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 return new PPCHazardRecognizer970();
135 // Include the pieces autogenerated from the target description.
136 #include "PPCGenDAGISel.inc"
139 SDOperand SelectSETCC(SDOperand Op);
140 SDOperand SelectCALL(SDOperand Op);
144 /// InstructionSelectBasicBlock - This callback is invoked by
145 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
146 void PPCDAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
149 // The selection process is inherently a bottom-up recursive process (users
150 // select their uses before themselves). Given infinite stack space, we
151 // could just start selecting on the root and traverse the whole graph. In
152 // practice however, this causes us to run out of stack space on large basic
153 // blocks. To avoid this problem, select the entry node, then all its uses,
154 // iteratively instead of recursively.
155 std::vector<SDOperand> Worklist;
156 Worklist.push_back(DAG.getEntryNode());
158 // Note that we can do this in the PPC target (scanning forward across token
159 // chain edges) because no nodes ever get folded across these edges. On a
160 // target like X86 which supports load/modify/store operations, this would
161 // have to be more careful.
162 while (!Worklist.empty()) {
163 SDOperand Node = Worklist.back();
166 // Chose from the least deep of the top two nodes.
167 if (!Worklist.empty() &&
168 Worklist.back().Val->getNodeDepth() < Node.Val->getNodeDepth())
169 std::swap(Worklist.back(), Node);
171 if ((Node.Val->getOpcode() >= ISD::BUILTIN_OP_END &&
172 Node.Val->getOpcode() < PPCISD::FIRST_NUMBER) ||
173 CodeGenMap.count(Node)) continue;
175 for (SDNode::use_iterator UI = Node.Val->use_begin(),
176 E = Node.Val->use_end(); UI != E; ++UI) {
177 // Scan the values. If this use has a value that is a token chain, add it
180 for (unsigned i = 0, e = User->getNumValues(); i != e; ++i)
181 if (User->getValueType(i) == MVT::Other) {
182 Worklist.push_back(SDOperand(User, i));
187 // Finally, legalize this node.
192 // Select target instructions for the DAG.
193 DAG.setRoot(SelectRoot(DAG.getRoot()));
195 DAG.RemoveDeadNodes();
197 // Emit machine code to BB.
198 ScheduleAndEmitDAG(DAG);
201 /// getGlobalBaseReg - Output the instructions required to put the
202 /// base address to use for accessing globals into a register.
204 SDOperand PPCDAGToDAGISel::getGlobalBaseReg() {
205 if (!GlobalBaseReg) {
206 // Insert the set of GlobalBaseReg into the first MBB of the function
207 MachineBasicBlock &FirstMBB = BB->getParent()->front();
208 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
209 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
210 // FIXME: when we get to LP64, we will need to create the appropriate
211 // type of register here.
212 GlobalBaseReg = RegMap->createVirtualRegister(PPC::GPRCRegisterClass);
213 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
214 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg);
216 return CurDAG->getRegister(GlobalBaseReg, MVT::i32);
220 // isIntImmediate - This method tests to see if a constant operand.
221 // If so Imm will receive the 32 bit value.
222 static bool isIntImmediate(SDNode *N, unsigned& Imm) {
223 if (N->getOpcode() == ISD::Constant) {
224 Imm = cast<ConstantSDNode>(N)->getValue();
230 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
231 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
232 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
233 // not, since all 1s are not contiguous.
234 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
235 if (isShiftedMask_32(Val)) {
236 // look for the first non-zero bit
237 MB = CountLeadingZeros_32(Val);
238 // look for the first zero bit after the run of ones
239 ME = CountLeadingZeros_32((Val - 1) ^ Val);
242 Val = ~Val; // invert mask
243 if (isShiftedMask_32(Val)) {
244 // effectively look for the first zero bit
245 ME = CountLeadingZeros_32(Val) - 1;
246 // effectively look for the first one bit after the run of zeros
247 MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
255 // isRotateAndMask - Returns true if Mask and Shift can be folded into a rotate
256 // and mask opcode and mask operation.
257 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool IsShiftMask,
258 unsigned &SH, unsigned &MB, unsigned &ME) {
259 // Don't even go down this path for i64, since different logic will be
260 // necessary for rldicl/rldicr/rldimi.
261 if (N->getValueType(0) != MVT::i32)
265 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
266 unsigned Opcode = N->getOpcode();
267 if (N->getNumOperands() != 2 ||
268 !isIntImmediate(N->getOperand(1).Val, Shift) || (Shift > 31))
271 if (Opcode == ISD::SHL) {
272 // apply shift left to mask if it comes first
273 if (IsShiftMask) Mask = Mask << Shift;
274 // determine which bits are made indeterminant by shift
275 Indeterminant = ~(0xFFFFFFFFu << Shift);
276 } else if (Opcode == ISD::SRL) {
277 // apply shift right to mask if it comes first
278 if (IsShiftMask) Mask = Mask >> Shift;
279 // determine which bits are made indeterminant by shift
280 Indeterminant = ~(0xFFFFFFFFu >> Shift);
281 // adjust for the left rotate
287 // if the mask doesn't intersect any Indeterminant bits
288 if (Mask && !(Mask & Indeterminant)) {
290 // make sure the mask is still a mask (wrap arounds may not be)
291 return isRunOfOnes(Mask, MB, ME);
296 // isOpcWithIntImmediate - This method tests to see if the node is a specific
297 // opcode and that it has a immediate integer right operand.
298 // If so Imm will receive the 32 bit value.
299 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
300 return N->getOpcode() == Opc && isIntImmediate(N->getOperand(1).Val, Imm);
303 // isIntImmediate - This method tests to see if a constant operand.
304 // If so Imm will receive the 32 bit value.
305 static bool isIntImmediate(SDOperand N, unsigned& Imm) {
306 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
307 Imm = (unsigned)CN->getSignExtended();
313 /// SelectBitfieldInsert - turn an or of two masked values into
314 /// the rotate left word immediate then mask insert (rlwimi) instruction.
315 /// Returns true on success, false if the caller still needs to select OR.
317 /// Patterns matched:
318 /// 1. or shl, and 5. or and, and
319 /// 2. or and, shl 6. or shl, shr
320 /// 3. or shr, and 7. or shr, shl
322 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
323 bool IsRotate = false;
324 unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, SH = 0;
327 SDOperand Op0 = N->getOperand(0);
328 SDOperand Op1 = N->getOperand(1);
330 unsigned Op0Opc = Op0.getOpcode();
331 unsigned Op1Opc = Op1.getOpcode();
333 // Verify that we have the correct opcodes
334 if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc)
336 if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc)
339 // Generate Mask value for Target
340 if (isIntImmediate(Op0.getOperand(1), Value)) {
342 case ISD::SHL: TgtMask <<= Value; break;
343 case ISD::SRL: TgtMask >>= Value; break;
344 case ISD::AND: TgtMask &= Value; break;
350 // Generate Mask value for Insert
351 if (!isIntImmediate(Op1.getOperand(1), Value))
358 if (Op0Opc == ISD::SRL) IsRotate = true;
364 if (Op0Opc == ISD::SHL) IsRotate = true;
371 // If both of the inputs are ANDs and one of them has a logical shift by
372 // constant as its input, make that AND the inserted value so that we can
373 // combine the shift into the rotate part of the rlwimi instruction
374 bool IsAndWithShiftOp = false;
375 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
376 if (Op1.getOperand(0).getOpcode() == ISD::SHL ||
377 Op1.getOperand(0).getOpcode() == ISD::SRL) {
378 if (isIntImmediate(Op1.getOperand(0).getOperand(1), Value)) {
379 SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
380 IsAndWithShiftOp = true;
382 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
383 Op0.getOperand(0).getOpcode() == ISD::SRL) {
384 if (isIntImmediate(Op0.getOperand(0).getOperand(1), Value)) {
386 std::swap(TgtMask, InsMask);
387 SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
388 IsAndWithShiftOp = true;
393 // Verify that the Target mask and Insert mask together form a full word mask
394 // and that the Insert mask is a run of set bits (which implies both are runs
395 // of set bits). Given that, Select the arguments and generate the rlwimi
398 if (((TgtMask & InsMask) == 0) && isRunOfOnes(InsMask, MB, ME)) {
399 bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF;
400 bool Op0IsAND = Op0Opc == ISD::AND;
401 // Check for rotlwi / rotrwi here, a special case of bitfield insert
402 // where both bitfield halves are sourced from the same value.
403 if (IsRotate && fullMask &&
404 N->getOperand(0).getOperand(0) == N->getOperand(1).getOperand(0)) {
406 Select(Tmp, N->getOperand(0).getOperand(0));
407 return CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, Tmp,
408 getI32Imm(SH), getI32Imm(0), getI32Imm(31));
410 SDOperand Tmp1, Tmp2;
411 Select(Tmp1, ((Op0IsAND && fullMask) ? Op0.getOperand(0) : Op0));
412 Select(Tmp2, (IsAndWithShiftOp ? Op1.getOperand(0).getOperand(0)
413 : Op1.getOperand(0)));
414 return CurDAG->getTargetNode(PPC::RLWIMI, MVT::i32, Tmp1, Tmp2,
415 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME));
420 /// SelectAddrImm - Returns true if the address N can be represented by
421 /// a base register plus a signed 16-bit displacement [r+imm].
422 bool PPCDAGToDAGISel::SelectAddrImm(SDOperand N, SDOperand &Disp,
424 // If this can be more profitably realized as r+r, fail.
425 if (SelectAddrIdx(N, Disp, Base))
428 if (N.getOpcode() == ISD::ADD) {
430 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm)) {
431 Disp = getI32Imm(imm & 0xFFFF);
432 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
433 Base = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
435 Base = N.getOperand(0);
437 return true; // [r+i]
438 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
439 // Match LOAD (ADD (X, Lo(G))).
440 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
441 && "Cannot handle constant offsets yet!");
442 Disp = N.getOperand(1).getOperand(0); // The global address.
443 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
444 Disp.getOpcode() == ISD::TargetConstantPool);
445 Base = N.getOperand(0);
446 return true; // [&g+r]
448 } else if (N.getOpcode() == ISD::OR) {
450 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm)) {
451 // If this is an or of disjoint bitfields, we can codegen this as an add
452 // (for better address arithmetic) if the LHS and RHS of the OR are
453 // provably disjoint.
454 uint64_t LHSKnownZero, LHSKnownOne;
455 PPCLowering.ComputeMaskedBits(N.getOperand(0), ~0U,
456 LHSKnownZero, LHSKnownOne);
457 if ((LHSKnownZero|~imm) == ~0U) {
458 // If all of the bits are known zero on the LHS or RHS, the add won't
460 Base = N.getOperand(0);
461 Disp = getI32Imm(imm & 0xFFFF);
467 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
468 Base = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
471 return true; // [r+0]
474 /// SelectAddrIdx - Given the specified addressed, check to see if it can be
475 /// represented as an indexed [r+r] operation. Returns false if it can
476 /// be represented by [r+imm], which are preferred.
477 bool PPCDAGToDAGISel::SelectAddrIdx(SDOperand N, SDOperand &Base,
480 if (N.getOpcode() == ISD::ADD) {
481 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm))
483 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
486 Base = N.getOperand(0);
487 Index = N.getOperand(1);
489 } else if (N.getOpcode() == ISD::OR) {
490 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm))
491 return false; // r+i can fold it if we can.
493 // If this is an or of disjoint bitfields, we can codegen this as an add
494 // (for better address arithmetic) if the LHS and RHS of the OR are provably
496 uint64_t LHSKnownZero, LHSKnownOne;
497 uint64_t RHSKnownZero, RHSKnownOne;
498 PPCLowering.ComputeMaskedBits(N.getOperand(0), ~0U,
499 LHSKnownZero, LHSKnownOne);
502 PPCLowering.ComputeMaskedBits(N.getOperand(1), ~0U,
503 RHSKnownZero, RHSKnownOne);
504 // If all of the bits are known zero on the LHS or RHS, the add won't
506 if ((LHSKnownZero | RHSKnownZero) == ~0U) {
507 Base = N.getOperand(0);
508 Index = N.getOperand(1);
517 /// SelectAddrIdxOnly - Given the specified addressed, force it to be
518 /// represented as an indexed [r+r] operation.
519 bool PPCDAGToDAGISel::SelectAddrIdxOnly(SDOperand N, SDOperand &Base,
521 // Check to see if we can easily represent this as an [r+r] address. This
522 // will fail if it thinks that the address is more profitably represented as
523 // reg+imm, e.g. where imm = 0.
524 if (!SelectAddrIdx(N, Base, Index)) {
525 // Nope, do it the hard way.
526 Base = CurDAG->getRegister(PPC::R0, MVT::i32);
532 /// SelectCC - Select a comparison of the specified values with the specified
533 /// condition code, returning the CR# of the expression.
534 SDOperand PPCDAGToDAGISel::SelectCC(SDOperand LHS, SDOperand RHS,
536 // Always select the LHS.
539 // Use U to determine whether the SETCC immediate range is signed or not.
540 if (MVT::isInteger(LHS.getValueType())) {
541 bool U = ISD::isUnsignedIntSetCC(CC);
543 if (isIntImmediate(RHS, Imm) &&
544 ((U && isUInt16(Imm)) || (!U && isInt16(Imm))))
545 return SDOperand(CurDAG->getTargetNode(U ? PPC::CMPLWI : PPC::CMPWI,
546 MVT::i32, LHS, getI32Imm(Imm & 0xFFFF)), 0);
548 return SDOperand(CurDAG->getTargetNode(U ? PPC::CMPLW : PPC::CMPW, MVT::i32,
550 } else if (LHS.getValueType() == MVT::f32) {
552 return SDOperand(CurDAG->getTargetNode(PPC::FCMPUS, MVT::i32, LHS, RHS), 0);
555 return SDOperand(CurDAG->getTargetNode(PPC::FCMPUD, MVT::i32, LHS, RHS), 0);
559 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
561 static unsigned getBCCForSetCC(ISD::CondCode CC) {
563 default: assert(0 && "Unknown condition!"); abort();
564 case ISD::SETOEQ: // FIXME: This is incorrect see PR642.
565 case ISD::SETEQ: return PPC::BEQ;
566 case ISD::SETONE: // FIXME: This is incorrect see PR642.
567 case ISD::SETNE: return PPC::BNE;
568 case ISD::SETOLT: // FIXME: This is incorrect see PR642.
570 case ISD::SETLT: return PPC::BLT;
571 case ISD::SETOLE: // FIXME: This is incorrect see PR642.
573 case ISD::SETLE: return PPC::BLE;
574 case ISD::SETOGT: // FIXME: This is incorrect see PR642.
576 case ISD::SETGT: return PPC::BGT;
577 case ISD::SETOGE: // FIXME: This is incorrect see PR642.
579 case ISD::SETGE: return PPC::BGE;
581 case ISD::SETO: return PPC::BUN;
582 case ISD::SETUO: return PPC::BNU;
587 /// getCRIdxForSetCC - Return the index of the condition register field
588 /// associated with the SetCC condition, and whether or not the field is
589 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
590 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool& Inv) {
592 default: assert(0 && "Unknown condition!"); abort();
593 case ISD::SETOLT: // FIXME: This is incorrect see PR642.
595 case ISD::SETLT: Inv = false; return 0;
596 case ISD::SETOGE: // FIXME: This is incorrect see PR642.
598 case ISD::SETGE: Inv = true; return 0;
599 case ISD::SETOGT: // FIXME: This is incorrect see PR642.
601 case ISD::SETGT: Inv = false; return 1;
602 case ISD::SETOLE: // FIXME: This is incorrect see PR642.
604 case ISD::SETLE: Inv = true; return 1;
605 case ISD::SETOEQ: // FIXME: This is incorrect see PR642.
606 case ISD::SETEQ: Inv = false; return 2;
607 case ISD::SETONE: // FIXME: This is incorrect see PR642.
608 case ISD::SETNE: Inv = true; return 2;
609 case ISD::SETO: Inv = true; return 3;
610 case ISD::SETUO: Inv = false; return 3;
615 SDOperand PPCDAGToDAGISel::SelectSETCC(SDOperand Op) {
618 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
619 if (isIntImmediate(N->getOperand(1), Imm)) {
620 // We can codegen setcc op, imm very efficiently compared to a brcond.
621 // Check for those cases here.
625 Select(Op, N->getOperand(0));
629 Op = SDOperand(CurDAG->getTargetNode(PPC::CNTLZW, MVT::i32, Op), 0);
630 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(27),
631 getI32Imm(5), getI32Imm(31));
634 SDOperand(CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
635 Op, getI32Imm(~0U)), 0);
636 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
640 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(1),
641 getI32Imm(31), getI32Imm(31));
644 SDOperand(CurDAG->getTargetNode(PPC::NEG, MVT::i32, Op), 0);
645 T = SDOperand(CurDAG->getTargetNode(PPC::ANDC, MVT::i32, T, Op), 0);
646 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, T, getI32Imm(1),
647 getI32Imm(31), getI32Imm(31));
650 } else if (Imm == ~0U) { // setcc op, -1
652 Select(Op, N->getOperand(0));
656 Op = SDOperand(CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
657 Op, getI32Imm(1)), 0);
658 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
659 SDOperand(CurDAG->getTargetNode(PPC::LI, MVT::i32,
663 Op = SDOperand(CurDAG->getTargetNode(PPC::NOR, MVT::i32, Op, Op), 0);
664 SDNode *AD = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
666 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDOperand(AD, 0), Op,
670 SDOperand AD = SDOperand(CurDAG->getTargetNode(PPC::ADDI, MVT::i32, Op,
672 SDOperand AN = SDOperand(CurDAG->getTargetNode(PPC::AND, MVT::i32, AD,
674 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, AN, getI32Imm(1),
675 getI32Imm(31), getI32Imm(31));
678 Op = SDOperand(CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, Op,
679 getI32Imm(1), getI32Imm(31),
681 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, getI32Imm(1));
687 unsigned Idx = getCRIdxForSetCC(CC, Inv);
688 SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
691 // Force the ccreg into CR7.
692 SDOperand CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
694 SDOperand InFlag(0, 0); // Null incoming flag value.
695 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), CR7Reg, CCReg,
698 if (TLI.getTargetMachine().getSubtarget<PPCSubtarget>().isGigaProcessor())
699 IntCR = SDOperand(CurDAG->getTargetNode(PPC::MFOCRF, MVT::i32, CR7Reg,
702 IntCR = SDOperand(CurDAG->getTargetNode(PPC::MFCR, MVT::i32, CCReg), 0);
705 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, IntCR,
706 getI32Imm((32-(3-Idx)) & 31),
707 getI32Imm(31), getI32Imm(31));
710 SDOperand(CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, IntCR,
711 getI32Imm((32-(3-Idx)) & 31),
712 getI32Imm(31),getI32Imm(31)), 0);
713 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
717 /// isCallCompatibleAddress - Return true if the specified 32-bit value is
718 /// representable in the immediate field of a Bx instruction.
719 static bool isCallCompatibleAddress(ConstantSDNode *C) {
720 int Addr = C->getValue();
721 if (Addr & 3) return false; // Low 2 bits are implicitly zero.
722 return (Addr << 6 >> 6) == Addr; // Top 6 bits have to be sext of immediate.
725 SDOperand PPCDAGToDAGISel::SelectCALL(SDOperand Op) {
728 Select(Chain, N->getOperand(0));
731 std::vector<SDOperand> CallOperands;
733 if (GlobalAddressSDNode *GASD =
734 dyn_cast<GlobalAddressSDNode>(N->getOperand(1))) {
735 CallOpcode = PPC::BL;
736 CallOperands.push_back(N->getOperand(1));
737 } else if (ExternalSymbolSDNode *ESSDN =
738 dyn_cast<ExternalSymbolSDNode>(N->getOperand(1))) {
739 CallOpcode = PPC::BL;
740 CallOperands.push_back(N->getOperand(1));
741 } else if (isa<ConstantSDNode>(N->getOperand(1)) &&
742 isCallCompatibleAddress(cast<ConstantSDNode>(N->getOperand(1)))) {
743 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(1));
744 CallOpcode = PPC::BLA;
745 CallOperands.push_back(getI32Imm((int)C->getValue() >> 2));
747 // Copy the callee address into the CTR register.
749 Select(Callee, N->getOperand(1));
750 Chain = SDOperand(CurDAG->getTargetNode(PPC::MTCTR, MVT::Other, Callee,
753 // Copy the callee address into R12 on darwin.
754 SDOperand R12 = CurDAG->getRegister(PPC::R12, MVT::i32);
755 Chain = CurDAG->getNode(ISD::CopyToReg, MVT::Other, Chain, R12, Callee);
757 CallOperands.push_back(R12);
758 CallOpcode = PPC::BCTRL;
761 unsigned GPR_idx = 0, FPR_idx = 0;
762 static const unsigned GPR[] = {
763 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
764 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
766 static const unsigned FPR[] = {
767 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
768 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
771 SDOperand InFlag; // Null incoming flag value.
773 for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i) {
774 unsigned DestReg = 0;
775 MVT::ValueType RegTy = N->getOperand(i).getValueType();
776 if (RegTy == MVT::i32) {
777 assert(GPR_idx < 8 && "Too many int args");
778 DestReg = GPR[GPR_idx++];
780 assert(MVT::isFloatingPoint(N->getOperand(i).getValueType()) &&
781 "Unpromoted integer arg?");
782 assert(FPR_idx < 13 && "Too many fp args");
783 DestReg = FPR[FPR_idx++];
786 if (N->getOperand(i).getOpcode() != ISD::UNDEF) {
788 Select(Val, N->getOperand(i));
789 Chain = CurDAG->getCopyToReg(Chain, DestReg, Val, InFlag);
790 InFlag = Chain.getValue(1);
791 CallOperands.push_back(CurDAG->getRegister(DestReg, RegTy));
795 // Finally, once everything is in registers to pass to the call, emit the
798 CallOperands.push_back(InFlag); // Strong dep on register copies.
800 CallOperands.push_back(Chain); // Weak dep on whatever occurs before
801 Chain = SDOperand(CurDAG->getTargetNode(CallOpcode, MVT::Other, MVT::Flag,
804 std::vector<SDOperand> CallResults;
806 // If the call has results, copy the values out of the ret val registers.
807 switch (N->getValueType(0)) {
808 default: assert(0 && "Unexpected ret value!");
809 case MVT::Other: break;
811 if (N->getValueType(1) == MVT::i32) {
812 Chain = CurDAG->getCopyFromReg(Chain, PPC::R4, MVT::i32,
813 Chain.getValue(1)).getValue(1);
814 CallResults.push_back(Chain.getValue(0));
815 Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
816 Chain.getValue(2)).getValue(1);
817 CallResults.push_back(Chain.getValue(0));
819 Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
820 Chain.getValue(1)).getValue(1);
821 CallResults.push_back(Chain.getValue(0));
826 Chain = CurDAG->getCopyFromReg(Chain, PPC::F1, N->getValueType(0),
827 Chain.getValue(1)).getValue(1);
828 CallResults.push_back(Chain.getValue(0));
832 CallResults.push_back(Chain);
833 for (unsigned i = 0, e = CallResults.size(); i != e; ++i)
834 CodeGenMap[Op.getValue(i)] = CallResults[i];
835 return CallResults[Op.ResNo];
838 // Select - Convert the specified operand from a target-independent to a
839 // target-specific node if it hasn't already been changed.
840 void PPCDAGToDAGISel::Select(SDOperand &Result, SDOperand Op) {
842 if (N->getOpcode() >= ISD::BUILTIN_OP_END &&
843 N->getOpcode() < PPCISD::FIRST_NUMBER) {
845 return; // Already selected.
848 // If this has already been converted, use it.
849 std::map<SDOperand, SDOperand>::iterator CGMI = CodeGenMap.find(Op);
850 if (CGMI != CodeGenMap.end()) {
851 Result = CGMI->second;
855 switch (N->getOpcode()) {
858 Result = SelectSETCC(Op);
861 Result = SelectCALL(Op);
863 case PPCISD::GlobalBaseReg:
864 Result = getGlobalBaseReg();
867 case ISD::FrameIndex: {
868 int FI = cast<FrameIndexSDNode>(N)->getIndex();
869 if (N->hasOneUse()) {
870 Result = CurDAG->SelectNodeTo(N, PPC::ADDI, MVT::i32,
871 CurDAG->getTargetFrameIndex(FI, MVT::i32),
875 Result = CodeGenMap[Op] =
876 SDOperand(CurDAG->getTargetNode(PPC::ADDI, MVT::i32,
877 CurDAG->getTargetFrameIndex(FI, MVT::i32),
882 // FIXME: since this depends on the setting of the carry flag from the srawi
883 // we should really be making notes about that for the scheduler.
884 // FIXME: It sure would be nice if we could cheaply recognize the
885 // srl/add/sra pattern the dag combiner will generate for this as
886 // sra/addze rather than having to handle sdiv ourselves. oh well.
888 if (isIntImmediate(N->getOperand(1), Imm)) {
890 Select(N0, N->getOperand(0));
891 if ((signed)Imm > 0 && isPowerOf2_32(Imm)) {
893 CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
894 N0, getI32Imm(Log2_32(Imm)));
895 Result = CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
896 SDOperand(Op, 0), SDOperand(Op, 1));
897 } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) {
899 CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
900 N0, getI32Imm(Log2_32(-Imm)));
902 SDOperand(CurDAG->getTargetNode(PPC::ADDZE, MVT::i32,
903 SDOperand(Op, 0), SDOperand(Op, 1)),
905 Result = CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT);
910 // Other cases are autogenerated.
915 // If this is an and of a value rotated between 0 and 31 bits and then and'd
916 // with a mask, emit rlwinm
917 if (isIntImmediate(N->getOperand(1), Imm) && (isShiftedMask_32(Imm) ||
918 isShiftedMask_32(~Imm))) {
921 if (isRotateAndMask(N->getOperand(0).Val, Imm, false, SH, MB, ME)) {
922 Select(Val, N->getOperand(0).getOperand(0));
923 } else if (Imm == 0) {
924 // AND X, 0 -> 0, not "rlwinm 32".
925 Select(Result, N->getOperand(1));
928 Select(Val, N->getOperand(0));
929 isRunOfOnes(Imm, MB, ME);
932 Result = CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Val,
933 getI32Imm(SH), getI32Imm(MB),
937 // ISD::OR doesn't get all the bitfield insertion fun.
938 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert
939 if (isIntImmediate(N->getOperand(1), Imm) &&
940 N->getOperand(0).getOpcode() == ISD::OR &&
941 isIntImmediate(N->getOperand(0).getOperand(1), Imm2)) {
944 if (isRunOfOnes(Imm, MB, ME)) {
945 SDOperand Tmp1, Tmp2;
946 Select(Tmp1, N->getOperand(0).getOperand(0));
947 Select(Tmp2, N->getOperand(0).getOperand(1));
948 Result = SDOperand(CurDAG->getTargetNode(PPC::RLWIMI, MVT::i32,
950 getI32Imm(0), getI32Imm(MB),
956 // Other cases are autogenerated.
960 if (SDNode *I = SelectBitfieldInsert(N)) {
961 Result = CodeGenMap[Op] = SDOperand(I, 0);
965 // Other cases are autogenerated.
968 unsigned Imm, SH, MB, ME;
969 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
970 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
972 Select(Val, N->getOperand(0).getOperand(0));
973 Result = CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
974 Val, getI32Imm(SH), getI32Imm(MB),
979 // Other cases are autogenerated.
983 unsigned Imm, SH, MB, ME;
984 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
985 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
987 Select(Val, N->getOperand(0).getOperand(0));
988 Result = CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
989 Val, getI32Imm(SH & 0x1F), getI32Imm(MB),
994 // Other cases are autogenerated.
997 case ISD::SELECT_CC: {
998 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1000 // handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
1001 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1002 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
1003 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
1004 if (N1C->isNullValue() && N3C->isNullValue() &&
1005 N2C->getValue() == 1ULL && CC == ISD::SETNE) {
1007 Select(LHS, N->getOperand(0));
1009 CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
1010 LHS, getI32Imm(~0U));
1011 Result = CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32,
1012 SDOperand(Tmp, 0), LHS,
1017 SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
1018 unsigned BROpc = getBCCForSetCC(CC);
1020 bool isFP = MVT::isFloatingPoint(N->getValueType(0));
1021 unsigned SelectCCOp;
1022 if (MVT::isInteger(N->getValueType(0)))
1023 SelectCCOp = PPC::SELECT_CC_Int;
1024 else if (N->getValueType(0) == MVT::f32)
1025 SelectCCOp = PPC::SELECT_CC_F4;
1027 SelectCCOp = PPC::SELECT_CC_F8;
1029 Select(N2, N->getOperand(2));
1030 Select(N3, N->getOperand(3));
1031 Result = CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), CCReg,
1032 N2, N3, getI32Imm(BROpc));
1036 case ISD::BRTWOWAY_CC: {
1038 Select(Chain, N->getOperand(0));
1039 MachineBasicBlock *Dest =
1040 cast<BasicBlockSDNode>(N->getOperand(4))->getBasicBlock();
1041 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
1042 SDOperand CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC);
1044 // If this is a two way branch, then grab the fallthrough basic block
1045 // argument and build a PowerPC branch pseudo-op, suitable for long branch
1046 // conversion if necessary by the branch selection pass. Otherwise, emit a
1047 // standard conditional branch.
1048 if (N->getOpcode() == ISD::BRTWOWAY_CC) {
1049 SDOperand CondTrueBlock = N->getOperand(4);
1050 SDOperand CondFalseBlock = N->getOperand(5);
1051 unsigned Opc = getBCCForSetCC(CC);
1053 SDOperand(CurDAG->getTargetNode(PPC::COND_BRANCH, MVT::Other,
1054 CondCode, getI32Imm(Opc),
1055 CondTrueBlock, CondFalseBlock,
1057 Result = CurDAG->SelectNodeTo(N, PPC::B, MVT::Other, CondFalseBlock, CB);
1059 // Iterate to the next basic block
1060 ilist<MachineBasicBlock>::iterator It = BB;
1063 // If the fallthrough path is off the end of the function, which would be
1064 // undefined behavior, set it to be the same as the current block because
1065 // we have nothing better to set it to, and leaving it alone will cause
1066 // the PowerPC Branch Selection pass to crash.
1067 if (It == BB->getParent()->end()) It = Dest;
1068 Result = CurDAG->SelectNodeTo(N, PPC::COND_BRANCH, MVT::Other, CondCode,
1069 getI32Imm(getBCCForSetCC(CC)),
1070 N->getOperand(4), CurDAG->getBasicBlock(It),
1077 SelectCode(Result, Op);
1081 /// createPPCISelDag - This pass converts a legalized DAG into a
1082 /// PowerPC-specific DAG, ready for instruction scheduling.
1084 FunctionPass *llvm::createPPCISelDag(TargetMachine &TM) {
1085 return new PPCDAGToDAGISel(TM);