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 "llvm/CodeGen/MachineInstrBuilder.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/SSARegMap.h"
21 #include "llvm/CodeGen/SelectionDAG.h"
22 #include "llvm/CodeGen/SelectionDAGISel.h"
23 #include "llvm/Target/TargetOptions.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Constants.h"
26 #include "llvm/GlobalValue.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/MathExtras.h"
34 Statistic<> FrameOff("ppc-codegen", "Number of frame idx offsets collapsed");
36 //===--------------------------------------------------------------------===//
37 /// PPCDAGToDAGISel - PPC specific code to select PPC machine
38 /// instructions for SelectionDAG operations.
40 class PPCDAGToDAGISel : public SelectionDAGISel {
41 PPCTargetLowering PPCLowering;
42 unsigned GlobalBaseReg;
44 PPCDAGToDAGISel(TargetMachine &TM)
45 : SelectionDAGISel(PPCLowering), PPCLowering(TM) {}
47 virtual bool runOnFunction(Function &Fn) {
48 // Make sure we re-emit a set of the global base reg if necessary
50 return SelectionDAGISel::runOnFunction(Fn);
53 /// getI32Imm - Return a target constant with the specified value, of type
55 inline SDOperand getI32Imm(unsigned Imm) {
56 return CurDAG->getTargetConstant(Imm, MVT::i32);
59 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
60 /// base register. Return the virtual register that holds this value.
61 SDOperand getGlobalBaseReg();
63 // Select - Convert the specified operand from a target-independent to a
64 // target-specific node if it hasn't already been changed.
65 SDOperand Select(SDOperand Op);
67 SDNode *SelectBitfieldInsert(SDNode *N);
69 /// SelectCC - Select a comparison of the specified values with the
70 /// specified condition code, returning the CR# of the expression.
71 SDOperand SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC);
73 /// SelectAddrImm - Returns true if the address N can be represented by
74 /// a base register plus a signed 16-bit displacement [r+imm].
75 bool SelectAddrImm(SDOperand N, SDOperand &Disp, SDOperand &Base);
77 /// SelectAddrIdx - Given the specified addressed, check to see if it can be
78 /// represented as an indexed [r+r] operation. Returns false if it can
79 /// be represented by [r+imm], which are preferred.
80 bool SelectAddrIdx(SDOperand N, SDOperand &Base, SDOperand &Index);
82 /// SelectAddrIdxOnly - Given the specified addressed, force it to be
83 /// represented as an indexed [r+r] operation.
84 bool SelectAddrIdxOnly(SDOperand N, SDOperand &Base, SDOperand &Index);
86 SDOperand BuildSDIVSequence(SDNode *N);
87 SDOperand BuildUDIVSequence(SDNode *N);
89 /// InstructionSelectBasicBlock - This callback is invoked by
90 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
91 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
93 virtual const char *getPassName() const {
94 return "PowerPC DAG->DAG Pattern Instruction Selection";
97 // Include the pieces autogenerated from the target description.
98 #include "PPCGenDAGISel.inc"
101 SDOperand SelectADD_PARTS(SDOperand Op);
102 SDOperand SelectSUB_PARTS(SDOperand Op);
103 SDOperand SelectSETCC(SDOperand Op);
104 SDOperand SelectCALL(SDOperand Op);
108 /// InstructionSelectBasicBlock - This callback is invoked by
109 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
110 void PPCDAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
113 // The selection process is inherently a bottom-up recursive process (users
114 // select their uses before themselves). Given infinite stack space, we
115 // could just start selecting on the root and traverse the whole graph. In
116 // practice however, this causes us to run out of stack space on large basic
117 // blocks. To avoid this problem, select the entry node, then all its uses,
118 // iteratively instead of recursively.
119 std::vector<SDOperand> Worklist;
120 Worklist.push_back(DAG.getEntryNode());
122 // Note that we can do this in the PPC target (scanning forward across token
123 // chain edges) because no nodes ever get folded across these edges. On a
124 // target like X86 which supports load/modify/store operations, this would
125 // have to be more careful.
126 while (!Worklist.empty()) {
127 SDOperand Node = Worklist.back();
130 // Chose from the least deep of the top two nodes.
131 if (!Worklist.empty() &&
132 Worklist.back().Val->getNodeDepth() < Node.Val->getNodeDepth())
133 std::swap(Worklist.back(), Node);
135 if ((Node.Val->getOpcode() >= ISD::BUILTIN_OP_END &&
136 Node.Val->getOpcode() < PPCISD::FIRST_NUMBER) ||
137 CodeGenMap.count(Node)) continue;
139 for (SDNode::use_iterator UI = Node.Val->use_begin(),
140 E = Node.Val->use_end(); UI != E; ++UI) {
141 // Scan the values. If this use has a value that is a token chain, add it
144 for (unsigned i = 0, e = User->getNumValues(); i != e; ++i)
145 if (User->getValueType(i) == MVT::Other) {
146 Worklist.push_back(SDOperand(User, i));
151 // Finally, legalize this node.
155 // Select target instructions for the DAG.
156 DAG.setRoot(SelectRoot(DAG.getRoot()));
158 DAG.RemoveDeadNodes();
160 // Emit machine code to BB.
161 ScheduleAndEmitDAG(DAG);
164 /// getGlobalBaseReg - Output the instructions required to put the
165 /// base address to use for accessing globals into a register.
167 SDOperand PPCDAGToDAGISel::getGlobalBaseReg() {
168 if (!GlobalBaseReg) {
169 // Insert the set of GlobalBaseReg into the first MBB of the function
170 MachineBasicBlock &FirstMBB = BB->getParent()->front();
171 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
172 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
173 // FIXME: when we get to LP64, we will need to create the appropriate
174 // type of register here.
175 GlobalBaseReg = RegMap->createVirtualRegister(PPC::GPRCRegisterClass);
176 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
177 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg);
179 return CurDAG->getRegister(GlobalBaseReg, MVT::i32);
183 // isIntImmediate - This method tests to see if a constant operand.
184 // If so Imm will receive the 32 bit value.
185 static bool isIntImmediate(SDNode *N, unsigned& Imm) {
186 if (N->getOpcode() == ISD::Constant) {
187 Imm = cast<ConstantSDNode>(N)->getValue();
193 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
194 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
195 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
196 // not, since all 1s are not contiguous.
197 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
198 if (isShiftedMask_32(Val)) {
199 // look for the first non-zero bit
200 MB = CountLeadingZeros_32(Val);
201 // look for the first zero bit after the run of ones
202 ME = CountLeadingZeros_32((Val - 1) ^ Val);
205 Val = ~Val; // invert mask
206 if (isShiftedMask_32(Val)) {
207 // effectively look for the first zero bit
208 ME = CountLeadingZeros_32(Val) - 1;
209 // effectively look for the first one bit after the run of zeros
210 MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
218 // isRotateAndMask - Returns true if Mask and Shift can be folded into a rotate
219 // and mask opcode and mask operation.
220 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool IsShiftMask,
221 unsigned &SH, unsigned &MB, unsigned &ME) {
222 // Don't even go down this path for i64, since different logic will be
223 // necessary for rldicl/rldicr/rldimi.
224 if (N->getValueType(0) != MVT::i32)
228 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
229 unsigned Opcode = N->getOpcode();
230 if (N->getNumOperands() != 2 ||
231 !isIntImmediate(N->getOperand(1).Val, Shift) || (Shift > 31))
234 if (Opcode == ISD::SHL) {
235 // apply shift left to mask if it comes first
236 if (IsShiftMask) Mask = Mask << Shift;
237 // determine which bits are made indeterminant by shift
238 Indeterminant = ~(0xFFFFFFFFu << Shift);
239 } else if (Opcode == ISD::SRL) {
240 // apply shift right to mask if it comes first
241 if (IsShiftMask) Mask = Mask >> Shift;
242 // determine which bits are made indeterminant by shift
243 Indeterminant = ~(0xFFFFFFFFu >> Shift);
244 // adjust for the left rotate
250 // if the mask doesn't intersect any Indeterminant bits
251 if (Mask && !(Mask & Indeterminant)) {
253 // make sure the mask is still a mask (wrap arounds may not be)
254 return isRunOfOnes(Mask, MB, ME);
259 // isOpcWithIntImmediate - This method tests to see if the node is a specific
260 // opcode and that it has a immediate integer right operand.
261 // If so Imm will receive the 32 bit value.
262 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
263 return N->getOpcode() == Opc && isIntImmediate(N->getOperand(1).Val, Imm);
266 // isIntImmediate - This method tests to see if a constant operand.
267 // If so Imm will receive the 32 bit value.
268 static bool isIntImmediate(SDOperand N, unsigned& Imm) {
269 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
270 Imm = (unsigned)CN->getSignExtended();
276 /// SelectBitfieldInsert - turn an or of two masked values into
277 /// the rotate left word immediate then mask insert (rlwimi) instruction.
278 /// Returns true on success, false if the caller still needs to select OR.
280 /// Patterns matched:
281 /// 1. or shl, and 5. or and, and
282 /// 2. or and, shl 6. or shl, shr
283 /// 3. or shr, and 7. or shr, shl
285 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
286 bool IsRotate = false;
287 unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, SH = 0;
290 SDOperand Op0 = N->getOperand(0);
291 SDOperand Op1 = N->getOperand(1);
293 unsigned Op0Opc = Op0.getOpcode();
294 unsigned Op1Opc = Op1.getOpcode();
296 // Verify that we have the correct opcodes
297 if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc)
299 if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc)
302 // Generate Mask value for Target
303 if (isIntImmediate(Op0.getOperand(1), Value)) {
305 case ISD::SHL: TgtMask <<= Value; break;
306 case ISD::SRL: TgtMask >>= Value; break;
307 case ISD::AND: TgtMask &= Value; break;
313 // Generate Mask value for Insert
314 if (!isIntImmediate(Op1.getOperand(1), Value))
321 if (Op0Opc == ISD::SRL) IsRotate = true;
327 if (Op0Opc == ISD::SHL) IsRotate = true;
334 // If both of the inputs are ANDs and one of them has a logical shift by
335 // constant as its input, make that AND the inserted value so that we can
336 // combine the shift into the rotate part of the rlwimi instruction
337 bool IsAndWithShiftOp = false;
338 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
339 if (Op1.getOperand(0).getOpcode() == ISD::SHL ||
340 Op1.getOperand(0).getOpcode() == ISD::SRL) {
341 if (isIntImmediate(Op1.getOperand(0).getOperand(1), Value)) {
342 SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
343 IsAndWithShiftOp = true;
345 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
346 Op0.getOperand(0).getOpcode() == ISD::SRL) {
347 if (isIntImmediate(Op0.getOperand(0).getOperand(1), Value)) {
349 std::swap(TgtMask, InsMask);
350 SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
351 IsAndWithShiftOp = true;
356 // Verify that the Target mask and Insert mask together form a full word mask
357 // and that the Insert mask is a run of set bits (which implies both are runs
358 // of set bits). Given that, Select the arguments and generate the rlwimi
361 if (((TgtMask & InsMask) == 0) && isRunOfOnes(InsMask, MB, ME)) {
362 bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF;
363 bool Op0IsAND = Op0Opc == ISD::AND;
364 // Check for rotlwi / rotrwi here, a special case of bitfield insert
365 // where both bitfield halves are sourced from the same value.
366 if (IsRotate && fullMask &&
367 N->getOperand(0).getOperand(0) == N->getOperand(1).getOperand(0)) {
368 Op0 = CurDAG->getTargetNode(PPC::RLWINM, MVT::i32,
369 Select(N->getOperand(0).getOperand(0)),
370 getI32Imm(SH), getI32Imm(0), getI32Imm(31));
373 SDOperand Tmp1 = (Op0IsAND && fullMask) ? Select(Op0.getOperand(0))
375 SDOperand Tmp2 = IsAndWithShiftOp ? Select(Op1.getOperand(0).getOperand(0))
376 : Select(Op1.getOperand(0));
377 Op0 = CurDAG->getTargetNode(PPC::RLWIMI, MVT::i32, Tmp1, Tmp2,
378 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME));
384 /// SelectAddrImm - Returns true if the address N can be represented by
385 /// a base register plus a signed 16-bit displacement [r+imm].
386 bool PPCDAGToDAGISel::SelectAddrImm(SDOperand N, SDOperand &Disp,
388 if (N.getOpcode() == ISD::ADD) {
390 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm)) {
391 Disp = getI32Imm(imm & 0xFFFF);
392 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
393 Base = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
395 Base = N.getOperand(0);
397 return true; // [r+i]
398 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
399 // Match LOAD (ADD (X, Lo(G))).
400 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
401 && "Cannot handle constant offsets yet!");
402 Disp = N.getOperand(1).getOperand(0); // The global address.
403 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
404 Disp.getOpcode() == ISD::TargetConstantPool);
405 Base = N.getOperand(0);
406 return true; // [&g+r]
408 return false; // [r+r]
411 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
412 Base = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
415 return true; // [r+0]
418 /// SelectAddrIdx - Given the specified addressed, check to see if it can be
419 /// represented as an indexed [r+r] operation. Returns false if it can
420 /// be represented by [r+imm], which are preferred.
421 bool PPCDAGToDAGISel::SelectAddrIdx(SDOperand N, SDOperand &Base,
423 // Check to see if we can represent this as an [r+imm] address instead,
424 // which will fail if the address is more profitably represented as an
426 if (SelectAddrImm(N, Base, Index))
429 if (N.getOpcode() == ISD::ADD) {
430 Base = N.getOperand(0);
431 Index = N.getOperand(1);
435 Base = CurDAG->getRegister(PPC::R0, MVT::i32);
440 /// SelectAddrIdxOnly - Given the specified addressed, force it to be
441 /// represented as an indexed [r+r] operation.
442 bool PPCDAGToDAGISel::SelectAddrIdxOnly(SDOperand N, SDOperand &Base,
444 if (N.getOpcode() == ISD::ADD) {
445 Base = N.getOperand(0);
446 Index = N.getOperand(1);
450 Base = CurDAG->getRegister(PPC::R0, MVT::i32);
455 /// SelectCC - Select a comparison of the specified values with the specified
456 /// condition code, returning the CR# of the expression.
457 SDOperand PPCDAGToDAGISel::SelectCC(SDOperand LHS, SDOperand RHS,
459 // Always select the LHS.
462 // Use U to determine whether the SETCC immediate range is signed or not.
463 if (MVT::isInteger(LHS.getValueType())) {
464 bool U = ISD::isUnsignedIntSetCC(CC);
466 if (isIntImmediate(RHS, Imm) &&
467 ((U && isUInt16(Imm)) || (!U && isInt16(Imm))))
468 return CurDAG->getTargetNode(U ? PPC::CMPLWI : PPC::CMPWI, MVT::i32,
469 LHS, getI32Imm(Imm & 0xFFFF));
470 return CurDAG->getTargetNode(U ? PPC::CMPLW : PPC::CMPW, MVT::i32,
472 } else if (LHS.getValueType() == MVT::f32) {
473 return CurDAG->getTargetNode(PPC::FCMPUS, MVT::i32, LHS, Select(RHS));
475 return CurDAG->getTargetNode(PPC::FCMPUD, MVT::i32, LHS, Select(RHS));
479 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
481 static unsigned getBCCForSetCC(ISD::CondCode CC) {
483 default: assert(0 && "Unknown condition!"); abort();
484 case ISD::SETOEQ: // FIXME: This is incorrect see PR642.
485 case ISD::SETEQ: return PPC::BEQ;
486 case ISD::SETONE: // FIXME: This is incorrect see PR642.
487 case ISD::SETNE: return PPC::BNE;
488 case ISD::SETOLT: // FIXME: This is incorrect see PR642.
490 case ISD::SETLT: return PPC::BLT;
491 case ISD::SETOLE: // FIXME: This is incorrect see PR642.
493 case ISD::SETLE: return PPC::BLE;
494 case ISD::SETOGT: // FIXME: This is incorrect see PR642.
496 case ISD::SETGT: return PPC::BGT;
497 case ISD::SETOGE: // FIXME: This is incorrect see PR642.
499 case ISD::SETGE: return PPC::BGE;
501 case ISD::SETO: return PPC::BUN;
502 case ISD::SETUO: return PPC::BNU;
507 /// getCRIdxForSetCC - Return the index of the condition register field
508 /// associated with the SetCC condition, and whether or not the field is
509 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
510 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool& Inv) {
512 default: assert(0 && "Unknown condition!"); abort();
513 case ISD::SETOLT: // FIXME: This is incorrect see PR642.
515 case ISD::SETLT: Inv = false; return 0;
516 case ISD::SETOGE: // FIXME: This is incorrect see PR642.
518 case ISD::SETGE: Inv = true; return 0;
519 case ISD::SETOGT: // FIXME: This is incorrect see PR642.
521 case ISD::SETGT: Inv = false; return 1;
522 case ISD::SETOLE: // FIXME: This is incorrect see PR642.
524 case ISD::SETLE: Inv = true; return 1;
525 case ISD::SETOEQ: // FIXME: This is incorrect see PR642.
526 case ISD::SETEQ: Inv = false; return 2;
527 case ISD::SETONE: // FIXME: This is incorrect see PR642.
528 case ISD::SETNE: Inv = true; return 2;
529 case ISD::SETO: Inv = true; return 3;
530 case ISD::SETUO: Inv = false; return 3;
536 SDOperand PPCDAGToDAGISel::SelectADD_PARTS(SDOperand Op) {
538 SDOperand LHSL = Select(N->getOperand(0));
539 SDOperand LHSH = Select(N->getOperand(1));
542 bool ME = false, ZE = false;
543 if (isIntImmediate(N->getOperand(3), Imm)) {
544 ME = (signed)Imm == -1;
548 std::vector<SDOperand> Result;
549 SDOperand CarryFromLo;
550 if (isIntImmediate(N->getOperand(2), Imm) &&
551 ((signed)Imm >= -32768 || (signed)Imm < 32768)) {
552 // Codegen the low 32 bits of the add. Interestingly, there is no
553 // shifted form of add immediate carrying.
554 CarryFromLo = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
555 LHSL, getI32Imm(Imm));
557 CarryFromLo = CurDAG->getTargetNode(PPC::ADDC, MVT::i32, MVT::Flag,
558 LHSL, Select(N->getOperand(2)));
560 CarryFromLo = CarryFromLo.getValue(1);
562 // Codegen the high 32 bits, adding zero, minus one, or the full value
563 // along with the carry flag produced by addc/addic.
566 ResultHi = CurDAG->getTargetNode(PPC::ADDZE, MVT::i32, LHSH, CarryFromLo);
568 ResultHi = CurDAG->getTargetNode(PPC::ADDME, MVT::i32, LHSH, CarryFromLo);
570 ResultHi = CurDAG->getTargetNode(PPC::ADDE, MVT::i32, LHSH,
571 Select(N->getOperand(3)), CarryFromLo);
572 Result.push_back(CarryFromLo.getValue(0));
573 Result.push_back(ResultHi);
575 CodeGenMap[Op.getValue(0)] = Result[0];
576 CodeGenMap[Op.getValue(1)] = Result[1];
577 return Result[Op.ResNo];
579 SDOperand PPCDAGToDAGISel::SelectSUB_PARTS(SDOperand Op) {
581 SDOperand LHSL = Select(N->getOperand(0));
582 SDOperand LHSH = Select(N->getOperand(1));
583 SDOperand RHSL = Select(N->getOperand(2));
584 SDOperand RHSH = Select(N->getOperand(3));
586 std::vector<SDOperand> Result;
587 Result.push_back(CurDAG->getTargetNode(PPC::SUBFC, MVT::i32, MVT::Flag,
589 Result.push_back(CurDAG->getTargetNode(PPC::SUBFE, MVT::i32, RHSH, LHSH,
590 Result[0].getValue(1)));
591 CodeGenMap[Op.getValue(0)] = Result[0];
592 CodeGenMap[Op.getValue(1)] = Result[1];
593 return Result[Op.ResNo];
596 SDOperand PPCDAGToDAGISel::SelectSETCC(SDOperand Op) {
599 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
600 if (isIntImmediate(N->getOperand(1), Imm)) {
601 // We can codegen setcc op, imm very efficiently compared to a brcond.
602 // Check for those cases here.
605 SDOperand Op = Select(N->getOperand(0));
609 Op = CurDAG->getTargetNode(PPC::CNTLZW, MVT::i32, Op);
610 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(27),
611 getI32Imm(5), getI32Imm(31));
613 SDOperand AD = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
615 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
619 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(1),
620 getI32Imm(31), getI32Imm(31));
622 SDOperand T = CurDAG->getTargetNode(PPC::NEG, MVT::i32, Op);
623 T = CurDAG->getTargetNode(PPC::ANDC, MVT::i32, T, Op);;
624 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, T, getI32Imm(1),
625 getI32Imm(31), getI32Imm(31));
628 } else if (Imm == ~0U) { // setcc op, -1
629 SDOperand Op = Select(N->getOperand(0));
633 Op = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
635 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
636 CurDAG->getTargetNode(PPC::LI, MVT::i32,
640 Op = CurDAG->getTargetNode(PPC::NOR, MVT::i32, Op, Op);
641 SDOperand AD = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
643 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
647 SDOperand AD = CurDAG->getTargetNode(PPC::ADDI, MVT::i32, Op,
649 SDOperand AN = CurDAG->getTargetNode(PPC::AND, MVT::i32, AD, Op);
650 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, AN, getI32Imm(1),
651 getI32Imm(31), getI32Imm(31));
654 Op = CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, Op, getI32Imm(1),
655 getI32Imm(31), getI32Imm(31));
656 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, getI32Imm(1));
662 unsigned Idx = getCRIdxForSetCC(CC, Inv);
663 SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
666 // Force the ccreg into CR7.
667 SDOperand CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
669 SDOperand InFlag(0, 0); // Null incoming flag value.
670 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), CR7Reg, CCReg,
673 if (TLI.getTargetMachine().getSubtarget<PPCSubtarget>().isGigaProcessor())
674 IntCR = CurDAG->getTargetNode(PPC::MFOCRF, MVT::i32, CR7Reg, CCReg);
676 IntCR = CurDAG->getTargetNode(PPC::MFCR, MVT::i32, CCReg);
679 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, IntCR,
680 getI32Imm((32-(3-Idx)) & 31),
681 getI32Imm(31), getI32Imm(31));
684 CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, IntCR,
685 getI32Imm((32-(3-Idx)) & 31),
686 getI32Imm(31),getI32Imm(31));
687 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
691 /// isCallCompatibleAddress - Return true if the specified 32-bit value is
692 /// representable in the immediate field of a Bx instruction.
693 static bool isCallCompatibleAddress(ConstantSDNode *C) {
694 int Addr = C->getValue();
695 if (Addr & 3) return false; // Low 2 bits are implicitly zero.
696 return (Addr << 6 >> 6) == Addr; // Top 6 bits have to be sext of immediate.
699 SDOperand PPCDAGToDAGISel::SelectCALL(SDOperand Op) {
701 SDOperand Chain = Select(N->getOperand(0));
704 std::vector<SDOperand> CallOperands;
706 if (GlobalAddressSDNode *GASD =
707 dyn_cast<GlobalAddressSDNode>(N->getOperand(1))) {
708 CallOpcode = PPC::BL;
709 CallOperands.push_back(N->getOperand(1));
710 } else if (ExternalSymbolSDNode *ESSDN =
711 dyn_cast<ExternalSymbolSDNode>(N->getOperand(1))) {
712 CallOpcode = PPC::BL;
713 CallOperands.push_back(N->getOperand(1));
714 } else if (isa<ConstantSDNode>(N->getOperand(1)) &&
715 isCallCompatibleAddress(cast<ConstantSDNode>(N->getOperand(1)))) {
716 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(1));
717 CallOpcode = PPC::BLA;
718 CallOperands.push_back(getI32Imm((int)C->getValue() >> 2));
720 // Copy the callee address into the CTR register.
721 SDOperand Callee = Select(N->getOperand(1));
722 Chain = CurDAG->getTargetNode(PPC::MTCTR, MVT::Other, Callee, Chain);
724 // Copy the callee address into R12 on darwin.
725 SDOperand R12 = CurDAG->getRegister(PPC::R12, MVT::i32);
726 Chain = CurDAG->getNode(ISD::CopyToReg, MVT::Other, Chain, R12, Callee);
728 CallOperands.push_back(R12);
729 CallOpcode = PPC::BCTRL;
732 unsigned GPR_idx = 0, FPR_idx = 0;
733 static const unsigned GPR[] = {
734 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
735 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
737 static const unsigned FPR[] = {
738 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
739 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
742 SDOperand InFlag; // Null incoming flag value.
744 for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i) {
745 unsigned DestReg = 0;
746 MVT::ValueType RegTy = N->getOperand(i).getValueType();
747 if (RegTy == MVT::i32) {
748 assert(GPR_idx < 8 && "Too many int args");
749 DestReg = GPR[GPR_idx++];
751 assert(MVT::isFloatingPoint(N->getOperand(i).getValueType()) &&
752 "Unpromoted integer arg?");
753 assert(FPR_idx < 13 && "Too many fp args");
754 DestReg = FPR[FPR_idx++];
757 if (N->getOperand(i).getOpcode() != ISD::UNDEF) {
758 SDOperand Val = Select(N->getOperand(i));
759 Chain = CurDAG->getCopyToReg(Chain, DestReg, Val, InFlag);
760 InFlag = Chain.getValue(1);
761 CallOperands.push_back(CurDAG->getRegister(DestReg, RegTy));
765 // Finally, once everything is in registers to pass to the call, emit the
768 CallOperands.push_back(InFlag); // Strong dep on register copies.
770 CallOperands.push_back(Chain); // Weak dep on whatever occurs before
771 Chain = CurDAG->getTargetNode(CallOpcode, MVT::Other, MVT::Flag,
774 std::vector<SDOperand> CallResults;
776 // If the call has results, copy the values out of the ret val registers.
777 switch (N->getValueType(0)) {
778 default: assert(0 && "Unexpected ret value!");
779 case MVT::Other: break;
781 if (N->getValueType(1) == MVT::i32) {
782 Chain = CurDAG->getCopyFromReg(Chain, PPC::R4, MVT::i32,
783 Chain.getValue(1)).getValue(1);
784 CallResults.push_back(Chain.getValue(0));
785 Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
786 Chain.getValue(2)).getValue(1);
787 CallResults.push_back(Chain.getValue(0));
789 Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
790 Chain.getValue(1)).getValue(1);
791 CallResults.push_back(Chain.getValue(0));
796 Chain = CurDAG->getCopyFromReg(Chain, PPC::F1, N->getValueType(0),
797 Chain.getValue(1)).getValue(1);
798 CallResults.push_back(Chain.getValue(0));
802 CallResults.push_back(Chain);
803 for (unsigned i = 0, e = CallResults.size(); i != e; ++i)
804 CodeGenMap[Op.getValue(i)] = CallResults[i];
805 return CallResults[Op.ResNo];
808 // Select - Convert the specified operand from a target-independent to a
809 // target-specific node if it hasn't already been changed.
810 SDOperand PPCDAGToDAGISel::Select(SDOperand Op) {
812 if (N->getOpcode() >= ISD::BUILTIN_OP_END &&
813 N->getOpcode() < PPCISD::FIRST_NUMBER)
814 return Op; // Already selected.
816 // If this has already been converted, use it.
817 std::map<SDOperand, SDOperand>::iterator CGMI = CodeGenMap.find(Op);
818 if (CGMI != CodeGenMap.end()) return CGMI->second;
820 switch (N->getOpcode()) {
822 case ISD::ADD_PARTS: return SelectADD_PARTS(Op);
823 case ISD::SUB_PARTS: return SelectSUB_PARTS(Op);
824 case ISD::SETCC: return SelectSETCC(Op);
825 case PPCISD::CALL: return SelectCALL(Op);
826 case PPCISD::GlobalBaseReg: return getGlobalBaseReg();
828 case ISD::FrameIndex: {
829 int FI = cast<FrameIndexSDNode>(N)->getIndex();
831 return CurDAG->SelectNodeTo(N, PPC::ADDI, MVT::i32,
832 CurDAG->getTargetFrameIndex(FI, MVT::i32),
834 return CodeGenMap[Op] =
835 CurDAG->getTargetNode(PPC::ADDI, MVT::i32,
836 CurDAG->getTargetFrameIndex(FI, MVT::i32),
840 // FIXME: since this depends on the setting of the carry flag from the srawi
841 // we should really be making notes about that for the scheduler.
842 // FIXME: It sure would be nice if we could cheaply recognize the
843 // srl/add/sra pattern the dag combiner will generate for this as
844 // sra/addze rather than having to handle sdiv ourselves. oh well.
846 if (isIntImmediate(N->getOperand(1), Imm)) {
847 if ((signed)Imm > 0 && isPowerOf2_32(Imm)) {
849 CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
850 Select(N->getOperand(0)),
851 getI32Imm(Log2_32(Imm)));
852 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
853 Op.getValue(0), Op.getValue(1));
854 } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) {
856 CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
857 Select(N->getOperand(0)),
858 getI32Imm(Log2_32(-Imm)));
860 CurDAG->getTargetNode(PPC::ADDZE, MVT::i32, Op.getValue(0),
862 return CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT);
866 // Other cases are autogenerated.
871 // If this is an and of a value rotated between 0 and 31 bits and then and'd
872 // with a mask, emit rlwinm
873 if (isIntImmediate(N->getOperand(1), Imm) && (isShiftedMask_32(Imm) ||
874 isShiftedMask_32(~Imm))) {
877 if (isRotateAndMask(N->getOperand(0).Val, Imm, false, SH, MB, ME)) {
878 Val = Select(N->getOperand(0).getOperand(0));
879 } else if (Imm == 0) {
880 // AND X, 0 -> 0, not "rlwinm 32".
881 return Select(N->getOperand(1));
883 Val = Select(N->getOperand(0));
884 isRunOfOnes(Imm, MB, ME);
887 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Val, getI32Imm(SH),
888 getI32Imm(MB), getI32Imm(ME));
890 // ISD::OR doesn't get all the bitfield insertion fun.
891 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert
892 if (isIntImmediate(N->getOperand(1), Imm) &&
893 N->getOperand(0).getOpcode() == ISD::OR &&
894 isIntImmediate(N->getOperand(0).getOperand(1), Imm2)) {
897 if (isRunOfOnes(Imm, MB, ME)) {
898 SDOperand Tmp1 = Select(N->getOperand(0).getOperand(0));
899 SDOperand Tmp2 = Select(N->getOperand(0).getOperand(1));
900 return CurDAG->getTargetNode(PPC::RLWIMI, MVT::i32, Tmp1, Tmp2,
901 getI32Imm(0), getI32Imm(MB), getI32Imm(ME));
905 // Other cases are autogenerated.
909 if (SDNode *I = SelectBitfieldInsert(N))
910 return CodeGenMap[Op] = SDOperand(I, 0);
912 // Other cases are autogenerated.
915 unsigned Imm, SH, MB, ME;
916 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
917 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
918 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
919 Select(N->getOperand(0).getOperand(0)),
920 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME));
923 // Other cases are autogenerated.
927 unsigned Imm, SH, MB, ME;
928 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
929 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
930 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
931 Select(N->getOperand(0).getOperand(0)),
932 getI32Imm(SH & 0x1F), getI32Imm(MB),
936 // Other cases are autogenerated.
939 case ISD::SELECT_CC: {
940 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
942 // handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
943 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
944 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
945 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
946 if (N1C->isNullValue() && N3C->isNullValue() &&
947 N2C->getValue() == 1ULL && CC == ISD::SETNE) {
948 SDOperand LHS = Select(N->getOperand(0));
950 CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
951 LHS, getI32Imm(~0U));
952 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, Tmp, LHS,
956 SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
957 unsigned BROpc = getBCCForSetCC(CC);
959 bool isFP = MVT::isFloatingPoint(N->getValueType(0));
961 if (MVT::isInteger(N->getValueType(0)))
962 SelectCCOp = PPC::SELECT_CC_Int;
963 else if (N->getValueType(0) == MVT::f32)
964 SelectCCOp = PPC::SELECT_CC_F4;
966 SelectCCOp = PPC::SELECT_CC_F8;
967 return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), CCReg,
968 Select(N->getOperand(2)),
969 Select(N->getOperand(3)),
973 case ISD::BRTWOWAY_CC: {
974 SDOperand Chain = Select(N->getOperand(0));
975 MachineBasicBlock *Dest =
976 cast<BasicBlockSDNode>(N->getOperand(4))->getBasicBlock();
977 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
978 SDOperand CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC);
980 // If this is a two way branch, then grab the fallthrough basic block
981 // argument and build a PowerPC branch pseudo-op, suitable for long branch
982 // conversion if necessary by the branch selection pass. Otherwise, emit a
983 // standard conditional branch.
984 if (N->getOpcode() == ISD::BRTWOWAY_CC) {
985 SDOperand CondTrueBlock = N->getOperand(4);
986 SDOperand CondFalseBlock = N->getOperand(5);
988 // If the false case is the current basic block, then this is a self loop.
989 // We do not want to emit "Loop: ... brcond Out; br Loop", as it adds an
990 // extra dispatch group to the loop. Instead, invert the condition and
991 // emit "Loop: ... br!cond Loop; br Out
992 if (cast<BasicBlockSDNode>(CondFalseBlock)->getBasicBlock() == BB) {
993 std::swap(CondTrueBlock, CondFalseBlock);
994 CC = getSetCCInverse(CC,
995 MVT::isInteger(N->getOperand(2).getValueType()));
998 unsigned Opc = getBCCForSetCC(CC);
999 SDOperand CB = CurDAG->getTargetNode(PPC::COND_BRANCH, MVT::Other,
1000 CondCode, getI32Imm(Opc),
1001 CondTrueBlock, CondFalseBlock,
1003 return CurDAG->SelectNodeTo(N, PPC::B, MVT::Other, CondFalseBlock, CB);
1005 // Iterate to the next basic block
1006 ilist<MachineBasicBlock>::iterator It = BB;
1009 // If the fallthrough path is off the end of the function, which would be
1010 // undefined behavior, set it to be the same as the current block because
1011 // we have nothing better to set it to, and leaving it alone will cause
1012 // the PowerPC Branch Selection pass to crash.
1013 if (It == BB->getParent()->end()) It = Dest;
1014 return CurDAG->SelectNodeTo(N, PPC::COND_BRANCH, MVT::Other, CondCode,
1015 getI32Imm(getBCCForSetCC(CC)),
1016 N->getOperand(4), CurDAG->getBasicBlock(It),
1022 return SelectCode(Op);
1026 /// createPPCISelDag - This pass converts a legalized DAG into a
1027 /// PowerPC-specific DAG, ready for instruction scheduling.
1029 FunctionPass *llvm::createPPCISelDag(TargetMachine &TM) {
1030 return new PPCDAGToDAGISel(TM);