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 //===----------------------------------------------------------------------===//
16 #include "MCTargetDesc/PPCPredicates.h"
17 #include "PPCMachineFunctionInfo.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/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetOptions.h"
38 #define DEBUG_TYPE "ppc-codegen"
40 // FIXME: Remove this once the bug has been fixed!
41 cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug",
42 cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden);
45 void initializePPCDAGToDAGISelPass(PassRegistry&);
49 //===--------------------------------------------------------------------===//
50 /// PPCDAGToDAGISel - PPC specific code to select PPC machine
51 /// instructions for SelectionDAG operations.
53 class PPCDAGToDAGISel : public SelectionDAGISel {
54 const PPCTargetMachine &TM;
55 const PPCTargetLowering *PPCLowering;
56 const PPCSubtarget *PPCSubTarget;
57 unsigned GlobalBaseReg;
59 explicit PPCDAGToDAGISel(PPCTargetMachine &tm)
60 : SelectionDAGISel(tm), TM(tm),
61 PPCLowering(TM.getSubtargetImpl()->getTargetLowering()),
62 PPCSubTarget(TM.getSubtargetImpl()) {
63 initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry());
66 bool runOnMachineFunction(MachineFunction &MF) override {
67 // Make sure we re-emit a set of the global base reg if necessary
69 PPCLowering = TM.getSubtargetImpl()->getTargetLowering();
70 PPCSubTarget = TM.getSubtargetImpl();
71 SelectionDAGISel::runOnMachineFunction(MF);
73 if (!PPCSubTarget->isSVR4ABI())
79 void PostprocessISelDAG() override;
81 /// getI32Imm - Return a target constant with the specified value, of type
83 inline SDValue getI32Imm(unsigned Imm) {
84 return CurDAG->getTargetConstant(Imm, MVT::i32);
87 /// getI64Imm - Return a target constant with the specified value, of type
89 inline SDValue getI64Imm(uint64_t Imm) {
90 return CurDAG->getTargetConstant(Imm, MVT::i64);
93 /// getSmallIPtrImm - Return a target constant of pointer type.
94 inline SDValue getSmallIPtrImm(unsigned Imm) {
95 return CurDAG->getTargetConstant(Imm, PPCLowering->getPointerTy());
98 /// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s
99 /// with any number of 0s on either side. The 1s are allowed to wrap from
100 /// LSB to MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs.
101 /// 0x0F0F0000 is not, since all 1s are not contiguous.
102 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME);
105 /// isRotateAndMask - Returns true if Mask and Shift can be folded into a
106 /// rotate and mask opcode and mask operation.
107 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask,
108 unsigned &SH, unsigned &MB, unsigned &ME);
110 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
111 /// base register. Return the virtual register that holds this value.
112 SDNode *getGlobalBaseReg();
114 // Select - Convert the specified operand from a target-independent to a
115 // target-specific node if it hasn't already been changed.
116 SDNode *Select(SDNode *N) override;
118 SDNode *SelectBitfieldInsert(SDNode *N);
120 /// SelectCC - Select a comparison of the specified values with the
121 /// specified condition code, returning the CR# of the expression.
122 SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, SDLoc dl);
124 /// SelectAddrImm - Returns true if the address N can be represented by
125 /// a base register plus a signed 16-bit displacement [r+imm].
126 bool SelectAddrImm(SDValue N, SDValue &Disp,
128 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, false);
131 /// SelectAddrImmOffs - Return true if the operand is valid for a preinc
132 /// immediate field. Note that the operand at this point is already the
133 /// result of a prior SelectAddressRegImm call.
134 bool SelectAddrImmOffs(SDValue N, SDValue &Out) const {
135 if (N.getOpcode() == ISD::TargetConstant ||
136 N.getOpcode() == ISD::TargetGlobalAddress) {
144 /// SelectAddrIdx - Given the specified addressed, check to see if it can be
145 /// represented as an indexed [r+r] operation. Returns false if it can
146 /// be represented by [r+imm], which are preferred.
147 bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) {
148 return PPCLowering->SelectAddressRegReg(N, Base, Index, *CurDAG);
151 /// SelectAddrIdxOnly - Given the specified addressed, force it to be
152 /// represented as an indexed [r+r] operation.
153 bool SelectAddrIdxOnly(SDValue N, SDValue &Base, SDValue &Index) {
154 return PPCLowering->SelectAddressRegRegOnly(N, Base, Index, *CurDAG);
157 /// SelectAddrImmX4 - Returns true if the address N can be represented by
158 /// a base register plus a signed 16-bit displacement that is a multiple of 4.
159 /// Suitable for use by STD and friends.
160 bool SelectAddrImmX4(SDValue N, SDValue &Disp, SDValue &Base) {
161 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, true);
164 // Select an address into a single register.
165 bool SelectAddr(SDValue N, SDValue &Base) {
170 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
171 /// inline asm expressions. It is always correct to compute the value into
172 /// a register. The case of adding a (possibly relocatable) constant to a
173 /// register can be improved, but it is wrong to substitute Reg+Reg for
174 /// Reg in an asm, because the load or store opcode would have to change.
175 bool SelectInlineAsmMemoryOperand(const SDValue &Op,
177 std::vector<SDValue> &OutOps) override {
178 OutOps.push_back(Op);
182 void InsertVRSaveCode(MachineFunction &MF);
184 const char *getPassName() const override {
185 return "PowerPC DAG->DAG Pattern Instruction Selection";
188 // Include the pieces autogenerated from the target description.
189 #include "PPCGenDAGISel.inc"
192 SDNode *SelectSETCC(SDNode *N);
194 void PeepholePPC64();
195 void PeepholeCROps();
197 bool AllUsersSelectZero(SDNode *N);
198 void SwapAllSelectUsers(SDNode *N);
202 /// InsertVRSaveCode - Once the entire function has been instruction selected,
203 /// all virtual registers are created and all machine instructions are built,
204 /// check to see if we need to save/restore VRSAVE. If so, do it.
205 void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) {
206 // Check to see if this function uses vector registers, which means we have to
207 // save and restore the VRSAVE register and update it with the regs we use.
209 // In this case, there will be virtual registers of vector type created
210 // by the scheduler. Detect them now.
211 bool HasVectorVReg = false;
212 for (unsigned i = 0, e = RegInfo->getNumVirtRegs(); i != e; ++i) {
213 unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
214 if (RegInfo->getRegClass(Reg) == &PPC::VRRCRegClass) {
215 HasVectorVReg = true;
219 if (!HasVectorVReg) return; // nothing to do.
221 // If we have a vector register, we want to emit code into the entry and exit
222 // blocks to save and restore the VRSAVE register. We do this here (instead
223 // of marking all vector instructions as clobbering VRSAVE) for two reasons:
225 // 1. This (trivially) reduces the load on the register allocator, by not
226 // having to represent the live range of the VRSAVE register.
227 // 2. This (more significantly) allows us to create a temporary virtual
228 // register to hold the saved VRSAVE value, allowing this temporary to be
229 // register allocated, instead of forcing it to be spilled to the stack.
231 // Create two vregs - one to hold the VRSAVE register that is live-in to the
232 // function and one for the value after having bits or'd into it.
233 unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
234 unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
236 const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
237 MachineBasicBlock &EntryBB = *Fn.begin();
239 // Emit the following code into the entry block:
240 // InVRSAVE = MFVRSAVE
241 // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE
242 // MTVRSAVE UpdatedVRSAVE
243 MachineBasicBlock::iterator IP = EntryBB.begin(); // Insert Point
244 BuildMI(EntryBB, IP, dl, TII.get(PPC::MFVRSAVE), InVRSAVE);
245 BuildMI(EntryBB, IP, dl, TII.get(PPC::UPDATE_VRSAVE),
246 UpdatedVRSAVE).addReg(InVRSAVE);
247 BuildMI(EntryBB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(UpdatedVRSAVE);
249 // Find all return blocks, outputting a restore in each epilog.
250 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
251 if (!BB->empty() && BB->back().isReturn()) {
252 IP = BB->end(); --IP;
254 // Skip over all terminator instructions, which are part of the return
256 MachineBasicBlock::iterator I2 = IP;
257 while (I2 != BB->begin() && (--I2)->isTerminator())
260 // Emit: MTVRSAVE InVRSave
261 BuildMI(*BB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(InVRSAVE);
267 /// getGlobalBaseReg - Output the instructions required to put the
268 /// base address to use for accessing globals into a register.
270 SDNode *PPCDAGToDAGISel::getGlobalBaseReg() {
271 if (!GlobalBaseReg) {
272 const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
273 // Insert the set of GlobalBaseReg into the first MBB of the function
274 MachineBasicBlock &FirstMBB = MF->front();
275 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
278 if (PPCLowering->getPointerTy() == MVT::i32) {
279 if (PPCSubTarget->isTargetELF())
280 GlobalBaseReg = PPC::R30;
283 RegInfo->createVirtualRegister(&PPC::GPRC_NOR0RegClass);
284 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
285 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
286 if (PPCSubTarget->isTargetELF()) {
287 unsigned TempReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
288 BuildMI(FirstMBB, MBBI, dl,
289 TII.get(PPC::GetGBRO), TempReg).addReg(GlobalBaseReg);
290 BuildMI(FirstMBB, MBBI, dl,
291 TII.get(PPC::UpdateGBR)).addReg(GlobalBaseReg).addReg(TempReg);
292 MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true);
295 GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RC_NOX0RegClass);
296 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8));
297 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg);
300 return CurDAG->getRegister(GlobalBaseReg,
301 PPCLowering->getPointerTy()).getNode();
304 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
305 /// or 64-bit immediate, and if the value can be accurately represented as a
306 /// sign extension from a 16-bit value. If so, this returns true and the
308 static bool isIntS16Immediate(SDNode *N, short &Imm) {
309 if (N->getOpcode() != ISD::Constant)
312 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
313 if (N->getValueType(0) == MVT::i32)
314 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
316 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
319 static bool isIntS16Immediate(SDValue Op, short &Imm) {
320 return isIntS16Immediate(Op.getNode(), Imm);
324 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
325 /// operand. If so Imm will receive the 32-bit value.
326 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
327 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
328 Imm = cast<ConstantSDNode>(N)->getZExtValue();
334 /// isInt64Immediate - This method tests to see if the node is a 64-bit constant
335 /// operand. If so Imm will receive the 64-bit value.
336 static bool isInt64Immediate(SDNode *N, uint64_t &Imm) {
337 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i64) {
338 Imm = cast<ConstantSDNode>(N)->getZExtValue();
344 // isInt32Immediate - This method tests to see if a constant operand.
345 // If so Imm will receive the 32 bit value.
346 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
347 return isInt32Immediate(N.getNode(), Imm);
351 // isOpcWithIntImmediate - This method tests to see if the node is a specific
352 // opcode and that it has a immediate integer right operand.
353 // If so Imm will receive the 32 bit value.
354 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
355 return N->getOpcode() == Opc
356 && isInt32Immediate(N->getOperand(1).getNode(), Imm);
359 bool PPCDAGToDAGISel::isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
363 if (isShiftedMask_32(Val)) {
364 // look for the first non-zero bit
365 MB = countLeadingZeros(Val);
366 // look for the first zero bit after the run of ones
367 ME = countLeadingZeros((Val - 1) ^ Val);
370 Val = ~Val; // invert mask
371 if (isShiftedMask_32(Val)) {
372 // effectively look for the first zero bit
373 ME = countLeadingZeros(Val) - 1;
374 // effectively look for the first one bit after the run of zeros
375 MB = countLeadingZeros((Val - 1) ^ Val) + 1;
383 bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask,
384 bool isShiftMask, unsigned &SH,
385 unsigned &MB, unsigned &ME) {
386 // Don't even go down this path for i64, since different logic will be
387 // necessary for rldicl/rldicr/rldimi.
388 if (N->getValueType(0) != MVT::i32)
392 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
393 unsigned Opcode = N->getOpcode();
394 if (N->getNumOperands() != 2 ||
395 !isInt32Immediate(N->getOperand(1).getNode(), Shift) || (Shift > 31))
398 if (Opcode == ISD::SHL) {
399 // apply shift left to mask if it comes first
400 if (isShiftMask) Mask = Mask << Shift;
401 // determine which bits are made indeterminant by shift
402 Indeterminant = ~(0xFFFFFFFFu << Shift);
403 } else if (Opcode == ISD::SRL) {
404 // apply shift right to mask if it comes first
405 if (isShiftMask) Mask = Mask >> Shift;
406 // determine which bits are made indeterminant by shift
407 Indeterminant = ~(0xFFFFFFFFu >> Shift);
408 // adjust for the left rotate
410 } else if (Opcode == ISD::ROTL) {
416 // if the mask doesn't intersect any Indeterminant bits
417 if (Mask && !(Mask & Indeterminant)) {
419 // make sure the mask is still a mask (wrap arounds may not be)
420 return isRunOfOnes(Mask, MB, ME);
425 /// SelectBitfieldInsert - turn an or of two masked values into
426 /// the rotate left word immediate then mask insert (rlwimi) instruction.
427 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
428 SDValue Op0 = N->getOperand(0);
429 SDValue Op1 = N->getOperand(1);
432 APInt LKZ, LKO, RKZ, RKO;
433 CurDAG->computeKnownBits(Op0, LKZ, LKO);
434 CurDAG->computeKnownBits(Op1, RKZ, RKO);
436 unsigned TargetMask = LKZ.getZExtValue();
437 unsigned InsertMask = RKZ.getZExtValue();
439 if ((TargetMask | InsertMask) == 0xFFFFFFFF) {
440 unsigned Op0Opc = Op0.getOpcode();
441 unsigned Op1Opc = Op1.getOpcode();
442 unsigned Value, SH = 0;
443 TargetMask = ~TargetMask;
444 InsertMask = ~InsertMask;
446 // If the LHS has a foldable shift and the RHS does not, then swap it to the
447 // RHS so that we can fold the shift into the insert.
448 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
449 if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
450 Op0.getOperand(0).getOpcode() == ISD::SRL) {
451 if (Op1.getOperand(0).getOpcode() != ISD::SHL &&
452 Op1.getOperand(0).getOpcode() != ISD::SRL) {
454 std::swap(Op0Opc, Op1Opc);
455 std::swap(TargetMask, InsertMask);
458 } else if (Op0Opc == ISD::SHL || Op0Opc == ISD::SRL) {
459 if (Op1Opc == ISD::AND && Op1.getOperand(0).getOpcode() != ISD::SHL &&
460 Op1.getOperand(0).getOpcode() != ISD::SRL) {
462 std::swap(Op0Opc, Op1Opc);
463 std::swap(TargetMask, InsertMask);
468 if (isRunOfOnes(InsertMask, MB, ME)) {
471 if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) &&
472 isInt32Immediate(Op1.getOperand(1), Value)) {
473 Op1 = Op1.getOperand(0);
474 SH = (Op1Opc == ISD::SHL) ? Value : 32 - Value;
476 if (Op1Opc == ISD::AND) {
477 // The AND mask might not be a constant, and we need to make sure that
478 // if we're going to fold the masking with the insert, all bits not
479 // know to be zero in the mask are known to be one.
481 CurDAG->computeKnownBits(Op1.getOperand(1), MKZ, MKO);
482 bool CanFoldMask = InsertMask == MKO.getZExtValue();
484 unsigned SHOpc = Op1.getOperand(0).getOpcode();
485 if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) && CanFoldMask &&
486 isInt32Immediate(Op1.getOperand(0).getOperand(1), Value)) {
487 // Note that Value must be in range here (less than 32) because
488 // otherwise there would not be any bits set in InsertMask.
489 Op1 = Op1.getOperand(0).getOperand(0);
490 SH = (SHOpc == ISD::SHL) ? Value : 32 - Value;
495 SDValue Ops[] = { Op0, Op1, getI32Imm(SH), getI32Imm(MB),
497 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
503 /// SelectCC - Select a comparison of the specified values with the specified
504 /// condition code, returning the CR# of the expression.
505 SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS,
506 ISD::CondCode CC, SDLoc dl) {
507 // Always select the LHS.
510 if (LHS.getValueType() == MVT::i32) {
512 if (CC == ISD::SETEQ || CC == ISD::SETNE) {
513 if (isInt32Immediate(RHS, Imm)) {
514 // SETEQ/SETNE comparison with 16-bit immediate, fold it.
516 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
517 getI32Imm(Imm & 0xFFFF)), 0);
518 // If this is a 16-bit signed immediate, fold it.
519 if (isInt<16>((int)Imm))
520 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
521 getI32Imm(Imm & 0xFFFF)), 0);
523 // For non-equality comparisons, the default code would materialize the
524 // constant, then compare against it, like this:
528 // Since we are just comparing for equality, we can emit this instead:
529 // xoris r0,r3,0x1234
530 // cmplwi cr0,r0,0x5678
532 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS,
533 getI32Imm(Imm >> 16)), 0);
534 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor,
535 getI32Imm(Imm & 0xFFFF)), 0);
538 } else if (ISD::isUnsignedIntSetCC(CC)) {
539 if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm))
540 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
541 getI32Imm(Imm & 0xFFFF)), 0);
545 if (isIntS16Immediate(RHS, SImm))
546 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
547 getI32Imm((int)SImm & 0xFFFF)),
551 } else if (LHS.getValueType() == MVT::i64) {
553 if (CC == ISD::SETEQ || CC == ISD::SETNE) {
554 if (isInt64Immediate(RHS.getNode(), Imm)) {
555 // SETEQ/SETNE comparison with 16-bit immediate, fold it.
557 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
558 getI32Imm(Imm & 0xFFFF)), 0);
559 // If this is a 16-bit signed immediate, fold it.
561 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
562 getI32Imm(Imm & 0xFFFF)), 0);
564 // For non-equality comparisons, the default code would materialize the
565 // constant, then compare against it, like this:
569 // Since we are just comparing for equality, we can emit this instead:
570 // xoris r0,r3,0x1234
571 // cmpldi cr0,r0,0x5678
573 if (isUInt<32>(Imm)) {
574 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS,
575 getI64Imm(Imm >> 16)), 0);
576 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor,
577 getI64Imm(Imm & 0xFFFF)), 0);
581 } else if (ISD::isUnsignedIntSetCC(CC)) {
582 if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm))
583 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
584 getI64Imm(Imm & 0xFFFF)), 0);
588 if (isIntS16Immediate(RHS, SImm))
589 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
590 getI64Imm(SImm & 0xFFFF)),
594 } else if (LHS.getValueType() == MVT::f32) {
597 assert(LHS.getValueType() == MVT::f64 && "Unknown vt!");
598 Opc = PPCSubTarget->hasVSX() ? PPC::XSCMPUDP : PPC::FCMPUD;
600 return SDValue(CurDAG->getMachineNode(Opc, dl, MVT::i32, LHS, RHS), 0);
603 static PPC::Predicate getPredicateForSetCC(ISD::CondCode CC) {
609 llvm_unreachable("Should be lowered by legalize!");
610 default: llvm_unreachable("Unknown condition!");
612 case ISD::SETEQ: return PPC::PRED_EQ;
614 case ISD::SETNE: return PPC::PRED_NE;
616 case ISD::SETLT: return PPC::PRED_LT;
618 case ISD::SETLE: return PPC::PRED_LE;
620 case ISD::SETGT: return PPC::PRED_GT;
622 case ISD::SETGE: return PPC::PRED_GE;
623 case ISD::SETO: return PPC::PRED_NU;
624 case ISD::SETUO: return PPC::PRED_UN;
625 // These two are invalid for floating point. Assume we have int.
626 case ISD::SETULT: return PPC::PRED_LT;
627 case ISD::SETUGT: return PPC::PRED_GT;
631 /// getCRIdxForSetCC - Return the index of the condition register field
632 /// associated with the SetCC condition, and whether or not the field is
633 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
634 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) {
637 default: llvm_unreachable("Unknown condition!");
639 case ISD::SETLT: return 0; // Bit #0 = SETOLT
641 case ISD::SETGT: return 1; // Bit #1 = SETOGT
643 case ISD::SETEQ: return 2; // Bit #2 = SETOEQ
644 case ISD::SETUO: return 3; // Bit #3 = SETUO
646 case ISD::SETGE: Invert = true; return 0; // !Bit #0 = SETUGE
648 case ISD::SETLE: Invert = true; return 1; // !Bit #1 = SETULE
650 case ISD::SETNE: Invert = true; return 2; // !Bit #2 = SETUNE
651 case ISD::SETO: Invert = true; return 3; // !Bit #3 = SETO
656 llvm_unreachable("Invalid branch code: should be expanded by legalize");
657 // These are invalid for floating point. Assume integer.
658 case ISD::SETULT: return 0;
659 case ISD::SETUGT: return 1;
663 // getVCmpInst: return the vector compare instruction for the specified
664 // vector type and condition code. Since this is for altivec specific code,
665 // only support the altivec types (v16i8, v8i16, v4i32, and v4f32).
666 static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC,
667 bool HasVSX, bool &Swap, bool &Negate) {
671 if (VecVT.isFloatingPoint()) {
672 /* Handle some cases by swapping input operands. */
674 case ISD::SETLE: CC = ISD::SETGE; Swap = true; break;
675 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break;
676 case ISD::SETOLE: CC = ISD::SETOGE; Swap = true; break;
677 case ISD::SETOLT: CC = ISD::SETOGT; Swap = true; break;
678 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break;
679 case ISD::SETUGT: CC = ISD::SETULT; Swap = true; break;
682 /* Handle some cases by negating the result. */
684 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break;
685 case ISD::SETUNE: CC = ISD::SETOEQ; Negate = true; break;
686 case ISD::SETULE: CC = ISD::SETOGT; Negate = true; break;
687 case ISD::SETULT: CC = ISD::SETOGE; Negate = true; break;
690 /* We have instructions implementing the remaining cases. */
694 if (VecVT == MVT::v4f32)
695 return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP;
696 else if (VecVT == MVT::v2f64)
697 return PPC::XVCMPEQDP;
701 if (VecVT == MVT::v4f32)
702 return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP;
703 else if (VecVT == MVT::v2f64)
704 return PPC::XVCMPGTDP;
708 if (VecVT == MVT::v4f32)
709 return HasVSX ? PPC::XVCMPGESP : PPC::VCMPGEFP;
710 else if (VecVT == MVT::v2f64)
711 return PPC::XVCMPGEDP;
716 llvm_unreachable("Invalid floating-point vector compare condition");
718 /* Handle some cases by swapping input operands. */
720 case ISD::SETGE: CC = ISD::SETLE; Swap = true; break;
721 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break;
722 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break;
723 case ISD::SETULT: CC = ISD::SETUGT; Swap = true; break;
726 /* Handle some cases by negating the result. */
728 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break;
729 case ISD::SETUNE: CC = ISD::SETUEQ; Negate = true; break;
730 case ISD::SETLE: CC = ISD::SETGT; Negate = true; break;
731 case ISD::SETULE: CC = ISD::SETUGT; Negate = true; break;
734 /* We have instructions implementing the remaining cases. */
738 if (VecVT == MVT::v16i8)
739 return PPC::VCMPEQUB;
740 else if (VecVT == MVT::v8i16)
741 return PPC::VCMPEQUH;
742 else if (VecVT == MVT::v4i32)
743 return PPC::VCMPEQUW;
746 if (VecVT == MVT::v16i8)
747 return PPC::VCMPGTSB;
748 else if (VecVT == MVT::v8i16)
749 return PPC::VCMPGTSH;
750 else if (VecVT == MVT::v4i32)
751 return PPC::VCMPGTSW;
754 if (VecVT == MVT::v16i8)
755 return PPC::VCMPGTUB;
756 else if (VecVT == MVT::v8i16)
757 return PPC::VCMPGTUH;
758 else if (VecVT == MVT::v4i32)
759 return PPC::VCMPGTUW;
764 llvm_unreachable("Invalid integer vector compare condition");
768 SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) {
771 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
772 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
773 bool isPPC64 = (PtrVT == MVT::i64);
775 if (!PPCSubTarget->useCRBits() &&
776 isInt32Immediate(N->getOperand(1), Imm)) {
777 // We can codegen setcc op, imm very efficiently compared to a brcond.
778 // Check for those cases here.
781 SDValue Op = N->getOperand(0);
785 Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0);
786 SDValue Ops[] = { Op, getI32Imm(27), getI32Imm(5), getI32Imm(31) };
787 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
792 SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
793 Op, getI32Imm(~0U)), 0);
794 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
798 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
799 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
803 SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0);
804 T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0);
805 SDValue Ops[] = { T, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
806 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
809 } else if (Imm == ~0U) { // setcc op, -1
810 SDValue Op = N->getOperand(0);
815 Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
816 Op, getI32Imm(1)), 0);
817 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
818 SDValue(CurDAG->getMachineNode(PPC::LI, dl,
824 Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0);
825 SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
827 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0),
831 SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op,
833 SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD,
835 SDValue Ops[] = { AN, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
836 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
839 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
840 Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops),
842 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op,
849 SDValue LHS = N->getOperand(0);
850 SDValue RHS = N->getOperand(1);
852 // Altivec Vector compare instructions do not set any CR register by default and
853 // vector compare operations return the same type as the operands.
854 if (LHS.getValueType().isVector()) {
855 EVT VecVT = LHS.getValueType();
857 unsigned int VCmpInst = getVCmpInst(VecVT.getSimpleVT(), CC,
858 PPCSubTarget->hasVSX(), Swap, Negate);
863 SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0);
864 return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLNOR :
869 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS);
872 if (PPCSubTarget->useCRBits())
876 unsigned Idx = getCRIdxForSetCC(CC, Inv);
877 SDValue CCReg = SelectCC(LHS, RHS, CC, dl);
880 // Force the ccreg into CR7.
881 SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
883 SDValue InFlag(nullptr, 0); // Null incoming flag value.
884 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg,
887 IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg,
890 SDValue Ops[] = { IntCR, getI32Imm((32-(3-Idx)) & 31),
891 getI32Imm(31), getI32Imm(31) };
893 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
895 // Get the specified bit.
897 SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
898 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
902 // Select - Convert the specified operand from a target-independent to a
903 // target-specific node if it hasn't already been changed.
904 SDNode *PPCDAGToDAGISel::Select(SDNode *N) {
906 if (N->isMachineOpcode()) {
908 return nullptr; // Already selected.
911 switch (N->getOpcode()) {
914 case ISD::Constant: {
915 if (N->getValueType(0) == MVT::i64) {
917 int64_t Imm = cast<ConstantSDNode>(N)->getZExtValue();
918 // Assume no remaining bits.
919 unsigned Remainder = 0;
920 // Assume no shift required.
923 // If it can't be represented as a 32 bit value.
924 if (!isInt<32>(Imm)) {
925 Shift = countTrailingZeros<uint64_t>(Imm);
926 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
928 // If the shifted value fits 32 bits.
929 if (isInt<32>(ImmSh)) {
930 // Go with the shifted value.
933 // Still stuck with a 64 bit value.
940 // Intermediate operand.
943 // Handle first 32 bits.
944 unsigned Lo = Imm & 0xFFFF;
945 unsigned Hi = (Imm >> 16) & 0xFFFF;
948 if (isInt<16>(Imm)) {
950 Result = CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(Lo));
952 // Handle the Hi bits.
953 unsigned OpC = Hi ? PPC::LIS8 : PPC::LI8;
954 Result = CurDAG->getMachineNode(OpC, dl, MVT::i64, getI32Imm(Hi));
956 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
957 SDValue(Result, 0), getI32Imm(Lo));
960 Result = CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(Hi));
963 // If no shift, we're done.
964 if (!Shift) return Result;
966 // Shift for next step if the upper 32-bits were not zero.
968 Result = CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64,
971 getI32Imm(63 - Shift));
974 // Add in the last bits as required.
975 if ((Hi = (Remainder >> 16) & 0xFFFF)) {
976 Result = CurDAG->getMachineNode(PPC::ORIS8, dl, MVT::i64,
977 SDValue(Result, 0), getI32Imm(Hi));
979 if ((Lo = Remainder & 0xFFFF)) {
980 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
981 SDValue(Result, 0), getI32Imm(Lo));
990 SDNode *SN = SelectSETCC(N);
995 case PPCISD::GlobalBaseReg:
996 return getGlobalBaseReg();
998 case ISD::FrameIndex: {
999 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1000 SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0));
1001 unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8;
1003 return CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), TFI,
1004 getSmallIPtrImm(0));
1005 return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI,
1006 getSmallIPtrImm(0));
1009 case PPCISD::MFOCRF: {
1010 SDValue InFlag = N->getOperand(1);
1011 return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32,
1012 N->getOperand(0), InFlag);
1016 // FIXME: since this depends on the setting of the carry flag from the srawi
1017 // we should really be making notes about that for the scheduler.
1018 // FIXME: It sure would be nice if we could cheaply recognize the
1019 // srl/add/sra pattern the dag combiner will generate for this as
1020 // sra/addze rather than having to handle sdiv ourselves. oh well.
1022 if (isInt32Immediate(N->getOperand(1), Imm)) {
1023 SDValue N0 = N->getOperand(0);
1024 if ((signed)Imm > 0 && isPowerOf2_32(Imm)) {
1026 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
1027 N0, getI32Imm(Log2_32(Imm)));
1028 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
1029 SDValue(Op, 0), SDValue(Op, 1));
1030 } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) {
1032 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
1033 N0, getI32Imm(Log2_32(-Imm)));
1035 SDValue(CurDAG->getMachineNode(PPC::ADDZE, dl, MVT::i32,
1036 SDValue(Op, 0), SDValue(Op, 1)),
1038 return CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT);
1042 // Other cases are autogenerated.
1047 // Handle preincrement loads.
1048 LoadSDNode *LD = cast<LoadSDNode>(N);
1049 EVT LoadedVT = LD->getMemoryVT();
1051 // Normal loads are handled by code generated from the .td file.
1052 if (LD->getAddressingMode() != ISD::PRE_INC)
1055 SDValue Offset = LD->getOffset();
1056 if (Offset.getOpcode() == ISD::TargetConstant ||
1057 Offset.getOpcode() == ISD::TargetGlobalAddress) {
1060 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
1061 if (LD->getValueType(0) != MVT::i64) {
1062 // Handle PPC32 integer and normal FP loads.
1063 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
1064 switch (LoadedVT.getSimpleVT().SimpleTy) {
1065 default: llvm_unreachable("Invalid PPC load type!");
1066 case MVT::f64: Opcode = PPC::LFDU; break;
1067 case MVT::f32: Opcode = PPC::LFSU; break;
1068 case MVT::i32: Opcode = PPC::LWZU; break;
1069 case MVT::i16: Opcode = isSExt ? PPC::LHAU : PPC::LHZU; break;
1071 case MVT::i8: Opcode = PPC::LBZU; break;
1074 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
1075 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
1076 switch (LoadedVT.getSimpleVT().SimpleTy) {
1077 default: llvm_unreachable("Invalid PPC load type!");
1078 case MVT::i64: Opcode = PPC::LDU; break;
1079 case MVT::i32: Opcode = PPC::LWZU8; break;
1080 case MVT::i16: Opcode = isSExt ? PPC::LHAU8 : PPC::LHZU8; break;
1082 case MVT::i8: Opcode = PPC::LBZU8; break;
1086 SDValue Chain = LD->getChain();
1087 SDValue Base = LD->getBasePtr();
1088 SDValue Ops[] = { Offset, Base, Chain };
1089 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
1090 PPCLowering->getPointerTy(),
1094 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
1095 if (LD->getValueType(0) != MVT::i64) {
1096 // Handle PPC32 integer and normal FP loads.
1097 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
1098 switch (LoadedVT.getSimpleVT().SimpleTy) {
1099 default: llvm_unreachable("Invalid PPC load type!");
1100 case MVT::f64: Opcode = PPC::LFDUX; break;
1101 case MVT::f32: Opcode = PPC::LFSUX; break;
1102 case MVT::i32: Opcode = PPC::LWZUX; break;
1103 case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break;
1105 case MVT::i8: Opcode = PPC::LBZUX; break;
1108 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
1109 assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) &&
1110 "Invalid sext update load");
1111 switch (LoadedVT.getSimpleVT().SimpleTy) {
1112 default: llvm_unreachable("Invalid PPC load type!");
1113 case MVT::i64: Opcode = PPC::LDUX; break;
1114 case MVT::i32: Opcode = isSExt ? PPC::LWAUX : PPC::LWZUX8; break;
1115 case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break;
1117 case MVT::i8: Opcode = PPC::LBZUX8; break;
1121 SDValue Chain = LD->getChain();
1122 SDValue Base = LD->getBasePtr();
1123 SDValue Ops[] = { Base, Offset, Chain };
1124 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
1125 PPCLowering->getPointerTy(),
1131 unsigned Imm, Imm2, SH, MB, ME;
1134 // If this is an and of a value rotated between 0 and 31 bits and then and'd
1135 // with a mask, emit rlwinm
1136 if (isInt32Immediate(N->getOperand(1), Imm) &&
1137 isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) {
1138 SDValue Val = N->getOperand(0).getOperand(0);
1139 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
1140 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
1142 // If this is just a masked value where the input is not handled above, and
1143 // is not a rotate-left (handled by a pattern in the .td file), emit rlwinm
1144 if (isInt32Immediate(N->getOperand(1), Imm) &&
1145 isRunOfOnes(Imm, MB, ME) &&
1146 N->getOperand(0).getOpcode() != ISD::ROTL) {
1147 SDValue Val = N->getOperand(0);
1148 SDValue Ops[] = { Val, getI32Imm(0), getI32Imm(MB), getI32Imm(ME) };
1149 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
1151 // If this is a 64-bit zero-extension mask, emit rldicl.
1152 if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) &&
1154 SDValue Val = N->getOperand(0);
1155 MB = 64 - CountTrailingOnes_64(Imm64);
1158 // If the operand is a logical right shift, we can fold it into this
1159 // instruction: rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb)
1160 // for n <= mb. The right shift is really a left rotate followed by a
1161 // mask, and this mask is a more-restrictive sub-mask of the mask implied
1163 if (Val.getOpcode() == ISD::SRL &&
1164 isInt32Immediate(Val.getOperand(1).getNode(), Imm) && Imm <= MB) {
1165 assert(Imm < 64 && "Illegal shift amount");
1166 Val = Val.getOperand(0);
1170 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB) };
1171 return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops);
1173 // AND X, 0 -> 0, not "rlwinm 32".
1174 if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) {
1175 ReplaceUses(SDValue(N, 0), N->getOperand(1));
1178 // ISD::OR doesn't get all the bitfield insertion fun.
1179 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert
1180 if (isInt32Immediate(N->getOperand(1), Imm) &&
1181 N->getOperand(0).getOpcode() == ISD::OR &&
1182 isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) {
1185 if (isRunOfOnes(Imm, MB, ME)) {
1186 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1187 N->getOperand(0).getOperand(1),
1188 getI32Imm(0), getI32Imm(MB),getI32Imm(ME) };
1189 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
1193 // Other cases are autogenerated.
1197 if (N->getValueType(0) == MVT::i32)
1198 if (SDNode *I = SelectBitfieldInsert(N))
1201 // Other cases are autogenerated.
1204 unsigned Imm, SH, MB, ME;
1205 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
1206 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
1207 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1208 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
1209 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
1212 // Other cases are autogenerated.
1216 unsigned Imm, SH, MB, ME;
1217 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
1218 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
1219 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1220 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
1221 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
1224 // Other cases are autogenerated.
1227 // FIXME: Remove this once the ANDI glue bug is fixed:
1228 case PPCISD::ANDIo_1_EQ_BIT:
1229 case PPCISD::ANDIo_1_GT_BIT: {
1233 EVT InVT = N->getOperand(0).getValueType();
1234 assert((InVT == MVT::i64 || InVT == MVT::i32) &&
1235 "Invalid input type for ANDIo_1_EQ_BIT");
1237 unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDIo8 : PPC::ANDIo;
1238 SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue,
1240 CurDAG->getTargetConstant(1, InVT)), 0);
1241 SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32);
1243 CurDAG->getTargetConstant(N->getOpcode() == PPCISD::ANDIo_1_EQ_BIT ?
1244 PPC::sub_eq : PPC::sub_gt, MVT::i32);
1246 return CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1,
1248 SDValue(AndI.getNode(), 1) /* glue */);
1250 case ISD::SELECT_CC: {
1251 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1252 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
1253 bool isPPC64 = (PtrVT == MVT::i64);
1255 // If this is a select of i1 operands, we'll pattern match it.
1256 if (PPCSubTarget->useCRBits() &&
1257 N->getOperand(0).getValueType() == MVT::i1)
1260 // Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
1262 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1263 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
1264 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
1265 if (N1C->isNullValue() && N3C->isNullValue() &&
1266 N2C->getZExtValue() == 1ULL && CC == ISD::SETNE &&
1267 // FIXME: Implement this optzn for PPC64.
1268 N->getValueType(0) == MVT::i32) {
1270 CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
1271 N->getOperand(0), getI32Imm(~0U));
1272 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32,
1273 SDValue(Tmp, 0), N->getOperand(0),
1277 SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl);
1279 if (N->getValueType(0) == MVT::i1) {
1280 // An i1 select is: (c & t) | (!c & f).
1282 unsigned Idx = getCRIdxForSetCC(CC, Inv);
1286 default: llvm_unreachable("Invalid CC index");
1287 case 0: SRI = PPC::sub_lt; break;
1288 case 1: SRI = PPC::sub_gt; break;
1289 case 2: SRI = PPC::sub_eq; break;
1290 case 3: SRI = PPC::sub_un; break;
1293 SDValue CCBit = CurDAG->getTargetExtractSubreg(SRI, dl, MVT::i1, CCReg);
1295 SDValue NotCCBit(CurDAG->getMachineNode(PPC::CRNOR, dl, MVT::i1,
1297 SDValue C = Inv ? NotCCBit : CCBit,
1298 NotC = Inv ? CCBit : NotCCBit;
1300 SDValue CAndT(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1,
1301 C, N->getOperand(2)), 0);
1302 SDValue NotCAndF(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1,
1303 NotC, N->getOperand(3)), 0);
1305 return CurDAG->SelectNodeTo(N, PPC::CROR, MVT::i1, CAndT, NotCAndF);
1308 unsigned BROpc = getPredicateForSetCC(CC);
1310 unsigned SelectCCOp;
1311 if (N->getValueType(0) == MVT::i32)
1312 SelectCCOp = PPC::SELECT_CC_I4;
1313 else if (N->getValueType(0) == MVT::i64)
1314 SelectCCOp = PPC::SELECT_CC_I8;
1315 else if (N->getValueType(0) == MVT::f32)
1316 SelectCCOp = PPC::SELECT_CC_F4;
1317 else if (N->getValueType(0) == MVT::f64)
1318 SelectCCOp = PPC::SELECT_CC_F8;
1320 SelectCCOp = PPC::SELECT_CC_VRRC;
1322 SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3),
1324 return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops);
1327 if (PPCSubTarget->hasVSX()) {
1328 SDValue Ops[] = { N->getOperand(2), N->getOperand(1), N->getOperand(0) };
1329 return CurDAG->SelectNodeTo(N, PPC::XXSEL, N->getValueType(0), Ops);
1333 case ISD::VECTOR_SHUFFLE:
1334 if (PPCSubTarget->hasVSX() && (N->getValueType(0) == MVT::v2f64 ||
1335 N->getValueType(0) == MVT::v2i64)) {
1336 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
1338 SDValue Op1 = N->getOperand(SVN->getMaskElt(0) < 2 ? 0 : 1),
1339 Op2 = N->getOperand(SVN->getMaskElt(1) < 2 ? 0 : 1);
1342 for (int i = 0; i < 2; ++i)
1343 if (SVN->getMaskElt(i) <= 0 || SVN->getMaskElt(i) == 2)
1348 SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), MVT::i32);
1350 if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 &&
1351 Op1.getOpcode() == ISD::SCALAR_TO_VECTOR &&
1352 isa<LoadSDNode>(Op1.getOperand(0))) {
1353 LoadSDNode *LD = cast<LoadSDNode>(Op1.getOperand(0));
1354 SDValue Base, Offset;
1356 if (LD->isUnindexed() &&
1357 SelectAddrIdxOnly(LD->getBasePtr(), Base, Offset)) {
1358 SDValue Chain = LD->getChain();
1359 SDValue Ops[] = { Base, Offset, Chain };
1360 return CurDAG->SelectNodeTo(N, PPC::LXVDSX,
1361 N->getValueType(0), Ops);
1365 SDValue Ops[] = { Op1, Op2, DMV };
1366 return CurDAG->SelectNodeTo(N, PPC::XXPERMDI, N->getValueType(0), Ops);
1372 bool IsPPC64 = PPCSubTarget->isPPC64();
1373 SDValue Ops[] = { N->getOperand(1), N->getOperand(0) };
1374 return CurDAG->SelectNodeTo(N, N->getOpcode() == PPCISD::BDNZ ?
1375 (IsPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
1376 (IsPPC64 ? PPC::BDZ8 : PPC::BDZ),
1379 case PPCISD::COND_BRANCH: {
1380 // Op #0 is the Chain.
1381 // Op #1 is the PPC::PRED_* number.
1383 // Op #3 is the Dest MBB
1384 // Op #4 is the Flag.
1385 // Prevent PPC::PRED_* from being selected into LI.
1387 getI32Imm(cast<ConstantSDNode>(N->getOperand(1))->getZExtValue());
1388 SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3),
1389 N->getOperand(0), N->getOperand(4) };
1390 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
1393 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
1394 unsigned PCC = getPredicateForSetCC(CC);
1396 if (N->getOperand(2).getValueType() == MVT::i1) {
1400 default: llvm_unreachable("Unexpected Boolean-operand predicate");
1401 case PPC::PRED_LT: Opc = PPC::CRANDC; Swap = true; break;
1402 case PPC::PRED_LE: Opc = PPC::CRORC; Swap = true; break;
1403 case PPC::PRED_EQ: Opc = PPC::CREQV; Swap = false; break;
1404 case PPC::PRED_GE: Opc = PPC::CRORC; Swap = false; break;
1405 case PPC::PRED_GT: Opc = PPC::CRANDC; Swap = false; break;
1406 case PPC::PRED_NE: Opc = PPC::CRXOR; Swap = false; break;
1409 SDValue BitComp(CurDAG->getMachineNode(Opc, dl, MVT::i1,
1410 N->getOperand(Swap ? 3 : 2),
1411 N->getOperand(Swap ? 2 : 3)), 0);
1412 return CurDAG->SelectNodeTo(N, PPC::BC, MVT::Other,
1413 BitComp, N->getOperand(4), N->getOperand(0));
1416 SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl);
1417 SDValue Ops[] = { getI32Imm(PCC), CondCode,
1418 N->getOperand(4), N->getOperand(0) };
1419 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
1422 // FIXME: Should custom lower this.
1423 SDValue Chain = N->getOperand(0);
1424 SDValue Target = N->getOperand(1);
1425 unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8;
1426 unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8;
1427 Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Target,
1429 return CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain);
1431 case PPCISD::TOC_ENTRY: {
1432 if (PPCSubTarget->isSVR4ABI() && !PPCSubTarget->isPPC64()) {
1433 SDValue GA = N->getOperand(0);
1434 return CurDAG->getMachineNode(PPC::LWZtoc, dl, MVT::i32, GA,
1437 assert (PPCSubTarget->isPPC64() &&
1438 "Only supported for 64-bit ABI and 32-bit SVR4");
1440 // For medium and large code model, we generate two instructions as
1441 // described below. Otherwise we allow SelectCodeCommon to handle this,
1442 // selecting one of LDtoc, LDtocJTI, and LDtocCPT.
1443 CodeModel::Model CModel = TM.getCodeModel();
1444 if (CModel != CodeModel::Medium && CModel != CodeModel::Large)
1447 // The first source operand is a TargetGlobalAddress or a TargetJumpTable.
1448 // If it is an externally defined symbol, a symbol with common linkage,
1449 // a non-local function address, or a jump table address, or if we are
1450 // generating code for large code model, we generate:
1451 // LDtocL(<ga:@sym>, ADDIStocHA(%X2, <ga:@sym>))
1452 // Otherwise we generate:
1453 // ADDItocL(ADDIStocHA(%X2, <ga:@sym>), <ga:@sym>)
1454 SDValue GA = N->getOperand(0);
1455 SDValue TOCbase = N->getOperand(1);
1456 SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64,
1459 if (isa<JumpTableSDNode>(GA) || CModel == CodeModel::Large)
1460 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
1463 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) {
1464 const GlobalValue *GValue = G->getGlobal();
1465 if ((GValue->getType()->getElementType()->isFunctionTy() &&
1466 (GValue->isDeclaration() || GValue->isWeakForLinker())) ||
1467 GValue->isDeclaration() || GValue->hasCommonLinkage() ||
1468 GValue->hasAvailableExternallyLinkage())
1469 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
1473 return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64,
1474 SDValue(Tmp, 0), GA);
1476 case PPCISD::PPC32_PICGOT: {
1477 // Generate a PIC-safe GOT reference.
1478 assert(!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI() &&
1479 "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4");
1480 return CurDAG->SelectNodeTo(N, PPC::PPC32PICGOT, PPCLowering->getPointerTy(), MVT::i32);
1482 case PPCISD::VADD_SPLAT: {
1483 // This expands into one of three sequences, depending on whether
1484 // the first operand is odd or even, positive or negative.
1485 assert(isa<ConstantSDNode>(N->getOperand(0)) &&
1486 isa<ConstantSDNode>(N->getOperand(1)) &&
1487 "Invalid operand on VADD_SPLAT!");
1489 int Elt = N->getConstantOperandVal(0);
1490 int EltSize = N->getConstantOperandVal(1);
1491 unsigned Opc1, Opc2, Opc3;
1495 Opc1 = PPC::VSPLTISB;
1496 Opc2 = PPC::VADDUBM;
1497 Opc3 = PPC::VSUBUBM;
1499 } else if (EltSize == 2) {
1500 Opc1 = PPC::VSPLTISH;
1501 Opc2 = PPC::VADDUHM;
1502 Opc3 = PPC::VSUBUHM;
1505 assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!");
1506 Opc1 = PPC::VSPLTISW;
1507 Opc2 = PPC::VADDUWM;
1508 Opc3 = PPC::VSUBUWM;
1512 if ((Elt & 1) == 0) {
1513 // Elt is even, in the range [-32,-18] + [16,30].
1515 // Convert: VADD_SPLAT elt, size
1516 // Into: tmp = VSPLTIS[BHW] elt
1517 // VADDU[BHW]M tmp, tmp
1518 // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4
1519 SDValue EltVal = getI32Imm(Elt >> 1);
1520 SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1521 SDValue TmpVal = SDValue(Tmp, 0);
1522 return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal);
1524 } else if (Elt > 0) {
1525 // Elt is odd and positive, in the range [17,31].
1527 // Convert: VADD_SPLAT elt, size
1528 // Into: tmp1 = VSPLTIS[BHW] elt-16
1529 // tmp2 = VSPLTIS[BHW] -16
1530 // VSUBU[BHW]M tmp1, tmp2
1531 SDValue EltVal = getI32Imm(Elt - 16);
1532 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1533 EltVal = getI32Imm(-16);
1534 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1535 return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0),
1539 // Elt is odd and negative, in the range [-31,-17].
1541 // Convert: VADD_SPLAT elt, size
1542 // Into: tmp1 = VSPLTIS[BHW] elt+16
1543 // tmp2 = VSPLTIS[BHW] -16
1544 // VADDU[BHW]M tmp1, tmp2
1545 SDValue EltVal = getI32Imm(Elt + 16);
1546 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1547 EltVal = getI32Imm(-16);
1548 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1549 return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0),
1555 return SelectCode(N);
1558 /// PostprocessISelDAG - Perform some late peephole optimizations
1559 /// on the DAG representation.
1560 void PPCDAGToDAGISel::PostprocessISelDAG() {
1562 // Skip peepholes at -O0.
1563 if (TM.getOptLevel() == CodeGenOpt::None)
1570 // Check if all users of this node will become isel where the second operand
1571 // is the constant zero. If this is so, and if we can negate the condition,
1572 // then we can flip the true and false operands. This will allow the zero to
1573 // be folded with the isel so that we don't need to materialize a register
1575 bool PPCDAGToDAGISel::AllUsersSelectZero(SDNode *N) {
1576 // If we're not using isel, then this does not matter.
1577 if (!PPCSubTarget->hasISEL())
1580 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
1583 if (!User->isMachineOpcode())
1585 if (User->getMachineOpcode() != PPC::SELECT_I4 &&
1586 User->getMachineOpcode() != PPC::SELECT_I8)
1589 SDNode *Op2 = User->getOperand(2).getNode();
1590 if (!Op2->isMachineOpcode())
1593 if (Op2->getMachineOpcode() != PPC::LI &&
1594 Op2->getMachineOpcode() != PPC::LI8)
1597 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op2->getOperand(0));
1601 if (!C->isNullValue())
1608 void PPCDAGToDAGISel::SwapAllSelectUsers(SDNode *N) {
1609 SmallVector<SDNode *, 4> ToReplace;
1610 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
1613 assert((User->getMachineOpcode() == PPC::SELECT_I4 ||
1614 User->getMachineOpcode() == PPC::SELECT_I8) &&
1615 "Must have all select users");
1616 ToReplace.push_back(User);
1619 for (SmallVector<SDNode *, 4>::iterator UI = ToReplace.begin(),
1620 UE = ToReplace.end(); UI != UE; ++UI) {
1623 CurDAG->getMachineNode(User->getMachineOpcode(), SDLoc(User),
1624 User->getValueType(0), User->getOperand(0),
1625 User->getOperand(2),
1626 User->getOperand(1));
1628 DEBUG(dbgs() << "CR Peephole replacing:\nOld: ");
1629 DEBUG(User->dump(CurDAG));
1630 DEBUG(dbgs() << "\nNew: ");
1631 DEBUG(ResNode->dump(CurDAG));
1632 DEBUG(dbgs() << "\n");
1634 ReplaceUses(User, ResNode);
1638 void PPCDAGToDAGISel::PeepholeCROps() {
1642 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
1643 E = CurDAG->allnodes_end(); I != E; ++I) {
1644 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(I);
1645 if (!MachineNode || MachineNode->use_empty())
1647 SDNode *ResNode = MachineNode;
1649 bool Op1Set = false, Op1Unset = false,
1651 Op2Set = false, Op2Unset = false,
1654 unsigned Opcode = MachineNode->getMachineOpcode();
1665 SDValue Op = MachineNode->getOperand(1);
1666 if (Op.isMachineOpcode()) {
1667 if (Op.getMachineOpcode() == PPC::CRSET)
1669 else if (Op.getMachineOpcode() == PPC::CRUNSET)
1671 else if (Op.getMachineOpcode() == PPC::CRNOR &&
1672 Op.getOperand(0) == Op.getOperand(1))
1678 case PPC::SELECT_I4:
1679 case PPC::SELECT_I8:
1680 case PPC::SELECT_F4:
1681 case PPC::SELECT_F8:
1682 case PPC::SELECT_VRRC: {
1683 SDValue Op = MachineNode->getOperand(0);
1684 if (Op.isMachineOpcode()) {
1685 if (Op.getMachineOpcode() == PPC::CRSET)
1687 else if (Op.getMachineOpcode() == PPC::CRUNSET)
1689 else if (Op.getMachineOpcode() == PPC::CRNOR &&
1690 Op.getOperand(0) == Op.getOperand(1))
1697 bool SelectSwap = false;
1701 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1703 ResNode = MachineNode->getOperand(0).getNode();
1706 ResNode = MachineNode->getOperand(1).getNode();
1709 ResNode = MachineNode->getOperand(0).getNode();
1710 else if (Op1Unset || Op2Unset)
1711 // x & 0 = 0 & y = 0
1712 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
1715 // ~x & y = andc(y, x)
1716 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
1717 MVT::i1, MachineNode->getOperand(1),
1718 MachineNode->getOperand(0).
1721 // x & ~y = andc(x, y)
1722 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
1723 MVT::i1, MachineNode->getOperand(0),
1724 MachineNode->getOperand(1).
1726 else if (AllUsersSelectZero(MachineNode))
1727 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode),
1728 MVT::i1, MachineNode->getOperand(0),
1729 MachineNode->getOperand(1)),
1733 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1734 // nand(x, x) -> nor(x, x)
1735 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1736 MVT::i1, MachineNode->getOperand(0),
1737 MachineNode->getOperand(0));
1739 // nand(1, y) -> nor(y, y)
1740 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1741 MVT::i1, MachineNode->getOperand(1),
1742 MachineNode->getOperand(1));
1744 // nand(x, 1) -> nor(x, x)
1745 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1746 MVT::i1, MachineNode->getOperand(0),
1747 MachineNode->getOperand(0));
1748 else if (Op1Unset || Op2Unset)
1749 // nand(x, 0) = nand(0, y) = 1
1750 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
1753 // nand(~x, y) = ~(~x & y) = x | ~y = orc(x, y)
1754 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
1755 MVT::i1, MachineNode->getOperand(0).
1757 MachineNode->getOperand(1));
1759 // nand(x, ~y) = ~x | y = orc(y, x)
1760 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
1761 MVT::i1, MachineNode->getOperand(1).
1763 MachineNode->getOperand(0));
1764 else if (AllUsersSelectZero(MachineNode))
1765 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode),
1766 MVT::i1, MachineNode->getOperand(0),
1767 MachineNode->getOperand(1)),
1771 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1773 ResNode = MachineNode->getOperand(0).getNode();
1774 else if (Op1Set || Op2Set)
1775 // x | 1 = 1 | y = 1
1776 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
1780 ResNode = MachineNode->getOperand(1).getNode();
1783 ResNode = MachineNode->getOperand(0).getNode();
1785 // ~x | y = orc(y, x)
1786 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
1787 MVT::i1, MachineNode->getOperand(1),
1788 MachineNode->getOperand(0).
1791 // x | ~y = orc(x, y)
1792 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
1793 MVT::i1, MachineNode->getOperand(0),
1794 MachineNode->getOperand(1).
1796 else if (AllUsersSelectZero(MachineNode))
1797 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1798 MVT::i1, MachineNode->getOperand(0),
1799 MachineNode->getOperand(1)),
1803 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1805 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
1808 // xor(1, y) -> nor(y, y)
1809 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1810 MVT::i1, MachineNode->getOperand(1),
1811 MachineNode->getOperand(1));
1813 // xor(x, 1) -> nor(x, x)
1814 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1815 MVT::i1, MachineNode->getOperand(0),
1816 MachineNode->getOperand(0));
1819 ResNode = MachineNode->getOperand(1).getNode();
1822 ResNode = MachineNode->getOperand(0).getNode();
1824 // xor(~x, y) = eqv(x, y)
1825 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
1826 MVT::i1, MachineNode->getOperand(0).
1828 MachineNode->getOperand(1));
1830 // xor(x, ~y) = eqv(x, y)
1831 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
1832 MVT::i1, MachineNode->getOperand(0),
1833 MachineNode->getOperand(1).
1835 else if (AllUsersSelectZero(MachineNode))
1836 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
1837 MVT::i1, MachineNode->getOperand(0),
1838 MachineNode->getOperand(1)),
1842 if (Op1Set || Op2Set)
1844 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
1847 // nor(0, y) = ~y -> nor(y, y)
1848 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1849 MVT::i1, MachineNode->getOperand(1),
1850 MachineNode->getOperand(1));
1853 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1854 MVT::i1, MachineNode->getOperand(0),
1855 MachineNode->getOperand(0));
1857 // nor(~x, y) = andc(x, y)
1858 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
1859 MVT::i1, MachineNode->getOperand(0).
1861 MachineNode->getOperand(1));
1863 // nor(x, ~y) = andc(y, x)
1864 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
1865 MVT::i1, MachineNode->getOperand(1).
1867 MachineNode->getOperand(0));
1868 else if (AllUsersSelectZero(MachineNode))
1869 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode),
1870 MVT::i1, MachineNode->getOperand(0),
1871 MachineNode->getOperand(1)),
1875 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1877 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
1881 ResNode = MachineNode->getOperand(1).getNode();
1884 ResNode = MachineNode->getOperand(0).getNode();
1886 // eqv(0, y) = ~y -> nor(y, y)
1887 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1888 MVT::i1, MachineNode->getOperand(1),
1889 MachineNode->getOperand(1));
1892 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1893 MVT::i1, MachineNode->getOperand(0),
1894 MachineNode->getOperand(0));
1896 // eqv(~x, y) = xor(x, y)
1897 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
1898 MVT::i1, MachineNode->getOperand(0).
1900 MachineNode->getOperand(1));
1902 // eqv(x, ~y) = xor(x, y)
1903 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
1904 MVT::i1, MachineNode->getOperand(0),
1905 MachineNode->getOperand(1).
1907 else if (AllUsersSelectZero(MachineNode))
1908 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
1909 MVT::i1, MachineNode->getOperand(0),
1910 MachineNode->getOperand(1)),
1914 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1916 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
1920 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1921 MVT::i1, MachineNode->getOperand(1),
1922 MachineNode->getOperand(1));
1923 else if (Op1Unset || Op2Set)
1924 // andc(0, y) = andc(x, 1) = 0
1925 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
1929 ResNode = MachineNode->getOperand(0).getNode();
1931 // andc(~x, y) = ~(x | y) = nor(x, y)
1932 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1933 MVT::i1, MachineNode->getOperand(0).
1935 MachineNode->getOperand(1));
1937 // andc(x, ~y) = x & y
1938 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode),
1939 MVT::i1, MachineNode->getOperand(0),
1940 MachineNode->getOperand(1).
1942 else if (AllUsersSelectZero(MachineNode))
1943 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
1944 MVT::i1, MachineNode->getOperand(1),
1945 MachineNode->getOperand(0)),
1949 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1951 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
1953 else if (Op1Set || Op2Unset)
1954 // orc(1, y) = orc(x, 0) = 1
1955 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
1959 ResNode = MachineNode->getOperand(0).getNode();
1962 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1963 MVT::i1, MachineNode->getOperand(1),
1964 MachineNode->getOperand(1));
1966 // orc(~x, y) = ~(x & y) = nand(x, y)
1967 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode),
1968 MVT::i1, MachineNode->getOperand(0).
1970 MachineNode->getOperand(1));
1972 // orc(x, ~y) = x | y
1973 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode),
1974 MVT::i1, MachineNode->getOperand(0),
1975 MachineNode->getOperand(1).
1977 else if (AllUsersSelectZero(MachineNode))
1978 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
1979 MVT::i1, MachineNode->getOperand(1),
1980 MachineNode->getOperand(0)),
1983 case PPC::SELECT_I4:
1984 case PPC::SELECT_I8:
1985 case PPC::SELECT_F4:
1986 case PPC::SELECT_F8:
1987 case PPC::SELECT_VRRC:
1989 ResNode = MachineNode->getOperand(1).getNode();
1991 ResNode = MachineNode->getOperand(2).getNode();
1993 ResNode = CurDAG->getMachineNode(MachineNode->getMachineOpcode(),
1995 MachineNode->getValueType(0),
1996 MachineNode->getOperand(0).
1998 MachineNode->getOperand(2),
1999 MachineNode->getOperand(1));
2004 ResNode = CurDAG->getMachineNode(Opcode == PPC::BC ? PPC::BCn :
2008 MachineNode->getOperand(0).
2010 MachineNode->getOperand(1),
2011 MachineNode->getOperand(2));
2012 // FIXME: Handle Op1Set, Op1Unset here too.
2016 // If we're inverting this node because it is used only by selects that
2017 // we'd like to swap, then swap the selects before the node replacement.
2019 SwapAllSelectUsers(MachineNode);
2021 if (ResNode != MachineNode) {
2022 DEBUG(dbgs() << "CR Peephole replacing:\nOld: ");
2023 DEBUG(MachineNode->dump(CurDAG));
2024 DEBUG(dbgs() << "\nNew: ");
2025 DEBUG(ResNode->dump(CurDAG));
2026 DEBUG(dbgs() << "\n");
2028 ReplaceUses(MachineNode, ResNode);
2033 CurDAG->RemoveDeadNodes();
2034 } while (IsModified);
2037 void PPCDAGToDAGISel::PeepholePPC64() {
2038 // These optimizations are currently supported only for 64-bit SVR4.
2039 if (PPCSubTarget->isDarwin() || !PPCSubTarget->isPPC64())
2042 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
2045 while (Position != CurDAG->allnodes_begin()) {
2046 SDNode *N = --Position;
2047 // Skip dead nodes and any non-machine opcodes.
2048 if (N->use_empty() || !N->isMachineOpcode())
2052 unsigned StorageOpcode = N->getMachineOpcode();
2054 switch (StorageOpcode) {
2085 // If this is a load or store with a zero offset, we may be able to
2086 // fold an add-immediate into the memory operation.
2087 if (!isa<ConstantSDNode>(N->getOperand(FirstOp)) ||
2088 N->getConstantOperandVal(FirstOp) != 0)
2091 SDValue Base = N->getOperand(FirstOp + 1);
2092 if (!Base.isMachineOpcode())
2096 bool ReplaceFlags = true;
2098 // When the feeding operation is an add-immediate of some sort,
2099 // determine whether we need to add relocation information to the
2100 // target flags on the immediate operand when we fold it into the
2101 // load instruction.
2103 // For something like ADDItocL, the relocation information is
2104 // inferred from the opcode; when we process it in the AsmPrinter,
2105 // we add the necessary relocation there. A load, though, can receive
2106 // relocation from various flavors of ADDIxxx, so we need to carry
2107 // the relocation information in the target flags.
2108 switch (Base.getMachineOpcode()) {
2113 // In some cases (such as TLS) the relocation information
2114 // is already in place on the operand, so copying the operand
2116 ReplaceFlags = false;
2117 // For these cases, the immediate may not be divisible by 4, in
2118 // which case the fold is illegal for DS-form instructions. (The
2119 // other cases provide aligned addresses and are always safe.)
2120 if ((StorageOpcode == PPC::LWA ||
2121 StorageOpcode == PPC::LD ||
2122 StorageOpcode == PPC::STD) &&
2123 (!isa<ConstantSDNode>(Base.getOperand(1)) ||
2124 Base.getConstantOperandVal(1) % 4 != 0))
2127 case PPC::ADDIdtprelL:
2128 Flags = PPCII::MO_DTPREL_LO;
2130 case PPC::ADDItlsldL:
2131 Flags = PPCII::MO_TLSLD_LO;
2134 Flags = PPCII::MO_TOC_LO;
2138 // We found an opportunity. Reverse the operands from the add
2139 // immediate and substitute them into the load or store. If
2140 // needed, update the target flags for the immediate operand to
2141 // reflect the necessary relocation information.
2142 DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase: ");
2143 DEBUG(Base->dump(CurDAG));
2144 DEBUG(dbgs() << "\nN: ");
2145 DEBUG(N->dump(CurDAG));
2146 DEBUG(dbgs() << "\n");
2148 SDValue ImmOpnd = Base.getOperand(1);
2150 // If the relocation information isn't already present on the
2151 // immediate operand, add it now.
2153 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) {
2155 const GlobalValue *GV = GA->getGlobal();
2156 // We can't perform this optimization for data whose alignment
2157 // is insufficient for the instruction encoding.
2158 if (GV->getAlignment() < 4 &&
2159 (StorageOpcode == PPC::LD || StorageOpcode == PPC::STD ||
2160 StorageOpcode == PPC::LWA)) {
2161 DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n");
2164 ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, 0, Flags);
2165 } else if (ConstantPoolSDNode *CP =
2166 dyn_cast<ConstantPoolSDNode>(ImmOpnd)) {
2167 const Constant *C = CP->getConstVal();
2168 ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64,
2174 if (FirstOp == 1) // Store
2175 (void)CurDAG->UpdateNodeOperands(N, N->getOperand(0), ImmOpnd,
2176 Base.getOperand(0), N->getOperand(3));
2178 (void)CurDAG->UpdateNodeOperands(N, ImmOpnd, Base.getOperand(0),
2181 // The add-immediate may now be dead, in which case remove it.
2182 if (Base.getNode()->use_empty())
2183 CurDAG->RemoveDeadNode(Base.getNode());
2188 /// createPPCISelDag - This pass converts a legalized DAG into a
2189 /// PowerPC-specific DAG, ready for instruction scheduling.
2191 FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM) {
2192 return new PPCDAGToDAGISel(TM);
2195 static void initializePassOnce(PassRegistry &Registry) {
2196 const char *Name = "PowerPC DAG->DAG Pattern Instruction Selection";
2197 PassInfo *PI = new PassInfo(Name, "ppc-codegen", &SelectionDAGISel::ID,
2198 nullptr, false, false);
2199 Registry.registerPass(*PI, true);
2202 void llvm::initializePPCDAGToDAGISelPass(PassRegistry &Registry) {
2203 CALL_ONCE_INITIALIZATION(initializePassOnce);