1 //===-- ARMISelDAGToDAG.cpp - A dag to dag inst selector for ARM ----------===//
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 an instruction selector for the ARM target.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "arm-isel"
16 #include "ARMBaseInstrInfo.h"
17 #include "ARMTargetMachine.h"
18 #include "MCTargetDesc/ARMAddressingModes.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/Intrinsics.h"
24 #include "llvm/LLVMContext.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/SelectionDAGISel.h"
30 #include "llvm/Target/TargetLowering.h"
31 #include "llvm/Target/TargetOptions.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Compiler.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/raw_ostream.h"
41 DisableShifterOp("disable-shifter-op", cl::Hidden,
42 cl::desc("Disable isel of shifter-op"),
46 CheckVMLxHazard("check-vmlx-hazard", cl::Hidden,
47 cl::desc("Check fp vmla / vmls hazard at isel time"),
50 //===--------------------------------------------------------------------===//
51 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
52 /// instructions for SelectionDAG operations.
57 AM2_BASE, // Simple AM2 (+-imm12)
58 AM2_SHOP // Shifter-op AM2
61 class ARMDAGToDAGISel : public SelectionDAGISel {
62 ARMBaseTargetMachine &TM;
63 const ARMBaseInstrInfo *TII;
65 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
66 /// make the right decision when generating code for different targets.
67 const ARMSubtarget *Subtarget;
70 explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
71 CodeGenOpt::Level OptLevel)
72 : SelectionDAGISel(tm, OptLevel), TM(tm),
73 TII(static_cast<const ARMBaseInstrInfo*>(TM.getInstrInfo())),
74 Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
77 virtual const char *getPassName() const {
78 return "ARM Instruction Selection";
81 /// getI32Imm - Return a target constant of type i32 with the specified
83 inline SDValue getI32Imm(unsigned Imm) {
84 return CurDAG->getTargetConstant(Imm, MVT::i32);
87 SDNode *Select(SDNode *N);
90 bool hasNoVMLxHazardUse(SDNode *N) const;
91 bool isShifterOpProfitable(const SDValue &Shift,
92 ARM_AM::ShiftOpc ShOpcVal, unsigned ShAmt);
93 bool SelectRegShifterOperand(SDValue N, SDValue &A,
94 SDValue &B, SDValue &C,
95 bool CheckProfitability = true);
96 bool SelectImmShifterOperand(SDValue N, SDValue &A,
97 SDValue &B, bool CheckProfitability = true);
98 bool SelectShiftRegShifterOperand(SDValue N, SDValue &A,
99 SDValue &B, SDValue &C) {
100 // Don't apply the profitability check
101 return SelectRegShifterOperand(N, A, B, C, false);
103 bool SelectShiftImmShifterOperand(SDValue N, SDValue &A,
105 // Don't apply the profitability check
106 return SelectImmShifterOperand(N, A, B, false);
109 bool SelectAddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm);
110 bool SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset, SDValue &Opc);
112 AddrMode2Type SelectAddrMode2Worker(SDValue N, SDValue &Base,
113 SDValue &Offset, SDValue &Opc);
114 bool SelectAddrMode2Base(SDValue N, SDValue &Base, SDValue &Offset,
116 return SelectAddrMode2Worker(N, Base, Offset, Opc) == AM2_BASE;
119 bool SelectAddrMode2ShOp(SDValue N, SDValue &Base, SDValue &Offset,
121 return SelectAddrMode2Worker(N, Base, Offset, Opc) == AM2_SHOP;
124 bool SelectAddrMode2(SDValue N, SDValue &Base, SDValue &Offset,
126 SelectAddrMode2Worker(N, Base, Offset, Opc);
127 // return SelectAddrMode2ShOp(N, Base, Offset, Opc);
128 // This always matches one way or another.
132 bool SelectAddrMode2OffsetReg(SDNode *Op, SDValue N,
133 SDValue &Offset, SDValue &Opc);
134 bool SelectAddrMode2OffsetImm(SDNode *Op, SDValue N,
135 SDValue &Offset, SDValue &Opc);
136 bool SelectAddrMode2OffsetImmPre(SDNode *Op, SDValue N,
137 SDValue &Offset, SDValue &Opc);
138 bool SelectAddrOffsetNone(SDValue N, SDValue &Base);
139 bool SelectAddrMode3(SDValue N, SDValue &Base,
140 SDValue &Offset, SDValue &Opc);
141 bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
142 SDValue &Offset, SDValue &Opc);
143 bool SelectAddrMode5(SDValue N, SDValue &Base,
145 bool SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,SDValue &Align);
146 bool SelectAddrMode6Offset(SDNode *Op, SDValue N, SDValue &Offset);
148 bool SelectAddrModePC(SDValue N, SDValue &Offset, SDValue &Label);
150 // Thumb Addressing Modes:
151 bool SelectThumbAddrModeRR(SDValue N, SDValue &Base, SDValue &Offset);
152 bool SelectThumbAddrModeRI(SDValue N, SDValue &Base, SDValue &Offset,
154 bool SelectThumbAddrModeRI5S1(SDValue N, SDValue &Base, SDValue &Offset);
155 bool SelectThumbAddrModeRI5S2(SDValue N, SDValue &Base, SDValue &Offset);
156 bool SelectThumbAddrModeRI5S4(SDValue N, SDValue &Base, SDValue &Offset);
157 bool SelectThumbAddrModeImm5S(SDValue N, unsigned Scale, SDValue &Base,
159 bool SelectThumbAddrModeImm5S1(SDValue N, SDValue &Base,
161 bool SelectThumbAddrModeImm5S2(SDValue N, SDValue &Base,
163 bool SelectThumbAddrModeImm5S4(SDValue N, SDValue &Base,
165 bool SelectThumbAddrModeSP(SDValue N, SDValue &Base, SDValue &OffImm);
167 // Thumb 2 Addressing Modes:
168 bool SelectT2ShifterOperandReg(SDValue N,
169 SDValue &BaseReg, SDValue &Opc);
170 bool SelectT2AddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm);
171 bool SelectT2AddrModeImm8(SDValue N, SDValue &Base,
173 bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
175 bool SelectT2AddrModeSoReg(SDValue N, SDValue &Base,
176 SDValue &OffReg, SDValue &ShImm);
178 inline bool is_so_imm(unsigned Imm) const {
179 return ARM_AM::getSOImmVal(Imm) != -1;
182 inline bool is_so_imm_not(unsigned Imm) const {
183 return ARM_AM::getSOImmVal(~Imm) != -1;
186 inline bool is_t2_so_imm(unsigned Imm) const {
187 return ARM_AM::getT2SOImmVal(Imm) != -1;
190 inline bool is_t2_so_imm_not(unsigned Imm) const {
191 return ARM_AM::getT2SOImmVal(~Imm) != -1;
194 // Include the pieces autogenerated from the target description.
195 #include "ARMGenDAGISel.inc"
198 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
200 SDNode *SelectARMIndexedLoad(SDNode *N);
201 SDNode *SelectT2IndexedLoad(SDNode *N);
203 /// SelectVLD - Select NEON load intrinsics. NumVecs should be
204 /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for
205 /// loads of D registers and even subregs and odd subregs of Q registers.
206 /// For NumVecs <= 2, QOpcodes1 is not used.
207 SDNode *SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs,
209 unsigned *QOpcodes0, unsigned *QOpcodes1);
211 /// SelectVST - Select NEON store intrinsics. NumVecs should
212 /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for
213 /// stores of D registers and even subregs and odd subregs of Q registers.
214 /// For NumVecs <= 2, QOpcodes1 is not used.
215 SDNode *SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs,
217 unsigned *QOpcodes0, unsigned *QOpcodes1);
219 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
220 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
221 /// load/store of D registers and Q registers.
222 SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad,
223 bool isUpdating, unsigned NumVecs,
224 unsigned *DOpcodes, unsigned *QOpcodes);
226 /// SelectVLDDup - Select NEON load-duplicate intrinsics. NumVecs
227 /// should be 2, 3 or 4. The opcode array specifies the instructions used
228 /// for loading D registers. (Q registers are not supported.)
229 SDNode *SelectVLDDup(SDNode *N, bool isUpdating, unsigned NumVecs,
232 /// SelectVTBL - Select NEON VTBL and VTBX intrinsics. NumVecs should be 2,
233 /// 3 or 4. These are custom-selected so that a REG_SEQUENCE can be
234 /// generated to force the table registers to be consecutive.
235 SDNode *SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs, unsigned Opc);
237 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
238 SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
240 /// SelectCMOVOp - Select CMOV instructions for ARM.
241 SDNode *SelectCMOVOp(SDNode *N);
242 SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
243 ARMCC::CondCodes CCVal, SDValue CCR,
245 SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
246 ARMCC::CondCodes CCVal, SDValue CCR,
248 SDNode *SelectT2CMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
249 ARMCC::CondCodes CCVal, SDValue CCR,
251 SDNode *SelectARMCMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
252 ARMCC::CondCodes CCVal, SDValue CCR,
255 SDNode *SelectConcatVector(SDNode *N);
257 SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
259 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
260 /// inline asm expressions.
261 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
263 std::vector<SDValue> &OutOps);
265 // Form pairs of consecutive S, D, or Q registers.
266 SDNode *PairSRegs(EVT VT, SDValue V0, SDValue V1);
267 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
268 SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1);
270 // Form sequences of 4 consecutive S, D, or Q registers.
271 SDNode *QuadSRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
272 SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
273 SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
275 // Get the alignment operand for a NEON VLD or VST instruction.
276 SDValue GetVLDSTAlign(SDValue Align, unsigned NumVecs, bool is64BitVector);
280 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
281 /// operand. If so Imm will receive the 32-bit value.
282 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
283 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
284 Imm = cast<ConstantSDNode>(N)->getZExtValue();
290 // isInt32Immediate - This method tests to see if a constant operand.
291 // If so Imm will receive the 32 bit value.
292 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
293 return isInt32Immediate(N.getNode(), Imm);
296 // isOpcWithIntImmediate - This method tests to see if the node is a specific
297 // opcode and that it has a immediate integer right operand.
298 // If so Imm will receive the 32 bit value.
299 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
300 return N->getOpcode() == Opc &&
301 isInt32Immediate(N->getOperand(1).getNode(), Imm);
304 /// \brief Check whether a particular node is a constant value representable as
305 /// (N * Scale) where (N in [\arg RangeMin, \arg RangeMax).
307 /// \param ScaledConstant [out] - On success, the pre-scaled constant value.
308 static bool isScaledConstantInRange(SDValue Node, int Scale,
309 int RangeMin, int RangeMax,
310 int &ScaledConstant) {
311 assert(Scale > 0 && "Invalid scale!");
313 // Check that this is a constant.
314 const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Node);
318 ScaledConstant = (int) C->getZExtValue();
319 if ((ScaledConstant % Scale) != 0)
322 ScaledConstant /= Scale;
323 return ScaledConstant >= RangeMin && ScaledConstant < RangeMax;
326 /// hasNoVMLxHazardUse - Return true if it's desirable to select a FP MLA / MLS
327 /// node. VFP / NEON fp VMLA / VMLS instructions have special RAW hazards (at
328 /// least on current ARM implementations) which should be avoidded.
329 bool ARMDAGToDAGISel::hasNoVMLxHazardUse(SDNode *N) const {
330 if (OptLevel == CodeGenOpt::None)
333 if (!CheckVMLxHazard)
336 if (!Subtarget->isCortexA8() && !Subtarget->isCortexA9())
342 SDNode *Use = *N->use_begin();
343 if (Use->getOpcode() == ISD::CopyToReg)
345 if (Use->isMachineOpcode()) {
346 const MCInstrDesc &MCID = TII->get(Use->getMachineOpcode());
349 unsigned Opcode = MCID.getOpcode();
350 if (Opcode == ARM::VMOVRS || Opcode == ARM::VMOVRRD)
352 // vmlx feeding into another vmlx. We actually want to unfold
353 // the use later in the MLxExpansion pass. e.g.
355 // vmla (stall 8 cycles)
360 // This adds up to about 18 - 19 cycles.
363 // vmul (stall 4 cycles)
364 // vadd adds up to about 14 cycles.
365 return TII->isFpMLxInstruction(Opcode);
371 bool ARMDAGToDAGISel::isShifterOpProfitable(const SDValue &Shift,
372 ARM_AM::ShiftOpc ShOpcVal,
374 if (!Subtarget->isCortexA9())
376 if (Shift.hasOneUse())
379 return ShOpcVal == ARM_AM::lsl && ShAmt == 2;
382 bool ARMDAGToDAGISel::SelectImmShifterOperand(SDValue N,
385 bool CheckProfitability) {
386 if (DisableShifterOp)
389 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
391 // Don't match base register only case. That is matched to a separate
392 // lower complexity pattern with explicit register operand.
393 if (ShOpcVal == ARM_AM::no_shift) return false;
395 BaseReg = N.getOperand(0);
396 unsigned ShImmVal = 0;
397 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1));
398 if (!RHS) return false;
399 ShImmVal = RHS->getZExtValue() & 31;
400 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
405 bool ARMDAGToDAGISel::SelectRegShifterOperand(SDValue N,
409 bool CheckProfitability) {
410 if (DisableShifterOp)
413 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
415 // Don't match base register only case. That is matched to a separate
416 // lower complexity pattern with explicit register operand.
417 if (ShOpcVal == ARM_AM::no_shift) return false;
419 BaseReg = N.getOperand(0);
420 unsigned ShImmVal = 0;
421 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1));
422 if (RHS) return false;
424 ShReg = N.getOperand(1);
425 if (CheckProfitability && !isShifterOpProfitable(N, ShOpcVal, ShImmVal))
427 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
433 bool ARMDAGToDAGISel::SelectAddrModeImm12(SDValue N,
436 // Match simple R + imm12 operands.
439 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
440 !CurDAG->isBaseWithConstantOffset(N)) {
441 if (N.getOpcode() == ISD::FrameIndex) {
442 // Match frame index.
443 int FI = cast<FrameIndexSDNode>(N)->getIndex();
444 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
445 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
449 if (N.getOpcode() == ARMISD::Wrapper &&
450 !(Subtarget->useMovt() &&
451 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
452 Base = N.getOperand(0);
455 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
459 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
460 int RHSC = (int)RHS->getZExtValue();
461 if (N.getOpcode() == ISD::SUB)
464 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
465 Base = N.getOperand(0);
466 if (Base.getOpcode() == ISD::FrameIndex) {
467 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
468 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
470 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
477 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
483 bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset,
485 if (N.getOpcode() == ISD::MUL &&
486 (!Subtarget->isCortexA9() || N.hasOneUse())) {
487 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
488 // X * [3,5,9] -> X + X * [2,4,8] etc.
489 int RHSC = (int)RHS->getZExtValue();
492 ARM_AM::AddrOpc AddSub = ARM_AM::add;
494 AddSub = ARM_AM::sub;
497 if (isPowerOf2_32(RHSC)) {
498 unsigned ShAmt = Log2_32(RHSC);
499 Base = Offset = N.getOperand(0);
500 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
509 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
510 // ISD::OR that is equivalent to an ISD::ADD.
511 !CurDAG->isBaseWithConstantOffset(N))
514 // Leave simple R +/- imm12 operands for LDRi12
515 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::OR) {
517 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1,
518 -0x1000+1, 0x1000, RHSC)) // 12 bits.
522 if (Subtarget->isCortexA9() && !N.hasOneUse())
523 // Compute R +/- (R << N) and reuse it.
526 // Otherwise this is R +/- [possibly shifted] R.
527 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::SUB ? ARM_AM::sub:ARM_AM::add;
528 ARM_AM::ShiftOpc ShOpcVal =
529 ARM_AM::getShiftOpcForNode(N.getOperand(1).getOpcode());
532 Base = N.getOperand(0);
533 Offset = N.getOperand(1);
535 if (ShOpcVal != ARM_AM::no_shift) {
536 // Check to see if the RHS of the shift is a constant, if not, we can't fold
538 if (ConstantSDNode *Sh =
539 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
540 ShAmt = Sh->getZExtValue();
541 if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt))
542 Offset = N.getOperand(1).getOperand(0);
545 ShOpcVal = ARM_AM::no_shift;
548 ShOpcVal = ARM_AM::no_shift;
552 // Try matching (R shl C) + (R).
553 if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift &&
554 !(Subtarget->isCortexA9() || N.getOperand(0).hasOneUse())) {
555 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0).getOpcode());
556 if (ShOpcVal != ARM_AM::no_shift) {
557 // Check to see if the RHS of the shift is a constant, if not, we can't
559 if (ConstantSDNode *Sh =
560 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
561 ShAmt = Sh->getZExtValue();
562 if (!Subtarget->isCortexA9() ||
564 isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt))) {
565 Offset = N.getOperand(0).getOperand(0);
566 Base = N.getOperand(1);
569 ShOpcVal = ARM_AM::no_shift;
572 ShOpcVal = ARM_AM::no_shift;
577 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
587 AddrMode2Type ARMDAGToDAGISel::SelectAddrMode2Worker(SDValue N,
591 if (N.getOpcode() == ISD::MUL &&
592 (!Subtarget->isCortexA9() || N.hasOneUse())) {
593 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
594 // X * [3,5,9] -> X + X * [2,4,8] etc.
595 int RHSC = (int)RHS->getZExtValue();
598 ARM_AM::AddrOpc AddSub = ARM_AM::add;
600 AddSub = ARM_AM::sub;
603 if (isPowerOf2_32(RHSC)) {
604 unsigned ShAmt = Log2_32(RHSC);
605 Base = Offset = N.getOperand(0);
606 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
615 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
616 // ISD::OR that is equivalent to an ADD.
617 !CurDAG->isBaseWithConstantOffset(N)) {
619 if (N.getOpcode() == ISD::FrameIndex) {
620 int FI = cast<FrameIndexSDNode>(N)->getIndex();
621 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
622 } else if (N.getOpcode() == ARMISD::Wrapper &&
623 !(Subtarget->useMovt() &&
624 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
625 Base = N.getOperand(0);
627 Offset = CurDAG->getRegister(0, MVT::i32);
628 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
634 // Match simple R +/- imm12 operands.
635 if (N.getOpcode() != ISD::SUB) {
637 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1,
638 -0x1000+1, 0x1000, RHSC)) { // 12 bits.
639 Base = N.getOperand(0);
640 if (Base.getOpcode() == ISD::FrameIndex) {
641 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
642 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
644 Offset = CurDAG->getRegister(0, MVT::i32);
646 ARM_AM::AddrOpc AddSub = ARM_AM::add;
648 AddSub = ARM_AM::sub;
651 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
658 if (Subtarget->isCortexA9() && !N.hasOneUse()) {
659 // Compute R +/- (R << N) and reuse it.
661 Offset = CurDAG->getRegister(0, MVT::i32);
662 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
668 // Otherwise this is R +/- [possibly shifted] R.
669 ARM_AM::AddrOpc AddSub = N.getOpcode() != ISD::SUB ? ARM_AM::add:ARM_AM::sub;
670 ARM_AM::ShiftOpc ShOpcVal =
671 ARM_AM::getShiftOpcForNode(N.getOperand(1).getOpcode());
674 Base = N.getOperand(0);
675 Offset = N.getOperand(1);
677 if (ShOpcVal != ARM_AM::no_shift) {
678 // Check to see if the RHS of the shift is a constant, if not, we can't fold
680 if (ConstantSDNode *Sh =
681 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
682 ShAmt = Sh->getZExtValue();
683 if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt))
684 Offset = N.getOperand(1).getOperand(0);
687 ShOpcVal = ARM_AM::no_shift;
690 ShOpcVal = ARM_AM::no_shift;
694 // Try matching (R shl C) + (R).
695 if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift &&
696 !(Subtarget->isCortexA9() || N.getOperand(0).hasOneUse())) {
697 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0).getOpcode());
698 if (ShOpcVal != ARM_AM::no_shift) {
699 // Check to see if the RHS of the shift is a constant, if not, we can't
701 if (ConstantSDNode *Sh =
702 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
703 ShAmt = Sh->getZExtValue();
704 if (!Subtarget->isCortexA9() ||
706 isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt))) {
707 Offset = N.getOperand(0).getOperand(0);
708 Base = N.getOperand(1);
711 ShOpcVal = ARM_AM::no_shift;
714 ShOpcVal = ARM_AM::no_shift;
719 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
724 bool ARMDAGToDAGISel::SelectAddrMode2OffsetReg(SDNode *Op, SDValue N,
725 SDValue &Offset, SDValue &Opc) {
726 unsigned Opcode = Op->getOpcode();
727 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
728 ? cast<LoadSDNode>(Op)->getAddressingMode()
729 : cast<StoreSDNode>(Op)->getAddressingMode();
730 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
731 ? ARM_AM::add : ARM_AM::sub;
733 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val))
737 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
739 if (ShOpcVal != ARM_AM::no_shift) {
740 // Check to see if the RHS of the shift is a constant, if not, we can't fold
742 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
743 ShAmt = Sh->getZExtValue();
744 if (isShifterOpProfitable(N, ShOpcVal, ShAmt))
745 Offset = N.getOperand(0);
748 ShOpcVal = ARM_AM::no_shift;
751 ShOpcVal = ARM_AM::no_shift;
755 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
760 bool ARMDAGToDAGISel::SelectAddrMode2OffsetImmPre(SDNode *Op, SDValue N,
761 SDValue &Offset, SDValue &Opc) {
762 unsigned Opcode = Op->getOpcode();
763 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
764 ? cast<LoadSDNode>(Op)->getAddressingMode()
765 : cast<StoreSDNode>(Op)->getAddressingMode();
766 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
767 ? ARM_AM::add : ARM_AM::sub;
769 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val)) { // 12 bits.
770 if (AddSub == ARM_AM::sub) Val *= -1;
771 Offset = CurDAG->getRegister(0, MVT::i32);
772 Opc = CurDAG->getTargetConstant(Val, MVT::i32);
780 bool ARMDAGToDAGISel::SelectAddrMode2OffsetImm(SDNode *Op, SDValue N,
781 SDValue &Offset, SDValue &Opc) {
782 unsigned Opcode = Op->getOpcode();
783 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
784 ? cast<LoadSDNode>(Op)->getAddressingMode()
785 : cast<StoreSDNode>(Op)->getAddressingMode();
786 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
787 ? ARM_AM::add : ARM_AM::sub;
789 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val)) { // 12 bits.
790 Offset = CurDAG->getRegister(0, MVT::i32);
791 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
800 bool ARMDAGToDAGISel::SelectAddrOffsetNone(SDValue N, SDValue &Base) {
805 bool ARMDAGToDAGISel::SelectAddrMode3(SDValue N,
806 SDValue &Base, SDValue &Offset,
808 if (N.getOpcode() == ISD::SUB) {
809 // X - C is canonicalize to X + -C, no need to handle it here.
810 Base = N.getOperand(0);
811 Offset = N.getOperand(1);
812 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
816 if (!CurDAG->isBaseWithConstantOffset(N)) {
818 if (N.getOpcode() == ISD::FrameIndex) {
819 int FI = cast<FrameIndexSDNode>(N)->getIndex();
820 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
822 Offset = CurDAG->getRegister(0, MVT::i32);
823 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
827 // If the RHS is +/- imm8, fold into addr mode.
829 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1,
830 -256 + 1, 256, RHSC)) { // 8 bits.
831 Base = N.getOperand(0);
832 if (Base.getOpcode() == ISD::FrameIndex) {
833 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
834 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
836 Offset = CurDAG->getRegister(0, MVT::i32);
838 ARM_AM::AddrOpc AddSub = ARM_AM::add;
840 AddSub = ARM_AM::sub;
843 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
847 Base = N.getOperand(0);
848 Offset = N.getOperand(1);
849 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
853 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
854 SDValue &Offset, SDValue &Opc) {
855 unsigned Opcode = Op->getOpcode();
856 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
857 ? cast<LoadSDNode>(Op)->getAddressingMode()
858 : cast<StoreSDNode>(Op)->getAddressingMode();
859 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
860 ? ARM_AM::add : ARM_AM::sub;
862 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 256, Val)) { // 12 bits.
863 Offset = CurDAG->getRegister(0, MVT::i32);
864 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
869 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
873 bool ARMDAGToDAGISel::SelectAddrMode5(SDValue N,
874 SDValue &Base, SDValue &Offset) {
875 if (!CurDAG->isBaseWithConstantOffset(N)) {
877 if (N.getOpcode() == ISD::FrameIndex) {
878 int FI = cast<FrameIndexSDNode>(N)->getIndex();
879 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
880 } else if (N.getOpcode() == ARMISD::Wrapper &&
881 !(Subtarget->useMovt() &&
882 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
883 Base = N.getOperand(0);
885 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
890 // If the RHS is +/- imm8, fold into addr mode.
892 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/4,
893 -256 + 1, 256, RHSC)) {
894 Base = N.getOperand(0);
895 if (Base.getOpcode() == ISD::FrameIndex) {
896 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
897 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
900 ARM_AM::AddrOpc AddSub = ARM_AM::add;
902 AddSub = ARM_AM::sub;
905 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
911 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
916 bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,
920 unsigned Alignment = 0;
921 if (LSBaseSDNode *LSN = dyn_cast<LSBaseSDNode>(Parent)) {
922 // This case occurs only for VLD1-lane/dup and VST1-lane instructions.
923 // The maximum alignment is equal to the memory size being referenced.
924 unsigned LSNAlign = LSN->getAlignment();
925 unsigned MemSize = LSN->getMemoryVT().getSizeInBits() / 8;
926 if (LSNAlign > MemSize && MemSize > 1)
929 // All other uses of addrmode6 are for intrinsics. For now just record
930 // the raw alignment value; it will be refined later based on the legal
931 // alignment operands for the intrinsic.
932 Alignment = cast<MemIntrinsicSDNode>(Parent)->getAlignment();
935 Align = CurDAG->getTargetConstant(Alignment, MVT::i32);
939 bool ARMDAGToDAGISel::SelectAddrMode6Offset(SDNode *Op, SDValue N,
941 LSBaseSDNode *LdSt = cast<LSBaseSDNode>(Op);
942 ISD::MemIndexedMode AM = LdSt->getAddressingMode();
943 if (AM != ISD::POST_INC)
946 if (ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N)) {
947 if (NC->getZExtValue() * 8 == LdSt->getMemoryVT().getSizeInBits())
948 Offset = CurDAG->getRegister(0, MVT::i32);
953 bool ARMDAGToDAGISel::SelectAddrModePC(SDValue N,
954 SDValue &Offset, SDValue &Label) {
955 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
956 Offset = N.getOperand(0);
957 SDValue N1 = N.getOperand(1);
958 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
967 //===----------------------------------------------------------------------===//
968 // Thumb Addressing Modes
969 //===----------------------------------------------------------------------===//
971 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue N,
972 SDValue &Base, SDValue &Offset){
973 if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(N)) {
974 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
975 if (!NC || !NC->isNullValue())
982 Base = N.getOperand(0);
983 Offset = N.getOperand(1);
988 ARMDAGToDAGISel::SelectThumbAddrModeRI(SDValue N, SDValue &Base,
989 SDValue &Offset, unsigned Scale) {
991 SDValue TmpBase, TmpOffImm;
992 if (SelectThumbAddrModeSP(N, TmpBase, TmpOffImm))
993 return false; // We want to select tLDRspi / tSTRspi instead.
995 if (N.getOpcode() == ARMISD::Wrapper &&
996 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
997 return false; // We want to select tLDRpci instead.
1000 if (!CurDAG->isBaseWithConstantOffset(N))
1003 // Thumb does not have [sp, r] address mode.
1004 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
1005 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
1006 if ((LHSR && LHSR->getReg() == ARM::SP) ||
1007 (RHSR && RHSR->getReg() == ARM::SP))
1010 // FIXME: Why do we explicitly check for a match here and then return false?
1011 // Presumably to allow something else to match, but shouldn't this be
1014 if (isScaledConstantInRange(N.getOperand(1), Scale, 0, 32, RHSC))
1017 Base = N.getOperand(0);
1018 Offset = N.getOperand(1);
1023 ARMDAGToDAGISel::SelectThumbAddrModeRI5S1(SDValue N,
1026 return SelectThumbAddrModeRI(N, Base, Offset, 1);
1030 ARMDAGToDAGISel::SelectThumbAddrModeRI5S2(SDValue N,
1033 return SelectThumbAddrModeRI(N, Base, Offset, 2);
1037 ARMDAGToDAGISel::SelectThumbAddrModeRI5S4(SDValue N,
1040 return SelectThumbAddrModeRI(N, Base, Offset, 4);
1044 ARMDAGToDAGISel::SelectThumbAddrModeImm5S(SDValue N, unsigned Scale,
1045 SDValue &Base, SDValue &OffImm) {
1047 SDValue TmpBase, TmpOffImm;
1048 if (SelectThumbAddrModeSP(N, TmpBase, TmpOffImm))
1049 return false; // We want to select tLDRspi / tSTRspi instead.
1051 if (N.getOpcode() == ARMISD::Wrapper &&
1052 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
1053 return false; // We want to select tLDRpci instead.
1056 if (!CurDAG->isBaseWithConstantOffset(N)) {
1057 if (N.getOpcode() == ARMISD::Wrapper &&
1058 !(Subtarget->useMovt() &&
1059 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
1060 Base = N.getOperand(0);
1065 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
1069 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
1070 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
1071 if ((LHSR && LHSR->getReg() == ARM::SP) ||
1072 (RHSR && RHSR->getReg() == ARM::SP)) {
1073 ConstantSDNode *LHS = dyn_cast<ConstantSDNode>(N.getOperand(0));
1074 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1));
1075 unsigned LHSC = LHS ? LHS->getZExtValue() : 0;
1076 unsigned RHSC = RHS ? RHS->getZExtValue() : 0;
1078 // Thumb does not have [sp, #imm5] address mode for non-zero imm5.
1079 if (LHSC != 0 || RHSC != 0) return false;
1082 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
1086 // If the RHS is + imm5 * scale, fold into addr mode.
1088 if (isScaledConstantInRange(N.getOperand(1), Scale, 0, 32, RHSC)) {
1089 Base = N.getOperand(0);
1090 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
1094 Base = N.getOperand(0);
1095 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
1100 ARMDAGToDAGISel::SelectThumbAddrModeImm5S4(SDValue N, SDValue &Base,
1102 return SelectThumbAddrModeImm5S(N, 4, Base, OffImm);
1106 ARMDAGToDAGISel::SelectThumbAddrModeImm5S2(SDValue N, SDValue &Base,
1108 return SelectThumbAddrModeImm5S(N, 2, Base, OffImm);
1112 ARMDAGToDAGISel::SelectThumbAddrModeImm5S1(SDValue N, SDValue &Base,
1114 return SelectThumbAddrModeImm5S(N, 1, Base, OffImm);
1117 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDValue N,
1118 SDValue &Base, SDValue &OffImm) {
1119 if (N.getOpcode() == ISD::FrameIndex) {
1120 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1121 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1122 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
1126 if (!CurDAG->isBaseWithConstantOffset(N))
1129 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
1130 if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
1131 (LHSR && LHSR->getReg() == ARM::SP)) {
1132 // If the RHS is + imm8 * scale, fold into addr mode.
1134 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/4, 0, 256, RHSC)) {
1135 Base = N.getOperand(0);
1136 if (Base.getOpcode() == ISD::FrameIndex) {
1137 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1138 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1140 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
1149 //===----------------------------------------------------------------------===//
1150 // Thumb 2 Addressing Modes
1151 //===----------------------------------------------------------------------===//
1154 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDValue N, SDValue &BaseReg,
1156 if (DisableShifterOp)
1159 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
1161 // Don't match base register only case. That is matched to a separate
1162 // lower complexity pattern with explicit register operand.
1163 if (ShOpcVal == ARM_AM::no_shift) return false;
1165 BaseReg = N.getOperand(0);
1166 unsigned ShImmVal = 0;
1167 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1168 ShImmVal = RHS->getZExtValue() & 31;
1169 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
1176 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDValue N,
1177 SDValue &Base, SDValue &OffImm) {
1178 // Match simple R + imm12 operands.
1181 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
1182 !CurDAG->isBaseWithConstantOffset(N)) {
1183 if (N.getOpcode() == ISD::FrameIndex) {
1184 // Match frame index.
1185 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1186 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1187 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
1191 if (N.getOpcode() == ARMISD::Wrapper &&
1192 !(Subtarget->useMovt() &&
1193 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
1194 Base = N.getOperand(0);
1195 if (Base.getOpcode() == ISD::TargetConstantPool)
1196 return false; // We want to select t2LDRpci instead.
1199 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
1203 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1204 if (SelectT2AddrModeImm8(N, Base, OffImm))
1205 // Let t2LDRi8 handle (R - imm8).
1208 int RHSC = (int)RHS->getZExtValue();
1209 if (N.getOpcode() == ISD::SUB)
1212 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
1213 Base = N.getOperand(0);
1214 if (Base.getOpcode() == ISD::FrameIndex) {
1215 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1216 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1218 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
1225 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
1229 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue N,
1230 SDValue &Base, SDValue &OffImm) {
1231 // Match simple R - imm8 operands.
1232 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
1233 !CurDAG->isBaseWithConstantOffset(N))
1236 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1237 int RHSC = (int)RHS->getSExtValue();
1238 if (N.getOpcode() == ISD::SUB)
1241 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
1242 Base = N.getOperand(0);
1243 if (Base.getOpcode() == ISD::FrameIndex) {
1244 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1245 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1247 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
1255 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
1257 unsigned Opcode = Op->getOpcode();
1258 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
1259 ? cast<LoadSDNode>(Op)->getAddressingMode()
1260 : cast<StoreSDNode>(Op)->getAddressingMode();
1262 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x100, RHSC)) { // 8 bits.
1263 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
1264 ? CurDAG->getTargetConstant(RHSC, MVT::i32)
1265 : CurDAG->getTargetConstant(-RHSC, MVT::i32);
1272 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue N,
1274 SDValue &OffReg, SDValue &ShImm) {
1275 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
1276 if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(N))
1279 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
1280 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1281 int RHSC = (int)RHS->getZExtValue();
1282 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
1284 else if (RHSC < 0 && RHSC >= -255) // 8 bits
1288 if (Subtarget->isCortexA9() && !N.hasOneUse()) {
1289 // Compute R + (R << [1,2,3]) and reuse it.
1293 // Look for (R + R) or (R + (R << [1,2,3])).
1295 Base = N.getOperand(0);
1296 OffReg = N.getOperand(1);
1298 // Swap if it is ((R << c) + R).
1299 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg.getOpcode());
1300 if (ShOpcVal != ARM_AM::lsl) {
1301 ShOpcVal = ARM_AM::getShiftOpcForNode(Base.getOpcode());
1302 if (ShOpcVal == ARM_AM::lsl)
1303 std::swap(Base, OffReg);
1306 if (ShOpcVal == ARM_AM::lsl) {
1307 // Check to see if the RHS of the shift is a constant, if not, we can't fold
1309 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
1310 ShAmt = Sh->getZExtValue();
1311 if (ShAmt < 4 && isShifterOpProfitable(OffReg, ShOpcVal, ShAmt))
1312 OffReg = OffReg.getOperand(0);
1315 ShOpcVal = ARM_AM::no_shift;
1318 ShOpcVal = ARM_AM::no_shift;
1322 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
1327 //===--------------------------------------------------------------------===//
1329 /// getAL - Returns a ARMCC::AL immediate node.
1330 static inline SDValue getAL(SelectionDAG *CurDAG) {
1331 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
1334 SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
1335 LoadSDNode *LD = cast<LoadSDNode>(N);
1336 ISD::MemIndexedMode AM = LD->getAddressingMode();
1337 if (AM == ISD::UNINDEXED)
1340 EVT LoadedVT = LD->getMemoryVT();
1341 SDValue Offset, AMOpc;
1342 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
1343 unsigned Opcode = 0;
1345 if (LoadedVT == MVT::i32 && isPre &&
1346 SelectAddrMode2OffsetImmPre(N, LD->getOffset(), Offset, AMOpc)) {
1347 Opcode = ARM::LDR_PRE_IMM;
1349 } else if (LoadedVT == MVT::i32 && !isPre &&
1350 SelectAddrMode2OffsetImm(N, LD->getOffset(), Offset, AMOpc)) {
1351 Opcode = ARM::LDR_POST_IMM;
1353 } else if (LoadedVT == MVT::i32 &&
1354 SelectAddrMode2OffsetReg(N, LD->getOffset(), Offset, AMOpc)) {
1355 Opcode = isPre ? ARM::LDR_PRE_REG : ARM::LDR_POST_REG;
1358 } else if (LoadedVT == MVT::i16 &&
1359 SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
1361 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
1362 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
1363 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
1364 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
1365 if (LD->getExtensionType() == ISD::SEXTLOAD) {
1366 if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
1368 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
1372 SelectAddrMode2OffsetImmPre(N, LD->getOffset(), Offset, AMOpc)) {
1374 Opcode = ARM::LDRB_PRE_IMM;
1375 } else if (!isPre &&
1376 SelectAddrMode2OffsetImm(N, LD->getOffset(), Offset, AMOpc)) {
1378 Opcode = ARM::LDRB_POST_IMM;
1379 } else if (SelectAddrMode2OffsetReg(N, LD->getOffset(), Offset, AMOpc)) {
1381 Opcode = isPre ? ARM::LDRB_PRE_REG : ARM::LDRB_POST_REG;
1387 if (Opcode == ARM::LDR_PRE_IMM || Opcode == ARM::LDRB_PRE_IMM) {
1388 SDValue Chain = LD->getChain();
1389 SDValue Base = LD->getBasePtr();
1390 SDValue Ops[]= { Base, AMOpc, getAL(CurDAG),
1391 CurDAG->getRegister(0, MVT::i32), Chain };
1392 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32,
1393 MVT::i32, MVT::Other, Ops, 5);
1395 SDValue Chain = LD->getChain();
1396 SDValue Base = LD->getBasePtr();
1397 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
1398 CurDAG->getRegister(0, MVT::i32), Chain };
1399 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32,
1400 MVT::i32, MVT::Other, Ops, 6);
1407 SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
1408 LoadSDNode *LD = cast<LoadSDNode>(N);
1409 ISD::MemIndexedMode AM = LD->getAddressingMode();
1410 if (AM == ISD::UNINDEXED)
1413 EVT LoadedVT = LD->getMemoryVT();
1414 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
1416 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
1417 unsigned Opcode = 0;
1419 if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
1420 switch (LoadedVT.getSimpleVT().SimpleTy) {
1422 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
1426 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
1428 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
1433 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
1435 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
1444 SDValue Chain = LD->getChain();
1445 SDValue Base = LD->getBasePtr();
1446 SDValue Ops[]= { Base, Offset, getAL(CurDAG),
1447 CurDAG->getRegister(0, MVT::i32), Chain };
1448 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
1449 MVT::Other, Ops, 5);
1455 /// PairSRegs - Form a D register from a pair of S registers.
1457 SDNode *ARMDAGToDAGISel::PairSRegs(EVT VT, SDValue V0, SDValue V1) {
1458 DebugLoc dl = V0.getNode()->getDebugLoc();
1460 CurDAG->getTargetConstant(ARM::DPR_VFP2RegClassID, MVT::i32);
1461 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
1462 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
1463 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1464 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 5);
1467 /// PairDRegs - Form a quad register from a pair of D registers.
1469 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
1470 DebugLoc dl = V0.getNode()->getDebugLoc();
1471 SDValue RegClass = CurDAG->getTargetConstant(ARM::QPRRegClassID, MVT::i32);
1472 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1473 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1474 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1475 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 5);
1478 /// PairQRegs - Form 4 consecutive D registers from a pair of Q registers.
1480 SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) {
1481 DebugLoc dl = V0.getNode()->getDebugLoc();
1482 SDValue RegClass = CurDAG->getTargetConstant(ARM::QQPRRegClassID, MVT::i32);
1483 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
1484 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
1485 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1486 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 5);
1489 /// QuadSRegs - Form 4 consecutive S registers.
1491 SDNode *ARMDAGToDAGISel::QuadSRegs(EVT VT, SDValue V0, SDValue V1,
1492 SDValue V2, SDValue V3) {
1493 DebugLoc dl = V0.getNode()->getDebugLoc();
1495 CurDAG->getTargetConstant(ARM::QPR_VFP2RegClassID, MVT::i32);
1496 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
1497 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
1498 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, MVT::i32);
1499 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, MVT::i32);
1500 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1,
1501 V2, SubReg2, V3, SubReg3 };
1502 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 9);
1505 /// QuadDRegs - Form 4 consecutive D registers.
1507 SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
1508 SDValue V2, SDValue V3) {
1509 DebugLoc dl = V0.getNode()->getDebugLoc();
1510 SDValue RegClass = CurDAG->getTargetConstant(ARM::QQPRRegClassID, MVT::i32);
1511 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1512 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1513 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
1514 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
1515 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1,
1516 V2, SubReg2, V3, SubReg3 };
1517 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 9);
1520 /// QuadQRegs - Form 4 consecutive Q registers.
1522 SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1,
1523 SDValue V2, SDValue V3) {
1524 DebugLoc dl = V0.getNode()->getDebugLoc();
1525 SDValue RegClass = CurDAG->getTargetConstant(ARM::QQQQPRRegClassID, MVT::i32);
1526 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
1527 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
1528 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, MVT::i32);
1529 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, MVT::i32);
1530 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1,
1531 V2, SubReg2, V3, SubReg3 };
1532 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 9);
1535 /// GetVLDSTAlign - Get the alignment (in bytes) for the alignment operand
1536 /// of a NEON VLD or VST instruction. The supported values depend on the
1537 /// number of registers being loaded.
1538 SDValue ARMDAGToDAGISel::GetVLDSTAlign(SDValue Align, unsigned NumVecs,
1539 bool is64BitVector) {
1540 unsigned NumRegs = NumVecs;
1541 if (!is64BitVector && NumVecs < 3)
1544 unsigned Alignment = cast<ConstantSDNode>(Align)->getZExtValue();
1545 if (Alignment >= 32 && NumRegs == 4)
1547 else if (Alignment >= 16 && (NumRegs == 2 || NumRegs == 4))
1549 else if (Alignment >= 8)
1554 return CurDAG->getTargetConstant(Alignment, MVT::i32);
1557 SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs,
1558 unsigned *DOpcodes, unsigned *QOpcodes0,
1559 unsigned *QOpcodes1) {
1560 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1561 DebugLoc dl = N->getDebugLoc();
1563 SDValue MemAddr, Align;
1564 unsigned AddrOpIdx = isUpdating ? 1 : 2;
1565 if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
1568 SDValue Chain = N->getOperand(0);
1569 EVT VT = N->getValueType(0);
1570 bool is64BitVector = VT.is64BitVector();
1571 Align = GetVLDSTAlign(Align, NumVecs, is64BitVector);
1573 unsigned OpcodeIndex;
1574 switch (VT.getSimpleVT().SimpleTy) {
1575 default: llvm_unreachable("unhandled vld type");
1576 // Double-register operations:
1577 case MVT::v8i8: OpcodeIndex = 0; break;
1578 case MVT::v4i16: OpcodeIndex = 1; break;
1580 case MVT::v2i32: OpcodeIndex = 2; break;
1581 case MVT::v1i64: OpcodeIndex = 3; break;
1582 // Quad-register operations:
1583 case MVT::v16i8: OpcodeIndex = 0; break;
1584 case MVT::v8i16: OpcodeIndex = 1; break;
1586 case MVT::v4i32: OpcodeIndex = 2; break;
1587 case MVT::v2i64: OpcodeIndex = 3;
1588 assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
1596 unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs;
1599 ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, ResTyElts);
1601 std::vector<EVT> ResTys;
1602 ResTys.push_back(ResTy);
1604 ResTys.push_back(MVT::i32);
1605 ResTys.push_back(MVT::Other);
1607 SDValue Pred = getAL(CurDAG);
1608 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1610 SmallVector<SDValue, 7> Ops;
1612 // Double registers and VLD1/VLD2 quad registers are directly supported.
1613 if (is64BitVector || NumVecs <= 2) {
1614 unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
1615 QOpcodes0[OpcodeIndex]);
1616 Ops.push_back(MemAddr);
1617 Ops.push_back(Align);
1619 SDValue Inc = N->getOperand(AddrOpIdx + 1);
1620 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc);
1622 Ops.push_back(Pred);
1623 Ops.push_back(Reg0);
1624 Ops.push_back(Chain);
1625 VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), Ops.size());
1628 // Otherwise, quad registers are loaded with two separate instructions,
1629 // where one loads the even registers and the other loads the odd registers.
1630 EVT AddrTy = MemAddr.getValueType();
1632 // Load the even subregs. This is always an updating load, so that it
1633 // provides the address to the second load for the odd subregs.
1635 SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, ResTy), 0);
1636 const SDValue OpsA[] = { MemAddr, Align, Reg0, ImplDef, Pred, Reg0, Chain };
1637 SDNode *VLdA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl,
1638 ResTy, AddrTy, MVT::Other, OpsA, 7);
1639 Chain = SDValue(VLdA, 2);
1641 // Load the odd subregs.
1642 Ops.push_back(SDValue(VLdA, 1));
1643 Ops.push_back(Align);
1645 SDValue Inc = N->getOperand(AddrOpIdx + 1);
1646 assert(isa<ConstantSDNode>(Inc.getNode()) &&
1647 "only constant post-increment update allowed for VLD3/4");
1649 Ops.push_back(Reg0);
1651 Ops.push_back(SDValue(VLdA, 0));
1652 Ops.push_back(Pred);
1653 Ops.push_back(Reg0);
1654 Ops.push_back(Chain);
1655 VLd = CurDAG->getMachineNode(QOpcodes1[OpcodeIndex], dl, ResTys,
1656 Ops.data(), Ops.size());
1659 // Transfer memoperands.
1660 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1661 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
1662 cast<MachineSDNode>(VLd)->setMemRefs(MemOp, MemOp + 1);
1667 // Extract out the subregisters.
1668 SDValue SuperReg = SDValue(VLd, 0);
1669 assert(ARM::dsub_7 == ARM::dsub_0+7 &&
1670 ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1671 unsigned Sub0 = (is64BitVector ? ARM::dsub_0 : ARM::qsub_0);
1672 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1673 ReplaceUses(SDValue(N, Vec),
1674 CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg));
1675 ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, 1));
1677 ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLd, 2));
1681 SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs,
1682 unsigned *DOpcodes, unsigned *QOpcodes0,
1683 unsigned *QOpcodes1) {
1684 assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range");
1685 DebugLoc dl = N->getDebugLoc();
1687 SDValue MemAddr, Align;
1688 unsigned AddrOpIdx = isUpdating ? 1 : 2;
1689 unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1)
1690 if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
1693 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1694 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
1696 SDValue Chain = N->getOperand(0);
1697 EVT VT = N->getOperand(Vec0Idx).getValueType();
1698 bool is64BitVector = VT.is64BitVector();
1699 Align = GetVLDSTAlign(Align, NumVecs, is64BitVector);
1701 unsigned OpcodeIndex;
1702 switch (VT.getSimpleVT().SimpleTy) {
1703 default: llvm_unreachable("unhandled vst type");
1704 // Double-register operations:
1705 case MVT::v8i8: OpcodeIndex = 0; break;
1706 case MVT::v4i16: OpcodeIndex = 1; break;
1708 case MVT::v2i32: OpcodeIndex = 2; break;
1709 case MVT::v1i64: OpcodeIndex = 3; break;
1710 // Quad-register operations:
1711 case MVT::v16i8: OpcodeIndex = 0; break;
1712 case MVT::v8i16: OpcodeIndex = 1; break;
1714 case MVT::v4i32: OpcodeIndex = 2; break;
1715 case MVT::v2i64: OpcodeIndex = 3;
1716 assert(NumVecs == 1 && "v2i64 type only supported for VST1");
1720 std::vector<EVT> ResTys;
1722 ResTys.push_back(MVT::i32);
1723 ResTys.push_back(MVT::Other);
1725 SDValue Pred = getAL(CurDAG);
1726 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1727 SmallVector<SDValue, 7> Ops;
1729 // Double registers and VST1/VST2 quad registers are directly supported.
1730 if (is64BitVector || NumVecs <= 2) {
1733 SrcReg = N->getOperand(Vec0Idx);
1734 } else if (is64BitVector) {
1735 // Form a REG_SEQUENCE to force register allocation.
1736 SDValue V0 = N->getOperand(Vec0Idx + 0);
1737 SDValue V1 = N->getOperand(Vec0Idx + 1);
1739 SrcReg = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1741 SDValue V2 = N->getOperand(Vec0Idx + 2);
1742 // If it's a vst3, form a quad D-register and leave the last part as
1744 SDValue V3 = (NumVecs == 3)
1745 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1746 : N->getOperand(Vec0Idx + 3);
1747 SrcReg = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1750 // Form a QQ register.
1751 SDValue Q0 = N->getOperand(Vec0Idx);
1752 SDValue Q1 = N->getOperand(Vec0Idx + 1);
1753 SrcReg = SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0);
1756 unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
1757 QOpcodes0[OpcodeIndex]);
1758 Ops.push_back(MemAddr);
1759 Ops.push_back(Align);
1761 SDValue Inc = N->getOperand(AddrOpIdx + 1);
1762 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc);
1764 Ops.push_back(SrcReg);
1765 Ops.push_back(Pred);
1766 Ops.push_back(Reg0);
1767 Ops.push_back(Chain);
1769 CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), Ops.size());
1771 // Transfer memoperands.
1772 cast<MachineSDNode>(VSt)->setMemRefs(MemOp, MemOp + 1);
1777 // Otherwise, quad registers are stored with two separate instructions,
1778 // where one stores the even registers and the other stores the odd registers.
1780 // Form the QQQQ REG_SEQUENCE.
1781 SDValue V0 = N->getOperand(Vec0Idx + 0);
1782 SDValue V1 = N->getOperand(Vec0Idx + 1);
1783 SDValue V2 = N->getOperand(Vec0Idx + 2);
1784 SDValue V3 = (NumVecs == 3)
1785 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
1786 : N->getOperand(Vec0Idx + 3);
1787 SDValue RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
1789 // Store the even D registers. This is always an updating store, so that it
1790 // provides the address to the second store for the odd subregs.
1791 const SDValue OpsA[] = { MemAddr, Align, Reg0, RegSeq, Pred, Reg0, Chain };
1792 SDNode *VStA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl,
1793 MemAddr.getValueType(),
1794 MVT::Other, OpsA, 7);
1795 cast<MachineSDNode>(VStA)->setMemRefs(MemOp, MemOp + 1);
1796 Chain = SDValue(VStA, 1);
1798 // Store the odd D registers.
1799 Ops.push_back(SDValue(VStA, 0));
1800 Ops.push_back(Align);
1802 SDValue Inc = N->getOperand(AddrOpIdx + 1);
1803 assert(isa<ConstantSDNode>(Inc.getNode()) &&
1804 "only constant post-increment update allowed for VST3/4");
1806 Ops.push_back(Reg0);
1808 Ops.push_back(RegSeq);
1809 Ops.push_back(Pred);
1810 Ops.push_back(Reg0);
1811 Ops.push_back(Chain);
1812 SDNode *VStB = CurDAG->getMachineNode(QOpcodes1[OpcodeIndex], dl, ResTys,
1813 Ops.data(), Ops.size());
1814 cast<MachineSDNode>(VStB)->setMemRefs(MemOp, MemOp + 1);
1818 SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1819 bool isUpdating, unsigned NumVecs,
1821 unsigned *QOpcodes) {
1822 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1823 DebugLoc dl = N->getDebugLoc();
1825 SDValue MemAddr, Align;
1826 unsigned AddrOpIdx = isUpdating ? 1 : 2;
1827 unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1)
1828 if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
1831 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1832 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
1834 SDValue Chain = N->getOperand(0);
1836 cast<ConstantSDNode>(N->getOperand(Vec0Idx + NumVecs))->getZExtValue();
1837 EVT VT = N->getOperand(Vec0Idx).getValueType();
1838 bool is64BitVector = VT.is64BitVector();
1840 unsigned Alignment = 0;
1842 Alignment = cast<ConstantSDNode>(Align)->getZExtValue();
1843 unsigned NumBytes = NumVecs * VT.getVectorElementType().getSizeInBits()/8;
1844 if (Alignment > NumBytes)
1845 Alignment = NumBytes;
1846 if (Alignment < 8 && Alignment < NumBytes)
1848 // Alignment must be a power of two; make sure of that.
1849 Alignment = (Alignment & -Alignment);
1853 Align = CurDAG->getTargetConstant(Alignment, MVT::i32);
1855 unsigned OpcodeIndex;
1856 switch (VT.getSimpleVT().SimpleTy) {
1857 default: llvm_unreachable("unhandled vld/vst lane type");
1858 // Double-register operations:
1859 case MVT::v8i8: OpcodeIndex = 0; break;
1860 case MVT::v4i16: OpcodeIndex = 1; break;
1862 case MVT::v2i32: OpcodeIndex = 2; break;
1863 // Quad-register operations:
1864 case MVT::v8i16: OpcodeIndex = 0; break;
1866 case MVT::v4i32: OpcodeIndex = 1; break;
1869 std::vector<EVT> ResTys;
1871 unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs;
1874 ResTys.push_back(EVT::getVectorVT(*CurDAG->getContext(),
1875 MVT::i64, ResTyElts));
1878 ResTys.push_back(MVT::i32);
1879 ResTys.push_back(MVT::Other);
1881 SDValue Pred = getAL(CurDAG);
1882 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1884 SmallVector<SDValue, 8> Ops;
1885 Ops.push_back(MemAddr);
1886 Ops.push_back(Align);
1888 SDValue Inc = N->getOperand(AddrOpIdx + 1);
1889 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc);
1893 SDValue V0 = N->getOperand(Vec0Idx + 0);
1894 SDValue V1 = N->getOperand(Vec0Idx + 1);
1897 SuperReg = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1899 SuperReg = SDValue(PairQRegs(MVT::v4i64, V0, V1), 0);
1901 SDValue V2 = N->getOperand(Vec0Idx + 2);
1902 SDValue V3 = (NumVecs == 3)
1903 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
1904 : N->getOperand(Vec0Idx + 3);
1906 SuperReg = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1908 SuperReg = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
1910 Ops.push_back(SuperReg);
1911 Ops.push_back(getI32Imm(Lane));
1912 Ops.push_back(Pred);
1913 Ops.push_back(Reg0);
1914 Ops.push_back(Chain);
1916 unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
1917 QOpcodes[OpcodeIndex]);
1918 SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys,
1919 Ops.data(), Ops.size());
1920 cast<MachineSDNode>(VLdLn)->setMemRefs(MemOp, MemOp + 1);
1924 // Extract the subregisters.
1925 SuperReg = SDValue(VLdLn, 0);
1926 assert(ARM::dsub_7 == ARM::dsub_0+7 &&
1927 ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1928 unsigned Sub0 = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
1929 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1930 ReplaceUses(SDValue(N, Vec),
1931 CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg));
1932 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, 1));
1934 ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdLn, 2));
1938 SDNode *ARMDAGToDAGISel::SelectVLDDup(SDNode *N, bool isUpdating,
1939 unsigned NumVecs, unsigned *Opcodes) {
1940 assert(NumVecs >=2 && NumVecs <= 4 && "VLDDup NumVecs out-of-range");
1941 DebugLoc dl = N->getDebugLoc();
1943 SDValue MemAddr, Align;
1944 if (!SelectAddrMode6(N, N->getOperand(1), MemAddr, Align))
1947 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1948 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
1950 SDValue Chain = N->getOperand(0);
1951 EVT VT = N->getValueType(0);
1953 unsigned Alignment = 0;
1955 Alignment = cast<ConstantSDNode>(Align)->getZExtValue();
1956 unsigned NumBytes = NumVecs * VT.getVectorElementType().getSizeInBits()/8;
1957 if (Alignment > NumBytes)
1958 Alignment = NumBytes;
1959 if (Alignment < 8 && Alignment < NumBytes)
1961 // Alignment must be a power of two; make sure of that.
1962 Alignment = (Alignment & -Alignment);
1966 Align = CurDAG->getTargetConstant(Alignment, MVT::i32);
1968 unsigned OpcodeIndex;
1969 switch (VT.getSimpleVT().SimpleTy) {
1970 default: llvm_unreachable("unhandled vld-dup type");
1971 case MVT::v8i8: OpcodeIndex = 0; break;
1972 case MVT::v4i16: OpcodeIndex = 1; break;
1974 case MVT::v2i32: OpcodeIndex = 2; break;
1977 SDValue Pred = getAL(CurDAG);
1978 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1980 unsigned Opc = Opcodes[OpcodeIndex];
1981 SmallVector<SDValue, 6> Ops;
1982 Ops.push_back(MemAddr);
1983 Ops.push_back(Align);
1985 SDValue Inc = N->getOperand(2);
1986 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc);
1988 Ops.push_back(Pred);
1989 Ops.push_back(Reg0);
1990 Ops.push_back(Chain);
1992 unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs;
1993 std::vector<EVT> ResTys;
1994 ResTys.push_back(EVT::getVectorVT(*CurDAG->getContext(), MVT::i64,ResTyElts));
1996 ResTys.push_back(MVT::i32);
1997 ResTys.push_back(MVT::Other);
1999 CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), Ops.size());
2000 cast<MachineSDNode>(VLdDup)->setMemRefs(MemOp, MemOp + 1);
2001 SuperReg = SDValue(VLdDup, 0);
2003 // Extract the subregisters.
2004 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
2005 unsigned SubIdx = ARM::dsub_0;
2006 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
2007 ReplaceUses(SDValue(N, Vec),
2008 CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, SuperReg));
2009 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdDup, 1));
2011 ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdDup, 2));
2015 SDNode *ARMDAGToDAGISel::SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs,
2017 assert(NumVecs >= 2 && NumVecs <= 4 && "VTBL NumVecs out-of-range");
2018 DebugLoc dl = N->getDebugLoc();
2019 EVT VT = N->getValueType(0);
2020 unsigned FirstTblReg = IsExt ? 2 : 1;
2022 // Form a REG_SEQUENCE to force register allocation.
2024 SDValue V0 = N->getOperand(FirstTblReg + 0);
2025 SDValue V1 = N->getOperand(FirstTblReg + 1);
2027 RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0);
2029 SDValue V2 = N->getOperand(FirstTblReg + 2);
2030 // If it's a vtbl3, form a quad D-register and leave the last part as
2032 SDValue V3 = (NumVecs == 3)
2033 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
2034 : N->getOperand(FirstTblReg + 3);
2035 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
2038 SmallVector<SDValue, 6> Ops;
2040 Ops.push_back(N->getOperand(1));
2041 Ops.push_back(RegSeq);
2042 Ops.push_back(N->getOperand(FirstTblReg + NumVecs));
2043 Ops.push_back(getAL(CurDAG)); // predicate
2044 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // predicate register
2045 return CurDAG->getMachineNode(Opc, dl, VT, Ops.data(), Ops.size());
2048 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
2050 if (!Subtarget->hasV6T2Ops())
2053 unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
2054 : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
2057 // For unsigned extracts, check for a shift right and mask
2058 unsigned And_imm = 0;
2059 if (N->getOpcode() == ISD::AND) {
2060 if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
2062 // The immediate is a mask of the low bits iff imm & (imm+1) == 0
2063 if (And_imm & (And_imm + 1))
2066 unsigned Srl_imm = 0;
2067 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
2069 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
2071 // Note: The width operand is encoded as width-1.
2072 unsigned Width = CountTrailingOnes_32(And_imm) - 1;
2073 unsigned LSB = Srl_imm;
2074 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2075 SDValue Ops[] = { N->getOperand(0).getOperand(0),
2076 CurDAG->getTargetConstant(LSB, MVT::i32),
2077 CurDAG->getTargetConstant(Width, MVT::i32),
2078 getAL(CurDAG), Reg0 };
2079 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
2085 // Otherwise, we're looking for a shift of a shift
2086 unsigned Shl_imm = 0;
2087 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
2088 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
2089 unsigned Srl_imm = 0;
2090 if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
2091 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
2092 // Note: The width operand is encoded as width-1.
2093 unsigned Width = 32 - Srl_imm - 1;
2094 int LSB = Srl_imm - Shl_imm;
2097 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2098 SDValue Ops[] = { N->getOperand(0).getOperand(0),
2099 CurDAG->getTargetConstant(LSB, MVT::i32),
2100 CurDAG->getTargetConstant(Width, MVT::i32),
2101 getAL(CurDAG), Reg0 };
2102 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
2108 SDNode *ARMDAGToDAGISel::
2109 SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
2110 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
2113 if (SelectT2ShifterOperandReg(TrueVal, CPTmp0, CPTmp1)) {
2114 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
2115 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
2118 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
2119 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
2120 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
2121 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
2123 llvm_unreachable("Unknown so_reg opcode!");
2127 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
2128 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
2129 SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
2130 return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
2135 SDNode *ARMDAGToDAGISel::
2136 SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
2137 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
2141 if (SelectImmShifterOperand(TrueVal, CPTmp0, CPTmp2)) {
2142 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
2143 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp2, CC, CCR, InFlag };
2144 return CurDAG->SelectNodeTo(N, ARM::MOVCCsi, MVT::i32, Ops, 6);
2147 if (SelectRegShifterOperand(TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
2148 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
2149 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
2150 return CurDAG->SelectNodeTo(N, ARM::MOVCCsr, MVT::i32, Ops, 7);
2155 SDNode *ARMDAGToDAGISel::
2156 SelectT2CMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
2157 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
2158 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
2163 unsigned TrueImm = T->getZExtValue();
2164 if (is_t2_so_imm(TrueImm)) {
2165 Opc = ARM::t2MOVCCi;
2166 } else if (TrueImm <= 0xffff) {
2167 Opc = ARM::t2MOVCCi16;
2168 } else if (is_t2_so_imm_not(TrueImm)) {
2170 Opc = ARM::t2MVNCCi;
2171 } else if (TrueVal.getNode()->hasOneUse() && Subtarget->hasV6T2Ops()) {
2173 Opc = ARM::t2MOVCCi32imm;
2177 SDValue True = CurDAG->getTargetConstant(TrueImm, MVT::i32);
2178 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
2179 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
2180 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
2186 SDNode *ARMDAGToDAGISel::
2187 SelectARMCMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
2188 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
2189 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
2194 unsigned TrueImm = T->getZExtValue();
2195 bool isSoImm = is_so_imm(TrueImm);
2198 } else if (Subtarget->hasV6T2Ops() && TrueImm <= 0xffff) {
2199 Opc = ARM::MOVCCi16;
2200 } else if (is_so_imm_not(TrueImm)) {
2203 } else if (TrueVal.getNode()->hasOneUse() &&
2204 (Subtarget->hasV6T2Ops() || ARM_AM::isSOImmTwoPartVal(TrueImm))) {
2206 Opc = ARM::MOVCCi32imm;
2210 SDValue True = CurDAG->getTargetConstant(TrueImm, MVT::i32);
2211 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
2212 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
2213 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
2219 SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
2220 EVT VT = N->getValueType(0);
2221 SDValue FalseVal = N->getOperand(0);
2222 SDValue TrueVal = N->getOperand(1);
2223 SDValue CC = N->getOperand(2);
2224 SDValue CCR = N->getOperand(3);
2225 SDValue InFlag = N->getOperand(4);
2226 assert(CC.getOpcode() == ISD::Constant);
2227 assert(CCR.getOpcode() == ISD::Register);
2228 ARMCC::CondCodes CCVal =
2229 (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
2231 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
2232 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
2233 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
2234 // Pattern complexity = 18 cost = 1 size = 0
2238 if (Subtarget->isThumb()) {
2239 SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
2240 CCVal, CCR, InFlag);
2242 Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
2243 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
2247 SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
2248 CCVal, CCR, InFlag);
2250 Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
2251 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
2256 // Pattern: (ARMcmov:i32 GPR:i32:$false,
2257 // (imm:i32)<<P:Pred_so_imm>>:$true,
2259 // Emits: (MOVCCi:i32 GPR:i32:$false,
2260 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
2261 // Pattern complexity = 10 cost = 1 size = 0
2262 if (Subtarget->isThumb()) {
2263 SDNode *Res = SelectT2CMOVImmOp(N, FalseVal, TrueVal,
2264 CCVal, CCR, InFlag);
2266 Res = SelectT2CMOVImmOp(N, TrueVal, FalseVal,
2267 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
2271 SDNode *Res = SelectARMCMOVImmOp(N, FalseVal, TrueVal,
2272 CCVal, CCR, InFlag);
2274 Res = SelectARMCMOVImmOp(N, TrueVal, FalseVal,
2275 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
2281 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
2282 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
2283 // Pattern complexity = 6 cost = 1 size = 0
2285 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
2286 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
2287 // Pattern complexity = 6 cost = 11 size = 0
2289 // Also VMOVScc and VMOVDcc.
2290 SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
2291 SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
2293 switch (VT.getSimpleVT().SimpleTy) {
2294 default: assert(false && "Illegal conditional move type!");
2297 Opc = Subtarget->isThumb()
2298 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
2308 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
2311 SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
2312 // The only time a CONCAT_VECTORS operation can have legal types is when
2313 // two 64-bit vectors are concatenated to a 128-bit vector.
2314 EVT VT = N->getValueType(0);
2315 if (!VT.is128BitVector() || N->getNumOperands() != 2)
2316 llvm_unreachable("unexpected CONCAT_VECTORS");
2317 return PairDRegs(VT, N->getOperand(0), N->getOperand(1));
2320 SDNode *ARMDAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
2321 SmallVector<SDValue, 6> Ops;
2322 Ops.push_back(Node->getOperand(1)); // Ptr
2323 Ops.push_back(Node->getOperand(2)); // Low part of Val1
2324 Ops.push_back(Node->getOperand(3)); // High part of Val1
2325 if (Opc == ARM::ATOMCMPXCHG6432) {
2326 Ops.push_back(Node->getOperand(4)); // Low part of Val2
2327 Ops.push_back(Node->getOperand(5)); // High part of Val2
2329 Ops.push_back(Node->getOperand(0)); // Chain
2330 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2331 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
2332 SDNode *ResNode = CurDAG->getMachineNode(Opc, Node->getDebugLoc(),
2333 MVT::i32, MVT::i32, MVT::Other,
2334 Ops.data() ,Ops.size());
2335 cast<MachineSDNode>(ResNode)->setMemRefs(MemOp, MemOp + 1);
2339 SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
2340 DebugLoc dl = N->getDebugLoc();
2342 if (N->isMachineOpcode())
2343 return NULL; // Already selected.
2345 switch (N->getOpcode()) {
2347 case ISD::Constant: {
2348 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
2350 if (Subtarget->hasThumb2())
2351 // Thumb2-aware targets have the MOVT instruction, so all immediates can
2352 // be done with MOV + MOVT, at worst.
2355 if (Subtarget->isThumb()) {
2356 UseCP = (Val > 255 && // MOV
2357 ~Val > 255 && // MOV + MVN
2358 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
2360 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
2361 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
2362 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
2367 CurDAG->getTargetConstantPool(ConstantInt::get(
2368 Type::getInt32Ty(*CurDAG->getContext()), Val),
2369 TLI.getPointerTy());
2372 if (Subtarget->isThumb1Only()) {
2373 SDValue Pred = getAL(CurDAG);
2374 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2375 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
2376 ResNode = CurDAG->getMachineNode(ARM::tLDRpci, dl, MVT::i32, MVT::Other,
2381 CurDAG->getTargetConstant(0, MVT::i32),
2383 CurDAG->getRegister(0, MVT::i32),
2384 CurDAG->getEntryNode()
2386 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
2389 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
2393 // Other cases are autogenerated.
2396 case ISD::FrameIndex: {
2397 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
2398 int FI = cast<FrameIndexSDNode>(N)->getIndex();
2399 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
2400 if (Subtarget->isThumb1Only()) {
2401 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
2402 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2403 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, Ops, 4);
2405 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
2406 ARM::t2ADDri : ARM::ADDri);
2407 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
2408 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2409 CurDAG->getRegister(0, MVT::i32) };
2410 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
2414 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
2418 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true))
2422 if (Subtarget->isThumb1Only())
2424 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
2425 unsigned RHSV = C->getZExtValue();
2427 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
2428 unsigned ShImm = Log2_32(RHSV-1);
2431 SDValue V = N->getOperand(0);
2432 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
2433 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
2434 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2435 if (Subtarget->isThumb()) {
2436 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
2437 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
2439 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
2440 return CurDAG->SelectNodeTo(N, ARM::ADDrsi, MVT::i32, Ops, 7);
2443 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
2444 unsigned ShImm = Log2_32(RHSV+1);
2447 SDValue V = N->getOperand(0);
2448 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
2449 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
2450 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2451 if (Subtarget->isThumb()) {
2452 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
2453 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 6);
2455 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
2456 return CurDAG->SelectNodeTo(N, ARM::RSBrsi, MVT::i32, Ops, 7);
2462 // Check for unsigned bitfield extract
2463 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
2466 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
2467 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
2468 // are entirely contributed by c2 and lower 16-bits are entirely contributed
2469 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
2470 // Select it to: "movt x, ((c1 & 0xffff) >> 16)
2471 EVT VT = N->getValueType(0);
2474 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
2476 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
2479 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2480 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2483 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
2484 SDValue N2 = N0.getOperand(1);
2485 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
2488 unsigned N1CVal = N1C->getZExtValue();
2489 unsigned N2CVal = N2C->getZExtValue();
2490 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
2491 (N1CVal & 0xffffU) == 0xffffU &&
2492 (N2CVal & 0xffffU) == 0x0U) {
2493 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
2495 SDValue Ops[] = { N0.getOperand(0), Imm16,
2496 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2497 return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
2502 case ARMISD::VMOVRRD:
2503 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
2504 N->getOperand(0), getAL(CurDAG),
2505 CurDAG->getRegister(0, MVT::i32));
2506 case ISD::UMUL_LOHI: {
2507 if (Subtarget->isThumb1Only())
2509 if (Subtarget->isThumb()) {
2510 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2511 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2512 CurDAG->getRegister(0, MVT::i32) };
2513 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32,Ops,4);
2515 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2516 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2517 CurDAG->getRegister(0, MVT::i32) };
2518 return CurDAG->getMachineNode(Subtarget->hasV6Ops() ?
2519 ARM::UMULL : ARM::UMULLv5,
2520 dl, MVT::i32, MVT::i32, Ops, 5);
2523 case ISD::SMUL_LOHI: {
2524 if (Subtarget->isThumb1Only())
2526 if (Subtarget->isThumb()) {
2527 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2528 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2529 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32,Ops,4);
2531 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2532 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2533 CurDAG->getRegister(0, MVT::i32) };
2534 return CurDAG->getMachineNode(Subtarget->hasV6Ops() ?
2535 ARM::SMULL : ARM::SMULLv5,
2536 dl, MVT::i32, MVT::i32, Ops, 5);
2540 SDNode *ResNode = 0;
2541 if (Subtarget->isThumb() && Subtarget->hasThumb2())
2542 ResNode = SelectT2IndexedLoad(N);
2544 ResNode = SelectARMIndexedLoad(N);
2547 // Other cases are autogenerated.
2550 case ARMISD::BRCOND: {
2551 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2552 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2553 // Pattern complexity = 6 cost = 1 size = 0
2555 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2556 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
2557 // Pattern complexity = 6 cost = 1 size = 0
2559 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2560 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2561 // Pattern complexity = 6 cost = 1 size = 0
2563 unsigned Opc = Subtarget->isThumb() ?
2564 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
2565 SDValue Chain = N->getOperand(0);
2566 SDValue N1 = N->getOperand(1);
2567 SDValue N2 = N->getOperand(2);
2568 SDValue N3 = N->getOperand(3);
2569 SDValue InFlag = N->getOperand(4);
2570 assert(N1.getOpcode() == ISD::BasicBlock);
2571 assert(N2.getOpcode() == ISD::Constant);
2572 assert(N3.getOpcode() == ISD::Register);
2574 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2575 cast<ConstantSDNode>(N2)->getZExtValue()),
2577 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
2578 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
2580 Chain = SDValue(ResNode, 0);
2581 if (N->getNumValues() == 2) {
2582 InFlag = SDValue(ResNode, 1);
2583 ReplaceUses(SDValue(N, 1), InFlag);
2585 ReplaceUses(SDValue(N, 0),
2586 SDValue(Chain.getNode(), Chain.getResNo()));
2590 return SelectCMOVOp(N);
2591 case ARMISD::VZIP: {
2593 EVT VT = N->getValueType(0);
2594 switch (VT.getSimpleVT().SimpleTy) {
2595 default: return NULL;
2596 case MVT::v8i8: Opc = ARM::VZIPd8; break;
2597 case MVT::v4i16: Opc = ARM::VZIPd16; break;
2599 case MVT::v2i32: Opc = ARM::VZIPd32; break;
2600 case MVT::v16i8: Opc = ARM::VZIPq8; break;
2601 case MVT::v8i16: Opc = ARM::VZIPq16; break;
2603 case MVT::v4i32: Opc = ARM::VZIPq32; break;
2605 SDValue Pred = getAL(CurDAG);
2606 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2607 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2608 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2610 case ARMISD::VUZP: {
2612 EVT VT = N->getValueType(0);
2613 switch (VT.getSimpleVT().SimpleTy) {
2614 default: return NULL;
2615 case MVT::v8i8: Opc = ARM::VUZPd8; break;
2616 case MVT::v4i16: Opc = ARM::VUZPd16; break;
2618 case MVT::v2i32: Opc = ARM::VUZPd32; break;
2619 case MVT::v16i8: Opc = ARM::VUZPq8; break;
2620 case MVT::v8i16: Opc = ARM::VUZPq16; break;
2622 case MVT::v4i32: Opc = ARM::VUZPq32; break;
2624 SDValue Pred = getAL(CurDAG);
2625 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2626 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2627 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2629 case ARMISD::VTRN: {
2631 EVT VT = N->getValueType(0);
2632 switch (VT.getSimpleVT().SimpleTy) {
2633 default: return NULL;
2634 case MVT::v8i8: Opc = ARM::VTRNd8; break;
2635 case MVT::v4i16: Opc = ARM::VTRNd16; break;
2637 case MVT::v2i32: Opc = ARM::VTRNd32; break;
2638 case MVT::v16i8: Opc = ARM::VTRNq8; break;
2639 case MVT::v8i16: Opc = ARM::VTRNq16; break;
2641 case MVT::v4i32: Opc = ARM::VTRNq32; break;
2643 SDValue Pred = getAL(CurDAG);
2644 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2645 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2646 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2648 case ARMISD::BUILD_VECTOR: {
2649 EVT VecVT = N->getValueType(0);
2650 EVT EltVT = VecVT.getVectorElementType();
2651 unsigned NumElts = VecVT.getVectorNumElements();
2652 if (EltVT == MVT::f64) {
2653 assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
2654 return PairDRegs(VecVT, N->getOperand(0), N->getOperand(1));
2656 assert(EltVT == MVT::f32 && "unexpected type for BUILD_VECTOR");
2658 return PairSRegs(VecVT, N->getOperand(0), N->getOperand(1));
2659 assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
2660 return QuadSRegs(VecVT, N->getOperand(0), N->getOperand(1),
2661 N->getOperand(2), N->getOperand(3));
2664 case ARMISD::VLD2DUP: {
2665 unsigned Opcodes[] = { ARM::VLD2DUPd8Pseudo, ARM::VLD2DUPd16Pseudo,
2666 ARM::VLD2DUPd32Pseudo };
2667 return SelectVLDDup(N, false, 2, Opcodes);
2670 case ARMISD::VLD3DUP: {
2671 unsigned Opcodes[] = { ARM::VLD3DUPd8Pseudo, ARM::VLD3DUPd16Pseudo,
2672 ARM::VLD3DUPd32Pseudo };
2673 return SelectVLDDup(N, false, 3, Opcodes);
2676 case ARMISD::VLD4DUP: {
2677 unsigned Opcodes[] = { ARM::VLD4DUPd8Pseudo, ARM::VLD4DUPd16Pseudo,
2678 ARM::VLD4DUPd32Pseudo };
2679 return SelectVLDDup(N, false, 4, Opcodes);
2682 case ARMISD::VLD2DUP_UPD: {
2683 unsigned Opcodes[] = { ARM::VLD2DUPd8Pseudo_UPD, ARM::VLD2DUPd16Pseudo_UPD,
2684 ARM::VLD2DUPd32Pseudo_UPD };
2685 return SelectVLDDup(N, true, 2, Opcodes);
2688 case ARMISD::VLD3DUP_UPD: {
2689 unsigned Opcodes[] = { ARM::VLD3DUPd8Pseudo_UPD, ARM::VLD3DUPd16Pseudo_UPD,
2690 ARM::VLD3DUPd32Pseudo_UPD };
2691 return SelectVLDDup(N, true, 3, Opcodes);
2694 case ARMISD::VLD4DUP_UPD: {
2695 unsigned Opcodes[] = { ARM::VLD4DUPd8Pseudo_UPD, ARM::VLD4DUPd16Pseudo_UPD,
2696 ARM::VLD4DUPd32Pseudo_UPD };
2697 return SelectVLDDup(N, true, 4, Opcodes);
2700 case ARMISD::VLD1_UPD: {
2701 unsigned DOpcodes[] = { ARM::VLD1d8_UPD, ARM::VLD1d16_UPD,
2702 ARM::VLD1d32_UPD, ARM::VLD1d64_UPD };
2703 unsigned QOpcodes[] = { ARM::VLD1q8Pseudo_UPD, ARM::VLD1q16Pseudo_UPD,
2704 ARM::VLD1q32Pseudo_UPD, ARM::VLD1q64Pseudo_UPD };
2705 return SelectVLD(N, true, 1, DOpcodes, QOpcodes, 0);
2708 case ARMISD::VLD2_UPD: {
2709 unsigned DOpcodes[] = { ARM::VLD2d8Pseudo_UPD, ARM::VLD2d16Pseudo_UPD,
2710 ARM::VLD2d32Pseudo_UPD, ARM::VLD1q64Pseudo_UPD };
2711 unsigned QOpcodes[] = { ARM::VLD2q8Pseudo_UPD, ARM::VLD2q16Pseudo_UPD,
2712 ARM::VLD2q32Pseudo_UPD };
2713 return SelectVLD(N, true, 2, DOpcodes, QOpcodes, 0);
2716 case ARMISD::VLD3_UPD: {
2717 unsigned DOpcodes[] = { ARM::VLD3d8Pseudo_UPD, ARM::VLD3d16Pseudo_UPD,
2718 ARM::VLD3d32Pseudo_UPD, ARM::VLD1d64TPseudo_UPD };
2719 unsigned QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD,
2720 ARM::VLD3q16Pseudo_UPD,
2721 ARM::VLD3q32Pseudo_UPD };
2722 unsigned QOpcodes1[] = { ARM::VLD3q8oddPseudo_UPD,
2723 ARM::VLD3q16oddPseudo_UPD,
2724 ARM::VLD3q32oddPseudo_UPD };
2725 return SelectVLD(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
2728 case ARMISD::VLD4_UPD: {
2729 unsigned DOpcodes[] = { ARM::VLD4d8Pseudo_UPD, ARM::VLD4d16Pseudo_UPD,
2730 ARM::VLD4d32Pseudo_UPD, ARM::VLD1d64QPseudo_UPD };
2731 unsigned QOpcodes0[] = { ARM::VLD4q8Pseudo_UPD,
2732 ARM::VLD4q16Pseudo_UPD,
2733 ARM::VLD4q32Pseudo_UPD };
2734 unsigned QOpcodes1[] = { ARM::VLD4q8oddPseudo_UPD,
2735 ARM::VLD4q16oddPseudo_UPD,
2736 ARM::VLD4q32oddPseudo_UPD };
2737 return SelectVLD(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
2740 case ARMISD::VLD2LN_UPD: {
2741 unsigned DOpcodes[] = { ARM::VLD2LNd8Pseudo_UPD, ARM::VLD2LNd16Pseudo_UPD,
2742 ARM::VLD2LNd32Pseudo_UPD };
2743 unsigned QOpcodes[] = { ARM::VLD2LNq16Pseudo_UPD,
2744 ARM::VLD2LNq32Pseudo_UPD };
2745 return SelectVLDSTLane(N, true, true, 2, DOpcodes, QOpcodes);
2748 case ARMISD::VLD3LN_UPD: {
2749 unsigned DOpcodes[] = { ARM::VLD3LNd8Pseudo_UPD, ARM::VLD3LNd16Pseudo_UPD,
2750 ARM::VLD3LNd32Pseudo_UPD };
2751 unsigned QOpcodes[] = { ARM::VLD3LNq16Pseudo_UPD,
2752 ARM::VLD3LNq32Pseudo_UPD };
2753 return SelectVLDSTLane(N, true, true, 3, DOpcodes, QOpcodes);
2756 case ARMISD::VLD4LN_UPD: {
2757 unsigned DOpcodes[] = { ARM::VLD4LNd8Pseudo_UPD, ARM::VLD4LNd16Pseudo_UPD,
2758 ARM::VLD4LNd32Pseudo_UPD };
2759 unsigned QOpcodes[] = { ARM::VLD4LNq16Pseudo_UPD,
2760 ARM::VLD4LNq32Pseudo_UPD };
2761 return SelectVLDSTLane(N, true, true, 4, DOpcodes, QOpcodes);
2764 case ARMISD::VST1_UPD: {
2765 unsigned DOpcodes[] = { ARM::VST1d8_UPD, ARM::VST1d16_UPD,
2766 ARM::VST1d32_UPD, ARM::VST1d64_UPD };
2767 unsigned QOpcodes[] = { ARM::VST1q8Pseudo_UPD, ARM::VST1q16Pseudo_UPD,
2768 ARM::VST1q32Pseudo_UPD, ARM::VST1q64Pseudo_UPD };
2769 return SelectVST(N, true, 1, DOpcodes, QOpcodes, 0);
2772 case ARMISD::VST2_UPD: {
2773 unsigned DOpcodes[] = { ARM::VST2d8Pseudo_UPD, ARM::VST2d16Pseudo_UPD,
2774 ARM::VST2d32Pseudo_UPD, ARM::VST1q64Pseudo_UPD };
2775 unsigned QOpcodes[] = { ARM::VST2q8Pseudo_UPD, ARM::VST2q16Pseudo_UPD,
2776 ARM::VST2q32Pseudo_UPD };
2777 return SelectVST(N, true, 2, DOpcodes, QOpcodes, 0);
2780 case ARMISD::VST3_UPD: {
2781 unsigned DOpcodes[] = { ARM::VST3d8Pseudo_UPD, ARM::VST3d16Pseudo_UPD,
2782 ARM::VST3d32Pseudo_UPD, ARM::VST1d64TPseudo_UPD };
2783 unsigned QOpcodes0[] = { ARM::VST3q8Pseudo_UPD,
2784 ARM::VST3q16Pseudo_UPD,
2785 ARM::VST3q32Pseudo_UPD };
2786 unsigned QOpcodes1[] = { ARM::VST3q8oddPseudo_UPD,
2787 ARM::VST3q16oddPseudo_UPD,
2788 ARM::VST3q32oddPseudo_UPD };
2789 return SelectVST(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
2792 case ARMISD::VST4_UPD: {
2793 unsigned DOpcodes[] = { ARM::VST4d8Pseudo_UPD, ARM::VST4d16Pseudo_UPD,
2794 ARM::VST4d32Pseudo_UPD, ARM::VST1d64QPseudo_UPD };
2795 unsigned QOpcodes0[] = { ARM::VST4q8Pseudo_UPD,
2796 ARM::VST4q16Pseudo_UPD,
2797 ARM::VST4q32Pseudo_UPD };
2798 unsigned QOpcodes1[] = { ARM::VST4q8oddPseudo_UPD,
2799 ARM::VST4q16oddPseudo_UPD,
2800 ARM::VST4q32oddPseudo_UPD };
2801 return SelectVST(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
2804 case ARMISD::VST2LN_UPD: {
2805 unsigned DOpcodes[] = { ARM::VST2LNd8Pseudo_UPD, ARM::VST2LNd16Pseudo_UPD,
2806 ARM::VST2LNd32Pseudo_UPD };
2807 unsigned QOpcodes[] = { ARM::VST2LNq16Pseudo_UPD,
2808 ARM::VST2LNq32Pseudo_UPD };
2809 return SelectVLDSTLane(N, false, true, 2, DOpcodes, QOpcodes);
2812 case ARMISD::VST3LN_UPD: {
2813 unsigned DOpcodes[] = { ARM::VST3LNd8Pseudo_UPD, ARM::VST3LNd16Pseudo_UPD,
2814 ARM::VST3LNd32Pseudo_UPD };
2815 unsigned QOpcodes[] = { ARM::VST3LNq16Pseudo_UPD,
2816 ARM::VST3LNq32Pseudo_UPD };
2817 return SelectVLDSTLane(N, false, true, 3, DOpcodes, QOpcodes);
2820 case ARMISD::VST4LN_UPD: {
2821 unsigned DOpcodes[] = { ARM::VST4LNd8Pseudo_UPD, ARM::VST4LNd16Pseudo_UPD,
2822 ARM::VST4LNd32Pseudo_UPD };
2823 unsigned QOpcodes[] = { ARM::VST4LNq16Pseudo_UPD,
2824 ARM::VST4LNq32Pseudo_UPD };
2825 return SelectVLDSTLane(N, false, true, 4, DOpcodes, QOpcodes);
2828 case ISD::INTRINSIC_VOID:
2829 case ISD::INTRINSIC_W_CHAIN: {
2830 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2835 case Intrinsic::arm_ldrexd: {
2836 SDValue MemAddr = N->getOperand(2);
2837 DebugLoc dl = N->getDebugLoc();
2838 SDValue Chain = N->getOperand(0);
2840 unsigned NewOpc = ARM::LDREXD;
2841 if (Subtarget->isThumb() && Subtarget->hasThumb2())
2842 NewOpc = ARM::t2LDREXD;
2844 // arm_ldrexd returns a i64 value in {i32, i32}
2845 std::vector<EVT> ResTys;
2846 ResTys.push_back(MVT::i32);
2847 ResTys.push_back(MVT::i32);
2848 ResTys.push_back(MVT::Other);
2850 // place arguments in the right order
2851 SmallVector<SDValue, 7> Ops;
2852 Ops.push_back(MemAddr);
2853 Ops.push_back(getAL(CurDAG));
2854 Ops.push_back(CurDAG->getRegister(0, MVT::i32));
2855 Ops.push_back(Chain);
2856 SDNode *Ld = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops.data(),
2858 // Transfer memoperands.
2859 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2860 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2861 cast<MachineSDNode>(Ld)->setMemRefs(MemOp, MemOp + 1);
2863 // Until there's support for specifing explicit register constraints
2864 // like the use of even/odd register pair, hardcode ldrexd to always
2865 // use the pair [R0, R1] to hold the load result.
2866 Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ARM::R0,
2867 SDValue(Ld, 0), SDValue(0,0));
2868 Chain = CurDAG->getCopyToReg(Chain, dl, ARM::R1,
2869 SDValue(Ld, 1), Chain.getValue(1));
2872 SDValue Glue = Chain.getValue(1);
2873 if (!SDValue(N, 0).use_empty()) {
2874 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2875 ARM::R0, MVT::i32, Glue);
2876 Glue = Result.getValue(2);
2877 ReplaceUses(SDValue(N, 0), Result);
2879 if (!SDValue(N, 1).use_empty()) {
2880 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2881 ARM::R1, MVT::i32, Glue);
2882 Glue = Result.getValue(2);
2883 ReplaceUses(SDValue(N, 1), Result);
2886 ReplaceUses(SDValue(N, 2), SDValue(Ld, 2));
2890 case Intrinsic::arm_strexd: {
2891 DebugLoc dl = N->getDebugLoc();
2892 SDValue Chain = N->getOperand(0);
2893 SDValue Val0 = N->getOperand(2);
2894 SDValue Val1 = N->getOperand(3);
2895 SDValue MemAddr = N->getOperand(4);
2897 // Until there's support for specifing explicit register constraints
2898 // like the use of even/odd register pair, hardcode strexd to always
2899 // use the pair [R2, R3] to hold the i64 (i32, i32) value to be stored.
2900 Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ARM::R2, Val0,
2902 Chain = CurDAG->getCopyToReg(Chain, dl, ARM::R3, Val1, Chain.getValue(1));
2904 SDValue Glue = Chain.getValue(1);
2905 Val0 = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2906 ARM::R2, MVT::i32, Glue);
2907 Glue = Val0.getValue(1);
2908 Val1 = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2909 ARM::R3, MVT::i32, Glue);
2911 // Store exclusive double return a i32 value which is the return status
2912 // of the issued store.
2913 std::vector<EVT> ResTys;
2914 ResTys.push_back(MVT::i32);
2915 ResTys.push_back(MVT::Other);
2917 // place arguments in the right order
2918 SmallVector<SDValue, 7> Ops;
2919 Ops.push_back(Val0);
2920 Ops.push_back(Val1);
2921 Ops.push_back(MemAddr);
2922 Ops.push_back(getAL(CurDAG));
2923 Ops.push_back(CurDAG->getRegister(0, MVT::i32));
2924 Ops.push_back(Chain);
2926 unsigned NewOpc = ARM::STREXD;
2927 if (Subtarget->isThumb() && Subtarget->hasThumb2())
2928 NewOpc = ARM::t2STREXD;
2930 SDNode *St = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops.data(),
2932 // Transfer memoperands.
2933 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2934 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2935 cast<MachineSDNode>(St)->setMemRefs(MemOp, MemOp + 1);
2940 case Intrinsic::arm_neon_vld1: {
2941 unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
2942 ARM::VLD1d32, ARM::VLD1d64 };
2943 unsigned QOpcodes[] = { ARM::VLD1q8Pseudo, ARM::VLD1q16Pseudo,
2944 ARM::VLD1q32Pseudo, ARM::VLD1q64Pseudo };
2945 return SelectVLD(N, false, 1, DOpcodes, QOpcodes, 0);
2948 case Intrinsic::arm_neon_vld2: {
2949 unsigned DOpcodes[] = { ARM::VLD2d8Pseudo, ARM::VLD2d16Pseudo,
2950 ARM::VLD2d32Pseudo, ARM::VLD1q64Pseudo };
2951 unsigned QOpcodes[] = { ARM::VLD2q8Pseudo, ARM::VLD2q16Pseudo,
2952 ARM::VLD2q32Pseudo };
2953 return SelectVLD(N, false, 2, DOpcodes, QOpcodes, 0);
2956 case Intrinsic::arm_neon_vld3: {
2957 unsigned DOpcodes[] = { ARM::VLD3d8Pseudo, ARM::VLD3d16Pseudo,
2958 ARM::VLD3d32Pseudo, ARM::VLD1d64TPseudo };
2959 unsigned QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD,
2960 ARM::VLD3q16Pseudo_UPD,
2961 ARM::VLD3q32Pseudo_UPD };
2962 unsigned QOpcodes1[] = { ARM::VLD3q8oddPseudo,
2963 ARM::VLD3q16oddPseudo,
2964 ARM::VLD3q32oddPseudo };
2965 return SelectVLD(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
2968 case Intrinsic::arm_neon_vld4: {
2969 unsigned DOpcodes[] = { ARM::VLD4d8Pseudo, ARM::VLD4d16Pseudo,
2970 ARM::VLD4d32Pseudo, ARM::VLD1d64QPseudo };
2971 unsigned QOpcodes0[] = { ARM::VLD4q8Pseudo_UPD,
2972 ARM::VLD4q16Pseudo_UPD,
2973 ARM::VLD4q32Pseudo_UPD };
2974 unsigned QOpcodes1[] = { ARM::VLD4q8oddPseudo,
2975 ARM::VLD4q16oddPseudo,
2976 ARM::VLD4q32oddPseudo };
2977 return SelectVLD(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
2980 case Intrinsic::arm_neon_vld2lane: {
2981 unsigned DOpcodes[] = { ARM::VLD2LNd8Pseudo, ARM::VLD2LNd16Pseudo,
2982 ARM::VLD2LNd32Pseudo };
2983 unsigned QOpcodes[] = { ARM::VLD2LNq16Pseudo, ARM::VLD2LNq32Pseudo };
2984 return SelectVLDSTLane(N, true, false, 2, DOpcodes, QOpcodes);
2987 case Intrinsic::arm_neon_vld3lane: {
2988 unsigned DOpcodes[] = { ARM::VLD3LNd8Pseudo, ARM::VLD3LNd16Pseudo,
2989 ARM::VLD3LNd32Pseudo };
2990 unsigned QOpcodes[] = { ARM::VLD3LNq16Pseudo, ARM::VLD3LNq32Pseudo };
2991 return SelectVLDSTLane(N, true, false, 3, DOpcodes, QOpcodes);
2994 case Intrinsic::arm_neon_vld4lane: {
2995 unsigned DOpcodes[] = { ARM::VLD4LNd8Pseudo, ARM::VLD4LNd16Pseudo,
2996 ARM::VLD4LNd32Pseudo };
2997 unsigned QOpcodes[] = { ARM::VLD4LNq16Pseudo, ARM::VLD4LNq32Pseudo };
2998 return SelectVLDSTLane(N, true, false, 4, DOpcodes, QOpcodes);
3001 case Intrinsic::arm_neon_vst1: {
3002 unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
3003 ARM::VST1d32, ARM::VST1d64 };
3004 unsigned QOpcodes[] = { ARM::VST1q8Pseudo, ARM::VST1q16Pseudo,
3005 ARM::VST1q32Pseudo, ARM::VST1q64Pseudo };
3006 return SelectVST(N, false, 1, DOpcodes, QOpcodes, 0);
3009 case Intrinsic::arm_neon_vst2: {
3010 unsigned DOpcodes[] = { ARM::VST2d8Pseudo, ARM::VST2d16Pseudo,
3011 ARM::VST2d32Pseudo, ARM::VST1q64Pseudo };
3012 unsigned QOpcodes[] = { ARM::VST2q8Pseudo, ARM::VST2q16Pseudo,
3013 ARM::VST2q32Pseudo };
3014 return SelectVST(N, false, 2, DOpcodes, QOpcodes, 0);
3017 case Intrinsic::arm_neon_vst3: {
3018 unsigned DOpcodes[] = { ARM::VST3d8Pseudo, ARM::VST3d16Pseudo,
3019 ARM::VST3d32Pseudo, ARM::VST1d64TPseudo };
3020 unsigned QOpcodes0[] = { ARM::VST3q8Pseudo_UPD,
3021 ARM::VST3q16Pseudo_UPD,
3022 ARM::VST3q32Pseudo_UPD };
3023 unsigned QOpcodes1[] = { ARM::VST3q8oddPseudo,
3024 ARM::VST3q16oddPseudo,
3025 ARM::VST3q32oddPseudo };
3026 return SelectVST(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
3029 case Intrinsic::arm_neon_vst4: {
3030 unsigned DOpcodes[] = { ARM::VST4d8Pseudo, ARM::VST4d16Pseudo,
3031 ARM::VST4d32Pseudo, ARM::VST1d64QPseudo };
3032 unsigned QOpcodes0[] = { ARM::VST4q8Pseudo_UPD,
3033 ARM::VST4q16Pseudo_UPD,
3034 ARM::VST4q32Pseudo_UPD };
3035 unsigned QOpcodes1[] = { ARM::VST4q8oddPseudo,
3036 ARM::VST4q16oddPseudo,
3037 ARM::VST4q32oddPseudo };
3038 return SelectVST(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
3041 case Intrinsic::arm_neon_vst2lane: {
3042 unsigned DOpcodes[] = { ARM::VST2LNd8Pseudo, ARM::VST2LNd16Pseudo,
3043 ARM::VST2LNd32Pseudo };
3044 unsigned QOpcodes[] = { ARM::VST2LNq16Pseudo, ARM::VST2LNq32Pseudo };
3045 return SelectVLDSTLane(N, false, false, 2, DOpcodes, QOpcodes);
3048 case Intrinsic::arm_neon_vst3lane: {
3049 unsigned DOpcodes[] = { ARM::VST3LNd8Pseudo, ARM::VST3LNd16Pseudo,
3050 ARM::VST3LNd32Pseudo };
3051 unsigned QOpcodes[] = { ARM::VST3LNq16Pseudo, ARM::VST3LNq32Pseudo };
3052 return SelectVLDSTLane(N, false, false, 3, DOpcodes, QOpcodes);
3055 case Intrinsic::arm_neon_vst4lane: {
3056 unsigned DOpcodes[] = { ARM::VST4LNd8Pseudo, ARM::VST4LNd16Pseudo,
3057 ARM::VST4LNd32Pseudo };
3058 unsigned QOpcodes[] = { ARM::VST4LNq16Pseudo, ARM::VST4LNq32Pseudo };
3059 return SelectVLDSTLane(N, false, false, 4, DOpcodes, QOpcodes);
3065 case ISD::INTRINSIC_WO_CHAIN: {
3066 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3071 case Intrinsic::arm_neon_vtbl2:
3072 return SelectVTBL(N, false, 2, ARM::VTBL2Pseudo);
3073 case Intrinsic::arm_neon_vtbl3:
3074 return SelectVTBL(N, false, 3, ARM::VTBL3Pseudo);
3075 case Intrinsic::arm_neon_vtbl4:
3076 return SelectVTBL(N, false, 4, ARM::VTBL4Pseudo);
3078 case Intrinsic::arm_neon_vtbx2:
3079 return SelectVTBL(N, true, 2, ARM::VTBX2Pseudo);
3080 case Intrinsic::arm_neon_vtbx3:
3081 return SelectVTBL(N, true, 3, ARM::VTBX3Pseudo);
3082 case Intrinsic::arm_neon_vtbx4:
3083 return SelectVTBL(N, true, 4, ARM::VTBX4Pseudo);
3088 case ARMISD::VTBL1: {
3089 DebugLoc dl = N->getDebugLoc();
3090 EVT VT = N->getValueType(0);
3091 SmallVector<SDValue, 6> Ops;
3093 Ops.push_back(N->getOperand(0));
3094 Ops.push_back(N->getOperand(1));
3095 Ops.push_back(getAL(CurDAG)); // Predicate
3096 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // Predicate Register
3097 return CurDAG->getMachineNode(ARM::VTBL1, dl, VT, Ops.data(), Ops.size());
3099 case ARMISD::VTBL2: {
3100 DebugLoc dl = N->getDebugLoc();
3101 EVT VT = N->getValueType(0);
3103 // Form a REG_SEQUENCE to force register allocation.
3104 SDValue V0 = N->getOperand(0);
3105 SDValue V1 = N->getOperand(1);
3106 SDValue RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0);
3108 SmallVector<SDValue, 6> Ops;
3109 Ops.push_back(RegSeq);
3110 Ops.push_back(N->getOperand(2));
3111 Ops.push_back(getAL(CurDAG)); // Predicate
3112 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // Predicate Register
3113 return CurDAG->getMachineNode(ARM::VTBL2Pseudo, dl, VT,
3114 Ops.data(), Ops.size());
3117 case ISD::CONCAT_VECTORS:
3118 return SelectConcatVector(N);
3120 case ARMISD::ATOMOR64_DAG:
3121 return SelectAtomic64(N, ARM::ATOMOR6432);
3122 case ARMISD::ATOMXOR64_DAG:
3123 return SelectAtomic64(N, ARM::ATOMXOR6432);
3124 case ARMISD::ATOMADD64_DAG:
3125 return SelectAtomic64(N, ARM::ATOMADD6432);
3126 case ARMISD::ATOMSUB64_DAG:
3127 return SelectAtomic64(N, ARM::ATOMSUB6432);
3128 case ARMISD::ATOMNAND64_DAG:
3129 return SelectAtomic64(N, ARM::ATOMNAND6432);
3130 case ARMISD::ATOMAND64_DAG:
3131 return SelectAtomic64(N, ARM::ATOMAND6432);
3132 case ARMISD::ATOMSWAP64_DAG:
3133 return SelectAtomic64(N, ARM::ATOMSWAP6432);
3134 case ARMISD::ATOMCMPXCHG64_DAG:
3135 return SelectAtomic64(N, ARM::ATOMCMPXCHG6432);
3138 return SelectCode(N);
3141 bool ARMDAGToDAGISel::
3142 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
3143 std::vector<SDValue> &OutOps) {
3144 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
3145 // Require the address to be in a register. That is safe for all ARM
3146 // variants and it is hard to do anything much smarter without knowing
3147 // how the operand is used.
3148 OutOps.push_back(Op);
3152 /// createARMISelDag - This pass converts a legalized DAG into a
3153 /// ARM-specific DAG, ready for instruction scheduling.
3155 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
3156 CodeGenOpt::Level OptLevel) {
3157 return new ARMDAGToDAGISel(TM, OptLevel);