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);
45 cl::opt<bool> UseBitPermRewriter("ppc-use-bit-perm-rewriter", cl::init(true),
46 cl::desc("use aggressive ppc isel for bit permutations"), cl::Hidden);
47 cl::opt<bool> BPermRewriterNoMasking("ppc-bit-perm-rewriter-stress-rotates",
48 cl::desc("stress rotate selection in aggressive ppc isel for "
49 "bit permutations"), cl::Hidden);
52 void initializePPCDAGToDAGISelPass(PassRegistry&);
56 //===--------------------------------------------------------------------===//
57 /// PPCDAGToDAGISel - PPC specific code to select PPC machine
58 /// instructions for SelectionDAG operations.
60 class PPCDAGToDAGISel : public SelectionDAGISel {
61 const PPCTargetMachine &TM;
62 const PPCTargetLowering *PPCLowering;
63 const PPCSubtarget *PPCSubTarget;
64 unsigned GlobalBaseReg;
66 explicit PPCDAGToDAGISel(PPCTargetMachine &tm)
67 : SelectionDAGISel(tm), TM(tm),
68 PPCLowering(TM.getSubtargetImpl()->getTargetLowering()),
69 PPCSubTarget(TM.getSubtargetImpl()) {
70 initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry());
73 bool runOnMachineFunction(MachineFunction &MF) override {
74 // Make sure we re-emit a set of the global base reg if necessary
76 PPCLowering = TM.getSubtargetImpl()->getTargetLowering();
77 PPCSubTarget = TM.getSubtargetImpl();
78 SelectionDAGISel::runOnMachineFunction(MF);
80 if (!PPCSubTarget->isSVR4ABI())
86 void PreprocessISelDAG() override;
87 void PostprocessISelDAG() override;
89 /// getI32Imm - Return a target constant with the specified value, of type
91 inline SDValue getI32Imm(unsigned Imm) {
92 return CurDAG->getTargetConstant(Imm, MVT::i32);
95 /// getI64Imm - Return a target constant with the specified value, of type
97 inline SDValue getI64Imm(uint64_t Imm) {
98 return CurDAG->getTargetConstant(Imm, MVT::i64);
101 /// getSmallIPtrImm - Return a target constant of pointer type.
102 inline SDValue getSmallIPtrImm(unsigned Imm) {
103 return CurDAG->getTargetConstant(Imm, PPCLowering->getPointerTy());
106 /// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s
107 /// with any number of 0s on either side. The 1s are allowed to wrap from
108 /// LSB to MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs.
109 /// 0x0F0F0000 is not, since all 1s are not contiguous.
110 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME);
113 /// isRotateAndMask - Returns true if Mask and Shift can be folded into a
114 /// rotate and mask opcode and mask operation.
115 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask,
116 unsigned &SH, unsigned &MB, unsigned &ME);
118 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
119 /// base register. Return the virtual register that holds this value.
120 SDNode *getGlobalBaseReg();
122 SDNode *getFrameIndex(SDNode *SN, SDNode *N, unsigned Offset = 0);
124 // Select - Convert the specified operand from a target-independent to a
125 // target-specific node if it hasn't already been changed.
126 SDNode *Select(SDNode *N) override;
128 SDNode *SelectBitfieldInsert(SDNode *N);
129 SDNode *SelectBitPermutation(SDNode *N);
131 /// SelectCC - Select a comparison of the specified values with the
132 /// specified condition code, returning the CR# of the expression.
133 SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, SDLoc dl);
135 /// SelectAddrImm - Returns true if the address N can be represented by
136 /// a base register plus a signed 16-bit displacement [r+imm].
137 bool SelectAddrImm(SDValue N, SDValue &Disp,
139 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, false);
142 /// SelectAddrImmOffs - Return true if the operand is valid for a preinc
143 /// immediate field. Note that the operand at this point is already the
144 /// result of a prior SelectAddressRegImm call.
145 bool SelectAddrImmOffs(SDValue N, SDValue &Out) const {
146 if (N.getOpcode() == ISD::TargetConstant ||
147 N.getOpcode() == ISD::TargetGlobalAddress) {
155 /// SelectAddrIdx - Given the specified addressed, check to see if it can be
156 /// represented as an indexed [r+r] operation. Returns false if it can
157 /// be represented by [r+imm], which are preferred.
158 bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) {
159 return PPCLowering->SelectAddressRegReg(N, Base, Index, *CurDAG);
162 /// SelectAddrIdxOnly - Given the specified addressed, force it to be
163 /// represented as an indexed [r+r] operation.
164 bool SelectAddrIdxOnly(SDValue N, SDValue &Base, SDValue &Index) {
165 return PPCLowering->SelectAddressRegRegOnly(N, Base, Index, *CurDAG);
168 /// SelectAddrImmX4 - Returns true if the address N can be represented by
169 /// a base register plus a signed 16-bit displacement that is a multiple of 4.
170 /// Suitable for use by STD and friends.
171 bool SelectAddrImmX4(SDValue N, SDValue &Disp, SDValue &Base) {
172 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, true);
175 // Select an address into a single register.
176 bool SelectAddr(SDValue N, SDValue &Base) {
181 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
182 /// inline asm expressions. It is always correct to compute the value into
183 /// a register. The case of adding a (possibly relocatable) constant to a
184 /// register can be improved, but it is wrong to substitute Reg+Reg for
185 /// Reg in an asm, because the load or store opcode would have to change.
186 bool SelectInlineAsmMemoryOperand(const SDValue &Op,
188 std::vector<SDValue> &OutOps) override {
189 // We need to make sure that this one operand does not end up in r0
190 // (because we might end up lowering this as 0(%op)).
191 const TargetRegisterInfo *TRI = TM.getSubtargetImpl()->getRegisterInfo();
192 const TargetRegisterClass *TRC = TRI->getPointerRegClass(*MF, /*Kind=*/1);
193 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
195 SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
196 SDLoc(Op), Op.getValueType(),
199 OutOps.push_back(NewOp);
203 void InsertVRSaveCode(MachineFunction &MF);
205 const char *getPassName() const override {
206 return "PowerPC DAG->DAG Pattern Instruction Selection";
209 // Include the pieces autogenerated from the target description.
210 #include "PPCGenDAGISel.inc"
213 SDNode *SelectSETCC(SDNode *N);
215 void PeepholePPC64();
216 void PeepholePPC64ZExt();
217 void PeepholeCROps();
219 SDValue combineToCMPB(SDNode *N);
221 bool AllUsersSelectZero(SDNode *N);
222 void SwapAllSelectUsers(SDNode *N);
226 /// InsertVRSaveCode - Once the entire function has been instruction selected,
227 /// all virtual registers are created and all machine instructions are built,
228 /// check to see if we need to save/restore VRSAVE. If so, do it.
229 void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) {
230 // Check to see if this function uses vector registers, which means we have to
231 // save and restore the VRSAVE register and update it with the regs we use.
233 // In this case, there will be virtual registers of vector type created
234 // by the scheduler. Detect them now.
235 bool HasVectorVReg = false;
236 for (unsigned i = 0, e = RegInfo->getNumVirtRegs(); i != e; ++i) {
237 unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
238 if (RegInfo->getRegClass(Reg) == &PPC::VRRCRegClass) {
239 HasVectorVReg = true;
243 if (!HasVectorVReg) return; // nothing to do.
245 // If we have a vector register, we want to emit code into the entry and exit
246 // blocks to save and restore the VRSAVE register. We do this here (instead
247 // of marking all vector instructions as clobbering VRSAVE) for two reasons:
249 // 1. This (trivially) reduces the load on the register allocator, by not
250 // having to represent the live range of the VRSAVE register.
251 // 2. This (more significantly) allows us to create a temporary virtual
252 // register to hold the saved VRSAVE value, allowing this temporary to be
253 // register allocated, instead of forcing it to be spilled to the stack.
255 // Create two vregs - one to hold the VRSAVE register that is live-in to the
256 // function and one for the value after having bits or'd into it.
257 unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
258 unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
260 const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
261 MachineBasicBlock &EntryBB = *Fn.begin();
263 // Emit the following code into the entry block:
264 // InVRSAVE = MFVRSAVE
265 // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE
266 // MTVRSAVE UpdatedVRSAVE
267 MachineBasicBlock::iterator IP = EntryBB.begin(); // Insert Point
268 BuildMI(EntryBB, IP, dl, TII.get(PPC::MFVRSAVE), InVRSAVE);
269 BuildMI(EntryBB, IP, dl, TII.get(PPC::UPDATE_VRSAVE),
270 UpdatedVRSAVE).addReg(InVRSAVE);
271 BuildMI(EntryBB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(UpdatedVRSAVE);
273 // Find all return blocks, outputting a restore in each epilog.
274 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
275 if (!BB->empty() && BB->back().isReturn()) {
276 IP = BB->end(); --IP;
278 // Skip over all terminator instructions, which are part of the return
280 MachineBasicBlock::iterator I2 = IP;
281 while (I2 != BB->begin() && (--I2)->isTerminator())
284 // Emit: MTVRSAVE InVRSave
285 BuildMI(*BB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(InVRSAVE);
291 /// getGlobalBaseReg - Output the instructions required to put the
292 /// base address to use for accessing globals into a register.
294 SDNode *PPCDAGToDAGISel::getGlobalBaseReg() {
295 if (!GlobalBaseReg) {
296 const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
297 // Insert the set of GlobalBaseReg into the first MBB of the function
298 MachineBasicBlock &FirstMBB = MF->front();
299 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
300 const Module *M = MF->getFunction()->getParent();
303 if (PPCLowering->getPointerTy() == MVT::i32) {
304 if (PPCSubTarget->isTargetELF()) {
305 GlobalBaseReg = PPC::R30;
306 if (M->getPICLevel() == PICLevel::Small) {
307 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MoveGOTtoLR));
308 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
310 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
311 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
312 unsigned TempReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
313 BuildMI(FirstMBB, MBBI, dl,
314 TII.get(PPC::UpdateGBR)).addReg(GlobalBaseReg)
315 .addReg(TempReg, RegState::Define).addReg(GlobalBaseReg);
316 MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true);
320 RegInfo->createVirtualRegister(&PPC::GPRC_NOR0RegClass);
321 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
322 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
325 GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RC_NOX0RegClass);
326 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8));
327 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg);
330 return CurDAG->getRegister(GlobalBaseReg,
331 PPCLowering->getPointerTy()).getNode();
334 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
335 /// or 64-bit immediate, and if the value can be accurately represented as a
336 /// sign extension from a 16-bit value. If so, this returns true and the
338 static bool isIntS16Immediate(SDNode *N, short &Imm) {
339 if (N->getOpcode() != ISD::Constant)
342 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
343 if (N->getValueType(0) == MVT::i32)
344 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
346 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
349 static bool isIntS16Immediate(SDValue Op, short &Imm) {
350 return isIntS16Immediate(Op.getNode(), Imm);
354 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
355 /// operand. If so Imm will receive the 32-bit value.
356 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
357 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
358 Imm = cast<ConstantSDNode>(N)->getZExtValue();
364 /// isInt64Immediate - This method tests to see if the node is a 64-bit constant
365 /// operand. If so Imm will receive the 64-bit value.
366 static bool isInt64Immediate(SDNode *N, uint64_t &Imm) {
367 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i64) {
368 Imm = cast<ConstantSDNode>(N)->getZExtValue();
374 // isInt32Immediate - This method tests to see if a constant operand.
375 // If so Imm will receive the 32 bit value.
376 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
377 return isInt32Immediate(N.getNode(), Imm);
381 // isOpcWithIntImmediate - This method tests to see if the node is a specific
382 // opcode and that it has a immediate integer right operand.
383 // If so Imm will receive the 32 bit value.
384 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
385 return N->getOpcode() == Opc
386 && isInt32Immediate(N->getOperand(1).getNode(), Imm);
389 SDNode *PPCDAGToDAGISel::getFrameIndex(SDNode *SN, SDNode *N, unsigned Offset) {
391 int FI = cast<FrameIndexSDNode>(N)->getIndex();
392 SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0));
393 unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8;
395 return CurDAG->SelectNodeTo(SN, Opc, N->getValueType(0), TFI,
396 getSmallIPtrImm(Offset));
397 return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI,
398 getSmallIPtrImm(Offset));
401 bool PPCDAGToDAGISel::isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
405 if (isShiftedMask_32(Val)) {
406 // look for the first non-zero bit
407 MB = countLeadingZeros(Val);
408 // look for the first zero bit after the run of ones
409 ME = countLeadingZeros((Val - 1) ^ Val);
412 Val = ~Val; // invert mask
413 if (isShiftedMask_32(Val)) {
414 // effectively look for the first zero bit
415 ME = countLeadingZeros(Val) - 1;
416 // effectively look for the first one bit after the run of zeros
417 MB = countLeadingZeros((Val - 1) ^ Val) + 1;
425 bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask,
426 bool isShiftMask, unsigned &SH,
427 unsigned &MB, unsigned &ME) {
428 // Don't even go down this path for i64, since different logic will be
429 // necessary for rldicl/rldicr/rldimi.
430 if (N->getValueType(0) != MVT::i32)
434 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
435 unsigned Opcode = N->getOpcode();
436 if (N->getNumOperands() != 2 ||
437 !isInt32Immediate(N->getOperand(1).getNode(), Shift) || (Shift > 31))
440 if (Opcode == ISD::SHL) {
441 // apply shift left to mask if it comes first
442 if (isShiftMask) Mask = Mask << Shift;
443 // determine which bits are made indeterminant by shift
444 Indeterminant = ~(0xFFFFFFFFu << Shift);
445 } else if (Opcode == ISD::SRL) {
446 // apply shift right to mask if it comes first
447 if (isShiftMask) Mask = Mask >> Shift;
448 // determine which bits are made indeterminant by shift
449 Indeterminant = ~(0xFFFFFFFFu >> Shift);
450 // adjust for the left rotate
452 } else if (Opcode == ISD::ROTL) {
458 // if the mask doesn't intersect any Indeterminant bits
459 if (Mask && !(Mask & Indeterminant)) {
461 // make sure the mask is still a mask (wrap arounds may not be)
462 return isRunOfOnes(Mask, MB, ME);
467 /// SelectBitfieldInsert - turn an or of two masked values into
468 /// the rotate left word immediate then mask insert (rlwimi) instruction.
469 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
470 SDValue Op0 = N->getOperand(0);
471 SDValue Op1 = N->getOperand(1);
474 APInt LKZ, LKO, RKZ, RKO;
475 CurDAG->computeKnownBits(Op0, LKZ, LKO);
476 CurDAG->computeKnownBits(Op1, RKZ, RKO);
478 unsigned TargetMask = LKZ.getZExtValue();
479 unsigned InsertMask = RKZ.getZExtValue();
481 if ((TargetMask | InsertMask) == 0xFFFFFFFF) {
482 unsigned Op0Opc = Op0.getOpcode();
483 unsigned Op1Opc = Op1.getOpcode();
484 unsigned Value, SH = 0;
485 TargetMask = ~TargetMask;
486 InsertMask = ~InsertMask;
488 // If the LHS has a foldable shift and the RHS does not, then swap it to the
489 // RHS so that we can fold the shift into the insert.
490 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
491 if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
492 Op0.getOperand(0).getOpcode() == ISD::SRL) {
493 if (Op1.getOperand(0).getOpcode() != ISD::SHL &&
494 Op1.getOperand(0).getOpcode() != ISD::SRL) {
496 std::swap(Op0Opc, Op1Opc);
497 std::swap(TargetMask, InsertMask);
500 } else if (Op0Opc == ISD::SHL || Op0Opc == ISD::SRL) {
501 if (Op1Opc == ISD::AND && Op1.getOperand(0).getOpcode() != ISD::SHL &&
502 Op1.getOperand(0).getOpcode() != ISD::SRL) {
504 std::swap(Op0Opc, Op1Opc);
505 std::swap(TargetMask, InsertMask);
510 if (isRunOfOnes(InsertMask, MB, ME)) {
513 if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) &&
514 isInt32Immediate(Op1.getOperand(1), Value)) {
515 Op1 = Op1.getOperand(0);
516 SH = (Op1Opc == ISD::SHL) ? Value : 32 - Value;
518 if (Op1Opc == ISD::AND) {
519 // The AND mask might not be a constant, and we need to make sure that
520 // if we're going to fold the masking with the insert, all bits not
521 // know to be zero in the mask are known to be one.
523 CurDAG->computeKnownBits(Op1.getOperand(1), MKZ, MKO);
524 bool CanFoldMask = InsertMask == MKO.getZExtValue();
526 unsigned SHOpc = Op1.getOperand(0).getOpcode();
527 if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) && CanFoldMask &&
528 isInt32Immediate(Op1.getOperand(0).getOperand(1), Value)) {
529 // Note that Value must be in range here (less than 32) because
530 // otherwise there would not be any bits set in InsertMask.
531 Op1 = Op1.getOperand(0).getOperand(0);
532 SH = (SHOpc == ISD::SHL) ? Value : 32 - Value;
537 SDValue Ops[] = { Op0, Op1, getI32Imm(SH), getI32Imm(MB),
539 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
545 // Predict the number of instructions that would be generated by calling
547 static unsigned SelectInt64CountDirect(int64_t Imm) {
548 // Assume no remaining bits.
549 unsigned Remainder = 0;
550 // Assume no shift required.
553 // If it can't be represented as a 32 bit value.
554 if (!isInt<32>(Imm)) {
555 Shift = countTrailingZeros<uint64_t>(Imm);
556 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
558 // If the shifted value fits 32 bits.
559 if (isInt<32>(ImmSh)) {
560 // Go with the shifted value.
563 // Still stuck with a 64 bit value.
570 // Intermediate operand.
573 // Handle first 32 bits.
574 unsigned Lo = Imm & 0xFFFF;
575 unsigned Hi = (Imm >> 16) & 0xFFFF;
578 if (isInt<16>(Imm)) {
582 // Handle the Hi bits and Lo bits.
589 // If no shift, we're done.
590 if (!Shift) return Result;
592 // Shift for next step if the upper 32-bits were not zero.
596 // Add in the last bits as required.
597 if ((Hi = (Remainder >> 16) & 0xFFFF))
599 if ((Lo = Remainder & 0xFFFF))
605 static uint64_t Rot64(uint64_t Imm, unsigned R) {
606 return (Imm << R) | (Imm >> (64 - R));
609 static unsigned SelectInt64Count(int64_t Imm) {
610 unsigned Count = SelectInt64CountDirect(Imm);
612 for (unsigned r = 1; r < 63; ++r) {
613 unsigned RCount = SelectInt64CountDirect(Rot64(Imm, r)) + 1;
614 Count = std::min(Count, RCount);
620 // Select a 64-bit constant. For cost-modeling purposes, SelectInt64Count
621 // (above) needs to be kept in sync with this function.
622 static SDNode *SelectInt64Direct(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) {
623 // Assume no remaining bits.
624 unsigned Remainder = 0;
625 // Assume no shift required.
628 // If it can't be represented as a 32 bit value.
629 if (!isInt<32>(Imm)) {
630 Shift = countTrailingZeros<uint64_t>(Imm);
631 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
633 // If the shifted value fits 32 bits.
634 if (isInt<32>(ImmSh)) {
635 // Go with the shifted value.
638 // Still stuck with a 64 bit value.
645 // Intermediate operand.
648 // Handle first 32 bits.
649 unsigned Lo = Imm & 0xFFFF;
650 unsigned Hi = (Imm >> 16) & 0xFFFF;
652 auto getI32Imm = [CurDAG](unsigned Imm) {
653 return CurDAG->getTargetConstant(Imm, MVT::i32);
657 if (isInt<16>(Imm)) {
659 Result = CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(Lo));
661 // Handle the Hi bits.
662 unsigned OpC = Hi ? PPC::LIS8 : PPC::LI8;
663 Result = CurDAG->getMachineNode(OpC, dl, MVT::i64, getI32Imm(Hi));
665 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
666 SDValue(Result, 0), getI32Imm(Lo));
669 Result = CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(Hi));
672 // If no shift, we're done.
673 if (!Shift) return Result;
675 // Shift for next step if the upper 32-bits were not zero.
677 Result = CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64,
680 getI32Imm(63 - Shift));
683 // Add in the last bits as required.
684 if ((Hi = (Remainder >> 16) & 0xFFFF)) {
685 Result = CurDAG->getMachineNode(PPC::ORIS8, dl, MVT::i64,
686 SDValue(Result, 0), getI32Imm(Hi));
688 if ((Lo = Remainder & 0xFFFF)) {
689 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
690 SDValue(Result, 0), getI32Imm(Lo));
696 static SDNode *SelectInt64(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) {
697 unsigned Count = SelectInt64CountDirect(Imm);
700 for (unsigned r = 1; r < 63; ++r) {
701 unsigned RCount = SelectInt64CountDirect(Rot64(Imm, r)) + 1;
702 if (RCount < Count) {
709 return SelectInt64Direct(CurDAG, dl, Imm);
711 auto getI32Imm = [CurDAG](unsigned Imm) {
712 return CurDAG->getTargetConstant(Imm, MVT::i32);
715 SDValue Val = SDValue(SelectInt64Direct(CurDAG, dl, Rot64(Imm, RMin)), 0);
716 return CurDAG->getMachineNode(PPC::RLDICL, dl, MVT::i64, Val,
717 getI32Imm(64 - RMin), getI32Imm(0));
720 // Select a 64-bit constant.
721 static SDNode *SelectInt64(SelectionDAG *CurDAG, SDNode *N) {
725 int64_t Imm = cast<ConstantSDNode>(N)->getZExtValue();
726 return SelectInt64(CurDAG, dl, Imm);
730 class BitPermutationSelector {
734 // The bit number in the value, using a convention where bit 0 is the
743 ValueBit(SDValue V, unsigned I, Kind K = Variable)
744 : V(V), Idx(I), K(K) {}
745 ValueBit(Kind K = Variable)
746 : V(SDValue(nullptr, 0)), Idx(UINT32_MAX), K(K) {}
748 bool isZero() const {
749 return K == ConstZero;
752 bool hasValue() const {
753 return K == Variable;
756 SDValue getValue() const {
757 assert(hasValue() && "Cannot get the value of a constant bit");
761 unsigned getValueBitIndex() const {
762 assert(hasValue() && "Cannot get the value bit index of a constant bit");
767 // A bit group has the same underlying value and the same rotate factor.
771 unsigned StartIdx, EndIdx;
773 // This rotation amount assumes that the lower 32 bits of the quantity are
774 // replicated in the high 32 bits by the rotation operator (which is done
775 // by rlwinm and friends in 64-bit mode).
777 // Did converting to Repl32 == true change the rotation factor? If it did,
778 // it decreased it by 32.
780 // Was this group coalesced after setting Repl32 to true?
781 bool Repl32Coalesced;
783 BitGroup(SDValue V, unsigned R, unsigned S, unsigned E)
784 : V(V), RLAmt(R), StartIdx(S), EndIdx(E), Repl32(false), Repl32CR(false),
785 Repl32Coalesced(false) {
786 DEBUG(dbgs() << "\tbit group for " << V.getNode() << " RLAmt = " << R <<
787 " [" << S << ", " << E << "]\n");
791 // Information on each (Value, RLAmt) pair (like the number of groups
792 // associated with each) used to choose the lowering method.
793 struct ValueRotInfo {
797 unsigned FirstGroupStartIdx;
801 : RLAmt(UINT32_MAX), NumGroups(0), FirstGroupStartIdx(UINT32_MAX),
804 // For sorting (in reverse order) by NumGroups, and then by
805 // FirstGroupStartIdx.
806 bool operator < (const ValueRotInfo &Other) const {
807 // We need to sort so that the non-Repl32 come first because, when we're
808 // doing masking, the Repl32 bit groups might be subsumed into the 64-bit
809 // masking operation.
810 if (Repl32 < Other.Repl32)
812 else if (Repl32 > Other.Repl32)
814 else if (NumGroups > Other.NumGroups)
816 else if (NumGroups < Other.NumGroups)
818 else if (FirstGroupStartIdx < Other.FirstGroupStartIdx)
824 // Return true if something interesting was deduced, return false if we're
825 // providing only a generic representation of V (or something else likewise
826 // uninteresting for instruction selection).
827 bool getValueBits(SDValue V, SmallVector<ValueBit, 64> &Bits) {
828 switch (V.getOpcode()) {
831 if (isa<ConstantSDNode>(V.getOperand(1))) {
832 unsigned RotAmt = V.getConstantOperandVal(1);
834 SmallVector<ValueBit, 64> LHSBits(Bits.size());
835 getValueBits(V.getOperand(0), LHSBits);
837 for (unsigned i = 0; i < Bits.size(); ++i)
838 Bits[i] = LHSBits[i < RotAmt ? i + (Bits.size() - RotAmt) : i - RotAmt];
844 if (isa<ConstantSDNode>(V.getOperand(1))) {
845 unsigned ShiftAmt = V.getConstantOperandVal(1);
847 SmallVector<ValueBit, 64> LHSBits(Bits.size());
848 getValueBits(V.getOperand(0), LHSBits);
850 for (unsigned i = ShiftAmt; i < Bits.size(); ++i)
851 Bits[i] = LHSBits[i - ShiftAmt];
853 for (unsigned i = 0; i < ShiftAmt; ++i)
854 Bits[i] = ValueBit(ValueBit::ConstZero);
860 if (isa<ConstantSDNode>(V.getOperand(1))) {
861 unsigned ShiftAmt = V.getConstantOperandVal(1);
863 SmallVector<ValueBit, 64> LHSBits(Bits.size());
864 getValueBits(V.getOperand(0), LHSBits);
866 for (unsigned i = 0; i < Bits.size() - ShiftAmt; ++i)
867 Bits[i] = LHSBits[i + ShiftAmt];
869 for (unsigned i = Bits.size() - ShiftAmt; i < Bits.size(); ++i)
870 Bits[i] = ValueBit(ValueBit::ConstZero);
876 if (isa<ConstantSDNode>(V.getOperand(1))) {
877 uint64_t Mask = V.getConstantOperandVal(1);
879 SmallVector<ValueBit, 64> LHSBits(Bits.size());
880 bool LHSTrivial = getValueBits(V.getOperand(0), LHSBits);
882 for (unsigned i = 0; i < Bits.size(); ++i)
883 if (((Mask >> i) & 1) == 1)
884 Bits[i] = LHSBits[i];
886 Bits[i] = ValueBit(ValueBit::ConstZero);
888 // Mark this as interesting, only if the LHS was also interesting. This
889 // prevents the overall procedure from matching a single immediate 'and'
890 // (which is non-optimal because such an and might be folded with other
891 // things if we don't select it here).
896 SmallVector<ValueBit, 64> LHSBits(Bits.size()), RHSBits(Bits.size());
897 getValueBits(V.getOperand(0), LHSBits);
898 getValueBits(V.getOperand(1), RHSBits);
900 bool AllDisjoint = true;
901 for (unsigned i = 0; i < Bits.size(); ++i)
902 if (LHSBits[i].isZero())
903 Bits[i] = RHSBits[i];
904 else if (RHSBits[i].isZero())
905 Bits[i] = LHSBits[i];
918 for (unsigned i = 0; i < Bits.size(); ++i)
919 Bits[i] = ValueBit(V, i);
924 // For each value (except the constant ones), compute the left-rotate amount
925 // to get it from its original to final position.
926 void computeRotationAmounts() {
928 RLAmt.resize(Bits.size());
929 for (unsigned i = 0; i < Bits.size(); ++i)
930 if (Bits[i].hasValue()) {
931 unsigned VBI = Bits[i].getValueBitIndex();
935 RLAmt[i] = Bits.size() - (VBI - i);
936 } else if (Bits[i].isZero()) {
938 RLAmt[i] = UINT32_MAX;
940 llvm_unreachable("Unknown value bit type");
944 // Collect groups of consecutive bits with the same underlying value and
945 // rotation factor. If we're doing late masking, we ignore zeros, otherwise
946 // they break up groups.
947 void collectBitGroups(bool LateMask) {
950 unsigned LastRLAmt = RLAmt[0];
951 SDValue LastValue = Bits[0].hasValue() ? Bits[0].getValue() : SDValue();
952 unsigned LastGroupStartIdx = 0;
953 for (unsigned i = 1; i < Bits.size(); ++i) {
954 unsigned ThisRLAmt = RLAmt[i];
955 SDValue ThisValue = Bits[i].hasValue() ? Bits[i].getValue() : SDValue();
956 if (LateMask && !ThisValue) {
957 ThisValue = LastValue;
958 ThisRLAmt = LastRLAmt;
959 // If we're doing late masking, then the first bit group always starts
960 // at zero (even if the first bits were zero).
961 if (BitGroups.empty())
962 LastGroupStartIdx = 0;
965 // If this bit has the same underlying value and the same rotate factor as
966 // the last one, then they're part of the same group.
967 if (ThisRLAmt == LastRLAmt && ThisValue == LastValue)
970 if (LastValue.getNode())
971 BitGroups.push_back(BitGroup(LastValue, LastRLAmt, LastGroupStartIdx,
973 LastRLAmt = ThisRLAmt;
974 LastValue = ThisValue;
975 LastGroupStartIdx = i;
977 if (LastValue.getNode())
978 BitGroups.push_back(BitGroup(LastValue, LastRLAmt, LastGroupStartIdx,
981 if (BitGroups.empty())
984 // We might be able to combine the first and last groups.
985 if (BitGroups.size() > 1) {
986 // If the first and last groups are the same, then remove the first group
987 // in favor of the last group, making the ending index of the last group
988 // equal to the ending index of the to-be-removed first group.
989 if (BitGroups[0].StartIdx == 0 &&
990 BitGroups[BitGroups.size()-1].EndIdx == Bits.size()-1 &&
991 BitGroups[0].V == BitGroups[BitGroups.size()-1].V &&
992 BitGroups[0].RLAmt == BitGroups[BitGroups.size()-1].RLAmt) {
993 DEBUG(dbgs() << "\tcombining final bit group with inital one\n");
994 BitGroups[BitGroups.size()-1].EndIdx = BitGroups[0].EndIdx;
995 BitGroups.erase(BitGroups.begin());
1000 // Take all (SDValue, RLAmt) pairs and sort them by the number of groups
1001 // associated with each. If there is a degeneracy, pick the one that occurs
1002 // first (in the final value).
1003 void collectValueRotInfo() {
1006 for (auto &BG : BitGroups) {
1007 unsigned RLAmtKey = BG.RLAmt + (BG.Repl32 ? 64 : 0);
1008 ValueRotInfo &VRI = ValueRots[std::make_pair(BG.V, RLAmtKey)];
1010 VRI.RLAmt = BG.RLAmt;
1011 VRI.Repl32 = BG.Repl32;
1013 VRI.FirstGroupStartIdx = std::min(VRI.FirstGroupStartIdx, BG.StartIdx);
1016 // Now that we've collected the various ValueRotInfo instances, we need to
1018 ValueRotsVec.clear();
1019 for (auto &I : ValueRots) {
1020 ValueRotsVec.push_back(I.second);
1022 std::sort(ValueRotsVec.begin(), ValueRotsVec.end());
1025 // In 64-bit mode, rlwinm and friends have a rotation operator that
1026 // replicates the low-order 32 bits into the high-order 32-bits. The mask
1027 // indices of these instructions can only be in the lower 32 bits, so they
1028 // can only represent some 64-bit bit groups. However, when they can be used,
1029 // the 32-bit replication can be used to represent, as a single bit group,
1030 // otherwise separate bit groups. We'll convert to replicated-32-bit bit
1031 // groups when possible. Returns true if any of the bit groups were
1033 void assignRepl32BitGroups() {
1034 // If we have bits like this:
1036 // Indices: 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
1037 // V bits: ... 7 6 5 4 3 2 1 0 31 30 29 28 27 26 25 24
1038 // Groups: | RLAmt = 8 | RLAmt = 40 |
1040 // But, making use of a 32-bit operation that replicates the low-order 32
1041 // bits into the high-order 32 bits, this can be one bit group with a RLAmt
1044 auto IsAllLow32 = [this](BitGroup & BG) {
1045 if (BG.StartIdx <= BG.EndIdx) {
1046 for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i) {
1047 if (!Bits[i].hasValue())
1049 if (Bits[i].getValueBitIndex() >= 32)
1053 for (unsigned i = BG.StartIdx; i < Bits.size(); ++i) {
1054 if (!Bits[i].hasValue())
1056 if (Bits[i].getValueBitIndex() >= 32)
1059 for (unsigned i = 0; i <= BG.EndIdx; ++i) {
1060 if (!Bits[i].hasValue())
1062 if (Bits[i].getValueBitIndex() >= 32)
1070 for (auto &BG : BitGroups) {
1071 if (BG.StartIdx < 32 && BG.EndIdx < 32) {
1072 if (IsAllLow32(BG)) {
1073 if (BG.RLAmt >= 32) {
1080 DEBUG(dbgs() << "\t32-bit replicated bit group for " <<
1081 BG.V.getNode() << " RLAmt = " << BG.RLAmt <<
1082 " [" << BG.StartIdx << ", " << BG.EndIdx << "]\n");
1087 // Now walk through the bit groups, consolidating where possible.
1088 for (auto I = BitGroups.begin(); I != BitGroups.end();) {
1089 // We might want to remove this bit group by merging it with the previous
1090 // group (which might be the ending group).
1091 auto IP = (I == BitGroups.begin()) ?
1092 std::prev(BitGroups.end()) : std::prev(I);
1093 if (I->Repl32 && IP->Repl32 && I->V == IP->V && I->RLAmt == IP->RLAmt &&
1094 I->StartIdx == (IP->EndIdx + 1) % 64 && I != IP) {
1096 DEBUG(dbgs() << "\tcombining 32-bit replicated bit group for " <<
1097 I->V.getNode() << " RLAmt = " << I->RLAmt <<
1098 " [" << I->StartIdx << ", " << I->EndIdx <<
1099 "] with group with range [" <<
1100 IP->StartIdx << ", " << IP->EndIdx << "]\n");
1102 IP->EndIdx = I->EndIdx;
1103 IP->Repl32CR = IP->Repl32CR || I->Repl32CR;
1104 IP->Repl32Coalesced = true;
1105 I = BitGroups.erase(I);
1108 // There is a special case worth handling: If there is a single group
1109 // covering the entire upper 32 bits, and it can be merged with both
1110 // the next and previous groups (which might be the same group), then
1111 // do so. If it is the same group (so there will be only one group in
1112 // total), then we need to reverse the order of the range so that it
1113 // covers the entire 64 bits.
1114 if (I->StartIdx == 32 && I->EndIdx == 63) {
1115 assert(std::next(I) == BitGroups.end() &&
1116 "bit group ends at index 63 but there is another?");
1117 auto IN = BitGroups.begin();
1119 if (IP->Repl32 && IN->Repl32 && I->V == IP->V && I->V == IN->V &&
1120 (I->RLAmt % 32) == IP->RLAmt && (I->RLAmt % 32) == IN->RLAmt &&
1121 IP->EndIdx == 31 && IN->StartIdx == 0 && I != IP &&
1124 DEBUG(dbgs() << "\tcombining bit group for " <<
1125 I->V.getNode() << " RLAmt = " << I->RLAmt <<
1126 " [" << I->StartIdx << ", " << I->EndIdx <<
1127 "] with 32-bit replicated groups with ranges [" <<
1128 IP->StartIdx << ", " << IP->EndIdx << "] and [" <<
1129 IN->StartIdx << ", " << IN->EndIdx << "]\n");
1132 // There is only one other group; change it to cover the whole
1133 // range (backward, so that it can still be Repl32 but cover the
1134 // whole 64-bit range).
1137 IP->Repl32CR = IP->Repl32CR || I->RLAmt >= 32;
1138 IP->Repl32Coalesced = true;
1139 I = BitGroups.erase(I);
1141 // There are two separate groups, one before this group and one
1142 // after us (at the beginning). We're going to remove this group,
1143 // but also the group at the very beginning.
1144 IP->EndIdx = IN->EndIdx;
1145 IP->Repl32CR = IP->Repl32CR || IN->Repl32CR || I->RLAmt >= 32;
1146 IP->Repl32Coalesced = true;
1147 I = BitGroups.erase(I);
1148 BitGroups.erase(BitGroups.begin());
1151 // This must be the last group in the vector (and we might have
1152 // just invalidated the iterator above), so break here.
1162 SDValue getI32Imm(unsigned Imm) {
1163 return CurDAG->getTargetConstant(Imm, MVT::i32);
1166 uint64_t getZerosMask() {
1168 for (unsigned i = 0; i < Bits.size(); ++i) {
1169 if (Bits[i].hasValue())
1171 Mask |= (UINT64_C(1) << i);
1177 // Depending on the number of groups for a particular value, it might be
1178 // better to rotate, mask explicitly (using andi/andis), and then or the
1179 // result. Select this part of the result first.
1180 void SelectAndParts32(SDLoc dl, SDValue &Res, unsigned *InstCnt) {
1181 if (BPermRewriterNoMasking)
1184 for (ValueRotInfo &VRI : ValueRotsVec) {
1186 for (unsigned i = 0; i < Bits.size(); ++i) {
1187 if (!Bits[i].hasValue() || Bits[i].getValue() != VRI.V)
1189 if (RLAmt[i] != VRI.RLAmt)
1194 // Compute the masks for andi/andis that would be necessary.
1195 unsigned ANDIMask = (Mask & UINT16_MAX), ANDISMask = Mask >> 16;
1196 assert((ANDIMask != 0 || ANDISMask != 0) &&
1197 "No set bits in mask for value bit groups");
1198 bool NeedsRotate = VRI.RLAmt != 0;
1200 // We're trying to minimize the number of instructions. If we have one
1201 // group, using one of andi/andis can break even. If we have three
1202 // groups, we can use both andi and andis and break even (to use both
1203 // andi and andis we also need to or the results together). We need four
1204 // groups if we also need to rotate. To use andi/andis we need to do more
1205 // than break even because rotate-and-mask instructions tend to be easier
1208 // FIXME: We've biased here against using andi/andis, which is right for
1209 // POWER cores, but not optimal everywhere. For example, on the A2,
1210 // andi/andis have single-cycle latency whereas the rotate-and-mask
1211 // instructions take two cycles, and it would be better to bias toward
1212 // andi/andis in break-even cases.
1214 unsigned NumAndInsts = (unsigned) NeedsRotate +
1215 (unsigned) (ANDIMask != 0) +
1216 (unsigned) (ANDISMask != 0) +
1217 (unsigned) (ANDIMask != 0 && ANDISMask != 0) +
1218 (unsigned) (bool) Res;
1220 DEBUG(dbgs() << "\t\trotation groups for " << VRI.V.getNode() <<
1221 " RL: " << VRI.RLAmt << ":" <<
1222 "\n\t\t\tisel using masking: " << NumAndInsts <<
1223 " using rotates: " << VRI.NumGroups << "\n");
1225 if (NumAndInsts >= VRI.NumGroups)
1228 DEBUG(dbgs() << "\t\t\t\tusing masking\n");
1230 if (InstCnt) *InstCnt += NumAndInsts;
1235 { VRI.V, getI32Imm(VRI.RLAmt), getI32Imm(0), getI32Imm(31) };
1236 VRot = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32,
1242 SDValue ANDIVal, ANDISVal;
1244 ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32,
1245 VRot, getI32Imm(ANDIMask)), 0);
1247 ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32,
1248 VRot, getI32Imm(ANDISMask)), 0);
1252 TotalVal = ANDISVal;
1256 TotalVal = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32,
1257 ANDIVal, ANDISVal), 0);
1262 Res = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32,
1265 // Now, remove all groups with this underlying value and rotation
1267 for (auto I = BitGroups.begin(); I != BitGroups.end();) {
1268 if (I->V == VRI.V && I->RLAmt == VRI.RLAmt)
1269 I = BitGroups.erase(I);
1276 // Instruction selection for the 32-bit case.
1277 SDNode *Select32(SDNode *N, bool LateMask, unsigned *InstCnt) {
1281 if (InstCnt) *InstCnt = 0;
1283 // Take care of cases that should use andi/andis first.
1284 SelectAndParts32(dl, Res, InstCnt);
1286 // If we've not yet selected a 'starting' instruction, and we have no zeros
1287 // to fill in, select the (Value, RLAmt) with the highest priority (largest
1288 // number of groups), and start with this rotated value.
1289 if ((!HasZeros || LateMask) && !Res) {
1290 ValueRotInfo &VRI = ValueRotsVec[0];
1292 if (InstCnt) *InstCnt += 1;
1294 { VRI.V, getI32Imm(VRI.RLAmt), getI32Imm(0), getI32Imm(31) };
1295 Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
1300 // Now, remove all groups with this underlying value and rotation factor.
1301 for (auto I = BitGroups.begin(); I != BitGroups.end();) {
1302 if (I->V == VRI.V && I->RLAmt == VRI.RLAmt)
1303 I = BitGroups.erase(I);
1309 if (InstCnt) *InstCnt += BitGroups.size();
1311 // Insert the other groups (one at a time).
1312 for (auto &BG : BitGroups) {
1315 { BG.V, getI32Imm(BG.RLAmt), getI32Imm(Bits.size() - BG.EndIdx - 1),
1316 getI32Imm(Bits.size() - BG.StartIdx - 1) };
1317 Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
1320 { Res, BG.V, getI32Imm(BG.RLAmt), getI32Imm(Bits.size() - BG.EndIdx - 1),
1321 getI32Imm(Bits.size() - BG.StartIdx - 1) };
1322 Res = SDValue(CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops), 0);
1327 unsigned Mask = (unsigned) getZerosMask();
1329 unsigned ANDIMask = (Mask & UINT16_MAX), ANDISMask = Mask >> 16;
1330 assert((ANDIMask != 0 || ANDISMask != 0) &&
1331 "No set bits in zeros mask?");
1333 if (InstCnt) *InstCnt += (unsigned) (ANDIMask != 0) +
1334 (unsigned) (ANDISMask != 0) +
1335 (unsigned) (ANDIMask != 0 && ANDISMask != 0);
1337 SDValue ANDIVal, ANDISVal;
1339 ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32,
1340 Res, getI32Imm(ANDIMask)), 0);
1342 ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32,
1343 Res, getI32Imm(ANDISMask)), 0);
1350 Res = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32,
1351 ANDIVal, ANDISVal), 0);
1354 return Res.getNode();
1357 unsigned SelectRotMask64Count(unsigned RLAmt, bool Repl32,
1358 unsigned MaskStart, unsigned MaskEnd,
1360 // In the notation used by the instructions, 'start' and 'end' are reversed
1361 // because bits are counted from high to low order.
1362 unsigned InstMaskStart = 64 - MaskEnd - 1,
1363 InstMaskEnd = 64 - MaskStart - 1;
1368 if ((!IsIns && (InstMaskEnd == 63 || InstMaskStart == 0)) ||
1369 InstMaskEnd == 63 - RLAmt)
1375 // For 64-bit values, not all combinations of rotates and masks are
1376 // available. Produce one if it is available.
1377 SDValue SelectRotMask64(SDValue V, SDLoc dl, unsigned RLAmt, bool Repl32,
1378 unsigned MaskStart, unsigned MaskEnd,
1379 unsigned *InstCnt = nullptr) {
1380 // In the notation used by the instructions, 'start' and 'end' are reversed
1381 // because bits are counted from high to low order.
1382 unsigned InstMaskStart = 64 - MaskEnd - 1,
1383 InstMaskEnd = 64 - MaskStart - 1;
1385 if (InstCnt) *InstCnt += 1;
1388 // This rotation amount assumes that the lower 32 bits of the quantity
1389 // are replicated in the high 32 bits by the rotation operator (which is
1390 // done by rlwinm and friends).
1391 assert(InstMaskStart >= 32 && "Mask cannot start out of range");
1392 assert(InstMaskEnd >= 32 && "Mask cannot end out of range");
1394 { V, getI32Imm(RLAmt), getI32Imm(InstMaskStart - 32),
1395 getI32Imm(InstMaskEnd - 32) };
1396 return SDValue(CurDAG->getMachineNode(PPC::RLWINM8, dl, MVT::i64,
1400 if (InstMaskEnd == 63) {
1402 { V, getI32Imm(RLAmt), getI32Imm(InstMaskStart) };
1403 return SDValue(CurDAG->getMachineNode(PPC::RLDICL, dl, MVT::i64, Ops), 0);
1406 if (InstMaskStart == 0) {
1408 { V, getI32Imm(RLAmt), getI32Imm(InstMaskEnd) };
1409 return SDValue(CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, Ops), 0);
1412 if (InstMaskEnd == 63 - RLAmt) {
1414 { V, getI32Imm(RLAmt), getI32Imm(InstMaskStart) };
1415 return SDValue(CurDAG->getMachineNode(PPC::RLDIC, dl, MVT::i64, Ops), 0);
1418 // We cannot do this with a single instruction, so we'll use two. The
1419 // problem is that we're not free to choose both a rotation amount and mask
1420 // start and end independently. We can choose an arbitrary mask start and
1421 // end, but then the rotation amount is fixed. Rotation, however, can be
1422 // inverted, and so by applying an "inverse" rotation first, we can get the
1424 if (InstCnt) *InstCnt += 1;
1426 // The rotation mask for the second instruction must be MaskStart.
1427 unsigned RLAmt2 = MaskStart;
1428 // The first instruction must rotate V so that the overall rotation amount
1430 unsigned RLAmt1 = (64 + RLAmt - RLAmt2) % 64;
1432 V = SelectRotMask64(V, dl, RLAmt1, false, 0, 63);
1433 return SelectRotMask64(V, dl, RLAmt2, false, MaskStart, MaskEnd);
1436 // For 64-bit values, not all combinations of rotates and masks are
1437 // available. Produce a rotate-mask-and-insert if one is available.
1438 SDValue SelectRotMaskIns64(SDValue Base, SDValue V, SDLoc dl, unsigned RLAmt,
1439 bool Repl32, unsigned MaskStart,
1440 unsigned MaskEnd, unsigned *InstCnt = nullptr) {
1441 // In the notation used by the instructions, 'start' and 'end' are reversed
1442 // because bits are counted from high to low order.
1443 unsigned InstMaskStart = 64 - MaskEnd - 1,
1444 InstMaskEnd = 64 - MaskStart - 1;
1446 if (InstCnt) *InstCnt += 1;
1449 // This rotation amount assumes that the lower 32 bits of the quantity
1450 // are replicated in the high 32 bits by the rotation operator (which is
1451 // done by rlwinm and friends).
1452 assert(InstMaskStart >= 32 && "Mask cannot start out of range");
1453 assert(InstMaskEnd >= 32 && "Mask cannot end out of range");
1455 { Base, V, getI32Imm(RLAmt), getI32Imm(InstMaskStart - 32),
1456 getI32Imm(InstMaskEnd - 32) };
1457 return SDValue(CurDAG->getMachineNode(PPC::RLWIMI8, dl, MVT::i64,
1461 if (InstMaskEnd == 63 - RLAmt) {
1463 { Base, V, getI32Imm(RLAmt), getI32Imm(InstMaskStart) };
1464 return SDValue(CurDAG->getMachineNode(PPC::RLDIMI, dl, MVT::i64, Ops), 0);
1467 // We cannot do this with a single instruction, so we'll use two. The
1468 // problem is that we're not free to choose both a rotation amount and mask
1469 // start and end independently. We can choose an arbitrary mask start and
1470 // end, but then the rotation amount is fixed. Rotation, however, can be
1471 // inverted, and so by applying an "inverse" rotation first, we can get the
1473 if (InstCnt) *InstCnt += 1;
1475 // The rotation mask for the second instruction must be MaskStart.
1476 unsigned RLAmt2 = MaskStart;
1477 // The first instruction must rotate V so that the overall rotation amount
1479 unsigned RLAmt1 = (64 + RLAmt - RLAmt2) % 64;
1481 V = SelectRotMask64(V, dl, RLAmt1, false, 0, 63);
1482 return SelectRotMaskIns64(Base, V, dl, RLAmt2, false, MaskStart, MaskEnd);
1485 void SelectAndParts64(SDLoc dl, SDValue &Res, unsigned *InstCnt) {
1486 if (BPermRewriterNoMasking)
1489 // The idea here is the same as in the 32-bit version, but with additional
1490 // complications from the fact that Repl32 might be true. Because we
1491 // aggressively convert bit groups to Repl32 form (which, for small
1492 // rotation factors, involves no other change), and then coalesce, it might
1493 // be the case that a single 64-bit masking operation could handle both
1494 // some Repl32 groups and some non-Repl32 groups. If converting to Repl32
1495 // form allowed coalescing, then we must use a 32-bit rotaton in order to
1496 // completely capture the new combined bit group.
1498 for (ValueRotInfo &VRI : ValueRotsVec) {
1501 // We need to add to the mask all bits from the associated bit groups.
1502 // If Repl32 is false, we need to add bits from bit groups that have
1503 // Repl32 true, but are trivially convertable to Repl32 false. Such a
1504 // group is trivially convertable if it overlaps only with the lower 32
1505 // bits, and the group has not been coalesced.
1506 auto MatchingBG = [VRI](BitGroup &BG) {
1510 unsigned EffRLAmt = BG.RLAmt;
1511 if (!VRI.Repl32 && BG.Repl32) {
1512 if (BG.StartIdx < 32 && BG.EndIdx < 32 && BG.StartIdx <= BG.EndIdx &&
1513 !BG.Repl32Coalesced) {
1519 } else if (VRI.Repl32 != BG.Repl32) {
1523 if (VRI.RLAmt != EffRLAmt)
1529 for (auto &BG : BitGroups) {
1530 if (!MatchingBG(BG))
1533 if (BG.StartIdx <= BG.EndIdx) {
1534 for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i)
1535 Mask |= (UINT64_C(1) << i);
1537 for (unsigned i = BG.StartIdx; i < Bits.size(); ++i)
1538 Mask |= (UINT64_C(1) << i);
1539 for (unsigned i = 0; i <= BG.EndIdx; ++i)
1540 Mask |= (UINT64_C(1) << i);
1544 // We can use the 32-bit andi/andis technique if the mask does not
1545 // require any higher-order bits. This can save an instruction compared
1546 // to always using the general 64-bit technique.
1547 bool Use32BitInsts = isUInt<32>(Mask);
1548 // Compute the masks for andi/andis that would be necessary.
1549 unsigned ANDIMask = (Mask & UINT16_MAX),
1550 ANDISMask = (Mask >> 16) & UINT16_MAX;
1552 bool NeedsRotate = VRI.RLAmt || (VRI.Repl32 && !isUInt<32>(Mask));
1554 unsigned NumAndInsts = (unsigned) NeedsRotate +
1555 (unsigned) (bool) Res;
1557 NumAndInsts += (unsigned) (ANDIMask != 0) + (unsigned) (ANDISMask != 0) +
1558 (unsigned) (ANDIMask != 0 && ANDISMask != 0);
1560 NumAndInsts += SelectInt64Count(Mask) + /* and */ 1;
1562 unsigned NumRLInsts = 0;
1563 bool FirstBG = true;
1564 for (auto &BG : BitGroups) {
1565 if (!MatchingBG(BG))
1568 SelectRotMask64Count(BG.RLAmt, BG.Repl32, BG.StartIdx, BG.EndIdx,
1573 DEBUG(dbgs() << "\t\trotation groups for " << VRI.V.getNode() <<
1574 " RL: " << VRI.RLAmt << (VRI.Repl32 ? " (32):" : ":") <<
1575 "\n\t\t\tisel using masking: " << NumAndInsts <<
1576 " using rotates: " << NumRLInsts << "\n");
1578 // When we'd use andi/andis, we bias toward using the rotates (andi only
1579 // has a record form, and is cracked on POWER cores). However, when using
1580 // general 64-bit constant formation, bias toward the constant form,
1581 // because that exposes more opportunities for CSE.
1582 if (NumAndInsts > NumRLInsts)
1584 if (Use32BitInsts && NumAndInsts == NumRLInsts)
1587 DEBUG(dbgs() << "\t\t\t\tusing masking\n");
1589 if (InstCnt) *InstCnt += NumAndInsts;
1592 // We actually need to generate a rotation if we have a non-zero rotation
1593 // factor or, in the Repl32 case, if we care about any of the
1594 // higher-order replicated bits. In the latter case, we generate a mask
1595 // backward so that it actually includes the entire 64 bits.
1596 if (VRI.RLAmt || (VRI.Repl32 && !isUInt<32>(Mask)))
1597 VRot = SelectRotMask64(VRI.V, dl, VRI.RLAmt, VRI.Repl32,
1598 VRI.Repl32 ? 31 : 0, VRI.Repl32 ? 30 : 63);
1603 if (Use32BitInsts) {
1604 assert((ANDIMask != 0 || ANDISMask != 0) &&
1605 "No set bits in mask when using 32-bit ands for 64-bit value");
1607 SDValue ANDIVal, ANDISVal;
1609 ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64,
1610 VRot, getI32Imm(ANDIMask)), 0);
1612 ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64,
1613 VRot, getI32Imm(ANDISMask)), 0);
1616 TotalVal = ANDISVal;
1620 TotalVal = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64,
1621 ANDIVal, ANDISVal), 0);
1623 TotalVal = SDValue(SelectInt64(CurDAG, dl, Mask), 0);
1625 SDValue(CurDAG->getMachineNode(PPC::AND8, dl, MVT::i64,
1626 VRot, TotalVal), 0);
1632 Res = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64,
1635 // Now, remove all groups with this underlying value and rotation
1637 for (auto I = BitGroups.begin(); I != BitGroups.end();) {
1639 I = BitGroups.erase(I);
1646 // Instruction selection for the 64-bit case.
1647 SDNode *Select64(SDNode *N, bool LateMask, unsigned *InstCnt) {
1651 if (InstCnt) *InstCnt = 0;
1653 // Take care of cases that should use andi/andis first.
1654 SelectAndParts64(dl, Res, InstCnt);
1656 // If we've not yet selected a 'starting' instruction, and we have no zeros
1657 // to fill in, select the (Value, RLAmt) with the highest priority (largest
1658 // number of groups), and start with this rotated value.
1659 if ((!HasZeros || LateMask) && !Res) {
1660 // If we have both Repl32 groups and non-Repl32 groups, the non-Repl32
1661 // groups will come first, and so the VRI representing the largest number
1662 // of groups might not be first (it might be the first Repl32 groups).
1663 unsigned MaxGroupsIdx = 0;
1664 if (!ValueRotsVec[0].Repl32) {
1665 for (unsigned i = 0, ie = ValueRotsVec.size(); i < ie; ++i)
1666 if (ValueRotsVec[i].Repl32) {
1667 if (ValueRotsVec[i].NumGroups > ValueRotsVec[0].NumGroups)
1673 ValueRotInfo &VRI = ValueRotsVec[MaxGroupsIdx];
1674 bool NeedsRotate = false;
1677 } else if (VRI.Repl32) {
1678 for (auto &BG : BitGroups) {
1679 if (BG.V != VRI.V || BG.RLAmt != VRI.RLAmt ||
1680 BG.Repl32 != VRI.Repl32)
1683 // We don't need a rotate if the bit group is confined to the lower
1685 if (BG.StartIdx < 32 && BG.EndIdx < 32 && BG.StartIdx < BG.EndIdx)
1694 Res = SelectRotMask64(VRI.V, dl, VRI.RLAmt, VRI.Repl32,
1695 VRI.Repl32 ? 31 : 0, VRI.Repl32 ? 30 : 63,
1700 // Now, remove all groups with this underlying value and rotation factor.
1702 for (auto I = BitGroups.begin(); I != BitGroups.end();) {
1703 if (I->V == VRI.V && I->RLAmt == VRI.RLAmt && I->Repl32 == VRI.Repl32)
1704 I = BitGroups.erase(I);
1710 // Because 64-bit rotates are more flexible than inserts, we might have a
1711 // preference regarding which one we do first (to save one instruction).
1713 for (auto I = BitGroups.begin(), IE = BitGroups.end(); I != IE; ++I) {
1714 if (SelectRotMask64Count(I->RLAmt, I->Repl32, I->StartIdx, I->EndIdx,
1716 SelectRotMask64Count(I->RLAmt, I->Repl32, I->StartIdx, I->EndIdx,
1718 if (I != BitGroups.begin()) {
1721 BitGroups.insert(BitGroups.begin(), BG);
1728 // Insert the other groups (one at a time).
1729 for (auto &BG : BitGroups) {
1731 Res = SelectRotMask64(BG.V, dl, BG.RLAmt, BG.Repl32, BG.StartIdx,
1732 BG.EndIdx, InstCnt);
1734 Res = SelectRotMaskIns64(Res, BG.V, dl, BG.RLAmt, BG.Repl32,
1735 BG.StartIdx, BG.EndIdx, InstCnt);
1739 uint64_t Mask = getZerosMask();
1741 // We can use the 32-bit andi/andis technique if the mask does not
1742 // require any higher-order bits. This can save an instruction compared
1743 // to always using the general 64-bit technique.
1744 bool Use32BitInsts = isUInt<32>(Mask);
1745 // Compute the masks for andi/andis that would be necessary.
1746 unsigned ANDIMask = (Mask & UINT16_MAX),
1747 ANDISMask = (Mask >> 16) & UINT16_MAX;
1749 if (Use32BitInsts) {
1750 assert((ANDIMask != 0 || ANDISMask != 0) &&
1751 "No set bits in mask when using 32-bit ands for 64-bit value");
1753 if (InstCnt) *InstCnt += (unsigned) (ANDIMask != 0) +
1754 (unsigned) (ANDISMask != 0) +
1755 (unsigned) (ANDIMask != 0 && ANDISMask != 0);
1757 SDValue ANDIVal, ANDISVal;
1759 ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64,
1760 Res, getI32Imm(ANDIMask)), 0);
1762 ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64,
1763 Res, getI32Imm(ANDISMask)), 0);
1770 Res = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64,
1771 ANDIVal, ANDISVal), 0);
1773 if (InstCnt) *InstCnt += SelectInt64Count(Mask) + /* and */ 1;
1775 SDValue MaskVal = SDValue(SelectInt64(CurDAG, dl, Mask), 0);
1777 SDValue(CurDAG->getMachineNode(PPC::AND8, dl, MVT::i64,
1782 return Res.getNode();
1785 SDNode *Select(SDNode *N, bool LateMask, unsigned *InstCnt = nullptr) {
1786 // Fill in BitGroups.
1787 collectBitGroups(LateMask);
1788 if (BitGroups.empty())
1791 // For 64-bit values, figure out when we can use 32-bit instructions.
1792 if (Bits.size() == 64)
1793 assignRepl32BitGroups();
1795 // Fill in ValueRotsVec.
1796 collectValueRotInfo();
1798 if (Bits.size() == 32) {
1799 return Select32(N, LateMask, InstCnt);
1801 assert(Bits.size() == 64 && "Not 64 bits here?");
1802 return Select64(N, LateMask, InstCnt);
1808 SmallVector<ValueBit, 64> Bits;
1811 SmallVector<unsigned, 64> RLAmt;
1813 SmallVector<BitGroup, 16> BitGroups;
1815 DenseMap<std::pair<SDValue, unsigned>, ValueRotInfo> ValueRots;
1816 SmallVector<ValueRotInfo, 16> ValueRotsVec;
1818 SelectionDAG *CurDAG;
1821 BitPermutationSelector(SelectionDAG *DAG)
1824 // Here we try to match complex bit permutations into a set of
1825 // rotate-and-shift/shift/and/or instructions, using a set of heuristics
1826 // known to produce optimial code for common cases (like i32 byte swapping).
1827 SDNode *Select(SDNode *N) {
1828 Bits.resize(N->getValueType(0).getSizeInBits());
1829 if (!getValueBits(SDValue(N, 0), Bits))
1832 DEBUG(dbgs() << "Considering bit-permutation-based instruction"
1833 " selection for: ");
1834 DEBUG(N->dump(CurDAG));
1836 // Fill it RLAmt and set HasZeros.
1837 computeRotationAmounts();
1840 return Select(N, false);
1842 // We currently have two techniques for handling results with zeros: early
1843 // masking (the default) and late masking. Late masking is sometimes more
1844 // efficient, but because the structure of the bit groups is different, it
1845 // is hard to tell without generating both and comparing the results. With
1846 // late masking, we ignore zeros in the resulting value when inserting each
1847 // set of bit groups, and then mask in the zeros at the end. With early
1848 // masking, we only insert the non-zero parts of the result at every step.
1850 unsigned InstCnt, InstCntLateMask;
1851 DEBUG(dbgs() << "\tEarly masking:\n");
1852 SDNode *RN = Select(N, false, &InstCnt);
1853 DEBUG(dbgs() << "\t\tisel would use " << InstCnt << " instructions\n");
1855 DEBUG(dbgs() << "\tLate masking:\n");
1856 SDNode *RNLM = Select(N, true, &InstCntLateMask);
1857 DEBUG(dbgs() << "\t\tisel would use " << InstCntLateMask <<
1860 if (InstCnt <= InstCntLateMask) {
1861 DEBUG(dbgs() << "\tUsing early-masking for isel\n");
1865 DEBUG(dbgs() << "\tUsing late-masking for isel\n");
1869 } // anonymous namespace
1871 SDNode *PPCDAGToDAGISel::SelectBitPermutation(SDNode *N) {
1872 if (N->getValueType(0) != MVT::i32 &&
1873 N->getValueType(0) != MVT::i64)
1876 if (!UseBitPermRewriter)
1879 switch (N->getOpcode()) {
1886 BitPermutationSelector BPS(CurDAG);
1887 return BPS.Select(N);
1894 /// SelectCC - Select a comparison of the specified values with the specified
1895 /// condition code, returning the CR# of the expression.
1896 SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS,
1897 ISD::CondCode CC, SDLoc dl) {
1898 // Always select the LHS.
1901 if (LHS.getValueType() == MVT::i32) {
1903 if (CC == ISD::SETEQ || CC == ISD::SETNE) {
1904 if (isInt32Immediate(RHS, Imm)) {
1905 // SETEQ/SETNE comparison with 16-bit immediate, fold it.
1906 if (isUInt<16>(Imm))
1907 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
1908 getI32Imm(Imm & 0xFFFF)), 0);
1909 // If this is a 16-bit signed immediate, fold it.
1910 if (isInt<16>((int)Imm))
1911 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
1912 getI32Imm(Imm & 0xFFFF)), 0);
1914 // For non-equality comparisons, the default code would materialize the
1915 // constant, then compare against it, like this:
1917 // ori r2, r2, 22136
1919 // Since we are just comparing for equality, we can emit this instead:
1920 // xoris r0,r3,0x1234
1921 // cmplwi cr0,r0,0x5678
1923 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS,
1924 getI32Imm(Imm >> 16)), 0);
1925 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor,
1926 getI32Imm(Imm & 0xFFFF)), 0);
1929 } else if (ISD::isUnsignedIntSetCC(CC)) {
1930 if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm))
1931 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
1932 getI32Imm(Imm & 0xFFFF)), 0);
1936 if (isIntS16Immediate(RHS, SImm))
1937 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
1938 getI32Imm((int)SImm & 0xFFFF)),
1942 } else if (LHS.getValueType() == MVT::i64) {
1944 if (CC == ISD::SETEQ || CC == ISD::SETNE) {
1945 if (isInt64Immediate(RHS.getNode(), Imm)) {
1946 // SETEQ/SETNE comparison with 16-bit immediate, fold it.
1947 if (isUInt<16>(Imm))
1948 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
1949 getI32Imm(Imm & 0xFFFF)), 0);
1950 // If this is a 16-bit signed immediate, fold it.
1952 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
1953 getI32Imm(Imm & 0xFFFF)), 0);
1955 // For non-equality comparisons, the default code would materialize the
1956 // constant, then compare against it, like this:
1958 // ori r2, r2, 22136
1960 // Since we are just comparing for equality, we can emit this instead:
1961 // xoris r0,r3,0x1234
1962 // cmpldi cr0,r0,0x5678
1964 if (isUInt<32>(Imm)) {
1965 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS,
1966 getI64Imm(Imm >> 16)), 0);
1967 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor,
1968 getI64Imm(Imm & 0xFFFF)), 0);
1972 } else if (ISD::isUnsignedIntSetCC(CC)) {
1973 if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm))
1974 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
1975 getI64Imm(Imm & 0xFFFF)), 0);
1979 if (isIntS16Immediate(RHS, SImm))
1980 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
1981 getI64Imm(SImm & 0xFFFF)),
1985 } else if (LHS.getValueType() == MVT::f32) {
1988 assert(LHS.getValueType() == MVT::f64 && "Unknown vt!");
1989 Opc = PPCSubTarget->hasVSX() ? PPC::XSCMPUDP : PPC::FCMPUD;
1991 return SDValue(CurDAG->getMachineNode(Opc, dl, MVT::i32, LHS, RHS), 0);
1994 static PPC::Predicate getPredicateForSetCC(ISD::CondCode CC) {
2000 llvm_unreachable("Should be lowered by legalize!");
2001 default: llvm_unreachable("Unknown condition!");
2003 case ISD::SETEQ: return PPC::PRED_EQ;
2005 case ISD::SETNE: return PPC::PRED_NE;
2007 case ISD::SETLT: return PPC::PRED_LT;
2009 case ISD::SETLE: return PPC::PRED_LE;
2011 case ISD::SETGT: return PPC::PRED_GT;
2013 case ISD::SETGE: return PPC::PRED_GE;
2014 case ISD::SETO: return PPC::PRED_NU;
2015 case ISD::SETUO: return PPC::PRED_UN;
2016 // These two are invalid for floating point. Assume we have int.
2017 case ISD::SETULT: return PPC::PRED_LT;
2018 case ISD::SETUGT: return PPC::PRED_GT;
2022 /// getCRIdxForSetCC - Return the index of the condition register field
2023 /// associated with the SetCC condition, and whether or not the field is
2024 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
2025 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) {
2028 default: llvm_unreachable("Unknown condition!");
2030 case ISD::SETLT: return 0; // Bit #0 = SETOLT
2032 case ISD::SETGT: return 1; // Bit #1 = SETOGT
2034 case ISD::SETEQ: return 2; // Bit #2 = SETOEQ
2035 case ISD::SETUO: return 3; // Bit #3 = SETUO
2037 case ISD::SETGE: Invert = true; return 0; // !Bit #0 = SETUGE
2039 case ISD::SETLE: Invert = true; return 1; // !Bit #1 = SETULE
2041 case ISD::SETNE: Invert = true; return 2; // !Bit #2 = SETUNE
2042 case ISD::SETO: Invert = true; return 3; // !Bit #3 = SETO
2047 llvm_unreachable("Invalid branch code: should be expanded by legalize");
2048 // These are invalid for floating point. Assume integer.
2049 case ISD::SETULT: return 0;
2050 case ISD::SETUGT: return 1;
2054 // getVCmpInst: return the vector compare instruction for the specified
2055 // vector type and condition code. Since this is for altivec specific code,
2056 // only support the altivec types (v16i8, v8i16, v4i32, and v4f32).
2057 static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC,
2058 bool HasVSX, bool &Swap, bool &Negate) {
2062 if (VecVT.isFloatingPoint()) {
2063 /* Handle some cases by swapping input operands. */
2065 case ISD::SETLE: CC = ISD::SETGE; Swap = true; break;
2066 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break;
2067 case ISD::SETOLE: CC = ISD::SETOGE; Swap = true; break;
2068 case ISD::SETOLT: CC = ISD::SETOGT; Swap = true; break;
2069 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break;
2070 case ISD::SETUGT: CC = ISD::SETULT; Swap = true; break;
2073 /* Handle some cases by negating the result. */
2075 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break;
2076 case ISD::SETUNE: CC = ISD::SETOEQ; Negate = true; break;
2077 case ISD::SETULE: CC = ISD::SETOGT; Negate = true; break;
2078 case ISD::SETULT: CC = ISD::SETOGE; Negate = true; break;
2081 /* We have instructions implementing the remaining cases. */
2085 if (VecVT == MVT::v4f32)
2086 return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP;
2087 else if (VecVT == MVT::v2f64)
2088 return PPC::XVCMPEQDP;
2092 if (VecVT == MVT::v4f32)
2093 return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP;
2094 else if (VecVT == MVT::v2f64)
2095 return PPC::XVCMPGTDP;
2099 if (VecVT == MVT::v4f32)
2100 return HasVSX ? PPC::XVCMPGESP : PPC::VCMPGEFP;
2101 else if (VecVT == MVT::v2f64)
2102 return PPC::XVCMPGEDP;
2107 llvm_unreachable("Invalid floating-point vector compare condition");
2109 /* Handle some cases by swapping input operands. */
2111 case ISD::SETGE: CC = ISD::SETLE; Swap = true; break;
2112 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break;
2113 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break;
2114 case ISD::SETULT: CC = ISD::SETUGT; Swap = true; break;
2117 /* Handle some cases by negating the result. */
2119 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break;
2120 case ISD::SETUNE: CC = ISD::SETUEQ; Negate = true; break;
2121 case ISD::SETLE: CC = ISD::SETGT; Negate = true; break;
2122 case ISD::SETULE: CC = ISD::SETUGT; Negate = true; break;
2125 /* We have instructions implementing the remaining cases. */
2129 if (VecVT == MVT::v16i8)
2130 return PPC::VCMPEQUB;
2131 else if (VecVT == MVT::v8i16)
2132 return PPC::VCMPEQUH;
2133 else if (VecVT == MVT::v4i32)
2134 return PPC::VCMPEQUW;
2137 if (VecVT == MVT::v16i8)
2138 return PPC::VCMPGTSB;
2139 else if (VecVT == MVT::v8i16)
2140 return PPC::VCMPGTSH;
2141 else if (VecVT == MVT::v4i32)
2142 return PPC::VCMPGTSW;
2145 if (VecVT == MVT::v16i8)
2146 return PPC::VCMPGTUB;
2147 else if (VecVT == MVT::v8i16)
2148 return PPC::VCMPGTUH;
2149 else if (VecVT == MVT::v4i32)
2150 return PPC::VCMPGTUW;
2155 llvm_unreachable("Invalid integer vector compare condition");
2159 SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) {
2162 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
2163 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
2164 bool isPPC64 = (PtrVT == MVT::i64);
2166 if (!PPCSubTarget->useCRBits() &&
2167 isInt32Immediate(N->getOperand(1), Imm)) {
2168 // We can codegen setcc op, imm very efficiently compared to a brcond.
2169 // Check for those cases here.
2172 SDValue Op = N->getOperand(0);
2176 Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0);
2177 SDValue Ops[] = { Op, getI32Imm(27), getI32Imm(5), getI32Imm(31) };
2178 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2183 SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
2184 Op, getI32Imm(~0U)), 0);
2185 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
2189 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
2190 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2194 SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0);
2195 T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0);
2196 SDValue Ops[] = { T, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
2197 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2200 } else if (Imm == ~0U) { // setcc op, -1
2201 SDValue Op = N->getOperand(0);
2206 Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
2207 Op, getI32Imm(1)), 0);
2208 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
2209 SDValue(CurDAG->getMachineNode(PPC::LI, dl,
2215 Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0);
2216 SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
2217 Op, getI32Imm(~0U));
2218 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0),
2219 Op, SDValue(AD, 1));
2222 SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op,
2224 SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD,
2226 SDValue Ops[] = { AN, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
2227 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2230 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
2231 Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops),
2233 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op,
2240 SDValue LHS = N->getOperand(0);
2241 SDValue RHS = N->getOperand(1);
2243 // Altivec Vector compare instructions do not set any CR register by default and
2244 // vector compare operations return the same type as the operands.
2245 if (LHS.getValueType().isVector()) {
2246 EVT VecVT = LHS.getValueType();
2248 unsigned int VCmpInst = getVCmpInst(VecVT.getSimpleVT(), CC,
2249 PPCSubTarget->hasVSX(), Swap, Negate);
2251 std::swap(LHS, RHS);
2254 SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0);
2255 return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLNOR :
2260 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS);
2263 if (PPCSubTarget->useCRBits())
2267 unsigned Idx = getCRIdxForSetCC(CC, Inv);
2268 SDValue CCReg = SelectCC(LHS, RHS, CC, dl);
2271 // Force the ccreg into CR7.
2272 SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
2274 SDValue InFlag(nullptr, 0); // Null incoming flag value.
2275 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg,
2276 InFlag).getValue(1);
2278 IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg,
2281 SDValue Ops[] = { IntCR, getI32Imm((32-(3-Idx)) & 31),
2282 getI32Imm(31), getI32Imm(31) };
2284 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2286 // Get the specified bit.
2288 SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
2289 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
2293 // Select - Convert the specified operand from a target-independent to a
2294 // target-specific node if it hasn't already been changed.
2295 SDNode *PPCDAGToDAGISel::Select(SDNode *N) {
2297 if (N->isMachineOpcode()) {
2299 return nullptr; // Already selected.
2302 // In case any misguided DAG-level optimizations form an ADD with a
2303 // TargetConstant operand, crash here instead of miscompiling (by selecting
2304 // an r+r add instead of some kind of r+i add).
2305 if (N->getOpcode() == ISD::ADD &&
2306 N->getOperand(1).getOpcode() == ISD::TargetConstant)
2307 llvm_unreachable("Invalid ADD with TargetConstant operand");
2309 // Try matching complex bit permutations before doing anything else.
2310 if (SDNode *NN = SelectBitPermutation(N))
2313 switch (N->getOpcode()) {
2316 case ISD::Constant: {
2317 if (N->getValueType(0) == MVT::i64)
2318 return SelectInt64(CurDAG, N);
2323 SDNode *SN = SelectSETCC(N);
2328 case PPCISD::GlobalBaseReg:
2329 return getGlobalBaseReg();
2331 case ISD::FrameIndex:
2332 return getFrameIndex(N, N);
2334 case PPCISD::MFOCRF: {
2335 SDValue InFlag = N->getOperand(1);
2336 return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32,
2337 N->getOperand(0), InFlag);
2340 case PPCISD::READ_TIME_BASE: {
2341 return CurDAG->getMachineNode(PPC::ReadTB, dl, MVT::i32, MVT::i32,
2342 MVT::Other, N->getOperand(0));
2345 case PPCISD::SRA_ADDZE: {
2346 SDValue N0 = N->getOperand(0);
2348 CurDAG->getTargetConstant(*cast<ConstantSDNode>(N->getOperand(1))->
2349 getConstantIntValue(), N->getValueType(0));
2350 if (N->getValueType(0) == MVT::i64) {
2352 CurDAG->getMachineNode(PPC::SRADI, dl, MVT::i64, MVT::Glue,
2354 return CurDAG->SelectNodeTo(N, PPC::ADDZE8, MVT::i64,
2355 SDValue(Op, 0), SDValue(Op, 1));
2357 assert(N->getValueType(0) == MVT::i32 &&
2358 "Expecting i64 or i32 in PPCISD::SRA_ADDZE");
2360 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
2362 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
2363 SDValue(Op, 0), SDValue(Op, 1));
2368 // Handle preincrement loads.
2369 LoadSDNode *LD = cast<LoadSDNode>(N);
2370 EVT LoadedVT = LD->getMemoryVT();
2372 // Normal loads are handled by code generated from the .td file.
2373 if (LD->getAddressingMode() != ISD::PRE_INC)
2376 SDValue Offset = LD->getOffset();
2377 if (Offset.getOpcode() == ISD::TargetConstant ||
2378 Offset.getOpcode() == ISD::TargetGlobalAddress) {
2381 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
2382 if (LD->getValueType(0) != MVT::i64) {
2383 // Handle PPC32 integer and normal FP loads.
2384 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
2385 switch (LoadedVT.getSimpleVT().SimpleTy) {
2386 default: llvm_unreachable("Invalid PPC load type!");
2387 case MVT::f64: Opcode = PPC::LFDU; break;
2388 case MVT::f32: Opcode = PPC::LFSU; break;
2389 case MVT::i32: Opcode = PPC::LWZU; break;
2390 case MVT::i16: Opcode = isSExt ? PPC::LHAU : PPC::LHZU; break;
2392 case MVT::i8: Opcode = PPC::LBZU; break;
2395 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
2396 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
2397 switch (LoadedVT.getSimpleVT().SimpleTy) {
2398 default: llvm_unreachable("Invalid PPC load type!");
2399 case MVT::i64: Opcode = PPC::LDU; break;
2400 case MVT::i32: Opcode = PPC::LWZU8; break;
2401 case MVT::i16: Opcode = isSExt ? PPC::LHAU8 : PPC::LHZU8; break;
2403 case MVT::i8: Opcode = PPC::LBZU8; break;
2407 SDValue Chain = LD->getChain();
2408 SDValue Base = LD->getBasePtr();
2409 SDValue Ops[] = { Offset, Base, Chain };
2410 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
2411 PPCLowering->getPointerTy(),
2415 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
2416 if (LD->getValueType(0) != MVT::i64) {
2417 // Handle PPC32 integer and normal FP loads.
2418 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
2419 switch (LoadedVT.getSimpleVT().SimpleTy) {
2420 default: llvm_unreachable("Invalid PPC load type!");
2421 case MVT::f64: Opcode = PPC::LFDUX; break;
2422 case MVT::f32: Opcode = PPC::LFSUX; break;
2423 case MVT::i32: Opcode = PPC::LWZUX; break;
2424 case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break;
2426 case MVT::i8: Opcode = PPC::LBZUX; break;
2429 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
2430 assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) &&
2431 "Invalid sext update load");
2432 switch (LoadedVT.getSimpleVT().SimpleTy) {
2433 default: llvm_unreachable("Invalid PPC load type!");
2434 case MVT::i64: Opcode = PPC::LDUX; break;
2435 case MVT::i32: Opcode = isSExt ? PPC::LWAUX : PPC::LWZUX8; break;
2436 case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break;
2438 case MVT::i8: Opcode = PPC::LBZUX8; break;
2442 SDValue Chain = LD->getChain();
2443 SDValue Base = LD->getBasePtr();
2444 SDValue Ops[] = { Base, Offset, Chain };
2445 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
2446 PPCLowering->getPointerTy(),
2452 unsigned Imm, Imm2, SH, MB, ME;
2455 // If this is an and of a value rotated between 0 and 31 bits and then and'd
2456 // with a mask, emit rlwinm
2457 if (isInt32Immediate(N->getOperand(1), Imm) &&
2458 isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) {
2459 SDValue Val = N->getOperand(0).getOperand(0);
2460 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
2461 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2463 // If this is just a masked value where the input is not handled above, and
2464 // is not a rotate-left (handled by a pattern in the .td file), emit rlwinm
2465 if (isInt32Immediate(N->getOperand(1), Imm) &&
2466 isRunOfOnes(Imm, MB, ME) &&
2467 N->getOperand(0).getOpcode() != ISD::ROTL) {
2468 SDValue Val = N->getOperand(0);
2469 SDValue Ops[] = { Val, getI32Imm(0), getI32Imm(MB), getI32Imm(ME) };
2470 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2472 // If this is a 64-bit zero-extension mask, emit rldicl.
2473 if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) &&
2475 SDValue Val = N->getOperand(0);
2476 MB = 64 - CountTrailingOnes_64(Imm64);
2479 // If the operand is a logical right shift, we can fold it into this
2480 // instruction: rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb)
2481 // for n <= mb. The right shift is really a left rotate followed by a
2482 // mask, and this mask is a more-restrictive sub-mask of the mask implied
2484 if (Val.getOpcode() == ISD::SRL &&
2485 isInt32Immediate(Val.getOperand(1).getNode(), Imm) && Imm <= MB) {
2486 assert(Imm < 64 && "Illegal shift amount");
2487 Val = Val.getOperand(0);
2491 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB) };
2492 return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops);
2494 // AND X, 0 -> 0, not "rlwinm 32".
2495 if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) {
2496 ReplaceUses(SDValue(N, 0), N->getOperand(1));
2499 // ISD::OR doesn't get all the bitfield insertion fun.
2500 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert
2501 if (isInt32Immediate(N->getOperand(1), Imm) &&
2502 N->getOperand(0).getOpcode() == ISD::OR &&
2503 isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) {
2506 if (isRunOfOnes(Imm, MB, ME)) {
2507 SDValue Ops[] = { N->getOperand(0).getOperand(0),
2508 N->getOperand(0).getOperand(1),
2509 getI32Imm(0), getI32Imm(MB),getI32Imm(ME) };
2510 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
2514 // Other cases are autogenerated.
2518 if (N->getValueType(0) == MVT::i32)
2519 if (SDNode *I = SelectBitfieldInsert(N))
2523 if (N->getOperand(0)->getOpcode() == ISD::FrameIndex &&
2524 isIntS16Immediate(N->getOperand(1), Imm)) {
2525 APInt LHSKnownZero, LHSKnownOne;
2526 CurDAG->computeKnownBits(N->getOperand(0), LHSKnownZero, LHSKnownOne);
2528 // If this is equivalent to an add, then we can fold it with the
2529 // FrameIndex calculation.
2530 if ((LHSKnownZero.getZExtValue()|~(uint64_t)Imm) == ~0ULL)
2531 return getFrameIndex(N, N->getOperand(0).getNode(), (int)Imm);
2534 // Other cases are autogenerated.
2539 if (N->getOperand(0)->getOpcode() == ISD::FrameIndex &&
2540 isIntS16Immediate(N->getOperand(1), Imm))
2541 return getFrameIndex(N, N->getOperand(0).getNode(), (int)Imm);
2546 unsigned Imm, SH, MB, ME;
2547 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
2548 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
2549 SDValue Ops[] = { N->getOperand(0).getOperand(0),
2550 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
2551 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2554 // Other cases are autogenerated.
2558 unsigned Imm, SH, MB, ME;
2559 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
2560 isRotateAndMask(N, Imm, true, SH, MB, ME)) {
2561 SDValue Ops[] = { N->getOperand(0).getOperand(0),
2562 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
2563 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
2566 // Other cases are autogenerated.
2569 // FIXME: Remove this once the ANDI glue bug is fixed:
2570 case PPCISD::ANDIo_1_EQ_BIT:
2571 case PPCISD::ANDIo_1_GT_BIT: {
2575 EVT InVT = N->getOperand(0).getValueType();
2576 assert((InVT == MVT::i64 || InVT == MVT::i32) &&
2577 "Invalid input type for ANDIo_1_EQ_BIT");
2579 unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDIo8 : PPC::ANDIo;
2580 SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue,
2582 CurDAG->getTargetConstant(1, InVT)), 0);
2583 SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32);
2585 CurDAG->getTargetConstant(N->getOpcode() == PPCISD::ANDIo_1_EQ_BIT ?
2586 PPC::sub_eq : PPC::sub_gt, MVT::i32);
2588 return CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1,
2590 SDValue(AndI.getNode(), 1) /* glue */);
2592 case ISD::SELECT_CC: {
2593 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
2594 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
2595 bool isPPC64 = (PtrVT == MVT::i64);
2597 // If this is a select of i1 operands, we'll pattern match it.
2598 if (PPCSubTarget->useCRBits() &&
2599 N->getOperand(0).getValueType() == MVT::i1)
2602 // Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
2604 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2605 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
2606 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
2607 if (N1C->isNullValue() && N3C->isNullValue() &&
2608 N2C->getZExtValue() == 1ULL && CC == ISD::SETNE &&
2609 // FIXME: Implement this optzn for PPC64.
2610 N->getValueType(0) == MVT::i32) {
2612 CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
2613 N->getOperand(0), getI32Imm(~0U));
2614 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32,
2615 SDValue(Tmp, 0), N->getOperand(0),
2619 SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl);
2621 if (N->getValueType(0) == MVT::i1) {
2622 // An i1 select is: (c & t) | (!c & f).
2624 unsigned Idx = getCRIdxForSetCC(CC, Inv);
2628 default: llvm_unreachable("Invalid CC index");
2629 case 0: SRI = PPC::sub_lt; break;
2630 case 1: SRI = PPC::sub_gt; break;
2631 case 2: SRI = PPC::sub_eq; break;
2632 case 3: SRI = PPC::sub_un; break;
2635 SDValue CCBit = CurDAG->getTargetExtractSubreg(SRI, dl, MVT::i1, CCReg);
2637 SDValue NotCCBit(CurDAG->getMachineNode(PPC::CRNOR, dl, MVT::i1,
2639 SDValue C = Inv ? NotCCBit : CCBit,
2640 NotC = Inv ? CCBit : NotCCBit;
2642 SDValue CAndT(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1,
2643 C, N->getOperand(2)), 0);
2644 SDValue NotCAndF(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1,
2645 NotC, N->getOperand(3)), 0);
2647 return CurDAG->SelectNodeTo(N, PPC::CROR, MVT::i1, CAndT, NotCAndF);
2650 unsigned BROpc = getPredicateForSetCC(CC);
2652 unsigned SelectCCOp;
2653 if (N->getValueType(0) == MVT::i32)
2654 SelectCCOp = PPC::SELECT_CC_I4;
2655 else if (N->getValueType(0) == MVT::i64)
2656 SelectCCOp = PPC::SELECT_CC_I8;
2657 else if (N->getValueType(0) == MVT::f32)
2658 SelectCCOp = PPC::SELECT_CC_F4;
2659 else if (N->getValueType(0) == MVT::f64)
2660 if (PPCSubTarget->hasVSX())
2661 SelectCCOp = PPC::SELECT_CC_VSFRC;
2663 SelectCCOp = PPC::SELECT_CC_F8;
2664 else if (N->getValueType(0) == MVT::v2f64 ||
2665 N->getValueType(0) == MVT::v2i64)
2666 SelectCCOp = PPC::SELECT_CC_VSRC;
2668 SelectCCOp = PPC::SELECT_CC_VRRC;
2670 SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3),
2672 return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops);
2675 if (PPCSubTarget->hasVSX()) {
2676 SDValue Ops[] = { N->getOperand(2), N->getOperand(1), N->getOperand(0) };
2677 return CurDAG->SelectNodeTo(N, PPC::XXSEL, N->getValueType(0), Ops);
2681 case ISD::VECTOR_SHUFFLE:
2682 if (PPCSubTarget->hasVSX() && (N->getValueType(0) == MVT::v2f64 ||
2683 N->getValueType(0) == MVT::v2i64)) {
2684 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
2686 SDValue Op1 = N->getOperand(SVN->getMaskElt(0) < 2 ? 0 : 1),
2687 Op2 = N->getOperand(SVN->getMaskElt(1) < 2 ? 0 : 1);
2690 for (int i = 0; i < 2; ++i)
2691 if (SVN->getMaskElt(i) <= 0 || SVN->getMaskElt(i) == 2)
2696 // For little endian, we must swap the input operands and adjust
2697 // the mask elements (reverse and invert them).
2698 if (PPCSubTarget->isLittleEndian()) {
2699 std::swap(Op1, Op2);
2700 unsigned tmp = DM[0];
2705 SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), MVT::i32);
2707 if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 &&
2708 Op1.getOpcode() == ISD::SCALAR_TO_VECTOR &&
2709 isa<LoadSDNode>(Op1.getOperand(0))) {
2710 LoadSDNode *LD = cast<LoadSDNode>(Op1.getOperand(0));
2711 SDValue Base, Offset;
2713 if (LD->isUnindexed() &&
2714 SelectAddrIdxOnly(LD->getBasePtr(), Base, Offset)) {
2715 SDValue Chain = LD->getChain();
2716 SDValue Ops[] = { Base, Offset, Chain };
2717 return CurDAG->SelectNodeTo(N, PPC::LXVDSX,
2718 N->getValueType(0), Ops);
2722 SDValue Ops[] = { Op1, Op2, DMV };
2723 return CurDAG->SelectNodeTo(N, PPC::XXPERMDI, N->getValueType(0), Ops);
2729 bool IsPPC64 = PPCSubTarget->isPPC64();
2730 SDValue Ops[] = { N->getOperand(1), N->getOperand(0) };
2731 return CurDAG->SelectNodeTo(N, N->getOpcode() == PPCISD::BDNZ ?
2732 (IsPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
2733 (IsPPC64 ? PPC::BDZ8 : PPC::BDZ),
2736 case PPCISD::COND_BRANCH: {
2737 // Op #0 is the Chain.
2738 // Op #1 is the PPC::PRED_* number.
2740 // Op #3 is the Dest MBB
2741 // Op #4 is the Flag.
2742 // Prevent PPC::PRED_* from being selected into LI.
2744 getI32Imm(cast<ConstantSDNode>(N->getOperand(1))->getZExtValue());
2745 SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3),
2746 N->getOperand(0), N->getOperand(4) };
2747 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
2750 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
2751 unsigned PCC = getPredicateForSetCC(CC);
2753 if (N->getOperand(2).getValueType() == MVT::i1) {
2757 default: llvm_unreachable("Unexpected Boolean-operand predicate");
2758 case PPC::PRED_LT: Opc = PPC::CRANDC; Swap = true; break;
2759 case PPC::PRED_LE: Opc = PPC::CRORC; Swap = true; break;
2760 case PPC::PRED_EQ: Opc = PPC::CREQV; Swap = false; break;
2761 case PPC::PRED_GE: Opc = PPC::CRORC; Swap = false; break;
2762 case PPC::PRED_GT: Opc = PPC::CRANDC; Swap = false; break;
2763 case PPC::PRED_NE: Opc = PPC::CRXOR; Swap = false; break;
2766 SDValue BitComp(CurDAG->getMachineNode(Opc, dl, MVT::i1,
2767 N->getOperand(Swap ? 3 : 2),
2768 N->getOperand(Swap ? 2 : 3)), 0);
2769 return CurDAG->SelectNodeTo(N, PPC::BC, MVT::Other,
2770 BitComp, N->getOperand(4), N->getOperand(0));
2773 SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl);
2774 SDValue Ops[] = { getI32Imm(PCC), CondCode,
2775 N->getOperand(4), N->getOperand(0) };
2776 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
2779 // FIXME: Should custom lower this.
2780 SDValue Chain = N->getOperand(0);
2781 SDValue Target = N->getOperand(1);
2782 unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8;
2783 unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8;
2784 Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Target,
2786 return CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain);
2788 case PPCISD::TOC_ENTRY: {
2789 assert ((PPCSubTarget->isPPC64() || PPCSubTarget->isSVR4ABI()) &&
2790 "Only supported for 64-bit ABI and 32-bit SVR4");
2791 if (PPCSubTarget->isSVR4ABI() && !PPCSubTarget->isPPC64()) {
2792 SDValue GA = N->getOperand(0);
2793 return CurDAG->getMachineNode(PPC::LWZtoc, dl, MVT::i32, GA,
2797 // For medium and large code model, we generate two instructions as
2798 // described below. Otherwise we allow SelectCodeCommon to handle this,
2799 // selecting one of LDtoc, LDtocJTI, LDtocCPT, and LDtocBA.
2800 CodeModel::Model CModel = TM.getCodeModel();
2801 if (CModel != CodeModel::Medium && CModel != CodeModel::Large)
2804 // The first source operand is a TargetGlobalAddress or a TargetJumpTable.
2805 // If it is an externally defined symbol, a symbol with common linkage,
2806 // a non-local function address, or a jump table address, or if we are
2807 // generating code for large code model, we generate:
2808 // LDtocL(<ga:@sym>, ADDIStocHA(%X2, <ga:@sym>))
2809 // Otherwise we generate:
2810 // ADDItocL(ADDIStocHA(%X2, <ga:@sym>), <ga:@sym>)
2811 SDValue GA = N->getOperand(0);
2812 SDValue TOCbase = N->getOperand(1);
2813 SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64,
2816 if (isa<JumpTableSDNode>(GA) || isa<BlockAddressSDNode>(GA) ||
2817 CModel == CodeModel::Large)
2818 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
2821 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) {
2822 const GlobalValue *GValue = G->getGlobal();
2823 if ((GValue->getType()->getElementType()->isFunctionTy() &&
2824 (GValue->isDeclaration() || GValue->isWeakForLinker())) ||
2825 GValue->isDeclaration() || GValue->hasCommonLinkage() ||
2826 GValue->hasAvailableExternallyLinkage())
2827 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
2831 return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64,
2832 SDValue(Tmp, 0), GA);
2834 case PPCISD::PPC32_PICGOT: {
2835 // Generate a PIC-safe GOT reference.
2836 assert(!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI() &&
2837 "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4");
2838 return CurDAG->SelectNodeTo(N, PPC::PPC32PICGOT, PPCLowering->getPointerTy(), MVT::i32);
2840 case PPCISD::VADD_SPLAT: {
2841 // This expands into one of three sequences, depending on whether
2842 // the first operand is odd or even, positive or negative.
2843 assert(isa<ConstantSDNode>(N->getOperand(0)) &&
2844 isa<ConstantSDNode>(N->getOperand(1)) &&
2845 "Invalid operand on VADD_SPLAT!");
2847 int Elt = N->getConstantOperandVal(0);
2848 int EltSize = N->getConstantOperandVal(1);
2849 unsigned Opc1, Opc2, Opc3;
2853 Opc1 = PPC::VSPLTISB;
2854 Opc2 = PPC::VADDUBM;
2855 Opc3 = PPC::VSUBUBM;
2857 } else if (EltSize == 2) {
2858 Opc1 = PPC::VSPLTISH;
2859 Opc2 = PPC::VADDUHM;
2860 Opc3 = PPC::VSUBUHM;
2863 assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!");
2864 Opc1 = PPC::VSPLTISW;
2865 Opc2 = PPC::VADDUWM;
2866 Opc3 = PPC::VSUBUWM;
2870 if ((Elt & 1) == 0) {
2871 // Elt is even, in the range [-32,-18] + [16,30].
2873 // Convert: VADD_SPLAT elt, size
2874 // Into: tmp = VSPLTIS[BHW] elt
2875 // VADDU[BHW]M tmp, tmp
2876 // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4
2877 SDValue EltVal = getI32Imm(Elt >> 1);
2878 SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
2879 SDValue TmpVal = SDValue(Tmp, 0);
2880 return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal);
2882 } else if (Elt > 0) {
2883 // Elt is odd and positive, in the range [17,31].
2885 // Convert: VADD_SPLAT elt, size
2886 // Into: tmp1 = VSPLTIS[BHW] elt-16
2887 // tmp2 = VSPLTIS[BHW] -16
2888 // VSUBU[BHW]M tmp1, tmp2
2889 SDValue EltVal = getI32Imm(Elt - 16);
2890 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
2891 EltVal = getI32Imm(-16);
2892 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
2893 return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0),
2897 // Elt is odd and negative, in the range [-31,-17].
2899 // Convert: VADD_SPLAT elt, size
2900 // Into: tmp1 = VSPLTIS[BHW] elt+16
2901 // tmp2 = VSPLTIS[BHW] -16
2902 // VADDU[BHW]M tmp1, tmp2
2903 SDValue EltVal = getI32Imm(Elt + 16);
2904 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
2905 EltVal = getI32Imm(-16);
2906 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
2907 return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0),
2913 return SelectCode(N);
2916 // If the target supports the cmpb instruction, do the idiom recognition here.
2917 // We don't do this as a DAG combine because we don't want to do it as nodes
2918 // are being combined (because we might miss part of the eventual idiom). We
2919 // don't want to do it during instruction selection because we want to reuse
2920 // the logic for lowering the masking operations already part of the
2921 // instruction selector.
2922 SDValue PPCDAGToDAGISel::combineToCMPB(SDNode *N) {
2925 assert(N->getOpcode() == ISD::OR &&
2926 "Only OR nodes are supported for CMPB");
2929 if (!PPCSubTarget->hasCMPB())
2932 if (N->getValueType(0) != MVT::i32 &&
2933 N->getValueType(0) != MVT::i64)
2936 EVT VT = N->getValueType(0);
2939 bool BytesFound[8] = { 0, 0, 0, 0, 0, 0, 0, 0 };
2940 uint64_t Mask = 0, Alt = 0;
2942 auto IsByteSelectCC = [this](SDValue O, unsigned &b,
2943 uint64_t &Mask, uint64_t &Alt,
2944 SDValue &LHS, SDValue &RHS) {
2945 if (O.getOpcode() != ISD::SELECT_CC)
2947 ISD::CondCode CC = cast<CondCodeSDNode>(O.getOperand(4))->get();
2949 if (!isa<ConstantSDNode>(O.getOperand(2)) ||
2950 !isa<ConstantSDNode>(O.getOperand(3)))
2953 uint64_t PM = O.getConstantOperandVal(2);
2954 uint64_t PAlt = O.getConstantOperandVal(3);
2955 for (b = 0; b < 8; ++b) {
2956 uint64_t Mask = UINT64_C(0xFF) << (8*b);
2957 if (PM && (PM & Mask) == PM && (PAlt & Mask) == PAlt)
2966 if (!isa<ConstantSDNode>(O.getOperand(1)) ||
2967 O.getConstantOperandVal(1) != 0) {
2968 SDValue Op0 = O.getOperand(0), Op1 = O.getOperand(1);
2969 if (Op0.getOpcode() == ISD::TRUNCATE)
2970 Op0 = Op0.getOperand(0);
2971 if (Op1.getOpcode() == ISD::TRUNCATE)
2972 Op1 = Op1.getOperand(0);
2974 if (Op0.getOpcode() == ISD::SRL && Op1.getOpcode() == ISD::SRL &&
2975 Op0.getOperand(1) == Op1.getOperand(1) && CC == ISD::SETEQ &&
2976 isa<ConstantSDNode>(Op0.getOperand(1))) {
2978 unsigned Bits = Op0.getValueType().getSizeInBits();
2981 if (Op0.getConstantOperandVal(1) != Bits-8)
2984 LHS = Op0.getOperand(0);
2985 RHS = Op1.getOperand(0);
2989 // When we have small integers (i16 to be specific), the form present
2990 // post-legalization uses SETULT in the SELECT_CC for the
2991 // higher-order byte, depending on the fact that the
2992 // even-higher-order bytes are known to all be zero, for example:
2993 // select_cc (xor $lhs, $rhs), 256, 65280, 0, setult
2994 // (so when the second byte is the same, because all higher-order
2995 // bits from bytes 3 and 4 are known to be zero, the result of the
2996 // xor can be at most 255)
2997 if (Op0.getOpcode() == ISD::XOR && CC == ISD::SETULT &&
2998 isa<ConstantSDNode>(O.getOperand(1))) {
3000 uint64_t ULim = O.getConstantOperandVal(1);
3001 if (ULim != (UINT64_C(1) << b*8))
3004 // Now we need to make sure that the upper bytes are known to be
3006 unsigned Bits = Op0.getValueType().getSizeInBits();
3007 if (!CurDAG->MaskedValueIsZero(Op0,
3008 APInt::getHighBitsSet(Bits, Bits - (b+1)*8)))
3011 LHS = Op0.getOperand(0);
3012 RHS = Op0.getOperand(1);
3019 if (CC != ISD::SETEQ)
3022 SDValue Op = O.getOperand(0);
3023 if (Op.getOpcode() == ISD::AND) {
3024 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3026 if (Op.getConstantOperandVal(1) != (UINT64_C(0xFF) << (8*b)))
3029 SDValue XOR = Op.getOperand(0);
3030 if (XOR.getOpcode() == ISD::TRUNCATE)
3031 XOR = XOR.getOperand(0);
3032 if (XOR.getOpcode() != ISD::XOR)
3035 LHS = XOR.getOperand(0);
3036 RHS = XOR.getOperand(1);
3038 } else if (Op.getOpcode() == ISD::SRL) {
3039 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3041 unsigned Bits = Op.getValueType().getSizeInBits();
3044 if (Op.getConstantOperandVal(1) != Bits-8)
3047 SDValue XOR = Op.getOperand(0);
3048 if (XOR.getOpcode() == ISD::TRUNCATE)
3049 XOR = XOR.getOperand(0);
3050 if (XOR.getOpcode() != ISD::XOR)
3053 LHS = XOR.getOperand(0);
3054 RHS = XOR.getOperand(1);
3061 SmallVector<SDValue, 8> Queue(1, SDValue(N, 0));
3062 while (!Queue.empty()) {
3063 SDValue V = Queue.pop_back_val();
3065 for (const SDValue &O : V.getNode()->ops()) {
3067 uint64_t M = 0, A = 0;
3069 if (O.getOpcode() == ISD::OR) {
3071 } else if (IsByteSelectCC(O, b, M, A, OLHS, ORHS)) {
3075 BytesFound[b] = true;
3078 } else if ((LHS == ORHS && RHS == OLHS) ||
3079 (RHS == ORHS && LHS == OLHS)) {
3080 BytesFound[b] = true;
3092 unsigned LastB = 0, BCnt = 0;
3093 for (unsigned i = 0; i < 8; ++i)
3094 if (BytesFound[LastB]) {
3099 if (!LastB || BCnt < 2)
3102 // Because we'll be zero-extending the output anyway if don't have a specific
3103 // value for each input byte (via the Mask), we can 'anyext' the inputs.
3104 if (LHS.getValueType() != VT) {
3105 LHS = CurDAG->getAnyExtOrTrunc(LHS, dl, VT);
3106 RHS = CurDAG->getAnyExtOrTrunc(RHS, dl, VT);
3109 Res = CurDAG->getNode(PPCISD::CMPB, dl, VT, LHS, RHS);
3111 bool NonTrivialMask = ((int64_t) Mask) != INT64_C(-1);
3112 if (NonTrivialMask && !Alt) {
3113 // Res = Mask & CMPB
3114 Res = CurDAG->getNode(ISD::AND, dl, VT, Res, CurDAG->getConstant(Mask, VT));
3116 // Res = (CMPB & Mask) | (~CMPB & Alt)
3117 // Which, as suggested here:
3118 // https://graphics.stanford.edu/~seander/bithacks.html#MaskedMerge
3119 // can be written as:
3120 // Res = Alt ^ ((Alt ^ Mask) & CMPB)
3121 // useful because the (Alt ^ Mask) can be pre-computed.
3122 Res = CurDAG->getNode(ISD::AND, dl, VT, Res,
3123 CurDAG->getConstant(Mask ^ Alt, VT));
3124 Res = CurDAG->getNode(ISD::XOR, dl, VT, Res, CurDAG->getConstant(Alt, VT));
3130 void PPCDAGToDAGISel::PreprocessISelDAG() {
3131 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
3134 bool MadeChange = false;
3135 while (Position != CurDAG->allnodes_begin()) {
3136 SDNode *N = --Position;
3141 switch (N->getOpcode()) {
3144 Res = combineToCMPB(N);
3149 DEBUG(dbgs() << "PPC DAG preprocessing replacing:\nOld: ");
3150 DEBUG(N->dump(CurDAG));
3151 DEBUG(dbgs() << "\nNew: ");
3152 DEBUG(Res.getNode()->dump(CurDAG));
3153 DEBUG(dbgs() << "\n");
3155 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Res);
3161 CurDAG->RemoveDeadNodes();
3164 /// PostprocessISelDAG - Perform some late peephole optimizations
3165 /// on the DAG representation.
3166 void PPCDAGToDAGISel::PostprocessISelDAG() {
3168 // Skip peepholes at -O0.
3169 if (TM.getOptLevel() == CodeGenOpt::None)
3174 PeepholePPC64ZExt();
3177 // Check if all users of this node will become isel where the second operand
3178 // is the constant zero. If this is so, and if we can negate the condition,
3179 // then we can flip the true and false operands. This will allow the zero to
3180 // be folded with the isel so that we don't need to materialize a register
3182 bool PPCDAGToDAGISel::AllUsersSelectZero(SDNode *N) {
3183 // If we're not using isel, then this does not matter.
3184 if (!PPCSubTarget->hasISEL())
3187 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
3190 if (!User->isMachineOpcode())
3192 if (User->getMachineOpcode() != PPC::SELECT_I4 &&
3193 User->getMachineOpcode() != PPC::SELECT_I8)
3196 SDNode *Op2 = User->getOperand(2).getNode();
3197 if (!Op2->isMachineOpcode())
3200 if (Op2->getMachineOpcode() != PPC::LI &&
3201 Op2->getMachineOpcode() != PPC::LI8)
3204 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op2->getOperand(0));
3208 if (!C->isNullValue())
3215 void PPCDAGToDAGISel::SwapAllSelectUsers(SDNode *N) {
3216 SmallVector<SDNode *, 4> ToReplace;
3217 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
3220 assert((User->getMachineOpcode() == PPC::SELECT_I4 ||
3221 User->getMachineOpcode() == PPC::SELECT_I8) &&
3222 "Must have all select users");
3223 ToReplace.push_back(User);
3226 for (SmallVector<SDNode *, 4>::iterator UI = ToReplace.begin(),
3227 UE = ToReplace.end(); UI != UE; ++UI) {
3230 CurDAG->getMachineNode(User->getMachineOpcode(), SDLoc(User),
3231 User->getValueType(0), User->getOperand(0),
3232 User->getOperand(2),
3233 User->getOperand(1));
3235 DEBUG(dbgs() << "CR Peephole replacing:\nOld: ");
3236 DEBUG(User->dump(CurDAG));
3237 DEBUG(dbgs() << "\nNew: ");
3238 DEBUG(ResNode->dump(CurDAG));
3239 DEBUG(dbgs() << "\n");
3241 ReplaceUses(User, ResNode);
3245 void PPCDAGToDAGISel::PeepholeCROps() {
3249 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
3250 E = CurDAG->allnodes_end(); I != E; ++I) {
3251 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(I);
3252 if (!MachineNode || MachineNode->use_empty())
3254 SDNode *ResNode = MachineNode;
3256 bool Op1Set = false, Op1Unset = false,
3258 Op2Set = false, Op2Unset = false,
3261 unsigned Opcode = MachineNode->getMachineOpcode();
3272 SDValue Op = MachineNode->getOperand(1);
3273 if (Op.isMachineOpcode()) {
3274 if (Op.getMachineOpcode() == PPC::CRSET)
3276 else if (Op.getMachineOpcode() == PPC::CRUNSET)
3278 else if (Op.getMachineOpcode() == PPC::CRNOR &&
3279 Op.getOperand(0) == Op.getOperand(1))
3285 case PPC::SELECT_I4:
3286 case PPC::SELECT_I8:
3287 case PPC::SELECT_F4:
3288 case PPC::SELECT_F8:
3289 case PPC::SELECT_VRRC:
3290 case PPC::SELECT_VSFRC:
3291 case PPC::SELECT_VSRC: {
3292 SDValue Op = MachineNode->getOperand(0);
3293 if (Op.isMachineOpcode()) {
3294 if (Op.getMachineOpcode() == PPC::CRSET)
3296 else if (Op.getMachineOpcode() == PPC::CRUNSET)
3298 else if (Op.getMachineOpcode() == PPC::CRNOR &&
3299 Op.getOperand(0) == Op.getOperand(1))
3306 bool SelectSwap = false;
3310 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3312 ResNode = MachineNode->getOperand(0).getNode();
3315 ResNode = MachineNode->getOperand(1).getNode();
3318 ResNode = MachineNode->getOperand(0).getNode();
3319 else if (Op1Unset || Op2Unset)
3320 // x & 0 = 0 & y = 0
3321 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
3324 // ~x & y = andc(y, x)
3325 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
3326 MVT::i1, MachineNode->getOperand(1),
3327 MachineNode->getOperand(0).
3330 // x & ~y = andc(x, y)
3331 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
3332 MVT::i1, MachineNode->getOperand(0),
3333 MachineNode->getOperand(1).
3335 else if (AllUsersSelectZero(MachineNode))
3336 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode),
3337 MVT::i1, MachineNode->getOperand(0),
3338 MachineNode->getOperand(1)),
3342 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3343 // nand(x, x) -> nor(x, x)
3344 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3345 MVT::i1, MachineNode->getOperand(0),
3346 MachineNode->getOperand(0));
3348 // nand(1, y) -> nor(y, y)
3349 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3350 MVT::i1, MachineNode->getOperand(1),
3351 MachineNode->getOperand(1));
3353 // nand(x, 1) -> nor(x, x)
3354 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3355 MVT::i1, MachineNode->getOperand(0),
3356 MachineNode->getOperand(0));
3357 else if (Op1Unset || Op2Unset)
3358 // nand(x, 0) = nand(0, y) = 1
3359 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
3362 // nand(~x, y) = ~(~x & y) = x | ~y = orc(x, y)
3363 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
3364 MVT::i1, MachineNode->getOperand(0).
3366 MachineNode->getOperand(1));
3368 // nand(x, ~y) = ~x | y = orc(y, x)
3369 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
3370 MVT::i1, MachineNode->getOperand(1).
3372 MachineNode->getOperand(0));
3373 else if (AllUsersSelectZero(MachineNode))
3374 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode),
3375 MVT::i1, MachineNode->getOperand(0),
3376 MachineNode->getOperand(1)),
3380 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3382 ResNode = MachineNode->getOperand(0).getNode();
3383 else if (Op1Set || Op2Set)
3384 // x | 1 = 1 | y = 1
3385 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
3389 ResNode = MachineNode->getOperand(1).getNode();
3392 ResNode = MachineNode->getOperand(0).getNode();
3394 // ~x | y = orc(y, x)
3395 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
3396 MVT::i1, MachineNode->getOperand(1),
3397 MachineNode->getOperand(0).
3400 // x | ~y = orc(x, y)
3401 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
3402 MVT::i1, MachineNode->getOperand(0),
3403 MachineNode->getOperand(1).
3405 else if (AllUsersSelectZero(MachineNode))
3406 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3407 MVT::i1, MachineNode->getOperand(0),
3408 MachineNode->getOperand(1)),
3412 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3414 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
3417 // xor(1, y) -> nor(y, y)
3418 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3419 MVT::i1, MachineNode->getOperand(1),
3420 MachineNode->getOperand(1));
3422 // xor(x, 1) -> nor(x, x)
3423 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3424 MVT::i1, MachineNode->getOperand(0),
3425 MachineNode->getOperand(0));
3428 ResNode = MachineNode->getOperand(1).getNode();
3431 ResNode = MachineNode->getOperand(0).getNode();
3433 // xor(~x, y) = eqv(x, y)
3434 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
3435 MVT::i1, MachineNode->getOperand(0).
3437 MachineNode->getOperand(1));
3439 // xor(x, ~y) = eqv(x, y)
3440 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
3441 MVT::i1, MachineNode->getOperand(0),
3442 MachineNode->getOperand(1).
3444 else if (AllUsersSelectZero(MachineNode))
3445 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
3446 MVT::i1, MachineNode->getOperand(0),
3447 MachineNode->getOperand(1)),
3451 if (Op1Set || Op2Set)
3453 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
3456 // nor(0, y) = ~y -> nor(y, y)
3457 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3458 MVT::i1, MachineNode->getOperand(1),
3459 MachineNode->getOperand(1));
3462 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3463 MVT::i1, MachineNode->getOperand(0),
3464 MachineNode->getOperand(0));
3466 // nor(~x, y) = andc(x, y)
3467 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
3468 MVT::i1, MachineNode->getOperand(0).
3470 MachineNode->getOperand(1));
3472 // nor(x, ~y) = andc(y, x)
3473 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
3474 MVT::i1, MachineNode->getOperand(1).
3476 MachineNode->getOperand(0));
3477 else if (AllUsersSelectZero(MachineNode))
3478 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode),
3479 MVT::i1, MachineNode->getOperand(0),
3480 MachineNode->getOperand(1)),
3484 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3486 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
3490 ResNode = MachineNode->getOperand(1).getNode();
3493 ResNode = MachineNode->getOperand(0).getNode();
3495 // eqv(0, y) = ~y -> nor(y, y)
3496 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3497 MVT::i1, MachineNode->getOperand(1),
3498 MachineNode->getOperand(1));
3501 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3502 MVT::i1, MachineNode->getOperand(0),
3503 MachineNode->getOperand(0));
3505 // eqv(~x, y) = xor(x, y)
3506 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
3507 MVT::i1, MachineNode->getOperand(0).
3509 MachineNode->getOperand(1));
3511 // eqv(x, ~y) = xor(x, y)
3512 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
3513 MVT::i1, MachineNode->getOperand(0),
3514 MachineNode->getOperand(1).
3516 else if (AllUsersSelectZero(MachineNode))
3517 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
3518 MVT::i1, MachineNode->getOperand(0),
3519 MachineNode->getOperand(1)),
3523 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3525 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
3529 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3530 MVT::i1, MachineNode->getOperand(1),
3531 MachineNode->getOperand(1));
3532 else if (Op1Unset || Op2Set)
3533 // andc(0, y) = andc(x, 1) = 0
3534 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
3538 ResNode = MachineNode->getOperand(0).getNode();
3540 // andc(~x, y) = ~(x | y) = nor(x, y)
3541 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3542 MVT::i1, MachineNode->getOperand(0).
3544 MachineNode->getOperand(1));
3546 // andc(x, ~y) = x & y
3547 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode),
3548 MVT::i1, MachineNode->getOperand(0),
3549 MachineNode->getOperand(1).
3551 else if (AllUsersSelectZero(MachineNode))
3552 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
3553 MVT::i1, MachineNode->getOperand(1),
3554 MachineNode->getOperand(0)),
3558 if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
3560 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
3562 else if (Op1Set || Op2Unset)
3563 // orc(1, y) = orc(x, 0) = 1
3564 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
3568 ResNode = MachineNode->getOperand(0).getNode();
3571 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
3572 MVT::i1, MachineNode->getOperand(1),
3573 MachineNode->getOperand(1));
3575 // orc(~x, y) = ~(x & y) = nand(x, y)
3576 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode),
3577 MVT::i1, MachineNode->getOperand(0).
3579 MachineNode->getOperand(1));
3581 // orc(x, ~y) = x | y
3582 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode),
3583 MVT::i1, MachineNode->getOperand(0),
3584 MachineNode->getOperand(1).
3586 else if (AllUsersSelectZero(MachineNode))
3587 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
3588 MVT::i1, MachineNode->getOperand(1),
3589 MachineNode->getOperand(0)),
3592 case PPC::SELECT_I4:
3593 case PPC::SELECT_I8:
3594 case PPC::SELECT_F4:
3595 case PPC::SELECT_F8:
3596 case PPC::SELECT_VRRC:
3597 case PPC::SELECT_VSFRC:
3598 case PPC::SELECT_VSRC:
3600 ResNode = MachineNode->getOperand(1).getNode();
3602 ResNode = MachineNode->getOperand(2).getNode();
3604 ResNode = CurDAG->getMachineNode(MachineNode->getMachineOpcode(),
3606 MachineNode->getValueType(0),
3607 MachineNode->getOperand(0).
3609 MachineNode->getOperand(2),
3610 MachineNode->getOperand(1));
3615 ResNode = CurDAG->getMachineNode(Opcode == PPC::BC ? PPC::BCn :
3619 MachineNode->getOperand(0).
3621 MachineNode->getOperand(1),
3622 MachineNode->getOperand(2));
3623 // FIXME: Handle Op1Set, Op1Unset here too.
3627 // If we're inverting this node because it is used only by selects that
3628 // we'd like to swap, then swap the selects before the node replacement.
3630 SwapAllSelectUsers(MachineNode);
3632 if (ResNode != MachineNode) {
3633 DEBUG(dbgs() << "CR Peephole replacing:\nOld: ");
3634 DEBUG(MachineNode->dump(CurDAG));
3635 DEBUG(dbgs() << "\nNew: ");
3636 DEBUG(ResNode->dump(CurDAG));
3637 DEBUG(dbgs() << "\n");
3639 ReplaceUses(MachineNode, ResNode);
3644 CurDAG->RemoveDeadNodes();
3645 } while (IsModified);
3648 // Gather the set of 32-bit operations that are known to have their
3649 // higher-order 32 bits zero, where ToPromote contains all such operations.
3650 static bool PeepholePPC64ZExtGather(SDValue Op32,
3651 SmallPtrSetImpl<SDNode *> &ToPromote) {
3652 if (!Op32.isMachineOpcode())
3655 // First, check for the "frontier" instructions (those that will clear the
3656 // higher-order 32 bits.
3658 // For RLWINM and RLWNM, we need to make sure that the mask does not wrap
3659 // around. If it does not, then these instructions will clear the
3660 // higher-order bits.
3661 if ((Op32.getMachineOpcode() == PPC::RLWINM ||
3662 Op32.getMachineOpcode() == PPC::RLWNM) &&
3663 Op32.getConstantOperandVal(2) <= Op32.getConstantOperandVal(3)) {
3664 ToPromote.insert(Op32.getNode());
3668 // SLW and SRW always clear the higher-order bits.
3669 if (Op32.getMachineOpcode() == PPC::SLW ||
3670 Op32.getMachineOpcode() == PPC::SRW) {
3671 ToPromote.insert(Op32.getNode());
3675 // For LI and LIS, we need the immediate to be positive (so that it is not
3677 if (Op32.getMachineOpcode() == PPC::LI ||
3678 Op32.getMachineOpcode() == PPC::LIS) {
3679 if (!isUInt<15>(Op32.getConstantOperandVal(0)))
3682 ToPromote.insert(Op32.getNode());
3686 // Next, check for those instructions we can look through.
3688 // Assuming the mask does not wrap around, then the higher-order bits are
3689 // taken directly from the first operand.
3690 if (Op32.getMachineOpcode() == PPC::RLWIMI &&
3691 Op32.getConstantOperandVal(3) <= Op32.getConstantOperandVal(4)) {
3692 SmallPtrSet<SDNode *, 16> ToPromote1;
3693 if (!PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1))
3696 ToPromote.insert(Op32.getNode());
3697 ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
3701 // For OR, the higher-order bits are zero if that is true for both operands.
3702 // For SELECT_I4, the same is true (but the relevant operand numbers are
3704 if (Op32.getMachineOpcode() == PPC::OR ||
3705 Op32.getMachineOpcode() == PPC::SELECT_I4) {
3706 unsigned B = Op32.getMachineOpcode() == PPC::SELECT_I4 ? 1 : 0;
3707 SmallPtrSet<SDNode *, 16> ToPromote1;
3708 if (!PeepholePPC64ZExtGather(Op32.getOperand(B+0), ToPromote1))
3710 if (!PeepholePPC64ZExtGather(Op32.getOperand(B+1), ToPromote1))
3713 ToPromote.insert(Op32.getNode());
3714 ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
3718 // For ORI and ORIS, we need the higher-order bits of the first operand to be
3719 // zero, and also for the constant to be positive (so that it is not sign
3721 if (Op32.getMachineOpcode() == PPC::ORI ||
3722 Op32.getMachineOpcode() == PPC::ORIS) {
3723 SmallPtrSet<SDNode *, 16> ToPromote1;
3724 if (!PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1))
3726 if (!isUInt<15>(Op32.getConstantOperandVal(1)))
3729 ToPromote.insert(Op32.getNode());
3730 ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
3734 // The higher-order bits of AND are zero if that is true for at least one of
3736 if (Op32.getMachineOpcode() == PPC::AND) {
3737 SmallPtrSet<SDNode *, 16> ToPromote1, ToPromote2;
3739 PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1);
3741 PeepholePPC64ZExtGather(Op32.getOperand(1), ToPromote2);
3742 if (!Op0OK && !Op1OK)
3745 ToPromote.insert(Op32.getNode());
3748 ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
3751 ToPromote.insert(ToPromote2.begin(), ToPromote2.end());
3756 // For ANDI and ANDIS, the higher-order bits are zero if either that is true
3757 // of the first operand, or if the second operand is positive (so that it is
3758 // not sign extended).
3759 if (Op32.getMachineOpcode() == PPC::ANDIo ||
3760 Op32.getMachineOpcode() == PPC::ANDISo) {
3761 SmallPtrSet<SDNode *, 16> ToPromote1;
3763 PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1);
3764 bool Op1OK = isUInt<15>(Op32.getConstantOperandVal(1));
3765 if (!Op0OK && !Op1OK)
3768 ToPromote.insert(Op32.getNode());
3771 ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
3779 void PPCDAGToDAGISel::PeepholePPC64ZExt() {
3780 if (!PPCSubTarget->isPPC64())
3783 // When we zero-extend from i32 to i64, we use a pattern like this:
3784 // def : Pat<(i64 (zext i32:$in)),
3785 // (RLDICL (INSERT_SUBREG (i64 (IMPLICIT_DEF)), $in, sub_32),
3787 // There are several 32-bit shift/rotate instructions, however, that will
3788 // clear the higher-order bits of their output, rendering the RLDICL
3789 // unnecessary. When that happens, we remove it here, and redefine the
3790 // relevant 32-bit operation to be a 64-bit operation.
3792 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
3795 bool MadeChange = false;
3796 while (Position != CurDAG->allnodes_begin()) {
3797 SDNode *N = --Position;
3798 // Skip dead nodes and any non-machine opcodes.
3799 if (N->use_empty() || !N->isMachineOpcode())
3802 if (N->getMachineOpcode() != PPC::RLDICL)
3805 if (N->getConstantOperandVal(1) != 0 ||
3806 N->getConstantOperandVal(2) != 32)
3809 SDValue ISR = N->getOperand(0);
3810 if (!ISR.isMachineOpcode() ||
3811 ISR.getMachineOpcode() != TargetOpcode::INSERT_SUBREG)
3814 if (!ISR.hasOneUse())
3817 if (ISR.getConstantOperandVal(2) != PPC::sub_32)
3820 SDValue IDef = ISR.getOperand(0);
3821 if (!IDef.isMachineOpcode() ||
3822 IDef.getMachineOpcode() != TargetOpcode::IMPLICIT_DEF)
3825 // We now know that we're looking at a canonical i32 -> i64 zext. See if we
3826 // can get rid of it.
3828 SDValue Op32 = ISR->getOperand(1);
3829 if (!Op32.isMachineOpcode())
3832 // There are some 32-bit instructions that always clear the high-order 32
3833 // bits, there are also some instructions (like AND) that we can look
3835 SmallPtrSet<SDNode *, 16> ToPromote;
3836 if (!PeepholePPC64ZExtGather(Op32, ToPromote))
3839 // If the ToPromote set contains nodes that have uses outside of the set
3840 // (except for the original INSERT_SUBREG), then abort the transformation.
3841 bool OutsideUse = false;
3842 for (SDNode *PN : ToPromote) {
3843 for (SDNode *UN : PN->uses()) {
3844 if (!ToPromote.count(UN) && UN != ISR.getNode()) {
3858 // We now know that this zero extension can be removed by promoting to
3859 // nodes in ToPromote to 64-bit operations, where for operations in the
3860 // frontier of the set, we need to insert INSERT_SUBREGs for their
3862 for (SDNode *PN : ToPromote) {
3864 switch (PN->getMachineOpcode()) {
3866 llvm_unreachable("Don't know the 64-bit variant of this instruction");
3867 case PPC::RLWINM: NewOpcode = PPC::RLWINM8; break;
3868 case PPC::RLWNM: NewOpcode = PPC::RLWNM8; break;
3869 case PPC::SLW: NewOpcode = PPC::SLW8; break;
3870 case PPC::SRW: NewOpcode = PPC::SRW8; break;
3871 case PPC::LI: NewOpcode = PPC::LI8; break;
3872 case PPC::LIS: NewOpcode = PPC::LIS8; break;
3873 case PPC::RLWIMI: NewOpcode = PPC::RLWIMI8; break;
3874 case PPC::OR: NewOpcode = PPC::OR8; break;
3875 case PPC::SELECT_I4: NewOpcode = PPC::SELECT_I8; break;
3876 case PPC::ORI: NewOpcode = PPC::ORI8; break;
3877 case PPC::ORIS: NewOpcode = PPC::ORIS8; break;
3878 case PPC::AND: NewOpcode = PPC::AND8; break;
3879 case PPC::ANDIo: NewOpcode = PPC::ANDIo8; break;
3880 case PPC::ANDISo: NewOpcode = PPC::ANDISo8; break;
3883 // Note: During the replacement process, the nodes will be in an
3884 // inconsistent state (some instructions will have operands with values
3885 // of the wrong type). Once done, however, everything should be right
3888 SmallVector<SDValue, 4> Ops;
3889 for (const SDValue &V : PN->ops()) {
3890 if (!ToPromote.count(V.getNode()) && V.getValueType() == MVT::i32 &&
3891 !isa<ConstantSDNode>(V)) {
3892 SDValue ReplOpOps[] = { ISR.getOperand(0), V, ISR.getOperand(2) };
3894 CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, SDLoc(V),
3895 ISR.getNode()->getVTList(), ReplOpOps);
3896 Ops.push_back(SDValue(ReplOp, 0));
3902 // Because all to-be-promoted nodes only have users that are other
3903 // promoted nodes (or the original INSERT_SUBREG), we can safely replace
3904 // the i32 result value type with i64.
3906 SmallVector<EVT, 2> NewVTs;
3907 SDVTList VTs = PN->getVTList();
3908 for (unsigned i = 0, ie = VTs.NumVTs; i != ie; ++i)
3909 if (VTs.VTs[i] == MVT::i32)
3910 NewVTs.push_back(MVT::i64);
3912 NewVTs.push_back(VTs.VTs[i]);
3914 DEBUG(dbgs() << "PPC64 ZExt Peephole morphing:\nOld: ");
3915 DEBUG(PN->dump(CurDAG));
3917 CurDAG->SelectNodeTo(PN, NewOpcode, CurDAG->getVTList(NewVTs), Ops);
3919 DEBUG(dbgs() << "\nNew: ");
3920 DEBUG(PN->dump(CurDAG));
3921 DEBUG(dbgs() << "\n");
3924 // Now we replace the original zero extend and its associated INSERT_SUBREG
3925 // with the value feeding the INSERT_SUBREG (which has now been promoted to
3928 DEBUG(dbgs() << "PPC64 ZExt Peephole replacing:\nOld: ");
3929 DEBUG(N->dump(CurDAG));
3930 DEBUG(dbgs() << "\nNew: ");
3931 DEBUG(Op32.getNode()->dump(CurDAG));
3932 DEBUG(dbgs() << "\n");
3934 ReplaceUses(N, Op32.getNode());
3938 CurDAG->RemoveDeadNodes();
3941 void PPCDAGToDAGISel::PeepholePPC64() {
3942 // These optimizations are currently supported only for 64-bit SVR4.
3943 if (PPCSubTarget->isDarwin() || !PPCSubTarget->isPPC64())
3946 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
3949 while (Position != CurDAG->allnodes_begin()) {
3950 SDNode *N = --Position;
3951 // Skip dead nodes and any non-machine opcodes.
3952 if (N->use_empty() || !N->isMachineOpcode())
3956 unsigned StorageOpcode = N->getMachineOpcode();
3958 switch (StorageOpcode) {
3989 // If this is a load or store with a zero offset, we may be able to
3990 // fold an add-immediate into the memory operation.
3991 if (!isa<ConstantSDNode>(N->getOperand(FirstOp)) ||
3992 N->getConstantOperandVal(FirstOp) != 0)
3995 SDValue Base = N->getOperand(FirstOp + 1);
3996 if (!Base.isMachineOpcode())
4000 bool ReplaceFlags = true;
4002 // When the feeding operation is an add-immediate of some sort,
4003 // determine whether we need to add relocation information to the
4004 // target flags on the immediate operand when we fold it into the
4005 // load instruction.
4007 // For something like ADDItocL, the relocation information is
4008 // inferred from the opcode; when we process it in the AsmPrinter,
4009 // we add the necessary relocation there. A load, though, can receive
4010 // relocation from various flavors of ADDIxxx, so we need to carry
4011 // the relocation information in the target flags.
4012 switch (Base.getMachineOpcode()) {
4017 // In some cases (such as TLS) the relocation information
4018 // is already in place on the operand, so copying the operand
4020 ReplaceFlags = false;
4021 // For these cases, the immediate may not be divisible by 4, in
4022 // which case the fold is illegal for DS-form instructions. (The
4023 // other cases provide aligned addresses and are always safe.)
4024 if ((StorageOpcode == PPC::LWA ||
4025 StorageOpcode == PPC::LD ||
4026 StorageOpcode == PPC::STD) &&
4027 (!isa<ConstantSDNode>(Base.getOperand(1)) ||
4028 Base.getConstantOperandVal(1) % 4 != 0))
4031 case PPC::ADDIdtprelL:
4032 Flags = PPCII::MO_DTPREL_LO;
4034 case PPC::ADDItlsldL:
4035 Flags = PPCII::MO_TLSLD_LO;
4038 Flags = PPCII::MO_TOC_LO;
4042 // We found an opportunity. Reverse the operands from the add
4043 // immediate and substitute them into the load or store. If
4044 // needed, update the target flags for the immediate operand to
4045 // reflect the necessary relocation information.
4046 DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase: ");
4047 DEBUG(Base->dump(CurDAG));
4048 DEBUG(dbgs() << "\nN: ");
4049 DEBUG(N->dump(CurDAG));
4050 DEBUG(dbgs() << "\n");
4052 SDValue ImmOpnd = Base.getOperand(1);
4054 // If the relocation information isn't already present on the
4055 // immediate operand, add it now.
4057 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) {
4059 const GlobalValue *GV = GA->getGlobal();
4060 // We can't perform this optimization for data whose alignment
4061 // is insufficient for the instruction encoding.
4062 if (GV->getAlignment() < 4 &&
4063 (StorageOpcode == PPC::LD || StorageOpcode == PPC::STD ||
4064 StorageOpcode == PPC::LWA)) {
4065 DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n");
4068 ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, 0, Flags);
4069 } else if (ConstantPoolSDNode *CP =
4070 dyn_cast<ConstantPoolSDNode>(ImmOpnd)) {
4071 const Constant *C = CP->getConstVal();
4072 ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64,
4078 if (FirstOp == 1) // Store
4079 (void)CurDAG->UpdateNodeOperands(N, N->getOperand(0), ImmOpnd,
4080 Base.getOperand(0), N->getOperand(3));
4082 (void)CurDAG->UpdateNodeOperands(N, ImmOpnd, Base.getOperand(0),
4085 // The add-immediate may now be dead, in which case remove it.
4086 if (Base.getNode()->use_empty())
4087 CurDAG->RemoveDeadNode(Base.getNode());
4092 /// createPPCISelDag - This pass converts a legalized DAG into a
4093 /// PowerPC-specific DAG, ready for instruction scheduling.
4095 FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM) {
4096 return new PPCDAGToDAGISel(TM);
4099 static void initializePassOnce(PassRegistry &Registry) {
4100 const char *Name = "PowerPC DAG->DAG Pattern Instruction Selection";
4101 PassInfo *PI = new PassInfo(Name, "ppc-codegen", &SelectionDAGISel::ID,
4102 nullptr, false, false);
4103 Registry.registerPass(*PI, true);
4106 void llvm::initializePPCDAGToDAGISelPass(PassRegistry &Registry) {
4107 CALL_ONCE_INITIALIZATION(initializePassOnce);