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"
38 //===----------------------------------------------------------------------===//
39 // PPC32TargetLowering - PPC32 Implementation of the TargetLowering interface
41 class PPC32TargetLowering : public TargetLowering {
42 int VarArgsFrameIndex; // FrameIndex for start of varargs area.
43 int ReturnAddrIndex; // FrameIndex for return slot.
45 PPC32TargetLowering(TargetMachine &TM) : TargetLowering(TM) {
46 // Fold away setcc operations if possible.
47 setSetCCIsExpensive();
49 // Set up the register classes.
50 addRegisterClass(MVT::i32, PPC32::GPRCRegisterClass);
51 addRegisterClass(MVT::f32, PPC32::FPRCRegisterClass);
52 addRegisterClass(MVT::f64, PPC32::FPRCRegisterClass);
54 // PowerPC has no intrinsics for these particular operations
55 setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
56 setOperationAction(ISD::MEMSET, MVT::Other, Expand);
57 setOperationAction(ISD::MEMCPY, MVT::Other, Expand);
59 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
60 setOperationAction(ISD::SEXTLOAD, MVT::i1, Expand);
61 setOperationAction(ISD::SEXTLOAD, MVT::i8, Expand);
63 // PowerPC has no SREM/UREM instructions
64 setOperationAction(ISD::SREM, MVT::i32, Expand);
65 setOperationAction(ISD::UREM, MVT::i32, Expand);
67 // We don't support sin/cos/sqrt/fmod
68 setOperationAction(ISD::FSIN , MVT::f64, Expand);
69 setOperationAction(ISD::FCOS , MVT::f64, Expand);
70 setOperationAction(ISD::SREM , MVT::f64, Expand);
71 setOperationAction(ISD::FSIN , MVT::f32, Expand);
72 setOperationAction(ISD::FCOS , MVT::f32, Expand);
73 setOperationAction(ISD::SREM , MVT::f32, Expand);
75 // If we're enabling GP optimizations, use hardware square root
77 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
78 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
81 //PowerPC does not have CTPOP or CTTZ
82 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
83 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
85 setSetCCResultContents(ZeroOrOneSetCCResult);
86 addLegalFPImmediate(+0.0); // Necessary for FSEL
87 addLegalFPImmediate(-0.0); //
89 computeRegisterProperties();
92 /// LowerArguments - This hook must be implemented to indicate how we should
93 /// lower the arguments for the specified function, into the specified DAG.
94 virtual std::vector<SDOperand>
95 LowerArguments(Function &F, SelectionDAG &DAG);
97 /// LowerCallTo - This hook lowers an abstract call to a function into an
99 virtual std::pair<SDOperand, SDOperand>
100 LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, unsigned CC,
101 bool isTailCall, SDOperand Callee, ArgListTy &Args,
104 virtual SDOperand LowerVAStart(SDOperand Chain, SDOperand VAListP,
105 Value *VAListV, SelectionDAG &DAG);
107 virtual std::pair<SDOperand,SDOperand>
108 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
109 const Type *ArgTy, SelectionDAG &DAG);
111 virtual std::pair<SDOperand, SDOperand>
112 LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
118 std::vector<SDOperand>
119 PPC32TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
121 // add beautiful description of PPC stack frame format, or at least some docs
123 MachineFunction &MF = DAG.getMachineFunction();
124 MachineFrameInfo *MFI = MF.getFrameInfo();
125 MachineBasicBlock& BB = MF.front();
126 std::vector<SDOperand> ArgValues;
128 // Due to the rather complicated nature of the PowerPC ABI, rather than a
129 // fixed size array of physical args, for the sake of simplicity let the STL
130 // handle tracking them for us.
131 std::vector<unsigned> argVR, argPR, argOp;
132 unsigned ArgOffset = 24;
133 unsigned GPR_remaining = 8;
134 unsigned FPR_remaining = 13;
135 unsigned GPR_idx = 0, FPR_idx = 0;
136 static const unsigned GPR[] = {
137 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
138 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
140 static const unsigned FPR[] = {
141 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
142 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
145 // Add DAG nodes to load the arguments... On entry to a function on PPC,
146 // the arguments start at offset 24, although they are likely to be passed
148 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
149 SDOperand newroot, argt;
151 bool needsLoad = false;
152 bool ArgLive = !I->use_empty();
153 MVT::ValueType ObjectVT = getValueType(I->getType());
156 default: assert(0 && "Unhandled argument type!");
163 if (GPR_remaining > 0) {
164 MF.addLiveIn(GPR[GPR_idx]);
165 argt = newroot = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32,
167 if (ObjectVT != MVT::i32)
168 argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, newroot);
173 case MVT::i64: ObjSize = 8;
175 if (GPR_remaining > 0) {
176 SDOperand argHi, argLo;
177 MF.addLiveIn(GPR[GPR_idx]);
178 argHi = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot());
179 // If we have two or more remaining argument registers, then both halves
180 // of the i64 can be sourced from there. Otherwise, the lower half will
181 // have to come off the stack. This can happen when an i64 is preceded
182 // by 28 bytes of arguments.
183 if (GPR_remaining > 1) {
184 MF.addLiveIn(GPR[GPR_idx+1]);
185 argLo = DAG.getCopyFromReg(GPR[GPR_idx+1], MVT::i32, argHi);
187 int FI = MFI->CreateFixedObject(4, ArgOffset+4);
188 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
189 argLo = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
190 DAG.getSrcValue(NULL));
192 // Build the outgoing arg thingy
193 argt = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, argLo, argHi);
201 ObjSize = (ObjectVT == MVT::f64) ? 8 : 4;
203 if (FPR_remaining > 0) {
204 MF.addLiveIn(FPR[FPR_idx]);
205 argt = newroot = DAG.getCopyFromReg(FPR[FPR_idx], ObjectVT,
215 // We need to load the argument to a virtual register if we determined above
216 // that we ran out of physical registers of the appropriate type
218 unsigned SubregOffset = 0;
219 if (ObjectVT == MVT::i8 || ObjectVT == MVT::i1) SubregOffset = 3;
220 if (ObjectVT == MVT::i16) SubregOffset = 2;
221 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
222 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
223 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN,
224 DAG.getConstant(SubregOffset, MVT::i32));
225 argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
226 DAG.getSrcValue(NULL));
229 // Every 4 bytes of argument space consumes one of the GPRs available for
231 if (GPR_remaining > 0) {
232 unsigned delta = (GPR_remaining > 1 && ObjSize == 8) ? 2 : 1;
233 GPR_remaining -= delta;
236 ArgOffset += ObjSize;
238 DAG.setRoot(newroot.getValue(1));
240 ArgValues.push_back(argt);
243 // If the function takes variable number of arguments, make a frame index for
244 // the start of the first vararg value... for expansion of llvm.va_start.
246 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
247 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
248 // If this function is vararg, store any remaining integer argument regs
249 // to their spots on the stack so that they may be loaded by deferencing the
250 // result of va_next.
251 std::vector<SDOperand> MemOps;
252 for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) {
253 MF.addLiveIn(GPR[GPR_idx]);
254 SDOperand Val = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot());
255 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
256 Val, FIN, DAG.getSrcValue(NULL));
257 MemOps.push_back(Store);
258 // Increment the address by four for the next argument to store
259 SDOperand PtrOff = DAG.getConstant(4, getPointerTy());
260 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff);
262 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps));
265 // Finally, inform the code generator which regs we return values in.
266 switch (getValueType(F.getReturnType())) {
267 default: assert(0 && "Unknown type!");
268 case MVT::isVoid: break;
273 MF.addLiveOut(PPC::R3);
276 MF.addLiveOut(PPC::R3);
277 MF.addLiveOut(PPC::R4);
281 MF.addLiveOut(PPC::F1);
288 std::pair<SDOperand, SDOperand>
289 PPC32TargetLowering::LowerCallTo(SDOperand Chain,
290 const Type *RetTy, bool isVarArg,
291 unsigned CallingConv, bool isTailCall,
292 SDOperand Callee, ArgListTy &Args,
294 // args_to_use will accumulate outgoing args for the ISD::CALL case in
295 // SelectExpr to use to put the arguments in the appropriate registers.
296 std::vector<SDOperand> args_to_use;
298 // Count how many bytes are to be pushed on the stack, including the linkage
299 // area, and parameter passing area.
300 unsigned NumBytes = 24;
303 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
304 DAG.getConstant(NumBytes, getPointerTy()));
306 for (unsigned i = 0, e = Args.size(); i != e; ++i)
307 switch (getValueType(Args[i].second)) {
308 default: assert(0 && "Unknown value type!");
322 // Just to be safe, we'll always reserve the full 24 bytes of linkage area
323 // plus 32 bytes of argument space in case any called code gets funky on us.
324 if (NumBytes < 56) NumBytes = 56;
326 // Adjust the stack pointer for the new arguments...
327 // These operations are automatically eliminated by the prolog/epilog pass
328 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
329 DAG.getConstant(NumBytes, getPointerTy()));
331 // Set up a copy of the stack pointer for use loading and storing any
332 // arguments that may not fit in the registers available for argument
334 SDOperand StackPtr = DAG.getCopyFromReg(PPC::R1, MVT::i32,
337 // Figure out which arguments are going to go in registers, and which in
338 // memory. Also, if this is a vararg function, floating point operations
339 // must be stored to our stack, and loaded into integer regs as well, if
340 // any integer regs are available for argument passing.
341 unsigned ArgOffset = 24;
342 unsigned GPR_remaining = 8;
343 unsigned FPR_remaining = 13;
345 std::vector<SDOperand> MemOps;
346 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
347 // PtrOff will be used to store the current argument to the stack if a
348 // register cannot be found for it.
349 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
350 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
351 MVT::ValueType ArgVT = getValueType(Args[i].second);
354 default: assert(0 && "Unexpected ValueType for argument!");
358 // Promote the integer to 32 bits. If the input type is signed use a
359 // sign extend, otherwise use a zero extend.
360 if (Args[i].second->isSigned())
361 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
363 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
366 if (GPR_remaining > 0) {
367 args_to_use.push_back(Args[i].first);
370 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
371 Args[i].first, PtrOff,
372 DAG.getSrcValue(NULL)));
377 // If we have one free GPR left, we can place the upper half of the i64
378 // in it, and store the other half to the stack. If we have two or more
379 // free GPRs, then we can pass both halves of the i64 in registers.
380 if (GPR_remaining > 0) {
381 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
382 Args[i].first, DAG.getConstant(1, MVT::i32));
383 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
384 Args[i].first, DAG.getConstant(0, MVT::i32));
385 args_to_use.push_back(Hi);
387 if (GPR_remaining > 0) {
388 args_to_use.push_back(Lo);
391 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
392 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
393 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
394 Lo, PtrOff, DAG.getSrcValue(NULL)));
397 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
398 Args[i].first, PtrOff,
399 DAG.getSrcValue(NULL)));
405 if (FPR_remaining > 0) {
406 args_to_use.push_back(Args[i].first);
409 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
410 Args[i].first, PtrOff,
411 DAG.getSrcValue(NULL));
412 MemOps.push_back(Store);
413 // Float varargs are always shadowed in available integer registers
414 if (GPR_remaining > 0) {
415 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
416 DAG.getSrcValue(NULL));
417 MemOps.push_back(Load);
418 args_to_use.push_back(Load);
421 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
422 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
423 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
424 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
425 DAG.getSrcValue(NULL));
426 MemOps.push_back(Load);
427 args_to_use.push_back(Load);
431 // If we have any FPRs remaining, we may also have GPRs remaining.
432 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
434 if (GPR_remaining > 0) {
435 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
438 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
439 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
444 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
445 Args[i].first, PtrOff,
446 DAG.getSrcValue(NULL)));
448 ArgOffset += (ArgVT == MVT::f32) ? 4 : 8;
453 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
456 std::vector<MVT::ValueType> RetVals;
457 MVT::ValueType RetTyVT = getValueType(RetTy);
458 if (RetTyVT != MVT::isVoid)
459 RetVals.push_back(RetTyVT);
460 RetVals.push_back(MVT::Other);
462 SDOperand TheCall = SDOperand(DAG.getCall(RetVals,
463 Chain, Callee, args_to_use), 0);
464 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
465 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
466 DAG.getConstant(NumBytes, getPointerTy()));
467 return std::make_pair(TheCall, Chain);
470 SDOperand PPC32TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
471 Value *VAListV, SelectionDAG &DAG) {
472 // vastart just stores the address of the VarArgsFrameIndex slot into the
473 // memory location argument.
474 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
475 return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR, VAListP,
476 DAG.getSrcValue(VAListV));
479 std::pair<SDOperand,SDOperand>
480 PPC32TargetLowering::LowerVAArg(SDOperand Chain,
481 SDOperand VAListP, Value *VAListV,
482 const Type *ArgTy, SelectionDAG &DAG) {
483 MVT::ValueType ArgVT = getValueType(ArgTy);
486 DAG.getLoad(MVT::i32, Chain, VAListP, DAG.getSrcValue(VAListV));
487 SDOperand Result = DAG.getLoad(ArgVT, Chain, VAList, DAG.getSrcValue(NULL));
489 if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
492 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
493 "Other types should have been promoted for varargs!");
496 VAList = DAG.getNode(ISD::ADD, VAList.getValueType(), VAList,
497 DAG.getConstant(Amt, VAList.getValueType()));
498 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
499 VAList, VAListP, DAG.getSrcValue(VAListV));
500 return std::make_pair(Result, Chain);
504 std::pair<SDOperand, SDOperand> PPC32TargetLowering::
505 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
507 assert(0 && "LowerFrameReturnAddress unimplemented");
512 Statistic<>Recorded("ppc-codegen", "Number of recording ops emitted");
513 Statistic<>FusedFP("ppc-codegen", "Number of fused fp operations");
514 Statistic<>MultiBranch("ppc-codegen", "Number of setcc logical ops collapsed");
515 //===--------------------------------------------------------------------===//
516 /// ISel - PPC32 specific code to select PPC32 machine instructions for
517 /// SelectionDAG operations.
518 //===--------------------------------------------------------------------===//
519 class ISel : public SelectionDAGISel {
520 PPC32TargetLowering PPC32Lowering;
521 SelectionDAG *ISelDAG; // Hack to support us having a dag->dag transform
522 // for sdiv and udiv until it is put into the future
525 /// ExprMap - As shared expressions are codegen'd, we keep track of which
526 /// vreg the value is produced in, so we only emit one copy of each compiled
528 std::map<SDOperand, unsigned> ExprMap;
530 unsigned GlobalBaseReg;
531 bool GlobalBaseInitialized;
534 ISel(TargetMachine &TM) : SelectionDAGISel(PPC32Lowering), PPC32Lowering(TM),
537 /// runOnFunction - Override this function in order to reset our per-function
539 virtual bool runOnFunction(Function &Fn) {
540 // Make sure we re-emit a set of the global base reg if necessary
541 GlobalBaseInitialized = false;
542 return SelectionDAGISel::runOnFunction(Fn);
545 /// InstructionSelectBasicBlock - This callback is invoked by
546 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
547 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG) {
549 // Codegen the basic block.
551 Select(DAG.getRoot());
553 // Clear state used for selection.
558 // dag -> dag expanders for integer divide by constant
559 SDOperand BuildSDIVSequence(SDOperand N);
560 SDOperand BuildUDIVSequence(SDOperand N);
562 unsigned getGlobalBaseReg();
563 unsigned getConstDouble(double floatVal, unsigned Result);
564 void MoveCRtoGPR(unsigned CCReg, bool Inv, unsigned Idx, unsigned Result);
565 bool SelectBitfieldInsert(SDOperand OR, unsigned Result);
566 unsigned FoldIfWideZeroExtend(SDOperand N);
567 unsigned SelectCC(SDOperand CC, unsigned &Opc, bool &Inv, unsigned &Idx);
568 unsigned SelectCCExpr(SDOperand N, unsigned& Opc, bool &Inv, unsigned &Idx);
569 unsigned SelectExpr(SDOperand N, bool Recording=false);
570 void Select(SDOperand N);
572 bool SelectAddr(SDOperand N, unsigned& Reg, int& offset);
573 void SelectBranchCC(SDOperand N);
576 /// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
577 /// returns zero when the input is not exactly a power of two.
578 static unsigned ExactLog2(unsigned Val) {
579 if (Val == 0 || (Val & (Val-1))) return 0;
588 // IsRunOfOnes - returns true if Val consists of one contiguous run of 1's with
589 // any number of 0's on either side. the 1's are allowed to wrap from LSB to
590 // MSB. so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
591 // not, since all 1's are not contiguous.
592 static bool IsRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
597 // look for first set bit
599 for (; i < 32; i++) {
600 if ((Val & (1 << (31 - i))) != 0) {
607 // look for last set bit
608 for (; i < 32; i++) {
609 if ((Val & (1 << (31 - i))) == 0)
614 // look for next set bit
615 for (; i < 32; i++) {
616 if ((Val & (1 << (31 - i))) != 0)
620 // if we exhausted all the bits, we found a match at this point for 0*1*0*
624 // since we just encountered more 1's, if it doesn't wrap around to the
625 // most significant bit of the word, then we did not find a match to 1*0*1* so
630 // look for last set bit
631 for (MB = i; i < 32; i++) {
632 if ((Val & (1 << (31 - i))) == 0)
636 // if we exhausted all the bits, then we found a match for 1*0*1*, otherwise,
637 // the value is not a run of ones.
643 /// getImmediateForOpcode - This method returns a value indicating whether
644 /// the ConstantSDNode N can be used as an immediate to Opcode. The return
645 /// values are either 0, 1 or 2. 0 indicates that either N is not a
646 /// ConstantSDNode, or is not suitable for use by that opcode.
647 /// Return value codes for turning into an enum someday:
648 /// 1: constant may be used in normal immediate form.
649 /// 2: constant may be used in shifted immediate form.
650 /// 3: log base 2 of the constant may be used.
651 /// 4: constant is suitable for integer division conversion
652 /// 5: constant is a bitfield mask
654 static unsigned getImmediateForOpcode(SDOperand N, unsigned Opcode,
655 unsigned& Imm, bool U = false) {
656 if (N.getOpcode() != ISD::Constant) return 0;
658 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
663 if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
664 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
668 if (IsRunOfOnes(v, MB, ME)) { Imm = MB << 16 | ME & 0xFFFF; return 5; }
669 if (v >= 0 && v <= 65535) { Imm = v & 0xFFFF; return 1; }
670 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
675 if (v >= 0 && v <= 65535) { Imm = v & 0xFFFF; return 1; }
676 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
679 if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
682 // handle subtract-from separately from subtract, since subi is really addi
683 if (U && v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
684 if (!U && v <= 32768 && v >= -32767) { Imm = (-v) & 0xFFFF; return 1; }
687 if (U && (v >= 0 && v <= 65535)) { Imm = v & 0xFFFF; return 1; }
688 if (!U && (v <= 32767 && v >= -32768)) { Imm = v & 0xFFFF; return 1; }
691 if ((Imm = ExactLog2(v))) { return 3; }
692 if ((Imm = ExactLog2(-v))) { Imm = -Imm; return 3; }
693 if (v <= -2 || v >= 2) { return 4; }
696 if (v > 1) { return 4; }
702 /// NodeHasRecordingVariant - If SelectExpr can always produce code for
703 /// NodeOpcode that also sets CR0 as a side effect, return true. Otherwise,
705 static bool NodeHasRecordingVariant(unsigned NodeOpcode) {
707 default: return false;
714 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
715 /// to Condition. If the Condition is unordered or unsigned, the bool argument
716 /// U is set to true, otherwise it is set to false.
717 static unsigned getBCCForSetCC(unsigned Condition, bool& U) {
720 default: assert(0 && "Unknown condition!"); abort();
721 case ISD::SETEQ: return PPC::BEQ;
722 case ISD::SETNE: return PPC::BNE;
723 case ISD::SETULT: U = true;
724 case ISD::SETLT: return PPC::BLT;
725 case ISD::SETULE: U = true;
726 case ISD::SETLE: return PPC::BLE;
727 case ISD::SETUGT: U = true;
728 case ISD::SETGT: return PPC::BGT;
729 case ISD::SETUGE: U = true;
730 case ISD::SETGE: return PPC::BGE;
735 /// getCROpForOp - Return the condition register opcode (or inverted opcode)
736 /// associated with the SelectionDAG opcode.
737 static unsigned getCROpForSetCC(unsigned Opcode, bool Inv1, bool Inv2) {
739 default: assert(0 && "Unknown opcode!"); abort();
741 if (Inv1 && Inv2) return PPC::CRNOR; // De Morgan's Law
742 if (!Inv1 && !Inv2) return PPC::CRAND;
743 if (Inv1 ^ Inv2) return PPC::CRANDC;
745 if (Inv1 && Inv2) return PPC::CRNAND; // De Morgan's Law
746 if (!Inv1 && !Inv2) return PPC::CROR;
747 if (Inv1 ^ Inv2) return PPC::CRORC;
752 /// getCRIdxForSetCC - Return the index of the condition register field
753 /// associated with the SetCC condition, and whether or not the field is
754 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
755 static unsigned getCRIdxForSetCC(unsigned Condition, bool& Inv) {
757 default: assert(0 && "Unknown condition!"); abort();
759 case ISD::SETLT: Inv = false; return 0;
761 case ISD::SETGE: Inv = true; return 0;
763 case ISD::SETGT: Inv = false; return 1;
765 case ISD::SETLE: Inv = true; return 1;
766 case ISD::SETEQ: Inv = false; return 2;
767 case ISD::SETNE: Inv = true; return 2;
772 /// IndexedOpForOp - Return the indexed variant for each of the PowerPC load
773 /// and store immediate instructions.
774 static unsigned IndexedOpForOp(unsigned Opcode) {
776 default: assert(0 && "Unknown opcode!"); abort();
777 case PPC::LBZ: return PPC::LBZX; case PPC::STB: return PPC::STBX;
778 case PPC::LHZ: return PPC::LHZX; case PPC::STH: return PPC::STHX;
779 case PPC::LHA: return PPC::LHAX; case PPC::STW: return PPC::STWX;
780 case PPC::LWZ: return PPC::LWZX; case PPC::STFS: return PPC::STFSX;
781 case PPC::LFS: return PPC::LFSX; case PPC::STFD: return PPC::STFDX;
782 case PPC::LFD: return PPC::LFDX;
787 // Structure used to return the necessary information to codegen an SDIV as
790 int m; // magic number
791 int s; // shift amount
795 unsigned int m; // magic number
796 int a; // add indicator
797 int s; // shift amount
800 /// magic - calculate the magic numbers required to codegen an integer sdiv as
801 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
803 static struct ms magic(int d) {
805 unsigned int ad, anc, delta, q1, r1, q2, r2, t;
806 const unsigned int two31 = 2147483648U; // 2^31
810 t = two31 + ((unsigned int)d >> 31);
811 anc = t - 1 - t%ad; // absolute value of nc
812 p = 31; // initialize p
813 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
814 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
815 q2 = two31/ad; // initialize q2 = 2p/abs(d)
816 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
819 q1 = 2*q1; // update q1 = 2p/abs(nc)
820 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
821 if (r1 >= anc) { // must be unsigned comparison
825 q2 = 2*q2; // update q2 = 2p/abs(d)
826 r2 = 2*r2; // update r2 = rem(2p/abs(d))
827 if (r2 >= ad) { // must be unsigned comparison
832 } while (q1 < delta || (q1 == delta && r1 == 0));
835 if (d < 0) mag.m = -mag.m; // resulting magic number
836 mag.s = p - 32; // resulting shift
840 /// magicu - calculate the magic numbers required to codegen an integer udiv as
841 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
842 static struct mu magicu(unsigned d)
845 unsigned int nc, delta, q1, r1, q2, r2;
847 magu.a = 0; // initialize "add" indicator
849 p = 31; // initialize p
850 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
851 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
852 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
853 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
856 if (r1 >= nc - r1 ) {
857 q1 = 2*q1 + 1; // update q1
858 r1 = 2*r1 - nc; // update r1
861 q1 = 2*q1; // update q1
862 r1 = 2*r1; // update r1
864 if (r2 + 1 >= d - r2) {
865 if (q2 >= 0x7FFFFFFF) magu.a = 1;
866 q2 = 2*q2 + 1; // update q2
867 r2 = 2*r2 + 1 - d; // update r2
870 if (q2 >= 0x80000000) magu.a = 1;
871 q2 = 2*q2; // update q2
872 r2 = 2*r2 + 1; // update r2
875 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
876 magu.m = q2 + 1; // resulting magic number
877 magu.s = p - 32; // resulting shift
882 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
883 /// return a DAG expression to select that will generate the same value by
884 /// multiplying by a magic number. See:
885 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
886 SDOperand ISel::BuildSDIVSequence(SDOperand N) {
887 int d = (int)cast<ConstantSDNode>(N.getOperand(1))->getSignExtended();
888 ms magics = magic(d);
889 // Multiply the numerator (operand 0) by the magic value
890 SDOperand Q = ISelDAG->getNode(ISD::MULHS, MVT::i32, N.getOperand(0),
891 ISelDAG->getConstant(magics.m, MVT::i32));
892 // If d > 0 and m < 0, add the numerator
893 if (d > 0 && magics.m < 0)
894 Q = ISelDAG->getNode(ISD::ADD, MVT::i32, Q, N.getOperand(0));
895 // If d < 0 and m > 0, subtract the numerator.
896 if (d < 0 && magics.m > 0)
897 Q = ISelDAG->getNode(ISD::SUB, MVT::i32, Q, N.getOperand(0));
898 // Shift right algebraic if shift value is nonzero
900 Q = ISelDAG->getNode(ISD::SRA, MVT::i32, Q,
901 ISelDAG->getConstant(magics.s, MVT::i32));
902 // Extract the sign bit and add it to the quotient
904 ISelDAG->getNode(ISD::SRL, MVT::i32, Q, ISelDAG->getConstant(31, MVT::i32));
905 return ISelDAG->getNode(ISD::ADD, MVT::i32, Q, T);
908 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
909 /// return a DAG expression to select that will generate the same value by
910 /// multiplying by a magic number. See:
911 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
912 SDOperand ISel::BuildUDIVSequence(SDOperand N) {
914 (unsigned)cast<ConstantSDNode>(N.getOperand(1))->getSignExtended();
915 mu magics = magicu(d);
916 // Multiply the numerator (operand 0) by the magic value
917 SDOperand Q = ISelDAG->getNode(ISD::MULHU, MVT::i32, N.getOperand(0),
918 ISelDAG->getConstant(magics.m, MVT::i32));
920 Q = ISelDAG->getNode(ISD::SRL, MVT::i32, Q,
921 ISelDAG->getConstant(magics.s, MVT::i32));
923 SDOperand NPQ = ISelDAG->getNode(ISD::SUB, MVT::i32, N.getOperand(0), Q);
924 NPQ = ISelDAG->getNode(ISD::SRL, MVT::i32, NPQ,
925 ISelDAG->getConstant(1, MVT::i32));
926 NPQ = ISelDAG->getNode(ISD::ADD, MVT::i32, NPQ, Q);
927 Q = ISelDAG->getNode(ISD::SRL, MVT::i32, NPQ,
928 ISelDAG->getConstant(magics.s-1, MVT::i32));
933 /// getGlobalBaseReg - Output the instructions required to put the
934 /// base address to use for accessing globals into a register.
936 unsigned ISel::getGlobalBaseReg() {
937 if (!GlobalBaseInitialized) {
938 // Insert the set of GlobalBaseReg into the first MBB of the function
939 MachineBasicBlock &FirstMBB = BB->getParent()->front();
940 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
941 GlobalBaseReg = MakeReg(MVT::i32);
942 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
943 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg).addReg(PPC::LR);
944 GlobalBaseInitialized = true;
946 return GlobalBaseReg;
949 /// getConstDouble - Loads a floating point value into a register, via the
950 /// Constant Pool. Optionally takes a register in which to load the value.
951 unsigned ISel::getConstDouble(double doubleVal, unsigned Result=0) {
952 unsigned Tmp1 = MakeReg(MVT::i32);
953 if (0 == Result) Result = MakeReg(MVT::f64);
954 MachineConstantPool *CP = BB->getParent()->getConstantPool();
955 ConstantFP *CFP = ConstantFP::get(Type::DoubleTy, doubleVal);
956 unsigned CPI = CP->getConstantPoolIndex(CFP);
958 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
959 .addConstantPoolIndex(CPI);
961 BuildMI(BB, PPC::LIS, 1, Tmp1).addConstantPoolIndex(CPI);
962 BuildMI(BB, PPC::LFD, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
966 /// MoveCRtoGPR - Move CCReg[Idx] to the least significant bit of Result. If
967 /// Inv is true, then invert the result.
968 void ISel::MoveCRtoGPR(unsigned CCReg, bool Inv, unsigned Idx, unsigned Result){
969 unsigned IntCR = MakeReg(MVT::i32);
970 BuildMI(BB, PPC::MCRF, 1, PPC::CR7).addReg(CCReg);
971 BuildMI(BB, GPOPT ? PPC::MFOCRF : PPC::MFCR, 1, IntCR).addReg(PPC::CR7);
973 unsigned Tmp1 = MakeReg(MVT::i32);
974 BuildMI(BB, PPC::RLWINM, 4, Tmp1).addReg(IntCR).addImm(32-(3-Idx))
975 .addImm(31).addImm(31);
976 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp1).addImm(1);
978 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(IntCR).addImm(32-(3-Idx))
979 .addImm(31).addImm(31);
983 /// SelectBitfieldInsert - turn an or of two masked values into
984 /// the rotate left word immediate then mask insert (rlwimi) instruction.
985 /// Returns true on success, false if the caller still needs to select OR.
987 /// Patterns matched:
988 /// 1. or shl, and 5. or and, and
989 /// 2. or and, shl 6. or shl, shr
990 /// 3. or shr, and 7. or shr, shl
992 bool ISel::SelectBitfieldInsert(SDOperand OR, unsigned Result) {
993 bool IsRotate = false;
994 unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, Amount = 0;
996 SDOperand Op0 = OR.getOperand(0);
997 SDOperand Op1 = OR.getOperand(1);
999 unsigned Op0Opc = Op0.getOpcode();
1000 unsigned Op1Opc = Op1.getOpcode();
1002 // Verify that we have the correct opcodes
1003 if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc)
1005 if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc)
1008 // Generate Mask value for Target
1009 if (ConstantSDNode *CN =
1010 dyn_cast<ConstantSDNode>(Op0.getOperand(1).Val)) {
1012 case ISD::SHL: TgtMask <<= (unsigned)CN->getValue(); break;
1013 case ISD::SRL: TgtMask >>= (unsigned)CN->getValue(); break;
1014 case ISD::AND: TgtMask &= (unsigned)CN->getValue(); break;
1020 // Generate Mask value for Insert
1021 if (ConstantSDNode *CN =
1022 dyn_cast<ConstantSDNode>(Op1.getOperand(1).Val)) {
1025 Amount = CN->getValue();
1027 if (Op0Opc == ISD::SRL) IsRotate = true;
1030 Amount = CN->getValue();
1033 if (Op0Opc == ISD::SHL) IsRotate = true;
1036 InsMask &= (unsigned)CN->getValue();
1045 // If both of the inputs are ANDs and one of them has a logical shift by
1046 // constant as its input, make that the inserted value so that we can combine
1047 // the shift into the rotate part of the rlwimi instruction
1048 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
1049 if (Op1.getOperand(0).getOpcode() == ISD::SHL ||
1050 Op1.getOperand(0).getOpcode() == ISD::SRL) {
1051 if (ConstantSDNode *CN =
1052 dyn_cast<ConstantSDNode>(Op1.getOperand(0).getOperand(1).Val)) {
1053 Amount = Op1.getOperand(0).getOpcode() == ISD::SHL ?
1054 CN->getValue() : 32 - CN->getValue();
1055 Tmp3 = SelectExpr(Op1.getOperand(0).getOperand(0));
1057 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
1058 Op0.getOperand(0).getOpcode() == ISD::SRL) {
1059 if (ConstantSDNode *CN =
1060 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(1).Val)) {
1061 std::swap(Op0, Op1);
1062 std::swap(TgtMask, InsMask);
1063 Amount = Op1.getOperand(0).getOpcode() == ISD::SHL ?
1064 CN->getValue() : 32 - CN->getValue();
1065 Tmp3 = SelectExpr(Op1.getOperand(0).getOperand(0));
1070 // Verify that the Target mask and Insert mask together form a full word mask
1071 // and that the Insert mask is a run of set bits (which implies both are runs
1072 // of set bits). Given that, Select the arguments and generate the rlwimi
1075 if (((TgtMask & InsMask) == 0) && IsRunOfOnes(InsMask, MB, ME)) {
1076 unsigned Tmp1, Tmp2;
1077 bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF;
1078 // Check for rotlwi / rotrwi here, a special case of bitfield insert
1079 // where both bitfield halves are sourced from the same value.
1080 if (IsRotate && fullMask &&
1081 OR.getOperand(0).getOperand(0) == OR.getOperand(1).getOperand(0)) {
1082 Tmp1 = SelectExpr(OR.getOperand(0).getOperand(0));
1083 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(Amount)
1084 .addImm(0).addImm(31);
1087 if (Op0Opc == ISD::AND && fullMask)
1088 Tmp1 = SelectExpr(Op0.getOperand(0));
1090 Tmp1 = SelectExpr(Op0);
1091 Tmp2 = Tmp3 ? Tmp3 : SelectExpr(Op1.getOperand(0));
1092 BuildMI(BB, PPC::RLWIMI, 5, Result).addReg(Tmp1).addReg(Tmp2)
1093 .addImm(Amount).addImm(MB).addImm(ME);
1099 /// FoldIfWideZeroExtend - 32 bit PowerPC implicit masks shift amounts to the
1100 /// low six bits. If the shift amount is an ISD::AND node with a mask that is
1101 /// wider than the implicit mask, then we can get rid of the AND and let the
1102 /// shift do the mask.
1103 unsigned ISel::FoldIfWideZeroExtend(SDOperand N) {
1105 if (N.getOpcode() == ISD::AND &&
1106 5 == getImmediateForOpcode(N.getOperand(1), ISD::AND, C) && // isMask
1107 31 == (C & 0xFFFF) && // ME
1108 26 >= (C >> 16)) // MB
1109 return SelectExpr(N.getOperand(0));
1111 return SelectExpr(N);
1114 unsigned ISel::SelectCC(SDOperand CC, unsigned& Opc, bool &Inv, unsigned& Idx) {
1115 unsigned Result, Tmp1, Tmp2;
1116 bool AlreadySelected = false;
1117 static const unsigned CompareOpcodes[] =
1118 { PPC::FCMPU, PPC::FCMPU, PPC::CMPW, PPC::CMPLW };
1120 // Allocate a condition register for this expression
1121 Result = RegMap->createVirtualRegister(PPC32::CRRCRegisterClass);
1123 // If the first operand to the select is a SETCC node, then we can fold it
1124 // into the branch that selects which value to return.
1125 if (SetCCSDNode* SetCC = dyn_cast<SetCCSDNode>(CC.Val)) {
1127 Opc = getBCCForSetCC(SetCC->getCondition(), U);
1128 Idx = getCRIdxForSetCC(SetCC->getCondition(), Inv);
1130 // Pass the optional argument U to getImmediateForOpcode for SETCC,
1131 // so that it knows whether the SETCC immediate range is signed or not.
1132 if (1 == getImmediateForOpcode(SetCC->getOperand(1), ISD::SETCC,
1134 // For comparisons against zero, we can implicity set CR0 if a recording
1135 // variant (e.g. 'or.' instead of 'or') of the instruction that defines
1136 // operand zero of the SetCC node is available.
1138 NodeHasRecordingVariant(SetCC->getOperand(0).getOpcode()) &&
1139 SetCC->getOperand(0).Val->hasOneUse()) {
1140 RecordSuccess = false;
1141 Tmp1 = SelectExpr(SetCC->getOperand(0), true);
1142 if (RecordSuccess) {
1144 BuildMI(BB, PPC::MCRF, 1, Result).addReg(PPC::CR0);
1147 AlreadySelected = true;
1149 // If we could not implicitly set CR0, then emit a compare immediate
1151 if (!AlreadySelected) Tmp1 = SelectExpr(SetCC->getOperand(0));
1153 BuildMI(BB, PPC::CMPLWI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1155 BuildMI(BB, PPC::CMPWI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1157 bool IsInteger = MVT::isInteger(SetCC->getOperand(0).getValueType());
1158 unsigned CompareOpc = CompareOpcodes[2 * IsInteger + U];
1159 Tmp1 = SelectExpr(SetCC->getOperand(0));
1160 Tmp2 = SelectExpr(SetCC->getOperand(1));
1161 BuildMI(BB, CompareOpc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1164 // If this isn't a SetCC, then select the value and compare it against zero,
1165 // treating it as if it were a boolean.
1167 Idx = getCRIdxForSetCC(ISD::SETNE, Inv);
1168 Tmp1 = SelectExpr(CC);
1169 BuildMI(BB, PPC::CMPLWI, 2, Result).addReg(Tmp1).addImm(0);
1174 unsigned ISel::SelectCCExpr(SDOperand N, unsigned& Opc, bool &Inv,
1177 unsigned Idx0, Idx1, CROpc, Opc1, Tmp1, Tmp2;
1179 // Allocate a condition register for this expression
1180 unsigned Result = RegMap->createVirtualRegister(PPC32::CRRCRegisterClass);
1182 // Check for the operations we support:
1183 switch(N.getOpcode()) {
1186 Idx = getCRIdxForSetCC(ISD::SETNE, Inv);
1187 Tmp1 = SelectExpr(N);
1188 BuildMI(BB, PPC::CMPLWI, 2, Result).addReg(Tmp1).addImm(0);
1193 Tmp1 = SelectCCExpr(N.getOperand(0), Opc, Inv0, Idx0);
1194 Tmp2 = SelectCCExpr(N.getOperand(1), Opc1, Inv1, Idx1);
1195 CROpc = getCROpForSetCC(N.getOpcode(), Inv0, Inv1);
1196 if (Inv0 && !Inv1) {
1197 std::swap(Tmp1, Tmp2);
1198 std::swap(Idx0, Idx1);
1201 if (Inv0 && Inv1) Opc = PPC32InstrInfo::invertPPCBranchOpcode(Opc);
1202 BuildMI(BB, CROpc, 5, Result).addImm(Idx0).addReg(Tmp1).addImm(Idx0)
1203 .addReg(Tmp2).addImm(Idx1);
1208 Tmp1 = SelectCC(N, Opc, Inv, Idx);
1215 /// Check to see if the load is a constant offset from a base register
1216 bool ISel::SelectAddr(SDOperand N, unsigned& Reg, int& offset)
1218 unsigned imm = 0, opcode = N.getOpcode();
1219 if (N.getOpcode() == ISD::ADD) {
1220 Reg = SelectExpr(N.getOperand(0));
1221 if (1 == getImmediateForOpcode(N.getOperand(1), opcode, imm)) {
1225 offset = SelectExpr(N.getOperand(1));
1228 Reg = SelectExpr(N);
1233 void ISel::SelectBranchCC(SDOperand N)
1235 MachineBasicBlock *Dest =
1236 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
1239 unsigned Opc, CCReg, Idx;
1240 Select(N.getOperand(0)); //chain
1241 CCReg = SelectCC(N.getOperand(1), Opc, Inv, Idx);
1243 // Iterate to the next basic block
1244 ilist<MachineBasicBlock>::iterator It = BB;
1247 // If this is a two way branch, then grab the fallthrough basic block argument
1248 // and build a PowerPC branch pseudo-op, suitable for long branch conversion
1249 // if necessary by the branch selection pass. Otherwise, emit a standard
1250 // conditional branch.
1251 if (N.getOpcode() == ISD::BRCONDTWOWAY) {
1252 MachineBasicBlock *Fallthrough =
1253 cast<BasicBlockSDNode>(N.getOperand(3))->getBasicBlock();
1255 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
1256 .addMBB(Dest).addMBB(Fallthrough);
1257 if (Fallthrough != It)
1258 BuildMI(BB, PPC::B, 1).addMBB(Fallthrough);
1260 if (Fallthrough != It) {
1261 Opc = PPC32InstrInfo::invertPPCBranchOpcode(Opc);
1262 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
1263 .addMBB(Fallthrough).addMBB(Dest);
1267 // If the fallthrough path is off the end of the function, which would be
1268 // undefined behavior, set it to be the same as the current block because
1269 // we have nothing better to set it to, and leaving it alone will cause the
1270 // PowerPC Branch Selection pass to crash.
1271 if (It == BB->getParent()->end()) It = Dest;
1272 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
1273 .addMBB(Dest).addMBB(It);
1278 unsigned ISel::SelectExpr(SDOperand N, bool Recording) {
1280 unsigned Tmp1, Tmp2, Tmp3;
1282 unsigned opcode = N.getOpcode();
1284 SDNode *Node = N.Val;
1285 MVT::ValueType DestType = N.getValueType();
1287 if (Node->getOpcode() == ISD::CopyFromReg &&
1288 MRegisterInfo::isVirtualRegister(cast<RegSDNode>(Node)->getReg()))
1289 // Just use the specified register as our input.
1290 return cast<RegSDNode>(Node)->getReg();
1292 unsigned &Reg = ExprMap[N];
1293 if (Reg) return Reg;
1295 switch (N.getOpcode()) {
1297 Reg = Result = (N.getValueType() != MVT::Other) ?
1298 MakeReg(N.getValueType()) : 1;
1302 // If this is a call instruction, make sure to prepare ALL of the result
1303 // values as well as the chain.
1304 if (Node->getNumValues() == 1)
1305 Reg = Result = 1; // Void call, just a chain.
1307 Result = MakeReg(Node->getValueType(0));
1308 ExprMap[N.getValue(0)] = Result;
1309 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
1310 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
1311 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
1314 case ISD::ADD_PARTS:
1315 case ISD::SUB_PARTS:
1316 case ISD::SHL_PARTS:
1317 case ISD::SRL_PARTS:
1318 case ISD::SRA_PARTS:
1319 Result = MakeReg(Node->getValueType(0));
1320 ExprMap[N.getValue(0)] = Result;
1321 for (unsigned i = 1, e = N.Val->getNumValues(); i != e; ++i)
1322 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
1329 assert(0 && "Node not handled!\n");
1331 BuildMI(BB, PPC::IMPLICIT_DEF, 0, Result);
1333 case ISD::DYNAMIC_STACKALLOC:
1334 // Generate both result values. FIXME: Need a better commment here?
1336 ExprMap[N.getValue(1)] = 1;
1338 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1340 // FIXME: We are currently ignoring the requested alignment for handling
1341 // greater than the stack alignment. This will need to be revisited at some
1342 // point. Align = N.getOperand(2);
1343 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
1344 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
1345 std::cerr << "Cannot allocate stack object with greater alignment than"
1346 << " the stack alignment yet!";
1349 Select(N.getOperand(0));
1350 Tmp1 = SelectExpr(N.getOperand(1));
1351 // Subtract size from stack pointer, thereby allocating some space.
1352 BuildMI(BB, PPC::SUBF, 2, PPC::R1).addReg(Tmp1).addReg(PPC::R1);
1353 // Put a pointer to the space into the result register by copying the SP
1354 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R1).addReg(PPC::R1);
1357 case ISD::ConstantPool:
1358 Tmp1 = cast<ConstantPoolSDNode>(N)->getIndex();
1359 Tmp2 = MakeReg(MVT::i32);
1361 BuildMI(BB, PPC::ADDIS, 2, Tmp2).addReg(getGlobalBaseReg())
1362 .addConstantPoolIndex(Tmp1);
1364 BuildMI(BB, PPC::LIS, 1, Tmp2).addConstantPoolIndex(Tmp1);
1365 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp2).addConstantPoolIndex(Tmp1);
1368 case ISD::FrameIndex:
1369 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
1370 addFrameReference(BuildMI(BB, PPC::ADDI, 2, Result), (int)Tmp1, 0, false);
1373 case ISD::GlobalAddress: {
1374 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
1375 Tmp1 = MakeReg(MVT::i32);
1377 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
1378 .addGlobalAddress(GV);
1380 BuildMI(BB, PPC::LIS, 2, Tmp1).addGlobalAddress(GV);
1381 if (GV->hasWeakLinkage() || GV->isExternal()) {
1382 BuildMI(BB, PPC::LWZ, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
1384 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp1).addGlobalAddress(GV);
1392 case ISD::SEXTLOAD: {
1393 MVT::ValueType TypeBeingLoaded = (ISD::LOAD == opcode) ?
1394 Node->getValueType(0) : cast<VTSDNode>(Node->getOperand(3))->getVT();
1395 bool sext = (ISD::SEXTLOAD == opcode);
1397 // Make sure we generate both values.
1399 ExprMap[N.getValue(1)] = 1; // Generate the token
1401 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1403 SDOperand Chain = N.getOperand(0);
1404 SDOperand Address = N.getOperand(1);
1407 switch (TypeBeingLoaded) {
1408 default: Node->dump(); assert(0 && "Cannot load this type!");
1409 case MVT::i1: Opc = PPC::LBZ; break;
1410 case MVT::i8: Opc = PPC::LBZ; break;
1411 case MVT::i16: Opc = sext ? PPC::LHA : PPC::LHZ; break;
1412 case MVT::i32: Opc = PPC::LWZ; break;
1413 case MVT::f32: Opc = PPC::LFS; break;
1414 case MVT::f64: Opc = PPC::LFD; break;
1417 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Address)) {
1418 Tmp1 = MakeReg(MVT::i32);
1419 int CPI = CP->getIndex();
1421 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
1422 .addConstantPoolIndex(CPI);
1424 BuildMI(BB, PPC::LIS, 1, Tmp1).addConstantPoolIndex(CPI);
1425 BuildMI(BB, Opc, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
1426 } else if (Address.getOpcode() == ISD::FrameIndex) {
1427 Tmp1 = cast<FrameIndexSDNode>(Address)->getIndex();
1428 addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1);
1429 } else if(GlobalAddressSDNode *GN = dyn_cast<GlobalAddressSDNode>(Address)){
1430 GlobalValue *GV = GN->getGlobal();
1431 Tmp1 = MakeReg(MVT::i32);
1433 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
1434 .addGlobalAddress(GV);
1436 BuildMI(BB, PPC::LIS, 2, Tmp1).addGlobalAddress(GV);
1437 if (GV->hasWeakLinkage() || GV->isExternal()) {
1438 Tmp2 = MakeReg(MVT::i32);
1439 BuildMI(BB, PPC::LWZ, 2, Tmp2).addGlobalAddress(GV).addReg(Tmp1);
1440 BuildMI(BB, Opc, 2, Result).addSImm(0).addReg(Tmp2);
1442 BuildMI(BB, Opc, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
1446 bool idx = SelectAddr(Address, Tmp1, offset);
1448 Opc = IndexedOpForOp(Opc);
1449 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(offset);
1451 BuildMI(BB, Opc, 2, Result).addSImm(offset).addReg(Tmp1);
1459 unsigned GPR_idx = 0, FPR_idx = 0;
1460 static const unsigned GPR[] = {
1461 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1462 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1464 static const unsigned FPR[] = {
1465 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1466 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1469 // Lower the chain for this call.
1470 Select(N.getOperand(0));
1471 ExprMap[N.getValue(Node->getNumValues()-1)] = 1;
1473 MachineInstr *CallMI;
1474 // Emit the correct call instruction based on the type of symbol called.
1475 if (GlobalAddressSDNode *GASD =
1476 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
1477 CallMI = BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(GASD->getGlobal(),
1479 } else if (ExternalSymbolSDNode *ESSDN =
1480 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
1481 CallMI = BuildMI(PPC::CALLpcrel, 1).addExternalSymbol(ESSDN->getSymbol(),
1484 Tmp1 = SelectExpr(N.getOperand(1));
1485 BuildMI(BB, PPC::OR, 2, PPC::R12).addReg(Tmp1).addReg(Tmp1);
1486 BuildMI(BB, PPC::MTCTR, 1).addReg(PPC::R12);
1487 CallMI = BuildMI(PPC::CALLindirect, 3).addImm(20).addImm(0)
1491 // Load the register args to virtual regs
1492 std::vector<unsigned> ArgVR;
1493 for(int i = 2, e = Node->getNumOperands(); i < e; ++i)
1494 ArgVR.push_back(SelectExpr(N.getOperand(i)));
1496 // Copy the virtual registers into the appropriate argument register
1497 for(int i = 0, e = ArgVR.size(); i < e; ++i) {
1498 switch(N.getOperand(i+2).getValueType()) {
1499 default: Node->dump(); assert(0 && "Unknown value type for call");
1504 assert(GPR_idx < 8 && "Too many int args");
1505 if (N.getOperand(i+2).getOpcode() != ISD::UNDEF) {
1506 BuildMI(BB, PPC::OR,2,GPR[GPR_idx]).addReg(ArgVR[i]).addReg(ArgVR[i]);
1507 CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use);
1513 assert(FPR_idx < 13 && "Too many fp args");
1514 BuildMI(BB, PPC::FMR, 1, FPR[FPR_idx]).addReg(ArgVR[i]);
1515 CallMI->addRegOperand(FPR[FPR_idx], MachineOperand::Use);
1521 // Put the call instruction in the correct place in the MachineBasicBlock
1522 BB->push_back(CallMI);
1524 switch (Node->getValueType(0)) {
1525 default: assert(0 && "Unknown value type for call result!");
1526 case MVT::Other: return 1;
1531 if (Node->getValueType(1) == MVT::i32) {
1532 BuildMI(BB, PPC::OR, 2, Result+1).addReg(PPC::R3).addReg(PPC::R3);
1533 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R4).addReg(PPC::R4);
1535 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R3).addReg(PPC::R3);
1540 BuildMI(BB, PPC::FMR, 1, Result).addReg(PPC::F1);
1543 return Result+N.ResNo;
1546 case ISD::SIGN_EXTEND:
1547 case ISD::SIGN_EXTEND_INREG:
1548 Tmp1 = SelectExpr(N.getOperand(0));
1549 switch(cast<VTSDNode>(Node->getOperand(1))->getVT()) {
1550 default: Node->dump(); assert(0 && "Unhandled SIGN_EXTEND type"); break;
1552 BuildMI(BB, PPC::EXTSH, 1, Result).addReg(Tmp1);
1555 BuildMI(BB, PPC::EXTSB, 1, Result).addReg(Tmp1);
1558 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp1).addSImm(0);
1563 case ISD::CopyFromReg:
1564 DestType = N.getValue(0).getValueType();
1566 Result = ExprMap[N.getValue(0)] = MakeReg(DestType);
1567 Tmp1 = dyn_cast<RegSDNode>(Node)->getReg();
1568 if (MVT::isInteger(DestType))
1569 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1571 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
1575 Tmp1 = SelectExpr(N.getOperand(0));
1576 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1577 Tmp2 = CN->getValue() & 0x1F;
1578 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(Tmp2).addImm(0)
1581 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1582 BuildMI(BB, PPC::SLW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1587 Tmp1 = SelectExpr(N.getOperand(0));
1588 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1589 Tmp2 = CN->getValue() & 0x1F;
1590 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(32-Tmp2)
1591 .addImm(Tmp2).addImm(31);
1593 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1594 BuildMI(BB, PPC::SRW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1599 Tmp1 = SelectExpr(N.getOperand(0));
1600 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1601 Tmp2 = CN->getValue() & 0x1F;
1602 BuildMI(BB, PPC::SRAWI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1604 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1605 BuildMI(BB, PPC::SRAW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1610 Tmp1 = SelectExpr(N.getOperand(0));
1611 BuildMI(BB, PPC::CNTLZW, 1, Result).addReg(Tmp1);
1615 if (!MVT::isInteger(DestType)) {
1616 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
1617 N.getOperand(0).Val->hasOneUse()) {
1618 ++FusedFP; // Statistic
1619 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1620 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1621 Tmp3 = SelectExpr(N.getOperand(1));
1622 Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
1623 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1626 if (!NoExcessFPPrecision && N.getOperand(1).getOpcode() == ISD::MUL &&
1627 N.getOperand(1).Val->hasOneUse()) {
1628 ++FusedFP; // Statistic
1629 Tmp1 = SelectExpr(N.getOperand(1).getOperand(0));
1630 Tmp2 = SelectExpr(N.getOperand(1).getOperand(1));
1631 Tmp3 = SelectExpr(N.getOperand(0));
1632 Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
1633 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1636 Opc = DestType == MVT::f64 ? PPC::FADD : PPC::FADDS;
1637 Tmp1 = SelectExpr(N.getOperand(0));
1638 Tmp2 = SelectExpr(N.getOperand(1));
1639 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1642 Tmp1 = SelectExpr(N.getOperand(0));
1643 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1644 default: assert(0 && "unhandled result code");
1645 case 0: // No immediate
1646 Tmp2 = SelectExpr(N.getOperand(1));
1647 BuildMI(BB, PPC::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1649 case 1: // Low immediate
1650 BuildMI(BB, PPC::ADDI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1652 case 2: // Shifted immediate
1653 BuildMI(BB, PPC::ADDIS, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1659 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1660 default: assert(0 && "unhandled result code");
1661 case 0: // No immediate
1662 // Check for andc: and, (xor a, -1), b
1663 if (N.getOperand(0).getOpcode() == ISD::XOR &&
1664 N.getOperand(0).getOperand(1).getOpcode() == ISD::Constant &&
1665 cast<ConstantSDNode>(N.getOperand(0).getOperand(1))->isAllOnesValue()) {
1666 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1667 Tmp2 = SelectExpr(N.getOperand(1));
1668 BuildMI(BB, PPC::ANDC, 2, Result).addReg(Tmp2).addReg(Tmp1);
1671 // It wasn't and-with-complement, emit a regular and
1672 Tmp1 = SelectExpr(N.getOperand(0));
1673 Tmp2 = SelectExpr(N.getOperand(1));
1674 Opc = Recording ? PPC::ANDo : PPC::AND;
1675 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1677 case 1: // Low immediate
1678 Tmp1 = SelectExpr(N.getOperand(0));
1679 BuildMI(BB, PPC::ANDIo, 2, Result).addReg(Tmp1).addImm(Tmp2);
1681 case 2: // Shifted immediate
1682 Tmp1 = SelectExpr(N.getOperand(0));
1683 BuildMI(BB, PPC::ANDISo, 2, Result).addReg(Tmp1).addImm(Tmp2);
1685 case 5: // Bitfield mask
1686 Opc = Recording ? PPC::RLWINMo : PPC::RLWINM;
1687 Tmp3 = Tmp2 >> 16; // MB
1688 Tmp2 &= 0xFFFF; // ME
1690 // FIXME: Catch SHL-AND in addition to SRL-AND in this block.
1691 if (N.getOperand(0).getOpcode() == ISD::SRL)
1692 if (ConstantSDNode *SA =
1693 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
1695 // We can fold the RLWINM and the SRL together if the mask is
1696 // clearing the top bits which are rotated around.
1697 unsigned RotAmt = 32-(SA->getValue() & 31);
1698 if (Tmp2 <= RotAmt) {
1699 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1700 BuildMI(BB, Opc, 4, Result).addReg(Tmp1).addImm(RotAmt)
1701 .addImm(Tmp3).addImm(Tmp2);
1706 Tmp1 = SelectExpr(N.getOperand(0));
1707 BuildMI(BB, Opc, 4, Result).addReg(Tmp1).addImm(0)
1708 .addImm(Tmp3).addImm(Tmp2);
1711 RecordSuccess = true;
1715 if (SelectBitfieldInsert(N, Result))
1717 Tmp1 = SelectExpr(N.getOperand(0));
1718 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1719 default: assert(0 && "unhandled result code");
1720 case 0: // No immediate
1721 Tmp2 = SelectExpr(N.getOperand(1));
1722 Opc = Recording ? PPC::ORo : PPC::OR;
1723 RecordSuccess = true;
1724 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1726 case 1: // Low immediate
1727 BuildMI(BB, PPC::ORI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1729 case 2: // Shifted immediate
1730 BuildMI(BB, PPC::ORIS, 2, Result).addReg(Tmp1).addImm(Tmp2);
1736 // Check for EQV: xor, (xor a, -1), b
1737 if (N.getOperand(0).getOpcode() == ISD::XOR &&
1738 N.getOperand(0).getOperand(1).getOpcode() == ISD::Constant &&
1739 cast<ConstantSDNode>(N.getOperand(0).getOperand(1))->isAllOnesValue()) {
1740 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1741 Tmp2 = SelectExpr(N.getOperand(1));
1742 BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1745 // Check for NOT, NOR, EQV, and NAND: xor (copy, or, xor, and), -1
1746 if (N.getOperand(1).getOpcode() == ISD::Constant &&
1747 cast<ConstantSDNode>(N.getOperand(1))->isAllOnesValue()) {
1748 switch(N.getOperand(0).getOpcode()) {
1750 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1751 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1752 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1755 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1756 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1757 BuildMI(BB, PPC::NAND, 2, Result).addReg(Tmp1).addReg(Tmp2);
1760 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1761 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1762 BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1765 Tmp1 = SelectExpr(N.getOperand(0));
1766 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1771 Tmp1 = SelectExpr(N.getOperand(0));
1772 switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1773 default: assert(0 && "unhandled result code");
1774 case 0: // No immediate
1775 Tmp2 = SelectExpr(N.getOperand(1));
1776 BuildMI(BB, PPC::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1778 case 1: // Low immediate
1779 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1781 case 2: // Shifted immediate
1782 BuildMI(BB, PPC::XORIS, 2, Result).addReg(Tmp1).addImm(Tmp2);
1789 if (!MVT::isInteger(DestType)) {
1790 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
1791 N.getOperand(0).Val->hasOneUse()) {
1792 ++FusedFP; // Statistic
1793 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1794 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1795 Tmp3 = SelectExpr(N.getOperand(1));
1796 Opc = DestType == MVT::f64 ? PPC::FMSUB : PPC::FMSUBS;
1797 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1800 if (!NoExcessFPPrecision && N.getOperand(1).getOpcode() == ISD::MUL &&
1801 N.getOperand(1).Val->hasOneUse()) {
1802 ++FusedFP; // Statistic
1803 Tmp1 = SelectExpr(N.getOperand(1).getOperand(0));
1804 Tmp2 = SelectExpr(N.getOperand(1).getOperand(1));
1805 Tmp3 = SelectExpr(N.getOperand(0));
1806 Opc = DestType == MVT::f64 ? PPC::FNMSUB : PPC::FNMSUBS;
1807 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1810 Opc = DestType == MVT::f64 ? PPC::FSUB : PPC::FSUBS;
1811 Tmp1 = SelectExpr(N.getOperand(0));
1812 Tmp2 = SelectExpr(N.getOperand(1));
1813 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1816 if (1 == getImmediateForOpcode(N.getOperand(0), opcode, Tmp1, true)) {
1817 Tmp2 = SelectExpr(N.getOperand(1));
1818 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp2).addSImm(Tmp1);
1819 } else if (1 == getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
1820 Tmp1 = SelectExpr(N.getOperand(0));
1821 BuildMI(BB, PPC::ADDI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1823 Tmp1 = SelectExpr(N.getOperand(0));
1824 Tmp2 = SelectExpr(N.getOperand(1));
1825 BuildMI(BB, PPC::SUBF, 2, Result).addReg(Tmp2).addReg(Tmp1);
1830 Tmp1 = SelectExpr(N.getOperand(0));
1831 if (1 == getImmediateForOpcode(N.getOperand(1), opcode, Tmp2))
1832 BuildMI(BB, PPC::MULLI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1834 Tmp2 = SelectExpr(N.getOperand(1));
1836 default: assert(0 && "Unknown type to ISD::MUL"); break;
1837 case MVT::i32: Opc = PPC::MULLW; break;
1838 case MVT::f32: Opc = PPC::FMULS; break;
1839 case MVT::f64: Opc = PPC::FMUL; break;
1841 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1847 Tmp1 = SelectExpr(N.getOperand(0));
1848 Tmp2 = SelectExpr(N.getOperand(1));
1849 Opc = (ISD::MULHU == opcode) ? PPC::MULHWU : PPC::MULHW;
1850 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1855 switch (getImmediateForOpcode(N.getOperand(1), opcode, Tmp3)) {
1857 // If this is an sdiv by a power of two, we can use an srawi/addze pair.
1859 Tmp1 = MakeReg(MVT::i32);
1860 Tmp2 = SelectExpr(N.getOperand(0));
1861 if ((int)Tmp3 < 0) {
1862 unsigned Tmp4 = MakeReg(MVT::i32);
1863 BuildMI(BB, PPC::SRAWI, 2, Tmp1).addReg(Tmp2).addImm(-Tmp3);
1864 BuildMI(BB, PPC::ADDZE, 1, Tmp4).addReg(Tmp1);
1865 BuildMI(BB, PPC::NEG, 1, Result).addReg(Tmp4);
1867 BuildMI(BB, PPC::SRAWI, 2, Tmp1).addReg(Tmp2).addImm(Tmp3);
1868 BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp1);
1871 // If this is a divide by constant, we can emit code using some magic
1872 // constants to implement it as a multiply instead.
1875 if (opcode == ISD::SDIV)
1876 return SelectExpr(BuildSDIVSequence(N));
1878 return SelectExpr(BuildUDIVSequence(N));
1880 Tmp1 = SelectExpr(N.getOperand(0));
1881 Tmp2 = SelectExpr(N.getOperand(1));
1883 default: assert(0 && "Unknown type to ISD::SDIV"); break;
1884 case MVT::i32: Opc = (ISD::UDIV == opcode) ? PPC::DIVWU : PPC::DIVW; break;
1885 case MVT::f32: Opc = PPC::FDIVS; break;
1886 case MVT::f64: Opc = PPC::FDIV; break;
1888 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1891 case ISD::ADD_PARTS:
1892 case ISD::SUB_PARTS: {
1893 assert(N.getNumOperands() == 4 && N.getValueType() == MVT::i32 &&
1894 "Not an i64 add/sub!");
1895 // Emit all of the operands.
1896 std::vector<unsigned> InVals;
1897 for (unsigned i = 0, e = N.getNumOperands(); i != e; ++i)
1898 InVals.push_back(SelectExpr(N.getOperand(i)));
1899 if (N.getOpcode() == ISD::ADD_PARTS) {
1900 BuildMI(BB, PPC::ADDC, 2, Result).addReg(InVals[0]).addReg(InVals[2]);
1901 BuildMI(BB, PPC::ADDE, 2, Result+1).addReg(InVals[1]).addReg(InVals[3]);
1903 BuildMI(BB, PPC::SUBFC, 2, Result).addReg(InVals[2]).addReg(InVals[0]);
1904 BuildMI(BB, PPC::SUBFE, 2, Result+1).addReg(InVals[3]).addReg(InVals[1]);
1906 return Result+N.ResNo;
1909 case ISD::SHL_PARTS:
1910 case ISD::SRA_PARTS:
1911 case ISD::SRL_PARTS: {
1912 assert(N.getNumOperands() == 3 && N.getValueType() == MVT::i32 &&
1913 "Not an i64 shift!");
1914 unsigned ShiftOpLo = SelectExpr(N.getOperand(0));
1915 unsigned ShiftOpHi = SelectExpr(N.getOperand(1));
1916 unsigned SHReg = FoldIfWideZeroExtend(N.getOperand(2));
1917 Tmp1 = MakeReg(MVT::i32);
1918 Tmp2 = MakeReg(MVT::i32);
1919 Tmp3 = MakeReg(MVT::i32);
1920 unsigned Tmp4 = MakeReg(MVT::i32);
1921 unsigned Tmp5 = MakeReg(MVT::i32);
1922 unsigned Tmp6 = MakeReg(MVT::i32);
1923 BuildMI(BB, PPC::SUBFIC, 2, Tmp1).addReg(SHReg).addSImm(32);
1924 if (ISD::SHL_PARTS == opcode) {
1925 BuildMI(BB, PPC::SLW, 2, Tmp2).addReg(ShiftOpHi).addReg(SHReg);
1926 BuildMI(BB, PPC::SRW, 2, Tmp3).addReg(ShiftOpLo).addReg(Tmp1);
1927 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1928 BuildMI(BB, PPC::ADDI, 2, Tmp5).addReg(SHReg).addSImm(-32);
1929 BuildMI(BB, PPC::SLW, 2, Tmp6).addReg(ShiftOpLo).addReg(Tmp5);
1930 BuildMI(BB, PPC::OR, 2, Result+1).addReg(Tmp4).addReg(Tmp6);
1931 BuildMI(BB, PPC::SLW, 2, Result).addReg(ShiftOpLo).addReg(SHReg);
1932 } else if (ISD::SRL_PARTS == opcode) {
1933 BuildMI(BB, PPC::SRW, 2, Tmp2).addReg(ShiftOpLo).addReg(SHReg);
1934 BuildMI(BB, PPC::SLW, 2, Tmp3).addReg(ShiftOpHi).addReg(Tmp1);
1935 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1936 BuildMI(BB, PPC::ADDI, 2, Tmp5).addReg(SHReg).addSImm(-32);
1937 BuildMI(BB, PPC::SRW, 2, Tmp6).addReg(ShiftOpHi).addReg(Tmp5);
1938 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp4).addReg(Tmp6);
1939 BuildMI(BB, PPC::SRW, 2, Result+1).addReg(ShiftOpHi).addReg(SHReg);
1941 MachineBasicBlock *TmpMBB = new MachineBasicBlock(BB->getBasicBlock());
1942 MachineBasicBlock *PhiMBB = new MachineBasicBlock(BB->getBasicBlock());
1943 MachineBasicBlock *OldMBB = BB;
1944 MachineFunction *F = BB->getParent();
1945 ilist<MachineBasicBlock>::iterator It = BB; ++It;
1946 F->getBasicBlockList().insert(It, TmpMBB);
1947 F->getBasicBlockList().insert(It, PhiMBB);
1948 BB->addSuccessor(TmpMBB);
1949 BB->addSuccessor(PhiMBB);
1950 BuildMI(BB, PPC::SRW, 2, Tmp2).addReg(ShiftOpLo).addReg(SHReg);
1951 BuildMI(BB, PPC::SLW, 2, Tmp3).addReg(ShiftOpHi).addReg(Tmp1);
1952 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1953 BuildMI(BB, PPC::ADDICo, 2, Tmp5).addReg(SHReg).addSImm(-32);
1954 BuildMI(BB, PPC::SRAW, 2, Tmp6).addReg(ShiftOpHi).addReg(Tmp5);
1955 BuildMI(BB, PPC::SRAW, 2, Result+1).addReg(ShiftOpHi).addReg(SHReg);
1956 BuildMI(BB, PPC::BLE, 2).addReg(PPC::CR0).addMBB(PhiMBB);
1957 // Select correct least significant half if the shift amount > 32
1959 unsigned Tmp7 = MakeReg(MVT::i32);
1960 BuildMI(BB, PPC::OR, 2, Tmp7).addReg(Tmp6).addReg(Tmp6);
1961 TmpMBB->addSuccessor(PhiMBB);
1963 BuildMI(BB, PPC::PHI, 4, Result).addReg(Tmp4).addMBB(OldMBB)
1964 .addReg(Tmp7).addMBB(TmpMBB);
1966 return Result+N.ResNo;
1969 case ISD::FP_TO_UINT:
1970 case ISD::FP_TO_SINT: {
1971 bool U = (ISD::FP_TO_UINT == opcode);
1972 Tmp1 = SelectExpr(N.getOperand(0));
1974 Tmp2 = MakeReg(MVT::f64);
1975 BuildMI(BB, PPC::FCTIWZ, 1, Tmp2).addReg(Tmp1);
1976 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
1977 addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(Tmp2), FrameIdx);
1978 addFrameReference(BuildMI(BB, PPC::LWZ, 2, Result), FrameIdx, 4);
1981 unsigned Zero = getConstDouble(0.0);
1982 unsigned MaxInt = getConstDouble((1LL << 32) - 1);
1983 unsigned Border = getConstDouble(1LL << 31);
1984 unsigned UseZero = MakeReg(MVT::f64);
1985 unsigned UseMaxInt = MakeReg(MVT::f64);
1986 unsigned UseChoice = MakeReg(MVT::f64);
1987 unsigned TmpReg = MakeReg(MVT::f64);
1988 unsigned TmpReg2 = MakeReg(MVT::f64);
1989 unsigned ConvReg = MakeReg(MVT::f64);
1990 unsigned IntTmp = MakeReg(MVT::i32);
1991 unsigned XorReg = MakeReg(MVT::i32);
1992 MachineFunction *F = BB->getParent();
1993 int FrameIdx = F->getFrameInfo()->CreateStackObject(8, 8);
1994 // Update machine-CFG edges
1995 MachineBasicBlock *XorMBB = new MachineBasicBlock(BB->getBasicBlock());
1996 MachineBasicBlock *PhiMBB = new MachineBasicBlock(BB->getBasicBlock());
1997 MachineBasicBlock *OldMBB = BB;
1998 ilist<MachineBasicBlock>::iterator It = BB; ++It;
1999 F->getBasicBlockList().insert(It, XorMBB);
2000 F->getBasicBlockList().insert(It, PhiMBB);
2001 BB->addSuccessor(XorMBB);
2002 BB->addSuccessor(PhiMBB);
2003 // Convert from floating point to unsigned 32-bit value
2004 // Use 0 if incoming value is < 0.0
2005 BuildMI(BB, PPC::FSEL, 3, UseZero).addReg(Tmp1).addReg(Tmp1).addReg(Zero);
2006 // Use 2**32 - 1 if incoming value is >= 2**32
2007 BuildMI(BB, PPC::FSUB, 2, UseMaxInt).addReg(MaxInt).addReg(Tmp1);
2008 BuildMI(BB, PPC::FSEL, 3, UseChoice).addReg(UseMaxInt).addReg(UseZero)
2011 BuildMI(BB, PPC::FSUB, 2, TmpReg).addReg(UseChoice).addReg(Border);
2012 // Use difference if >= 2**31
2013 BuildMI(BB, PPC::FCMPU, 2, PPC::CR0).addReg(UseChoice).addReg(Border);
2014 BuildMI(BB, PPC::FSEL, 3, TmpReg2).addReg(TmpReg).addReg(TmpReg)
2016 // Convert to integer
2017 BuildMI(BB, PPC::FCTIWZ, 1, ConvReg).addReg(TmpReg2);
2018 addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(ConvReg), FrameIdx);
2019 addFrameReference(BuildMI(BB, PPC::LWZ, 2, IntTmp), FrameIdx, 4);
2020 BuildMI(BB, PPC::BLT, 2).addReg(PPC::CR0).addMBB(PhiMBB);
2021 BuildMI(BB, PPC::B, 1).addMBB(XorMBB);
2024 // add 2**31 if input was >= 2**31
2026 BuildMI(BB, PPC::XORIS, 2, XorReg).addReg(IntTmp).addImm(0x8000);
2027 XorMBB->addSuccessor(PhiMBB);
2030 // DestReg = phi [ IntTmp, OldMBB ], [ XorReg, XorMBB ]
2032 BuildMI(BB, PPC::PHI, 4, Result).addReg(IntTmp).addMBB(OldMBB)
2033 .addReg(XorReg).addMBB(XorMBB);
2036 assert(0 && "Should never get here");
2041 if (SetCCSDNode *SetCC = dyn_cast<SetCCSDNode>(Node)) {
2042 if (ConstantSDNode *CN =
2043 dyn_cast<ConstantSDNode>(SetCC->getOperand(1).Val)) {
2044 // We can codegen setcc op, imm very efficiently compared to a brcond.
2045 // Check for those cases here.
2047 if (CN->getValue() == 0) {
2048 Tmp1 = SelectExpr(SetCC->getOperand(0));
2049 switch (SetCC->getCondition()) {
2050 default: SetCC->dump(); assert(0 && "Unhandled SetCC condition"); abort();
2052 Tmp2 = MakeReg(MVT::i32);
2053 BuildMI(BB, PPC::CNTLZW, 1, Tmp2).addReg(Tmp1);
2054 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp2).addImm(27)
2055 .addImm(5).addImm(31);
2058 Tmp2 = MakeReg(MVT::i32);
2059 BuildMI(BB, PPC::ADDIC, 2, Tmp2).addReg(Tmp1).addSImm(-1);
2060 BuildMI(BB, PPC::SUBFE, 2, Result).addReg(Tmp2).addReg(Tmp1);
2063 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(1)
2064 .addImm(31).addImm(31);
2067 Tmp2 = MakeReg(MVT::i32);
2068 Tmp3 = MakeReg(MVT::i32);
2069 BuildMI(BB, PPC::NEG, 2, Tmp2).addReg(Tmp1);
2070 BuildMI(BB, PPC::ANDC, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2071 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp3).addImm(1)
2072 .addImm(31).addImm(31);
2078 if (CN->isAllOnesValue()) {
2079 Tmp1 = SelectExpr(SetCC->getOperand(0));
2080 switch (SetCC->getCondition()) {
2081 default: assert(0 && "Unhandled SetCC condition"); abort();
2083 Tmp2 = MakeReg(MVT::i32);
2084 Tmp3 = MakeReg(MVT::i32);
2085 BuildMI(BB, PPC::ADDIC, 2, Tmp2).addReg(Tmp1).addSImm(1);
2086 BuildMI(BB, PPC::LI, 1, Tmp3).addSImm(0);
2087 BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp3);
2090 Tmp2 = MakeReg(MVT::i32);
2091 Tmp3 = MakeReg(MVT::i32);
2092 BuildMI(BB, PPC::NOR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
2093 BuildMI(BB, PPC::ADDIC, 2, Tmp3).addReg(Tmp2).addSImm(-1);
2094 BuildMI(BB, PPC::SUBFE, 2, Result).addReg(Tmp3).addReg(Tmp2);
2097 Tmp2 = MakeReg(MVT::i32);
2098 Tmp3 = MakeReg(MVT::i32);
2099 BuildMI(BB, PPC::ADDI, 2, Tmp2).addReg(Tmp1).addSImm(1);
2100 BuildMI(BB, PPC::AND, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2101 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp3).addImm(1)
2102 .addImm(31).addImm(31);
2105 Tmp2 = MakeReg(MVT::i32);
2106 BuildMI(BB, PPC::RLWINM, 4, Tmp2).addReg(Tmp1).addImm(1)
2107 .addImm(31).addImm(31);
2108 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp2).addImm(1);
2116 unsigned CCReg = SelectCC(N, Opc, Inv, Tmp2);
2117 MoveCRtoGPR(CCReg, Inv, Tmp2, Result);
2120 assert(0 && "Is this legal?");
2124 SetCCSDNode* SetCC = dyn_cast<SetCCSDNode>(N.getOperand(0).Val);
2125 if (SetCC && N.getOperand(0).getOpcode() == ISD::SETCC &&
2126 !MVT::isInteger(SetCC->getOperand(0).getValueType()) &&
2127 !MVT::isInteger(N.getOperand(1).getValueType()) &&
2128 !MVT::isInteger(N.getOperand(2).getValueType()) &&
2129 SetCC->getCondition() != ISD::SETEQ &&
2130 SetCC->getCondition() != ISD::SETNE) {
2131 MVT::ValueType VT = SetCC->getOperand(0).getValueType();
2132 unsigned TV = SelectExpr(N.getOperand(1)); // Use if TRUE
2133 unsigned FV = SelectExpr(N.getOperand(2)); // Use if FALSE
2135 ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(SetCC->getOperand(1));
2136 if (CN && (CN->isExactlyValue(-0.0) || CN->isExactlyValue(0.0))) {
2137 switch(SetCC->getCondition()) {
2138 default: assert(0 && "Invalid FSEL condition"); abort();
2141 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2144 Tmp1 = SelectExpr(SetCC->getOperand(0)); // Val to compare against
2145 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp1).addReg(TV).addReg(FV);
2149 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2152 if (SetCC->getOperand(0).getOpcode() == ISD::FNEG) {
2153 Tmp2 = SelectExpr(SetCC->getOperand(0).getOperand(0));
2156 Tmp1 = SelectExpr(SetCC->getOperand(0)); // Val to compare against
2157 BuildMI(BB, PPC::FNEG, 1, Tmp2).addReg(Tmp1);
2159 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp2).addReg(TV).addReg(FV);
2164 Opc = (MVT::f64 == VT) ? PPC::FSUB : PPC::FSUBS;
2165 Tmp1 = SelectExpr(SetCC->getOperand(0)); // Val to compare against
2166 Tmp2 = SelectExpr(SetCC->getOperand(1));
2168 switch(SetCC->getCondition()) {
2169 default: assert(0 && "Invalid FSEL condition"); abort();
2172 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
2173 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
2177 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
2178 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
2182 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2183 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
2187 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
2188 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
2192 assert(0 && "Should never get here");
2197 unsigned TrueValue = SelectExpr(N.getOperand(1)); //Use if TRUE
2198 unsigned FalseValue = SelectExpr(N.getOperand(2)); //Use if FALSE
2199 unsigned CCReg = SelectCC(N.getOperand(0), Opc, Inv, Tmp3);
2201 // Create an iterator with which to insert the MBB for copying the false
2202 // value and the MBB to hold the PHI instruction for this SetCC.
2203 MachineBasicBlock *thisMBB = BB;
2204 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2205 ilist<MachineBasicBlock>::iterator It = BB;
2211 // cmpTY ccX, r1, r2
2213 // fallthrough --> copy0MBB
2214 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
2215 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
2216 BuildMI(BB, Opc, 2).addReg(CCReg).addMBB(sinkMBB);
2217 MachineFunction *F = BB->getParent();
2218 F->getBasicBlockList().insert(It, copy0MBB);
2219 F->getBasicBlockList().insert(It, sinkMBB);
2220 // Update machine-CFG edges
2221 BB->addSuccessor(copy0MBB);
2222 BB->addSuccessor(sinkMBB);
2225 // %FalseValue = ...
2226 // # fallthrough to sinkMBB
2228 // Update machine-CFG edges
2229 BB->addSuccessor(sinkMBB);
2232 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2235 BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue)
2236 .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB);
2241 switch (N.getValueType()) {
2242 default: assert(0 && "Cannot use constants of this type!");
2244 BuildMI(BB, PPC::LI, 1, Result)
2245 .addSImm(!cast<ConstantSDNode>(N)->isNullValue());
2249 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
2250 if (v < 32768 && v >= -32768) {
2251 BuildMI(BB, PPC::LI, 1, Result).addSImm(v);
2253 Tmp1 = MakeReg(MVT::i32);
2254 BuildMI(BB, PPC::LIS, 1, Tmp1).addSImm(v >> 16);
2255 BuildMI(BB, PPC::ORI, 2, Result).addReg(Tmp1).addImm(v & 0xFFFF);
2261 case ISD::ConstantFP: {
2262 ConstantFPSDNode *CN = cast<ConstantFPSDNode>(N);
2263 Result = getConstDouble(CN->getValue(), Result);
2268 if (!NoExcessFPPrecision &&
2269 ISD::ADD == N.getOperand(0).getOpcode() &&
2270 N.getOperand(0).Val->hasOneUse() &&
2271 ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() &&
2272 N.getOperand(0).getOperand(0).Val->hasOneUse()) {
2273 ++FusedFP; // Statistic
2274 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0));
2275 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1));
2276 Tmp3 = SelectExpr(N.getOperand(0).getOperand(1));
2277 Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
2278 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
2279 } else if (!NoExcessFPPrecision &&
2280 ISD::ADD == N.getOperand(0).getOpcode() &&
2281 N.getOperand(0).Val->hasOneUse() &&
2282 ISD::MUL == N.getOperand(0).getOperand(1).getOpcode() &&
2283 N.getOperand(0).getOperand(1).Val->hasOneUse()) {
2284 ++FusedFP; // Statistic
2285 Tmp1 = SelectExpr(N.getOperand(0).getOperand(1).getOperand(0));
2286 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1).getOperand(1));
2287 Tmp3 = SelectExpr(N.getOperand(0).getOperand(0));
2288 Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
2289 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
2290 } else if (ISD::FABS == N.getOperand(0).getOpcode()) {
2291 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
2292 BuildMI(BB, PPC::FNABS, 1, Result).addReg(Tmp1);
2294 Tmp1 = SelectExpr(N.getOperand(0));
2295 BuildMI(BB, PPC::FNEG, 1, Result).addReg(Tmp1);
2300 Tmp1 = SelectExpr(N.getOperand(0));
2301 BuildMI(BB, PPC::FABS, 1, Result).addReg(Tmp1);
2305 Tmp1 = SelectExpr(N.getOperand(0));
2306 Opc = DestType == MVT::f64 ? PPC::FSQRT : PPC::FSQRTS;
2307 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
2311 assert (DestType == MVT::f32 &&
2312 N.getOperand(0).getValueType() == MVT::f64 &&
2313 "only f64 to f32 conversion supported here");
2314 Tmp1 = SelectExpr(N.getOperand(0));
2315 BuildMI(BB, PPC::FRSP, 1, Result).addReg(Tmp1);
2318 case ISD::FP_EXTEND:
2319 assert (DestType == MVT::f64 &&
2320 N.getOperand(0).getValueType() == MVT::f32 &&
2321 "only f32 to f64 conversion supported here");
2322 Tmp1 = SelectExpr(N.getOperand(0));
2323 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
2326 case ISD::UINT_TO_FP:
2327 case ISD::SINT_TO_FP: {
2328 assert (N.getOperand(0).getValueType() == MVT::i32
2329 && "int to float must operate on i32");
2330 bool IsUnsigned = (ISD::UINT_TO_FP == opcode);
2331 Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
2332 Tmp2 = MakeReg(MVT::f64); // temp reg to load the integer value into
2333 Tmp3 = MakeReg(MVT::i32); // temp reg to hold the conversion constant
2335 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
2336 MachineConstantPool *CP = BB->getParent()->getConstantPool();
2339 unsigned ConstF = getConstDouble(0x1.000000p52);
2340 // Store the hi & low halves of the fp value, currently in int regs
2341 BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330);
2342 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx);
2343 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp1), FrameIdx, 4);
2344 addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
2345 // Generate the return value with a subtract
2346 BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF);
2348 unsigned ConstF = getConstDouble(0x1.000008p52);
2349 unsigned TmpL = MakeReg(MVT::i32);
2350 // Store the hi & low halves of the fp value, currently in int regs
2351 BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330);
2352 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx);
2353 BuildMI(BB, PPC::XORIS, 2, TmpL).addReg(Tmp1).addImm(0x8000);
2354 addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(TmpL), FrameIdx, 4);
2355 addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
2356 // Generate the return value with a subtract
2357 BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF);
2365 void ISel::Select(SDOperand N) {
2366 unsigned Tmp1, Tmp2, Tmp3, Opc;
2367 unsigned opcode = N.getOpcode();
2369 if (!ExprMap.insert(std::make_pair(N, 1)).second)
2370 return; // Already selected.
2372 SDNode *Node = N.Val;
2374 switch (Node->getOpcode()) {
2376 Node->dump(); std::cerr << "\n";
2377 assert(0 && "Node not handled yet!");
2378 case ISD::EntryToken: return; // Noop
2379 case ISD::TokenFactor:
2380 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
2381 Select(Node->getOperand(i));
2383 case ISD::CALLSEQ_START:
2384 case ISD::CALLSEQ_END:
2385 Select(N.getOperand(0));
2386 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
2387 Opc = N.getOpcode() == ISD::CALLSEQ_START ? PPC::ADJCALLSTACKDOWN :
2388 PPC::ADJCALLSTACKUP;
2389 BuildMI(BB, Opc, 1).addImm(Tmp1);
2392 MachineBasicBlock *Dest =
2393 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
2394 Select(N.getOperand(0));
2395 BuildMI(BB, PPC::B, 1).addMBB(Dest);
2399 case ISD::BRCONDTWOWAY:
2402 case ISD::CopyToReg:
2403 Select(N.getOperand(0));
2404 Tmp1 = SelectExpr(N.getOperand(1));
2405 Tmp2 = cast<RegSDNode>(N)->getReg();
2408 if (N.getOperand(1).getValueType() == MVT::f64 ||
2409 N.getOperand(1).getValueType() == MVT::f32)
2410 BuildMI(BB, PPC::FMR, 1, Tmp2).addReg(Tmp1);
2412 BuildMI(BB, PPC::OR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
2415 case ISD::ImplicitDef:
2416 Select(N.getOperand(0));
2417 BuildMI(BB, PPC::IMPLICIT_DEF, 0, cast<RegSDNode>(N)->getReg());
2420 switch (N.getNumOperands()) {
2422 assert(0 && "Unknown return instruction!");
2424 assert(N.getOperand(1).getValueType() == MVT::i32 &&
2425 N.getOperand(2).getValueType() == MVT::i32 &&
2426 "Unknown two-register value!");
2427 Select(N.getOperand(0));
2428 Tmp1 = SelectExpr(N.getOperand(1));
2429 Tmp2 = SelectExpr(N.getOperand(2));
2430 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp2).addReg(Tmp2);
2431 BuildMI(BB, PPC::OR, 2, PPC::R4).addReg(Tmp1).addReg(Tmp1);
2434 Select(N.getOperand(0));
2435 Tmp1 = SelectExpr(N.getOperand(1));
2436 switch (N.getOperand(1).getValueType()) {
2438 assert(0 && "Unknown return type!");
2441 BuildMI(BB, PPC::FMR, 1, PPC::F1).addReg(Tmp1);
2444 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp1).addReg(Tmp1);
2448 Select(N.getOperand(0));
2451 BuildMI(BB, PPC::BLR, 0); // Just emit a 'ret' instruction
2453 case ISD::TRUNCSTORE:
2455 SDOperand Chain = N.getOperand(0);
2456 SDOperand Value = N.getOperand(1);
2457 SDOperand Address = N.getOperand(2);
2460 Tmp1 = SelectExpr(Value); //value
2462 if (opcode == ISD::STORE) {
2463 switch(Value.getValueType()) {
2464 default: assert(0 && "unknown Type in store");
2465 case MVT::i32: Opc = PPC::STW; break;
2466 case MVT::f64: Opc = PPC::STFD; break;
2467 case MVT::f32: Opc = PPC::STFS; break;
2469 } else { //ISD::TRUNCSTORE
2470 switch(cast<VTSDNode>(Node->getOperand(4))->getVT()) {
2471 default: assert(0 && "unknown Type in store");
2473 case MVT::i8: Opc = PPC::STB; break;
2474 case MVT::i16: Opc = PPC::STH; break;
2478 if(Address.getOpcode() == ISD::FrameIndex) {
2479 Tmp2 = cast<FrameIndexSDNode>(Address)->getIndex();
2480 addFrameReference(BuildMI(BB, Opc, 3).addReg(Tmp1), (int)Tmp2);
2481 } else if(GlobalAddressSDNode *GN = dyn_cast<GlobalAddressSDNode>(Address)){
2482 GlobalValue *GV = GN->getGlobal();
2483 Tmp2 = MakeReg(MVT::i32);
2485 BuildMI(BB, PPC::ADDIS, 2, Tmp2).addReg(getGlobalBaseReg())
2486 .addGlobalAddress(GV);
2488 BuildMI(BB, PPC::LIS, 2, Tmp2).addGlobalAddress(GV);
2489 if (GV->hasWeakLinkage() || GV->isExternal()) {
2490 Tmp3 = MakeReg(MVT::i32);
2491 BuildMI(BB, PPC::LWZ, 2, Tmp3).addGlobalAddress(GV).addReg(Tmp2);
2492 BuildMI(BB, Opc, 3).addReg(Tmp1).addSImm(0).addReg(Tmp3);
2494 BuildMI(BB, Opc, 3).addReg(Tmp1).addGlobalAddress(GV).addReg(Tmp2);
2498 bool idx = SelectAddr(Address, Tmp2, offset);
2500 Opc = IndexedOpForOp(Opc);
2501 BuildMI(BB, Opc, 3).addReg(Tmp1).addReg(Tmp2).addReg(offset);
2503 BuildMI(BB, Opc, 3).addReg(Tmp1).addImm(offset).addReg(Tmp2);
2512 case ISD::CopyFromReg:
2515 case ISD::DYNAMIC_STACKALLOC:
2520 assert(0 && "Should not be reached!");
2524 /// createPPC32PatternInstructionSelector - This pass converts an LLVM function
2525 /// into a machine code representation using pattern matching and a machine
2526 /// description file.
2528 FunctionPass *llvm::createPPC32ISelPattern(TargetMachine &TM) {
2529 return new ISel(TM);