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 void Select(SDOperand &Result, 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.
156 // Select target instructions for the DAG.
157 DAG.setRoot(SelectRoot(DAG.getRoot()));
159 DAG.RemoveDeadNodes();
161 // Emit machine code to BB.
162 ScheduleAndEmitDAG(DAG);
165 /// getGlobalBaseReg - Output the instructions required to put the
166 /// base address to use for accessing globals into a register.
168 SDOperand PPCDAGToDAGISel::getGlobalBaseReg() {
169 if (!GlobalBaseReg) {
170 // Insert the set of GlobalBaseReg into the first MBB of the function
171 MachineBasicBlock &FirstMBB = BB->getParent()->front();
172 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
173 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
174 // FIXME: when we get to LP64, we will need to create the appropriate
175 // type of register here.
176 GlobalBaseReg = RegMap->createVirtualRegister(PPC::GPRCRegisterClass);
177 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
178 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg);
180 return CurDAG->getRegister(GlobalBaseReg, MVT::i32);
184 // isIntImmediate - This method tests to see if a constant operand.
185 // If so Imm will receive the 32 bit value.
186 static bool isIntImmediate(SDNode *N, unsigned& Imm) {
187 if (N->getOpcode() == ISD::Constant) {
188 Imm = cast<ConstantSDNode>(N)->getValue();
194 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
195 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
196 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
197 // not, since all 1s are not contiguous.
198 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
199 if (isShiftedMask_32(Val)) {
200 // look for the first non-zero bit
201 MB = CountLeadingZeros_32(Val);
202 // look for the first zero bit after the run of ones
203 ME = CountLeadingZeros_32((Val - 1) ^ Val);
206 Val = ~Val; // invert mask
207 if (isShiftedMask_32(Val)) {
208 // effectively look for the first zero bit
209 ME = CountLeadingZeros_32(Val) - 1;
210 // effectively look for the first one bit after the run of zeros
211 MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
219 // isRotateAndMask - Returns true if Mask and Shift can be folded into a rotate
220 // and mask opcode and mask operation.
221 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool IsShiftMask,
222 unsigned &SH, unsigned &MB, unsigned &ME) {
223 // Don't even go down this path for i64, since different logic will be
224 // necessary for rldicl/rldicr/rldimi.
225 if (N->getValueType(0) != MVT::i32)
229 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
230 unsigned Opcode = N->getOpcode();
231 if (N->getNumOperands() != 2 ||
232 !isIntImmediate(N->getOperand(1).Val, Shift) || (Shift > 31))
235 if (Opcode == ISD::SHL) {
236 // apply shift left to mask if it comes first
237 if (IsShiftMask) Mask = Mask << Shift;
238 // determine which bits are made indeterminant by shift
239 Indeterminant = ~(0xFFFFFFFFu << Shift);
240 } else if (Opcode == ISD::SRL) {
241 // apply shift right to mask if it comes first
242 if (IsShiftMask) Mask = Mask >> Shift;
243 // determine which bits are made indeterminant by shift
244 Indeterminant = ~(0xFFFFFFFFu >> Shift);
245 // adjust for the left rotate
251 // if the mask doesn't intersect any Indeterminant bits
252 if (Mask && !(Mask & Indeterminant)) {
254 // make sure the mask is still a mask (wrap arounds may not be)
255 return isRunOfOnes(Mask, MB, ME);
260 // isOpcWithIntImmediate - This method tests to see if the node is a specific
261 // opcode and that it has a immediate integer right operand.
262 // If so Imm will receive the 32 bit value.
263 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
264 return N->getOpcode() == Opc && isIntImmediate(N->getOperand(1).Val, Imm);
267 // isIntImmediate - This method tests to see if a constant operand.
268 // If so Imm will receive the 32 bit value.
269 static bool isIntImmediate(SDOperand N, unsigned& Imm) {
270 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
271 Imm = (unsigned)CN->getSignExtended();
277 /// SelectBitfieldInsert - turn an or of two masked values into
278 /// the rotate left word immediate then mask insert (rlwimi) instruction.
279 /// Returns true on success, false if the caller still needs to select OR.
281 /// Patterns matched:
282 /// 1. or shl, and 5. or and, and
283 /// 2. or and, shl 6. or shl, shr
284 /// 3. or shr, and 7. or shr, shl
286 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
287 bool IsRotate = false;
288 unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, SH = 0;
291 SDOperand Op0 = N->getOperand(0);
292 SDOperand Op1 = N->getOperand(1);
294 unsigned Op0Opc = Op0.getOpcode();
295 unsigned Op1Opc = Op1.getOpcode();
297 // Verify that we have the correct opcodes
298 if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc)
300 if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc)
303 // Generate Mask value for Target
304 if (isIntImmediate(Op0.getOperand(1), Value)) {
306 case ISD::SHL: TgtMask <<= Value; break;
307 case ISD::SRL: TgtMask >>= Value; break;
308 case ISD::AND: TgtMask &= Value; break;
314 // Generate Mask value for Insert
315 if (!isIntImmediate(Op1.getOperand(1), Value))
322 if (Op0Opc == ISD::SRL) IsRotate = true;
328 if (Op0Opc == ISD::SHL) IsRotate = true;
335 // If both of the inputs are ANDs and one of them has a logical shift by
336 // constant as its input, make that AND the inserted value so that we can
337 // combine the shift into the rotate part of the rlwimi instruction
338 bool IsAndWithShiftOp = false;
339 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
340 if (Op1.getOperand(0).getOpcode() == ISD::SHL ||
341 Op1.getOperand(0).getOpcode() == ISD::SRL) {
342 if (isIntImmediate(Op1.getOperand(0).getOperand(1), Value)) {
343 SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
344 IsAndWithShiftOp = true;
346 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
347 Op0.getOperand(0).getOpcode() == ISD::SRL) {
348 if (isIntImmediate(Op0.getOperand(0).getOperand(1), Value)) {
350 std::swap(TgtMask, InsMask);
351 SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
352 IsAndWithShiftOp = true;
357 // Verify that the Target mask and Insert mask together form a full word mask
358 // and that the Insert mask is a run of set bits (which implies both are runs
359 // of set bits). Given that, Select the arguments and generate the rlwimi
362 if (((TgtMask & InsMask) == 0) && isRunOfOnes(InsMask, MB, ME)) {
363 bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF;
364 bool Op0IsAND = Op0Opc == ISD::AND;
365 // Check for rotlwi / rotrwi here, a special case of bitfield insert
366 // where both bitfield halves are sourced from the same value.
367 if (IsRotate && fullMask &&
368 N->getOperand(0).getOperand(0) == N->getOperand(1).getOperand(0)) {
370 Select(Tmp, N->getOperand(0).getOperand(0));
371 Op0 = CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, Tmp,
372 getI32Imm(SH), getI32Imm(0), getI32Imm(31));
375 SDOperand Tmp1, Tmp2;
376 Select(Tmp1, ((Op0IsAND && fullMask) ? Op0.getOperand(0) : Op0));
377 Select(Tmp2, (IsAndWithShiftOp ? Op1.getOperand(0).getOperand(0)
378 : Op1.getOperand(0)));
379 Op0 = CurDAG->getTargetNode(PPC::RLWIMI, MVT::i32, Tmp1, Tmp2,
380 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME));
386 /// SelectAddrImm - Returns true if the address N can be represented by
387 /// a base register plus a signed 16-bit displacement [r+imm].
388 bool PPCDAGToDAGISel::SelectAddrImm(SDOperand N, SDOperand &Disp,
390 if (N.getOpcode() == ISD::ADD) {
392 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm)) {
393 Disp = getI32Imm(imm & 0xFFFF);
394 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
395 Base = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
397 Base = N.getOperand(0);
399 return true; // [r+i]
400 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
401 // Match LOAD (ADD (X, Lo(G))).
402 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
403 && "Cannot handle constant offsets yet!");
404 Disp = N.getOperand(1).getOperand(0); // The global address.
405 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
406 Disp.getOpcode() == ISD::TargetConstantPool);
407 Base = N.getOperand(0);
408 return true; // [&g+r]
410 return false; // [r+r]
413 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
414 Base = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
417 return true; // [r+0]
420 /// SelectAddrIdx - Given the specified addressed, check to see if it can be
421 /// represented as an indexed [r+r] operation. Returns false if it can
422 /// be represented by [r+imm], which are preferred.
423 bool PPCDAGToDAGISel::SelectAddrIdx(SDOperand N, SDOperand &Base,
425 // Check to see if we can represent this as an [r+imm] address instead,
426 // which will fail if the address is more profitably represented as an
428 if (SelectAddrImm(N, Base, Index))
431 if (N.getOpcode() == ISD::ADD) {
432 Base = N.getOperand(0);
433 Index = N.getOperand(1);
437 Base = CurDAG->getRegister(PPC::R0, MVT::i32);
442 /// SelectAddrIdxOnly - Given the specified addressed, force it to be
443 /// represented as an indexed [r+r] operation.
444 bool PPCDAGToDAGISel::SelectAddrIdxOnly(SDOperand N, SDOperand &Base,
446 if (N.getOpcode() == ISD::ADD) {
447 Base = N.getOperand(0);
448 Index = N.getOperand(1);
452 Base = CurDAG->getRegister(PPC::R0, MVT::i32);
457 /// SelectCC - Select a comparison of the specified values with the specified
458 /// condition code, returning the CR# of the expression.
459 SDOperand PPCDAGToDAGISel::SelectCC(SDOperand LHS, SDOperand RHS,
461 // Always select the LHS.
464 // Use U to determine whether the SETCC immediate range is signed or not.
465 if (MVT::isInteger(LHS.getValueType())) {
466 bool U = ISD::isUnsignedIntSetCC(CC);
468 if (isIntImmediate(RHS, Imm) &&
469 ((U && isUInt16(Imm)) || (!U && isInt16(Imm))))
470 return CurDAG->getTargetNode(U ? PPC::CMPLWI : PPC::CMPWI, MVT::i32,
471 LHS, getI32Imm(Imm & 0xFFFF));
473 return CurDAG->getTargetNode(U ? PPC::CMPLW : PPC::CMPW, MVT::i32,
475 } else if (LHS.getValueType() == MVT::f32) {
477 return CurDAG->getTargetNode(PPC::FCMPUS, MVT::i32, LHS, RHS);
480 return CurDAG->getTargetNode(PPC::FCMPUD, MVT::i32, LHS, RHS);
484 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
486 static unsigned getBCCForSetCC(ISD::CondCode CC) {
488 default: assert(0 && "Unknown condition!"); abort();
489 case ISD::SETOEQ: // FIXME: This is incorrect see PR642.
490 case ISD::SETEQ: return PPC::BEQ;
491 case ISD::SETONE: // FIXME: This is incorrect see PR642.
492 case ISD::SETNE: return PPC::BNE;
493 case ISD::SETOLT: // FIXME: This is incorrect see PR642.
495 case ISD::SETLT: return PPC::BLT;
496 case ISD::SETOLE: // FIXME: This is incorrect see PR642.
498 case ISD::SETLE: return PPC::BLE;
499 case ISD::SETOGT: // FIXME: This is incorrect see PR642.
501 case ISD::SETGT: return PPC::BGT;
502 case ISD::SETOGE: // FIXME: This is incorrect see PR642.
504 case ISD::SETGE: return PPC::BGE;
506 case ISD::SETO: return PPC::BUN;
507 case ISD::SETUO: return PPC::BNU;
512 /// getCRIdxForSetCC - Return the index of the condition register field
513 /// associated with the SetCC condition, and whether or not the field is
514 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
515 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool& Inv) {
517 default: assert(0 && "Unknown condition!"); abort();
518 case ISD::SETOLT: // FIXME: This is incorrect see PR642.
520 case ISD::SETLT: Inv = false; return 0;
521 case ISD::SETOGE: // FIXME: This is incorrect see PR642.
523 case ISD::SETGE: Inv = true; return 0;
524 case ISD::SETOGT: // FIXME: This is incorrect see PR642.
526 case ISD::SETGT: Inv = false; return 1;
527 case ISD::SETOLE: // FIXME: This is incorrect see PR642.
529 case ISD::SETLE: Inv = true; return 1;
530 case ISD::SETOEQ: // FIXME: This is incorrect see PR642.
531 case ISD::SETEQ: Inv = false; return 2;
532 case ISD::SETONE: // FIXME: This is incorrect see PR642.
533 case ISD::SETNE: Inv = true; return 2;
534 case ISD::SETO: Inv = true; return 3;
535 case ISD::SETUO: Inv = false; return 3;
541 SDOperand PPCDAGToDAGISel::SelectADD_PARTS(SDOperand Op) {
543 SDOperand LHSL, LHSH;
544 Select(LHSL, N->getOperand(0));
545 Select(LHSH, N->getOperand(1));
548 bool ME = false, ZE = false;
549 if (isIntImmediate(N->getOperand(3), Imm)) {
550 ME = (signed)Imm == -1;
554 std::vector<SDOperand> Result;
555 SDOperand CarryFromLo, Tmp;
556 if (isIntImmediate(N->getOperand(2), Imm) &&
557 ((signed)Imm >= -32768 || (signed)Imm < 32768)) {
558 // Codegen the low 32 bits of the add. Interestingly, there is no
559 // shifted form of add immediate carrying.
560 CarryFromLo = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
561 LHSL, getI32Imm(Imm));
563 Select(Tmp, N->getOperand(2));
564 CarryFromLo = CurDAG->getTargetNode(PPC::ADDC, MVT::i32, MVT::Flag,
567 CarryFromLo = CarryFromLo.getValue(1);
569 // Codegen the high 32 bits, adding zero, minus one, or the full value
570 // along with the carry flag produced by addc/addic.
573 ResultHi = CurDAG->getTargetNode(PPC::ADDZE, MVT::i32, LHSH, CarryFromLo);
575 ResultHi = CurDAG->getTargetNode(PPC::ADDME, MVT::i32, LHSH, CarryFromLo);
577 Select(Tmp, N->getOperand(3));
578 ResultHi = CurDAG->getTargetNode(PPC::ADDE, MVT::i32, LHSH,
581 Result.push_back(CarryFromLo.getValue(0));
582 Result.push_back(ResultHi);
584 CodeGenMap[Op.getValue(0)] = Result[0];
585 CodeGenMap[Op.getValue(1)] = Result[1];
586 return Result[Op.ResNo];
588 SDOperand PPCDAGToDAGISel::SelectSUB_PARTS(SDOperand Op) {
590 SDOperand LHSL, LHSH, RHSL, RHSH;
591 Select(LHSL, N->getOperand(0));
592 Select(LHSH, N->getOperand(1));
593 Select(RHSL, N->getOperand(2));
594 Select(RHSH, N->getOperand(3));
596 std::vector<SDOperand> Result;
597 Result.push_back(CurDAG->getTargetNode(PPC::SUBFC, MVT::i32, MVT::Flag,
599 Result.push_back(CurDAG->getTargetNode(PPC::SUBFE, MVT::i32, RHSH, LHSH,
600 Result[0].getValue(1)));
601 CodeGenMap[Op.getValue(0)] = Result[0];
602 CodeGenMap[Op.getValue(1)] = Result[1];
603 return Result[Op.ResNo];
606 SDOperand PPCDAGToDAGISel::SelectSETCC(SDOperand Op) {
609 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
610 if (isIntImmediate(N->getOperand(1), Imm)) {
611 // We can codegen setcc op, imm very efficiently compared to a brcond.
612 // Check for those cases here.
616 Select(Op, N->getOperand(0));
620 Op = CurDAG->getTargetNode(PPC::CNTLZW, MVT::i32, Op);
621 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(27),
622 getI32Imm(5), getI32Imm(31));
624 SDOperand AD = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
626 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
630 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(1),
631 getI32Imm(31), getI32Imm(31));
633 SDOperand T = CurDAG->getTargetNode(PPC::NEG, MVT::i32, Op);
634 T = CurDAG->getTargetNode(PPC::ANDC, MVT::i32, T, Op);;
635 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, T, getI32Imm(1),
636 getI32Imm(31), getI32Imm(31));
639 } else if (Imm == ~0U) { // setcc op, -1
641 Select(Op, N->getOperand(0));
645 Op = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
647 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
648 CurDAG->getTargetNode(PPC::LI, MVT::i32,
652 Op = CurDAG->getTargetNode(PPC::NOR, MVT::i32, Op, Op);
653 SDOperand AD = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
655 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
659 SDOperand AD = CurDAG->getTargetNode(PPC::ADDI, MVT::i32, Op,
661 SDOperand AN = CurDAG->getTargetNode(PPC::AND, MVT::i32, AD, Op);
662 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, AN, getI32Imm(1),
663 getI32Imm(31), getI32Imm(31));
666 Op = CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, Op, getI32Imm(1),
667 getI32Imm(31), getI32Imm(31));
668 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, getI32Imm(1));
674 unsigned Idx = getCRIdxForSetCC(CC, Inv);
675 SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
678 // Force the ccreg into CR7.
679 SDOperand CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
681 SDOperand InFlag(0, 0); // Null incoming flag value.
682 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), CR7Reg, CCReg,
685 if (TLI.getTargetMachine().getSubtarget<PPCSubtarget>().isGigaProcessor())
686 IntCR = CurDAG->getTargetNode(PPC::MFOCRF, MVT::i32, CR7Reg, CCReg);
688 IntCR = CurDAG->getTargetNode(PPC::MFCR, MVT::i32, CCReg);
691 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, IntCR,
692 getI32Imm((32-(3-Idx)) & 31),
693 getI32Imm(31), getI32Imm(31));
696 CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, IntCR,
697 getI32Imm((32-(3-Idx)) & 31),
698 getI32Imm(31),getI32Imm(31));
699 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
703 /// isCallCompatibleAddress - Return true if the specified 32-bit value is
704 /// representable in the immediate field of a Bx instruction.
705 static bool isCallCompatibleAddress(ConstantSDNode *C) {
706 int Addr = C->getValue();
707 if (Addr & 3) return false; // Low 2 bits are implicitly zero.
708 return (Addr << 6 >> 6) == Addr; // Top 6 bits have to be sext of immediate.
711 SDOperand PPCDAGToDAGISel::SelectCALL(SDOperand Op) {
714 Select(Chain, N->getOperand(0));
717 std::vector<SDOperand> CallOperands;
719 if (GlobalAddressSDNode *GASD =
720 dyn_cast<GlobalAddressSDNode>(N->getOperand(1))) {
721 CallOpcode = PPC::BL;
722 CallOperands.push_back(N->getOperand(1));
723 } else if (ExternalSymbolSDNode *ESSDN =
724 dyn_cast<ExternalSymbolSDNode>(N->getOperand(1))) {
725 CallOpcode = PPC::BL;
726 CallOperands.push_back(N->getOperand(1));
727 } else if (isa<ConstantSDNode>(N->getOperand(1)) &&
728 isCallCompatibleAddress(cast<ConstantSDNode>(N->getOperand(1)))) {
729 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(1));
730 CallOpcode = PPC::BLA;
731 CallOperands.push_back(getI32Imm((int)C->getValue() >> 2));
733 // Copy the callee address into the CTR register.
735 Select(Callee, N->getOperand(1));
736 Chain = CurDAG->getTargetNode(PPC::MTCTR, MVT::Other, Callee, Chain);
738 // Copy the callee address into R12 on darwin.
739 SDOperand R12 = CurDAG->getRegister(PPC::R12, MVT::i32);
740 Chain = CurDAG->getNode(ISD::CopyToReg, MVT::Other, Chain, R12, Callee);
742 CallOperands.push_back(R12);
743 CallOpcode = PPC::BCTRL;
746 unsigned GPR_idx = 0, FPR_idx = 0;
747 static const unsigned GPR[] = {
748 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
749 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
751 static const unsigned FPR[] = {
752 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
753 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
756 SDOperand InFlag; // Null incoming flag value.
758 for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i) {
759 unsigned DestReg = 0;
760 MVT::ValueType RegTy = N->getOperand(i).getValueType();
761 if (RegTy == MVT::i32) {
762 assert(GPR_idx < 8 && "Too many int args");
763 DestReg = GPR[GPR_idx++];
765 assert(MVT::isFloatingPoint(N->getOperand(i).getValueType()) &&
766 "Unpromoted integer arg?");
767 assert(FPR_idx < 13 && "Too many fp args");
768 DestReg = FPR[FPR_idx++];
771 if (N->getOperand(i).getOpcode() != ISD::UNDEF) {
773 Select(Val, N->getOperand(i));
774 Chain = CurDAG->getCopyToReg(Chain, DestReg, Val, InFlag);
775 InFlag = Chain.getValue(1);
776 CallOperands.push_back(CurDAG->getRegister(DestReg, RegTy));
780 // Finally, once everything is in registers to pass to the call, emit the
783 CallOperands.push_back(InFlag); // Strong dep on register copies.
785 CallOperands.push_back(Chain); // Weak dep on whatever occurs before
786 Chain = CurDAG->getTargetNode(CallOpcode, MVT::Other, MVT::Flag,
789 std::vector<SDOperand> CallResults;
791 // If the call has results, copy the values out of the ret val registers.
792 switch (N->getValueType(0)) {
793 default: assert(0 && "Unexpected ret value!");
794 case MVT::Other: break;
796 if (N->getValueType(1) == MVT::i32) {
797 Chain = CurDAG->getCopyFromReg(Chain, PPC::R4, MVT::i32,
798 Chain.getValue(1)).getValue(1);
799 CallResults.push_back(Chain.getValue(0));
800 Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
801 Chain.getValue(2)).getValue(1);
802 CallResults.push_back(Chain.getValue(0));
804 Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
805 Chain.getValue(1)).getValue(1);
806 CallResults.push_back(Chain.getValue(0));
811 Chain = CurDAG->getCopyFromReg(Chain, PPC::F1, N->getValueType(0),
812 Chain.getValue(1)).getValue(1);
813 CallResults.push_back(Chain.getValue(0));
817 CallResults.push_back(Chain);
818 for (unsigned i = 0, e = CallResults.size(); i != e; ++i)
819 CodeGenMap[Op.getValue(i)] = CallResults[i];
820 return CallResults[Op.ResNo];
823 // Select - Convert the specified operand from a target-independent to a
824 // target-specific node if it hasn't already been changed.
825 void PPCDAGToDAGISel::Select(SDOperand &Result, SDOperand Op) {
827 if (N->getOpcode() >= ISD::BUILTIN_OP_END &&
828 N->getOpcode() < PPCISD::FIRST_NUMBER) {
830 return; // Already selected.
833 // If this has already been converted, use it.
834 std::map<SDOperand, SDOperand>::iterator CGMI = CodeGenMap.find(Op);
835 if (CGMI != CodeGenMap.end()) {
836 Result = CGMI->second;
840 switch (N->getOpcode()) {
843 Result = SelectADD_PARTS(Op);
846 Result = SelectSUB_PARTS(Op);
849 Result = SelectSETCC(Op);
852 Result = SelectCALL(Op);
854 case PPCISD::GlobalBaseReg:
855 Result = getGlobalBaseReg();
858 case ISD::FrameIndex: {
859 int FI = cast<FrameIndexSDNode>(N)->getIndex();
860 if (N->hasOneUse()) {
861 Result = CurDAG->SelectNodeTo(N, PPC::ADDI, MVT::i32,
862 CurDAG->getTargetFrameIndex(FI, MVT::i32),
866 Result = CodeGenMap[Op] =
867 CurDAG->getTargetNode(PPC::ADDI, MVT::i32,
868 CurDAG->getTargetFrameIndex(FI, MVT::i32),
873 // FIXME: since this depends on the setting of the carry flag from the srawi
874 // we should really be making notes about that for the scheduler.
875 // FIXME: It sure would be nice if we could cheaply recognize the
876 // srl/add/sra pattern the dag combiner will generate for this as
877 // sra/addze rather than having to handle sdiv ourselves. oh well.
879 if (isIntImmediate(N->getOperand(1), Imm)) {
881 Select(N0, N->getOperand(0));
882 if ((signed)Imm > 0 && isPowerOf2_32(Imm)) {
884 CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
885 N0, getI32Imm(Log2_32(Imm)));
886 Result = CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
887 Op.getValue(0), Op.getValue(1));
888 } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) {
890 CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
891 N0, getI32Imm(Log2_32(-Imm)));
893 CurDAG->getTargetNode(PPC::ADDZE, MVT::i32, Op.getValue(0),
895 Result = CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT);
900 // Other cases are autogenerated.
905 // If this is an and of a value rotated between 0 and 31 bits and then and'd
906 // with a mask, emit rlwinm
907 if (isIntImmediate(N->getOperand(1), Imm) && (isShiftedMask_32(Imm) ||
908 isShiftedMask_32(~Imm))) {
911 if (isRotateAndMask(N->getOperand(0).Val, Imm, false, SH, MB, ME)) {
912 Select(Val, N->getOperand(0).getOperand(0));
913 } else if (Imm == 0) {
914 // AND X, 0 -> 0, not "rlwinm 32".
915 Select(Result, N->getOperand(1));
918 Select(Val, N->getOperand(0));
919 isRunOfOnes(Imm, MB, ME);
922 Result = CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Val,
923 getI32Imm(SH), getI32Imm(MB),
927 // ISD::OR doesn't get all the bitfield insertion fun.
928 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert
929 if (isIntImmediate(N->getOperand(1), Imm) &&
930 N->getOperand(0).getOpcode() == ISD::OR &&
931 isIntImmediate(N->getOperand(0).getOperand(1), Imm2)) {
934 if (isRunOfOnes(Imm, MB, ME)) {
935 SDOperand Tmp1, Tmp2;
936 Select(Tmp1, N->getOperand(0).getOperand(0));
937 Select(Tmp2, N->getOperand(0).getOperand(1));
938 Result = CurDAG->getTargetNode(PPC::RLWIMI, MVT::i32, Tmp1, Tmp2,
939 getI32Imm(0), getI32Imm(MB),
945 // Other cases are autogenerated.
949 if (SDNode *I = SelectBitfieldInsert(N)) {
950 Result = CodeGenMap[Op] = SDOperand(I, 0);
954 // Other cases are autogenerated.
957 unsigned Imm, SH, MB, ME;
958 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
959 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
961 Select(Val, N->getOperand(0).getOperand(0));
962 Result = CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
963 Val, getI32Imm(SH), getI32Imm(MB),
968 // Other cases are autogenerated.
972 unsigned Imm, SH, MB, ME;
973 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
974 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
976 Select(Val, N->getOperand(0).getOperand(0));
977 Result = CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
978 Val, getI32Imm(SH & 0x1F), getI32Imm(MB),
983 // Other cases are autogenerated.
986 case ISD::SELECT_CC: {
987 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
989 // handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
990 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
991 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
992 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
993 if (N1C->isNullValue() && N3C->isNullValue() &&
994 N2C->getValue() == 1ULL && CC == ISD::SETNE) {
996 Select(LHS, N->getOperand(0));
998 CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
999 LHS, getI32Imm(~0U));
1000 Result = CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, Tmp, LHS,
1005 SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
1006 unsigned BROpc = getBCCForSetCC(CC);
1008 bool isFP = MVT::isFloatingPoint(N->getValueType(0));
1009 unsigned SelectCCOp;
1010 if (MVT::isInteger(N->getValueType(0)))
1011 SelectCCOp = PPC::SELECT_CC_Int;
1012 else if (N->getValueType(0) == MVT::f32)
1013 SelectCCOp = PPC::SELECT_CC_F4;
1015 SelectCCOp = PPC::SELECT_CC_F8;
1017 Select(N2, N->getOperand(2));
1018 Select(N3, N->getOperand(3));
1019 Result = CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), CCReg,
1020 N2, N3, getI32Imm(BROpc));
1024 case ISD::BRTWOWAY_CC: {
1026 Select(Chain, N->getOperand(0));
1027 MachineBasicBlock *Dest =
1028 cast<BasicBlockSDNode>(N->getOperand(4))->getBasicBlock();
1029 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
1030 SDOperand CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC);
1032 // If this is a two way branch, then grab the fallthrough basic block
1033 // argument and build a PowerPC branch pseudo-op, suitable for long branch
1034 // conversion if necessary by the branch selection pass. Otherwise, emit a
1035 // standard conditional branch.
1036 if (N->getOpcode() == ISD::BRTWOWAY_CC) {
1037 SDOperand CondTrueBlock = N->getOperand(4);
1038 SDOperand CondFalseBlock = N->getOperand(5);
1040 // If the false case is the current basic block, then this is a self loop.
1041 // We do not want to emit "Loop: ... brcond Out; br Loop", as it adds an
1042 // extra dispatch group to the loop. Instead, invert the condition and
1043 // emit "Loop: ... br!cond Loop; br Out
1044 if (cast<BasicBlockSDNode>(CondFalseBlock)->getBasicBlock() == BB) {
1045 std::swap(CondTrueBlock, CondFalseBlock);
1046 CC = getSetCCInverse(CC,
1047 MVT::isInteger(N->getOperand(2).getValueType()));
1050 unsigned Opc = getBCCForSetCC(CC);
1051 SDOperand CB = CurDAG->getTargetNode(PPC::COND_BRANCH, MVT::Other,
1052 CondCode, getI32Imm(Opc),
1053 CondTrueBlock, CondFalseBlock,
1055 Result = CurDAG->SelectNodeTo(N, PPC::B, MVT::Other, CondFalseBlock, CB);
1057 // Iterate to the next basic block
1058 ilist<MachineBasicBlock>::iterator It = BB;
1061 // If the fallthrough path is off the end of the function, which would be
1062 // undefined behavior, set it to be the same as the current block because
1063 // we have nothing better to set it to, and leaving it alone will cause
1064 // the PowerPC Branch Selection pass to crash.
1065 if (It == BB->getParent()->end()) It = Dest;
1066 Result = CurDAG->SelectNodeTo(N, PPC::COND_BRANCH, MVT::Other, CondCode,
1067 getI32Imm(getBCCForSetCC(CC)),
1068 N->getOperand(4), CurDAG->getBasicBlock(It),
1075 SelectCode(Result, Op);
1079 /// createPPCISelDag - This pass converts a legalized DAG into a
1080 /// PowerPC-specific DAG, ready for instruction scheduling.
1082 FunctionPass *llvm::createPPCISelDag(TargetMachine &TM) {
1083 return new PPCDAGToDAGISel(TM);