1 //===-- PPCISelDAGToDAG.cpp - PPC --pattern matching inst selector --------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a pattern matching instruction selector for PowerPC,
11 // converting from a legalized dag to a PPC dag.
13 //===----------------------------------------------------------------------===//
16 #include "PPCTargetMachine.h"
17 #include "PPCISelLowering.h"
18 #include "PPCHazardRecognizers.h"
19 #include "llvm/CodeGen/MachineInstrBuilder.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/SSARegMap.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/CodeGen/SelectionDAGISel.h"
24 #include "llvm/Target/TargetOptions.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/Constants.h"
27 #include "llvm/GlobalValue.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/MathExtras.h"
35 Statistic<> FrameOff("ppc-codegen", "Number of frame idx offsets collapsed");
37 //===--------------------------------------------------------------------===//
38 /// PPCDAGToDAGISel - PPC specific code to select PPC machine
39 /// instructions for SelectionDAG operations.
41 class PPCDAGToDAGISel : public SelectionDAGISel {
42 PPCTargetLowering PPCLowering;
43 unsigned GlobalBaseReg;
45 PPCDAGToDAGISel(TargetMachine &TM)
46 : SelectionDAGISel(PPCLowering), PPCLowering(TM) {}
48 virtual bool runOnFunction(Function &Fn) {
49 // Make sure we re-emit a set of the global base reg if necessary
51 return SelectionDAGISel::runOnFunction(Fn);
54 /// getI32Imm - Return a target constant with the specified value, of type
56 inline SDOperand getI32Imm(unsigned Imm) {
57 return CurDAG->getTargetConstant(Imm, MVT::i32);
60 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
61 /// base register. Return the virtual register that holds this value.
62 SDOperand getGlobalBaseReg();
64 // Select - Convert the specified operand from a target-independent to a
65 // target-specific node if it hasn't already been changed.
66 void Select(SDOperand &Result, SDOperand Op);
68 SDNode *SelectBitfieldInsert(SDNode *N);
70 /// SelectCC - Select a comparison of the specified values with the
71 /// specified condition code, returning the CR# of the expression.
72 SDOperand SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC);
74 /// SelectAddrImm - Returns true if the address N can be represented by
75 /// a base register plus a signed 16-bit displacement [r+imm].
76 bool SelectAddrImm(SDOperand N, SDOperand &Disp, SDOperand &Base);
78 /// SelectAddrIdx - Given the specified addressed, check to see if it can be
79 /// represented as an indexed [r+r] operation. Returns false if it can
80 /// be represented by [r+imm], which are preferred.
81 bool SelectAddrIdx(SDOperand N, SDOperand &Base, SDOperand &Index);
83 /// SelectAddrIdxOnly - Given the specified addressed, force it to be
84 /// represented as an indexed [r+r] operation.
85 bool SelectAddrIdxOnly(SDOperand N, SDOperand &Base, SDOperand &Index);
87 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
88 /// inline asm expressions.
89 virtual bool SelectInlineAsmMemoryOperand(const SDOperand &Op,
91 std::vector<SDOperand> &OutOps,
94 switch (ConstraintCode) {
97 if (!SelectAddrIdx(Op, Op0, Op1))
98 SelectAddrImm(Op, Op0, Op1);
100 case 'o': // offsetable
101 if (!SelectAddrImm(Op, Op0, Op1)) {
102 Select(Op0, Op); // r+0.
106 case 'v': // not offsetable
107 SelectAddrIdxOnly(Op, Op0, Op1);
111 OutOps.push_back(Op0);
112 OutOps.push_back(Op1);
116 SDOperand BuildSDIVSequence(SDNode *N);
117 SDOperand BuildUDIVSequence(SDNode *N);
119 /// InstructionSelectBasicBlock - This callback is invoked by
120 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
121 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
123 virtual const char *getPassName() const {
124 return "PowerPC DAG->DAG Pattern Instruction Selection";
127 /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for this
128 /// target when scheduling the DAG.
129 virtual HazardRecognizer *CreateTargetHazardRecognizer() {
130 // Should use subtarget info to pick the right hazard recognizer. For
131 // now, always return a PPC970 recognizer.
132 const TargetInstrInfo *II = PPCLowering.getTargetMachine().getInstrInfo();
133 assert(II && "No InstrInfo?");
134 return new PPCHazardRecognizer970(*II);
137 // Include the pieces autogenerated from the target description.
138 #include "PPCGenDAGISel.inc"
141 SDOperand SelectSETCC(SDOperand Op);
142 SDOperand SelectCALL(SDOperand Op);
146 /// InstructionSelectBasicBlock - This callback is invoked by
147 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
148 void PPCDAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
151 // The selection process is inherently a bottom-up recursive process (users
152 // select their uses before themselves). Given infinite stack space, we
153 // could just start selecting on the root and traverse the whole graph. In
154 // practice however, this causes us to run out of stack space on large basic
155 // blocks. To avoid this problem, select the entry node, then all its uses,
156 // iteratively instead of recursively.
157 std::vector<SDOperand> Worklist;
158 Worklist.push_back(DAG.getEntryNode());
160 // Note that we can do this in the PPC target (scanning forward across token
161 // chain edges) because no nodes ever get folded across these edges. On a
162 // target like X86 which supports load/modify/store operations, this would
163 // have to be more careful.
164 while (!Worklist.empty()) {
165 SDOperand Node = Worklist.back();
168 // Chose from the least deep of the top two nodes.
169 if (!Worklist.empty() &&
170 Worklist.back().Val->getNodeDepth() < Node.Val->getNodeDepth())
171 std::swap(Worklist.back(), Node);
173 if ((Node.Val->getOpcode() >= ISD::BUILTIN_OP_END &&
174 Node.Val->getOpcode() < PPCISD::FIRST_NUMBER) ||
175 CodeGenMap.count(Node)) continue;
177 for (SDNode::use_iterator UI = Node.Val->use_begin(),
178 E = Node.Val->use_end(); UI != E; ++UI) {
179 // Scan the values. If this use has a value that is a token chain, add it
182 for (unsigned i = 0, e = User->getNumValues(); i != e; ++i)
183 if (User->getValueType(i) == MVT::Other) {
184 Worklist.push_back(SDOperand(User, i));
189 // Finally, legalize this node.
194 // Select target instructions for the DAG.
195 DAG.setRoot(SelectRoot(DAG.getRoot()));
197 DAG.RemoveDeadNodes();
199 // Emit machine code to BB.
200 ScheduleAndEmitDAG(DAG);
202 // Check to see if this function uses vector registers, which means we have to
203 // save and restore the VRSAVE register and update it with the regs we use.
205 // In this case, there will be virtual registers of vector type type created
206 // by the scheduler. Detect them now.
207 SSARegMap *RegMap = DAG.getMachineFunction().getSSARegMap();
208 bool HasVectorVReg = false;
209 for (unsigned i = MRegisterInfo::FirstVirtualRegister,
210 e = RegMap->getLastVirtReg(); i != e; ++i)
211 if (RegMap->getRegClass(i) == &PPC::VRRCRegClass) {
212 HasVectorVReg = true;
216 // If we have a vector register, we want to emit code into the entry and exit
217 // blocks to save and restore the VRSAVE register. We do this here (instead
218 // of marking all vector instructions as clobbering VRSAVE) for two reasons:
220 // 1. This (trivially) reduces the load on the register allocator, by not
221 // having to represent the live range of the VRSAVE register.
222 // 2. This (more significantly) allows us to create a temporary virtual
223 // register to hold the saved VRSAVE value, allowing this temporary to be
224 // register allocated, instead of forcing it to be spilled to the stack.
226 // Create two vregs - one to hold the VRSAVE register that is live-in to the
227 // function and one for the value after having bits or'd into it.
228 unsigned InVRSAVE = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
229 unsigned UpdatedVRSAVE = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
231 MachineFunction &MF = DAG.getMachineFunction();
232 MachineBasicBlock &EntryBB = *MF.begin();
233 // Emit the following code into the entry block:
234 // InVRSAVE = MFVRSAVE
235 // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE
236 // MTVRSAVE UpdatedVRSAVE
237 MachineBasicBlock::iterator IP = EntryBB.begin(); // Insert Point
238 BuildMI(EntryBB, IP, PPC::MFVRSAVE, 0, InVRSAVE);
239 BuildMI(EntryBB, IP, PPC::UPDATE_VRSAVE, 1, UpdatedVRSAVE).addReg(InVRSAVE);
240 BuildMI(EntryBB, IP, PPC::MTVRSAVE, 1).addReg(UpdatedVRSAVE);
242 // Find all return blocks, outputting a restore in each epilog.
243 const TargetInstrInfo &TII = *DAG.getTarget().getInstrInfo();
244 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); BB != E; ++BB)
245 if (!BB->empty() && TII.isReturn(BB->back().getOpcode())) {
246 IP = BB->end(); --IP;
248 // Skip over all terminator instructions, which are part of the return
250 MachineBasicBlock::iterator I2 = IP;
251 while (I2 != BB->begin() && TII.isTerminatorInstr((--I2)->getOpcode()))
254 // Emit: MTVRSAVE InVRSave
255 BuildMI(*BB, IP, PPC::MTVRSAVE, 1).addReg(InVRSAVE);
260 /// getGlobalBaseReg - Output the instructions required to put the
261 /// base address to use for accessing globals into a register.
263 SDOperand PPCDAGToDAGISel::getGlobalBaseReg() {
264 if (!GlobalBaseReg) {
265 // Insert the set of GlobalBaseReg into the first MBB of the function
266 MachineBasicBlock &FirstMBB = BB->getParent()->front();
267 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
268 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
269 // FIXME: when we get to LP64, we will need to create the appropriate
270 // type of register here.
271 GlobalBaseReg = RegMap->createVirtualRegister(PPC::GPRCRegisterClass);
272 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
273 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg);
275 return CurDAG->getRegister(GlobalBaseReg, MVT::i32);
279 // isIntImmediate - This method tests to see if a constant operand.
280 // If so Imm will receive the 32 bit value.
281 static bool isIntImmediate(SDNode *N, unsigned& Imm) {
282 if (N->getOpcode() == ISD::Constant) {
283 Imm = cast<ConstantSDNode>(N)->getValue();
289 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
290 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
291 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
292 // not, since all 1s are not contiguous.
293 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
294 if (isShiftedMask_32(Val)) {
295 // look for the first non-zero bit
296 MB = CountLeadingZeros_32(Val);
297 // look for the first zero bit after the run of ones
298 ME = CountLeadingZeros_32((Val - 1) ^ Val);
301 Val = ~Val; // invert mask
302 if (isShiftedMask_32(Val)) {
303 // effectively look for the first zero bit
304 ME = CountLeadingZeros_32(Val) - 1;
305 // effectively look for the first one bit after the run of zeros
306 MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
314 // isRotateAndMask - Returns true if Mask and Shift can be folded into a rotate
315 // and mask opcode and mask operation.
316 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool IsShiftMask,
317 unsigned &SH, unsigned &MB, unsigned &ME) {
318 // Don't even go down this path for i64, since different logic will be
319 // necessary for rldicl/rldicr/rldimi.
320 if (N->getValueType(0) != MVT::i32)
324 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
325 unsigned Opcode = N->getOpcode();
326 if (N->getNumOperands() != 2 ||
327 !isIntImmediate(N->getOperand(1).Val, Shift) || (Shift > 31))
330 if (Opcode == ISD::SHL) {
331 // apply shift left to mask if it comes first
332 if (IsShiftMask) Mask = Mask << Shift;
333 // determine which bits are made indeterminant by shift
334 Indeterminant = ~(0xFFFFFFFFu << Shift);
335 } else if (Opcode == ISD::SRL) {
336 // apply shift right to mask if it comes first
337 if (IsShiftMask) Mask = Mask >> Shift;
338 // determine which bits are made indeterminant by shift
339 Indeterminant = ~(0xFFFFFFFFu >> Shift);
340 // adjust for the left rotate
346 // if the mask doesn't intersect any Indeterminant bits
347 if (Mask && !(Mask & Indeterminant)) {
349 // make sure the mask is still a mask (wrap arounds may not be)
350 return isRunOfOnes(Mask, MB, ME);
355 // isOpcWithIntImmediate - This method tests to see if the node is a specific
356 // opcode and that it has a immediate integer right operand.
357 // If so Imm will receive the 32 bit value.
358 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
359 return N->getOpcode() == Opc && isIntImmediate(N->getOperand(1).Val, Imm);
362 // isIntImmediate - This method tests to see if a constant operand.
363 // If so Imm will receive the 32 bit value.
364 static bool isIntImmediate(SDOperand N, unsigned& Imm) {
365 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
366 Imm = (unsigned)CN->getSignExtended();
372 /// SelectBitfieldInsert - turn an or of two masked values into
373 /// the rotate left word immediate then mask insert (rlwimi) instruction.
374 /// Returns true on success, false if the caller still needs to select OR.
376 /// Patterns matched:
377 /// 1. or shl, and 5. or and, and
378 /// 2. or and, shl 6. or shl, shr
379 /// 3. or shr, and 7. or shr, shl
381 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
382 bool IsRotate = false;
383 unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, SH = 0;
386 SDOperand Op0 = N->getOperand(0);
387 SDOperand Op1 = N->getOperand(1);
389 unsigned Op0Opc = Op0.getOpcode();
390 unsigned Op1Opc = Op1.getOpcode();
392 // Verify that we have the correct opcodes
393 if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc)
395 if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc)
398 // Generate Mask value for Target
399 if (isIntImmediate(Op0.getOperand(1), Value)) {
401 case ISD::SHL: TgtMask <<= Value; break;
402 case ISD::SRL: TgtMask >>= Value; break;
403 case ISD::AND: TgtMask &= Value; break;
409 // Generate Mask value for Insert
410 if (!isIntImmediate(Op1.getOperand(1), Value))
417 if (Op0Opc == ISD::SRL) IsRotate = true;
423 if (Op0Opc == ISD::SHL) IsRotate = true;
430 // If both of the inputs are ANDs and one of them has a logical shift by
431 // constant as its input, make that AND the inserted value so that we can
432 // combine the shift into the rotate part of the rlwimi instruction
433 bool IsAndWithShiftOp = false;
434 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
435 if (Op1.getOperand(0).getOpcode() == ISD::SHL ||
436 Op1.getOperand(0).getOpcode() == ISD::SRL) {
437 if (isIntImmediate(Op1.getOperand(0).getOperand(1), Value)) {
438 SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
439 IsAndWithShiftOp = true;
441 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
442 Op0.getOperand(0).getOpcode() == ISD::SRL) {
443 if (isIntImmediate(Op0.getOperand(0).getOperand(1), Value)) {
445 std::swap(TgtMask, InsMask);
446 SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
447 IsAndWithShiftOp = true;
452 // Verify that the Target mask and Insert mask together form a full word mask
453 // and that the Insert mask is a run of set bits (which implies both are runs
454 // of set bits). Given that, Select the arguments and generate the rlwimi
457 if (((TgtMask & InsMask) == 0) && isRunOfOnes(InsMask, MB, ME)) {
458 bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF;
459 bool Op0IsAND = Op0Opc == ISD::AND;
460 // Check for rotlwi / rotrwi here, a special case of bitfield insert
461 // where both bitfield halves are sourced from the same value.
462 if (IsRotate && fullMask &&
463 N->getOperand(0).getOperand(0) == N->getOperand(1).getOperand(0)) {
465 Select(Tmp, N->getOperand(0).getOperand(0));
466 return CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, Tmp,
467 getI32Imm(SH), getI32Imm(0), getI32Imm(31));
469 SDOperand Tmp1, Tmp2;
470 Select(Tmp1, ((Op0IsAND && fullMask) ? Op0.getOperand(0) : Op0));
471 Select(Tmp2, (IsAndWithShiftOp ? Op1.getOperand(0).getOperand(0)
472 : Op1.getOperand(0)));
473 return CurDAG->getTargetNode(PPC::RLWIMI, MVT::i32, Tmp1, Tmp2,
474 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME));
479 /// SelectAddrImm - Returns true if the address N can be represented by
480 /// a base register plus a signed 16-bit displacement [r+imm].
481 bool PPCDAGToDAGISel::SelectAddrImm(SDOperand N, SDOperand &Disp,
483 // If this can be more profitably realized as r+r, fail.
484 if (SelectAddrIdx(N, Disp, Base))
487 if (N.getOpcode() == ISD::ADD) {
489 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm)) {
490 Disp = getI32Imm(imm & 0xFFFF);
491 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
492 Base = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
494 Base = N.getOperand(0);
496 return true; // [r+i]
497 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
498 // Match LOAD (ADD (X, Lo(G))).
499 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
500 && "Cannot handle constant offsets yet!");
501 Disp = N.getOperand(1).getOperand(0); // The global address.
502 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
503 Disp.getOpcode() == ISD::TargetConstantPool);
504 Base = N.getOperand(0);
505 return true; // [&g+r]
507 } else if (N.getOpcode() == ISD::OR) {
509 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm)) {
510 // If this is an or of disjoint bitfields, we can codegen this as an add
511 // (for better address arithmetic) if the LHS and RHS of the OR are
512 // provably disjoint.
513 uint64_t LHSKnownZero, LHSKnownOne;
514 PPCLowering.ComputeMaskedBits(N.getOperand(0), ~0U,
515 LHSKnownZero, LHSKnownOne);
516 if ((LHSKnownZero|~imm) == ~0U) {
517 // If all of the bits are known zero on the LHS or RHS, the add won't
519 Base = N.getOperand(0);
520 Disp = getI32Imm(imm & 0xFFFF);
526 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
527 Base = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
530 return true; // [r+0]
533 /// SelectAddrIdx - Given the specified addressed, check to see if it can be
534 /// represented as an indexed [r+r] operation. Returns false if it can
535 /// be represented by [r+imm], which are preferred.
536 bool PPCDAGToDAGISel::SelectAddrIdx(SDOperand N, SDOperand &Base,
539 if (N.getOpcode() == ISD::ADD) {
540 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm))
542 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
545 Base = N.getOperand(0);
546 Index = N.getOperand(1);
548 } else if (N.getOpcode() == ISD::OR) {
549 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm))
550 return false; // r+i can fold it if we can.
552 // If this is an or of disjoint bitfields, we can codegen this as an add
553 // (for better address arithmetic) if the LHS and RHS of the OR are provably
555 uint64_t LHSKnownZero, LHSKnownOne;
556 uint64_t RHSKnownZero, RHSKnownOne;
557 PPCLowering.ComputeMaskedBits(N.getOperand(0), ~0U,
558 LHSKnownZero, LHSKnownOne);
561 PPCLowering.ComputeMaskedBits(N.getOperand(1), ~0U,
562 RHSKnownZero, RHSKnownOne);
563 // If all of the bits are known zero on the LHS or RHS, the add won't
565 if ((LHSKnownZero | RHSKnownZero) == ~0U) {
566 Base = N.getOperand(0);
567 Index = N.getOperand(1);
576 /// SelectAddrIdxOnly - Given the specified addressed, force it to be
577 /// represented as an indexed [r+r] operation.
578 bool PPCDAGToDAGISel::SelectAddrIdxOnly(SDOperand N, SDOperand &Base,
580 // Check to see if we can easily represent this as an [r+r] address. This
581 // will fail if it thinks that the address is more profitably represented as
582 // reg+imm, e.g. where imm = 0.
583 if (!SelectAddrIdx(N, Base, Index)) {
584 // Nope, do it the hard way.
585 Base = CurDAG->getRegister(PPC::R0, MVT::i32);
591 /// SelectCC - Select a comparison of the specified values with the specified
592 /// condition code, returning the CR# of the expression.
593 SDOperand PPCDAGToDAGISel::SelectCC(SDOperand LHS, SDOperand RHS,
595 // Always select the LHS.
598 // Use U to determine whether the SETCC immediate range is signed or not.
599 if (MVT::isInteger(LHS.getValueType())) {
600 bool U = ISD::isUnsignedIntSetCC(CC);
602 if (isIntImmediate(RHS, Imm) &&
603 ((U && isUInt16(Imm)) || (!U && isInt16(Imm))))
604 return SDOperand(CurDAG->getTargetNode(U ? PPC::CMPLWI : PPC::CMPWI,
605 MVT::i32, LHS, getI32Imm(Imm & 0xFFFF)), 0);
607 return SDOperand(CurDAG->getTargetNode(U ? PPC::CMPLW : PPC::CMPW, MVT::i32,
609 } else if (LHS.getValueType() == MVT::f32) {
611 return SDOperand(CurDAG->getTargetNode(PPC::FCMPUS, MVT::i32, LHS, RHS), 0);
614 return SDOperand(CurDAG->getTargetNode(PPC::FCMPUD, MVT::i32, LHS, RHS), 0);
618 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
620 static unsigned getBCCForSetCC(ISD::CondCode CC) {
622 default: assert(0 && "Unknown condition!"); abort();
623 case ISD::SETOEQ: // FIXME: This is incorrect see PR642.
624 case ISD::SETEQ: return PPC::BEQ;
625 case ISD::SETONE: // FIXME: This is incorrect see PR642.
626 case ISD::SETNE: return PPC::BNE;
627 case ISD::SETOLT: // FIXME: This is incorrect see PR642.
629 case ISD::SETLT: return PPC::BLT;
630 case ISD::SETOLE: // FIXME: This is incorrect see PR642.
632 case ISD::SETLE: return PPC::BLE;
633 case ISD::SETOGT: // FIXME: This is incorrect see PR642.
635 case ISD::SETGT: return PPC::BGT;
636 case ISD::SETOGE: // FIXME: This is incorrect see PR642.
638 case ISD::SETGE: return PPC::BGE;
640 case ISD::SETO: return PPC::BUN;
641 case ISD::SETUO: return PPC::BNU;
646 /// getCRIdxForSetCC - Return the index of the condition register field
647 /// associated with the SetCC condition, and whether or not the field is
648 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
649 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool& Inv) {
651 default: assert(0 && "Unknown condition!"); abort();
652 case ISD::SETOLT: // FIXME: This is incorrect see PR642.
654 case ISD::SETLT: Inv = false; return 0;
655 case ISD::SETOGE: // FIXME: This is incorrect see PR642.
657 case ISD::SETGE: Inv = true; return 0;
658 case ISD::SETOGT: // FIXME: This is incorrect see PR642.
660 case ISD::SETGT: Inv = false; return 1;
661 case ISD::SETOLE: // FIXME: This is incorrect see PR642.
663 case ISD::SETLE: Inv = true; return 1;
664 case ISD::SETOEQ: // FIXME: This is incorrect see PR642.
665 case ISD::SETEQ: Inv = false; return 2;
666 case ISD::SETONE: // FIXME: This is incorrect see PR642.
667 case ISD::SETNE: Inv = true; return 2;
668 case ISD::SETO: Inv = true; return 3;
669 case ISD::SETUO: Inv = false; return 3;
674 SDOperand PPCDAGToDAGISel::SelectSETCC(SDOperand Op) {
677 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
678 if (isIntImmediate(N->getOperand(1), Imm)) {
679 // We can codegen setcc op, imm very efficiently compared to a brcond.
680 // Check for those cases here.
684 Select(Op, N->getOperand(0));
688 Op = SDOperand(CurDAG->getTargetNode(PPC::CNTLZW, MVT::i32, Op), 0);
689 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(27),
690 getI32Imm(5), getI32Imm(31));
693 SDOperand(CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
694 Op, getI32Imm(~0U)), 0);
695 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
699 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(1),
700 getI32Imm(31), getI32Imm(31));
703 SDOperand(CurDAG->getTargetNode(PPC::NEG, MVT::i32, Op), 0);
704 T = SDOperand(CurDAG->getTargetNode(PPC::ANDC, MVT::i32, T, Op), 0);
705 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, T, getI32Imm(1),
706 getI32Imm(31), getI32Imm(31));
709 } else if (Imm == ~0U) { // setcc op, -1
711 Select(Op, N->getOperand(0));
715 Op = SDOperand(CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
716 Op, getI32Imm(1)), 0);
717 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
718 SDOperand(CurDAG->getTargetNode(PPC::LI, MVT::i32,
722 Op = SDOperand(CurDAG->getTargetNode(PPC::NOR, MVT::i32, Op, Op), 0);
723 SDNode *AD = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
725 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDOperand(AD, 0), Op,
729 SDOperand AD = SDOperand(CurDAG->getTargetNode(PPC::ADDI, MVT::i32, Op,
731 SDOperand AN = SDOperand(CurDAG->getTargetNode(PPC::AND, MVT::i32, AD,
733 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, AN, getI32Imm(1),
734 getI32Imm(31), getI32Imm(31));
737 Op = SDOperand(CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, Op,
738 getI32Imm(1), getI32Imm(31),
740 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, getI32Imm(1));
746 unsigned Idx = getCRIdxForSetCC(CC, Inv);
747 SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
750 // Force the ccreg into CR7.
751 SDOperand CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
753 SDOperand InFlag(0, 0); // Null incoming flag value.
754 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), CR7Reg, CCReg,
757 if (TLI.getTargetMachine().getSubtarget<PPCSubtarget>().isGigaProcessor())
758 IntCR = SDOperand(CurDAG->getTargetNode(PPC::MFOCRF, MVT::i32, CR7Reg,
761 IntCR = SDOperand(CurDAG->getTargetNode(PPC::MFCR, MVT::i32, CCReg), 0);
764 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, IntCR,
765 getI32Imm((32-(3-Idx)) & 31),
766 getI32Imm(31), getI32Imm(31));
769 SDOperand(CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, IntCR,
770 getI32Imm((32-(3-Idx)) & 31),
771 getI32Imm(31),getI32Imm(31)), 0);
772 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
776 /// isCallCompatibleAddress - Return true if the specified 32-bit value is
777 /// representable in the immediate field of a Bx instruction.
778 static bool isCallCompatibleAddress(ConstantSDNode *C) {
779 int Addr = C->getValue();
780 if (Addr & 3) return false; // Low 2 bits are implicitly zero.
781 return (Addr << 6 >> 6) == Addr; // Top 6 bits have to be sext of immediate.
784 SDOperand PPCDAGToDAGISel::SelectCALL(SDOperand Op) {
787 Select(Chain, N->getOperand(0));
790 std::vector<SDOperand> CallOperands;
792 if (GlobalAddressSDNode *GASD =
793 dyn_cast<GlobalAddressSDNode>(N->getOperand(1))) {
794 CallOpcode = PPC::BL;
795 CallOperands.push_back(N->getOperand(1));
796 } else if (ExternalSymbolSDNode *ESSDN =
797 dyn_cast<ExternalSymbolSDNode>(N->getOperand(1))) {
798 CallOpcode = PPC::BL;
799 CallOperands.push_back(N->getOperand(1));
800 } else if (isa<ConstantSDNode>(N->getOperand(1)) &&
801 isCallCompatibleAddress(cast<ConstantSDNode>(N->getOperand(1)))) {
802 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(1));
803 CallOpcode = PPC::BLA;
804 CallOperands.push_back(getI32Imm((int)C->getValue() >> 2));
806 // Copy the callee address into the CTR register.
808 Select(Callee, N->getOperand(1));
809 Chain = SDOperand(CurDAG->getTargetNode(PPC::MTCTR, MVT::Other, Callee,
812 // Copy the callee address into R12 on darwin.
813 SDOperand R12 = CurDAG->getRegister(PPC::R12, MVT::i32);
814 Chain = CurDAG->getNode(ISD::CopyToReg, MVT::Other, Chain, R12, Callee);
816 CallOperands.push_back(R12);
817 CallOpcode = PPC::BCTRL;
820 unsigned GPR_idx = 0, FPR_idx = 0;
821 static const unsigned GPR[] = {
822 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
823 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
825 static const unsigned FPR[] = {
826 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
827 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
830 SDOperand InFlag; // Null incoming flag value.
832 for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i) {
833 unsigned DestReg = 0;
834 MVT::ValueType RegTy = N->getOperand(i).getValueType();
835 if (RegTy == MVT::i32) {
836 assert(GPR_idx < 8 && "Too many int args");
837 DestReg = GPR[GPR_idx++];
839 assert(MVT::isFloatingPoint(N->getOperand(i).getValueType()) &&
840 "Unpromoted integer arg?");
841 assert(FPR_idx < 13 && "Too many fp args");
842 DestReg = FPR[FPR_idx++];
845 if (N->getOperand(i).getOpcode() != ISD::UNDEF) {
847 Select(Val, N->getOperand(i));
848 Chain = CurDAG->getCopyToReg(Chain, DestReg, Val, InFlag);
849 InFlag = Chain.getValue(1);
850 CallOperands.push_back(CurDAG->getRegister(DestReg, RegTy));
854 // Finally, once everything is in registers to pass to the call, emit the
857 CallOperands.push_back(InFlag); // Strong dep on register copies.
859 CallOperands.push_back(Chain); // Weak dep on whatever occurs before
860 Chain = SDOperand(CurDAG->getTargetNode(CallOpcode, MVT::Other, MVT::Flag,
863 std::vector<SDOperand> CallResults;
865 // If the call has results, copy the values out of the ret val registers.
866 switch (N->getValueType(0)) {
867 default: assert(0 && "Unexpected ret value!");
868 case MVT::Other: break;
870 if (N->getValueType(1) == MVT::i32) {
871 Chain = CurDAG->getCopyFromReg(Chain, PPC::R4, MVT::i32,
872 Chain.getValue(1)).getValue(1);
873 CallResults.push_back(Chain.getValue(0));
874 Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
875 Chain.getValue(2)).getValue(1);
876 CallResults.push_back(Chain.getValue(0));
878 Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
879 Chain.getValue(1)).getValue(1);
880 CallResults.push_back(Chain.getValue(0));
885 Chain = CurDAG->getCopyFromReg(Chain, PPC::F1, N->getValueType(0),
886 Chain.getValue(1)).getValue(1);
887 CallResults.push_back(Chain.getValue(0));
891 CallResults.push_back(Chain);
892 for (unsigned i = 0, e = CallResults.size(); i != e; ++i)
893 CodeGenMap[Op.getValue(i)] = CallResults[i];
894 return CallResults[Op.ResNo];
897 // Select - Convert the specified operand from a target-independent to a
898 // target-specific node if it hasn't already been changed.
899 void PPCDAGToDAGISel::Select(SDOperand &Result, SDOperand Op) {
901 if (N->getOpcode() >= ISD::BUILTIN_OP_END &&
902 N->getOpcode() < PPCISD::FIRST_NUMBER) {
904 return; // Already selected.
907 // If this has already been converted, use it.
908 std::map<SDOperand, SDOperand>::iterator CGMI = CodeGenMap.find(Op);
909 if (CGMI != CodeGenMap.end()) {
910 Result = CGMI->second;
914 switch (N->getOpcode()) {
917 Result = SelectSETCC(Op);
920 Result = SelectCALL(Op);
922 case PPCISD::GlobalBaseReg:
923 Result = getGlobalBaseReg();
926 case ISD::FrameIndex: {
927 int FI = cast<FrameIndexSDNode>(N)->getIndex();
928 if (N->hasOneUse()) {
929 Result = CurDAG->SelectNodeTo(N, PPC::ADDI, MVT::i32,
930 CurDAG->getTargetFrameIndex(FI, MVT::i32),
934 Result = CodeGenMap[Op] =
935 SDOperand(CurDAG->getTargetNode(PPC::ADDI, MVT::i32,
936 CurDAG->getTargetFrameIndex(FI, MVT::i32),
941 // FIXME: since this depends on the setting of the carry flag from the srawi
942 // we should really be making notes about that for the scheduler.
943 // FIXME: It sure would be nice if we could cheaply recognize the
944 // srl/add/sra pattern the dag combiner will generate for this as
945 // sra/addze rather than having to handle sdiv ourselves. oh well.
947 if (isIntImmediate(N->getOperand(1), Imm)) {
949 Select(N0, N->getOperand(0));
950 if ((signed)Imm > 0 && isPowerOf2_32(Imm)) {
952 CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
953 N0, getI32Imm(Log2_32(Imm)));
954 Result = CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
955 SDOperand(Op, 0), SDOperand(Op, 1));
956 } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) {
958 CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
959 N0, getI32Imm(Log2_32(-Imm)));
961 SDOperand(CurDAG->getTargetNode(PPC::ADDZE, MVT::i32,
962 SDOperand(Op, 0), SDOperand(Op, 1)),
964 Result = CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT);
969 // Other cases are autogenerated.
974 // If this is an and of a value rotated between 0 and 31 bits and then and'd
975 // with a mask, emit rlwinm
976 if (isIntImmediate(N->getOperand(1), Imm) && (isShiftedMask_32(Imm) ||
977 isShiftedMask_32(~Imm))) {
980 if (isRotateAndMask(N->getOperand(0).Val, Imm, false, SH, MB, ME)) {
981 Select(Val, N->getOperand(0).getOperand(0));
982 } else if (Imm == 0) {
983 // AND X, 0 -> 0, not "rlwinm 32".
984 Select(Result, N->getOperand(1));
987 Select(Val, N->getOperand(0));
988 isRunOfOnes(Imm, MB, ME);
991 Result = CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Val,
992 getI32Imm(SH), getI32Imm(MB),
996 // ISD::OR doesn't get all the bitfield insertion fun.
997 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert
998 if (isIntImmediate(N->getOperand(1), Imm) &&
999 N->getOperand(0).getOpcode() == ISD::OR &&
1000 isIntImmediate(N->getOperand(0).getOperand(1), Imm2)) {
1003 if (isRunOfOnes(Imm, MB, ME)) {
1004 SDOperand Tmp1, Tmp2;
1005 Select(Tmp1, N->getOperand(0).getOperand(0));
1006 Select(Tmp2, N->getOperand(0).getOperand(1));
1007 Result = SDOperand(CurDAG->getTargetNode(PPC::RLWIMI, MVT::i32,
1009 getI32Imm(0), getI32Imm(MB),
1015 // Other cases are autogenerated.
1019 if (SDNode *I = SelectBitfieldInsert(N)) {
1020 Result = CodeGenMap[Op] = SDOperand(I, 0);
1024 // Other cases are autogenerated.
1027 unsigned Imm, SH, MB, ME;
1028 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
1029 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
1031 Select(Val, N->getOperand(0).getOperand(0));
1032 Result = CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
1033 Val, getI32Imm(SH), getI32Imm(MB),
1038 // Other cases are autogenerated.
1042 unsigned Imm, SH, MB, ME;
1043 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
1044 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
1046 Select(Val, N->getOperand(0).getOperand(0));
1047 Result = CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
1048 Val, getI32Imm(SH & 0x1F), getI32Imm(MB),
1053 // Other cases are autogenerated.
1056 case ISD::SELECT_CC: {
1057 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1059 // handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
1060 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1061 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
1062 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
1063 if (N1C->isNullValue() && N3C->isNullValue() &&
1064 N2C->getValue() == 1ULL && CC == ISD::SETNE) {
1066 Select(LHS, N->getOperand(0));
1068 CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
1069 LHS, getI32Imm(~0U));
1070 Result = CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32,
1071 SDOperand(Tmp, 0), LHS,
1076 SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
1077 unsigned BROpc = getBCCForSetCC(CC);
1079 bool isFP = MVT::isFloatingPoint(N->getValueType(0));
1080 unsigned SelectCCOp;
1081 if (MVT::isInteger(N->getValueType(0)))
1082 SelectCCOp = PPC::SELECT_CC_Int;
1083 else if (N->getValueType(0) == MVT::f32)
1084 SelectCCOp = PPC::SELECT_CC_F4;
1086 SelectCCOp = PPC::SELECT_CC_F8;
1088 Select(N2, N->getOperand(2));
1089 Select(N3, N->getOperand(3));
1090 Result = CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), CCReg,
1091 N2, N3, getI32Imm(BROpc));
1095 case ISD::BRTWOWAY_CC: {
1097 Select(Chain, N->getOperand(0));
1098 MachineBasicBlock *Dest =
1099 cast<BasicBlockSDNode>(N->getOperand(4))->getBasicBlock();
1100 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
1101 SDOperand CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC);
1103 // If this is a two way branch, then grab the fallthrough basic block
1104 // argument and build a PowerPC branch pseudo-op, suitable for long branch
1105 // conversion if necessary by the branch selection pass. Otherwise, emit a
1106 // standard conditional branch.
1107 if (N->getOpcode() == ISD::BRTWOWAY_CC) {
1108 SDOperand CondTrueBlock = N->getOperand(4);
1109 SDOperand CondFalseBlock = N->getOperand(5);
1110 unsigned Opc = getBCCForSetCC(CC);
1112 SDOperand(CurDAG->getTargetNode(PPC::COND_BRANCH, MVT::Other,
1113 CondCode, getI32Imm(Opc),
1114 CondTrueBlock, CondFalseBlock,
1116 Result = CurDAG->SelectNodeTo(N, PPC::B, MVT::Other, CondFalseBlock, CB);
1118 // Iterate to the next basic block
1119 ilist<MachineBasicBlock>::iterator It = BB;
1122 // If the fallthrough path is off the end of the function, which would be
1123 // undefined behavior, set it to be the same as the current block because
1124 // we have nothing better to set it to, and leaving it alone will cause
1125 // the PowerPC Branch Selection pass to crash.
1126 if (It == BB->getParent()->end()) It = Dest;
1127 Result = CurDAG->SelectNodeTo(N, PPC::COND_BRANCH, MVT::Other, CondCode,
1128 getI32Imm(getBCCForSetCC(CC)),
1129 N->getOperand(4), CurDAG->getBasicBlock(It),
1136 SelectCode(Result, Op);
1140 /// createPPCISelDag - This pass converts a legalized DAG into a
1141 /// PowerPC-specific DAG, ready for instruction scheduling.
1143 FunctionPass *llvm::createPPCISelDag(TargetMachine &TM) {
1144 return new PPCDAGToDAGISel(TM);