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/IR/Module.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetOptions.h"
39 #define DEBUG_TYPE "ppc-codegen"
41 // FIXME: Remove this once the bug has been fixed!
42 cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug",
43 cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden);
46 void initializePPCDAGToDAGISelPass(PassRegistry&);
50 //===--------------------------------------------------------------------===//
51 /// PPCDAGToDAGISel - PPC specific code to select PPC machine
52 /// instructions for SelectionDAG operations.
54 class PPCDAGToDAGISel : public SelectionDAGISel {
55 const PPCTargetMachine &TM;
56 const PPCTargetLowering *PPCLowering;
57 const PPCSubtarget *PPCSubTarget;
58 unsigned GlobalBaseReg;
60 explicit PPCDAGToDAGISel(PPCTargetMachine &tm)
61 : SelectionDAGISel(tm), TM(tm),
62 PPCLowering(TM.getSubtargetImpl()->getTargetLowering()),
63 PPCSubTarget(TM.getSubtargetImpl()) {
64 initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry());
67 bool runOnMachineFunction(MachineFunction &MF) override {
68 // Make sure we re-emit a set of the global base reg if necessary
70 PPCLowering = TM.getSubtargetImpl()->getTargetLowering();
71 PPCSubTarget = TM.getSubtargetImpl();
72 SelectionDAGISel::runOnMachineFunction(MF);
74 if (!PPCSubTarget->isSVR4ABI())
80 void PostprocessISelDAG() override;
82 /// getI32Imm - Return a target constant with the specified value, of type
84 inline SDValue getI32Imm(unsigned Imm) {
85 return CurDAG->getTargetConstant(Imm, MVT::i32);
88 /// getI64Imm - Return a target constant with the specified value, of type
90 inline SDValue getI64Imm(uint64_t Imm) {
91 return CurDAG->getTargetConstant(Imm, MVT::i64);
94 /// getSmallIPtrImm - Return a target constant of pointer type.
95 inline SDValue getSmallIPtrImm(unsigned Imm) {
96 return CurDAG->getTargetConstant(Imm, PPCLowering->getPointerTy());
99 /// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s
100 /// with any number of 0s on either side. The 1s are allowed to wrap from
101 /// LSB to MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs.
102 /// 0x0F0F0000 is not, since all 1s are not contiguous.
103 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME);
106 /// isRotateAndMask - Returns true if Mask and Shift can be folded into a
107 /// rotate and mask opcode and mask operation.
108 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask,
109 unsigned &SH, unsigned &MB, unsigned &ME);
111 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
112 /// base register. Return the virtual register that holds this value.
113 SDNode *getGlobalBaseReg();
115 // Select - Convert the specified operand from a target-independent to a
116 // target-specific node if it hasn't already been changed.
117 SDNode *Select(SDNode *N) override;
119 SDNode *SelectBitfieldInsert(SDNode *N);
121 /// SelectCC - Select a comparison of the specified values with the
122 /// specified condition code, returning the CR# of the expression.
123 SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, SDLoc dl);
125 /// SelectAddrImm - Returns true if the address N can be represented by
126 /// a base register plus a signed 16-bit displacement [r+imm].
127 bool SelectAddrImm(SDValue N, SDValue &Disp,
129 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, false);
132 /// SelectAddrImmOffs - Return true if the operand is valid for a preinc
133 /// immediate field. Note that the operand at this point is already the
134 /// result of a prior SelectAddressRegImm call.
135 bool SelectAddrImmOffs(SDValue N, SDValue &Out) const {
136 if (N.getOpcode() == ISD::TargetConstant ||
137 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 /// SelectAddrImmX4 - Returns true if the address N can be represented by
159 /// a base register plus a signed 16-bit displacement that is a multiple of 4.
160 /// Suitable for use by STD and friends.
161 bool SelectAddrImmX4(SDValue N, SDValue &Disp, SDValue &Base) {
162 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, true);
165 // Select an address into a single register.
166 bool SelectAddr(SDValue N, SDValue &Base) {
171 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
172 /// inline asm expressions. It is always correct to compute the value into
173 /// a register. The case of adding a (possibly relocatable) constant to a
174 /// register can be improved, but it is wrong to substitute Reg+Reg for
175 /// Reg in an asm, because the load or store opcode would have to change.
176 bool SelectInlineAsmMemoryOperand(const SDValue &Op,
178 std::vector<SDValue> &OutOps) override {
179 OutOps.push_back(Op);
183 void InsertVRSaveCode(MachineFunction &MF);
185 const char *getPassName() const override {
186 return "PowerPC DAG->DAG Pattern Instruction Selection";
189 // Include the pieces autogenerated from the target description.
190 #include "PPCGenDAGISel.inc"
193 SDNode *SelectSETCC(SDNode *N);
195 void PeepholePPC64();
196 void PeepholeCROps();
198 bool AllUsersSelectZero(SDNode *N);
199 void SwapAllSelectUsers(SDNode *N);
203 /// InsertVRSaveCode - Once the entire function has been instruction selected,
204 /// all virtual registers are created and all machine instructions are built,
205 /// check to see if we need to save/restore VRSAVE. If so, do it.
206 void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) {
207 // Check to see if this function uses vector registers, which means we have to
208 // save and restore the VRSAVE register and update it with the regs we use.
210 // In this case, there will be virtual registers of vector type created
211 // by the scheduler. Detect them now.
212 bool HasVectorVReg = false;
213 for (unsigned i = 0, e = RegInfo->getNumVirtRegs(); i != e; ++i) {
214 unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
215 if (RegInfo->getRegClass(Reg) == &PPC::VRRCRegClass) {
216 HasVectorVReg = true;
220 if (!HasVectorVReg) return; // nothing to do.
222 // If we have a vector register, we want to emit code into the entry and exit
223 // blocks to save and restore the VRSAVE register. We do this here (instead
224 // of marking all vector instructions as clobbering VRSAVE) for two reasons:
226 // 1. This (trivially) reduces the load on the register allocator, by not
227 // having to represent the live range of the VRSAVE register.
228 // 2. This (more significantly) allows us to create a temporary virtual
229 // register to hold the saved VRSAVE value, allowing this temporary to be
230 // register allocated, instead of forcing it to be spilled to the stack.
232 // Create two vregs - one to hold the VRSAVE register that is live-in to the
233 // function and one for the value after having bits or'd into it.
234 unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
235 unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
237 const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
238 MachineBasicBlock &EntryBB = *Fn.begin();
240 // Emit the following code into the entry block:
241 // InVRSAVE = MFVRSAVE
242 // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE
243 // MTVRSAVE UpdatedVRSAVE
244 MachineBasicBlock::iterator IP = EntryBB.begin(); // Insert Point
245 BuildMI(EntryBB, IP, dl, TII.get(PPC::MFVRSAVE), InVRSAVE);
246 BuildMI(EntryBB, IP, dl, TII.get(PPC::UPDATE_VRSAVE),
247 UpdatedVRSAVE).addReg(InVRSAVE);
248 BuildMI(EntryBB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(UpdatedVRSAVE);
250 // Find all return blocks, outputting a restore in each epilog.
251 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
252 if (!BB->empty() && BB->back().isReturn()) {
253 IP = BB->end(); --IP;
255 // Skip over all terminator instructions, which are part of the return
257 MachineBasicBlock::iterator I2 = IP;
258 while (I2 != BB->begin() && (--I2)->isTerminator())
261 // Emit: MTVRSAVE InVRSave
262 BuildMI(*BB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(InVRSAVE);
268 /// getGlobalBaseReg - Output the instructions required to put the
269 /// base address to use for accessing globals into a register.
271 SDNode *PPCDAGToDAGISel::getGlobalBaseReg() {
272 if (!GlobalBaseReg) {
273 const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
274 // Insert the set of GlobalBaseReg into the first MBB of the function
275 MachineBasicBlock &FirstMBB = MF->front();
276 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
277 const Module *M = MF->getFunction()->getParent();
280 if (PPCLowering->getPointerTy() == MVT::i32) {
281 if (PPCSubTarget->isTargetELF()) {
282 GlobalBaseReg = PPC::R30;
283 if (M->getPICLevel() == PICLevel::Small) {
284 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MoveGOTtoLR));
285 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
287 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
288 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
289 unsigned TempReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
290 BuildMI(FirstMBB, MBBI, dl,
291 TII.get(PPC::UpdateGBR)).addReg(GlobalBaseReg)
292 .addReg(TempReg, RegState::Define).addReg(GlobalBaseReg);
293 MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true);
297 RegInfo->createVirtualRegister(&PPC::GPRC_NOR0RegClass);
298 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
299 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
302 GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RC_NOX0RegClass);
303 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8));
304 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg);
307 return CurDAG->getRegister(GlobalBaseReg,
308 PPCLowering->getPointerTy()).getNode();
311 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
312 /// or 64-bit immediate, and if the value can be accurately represented as a
313 /// sign extension from a 16-bit value. If so, this returns true and the
315 static bool isIntS16Immediate(SDNode *N, short &Imm) {
316 if (N->getOpcode() != ISD::Constant)
319 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
320 if (N->getValueType(0) == MVT::i32)
321 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
323 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
326 static bool isIntS16Immediate(SDValue Op, short &Imm) {
327 return isIntS16Immediate(Op.getNode(), Imm);
331 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
332 /// operand. If so Imm will receive the 32-bit value.
333 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
334 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
335 Imm = cast<ConstantSDNode>(N)->getZExtValue();
341 /// isInt64Immediate - This method tests to see if the node is a 64-bit constant
342 /// operand. If so Imm will receive the 64-bit value.
343 static bool isInt64Immediate(SDNode *N, uint64_t &Imm) {
344 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i64) {
345 Imm = cast<ConstantSDNode>(N)->getZExtValue();
351 // isInt32Immediate - This method tests to see if a constant operand.
352 // If so Imm will receive the 32 bit value.
353 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
354 return isInt32Immediate(N.getNode(), Imm);
358 // isOpcWithIntImmediate - This method tests to see if the node is a specific
359 // opcode and that it has a immediate integer right operand.
360 // If so Imm will receive the 32 bit value.
361 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
362 return N->getOpcode() == Opc
363 && isInt32Immediate(N->getOperand(1).getNode(), Imm);
366 bool PPCDAGToDAGISel::isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
370 if (isShiftedMask_32(Val)) {
371 // look for the first non-zero bit
372 MB = countLeadingZeros(Val);
373 // look for the first zero bit after the run of ones
374 ME = countLeadingZeros((Val - 1) ^ Val);
377 Val = ~Val; // invert mask
378 if (isShiftedMask_32(Val)) {
379 // effectively look for the first zero bit
380 ME = countLeadingZeros(Val) - 1;
381 // effectively look for the first one bit after the run of zeros
382 MB = countLeadingZeros((Val - 1) ^ Val) + 1;
390 bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask,
391 bool isShiftMask, unsigned &SH,
392 unsigned &MB, unsigned &ME) {
393 // Don't even go down this path for i64, since different logic will be
394 // necessary for rldicl/rldicr/rldimi.
395 if (N->getValueType(0) != MVT::i32)
399 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
400 unsigned Opcode = N->getOpcode();
401 if (N->getNumOperands() != 2 ||
402 !isInt32Immediate(N->getOperand(1).getNode(), Shift) || (Shift > 31))
405 if (Opcode == ISD::SHL) {
406 // apply shift left to mask if it comes first
407 if (isShiftMask) Mask = Mask << Shift;
408 // determine which bits are made indeterminant by shift
409 Indeterminant = ~(0xFFFFFFFFu << Shift);
410 } else if (Opcode == ISD::SRL) {
411 // apply shift right to mask if it comes first
412 if (isShiftMask) Mask = Mask >> Shift;
413 // determine which bits are made indeterminant by shift
414 Indeterminant = ~(0xFFFFFFFFu >> Shift);
415 // adjust for the left rotate
417 } else if (Opcode == ISD::ROTL) {
423 // if the mask doesn't intersect any Indeterminant bits
424 if (Mask && !(Mask & Indeterminant)) {
426 // make sure the mask is still a mask (wrap arounds may not be)
427 return isRunOfOnes(Mask, MB, ME);
432 /// SelectBitfieldInsert - turn an or of two masked values into
433 /// the rotate left word immediate then mask insert (rlwimi) instruction.
434 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
435 SDValue Op0 = N->getOperand(0);
436 SDValue Op1 = N->getOperand(1);
439 APInt LKZ, LKO, RKZ, RKO;
440 CurDAG->computeKnownBits(Op0, LKZ, LKO);
441 CurDAG->computeKnownBits(Op1, RKZ, RKO);
443 unsigned TargetMask = LKZ.getZExtValue();
444 unsigned InsertMask = RKZ.getZExtValue();
446 if ((TargetMask | InsertMask) == 0xFFFFFFFF) {
447 unsigned Op0Opc = Op0.getOpcode();
448 unsigned Op1Opc = Op1.getOpcode();
449 unsigned Value, SH = 0;
450 TargetMask = ~TargetMask;
451 InsertMask = ~InsertMask;
453 // If the LHS has a foldable shift and the RHS does not, then swap it to the
454 // RHS so that we can fold the shift into the insert.
455 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
456 if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
457 Op0.getOperand(0).getOpcode() == ISD::SRL) {
458 if (Op1.getOperand(0).getOpcode() != ISD::SHL &&
459 Op1.getOperand(0).getOpcode() != ISD::SRL) {
461 std::swap(Op0Opc, Op1Opc);
462 std::swap(TargetMask, InsertMask);
465 } else if (Op0Opc == ISD::SHL || Op0Opc == ISD::SRL) {
466 if (Op1Opc == ISD::AND && Op1.getOperand(0).getOpcode() != ISD::SHL &&
467 Op1.getOperand(0).getOpcode() != ISD::SRL) {
469 std::swap(Op0Opc, Op1Opc);
470 std::swap(TargetMask, InsertMask);
475 if (isRunOfOnes(InsertMask, MB, ME)) {
478 if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) &&
479 isInt32Immediate(Op1.getOperand(1), Value)) {
480 Op1 = Op1.getOperand(0);
481 SH = (Op1Opc == ISD::SHL) ? Value : 32 - Value;
483 if (Op1Opc == ISD::AND) {
484 // The AND mask might not be a constant, and we need to make sure that
485 // if we're going to fold the masking with the insert, all bits not
486 // know to be zero in the mask are known to be one.
488 CurDAG->computeKnownBits(Op1.getOperand(1), MKZ, MKO);
489 bool CanFoldMask = InsertMask == MKO.getZExtValue();
491 unsigned SHOpc = Op1.getOperand(0).getOpcode();
492 if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) && CanFoldMask &&
493 isInt32Immediate(Op1.getOperand(0).getOperand(1), Value)) {
494 // Note that Value must be in range here (less than 32) because
495 // otherwise there would not be any bits set in InsertMask.
496 Op1 = Op1.getOperand(0).getOperand(0);
497 SH = (SHOpc == ISD::SHL) ? Value : 32 - Value;
502 SDValue Ops[] = { Op0, Op1, getI32Imm(SH), getI32Imm(MB),
504 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
510 /// SelectCC - Select a comparison of the specified values with the specified
511 /// condition code, returning the CR# of the expression.
512 SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS,
513 ISD::CondCode CC, SDLoc dl) {
514 // Always select the LHS.
517 if (LHS.getValueType() == MVT::i32) {
519 if (CC == ISD::SETEQ || CC == ISD::SETNE) {
520 if (isInt32Immediate(RHS, Imm)) {
521 // SETEQ/SETNE comparison with 16-bit immediate, fold it.
523 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
524 getI32Imm(Imm & 0xFFFF)), 0);
525 // If this is a 16-bit signed immediate, fold it.
526 if (isInt<16>((int)Imm))
527 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
528 getI32Imm(Imm & 0xFFFF)), 0);
530 // For non-equality comparisons, the default code would materialize the
531 // constant, then compare against it, like this:
535 // Since we are just comparing for equality, we can emit this instead:
536 // xoris r0,r3,0x1234
537 // cmplwi cr0,r0,0x5678
539 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS,
540 getI32Imm(Imm >> 16)), 0);
541 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor,
542 getI32Imm(Imm & 0xFFFF)), 0);
545 } else if (ISD::isUnsignedIntSetCC(CC)) {
546 if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm))
547 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
548 getI32Imm(Imm & 0xFFFF)), 0);
552 if (isIntS16Immediate(RHS, SImm))
553 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
554 getI32Imm((int)SImm & 0xFFFF)),
558 } else if (LHS.getValueType() == MVT::i64) {
560 if (CC == ISD::SETEQ || CC == ISD::SETNE) {
561 if (isInt64Immediate(RHS.getNode(), Imm)) {
562 // SETEQ/SETNE comparison with 16-bit immediate, fold it.
564 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
565 getI32Imm(Imm & 0xFFFF)), 0);
566 // If this is a 16-bit signed immediate, fold it.
568 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
569 getI32Imm(Imm & 0xFFFF)), 0);
571 // For non-equality comparisons, the default code would materialize the
572 // constant, then compare against it, like this:
576 // Since we are just comparing for equality, we can emit this instead:
577 // xoris r0,r3,0x1234
578 // cmpldi cr0,r0,0x5678
580 if (isUInt<32>(Imm)) {
581 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS,
582 getI64Imm(Imm >> 16)), 0);
583 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor,
584 getI64Imm(Imm & 0xFFFF)), 0);
588 } else if (ISD::isUnsignedIntSetCC(CC)) {
589 if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm))
590 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
591 getI64Imm(Imm & 0xFFFF)), 0);
595 if (isIntS16Immediate(RHS, SImm))
596 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
597 getI64Imm(SImm & 0xFFFF)),
601 } else if (LHS.getValueType() == MVT::f32) {
604 assert(LHS.getValueType() == MVT::f64 && "Unknown vt!");
605 Opc = PPCSubTarget->hasVSX() ? PPC::XSCMPUDP : PPC::FCMPUD;
607 return SDValue(CurDAG->getMachineNode(Opc, dl, MVT::i32, LHS, RHS), 0);
610 static PPC::Predicate getPredicateForSetCC(ISD::CondCode CC) {
616 llvm_unreachable("Should be lowered by legalize!");
617 default: llvm_unreachable("Unknown condition!");
619 case ISD::SETEQ: return PPC::PRED_EQ;
621 case ISD::SETNE: return PPC::PRED_NE;
623 case ISD::SETLT: return PPC::PRED_LT;
625 case ISD::SETLE: return PPC::PRED_LE;
627 case ISD::SETGT: return PPC::PRED_GT;
629 case ISD::SETGE: return PPC::PRED_GE;
630 case ISD::SETO: return PPC::PRED_NU;
631 case ISD::SETUO: return PPC::PRED_UN;
632 // These two are invalid for floating point. Assume we have int.
633 case ISD::SETULT: return PPC::PRED_LT;
634 case ISD::SETUGT: return PPC::PRED_GT;
638 /// getCRIdxForSetCC - Return the index of the condition register field
639 /// associated with the SetCC condition, and whether or not the field is
640 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
641 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) {
644 default: llvm_unreachable("Unknown condition!");
646 case ISD::SETLT: return 0; // Bit #0 = SETOLT
648 case ISD::SETGT: return 1; // Bit #1 = SETOGT
650 case ISD::SETEQ: return 2; // Bit #2 = SETOEQ
651 case ISD::SETUO: return 3; // Bit #3 = SETUO
653 case ISD::SETGE: Invert = true; return 0; // !Bit #0 = SETUGE
655 case ISD::SETLE: Invert = true; return 1; // !Bit #1 = SETULE
657 case ISD::SETNE: Invert = true; return 2; // !Bit #2 = SETUNE
658 case ISD::SETO: Invert = true; return 3; // !Bit #3 = SETO
663 llvm_unreachable("Invalid branch code: should be expanded by legalize");
664 // These are invalid for floating point. Assume integer.
665 case ISD::SETULT: return 0;
666 case ISD::SETUGT: return 1;
670 // getVCmpInst: return the vector compare instruction for the specified
671 // vector type and condition code. Since this is for altivec specific code,
672 // only support the altivec types (v16i8, v8i16, v4i32, and v4f32).
673 static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC,
674 bool HasVSX, bool &Swap, bool &Negate) {
678 if (VecVT.isFloatingPoint()) {
679 /* Handle some cases by swapping input operands. */
681 case ISD::SETLE: CC = ISD::SETGE; Swap = true; break;
682 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break;
683 case ISD::SETOLE: CC = ISD::SETOGE; Swap = true; break;
684 case ISD::SETOLT: CC = ISD::SETOGT; Swap = true; break;
685 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break;
686 case ISD::SETUGT: CC = ISD::SETULT; Swap = true; break;
689 /* Handle some cases by negating the result. */
691 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break;
692 case ISD::SETUNE: CC = ISD::SETOEQ; Negate = true; break;
693 case ISD::SETULE: CC = ISD::SETOGT; Negate = true; break;
694 case ISD::SETULT: CC = ISD::SETOGE; Negate = true; break;
697 /* We have instructions implementing the remaining cases. */
701 if (VecVT == MVT::v4f32)
702 return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP;
703 else if (VecVT == MVT::v2f64)
704 return PPC::XVCMPEQDP;
708 if (VecVT == MVT::v4f32)
709 return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP;
710 else if (VecVT == MVT::v2f64)
711 return PPC::XVCMPGTDP;
715 if (VecVT == MVT::v4f32)
716 return HasVSX ? PPC::XVCMPGESP : PPC::VCMPGEFP;
717 else if (VecVT == MVT::v2f64)
718 return PPC::XVCMPGEDP;
723 llvm_unreachable("Invalid floating-point vector compare condition");
725 /* Handle some cases by swapping input operands. */
727 case ISD::SETGE: CC = ISD::SETLE; Swap = true; break;
728 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break;
729 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break;
730 case ISD::SETULT: CC = ISD::SETUGT; Swap = true; break;
733 /* Handle some cases by negating the result. */
735 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break;
736 case ISD::SETUNE: CC = ISD::SETUEQ; Negate = true; break;
737 case ISD::SETLE: CC = ISD::SETGT; Negate = true; break;
738 case ISD::SETULE: CC = ISD::SETUGT; Negate = true; break;
741 /* We have instructions implementing the remaining cases. */
745 if (VecVT == MVT::v16i8)
746 return PPC::VCMPEQUB;
747 else if (VecVT == MVT::v8i16)
748 return PPC::VCMPEQUH;
749 else if (VecVT == MVT::v4i32)
750 return PPC::VCMPEQUW;
753 if (VecVT == MVT::v16i8)
754 return PPC::VCMPGTSB;
755 else if (VecVT == MVT::v8i16)
756 return PPC::VCMPGTSH;
757 else if (VecVT == MVT::v4i32)
758 return PPC::VCMPGTSW;
761 if (VecVT == MVT::v16i8)
762 return PPC::VCMPGTUB;
763 else if (VecVT == MVT::v8i16)
764 return PPC::VCMPGTUH;
765 else if (VecVT == MVT::v4i32)
766 return PPC::VCMPGTUW;
771 llvm_unreachable("Invalid integer vector compare condition");
775 SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) {
778 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
779 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
780 bool isPPC64 = (PtrVT == MVT::i64);
782 if (!PPCSubTarget->useCRBits() &&
783 isInt32Immediate(N->getOperand(1), Imm)) {
784 // We can codegen setcc op, imm very efficiently compared to a brcond.
785 // Check for those cases here.
788 SDValue Op = N->getOperand(0);
792 Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0);
793 SDValue Ops[] = { Op, getI32Imm(27), getI32Imm(5), getI32Imm(31) };
794 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
799 SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
800 Op, getI32Imm(~0U)), 0);
801 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
805 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
806 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
810 SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0);
811 T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0);
812 SDValue Ops[] = { T, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
813 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
816 } else if (Imm == ~0U) { // setcc op, -1
817 SDValue Op = N->getOperand(0);
822 Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
823 Op, getI32Imm(1)), 0);
824 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
825 SDValue(CurDAG->getMachineNode(PPC::LI, dl,
831 Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0);
832 SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
834 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0),
838 SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op,
840 SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD,
842 SDValue Ops[] = { AN, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
843 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
846 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
847 Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops),
849 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op,
856 SDValue LHS = N->getOperand(0);
857 SDValue RHS = N->getOperand(1);
859 // Altivec Vector compare instructions do not set any CR register by default and
860 // vector compare operations return the same type as the operands.
861 if (LHS.getValueType().isVector()) {
862 EVT VecVT = LHS.getValueType();
864 unsigned int VCmpInst = getVCmpInst(VecVT.getSimpleVT(), CC,
865 PPCSubTarget->hasVSX(), Swap, Negate);
870 SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0);
871 return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLNOR :
876 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS);
879 if (PPCSubTarget->useCRBits())
883 unsigned Idx = getCRIdxForSetCC(CC, Inv);
884 SDValue CCReg = SelectCC(LHS, RHS, CC, dl);
887 // Force the ccreg into CR7.
888 SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
890 SDValue InFlag(nullptr, 0); // Null incoming flag value.
891 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg,
894 IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg,
897 SDValue Ops[] = { IntCR, getI32Imm((32-(3-Idx)) & 31),
898 getI32Imm(31), getI32Imm(31) };
900 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
902 // Get the specified bit.
904 SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
905 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
909 // Select - Convert the specified operand from a target-independent to a
910 // target-specific node if it hasn't already been changed.
911 SDNode *PPCDAGToDAGISel::Select(SDNode *N) {
913 if (N->isMachineOpcode()) {
915 return nullptr; // Already selected.
918 // In case any misguided DAG-level optimizations form an ADD with a
919 // TargetConstant operand, crash here instead of miscompiling (by selecting
920 // an r+r add instead of some kind of r+i add).
921 if (N->getOpcode() == ISD::ADD &&
922 N->getOperand(1).getOpcode() == ISD::TargetConstant)
923 llvm_unreachable("Invalid ADD with TargetConstant operand");
925 switch (N->getOpcode()) {
928 case ISD::Constant: {
929 if (N->getValueType(0) == MVT::i64) {
931 int64_t Imm = cast<ConstantSDNode>(N)->getZExtValue();
932 // Assume no remaining bits.
933 unsigned Remainder = 0;
934 // Assume no shift required.
937 // If it can't be represented as a 32 bit value.
938 if (!isInt<32>(Imm)) {
939 Shift = countTrailingZeros<uint64_t>(Imm);
940 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
942 // If the shifted value fits 32 bits.
943 if (isInt<32>(ImmSh)) {
944 // Go with the shifted value.
947 // Still stuck with a 64 bit value.
954 // Intermediate operand.
957 // Handle first 32 bits.
958 unsigned Lo = Imm & 0xFFFF;
959 unsigned Hi = (Imm >> 16) & 0xFFFF;
962 if (isInt<16>(Imm)) {
964 Result = CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(Lo));
966 // Handle the Hi bits.
967 unsigned OpC = Hi ? PPC::LIS8 : PPC::LI8;
968 Result = CurDAG->getMachineNode(OpC, dl, MVT::i64, getI32Imm(Hi));
970 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
971 SDValue(Result, 0), getI32Imm(Lo));
974 Result = CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(Hi));
977 // If no shift, we're done.
978 if (!Shift) return Result;
980 // Shift for next step if the upper 32-bits were not zero.
982 Result = CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64,
985 getI32Imm(63 - Shift));
988 // Add in the last bits as required.
989 if ((Hi = (Remainder >> 16) & 0xFFFF)) {
990 Result = CurDAG->getMachineNode(PPC::ORIS8, dl, MVT::i64,
991 SDValue(Result, 0), getI32Imm(Hi));
993 if ((Lo = Remainder & 0xFFFF)) {
994 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
995 SDValue(Result, 0), getI32Imm(Lo));
1004 SDNode *SN = SelectSETCC(N);
1009 case PPCISD::GlobalBaseReg:
1010 return getGlobalBaseReg();
1012 case ISD::FrameIndex: {
1013 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1014 SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0));
1015 unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8;
1017 return CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), TFI,
1018 getSmallIPtrImm(0));
1019 return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI,
1020 getSmallIPtrImm(0));
1023 case PPCISD::MFOCRF: {
1024 SDValue InFlag = N->getOperand(1);
1025 return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32,
1026 N->getOperand(0), InFlag);
1029 case PPCISD::READ_TIME_BASE: {
1030 return CurDAG->getMachineNode(PPC::ReadTB, dl, MVT::i32, MVT::i32,
1031 MVT::Other, N->getOperand(0));
1035 // FIXME: since this depends on the setting of the carry flag from the srawi
1036 // we should really be making notes about that for the scheduler.
1037 // FIXME: It sure would be nice if we could cheaply recognize the
1038 // srl/add/sra pattern the dag combiner will generate for this as
1039 // sra/addze rather than having to handle sdiv ourselves. oh well.
1041 if (isInt32Immediate(N->getOperand(1), Imm)) {
1042 SDValue N0 = N->getOperand(0);
1043 if ((signed)Imm > 0 && isPowerOf2_32(Imm)) {
1045 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
1046 N0, getI32Imm(Log2_32(Imm)));
1047 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
1048 SDValue(Op, 0), SDValue(Op, 1));
1049 } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) {
1051 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
1052 N0, getI32Imm(Log2_32(-Imm)));
1054 SDValue(CurDAG->getMachineNode(PPC::ADDZE, dl, MVT::i32,
1055 SDValue(Op, 0), SDValue(Op, 1)),
1057 return CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT);
1061 // Other cases are autogenerated.
1066 // Handle preincrement loads.
1067 LoadSDNode *LD = cast<LoadSDNode>(N);
1068 EVT LoadedVT = LD->getMemoryVT();
1070 // Normal loads are handled by code generated from the .td file.
1071 if (LD->getAddressingMode() != ISD::PRE_INC)
1074 SDValue Offset = LD->getOffset();
1075 if (Offset.getOpcode() == ISD::TargetConstant ||
1076 Offset.getOpcode() == ISD::TargetGlobalAddress) {
1079 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
1080 if (LD->getValueType(0) != MVT::i64) {
1081 // Handle PPC32 integer and normal FP loads.
1082 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
1083 switch (LoadedVT.getSimpleVT().SimpleTy) {
1084 default: llvm_unreachable("Invalid PPC load type!");
1085 case MVT::f64: Opcode = PPC::LFDU; break;
1086 case MVT::f32: Opcode = PPC::LFSU; break;
1087 case MVT::i32: Opcode = PPC::LWZU; break;
1088 case MVT::i16: Opcode = isSExt ? PPC::LHAU : PPC::LHZU; break;
1090 case MVT::i8: Opcode = PPC::LBZU; break;
1093 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
1094 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
1095 switch (LoadedVT.getSimpleVT().SimpleTy) {
1096 default: llvm_unreachable("Invalid PPC load type!");
1097 case MVT::i64: Opcode = PPC::LDU; break;
1098 case MVT::i32: Opcode = PPC::LWZU8; break;
1099 case MVT::i16: Opcode = isSExt ? PPC::LHAU8 : PPC::LHZU8; break;
1101 case MVT::i8: Opcode = PPC::LBZU8; break;
1105 SDValue Chain = LD->getChain();
1106 SDValue Base = LD->getBasePtr();
1107 SDValue Ops[] = { Offset, Base, Chain };
1108 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
1109 PPCLowering->getPointerTy(),
1113 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
1114 if (LD->getValueType(0) != MVT::i64) {
1115 // Handle PPC32 integer and normal FP loads.
1116 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
1117 switch (LoadedVT.getSimpleVT().SimpleTy) {
1118 default: llvm_unreachable("Invalid PPC load type!");
1119 case MVT::f64: Opcode = PPC::LFDUX; break;
1120 case MVT::f32: Opcode = PPC::LFSUX; break;
1121 case MVT::i32: Opcode = PPC::LWZUX; break;
1122 case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break;
1124 case MVT::i8: Opcode = PPC::LBZUX; break;
1127 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
1128 assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) &&
1129 "Invalid sext update load");
1130 switch (LoadedVT.getSimpleVT().SimpleTy) {
1131 default: llvm_unreachable("Invalid PPC load type!");
1132 case MVT::i64: Opcode = PPC::LDUX; break;
1133 case MVT::i32: Opcode = isSExt ? PPC::LWAUX : PPC::LWZUX8; break;
1134 case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break;
1136 case MVT::i8: Opcode = PPC::LBZUX8; break;
1140 SDValue Chain = LD->getChain();
1141 SDValue Base = LD->getBasePtr();
1142 SDValue Ops[] = { Base, Offset, Chain };
1143 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
1144 PPCLowering->getPointerTy(),
1150 unsigned Imm, Imm2, SH, MB, ME;
1153 // If this is an and of a value rotated between 0 and 31 bits and then and'd
1154 // with a mask, emit rlwinm
1155 if (isInt32Immediate(N->getOperand(1), Imm) &&
1156 isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) {
1157 SDValue Val = N->getOperand(0).getOperand(0);
1158 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
1159 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
1161 // If this is just a masked value where the input is not handled above, and
1162 // is not a rotate-left (handled by a pattern in the .td file), emit rlwinm
1163 if (isInt32Immediate(N->getOperand(1), Imm) &&
1164 isRunOfOnes(Imm, MB, ME) &&
1165 N->getOperand(0).getOpcode() != ISD::ROTL) {
1166 SDValue Val = N->getOperand(0);
1167 SDValue Ops[] = { Val, getI32Imm(0), getI32Imm(MB), getI32Imm(ME) };
1168 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
1170 // If this is a 64-bit zero-extension mask, emit rldicl.
1171 if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) &&
1173 SDValue Val = N->getOperand(0);
1174 MB = 64 - CountTrailingOnes_64(Imm64);
1177 // If the operand is a logical right shift, we can fold it into this
1178 // instruction: rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb)
1179 // for n <= mb. The right shift is really a left rotate followed by a
1180 // mask, and this mask is a more-restrictive sub-mask of the mask implied
1182 if (Val.getOpcode() == ISD::SRL &&
1183 isInt32Immediate(Val.getOperand(1).getNode(), Imm) && Imm <= MB) {
1184 assert(Imm < 64 && "Illegal shift amount");
1185 Val = Val.getOperand(0);
1189 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB) };
1190 return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops);
1192 // AND X, 0 -> 0, not "rlwinm 32".
1193 if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) {
1194 ReplaceUses(SDValue(N, 0), N->getOperand(1));
1197 // ISD::OR doesn't get all the bitfield insertion fun.
1198 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert
1199 if (isInt32Immediate(N->getOperand(1), Imm) &&
1200 N->getOperand(0).getOpcode() == ISD::OR &&
1201 isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) {
1204 if (isRunOfOnes(Imm, MB, ME)) {
1205 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1206 N->getOperand(0).getOperand(1),
1207 getI32Imm(0), getI32Imm(MB),getI32Imm(ME) };
1208 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
1212 // Other cases are autogenerated.
1216 if (N->getValueType(0) == MVT::i32)
1217 if (SDNode *I = SelectBitfieldInsert(N))
1220 // Other cases are autogenerated.
1223 unsigned Imm, SH, MB, ME;
1224 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
1225 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
1226 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1227 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
1228 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
1231 // Other cases are autogenerated.
1235 unsigned Imm, SH, MB, ME;
1236 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
1237 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
1238 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1239 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
1240 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
1243 // Other cases are autogenerated.
1246 // FIXME: Remove this once the ANDI glue bug is fixed:
1247 case PPCISD::ANDIo_1_EQ_BIT:
1248 case PPCISD::ANDIo_1_GT_BIT: {
1252 EVT InVT = N->getOperand(0).getValueType();
1253 assert((InVT == MVT::i64 || InVT == MVT::i32) &&
1254 "Invalid input type for ANDIo_1_EQ_BIT");
1256 unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDIo8 : PPC::ANDIo;
1257 SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue,
1259 CurDAG->getTargetConstant(1, InVT)), 0);
1260 SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32);
1262 CurDAG->getTargetConstant(N->getOpcode() == PPCISD::ANDIo_1_EQ_BIT ?
1263 PPC::sub_eq : PPC::sub_gt, MVT::i32);
1265 return CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1,
1267 SDValue(AndI.getNode(), 1) /* glue */);
1269 case ISD::SELECT_CC: {
1270 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1271 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
1272 bool isPPC64 = (PtrVT == MVT::i64);
1274 // If this is a select of i1 operands, we'll pattern match it.
1275 if (PPCSubTarget->useCRBits() &&
1276 N->getOperand(0).getValueType() == MVT::i1)
1279 // Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
1281 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1282 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
1283 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
1284 if (N1C->isNullValue() && N3C->isNullValue() &&
1285 N2C->getZExtValue() == 1ULL && CC == ISD::SETNE &&
1286 // FIXME: Implement this optzn for PPC64.
1287 N->getValueType(0) == MVT::i32) {
1289 CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
1290 N->getOperand(0), getI32Imm(~0U));
1291 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32,
1292 SDValue(Tmp, 0), N->getOperand(0),
1296 SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl);
1298 if (N->getValueType(0) == MVT::i1) {
1299 // An i1 select is: (c & t) | (!c & f).
1301 unsigned Idx = getCRIdxForSetCC(CC, Inv);
1305 default: llvm_unreachable("Invalid CC index");
1306 case 0: SRI = PPC::sub_lt; break;
1307 case 1: SRI = PPC::sub_gt; break;
1308 case 2: SRI = PPC::sub_eq; break;
1309 case 3: SRI = PPC::sub_un; break;
1312 SDValue CCBit = CurDAG->getTargetExtractSubreg(SRI, dl, MVT::i1, CCReg);
1314 SDValue NotCCBit(CurDAG->getMachineNode(PPC::CRNOR, dl, MVT::i1,
1316 SDValue C = Inv ? NotCCBit : CCBit,
1317 NotC = Inv ? CCBit : NotCCBit;
1319 SDValue CAndT(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1,
1320 C, N->getOperand(2)), 0);
1321 SDValue NotCAndF(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1,
1322 NotC, N->getOperand(3)), 0);
1324 return CurDAG->SelectNodeTo(N, PPC::CROR, MVT::i1, CAndT, NotCAndF);
1327 unsigned BROpc = getPredicateForSetCC(CC);
1329 unsigned SelectCCOp;
1330 if (N->getValueType(0) == MVT::i32)
1331 SelectCCOp = PPC::SELECT_CC_I4;
1332 else if (N->getValueType(0) == MVT::i64)
1333 SelectCCOp = PPC::SELECT_CC_I8;
1334 else if (N->getValueType(0) == MVT::f32)
1335 SelectCCOp = PPC::SELECT_CC_F4;
1336 else if (N->getValueType(0) == MVT::f64)
1337 if (PPCSubTarget->hasVSX())
1338 SelectCCOp = PPC::SELECT_CC_VSFRC;
1340 SelectCCOp = PPC::SELECT_CC_F8;
1341 else if (N->getValueType(0) == MVT::v2f64 ||
1342 N->getValueType(0) == MVT::v2i64)
1343 SelectCCOp = PPC::SELECT_CC_VSRC;
1345 SelectCCOp = PPC::SELECT_CC_VRRC;
1347 SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3),
1349 return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops);
1352 if (PPCSubTarget->hasVSX()) {
1353 SDValue Ops[] = { N->getOperand(2), N->getOperand(1), N->getOperand(0) };
1354 return CurDAG->SelectNodeTo(N, PPC::XXSEL, N->getValueType(0), Ops);
1358 case ISD::VECTOR_SHUFFLE:
1359 if (PPCSubTarget->hasVSX() && (N->getValueType(0) == MVT::v2f64 ||
1360 N->getValueType(0) == MVT::v2i64)) {
1361 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
1363 SDValue Op1 = N->getOperand(SVN->getMaskElt(0) < 2 ? 0 : 1),
1364 Op2 = N->getOperand(SVN->getMaskElt(1) < 2 ? 0 : 1);
1367 for (int i = 0; i < 2; ++i)
1368 if (SVN->getMaskElt(i) <= 0 || SVN->getMaskElt(i) == 2)
1373 SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), MVT::i32);
1375 if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 &&
1376 Op1.getOpcode() == ISD::SCALAR_TO_VECTOR &&
1377 isa<LoadSDNode>(Op1.getOperand(0))) {
1378 LoadSDNode *LD = cast<LoadSDNode>(Op1.getOperand(0));
1379 SDValue Base, Offset;
1381 if (LD->isUnindexed() &&
1382 SelectAddrIdxOnly(LD->getBasePtr(), Base, Offset)) {
1383 SDValue Chain = LD->getChain();
1384 SDValue Ops[] = { Base, Offset, Chain };
1385 return CurDAG->SelectNodeTo(N, PPC::LXVDSX,
1386 N->getValueType(0), Ops);
1390 SDValue Ops[] = { Op1, Op2, DMV };
1391 return CurDAG->SelectNodeTo(N, PPC::XXPERMDI, N->getValueType(0), Ops);
1397 bool IsPPC64 = PPCSubTarget->isPPC64();
1398 SDValue Ops[] = { N->getOperand(1), N->getOperand(0) };
1399 return CurDAG->SelectNodeTo(N, N->getOpcode() == PPCISD::BDNZ ?
1400 (IsPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
1401 (IsPPC64 ? PPC::BDZ8 : PPC::BDZ),
1404 case PPCISD::COND_BRANCH: {
1405 // Op #0 is the Chain.
1406 // Op #1 is the PPC::PRED_* number.
1408 // Op #3 is the Dest MBB
1409 // Op #4 is the Flag.
1410 // Prevent PPC::PRED_* from being selected into LI.
1412 getI32Imm(cast<ConstantSDNode>(N->getOperand(1))->getZExtValue());
1413 SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3),
1414 N->getOperand(0), N->getOperand(4) };
1415 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
1418 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
1419 unsigned PCC = getPredicateForSetCC(CC);
1421 if (N->getOperand(2).getValueType() == MVT::i1) {
1425 default: llvm_unreachable("Unexpected Boolean-operand predicate");
1426 case PPC::PRED_LT: Opc = PPC::CRANDC; Swap = true; break;
1427 case PPC::PRED_LE: Opc = PPC::CRORC; Swap = true; break;
1428 case PPC::PRED_EQ: Opc = PPC::CREQV; Swap = false; break;
1429 case PPC::PRED_GE: Opc = PPC::CRORC; Swap = false; break;
1430 case PPC::PRED_GT: Opc = PPC::CRANDC; Swap = false; break;
1431 case PPC::PRED_NE: Opc = PPC::CRXOR; Swap = false; break;
1434 SDValue BitComp(CurDAG->getMachineNode(Opc, dl, MVT::i1,
1435 N->getOperand(Swap ? 3 : 2),
1436 N->getOperand(Swap ? 2 : 3)), 0);
1437 return CurDAG->SelectNodeTo(N, PPC::BC, MVT::Other,
1438 BitComp, N->getOperand(4), N->getOperand(0));
1441 SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl);
1442 SDValue Ops[] = { getI32Imm(PCC), CondCode,
1443 N->getOperand(4), N->getOperand(0) };
1444 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
1447 // FIXME: Should custom lower this.
1448 SDValue Chain = N->getOperand(0);
1449 SDValue Target = N->getOperand(1);
1450 unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8;
1451 unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8;
1452 Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Target,
1454 return CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain);
1456 case PPCISD::TOC_ENTRY: {
1457 assert ((PPCSubTarget->isPPC64() || PPCSubTarget->isSVR4ABI()) &&
1458 "Only supported for 64-bit ABI and 32-bit SVR4");
1459 if (PPCSubTarget->isSVR4ABI() && !PPCSubTarget->isPPC64()) {
1460 SDValue GA = N->getOperand(0);
1461 return CurDAG->getMachineNode(PPC::LWZtoc, dl, MVT::i32, GA,
1465 // For medium and large code model, we generate two instructions as
1466 // described below. Otherwise we allow SelectCodeCommon to handle this,
1467 // selecting one of LDtoc, LDtocJTI, LDtocCPT, and LDtocBA.
1468 CodeModel::Model CModel = TM.getCodeModel();
1469 if (CModel != CodeModel::Medium && CModel != CodeModel::Large)
1472 // The first source operand is a TargetGlobalAddress or a TargetJumpTable.
1473 // If it is an externally defined symbol, a symbol with common linkage,
1474 // a non-local function address, or a jump table address, or if we are
1475 // generating code for large code model, we generate:
1476 // LDtocL(<ga:@sym>, ADDIStocHA(%X2, <ga:@sym>))
1477 // Otherwise we generate:
1478 // ADDItocL(ADDIStocHA(%X2, <ga:@sym>), <ga:@sym>)
1479 SDValue GA = N->getOperand(0);
1480 SDValue TOCbase = N->getOperand(1);
1481 SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64,
1484 if (isa<JumpTableSDNode>(GA) || isa<BlockAddressSDNode>(GA) ||
1485 CModel == CodeModel::Large)
1486 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
1489 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) {
1490 const GlobalValue *GValue = G->getGlobal();
1491 if ((GValue->getType()->getElementType()->isFunctionTy() &&
1492 (GValue->isDeclaration() || GValue->isWeakForLinker())) ||
1493 GValue->isDeclaration() || GValue->hasCommonLinkage() ||
1494 GValue->hasAvailableExternallyLinkage())
1495 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
1499 return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64,
1500 SDValue(Tmp, 0), GA);
1502 case PPCISD::PPC32_PICGOT: {
1503 // Generate a PIC-safe GOT reference.
1504 assert(!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI() &&
1505 "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4");
1506 return CurDAG->SelectNodeTo(N, PPC::PPC32PICGOT, PPCLowering->getPointerTy(), MVT::i32);
1508 case PPCISD::VADD_SPLAT: {
1509 // This expands into one of three sequences, depending on whether
1510 // the first operand is odd or even, positive or negative.
1511 assert(isa<ConstantSDNode>(N->getOperand(0)) &&
1512 isa<ConstantSDNode>(N->getOperand(1)) &&
1513 "Invalid operand on VADD_SPLAT!");
1515 int Elt = N->getConstantOperandVal(0);
1516 int EltSize = N->getConstantOperandVal(1);
1517 unsigned Opc1, Opc2, Opc3;
1521 Opc1 = PPC::VSPLTISB;
1522 Opc2 = PPC::VADDUBM;
1523 Opc3 = PPC::VSUBUBM;
1525 } else if (EltSize == 2) {
1526 Opc1 = PPC::VSPLTISH;
1527 Opc2 = PPC::VADDUHM;
1528 Opc3 = PPC::VSUBUHM;
1531 assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!");
1532 Opc1 = PPC::VSPLTISW;
1533 Opc2 = PPC::VADDUWM;
1534 Opc3 = PPC::VSUBUWM;
1538 if ((Elt & 1) == 0) {
1539 // Elt is even, in the range [-32,-18] + [16,30].
1541 // Convert: VADD_SPLAT elt, size
1542 // Into: tmp = VSPLTIS[BHW] elt
1543 // VADDU[BHW]M tmp, tmp
1544 // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4
1545 SDValue EltVal = getI32Imm(Elt >> 1);
1546 SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1547 SDValue TmpVal = SDValue(Tmp, 0);
1548 return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal);
1550 } else if (Elt > 0) {
1551 // Elt is odd and positive, in the range [17,31].
1553 // Convert: VADD_SPLAT elt, size
1554 // Into: tmp1 = VSPLTIS[BHW] elt-16
1555 // tmp2 = VSPLTIS[BHW] -16
1556 // VSUBU[BHW]M tmp1, tmp2
1557 SDValue EltVal = getI32Imm(Elt - 16);
1558 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1559 EltVal = getI32Imm(-16);
1560 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1561 return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0),
1565 // Elt is odd and negative, in the range [-31,-17].
1567 // Convert: VADD_SPLAT elt, size
1568 // Into: tmp1 = VSPLTIS[BHW] elt+16
1569 // tmp2 = VSPLTIS[BHW] -16
1570 // VADDU[BHW]M tmp1, tmp2
1571 SDValue EltVal = getI32Imm(Elt + 16);
1572 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1573 EltVal = getI32Imm(-16);
1574 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
1575 return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0),
1581 return SelectCode(N);
1584 /// PostprocessISelDAG - Perform some late peephole optimizations
1585 /// on the DAG representation.
1586 void PPCDAGToDAGISel::PostprocessISelDAG() {
1588 // Skip peepholes at -O0.
1589 if (TM.getOptLevel() == CodeGenOpt::None)
1596 // Check if all users of this node will become isel where the second operand
1597 // is the constant zero. If this is so, and if we can negate the condition,
1598 // then we can flip the true and false operands. This will allow the zero to
1599 // be folded with the isel so that we don't need to materialize a register
1601 bool PPCDAGToDAGISel::AllUsersSelectZero(SDNode *N) {
1602 // If we're not using isel, then this does not matter.
1603 if (!PPCSubTarget->hasISEL())
1606 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
1609 if (!User->isMachineOpcode())
1611 if (User->getMachineOpcode() != PPC::SELECT_I4 &&
1612 User->getMachineOpcode() != PPC::SELECT_I8)
1615 SDNode *Op2 = User->getOperand(2).getNode();
1616 if (!Op2->isMachineOpcode())
1619 if (Op2->getMachineOpcode() != PPC::LI &&
1620 Op2->getMachineOpcode() != PPC::LI8)
1623 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op2->getOperand(0));
1627 if (!C->isNullValue())
1634 void PPCDAGToDAGISel::SwapAllSelectUsers(SDNode *N) {
1635 SmallVector<SDNode *, 4> ToReplace;
1636 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
1639 assert((User->getMachineOpcode() == PPC::SELECT_I4 ||
1640 User->getMachineOpcode() == PPC::SELECT_I8) &&
1641 "Must have all select users");
1642 ToReplace.push_back(User);
1645 for (SmallVector<SDNode *, 4>::iterator UI = ToReplace.begin(),
1646 UE = ToReplace.end(); UI != UE; ++UI) {
1649 CurDAG->getMachineNode(User->getMachineOpcode(), SDLoc(User),
1650 User->getValueType(0), User->getOperand(0),
1651 User->getOperand(2),
1652 User->getOperand(1));
1654 DEBUG(dbgs() << "CR Peephole replacing:\nOld: ");
1655 DEBUG(User->dump(CurDAG));
1656 DEBUG(dbgs() << "\nNew: ");
1657 DEBUG(ResNode->dump(CurDAG));
1658 DEBUG(dbgs() << "\n");
1660 ReplaceUses(User, ResNode);
1664 void PPCDAGToDAGISel::PeepholeCROps() {
1668 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
1669 E = CurDAG->allnodes_end(); I != E; ++I) {
1670 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(I);
1671 if (!MachineNode || MachineNode->use_empty())
1673 SDNode *ResNode = MachineNode;
1675 bool Op1Set = false, Op1Unset = false,
1677 Op2Set = false, Op2Unset = false,
1680 unsigned Opcode = MachineNode->getMachineOpcode();
1691 SDValue Op = MachineNode->getOperand(1);
1692 if (Op.isMachineOpcode()) {
1693 if (Op.getMachineOpcode() == PPC::CRSET)
1695 else if (Op.getMachineOpcode() == PPC::CRUNSET)
1697 else if (Op.getMachineOpcode() == PPC::CRNOR &&
1698 Op.getOperand(0) == Op.getOperand(1))
1704 case PPC::SELECT_I4:
1705 case PPC::SELECT_I8:
1706 case PPC::SELECT_F4:
1707 case PPC::SELECT_F8:
1708 case PPC::SELECT_VRRC:
1709 case PPC::SELECT_VSFRC:
1710 case PPC::SELECT_VSRC: {
1711 SDValue Op = MachineNode->getOperand(0);
1712 if (Op.isMachineOpcode()) {
1713 if (Op.getMachineOpcode() == PPC::CRSET)
1715 else if (Op.getMachineOpcode() == PPC::CRUNSET)
1717 else if (Op.getMachineOpcode() == PPC::CRNOR &&
1718 Op.getOperand(0) == Op.getOperand(1))
1725 bool SelectSwap = false;
1729 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1731 ResNode = MachineNode->getOperand(0).getNode();
1734 ResNode = MachineNode->getOperand(1).getNode();
1737 ResNode = MachineNode->getOperand(0).getNode();
1738 else if (Op1Unset || Op2Unset)
1739 // x & 0 = 0 & y = 0
1740 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
1743 // ~x & y = andc(y, x)
1744 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
1745 MVT::i1, MachineNode->getOperand(1),
1746 MachineNode->getOperand(0).
1749 // x & ~y = andc(x, y)
1750 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
1751 MVT::i1, MachineNode->getOperand(0),
1752 MachineNode->getOperand(1).
1754 else if (AllUsersSelectZero(MachineNode))
1755 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode),
1756 MVT::i1, MachineNode->getOperand(0),
1757 MachineNode->getOperand(1)),
1761 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1762 // nand(x, x) -> nor(x, x)
1763 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1764 MVT::i1, MachineNode->getOperand(0),
1765 MachineNode->getOperand(0));
1767 // nand(1, y) -> nor(y, y)
1768 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1769 MVT::i1, MachineNode->getOperand(1),
1770 MachineNode->getOperand(1));
1772 // nand(x, 1) -> nor(x, x)
1773 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1774 MVT::i1, MachineNode->getOperand(0),
1775 MachineNode->getOperand(0));
1776 else if (Op1Unset || Op2Unset)
1777 // nand(x, 0) = nand(0, y) = 1
1778 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
1781 // nand(~x, y) = ~(~x & y) = x | ~y = orc(x, y)
1782 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
1783 MVT::i1, MachineNode->getOperand(0).
1785 MachineNode->getOperand(1));
1787 // nand(x, ~y) = ~x | y = orc(y, x)
1788 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
1789 MVT::i1, MachineNode->getOperand(1).
1791 MachineNode->getOperand(0));
1792 else if (AllUsersSelectZero(MachineNode))
1793 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode),
1794 MVT::i1, MachineNode->getOperand(0),
1795 MachineNode->getOperand(1)),
1799 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1801 ResNode = MachineNode->getOperand(0).getNode();
1802 else if (Op1Set || Op2Set)
1803 // x | 1 = 1 | y = 1
1804 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
1808 ResNode = MachineNode->getOperand(1).getNode();
1811 ResNode = MachineNode->getOperand(0).getNode();
1813 // ~x | y = orc(y, x)
1814 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
1815 MVT::i1, MachineNode->getOperand(1),
1816 MachineNode->getOperand(0).
1819 // x | ~y = orc(x, y)
1820 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
1821 MVT::i1, MachineNode->getOperand(0),
1822 MachineNode->getOperand(1).
1824 else if (AllUsersSelectZero(MachineNode))
1825 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1826 MVT::i1, MachineNode->getOperand(0),
1827 MachineNode->getOperand(1)),
1831 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1833 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
1836 // xor(1, y) -> nor(y, y)
1837 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1838 MVT::i1, MachineNode->getOperand(1),
1839 MachineNode->getOperand(1));
1841 // xor(x, 1) -> nor(x, x)
1842 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1843 MVT::i1, MachineNode->getOperand(0),
1844 MachineNode->getOperand(0));
1847 ResNode = MachineNode->getOperand(1).getNode();
1850 ResNode = MachineNode->getOperand(0).getNode();
1852 // xor(~x, y) = eqv(x, y)
1853 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
1854 MVT::i1, MachineNode->getOperand(0).
1856 MachineNode->getOperand(1));
1858 // xor(x, ~y) = eqv(x, y)
1859 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
1860 MVT::i1, MachineNode->getOperand(0),
1861 MachineNode->getOperand(1).
1863 else if (AllUsersSelectZero(MachineNode))
1864 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
1865 MVT::i1, MachineNode->getOperand(0),
1866 MachineNode->getOperand(1)),
1870 if (Op1Set || Op2Set)
1872 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
1875 // nor(0, y) = ~y -> nor(y, y)
1876 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1877 MVT::i1, MachineNode->getOperand(1),
1878 MachineNode->getOperand(1));
1881 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1882 MVT::i1, MachineNode->getOperand(0),
1883 MachineNode->getOperand(0));
1885 // nor(~x, y) = andc(x, y)
1886 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
1887 MVT::i1, MachineNode->getOperand(0).
1889 MachineNode->getOperand(1));
1891 // nor(x, ~y) = andc(y, x)
1892 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
1893 MVT::i1, MachineNode->getOperand(1).
1895 MachineNode->getOperand(0));
1896 else if (AllUsersSelectZero(MachineNode))
1897 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode),
1898 MVT::i1, MachineNode->getOperand(0),
1899 MachineNode->getOperand(1)),
1903 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1905 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
1909 ResNode = MachineNode->getOperand(1).getNode();
1912 ResNode = MachineNode->getOperand(0).getNode();
1914 // eqv(0, y) = ~y -> nor(y, y)
1915 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1916 MVT::i1, MachineNode->getOperand(1),
1917 MachineNode->getOperand(1));
1920 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1921 MVT::i1, MachineNode->getOperand(0),
1922 MachineNode->getOperand(0));
1924 // eqv(~x, y) = xor(x, y)
1925 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
1926 MVT::i1, MachineNode->getOperand(0).
1928 MachineNode->getOperand(1));
1930 // eqv(x, ~y) = xor(x, y)
1931 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
1932 MVT::i1, MachineNode->getOperand(0),
1933 MachineNode->getOperand(1).
1935 else if (AllUsersSelectZero(MachineNode))
1936 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
1937 MVT::i1, MachineNode->getOperand(0),
1938 MachineNode->getOperand(1)),
1942 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1944 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
1948 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1949 MVT::i1, MachineNode->getOperand(1),
1950 MachineNode->getOperand(1));
1951 else if (Op1Unset || Op2Set)
1952 // andc(0, y) = andc(x, 1) = 0
1953 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
1957 ResNode = MachineNode->getOperand(0).getNode();
1959 // andc(~x, y) = ~(x | y) = nor(x, y)
1960 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1961 MVT::i1, MachineNode->getOperand(0).
1963 MachineNode->getOperand(1));
1965 // andc(x, ~y) = x & y
1966 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode),
1967 MVT::i1, MachineNode->getOperand(0),
1968 MachineNode->getOperand(1).
1970 else if (AllUsersSelectZero(MachineNode))
1971 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
1972 MVT::i1, MachineNode->getOperand(1),
1973 MachineNode->getOperand(0)),
1977 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
1979 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
1981 else if (Op1Set || Op2Unset)
1982 // orc(1, y) = orc(x, 0) = 1
1983 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
1987 ResNode = MachineNode->getOperand(0).getNode();
1990 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
1991 MVT::i1, MachineNode->getOperand(1),
1992 MachineNode->getOperand(1));
1994 // orc(~x, y) = ~(x & y) = nand(x, y)
1995 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode),
1996 MVT::i1, MachineNode->getOperand(0).
1998 MachineNode->getOperand(1));
2000 // orc(x, ~y) = x | y
2001 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode),
2002 MVT::i1, MachineNode->getOperand(0),
2003 MachineNode->getOperand(1).
2005 else if (AllUsersSelectZero(MachineNode))
2006 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
2007 MVT::i1, MachineNode->getOperand(1),
2008 MachineNode->getOperand(0)),
2011 case PPC::SELECT_I4:
2012 case PPC::SELECT_I8:
2013 case PPC::SELECT_F4:
2014 case PPC::SELECT_F8:
2015 case PPC::SELECT_VRRC:
2016 case PPC::SELECT_VSFRC:
2017 case PPC::SELECT_VSRC:
2019 ResNode = MachineNode->getOperand(1).getNode();
2021 ResNode = MachineNode->getOperand(2).getNode();
2023 ResNode = CurDAG->getMachineNode(MachineNode->getMachineOpcode(),
2025 MachineNode->getValueType(0),
2026 MachineNode->getOperand(0).
2028 MachineNode->getOperand(2),
2029 MachineNode->getOperand(1));
2034 ResNode = CurDAG->getMachineNode(Opcode == PPC::BC ? PPC::BCn :
2038 MachineNode->getOperand(0).
2040 MachineNode->getOperand(1),
2041 MachineNode->getOperand(2));
2042 // FIXME: Handle Op1Set, Op1Unset here too.
2046 // If we're inverting this node because it is used only by selects that
2047 // we'd like to swap, then swap the selects before the node replacement.
2049 SwapAllSelectUsers(MachineNode);
2051 if (ResNode != MachineNode) {
2052 DEBUG(dbgs() << "CR Peephole replacing:\nOld: ");
2053 DEBUG(MachineNode->dump(CurDAG));
2054 DEBUG(dbgs() << "\nNew: ");
2055 DEBUG(ResNode->dump(CurDAG));
2056 DEBUG(dbgs() << "\n");
2058 ReplaceUses(MachineNode, ResNode);
2063 CurDAG->RemoveDeadNodes();
2064 } while (IsModified);
2067 void PPCDAGToDAGISel::PeepholePPC64() {
2068 // These optimizations are currently supported only for 64-bit SVR4.
2069 if (PPCSubTarget->isDarwin() || !PPCSubTarget->isPPC64())
2072 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
2075 while (Position != CurDAG->allnodes_begin()) {
2076 SDNode *N = --Position;
2077 // Skip dead nodes and any non-machine opcodes.
2078 if (N->use_empty() || !N->isMachineOpcode())
2082 unsigned StorageOpcode = N->getMachineOpcode();
2084 switch (StorageOpcode) {
2115 // If this is a load or store with a zero offset, we may be able to
2116 // fold an add-immediate into the memory operation.
2117 if (!isa<ConstantSDNode>(N->getOperand(FirstOp)) ||
2118 N->getConstantOperandVal(FirstOp) != 0)
2121 SDValue Base = N->getOperand(FirstOp + 1);
2122 if (!Base.isMachineOpcode())
2126 bool ReplaceFlags = true;
2128 // When the feeding operation is an add-immediate of some sort,
2129 // determine whether we need to add relocation information to the
2130 // target flags on the immediate operand when we fold it into the
2131 // load instruction.
2133 // For something like ADDItocL, the relocation information is
2134 // inferred from the opcode; when we process it in the AsmPrinter,
2135 // we add the necessary relocation there. A load, though, can receive
2136 // relocation from various flavors of ADDIxxx, so we need to carry
2137 // the relocation information in the target flags.
2138 switch (Base.getMachineOpcode()) {
2143 // In some cases (such as TLS) the relocation information
2144 // is already in place on the operand, so copying the operand
2146 ReplaceFlags = false;
2147 // For these cases, the immediate may not be divisible by 4, in
2148 // which case the fold is illegal for DS-form instructions. (The
2149 // other cases provide aligned addresses and are always safe.)
2150 if ((StorageOpcode == PPC::LWA ||
2151 StorageOpcode == PPC::LD ||
2152 StorageOpcode == PPC::STD) &&
2153 (!isa<ConstantSDNode>(Base.getOperand(1)) ||
2154 Base.getConstantOperandVal(1) % 4 != 0))
2157 case PPC::ADDIdtprelL:
2158 Flags = PPCII::MO_DTPREL_LO;
2160 case PPC::ADDItlsldL:
2161 Flags = PPCII::MO_TLSLD_LO;
2164 Flags = PPCII::MO_TOC_LO;
2168 // We found an opportunity. Reverse the operands from the add
2169 // immediate and substitute them into the load or store. If
2170 // needed, update the target flags for the immediate operand to
2171 // reflect the necessary relocation information.
2172 DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase: ");
2173 DEBUG(Base->dump(CurDAG));
2174 DEBUG(dbgs() << "\nN: ");
2175 DEBUG(N->dump(CurDAG));
2176 DEBUG(dbgs() << "\n");
2178 SDValue ImmOpnd = Base.getOperand(1);
2180 // If the relocation information isn't already present on the
2181 // immediate operand, add it now.
2183 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) {
2185 const GlobalValue *GV = GA->getGlobal();
2186 // We can't perform this optimization for data whose alignment
2187 // is insufficient for the instruction encoding.
2188 if (GV->getAlignment() < 4 &&
2189 (StorageOpcode == PPC::LD || StorageOpcode == PPC::STD ||
2190 StorageOpcode == PPC::LWA)) {
2191 DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n");
2194 ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, 0, Flags);
2195 } else if (ConstantPoolSDNode *CP =
2196 dyn_cast<ConstantPoolSDNode>(ImmOpnd)) {
2197 const Constant *C = CP->getConstVal();
2198 ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64,
2204 if (FirstOp == 1) // Store
2205 (void)CurDAG->UpdateNodeOperands(N, N->getOperand(0), ImmOpnd,
2206 Base.getOperand(0), N->getOperand(3));
2208 (void)CurDAG->UpdateNodeOperands(N, ImmOpnd, Base.getOperand(0),
2211 // The add-immediate may now be dead, in which case remove it.
2212 if (Base.getNode()->use_empty())
2213 CurDAG->RemoveDeadNode(Base.getNode());
2218 /// createPPCISelDag - This pass converts a legalized DAG into a
2219 /// PowerPC-specific DAG, ready for instruction scheduling.
2221 FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM) {
2222 return new PPCDAGToDAGISel(TM);
2225 static void initializePassOnce(PassRegistry &Registry) {
2226 const char *Name = "PowerPC DAG->DAG Pattern Instruction Selection";
2227 PassInfo *PI = new PassInfo(Name, "ppc-codegen", &SelectionDAGISel::ID,
2228 nullptr, false, false);
2229 Registry.registerPass(*PI, true);
2232 void llvm::initializePPCDAGToDAGISelPass(PassRegistry &Registry) {
2233 CALL_ONCE_INITIALIZATION(initializePassOnce);