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 // IsRunOfOnes - returns true if Val consists of one contiguous run of 1's with
40 // any number of 0's on either side. the 1's are allowed to wrap from LSB to
41 // MSB. so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
42 // not, since all 1's are not contiguous.
43 static bool IsRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
44 if (isShiftedMask_32(Val)) {
45 // look for the first non-zero bit
46 MB = CountLeadingZeros_32(Val);
47 // look for the first zero bit after the run of ones
48 ME = CountLeadingZeros_32((Val - 1) ^ Val);
50 } else if (isShiftedMask_32(Val = ~Val)) { // invert mask
51 // effectively look for the first zero bit
52 ME = CountLeadingZeros_32(Val) - 1;
53 // effectively look for the first one bit after the run of zeros
54 MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
61 //===----------------------------------------------------------------------===//
62 // PPC32TargetLowering - PPC32 Implementation of the TargetLowering interface
64 class PPC32TargetLowering : public TargetLowering {
65 int VarArgsFrameIndex; // FrameIndex for start of varargs area.
66 int ReturnAddrIndex; // FrameIndex for return slot.
68 PPC32TargetLowering(TargetMachine &TM) : TargetLowering(TM) {
69 // Fold away setcc operations if possible.
70 setSetCCIsExpensive();
72 // Set up the register classes.
73 addRegisterClass(MVT::i32, PPC32::GPRCRegisterClass);
74 addRegisterClass(MVT::f32, PPC32::FPRCRegisterClass);
75 addRegisterClass(MVT::f64, PPC32::FPRCRegisterClass);
77 // PowerPC has no intrinsics for these particular operations
78 setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
79 setOperationAction(ISD::MEMSET, MVT::Other, Expand);
80 setOperationAction(ISD::MEMCPY, MVT::Other, Expand);
82 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
83 setOperationAction(ISD::SEXTLOAD, MVT::i1, Expand);
84 setOperationAction(ISD::SEXTLOAD, MVT::i8, Expand);
86 // PowerPC has no SREM/UREM instructions
87 setOperationAction(ISD::SREM, MVT::i32, Expand);
88 setOperationAction(ISD::UREM, MVT::i32, Expand);
90 // We don't support sin/cos/sqrt/fmod
91 setOperationAction(ISD::FSIN , MVT::f64, Expand);
92 setOperationAction(ISD::FCOS , MVT::f64, Expand);
93 setOperationAction(ISD::SREM , MVT::f64, Expand);
94 setOperationAction(ISD::FSIN , MVT::f32, Expand);
95 setOperationAction(ISD::FCOS , MVT::f32, Expand);
96 setOperationAction(ISD::SREM , MVT::f32, Expand);
98 // If we're enabling GP optimizations, use hardware square root
100 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
101 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
104 //PowerPC does not have CTPOP or CTTZ
105 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
106 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
108 setSetCCResultContents(ZeroOrOneSetCCResult);
109 addLegalFPImmediate(+0.0); // Necessary for FSEL
110 addLegalFPImmediate(-0.0); //
112 computeRegisterProperties();
115 /// LowerArguments - This hook must be implemented to indicate how we should
116 /// lower the arguments for the specified function, into the specified DAG.
117 virtual std::vector<SDOperand>
118 LowerArguments(Function &F, SelectionDAG &DAG);
120 /// LowerCallTo - This hook lowers an abstract call to a function into an
122 virtual std::pair<SDOperand, SDOperand>
123 LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, unsigned CC,
124 bool isTailCall, SDOperand Callee, ArgListTy &Args,
127 virtual SDOperand LowerVAStart(SDOperand Chain, SDOperand VAListP,
128 Value *VAListV, SelectionDAG &DAG);
130 virtual std::pair<SDOperand,SDOperand>
131 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
132 const Type *ArgTy, SelectionDAG &DAG);
134 virtual std::pair<SDOperand, SDOperand>
135 LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
141 std::vector<SDOperand>
142 PPC32TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
144 // add beautiful description of PPC stack frame format, or at least some docs
146 MachineFunction &MF = DAG.getMachineFunction();
147 MachineFrameInfo *MFI = MF.getFrameInfo();
148 MachineBasicBlock& BB = MF.front();
149 std::vector<SDOperand> ArgValues;
151 // Due to the rather complicated nature of the PowerPC ABI, rather than a
152 // fixed size array of physical args, for the sake of simplicity let the STL
153 // handle tracking them for us.
154 std::vector<unsigned> argVR, argPR, argOp;
155 unsigned ArgOffset = 24;
156 unsigned GPR_remaining = 8;
157 unsigned FPR_remaining = 13;
158 unsigned GPR_idx = 0, FPR_idx = 0;
159 static const unsigned GPR[] = {
160 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
161 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
163 static const unsigned FPR[] = {
164 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
165 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
168 // Add DAG nodes to load the arguments... On entry to a function on PPC,
169 // the arguments start at offset 24, although they are likely to be passed
171 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
172 SDOperand newroot, argt;
174 bool needsLoad = false;
175 bool ArgLive = !I->use_empty();
176 MVT::ValueType ObjectVT = getValueType(I->getType());
179 default: assert(0 && "Unhandled argument type!");
186 if (GPR_remaining > 0) {
187 MF.addLiveIn(GPR[GPR_idx]);
188 argt = newroot = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32,
190 if (ObjectVT != MVT::i32)
191 argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, newroot);
196 case MVT::i64: ObjSize = 8;
198 if (GPR_remaining > 0) {
199 SDOperand argHi, argLo;
200 MF.addLiveIn(GPR[GPR_idx]);
201 argHi = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot());
202 // If we have two or more remaining argument registers, then both halves
203 // of the i64 can be sourced from there. Otherwise, the lower half will
204 // have to come off the stack. This can happen when an i64 is preceded
205 // by 28 bytes of arguments.
206 if (GPR_remaining > 1) {
207 MF.addLiveIn(GPR[GPR_idx+1]);
208 argLo = DAG.getCopyFromReg(GPR[GPR_idx+1], MVT::i32, argHi);
210 int FI = MFI->CreateFixedObject(4, ArgOffset+4);
211 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
212 argLo = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
213 DAG.getSrcValue(NULL));
215 // Build the outgoing arg thingy
216 argt = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, argLo, argHi);
224 ObjSize = (ObjectVT == MVT::f64) ? 8 : 4;
226 if (FPR_remaining > 0) {
227 MF.addLiveIn(FPR[FPR_idx]);
228 argt = newroot = DAG.getCopyFromReg(FPR[FPR_idx], ObjectVT,
238 // We need to load the argument to a virtual register if we determined above
239 // that we ran out of physical registers of the appropriate type
241 unsigned SubregOffset = 0;
242 if (ObjectVT == MVT::i8 || ObjectVT == MVT::i1) SubregOffset = 3;
243 if (ObjectVT == MVT::i16) SubregOffset = 2;
244 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
245 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
246 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN,
247 DAG.getConstant(SubregOffset, MVT::i32));
248 argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
249 DAG.getSrcValue(NULL));
252 // Every 4 bytes of argument space consumes one of the GPRs available for
254 if (GPR_remaining > 0) {
255 unsigned delta = (GPR_remaining > 1 && ObjSize == 8) ? 2 : 1;
256 GPR_remaining -= delta;
259 ArgOffset += ObjSize;
261 DAG.setRoot(newroot.getValue(1));
263 ArgValues.push_back(argt);
266 // If the function takes variable number of arguments, make a frame index for
267 // the start of the first vararg value... for expansion of llvm.va_start.
269 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
270 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
271 // If this function is vararg, store any remaining integer argument regs
272 // to their spots on the stack so that they may be loaded by deferencing the
273 // result of va_next.
274 std::vector<SDOperand> MemOps;
275 for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) {
276 MF.addLiveIn(GPR[GPR_idx]);
277 SDOperand Val = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot());
278 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
279 Val, FIN, DAG.getSrcValue(NULL));
280 MemOps.push_back(Store);
281 // Increment the address by four for the next argument to store
282 SDOperand PtrOff = DAG.getConstant(4, getPointerTy());
283 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff);
285 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps));
288 // Finally, inform the code generator which regs we return values in.
289 switch (getValueType(F.getReturnType())) {
290 default: assert(0 && "Unknown type!");
291 case MVT::isVoid: break;
296 MF.addLiveOut(PPC::R3);
299 MF.addLiveOut(PPC::R3);
300 MF.addLiveOut(PPC::R4);
304 MF.addLiveOut(PPC::F1);
311 std::pair<SDOperand, SDOperand>
312 PPC32TargetLowering::LowerCallTo(SDOperand Chain,
313 const Type *RetTy, bool isVarArg,
314 unsigned CallingConv, bool isTailCall,
315 SDOperand Callee, ArgListTy &Args,
317 // args_to_use will accumulate outgoing args for the ISD::CALL case in
318 // SelectExpr to use to put the arguments in the appropriate registers.
319 std::vector<SDOperand> args_to_use;
321 // Count how many bytes are to be pushed on the stack, including the linkage
322 // area, and parameter passing area.
323 unsigned NumBytes = 24;
326 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
327 DAG.getConstant(NumBytes, getPointerTy()));
329 for (unsigned i = 0, e = Args.size(); i != e; ++i)
330 switch (getValueType(Args[i].second)) {
331 default: assert(0 && "Unknown value type!");
345 // Just to be safe, we'll always reserve the full 24 bytes of linkage area
346 // plus 32 bytes of argument space in case any called code gets funky on us.
347 // (Required by ABI to support var arg)
348 if (NumBytes < 56) NumBytes = 56;
350 // Adjust the stack pointer for the new arguments...
351 // These operations are automatically eliminated by the prolog/epilog pass
352 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
353 DAG.getConstant(NumBytes, getPointerTy()));
355 // Set up a copy of the stack pointer for use loading and storing any
356 // arguments that may not fit in the registers available for argument
358 SDOperand StackPtr = DAG.getCopyFromReg(PPC::R1, MVT::i32,
361 // Figure out which arguments are going to go in registers, and which in
362 // memory. Also, if this is a vararg function, floating point operations
363 // must be stored to our stack, and loaded into integer regs as well, if
364 // any integer regs are available for argument passing.
365 unsigned ArgOffset = 24;
366 unsigned GPR_remaining = 8;
367 unsigned FPR_remaining = 13;
369 std::vector<SDOperand> MemOps;
370 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
371 // PtrOff will be used to store the current argument to the stack if a
372 // register cannot be found for it.
373 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
374 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
375 MVT::ValueType ArgVT = getValueType(Args[i].second);
378 default: assert(0 && "Unexpected ValueType for argument!");
382 // Promote the integer to 32 bits. If the input type is signed use a
383 // sign extend, otherwise use a zero extend.
384 if (Args[i].second->isSigned())
385 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
387 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
390 if (GPR_remaining > 0) {
391 args_to_use.push_back(Args[i].first);
394 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
395 Args[i].first, PtrOff,
396 DAG.getSrcValue(NULL)));
401 // If we have one free GPR left, we can place the upper half of the i64
402 // in it, and store the other half to the stack. If we have two or more
403 // free GPRs, then we can pass both halves of the i64 in registers.
404 if (GPR_remaining > 0) {
405 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
406 Args[i].first, DAG.getConstant(1, MVT::i32));
407 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
408 Args[i].first, DAG.getConstant(0, MVT::i32));
409 args_to_use.push_back(Hi);
411 if (GPR_remaining > 0) {
412 args_to_use.push_back(Lo);
415 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
416 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
417 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
418 Lo, PtrOff, DAG.getSrcValue(NULL)));
421 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
422 Args[i].first, PtrOff,
423 DAG.getSrcValue(NULL)));
429 if (FPR_remaining > 0) {
430 args_to_use.push_back(Args[i].first);
433 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
434 Args[i].first, PtrOff,
435 DAG.getSrcValue(NULL));
436 MemOps.push_back(Store);
437 // Float varargs are always shadowed in available integer registers
438 if (GPR_remaining > 0) {
439 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
440 DAG.getSrcValue(NULL));
441 MemOps.push_back(Load);
442 args_to_use.push_back(Load);
445 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
446 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
447 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
448 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
449 DAG.getSrcValue(NULL));
450 MemOps.push_back(Load);
451 args_to_use.push_back(Load);
455 // If we have any FPRs remaining, we may also have GPRs remaining.
456 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
458 if (GPR_remaining > 0) {
459 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
462 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
463 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
468 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
469 Args[i].first, PtrOff,
470 DAG.getSrcValue(NULL)));
472 ArgOffset += (ArgVT == MVT::f32) ? 4 : 8;
477 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
480 std::vector<MVT::ValueType> RetVals;
481 MVT::ValueType RetTyVT = getValueType(RetTy);
482 if (RetTyVT != MVT::isVoid)
483 RetVals.push_back(RetTyVT);
484 RetVals.push_back(MVT::Other);
486 SDOperand TheCall = SDOperand(DAG.getCall(RetVals,
487 Chain, Callee, args_to_use), 0);
488 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
489 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
490 DAG.getConstant(NumBytes, getPointerTy()));
491 return std::make_pair(TheCall, Chain);
494 SDOperand PPC32TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
495 Value *VAListV, SelectionDAG &DAG) {
496 // vastart just stores the address of the VarArgsFrameIndex slot into the
497 // memory location argument.
498 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
499 return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR, VAListP,
500 DAG.getSrcValue(VAListV));
503 std::pair<SDOperand,SDOperand>
504 PPC32TargetLowering::LowerVAArg(SDOperand Chain,
505 SDOperand VAListP, Value *VAListV,
506 const Type *ArgTy, SelectionDAG &DAG) {
507 MVT::ValueType ArgVT = getValueType(ArgTy);
510 DAG.getLoad(MVT::i32, Chain, VAListP, DAG.getSrcValue(VAListV));
511 SDOperand Result = DAG.getLoad(ArgVT, Chain, VAList, DAG.getSrcValue(NULL));
513 if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
516 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
517 "Other types should have been promoted for varargs!");
520 VAList = DAG.getNode(ISD::ADD, VAList.getValueType(), VAList,
521 DAG.getConstant(Amt, VAList.getValueType()));
522 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
523 VAList, VAListP, DAG.getSrcValue(VAListV));
524 return std::make_pair(Result, Chain);
528 std::pair<SDOperand, SDOperand> PPC32TargetLowering::
529 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
531 assert(0 && "LowerFrameReturnAddress unimplemented");
536 Statistic<>Recorded("ppc-codegen", "Number of recording ops emitted");
537 Statistic<>FusedFP("ppc-codegen", "Number of fused fp operations");
538 Statistic<>FrameOff("ppc-codegen", "Number of frame idx offsets collapsed");
539 //===--------------------------------------------------------------------===//
540 /// ISel - PPC32 specific code to select PPC32 machine instructions for
541 /// SelectionDAG operations.
542 //===--------------------------------------------------------------------===//
543 class ISel : public SelectionDAGISel {
544 PPC32TargetLowering PPC32Lowering;
545 SelectionDAG *ISelDAG; // Hack to support us having a dag->dag transform
546 // for sdiv and udiv until it is put into the future
549 /// ExprMap - As shared expressions are codegen'd, we keep track of which
550 /// vreg the value is produced in, so we only emit one copy of each compiled
552 std::map<SDOperand, unsigned> ExprMap;
554 unsigned GlobalBaseReg;
555 bool GlobalBaseInitialized;
558 ISel(TargetMachine &TM) : SelectionDAGISel(PPC32Lowering), PPC32Lowering(TM),
561 /// runOnFunction - Override this function in order to reset our per-function
563 virtual bool runOnFunction(Function &Fn) {
564 // Make sure we re-emit a set of the global base reg if necessary
565 GlobalBaseInitialized = false;
566 return SelectionDAGISel::runOnFunction(Fn);
569 /// InstructionSelectBasicBlock - This callback is invoked by
570 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
571 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG) {
573 // Codegen the basic block.
575 Select(DAG.getRoot());
577 // Clear state used for selection.
582 // dag -> dag expanders for integer divide by constant
583 SDOperand BuildSDIVSequence(SDOperand N);
584 SDOperand BuildUDIVSequence(SDOperand N);
586 unsigned getGlobalBaseReg();
587 unsigned getConstDouble(double floatVal, unsigned Result);
588 void MoveCRtoGPR(unsigned CCReg, bool Inv, unsigned Idx, unsigned Result);
589 bool SelectBitfieldInsert(SDOperand OR, unsigned Result);
590 unsigned FoldIfWideZeroExtend(SDOperand N);
591 unsigned SelectCC(SDOperand CC, unsigned &Opc, bool &Inv, unsigned &Idx);
592 unsigned SelectCCExpr(SDOperand N, unsigned& Opc, bool &Inv, unsigned &Idx);
593 unsigned SelectExpr(SDOperand N, bool Recording=false);
594 void Select(SDOperand N);
596 unsigned SelectAddr(SDOperand N, unsigned& Reg, int& offset);
597 void SelectBranchCC(SDOperand N);
599 virtual const char *getPassName() const {
600 return "PowerPC Pattern Instruction Selection";
604 /// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
605 /// returns zero when the input is not exactly a power of two.
606 static unsigned ExactLog2(unsigned Val) {
607 if (Val == 0 || (Val & (Val-1))) return 0;
616 // IsRunOfOnes - returns true if Val consists of one contiguous run of 1's with
617 // any number of 0's on either side. the 1's are allowed to wrap from LSB to
618 // MSB. so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
619 // not, since all 1's are not contiguous.
620 static bool IsRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
625 // look for first set bit
627 for (; i < 32; i++) {
628 if ((Val & (1 << (31 - i))) != 0) {
635 // look for last set bit
636 for (; i < 32; i++) {
637 if ((Val & (1 << (31 - i))) == 0)
642 // look for next set bit
643 for (; i < 32; i++) {
644 if ((Val & (1 << (31 - i))) != 0)
648 // if we exhausted all the bits, we found a match at this point for 0*1*0*
652 // since we just encountered more 1's, if it doesn't wrap around to the
653 // most significant bit of the word, then we did not find a match to 1*0*1* so
658 // look for last set bit
659 for (MB = i; i < 32; i++) {
660 if ((Val & (1 << (31 - i))) == 0)
664 // if we exhausted all the bits, then we found a match for 1*0*1*, otherwise,
665 // the value is not a run of ones.
671 /// getImmediateForOpcode - This method returns a value indicating whether
672 /// the ConstantSDNode N can be used as an immediate to Opcode. The return
673 /// values are either 0, 1 or 2. 0 indicates that either N is not a
674 /// ConstantSDNode, or is not suitable for use by that opcode.
675 /// Return value codes for turning into an enum someday:
676 /// 1: constant may be used in normal immediate form.
677 /// 2: constant may be used in shifted immediate form.
678 /// 3: log base 2 of the constant may be used.
679 /// 4: constant is suitable for integer division conversion
680 /// 5: constant is a bitfield mask
682 static unsigned getImmediateForOpcode(SDOperand N, unsigned Opcode,
683 unsigned& Imm, bool U = false) {
684 if (N.getOpcode() != ISD::Constant) return 0;
686 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
691 if (isInt16(v)) { Imm = v & 0xFFFF; return 1; }
692 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
696 if (IsRunOfOnes(v, MB, ME)) { Imm = MB << 16 | ME & 0xFFFF; return 5; }
697 if (isUInt16(v)) { Imm = v & 0xFFFF; return 1; }
698 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
703 if (isUInt16(v)) { Imm = v & 0xFFFF; return 1; }
704 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
707 if (isInt16(v)) { Imm = v & 0xFFFF; return 1; }
710 // handle subtract-from separately from subtract, since subi is really addi
711 if (U && isInt16(v)) { Imm = v & 0xFFFF; return 1; }
712 if (!U && isInt16(-v)) { Imm = (-v) & 0xFFFF; return 1; }
715 if (U && isUInt16(v)) { Imm = v & 0xFFFF; return 1; }
716 if (!U && isInt16(v)) { Imm = v & 0xFFFF; return 1; }
719 if (isPowerOf2_32(v)) { Imm = Log2_32(v); return 3; }
720 if (isPowerOf2_32(-v)) { Imm = Log2_32(-v); return 3; }
721 if (v <= -2 || v >= 2) { return 4; }
724 if (v > 1) { return 4; }
730 /// NodeHasRecordingVariant - If SelectExpr can always produce code for
731 /// NodeOpcode that also sets CR0 as a side effect, return true. Otherwise,
733 static bool NodeHasRecordingVariant(unsigned NodeOpcode) {
735 default: return false;
742 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
743 /// to Condition. If the Condition is unordered or unsigned, the bool argument
744 /// U is set to true, otherwise it is set to false.
745 static unsigned getBCCForSetCC(unsigned Condition, bool& U) {
748 default: assert(0 && "Unknown condition!"); abort();
749 case ISD::SETEQ: return PPC::BEQ;
750 case ISD::SETNE: return PPC::BNE;
751 case ISD::SETULT: U = true;
752 case ISD::SETLT: return PPC::BLT;
753 case ISD::SETULE: U = true;
754 case ISD::SETLE: return PPC::BLE;
755 case ISD::SETUGT: U = true;
756 case ISD::SETGT: return PPC::BGT;
757 case ISD::SETUGE: U = true;
758 case ISD::SETGE: return PPC::BGE;
763 /// getCROpForOp - Return the condition register opcode (or inverted opcode)
764 /// associated with the SelectionDAG opcode.
765 static unsigned getCROpForSetCC(unsigned Opcode, bool Inv1, bool Inv2) {
767 default: assert(0 && "Unknown opcode!"); abort();
769 if (Inv1 && Inv2) return PPC::CRNOR; // De Morgan's Law
770 if (!Inv1 && !Inv2) return PPC::CRAND;
771 if (Inv1 ^ Inv2) return PPC::CRANDC;
773 if (Inv1 && Inv2) return PPC::CRNAND; // De Morgan's Law
774 if (!Inv1 && !Inv2) return PPC::CROR;
775 if (Inv1 ^ Inv2) return PPC::CRORC;
780 /// getCRIdxForSetCC - Return the index of the condition register field
781 /// associated with the SetCC condition, and whether or not the field is
782 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
783 static unsigned getCRIdxForSetCC(unsigned Condition, bool& Inv) {
785 default: assert(0 && "Unknown condition!"); abort();
787 case ISD::SETLT: Inv = false; return 0;
789 case ISD::SETGE: Inv = true; return 0;
791 case ISD::SETGT: Inv = false; return 1;
793 case ISD::SETLE: Inv = true; return 1;
794 case ISD::SETEQ: Inv = false; return 2;
795 case ISD::SETNE: Inv = true; return 2;
800 /// IndexedOpForOp - Return the indexed variant for each of the PowerPC load
801 /// and store immediate instructions.
802 static unsigned IndexedOpForOp(unsigned Opcode) {
804 default: assert(0 && "Unknown opcode!"); abort();
805 case PPC::LBZ: return PPC::LBZX; case PPC::STB: return PPC::STBX;
806 case PPC::LHZ: return PPC::LHZX; case PPC::STH: return PPC::STHX;
807 case PPC::LHA: return PPC::LHAX; case PPC::STW: return PPC::STWX;
808 case PPC::LWZ: return PPC::LWZX; case PPC::STFS: return PPC::STFSX;
809 case PPC::LFS: return PPC::LFSX; case PPC::STFD: return PPC::STFDX;
810 case PPC::LFD: return PPC::LFDX;
815 // Structure used to return the necessary information to codegen an SDIV as
818 int m; // magic number
819 int s; // shift amount
823 unsigned int m; // magic number
824 int a; // add indicator
825 int s; // shift amount
828 /// magic - calculate the magic numbers required to codegen an integer sdiv as
829 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
831 static struct ms magic(int d) {
833 unsigned int ad, anc, delta, q1, r1, q2, r2, t;
834 const unsigned int two31 = 0x80000000U;
838 t = two31 + ((unsigned int)d >> 31);
839 anc = t - 1 - t%ad; // absolute value of nc
840 p = 31; // initialize p
841 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
842 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
843 q2 = two31/ad; // initialize q2 = 2p/abs(d)
844 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
847 q1 = 2*q1; // update q1 = 2p/abs(nc)
848 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
849 if (r1 >= anc) { // must be unsigned comparison
853 q2 = 2*q2; // update q2 = 2p/abs(d)
854 r2 = 2*r2; // update r2 = rem(2p/abs(d))
855 if (r2 >= ad) { // must be unsigned comparison
860 } while (q1 < delta || (q1 == delta && r1 == 0));
863 if (d < 0) mag.m = -mag.m; // resulting magic number
864 mag.s = p - 32; // resulting shift
868 /// magicu - calculate the magic numbers required to codegen an integer udiv as
869 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
870 static struct mu magicu(unsigned d)
873 unsigned int nc, delta, q1, r1, q2, r2;
875 magu.a = 0; // initialize "add" indicator
877 p = 31; // initialize p
878 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
879 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
880 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
881 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
884 if (r1 >= nc - r1 ) {
885 q1 = 2*q1 + 1; // update q1
886 r1 = 2*r1 - nc; // update r1
889 q1 = 2*q1; // update q1
890 r1 = 2*r1; // update r1
892 if (r2 + 1 >= d - r2) {
893 if (q2 >= 0x7FFFFFFF) magu.a = 1;
894 q2 = 2*q2 + 1; // update q2
895 r2 = 2*r2 + 1 - d; // update r2
898 if (q2 >= 0x80000000) magu.a = 1;
899 q2 = 2*q2; // update q2
900 r2 = 2*r2 + 1; // update r2
903 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
904 magu.m = q2 + 1; // resulting magic number
905 magu.s = p - 32; // resulting shift
910 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
911 /// return a DAG expression to select that will generate the same value by
912 /// multiplying by a magic number. See:
913 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
914 SDOperand ISel::BuildSDIVSequence(SDOperand N) {
915 int d = (int)cast<ConstantSDNode>(N.getOperand(1))->getSignExtended();
916 ms magics = magic(d);
917 // Multiply the numerator (operand 0) by the magic value
918 SDOperand Q = ISelDAG->getNode(ISD::MULHS, MVT::i32, N.getOperand(0),
919 ISelDAG->getConstant(magics.m, MVT::i32));
920 // If d > 0 and m < 0, add the numerator
921 if (d > 0 && magics.m < 0)
922 Q = ISelDAG->getNode(ISD::ADD, MVT::i32, Q, N.getOperand(0));
923 // If d < 0 and m > 0, subtract the numerator.
924 if (d < 0 && magics.m > 0)
925 Q = ISelDAG->getNode(ISD::SUB, MVT::i32, Q, N.getOperand(0));
926 // Shift right algebraic if shift value is nonzero
928 Q = ISelDAG->getNode(ISD::SRA, MVT::i32, Q,
929 ISelDAG->getConstant(magics.s, MVT::i32));
930 // Extract the sign bit and add it to the quotient
932 ISelDAG->getNode(ISD::SRL, MVT::i32, Q, ISelDAG->getConstant(31, MVT::i32));
933 return ISelDAG->getNode(ISD::ADD, MVT::i32, Q, T);
936 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
937 /// return a DAG expression to select that will generate the same value by
938 /// multiplying by a magic number. See:
939 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
940 SDOperand ISel::BuildUDIVSequence(SDOperand N) {
942 (unsigned)cast<ConstantSDNode>(N.getOperand(1))->getSignExtended();
943 mu magics = magicu(d);
944 // Multiply the numerator (operand 0) by the magic value
945 SDOperand Q = ISelDAG->getNode(ISD::MULHU, MVT::i32, N.getOperand(0),
946 ISelDAG->getConstant(magics.m, MVT::i32));
948 Q = ISelDAG->getNode(ISD::SRL, MVT::i32, Q,
949 ISelDAG->getConstant(magics.s, MVT::i32));
951 SDOperand NPQ = ISelDAG->getNode(ISD::SUB, MVT::i32, N.getOperand(0), Q);
952 NPQ = ISelDAG->getNode(ISD::SRL, MVT::i32, NPQ,
953 ISelDAG->getConstant(1, MVT::i32));
954 NPQ = ISelDAG->getNode(ISD::ADD, MVT::i32, NPQ, Q);
955 Q = ISelDAG->getNode(ISD::SRL, MVT::i32, NPQ,
956 ISelDAG->getConstant(magics.s-1, MVT::i32));
961 /// getGlobalBaseReg - Output the instructions required to put the
962 /// base address to use for accessing globals into a register.
964 unsigned ISel::getGlobalBaseReg() {
965 if (!GlobalBaseInitialized) {
966 // Insert the set of GlobalBaseReg into the first MBB of the function
967 MachineBasicBlock &FirstMBB = BB->getParent()->front();
968 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
969 GlobalBaseReg = MakeReg(MVT::i32);
970 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
971 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg).addReg(PPC::LR);
972 GlobalBaseInitialized = true;
974 return GlobalBaseReg;
977 /// getConstDouble - Loads a floating point value into a register, via the
978 /// Constant Pool. Optionally takes a register in which to load the value.
979 unsigned ISel::getConstDouble(double doubleVal, unsigned Result=0) {
980 unsigned Tmp1 = MakeReg(MVT::i32);
981 if (0 == Result) Result = MakeReg(MVT::f64);
982 MachineConstantPool *CP = BB->getParent()->getConstantPool();
983 ConstantFP *CFP = ConstantFP::get(Type::DoubleTy, doubleVal);
984 unsigned CPI = CP->getConstantPoolIndex(CFP);
986 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
987 .addConstantPoolIndex(CPI);
989 BuildMI(BB, PPC::LIS, 1, Tmp1).addConstantPoolIndex(CPI);
990 BuildMI(BB, PPC::LFD, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
994 /// MoveCRtoGPR - Move CCReg[Idx] to the least significant bit of Result. If
995 /// Inv is true, then invert the result.
996 void ISel::MoveCRtoGPR(unsigned CCReg, bool Inv, unsigned Idx, unsigned Result){
997 unsigned IntCR = MakeReg(MVT::i32);
998 BuildMI(BB, PPC::MCRF, 1, PPC::CR7).addReg(CCReg);
999 BuildMI(BB, GPOPT ? PPC::MFOCRF : PPC::MFCR, 1, IntCR).addReg(PPC::CR7);
1001 unsigned Tmp1 = MakeReg(MVT::i32);
1002 BuildMI(BB, PPC::RLWINM, 4, Tmp1).addReg(IntCR).addImm(32-(3-Idx))
1003 .addImm(31).addImm(31);
1004 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp1).addImm(1);
1006 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(IntCR).addImm(32-(3-Idx))
1007 .addImm(31).addImm(31);
1011 /// SelectBitfieldInsert - turn an or of two masked values into
1012 /// the rotate left word immediate then mask insert (rlwimi) instruction.
1013 /// Returns true on success, false if the caller still needs to select OR.
1015 /// Patterns matched:
1016 /// 1. or shl, and 5. or and, and
1017 /// 2. or and, shl 6. or shl, shr
1018 /// 3. or shr, and 7. or shr, shl
1020 bool ISel::SelectBitfieldInsert(SDOperand OR, unsigned Result) {
1021 bool IsRotate = false;
1022 unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, Amount = 0;
1024 SDOperand Op0 = OR.getOperand(0);
1025 SDOperand Op1 = OR.getOperand(1);
1027 unsigned Op0Opc = Op0.getOpcode();
1028 unsigned Op1Opc = Op1.getOpcode();
1030 // Verify that we have the correct opcodes
1031 if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc)
1033 if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc)
1036 // Generate Mask value for Target
1037 if (ConstantSDNode *CN =
1038 dyn_cast<ConstantSDNode>(Op0.getOperand(1).Val)) {
1040 case ISD::SHL: TgtMask <<= (unsigned)CN->getValue(); break;
1041 case ISD::SRL: TgtMask >>= (unsigned)CN->getValue(); break;
1042 case ISD::AND: TgtMask &= (unsigned)CN->getValue(); break;
1048 // Generate Mask value for Insert
1049 if (ConstantSDNode *CN =
1050 dyn_cast<ConstantSDNode>(Op1.getOperand(1).Val)) {
1053 Amount = CN->getValue();
1055 if (Op0Opc == ISD::SRL) IsRotate = true;
1058 Amount = CN->getValue();
1061 if (Op0Opc == ISD::SHL) IsRotate = true;
1064 InsMask &= (unsigned)CN->getValue();
1073 // If both of the inputs are ANDs and one of them has a logical shift by
1074 // constant as its input, make that the inserted value so that we can combine
1075 // the shift into the rotate part of the rlwimi instruction
1076 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
1077 if (Op1.getOperand(0).getOpcode() == ISD::SHL ||
1078 Op1.getOperand(0).getOpcode() == ISD::SRL) {
1079 if (ConstantSDNode *CN =
1080 dyn_cast<ConstantSDNode>(Op1.getOperand(0).getOperand(1).Val)) {
1081 Amount = Op1.getOperand(0).getOpcode() == ISD::SHL ?
1082 CN->getValue() : 32 - CN->getValue();
1083 Tmp3 = SelectExpr(Op1.getOperand(0).getOperand(0));
1085 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
1086 Op0.getOperand(0).getOpcode() == ISD::SRL) {
1087 if (ConstantSDNode *CN =
1088 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(1).Val)) {
1089 std::swap(Op0, Op1);
1090 std::swap(TgtMask, InsMask);
1091 Amount = Op1.getOperand(0).getOpcode() == ISD::SHL ?
1092 CN->getValue() : 32 - CN->getValue();
1093 Tmp3 = SelectExpr(Op1.getOperand(0).getOperand(0));
1098 // Verify that the Target mask and Insert mask together form a full word mask
1099 // and that the Insert mask is a run of set bits (which implies both are runs
1100 // of set bits). Given that, Select the arguments and generate the rlwimi
1103 if (((TgtMask & InsMask) == 0) && IsRunOfOnes(InsMask, MB, ME)) {
1104 unsigned Tmp1, Tmp2;
1105 bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF;
1106 // Check for rotlwi / rotrwi here, a special case of bitfield insert
1107 // where both bitfield halves are sourced from the same value.
1108 if (IsRotate && fullMask &&
1109 OR.getOperand(0).getOperand(0) == OR.getOperand(1).getOperand(0)) {
1110 Tmp1 = SelectExpr(OR.getOperand(0).getOperand(0));
1111 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(Amount)
1112 .addImm(0).addImm(31);
1115 if (Op0Opc == ISD::AND && fullMask)
1116 Tmp1 = SelectExpr(Op0.getOperand(0));
1118 Tmp1 = SelectExpr(Op0);
1119 Tmp2 = Tmp3 ? Tmp3 : SelectExpr(Op1.getOperand(0));
1120 BuildMI(BB, PPC::RLWIMI, 5, Result).addReg(Tmp1).addReg(Tmp2)
1121 .addImm(Amount).addImm(MB).addImm(ME);
1127 /// FoldIfWideZeroExtend - 32 bit PowerPC implicit masks shift amounts to the
1128 /// low six bits. If the shift amount is an ISD::AND node with a mask that is
1129 /// wider than the implicit mask, then we can get rid of the AND and let the
1130 /// shift do the mask.
1131 unsigned ISel::FoldIfWideZeroExtend(SDOperand N) {
1133 if (N.getOpcode() == ISD::AND &&
1134 5 == getImmediateForOpcode(N.getOperand(1), ISD::AND, C) && // isMask
1135 31 == (C & 0xFFFF) && // ME
1136 26 >= (C >> 16)) // MB
1137 return SelectExpr(N.getOperand(0));
1139 return SelectExpr(N);
1142 unsigned ISel::SelectCC(SDOperand CC, unsigned& Opc, bool &Inv, unsigned& Idx) {
1143 unsigned Result, Tmp1, Tmp2;
1144 bool AlreadySelected = false;
1145 static const unsigned CompareOpcodes[] =
1146 { PPC::FCMPU, PPC::FCMPU, PPC::CMPW, PPC::CMPLW };
1148 // Allocate a condition register for this expression
1149 Result = RegMap->createVirtualRegister(PPC32::CRRCRegisterClass);
1151 // If the first operand to the select is a SETCC node, then we can fold it
1152 // into the branch that selects which value to return.
1153 if (SetCCSDNode* SetCC = dyn_cast<SetCCSDNode>(CC.Val)) {
1155 Opc = getBCCForSetCC(SetCC->getCondition(), U);
1156 Idx = getCRIdxForSetCC(SetCC->getCondition(), Inv);
1158 // Pass the optional argument U to getImmediateForOpcode for SETCC,
1159 // so that it knows whether the SETCC immediate range is signed or not.
1160 if (1 == getImmediateForOpcode(SetCC->getOperand(1), ISD::SETCC,
1162 // For comparisons against zero, we can implicity set CR0 if a recording
1163 // variant (e.g. 'or.' instead of 'or') of the instruction that defines
1164 // operand zero of the SetCC node is available.
1166 NodeHasRecordingVariant(SetCC->getOperand(0).getOpcode()) &&
1167 SetCC->getOperand(0).Val->hasOneUse()) {
1168 RecordSuccess = false;
1169 Tmp1 = SelectExpr(SetCC->getOperand(0), true);
1170 if (RecordSuccess) {
1172 BuildMI(BB, PPC::MCRF, 1, Result).addReg(PPC::CR0);
1175 AlreadySelected = true;
1177 // If we could not implicitly set CR0, then emit a compare immediate
1179 if (!AlreadySelected) Tmp1 = SelectExpr(SetCC->getOperand(0));
1181 BuildMI(BB, PPC::CMPLWI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1183 BuildMI(BB, PPC::CMPWI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1185 bool IsInteger = MVT::isInteger(SetCC->getOperand(0).getValueType());
1186 unsigned CompareOpc = CompareOpcodes[2 * IsInteger + U];
1187 Tmp1 = SelectExpr(SetCC->getOperand(0));
1188 Tmp2 = SelectExpr(SetCC->getOperand(1));
1189 BuildMI(BB, CompareOpc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1192 // If this isn't a SetCC, then select the value and compare it against zero,
1193 // treating it as if it were a boolean.
1195 Idx = getCRIdxForSetCC(ISD::SETNE, Inv);
1196 Tmp1 = SelectExpr(CC);
1197 BuildMI(BB, PPC::CMPLWI, 2, Result).addReg(Tmp1).addImm(0);
1202 unsigned ISel::SelectCCExpr(SDOperand N, unsigned& Opc, bool &Inv,
1205 unsigned Idx0, Idx1, CROpc, Opc1, Tmp1, Tmp2;
1207 // Allocate a condition register for this expression
1208 unsigned Result = RegMap->createVirtualRegister(PPC32::CRRCRegisterClass);
1210 // Check for the operations we support:
1211 switch(N.getOpcode()) {
1214 Idx = getCRIdxForSetCC(ISD::SETNE, Inv);
1215 Tmp1 = SelectExpr(N);
1216 BuildMI(BB, PPC::CMPLWI, 2, Result).addReg(Tmp1).addImm(0);
1220 Tmp1 = SelectCCExpr(N.getOperand(0), Opc, Inv0, Idx0);
1221 Tmp2 = SelectCCExpr(N.getOperand(1), Opc1, Inv1, Idx1);
1222 CROpc = getCROpForSetCC(N.getOpcode(), Inv0, Inv1);
1223 if (Inv0 && !Inv1) {
1224 std::swap(Tmp1, Tmp2);
1225 std::swap(Idx0, Idx1);
1228 if (Inv0 && Inv1) Opc = PPC32InstrInfo::invertPPCBranchOpcode(Opc);
1229 BuildMI(BB, CROpc, 5, Result).addImm(Idx0).addReg(Tmp1).addImm(Idx0)
1230 .addReg(Tmp2).addImm(Idx1);
1235 Tmp1 = SelectCC(N, Opc, Inv, Idx);
1242 /// Check to see if the load is a constant offset from a base register
1243 unsigned ISel::SelectAddr(SDOperand N, unsigned& Reg, int& offset)
1245 unsigned imm = 0, opcode = N.getOpcode();
1246 if (N.getOpcode() == ISD::ADD) {
1247 bool isFrame = N.getOperand(0).getOpcode() == ISD::FrameIndex;
1248 if (1 == getImmediateForOpcode(N.getOperand(1), opcode, imm)) {
1252 Reg = cast<FrameIndexSDNode>(N.getOperand(0))->getIndex();
1255 Reg = SelectExpr(N.getOperand(0));
1259 Reg = SelectExpr(N.getOperand(0));
1260 offset = SelectExpr(N.getOperand(1));
1264 Reg = SelectExpr(N);
1269 void ISel::SelectBranchCC(SDOperand N)
1271 MachineBasicBlock *Dest =
1272 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
1275 unsigned Opc, CCReg, Idx;
1276 Select(N.getOperand(0)); //chain
1277 CCReg = SelectCC(N.getOperand(1), Opc, Inv, Idx);
1279 // Iterate to the next basic block
1280 ilist<MachineBasicBlock>::iterator It = BB;
1283 // If this is a two way branch, then grab the fallthrough basic block argument
1284 // and build a PowerPC branch pseudo-op, suitable for long branch conversion
1285 // if necessary by the branch selection pass. Otherwise, emit a standard
1286 // conditional branch.
1287 if (N.getOpcode() == ISD::BRCONDTWOWAY) {
1288 MachineBasicBlock *Fallthrough =
1289 cast<BasicBlockSDNode>(N.getOperand(3))->getBasicBlock();
1291 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
1292 .addMBB(Dest).addMBB(Fallthrough);
1293 if (Fallthrough != It)
1294 BuildMI(BB, PPC::B, 1).addMBB(Fallthrough);
1296 if (Fallthrough != It) {
1297 Opc = PPC32InstrInfo::invertPPCBranchOpcode(Opc);
1298 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
1299 .addMBB(Fallthrough).addMBB(Dest);
1303 // If the fallthrough path is off the end of the function, which would be
1304 // undefined behavior, set it to be the same as the current block because
1305 // we have nothing better to set it to, and leaving it alone will cause the
1306 // PowerPC Branch Selection pass to crash.
1307 if (It == BB->getParent()->end()) It = Dest;
1308 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
1309 .addMBB(Dest).addMBB(It);
1314 unsigned ISel::SelectExpr(SDOperand N, bool Recording) {
1316 unsigned Tmp1, Tmp2, Tmp3;
1318 unsigned opcode = N.getOpcode();
1320 SDNode *Node = N.Val;
1321 MVT::ValueType DestType = N.getValueType();
1323 if (Node->getOpcode() == ISD::CopyFromReg &&
1324 (MRegisterInfo::isVirtualRegister(cast<RegSDNode>(Node)->getReg()) ||
1325 cast<RegSDNode>(Node)->getReg() == PPC::R1))
1326 // Just use the specified register as our input.
1327 return cast<RegSDNode>(Node)->getReg();
1329 unsigned &Reg = ExprMap[N];
1330 if (Reg) return Reg;
1332 switch (N.getOpcode()) {
1334 Reg = Result = (N.getValueType() != MVT::Other) ?
1335 MakeReg(N.getValueType()) : 1;
1339 // If this is a call instruction, make sure to prepare ALL of the result
1340 // values as well as the chain.
1341 if (Node->getNumValues() == 1)
1342 Reg = Result = 1; // Void call, just a chain.
1344 Result = MakeReg(Node->getValueType(0));
1345 ExprMap[N.getValue(0)] = Result;
1346 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
1347 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
1348 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
1351 case ISD::ADD_PARTS:
1352 case ISD::SUB_PARTS:
1353 case ISD::SHL_PARTS:
1354 case ISD::SRL_PARTS:
1355 case ISD::SRA_PARTS:
1356 Result = MakeReg(Node->getValueType(0));
1357 ExprMap[N.getValue(0)] = Result;
1358 for (unsigned i = 1, e = N.Val->getNumValues(); i != e; ++i)
1359 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
1366 assert(0 && "Node not handled!\n");
1368 BuildMI(BB, PPC::IMPLICIT_DEF, 0, Result);
1370 case ISD::DYNAMIC_STACKALLOC:
1371 // Generate both result values. FIXME: Need a better commment here?
1373 ExprMap[N.getValue(1)] = 1;
1375 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1377 // FIXME: We are currently ignoring the requested alignment for handling
1378 // greater than the stack alignment. This will need to be revisited at some
1379 // point. Align = N.getOperand(2);
1380 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
1381 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
1382 std::cerr << "Cannot allocate stack object with greater alignment than"
1383 << " the stack alignment yet!";
1386 Select(N.getOperand(0));
1387 Tmp1 = SelectExpr(N.getOperand(1));
1388 // Subtract size from stack pointer, thereby allocating some space.
1389 BuildMI(BB, PPC::SUBF, 2, PPC::R1).addReg(Tmp1).addReg(PPC::R1);
1390 // Put a pointer to the space into the result register by copying the SP
1391 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R1).addReg(PPC::R1);
1394 case ISD::ConstantPool:
1395 Tmp1 = cast<ConstantPoolSDNode>(N)->getIndex();
1396 Tmp2 = MakeReg(MVT::i32);
1398 BuildMI(BB, PPC::ADDIS, 2, Tmp2).addReg(getGlobalBaseReg())
1399 .addConstantPoolIndex(Tmp1);
1401 BuildMI(BB, PPC::LIS, 1, Tmp2).addConstantPoolIndex(Tmp1);
1402 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp2).addConstantPoolIndex(Tmp1);
1405 case ISD::FrameIndex:
1406 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
1407 addFrameReference(BuildMI(BB, PPC::ADDI, 2, Result), (int)Tmp1, 0, false);
1410 case ISD::GlobalAddress: {
1411 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
1412 Tmp1 = MakeReg(MVT::i32);
1414 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
1415 .addGlobalAddress(GV);
1417 BuildMI(BB, PPC::LIS, 1, Tmp1).addGlobalAddress(GV);
1418 if (GV->hasWeakLinkage() || GV->isExternal()) {
1419 BuildMI(BB, PPC::LWZ, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
1421 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp1).addGlobalAddress(GV);
1429 case ISD::SEXTLOAD: {
1430 MVT::ValueType TypeBeingLoaded = (ISD::LOAD == opcode) ?
1431 Node->getValueType(0) : cast<VTSDNode>(Node->getOperand(3))->getVT();
1432 bool sext = (ISD::SEXTLOAD == opcode);
1434 // Make sure we generate both values.
1436 ExprMap[N.getValue(1)] = 1; // Generate the token
1438 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1440 SDOperand Chain = N.getOperand(0);
1441 SDOperand Address = N.getOperand(1);
1444 switch (TypeBeingLoaded) {
1445 default: Node->dump(); assert(0 && "Cannot load this type!");
1446 case MVT::i1: Opc = PPC::LBZ; break;
1447 case MVT::i8: Opc = PPC::LBZ; break;
1448 case MVT::i16: Opc = sext ? PPC::LHA : PPC::LHZ; break;
1449 case MVT::i32: Opc = PPC::LWZ; break;
1450 case MVT::f32: Opc = PPC::LFS; break;
1451 case MVT::f64: Opc = PPC::LFD; break;
1454 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Address)) {
1455 Tmp1 = MakeReg(MVT::i32);
1456 int CPI = CP->getIndex();
1458 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
1459 .addConstantPoolIndex(CPI);
1461 BuildMI(BB, PPC::LIS, 1, Tmp1).addConstantPoolIndex(CPI);
1462 BuildMI(BB, Opc, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
1463 } else if (Address.getOpcode() == ISD::FrameIndex) {
1464 Tmp1 = cast<FrameIndexSDNode>(Address)->getIndex();
1465 addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1);
1466 } else if(GlobalAddressSDNode *GN = dyn_cast<GlobalAddressSDNode>(Address)){
1467 GlobalValue *GV = GN->getGlobal();
1468 Tmp1 = MakeReg(MVT::i32);
1470 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
1471 .addGlobalAddress(GV);
1473 BuildMI(BB, PPC::LIS, 1, Tmp1).addGlobalAddress(GV);
1474 if (GV->hasWeakLinkage() || GV->isExternal()) {
1475 Tmp2 = MakeReg(MVT::i32);
1476 BuildMI(BB, PPC::LWZ, 2, Tmp2).addGlobalAddress(GV).addReg(Tmp1);
1477 BuildMI(BB, Opc, 2, Result).addSImm(0).addReg(Tmp2);
1479 BuildMI(BB, Opc, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
1483 switch(SelectAddr(Address, Tmp1, offset)) {
1484 default: assert(0 && "Unhandled return value from SelectAddr");
1485 case 0: // imm offset, no frame, no index
1486 BuildMI(BB, Opc, 2, Result).addSImm(offset).addReg(Tmp1);
1488 case 1: // imm offset + frame index
1489 addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1, offset);
1491 case 2: // base+index addressing
1492 Opc = IndexedOpForOp(Opc);
1493 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(offset);
1502 unsigned GPR_idx = 0, FPR_idx = 0;
1503 static const unsigned GPR[] = {
1504 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1505 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1507 static const unsigned FPR[] = {
1508 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1509 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1512 // Lower the chain for this call.
1513 Select(N.getOperand(0));
1514 ExprMap[N.getValue(Node->getNumValues()-1)] = 1;
1516 MachineInstr *CallMI;
1517 // Emit the correct call instruction based on the type of symbol called.
1518 if (GlobalAddressSDNode *GASD =
1519 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
1520 CallMI = BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(GASD->getGlobal(),
1522 } else if (ExternalSymbolSDNode *ESSDN =
1523 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
1524 CallMI = BuildMI(PPC::CALLpcrel, 1).addExternalSymbol(ESSDN->getSymbol(),
1527 Tmp1 = SelectExpr(N.getOperand(1));
1528 BuildMI(BB, PPC::OR, 2, PPC::R12).addReg(Tmp1).addReg(Tmp1);
1529 BuildMI(BB, PPC::MTCTR, 1).addReg(PPC::R12);
1530 CallMI = BuildMI(PPC::CALLindirect, 3).addImm(20).addImm(0)
1534 // Load the register args to virtual regs
1535 std::vector<unsigned> ArgVR;
1536 for(int i = 2, e = Node->getNumOperands(); i < e; ++i)
1537 ArgVR.push_back(SelectExpr(N.getOperand(i)));
1539 // Copy the virtual registers into the appropriate argument register
1540 for(int i = 0, e = ArgVR.size(); i < e; ++i) {
1541 switch(N.getOperand(i+2).getValueType()) {
1542 default: Node->dump(); assert(0 && "Unknown value type for call");
1547 assert(GPR_idx < 8 && "Too many int args");
1548 if (N.getOperand(i+2).getOpcode() != ISD::UNDEF) {
1549 BuildMI(BB, PPC::OR,2,GPR[GPR_idx]).addReg(ArgVR[i]).addReg(ArgVR[i]);
1550 CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use);
1556 assert(FPR_idx < 13 && "Too many fp args");
1557 BuildMI(BB, PPC::FMR, 1, FPR[FPR_idx]).addReg(ArgVR[i]);
1558 CallMI->addRegOperand(FPR[FPR_idx], MachineOperand::Use);
1564 // Put the call instruction in the correct place in the MachineBasicBlock
1565 BB->push_back(CallMI);
1567 switch (Node->getValueType(0)) {
1568 default: assert(0 && "Unknown value type for call result!");
1569 case MVT::Other: return 1;
1574 if (Node->getValueType(1) == MVT::i32) {
1575 BuildMI(BB, PPC::OR, 2, Result+1).addReg(PPC::R3).addReg(PPC::R3);
1576 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R4).addReg(PPC::R4);
1578 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R3).addReg(PPC::R3);
1583 BuildMI(BB, PPC::FMR, 1, Result).addReg(PPC::F1);
1586 return Result+N.ResNo;
1589 case ISD::SIGN_EXTEND:
1590 case ISD::SIGN_EXTEND_INREG:
1591 Tmp1 = SelectExpr(N.getOperand(0));
1592 switch(cast<VTSDNode>(Node->getOperand(1))->getVT()) {
1593 default: Node->dump(); assert(0 && "Unhandled SIGN_EXTEND type"); break;
1595 BuildMI(BB, PPC::EXTSH, 1, Result).addReg(Tmp1);
1598 BuildMI(BB, PPC::EXTSB, 1, Result).addReg(Tmp1);
1601 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp1).addSImm(0);
1606 case ISD::CopyFromReg:
1607 DestType = N.getValue(0).getValueType();
1609 Result = ExprMap[N.getValue(0)] = MakeReg(DestType);
1610 Tmp1 = dyn_cast<RegSDNode>(Node)->getReg();
1611 if (MVT::isInteger(DestType))
1612 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1614 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
1618 Tmp1 = SelectExpr(N.getOperand(0));
1619 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1620 Tmp2 = CN->getValue() & 0x1F;
1621 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(Tmp2).addImm(0)
1624 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1625 BuildMI(BB, PPC::SLW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1630 Tmp1 = SelectExpr(N.getOperand(0));
1631 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1632 Tmp2 = CN->getValue() & 0x1F;
1633 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(32-Tmp2)
1634 .addImm(Tmp2).addImm(31);
1636 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1637 BuildMI(BB, PPC::SRW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1642 Tmp1 = SelectExpr(N.getOperand(0));
1643 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1644 Tmp2 = CN->getValue() & 0x1F;
1645 BuildMI(BB, PPC::SRAWI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1647 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1648 BuildMI(BB, PPC::SRAW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1653 Tmp1 = SelectExpr(N.getOperand(0));
1654 BuildMI(BB, PPC::CNTLZW, 1, Result).addReg(Tmp1);
1658 if (!MVT::isInteger(DestType)) {
1659 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
1660 N.getOperand(0).Val->hasOneUse()) {
1661 ++FusedFP; // Statistic
1662 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1663 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1664 Tmp3 = SelectExpr(N.getOperand(1));
1665 Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
1666 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1669 if (!NoExcessFPPrecision && N.getOperand(1).getOpcode() == ISD::MUL &&
1670 N.getOperand(1).Val->hasOneUse()) {
1671 ++FusedFP; // Statistic
1672 Tmp1 = SelectExpr(N.getOperand(1).getOperand(0));
1673 Tmp2 = SelectExpr(N.getOperand(1).getOperand(1));
1674 Tmp3 = SelectExpr(N.getOperand(0));
1675 Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
1676 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1679 Opc = DestType == MVT::f64 ? PPC::FADD : PPC::FADDS;
1680 Tmp1 = SelectExpr(N.getOperand(0));
1681 Tmp2 = SelectExpr(N.getOperand(1));
1682 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1685 Tmp1 = SelectExpr(N.getOperand(0));
1686 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1687 default: assert(0 && "unhandled result code");
1688 case 0: // No immediate
1689 Tmp2 = SelectExpr(N.getOperand(1));
1690 BuildMI(BB, PPC::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1692 case 1: // Low immediate
1693 BuildMI(BB, PPC::ADDI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1695 case 2: // Shifted immediate
1696 BuildMI(BB, PPC::ADDIS, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1702 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1703 default: assert(0 && "unhandled result code");
1704 case 0: // No immediate
1705 // Check for andc: and, (xor a, -1), b
1706 if (N.getOperand(0).getOpcode() == ISD::XOR &&
1707 N.getOperand(0).getOperand(1).getOpcode() == ISD::Constant &&
1708 cast<ConstantSDNode>(N.getOperand(0).getOperand(1))->isAllOnesValue()) {
1709 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1710 Tmp2 = SelectExpr(N.getOperand(1));
1711 BuildMI(BB, PPC::ANDC, 2, Result).addReg(Tmp2).addReg(Tmp1);
1714 // It wasn't and-with-complement, emit a regular and
1715 Tmp1 = SelectExpr(N.getOperand(0));
1716 Tmp2 = SelectExpr(N.getOperand(1));
1717 Opc = Recording ? PPC::ANDo : PPC::AND;
1718 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1720 case 1: // Low immediate
1721 Tmp1 = SelectExpr(N.getOperand(0));
1722 BuildMI(BB, PPC::ANDIo, 2, Result).addReg(Tmp1).addImm(Tmp2);
1724 case 2: // Shifted immediate
1725 Tmp1 = SelectExpr(N.getOperand(0));
1726 BuildMI(BB, PPC::ANDISo, 2, Result).addReg(Tmp1).addImm(Tmp2);
1728 case 5: // Bitfield mask
1729 Opc = Recording ? PPC::RLWINMo : PPC::RLWINM;
1730 Tmp3 = Tmp2 >> 16; // MB
1731 Tmp2 &= 0xFFFF; // ME
1733 // FIXME: Catch SHL-AND in addition to SRL-AND in this block.
1734 if (N.getOperand(0).getOpcode() == ISD::SRL)
1735 if (ConstantSDNode *SA =
1736 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
1738 // We can fold the RLWINM and the SRL together if the mask is
1739 // clearing the top bits which are rotated around.
1740 unsigned RotAmt = 32-(SA->getValue() & 31);
1741 if (Tmp2 <= RotAmt) {
1742 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1743 BuildMI(BB, Opc, 4, Result).addReg(Tmp1).addImm(RotAmt)
1744 .addImm(Tmp3).addImm(Tmp2);
1749 Tmp1 = SelectExpr(N.getOperand(0));
1750 BuildMI(BB, Opc, 4, Result).addReg(Tmp1).addImm(0)
1751 .addImm(Tmp3).addImm(Tmp2);
1754 RecordSuccess = true;
1758 if (SelectBitfieldInsert(N, Result))
1760 Tmp1 = SelectExpr(N.getOperand(0));
1761 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1762 default: assert(0 && "unhandled result code");
1763 case 0: // No immediate
1764 Tmp2 = SelectExpr(N.getOperand(1));
1765 Opc = Recording ? PPC::ORo : PPC::OR;
1766 RecordSuccess = true;
1767 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1769 case 1: // Low immediate
1770 BuildMI(BB, PPC::ORI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1772 case 2: // Shifted immediate
1773 BuildMI(BB, PPC::ORIS, 2, Result).addReg(Tmp1).addImm(Tmp2);
1779 // Check for EQV: xor, (xor a, -1), b
1780 if (N.getOperand(0).getOpcode() == ISD::XOR &&
1781 N.getOperand(0).getOperand(1).getOpcode() == ISD::Constant &&
1782 cast<ConstantSDNode>(N.getOperand(0).getOperand(1))->isAllOnesValue()) {
1783 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1784 Tmp2 = SelectExpr(N.getOperand(1));
1785 BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1788 // Check for NOT, NOR, EQV, and NAND: xor (copy, or, xor, and), -1
1789 if (N.getOperand(1).getOpcode() == ISD::Constant &&
1790 cast<ConstantSDNode>(N.getOperand(1))->isAllOnesValue()) {
1791 switch(N.getOperand(0).getOpcode()) {
1793 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1794 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1795 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1798 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1799 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1800 BuildMI(BB, PPC::NAND, 2, Result).addReg(Tmp1).addReg(Tmp2);
1803 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1804 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1805 BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1808 Tmp1 = SelectExpr(N.getOperand(0));
1809 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1814 Tmp1 = SelectExpr(N.getOperand(0));
1815 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1816 default: assert(0 && "unhandled result code");
1817 case 0: // No immediate
1818 Tmp2 = SelectExpr(N.getOperand(1));
1819 BuildMI(BB, PPC::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1821 case 1: // Low immediate
1822 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1824 case 2: // Shifted immediate
1825 BuildMI(BB, PPC::XORIS, 2, Result).addReg(Tmp1).addImm(Tmp2);
1832 if (!MVT::isInteger(DestType)) {
1833 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
1834 N.getOperand(0).Val->hasOneUse()) {
1835 ++FusedFP; // Statistic
1836 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1837 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1838 Tmp3 = SelectExpr(N.getOperand(1));
1839 Opc = DestType == MVT::f64 ? PPC::FMSUB : PPC::FMSUBS;
1840 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1843 if (!NoExcessFPPrecision && N.getOperand(1).getOpcode() == ISD::MUL &&
1844 N.getOperand(1).Val->hasOneUse()) {
1845 ++FusedFP; // Statistic
1846 Tmp1 = SelectExpr(N.getOperand(1).getOperand(0));
1847 Tmp2 = SelectExpr(N.getOperand(1).getOperand(1));
1848 Tmp3 = SelectExpr(N.getOperand(0));
1849 Opc = DestType == MVT::f64 ? PPC::FNMSUB : PPC::FNMSUBS;
1850 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1853 Opc = DestType == MVT::f64 ? PPC::FSUB : PPC::FSUBS;
1854 Tmp1 = SelectExpr(N.getOperand(0));
1855 Tmp2 = SelectExpr(N.getOperand(1));
1856 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1859 if (1 == getImmediateForOpcode(N.getOperand(0), opcode, Tmp1, true)) {
1860 Tmp2 = SelectExpr(N.getOperand(1));
1861 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp2).addSImm(Tmp1);
1862 } else if (1 == getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1863 Tmp1 = SelectExpr(N.getOperand(0));
1864 BuildMI(BB, PPC::ADDI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1866 Tmp1 = SelectExpr(N.getOperand(0));
1867 Tmp2 = SelectExpr(N.getOperand(1));
1868 BuildMI(BB, PPC::SUBF, 2, Result).addReg(Tmp2).addReg(Tmp1);
1873 Tmp1 = SelectExpr(N.getOperand(0));
1874 if (1 == getImmediateForOpcode(N.getOperand(1), opcode, Tmp2))
1875 BuildMI(BB, PPC::MULLI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1877 Tmp2 = SelectExpr(N.getOperand(1));
1879 default: assert(0 && "Unknown type to ISD::MUL"); break;
1880 case MVT::i32: Opc = PPC::MULLW; break;
1881 case MVT::f32: Opc = PPC::FMULS; break;
1882 case MVT::f64: Opc = PPC::FMUL; break;
1884 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1890 Tmp1 = SelectExpr(N.getOperand(0));
1891 Tmp2 = SelectExpr(N.getOperand(1));
1892 Opc = (ISD::MULHU == opcode) ? PPC::MULHWU : PPC::MULHW;
1893 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1898 switch (getImmediateForOpcode(N.getOperand(1), opcode, Tmp3)) {
1900 // If this is an sdiv by a power of two, we can use an srawi/addze pair.
1902 Tmp1 = MakeReg(MVT::i32);
1903 Tmp2 = SelectExpr(N.getOperand(0));
1904 if ((int)Tmp3 < 0) {
1905 unsigned Tmp4 = MakeReg(MVT::i32);
1906 BuildMI(BB, PPC::SRAWI, 2, Tmp1).addReg(Tmp2).addImm(-Tmp3);
1907 BuildMI(BB, PPC::ADDZE, 1, Tmp4).addReg(Tmp1);
1908 BuildMI(BB, PPC::NEG, 1, Result).addReg(Tmp4);
1910 BuildMI(BB, PPC::SRAWI, 2, Tmp1).addReg(Tmp2).addImm(Tmp3);
1911 BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp1);
1914 // If this is a divide by constant, we can emit code using some magic
1915 // constants to implement it as a multiply instead.
1918 if (opcode == ISD::SDIV)
1919 return SelectExpr(BuildSDIVSequence(N));
1921 return SelectExpr(BuildUDIVSequence(N));
1923 Tmp1 = SelectExpr(N.getOperand(0));
1924 Tmp2 = SelectExpr(N.getOperand(1));
1926 default: assert(0 && "Unknown type to ISD::SDIV"); break;
1927 case MVT::i32: Opc = (ISD::UDIV == opcode) ? PPC::DIVWU : PPC::DIVW; break;
1928 case MVT::f32: Opc = PPC::FDIVS; break;
1929 case MVT::f64: Opc = PPC::FDIV; break;
1931 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1934 case ISD::ADD_PARTS:
1935 case ISD::SUB_PARTS: {
1936 assert(N.getNumOperands() == 4 && N.getValueType() == MVT::i32 &&
1937 "Not an i64 add/sub!");
1938 // Emit all of the operands.
1939 std::vector<unsigned> InVals;
1940 for (unsigned i = 0, e = N.getNumOperands(); i != e; ++i)
1941 InVals.push_back(SelectExpr(N.getOperand(i)));
1942 if (N.getOpcode() == ISD::ADD_PARTS) {
1943 BuildMI(BB, PPC::ADDC, 2, Result).addReg(InVals[0]).addReg(InVals[2]);
1944 BuildMI(BB, PPC::ADDE, 2, Result+1).addReg(InVals[1]).addReg(InVals[3]);
1946 BuildMI(BB, PPC::SUBFC, 2, Result).addReg(InVals[2]).addReg(InVals[0]);
1947 BuildMI(BB, PPC::SUBFE, 2, Result+1).addReg(InVals[3]).addReg(InVals[1]);
1949 return Result+N.ResNo;
1952 case ISD::SHL_PARTS:
1953 case ISD::SRA_PARTS:
1954 case ISD::SRL_PARTS: {
1955 assert(N.getNumOperands() == 3 && N.getValueType() == MVT::i32 &&
1956 "Not an i64 shift!");
1957 unsigned ShiftOpLo = SelectExpr(N.getOperand(0));
1958 unsigned ShiftOpHi = SelectExpr(N.getOperand(1));
1959 unsigned SHReg = FoldIfWideZeroExtend(N.getOperand(2));
1960 Tmp1 = MakeReg(MVT::i32);
1961 Tmp2 = MakeReg(MVT::i32);
1962 Tmp3 = MakeReg(MVT::i32);
1963 unsigned Tmp4 = MakeReg(MVT::i32);
1964 unsigned Tmp5 = MakeReg(MVT::i32);
1965 unsigned Tmp6 = MakeReg(MVT::i32);
1966 BuildMI(BB, PPC::SUBFIC, 2, Tmp1).addReg(SHReg).addSImm(32);
1967 if (ISD::SHL_PARTS == opcode) {
1968 BuildMI(BB, PPC::SLW, 2, Tmp2).addReg(ShiftOpHi).addReg(SHReg);
1969 BuildMI(BB, PPC::SRW, 2, Tmp3).addReg(ShiftOpLo).addReg(Tmp1);
1970 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1971 BuildMI(BB, PPC::ADDI, 2, Tmp5).addReg(SHReg).addSImm(-32);
1972 BuildMI(BB, PPC::SLW, 2, Tmp6).addReg(ShiftOpLo).addReg(Tmp5);
1973 BuildMI(BB, PPC::OR, 2, Result+1).addReg(Tmp4).addReg(Tmp6);
1974 BuildMI(BB, PPC::SLW, 2, Result).addReg(ShiftOpLo).addReg(SHReg);
1975 } else if (ISD::SRL_PARTS == opcode) {
1976 BuildMI(BB, PPC::SRW, 2, Tmp2).addReg(ShiftOpLo).addReg(SHReg);
1977 BuildMI(BB, PPC::SLW, 2, Tmp3).addReg(ShiftOpHi).addReg(Tmp1);
1978 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1979 BuildMI(BB, PPC::ADDI, 2, Tmp5).addReg(SHReg).addSImm(-32);
1980 BuildMI(BB, PPC::SRW, 2, Tmp6).addReg(ShiftOpHi).addReg(Tmp5);
1981 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp4).addReg(Tmp6);
1982 BuildMI(BB, PPC::SRW, 2, Result+1).addReg(ShiftOpHi).addReg(SHReg);
1984 MachineBasicBlock *TmpMBB = new MachineBasicBlock(BB->getBasicBlock());
1985 MachineBasicBlock *PhiMBB = new MachineBasicBlock(BB->getBasicBlock());
1986 MachineBasicBlock *OldMBB = BB;
1987 MachineFunction *F = BB->getParent();
1988 ilist<MachineBasicBlock>::iterator It = BB; ++It;
1989 F->getBasicBlockList().insert(It, TmpMBB);
1990 F->getBasicBlockList().insert(It, PhiMBB);
1991 BB->addSuccessor(TmpMBB);
1992 BB->addSuccessor(PhiMBB);
1993 BuildMI(BB, PPC::SRW, 2, Tmp2).addReg(ShiftOpLo).addReg(SHReg);
1994 BuildMI(BB, PPC::SLW, 2, Tmp3).addReg(ShiftOpHi).addReg(Tmp1);
1995 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1996 BuildMI(BB, PPC::ADDICo, 2, Tmp5).addReg(SHReg).addSImm(-32);
1997 BuildMI(BB, PPC::SRAW, 2, Tmp6).addReg(ShiftOpHi).addReg(Tmp5);
1998 BuildMI(BB, PPC::SRAW, 2, Result+1).addReg(ShiftOpHi).addReg(SHReg);
1999 BuildMI(BB, PPC::BLE, 2).addReg(PPC::CR0).addMBB(PhiMBB);
2000 // Select correct least significant half if the shift amount > 32
2002 unsigned Tmp7 = MakeReg(MVT::i32);
2003 BuildMI(BB, PPC::OR, 2, Tmp7).addReg(Tmp6).addReg(Tmp6);
2004 TmpMBB->addSuccessor(PhiMBB);
2006 BuildMI(BB, PPC::PHI, 4, Result).addReg(Tmp4).addMBB(OldMBB)
2007 .addReg(Tmp7).addMBB(TmpMBB);
2009 return Result+N.ResNo;
2012 case ISD::FP_TO_UINT:
2013 case ISD::FP_TO_SINT: {
2014 bool U = (ISD::FP_TO_UINT == opcode);
2015 Tmp1 = SelectExpr(N.getOperand(0));
2017 Tmp2 = MakeReg(MVT::f64);
2018 BuildMI(BB, PPC::FCTIWZ, 1, Tmp2).addReg(Tmp1);
2019 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
2020 addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(Tmp2), FrameIdx);
2021 addFrameReference(BuildMI(BB, PPC::LWZ, 2, Result), FrameIdx, 4);
2024 unsigned Zero = getConstDouble(0.0);
2025 unsigned MaxInt = getConstDouble((1LL << 32) - 1);
2026 unsigned Border = getConstDouble(1LL << 31);
2027 unsigned UseZero = MakeReg(MVT::f64);
2028 unsigned UseMaxInt = MakeReg(MVT::f64);
2029 unsigned UseChoice = MakeReg(MVT::f64);
2030 unsigned TmpReg = MakeReg(MVT::f64);
2031 unsigned TmpReg2 = MakeReg(MVT::f64);
2032 unsigned ConvReg = MakeReg(MVT::f64);
2033 unsigned IntTmp = MakeReg(MVT::i32);
2034 unsigned XorReg = MakeReg(MVT::i32);
2035 MachineFunction *F = BB->getParent();
2036 int FrameIdx = F->getFrameInfo()->CreateStackObject(8, 8);
2037 // Update machine-CFG edges
2038 MachineBasicBlock *XorMBB = new MachineBasicBlock(BB->getBasicBlock());
2039 MachineBasicBlock *PhiMBB = new MachineBasicBlock(BB->getBasicBlock());
2040 MachineBasicBlock *OldMBB = BB;
2041 ilist<MachineBasicBlock>::iterator It = BB; ++It;
2042 F->getBasicBlockList().insert(It, XorMBB);
2043 F->getBasicBlockList().insert(It, PhiMBB);
2044 BB->addSuccessor(XorMBB);
2045 BB->addSuccessor(PhiMBB);
2046 // Convert from floating point to unsigned 32-bit value
2047 // Use 0 if incoming value is < 0.0
2048 BuildMI(BB, PPC::FSEL, 3, UseZero).addReg(Tmp1).addReg(Tmp1).addReg(Zero);
2049 // Use 2**32 - 1 if incoming value is >= 2**32
2050 BuildMI(BB, PPC::FSUB, 2, UseMaxInt).addReg(MaxInt).addReg(Tmp1);
2051 BuildMI(BB, PPC::FSEL, 3, UseChoice).addReg(UseMaxInt).addReg(UseZero)
2054 BuildMI(BB, PPC::FSUB, 2, TmpReg).addReg(UseChoice).addReg(Border);
2055 // Use difference if >= 2**31
2056 BuildMI(BB, PPC::FCMPU, 2, PPC::CR0).addReg(UseChoice).addReg(Border);
2057 BuildMI(BB, PPC::FSEL, 3, TmpReg2).addReg(TmpReg).addReg(TmpReg)
2059 // Convert to integer
2060 BuildMI(BB, PPC::FCTIWZ, 1, ConvReg).addReg(TmpReg2);
2061 addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(ConvReg), FrameIdx);
2062 addFrameReference(BuildMI(BB, PPC::LWZ, 2, IntTmp), FrameIdx, 4);
2063 BuildMI(BB, PPC::BLT, 2).addReg(PPC::CR0).addMBB(PhiMBB);
2064 BuildMI(BB, PPC::B, 1).addMBB(XorMBB);
2067 // add 2**31 if input was >= 2**31
2069 BuildMI(BB, PPC::XORIS, 2, XorReg).addReg(IntTmp).addImm(0x8000);
2070 XorMBB->addSuccessor(PhiMBB);
2073 // DestReg = phi [ IntTmp, OldMBB ], [ XorReg, XorMBB ]
2075 BuildMI(BB, PPC::PHI, 4, Result).addReg(IntTmp).addMBB(OldMBB)
2076 .addReg(XorReg).addMBB(XorMBB);
2079 assert(0 && "Should never get here");
2084 if (SetCCSDNode *SetCC = dyn_cast<SetCCSDNode>(Node)) {
2085 if (ConstantSDNode *CN =
2086 dyn_cast<ConstantSDNode>(SetCC->getOperand(1).Val)) {
2087 // We can codegen setcc op, imm very efficiently compared to a brcond.
2088 // Check for those cases here.
2090 if (CN->getValue() == 0) {
2091 Tmp1 = SelectExpr(SetCC->getOperand(0));
2092 switch (SetCC->getCondition()) {
2093 default: SetCC->dump(); assert(0 && "Unhandled SetCC condition"); abort();
2095 Tmp2 = MakeReg(MVT::i32);
2096 BuildMI(BB, PPC::CNTLZW, 1, Tmp2).addReg(Tmp1);
2097 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp2).addImm(27)
2098 .addImm(5).addImm(31);
2101 Tmp2 = MakeReg(MVT::i32);
2102 BuildMI(BB, PPC::ADDIC, 2, Tmp2).addReg(Tmp1).addSImm(-1);
2103 BuildMI(BB, PPC::SUBFE, 2, Result).addReg(Tmp2).addReg(Tmp1);
2106 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(1)
2107 .addImm(31).addImm(31);
2110 Tmp2 = MakeReg(MVT::i32);
2111 Tmp3 = MakeReg(MVT::i32);
2112 BuildMI(BB, PPC::NEG, 2, Tmp2).addReg(Tmp1);
2113 BuildMI(BB, PPC::ANDC, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2114 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp3).addImm(1)
2115 .addImm(31).addImm(31);
2121 if (CN->isAllOnesValue()) {
2122 Tmp1 = SelectExpr(SetCC->getOperand(0));
2123 switch (SetCC->getCondition()) {
2124 default: assert(0 && "Unhandled SetCC condition"); abort();
2126 Tmp2 = MakeReg(MVT::i32);
2127 Tmp3 = MakeReg(MVT::i32);
2128 BuildMI(BB, PPC::ADDIC, 2, Tmp2).addReg(Tmp1).addSImm(1);
2129 BuildMI(BB, PPC::LI, 1, Tmp3).addSImm(0);
2130 BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp3);
2133 Tmp2 = MakeReg(MVT::i32);
2134 Tmp3 = MakeReg(MVT::i32);
2135 BuildMI(BB, PPC::NOR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
2136 BuildMI(BB, PPC::ADDIC, 2, Tmp3).addReg(Tmp2).addSImm(-1);
2137 BuildMI(BB, PPC::SUBFE, 2, Result).addReg(Tmp3).addReg(Tmp2);
2140 Tmp2 = MakeReg(MVT::i32);
2141 Tmp3 = MakeReg(MVT::i32);
2142 BuildMI(BB, PPC::ADDI, 2, Tmp2).addReg(Tmp1).addSImm(1);
2143 BuildMI(BB, PPC::AND, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2144 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp3).addImm(1)
2145 .addImm(31).addImm(31);
2148 Tmp2 = MakeReg(MVT::i32);
2149 BuildMI(BB, PPC::RLWINM, 4, Tmp2).addReg(Tmp1).addImm(1)
2150 .addImm(31).addImm(31);
2151 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp2).addImm(1);
2159 unsigned CCReg = SelectCC(N, Opc, Inv, Tmp2);
2160 MoveCRtoGPR(CCReg, Inv, Tmp2, Result);
2163 assert(0 && "Is this legal?");
2167 SetCCSDNode* SetCC = dyn_cast<SetCCSDNode>(N.getOperand(0).Val);
2168 if (SetCC && N.getOperand(0).getOpcode() == ISD::SETCC &&
2169 !MVT::isInteger(SetCC->getOperand(0).getValueType()) &&
2170 !MVT::isInteger(N.getOperand(1).getValueType()) &&
2171 !MVT::isInteger(N.getOperand(2).getValueType()) &&
2172 SetCC->getCondition() != ISD::SETEQ &&
2173 SetCC->getCondition() != ISD::SETNE) {
2174 MVT::ValueType VT = SetCC->getOperand(0).getValueType();
2175 unsigned TV = SelectExpr(N.getOperand(1)); // Use if TRUE
2176 unsigned FV = SelectExpr(N.getOperand(2)); // Use if FALSE
2178 ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(SetCC->getOperand(1));
2179 if (CN && (CN->isExactlyValue(-0.0) || CN->isExactlyValue(0.0))) {
2180 switch(SetCC->getCondition()) {
2181 default: assert(0 && "Invalid FSEL condition"); abort();
2184 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2187 Tmp1 = SelectExpr(SetCC->getOperand(0)); // Val to compare against
2188 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp1).addReg(TV).addReg(FV);
2192 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2195 if (SetCC->getOperand(0).getOpcode() == ISD::FNEG) {
2196 Tmp2 = SelectExpr(SetCC->getOperand(0).getOperand(0));
2199 Tmp1 = SelectExpr(SetCC->getOperand(0)); // Val to compare against
2200 BuildMI(BB, PPC::FNEG, 1, Tmp2).addReg(Tmp1);
2202 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp2).addReg(TV).addReg(FV);
2207 Opc = (MVT::f64 == VT) ? PPC::FSUB : PPC::FSUBS;
2208 Tmp1 = SelectExpr(SetCC->getOperand(0)); // Val to compare against
2209 Tmp2 = SelectExpr(SetCC->getOperand(1));
2211 switch(SetCC->getCondition()) {
2212 default: assert(0 && "Invalid FSEL condition"); abort();
2215 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
2216 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
2220 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
2221 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
2225 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2226 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
2230 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2231 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
2235 assert(0 && "Should never get here");
2240 unsigned TrueValue = SelectExpr(N.getOperand(1)); //Use if TRUE
2241 unsigned FalseValue = SelectExpr(N.getOperand(2)); //Use if FALSE
2242 unsigned CCReg = SelectCC(N.getOperand(0), Opc, Inv, Tmp3);
2244 // Create an iterator with which to insert the MBB for copying the false
2245 // value and the MBB to hold the PHI instruction for this SetCC.
2246 MachineBasicBlock *thisMBB = BB;
2247 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2248 ilist<MachineBasicBlock>::iterator It = BB;
2254 // cmpTY ccX, r1, r2
2256 // fallthrough --> copy0MBB
2257 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
2258 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
2259 BuildMI(BB, Opc, 2).addReg(CCReg).addMBB(sinkMBB);
2260 MachineFunction *F = BB->getParent();
2261 F->getBasicBlockList().insert(It, copy0MBB);
2262 F->getBasicBlockList().insert(It, sinkMBB);
2263 // Update machine-CFG edges
2264 BB->addSuccessor(copy0MBB);
2265 BB->addSuccessor(sinkMBB);
2268 // %FalseValue = ...
2269 // # fallthrough to sinkMBB
2271 // Update machine-CFG edges
2272 BB->addSuccessor(sinkMBB);
2275 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2278 BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue)
2279 .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB);
2284 switch (N.getValueType()) {
2285 default: assert(0 && "Cannot use constants of this type!");
2287 BuildMI(BB, PPC::LI, 1, Result)
2288 .addSImm(!cast<ConstantSDNode>(N)->isNullValue());
2292 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
2293 if (v < 32768 && v >= -32768) {
2294 BuildMI(BB, PPC::LI, 1, Result).addSImm(v);
2296 Tmp1 = MakeReg(MVT::i32);
2297 BuildMI(BB, PPC::LIS, 1, Tmp1).addSImm(v >> 16);
2298 BuildMI(BB, PPC::ORI, 2, Result).addReg(Tmp1).addImm(v & 0xFFFF);
2304 case ISD::ConstantFP: {
2305 ConstantFPSDNode *CN = cast<ConstantFPSDNode>(N);
2306 Result = getConstDouble(CN->getValue(), Result);
2311 if (!NoExcessFPPrecision &&
2312 ISD::ADD == N.getOperand(0).getOpcode() &&
2313 N.getOperand(0).Val->hasOneUse() &&
2314 ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() &&
2315 N.getOperand(0).getOperand(0).Val->hasOneUse()) {
2316 ++FusedFP; // Statistic
2317 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0));
2318 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1));
2319 Tmp3 = SelectExpr(N.getOperand(0).getOperand(1));
2320 Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
2321 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
2322 } else if (!NoExcessFPPrecision &&
2323 ISD::ADD == N.getOperand(0).getOpcode() &&
2324 N.getOperand(0).Val->hasOneUse() &&
2325 ISD::MUL == N.getOperand(0).getOperand(1).getOpcode() &&
2326 N.getOperand(0).getOperand(1).Val->hasOneUse()) {
2327 ++FusedFP; // Statistic
2328 Tmp1 = SelectExpr(N.getOperand(0).getOperand(1).getOperand(0));
2329 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1).getOperand(1));
2330 Tmp3 = SelectExpr(N.getOperand(0).getOperand(0));
2331 Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
2332 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
2333 } else if (ISD::FABS == N.getOperand(0).getOpcode()) {
2334 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
2335 BuildMI(BB, PPC::FNABS, 1, Result).addReg(Tmp1);
2337 Tmp1 = SelectExpr(N.getOperand(0));
2338 BuildMI(BB, PPC::FNEG, 1, Result).addReg(Tmp1);
2343 Tmp1 = SelectExpr(N.getOperand(0));
2344 BuildMI(BB, PPC::FABS, 1, Result).addReg(Tmp1);
2348 Tmp1 = SelectExpr(N.getOperand(0));
2349 Opc = DestType == MVT::f64 ? PPC::FSQRT : PPC::FSQRTS;
2350 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
2354 assert (DestType == MVT::f32 &&
2355 N.getOperand(0).getValueType() == MVT::f64 &&
2356 "only f64 to f32 conversion supported here");
2357 Tmp1 = SelectExpr(N.getOperand(0));
2358 BuildMI(BB, PPC::FRSP, 1, Result).addReg(Tmp1);
2361 case ISD::FP_EXTEND:
2362 assert (DestType == MVT::f64 &&
2363 N.getOperand(0).getValueType() == MVT::f32 &&
2364 "only f32 to f64 conversion supported here");
2365 Tmp1 = SelectExpr(N.getOperand(0));
2366 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
2369 case ISD::UINT_TO_FP:
2370 case ISD::SINT_TO_FP: {
2371 assert (N.getOperand(0).getValueType() == MVT::i32
2372 && "int to float must operate on i32");
2373 bool IsUnsigned = (ISD::UINT_TO_FP == opcode);
2374 Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
2375 Tmp2 = MakeReg(MVT::f64); // temp reg to load the integer value into
2376 Tmp3 = MakeReg(MVT::i32); // temp reg to hold the conversion constant
2378 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
2379 MachineConstantPool *CP = BB->getParent()->getConstantPool();
2382 unsigned ConstF = getConstDouble(0x1.000000p52);
2383 // Store the hi & low halves of the fp value, currently in int regs
2384 BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330);
2385 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx);
2386 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp1), FrameIdx, 4);
2387 addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
2388 // Generate the return value with a subtract
2389 BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF);
2391 unsigned ConstF = getConstDouble(0x1.000008p52);
2392 unsigned TmpL = MakeReg(MVT::i32);
2393 // Store the hi & low halves of the fp value, currently in int regs
2394 BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330);
2395 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx);
2396 BuildMI(BB, PPC::XORIS, 2, TmpL).addReg(Tmp1).addImm(0x8000);
2397 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(TmpL), FrameIdx, 4);
2398 addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
2399 // Generate the return value with a subtract
2400 BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF);
2408 void ISel::Select(SDOperand N) {
2409 unsigned Tmp1, Tmp2, Tmp3, Opc;
2410 unsigned opcode = N.getOpcode();
2412 if (!ExprMap.insert(std::make_pair(N, 1)).second)
2413 return; // Already selected.
2415 SDNode *Node = N.Val;
2417 switch (Node->getOpcode()) {
2419 Node->dump(); std::cerr << "\n";
2420 assert(0 && "Node not handled yet!");
2421 case ISD::EntryToken: return; // Noop
2422 case ISD::TokenFactor:
2423 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
2424 Select(Node->getOperand(i));
2426 case ISD::CALLSEQ_START:
2427 case ISD::CALLSEQ_END:
2428 Select(N.getOperand(0));
2429 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
2430 Opc = N.getOpcode() == ISD::CALLSEQ_START ? PPC::ADJCALLSTACKDOWN :
2431 PPC::ADJCALLSTACKUP;
2432 BuildMI(BB, Opc, 1).addImm(Tmp1);
2435 MachineBasicBlock *Dest =
2436 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
2437 Select(N.getOperand(0));
2438 BuildMI(BB, PPC::B, 1).addMBB(Dest);
2442 case ISD::BRCONDTWOWAY:
2445 case ISD::CopyToReg:
2446 Select(N.getOperand(0));
2447 Tmp1 = SelectExpr(N.getOperand(1));
2448 Tmp2 = cast<RegSDNode>(N)->getReg();
2451 if (N.getOperand(1).getValueType() == MVT::f64 ||
2452 N.getOperand(1).getValueType() == MVT::f32)
2453 BuildMI(BB, PPC::FMR, 1, Tmp2).addReg(Tmp1);
2455 BuildMI(BB, PPC::OR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
2458 case ISD::ImplicitDef:
2459 Select(N.getOperand(0));
2460 BuildMI(BB, PPC::IMPLICIT_DEF, 0, cast<RegSDNode>(N)->getReg());
2463 switch (N.getNumOperands()) {
2465 assert(0 && "Unknown return instruction!");
2467 assert(N.getOperand(1).getValueType() == MVT::i32 &&
2468 N.getOperand(2).getValueType() == MVT::i32 &&
2469 "Unknown two-register value!");
2470 Select(N.getOperand(0));
2471 Tmp1 = SelectExpr(N.getOperand(1));
2472 Tmp2 = SelectExpr(N.getOperand(2));
2473 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp2).addReg(Tmp2);
2474 BuildMI(BB, PPC::OR, 2, PPC::R4).addReg(Tmp1).addReg(Tmp1);
2477 Select(N.getOperand(0));
2478 Tmp1 = SelectExpr(N.getOperand(1));
2479 switch (N.getOperand(1).getValueType()) {
2481 assert(0 && "Unknown return type!");
2484 BuildMI(BB, PPC::FMR, 1, PPC::F1).addReg(Tmp1);
2487 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp1).addReg(Tmp1);
2491 Select(N.getOperand(0));
2494 BuildMI(BB, PPC::BLR, 0); // Just emit a 'ret' instruction
2496 case ISD::TRUNCSTORE:
2498 SDOperand Chain = N.getOperand(0);
2499 SDOperand Value = N.getOperand(1);
2500 SDOperand Address = N.getOperand(2);
2503 Tmp1 = SelectExpr(Value); //value
2505 if (opcode == ISD::STORE) {
2506 switch(Value.getValueType()) {
2507 default: assert(0 && "unknown Type in store");
2508 case MVT::i32: Opc = PPC::STW; break;
2509 case MVT::f64: Opc = PPC::STFD; break;
2510 case MVT::f32: Opc = PPC::STFS; break;
2512 } else { //ISD::TRUNCSTORE
2513 switch(cast<VTSDNode>(Node->getOperand(4))->getVT()) {
2514 default: assert(0 && "unknown Type in store");
2516 case MVT::i8: Opc = PPC::STB; break;
2517 case MVT::i16: Opc = PPC::STH; break;
2521 if(Address.getOpcode() == ISD::FrameIndex) {
2522 Tmp2 = cast<FrameIndexSDNode>(Address)->getIndex();
2523 addFrameReference(BuildMI(BB, Opc, 3).addReg(Tmp1), (int)Tmp2);
2524 } else if(GlobalAddressSDNode *GN = dyn_cast<GlobalAddressSDNode>(Address)){
2525 GlobalValue *GV = GN->getGlobal();
2526 Tmp2 = MakeReg(MVT::i32);
2528 BuildMI(BB, PPC::ADDIS, 2, Tmp2).addReg(getGlobalBaseReg())
2529 .addGlobalAddress(GV);
2531 BuildMI(BB, PPC::LIS, 1, Tmp2).addGlobalAddress(GV);
2532 if (GV->hasWeakLinkage() || GV->isExternal()) {
2533 Tmp3 = MakeReg(MVT::i32);
2534 BuildMI(BB, PPC::LWZ, 2, Tmp3).addGlobalAddress(GV).addReg(Tmp2);
2535 BuildMI(BB, Opc, 3).addReg(Tmp1).addSImm(0).addReg(Tmp3);
2537 BuildMI(BB, Opc, 3).addReg(Tmp1).addGlobalAddress(GV).addReg(Tmp2);
2541 switch(SelectAddr(Address, Tmp2, offset)) {
2542 default: assert(0 && "Unhandled return value from SelectAddr");
2543 case 0: // imm offset, no frame, no index
2544 BuildMI(BB, Opc, 3).addReg(Tmp1).addSImm(offset).addReg(Tmp2);
2546 case 1: // imm offset + frame index
2547 addFrameReference(BuildMI(BB, Opc, 3).addReg(Tmp1), (int)Tmp2, offset);
2549 case 2: // base+index addressing
2550 Opc = IndexedOpForOp(Opc);
2551 BuildMI(BB, Opc, 3).addReg(Tmp1).addReg(Tmp2).addReg(offset);
2561 case ISD::CopyFromReg:
2564 case ISD::DYNAMIC_STACKALLOC:
2569 assert(0 && "Should not be reached!");
2573 /// createPPC32PatternInstructionSelector - This pass converts an LLVM function
2574 /// into a machine code representation using pattern matching and a machine
2575 /// description file.
2577 FunctionPass *llvm::createPPC32ISelPattern(TargetMachine &TM) {
2578 return new ISel(TM);