1 //===-- PPCISelDAGToDAG.cpp - PPC --pattern matching inst selector --------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner 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 PowerPC,
11 // converting from a legalized dag to a PPC dag.
13 //===----------------------------------------------------------------------===//
16 #include "PPCTargetMachine.h"
17 #include "PPCISelLowering.h"
18 #include "PPCHazardRecognizers.h"
19 #include "llvm/CodeGen/MachineInstrBuilder.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/SSARegMap.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/CodeGen/SelectionDAGISel.h"
24 #include "llvm/Target/TargetOptions.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/Constants.h"
27 #include "llvm/GlobalValue.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/MathExtras.h"
35 Statistic<> FrameOff("ppc-codegen", "Number of frame idx offsets collapsed");
37 //===--------------------------------------------------------------------===//
38 /// PPCDAGToDAGISel - PPC specific code to select PPC machine
39 /// instructions for SelectionDAG operations.
41 class PPCDAGToDAGISel : public SelectionDAGISel {
42 PPCTargetLowering PPCLowering;
43 unsigned GlobalBaseReg;
45 PPCDAGToDAGISel(TargetMachine &TM)
46 : SelectionDAGISel(PPCLowering), PPCLowering(TM) {}
48 virtual bool runOnFunction(Function &Fn) {
49 // Make sure we re-emit a set of the global base reg if necessary
51 return SelectionDAGISel::runOnFunction(Fn);
54 /// getI32Imm - Return a target constant with the specified value, of type
56 inline SDOperand getI32Imm(unsigned Imm) {
57 return CurDAG->getTargetConstant(Imm, MVT::i32);
60 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
61 /// base register. Return the virtual register that holds this value.
62 SDOperand getGlobalBaseReg();
64 // Select - Convert the specified operand from a target-independent to a
65 // target-specific node if it hasn't already been changed.
66 void Select(SDOperand &Result, SDOperand Op);
68 SDNode *SelectBitfieldInsert(SDNode *N);
70 /// SelectCC - Select a comparison of the specified values with the
71 /// specified condition code, returning the CR# of the expression.
72 SDOperand SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC);
74 /// SelectAddrImm - Returns true if the address N can be represented by
75 /// a base register plus a signed 16-bit displacement [r+imm].
76 bool SelectAddrImm(SDOperand N, SDOperand &Disp, SDOperand &Base);
78 /// SelectAddrIdx - Given the specified addressed, check to see if it can be
79 /// represented as an indexed [r+r] operation. Returns false if it can
80 /// be represented by [r+imm], which are preferred.
81 bool SelectAddrIdx(SDOperand N, SDOperand &Base, SDOperand &Index);
83 /// SelectAddrIdxOnly - Given the specified addressed, force it to be
84 /// represented as an indexed [r+r] operation.
85 bool SelectAddrIdxOnly(SDOperand N, SDOperand &Base, SDOperand &Index);
87 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
88 /// inline asm expressions.
89 virtual bool SelectInlineAsmMemoryOperand(const SDOperand &Op,
91 std::vector<SDOperand> &OutOps,
94 switch (ConstraintCode) {
97 if (!SelectAddrIdx(Op, Op0, Op1))
98 SelectAddrImm(Op, Op0, Op1);
100 case 'o': // offsetable
101 if (!SelectAddrImm(Op, Op0, Op1)) {
102 Select(Op0, Op); // r+0.
106 case 'v': // not offsetable
107 SelectAddrIdxOnly(Op, Op0, Op1);
111 OutOps.push_back(Op0);
112 OutOps.push_back(Op1);
116 SDOperand BuildSDIVSequence(SDNode *N);
117 SDOperand BuildUDIVSequence(SDNode *N);
119 /// InstructionSelectBasicBlock - This callback is invoked by
120 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
121 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
123 virtual const char *getPassName() const {
124 return "PowerPC DAG->DAG Pattern Instruction Selection";
127 /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for this
128 /// target when scheduling the DAG.
129 virtual HazardRecognizer *CreateTargetHazardRecognizer() {
130 // Should use subtarget info to pick the right hazard recognizer. For
131 // now, always return a PPC970 recognizer.
132 const TargetInstrInfo *II = PPCLowering.getTargetMachine().getInstrInfo();
133 assert(II && "No InstrInfo?");
134 return new PPCHazardRecognizer970(*II);
137 // Include the pieces autogenerated from the target description.
138 #include "PPCGenDAGISel.inc"
141 SDOperand SelectSETCC(SDOperand Op);
142 SDOperand SelectCALL(SDOperand Op);
146 /// InstructionSelectBasicBlock - This callback is invoked by
147 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
148 void PPCDAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
151 // The selection process is inherently a bottom-up recursive process (users
152 // select their uses before themselves). Given infinite stack space, we
153 // could just start selecting on the root and traverse the whole graph. In
154 // practice however, this causes us to run out of stack space on large basic
155 // blocks. To avoid this problem, select the entry node, then all its uses,
156 // iteratively instead of recursively.
157 std::vector<SDOperand> Worklist;
158 Worklist.push_back(DAG.getEntryNode());
160 // Note that we can do this in the PPC target (scanning forward across token
161 // chain edges) because no nodes ever get folded across these edges. On a
162 // target like X86 which supports load/modify/store operations, this would
163 // have to be more careful.
164 while (!Worklist.empty()) {
165 SDOperand Node = Worklist.back();
168 // Chose from the least deep of the top two nodes.
169 if (!Worklist.empty() &&
170 Worklist.back().Val->getNodeDepth() < Node.Val->getNodeDepth())
171 std::swap(Worklist.back(), Node);
173 if ((Node.Val->getOpcode() >= ISD::BUILTIN_OP_END &&
174 Node.Val->getOpcode() < PPCISD::FIRST_NUMBER) ||
175 CodeGenMap.count(Node)) continue;
177 for (SDNode::use_iterator UI = Node.Val->use_begin(),
178 E = Node.Val->use_end(); UI != E; ++UI) {
179 // Scan the values. If this use has a value that is a token chain, add it
182 for (unsigned i = 0, e = User->getNumValues(); i != e; ++i)
183 if (User->getValueType(i) == MVT::Other) {
184 Worklist.push_back(SDOperand(User, i));
189 // Finally, legalize this node.
194 // Select target instructions for the DAG.
195 DAG.setRoot(SelectRoot(DAG.getRoot()));
197 DAG.RemoveDeadNodes();
199 // Emit machine code to BB.
200 ScheduleAndEmitDAG(DAG);
203 /// getGlobalBaseReg - Output the instructions required to put the
204 /// base address to use for accessing globals into a register.
206 SDOperand PPCDAGToDAGISel::getGlobalBaseReg() {
207 if (!GlobalBaseReg) {
208 // Insert the set of GlobalBaseReg into the first MBB of the function
209 MachineBasicBlock &FirstMBB = BB->getParent()->front();
210 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
211 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
212 // FIXME: when we get to LP64, we will need to create the appropriate
213 // type of register here.
214 GlobalBaseReg = RegMap->createVirtualRegister(PPC::GPRCRegisterClass);
215 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
216 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg);
218 return CurDAG->getRegister(GlobalBaseReg, MVT::i32);
222 // isIntImmediate - This method tests to see if a constant operand.
223 // If so Imm will receive the 32 bit value.
224 static bool isIntImmediate(SDNode *N, unsigned& Imm) {
225 if (N->getOpcode() == ISD::Constant) {
226 Imm = cast<ConstantSDNode>(N)->getValue();
232 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
233 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
234 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
235 // not, since all 1s are not contiguous.
236 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
237 if (isShiftedMask_32(Val)) {
238 // look for the first non-zero bit
239 MB = CountLeadingZeros_32(Val);
240 // look for the first zero bit after the run of ones
241 ME = CountLeadingZeros_32((Val - 1) ^ Val);
244 Val = ~Val; // invert mask
245 if (isShiftedMask_32(Val)) {
246 // effectively look for the first zero bit
247 ME = CountLeadingZeros_32(Val) - 1;
248 // effectively look for the first one bit after the run of zeros
249 MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
257 // isRotateAndMask - Returns true if Mask and Shift can be folded into a rotate
258 // and mask opcode and mask operation.
259 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool IsShiftMask,
260 unsigned &SH, unsigned &MB, unsigned &ME) {
261 // Don't even go down this path for i64, since different logic will be
262 // necessary for rldicl/rldicr/rldimi.
263 if (N->getValueType(0) != MVT::i32)
267 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
268 unsigned Opcode = N->getOpcode();
269 if (N->getNumOperands() != 2 ||
270 !isIntImmediate(N->getOperand(1).Val, Shift) || (Shift > 31))
273 if (Opcode == ISD::SHL) {
274 // apply shift left to mask if it comes first
275 if (IsShiftMask) Mask = Mask << Shift;
276 // determine which bits are made indeterminant by shift
277 Indeterminant = ~(0xFFFFFFFFu << Shift);
278 } else if (Opcode == ISD::SRL) {
279 // apply shift right to mask if it comes first
280 if (IsShiftMask) Mask = Mask >> Shift;
281 // determine which bits are made indeterminant by shift
282 Indeterminant = ~(0xFFFFFFFFu >> Shift);
283 // adjust for the left rotate
289 // if the mask doesn't intersect any Indeterminant bits
290 if (Mask && !(Mask & Indeterminant)) {
292 // make sure the mask is still a mask (wrap arounds may not be)
293 return isRunOfOnes(Mask, MB, ME);
298 // isOpcWithIntImmediate - This method tests to see if the node is a specific
299 // opcode and that it has a immediate integer right operand.
300 // If so Imm will receive the 32 bit value.
301 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
302 return N->getOpcode() == Opc && isIntImmediate(N->getOperand(1).Val, Imm);
305 // isIntImmediate - This method tests to see if a constant operand.
306 // If so Imm will receive the 32 bit value.
307 static bool isIntImmediate(SDOperand N, unsigned& Imm) {
308 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
309 Imm = (unsigned)CN->getSignExtended();
315 /// SelectBitfieldInsert - turn an or of two masked values into
316 /// the rotate left word immediate then mask insert (rlwimi) instruction.
317 /// Returns true on success, false if the caller still needs to select OR.
319 /// Patterns matched:
320 /// 1. or shl, and 5. or and, and
321 /// 2. or and, shl 6. or shl, shr
322 /// 3. or shr, and 7. or shr, shl
324 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
325 bool IsRotate = false;
326 unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, SH = 0;
329 SDOperand Op0 = N->getOperand(0);
330 SDOperand Op1 = N->getOperand(1);
332 unsigned Op0Opc = Op0.getOpcode();
333 unsigned Op1Opc = Op1.getOpcode();
335 // Verify that we have the correct opcodes
336 if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc)
338 if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc)
341 // Generate Mask value for Target
342 if (isIntImmediate(Op0.getOperand(1), Value)) {
344 case ISD::SHL: TgtMask <<= Value; break;
345 case ISD::SRL: TgtMask >>= Value; break;
346 case ISD::AND: TgtMask &= Value; break;
352 // Generate Mask value for Insert
353 if (!isIntImmediate(Op1.getOperand(1), Value))
360 if (Op0Opc == ISD::SRL) IsRotate = true;
366 if (Op0Opc == ISD::SHL) IsRotate = true;
373 // If both of the inputs are ANDs and one of them has a logical shift by
374 // constant as its input, make that AND the inserted value so that we can
375 // combine the shift into the rotate part of the rlwimi instruction
376 bool IsAndWithShiftOp = false;
377 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
378 if (Op1.getOperand(0).getOpcode() == ISD::SHL ||
379 Op1.getOperand(0).getOpcode() == ISD::SRL) {
380 if (isIntImmediate(Op1.getOperand(0).getOperand(1), Value)) {
381 SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
382 IsAndWithShiftOp = true;
384 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
385 Op0.getOperand(0).getOpcode() == ISD::SRL) {
386 if (isIntImmediate(Op0.getOperand(0).getOperand(1), Value)) {
388 std::swap(TgtMask, InsMask);
389 SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
390 IsAndWithShiftOp = true;
395 // Verify that the Target mask and Insert mask together form a full word mask
396 // and that the Insert mask is a run of set bits (which implies both are runs
397 // of set bits). Given that, Select the arguments and generate the rlwimi
400 if (((TgtMask & InsMask) == 0) && isRunOfOnes(InsMask, MB, ME)) {
401 bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF;
402 bool Op0IsAND = Op0Opc == ISD::AND;
403 // Check for rotlwi / rotrwi here, a special case of bitfield insert
404 // where both bitfield halves are sourced from the same value.
405 if (IsRotate && fullMask &&
406 N->getOperand(0).getOperand(0) == N->getOperand(1).getOperand(0)) {
408 Select(Tmp, N->getOperand(0).getOperand(0));
409 return CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, Tmp,
410 getI32Imm(SH), getI32Imm(0), getI32Imm(31));
412 SDOperand Tmp1, Tmp2;
413 Select(Tmp1, ((Op0IsAND && fullMask) ? Op0.getOperand(0) : Op0));
414 Select(Tmp2, (IsAndWithShiftOp ? Op1.getOperand(0).getOperand(0)
415 : Op1.getOperand(0)));
416 return CurDAG->getTargetNode(PPC::RLWIMI, MVT::i32, Tmp1, Tmp2,
417 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME));
422 /// SelectAddrImm - Returns true if the address N can be represented by
423 /// a base register plus a signed 16-bit displacement [r+imm].
424 bool PPCDAGToDAGISel::SelectAddrImm(SDOperand N, SDOperand &Disp,
426 // If this can be more profitably realized as r+r, fail.
427 if (SelectAddrIdx(N, Disp, Base))
430 if (N.getOpcode() == ISD::ADD) {
432 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm)) {
433 Disp = getI32Imm(imm & 0xFFFF);
434 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
435 Base = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
437 Base = N.getOperand(0);
439 return true; // [r+i]
440 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
441 // Match LOAD (ADD (X, Lo(G))).
442 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
443 && "Cannot handle constant offsets yet!");
444 Disp = N.getOperand(1).getOperand(0); // The global address.
445 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
446 Disp.getOpcode() == ISD::TargetConstantPool);
447 Base = N.getOperand(0);
448 return true; // [&g+r]
450 } else if (N.getOpcode() == ISD::OR) {
452 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm)) {
453 // If this is an or of disjoint bitfields, we can codegen this as an add
454 // (for better address arithmetic) if the LHS and RHS of the OR are
455 // provably disjoint.
456 uint64_t LHSKnownZero, LHSKnownOne;
457 PPCLowering.ComputeMaskedBits(N.getOperand(0), ~0U,
458 LHSKnownZero, LHSKnownOne);
459 if ((LHSKnownZero|~imm) == ~0U) {
460 // If all of the bits are known zero on the LHS or RHS, the add won't
462 Base = N.getOperand(0);
463 Disp = getI32Imm(imm & 0xFFFF);
469 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
470 Base = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
473 return true; // [r+0]
476 /// SelectAddrIdx - Given the specified addressed, check to see if it can be
477 /// represented as an indexed [r+r] operation. Returns false if it can
478 /// be represented by [r+imm], which are preferred.
479 bool PPCDAGToDAGISel::SelectAddrIdx(SDOperand N, SDOperand &Base,
482 if (N.getOpcode() == ISD::ADD) {
483 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm))
485 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
488 Base = N.getOperand(0);
489 Index = N.getOperand(1);
491 } else if (N.getOpcode() == ISD::OR) {
492 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm))
493 return false; // r+i can fold it if we can.
495 // If this is an or of disjoint bitfields, we can codegen this as an add
496 // (for better address arithmetic) if the LHS and RHS of the OR are provably
498 uint64_t LHSKnownZero, LHSKnownOne;
499 uint64_t RHSKnownZero, RHSKnownOne;
500 PPCLowering.ComputeMaskedBits(N.getOperand(0), ~0U,
501 LHSKnownZero, LHSKnownOne);
504 PPCLowering.ComputeMaskedBits(N.getOperand(1), ~0U,
505 RHSKnownZero, RHSKnownOne);
506 // If all of the bits are known zero on the LHS or RHS, the add won't
508 if ((LHSKnownZero | RHSKnownZero) == ~0U) {
509 Base = N.getOperand(0);
510 Index = N.getOperand(1);
519 /// SelectAddrIdxOnly - Given the specified addressed, force it to be
520 /// represented as an indexed [r+r] operation.
521 bool PPCDAGToDAGISel::SelectAddrIdxOnly(SDOperand N, SDOperand &Base,
523 // Check to see if we can easily represent this as an [r+r] address. This
524 // will fail if it thinks that the address is more profitably represented as
525 // reg+imm, e.g. where imm = 0.
526 if (!SelectAddrIdx(N, Base, Index)) {
527 // Nope, do it the hard way.
528 Base = CurDAG->getRegister(PPC::R0, MVT::i32);
534 /// SelectCC - Select a comparison of the specified values with the specified
535 /// condition code, returning the CR# of the expression.
536 SDOperand PPCDAGToDAGISel::SelectCC(SDOperand LHS, SDOperand RHS,
538 // Always select the LHS.
541 // Use U to determine whether the SETCC immediate range is signed or not.
542 if (MVT::isInteger(LHS.getValueType())) {
543 bool U = ISD::isUnsignedIntSetCC(CC);
545 if (isIntImmediate(RHS, Imm) &&
546 ((U && isUInt16(Imm)) || (!U && isInt16(Imm))))
547 return SDOperand(CurDAG->getTargetNode(U ? PPC::CMPLWI : PPC::CMPWI,
548 MVT::i32, LHS, getI32Imm(Imm & 0xFFFF)), 0);
550 return SDOperand(CurDAG->getTargetNode(U ? PPC::CMPLW : PPC::CMPW, MVT::i32,
552 } else if (LHS.getValueType() == MVT::f32) {
554 return SDOperand(CurDAG->getTargetNode(PPC::FCMPUS, MVT::i32, LHS, RHS), 0);
557 return SDOperand(CurDAG->getTargetNode(PPC::FCMPUD, MVT::i32, LHS, RHS), 0);
561 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
563 static unsigned getBCCForSetCC(ISD::CondCode CC) {
565 default: assert(0 && "Unknown condition!"); abort();
566 case ISD::SETOEQ: // FIXME: This is incorrect see PR642.
567 case ISD::SETEQ: return PPC::BEQ;
568 case ISD::SETONE: // FIXME: This is incorrect see PR642.
569 case ISD::SETNE: return PPC::BNE;
570 case ISD::SETOLT: // FIXME: This is incorrect see PR642.
572 case ISD::SETLT: return PPC::BLT;
573 case ISD::SETOLE: // FIXME: This is incorrect see PR642.
575 case ISD::SETLE: return PPC::BLE;
576 case ISD::SETOGT: // FIXME: This is incorrect see PR642.
578 case ISD::SETGT: return PPC::BGT;
579 case ISD::SETOGE: // FIXME: This is incorrect see PR642.
581 case ISD::SETGE: return PPC::BGE;
583 case ISD::SETO: return PPC::BUN;
584 case ISD::SETUO: return PPC::BNU;
589 /// getCRIdxForSetCC - Return the index of the condition register field
590 /// associated with the SetCC condition, and whether or not the field is
591 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
592 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool& Inv) {
594 default: assert(0 && "Unknown condition!"); abort();
595 case ISD::SETOLT: // FIXME: This is incorrect see PR642.
597 case ISD::SETLT: Inv = false; return 0;
598 case ISD::SETOGE: // FIXME: This is incorrect see PR642.
600 case ISD::SETGE: Inv = true; return 0;
601 case ISD::SETOGT: // FIXME: This is incorrect see PR642.
603 case ISD::SETGT: Inv = false; return 1;
604 case ISD::SETOLE: // FIXME: This is incorrect see PR642.
606 case ISD::SETLE: Inv = true; return 1;
607 case ISD::SETOEQ: // FIXME: This is incorrect see PR642.
608 case ISD::SETEQ: Inv = false; return 2;
609 case ISD::SETONE: // FIXME: This is incorrect see PR642.
610 case ISD::SETNE: Inv = true; return 2;
611 case ISD::SETO: Inv = true; return 3;
612 case ISD::SETUO: Inv = false; return 3;
617 SDOperand PPCDAGToDAGISel::SelectSETCC(SDOperand Op) {
620 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
621 if (isIntImmediate(N->getOperand(1), Imm)) {
622 // We can codegen setcc op, imm very efficiently compared to a brcond.
623 // Check for those cases here.
627 Select(Op, N->getOperand(0));
631 Op = SDOperand(CurDAG->getTargetNode(PPC::CNTLZW, MVT::i32, Op), 0);
632 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(27),
633 getI32Imm(5), getI32Imm(31));
636 SDOperand(CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
637 Op, getI32Imm(~0U)), 0);
638 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
642 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(1),
643 getI32Imm(31), getI32Imm(31));
646 SDOperand(CurDAG->getTargetNode(PPC::NEG, MVT::i32, Op), 0);
647 T = SDOperand(CurDAG->getTargetNode(PPC::ANDC, MVT::i32, T, Op), 0);
648 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, T, getI32Imm(1),
649 getI32Imm(31), getI32Imm(31));
652 } else if (Imm == ~0U) { // setcc op, -1
654 Select(Op, N->getOperand(0));
658 Op = SDOperand(CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
659 Op, getI32Imm(1)), 0);
660 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
661 SDOperand(CurDAG->getTargetNode(PPC::LI, MVT::i32,
665 Op = SDOperand(CurDAG->getTargetNode(PPC::NOR, MVT::i32, Op, Op), 0);
666 SDNode *AD = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
668 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDOperand(AD, 0), Op,
672 SDOperand AD = SDOperand(CurDAG->getTargetNode(PPC::ADDI, MVT::i32, Op,
674 SDOperand AN = SDOperand(CurDAG->getTargetNode(PPC::AND, MVT::i32, AD,
676 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, AN, getI32Imm(1),
677 getI32Imm(31), getI32Imm(31));
680 Op = SDOperand(CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, Op,
681 getI32Imm(1), getI32Imm(31),
683 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, getI32Imm(1));
689 unsigned Idx = getCRIdxForSetCC(CC, Inv);
690 SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
693 // Force the ccreg into CR7.
694 SDOperand CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
696 SDOperand InFlag(0, 0); // Null incoming flag value.
697 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), CR7Reg, CCReg,
700 if (TLI.getTargetMachine().getSubtarget<PPCSubtarget>().isGigaProcessor())
701 IntCR = SDOperand(CurDAG->getTargetNode(PPC::MFOCRF, MVT::i32, CR7Reg,
704 IntCR = SDOperand(CurDAG->getTargetNode(PPC::MFCR, MVT::i32, CCReg), 0);
707 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, IntCR,
708 getI32Imm((32-(3-Idx)) & 31),
709 getI32Imm(31), getI32Imm(31));
712 SDOperand(CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, IntCR,
713 getI32Imm((32-(3-Idx)) & 31),
714 getI32Imm(31),getI32Imm(31)), 0);
715 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
719 /// isCallCompatibleAddress - Return true if the specified 32-bit value is
720 /// representable in the immediate field of a Bx instruction.
721 static bool isCallCompatibleAddress(ConstantSDNode *C) {
722 int Addr = C->getValue();
723 if (Addr & 3) return false; // Low 2 bits are implicitly zero.
724 return (Addr << 6 >> 6) == Addr; // Top 6 bits have to be sext of immediate.
727 SDOperand PPCDAGToDAGISel::SelectCALL(SDOperand Op) {
730 Select(Chain, N->getOperand(0));
733 std::vector<SDOperand> CallOperands;
735 if (GlobalAddressSDNode *GASD =
736 dyn_cast<GlobalAddressSDNode>(N->getOperand(1))) {
737 CallOpcode = PPC::BL;
738 CallOperands.push_back(N->getOperand(1));
739 } else if (ExternalSymbolSDNode *ESSDN =
740 dyn_cast<ExternalSymbolSDNode>(N->getOperand(1))) {
741 CallOpcode = PPC::BL;
742 CallOperands.push_back(N->getOperand(1));
743 } else if (isa<ConstantSDNode>(N->getOperand(1)) &&
744 isCallCompatibleAddress(cast<ConstantSDNode>(N->getOperand(1)))) {
745 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(1));
746 CallOpcode = PPC::BLA;
747 CallOperands.push_back(getI32Imm((int)C->getValue() >> 2));
749 // Copy the callee address into the CTR register.
751 Select(Callee, N->getOperand(1));
752 Chain = SDOperand(CurDAG->getTargetNode(PPC::MTCTR, MVT::Other, Callee,
755 // Copy the callee address into R12 on darwin.
756 SDOperand R12 = CurDAG->getRegister(PPC::R12, MVT::i32);
757 Chain = CurDAG->getNode(ISD::CopyToReg, MVT::Other, Chain, R12, Callee);
759 CallOperands.push_back(R12);
760 CallOpcode = PPC::BCTRL;
763 unsigned GPR_idx = 0, FPR_idx = 0;
764 static const unsigned GPR[] = {
765 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
766 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
768 static const unsigned FPR[] = {
769 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
770 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
773 SDOperand InFlag; // Null incoming flag value.
775 for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i) {
776 unsigned DestReg = 0;
777 MVT::ValueType RegTy = N->getOperand(i).getValueType();
778 if (RegTy == MVT::i32) {
779 assert(GPR_idx < 8 && "Too many int args");
780 DestReg = GPR[GPR_idx++];
782 assert(MVT::isFloatingPoint(N->getOperand(i).getValueType()) &&
783 "Unpromoted integer arg?");
784 assert(FPR_idx < 13 && "Too many fp args");
785 DestReg = FPR[FPR_idx++];
788 if (N->getOperand(i).getOpcode() != ISD::UNDEF) {
790 Select(Val, N->getOperand(i));
791 Chain = CurDAG->getCopyToReg(Chain, DestReg, Val, InFlag);
792 InFlag = Chain.getValue(1);
793 CallOperands.push_back(CurDAG->getRegister(DestReg, RegTy));
797 // Finally, once everything is in registers to pass to the call, emit the
800 CallOperands.push_back(InFlag); // Strong dep on register copies.
802 CallOperands.push_back(Chain); // Weak dep on whatever occurs before
803 Chain = SDOperand(CurDAG->getTargetNode(CallOpcode, MVT::Other, MVT::Flag,
806 std::vector<SDOperand> CallResults;
808 // If the call has results, copy the values out of the ret val registers.
809 switch (N->getValueType(0)) {
810 default: assert(0 && "Unexpected ret value!");
811 case MVT::Other: break;
813 if (N->getValueType(1) == MVT::i32) {
814 Chain = CurDAG->getCopyFromReg(Chain, PPC::R4, MVT::i32,
815 Chain.getValue(1)).getValue(1);
816 CallResults.push_back(Chain.getValue(0));
817 Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
818 Chain.getValue(2)).getValue(1);
819 CallResults.push_back(Chain.getValue(0));
821 Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
822 Chain.getValue(1)).getValue(1);
823 CallResults.push_back(Chain.getValue(0));
828 Chain = CurDAG->getCopyFromReg(Chain, PPC::F1, N->getValueType(0),
829 Chain.getValue(1)).getValue(1);
830 CallResults.push_back(Chain.getValue(0));
834 CallResults.push_back(Chain);
835 for (unsigned i = 0, e = CallResults.size(); i != e; ++i)
836 CodeGenMap[Op.getValue(i)] = CallResults[i];
837 return CallResults[Op.ResNo];
840 // Select - Convert the specified operand from a target-independent to a
841 // target-specific node if it hasn't already been changed.
842 void PPCDAGToDAGISel::Select(SDOperand &Result, SDOperand Op) {
844 if (N->getOpcode() >= ISD::BUILTIN_OP_END &&
845 N->getOpcode() < PPCISD::FIRST_NUMBER) {
847 return; // Already selected.
850 // If this has already been converted, use it.
851 std::map<SDOperand, SDOperand>::iterator CGMI = CodeGenMap.find(Op);
852 if (CGMI != CodeGenMap.end()) {
853 Result = CGMI->second;
857 switch (N->getOpcode()) {
860 Result = SelectSETCC(Op);
863 Result = SelectCALL(Op);
865 case PPCISD::GlobalBaseReg:
866 Result = getGlobalBaseReg();
869 case ISD::FrameIndex: {
870 int FI = cast<FrameIndexSDNode>(N)->getIndex();
871 if (N->hasOneUse()) {
872 Result = CurDAG->SelectNodeTo(N, PPC::ADDI, MVT::i32,
873 CurDAG->getTargetFrameIndex(FI, MVT::i32),
877 Result = CodeGenMap[Op] =
878 SDOperand(CurDAG->getTargetNode(PPC::ADDI, MVT::i32,
879 CurDAG->getTargetFrameIndex(FI, MVT::i32),
884 // FIXME: since this depends on the setting of the carry flag from the srawi
885 // we should really be making notes about that for the scheduler.
886 // FIXME: It sure would be nice if we could cheaply recognize the
887 // srl/add/sra pattern the dag combiner will generate for this as
888 // sra/addze rather than having to handle sdiv ourselves. oh well.
890 if (isIntImmediate(N->getOperand(1), Imm)) {
892 Select(N0, N->getOperand(0));
893 if ((signed)Imm > 0 && isPowerOf2_32(Imm)) {
895 CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
896 N0, getI32Imm(Log2_32(Imm)));
897 Result = CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
898 SDOperand(Op, 0), SDOperand(Op, 1));
899 } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) {
901 CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
902 N0, getI32Imm(Log2_32(-Imm)));
904 SDOperand(CurDAG->getTargetNode(PPC::ADDZE, MVT::i32,
905 SDOperand(Op, 0), SDOperand(Op, 1)),
907 Result = CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT);
912 // Other cases are autogenerated.
917 // If this is an and of a value rotated between 0 and 31 bits and then and'd
918 // with a mask, emit rlwinm
919 if (isIntImmediate(N->getOperand(1), Imm) && (isShiftedMask_32(Imm) ||
920 isShiftedMask_32(~Imm))) {
923 if (isRotateAndMask(N->getOperand(0).Val, Imm, false, SH, MB, ME)) {
924 Select(Val, N->getOperand(0).getOperand(0));
925 } else if (Imm == 0) {
926 // AND X, 0 -> 0, not "rlwinm 32".
927 Select(Result, N->getOperand(1));
930 Select(Val, N->getOperand(0));
931 isRunOfOnes(Imm, MB, ME);
934 Result = CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Val,
935 getI32Imm(SH), getI32Imm(MB),
939 // ISD::OR doesn't get all the bitfield insertion fun.
940 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert
941 if (isIntImmediate(N->getOperand(1), Imm) &&
942 N->getOperand(0).getOpcode() == ISD::OR &&
943 isIntImmediate(N->getOperand(0).getOperand(1), Imm2)) {
946 if (isRunOfOnes(Imm, MB, ME)) {
947 SDOperand Tmp1, Tmp2;
948 Select(Tmp1, N->getOperand(0).getOperand(0));
949 Select(Tmp2, N->getOperand(0).getOperand(1));
950 Result = SDOperand(CurDAG->getTargetNode(PPC::RLWIMI, MVT::i32,
952 getI32Imm(0), getI32Imm(MB),
958 // Other cases are autogenerated.
962 if (SDNode *I = SelectBitfieldInsert(N)) {
963 Result = CodeGenMap[Op] = SDOperand(I, 0);
967 // Other cases are autogenerated.
970 unsigned Imm, SH, MB, ME;
971 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
972 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
974 Select(Val, N->getOperand(0).getOperand(0));
975 Result = CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
976 Val, getI32Imm(SH), getI32Imm(MB),
981 // Other cases are autogenerated.
985 unsigned Imm, SH, MB, ME;
986 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
987 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
989 Select(Val, N->getOperand(0).getOperand(0));
990 Result = CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
991 Val, getI32Imm(SH & 0x1F), getI32Imm(MB),
996 // Other cases are autogenerated.
999 case ISD::SELECT_CC: {
1000 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1002 // handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
1003 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1004 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
1005 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
1006 if (N1C->isNullValue() && N3C->isNullValue() &&
1007 N2C->getValue() == 1ULL && CC == ISD::SETNE) {
1009 Select(LHS, N->getOperand(0));
1011 CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
1012 LHS, getI32Imm(~0U));
1013 Result = CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32,
1014 SDOperand(Tmp, 0), LHS,
1019 SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
1020 unsigned BROpc = getBCCForSetCC(CC);
1022 bool isFP = MVT::isFloatingPoint(N->getValueType(0));
1023 unsigned SelectCCOp;
1024 if (MVT::isInteger(N->getValueType(0)))
1025 SelectCCOp = PPC::SELECT_CC_Int;
1026 else if (N->getValueType(0) == MVT::f32)
1027 SelectCCOp = PPC::SELECT_CC_F4;
1029 SelectCCOp = PPC::SELECT_CC_F8;
1031 Select(N2, N->getOperand(2));
1032 Select(N3, N->getOperand(3));
1033 Result = CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), CCReg,
1034 N2, N3, getI32Imm(BROpc));
1038 case ISD::BRTWOWAY_CC: {
1040 Select(Chain, N->getOperand(0));
1041 MachineBasicBlock *Dest =
1042 cast<BasicBlockSDNode>(N->getOperand(4))->getBasicBlock();
1043 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
1044 SDOperand CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC);
1046 // If this is a two way branch, then grab the fallthrough basic block
1047 // argument and build a PowerPC branch pseudo-op, suitable for long branch
1048 // conversion if necessary by the branch selection pass. Otherwise, emit a
1049 // standard conditional branch.
1050 if (N->getOpcode() == ISD::BRTWOWAY_CC) {
1051 SDOperand CondTrueBlock = N->getOperand(4);
1052 SDOperand CondFalseBlock = N->getOperand(5);
1053 unsigned Opc = getBCCForSetCC(CC);
1055 SDOperand(CurDAG->getTargetNode(PPC::COND_BRANCH, MVT::Other,
1056 CondCode, getI32Imm(Opc),
1057 CondTrueBlock, CondFalseBlock,
1059 Result = CurDAG->SelectNodeTo(N, PPC::B, MVT::Other, CondFalseBlock, CB);
1061 // Iterate to the next basic block
1062 ilist<MachineBasicBlock>::iterator It = BB;
1065 // If the fallthrough path is off the end of the function, which would be
1066 // undefined behavior, set it to be the same as the current block because
1067 // we have nothing better to set it to, and leaving it alone will cause
1068 // the PowerPC Branch Selection pass to crash.
1069 if (It == BB->getParent()->end()) It = Dest;
1070 Result = CurDAG->SelectNodeTo(N, PPC::COND_BRANCH, MVT::Other, CondCode,
1071 getI32Imm(getBCCForSetCC(CC)),
1072 N->getOperand(4), CurDAG->getBasicBlock(It),
1079 SelectCode(Result, Op);
1083 /// createPPCISelDag - This pass converts a legalized DAG into a
1084 /// PowerPC-specific DAG, ready for instruction scheduling.
1086 FunctionPass *llvm::createPPCISelDag(TargetMachine &TM) {
1087 return new PPCDAGToDAGISel(TM);