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 // PowerPC does not have Select
87 setOperationAction(ISD::SELECT, MVT::i32, Expand);
88 setOperationAction(ISD::SELECT, MVT::f32, Expand);
89 setOperationAction(ISD::SELECT, MVT::f64, Expand);
91 setSetCCResultContents(ZeroOrOneSetCCResult);
92 addLegalFPImmediate(+0.0); // Necessary for FSEL
93 addLegalFPImmediate(-0.0); //
95 computeRegisterProperties();
98 /// LowerArguments - This hook must be implemented to indicate how we should
99 /// lower the arguments for the specified function, into the specified DAG.
100 virtual std::vector<SDOperand>
101 LowerArguments(Function &F, SelectionDAG &DAG);
103 /// LowerCallTo - This hook lowers an abstract call to a function into an
105 virtual std::pair<SDOperand, SDOperand>
106 LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, unsigned CC,
107 bool isTailCall, SDOperand Callee, ArgListTy &Args,
110 virtual SDOperand LowerVAStart(SDOperand Chain, SDOperand VAListP,
111 Value *VAListV, SelectionDAG &DAG);
113 virtual std::pair<SDOperand,SDOperand>
114 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
115 const Type *ArgTy, SelectionDAG &DAG);
117 virtual std::pair<SDOperand, SDOperand>
118 LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
124 std::vector<SDOperand>
125 PPC32TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
127 // add beautiful description of PPC stack frame format, or at least some docs
129 MachineFunction &MF = DAG.getMachineFunction();
130 MachineFrameInfo *MFI = MF.getFrameInfo();
131 MachineBasicBlock& BB = MF.front();
132 std::vector<SDOperand> ArgValues;
134 // Due to the rather complicated nature of the PowerPC ABI, rather than a
135 // fixed size array of physical args, for the sake of simplicity let the STL
136 // handle tracking them for us.
137 std::vector<unsigned> argVR, argPR, argOp;
138 unsigned ArgOffset = 24;
139 unsigned GPR_remaining = 8;
140 unsigned FPR_remaining = 13;
141 unsigned GPR_idx = 0, FPR_idx = 0;
142 static const unsigned GPR[] = {
143 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
144 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
146 static const unsigned FPR[] = {
147 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
148 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
151 // Add DAG nodes to load the arguments... On entry to a function on PPC,
152 // the arguments start at offset 24, although they are likely to be passed
154 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
155 SDOperand newroot, argt;
157 bool needsLoad = false;
158 bool ArgLive = !I->use_empty();
159 MVT::ValueType ObjectVT = getValueType(I->getType());
162 default: assert(0 && "Unhandled argument type!");
169 if (GPR_remaining > 0) {
170 MF.addLiveIn(GPR[GPR_idx]);
171 argt = newroot = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32,
173 if (ObjectVT != MVT::i32)
174 argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, newroot);
179 case MVT::i64: ObjSize = 8;
181 if (GPR_remaining > 0) {
182 SDOperand argHi, argLo;
183 MF.addLiveIn(GPR[GPR_idx]);
184 argHi = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot());
185 // If we have two or more remaining argument registers, then both halves
186 // of the i64 can be sourced from there. Otherwise, the lower half will
187 // have to come off the stack. This can happen when an i64 is preceded
188 // by 28 bytes of arguments.
189 if (GPR_remaining > 1) {
190 MF.addLiveIn(GPR[GPR_idx+1]);
191 argLo = DAG.getCopyFromReg(GPR[GPR_idx+1], MVT::i32, argHi);
193 int FI = MFI->CreateFixedObject(4, ArgOffset+4);
194 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
195 argLo = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
196 DAG.getSrcValue(NULL));
198 // Build the outgoing arg thingy
199 argt = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, argLo, argHi);
207 ObjSize = (ObjectVT == MVT::f64) ? 8 : 4;
209 if (FPR_remaining > 0) {
210 MF.addLiveIn(FPR[FPR_idx]);
211 argt = newroot = DAG.getCopyFromReg(FPR[FPR_idx], ObjectVT,
221 // We need to load the argument to a virtual register if we determined above
222 // that we ran out of physical registers of the appropriate type
224 unsigned SubregOffset = 0;
225 if (ObjectVT == MVT::i8 || ObjectVT == MVT::i1) SubregOffset = 3;
226 if (ObjectVT == MVT::i16) SubregOffset = 2;
227 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
228 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
229 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN,
230 DAG.getConstant(SubregOffset, MVT::i32));
231 argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
232 DAG.getSrcValue(NULL));
235 // Every 4 bytes of argument space consumes one of the GPRs available for
237 if (GPR_remaining > 0) {
238 unsigned delta = (GPR_remaining > 1 && ObjSize == 8) ? 2 : 1;
239 GPR_remaining -= delta;
242 ArgOffset += ObjSize;
244 DAG.setRoot(newroot.getValue(1));
246 ArgValues.push_back(argt);
249 // If the function takes variable number of arguments, make a frame index for
250 // the start of the first vararg value... for expansion of llvm.va_start.
252 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
253 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
254 // If this function is vararg, store any remaining integer argument regs
255 // to their spots on the stack so that they may be loaded by deferencing the
256 // result of va_next.
257 std::vector<SDOperand> MemOps;
258 for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) {
259 MF.addLiveIn(GPR[GPR_idx]);
260 SDOperand Val = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot());
261 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
262 Val, FIN, DAG.getSrcValue(NULL));
263 MemOps.push_back(Store);
264 // Increment the address by four for the next argument to store
265 SDOperand PtrOff = DAG.getConstant(4, getPointerTy());
266 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff);
268 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps));
271 // Finally, inform the code generator which regs we return values in.
272 switch (getValueType(F.getReturnType())) {
273 default: assert(0 && "Unknown type!");
274 case MVT::isVoid: break;
279 MF.addLiveOut(PPC::R3);
282 MF.addLiveOut(PPC::R3);
283 MF.addLiveOut(PPC::R4);
287 MF.addLiveOut(PPC::F1);
294 std::pair<SDOperand, SDOperand>
295 PPC32TargetLowering::LowerCallTo(SDOperand Chain,
296 const Type *RetTy, bool isVarArg,
297 unsigned CallingConv, bool isTailCall,
298 SDOperand Callee, ArgListTy &Args,
300 // args_to_use will accumulate outgoing args for the ISD::CALL case in
301 // SelectExpr to use to put the arguments in the appropriate registers.
302 std::vector<SDOperand> args_to_use;
304 // Count how many bytes are to be pushed on the stack, including the linkage
305 // area, and parameter passing area.
306 unsigned NumBytes = 24;
309 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
310 DAG.getConstant(NumBytes, getPointerTy()));
312 for (unsigned i = 0, e = Args.size(); i != e; ++i)
313 switch (getValueType(Args[i].second)) {
314 default: assert(0 && "Unknown value type!");
328 // Just to be safe, we'll always reserve the full 24 bytes of linkage area
329 // plus 32 bytes of argument space in case any called code gets funky on us.
330 // (Required by ABI to support var arg)
331 if (NumBytes < 56) NumBytes = 56;
333 // Adjust the stack pointer for the new arguments...
334 // These operations are automatically eliminated by the prolog/epilog pass
335 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
336 DAG.getConstant(NumBytes, getPointerTy()));
338 // Set up a copy of the stack pointer for use loading and storing any
339 // arguments that may not fit in the registers available for argument
341 SDOperand StackPtr = DAG.getCopyFromReg(PPC::R1, MVT::i32,
344 // Figure out which arguments are going to go in registers, and which in
345 // memory. Also, if this is a vararg function, floating point operations
346 // must be stored to our stack, and loaded into integer regs as well, if
347 // any integer regs are available for argument passing.
348 unsigned ArgOffset = 24;
349 unsigned GPR_remaining = 8;
350 unsigned FPR_remaining = 13;
352 std::vector<SDOperand> MemOps;
353 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
354 // PtrOff will be used to store the current argument to the stack if a
355 // register cannot be found for it.
356 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
357 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
358 MVT::ValueType ArgVT = getValueType(Args[i].second);
361 default: assert(0 && "Unexpected ValueType for argument!");
365 // Promote the integer to 32 bits. If the input type is signed use a
366 // sign extend, otherwise use a zero extend.
367 if (Args[i].second->isSigned())
368 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
370 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
373 if (GPR_remaining > 0) {
374 args_to_use.push_back(Args[i].first);
377 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
378 Args[i].first, PtrOff,
379 DAG.getSrcValue(NULL)));
384 // If we have one free GPR left, we can place the upper half of the i64
385 // in it, and store the other half to the stack. If we have two or more
386 // free GPRs, then we can pass both halves of the i64 in registers.
387 if (GPR_remaining > 0) {
388 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
389 Args[i].first, DAG.getConstant(1, MVT::i32));
390 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
391 Args[i].first, DAG.getConstant(0, MVT::i32));
392 args_to_use.push_back(Hi);
394 if (GPR_remaining > 0) {
395 args_to_use.push_back(Lo);
398 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
399 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
400 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
401 Lo, PtrOff, DAG.getSrcValue(NULL)));
404 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
405 Args[i].first, PtrOff,
406 DAG.getSrcValue(NULL)));
412 if (FPR_remaining > 0) {
413 args_to_use.push_back(Args[i].first);
416 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
417 Args[i].first, PtrOff,
418 DAG.getSrcValue(NULL));
419 MemOps.push_back(Store);
420 // Float varargs are always shadowed in available integer registers
421 if (GPR_remaining > 0) {
422 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
423 DAG.getSrcValue(NULL));
424 MemOps.push_back(Load);
425 args_to_use.push_back(Load);
428 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
429 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
430 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
431 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
432 DAG.getSrcValue(NULL));
433 MemOps.push_back(Load);
434 args_to_use.push_back(Load);
438 // If we have any FPRs remaining, we may also have GPRs remaining.
439 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
441 if (GPR_remaining > 0) {
442 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
445 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
446 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
451 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
452 Args[i].first, PtrOff,
453 DAG.getSrcValue(NULL)));
455 ArgOffset += (ArgVT == MVT::f32) ? 4 : 8;
460 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
463 std::vector<MVT::ValueType> RetVals;
464 MVT::ValueType RetTyVT = getValueType(RetTy);
465 if (RetTyVT != MVT::isVoid)
466 RetVals.push_back(RetTyVT);
467 RetVals.push_back(MVT::Other);
469 SDOperand TheCall = SDOperand(DAG.getCall(RetVals,
470 Chain, Callee, args_to_use), 0);
471 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
472 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
473 DAG.getConstant(NumBytes, getPointerTy()));
474 return std::make_pair(TheCall, Chain);
477 SDOperand PPC32TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
478 Value *VAListV, SelectionDAG &DAG) {
479 // vastart just stores the address of the VarArgsFrameIndex slot into the
480 // memory location argument.
481 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
482 return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR, VAListP,
483 DAG.getSrcValue(VAListV));
486 std::pair<SDOperand,SDOperand>
487 PPC32TargetLowering::LowerVAArg(SDOperand Chain,
488 SDOperand VAListP, Value *VAListV,
489 const Type *ArgTy, SelectionDAG &DAG) {
490 MVT::ValueType ArgVT = getValueType(ArgTy);
493 DAG.getLoad(MVT::i32, Chain, VAListP, DAG.getSrcValue(VAListV));
494 SDOperand Result = DAG.getLoad(ArgVT, Chain, VAList, DAG.getSrcValue(NULL));
496 if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
499 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
500 "Other types should have been promoted for varargs!");
503 VAList = DAG.getNode(ISD::ADD, VAList.getValueType(), VAList,
504 DAG.getConstant(Amt, VAList.getValueType()));
505 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
506 VAList, VAListP, DAG.getSrcValue(VAListV));
507 return std::make_pair(Result, Chain);
511 std::pair<SDOperand, SDOperand> PPC32TargetLowering::
512 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
514 assert(0 && "LowerFrameReturnAddress unimplemented");
519 Statistic<>Recorded("ppc-codegen", "Number of recording ops emitted");
520 Statistic<>FusedFP("ppc-codegen", "Number of fused fp operations");
521 Statistic<>FrameOff("ppc-codegen", "Number of frame idx offsets collapsed");
523 //===--------------------------------------------------------------------===//
524 /// ISel - PPC32 specific code to select PPC32 machine instructions for
525 /// SelectionDAG operations.
526 //===--------------------------------------------------------------------===//
527 class ISel : public SelectionDAGISel {
528 PPC32TargetLowering PPC32Lowering;
529 SelectionDAG *ISelDAG; // Hack to support us having a dag->dag transform
530 // for sdiv and udiv until it is put into the future
533 /// ExprMap - As shared expressions are codegen'd, we keep track of which
534 /// vreg the value is produced in, so we only emit one copy of each compiled
536 std::map<SDOperand, unsigned> ExprMap;
538 unsigned GlobalBaseReg;
539 bool GlobalBaseInitialized;
542 ISel(TargetMachine &TM) : SelectionDAGISel(PPC32Lowering), PPC32Lowering(TM),
545 /// runOnFunction - Override this function in order to reset our per-function
547 virtual bool runOnFunction(Function &Fn) {
548 // Make sure we re-emit a set of the global base reg if necessary
549 GlobalBaseInitialized = false;
550 return SelectionDAGISel::runOnFunction(Fn);
553 /// InstructionSelectBasicBlock - This callback is invoked by
554 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
555 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG) {
557 // Codegen the basic block.
559 Select(DAG.getRoot());
561 // Clear state used for selection.
566 // dag -> dag expanders for integer divide by constant
567 SDOperand BuildSDIVSequence(SDOperand N);
568 SDOperand BuildUDIVSequence(SDOperand N);
570 unsigned getGlobalBaseReg();
571 unsigned getConstDouble(double floatVal, unsigned Result);
572 void MoveCRtoGPR(unsigned CCReg, ISD::CondCode CC, unsigned Result);
573 bool SelectBitfieldInsert(SDOperand OR, unsigned Result);
574 unsigned FoldIfWideZeroExtend(SDOperand N);
575 unsigned SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC);
576 bool SelectIntImmediateExpr(SDOperand N, unsigned Result,
577 unsigned OCHi, unsigned OCLo,
578 bool IsArithmetic = false, bool Negate = false);
579 unsigned SelectExpr(SDOperand N, bool Recording=false);
580 void Select(SDOperand N);
582 unsigned SelectAddr(SDOperand N, unsigned& Reg, int& offset);
583 void SelectBranchCC(SDOperand N);
585 virtual const char *getPassName() const {
586 return "PowerPC Pattern Instruction Selection";
590 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
591 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
592 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
593 // not, since all 1s are not contiguous.
594 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
595 if (isShiftedMask_32(Val)) {
596 // look for the first non-zero bit
597 MB = CountLeadingZeros_32(Val);
598 // look for the first zero bit after the run of ones
599 ME = CountLeadingZeros_32((Val - 1) ^ Val);
601 } else if (isShiftedMask_32(Val = ~Val)) { // invert mask
602 // effectively look for the first zero bit
603 ME = CountLeadingZeros_32(Val) - 1;
604 // effectively look for the first one bit after the run of zeros
605 MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
612 // isRotateAndMask - Returns true if Mask and Shift can be folded in to a rotate
613 // and mask opcode and mask operation.
614 static bool isRotateAndMask(unsigned Opcode, unsigned Shift, unsigned Mask,
616 unsigned &SH, unsigned &MB, unsigned &ME) {
617 if (Shift > 31) return false;
618 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
620 if (Opcode == ISD::SHL) { // shift left
621 // apply shift to mask if it comes first
622 if (IsShiftMask) Mask = Mask << Shift;
623 // determine which bits are made indeterminant by shift
624 Indeterminant = ~(0xFFFFFFFFu << Shift);
625 } else if (Opcode == ISD::SRA || Opcode == ISD::SRL) { // shift rights
626 // apply shift to mask if it comes first
627 if (IsShiftMask) Mask = Mask >> Shift;
628 // determine which bits are made indeterminant by shift
629 Indeterminant = ~(0xFFFFFFFFu >> Shift);
630 // adjust for the left rotate
634 // if the mask doesn't intersect any Indeterminant bits
635 if (!(Mask & Indeterminant)) {
637 // make sure the mask is still a mask (wrap arounds may not be)
638 return isRunOfOnes(Mask, MB, ME);
645 // isIntImmediate - This method tests to see if a constant operand.
646 // If so Imm will receive the 32 bit value.
647 static bool isIntImmediate(SDOperand N, unsigned& Imm) {
649 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
651 Imm = (unsigned)CN->getSignExtended();
659 // isOprShiftImm - Returns true if the specified operand is a shift opcode with
660 // a immediate shift count less than 32.
661 static bool isOprShiftImm(SDOperand N, unsigned& Opc, unsigned& SH) {
663 return (Opc == ISD::SHL || Opc == ISD::SRL || Opc == ISD::SRA) &&
664 isIntImmediate(N.getOperand(1), SH) && SH < 32;
667 // isOprNot - Returns true if the specified operand is an xor with immediate -1.
668 static bool isOprNot(SDOperand N) {
670 return N.getOpcode() == ISD::XOR &&
671 isIntImmediate(N.getOperand(1), Imm) && (signed)Imm == -1;
674 // Immediate constant composers.
675 // Lo16 - grabs the lo 16 bits from a 32 bit constant.
676 // Hi16 - grabs the hi 16 bits from a 32 bit constant.
677 // HA16 - computes the hi bits required if the lo bits are add/subtracted in
679 static unsigned Lo16(unsigned x) { return x & 0x0000FFFF; }
680 static unsigned Hi16(unsigned x) { return Lo16(x >> 16); }
681 static unsigned HA16(unsigned x) { return Hi16((signed)x - (signed short)x); }
683 /// NodeHasRecordingVariant - If SelectExpr can always produce code for
684 /// NodeOpcode that also sets CR0 as a side effect, return true. Otherwise,
686 static bool NodeHasRecordingVariant(unsigned NodeOpcode) {
688 default: return false;
695 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
697 static unsigned getBCCForSetCC(ISD::CondCode CC) {
699 default: assert(0 && "Unknown condition!"); abort();
700 case ISD::SETEQ: return PPC::BEQ;
701 case ISD::SETNE: return PPC::BNE;
703 case ISD::SETLT: return PPC::BLT;
705 case ISD::SETLE: return PPC::BLE;
707 case ISD::SETGT: return PPC::BGT;
709 case ISD::SETGE: return PPC::BGE;
714 /// getCROpForOp - Return the condition register opcode (or inverted opcode)
715 /// associated with the SelectionDAG opcode.
716 static unsigned getCROpForSetCC(unsigned Opcode, bool Inv1, bool Inv2) {
718 default: assert(0 && "Unknown opcode!"); abort();
720 if (Inv1 && Inv2) return PPC::CRNOR; // De Morgan's Law
721 if (!Inv1 && !Inv2) return PPC::CRAND;
722 if (Inv1 ^ Inv2) return PPC::CRANDC;
724 if (Inv1 && Inv2) return PPC::CRNAND; // De Morgan's Law
725 if (!Inv1 && !Inv2) return PPC::CROR;
726 if (Inv1 ^ Inv2) return PPC::CRORC;
731 /// getCRIdxForSetCC - Return the index of the condition register field
732 /// associated with the SetCC condition, and whether or not the field is
733 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
734 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool& Inv) {
736 default: assert(0 && "Unknown condition!"); abort();
738 case ISD::SETLT: Inv = false; return 0;
740 case ISD::SETGE: Inv = true; return 0;
742 case ISD::SETGT: Inv = false; return 1;
744 case ISD::SETLE: Inv = true; return 1;
745 case ISD::SETEQ: Inv = false; return 2;
746 case ISD::SETNE: Inv = true; return 2;
751 /// IndexedOpForOp - Return the indexed variant for each of the PowerPC load
752 /// and store immediate instructions.
753 static unsigned IndexedOpForOp(unsigned Opcode) {
755 default: assert(0 && "Unknown opcode!"); abort();
756 case PPC::LBZ: return PPC::LBZX; case PPC::STB: return PPC::STBX;
757 case PPC::LHZ: return PPC::LHZX; case PPC::STH: return PPC::STHX;
758 case PPC::LHA: return PPC::LHAX; case PPC::STW: return PPC::STWX;
759 case PPC::LWZ: return PPC::LWZX; case PPC::STFS: return PPC::STFSX;
760 case PPC::LFS: return PPC::LFSX; case PPC::STFD: return PPC::STFDX;
761 case PPC::LFD: return PPC::LFDX;
766 // Structure used to return the necessary information to codegen an SDIV as
769 int m; // magic number
770 int s; // shift amount
774 unsigned int m; // magic number
775 int a; // add indicator
776 int s; // shift amount
779 /// magic - calculate the magic numbers required to codegen an integer sdiv as
780 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
782 static struct ms magic(int d) {
784 unsigned int ad, anc, delta, q1, r1, q2, r2, t;
785 const unsigned int two31 = 0x80000000U;
789 t = two31 + ((unsigned int)d >> 31);
790 anc = t - 1 - t%ad; // absolute value of nc
791 p = 31; // initialize p
792 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
793 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
794 q2 = two31/ad; // initialize q2 = 2p/abs(d)
795 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
798 q1 = 2*q1; // update q1 = 2p/abs(nc)
799 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
800 if (r1 >= anc) { // must be unsigned comparison
804 q2 = 2*q2; // update q2 = 2p/abs(d)
805 r2 = 2*r2; // update r2 = rem(2p/abs(d))
806 if (r2 >= ad) { // must be unsigned comparison
811 } while (q1 < delta || (q1 == delta && r1 == 0));
814 if (d < 0) mag.m = -mag.m; // resulting magic number
815 mag.s = p - 32; // resulting shift
819 /// magicu - calculate the magic numbers required to codegen an integer udiv as
820 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
821 static struct mu magicu(unsigned d)
824 unsigned int nc, delta, q1, r1, q2, r2;
826 magu.a = 0; // initialize "add" indicator
828 p = 31; // initialize p
829 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
830 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
831 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
832 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
835 if (r1 >= nc - r1 ) {
836 q1 = 2*q1 + 1; // update q1
837 r1 = 2*r1 - nc; // update r1
840 q1 = 2*q1; // update q1
841 r1 = 2*r1; // update r1
843 if (r2 + 1 >= d - r2) {
844 if (q2 >= 0x7FFFFFFF) magu.a = 1;
845 q2 = 2*q2 + 1; // update q2
846 r2 = 2*r2 + 1 - d; // update r2
849 if (q2 >= 0x80000000) magu.a = 1;
850 q2 = 2*q2; // update q2
851 r2 = 2*r2 + 1; // update r2
854 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
855 magu.m = q2 + 1; // resulting magic number
856 magu.s = p - 32; // resulting shift
861 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
862 /// return a DAG expression to select that will generate the same value by
863 /// multiplying by a magic number. See:
864 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
865 SDOperand ISel::BuildSDIVSequence(SDOperand N) {
866 int d = (int)cast<ConstantSDNode>(N.getOperand(1))->getSignExtended();
867 ms magics = magic(d);
868 // Multiply the numerator (operand 0) by the magic value
869 SDOperand Q = ISelDAG->getNode(ISD::MULHS, MVT::i32, N.getOperand(0),
870 ISelDAG->getConstant(magics.m, MVT::i32));
871 // If d > 0 and m < 0, add the numerator
872 if (d > 0 && magics.m < 0)
873 Q = ISelDAG->getNode(ISD::ADD, MVT::i32, Q, N.getOperand(0));
874 // If d < 0 and m > 0, subtract the numerator.
875 if (d < 0 && magics.m > 0)
876 Q = ISelDAG->getNode(ISD::SUB, MVT::i32, Q, N.getOperand(0));
877 // Shift right algebraic if shift value is nonzero
879 Q = ISelDAG->getNode(ISD::SRA, MVT::i32, Q,
880 ISelDAG->getConstant(magics.s, MVT::i32));
881 // Extract the sign bit and add it to the quotient
883 ISelDAG->getNode(ISD::SRL, MVT::i32, Q, ISelDAG->getConstant(31, MVT::i32));
884 return ISelDAG->getNode(ISD::ADD, MVT::i32, Q, T);
887 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
888 /// return a DAG expression to select that will generate the same value by
889 /// multiplying by a magic number. See:
890 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
891 SDOperand ISel::BuildUDIVSequence(SDOperand N) {
893 (unsigned)cast<ConstantSDNode>(N.getOperand(1))->getSignExtended();
894 mu magics = magicu(d);
895 // Multiply the numerator (operand 0) by the magic value
896 SDOperand Q = ISelDAG->getNode(ISD::MULHU, MVT::i32, N.getOperand(0),
897 ISelDAG->getConstant(magics.m, MVT::i32));
899 Q = ISelDAG->getNode(ISD::SRL, MVT::i32, Q,
900 ISelDAG->getConstant(magics.s, MVT::i32));
902 SDOperand NPQ = ISelDAG->getNode(ISD::SUB, MVT::i32, N.getOperand(0), Q);
903 NPQ = ISelDAG->getNode(ISD::SRL, MVT::i32, NPQ,
904 ISelDAG->getConstant(1, MVT::i32));
905 NPQ = ISelDAG->getNode(ISD::ADD, MVT::i32, NPQ, Q);
906 Q = ISelDAG->getNode(ISD::SRL, MVT::i32, NPQ,
907 ISelDAG->getConstant(magics.s-1, MVT::i32));
912 /// getGlobalBaseReg - Output the instructions required to put the
913 /// base address to use for accessing globals into a register.
915 unsigned ISel::getGlobalBaseReg() {
916 if (!GlobalBaseInitialized) {
917 // Insert the set of GlobalBaseReg into the first MBB of the function
918 MachineBasicBlock &FirstMBB = BB->getParent()->front();
919 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
920 GlobalBaseReg = MakeReg(MVT::i32);
921 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
922 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg).addReg(PPC::LR);
923 GlobalBaseInitialized = true;
925 return GlobalBaseReg;
928 /// getConstDouble - Loads a floating point value into a register, via the
929 /// Constant Pool. Optionally takes a register in which to load the value.
930 unsigned ISel::getConstDouble(double doubleVal, unsigned Result=0) {
931 unsigned Tmp1 = MakeReg(MVT::i32);
932 if (0 == Result) Result = MakeReg(MVT::f64);
933 MachineConstantPool *CP = BB->getParent()->getConstantPool();
934 ConstantFP *CFP = ConstantFP::get(Type::DoubleTy, doubleVal);
935 unsigned CPI = CP->getConstantPoolIndex(CFP);
937 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
938 .addConstantPoolIndex(CPI);
940 BuildMI(BB, PPC::LIS, 1, Tmp1).addConstantPoolIndex(CPI);
941 BuildMI(BB, PPC::LFD, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
945 /// MoveCRtoGPR - Move CCReg[Idx] to the least significant bit of Result. If
946 /// Inv is true, then invert the result.
947 void ISel::MoveCRtoGPR(unsigned CCReg, ISD::CondCode CC, unsigned Result){
949 unsigned IntCR = MakeReg(MVT::i32);
950 unsigned Idx = getCRIdxForSetCC(CC, Inv);
951 BuildMI(BB, PPC::MCRF, 1, PPC::CR7).addReg(CCReg);
953 TLI.getTargetMachine().getSubtarget<PPCSubtarget>().isGigaProcessor();
954 BuildMI(BB, GPOpt ? PPC::MFOCRF : PPC::MFCR, 1, IntCR).addReg(PPC::CR7);
956 unsigned Tmp1 = MakeReg(MVT::i32);
957 BuildMI(BB, PPC::RLWINM, 4, Tmp1).addReg(IntCR).addImm(32-(3-Idx))
958 .addImm(31).addImm(31);
959 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp1).addImm(1);
961 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(IntCR).addImm(32-(3-Idx))
962 .addImm(31).addImm(31);
966 /// SelectBitfieldInsert - turn an or of two masked values into
967 /// the rotate left word immediate then mask insert (rlwimi) instruction.
968 /// Returns true on success, false if the caller still needs to select OR.
970 /// Patterns matched:
971 /// 1. or shl, and 5. or and, and
972 /// 2. or and, shl 6. or shl, shr
973 /// 3. or shr, and 7. or shr, shl
975 bool ISel::SelectBitfieldInsert(SDOperand OR, unsigned Result) {
976 bool IsRotate = false;
977 unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, Amount = 0;
979 SDOperand Op0 = OR.getOperand(0);
980 SDOperand Op1 = OR.getOperand(1);
982 unsigned Op0Opc = Op0.getOpcode();
983 unsigned Op1Opc = Op1.getOpcode();
985 // Verify that we have the correct opcodes
986 if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc)
988 if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc)
991 // Generate Mask value for Target
992 if (ConstantSDNode *CN =
993 dyn_cast<ConstantSDNode>(Op0.getOperand(1).Val)) {
995 case ISD::SHL: TgtMask <<= (unsigned)CN->getValue(); break;
996 case ISD::SRL: TgtMask >>= (unsigned)CN->getValue(); break;
997 case ISD::AND: TgtMask &= (unsigned)CN->getValue(); break;
1003 // Generate Mask value for Insert
1004 if (ConstantSDNode *CN =
1005 dyn_cast<ConstantSDNode>(Op1.getOperand(1).Val)) {
1008 Amount = CN->getValue();
1010 if (Op0Opc == ISD::SRL) IsRotate = true;
1013 Amount = CN->getValue();
1016 if (Op0Opc == ISD::SHL) IsRotate = true;
1019 InsMask &= (unsigned)CN->getValue();
1028 // If both of the inputs are ANDs and one of them has a logical shift by
1029 // constant as its input, make that the inserted value so that we can combine
1030 // the shift into the rotate part of the rlwimi instruction
1031 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
1032 if (Op1.getOperand(0).getOpcode() == ISD::SHL ||
1033 Op1.getOperand(0).getOpcode() == ISD::SRL) {
1034 if (ConstantSDNode *CN =
1035 dyn_cast<ConstantSDNode>(Op1.getOperand(0).getOperand(1).Val)) {
1036 Amount = Op1.getOperand(0).getOpcode() == ISD::SHL ?
1037 CN->getValue() : 32 - CN->getValue();
1038 Tmp3 = SelectExpr(Op1.getOperand(0).getOperand(0));
1040 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
1041 Op0.getOperand(0).getOpcode() == ISD::SRL) {
1042 if (ConstantSDNode *CN =
1043 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(1).Val)) {
1044 std::swap(Op0, Op1);
1045 std::swap(TgtMask, InsMask);
1046 Amount = Op1.getOperand(0).getOpcode() == ISD::SHL ?
1047 CN->getValue() : 32 - CN->getValue();
1048 Tmp3 = SelectExpr(Op1.getOperand(0).getOperand(0));
1053 // Verify that the Target mask and Insert mask together form a full word mask
1054 // and that the Insert mask is a run of set bits (which implies both are runs
1055 // of set bits). Given that, Select the arguments and generate the rlwimi
1058 if (((TgtMask & InsMask) == 0) && isRunOfOnes(InsMask, MB, ME)) {
1059 unsigned Tmp1, Tmp2;
1060 bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF;
1061 // Check for rotlwi / rotrwi here, a special case of bitfield insert
1062 // where both bitfield halves are sourced from the same value.
1063 if (IsRotate && fullMask &&
1064 OR.getOperand(0).getOperand(0) == OR.getOperand(1).getOperand(0)) {
1065 Tmp1 = SelectExpr(OR.getOperand(0).getOperand(0));
1066 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(Amount)
1067 .addImm(0).addImm(31);
1070 if (Op0Opc == ISD::AND && fullMask)
1071 Tmp1 = SelectExpr(Op0.getOperand(0));
1073 Tmp1 = SelectExpr(Op0);
1074 Tmp2 = Tmp3 ? Tmp3 : SelectExpr(Op1.getOperand(0));
1075 BuildMI(BB, PPC::RLWIMI, 5, Result).addReg(Tmp1).addReg(Tmp2)
1076 .addImm(Amount).addImm(MB).addImm(ME);
1082 /// FoldIfWideZeroExtend - 32 bit PowerPC implicit masks shift amounts to the
1083 /// low six bits. If the shift amount is an ISD::AND node with a mask that is
1084 /// wider than the implicit mask, then we can get rid of the AND and let the
1085 /// shift do the mask.
1086 unsigned ISel::FoldIfWideZeroExtend(SDOperand N) {
1088 if (N.getOpcode() == ISD::AND &&
1089 isIntImmediate(N.getOperand(1), C) && isRunOfOnes(C, MB, ME) &&
1090 MB <= 26 && ME == 31)
1091 return SelectExpr(N.getOperand(0));
1093 return SelectExpr(N);
1096 unsigned ISel::SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC) {
1097 unsigned Result, Tmp1, Tmp2;
1098 bool AlreadySelected = false;
1099 static const unsigned CompareOpcodes[] =
1100 { PPC::FCMPU, PPC::FCMPU, PPC::CMPW, PPC::CMPLW };
1102 // Allocate a condition register for this expression
1103 Result = RegMap->createVirtualRegister(PPC32::CRRCRegisterClass);
1105 // Use U to determine whether the SETCC immediate range is signed or not.
1106 bool U = ISD::isUnsignedIntSetCC(CC);
1107 if (isIntImmediate(RHS, Tmp2) &&
1108 ((U && isUInt16(Tmp2)) || (!U && isInt16(Tmp2)))) {
1110 // For comparisons against zero, we can implicity set CR0 if a recording
1111 // variant (e.g. 'or.' instead of 'or') of the instruction that defines
1112 // operand zero of the SetCC node is available.
1114 NodeHasRecordingVariant(LHS.getOpcode()) && LHS.Val->hasOneUse()) {
1115 RecordSuccess = false;
1116 Tmp1 = SelectExpr(LHS, true);
1117 if (RecordSuccess) {
1119 BuildMI(BB, PPC::MCRF, 1, Result).addReg(PPC::CR0);
1122 AlreadySelected = true;
1124 // If we could not implicitly set CR0, then emit a compare immediate
1126 if (!AlreadySelected) Tmp1 = SelectExpr(LHS);
1128 BuildMI(BB, PPC::CMPLWI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1130 BuildMI(BB, PPC::CMPWI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1132 bool IsInteger = MVT::isInteger(LHS.getValueType());
1133 unsigned CompareOpc = CompareOpcodes[2 * IsInteger + U];
1134 Tmp1 = SelectExpr(LHS);
1135 Tmp2 = SelectExpr(RHS);
1136 BuildMI(BB, CompareOpc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1141 /// Check to see if the load is a constant offset from a base register.
1142 unsigned ISel::SelectAddr(SDOperand N, unsigned& Reg, int& offset)
1144 unsigned imm = 0, opcode = N.getOpcode();
1145 if (N.getOpcode() == ISD::ADD) {
1146 bool isFrame = N.getOperand(0).getOpcode() == ISD::FrameIndex;
1147 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm)) {
1151 Reg = cast<FrameIndexSDNode>(N.getOperand(0))->getIndex();
1154 Reg = SelectExpr(N.getOperand(0));
1158 Reg = SelectExpr(N.getOperand(0));
1159 offset = SelectExpr(N.getOperand(1));
1163 // Now check if we're dealing with a global, and whether or not we should emit
1164 // an optimized load or store for statics.
1165 if(GlobalAddressSDNode *GN = dyn_cast<GlobalAddressSDNode>(N)) {
1166 GlobalValue *GV = GN->getGlobal();
1167 if (!GV->hasWeakLinkage() && !GV->isExternal()) {
1168 unsigned GlobalHi = MakeReg(MVT::i32);
1170 BuildMI(BB, PPC::ADDIS, 2, GlobalHi).addReg(getGlobalBaseReg())
1171 .addGlobalAddress(GV);
1173 BuildMI(BB, PPC::LIS, 1, GlobalHi).addGlobalAddress(GV);
1179 Reg = SelectExpr(N);
1184 void ISel::SelectBranchCC(SDOperand N)
1186 MachineBasicBlock *Dest =
1187 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
1189 Select(N.getOperand(0)); //chain
1191 // FIXME: Until we have Branch_CC and Branch_Twoway_CC, we're going to have to
1192 // Fake it up by hand by checking to see if op 1 is a SetCC, or a boolean.
1195 SDOperand Cond = N.getOperand(1);
1196 if (Cond.getOpcode() == ISD::SETCC) {
1197 CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
1198 CCReg = SelectCC(Cond.getOperand(0), Cond.getOperand(1), CC);
1201 CCReg = SelectCC(Cond, ISelDAG->getConstant(0, Cond.getValueType()), CC);
1203 unsigned Opc = getBCCForSetCC(CC);
1205 // Iterate to the next basic block
1206 ilist<MachineBasicBlock>::iterator It = BB;
1209 // If this is a two way branch, then grab the fallthrough basic block argument
1210 // and build a PowerPC branch pseudo-op, suitable for long branch conversion
1211 // if necessary by the branch selection pass. Otherwise, emit a standard
1212 // conditional branch.
1213 if (N.getOpcode() == ISD::BRCONDTWOWAY) {
1214 MachineBasicBlock *Fallthrough =
1215 cast<BasicBlockSDNode>(N.getOperand(3))->getBasicBlock();
1217 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
1218 .addMBB(Dest).addMBB(Fallthrough);
1219 if (Fallthrough != It)
1220 BuildMI(BB, PPC::B, 1).addMBB(Fallthrough);
1222 if (Fallthrough != It) {
1223 Opc = PPC32InstrInfo::invertPPCBranchOpcode(Opc);
1224 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
1225 .addMBB(Fallthrough).addMBB(Dest);
1229 // If the fallthrough path is off the end of the function, which would be
1230 // undefined behavior, set it to be the same as the current block because
1231 // we have nothing better to set it to, and leaving it alone will cause the
1232 // PowerPC Branch Selection pass to crash.
1233 if (It == BB->getParent()->end()) It = Dest;
1234 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
1235 .addMBB(Dest).addMBB(It);
1240 // SelectIntImmediateExpr - Choose code for opcodes with immediate value.
1241 bool ISel::SelectIntImmediateExpr(SDOperand N, unsigned Result,
1242 unsigned OCHi, unsigned OCLo,
1243 bool IsArithmetic, bool Negate) {
1245 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1));
1246 // exit if not a constant
1247 if (!CN) return false;
1248 // extract immediate
1249 unsigned C = (unsigned)CN->getSignExtended();
1250 // negate if required (ISD::SUB)
1252 // get the hi and lo portions of constant
1253 unsigned Hi = IsArithmetic ? HA16(C) : Hi16(C);
1254 unsigned Lo = Lo16(C);
1255 // assume no intermediate result from lo instruction (same as final result)
1256 unsigned Tmp = Result;
1257 // check if two instructions are needed
1259 // exit if usage indicates it would be better to load immediate into a
1261 if (CN->use_size() > 2) return false;
1262 // need intermediate result for two instructions
1263 Tmp = MakeReg(MVT::i32);
1265 // get first operand
1266 unsigned Opr0 = SelectExpr(N.getOperand(0));
1267 // is a lo instruction needed
1269 // generate instruction for hi portion
1270 const MachineInstrBuilder &MIBLo = BuildMI(BB, OCLo, 2, Tmp).addReg(Opr0);
1271 if (IsArithmetic) MIBLo.addSImm(Lo); else MIBLo.addImm(Lo);
1272 // need to switch out first operand for hi instruction
1275 // is a ho instruction needed
1277 // generate instruction for hi portion
1278 const MachineInstrBuilder &MIBHi = BuildMI(BB, OCHi, 2, Result).addReg(Opr0);
1279 if (IsArithmetic) MIBHi.addSImm(Hi); else MIBHi.addImm(Hi);
1284 unsigned ISel::SelectExpr(SDOperand N, bool Recording) {
1286 unsigned Tmp1, Tmp2, Tmp3;
1288 unsigned opcode = N.getOpcode();
1290 SDNode *Node = N.Val;
1291 MVT::ValueType DestType = N.getValueType();
1293 if (Node->getOpcode() == ISD::CopyFromReg &&
1294 (MRegisterInfo::isVirtualRegister(cast<RegSDNode>(Node)->getReg()) ||
1295 cast<RegSDNode>(Node)->getReg() == PPC::R1))
1296 // Just use the specified register as our input.
1297 return cast<RegSDNode>(Node)->getReg();
1299 unsigned &Reg = ExprMap[N];
1300 if (Reg) return Reg;
1302 switch (N.getOpcode()) {
1304 Reg = Result = (N.getValueType() != MVT::Other) ?
1305 MakeReg(N.getValueType()) : 1;
1309 // If this is a call instruction, make sure to prepare ALL of the result
1310 // values as well as the chain.
1311 if (Node->getNumValues() == 1)
1312 Reg = Result = 1; // Void call, just a chain.
1314 Result = MakeReg(Node->getValueType(0));
1315 ExprMap[N.getValue(0)] = Result;
1316 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
1317 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
1318 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
1321 case ISD::ADD_PARTS:
1322 case ISD::SUB_PARTS:
1323 case ISD::SHL_PARTS:
1324 case ISD::SRL_PARTS:
1325 case ISD::SRA_PARTS:
1326 Result = MakeReg(Node->getValueType(0));
1327 ExprMap[N.getValue(0)] = Result;
1328 for (unsigned i = 1, e = N.Val->getNumValues(); i != e; ++i)
1329 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
1336 assert(0 && "\nNode not handled!\n");
1338 BuildMI(BB, PPC::IMPLICIT_DEF, 0, Result);
1340 case ISD::DYNAMIC_STACKALLOC:
1341 // Generate both result values. FIXME: Need a better commment here?
1343 ExprMap[N.getValue(1)] = 1;
1345 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1347 // FIXME: We are currently ignoring the requested alignment for handling
1348 // greater than the stack alignment. This will need to be revisited at some
1349 // point. Align = N.getOperand(2);
1350 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
1351 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
1352 std::cerr << "Cannot allocate stack object with greater alignment than"
1353 << " the stack alignment yet!";
1356 Select(N.getOperand(0));
1357 Tmp1 = SelectExpr(N.getOperand(1));
1358 // Subtract size from stack pointer, thereby allocating some space.
1359 BuildMI(BB, PPC::SUBF, 2, PPC::R1).addReg(Tmp1).addReg(PPC::R1);
1360 // Put a pointer to the space into the result register by copying the SP
1361 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R1).addReg(PPC::R1);
1364 case ISD::ConstantPool:
1365 Tmp1 = cast<ConstantPoolSDNode>(N)->getIndex();
1366 Tmp2 = MakeReg(MVT::i32);
1368 BuildMI(BB, PPC::ADDIS, 2, Tmp2).addReg(getGlobalBaseReg())
1369 .addConstantPoolIndex(Tmp1);
1371 BuildMI(BB, PPC::LIS, 1, Tmp2).addConstantPoolIndex(Tmp1);
1372 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp2).addConstantPoolIndex(Tmp1);
1375 case ISD::FrameIndex:
1376 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
1377 addFrameReference(BuildMI(BB, PPC::ADDI, 2, Result), (int)Tmp1, 0, false);
1380 case ISD::GlobalAddress: {
1381 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
1382 Tmp1 = MakeReg(MVT::i32);
1384 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
1385 .addGlobalAddress(GV);
1387 BuildMI(BB, PPC::LIS, 1, Tmp1).addGlobalAddress(GV);
1388 if (GV->hasWeakLinkage() || GV->isExternal()) {
1389 BuildMI(BB, PPC::LWZ, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
1391 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp1).addGlobalAddress(GV);
1399 case ISD::SEXTLOAD: {
1400 MVT::ValueType TypeBeingLoaded = (ISD::LOAD == opcode) ?
1401 Node->getValueType(0) : cast<VTSDNode>(Node->getOperand(3))->getVT();
1402 bool sext = (ISD::SEXTLOAD == opcode);
1404 // Make sure we generate both values.
1406 ExprMap[N.getValue(1)] = 1; // Generate the token
1408 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1410 SDOperand Chain = N.getOperand(0);
1411 SDOperand Address = N.getOperand(1);
1414 switch (TypeBeingLoaded) {
1415 default: Node->dump(); assert(0 && "Cannot load this type!");
1416 case MVT::i1: Opc = PPC::LBZ; break;
1417 case MVT::i8: Opc = PPC::LBZ; break;
1418 case MVT::i16: Opc = sext ? PPC::LHA : PPC::LHZ; break;
1419 case MVT::i32: Opc = PPC::LWZ; break;
1420 case MVT::f32: Opc = PPC::LFS; break;
1421 case MVT::f64: Opc = PPC::LFD; break;
1424 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Address)) {
1425 Tmp1 = MakeReg(MVT::i32);
1426 int CPI = CP->getIndex();
1428 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
1429 .addConstantPoolIndex(CPI);
1431 BuildMI(BB, PPC::LIS, 1, Tmp1).addConstantPoolIndex(CPI);
1432 BuildMI(BB, Opc, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
1433 } else if (Address.getOpcode() == ISD::FrameIndex) {
1434 Tmp1 = cast<FrameIndexSDNode>(Address)->getIndex();
1435 addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1);
1438 switch(SelectAddr(Address, Tmp1, offset)) {
1439 default: assert(0 && "Unhandled return value from SelectAddr");
1440 case 0: // imm offset, no frame, no index
1441 BuildMI(BB, Opc, 2, Result).addSImm(offset).addReg(Tmp1);
1443 case 1: // imm offset + frame index
1444 addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1, offset);
1446 case 2: // base+index addressing
1447 Opc = IndexedOpForOp(Opc);
1448 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(offset);
1451 GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Address);
1452 GlobalValue *GV = GN->getGlobal();
1453 BuildMI(BB, Opc, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
1462 unsigned GPR_idx = 0, FPR_idx = 0;
1463 static const unsigned GPR[] = {
1464 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1465 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1467 static const unsigned FPR[] = {
1468 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1469 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1472 // Lower the chain for this call.
1473 Select(N.getOperand(0));
1474 ExprMap[N.getValue(Node->getNumValues()-1)] = 1;
1476 MachineInstr *CallMI;
1477 // Emit the correct call instruction based on the type of symbol called.
1478 if (GlobalAddressSDNode *GASD =
1479 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
1480 CallMI = BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(GASD->getGlobal(),
1482 } else if (ExternalSymbolSDNode *ESSDN =
1483 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
1484 CallMI = BuildMI(PPC::CALLpcrel, 1).addExternalSymbol(ESSDN->getSymbol(),
1487 Tmp1 = SelectExpr(N.getOperand(1));
1488 BuildMI(BB, PPC::OR, 2, PPC::R12).addReg(Tmp1).addReg(Tmp1);
1489 BuildMI(BB, PPC::MTCTR, 1).addReg(PPC::R12);
1490 CallMI = BuildMI(PPC::CALLindirect, 3).addImm(20).addImm(0)
1494 // Load the register args to virtual regs
1495 std::vector<unsigned> ArgVR;
1496 for(int i = 2, e = Node->getNumOperands(); i < e; ++i)
1497 ArgVR.push_back(SelectExpr(N.getOperand(i)));
1499 // Copy the virtual registers into the appropriate argument register
1500 for(int i = 0, e = ArgVR.size(); i < e; ++i) {
1501 switch(N.getOperand(i+2).getValueType()) {
1502 default: Node->dump(); assert(0 && "Unknown value type for call");
1507 assert(GPR_idx < 8 && "Too many int args");
1508 if (N.getOperand(i+2).getOpcode() != ISD::UNDEF) {
1509 BuildMI(BB, PPC::OR,2,GPR[GPR_idx]).addReg(ArgVR[i]).addReg(ArgVR[i]);
1510 CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use);
1516 assert(FPR_idx < 13 && "Too many fp args");
1517 BuildMI(BB, PPC::FMR, 1, FPR[FPR_idx]).addReg(ArgVR[i]);
1518 CallMI->addRegOperand(FPR[FPR_idx], MachineOperand::Use);
1524 // Put the call instruction in the correct place in the MachineBasicBlock
1525 BB->push_back(CallMI);
1527 switch (Node->getValueType(0)) {
1528 default: assert(0 && "Unknown value type for call result!");
1529 case MVT::Other: return 1;
1534 if (Node->getValueType(1) == MVT::i32) {
1535 BuildMI(BB, PPC::OR, 2, Result+1).addReg(PPC::R3).addReg(PPC::R3);
1536 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R4).addReg(PPC::R4);
1538 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R3).addReg(PPC::R3);
1543 BuildMI(BB, PPC::FMR, 1, Result).addReg(PPC::F1);
1546 return Result+N.ResNo;
1549 case ISD::SIGN_EXTEND:
1550 case ISD::SIGN_EXTEND_INREG:
1551 Tmp1 = SelectExpr(N.getOperand(0));
1552 switch(cast<VTSDNode>(Node->getOperand(1))->getVT()) {
1553 default: Node->dump(); assert(0 && "Unhandled SIGN_EXTEND type"); break;
1555 BuildMI(BB, PPC::EXTSH, 1, Result).addReg(Tmp1);
1558 BuildMI(BB, PPC::EXTSB, 1, Result).addReg(Tmp1);
1561 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp1).addSImm(0);
1566 case ISD::CopyFromReg:
1567 DestType = N.getValue(0).getValueType();
1569 Result = ExprMap[N.getValue(0)] = MakeReg(DestType);
1570 Tmp1 = dyn_cast<RegSDNode>(Node)->getReg();
1571 if (MVT::isInteger(DestType))
1572 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1574 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
1578 Tmp1 = SelectExpr(N.getOperand(0));
1579 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1580 Tmp2 = CN->getValue() & 0x1F;
1581 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(Tmp2).addImm(0)
1584 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1585 BuildMI(BB, PPC::SLW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1590 Tmp1 = SelectExpr(N.getOperand(0));
1591 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1592 Tmp2 = CN->getValue() & 0x1F;
1593 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(32-Tmp2)
1594 .addImm(Tmp2).addImm(31);
1596 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1597 BuildMI(BB, PPC::SRW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1602 Tmp1 = SelectExpr(N.getOperand(0));
1603 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1604 Tmp2 = CN->getValue() & 0x1F;
1605 BuildMI(BB, PPC::SRAWI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1607 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1608 BuildMI(BB, PPC::SRAW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1613 Tmp1 = SelectExpr(N.getOperand(0));
1614 BuildMI(BB, PPC::CNTLZW, 1, Result).addReg(Tmp1);
1618 if (!MVT::isInteger(DestType)) {
1619 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
1620 N.getOperand(0).Val->hasOneUse()) {
1621 ++FusedFP; // Statistic
1622 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1623 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1624 Tmp3 = SelectExpr(N.getOperand(1));
1625 Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
1626 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1629 if (!NoExcessFPPrecision && N.getOperand(1).getOpcode() == ISD::MUL &&
1630 N.getOperand(1).Val->hasOneUse()) {
1631 ++FusedFP; // Statistic
1632 Tmp1 = SelectExpr(N.getOperand(1).getOperand(0));
1633 Tmp2 = SelectExpr(N.getOperand(1).getOperand(1));
1634 Tmp3 = SelectExpr(N.getOperand(0));
1635 Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
1636 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1639 Opc = DestType == MVT::f64 ? PPC::FADD : PPC::FADDS;
1640 Tmp1 = SelectExpr(N.getOperand(0));
1641 Tmp2 = SelectExpr(N.getOperand(1));
1642 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1645 if (SelectIntImmediateExpr(N, Result, PPC::ADDIS, PPC::ADDI, true))
1647 Tmp1 = SelectExpr(N.getOperand(0));
1648 Tmp2 = SelectExpr(N.getOperand(1));
1649 BuildMI(BB, PPC::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1653 if (isIntImmediate(N.getOperand(1), Tmp2)) {
1654 if (isShiftedMask_32(Tmp2) || isShiftedMask_32(~Tmp2)) {
1655 unsigned SH, MB, ME;
1656 Opc = Recording ? PPC::RLWINMo : PPC::RLWINM;
1658 if (isOprShiftImm(N.getOperand(0), OprOpc, Tmp3) &&
1659 isRotateAndMask(OprOpc, Tmp3, Tmp2, false, SH, MB, ME)) {
1660 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1662 Tmp1 = SelectExpr(N.getOperand(0));
1663 isRunOfOnes(Tmp2, MB, ME);
1666 BuildMI(BB, Opc, 4, Result).addReg(Tmp1).addImm(SH)
1667 .addImm(MB).addImm(ME);
1668 RecordSuccess = true;
1670 } else if (isUInt16(Tmp2)) {
1672 Tmp1 = SelectExpr(N.getOperand(0));
1673 BuildMI(BB, PPC::ANDIo, 2, Result).addReg(Tmp1).addImm(Tmp2);
1674 RecordSuccess = true;
1676 } else if (isUInt16(Tmp2)) {
1678 Tmp1 = SelectExpr(N.getOperand(0));
1679 BuildMI(BB, PPC::ANDISo, 2, Result).addReg(Tmp1).addImm(Tmp2);
1680 RecordSuccess = true;
1684 if (isOprNot(N.getOperand(0))) {
1685 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1686 Tmp2 = SelectExpr(N.getOperand(1));
1687 BuildMI(BB, PPC::ANDC, 2, Result).addReg(Tmp2).addReg(Tmp1);
1688 RecordSuccess = false;
1691 // emit a regular and
1692 Tmp1 = SelectExpr(N.getOperand(0));
1693 Tmp2 = SelectExpr(N.getOperand(1));
1694 Opc = Recording ? PPC::ANDo : PPC::AND;
1695 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1696 RecordSuccess = true;
1700 if (SelectBitfieldInsert(N, Result))
1702 if (SelectIntImmediateExpr(N, Result, PPC::ORIS, PPC::ORI))
1705 Tmp1 = SelectExpr(N.getOperand(0));
1706 Tmp2 = SelectExpr(N.getOperand(1));
1707 Opc = Recording ? PPC::ORo : PPC::OR;
1708 RecordSuccess = true;
1709 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1713 // Check for EQV: xor, (xor a, -1), b
1714 if (isOprNot(N.getOperand(0))) {
1715 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1716 Tmp2 = SelectExpr(N.getOperand(1));
1717 BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1720 // Check for NOT, NOR, EQV, and NAND: xor (copy, or, xor, and), -1
1722 switch(N.getOperand(0).getOpcode()) {
1724 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1725 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1726 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1729 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1730 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1731 BuildMI(BB, PPC::NAND, 2, Result).addReg(Tmp1).addReg(Tmp2);
1734 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1735 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1736 BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1739 Tmp1 = SelectExpr(N.getOperand(0));
1740 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1745 if (SelectIntImmediateExpr(N, Result, PPC::XORIS, PPC::XORI))
1748 Tmp1 = SelectExpr(N.getOperand(0));
1749 Tmp2 = SelectExpr(N.getOperand(1));
1750 BuildMI(BB, PPC::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1755 if (!MVT::isInteger(DestType)) {
1756 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
1757 N.getOperand(0).Val->hasOneUse()) {
1758 ++FusedFP; // Statistic
1759 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1760 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1761 Tmp3 = SelectExpr(N.getOperand(1));
1762 Opc = DestType == MVT::f64 ? PPC::FMSUB : PPC::FMSUBS;
1763 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1766 if (!NoExcessFPPrecision && N.getOperand(1).getOpcode() == ISD::MUL &&
1767 N.getOperand(1).Val->hasOneUse()) {
1768 ++FusedFP; // Statistic
1769 Tmp1 = SelectExpr(N.getOperand(1).getOperand(0));
1770 Tmp2 = SelectExpr(N.getOperand(1).getOperand(1));
1771 Tmp3 = SelectExpr(N.getOperand(0));
1772 Opc = DestType == MVT::f64 ? PPC::FNMSUB : PPC::FNMSUBS;
1773 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1776 Opc = DestType == MVT::f64 ? PPC::FSUB : PPC::FSUBS;
1777 Tmp1 = SelectExpr(N.getOperand(0));
1778 Tmp2 = SelectExpr(N.getOperand(1));
1779 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1782 if (isIntImmediate(N.getOperand(0), Tmp1) && isInt16(Tmp1)) {
1784 Tmp2 = SelectExpr(N.getOperand(1));
1785 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp2).addSImm(Tmp1);
1788 if (SelectIntImmediateExpr(N, Result, PPC::ADDIS, PPC::ADDI, true, true))
1790 Tmp1 = SelectExpr(N.getOperand(0));
1791 Tmp2 = SelectExpr(N.getOperand(1));
1792 BuildMI(BB, PPC::SUBF, 2, Result).addReg(Tmp2).addReg(Tmp1);
1796 Tmp1 = SelectExpr(N.getOperand(0));
1797 if (isIntImmediate(N.getOperand(1), Tmp2) && isInt16(Tmp2)) {
1799 BuildMI(BB, PPC::MULLI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1801 Tmp2 = SelectExpr(N.getOperand(1));
1803 default: assert(0 && "Unknown type to ISD::MUL"); break;
1804 case MVT::i32: Opc = PPC::MULLW; break;
1805 case MVT::f32: Opc = PPC::FMULS; break;
1806 case MVT::f64: Opc = PPC::FMUL; break;
1808 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1814 Tmp1 = SelectExpr(N.getOperand(0));
1815 Tmp2 = SelectExpr(N.getOperand(1));
1816 Opc = (ISD::MULHU == opcode) ? PPC::MULHWU : PPC::MULHW;
1817 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1821 if (isIntImmediate(N.getOperand(1), Tmp3)) {
1822 if ((signed)Tmp3 > 0 && isPowerOf2_32(Tmp3)) {
1823 Tmp3 = Log2_32(Tmp3);
1824 Tmp1 = MakeReg(MVT::i32);
1825 Tmp2 = SelectExpr(N.getOperand(0));
1826 BuildMI(BB, PPC::SRAWI, 2, Tmp1).addReg(Tmp2).addImm(Tmp3);
1827 BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp1);
1829 } else if ((signed)Tmp3 < 0 && isPowerOf2_32(-Tmp3)) {
1830 Tmp3 = Log2_32(-Tmp3);
1831 Tmp2 = SelectExpr(N.getOperand(0));
1832 Tmp1 = MakeReg(MVT::i32);
1833 unsigned Tmp4 = MakeReg(MVT::i32);
1834 BuildMI(BB, PPC::SRAWI, 2, Tmp1).addReg(Tmp2).addImm(Tmp3);
1835 BuildMI(BB, PPC::ADDZE, 1, Tmp4).addReg(Tmp1);
1836 BuildMI(BB, PPC::NEG, 1, Result).addReg(Tmp4);
1842 // If this is a divide by constant, we can emit code using some magic
1843 // constants to implement it as a multiply instead.
1844 if (isIntImmediate(N.getOperand(1), Tmp3)) {
1845 if (opcode == ISD::SDIV) {
1846 if ((signed)Tmp3 < -1 || (signed)Tmp3 > 1) {
1848 return SelectExpr(BuildSDIVSequence(N));
1851 if ((signed)Tmp3 > 1) {
1853 return SelectExpr(BuildUDIVSequence(N));
1857 Tmp1 = SelectExpr(N.getOperand(0));
1858 Tmp2 = SelectExpr(N.getOperand(1));
1860 default: assert(0 && "Unknown type to ISD::SDIV"); break;
1861 case MVT::i32: Opc = (ISD::UDIV == opcode) ? PPC::DIVWU : PPC::DIVW; break;
1862 case MVT::f32: Opc = PPC::FDIVS; break;
1863 case MVT::f64: Opc = PPC::FDIV; break;
1865 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1868 case ISD::ADD_PARTS:
1869 case ISD::SUB_PARTS: {
1870 assert(N.getNumOperands() == 4 && N.getValueType() == MVT::i32 &&
1871 "Not an i64 add/sub!");
1872 // Emit all of the operands.
1873 std::vector<unsigned> InVals;
1874 for (unsigned i = 0, e = N.getNumOperands(); i != e; ++i)
1875 InVals.push_back(SelectExpr(N.getOperand(i)));
1876 if (N.getOpcode() == ISD::ADD_PARTS) {
1877 BuildMI(BB, PPC::ADDC, 2, Result).addReg(InVals[0]).addReg(InVals[2]);
1878 BuildMI(BB, PPC::ADDE, 2, Result+1).addReg(InVals[1]).addReg(InVals[3]);
1880 BuildMI(BB, PPC::SUBFC, 2, Result).addReg(InVals[2]).addReg(InVals[0]);
1881 BuildMI(BB, PPC::SUBFE, 2, Result+1).addReg(InVals[3]).addReg(InVals[1]);
1883 return Result+N.ResNo;
1886 case ISD::SHL_PARTS:
1887 case ISD::SRA_PARTS:
1888 case ISD::SRL_PARTS: {
1889 assert(N.getNumOperands() == 3 && N.getValueType() == MVT::i32 &&
1890 "Not an i64 shift!");
1891 unsigned ShiftOpLo = SelectExpr(N.getOperand(0));
1892 unsigned ShiftOpHi = SelectExpr(N.getOperand(1));
1893 unsigned SHReg = FoldIfWideZeroExtend(N.getOperand(2));
1894 Tmp1 = MakeReg(MVT::i32);
1895 Tmp2 = MakeReg(MVT::i32);
1896 Tmp3 = MakeReg(MVT::i32);
1897 unsigned Tmp4 = MakeReg(MVT::i32);
1898 unsigned Tmp5 = MakeReg(MVT::i32);
1899 unsigned Tmp6 = MakeReg(MVT::i32);
1900 BuildMI(BB, PPC::SUBFIC, 2, Tmp1).addReg(SHReg).addSImm(32);
1901 if (ISD::SHL_PARTS == opcode) {
1902 BuildMI(BB, PPC::SLW, 2, Tmp2).addReg(ShiftOpHi).addReg(SHReg);
1903 BuildMI(BB, PPC::SRW, 2, Tmp3).addReg(ShiftOpLo).addReg(Tmp1);
1904 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1905 BuildMI(BB, PPC::ADDI, 2, Tmp5).addReg(SHReg).addSImm(-32);
1906 BuildMI(BB, PPC::SLW, 2, Tmp6).addReg(ShiftOpLo).addReg(Tmp5);
1907 BuildMI(BB, PPC::OR, 2, Result+1).addReg(Tmp4).addReg(Tmp6);
1908 BuildMI(BB, PPC::SLW, 2, Result).addReg(ShiftOpLo).addReg(SHReg);
1909 } else if (ISD::SRL_PARTS == opcode) {
1910 BuildMI(BB, PPC::SRW, 2, Tmp2).addReg(ShiftOpLo).addReg(SHReg);
1911 BuildMI(BB, PPC::SLW, 2, Tmp3).addReg(ShiftOpHi).addReg(Tmp1);
1912 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1913 BuildMI(BB, PPC::ADDI, 2, Tmp5).addReg(SHReg).addSImm(-32);
1914 BuildMI(BB, PPC::SRW, 2, Tmp6).addReg(ShiftOpHi).addReg(Tmp5);
1915 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp4).addReg(Tmp6);
1916 BuildMI(BB, PPC::SRW, 2, Result+1).addReg(ShiftOpHi).addReg(SHReg);
1918 MachineBasicBlock *TmpMBB = new MachineBasicBlock(BB->getBasicBlock());
1919 MachineBasicBlock *PhiMBB = new MachineBasicBlock(BB->getBasicBlock());
1920 MachineBasicBlock *OldMBB = BB;
1921 MachineFunction *F = BB->getParent();
1922 ilist<MachineBasicBlock>::iterator It = BB; ++It;
1923 F->getBasicBlockList().insert(It, TmpMBB);
1924 F->getBasicBlockList().insert(It, PhiMBB);
1925 BB->addSuccessor(TmpMBB);
1926 BB->addSuccessor(PhiMBB);
1927 BuildMI(BB, PPC::SRW, 2, Tmp2).addReg(ShiftOpLo).addReg(SHReg);
1928 BuildMI(BB, PPC::SLW, 2, Tmp3).addReg(ShiftOpHi).addReg(Tmp1);
1929 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1930 BuildMI(BB, PPC::ADDICo, 2, Tmp5).addReg(SHReg).addSImm(-32);
1931 BuildMI(BB, PPC::SRAW, 2, Tmp6).addReg(ShiftOpHi).addReg(Tmp5);
1932 BuildMI(BB, PPC::SRAW, 2, Result+1).addReg(ShiftOpHi).addReg(SHReg);
1933 BuildMI(BB, PPC::BLE, 2).addReg(PPC::CR0).addMBB(PhiMBB);
1934 // Select correct least significant half if the shift amount > 32
1936 unsigned Tmp7 = MakeReg(MVT::i32);
1937 BuildMI(BB, PPC::OR, 2, Tmp7).addReg(Tmp6).addReg(Tmp6);
1938 TmpMBB->addSuccessor(PhiMBB);
1940 BuildMI(BB, PPC::PHI, 4, Result).addReg(Tmp4).addMBB(OldMBB)
1941 .addReg(Tmp7).addMBB(TmpMBB);
1943 return Result+N.ResNo;
1946 case ISD::FP_TO_UINT:
1947 case ISD::FP_TO_SINT: {
1948 bool U = (ISD::FP_TO_UINT == opcode);
1949 Tmp1 = SelectExpr(N.getOperand(0));
1951 Tmp2 = MakeReg(MVT::f64);
1952 BuildMI(BB, PPC::FCTIWZ, 1, Tmp2).addReg(Tmp1);
1953 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
1954 addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(Tmp2), FrameIdx);
1955 addFrameReference(BuildMI(BB, PPC::LWZ, 2, Result), FrameIdx, 4);
1958 unsigned Zero = getConstDouble(0.0);
1959 unsigned MaxInt = getConstDouble((1LL << 32) - 1);
1960 unsigned Border = getConstDouble(1LL << 31);
1961 unsigned UseZero = MakeReg(MVT::f64);
1962 unsigned UseMaxInt = MakeReg(MVT::f64);
1963 unsigned UseChoice = MakeReg(MVT::f64);
1964 unsigned TmpReg = MakeReg(MVT::f64);
1965 unsigned TmpReg2 = MakeReg(MVT::f64);
1966 unsigned ConvReg = MakeReg(MVT::f64);
1967 unsigned IntTmp = MakeReg(MVT::i32);
1968 unsigned XorReg = MakeReg(MVT::i32);
1969 MachineFunction *F = BB->getParent();
1970 int FrameIdx = F->getFrameInfo()->CreateStackObject(8, 8);
1971 // Update machine-CFG edges
1972 MachineBasicBlock *XorMBB = new MachineBasicBlock(BB->getBasicBlock());
1973 MachineBasicBlock *PhiMBB = new MachineBasicBlock(BB->getBasicBlock());
1974 MachineBasicBlock *OldMBB = BB;
1975 ilist<MachineBasicBlock>::iterator It = BB; ++It;
1976 F->getBasicBlockList().insert(It, XorMBB);
1977 F->getBasicBlockList().insert(It, PhiMBB);
1978 BB->addSuccessor(XorMBB);
1979 BB->addSuccessor(PhiMBB);
1980 // Convert from floating point to unsigned 32-bit value
1981 // Use 0 if incoming value is < 0.0
1982 BuildMI(BB, PPC::FSEL, 3, UseZero).addReg(Tmp1).addReg(Tmp1).addReg(Zero);
1983 // Use 2**32 - 1 if incoming value is >= 2**32
1984 BuildMI(BB, PPC::FSUB, 2, UseMaxInt).addReg(MaxInt).addReg(Tmp1);
1985 BuildMI(BB, PPC::FSEL, 3, UseChoice).addReg(UseMaxInt).addReg(UseZero)
1988 BuildMI(BB, PPC::FSUB, 2, TmpReg).addReg(UseChoice).addReg(Border);
1989 // Use difference if >= 2**31
1990 BuildMI(BB, PPC::FCMPU, 2, PPC::CR0).addReg(UseChoice).addReg(Border);
1991 BuildMI(BB, PPC::FSEL, 3, TmpReg2).addReg(TmpReg).addReg(TmpReg)
1993 // Convert to integer
1994 BuildMI(BB, PPC::FCTIWZ, 1, ConvReg).addReg(TmpReg2);
1995 addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(ConvReg), FrameIdx);
1996 addFrameReference(BuildMI(BB, PPC::LWZ, 2, IntTmp), FrameIdx, 4);
1997 BuildMI(BB, PPC::BLT, 2).addReg(PPC::CR0).addMBB(PhiMBB);
1998 BuildMI(BB, PPC::B, 1).addMBB(XorMBB);
2001 // add 2**31 if input was >= 2**31
2003 BuildMI(BB, PPC::XORIS, 2, XorReg).addReg(IntTmp).addImm(0x8000);
2004 XorMBB->addSuccessor(PhiMBB);
2007 // DestReg = phi [ IntTmp, OldMBB ], [ XorReg, XorMBB ]
2009 BuildMI(BB, PPC::PHI, 4, Result).addReg(IntTmp).addMBB(OldMBB)
2010 .addReg(XorReg).addMBB(XorMBB);
2013 assert(0 && "Should never get here");
2018 ISD::CondCode CC = cast<CondCodeSDNode>(Node->getOperand(2))->get();
2019 if (isIntImmediate(Node->getOperand(1), Tmp3)) {
2020 // We can codegen setcc op, imm very efficiently compared to a brcond.
2021 // Check for those cases here.
2024 Tmp1 = SelectExpr(Node->getOperand(0));
2026 default: Node->dump(); assert(0 && "Unhandled SetCC condition"); abort();
2028 Tmp2 = MakeReg(MVT::i32);
2029 BuildMI(BB, PPC::CNTLZW, 1, Tmp2).addReg(Tmp1);
2030 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp2).addImm(27)
2031 .addImm(5).addImm(31);
2034 Tmp2 = MakeReg(MVT::i32);
2035 BuildMI(BB, PPC::ADDIC, 2, Tmp2).addReg(Tmp1).addSImm(-1);
2036 BuildMI(BB, PPC::SUBFE, 2, Result).addReg(Tmp2).addReg(Tmp1);
2039 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(1)
2040 .addImm(31).addImm(31);
2043 Tmp2 = MakeReg(MVT::i32);
2044 Tmp3 = MakeReg(MVT::i32);
2045 BuildMI(BB, PPC::NEG, 2, Tmp2).addReg(Tmp1);
2046 BuildMI(BB, PPC::ANDC, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2047 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp3).addImm(1)
2048 .addImm(31).addImm(31);
2052 } else if (Tmp3 == ~0U) { // setcc op, -1
2053 Tmp1 = SelectExpr(Node->getOperand(0));
2055 default: assert(0 && "Unhandled SetCC condition"); abort();
2057 Tmp2 = MakeReg(MVT::i32);
2058 Tmp3 = MakeReg(MVT::i32);
2059 BuildMI(BB, PPC::ADDIC, 2, Tmp2).addReg(Tmp1).addSImm(1);
2060 BuildMI(BB, PPC::LI, 1, Tmp3).addSImm(0);
2061 BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp3);
2064 Tmp2 = MakeReg(MVT::i32);
2065 Tmp3 = MakeReg(MVT::i32);
2066 BuildMI(BB, PPC::NOR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
2067 BuildMI(BB, PPC::ADDIC, 2, Tmp3).addReg(Tmp2).addSImm(-1);
2068 BuildMI(BB, PPC::SUBFE, 2, Result).addReg(Tmp3).addReg(Tmp2);
2071 Tmp2 = MakeReg(MVT::i32);
2072 Tmp3 = MakeReg(MVT::i32);
2073 BuildMI(BB, PPC::ADDI, 2, Tmp2).addReg(Tmp1).addSImm(1);
2074 BuildMI(BB, PPC::AND, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2075 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp3).addImm(1)
2076 .addImm(31).addImm(31);
2079 Tmp2 = MakeReg(MVT::i32);
2080 BuildMI(BB, PPC::RLWINM, 4, Tmp2).addReg(Tmp1).addImm(1)
2081 .addImm(31).addImm(31);
2082 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp2).addImm(1);
2089 unsigned CCReg = SelectCC(N.getOperand(0), N.getOperand(1), CC);
2090 MoveCRtoGPR(CCReg, CC, Result);
2094 case ISD::SELECT_CC: {
2095 ISD::CondCode CC = cast<CondCodeSDNode>(N.getOperand(4))->get();
2096 if (!MVT::isInteger(N.getOperand(0).getValueType()) &&
2097 !MVT::isInteger(N.getOperand(2).getValueType()) &&
2098 CC != ISD::SETEQ && CC != ISD::SETNE) {
2099 MVT::ValueType VT = N.getOperand(0).getValueType();
2100 unsigned TV = SelectExpr(N.getOperand(2)); // Use if TRUE
2101 unsigned FV = SelectExpr(N.getOperand(3)); // Use if FALSE
2103 ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N.getOperand(1));
2104 if (CN && (CN->isExactlyValue(-0.0) || CN->isExactlyValue(0.0))) {
2106 default: assert(0 && "Invalid FSEL condition"); abort();
2109 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2112 Tmp1 = SelectExpr(N.getOperand(0)); // Val to compare against
2113 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp1).addReg(TV).addReg(FV);
2117 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2120 if (N.getOperand(0).getOpcode() == ISD::FNEG) {
2121 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0));
2124 Tmp1 = SelectExpr(N.getOperand(0)); // Val to compare against
2125 BuildMI(BB, PPC::FNEG, 1, Tmp2).addReg(Tmp1);
2127 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp2).addReg(TV).addReg(FV);
2132 Opc = (MVT::f64 == VT) ? PPC::FSUB : PPC::FSUBS;
2133 Tmp1 = SelectExpr(N.getOperand(0)); // Val to compare against
2134 Tmp2 = SelectExpr(N.getOperand(1));
2137 default: assert(0 && "Invalid FSEL condition"); abort();
2140 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
2141 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
2145 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
2146 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
2150 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2151 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
2155 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2156 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
2160 assert(0 && "Should never get here");
2163 unsigned TrueValue = SelectExpr(N.getOperand(2)); //Use if TRUE
2164 unsigned FalseValue = SelectExpr(N.getOperand(3)); //Use if FALSE
2165 unsigned CCReg = SelectCC(N.getOperand(0), N.getOperand(1), CC);
2166 Opc = getBCCForSetCC(CC);
2168 // Create an iterator with which to insert the MBB for copying the false
2169 // value and the MBB to hold the PHI instruction for this SetCC.
2170 MachineBasicBlock *thisMBB = BB;
2171 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2172 ilist<MachineBasicBlock>::iterator It = BB;
2178 // cmpTY ccX, r1, r2
2180 // fallthrough --> copy0MBB
2181 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
2182 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
2183 BuildMI(BB, Opc, 2).addReg(CCReg).addMBB(sinkMBB);
2184 MachineFunction *F = BB->getParent();
2185 F->getBasicBlockList().insert(It, copy0MBB);
2186 F->getBasicBlockList().insert(It, sinkMBB);
2187 // Update machine-CFG edges
2188 BB->addSuccessor(copy0MBB);
2189 BB->addSuccessor(sinkMBB);
2192 // %FalseValue = ...
2193 // # fallthrough to sinkMBB
2195 // Update machine-CFG edges
2196 BB->addSuccessor(sinkMBB);
2199 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2202 BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue)
2203 .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB);
2208 switch (N.getValueType()) {
2209 default: assert(0 && "Cannot use constants of this type!");
2211 BuildMI(BB, PPC::LI, 1, Result)
2212 .addSImm(!cast<ConstantSDNode>(N)->isNullValue());
2216 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
2217 if (v < 32768 && v >= -32768) {
2218 BuildMI(BB, PPC::LI, 1, Result).addSImm(v);
2220 Tmp1 = MakeReg(MVT::i32);
2221 BuildMI(BB, PPC::LIS, 1, Tmp1).addSImm(v >> 16);
2222 BuildMI(BB, PPC::ORI, 2, Result).addReg(Tmp1).addImm(v & 0xFFFF);
2228 case ISD::ConstantFP: {
2229 ConstantFPSDNode *CN = cast<ConstantFPSDNode>(N);
2230 Result = getConstDouble(CN->getValue(), Result);
2235 if (!NoExcessFPPrecision &&
2236 ISD::ADD == N.getOperand(0).getOpcode() &&
2237 N.getOperand(0).Val->hasOneUse() &&
2238 ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() &&
2239 N.getOperand(0).getOperand(0).Val->hasOneUse()) {
2240 ++FusedFP; // Statistic
2241 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0));
2242 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1));
2243 Tmp3 = SelectExpr(N.getOperand(0).getOperand(1));
2244 Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
2245 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
2246 } else if (!NoExcessFPPrecision &&
2247 ISD::ADD == N.getOperand(0).getOpcode() &&
2248 N.getOperand(0).Val->hasOneUse() &&
2249 ISD::MUL == N.getOperand(0).getOperand(1).getOpcode() &&
2250 N.getOperand(0).getOperand(1).Val->hasOneUse()) {
2251 ++FusedFP; // Statistic
2252 Tmp1 = SelectExpr(N.getOperand(0).getOperand(1).getOperand(0));
2253 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1).getOperand(1));
2254 Tmp3 = SelectExpr(N.getOperand(0).getOperand(0));
2255 Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
2256 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
2257 } else if (ISD::FABS == N.getOperand(0).getOpcode()) {
2258 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
2259 BuildMI(BB, PPC::FNABS, 1, Result).addReg(Tmp1);
2261 Tmp1 = SelectExpr(N.getOperand(0));
2262 BuildMI(BB, PPC::FNEG, 1, Result).addReg(Tmp1);
2267 Tmp1 = SelectExpr(N.getOperand(0));
2268 BuildMI(BB, PPC::FABS, 1, Result).addReg(Tmp1);
2272 Tmp1 = SelectExpr(N.getOperand(0));
2273 Opc = DestType == MVT::f64 ? PPC::FSQRT : PPC::FSQRTS;
2274 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
2278 assert (DestType == MVT::f32 &&
2279 N.getOperand(0).getValueType() == MVT::f64 &&
2280 "only f64 to f32 conversion supported here");
2281 Tmp1 = SelectExpr(N.getOperand(0));
2282 BuildMI(BB, PPC::FRSP, 1, Result).addReg(Tmp1);
2285 case ISD::FP_EXTEND:
2286 assert (DestType == MVT::f64 &&
2287 N.getOperand(0).getValueType() == MVT::f32 &&
2288 "only f32 to f64 conversion supported here");
2289 Tmp1 = SelectExpr(N.getOperand(0));
2290 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
2293 case ISD::UINT_TO_FP:
2294 case ISD::SINT_TO_FP: {
2295 assert (N.getOperand(0).getValueType() == MVT::i32
2296 && "int to float must operate on i32");
2297 bool IsUnsigned = (ISD::UINT_TO_FP == opcode);
2298 Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
2299 Tmp2 = MakeReg(MVT::f64); // temp reg to load the integer value into
2300 Tmp3 = MakeReg(MVT::i32); // temp reg to hold the conversion constant
2302 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
2303 MachineConstantPool *CP = BB->getParent()->getConstantPool();
2306 unsigned ConstF = getConstDouble(0x1.000000p52);
2307 // Store the hi & low halves of the fp value, currently in int regs
2308 BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330);
2309 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx);
2310 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp1), FrameIdx, 4);
2311 addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
2312 // Generate the return value with a subtract
2313 BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF);
2315 unsigned ConstF = getConstDouble(0x1.000008p52);
2316 unsigned TmpL = MakeReg(MVT::i32);
2317 // Store the hi & low halves of the fp value, currently in int regs
2318 BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330);
2319 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx);
2320 BuildMI(BB, PPC::XORIS, 2, TmpL).addReg(Tmp1).addImm(0x8000);
2321 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(TmpL), FrameIdx, 4);
2322 addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
2323 // Generate the return value with a subtract
2324 BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF);
2332 void ISel::Select(SDOperand N) {
2333 unsigned Tmp1, Tmp2, Tmp3, Opc;
2334 unsigned opcode = N.getOpcode();
2336 if (!ExprMap.insert(std::make_pair(N, 1)).second)
2337 return; // Already selected.
2339 SDNode *Node = N.Val;
2341 switch (Node->getOpcode()) {
2343 Node->dump(); std::cerr << "\n";
2344 assert(0 && "Node not handled yet!");
2345 case ISD::EntryToken: return; // Noop
2346 case ISD::TokenFactor:
2347 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
2348 Select(Node->getOperand(i));
2350 case ISD::CALLSEQ_START:
2351 case ISD::CALLSEQ_END:
2352 Select(N.getOperand(0));
2353 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
2354 Opc = N.getOpcode() == ISD::CALLSEQ_START ? PPC::ADJCALLSTACKDOWN :
2355 PPC::ADJCALLSTACKUP;
2356 BuildMI(BB, Opc, 1).addImm(Tmp1);
2359 MachineBasicBlock *Dest =
2360 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
2361 Select(N.getOperand(0));
2362 BuildMI(BB, PPC::B, 1).addMBB(Dest);
2366 case ISD::BRCONDTWOWAY:
2369 case ISD::CopyToReg:
2370 Select(N.getOperand(0));
2371 Tmp1 = SelectExpr(N.getOperand(1));
2372 Tmp2 = cast<RegSDNode>(N)->getReg();
2375 if (N.getOperand(1).getValueType() == MVT::f64 ||
2376 N.getOperand(1).getValueType() == MVT::f32)
2377 BuildMI(BB, PPC::FMR, 1, Tmp2).addReg(Tmp1);
2379 BuildMI(BB, PPC::OR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
2382 case ISD::ImplicitDef:
2383 Select(N.getOperand(0));
2384 BuildMI(BB, PPC::IMPLICIT_DEF, 0, cast<RegSDNode>(N)->getReg());
2387 switch (N.getNumOperands()) {
2389 assert(0 && "Unknown return instruction!");
2391 assert(N.getOperand(1).getValueType() == MVT::i32 &&
2392 N.getOperand(2).getValueType() == MVT::i32 &&
2393 "Unknown two-register value!");
2394 Select(N.getOperand(0));
2395 Tmp1 = SelectExpr(N.getOperand(1));
2396 Tmp2 = SelectExpr(N.getOperand(2));
2397 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp2).addReg(Tmp2);
2398 BuildMI(BB, PPC::OR, 2, PPC::R4).addReg(Tmp1).addReg(Tmp1);
2401 Select(N.getOperand(0));
2402 Tmp1 = SelectExpr(N.getOperand(1));
2403 switch (N.getOperand(1).getValueType()) {
2405 assert(0 && "Unknown return type!");
2408 BuildMI(BB, PPC::FMR, 1, PPC::F1).addReg(Tmp1);
2411 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp1).addReg(Tmp1);
2415 Select(N.getOperand(0));
2418 BuildMI(BB, PPC::BLR, 0); // Just emit a 'ret' instruction
2420 case ISD::TRUNCSTORE:
2422 SDOperand Chain = N.getOperand(0);
2423 SDOperand Value = N.getOperand(1);
2424 SDOperand Address = N.getOperand(2);
2427 Tmp1 = SelectExpr(Value); //value
2429 if (opcode == ISD::STORE) {
2430 switch(Value.getValueType()) {
2431 default: assert(0 && "unknown Type in store");
2432 case MVT::i32: Opc = PPC::STW; break;
2433 case MVT::f64: Opc = PPC::STFD; break;
2434 case MVT::f32: Opc = PPC::STFS; break;
2436 } else { //ISD::TRUNCSTORE
2437 switch(cast<VTSDNode>(Node->getOperand(4))->getVT()) {
2438 default: assert(0 && "unknown Type in store");
2440 case MVT::i8: Opc = PPC::STB; break;
2441 case MVT::i16: Opc = PPC::STH; break;
2445 if(Address.getOpcode() == ISD::FrameIndex) {
2446 Tmp2 = cast<FrameIndexSDNode>(Address)->getIndex();
2447 addFrameReference(BuildMI(BB, Opc, 3).addReg(Tmp1), (int)Tmp2);
2450 switch(SelectAddr(Address, Tmp2, offset)) {
2451 default: assert(0 && "Unhandled return value from SelectAddr");
2452 case 0: // imm offset, no frame, no index
2453 BuildMI(BB, Opc, 3).addReg(Tmp1).addSImm(offset).addReg(Tmp2);
2455 case 1: // imm offset + frame index
2456 addFrameReference(BuildMI(BB, Opc, 3).addReg(Tmp1), (int)Tmp2, offset);
2458 case 2: // base+index addressing
2459 Opc = IndexedOpForOp(Opc);
2460 BuildMI(BB, Opc, 3).addReg(Tmp1).addReg(Tmp2).addReg(offset);
2463 GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Address);
2464 GlobalValue *GV = GN->getGlobal();
2465 BuildMI(BB, Opc, 3).addReg(Tmp1).addGlobalAddress(GV).addReg(Tmp2);
2475 case ISD::CopyFromReg:
2478 case ISD::DYNAMIC_STACKALLOC:
2483 assert(0 && "Should not be reached!");
2487 /// createPPC32PatternInstructionSelector - This pass converts an LLVM function
2488 /// into a machine code representation using pattern matching and a machine
2489 /// description file.
2491 FunctionPass *llvm::createPPC32ISelPattern(TargetMachine &TM) {
2492 return new ISel(TM);