1 //===-- PPCISelDAGToDAG.cpp - PPC --pattern matching inst selector --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "ppc-codegen"
17 #include "MCTargetDesc/PPCPredicates.h"
18 #include "PPCTargetMachine.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/MachineRegisterInfo.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/CodeGen/SelectionDAGISel.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/Function.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/GlobalValue.h"
28 #include "llvm/IR/GlobalVariable.h"
29 #include "llvm/IR/Intrinsics.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetOptions.h"
38 void initializePPCDAGToDAGISelPass(PassRegistry&);
42 //===--------------------------------------------------------------------===//
43 /// PPCDAGToDAGISel - PPC specific code to select PPC machine
44 /// instructions for SelectionDAG operations.
46 class PPCDAGToDAGISel : public SelectionDAGISel {
47 const PPCTargetMachine &TM;
48 const PPCTargetLowering &PPCLowering;
49 const PPCSubtarget &PPCSubTarget;
50 unsigned GlobalBaseReg;
52 explicit PPCDAGToDAGISel(PPCTargetMachine &tm)
53 : SelectionDAGISel(tm), TM(tm),
54 PPCLowering(*TM.getTargetLowering()),
55 PPCSubTarget(*TM.getSubtargetImpl()) {
56 initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry());
59 virtual bool runOnMachineFunction(MachineFunction &MF) {
60 // Make sure we re-emit a set of the global base reg if necessary
62 SelectionDAGISel::runOnMachineFunction(MF);
64 if (!PPCSubTarget.isSVR4ABI())
70 /// getI32Imm - Return a target constant with the specified value, of type
72 inline SDValue getI32Imm(unsigned Imm) {
73 return CurDAG->getTargetConstant(Imm, MVT::i32);
76 /// getI64Imm - Return a target constant with the specified value, of type
78 inline SDValue getI64Imm(uint64_t Imm) {
79 return CurDAG->getTargetConstant(Imm, MVT::i64);
82 /// getSmallIPtrImm - Return a target constant of pointer type.
83 inline SDValue getSmallIPtrImm(unsigned Imm) {
84 return CurDAG->getTargetConstant(Imm, PPCLowering.getPointerTy());
87 /// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s
88 /// with any number of 0s on either side. The 1s are allowed to wrap from
89 /// LSB to MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs.
90 /// 0x0F0F0000 is not, since all 1s are not contiguous.
91 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME);
94 /// isRotateAndMask - Returns true if Mask and Shift can be folded into a
95 /// rotate and mask opcode and mask operation.
96 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask,
97 unsigned &SH, unsigned &MB, unsigned &ME);
99 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
100 /// base register. Return the virtual register that holds this value.
101 SDNode *getGlobalBaseReg();
103 // Select - Convert the specified operand from a target-independent to a
104 // target-specific node if it hasn't already been changed.
105 SDNode *Select(SDNode *N);
107 SDNode *SelectBitfieldInsert(SDNode *N);
109 /// SelectCC - Select a comparison of the specified values with the
110 /// specified condition code, returning the CR# of the expression.
111 SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, DebugLoc dl);
113 /// SelectAddrImm - Returns true if the address N can be represented by
114 /// a base register plus a signed 16-bit displacement [r+imm].
115 bool SelectAddrImm(SDValue N, SDValue &Disp,
117 return PPCLowering.SelectAddressRegImm(N, Disp, Base, *CurDAG);
120 /// SelectAddrImmOffs - Return true if the operand is valid for a preinc
121 /// immediate field. Because preinc imms have already been validated, just
123 bool SelectAddrImmOffs(SDValue N, SDValue &Out) const {
124 if (isa<ConstantSDNode>(N) || N.getOpcode() == PPCISD::Lo ||
125 N.getOpcode() == ISD::TargetGlobalAddress) {
133 /// SelectAddrIdxOffs - Return true if the operand is valid for a preinc
134 /// index field. Because preinc imms have already been validated, just
136 bool SelectAddrIdxOffs(SDValue N, SDValue &Out) const {
137 if (isa<ConstantSDNode>(N) || N.getOpcode() == PPCISD::Lo ||
138 N.getOpcode() == ISD::TargetGlobalAddress)
145 /// SelectAddrIdx - Given the specified addressed, check to see if it can be
146 /// represented as an indexed [r+r] operation. Returns false if it can
147 /// be represented by [r+imm], which are preferred.
148 bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) {
149 return PPCLowering.SelectAddressRegReg(N, Base, Index, *CurDAG);
152 /// SelectAddrIdxOnly - Given the specified addressed, force it to be
153 /// represented as an indexed [r+r] operation.
154 bool SelectAddrIdxOnly(SDValue N, SDValue &Base, SDValue &Index) {
155 return PPCLowering.SelectAddressRegRegOnly(N, Base, Index, *CurDAG);
158 /// SelectAddrImmShift - Returns true if the address N can be represented by
159 /// a base register plus a signed 14-bit displacement [r+imm*4]. Suitable
160 /// for use by STD and friends.
161 bool SelectAddrImmShift(SDValue N, SDValue &Disp, SDValue &Base) {
162 return PPCLowering.SelectAddressRegImmShift(N, Disp, Base, *CurDAG);
165 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
166 /// inline asm expressions. It is always correct to compute the value into
167 /// a register. The case of adding a (possibly relocatable) constant to a
168 /// register can be improved, but it is wrong to substitute Reg+Reg for
169 /// Reg in an asm, because the load or store opcode would have to change.
170 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
172 std::vector<SDValue> &OutOps) {
173 OutOps.push_back(Op);
177 void InsertVRSaveCode(MachineFunction &MF);
179 virtual const char *getPassName() const {
180 return "PowerPC DAG->DAG Pattern Instruction Selection";
183 // Include the pieces autogenerated from the target description.
184 #include "PPCGenDAGISel.inc"
187 SDNode *SelectSETCC(SDNode *N);
191 /// InsertVRSaveCode - Once the entire function has been instruction selected,
192 /// all virtual registers are created and all machine instructions are built,
193 /// check to see if we need to save/restore VRSAVE. If so, do it.
194 void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) {
195 // Check to see if this function uses vector registers, which means we have to
196 // save and restore the VRSAVE register and update it with the regs we use.
198 // In this case, there will be virtual registers of vector type created
199 // by the scheduler. Detect them now.
200 bool HasVectorVReg = false;
201 for (unsigned i = 0, e = RegInfo->getNumVirtRegs(); i != e; ++i) {
202 unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
203 if (RegInfo->getRegClass(Reg) == &PPC::VRRCRegClass) {
204 HasVectorVReg = true;
208 if (!HasVectorVReg) return; // nothing to do.
210 // If we have a vector register, we want to emit code into the entry and exit
211 // blocks to save and restore the VRSAVE register. We do this here (instead
212 // of marking all vector instructions as clobbering VRSAVE) for two reasons:
214 // 1. This (trivially) reduces the load on the register allocator, by not
215 // having to represent the live range of the VRSAVE register.
216 // 2. This (more significantly) allows us to create a temporary virtual
217 // register to hold the saved VRSAVE value, allowing this temporary to be
218 // register allocated, instead of forcing it to be spilled to the stack.
220 // Create two vregs - one to hold the VRSAVE register that is live-in to the
221 // function and one for the value after having bits or'd into it.
222 unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
223 unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
225 const TargetInstrInfo &TII = *TM.getInstrInfo();
226 MachineBasicBlock &EntryBB = *Fn.begin();
228 // Emit the following code into the entry block:
229 // InVRSAVE = MFVRSAVE
230 // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE
231 // MTVRSAVE UpdatedVRSAVE
232 MachineBasicBlock::iterator IP = EntryBB.begin(); // Insert Point
233 BuildMI(EntryBB, IP, dl, TII.get(PPC::MFVRSAVE), InVRSAVE);
234 BuildMI(EntryBB, IP, dl, TII.get(PPC::UPDATE_VRSAVE),
235 UpdatedVRSAVE).addReg(InVRSAVE);
236 BuildMI(EntryBB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(UpdatedVRSAVE);
238 // Find all return blocks, outputting a restore in each epilog.
239 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
240 if (!BB->empty() && BB->back().isReturn()) {
241 IP = BB->end(); --IP;
243 // Skip over all terminator instructions, which are part of the return
245 MachineBasicBlock::iterator I2 = IP;
246 while (I2 != BB->begin() && (--I2)->isTerminator())
249 // Emit: MTVRSAVE InVRSave
250 BuildMI(*BB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(InVRSAVE);
256 /// getGlobalBaseReg - Output the instructions required to put the
257 /// base address to use for accessing globals into a register.
259 SDNode *PPCDAGToDAGISel::getGlobalBaseReg() {
260 if (!GlobalBaseReg) {
261 const TargetInstrInfo &TII = *TM.getInstrInfo();
262 // Insert the set of GlobalBaseReg into the first MBB of the function
263 MachineBasicBlock &FirstMBB = MF->front();
264 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
267 if (PPCLowering.getPointerTy() == MVT::i32) {
268 GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
269 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
270 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
272 GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RCRegClass);
273 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8));
274 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg);
277 return CurDAG->getRegister(GlobalBaseReg,
278 PPCLowering.getPointerTy()).getNode();
281 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
282 /// or 64-bit immediate, and if the value can be accurately represented as a
283 /// sign extension from a 16-bit value. If so, this returns true and the
285 static bool isIntS16Immediate(SDNode *N, short &Imm) {
286 if (N->getOpcode() != ISD::Constant)
289 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
290 if (N->getValueType(0) == MVT::i32)
291 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
293 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
296 static bool isIntS16Immediate(SDValue Op, short &Imm) {
297 return isIntS16Immediate(Op.getNode(), Imm);
301 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
302 /// operand. If so Imm will receive the 32-bit value.
303 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
304 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
305 Imm = cast<ConstantSDNode>(N)->getZExtValue();
311 /// isInt64Immediate - This method tests to see if the node is a 64-bit constant
312 /// operand. If so Imm will receive the 64-bit value.
313 static bool isInt64Immediate(SDNode *N, uint64_t &Imm) {
314 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i64) {
315 Imm = cast<ConstantSDNode>(N)->getZExtValue();
321 // isInt32Immediate - This method tests to see if a constant operand.
322 // If so Imm will receive the 32 bit value.
323 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
324 return isInt32Immediate(N.getNode(), Imm);
328 // isOpcWithIntImmediate - This method tests to see if the node is a specific
329 // opcode and that it has a immediate integer right operand.
330 // If so Imm will receive the 32 bit value.
331 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
332 return N->getOpcode() == Opc
333 && isInt32Immediate(N->getOperand(1).getNode(), Imm);
336 bool PPCDAGToDAGISel::isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
337 if (isShiftedMask_32(Val)) {
338 // look for the first non-zero bit
339 MB = CountLeadingZeros_32(Val);
340 // look for the first zero bit after the run of ones
341 ME = CountLeadingZeros_32((Val - 1) ^ Val);
344 Val = ~Val; // invert mask
345 if (isShiftedMask_32(Val)) {
346 // effectively look for the first zero bit
347 ME = CountLeadingZeros_32(Val) - 1;
348 // effectively look for the first one bit after the run of zeros
349 MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
357 bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask,
358 bool isShiftMask, unsigned &SH,
359 unsigned &MB, unsigned &ME) {
360 // Don't even go down this path for i64, since different logic will be
361 // necessary for rldicl/rldicr/rldimi.
362 if (N->getValueType(0) != MVT::i32)
366 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
367 unsigned Opcode = N->getOpcode();
368 if (N->getNumOperands() != 2 ||
369 !isInt32Immediate(N->getOperand(1).getNode(), Shift) || (Shift > 31))
372 if (Opcode == ISD::SHL) {
373 // apply shift left to mask if it comes first
374 if (isShiftMask) Mask = Mask << Shift;
375 // determine which bits are made indeterminant by shift
376 Indeterminant = ~(0xFFFFFFFFu << Shift);
377 } else if (Opcode == ISD::SRL) {
378 // apply shift right to mask if it comes first
379 if (isShiftMask) Mask = Mask >> Shift;
380 // determine which bits are made indeterminant by shift
381 Indeterminant = ~(0xFFFFFFFFu >> Shift);
382 // adjust for the left rotate
384 } else if (Opcode == ISD::ROTL) {
390 // if the mask doesn't intersect any Indeterminant bits
391 if (Mask && !(Mask & Indeterminant)) {
393 // make sure the mask is still a mask (wrap arounds may not be)
394 return isRunOfOnes(Mask, MB, ME);
399 /// SelectBitfieldInsert - turn an or of two masked values into
400 /// the rotate left word immediate then mask insert (rlwimi) instruction.
401 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
402 SDValue Op0 = N->getOperand(0);
403 SDValue Op1 = N->getOperand(1);
404 DebugLoc dl = N->getDebugLoc();
406 APInt LKZ, LKO, RKZ, RKO;
407 CurDAG->ComputeMaskedBits(Op0, LKZ, LKO);
408 CurDAG->ComputeMaskedBits(Op1, RKZ, RKO);
410 unsigned TargetMask = LKZ.getZExtValue();
411 unsigned InsertMask = RKZ.getZExtValue();
413 if ((TargetMask | InsertMask) == 0xFFFFFFFF) {
414 unsigned Op0Opc = Op0.getOpcode();
415 unsigned Op1Opc = Op1.getOpcode();
416 unsigned Value, SH = 0;
417 TargetMask = ~TargetMask;
418 InsertMask = ~InsertMask;
420 // If the LHS has a foldable shift and the RHS does not, then swap it to the
421 // RHS so that we can fold the shift into the insert.
422 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
423 if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
424 Op0.getOperand(0).getOpcode() == ISD::SRL) {
425 if (Op1.getOperand(0).getOpcode() != ISD::SHL &&
426 Op1.getOperand(0).getOpcode() != ISD::SRL) {
428 std::swap(Op0Opc, Op1Opc);
429 std::swap(TargetMask, InsertMask);
432 } else if (Op0Opc == ISD::SHL || Op0Opc == ISD::SRL) {
433 if (Op1Opc == ISD::AND && Op1.getOperand(0).getOpcode() != ISD::SHL &&
434 Op1.getOperand(0).getOpcode() != ISD::SRL) {
436 std::swap(Op0Opc, Op1Opc);
437 std::swap(TargetMask, InsertMask);
442 if (InsertMask && isRunOfOnes(InsertMask, MB, ME)) {
445 if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) &&
446 isInt32Immediate(Op1.getOperand(1), Value)) {
447 Op1 = Op1.getOperand(0);
448 SH = (Op1Opc == ISD::SHL) ? Value : 32 - Value;
450 if (Op1Opc == ISD::AND) {
451 unsigned SHOpc = Op1.getOperand(0).getOpcode();
452 if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) &&
453 isInt32Immediate(Op1.getOperand(0).getOperand(1), Value)) {
454 Op1 = Op1.getOperand(0).getOperand(0);
455 SH = (SHOpc == ISD::SHL) ? Value : 32 - Value;
457 Op1 = Op1.getOperand(0);
462 SDValue Ops[] = { Op0, Op1, getI32Imm(SH), getI32Imm(MB),
464 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops, 5);
470 /// SelectCC - Select a comparison of the specified values with the specified
471 /// condition code, returning the CR# of the expression.
472 SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS,
473 ISD::CondCode CC, DebugLoc dl) {
474 // Always select the LHS.
477 if (LHS.getValueType() == MVT::i32) {
479 if (CC == ISD::SETEQ || CC == ISD::SETNE) {
480 if (isInt32Immediate(RHS, Imm)) {
481 // SETEQ/SETNE comparison with 16-bit immediate, fold it.
483 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
484 getI32Imm(Imm & 0xFFFF)), 0);
485 // If this is a 16-bit signed immediate, fold it.
486 if (isInt<16>((int)Imm))
487 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
488 getI32Imm(Imm & 0xFFFF)), 0);
490 // For non-equality comparisons, the default code would materialize the
491 // constant, then compare against it, like this:
495 // Since we are just comparing for equality, we can emit this instead:
496 // xoris r0,r3,0x1234
497 // cmplwi cr0,r0,0x5678
499 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS,
500 getI32Imm(Imm >> 16)), 0);
501 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor,
502 getI32Imm(Imm & 0xFFFF)), 0);
505 } else if (ISD::isUnsignedIntSetCC(CC)) {
506 if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm))
507 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
508 getI32Imm(Imm & 0xFFFF)), 0);
512 if (isIntS16Immediate(RHS, SImm))
513 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
514 getI32Imm((int)SImm & 0xFFFF)),
518 } else if (LHS.getValueType() == MVT::i64) {
520 if (CC == ISD::SETEQ || CC == ISD::SETNE) {
521 if (isInt64Immediate(RHS.getNode(), Imm)) {
522 // SETEQ/SETNE comparison with 16-bit immediate, fold it.
524 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
525 getI32Imm(Imm & 0xFFFF)), 0);
526 // If this is a 16-bit signed immediate, fold it.
528 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
529 getI32Imm(Imm & 0xFFFF)), 0);
531 // For non-equality comparisons, the default code would materialize the
532 // constant, then compare against it, like this:
536 // Since we are just comparing for equality, we can emit this instead:
537 // xoris r0,r3,0x1234
538 // cmpldi cr0,r0,0x5678
540 if (isUInt<32>(Imm)) {
541 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS,
542 getI64Imm(Imm >> 16)), 0);
543 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor,
544 getI64Imm(Imm & 0xFFFF)), 0);
548 } else if (ISD::isUnsignedIntSetCC(CC)) {
549 if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm))
550 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
551 getI64Imm(Imm & 0xFFFF)), 0);
555 if (isIntS16Immediate(RHS, SImm))
556 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
557 getI64Imm(SImm & 0xFFFF)),
561 } else if (LHS.getValueType() == MVT::f32) {
564 assert(LHS.getValueType() == MVT::f64 && "Unknown vt!");
567 return SDValue(CurDAG->getMachineNode(Opc, dl, MVT::i32, LHS, RHS), 0);
570 static PPC::Predicate getPredicateForSetCC(ISD::CondCode CC) {
576 llvm_unreachable("Should be lowered by legalize!");
577 default: llvm_unreachable("Unknown condition!");
579 case ISD::SETEQ: return PPC::PRED_EQ;
581 case ISD::SETNE: return PPC::PRED_NE;
583 case ISD::SETLT: return PPC::PRED_LT;
585 case ISD::SETLE: return PPC::PRED_LE;
587 case ISD::SETGT: return PPC::PRED_GT;
589 case ISD::SETGE: return PPC::PRED_GE;
590 case ISD::SETO: return PPC::PRED_NU;
591 case ISD::SETUO: return PPC::PRED_UN;
592 // These two are invalid for floating point. Assume we have int.
593 case ISD::SETULT: return PPC::PRED_LT;
594 case ISD::SETUGT: return PPC::PRED_GT;
598 /// getCRIdxForSetCC - Return the index of the condition register field
599 /// associated with the SetCC condition, and whether or not the field is
600 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
602 /// If this returns with Other != -1, then the returned comparison is an or of
603 /// two simpler comparisons. In this case, Invert is guaranteed to be false.
604 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert, int &Other) {
608 default: llvm_unreachable("Unknown condition!");
610 case ISD::SETLT: return 0; // Bit #0 = SETOLT
612 case ISD::SETGT: return 1; // Bit #1 = SETOGT
614 case ISD::SETEQ: return 2; // Bit #2 = SETOEQ
615 case ISD::SETUO: return 3; // Bit #3 = SETUO
617 case ISD::SETGE: Invert = true; return 0; // !Bit #0 = SETUGE
619 case ISD::SETLE: Invert = true; return 1; // !Bit #1 = SETULE
621 case ISD::SETNE: Invert = true; return 2; // !Bit #2 = SETUNE
622 case ISD::SETO: Invert = true; return 3; // !Bit #3 = SETO
627 llvm_unreachable("Invalid branch code: should be expanded by legalize");
628 // These are invalid for floating point. Assume integer.
629 case ISD::SETULT: return 0;
630 case ISD::SETUGT: return 1;
634 // getVCmpInst: return the vector compare instruction for the specified
635 // vector type and condition code. Since this is for altivec specific code,
636 // only support the altivec types (v16i8, v8i16, v4i32, and v4f32).
637 static unsigned int getVCmpInst(MVT::SimpleValueType VecVT, ISD::CondCode CC) {
643 if (VecVT == MVT::v16i8)
644 return PPC::VCMPEQUB;
645 else if (VecVT == MVT::v8i16)
646 return PPC::VCMPEQUH;
647 else if (VecVT == MVT::v4i32)
648 return PPC::VCMPEQUW;
649 // v4f32 != v4f32 could be translate to unordered not equal
650 else if (VecVT == MVT::v4f32)
651 return PPC::VCMPEQFP;
657 if (VecVT == MVT::v16i8)
658 return PPC::VCMPGTSB;
659 else if (VecVT == MVT::v8i16)
660 return PPC::VCMPGTSH;
661 else if (VecVT == MVT::v4i32)
662 return PPC::VCMPGTSW;
663 else if (VecVT == MVT::v4f32)
664 return PPC::VCMPGTFP;
670 if (VecVT == MVT::v16i8)
671 return PPC::VCMPGTUB;
672 else if (VecVT == MVT::v8i16)
673 return PPC::VCMPGTUH;
674 else if (VecVT == MVT::v4i32)
675 return PPC::VCMPGTUW;
678 if (VecVT == MVT::v4f32)
679 return PPC::VCMPEQFP;
684 if (VecVT == MVT::v4f32)
685 return PPC::VCMPGTFP;
688 if (VecVT == MVT::v4f32)
689 return PPC::VCMPGEFP;
694 llvm_unreachable("Invalid integer vector compare condition");
697 // getVCmpEQInst: return the equal compare instruction for the specified vector
698 // type. Since this is for altivec specific code, only support the altivec
699 // types (v16i8, v8i16, v4i32, and v4f32).
700 static unsigned int getVCmpEQInst(MVT::SimpleValueType VecVT) {
703 return PPC::VCMPEQUB;
705 return PPC::VCMPEQUH;
707 return PPC::VCMPEQUW;
709 return PPC::VCMPEQFP;
711 llvm_unreachable("Invalid integer vector compare condition");
716 SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) {
717 DebugLoc dl = N->getDebugLoc();
719 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
720 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
721 bool isPPC64 = (PtrVT == MVT::i64);
723 if (isInt32Immediate(N->getOperand(1), Imm)) {
724 // We can codegen setcc op, imm very efficiently compared to a brcond.
725 // Check for those cases here.
728 SDValue Op = N->getOperand(0);
732 Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0);
733 SDValue Ops[] = { Op, getI32Imm(27), getI32Imm(5), getI32Imm(31) };
734 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
739 SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
740 Op, getI32Imm(~0U)), 0);
741 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
745 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
746 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
750 SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0);
751 T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0);
752 SDValue Ops[] = { T, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
753 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
756 } else if (Imm == ~0U) { // setcc op, -1
757 SDValue Op = N->getOperand(0);
762 Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
763 Op, getI32Imm(1)), 0);
764 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
765 SDValue(CurDAG->getMachineNode(PPC::LI, dl,
771 Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0);
772 SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
774 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0),
778 SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op,
780 SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD,
782 SDValue Ops[] = { AN, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
783 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
786 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
787 Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops, 4),
789 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op,
796 SDValue LHS = N->getOperand(0);
797 SDValue RHS = N->getOperand(1);
799 // Altivec Vector compare instructions do not set any CR register by default and
800 // vector compare operations return the same type as the operands.
801 if (LHS.getValueType().isVector()) {
802 EVT VecVT = LHS.getValueType();
803 MVT::SimpleValueType VT = VecVT.getSimpleVT().SimpleTy;
804 unsigned int VCmpInst = getVCmpInst(VT, CC);
810 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS);
814 SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0);
815 return CurDAG->SelectNodeTo(N, PPC::VNOR, VecVT, VCmp, VCmp);
820 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, RHS, LHS);
824 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS);
828 // Small optimization: Altivec provides a 'Vector Compare Greater Than
829 // or Equal To' instruction (vcmpgefp), so in this case there is no
830 // need for extra logic for the equal compare.
831 if (VecVT.getSimpleVT().isFloatingPoint()) {
832 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS);
834 SDValue VCmpGT(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0);
835 unsigned int VCmpEQInst = getVCmpEQInst(VT);
836 SDValue VCmpEQ(CurDAG->getMachineNode(VCmpEQInst, dl, VecVT, LHS, RHS), 0);
837 return CurDAG->SelectNodeTo(N, PPC::VOR, VecVT, VCmpGT, VCmpEQ);
843 SDValue VCmpLE(CurDAG->getMachineNode(VCmpInst, dl, VecVT, RHS, LHS), 0);
844 unsigned int VCmpEQInst = getVCmpEQInst(VT);
845 SDValue VCmpEQ(CurDAG->getMachineNode(VCmpEQInst, dl, VecVT, LHS, RHS), 0);
846 return CurDAG->SelectNodeTo(N, PPC::VOR, VecVT, VCmpLE, VCmpEQ);
849 llvm_unreachable("Invalid vector compare type: should be expanded by legalize");
855 unsigned Idx = getCRIdxForSetCC(CC, Inv, OtherCondIdx);
856 SDValue CCReg = SelectCC(LHS, RHS, CC, dl);
859 // Force the ccreg into CR7.
860 SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
862 SDValue InFlag(0, 0); // Null incoming flag value.
863 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg,
866 if (PPCSubTarget.hasMFOCRF() && OtherCondIdx == -1)
867 IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg,
870 IntCR = SDValue(CurDAG->getMachineNode(PPC::MFCRpseud, dl, MVT::i32,
873 SDValue Ops[] = { IntCR, getI32Imm((32-(3-Idx)) & 31),
874 getI32Imm(31), getI32Imm(31) };
875 if (OtherCondIdx == -1 && !Inv)
876 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
878 // Get the specified bit.
880 SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops, 4), 0);
882 assert(OtherCondIdx == -1 && "Can't have split plus negation");
883 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
886 // Otherwise, we have to turn an operation like SETONE -> SETOLT | SETOGT.
887 // We already got the bit for the first part of the comparison (e.g. SETULE).
889 // Get the other bit of the comparison.
890 Ops[1] = getI32Imm((32-(3-OtherCondIdx)) & 31);
892 SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops, 4), 0);
894 return CurDAG->SelectNodeTo(N, PPC::OR, MVT::i32, Tmp, OtherCond);
898 // Select - Convert the specified operand from a target-independent to a
899 // target-specific node if it hasn't already been changed.
900 SDNode *PPCDAGToDAGISel::Select(SDNode *N) {
901 DebugLoc dl = N->getDebugLoc();
902 if (N->isMachineOpcode())
903 return NULL; // Already selected.
905 switch (N->getOpcode()) {
908 case ISD::Constant: {
909 if (N->getValueType(0) == MVT::i64) {
911 int64_t Imm = cast<ConstantSDNode>(N)->getZExtValue();
912 // Assume no remaining bits.
913 unsigned Remainder = 0;
914 // Assume no shift required.
917 // If it can't be represented as a 32 bit value.
918 if (!isInt<32>(Imm)) {
919 Shift = CountTrailingZeros_64(Imm);
920 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
922 // If the shifted value fits 32 bits.
923 if (isInt<32>(ImmSh)) {
924 // Go with the shifted value.
927 // Still stuck with a 64 bit value.
934 // Intermediate operand.
937 // Handle first 32 bits.
938 unsigned Lo = Imm & 0xFFFF;
939 unsigned Hi = (Imm >> 16) & 0xFFFF;
942 if (isInt<16>(Imm)) {
944 Result = CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(Lo));
946 // Handle the Hi bits.
947 unsigned OpC = Hi ? PPC::LIS8 : PPC::LI8;
948 Result = CurDAG->getMachineNode(OpC, dl, MVT::i64, getI32Imm(Hi));
950 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
951 SDValue(Result, 0), getI32Imm(Lo));
954 Result = CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(Hi));
957 // If no shift, we're done.
958 if (!Shift) return Result;
960 // Shift for next step if the upper 32-bits were not zero.
962 Result = CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64,
965 getI32Imm(63 - Shift));
968 // Add in the last bits as required.
969 if ((Hi = (Remainder >> 16) & 0xFFFF)) {
970 Result = CurDAG->getMachineNode(PPC::ORIS8, dl, MVT::i64,
971 SDValue(Result, 0), getI32Imm(Hi));
973 if ((Lo = Remainder & 0xFFFF)) {
974 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
975 SDValue(Result, 0), getI32Imm(Lo));
984 return SelectSETCC(N);
985 case PPCISD::GlobalBaseReg:
986 return getGlobalBaseReg();
988 case ISD::FrameIndex: {
989 int FI = cast<FrameIndexSDNode>(N)->getIndex();
990 SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0));
991 unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8;
993 return CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), TFI,
995 return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI,
1000 SDValue InFlag = N->getOperand(1);
1001 // Use MFOCRF if supported.
1002 if (PPCSubTarget.hasMFOCRF())
1003 return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32,
1004 N->getOperand(0), InFlag);
1006 return CurDAG->getMachineNode(PPC::MFCRpseud, dl, MVT::i32,
1007 N->getOperand(0), InFlag);
1011 // FIXME: since this depends on the setting of the carry flag from the srawi
1012 // we should really be making notes about that for the scheduler.
1013 // FIXME: It sure would be nice if we could cheaply recognize the
1014 // srl/add/sra pattern the dag combiner will generate for this as
1015 // sra/addze rather than having to handle sdiv ourselves. oh well.
1017 if (isInt32Immediate(N->getOperand(1), Imm)) {
1018 SDValue N0 = N->getOperand(0);
1019 if ((signed)Imm > 0 && isPowerOf2_32(Imm)) {
1021 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
1022 N0, getI32Imm(Log2_32(Imm)));
1023 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
1024 SDValue(Op, 0), SDValue(Op, 1));
1025 } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) {
1027 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
1028 N0, getI32Imm(Log2_32(-Imm)));
1030 SDValue(CurDAG->getMachineNode(PPC::ADDZE, dl, MVT::i32,
1031 SDValue(Op, 0), SDValue(Op, 1)),
1033 return CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT);
1037 // Other cases are autogenerated.
1042 // Handle preincrement loads.
1043 LoadSDNode *LD = cast<LoadSDNode>(N);
1044 EVT LoadedVT = LD->getMemoryVT();
1046 // Normal loads are handled by code generated from the .td file.
1047 if (LD->getAddressingMode() != ISD::PRE_INC)
1050 SDValue Offset = LD->getOffset();
1051 if (isa<ConstantSDNode>(Offset) ||
1052 Offset.getOpcode() == ISD::TargetGlobalAddress) {
1055 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
1056 if (LD->getValueType(0) != MVT::i64) {
1057 // Handle PPC32 integer and normal FP loads.
1058 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
1059 switch (LoadedVT.getSimpleVT().SimpleTy) {
1060 default: llvm_unreachable("Invalid PPC load type!");
1061 case MVT::f64: Opcode = PPC::LFDU; break;
1062 case MVT::f32: Opcode = PPC::LFSU; break;
1063 case MVT::i32: Opcode = PPC::LWZU; break;
1064 case MVT::i16: Opcode = isSExt ? PPC::LHAU : PPC::LHZU; break;
1066 case MVT::i8: Opcode = PPC::LBZU; break;
1069 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
1070 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
1071 switch (LoadedVT.getSimpleVT().SimpleTy) {
1072 default: llvm_unreachable("Invalid PPC load type!");
1073 case MVT::i64: Opcode = PPC::LDU; break;
1074 case MVT::i32: Opcode = PPC::LWZU8; break;
1075 case MVT::i16: Opcode = isSExt ? PPC::LHAU8 : PPC::LHZU8; break;
1077 case MVT::i8: Opcode = PPC::LBZU8; break;
1081 SDValue Chain = LD->getChain();
1082 SDValue Base = LD->getBasePtr();
1083 SDValue Ops[] = { Offset, Base, Chain };
1084 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
1085 PPCLowering.getPointerTy(),
1086 MVT::Other, Ops, 3);
1089 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
1090 if (LD->getValueType(0) != MVT::i64) {
1091 // Handle PPC32 integer and normal FP loads.
1092 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
1093 switch (LoadedVT.getSimpleVT().SimpleTy) {
1094 default: llvm_unreachable("Invalid PPC load type!");
1095 case MVT::f64: Opcode = PPC::LFDUX; break;
1096 case MVT::f32: Opcode = PPC::LFSUX; break;
1097 case MVT::i32: Opcode = PPC::LWZUX; break;
1098 case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break;
1100 case MVT::i8: Opcode = PPC::LBZUX; break;
1103 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
1104 assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) &&
1105 "Invalid sext update load");
1106 switch (LoadedVT.getSimpleVT().SimpleTy) {
1107 default: llvm_unreachable("Invalid PPC load type!");
1108 case MVT::i64: Opcode = PPC::LDUX; break;
1109 case MVT::i32: Opcode = isSExt ? PPC::LWAUX : PPC::LWZUX8; break;
1110 case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break;
1112 case MVT::i8: Opcode = PPC::LBZUX8; break;
1116 SDValue Chain = LD->getChain();
1117 SDValue Base = LD->getBasePtr();
1118 SDValue Ops[] = { Offset, Base, Chain };
1119 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
1120 PPCLowering.getPointerTy(),
1121 MVT::Other, Ops, 3);
1126 unsigned Imm, Imm2, SH, MB, ME;
1129 // If this is an and of a value rotated between 0 and 31 bits and then and'd
1130 // with a mask, emit rlwinm
1131 if (isInt32Immediate(N->getOperand(1), Imm) &&
1132 isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) {
1133 SDValue Val = N->getOperand(0).getOperand(0);
1134 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
1135 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
1137 // If this is just a masked value where the input is not handled above, and
1138 // is not a rotate-left (handled by a pattern in the .td file), emit rlwinm
1139 if (isInt32Immediate(N->getOperand(1), Imm) &&
1140 isRunOfOnes(Imm, MB, ME) &&
1141 N->getOperand(0).getOpcode() != ISD::ROTL) {
1142 SDValue Val = N->getOperand(0);
1143 SDValue Ops[] = { Val, getI32Imm(0), getI32Imm(MB), getI32Imm(ME) };
1144 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
1146 // If this is a 64-bit zero-extension mask, emit rldicl.
1147 if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) &&
1149 SDValue Val = N->getOperand(0);
1150 MB = 64 - CountTrailingOnes_64(Imm64);
1151 SDValue Ops[] = { Val, getI32Imm(0), getI32Imm(MB) };
1152 return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops, 3);
1154 // AND X, 0 -> 0, not "rlwinm 32".
1155 if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) {
1156 ReplaceUses(SDValue(N, 0), N->getOperand(1));
1159 // ISD::OR doesn't get all the bitfield insertion fun.
1160 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert
1161 if (isInt32Immediate(N->getOperand(1), Imm) &&
1162 N->getOperand(0).getOpcode() == ISD::OR &&
1163 isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) {
1166 if (isRunOfOnes(Imm, MB, ME)) {
1167 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1168 N->getOperand(0).getOperand(1),
1169 getI32Imm(0), getI32Imm(MB),getI32Imm(ME) };
1170 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops, 5);
1174 // Other cases are autogenerated.
1178 if (N->getValueType(0) == MVT::i32)
1179 if (SDNode *I = SelectBitfieldInsert(N))
1182 // Other cases are autogenerated.
1185 unsigned Imm, SH, MB, ME;
1186 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
1187 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
1188 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1189 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
1190 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
1193 // Other cases are autogenerated.
1197 unsigned Imm, SH, MB, ME;
1198 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
1199 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
1200 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1201 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
1202 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
1205 // Other cases are autogenerated.
1208 case ISD::SELECT_CC: {
1209 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1210 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
1211 bool isPPC64 = (PtrVT == MVT::i64);
1213 // Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
1215 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1216 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
1217 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
1218 if (N1C->isNullValue() && N3C->isNullValue() &&
1219 N2C->getZExtValue() == 1ULL && CC == ISD::SETNE &&
1220 // FIXME: Implement this optzn for PPC64.
1221 N->getValueType(0) == MVT::i32) {
1223 CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
1224 N->getOperand(0), getI32Imm(~0U));
1225 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32,
1226 SDValue(Tmp, 0), N->getOperand(0),
1230 SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl);
1231 unsigned BROpc = getPredicateForSetCC(CC);
1233 unsigned SelectCCOp;
1234 if (N->getValueType(0) == MVT::i32)
1235 SelectCCOp = PPC::SELECT_CC_I4;
1236 else if (N->getValueType(0) == MVT::i64)
1237 SelectCCOp = PPC::SELECT_CC_I8;
1238 else if (N->getValueType(0) == MVT::f32)
1239 SelectCCOp = PPC::SELECT_CC_F4;
1240 else if (N->getValueType(0) == MVT::f64)
1241 SelectCCOp = PPC::SELECT_CC_F8;
1243 SelectCCOp = PPC::SELECT_CC_VRRC;
1245 SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3),
1247 return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops, 4);
1249 case PPCISD::COND_BRANCH: {
1250 // Op #0 is the Chain.
1251 // Op #1 is the PPC::PRED_* number.
1253 // Op #3 is the Dest MBB
1254 // Op #4 is the Flag.
1255 // Prevent PPC::PRED_* from being selected into LI.
1257 getI32Imm(cast<ConstantSDNode>(N->getOperand(1))->getZExtValue());
1258 SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3),
1259 N->getOperand(0), N->getOperand(4) };
1260 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops, 5);
1263 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
1264 SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl);
1265 SDValue Ops[] = { getI32Imm(getPredicateForSetCC(CC)), CondCode,
1266 N->getOperand(4), N->getOperand(0) };
1267 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops, 4);
1270 // FIXME: Should custom lower this.
1271 SDValue Chain = N->getOperand(0);
1272 SDValue Target = N->getOperand(1);
1273 unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8;
1274 unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8;
1275 Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Target,
1277 return CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain);
1279 case PPCISD::TOC_ENTRY: {
1280 assert (PPCSubTarget.isPPC64() && "Only supported for 64-bit ABI");
1282 // For medium code model, we generate two instructions as described
1283 // below. Otherwise we allow SelectCodeCommon to handle this, selecting
1284 // one of LDtoc, LDtocJTI, and LDtocCPT.
1285 if (TM.getCodeModel() != CodeModel::Medium)
1288 // The first source operand is a TargetGlobalAddress or a
1289 // TargetJumpTable. If it is an externally defined symbol, a symbol
1290 // with common linkage, a function address, or a jump table address,
1292 // LDtocL(<ga:@sym>, ADDIStocHA(%X2, <ga:@sym>))
1293 // Otherwise we generate:
1294 // ADDItocL(ADDIStocHA(%X2, <ga:@sym>), <ga:@sym>)
1295 SDValue GA = N->getOperand(0);
1296 SDValue TOCbase = N->getOperand(1);
1297 SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64,
1300 if (isa<JumpTableSDNode>(GA))
1301 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
1304 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) {
1305 const GlobalValue *GValue = G->getGlobal();
1306 const GlobalAlias *GAlias = dyn_cast<GlobalAlias>(GValue);
1307 const GlobalValue *RealGValue = GAlias ?
1308 GAlias->resolveAliasedGlobal(false) : GValue;
1309 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(RealGValue);
1310 assert((GVar || isa<Function>(RealGValue)) &&
1311 "Unexpected global value subclass!");
1313 // An external variable is one without an initializer. For these,
1314 // for variables with common linkage, and for Functions, generate
1316 if (!GVar || !GVar->hasInitializer() || RealGValue->hasCommonLinkage() ||
1317 RealGValue->hasAvailableExternallyLinkage())
1318 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
1322 return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64,
1323 SDValue(Tmp, 0), GA);
1325 case PPCISD::VADD_SPLAT: {
1326 // This expands into one of three sequences, depending on whether
1327 // the first operand is odd or even, positive or negative.
1328 assert(isa<ConstantSDNode>(N->getOperand(0)) &&
1329 isa<ConstantSDNode>(N->getOperand(1)) &&
1330 "Invalid operand on VADD_SPLAT!");
1332 int Elt = N->getConstantOperandVal(0);
1333 int EltSize = N->getConstantOperandVal(1);
1334 unsigned Opc1, Opc2, Opc3;
1338 Opc1 = PPC::VSPLTISB;
1339 Opc2 = PPC::VADDUBM;
1340 Opc3 = PPC::VSUBUBM;
1342 } else if (EltSize == 2) {
1343 Opc1 = PPC::VSPLTISH;
1344 Opc2 = PPC::VADDUHM;
1345 Opc3 = PPC::VSUBUHM;
1348 assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!");
1349 Opc1 = PPC::VSPLTISW;
1350 Opc2 = PPC::VADDUWM;
1351 Opc3 = PPC::VSUBUWM;
1355 if ((Elt & 1) == 0) {
1356 // Elt is even, in the range [-32,-18] + [16,30].
1358 // Convert: VADD_SPLAT elt, size
1359 // Into: tmp = VSPLTIS[BHW] elt
1360 // VADDU[BHW]M tmp, tmp
1361 // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4
1362 SDValue EltVal = getI32Imm(Elt >> 1);
1363 SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1364 SDValue TmpVal = SDValue(Tmp, 0);
1365 return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal);
1367 } else if (Elt > 0) {
1368 // Elt is odd and positive, in the range [17,31].
1370 // Convert: VADD_SPLAT elt, size
1371 // Into: tmp1 = VSPLTIS[BHW] elt-16
1372 // tmp2 = VSPLTIS[BHW] -16
1373 // VSUBU[BHW]M tmp1, tmp2
1374 SDValue EltVal = getI32Imm(Elt - 16);
1375 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1376 EltVal = getI32Imm(-16);
1377 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1378 return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0),
1382 // Elt is odd and negative, in the range [-31,-17].
1384 // Convert: VADD_SPLAT elt, size
1385 // Into: tmp1 = VSPLTIS[BHW] elt+16
1386 // tmp2 = VSPLTIS[BHW] -16
1387 // VADDU[BHW]M tmp1, tmp2
1388 SDValue EltVal = getI32Imm(Elt + 16);
1389 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1390 EltVal = getI32Imm(-16);
1391 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1392 return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0),
1398 return SelectCode(N);
1403 /// createPPCISelDag - This pass converts a legalized DAG into a
1404 /// PowerPC-specific DAG, ready for instruction scheduling.
1406 FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM) {
1407 return new PPCDAGToDAGISel(TM);
1410 static void initializePassOnce(PassRegistry &Registry) {
1411 const char *Name = "PowerPC DAG->DAG Pattern Instruction Selection";
1412 PassInfo *PI = new PassInfo(Name, "ppc-codegen", &SelectionDAGISel::ID, 0,
1414 Registry.registerPass(*PI, true);
1417 void llvm::initializePPCDAGToDAGISelPass(PassRegistry &Registry) {
1418 CALL_ONCE_INITIALIZATION(initializePassOnce);