImm(Tmp2);
1698 BuildMI(BB, PPC::ADDIS, 2, Result).addReg(Reg).addSImm(Tmp3);
1700 BuildMI(BB, PPC::ADDI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1702 BuildMI(BB, PPC::ADDIS, 2, Result).addReg(Tmp1).addSImm(Tmp3);
1707 Tmp2 = SelectExpr(N.getOperand(1));
1708 BuildMI(BB, PPC::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1712 if (isImmediate(N.getOperand(1), Tmp2)) {
1713 if (isShiftedMask_32(Tmp2) || isShiftedMask_32(~Tmp2)) {
1714 unsigned SH, MB, ME;
1715 Opc = Recording ? PPC::RLWINMo : PPC::RLWINM;
1717 if (isOprShiftImm(N.getOperand(0), OprOpc, Tmp3) &&
1718 isRotateAndMask(OprOpc, Tmp3, Tmp2, false, SH, MB, ME)) {
1719 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1721 Tmp1 = SelectExpr(N.getOperand(0));
1722 isRunOfOnes(Tmp2, MB, ME);
1725 BuildMI(BB, Opc, 4, Result).addReg(Tmp1).addImm(SH)
1726 .addImm(MB).addImm(ME);
1727 RecordSuccess = true;
1729 } else if (isUInt16(Tmp2)) {
1731 Tmp1 = SelectExpr(N.getOperand(0));
1732 BuildMI(BB, PPC::ANDIo, 2, Result).addReg(Tmp1).addImm(Tmp2);
1733 RecordSuccess = true;
1735 } else if (isUInt16(Tmp2)) {
1737 Tmp1 = SelectExpr(N.getOperand(0));
1738 BuildMI(BB, PPC::ANDISo, 2, Result).addReg(Tmp1).addImm(Tmp2);
1739 RecordSuccess = true;
1743 if (isOprNot(N.getOperand(0))) {
1744 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1745 Tmp2 = SelectExpr(N.getOperand(1));
1746 BuildMI(BB, PPC::ANDC, 2, Result).addReg(Tmp2).addReg(Tmp1);
1747 RecordSuccess = false;
1750 // emit a regular and
1751 Tmp1 = SelectExpr(N.getOperand(0));
1752 Tmp2 = SelectExpr(N.getOperand(1));
1753 Opc = Recording ? PPC::ANDo : PPC::AND;
1754 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1755 RecordSuccess = true;
1759 if (SelectBitfieldInsert(N, Result))
1761 Tmp1 = SelectExpr(N.getOperand(0));
1762 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1763 default: assert(0 && "unhandled result code");
1764 case 0: // No immediate
1765 Tmp2 = SelectExpr(N.getOperand(1));
1766 Opc = Recording ? PPC::ORo : PPC::OR;
1767 RecordSuccess = true;
1768 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1770 case 1: // Low immediate
1771 BuildMI(BB, PPC::ORI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1773 case 2: // Shifted immediate
1774 BuildMI(BB, PPC::ORIS, 2, Result).addReg(Tmp1).addImm(Tmp2);
1780 // Check for EQV: xor, (xor a, -1), b
1781 if (N.getOperand(0).getOpcode() == ISD::XOR &&
1782 N.getOperand(0).getOperand(1).getOpcode() == ISD::Constant &&
1783 cast<ConstantSDNode>(N.getOperand(0).getOperand(1))->isAllOnesValue()) {
1784 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1785 Tmp2 = SelectExpr(N.getOperand(1));
1786 BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1789 // Check for NOT, NOR, EQV, and NAND: xor (copy, or, xor, and), -1
1790 if (N.getOperand(1).getOpcode() == ISD::Constant &&
1791 cast<ConstantSDNode>(N.getOperand(1))->isAllOnesValue()) {
1792 switch(N.getOperand(0).getOpcode()) {
1794 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1795 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1796 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1799 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1800 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1801 BuildMI(BB, PPC::NAND, 2, Result).addReg(Tmp1).addReg(Tmp2);
1804 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1805 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1806 BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1809 Tmp1 = SelectExpr(N.getOperand(0));
1810 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1815 Tmp1 = SelectExpr(N.getOperand(0));
1816 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1817 default: assert(0 && "unhandled result code");
1818 case 0: // No immediate
1819 Tmp2 = SelectExpr(N.getOperand(1));
1820 BuildMI(BB, PPC::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1822 case 1: // Low immediate
1823 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1825 case 2: // Shifted immediate
1826 BuildMI(BB, PPC::XORIS, 2, Result).addReg(Tmp1).addImm(Tmp2);
1833 if (!MVT::isInteger(DestType)) {
1834 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
1835 N.getOperand(0).Val->hasOneUse()) {
1836 ++FusedFP; // Statistic
1837 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1838 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1839 Tmp3 = SelectExpr(N.getOperand(1));
1840 Opc = DestType == MVT::f64 ? PPC::FMSUB : PPC::FMSUBS;
1841 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1844 if (!NoExcessFPPrecision && N.getOperand(1).getOpcode() == ISD::MUL &&
1845 N.getOperand(1).Val->hasOneUse()) {
1846 ++FusedFP; // Statistic
1847 Tmp1 = SelectExpr(N.getOperand(1).getOperand(0));
1848 Tmp2 = SelectExpr(N.getOperand(1).getOperand(1));
1849 Tmp3 = SelectExpr(N.getOperand(0));
1850 Opc = DestType == MVT::f64 ? PPC::FNMSUB : PPC::FNMSUBS;
1851 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1854 Opc = DestType == MVT::f64 ? PPC::FSUB : PPC::FSUBS;
1855 Tmp1 = SelectExpr(N.getOperand(0));
1856 Tmp2 = SelectExpr(N.getOperand(1));
1857 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1860 if (1 == getImmediateForOpcode(N.getOperand(0), opcode, Tmp1, true)) {
1861 Tmp2 = SelectExpr(N.getOperand(1));
1862 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp2).addSImm(Tmp1);
1863 } else if (1 == getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1864 Tmp1 = SelectExpr(N.getOperand(0));
1865 BuildMI(BB, PPC::ADDI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1867 Tmp1 = SelectExpr(N.getOperand(0));
1868 Tmp2 = SelectExpr(N.getOperand(1));
1869 BuildMI(BB, PPC::SUBF, 2, Result).addReg(Tmp2).addReg(Tmp1);
1874 Tmp1 = SelectExpr(N.getOperand(0));
1875 if (1 == getImmediateForOpcode(N.getOperand(1), opcode, Tmp2))
1876 BuildMI(BB, PPC::MULLI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1878 Tmp2 = SelectExpr(N.getOperand(1));
1880 default: assert(0 && "Unknown type to ISD::MUL"); break;
1881 case MVT::i32: Opc = PPC::MULLW; break;
1882 case MVT::f32: Opc = PPC::FMULS; break;
1883 case MVT::f64: Opc = PPC::FMUL; break;
1885 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1891 Tmp1 = SelectExpr(N.getOperand(0));
1892 Tmp2 = SelectExpr(N.getOperand(1));
1893 Opc = (ISD::MULHU == opcode) ? PPC::MULHWU : PPC::MULHW;
1894 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1899 switch (getImmediateForOpcode(N.getOperand(1), opcode, Tmp3)) {
1901 // If this is an sdiv by a power of two, we can use an srawi/addze pair.
1903 Tmp1 = MakeReg(MVT::i32);
1904 Tmp2 = SelectExpr(N.getOperand(0));
1905 if ((int)Tmp3 < 0) {
1906 unsigned Tmp4 = MakeReg(MVT::i32);
1907 BuildMI(BB, PPC::SRAWI, 2, Tmp1).addReg(Tmp2).addImm(-Tmp3);
1908 BuildMI(BB, PPC::ADDZE, 1, Tmp4).addReg(Tmp1);
1909 BuildMI(BB, PPC::NEG, 1, Result).addReg(Tmp4);
1911 BuildMI(BB, PPC::SRAWI, 2, Tmp1).addReg(Tmp2).addImm(Tmp3);
1912 BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp1);
1915 // If this is a divide by constant, we can emit code using some magic
1916 // constants to implement it as a multiply instead.
1919 if (opcode == ISD::SDIV)
1920 return SelectExpr(BuildSDIVSequence(N));
1922 return SelectExpr(BuildUDIVSequence(N));
1924 Tmp1 = SelectExpr(N.getOperand(0));
1925 Tmp2 = SelectExpr(N.getOperand(1));
1927 default: assert(0 && "Unknown type to ISD::SDIV"); break;
1928 case MVT::i32: Opc = (ISD::UDIV == opcode) ? PPC::DIVWU : PPC::DIVW; break;
1929 case MVT::f32: Opc = PPC::FDIVS; break;
1930 case MVT::f64: Opc = PPC::FDIV; break;
1932 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1935 case ISD::ADD_PARTS:
1936 case ISD::SUB_PARTS: {
1937 assert(N.getNumOperands() == 4 && N.getValueType() == MVT::i32 &&
1938 "Not an i64 add/sub!");
1939 // Emit all of the operands.
1940 std::vector<unsigned> InVals;
1941 for (unsigned i = 0, e = N.getNumOperands(); i != e; ++i)
1942 InVals.push_back(SelectExpr(N.getOperand(i)));
1943 if (N.getOpcode() == ISD::ADD_PARTS) {
1944 BuildMI(BB, PPC::ADDC, 2, Result).addReg(InVals[0]).addReg(InVals[2]);
1945 BuildMI(BB, PPC::ADDE, 2, Result+1).addReg(InVals[1]).addReg(InVals[3]);
1947 BuildMI(BB, PPC::SUBFC, 2, Result).addReg(InVals[2]).addReg(InVals[0]);
1948 BuildMI(BB, PPC::SUBFE, 2, Result+1).addReg(InVals[3]).addReg(InVals[1]);
1950 return Result+N.ResNo;
1953 case ISD::SHL_PARTS:
1954 case ISD::SRA_PARTS:
1955 case ISD::SRL_PARTS: {
1956 assert(N.getNumOperands() == 3 && N.getValueType() == MVT::i32 &&
1957 "Not an i64 shift!");
1958 unsigned ShiftOpLo = SelectExpr(N.getOperand(0));
1959 unsigned ShiftOpHi = SelectExpr(N.getOperand(1));
1960 unsigned SHReg = FoldIfWideZeroExtend(N.getOperand(2));
1961 Tmp1 = MakeReg(MVT::i32);
1962 Tmp2 = MakeReg(MVT::i32);
1963 Tmp3 = MakeReg(MVT::i32);
1964 unsigned Tmp4 = MakeReg(MVT::i32);
1965 unsigned Tmp5 = MakeReg(MVT::i32);
1966 unsigned Tmp6 = MakeReg(MVT::i32);
1967 BuildMI(BB, PPC::SUBFIC, 2, Tmp1).addReg(SHReg).addSImm(32);
1968 if (ISD::SHL_PARTS == opcode) {
1969 BuildMI(BB, PPC::SLW, 2, Tmp2).addReg(ShiftOpHi).addReg(SHReg);
1970 BuildMI(BB, PPC::SRW, 2, Tmp3).addReg(ShiftOpLo).addReg(Tmp1);
1971 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1972 BuildMI(BB, PPC::ADDI, 2, Tmp5).addReg(SHReg).addSImm(-32);
1973 BuildMI(BB, PPC::SLW, 2, Tmp6).addReg(ShiftOpLo).addReg(Tmp5);
1974 BuildMI(BB, PPC::OR, 2, Result+1).addReg(Tmp4).addReg(Tmp6);
1975 BuildMI(BB, PPC::SLW, 2, Result).addReg(ShiftOpLo).addReg(SHReg);
1976 } else if (ISD::SRL_PARTS == opcode) {
1977 BuildMI(BB, PPC::SRW, 2, Tmp2).addReg(ShiftOpLo).addReg(SHReg);
1978 BuildMI(BB, PPC::SLW, 2, Tmp3).addReg(ShiftOpHi).addReg(Tmp1);
1979 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1980 BuildMI(BB, PPC::ADDI, 2, Tmp5).addReg(SHReg).addSImm(-32);
1981 BuildMI(BB, PPC::SRW, 2, Tmp6).addReg(ShiftOpHi).addReg(Tmp5);
1982 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp4).addReg(Tmp6);
1983 BuildMI(BB, PPC::SRW, 2, Result+1).addReg(ShiftOpHi).addReg(SHReg);
1985 MachineBasicBlock *TmpMBB = new MachineBasicBlock(BB->getBasicBlock());
1986 MachineBasicBlock *PhiMBB = new MachineBasicBlock(BB->getBasicBlock());
1987 MachineBasicBlock *OldMBB = BB;
1988 MachineFunction *F = BB->getParent();
1989 ilist<MachineBasicBlock>::iterator It = BB; ++It;
1990 F->getBasicBlockList().insert(It, TmpMBB);
1991 F->getBasicBlockList().insert(It, PhiMBB);
1992 BB->addSuccessor(TmpMBB);
1993 BB->addSuccessor(PhiMBB);
1994 BuildMI(BB, PPC::SRW, 2, Tmp2).addReg(ShiftOpLo).addReg(SHReg);
1995 BuildMI(BB, PPC::SLW, 2, Tmp3).addReg(ShiftOpHi).addReg(Tmp1);
1996 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1997 BuildMI(BB, PPC::ADDICo, 2, Tmp5).addReg(SHReg).addSImm(-32);
1998 BuildMI(BB, PPC::SRAW, 2, Tmp6).addReg(ShiftOpHi).addReg(Tmp5);
1999 BuildMI(BB, PPC::SRAW, 2, Result+1).addReg(ShiftOpHi).addReg(SHReg);
2000 BuildMI(BB, PPC::BLE, 2).addReg(PPC::CR0).addMBB(PhiMBB);
2001 // Select correct least significant half if the shift amount > 32
2003 unsigned Tmp7 = MakeReg(MVT::i32);
2004 BuildMI(BB, PPC::OR, 2, Tmp7).addReg(Tmp6).addReg(Tmp6);
2005 TmpMBB->addSuccessor(PhiMBB);
2007 BuildMI(BB, PPC::PHI, 4, Result).addReg(Tmp4).addMBB(OldMBB)
2008 .addReg(Tmp7).addMBB(TmpMBB);
2010 return Result+N.ResNo;
2013 case ISD::FP_TO_UINT:
2014 case ISD::FP_TO_SINT: {
2015 bool U = (ISD::FP_TO_UINT == opcode);
2016 Tmp1 = SelectExpr(N.getOperand(0));
2018 Tmp2 = MakeReg(MVT::f64);
2019 BuildMI(BB, PPC::FCTIWZ, 1, Tmp2).addReg(Tmp1);
2020 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
2021 addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(Tmp2), FrameIdx);
2022 addFrameReference(BuildMI(BB, PPC::LWZ, 2, Result), FrameIdx, 4);
2025 unsigned Zero = getConstDouble(0.0);
2026 unsigned MaxInt = getConstDouble((1LL << 32) - 1);
2027 unsigned Border = getConstDouble(1LL << 31);
2028 unsigned UseZero = MakeReg(MVT::f64);
2029 unsigned UseMaxInt = MakeReg(MVT::f64);
2030 unsigned UseChoice = MakeReg(MVT::f64);
2031 unsigned TmpReg = MakeReg(MVT::f64);
2032 unsigned TmpReg2 = MakeReg(MVT::f64);
2033 unsigned ConvReg = MakeReg(MVT::f64);
2034 unsigned IntTmp = MakeReg(MVT::i32);
2035 unsigned XorReg = MakeReg(MVT::i32);
2036 MachineFunction *F = BB->getParent();
2037 int FrameIdx = F->getFrameInfo()->CreateStackObject(8, 8);
2038 // Update machine-CFG edges
2039 MachineBasicBlock *XorMBB = new MachineBasicBlock(BB->getBasicBlock());
2040 MachineBasicBlock *PhiMBB = new MachineBasicBlock(BB->getBasicBlock());
2041 MachineBasicBlock *OldMBB = BB;
2042 ilist<MachineBasicBlock>::iterator It = BB; ++It;
2043 F->getBasicBlockList().insert(It, XorMBB);
2044 F->getBasicBlockList().insert(It, PhiMBB);
2045 BB->addSuccessor(XorMBB);
2046 BB->addSuccessor(PhiMBB);
2047 // Convert from floating point to unsigned 32-bit value
2048 // Use 0 if incoming value is < 0.0
2049 BuildMI(BB, PPC::FSEL, 3, UseZero).addReg(Tmp1).addReg(Tmp1).addReg(Zero);
2050 // Use 2**32 - 1 if incoming value is >= 2**32
2051 BuildMI(BB, PPC::FSUB, 2, UseMaxInt).addReg(MaxInt).addReg(Tmp1);
2052 BuildMI(BB, PPC::FSEL, 3, UseChoice).addReg(UseMaxInt).addReg(UseZero)
2055 BuildMI(BB, PPC::FSUB, 2, TmpReg).addReg(UseChoice).addReg(Border);
2056 // Use difference if >= 2**31
2057 BuildMI(BB, PPC::FCMPU, 2, PPC::CR0).addReg(UseChoice).addReg(Border);
2058 BuildMI(BB, PPC::FSEL, 3, TmpReg2).addReg(TmpReg).addReg(TmpReg)
2060 // Convert to integer
2061 BuildMI(BB, PPC::FCTIWZ, 1, ConvReg).addReg(TmpReg2);
2062 addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(ConvReg), FrameIdx);
2063 addFrameReference(BuildMI(BB, PPC::LWZ, 2, IntTmp), FrameIdx, 4);
2064 BuildMI(BB, PPC::BLT, 2).addReg(PPC::CR0).addMBB(PhiMBB);
2065 BuildMI(BB, PPC::B, 1).addMBB(XorMBB);
2068 // add 2**31 if input was >= 2**31
2070 BuildMI(BB, PPC::XORIS, 2, XorReg).addReg(IntTmp).addImm(0x8000);
2071 XorMBB->addSuccessor(PhiMBB);
2074 // DestReg = phi [ IntTmp, OldMBB ], [ XorReg, XorMBB ]
2076 BuildMI(BB, PPC::PHI, 4, Result).addReg(IntTmp).addMBB(OldMBB)
2077 .addReg(XorReg).addMBB(XorMBB);
2080 assert(0 && "Should never get here");
2085 if (SetCCSDNode *SetCC = dyn_cast<SetCCSDNode>(Node)) {
2086 if (ConstantSDNode *CN =
2087 dyn_cast<ConstantSDNode>(SetCC->getOperand(1).Val)) {
2088 // We can codegen setcc op, imm very efficiently compared to a brcond.
2089 // Check for those cases here.
2091 if (CN->getValue() == 0) {
2092 Tmp1 = SelectExpr(SetCC->getOperand(0));
2093 switch (SetCC->getCondition()) {
2094 default: SetCC->dump(); assert(0 && "Unhandled SetCC condition"); abort();
2096 Tmp2 = MakeReg(MVT::i32);
2097 BuildMI(BB, PPC::CNTLZW, 1, Tmp2).addReg(Tmp1);
2098 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp2).addImm(27)
2099 .addImm(5).addImm(31);
2102 Tmp2 = MakeReg(MVT::i32);
2103 BuildMI(BB, PPC::ADDIC, 2, Tmp2).addReg(Tmp1).addSImm(-1);
2104 BuildMI(BB, PPC::SUBFE, 2, Result).addReg(Tmp2).addReg(Tmp1);
2107 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(1)
2108 .addImm(31).addImm(31);
2111 Tmp2 = MakeReg(MVT::i32);
2112 Tmp3 = MakeReg(MVT::i32);
2113 BuildMI(BB, PPC::NEG, 2, Tmp2).addReg(Tmp1);
2114 BuildMI(BB, PPC::ANDC, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2115 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp3).addImm(1)
2116 .addImm(31).addImm(31);
2122 if (CN->isAllOnesValue()) {
2123 Tmp1 = SelectExpr(SetCC->getOperand(0));
2124 switch (SetCC->getCondition()) {
2125 default: assert(0 && "Unhandled SetCC condition"); abort();
2127 Tmp2 = MakeReg(MVT::i32);
2128 Tmp3 = MakeReg(MVT::i32);
2129 BuildMI(BB, PPC::ADDIC, 2, Tmp2).addReg(Tmp1).addSImm(1);
2130 BuildMI(BB, PPC::LI, 1, Tmp3).addSImm(0);
2131 BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp3);
2134 Tmp2 = MakeReg(MVT::i32);
2135 Tmp3 = MakeReg(MVT::i32);
2136 BuildMI(BB, PPC::NOR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
2137 BuildMI(BB, PPC::ADDIC, 2, Tmp3).addReg(Tmp2).addSImm(-1);
2138 BuildMI(BB, PPC::SUBFE, 2, Result).addReg(Tmp3).addReg(Tmp2);
2141 Tmp2 = MakeReg(MVT::i32);
2142 Tmp3 = MakeReg(MVT::i32);
2143 BuildMI(BB, PPC::ADDI, 2, Tmp2).addReg(Tmp1).addSImm(1);
2144 BuildMI(BB, PPC::AND, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2145 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp3).addImm(1)
2146 .addImm(31).addImm(31);
2149 Tmp2 = MakeReg(MVT::i32);
2150 BuildMI(BB, PPC::RLWINM, 4, Tmp2).addReg(Tmp1).addImm(1)
2151 .addImm(31).addImm(31);
2152 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp2).addImm(1);
2160 unsigned CCReg = SelectCC(N, Opc, Inv, Tmp2);
2161 MoveCRtoGPR(CCReg, Inv, Tmp2, Result);
2164 assert(0 && "Is this legal?");
2168 SetCCSDNode* SetCC = dyn_cast<SetCCSDNode>(N.getOperand(0).Val);
2169 if (SetCC && N.getOperand(0).getOpcode() == ISD::SETCC &&
2170 !MVT::isInteger(SetCC->getOperand(0).getValueType()) &&
2171 !MVT::isInteger(N.getOperand(1).getValueType()) &&
2172 !MVT::isInteger(N.getOperand(2).getValueType()) &&
2173 SetCC->getCondition() != ISD::SETEQ &&
2174 SetCC->getCondition() != ISD::SETNE) {
2175 MVT::ValueType VT = SetCC->getOperand(0).getValueType();
2176 unsigned TV = SelectExpr(N.getOperand(1)); // Use if TRUE
2177 unsigned FV = SelectExpr(N.getOperand(2)); // Use if FALSE
2179 ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(SetCC->getOperand(1));
2180 if (CN && (CN->isExactlyValue(-0.0) || CN->isExactlyValue(0.0))) {
2181 switch(SetCC->getCondition()) {
2182 default: assert(0 && "Invalid FSEL condition"); abort();
2185 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2188 Tmp1 = SelectExpr(SetCC->getOperand(0)); // Val to compare against
2189 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp1).addReg(TV).addReg(FV);
2193 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2196 if (SetCC->getOperand(0).getOpcode() == ISD::FNEG) {
2197 Tmp2 = SelectExpr(SetCC->getOperand(0).getOperand(0));
2200 Tmp1 = SelectExpr(SetCC->getOperand(0)); // Val to compare against
2201 BuildMI(BB, PPC::FNEG, 1, Tmp2).addReg(Tmp1);
2203 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp2).addReg(TV).addReg(FV);
2208 Opc = (MVT::f64 == VT) ? PPC::FSUB : PPC::FSUBS;
2209 Tmp1 = SelectExpr(SetCC->getOperand(0)); // Val to compare against
2210 Tmp2 = SelectExpr(SetCC->getOperand(1));
2212 switch(SetCC->getCondition()) {
2213 default: assert(0 && "Invalid FSEL condition"); abort();
2216 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
2217 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
2221 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
2222 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
2226 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2227 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
2231 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2232 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
2236 assert(0 && "Should never get here");
2241 unsigned TrueValue = SelectExpr(N.getOperand(1)); //Use if TRUE
2242 unsigned FalseValue = SelectExpr(N.getOperand(2)); //Use if FALSE
2243 unsigned CCReg = SelectCC(N.getOperand(0), Opc, Inv, Tmp3);
2245 // Create an iterator with which to insert the MBB for copying the false
2246 // value and the MBB to hold the PHI instruction for this SetCC.
2247 MachineBasicBlock *thisMBB = BB;
2248 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2249 ilist<MachineBasicBlock>::iterator It = BB;
2255 // cmpTY ccX, r1, r2
2257 // fallthrough --> copy0MBB
2258 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
2259 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
2260 BuildMI(BB, Opc, 2).addReg(CCReg).addMBB(sinkMBB);
2261 MachineFunction *F = BB->getParent();
2262 F->getBasicBlockList().insert(It, copy0MBB);
2263 F->getBasicBlockList().insert(It, sinkMBB);
2264 // Update machine-CFG edges
2265 BB->addSuccessor(copy0MBB);
2266 BB->addSuccessor(sinkMBB);
2269 // %FalseValue = ...
2270 // # fallthrough to sinkMBB
2272 // Update machine-CFG edges
2273 BB->addSuccessor(sinkMBB);
2276 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2279 BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue)
2280 .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB);
2285 switch (N.getValueType()) {
2286 default: assert(0 && "Cannot use constants of this type!");
2288 BuildMI(BB, PPC::LI, 1, Result)
2289 .addSImm(!cast<ConstantSDNode>(N)->isNullValue());
2293 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
2294 if (v < 32768 && v >= -32768) {
2295 BuildMI(BB, PPC::LI, 1, Result).addSImm(v);
2297 Tmp1 = MakeReg(MVT::i32);
2298 BuildMI(BB, PPC::LIS, 1, Tmp1).addSImm(v >> 16);
2299 BuildMI(BB, PPC::ORI, 2, Result).addReg(Tmp1).addImm(v & 0xFFFF);
2305 case ISD::ConstantFP: {
2306 ConstantFPSDNode *CN = cast<ConstantFPSDNode>(N);
2307 Result = getConstDouble(CN->getValue(), Result);
2312 if (!NoExcessFPPrecision &&
2313 ISD::ADD == N.getOperand(0).getOpcode() &&
2314 N.getOperand(0).Val->hasOneUse() &&
2315 ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() &&
2316 N.getOperand(0).getOperand(0).Val->hasOneUse()) {
2317 ++FusedFP; // Statistic
2318 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0));
2319 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1));
2320 Tmp3 = SelectExpr(N.getOperand(0).getOperand(1));
2321 Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
2322 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
2323 } else if (!NoExcessFPPrecision &&
2324 ISD::ADD == N.getOperand(0).getOpcode() &&
2325 N.getOperand(0).Val->hasOneUse() &&
2326 ISD::MUL == N.getOperand(0).getOperand(1).getOpcode() &&
2327 N.getOperand(0).getOperand(1).Val->hasOneUse()) {
2328 ++FusedFP; // Statistic
2329 Tmp1 = SelectExpr(N.getOperand(0).getOperand(1).getOperand(0));
2330 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1).getOperand(1));
2331 Tmp3 = SelectExpr(N.getOperand(0).getOperand(0));
2332 Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
2333 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
2334 } else if (ISD::FABS == N.getOperand(0).getOpcode()) {
2335 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
2336 BuildMI(BB, PPC::FNABS, 1, Result).addReg(Tmp1);
2338 Tmp1 = SelectExpr(N.getOperand(0));
2339 BuildMI(BB, PPC::FNEG, 1, Result).addReg(Tmp1);
2344 Tmp1 = SelectExpr(N.getOperand(0));
2345 BuildMI(BB, PPC::FABS, 1, Result).addReg(Tmp1);
2349 Tmp1 = SelectExpr(N.getOperand(0));
2350 Opc = DestType == MVT::f64 ? PPC::FSQRT : PPC::FSQRTS;
2351 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
2355 assert (DestType == MVT::f32 &&
2356 N.getOperand(0).getValueType() == MVT::f64 &&
2357 "only f64 to f32 conversion supported here");
2358 Tmp1 = SelectExpr(N.getOperand(0));
2359 BuildMI(BB, PPC::FRSP, 1, Result).addReg(Tmp1);
2362 case ISD::FP_EXTEND:
2363 assert (DestType == MVT::f64 &&
2364 N.getOperand(0).getValueType() == MVT::f32 &&
2365 "only f32 to f64 conversion supported here");
2366 Tmp1 = SelectExpr(N.getOperand(0));
2367 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
2370 case ISD::UINT_TO_FP:
2371 case ISD::SINT_TO_FP: {
2372 assert (N.getOperand(0).getValueType() == MVT::i32
2373 && "int to float must operate on i32");
2374 bool IsUnsigned = (ISD::UINT_TO_FP == opcode);
2375 Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
2376 Tmp2 = MakeReg(MVT::f64); // temp reg to load the integer value into
2377 Tmp3 = MakeReg(MVT::i32); // temp reg to hold the conversion constant
2379 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
2380 MachineConstantPool *CP = BB->getParent()->getConstantPool();
2383 unsigned ConstF = getConstDouble(0x1.000000p52);
2384 // Store the hi & low halves of the fp value, currently in int regs
2385 BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330);
2386 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx);
2387 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp1), FrameIdx, 4);
2388 addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
2389 // Generate the return value with a subtract
2390 BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF);
2392 unsigned ConstF = getConstDouble(0x1.000008p52);
2393 unsigned TmpL = MakeReg(MVT::i32);
2394 // Store the hi & low halves of the fp value, currently in int regs
2395 BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330);
2396 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx);
2397 BuildMI(BB, PPC::XORIS, 2, TmpL).addReg(Tmp1).addImm(0x8000);
2398 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(TmpL), FrameIdx, 4);
2399 addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
2400 // Generate the return value with a subtract
2401 BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF);
2409 void ISel::Select(SDOperand N) {
2410 unsigned Tmp1, Tmp2, Tmp3, Opc;
2411 unsigned opcode = N.getOpcode();
2413 if (!ExprMap.insert(std::make_pair(N, 1)).second)
2414 return; // Already selected.
2416 SDNode *Node = N.Val;
2418 switch (Node->getOpcode()) {
2420 Node->dump(); std::cerr << "\n";
2421 assert(0 && "Node not handled yet!");
2422 case ISD::EntryToken: return; // Noop
2423 case ISD::TokenFactor:
2424 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
2425 Select(Node->getOperand(i));
2427 case ISD::CALLSEQ_START:
2428 case ISD::CALLSEQ_END:
2429 Select(N.getOperand(0));
2430 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
2431 Opc = N.getOpcode() == ISD::CALLSEQ_START ? PPC::ADJCALLSTACKDOWN :
2432 PPC::ADJCALLSTACKUP;
2433 BuildMI(BB, Opc, 1).addImm(Tmp1);
2436 MachineBasicBlock *Dest =
2437 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
2438 Select(N.getOperand(0));
2439 BuildMI(BB, PPC::B, 1).addMBB(Dest);
2443 case ISD::BRCONDTWOWAY:
2446 case ISD::CopyToReg:
2447 Select(N.getOperand(0));
2448 Tmp1 = SelectExpr(N.getOperand(1));
2449 Tmp2 = cast<RegSDNode>(N)->getReg();
2452 if (N.getOperand(1).getValueType() == MVT::f64 ||
2453 N.getOperand(1).getValueType() == MVT::f32)
2454 BuildMI(BB, PPC::FMR, 1, Tmp2).addReg(Tmp1);
2456 BuildMI(BB, PPC::OR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
2459 case ISD::ImplicitDef:
2460 Select(N.getOperand(0));
2461 BuildMI(BB, PPC::IMPLICIT_DEF, 0, cast<RegSDNode>(N)->getReg());
2464 switch (N.getNumOperands()) {
2466 assert(0 && "Unknown return instruction!");
2468 assert(N.getOperand(1).getValueType() == MVT::i32 &&
2469 N.getOperand(2).getValueType() == MVT::i32 &&
2470 "Unknown two-register value!");
2471 Select(N.getOperand(0));
2472 Tmp1 = SelectExpr(N.getOperand(1));
2473 Tmp2 = SelectExpr(N.getOperand(2));
2474 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp2).addReg(Tmp2);
2475 BuildMI(BB, PPC::OR, 2, PPC::R4).addReg(Tmp1).addReg(Tmp1);
2478 Select(N.getOperand(0));
2479 Tmp1 = SelectExpr(N.getOperand(1));
2480 switch (N.getOperand(1).getValueType()) {
2482 assert(0 && "Unknown return type!");
2485 BuildMI(BB, PPC::FMR, 1, PPC::F1).addReg(Tmp1);
2488 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp1).addReg(Tmp1);
2492 Select(N.getOperand(0));
2495 BuildMI(BB, PPC::BLR, 0); // Just emit a 'ret' instruction
2497 case ISD::TRUNCSTORE:
2499 SDOperand Chain = N.getOperand(0);
2500 SDOperand Value = N.getOperand(1);
2501 SDOperand Address = N.getOperand(2);
2504 Tmp1 = SelectExpr(Value); //value
2506 if (opcode == ISD::STORE) {
2507 switch(Value.getValueType()) {
2508 default: assert(0 && "unknown Type in store");
2509 case MVT::i32: Opc = PPC::STW; break;
2510 case MVT::f64: Opc = PPC::STFD; break;
2511 case MVT::f32: Opc = PPC::STFS; break;
2513 } else { //ISD::TRUNCSTORE
2514 switch(cast<VTSDNode>(Node->getOperand(4))->getVT()) {
2515 default: assert(0 && "unknown Type in store");
2517 case MVT::i8: Opc = PPC::STB; break;
2518 case MVT::i16: Opc = PPC::STH; break;
2522 if(Address.getOpcode() == ISD::FrameIndex) {
2523 Tmp2 = cast<FrameIndexSDNode>(Address)->getIndex();
2524 addFrameReference(BuildMI(BB, Opc, 3).addReg(Tmp1), (int)Tmp2);
2525 } else if(GlobalAddressSDNode *GN = dyn_cast<GlobalAddressSDNode>(Address)){
2526 GlobalValue *GV = GN->getGlobal();
2527 Tmp2 = MakeReg(MVT::i32);
2529 BuildMI(BB, PPC::ADDIS, 2, Tmp2).addReg(getGlobalBaseReg())
2530 .addGlobalAddress(GV);
2532 BuildMI(BB, PPC::LIS, 1, Tmp2).addGlobalAddress(GV);
2533 if (GV->hasWeakLinkage() || GV->isExternal()) {
2534 Tmp3 = MakeReg(MVT::i32);
2535 BuildMI(BB, PPC::LWZ, 2, Tmp3).addGlobalAddress(GV).addReg(Tmp2);
2536 BuildMI(BB, Opc, 3).addReg(Tmp1).addSImm(0).addReg(Tmp3);
2538 BuildMI(BB, Opc, 3).addReg(Tmp1).addGlobalAddress(GV).addReg(Tmp2);
2542 switch(SelectAddr(Address, Tmp2, offset)) {
2543 default: assert(0 && "Unhandled return value from SelectAddr");
2544 case 0: // imm offset, no frame, no index
2545 BuildMI(BB, Opc, 3).addReg(Tmp1).addSImm(offset).addReg(Tmp2);
2547 case 1: // imm offset + frame index
2548 addFrameReference(BuildMI(BB, Opc, 3).addReg(Tmp1), (int)Tmp2, offset);
2550 case 2: // base+index addressing
2551 Opc = IndexedOpForOp(Opc);
2552 BuildMI(BB, Opc, 3).addReg(Tmp1).addReg(Tmp2).addReg(offset);
2562 case ISD::CopyFromReg:
2565 case ISD::DYNAMIC_STACKALLOC:
2570 assert(0 && "Should not be reached!");
2574 /// createPPC32PatternInstructionSelector - This pass converts an LLVM function
2575 /// into a machine code representation using pattern matching and a machine
2576 /// description file.
2578 FunctionPass *llvm::createPPC32ISelPattern(TargetMachine &TM) {
2579 return new ISel(TM);