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"),
51 DisableARMIntABS("disable-arm-int-abs", cl::Hidden,
52 cl::desc("Enable / disable ARM integer abs transform"),
55 //===--------------------------------------------------------------------===//
56 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
57 /// instructions for SelectionDAG operations.
62 AM2_BASE, // Simple AM2 (+-imm12)
63 AM2_SHOP // Shifter-op AM2
66 class ARMDAGToDAGISel : public SelectionDAGISel {
67 ARMBaseTargetMachine &TM;
68 const ARMBaseInstrInfo *TII;
70 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
71 /// make the right decision when generating code for different targets.
72 const ARMSubtarget *Subtarget;
75 explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
76 CodeGenOpt::Level OptLevel)
77 : SelectionDAGISel(tm, OptLevel), TM(tm),
78 TII(static_cast<const ARMBaseInstrInfo*>(TM.getInstrInfo())),
79 Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
82 virtual const char *getPassName() const {
83 return "ARM Instruction Selection";
86 /// getI32Imm - Return a target constant of type i32 with the specified
88 inline SDValue getI32Imm(unsigned Imm) {
89 return CurDAG->getTargetConstant(Imm, MVT::i32);
92 SDNode *Select(SDNode *N);
95 bool hasNoVMLxHazardUse(SDNode *N) const;
96 bool isShifterOpProfitable(const SDValue &Shift,
97 ARM_AM::ShiftOpc ShOpcVal, unsigned ShAmt);
98 bool SelectRegShifterOperand(SDValue N, SDValue &A,
99 SDValue &B, SDValue &C,
100 bool CheckProfitability = true);
101 bool SelectImmShifterOperand(SDValue N, SDValue &A,
102 SDValue &B, bool CheckProfitability = true);
103 bool SelectShiftRegShifterOperand(SDValue N, SDValue &A,
104 SDValue &B, SDValue &C) {
105 // Don't apply the profitability check
106 return SelectRegShifterOperand(N, A, B, C, false);
108 bool SelectShiftImmShifterOperand(SDValue N, SDValue &A,
110 // Don't apply the profitability check
111 return SelectImmShifterOperand(N, A, B, false);
114 bool SelectAddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm);
115 bool SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset, SDValue &Opc);
117 AddrMode2Type SelectAddrMode2Worker(SDValue N, SDValue &Base,
118 SDValue &Offset, SDValue &Opc);
119 bool SelectAddrMode2Base(SDValue N, SDValue &Base, SDValue &Offset,
121 return SelectAddrMode2Worker(N, Base, Offset, Opc) == AM2_BASE;
124 bool SelectAddrMode2ShOp(SDValue N, SDValue &Base, SDValue &Offset,
126 return SelectAddrMode2Worker(N, Base, Offset, Opc) == AM2_SHOP;
129 bool SelectAddrMode2(SDValue N, SDValue &Base, SDValue &Offset,
131 SelectAddrMode2Worker(N, Base, Offset, Opc);
132 // return SelectAddrMode2ShOp(N, Base, Offset, Opc);
133 // This always matches one way or another.
137 bool SelectAddrMode2OffsetReg(SDNode *Op, SDValue N,
138 SDValue &Offset, SDValue &Opc);
139 bool SelectAddrMode2OffsetImm(SDNode *Op, SDValue N,
140 SDValue &Offset, SDValue &Opc);
141 bool SelectAddrMode2OffsetImmPre(SDNode *Op, SDValue N,
142 SDValue &Offset, SDValue &Opc);
143 bool SelectAddrOffsetNone(SDValue N, SDValue &Base);
144 bool SelectAddrMode3(SDValue N, SDValue &Base,
145 SDValue &Offset, SDValue &Opc);
146 bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
147 SDValue &Offset, SDValue &Opc);
148 bool SelectAddrMode5(SDValue N, SDValue &Base,
150 bool SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,SDValue &Align);
151 bool SelectAddrMode6Offset(SDNode *Op, SDValue N, SDValue &Offset);
153 bool SelectAddrModePC(SDValue N, SDValue &Offset, SDValue &Label);
155 // Thumb Addressing Modes:
156 bool SelectThumbAddrModeRR(SDValue N, SDValue &Base, SDValue &Offset);
157 bool SelectThumbAddrModeRI(SDValue N, SDValue &Base, SDValue &Offset,
159 bool SelectThumbAddrModeRI5S1(SDValue N, SDValue &Base, SDValue &Offset);
160 bool SelectThumbAddrModeRI5S2(SDValue N, SDValue &Base, SDValue &Offset);
161 bool SelectThumbAddrModeRI5S4(SDValue N, SDValue &Base, SDValue &Offset);
162 bool SelectThumbAddrModeImm5S(SDValue N, unsigned Scale, SDValue &Base,
164 bool SelectThumbAddrModeImm5S1(SDValue N, SDValue &Base,
166 bool SelectThumbAddrModeImm5S2(SDValue N, SDValue &Base,
168 bool SelectThumbAddrModeImm5S4(SDValue N, SDValue &Base,
170 bool SelectThumbAddrModeSP(SDValue N, SDValue &Base, SDValue &OffImm);
172 // Thumb 2 Addressing Modes:
173 bool SelectT2ShifterOperandReg(SDValue N,
174 SDValue &BaseReg, SDValue &Opc);
175 bool SelectT2AddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm);
176 bool SelectT2AddrModeImm8(SDValue N, SDValue &Base,
178 bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
180 bool SelectT2AddrModeSoReg(SDValue N, SDValue &Base,
181 SDValue &OffReg, SDValue &ShImm);
183 inline bool is_so_imm(unsigned Imm) const {
184 return ARM_AM::getSOImmVal(Imm) != -1;
187 inline bool is_so_imm_not(unsigned Imm) const {
188 return ARM_AM::getSOImmVal(~Imm) != -1;
191 inline bool is_t2_so_imm(unsigned Imm) const {
192 return ARM_AM::getT2SOImmVal(Imm) != -1;
195 inline bool is_t2_so_imm_not(unsigned Imm) const {
196 return ARM_AM::getT2SOImmVal(~Imm) != -1;
199 // Include the pieces autogenerated from the target description.
200 #include "ARMGenDAGISel.inc"
203 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
205 SDNode *SelectARMIndexedLoad(SDNode *N);
206 SDNode *SelectT2IndexedLoad(SDNode *N);
208 /// SelectVLD - Select NEON load intrinsics. NumVecs should be
209 /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for
210 /// loads of D registers and even subregs and odd subregs of Q registers.
211 /// For NumVecs <= 2, QOpcodes1 is not used.
212 SDNode *SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs,
214 unsigned *QOpcodes0, unsigned *QOpcodes1);
216 /// SelectVST - Select NEON store intrinsics. NumVecs should
217 /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for
218 /// stores of D registers and even subregs and odd subregs of Q registers.
219 /// For NumVecs <= 2, QOpcodes1 is not used.
220 SDNode *SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs,
222 unsigned *QOpcodes0, unsigned *QOpcodes1);
224 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
225 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
226 /// load/store of D registers and Q registers.
227 SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad,
228 bool isUpdating, unsigned NumVecs,
229 unsigned *DOpcodes, unsigned *QOpcodes);
231 /// SelectVLDDup - Select NEON load-duplicate intrinsics. NumVecs
232 /// should be 2, 3 or 4. The opcode array specifies the instructions used
233 /// for loading D registers. (Q registers are not supported.)
234 SDNode *SelectVLDDup(SDNode *N, bool isUpdating, unsigned NumVecs,
237 /// SelectVTBL - Select NEON VTBL and VTBX intrinsics. NumVecs should be 2,
238 /// 3 or 4. These are custom-selected so that a REG_SEQUENCE can be
239 /// generated to force the table registers to be consecutive.
240 SDNode *SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs, unsigned Opc);
242 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
243 SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
245 /// SelectCMOVOp - Select CMOV instructions for ARM.
246 SDNode *SelectCMOVOp(SDNode *N);
247 SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
248 ARMCC::CondCodes CCVal, SDValue CCR,
250 SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
251 ARMCC::CondCodes CCVal, SDValue CCR,
253 SDNode *SelectT2CMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
254 ARMCC::CondCodes CCVal, SDValue CCR,
256 SDNode *SelectARMCMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
257 ARMCC::CondCodes CCVal, SDValue CCR,
260 // Select special operations if node forms integer ABS pattern
261 SDNode *SelectABSOp(SDNode *N);
263 SDNode *SelectConcatVector(SDNode *N);
265 SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
267 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
268 /// inline asm expressions.
269 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
271 std::vector<SDValue> &OutOps);
273 // Form pairs of consecutive S, D, or Q registers.
274 SDNode *PairSRegs(EVT VT, SDValue V0, SDValue V1);
275 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
276 SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1);
278 // Form sequences of 4 consecutive S, D, or Q registers.
279 SDNode *QuadSRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
280 SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
281 SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
283 // Get the alignment operand for a NEON VLD or VST instruction.
284 SDValue GetVLDSTAlign(SDValue Align, unsigned NumVecs, bool is64BitVector);
288 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
289 /// operand. If so Imm will receive the 32-bit value.
290 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
291 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
292 Imm = cast<ConstantSDNode>(N)->getZExtValue();
298 // isInt32Immediate - This method tests to see if a constant operand.
299 // If so Imm will receive the 32 bit value.
300 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
301 return isInt32Immediate(N.getNode(), Imm);
304 // isOpcWithIntImmediate - This method tests to see if the node is a specific
305 // opcode and that it has a immediate integer right operand.
306 // If so Imm will receive the 32 bit value.
307 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
308 return N->getOpcode() == Opc &&
309 isInt32Immediate(N->getOperand(1).getNode(), Imm);
312 /// \brief Check whether a particular node is a constant value representable as
313 /// (N * Scale) where (N in [\arg RangeMin, \arg RangeMax).
315 /// \param ScaledConstant [out] - On success, the pre-scaled constant value.
316 static bool isScaledConstantInRange(SDValue Node, int Scale,
317 int RangeMin, int RangeMax,
318 int &ScaledConstant) {
319 assert(Scale > 0 && "Invalid scale!");
321 // Check that this is a constant.
322 const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Node);
326 ScaledConstant = (int) C->getZExtValue();
327 if ((ScaledConstant % Scale) != 0)
330 ScaledConstant /= Scale;
331 return ScaledConstant >= RangeMin && ScaledConstant < RangeMax;
334 /// hasNoVMLxHazardUse - Return true if it's desirable to select a FP MLA / MLS
335 /// node. VFP / NEON fp VMLA / VMLS instructions have special RAW hazards (at
336 /// least on current ARM implementations) which should be avoidded.
337 bool ARMDAGToDAGISel::hasNoVMLxHazardUse(SDNode *N) const {
338 if (OptLevel == CodeGenOpt::None)
341 if (!CheckVMLxHazard)
344 if (!Subtarget->isCortexA8() && !Subtarget->isCortexA9())
350 SDNode *Use = *N->use_begin();
351 if (Use->getOpcode() == ISD::CopyToReg)
353 if (Use->isMachineOpcode()) {
354 const MCInstrDesc &MCID = TII->get(Use->getMachineOpcode());
357 unsigned Opcode = MCID.getOpcode();
358 if (Opcode == ARM::VMOVRS || Opcode == ARM::VMOVRRD)
360 // vmlx feeding into another vmlx. We actually want to unfold
361 // the use later in the MLxExpansion pass. e.g.
363 // vmla (stall 8 cycles)
368 // This adds up to about 18 - 19 cycles.
371 // vmul (stall 4 cycles)
372 // vadd adds up to about 14 cycles.
373 return TII->isFpMLxInstruction(Opcode);
379 bool ARMDAGToDAGISel::isShifterOpProfitable(const SDValue &Shift,
380 ARM_AM::ShiftOpc ShOpcVal,
382 if (!Subtarget->isCortexA9())
384 if (Shift.hasOneUse())
387 return ShOpcVal == ARM_AM::lsl && ShAmt == 2;
390 bool ARMDAGToDAGISel::SelectImmShifterOperand(SDValue N,
393 bool CheckProfitability) {
394 if (DisableShifterOp)
397 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
399 // Don't match base register only case. That is matched to a separate
400 // lower complexity pattern with explicit register operand.
401 if (ShOpcVal == ARM_AM::no_shift) return false;
403 BaseReg = N.getOperand(0);
404 unsigned ShImmVal = 0;
405 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1));
406 if (!RHS) return false;
407 ShImmVal = RHS->getZExtValue() & 31;
408 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
413 bool ARMDAGToDAGISel::SelectRegShifterOperand(SDValue N,
417 bool CheckProfitability) {
418 if (DisableShifterOp)
421 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
423 // Don't match base register only case. That is matched to a separate
424 // lower complexity pattern with explicit register operand.
425 if (ShOpcVal == ARM_AM::no_shift) return false;
427 BaseReg = N.getOperand(0);
428 unsigned ShImmVal = 0;
429 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1));
430 if (RHS) return false;
432 ShReg = N.getOperand(1);
433 if (CheckProfitability && !isShifterOpProfitable(N, ShOpcVal, ShImmVal))
435 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
441 bool ARMDAGToDAGISel::SelectAddrModeImm12(SDValue N,
444 // Match simple R + imm12 operands.
447 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
448 !CurDAG->isBaseWithConstantOffset(N)) {
449 if (N.getOpcode() == ISD::FrameIndex) {
450 // Match frame index.
451 int FI = cast<FrameIndexSDNode>(N)->getIndex();
452 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
453 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
457 if (N.getOpcode() == ARMISD::Wrapper &&
458 !(Subtarget->useMovt() &&
459 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
460 Base = N.getOperand(0);
463 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
467 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
468 int RHSC = (int)RHS->getZExtValue();
469 if (N.getOpcode() == ISD::SUB)
472 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
473 Base = N.getOperand(0);
474 if (Base.getOpcode() == ISD::FrameIndex) {
475 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
476 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
478 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
485 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
491 bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset,
493 if (N.getOpcode() == ISD::MUL &&
494 (!Subtarget->isCortexA9() || N.hasOneUse())) {
495 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
496 // X * [3,5,9] -> X + X * [2,4,8] etc.
497 int RHSC = (int)RHS->getZExtValue();
500 ARM_AM::AddrOpc AddSub = ARM_AM::add;
502 AddSub = ARM_AM::sub;
505 if (isPowerOf2_32(RHSC)) {
506 unsigned ShAmt = Log2_32(RHSC);
507 Base = Offset = N.getOperand(0);
508 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
517 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
518 // ISD::OR that is equivalent to an ISD::ADD.
519 !CurDAG->isBaseWithConstantOffset(N))
522 // Leave simple R +/- imm12 operands for LDRi12
523 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::OR) {
525 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1,
526 -0x1000+1, 0x1000, RHSC)) // 12 bits.
530 // Otherwise this is R +/- [possibly shifted] R.
531 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::SUB ? ARM_AM::sub:ARM_AM::add;
532 ARM_AM::ShiftOpc ShOpcVal =
533 ARM_AM::getShiftOpcForNode(N.getOperand(1).getOpcode());
536 Base = N.getOperand(0);
537 Offset = N.getOperand(1);
539 if (ShOpcVal != ARM_AM::no_shift) {
540 // Check to see if the RHS of the shift is a constant, if not, we can't fold
542 if (ConstantSDNode *Sh =
543 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
544 ShAmt = Sh->getZExtValue();
545 if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt))
546 Offset = N.getOperand(1).getOperand(0);
549 ShOpcVal = ARM_AM::no_shift;
552 ShOpcVal = ARM_AM::no_shift;
556 // Try matching (R shl C) + (R).
557 if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift &&
558 !(Subtarget->isCortexA9() || N.getOperand(0).hasOneUse())) {
559 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0).getOpcode());
560 if (ShOpcVal != ARM_AM::no_shift) {
561 // Check to see if the RHS of the shift is a constant, if not, we can't
563 if (ConstantSDNode *Sh =
564 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
565 ShAmt = Sh->getZExtValue();
566 if (isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt)) {
567 Offset = N.getOperand(0).getOperand(0);
568 Base = N.getOperand(1);
571 ShOpcVal = ARM_AM::no_shift;
574 ShOpcVal = ARM_AM::no_shift;
579 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
589 AddrMode2Type ARMDAGToDAGISel::SelectAddrMode2Worker(SDValue N,
593 if (N.getOpcode() == ISD::MUL &&
594 (!Subtarget->isCortexA9() || N.hasOneUse())) {
595 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
596 // X * [3,5,9] -> X + X * [2,4,8] etc.
597 int RHSC = (int)RHS->getZExtValue();
600 ARM_AM::AddrOpc AddSub = ARM_AM::add;
602 AddSub = ARM_AM::sub;
605 if (isPowerOf2_32(RHSC)) {
606 unsigned ShAmt = Log2_32(RHSC);
607 Base = Offset = N.getOperand(0);
608 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
617 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
618 // ISD::OR that is equivalent to an ADD.
619 !CurDAG->isBaseWithConstantOffset(N)) {
621 if (N.getOpcode() == ISD::FrameIndex) {
622 int FI = cast<FrameIndexSDNode>(N)->getIndex();
623 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
624 } else if (N.getOpcode() == ARMISD::Wrapper &&
625 !(Subtarget->useMovt() &&
626 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
627 Base = N.getOperand(0);
629 Offset = CurDAG->getRegister(0, MVT::i32);
630 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
636 // Match simple R +/- imm12 operands.
637 if (N.getOpcode() != ISD::SUB) {
639 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1,
640 -0x1000+1, 0x1000, RHSC)) { // 12 bits.
641 Base = N.getOperand(0);
642 if (Base.getOpcode() == ISD::FrameIndex) {
643 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
644 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
646 Offset = CurDAG->getRegister(0, MVT::i32);
648 ARM_AM::AddrOpc AddSub = ARM_AM::add;
650 AddSub = ARM_AM::sub;
653 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
660 if (Subtarget->isCortexA9() && !N.hasOneUse()) {
661 // Compute R +/- (R << N) and reuse it.
663 Offset = CurDAG->getRegister(0, MVT::i32);
664 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
670 // Otherwise this is R +/- [possibly shifted] R.
671 ARM_AM::AddrOpc AddSub = N.getOpcode() != ISD::SUB ? ARM_AM::add:ARM_AM::sub;
672 ARM_AM::ShiftOpc ShOpcVal =
673 ARM_AM::getShiftOpcForNode(N.getOperand(1).getOpcode());
676 Base = N.getOperand(0);
677 Offset = N.getOperand(1);
679 if (ShOpcVal != ARM_AM::no_shift) {
680 // Check to see if the RHS of the shift is a constant, if not, we can't fold
682 if (ConstantSDNode *Sh =
683 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
684 ShAmt = Sh->getZExtValue();
685 if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt))
686 Offset = N.getOperand(1).getOperand(0);
689 ShOpcVal = ARM_AM::no_shift;
692 ShOpcVal = ARM_AM::no_shift;
696 // Try matching (R shl C) + (R).
697 if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift &&
698 !(Subtarget->isCortexA9() || N.getOperand(0).hasOneUse())) {
699 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0).getOpcode());
700 if (ShOpcVal != ARM_AM::no_shift) {
701 // Check to see if the RHS of the shift is a constant, if not, we can't
703 if (ConstantSDNode *Sh =
704 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
705 ShAmt = Sh->getZExtValue();
706 if (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 // Look for (R + R) or (R + (R << [1,2,3])).
1290 Base = N.getOperand(0);
1291 OffReg = N.getOperand(1);
1293 // Swap if it is ((R << c) + R).
1294 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg.getOpcode());
1295 if (ShOpcVal != ARM_AM::lsl) {
1296 ShOpcVal = ARM_AM::getShiftOpcForNode(Base.getOpcode());
1297 if (ShOpcVal == ARM_AM::lsl)
1298 std::swap(Base, OffReg);
1301 if (ShOpcVal == ARM_AM::lsl) {
1302 // Check to see if the RHS of the shift is a constant, if not, we can't fold
1304 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
1305 ShAmt = Sh->getZExtValue();
1306 if (ShAmt < 4 && isShifterOpProfitable(OffReg, ShOpcVal, ShAmt))
1307 OffReg = OffReg.getOperand(0);
1310 ShOpcVal = ARM_AM::no_shift;
1313 ShOpcVal = ARM_AM::no_shift;
1317 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
1322 //===--------------------------------------------------------------------===//
1324 /// getAL - Returns a ARMCC::AL immediate node.
1325 static inline SDValue getAL(SelectionDAG *CurDAG) {
1326 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
1329 SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
1330 LoadSDNode *LD = cast<LoadSDNode>(N);
1331 ISD::MemIndexedMode AM = LD->getAddressingMode();
1332 if (AM == ISD::UNINDEXED)
1335 EVT LoadedVT = LD->getMemoryVT();
1336 SDValue Offset, AMOpc;
1337 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
1338 unsigned Opcode = 0;
1340 if (LoadedVT == MVT::i32 && isPre &&
1341 SelectAddrMode2OffsetImmPre(N, LD->getOffset(), Offset, AMOpc)) {
1342 Opcode = ARM::LDR_PRE_IMM;
1344 } else if (LoadedVT == MVT::i32 && !isPre &&
1345 SelectAddrMode2OffsetImm(N, LD->getOffset(), Offset, AMOpc)) {
1346 Opcode = ARM::LDR_POST_IMM;
1348 } else if (LoadedVT == MVT::i32 &&
1349 SelectAddrMode2OffsetReg(N, LD->getOffset(), Offset, AMOpc)) {
1350 Opcode = isPre ? ARM::LDR_PRE_REG : ARM::LDR_POST_REG;
1353 } else if (LoadedVT == MVT::i16 &&
1354 SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
1356 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
1357 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
1358 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
1359 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
1360 if (LD->getExtensionType() == ISD::SEXTLOAD) {
1361 if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
1363 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
1367 SelectAddrMode2OffsetImmPre(N, LD->getOffset(), Offset, AMOpc)) {
1369 Opcode = ARM::LDRB_PRE_IMM;
1370 } else if (!isPre &&
1371 SelectAddrMode2OffsetImm(N, LD->getOffset(), Offset, AMOpc)) {
1373 Opcode = ARM::LDRB_POST_IMM;
1374 } else if (SelectAddrMode2OffsetReg(N, LD->getOffset(), Offset, AMOpc)) {
1376 Opcode = isPre ? ARM::LDRB_PRE_REG : ARM::LDRB_POST_REG;
1382 if (Opcode == ARM::LDR_PRE_IMM || Opcode == ARM::LDRB_PRE_IMM) {
1383 SDValue Chain = LD->getChain();
1384 SDValue Base = LD->getBasePtr();
1385 SDValue Ops[]= { Base, AMOpc, getAL(CurDAG),
1386 CurDAG->getRegister(0, MVT::i32), Chain };
1387 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32,
1388 MVT::i32, MVT::Other, Ops, 5);
1390 SDValue Chain = LD->getChain();
1391 SDValue Base = LD->getBasePtr();
1392 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
1393 CurDAG->getRegister(0, MVT::i32), Chain };
1394 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32,
1395 MVT::i32, MVT::Other, Ops, 6);
1402 SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
1403 LoadSDNode *LD = cast<LoadSDNode>(N);
1404 ISD::MemIndexedMode AM = LD->getAddressingMode();
1405 if (AM == ISD::UNINDEXED)
1408 EVT LoadedVT = LD->getMemoryVT();
1409 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
1411 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
1412 unsigned Opcode = 0;
1414 if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
1415 switch (LoadedVT.getSimpleVT().SimpleTy) {
1417 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
1421 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
1423 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
1428 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
1430 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
1439 SDValue Chain = LD->getChain();
1440 SDValue Base = LD->getBasePtr();
1441 SDValue Ops[]= { Base, Offset, getAL(CurDAG),
1442 CurDAG->getRegister(0, MVT::i32), Chain };
1443 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
1444 MVT::Other, Ops, 5);
1450 /// PairSRegs - Form a D register from a pair of S registers.
1452 SDNode *ARMDAGToDAGISel::PairSRegs(EVT VT, SDValue V0, SDValue V1) {
1453 DebugLoc dl = V0.getNode()->getDebugLoc();
1455 CurDAG->getTargetConstant(ARM::DPR_VFP2RegClassID, MVT::i32);
1456 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
1457 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
1458 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1459 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 5);
1462 /// PairDRegs - Form a quad register from a pair of D registers.
1464 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
1465 DebugLoc dl = V0.getNode()->getDebugLoc();
1466 SDValue RegClass = CurDAG->getTargetConstant(ARM::QPRRegClassID, MVT::i32);
1467 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1468 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1469 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1470 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 5);
1473 /// PairQRegs - Form 4 consecutive D registers from a pair of Q registers.
1475 SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) {
1476 DebugLoc dl = V0.getNode()->getDebugLoc();
1477 SDValue RegClass = CurDAG->getTargetConstant(ARM::QQPRRegClassID, MVT::i32);
1478 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
1479 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
1480 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1481 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 5);
1484 /// QuadSRegs - Form 4 consecutive S registers.
1486 SDNode *ARMDAGToDAGISel::QuadSRegs(EVT VT, SDValue V0, SDValue V1,
1487 SDValue V2, SDValue V3) {
1488 DebugLoc dl = V0.getNode()->getDebugLoc();
1490 CurDAG->getTargetConstant(ARM::QPR_VFP2RegClassID, MVT::i32);
1491 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
1492 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
1493 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, MVT::i32);
1494 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, MVT::i32);
1495 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1,
1496 V2, SubReg2, V3, SubReg3 };
1497 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 9);
1500 /// QuadDRegs - Form 4 consecutive D registers.
1502 SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
1503 SDValue V2, SDValue V3) {
1504 DebugLoc dl = V0.getNode()->getDebugLoc();
1505 SDValue RegClass = CurDAG->getTargetConstant(ARM::QQPRRegClassID, MVT::i32);
1506 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1507 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1508 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
1509 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
1510 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1,
1511 V2, SubReg2, V3, SubReg3 };
1512 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 9);
1515 /// QuadQRegs - Form 4 consecutive Q registers.
1517 SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1,
1518 SDValue V2, SDValue V3) {
1519 DebugLoc dl = V0.getNode()->getDebugLoc();
1520 SDValue RegClass = CurDAG->getTargetConstant(ARM::QQQQPRRegClassID, MVT::i32);
1521 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
1522 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
1523 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, MVT::i32);
1524 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, MVT::i32);
1525 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1,
1526 V2, SubReg2, V3, SubReg3 };
1527 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 9);
1530 /// GetVLDSTAlign - Get the alignment (in bytes) for the alignment operand
1531 /// of a NEON VLD or VST instruction. The supported values depend on the
1532 /// number of registers being loaded.
1533 SDValue ARMDAGToDAGISel::GetVLDSTAlign(SDValue Align, unsigned NumVecs,
1534 bool is64BitVector) {
1535 unsigned NumRegs = NumVecs;
1536 if (!is64BitVector && NumVecs < 3)
1539 unsigned Alignment = cast<ConstantSDNode>(Align)->getZExtValue();
1540 if (Alignment >= 32 && NumRegs == 4)
1542 else if (Alignment >= 16 && (NumRegs == 2 || NumRegs == 4))
1544 else if (Alignment >= 8)
1549 return CurDAG->getTargetConstant(Alignment, MVT::i32);
1552 SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs,
1553 unsigned *DOpcodes, unsigned *QOpcodes0,
1554 unsigned *QOpcodes1) {
1555 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1556 DebugLoc dl = N->getDebugLoc();
1558 SDValue MemAddr, Align;
1559 unsigned AddrOpIdx = isUpdating ? 1 : 2;
1560 if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
1563 SDValue Chain = N->getOperand(0);
1564 EVT VT = N->getValueType(0);
1565 bool is64BitVector = VT.is64BitVector();
1566 Align = GetVLDSTAlign(Align, NumVecs, is64BitVector);
1568 unsigned OpcodeIndex;
1569 switch (VT.getSimpleVT().SimpleTy) {
1570 default: llvm_unreachable("unhandled vld type");
1571 // Double-register operations:
1572 case MVT::v8i8: OpcodeIndex = 0; break;
1573 case MVT::v4i16: OpcodeIndex = 1; break;
1575 case MVT::v2i32: OpcodeIndex = 2; break;
1576 case MVT::v1i64: OpcodeIndex = 3; break;
1577 // Quad-register operations:
1578 case MVT::v16i8: OpcodeIndex = 0; break;
1579 case MVT::v8i16: OpcodeIndex = 1; break;
1581 case MVT::v4i32: OpcodeIndex = 2; break;
1582 case MVT::v2i64: OpcodeIndex = 3;
1583 assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
1591 unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs;
1594 ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, ResTyElts);
1596 std::vector<EVT> ResTys;
1597 ResTys.push_back(ResTy);
1599 ResTys.push_back(MVT::i32);
1600 ResTys.push_back(MVT::Other);
1602 SDValue Pred = getAL(CurDAG);
1603 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1605 SmallVector<SDValue, 7> Ops;
1607 // Double registers and VLD1/VLD2 quad registers are directly supported.
1608 if (is64BitVector || NumVecs <= 2) {
1609 unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
1610 QOpcodes0[OpcodeIndex]);
1611 Ops.push_back(MemAddr);
1612 Ops.push_back(Align);
1614 SDValue Inc = N->getOperand(AddrOpIdx + 1);
1615 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc);
1617 Ops.push_back(Pred);
1618 Ops.push_back(Reg0);
1619 Ops.push_back(Chain);
1620 VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), Ops.size());
1623 // Otherwise, quad registers are loaded with two separate instructions,
1624 // where one loads the even registers and the other loads the odd registers.
1625 EVT AddrTy = MemAddr.getValueType();
1627 // Load the even subregs. This is always an updating load, so that it
1628 // provides the address to the second load for the odd subregs.
1630 SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, ResTy), 0);
1631 const SDValue OpsA[] = { MemAddr, Align, Reg0, ImplDef, Pred, Reg0, Chain };
1632 SDNode *VLdA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl,
1633 ResTy, AddrTy, MVT::Other, OpsA, 7);
1634 Chain = SDValue(VLdA, 2);
1636 // Load the odd subregs.
1637 Ops.push_back(SDValue(VLdA, 1));
1638 Ops.push_back(Align);
1640 SDValue Inc = N->getOperand(AddrOpIdx + 1);
1641 assert(isa<ConstantSDNode>(Inc.getNode()) &&
1642 "only constant post-increment update allowed for VLD3/4");
1644 Ops.push_back(Reg0);
1646 Ops.push_back(SDValue(VLdA, 0));
1647 Ops.push_back(Pred);
1648 Ops.push_back(Reg0);
1649 Ops.push_back(Chain);
1650 VLd = CurDAG->getMachineNode(QOpcodes1[OpcodeIndex], dl, ResTys,
1651 Ops.data(), Ops.size());
1654 // Transfer memoperands.
1655 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1656 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
1657 cast<MachineSDNode>(VLd)->setMemRefs(MemOp, MemOp + 1);
1662 // Extract out the subregisters.
1663 SDValue SuperReg = SDValue(VLd, 0);
1664 assert(ARM::dsub_7 == ARM::dsub_0+7 &&
1665 ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1666 unsigned Sub0 = (is64BitVector ? ARM::dsub_0 : ARM::qsub_0);
1667 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1668 ReplaceUses(SDValue(N, Vec),
1669 CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg));
1670 ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, 1));
1672 ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLd, 2));
1676 SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs,
1677 unsigned *DOpcodes, unsigned *QOpcodes0,
1678 unsigned *QOpcodes1) {
1679 assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range");
1680 DebugLoc dl = N->getDebugLoc();
1682 SDValue MemAddr, Align;
1683 unsigned AddrOpIdx = isUpdating ? 1 : 2;
1684 unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1)
1685 if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
1688 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1689 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
1691 SDValue Chain = N->getOperand(0);
1692 EVT VT = N->getOperand(Vec0Idx).getValueType();
1693 bool is64BitVector = VT.is64BitVector();
1694 Align = GetVLDSTAlign(Align, NumVecs, is64BitVector);
1696 unsigned OpcodeIndex;
1697 switch (VT.getSimpleVT().SimpleTy) {
1698 default: llvm_unreachable("unhandled vst type");
1699 // Double-register operations:
1700 case MVT::v8i8: OpcodeIndex = 0; break;
1701 case MVT::v4i16: OpcodeIndex = 1; break;
1703 case MVT::v2i32: OpcodeIndex = 2; break;
1704 case MVT::v1i64: OpcodeIndex = 3; break;
1705 // Quad-register operations:
1706 case MVT::v16i8: OpcodeIndex = 0; break;
1707 case MVT::v8i16: OpcodeIndex = 1; break;
1709 case MVT::v4i32: OpcodeIndex = 2; break;
1710 case MVT::v2i64: OpcodeIndex = 3;
1711 assert(NumVecs == 1 && "v2i64 type only supported for VST1");
1715 std::vector<EVT> ResTys;
1717 ResTys.push_back(MVT::i32);
1718 ResTys.push_back(MVT::Other);
1720 SDValue Pred = getAL(CurDAG);
1721 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1722 SmallVector<SDValue, 7> Ops;
1724 // Double registers and VST1/VST2 quad registers are directly supported.
1725 if (is64BitVector || NumVecs <= 2) {
1728 SrcReg = N->getOperand(Vec0Idx);
1729 } else if (is64BitVector) {
1730 // Form a REG_SEQUENCE to force register allocation.
1731 SDValue V0 = N->getOperand(Vec0Idx + 0);
1732 SDValue V1 = N->getOperand(Vec0Idx + 1);
1734 SrcReg = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1736 SDValue V2 = N->getOperand(Vec0Idx + 2);
1737 // If it's a vst3, form a quad D-register and leave the last part as
1739 SDValue V3 = (NumVecs == 3)
1740 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1741 : N->getOperand(Vec0Idx + 3);
1742 SrcReg = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1745 // Form a QQ register.
1746 SDValue Q0 = N->getOperand(Vec0Idx);
1747 SDValue Q1 = N->getOperand(Vec0Idx + 1);
1748 SrcReg = SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0);
1751 unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
1752 QOpcodes0[OpcodeIndex]);
1753 Ops.push_back(MemAddr);
1754 Ops.push_back(Align);
1756 SDValue Inc = N->getOperand(AddrOpIdx + 1);
1757 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc);
1759 Ops.push_back(SrcReg);
1760 Ops.push_back(Pred);
1761 Ops.push_back(Reg0);
1762 Ops.push_back(Chain);
1764 CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), Ops.size());
1766 // Transfer memoperands.
1767 cast<MachineSDNode>(VSt)->setMemRefs(MemOp, MemOp + 1);
1772 // Otherwise, quad registers are stored with two separate instructions,
1773 // where one stores the even registers and the other stores the odd registers.
1775 // Form the QQQQ REG_SEQUENCE.
1776 SDValue V0 = N->getOperand(Vec0Idx + 0);
1777 SDValue V1 = N->getOperand(Vec0Idx + 1);
1778 SDValue V2 = N->getOperand(Vec0Idx + 2);
1779 SDValue V3 = (NumVecs == 3)
1780 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
1781 : N->getOperand(Vec0Idx + 3);
1782 SDValue RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
1784 // Store the even D registers. This is always an updating store, so that it
1785 // provides the address to the second store for the odd subregs.
1786 const SDValue OpsA[] = { MemAddr, Align, Reg0, RegSeq, Pred, Reg0, Chain };
1787 SDNode *VStA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl,
1788 MemAddr.getValueType(),
1789 MVT::Other, OpsA, 7);
1790 cast<MachineSDNode>(VStA)->setMemRefs(MemOp, MemOp + 1);
1791 Chain = SDValue(VStA, 1);
1793 // Store the odd D registers.
1794 Ops.push_back(SDValue(VStA, 0));
1795 Ops.push_back(Align);
1797 SDValue Inc = N->getOperand(AddrOpIdx + 1);
1798 assert(isa<ConstantSDNode>(Inc.getNode()) &&
1799 "only constant post-increment update allowed for VST3/4");
1801 Ops.push_back(Reg0);
1803 Ops.push_back(RegSeq);
1804 Ops.push_back(Pred);
1805 Ops.push_back(Reg0);
1806 Ops.push_back(Chain);
1807 SDNode *VStB = CurDAG->getMachineNode(QOpcodes1[OpcodeIndex], dl, ResTys,
1808 Ops.data(), Ops.size());
1809 cast<MachineSDNode>(VStB)->setMemRefs(MemOp, MemOp + 1);
1813 SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1814 bool isUpdating, unsigned NumVecs,
1816 unsigned *QOpcodes) {
1817 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1818 DebugLoc dl = N->getDebugLoc();
1820 SDValue MemAddr, Align;
1821 unsigned AddrOpIdx = isUpdating ? 1 : 2;
1822 unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1)
1823 if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
1826 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1827 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
1829 SDValue Chain = N->getOperand(0);
1831 cast<ConstantSDNode>(N->getOperand(Vec0Idx + NumVecs))->getZExtValue();
1832 EVT VT = N->getOperand(Vec0Idx).getValueType();
1833 bool is64BitVector = VT.is64BitVector();
1835 unsigned Alignment = 0;
1837 Alignment = cast<ConstantSDNode>(Align)->getZExtValue();
1838 unsigned NumBytes = NumVecs * VT.getVectorElementType().getSizeInBits()/8;
1839 if (Alignment > NumBytes)
1840 Alignment = NumBytes;
1841 if (Alignment < 8 && Alignment < NumBytes)
1843 // Alignment must be a power of two; make sure of that.
1844 Alignment = (Alignment & -Alignment);
1848 Align = CurDAG->getTargetConstant(Alignment, MVT::i32);
1850 unsigned OpcodeIndex;
1851 switch (VT.getSimpleVT().SimpleTy) {
1852 default: llvm_unreachable("unhandled vld/vst lane type");
1853 // Double-register operations:
1854 case MVT::v8i8: OpcodeIndex = 0; break;
1855 case MVT::v4i16: OpcodeIndex = 1; break;
1857 case MVT::v2i32: OpcodeIndex = 2; break;
1858 // Quad-register operations:
1859 case MVT::v8i16: OpcodeIndex = 0; break;
1861 case MVT::v4i32: OpcodeIndex = 1; break;
1864 std::vector<EVT> ResTys;
1866 unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs;
1869 ResTys.push_back(EVT::getVectorVT(*CurDAG->getContext(),
1870 MVT::i64, ResTyElts));
1873 ResTys.push_back(MVT::i32);
1874 ResTys.push_back(MVT::Other);
1876 SDValue Pred = getAL(CurDAG);
1877 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1879 SmallVector<SDValue, 8> Ops;
1880 Ops.push_back(MemAddr);
1881 Ops.push_back(Align);
1883 SDValue Inc = N->getOperand(AddrOpIdx + 1);
1884 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc);
1888 SDValue V0 = N->getOperand(Vec0Idx + 0);
1889 SDValue V1 = N->getOperand(Vec0Idx + 1);
1892 SuperReg = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1894 SuperReg = SDValue(PairQRegs(MVT::v4i64, V0, V1), 0);
1896 SDValue V2 = N->getOperand(Vec0Idx + 2);
1897 SDValue V3 = (NumVecs == 3)
1898 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
1899 : N->getOperand(Vec0Idx + 3);
1901 SuperReg = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1903 SuperReg = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
1905 Ops.push_back(SuperReg);
1906 Ops.push_back(getI32Imm(Lane));
1907 Ops.push_back(Pred);
1908 Ops.push_back(Reg0);
1909 Ops.push_back(Chain);
1911 unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
1912 QOpcodes[OpcodeIndex]);
1913 SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys,
1914 Ops.data(), Ops.size());
1915 cast<MachineSDNode>(VLdLn)->setMemRefs(MemOp, MemOp + 1);
1919 // Extract the subregisters.
1920 SuperReg = SDValue(VLdLn, 0);
1921 assert(ARM::dsub_7 == ARM::dsub_0+7 &&
1922 ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1923 unsigned Sub0 = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
1924 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1925 ReplaceUses(SDValue(N, Vec),
1926 CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg));
1927 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, 1));
1929 ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdLn, 2));
1933 SDNode *ARMDAGToDAGISel::SelectVLDDup(SDNode *N, bool isUpdating,
1934 unsigned NumVecs, unsigned *Opcodes) {
1935 assert(NumVecs >=2 && NumVecs <= 4 && "VLDDup NumVecs out-of-range");
1936 DebugLoc dl = N->getDebugLoc();
1938 SDValue MemAddr, Align;
1939 if (!SelectAddrMode6(N, N->getOperand(1), MemAddr, Align))
1942 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1943 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
1945 SDValue Chain = N->getOperand(0);
1946 EVT VT = N->getValueType(0);
1948 unsigned Alignment = 0;
1950 Alignment = cast<ConstantSDNode>(Align)->getZExtValue();
1951 unsigned NumBytes = NumVecs * VT.getVectorElementType().getSizeInBits()/8;
1952 if (Alignment > NumBytes)
1953 Alignment = NumBytes;
1954 if (Alignment < 8 && Alignment < NumBytes)
1956 // Alignment must be a power of two; make sure of that.
1957 Alignment = (Alignment & -Alignment);
1961 Align = CurDAG->getTargetConstant(Alignment, MVT::i32);
1963 unsigned OpcodeIndex;
1964 switch (VT.getSimpleVT().SimpleTy) {
1965 default: llvm_unreachable("unhandled vld-dup type");
1966 case MVT::v8i8: OpcodeIndex = 0; break;
1967 case MVT::v4i16: OpcodeIndex = 1; break;
1969 case MVT::v2i32: OpcodeIndex = 2; break;
1972 SDValue Pred = getAL(CurDAG);
1973 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1975 unsigned Opc = Opcodes[OpcodeIndex];
1976 SmallVector<SDValue, 6> Ops;
1977 Ops.push_back(MemAddr);
1978 Ops.push_back(Align);
1980 SDValue Inc = N->getOperand(2);
1981 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc);
1983 Ops.push_back(Pred);
1984 Ops.push_back(Reg0);
1985 Ops.push_back(Chain);
1987 unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs;
1988 std::vector<EVT> ResTys;
1989 ResTys.push_back(EVT::getVectorVT(*CurDAG->getContext(), MVT::i64,ResTyElts));
1991 ResTys.push_back(MVT::i32);
1992 ResTys.push_back(MVT::Other);
1994 CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), Ops.size());
1995 cast<MachineSDNode>(VLdDup)->setMemRefs(MemOp, MemOp + 1);
1996 SuperReg = SDValue(VLdDup, 0);
1998 // Extract the subregisters.
1999 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
2000 unsigned SubIdx = ARM::dsub_0;
2001 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
2002 ReplaceUses(SDValue(N, Vec),
2003 CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, SuperReg));
2004 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdDup, 1));
2006 ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdDup, 2));
2010 SDNode *ARMDAGToDAGISel::SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs,
2012 assert(NumVecs >= 2 && NumVecs <= 4 && "VTBL NumVecs out-of-range");
2013 DebugLoc dl = N->getDebugLoc();
2014 EVT VT = N->getValueType(0);
2015 unsigned FirstTblReg = IsExt ? 2 : 1;
2017 // Form a REG_SEQUENCE to force register allocation.
2019 SDValue V0 = N->getOperand(FirstTblReg + 0);
2020 SDValue V1 = N->getOperand(FirstTblReg + 1);
2022 RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0);
2024 SDValue V2 = N->getOperand(FirstTblReg + 2);
2025 // If it's a vtbl3, form a quad D-register and leave the last part as
2027 SDValue V3 = (NumVecs == 3)
2028 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
2029 : N->getOperand(FirstTblReg + 3);
2030 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
2033 SmallVector<SDValue, 6> Ops;
2035 Ops.push_back(N->getOperand(1));
2036 Ops.push_back(RegSeq);
2037 Ops.push_back(N->getOperand(FirstTblReg + NumVecs));
2038 Ops.push_back(getAL(CurDAG)); // predicate
2039 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // predicate register
2040 return CurDAG->getMachineNode(Opc, dl, VT, Ops.data(), Ops.size());
2043 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
2045 if (!Subtarget->hasV6T2Ops())
2048 unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
2049 : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
2052 // For unsigned extracts, check for a shift right and mask
2053 unsigned And_imm = 0;
2054 if (N->getOpcode() == ISD::AND) {
2055 if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
2057 // The immediate is a mask of the low bits iff imm & (imm+1) == 0
2058 if (And_imm & (And_imm + 1))
2061 unsigned Srl_imm = 0;
2062 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
2064 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
2066 // Note: The width operand is encoded as width-1.
2067 unsigned Width = CountTrailingOnes_32(And_imm) - 1;
2068 unsigned LSB = Srl_imm;
2069 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2070 SDValue Ops[] = { N->getOperand(0).getOperand(0),
2071 CurDAG->getTargetConstant(LSB, MVT::i32),
2072 CurDAG->getTargetConstant(Width, MVT::i32),
2073 getAL(CurDAG), Reg0 };
2074 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
2080 // Otherwise, we're looking for a shift of a shift
2081 unsigned Shl_imm = 0;
2082 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
2083 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
2084 unsigned Srl_imm = 0;
2085 if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
2086 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
2087 // Note: The width operand is encoded as width-1.
2088 unsigned Width = 32 - Srl_imm - 1;
2089 int LSB = Srl_imm - Shl_imm;
2092 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2093 SDValue Ops[] = { N->getOperand(0).getOperand(0),
2094 CurDAG->getTargetConstant(LSB, MVT::i32),
2095 CurDAG->getTargetConstant(Width, MVT::i32),
2096 getAL(CurDAG), Reg0 };
2097 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
2103 SDNode *ARMDAGToDAGISel::
2104 SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
2105 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
2108 if (SelectT2ShifterOperandReg(TrueVal, CPTmp0, CPTmp1)) {
2109 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
2110 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
2113 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
2114 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
2115 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
2116 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
2118 llvm_unreachable("Unknown so_reg opcode!");
2122 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
2123 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
2124 SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
2125 return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
2130 SDNode *ARMDAGToDAGISel::
2131 SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
2132 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
2136 if (SelectImmShifterOperand(TrueVal, CPTmp0, CPTmp2)) {
2137 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
2138 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp2, CC, CCR, InFlag };
2139 return CurDAG->SelectNodeTo(N, ARM::MOVCCsi, MVT::i32, Ops, 6);
2142 if (SelectRegShifterOperand(TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
2143 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
2144 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
2145 return CurDAG->SelectNodeTo(N, ARM::MOVCCsr, MVT::i32, Ops, 7);
2150 SDNode *ARMDAGToDAGISel::
2151 SelectT2CMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
2152 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
2153 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
2158 unsigned TrueImm = T->getZExtValue();
2159 if (is_t2_so_imm(TrueImm)) {
2160 Opc = ARM::t2MOVCCi;
2161 } else if (TrueImm <= 0xffff) {
2162 Opc = ARM::t2MOVCCi16;
2163 } else if (is_t2_so_imm_not(TrueImm)) {
2165 Opc = ARM::t2MVNCCi;
2166 } else if (TrueVal.getNode()->hasOneUse() && Subtarget->hasV6T2Ops()) {
2168 Opc = ARM::t2MOVCCi32imm;
2172 SDValue True = CurDAG->getTargetConstant(TrueImm, MVT::i32);
2173 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
2174 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
2175 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
2181 SDNode *ARMDAGToDAGISel::
2182 SelectARMCMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
2183 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
2184 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
2189 unsigned TrueImm = T->getZExtValue();
2190 bool isSoImm = is_so_imm(TrueImm);
2193 } else if (Subtarget->hasV6T2Ops() && TrueImm <= 0xffff) {
2194 Opc = ARM::MOVCCi16;
2195 } else if (is_so_imm_not(TrueImm)) {
2198 } else if (TrueVal.getNode()->hasOneUse() &&
2199 (Subtarget->hasV6T2Ops() || ARM_AM::isSOImmTwoPartVal(TrueImm))) {
2201 Opc = ARM::MOVCCi32imm;
2205 SDValue True = CurDAG->getTargetConstant(TrueImm, MVT::i32);
2206 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
2207 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
2208 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
2214 SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
2215 EVT VT = N->getValueType(0);
2216 SDValue FalseVal = N->getOperand(0);
2217 SDValue TrueVal = N->getOperand(1);
2218 SDValue CC = N->getOperand(2);
2219 SDValue CCR = N->getOperand(3);
2220 SDValue InFlag = N->getOperand(4);
2221 assert(CC.getOpcode() == ISD::Constant);
2222 assert(CCR.getOpcode() == ISD::Register);
2223 ARMCC::CondCodes CCVal =
2224 (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
2226 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
2227 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
2228 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
2229 // Pattern complexity = 18 cost = 1 size = 0
2233 if (Subtarget->isThumb()) {
2234 SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
2235 CCVal, CCR, InFlag);
2237 Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
2238 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
2242 SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
2243 CCVal, CCR, InFlag);
2245 Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
2246 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
2251 // Pattern: (ARMcmov:i32 GPR:i32:$false,
2252 // (imm:i32)<<P:Pred_so_imm>>:$true,
2254 // Emits: (MOVCCi:i32 GPR:i32:$false,
2255 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
2256 // Pattern complexity = 10 cost = 1 size = 0
2257 if (Subtarget->isThumb()) {
2258 SDNode *Res = SelectT2CMOVImmOp(N, FalseVal, TrueVal,
2259 CCVal, CCR, InFlag);
2261 Res = SelectT2CMOVImmOp(N, TrueVal, FalseVal,
2262 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
2266 SDNode *Res = SelectARMCMOVImmOp(N, FalseVal, TrueVal,
2267 CCVal, CCR, InFlag);
2269 Res = SelectARMCMOVImmOp(N, TrueVal, FalseVal,
2270 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
2276 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
2277 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
2278 // Pattern complexity = 6 cost = 1 size = 0
2280 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
2281 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
2282 // Pattern complexity = 6 cost = 11 size = 0
2284 // Also VMOVScc and VMOVDcc.
2285 SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
2286 SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
2288 switch (VT.getSimpleVT().SimpleTy) {
2289 default: assert(false && "Illegal conditional move type!");
2292 Opc = Subtarget->isThumb()
2293 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
2303 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
2306 /// Target-specific DAG combining for ISD::XOR.
2307 /// Target-independent combining lowers SELECT_CC nodes of the form
2308 /// select_cc setg[ge] X, 0, X, -X
2309 /// select_cc setgt X, -1, X, -X
2310 /// select_cc setl[te] X, 0, -X, X
2311 /// select_cc setlt X, 1, -X, X
2312 /// which represent Integer ABS into:
2313 /// Y = sra (X, size(X)-1); xor (add (X, Y), Y)
2314 /// ARM instruction selection detects the latter and matches it to
2315 /// ARM::ABS or ARM::t2ABS machine node.
2316 SDNode *ARMDAGToDAGISel::SelectABSOp(SDNode *N){
2317 SDValue XORSrc0 = N->getOperand(0);
2318 SDValue XORSrc1 = N->getOperand(1);
2319 DebugLoc DL = N->getDebugLoc();
2320 EVT VT = N->getValueType(0);
2322 if (DisableARMIntABS)
2325 if (Subtarget->isThumb1Only())
2328 if (XORSrc0.getOpcode() != ISD::ADD ||
2329 XORSrc1.getOpcode() != ISD::SRA)
2332 SDValue ADDSrc0 = XORSrc0.getOperand(0);
2333 SDValue ADDSrc1 = XORSrc0.getOperand(1);
2334 SDValue SRASrc0 = XORSrc1.getOperand(0);
2335 SDValue SRASrc1 = XORSrc1.getOperand(1);
2336 ConstantSDNode *SRAConstant = dyn_cast<ConstantSDNode>(SRASrc1);
2337 EVT XType = SRASrc0.getValueType();
2338 unsigned Size = XType.getSizeInBits() - 1;
2340 if (ADDSrc1 == XORSrc1 &&
2341 ADDSrc0 == SRASrc0 &&
2342 XType.isInteger() &&
2343 SRAConstant != NULL &&
2344 Size == SRAConstant->getZExtValue()) {
2346 unsigned Opcode = ARM::ABS;
2347 if (Subtarget->isThumb2())
2348 Opcode = ARM::t2ABS;
2350 return CurDAG->SelectNodeTo(N, Opcode, VT, ADDSrc0);
2356 SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
2357 // The only time a CONCAT_VECTORS operation can have legal types is when
2358 // two 64-bit vectors are concatenated to a 128-bit vector.
2359 EVT VT = N->getValueType(0);
2360 if (!VT.is128BitVector() || N->getNumOperands() != 2)
2361 llvm_unreachable("unexpected CONCAT_VECTORS");
2362 return PairDRegs(VT, N->getOperand(0), N->getOperand(1));
2365 SDNode *ARMDAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
2366 SmallVector<SDValue, 6> Ops;
2367 Ops.push_back(Node->getOperand(1)); // Ptr
2368 Ops.push_back(Node->getOperand(2)); // Low part of Val1
2369 Ops.push_back(Node->getOperand(3)); // High part of Val1
2370 if (Opc == ARM::ATOMCMPXCHG6432) {
2371 Ops.push_back(Node->getOperand(4)); // Low part of Val2
2372 Ops.push_back(Node->getOperand(5)); // High part of Val2
2374 Ops.push_back(Node->getOperand(0)); // Chain
2375 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2376 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
2377 SDNode *ResNode = CurDAG->getMachineNode(Opc, Node->getDebugLoc(),
2378 MVT::i32, MVT::i32, MVT::Other,
2379 Ops.data() ,Ops.size());
2380 cast<MachineSDNode>(ResNode)->setMemRefs(MemOp, MemOp + 1);
2384 SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
2385 DebugLoc dl = N->getDebugLoc();
2387 if (N->isMachineOpcode())
2388 return NULL; // Already selected.
2390 switch (N->getOpcode()) {
2393 // Select special operations if XOR node forms integer ABS pattern
2394 SDNode *ResNode = SelectABSOp(N);
2397 // Other cases are autogenerated.
2400 case ISD::Constant: {
2401 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
2403 if (Subtarget->hasThumb2())
2404 // Thumb2-aware targets have the MOVT instruction, so all immediates can
2405 // be done with MOV + MOVT, at worst.
2408 if (Subtarget->isThumb()) {
2409 UseCP = (Val > 255 && // MOV
2410 ~Val > 255 && // MOV + MVN
2411 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
2413 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
2414 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
2415 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
2420 CurDAG->getTargetConstantPool(ConstantInt::get(
2421 Type::getInt32Ty(*CurDAG->getContext()), Val),
2422 TLI.getPointerTy());
2425 if (Subtarget->isThumb1Only()) {
2426 SDValue Pred = getAL(CurDAG);
2427 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2428 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
2429 ResNode = CurDAG->getMachineNode(ARM::tLDRpci, dl, MVT::i32, MVT::Other,
2434 CurDAG->getTargetConstant(0, MVT::i32),
2436 CurDAG->getRegister(0, MVT::i32),
2437 CurDAG->getEntryNode()
2439 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
2442 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
2446 // Other cases are autogenerated.
2449 case ISD::FrameIndex: {
2450 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
2451 int FI = cast<FrameIndexSDNode>(N)->getIndex();
2452 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
2453 if (Subtarget->isThumb1Only()) {
2454 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
2455 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2456 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, Ops, 4);
2458 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
2459 ARM::t2ADDri : ARM::ADDri);
2460 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
2461 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2462 CurDAG->getRegister(0, MVT::i32) };
2463 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
2467 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
2471 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true))
2475 if (Subtarget->isThumb1Only())
2477 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
2478 unsigned RHSV = C->getZExtValue();
2480 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
2481 unsigned ShImm = Log2_32(RHSV-1);
2484 SDValue V = N->getOperand(0);
2485 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
2486 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
2487 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2488 if (Subtarget->isThumb()) {
2489 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
2490 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
2492 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
2493 return CurDAG->SelectNodeTo(N, ARM::ADDrsi, MVT::i32, Ops, 7);
2496 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
2497 unsigned ShImm = Log2_32(RHSV+1);
2500 SDValue V = N->getOperand(0);
2501 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
2502 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
2503 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2504 if (Subtarget->isThumb()) {
2505 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
2506 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 6);
2508 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
2509 return CurDAG->SelectNodeTo(N, ARM::RSBrsi, MVT::i32, Ops, 7);
2515 // Check for unsigned bitfield extract
2516 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
2519 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
2520 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
2521 // are entirely contributed by c2 and lower 16-bits are entirely contributed
2522 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
2523 // Select it to: "movt x, ((c1 & 0xffff) >> 16)
2524 EVT VT = N->getValueType(0);
2527 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
2529 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
2532 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2533 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2536 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
2537 SDValue N2 = N0.getOperand(1);
2538 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
2541 unsigned N1CVal = N1C->getZExtValue();
2542 unsigned N2CVal = N2C->getZExtValue();
2543 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
2544 (N1CVal & 0xffffU) == 0xffffU &&
2545 (N2CVal & 0xffffU) == 0x0U) {
2546 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
2548 SDValue Ops[] = { N0.getOperand(0), Imm16,
2549 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2550 return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
2555 case ARMISD::VMOVRRD:
2556 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
2557 N->getOperand(0), getAL(CurDAG),
2558 CurDAG->getRegister(0, MVT::i32));
2559 case ISD::UMUL_LOHI: {
2560 if (Subtarget->isThumb1Only())
2562 if (Subtarget->isThumb()) {
2563 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2564 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2565 CurDAG->getRegister(0, MVT::i32) };
2566 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32,Ops,4);
2568 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2569 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2570 CurDAG->getRegister(0, MVT::i32) };
2571 return CurDAG->getMachineNode(Subtarget->hasV6Ops() ?
2572 ARM::UMULL : ARM::UMULLv5,
2573 dl, MVT::i32, MVT::i32, Ops, 5);
2576 case ISD::SMUL_LOHI: {
2577 if (Subtarget->isThumb1Only())
2579 if (Subtarget->isThumb()) {
2580 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2581 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2582 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32,Ops,4);
2584 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2585 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2586 CurDAG->getRegister(0, MVT::i32) };
2587 return CurDAG->getMachineNode(Subtarget->hasV6Ops() ?
2588 ARM::SMULL : ARM::SMULLv5,
2589 dl, MVT::i32, MVT::i32, Ops, 5);
2593 SDNode *ResNode = 0;
2594 if (Subtarget->isThumb() && Subtarget->hasThumb2())
2595 ResNode = SelectT2IndexedLoad(N);
2597 ResNode = SelectARMIndexedLoad(N);
2600 // Other cases are autogenerated.
2603 case ARMISD::BRCOND: {
2604 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2605 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2606 // Pattern complexity = 6 cost = 1 size = 0
2608 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2609 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
2610 // Pattern complexity = 6 cost = 1 size = 0
2612 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2613 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2614 // Pattern complexity = 6 cost = 1 size = 0
2616 unsigned Opc = Subtarget->isThumb() ?
2617 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
2618 SDValue Chain = N->getOperand(0);
2619 SDValue N1 = N->getOperand(1);
2620 SDValue N2 = N->getOperand(2);
2621 SDValue N3 = N->getOperand(3);
2622 SDValue InFlag = N->getOperand(4);
2623 assert(N1.getOpcode() == ISD::BasicBlock);
2624 assert(N2.getOpcode() == ISD::Constant);
2625 assert(N3.getOpcode() == ISD::Register);
2627 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2628 cast<ConstantSDNode>(N2)->getZExtValue()),
2630 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
2631 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
2633 Chain = SDValue(ResNode, 0);
2634 if (N->getNumValues() == 2) {
2635 InFlag = SDValue(ResNode, 1);
2636 ReplaceUses(SDValue(N, 1), InFlag);
2638 ReplaceUses(SDValue(N, 0),
2639 SDValue(Chain.getNode(), Chain.getResNo()));
2643 return SelectCMOVOp(N);
2644 case ARMISD::VZIP: {
2646 EVT VT = N->getValueType(0);
2647 switch (VT.getSimpleVT().SimpleTy) {
2648 default: return NULL;
2649 case MVT::v8i8: Opc = ARM::VZIPd8; break;
2650 case MVT::v4i16: Opc = ARM::VZIPd16; break;
2652 case MVT::v2i32: Opc = ARM::VZIPd32; break;
2653 case MVT::v16i8: Opc = ARM::VZIPq8; break;
2654 case MVT::v8i16: Opc = ARM::VZIPq16; break;
2656 case MVT::v4i32: Opc = ARM::VZIPq32; break;
2658 SDValue Pred = getAL(CurDAG);
2659 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2660 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2661 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2663 case ARMISD::VUZP: {
2665 EVT VT = N->getValueType(0);
2666 switch (VT.getSimpleVT().SimpleTy) {
2667 default: return NULL;
2668 case MVT::v8i8: Opc = ARM::VUZPd8; break;
2669 case MVT::v4i16: Opc = ARM::VUZPd16; break;
2671 case MVT::v2i32: Opc = ARM::VUZPd32; break;
2672 case MVT::v16i8: Opc = ARM::VUZPq8; break;
2673 case MVT::v8i16: Opc = ARM::VUZPq16; break;
2675 case MVT::v4i32: Opc = ARM::VUZPq32; break;
2677 SDValue Pred = getAL(CurDAG);
2678 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2679 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2680 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2682 case ARMISD::VTRN: {
2684 EVT VT = N->getValueType(0);
2685 switch (VT.getSimpleVT().SimpleTy) {
2686 default: return NULL;
2687 case MVT::v8i8: Opc = ARM::VTRNd8; break;
2688 case MVT::v4i16: Opc = ARM::VTRNd16; break;
2690 case MVT::v2i32: Opc = ARM::VTRNd32; break;
2691 case MVT::v16i8: Opc = ARM::VTRNq8; break;
2692 case MVT::v8i16: Opc = ARM::VTRNq16; break;
2694 case MVT::v4i32: Opc = ARM::VTRNq32; break;
2696 SDValue Pred = getAL(CurDAG);
2697 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2698 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2699 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2701 case ARMISD::BUILD_VECTOR: {
2702 EVT VecVT = N->getValueType(0);
2703 EVT EltVT = VecVT.getVectorElementType();
2704 unsigned NumElts = VecVT.getVectorNumElements();
2705 if (EltVT == MVT::f64) {
2706 assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
2707 return PairDRegs(VecVT, N->getOperand(0), N->getOperand(1));
2709 assert(EltVT == MVT::f32 && "unexpected type for BUILD_VECTOR");
2711 return PairSRegs(VecVT, N->getOperand(0), N->getOperand(1));
2712 assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
2713 return QuadSRegs(VecVT, N->getOperand(0), N->getOperand(1),
2714 N->getOperand(2), N->getOperand(3));
2717 case ARMISD::VLD2DUP: {
2718 unsigned Opcodes[] = { ARM::VLD2DUPd8Pseudo, ARM::VLD2DUPd16Pseudo,
2719 ARM::VLD2DUPd32Pseudo };
2720 return SelectVLDDup(N, false, 2, Opcodes);
2723 case ARMISD::VLD3DUP: {
2724 unsigned Opcodes[] = { ARM::VLD3DUPd8Pseudo, ARM::VLD3DUPd16Pseudo,
2725 ARM::VLD3DUPd32Pseudo };
2726 return SelectVLDDup(N, false, 3, Opcodes);
2729 case ARMISD::VLD4DUP: {
2730 unsigned Opcodes[] = { ARM::VLD4DUPd8Pseudo, ARM::VLD4DUPd16Pseudo,
2731 ARM::VLD4DUPd32Pseudo };
2732 return SelectVLDDup(N, false, 4, Opcodes);
2735 case ARMISD::VLD2DUP_UPD: {
2736 unsigned Opcodes[] = { ARM::VLD2DUPd8Pseudo_UPD, ARM::VLD2DUPd16Pseudo_UPD,
2737 ARM::VLD2DUPd32Pseudo_UPD };
2738 return SelectVLDDup(N, true, 2, Opcodes);
2741 case ARMISD::VLD3DUP_UPD: {
2742 unsigned Opcodes[] = { ARM::VLD3DUPd8Pseudo_UPD, ARM::VLD3DUPd16Pseudo_UPD,
2743 ARM::VLD3DUPd32Pseudo_UPD };
2744 return SelectVLDDup(N, true, 3, Opcodes);
2747 case ARMISD::VLD4DUP_UPD: {
2748 unsigned Opcodes[] = { ARM::VLD4DUPd8Pseudo_UPD, ARM::VLD4DUPd16Pseudo_UPD,
2749 ARM::VLD4DUPd32Pseudo_UPD };
2750 return SelectVLDDup(N, true, 4, Opcodes);
2753 case ARMISD::VLD1_UPD: {
2754 unsigned DOpcodes[] = { ARM::VLD1d8_UPD, ARM::VLD1d16_UPD,
2755 ARM::VLD1d32_UPD, ARM::VLD1d64_UPD };
2756 unsigned QOpcodes[] = { ARM::VLD1q8Pseudo_UPD, ARM::VLD1q16Pseudo_UPD,
2757 ARM::VLD1q32Pseudo_UPD, ARM::VLD1q64Pseudo_UPD };
2758 return SelectVLD(N, true, 1, DOpcodes, QOpcodes, 0);
2761 case ARMISD::VLD2_UPD: {
2762 unsigned DOpcodes[] = { ARM::VLD2d8Pseudo_UPD, ARM::VLD2d16Pseudo_UPD,
2763 ARM::VLD2d32Pseudo_UPD, ARM::VLD1q64Pseudo_UPD };
2764 unsigned QOpcodes[] = { ARM::VLD2q8Pseudo_UPD, ARM::VLD2q16Pseudo_UPD,
2765 ARM::VLD2q32Pseudo_UPD };
2766 return SelectVLD(N, true, 2, DOpcodes, QOpcodes, 0);
2769 case ARMISD::VLD3_UPD: {
2770 unsigned DOpcodes[] = { ARM::VLD3d8Pseudo_UPD, ARM::VLD3d16Pseudo_UPD,
2771 ARM::VLD3d32Pseudo_UPD, ARM::VLD1d64TPseudo_UPD };
2772 unsigned QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD,
2773 ARM::VLD3q16Pseudo_UPD,
2774 ARM::VLD3q32Pseudo_UPD };
2775 unsigned QOpcodes1[] = { ARM::VLD3q8oddPseudo_UPD,
2776 ARM::VLD3q16oddPseudo_UPD,
2777 ARM::VLD3q32oddPseudo_UPD };
2778 return SelectVLD(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
2781 case ARMISD::VLD4_UPD: {
2782 unsigned DOpcodes[] = { ARM::VLD4d8Pseudo_UPD, ARM::VLD4d16Pseudo_UPD,
2783 ARM::VLD4d32Pseudo_UPD, ARM::VLD1d64QPseudo_UPD };
2784 unsigned QOpcodes0[] = { ARM::VLD4q8Pseudo_UPD,
2785 ARM::VLD4q16Pseudo_UPD,
2786 ARM::VLD4q32Pseudo_UPD };
2787 unsigned QOpcodes1[] = { ARM::VLD4q8oddPseudo_UPD,
2788 ARM::VLD4q16oddPseudo_UPD,
2789 ARM::VLD4q32oddPseudo_UPD };
2790 return SelectVLD(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
2793 case ARMISD::VLD2LN_UPD: {
2794 unsigned DOpcodes[] = { ARM::VLD2LNd8Pseudo_UPD, ARM::VLD2LNd16Pseudo_UPD,
2795 ARM::VLD2LNd32Pseudo_UPD };
2796 unsigned QOpcodes[] = { ARM::VLD2LNq16Pseudo_UPD,
2797 ARM::VLD2LNq32Pseudo_UPD };
2798 return SelectVLDSTLane(N, true, true, 2, DOpcodes, QOpcodes);
2801 case ARMISD::VLD3LN_UPD: {
2802 unsigned DOpcodes[] = { ARM::VLD3LNd8Pseudo_UPD, ARM::VLD3LNd16Pseudo_UPD,
2803 ARM::VLD3LNd32Pseudo_UPD };
2804 unsigned QOpcodes[] = { ARM::VLD3LNq16Pseudo_UPD,
2805 ARM::VLD3LNq32Pseudo_UPD };
2806 return SelectVLDSTLane(N, true, true, 3, DOpcodes, QOpcodes);
2809 case ARMISD::VLD4LN_UPD: {
2810 unsigned DOpcodes[] = { ARM::VLD4LNd8Pseudo_UPD, ARM::VLD4LNd16Pseudo_UPD,
2811 ARM::VLD4LNd32Pseudo_UPD };
2812 unsigned QOpcodes[] = { ARM::VLD4LNq16Pseudo_UPD,
2813 ARM::VLD4LNq32Pseudo_UPD };
2814 return SelectVLDSTLane(N, true, true, 4, DOpcodes, QOpcodes);
2817 case ARMISD::VST1_UPD: {
2818 unsigned DOpcodes[] = { ARM::VST1d8_UPD, ARM::VST1d16_UPD,
2819 ARM::VST1d32_UPD, ARM::VST1d64_UPD };
2820 unsigned QOpcodes[] = { ARM::VST1q8Pseudo_UPD, ARM::VST1q16Pseudo_UPD,
2821 ARM::VST1q32Pseudo_UPD, ARM::VST1q64Pseudo_UPD };
2822 return SelectVST(N, true, 1, DOpcodes, QOpcodes, 0);
2825 case ARMISD::VST2_UPD: {
2826 unsigned DOpcodes[] = { ARM::VST2d8Pseudo_UPD, ARM::VST2d16Pseudo_UPD,
2827 ARM::VST2d32Pseudo_UPD, ARM::VST1q64Pseudo_UPD };
2828 unsigned QOpcodes[] = { ARM::VST2q8Pseudo_UPD, ARM::VST2q16Pseudo_UPD,
2829 ARM::VST2q32Pseudo_UPD };
2830 return SelectVST(N, true, 2, DOpcodes, QOpcodes, 0);
2833 case ARMISD::VST3_UPD: {
2834 unsigned DOpcodes[] = { ARM::VST3d8Pseudo_UPD, ARM::VST3d16Pseudo_UPD,
2835 ARM::VST3d32Pseudo_UPD, ARM::VST1d64TPseudo_UPD };
2836 unsigned QOpcodes0[] = { ARM::VST3q8Pseudo_UPD,
2837 ARM::VST3q16Pseudo_UPD,
2838 ARM::VST3q32Pseudo_UPD };
2839 unsigned QOpcodes1[] = { ARM::VST3q8oddPseudo_UPD,
2840 ARM::VST3q16oddPseudo_UPD,
2841 ARM::VST3q32oddPseudo_UPD };
2842 return SelectVST(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
2845 case ARMISD::VST4_UPD: {
2846 unsigned DOpcodes[] = { ARM::VST4d8Pseudo_UPD, ARM::VST4d16Pseudo_UPD,
2847 ARM::VST4d32Pseudo_UPD, ARM::VST1d64QPseudo_UPD };
2848 unsigned QOpcodes0[] = { ARM::VST4q8Pseudo_UPD,
2849 ARM::VST4q16Pseudo_UPD,
2850 ARM::VST4q32Pseudo_UPD };
2851 unsigned QOpcodes1[] = { ARM::VST4q8oddPseudo_UPD,
2852 ARM::VST4q16oddPseudo_UPD,
2853 ARM::VST4q32oddPseudo_UPD };
2854 return SelectVST(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
2857 case ARMISD::VST2LN_UPD: {
2858 unsigned DOpcodes[] = { ARM::VST2LNd8Pseudo_UPD, ARM::VST2LNd16Pseudo_UPD,
2859 ARM::VST2LNd32Pseudo_UPD };
2860 unsigned QOpcodes[] = { ARM::VST2LNq16Pseudo_UPD,
2861 ARM::VST2LNq32Pseudo_UPD };
2862 return SelectVLDSTLane(N, false, true, 2, DOpcodes, QOpcodes);
2865 case ARMISD::VST3LN_UPD: {
2866 unsigned DOpcodes[] = { ARM::VST3LNd8Pseudo_UPD, ARM::VST3LNd16Pseudo_UPD,
2867 ARM::VST3LNd32Pseudo_UPD };
2868 unsigned QOpcodes[] = { ARM::VST3LNq16Pseudo_UPD,
2869 ARM::VST3LNq32Pseudo_UPD };
2870 return SelectVLDSTLane(N, false, true, 3, DOpcodes, QOpcodes);
2873 case ARMISD::VST4LN_UPD: {
2874 unsigned DOpcodes[] = { ARM::VST4LNd8Pseudo_UPD, ARM::VST4LNd16Pseudo_UPD,
2875 ARM::VST4LNd32Pseudo_UPD };
2876 unsigned QOpcodes[] = { ARM::VST4LNq16Pseudo_UPD,
2877 ARM::VST4LNq32Pseudo_UPD };
2878 return SelectVLDSTLane(N, false, true, 4, DOpcodes, QOpcodes);
2881 case ISD::INTRINSIC_VOID:
2882 case ISD::INTRINSIC_W_CHAIN: {
2883 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2888 case Intrinsic::arm_ldrexd: {
2889 SDValue MemAddr = N->getOperand(2);
2890 DebugLoc dl = N->getDebugLoc();
2891 SDValue Chain = N->getOperand(0);
2893 unsigned NewOpc = ARM::LDREXD;
2894 if (Subtarget->isThumb() && Subtarget->hasThumb2())
2895 NewOpc = ARM::t2LDREXD;
2897 // arm_ldrexd returns a i64 value in {i32, i32}
2898 std::vector<EVT> ResTys;
2899 ResTys.push_back(MVT::i32);
2900 ResTys.push_back(MVT::i32);
2901 ResTys.push_back(MVT::Other);
2903 // place arguments in the right order
2904 SmallVector<SDValue, 7> Ops;
2905 Ops.push_back(MemAddr);
2906 Ops.push_back(getAL(CurDAG));
2907 Ops.push_back(CurDAG->getRegister(0, MVT::i32));
2908 Ops.push_back(Chain);
2909 SDNode *Ld = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops.data(),
2911 // Transfer memoperands.
2912 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2913 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2914 cast<MachineSDNode>(Ld)->setMemRefs(MemOp, MemOp + 1);
2916 // Until there's support for specifing explicit register constraints
2917 // like the use of even/odd register pair, hardcode ldrexd to always
2918 // use the pair [R0, R1] to hold the load result.
2919 Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ARM::R0,
2920 SDValue(Ld, 0), SDValue(0,0));
2921 Chain = CurDAG->getCopyToReg(Chain, dl, ARM::R1,
2922 SDValue(Ld, 1), Chain.getValue(1));
2925 SDValue Glue = Chain.getValue(1);
2926 if (!SDValue(N, 0).use_empty()) {
2927 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2928 ARM::R0, MVT::i32, Glue);
2929 Glue = Result.getValue(2);
2930 ReplaceUses(SDValue(N, 0), Result);
2932 if (!SDValue(N, 1).use_empty()) {
2933 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2934 ARM::R1, MVT::i32, Glue);
2935 Glue = Result.getValue(2);
2936 ReplaceUses(SDValue(N, 1), Result);
2939 ReplaceUses(SDValue(N, 2), SDValue(Ld, 2));
2943 case Intrinsic::arm_strexd: {
2944 DebugLoc dl = N->getDebugLoc();
2945 SDValue Chain = N->getOperand(0);
2946 SDValue Val0 = N->getOperand(2);
2947 SDValue Val1 = N->getOperand(3);
2948 SDValue MemAddr = N->getOperand(4);
2950 // Until there's support for specifing explicit register constraints
2951 // like the use of even/odd register pair, hardcode strexd to always
2952 // use the pair [R2, R3] to hold the i64 (i32, i32) value to be stored.
2953 Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ARM::R2, Val0,
2955 Chain = CurDAG->getCopyToReg(Chain, dl, ARM::R3, Val1, Chain.getValue(1));
2957 SDValue Glue = Chain.getValue(1);
2958 Val0 = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2959 ARM::R2, MVT::i32, Glue);
2960 Glue = Val0.getValue(1);
2961 Val1 = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2962 ARM::R3, MVT::i32, Glue);
2964 // Store exclusive double return a i32 value which is the return status
2965 // of the issued store.
2966 std::vector<EVT> ResTys;
2967 ResTys.push_back(MVT::i32);
2968 ResTys.push_back(MVT::Other);
2970 // place arguments in the right order
2971 SmallVector<SDValue, 7> Ops;
2972 Ops.push_back(Val0);
2973 Ops.push_back(Val1);
2974 Ops.push_back(MemAddr);
2975 Ops.push_back(getAL(CurDAG));
2976 Ops.push_back(CurDAG->getRegister(0, MVT::i32));
2977 Ops.push_back(Chain);
2979 unsigned NewOpc = ARM::STREXD;
2980 if (Subtarget->isThumb() && Subtarget->hasThumb2())
2981 NewOpc = ARM::t2STREXD;
2983 SDNode *St = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops.data(),
2985 // Transfer memoperands.
2986 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2987 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2988 cast<MachineSDNode>(St)->setMemRefs(MemOp, MemOp + 1);
2993 case Intrinsic::arm_neon_vld1: {
2994 unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
2995 ARM::VLD1d32, ARM::VLD1d64 };
2996 unsigned QOpcodes[] = { ARM::VLD1q8Pseudo, ARM::VLD1q16Pseudo,
2997 ARM::VLD1q32Pseudo, ARM::VLD1q64Pseudo };
2998 return SelectVLD(N, false, 1, DOpcodes, QOpcodes, 0);
3001 case Intrinsic::arm_neon_vld2: {
3002 unsigned DOpcodes[] = { ARM::VLD2d8Pseudo, ARM::VLD2d16Pseudo,
3003 ARM::VLD2d32Pseudo, ARM::VLD1q64Pseudo };
3004 unsigned QOpcodes[] = { ARM::VLD2q8Pseudo, ARM::VLD2q16Pseudo,
3005 ARM::VLD2q32Pseudo };
3006 return SelectVLD(N, false, 2, DOpcodes, QOpcodes, 0);
3009 case Intrinsic::arm_neon_vld3: {
3010 unsigned DOpcodes[] = { ARM::VLD3d8Pseudo, ARM::VLD3d16Pseudo,
3011 ARM::VLD3d32Pseudo, ARM::VLD1d64TPseudo };
3012 unsigned QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD,
3013 ARM::VLD3q16Pseudo_UPD,
3014 ARM::VLD3q32Pseudo_UPD };
3015 unsigned QOpcodes1[] = { ARM::VLD3q8oddPseudo,
3016 ARM::VLD3q16oddPseudo,
3017 ARM::VLD3q32oddPseudo };
3018 return SelectVLD(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
3021 case Intrinsic::arm_neon_vld4: {
3022 unsigned DOpcodes[] = { ARM::VLD4d8Pseudo, ARM::VLD4d16Pseudo,
3023 ARM::VLD4d32Pseudo, ARM::VLD1d64QPseudo };
3024 unsigned QOpcodes0[] = { ARM::VLD4q8Pseudo_UPD,
3025 ARM::VLD4q16Pseudo_UPD,
3026 ARM::VLD4q32Pseudo_UPD };
3027 unsigned QOpcodes1[] = { ARM::VLD4q8oddPseudo,
3028 ARM::VLD4q16oddPseudo,
3029 ARM::VLD4q32oddPseudo };
3030 return SelectVLD(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
3033 case Intrinsic::arm_neon_vld2lane: {
3034 unsigned DOpcodes[] = { ARM::VLD2LNd8Pseudo, ARM::VLD2LNd16Pseudo,
3035 ARM::VLD2LNd32Pseudo };
3036 unsigned QOpcodes[] = { ARM::VLD2LNq16Pseudo, ARM::VLD2LNq32Pseudo };
3037 return SelectVLDSTLane(N, true, false, 2, DOpcodes, QOpcodes);
3040 case Intrinsic::arm_neon_vld3lane: {
3041 unsigned DOpcodes[] = { ARM::VLD3LNd8Pseudo, ARM::VLD3LNd16Pseudo,
3042 ARM::VLD3LNd32Pseudo };
3043 unsigned QOpcodes[] = { ARM::VLD3LNq16Pseudo, ARM::VLD3LNq32Pseudo };
3044 return SelectVLDSTLane(N, true, false, 3, DOpcodes, QOpcodes);
3047 case Intrinsic::arm_neon_vld4lane: {
3048 unsigned DOpcodes[] = { ARM::VLD4LNd8Pseudo, ARM::VLD4LNd16Pseudo,
3049 ARM::VLD4LNd32Pseudo };
3050 unsigned QOpcodes[] = { ARM::VLD4LNq16Pseudo, ARM::VLD4LNq32Pseudo };
3051 return SelectVLDSTLane(N, true, false, 4, DOpcodes, QOpcodes);
3054 case Intrinsic::arm_neon_vst1: {
3055 unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
3056 ARM::VST1d32, ARM::VST1d64 };
3057 unsigned QOpcodes[] = { ARM::VST1q8Pseudo, ARM::VST1q16Pseudo,
3058 ARM::VST1q32Pseudo, ARM::VST1q64Pseudo };
3059 return SelectVST(N, false, 1, DOpcodes, QOpcodes, 0);
3062 case Intrinsic::arm_neon_vst2: {
3063 unsigned DOpcodes[] = { ARM::VST2d8Pseudo, ARM::VST2d16Pseudo,
3064 ARM::VST2d32Pseudo, ARM::VST1q64Pseudo };
3065 unsigned QOpcodes[] = { ARM::VST2q8Pseudo, ARM::VST2q16Pseudo,
3066 ARM::VST2q32Pseudo };
3067 return SelectVST(N, false, 2, DOpcodes, QOpcodes, 0);
3070 case Intrinsic::arm_neon_vst3: {
3071 unsigned DOpcodes[] = { ARM::VST3d8Pseudo, ARM::VST3d16Pseudo,
3072 ARM::VST3d32Pseudo, ARM::VST1d64TPseudo };
3073 unsigned QOpcodes0[] = { ARM::VST3q8Pseudo_UPD,
3074 ARM::VST3q16Pseudo_UPD,
3075 ARM::VST3q32Pseudo_UPD };
3076 unsigned QOpcodes1[] = { ARM::VST3q8oddPseudo,
3077 ARM::VST3q16oddPseudo,
3078 ARM::VST3q32oddPseudo };
3079 return SelectVST(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
3082 case Intrinsic::arm_neon_vst4: {
3083 unsigned DOpcodes[] = { ARM::VST4d8Pseudo, ARM::VST4d16Pseudo,
3084 ARM::VST4d32Pseudo, ARM::VST1d64QPseudo };
3085 unsigned QOpcodes0[] = { ARM::VST4q8Pseudo_UPD,
3086 ARM::VST4q16Pseudo_UPD,
3087 ARM::VST4q32Pseudo_UPD };
3088 unsigned QOpcodes1[] = { ARM::VST4q8oddPseudo,
3089 ARM::VST4q16oddPseudo,
3090 ARM::VST4q32oddPseudo };
3091 return SelectVST(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
3094 case Intrinsic::arm_neon_vst2lane: {
3095 unsigned DOpcodes[] = { ARM::VST2LNd8Pseudo, ARM::VST2LNd16Pseudo,
3096 ARM::VST2LNd32Pseudo };
3097 unsigned QOpcodes[] = { ARM::VST2LNq16Pseudo, ARM::VST2LNq32Pseudo };
3098 return SelectVLDSTLane(N, false, false, 2, DOpcodes, QOpcodes);
3101 case Intrinsic::arm_neon_vst3lane: {
3102 unsigned DOpcodes[] = { ARM::VST3LNd8Pseudo, ARM::VST3LNd16Pseudo,
3103 ARM::VST3LNd32Pseudo };
3104 unsigned QOpcodes[] = { ARM::VST3LNq16Pseudo, ARM::VST3LNq32Pseudo };
3105 return SelectVLDSTLane(N, false, false, 3, DOpcodes, QOpcodes);
3108 case Intrinsic::arm_neon_vst4lane: {
3109 unsigned DOpcodes[] = { ARM::VST4LNd8Pseudo, ARM::VST4LNd16Pseudo,
3110 ARM::VST4LNd32Pseudo };
3111 unsigned QOpcodes[] = { ARM::VST4LNq16Pseudo, ARM::VST4LNq32Pseudo };
3112 return SelectVLDSTLane(N, false, false, 4, DOpcodes, QOpcodes);
3118 case ISD::INTRINSIC_WO_CHAIN: {
3119 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3124 case Intrinsic::arm_neon_vtbl2:
3125 return SelectVTBL(N, false, 2, ARM::VTBL2Pseudo);
3126 case Intrinsic::arm_neon_vtbl3:
3127 return SelectVTBL(N, false, 3, ARM::VTBL3Pseudo);
3128 case Intrinsic::arm_neon_vtbl4:
3129 return SelectVTBL(N, false, 4, ARM::VTBL4Pseudo);
3131 case Intrinsic::arm_neon_vtbx2:
3132 return SelectVTBL(N, true, 2, ARM::VTBX2Pseudo);
3133 case Intrinsic::arm_neon_vtbx3:
3134 return SelectVTBL(N, true, 3, ARM::VTBX3Pseudo);
3135 case Intrinsic::arm_neon_vtbx4:
3136 return SelectVTBL(N, true, 4, ARM::VTBX4Pseudo);
3141 case ARMISD::VTBL1: {
3142 DebugLoc dl = N->getDebugLoc();
3143 EVT VT = N->getValueType(0);
3144 SmallVector<SDValue, 6> Ops;
3146 Ops.push_back(N->getOperand(0));
3147 Ops.push_back(N->getOperand(1));
3148 Ops.push_back(getAL(CurDAG)); // Predicate
3149 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // Predicate Register
3150 return CurDAG->getMachineNode(ARM::VTBL1, dl, VT, Ops.data(), Ops.size());
3152 case ARMISD::VTBL2: {
3153 DebugLoc dl = N->getDebugLoc();
3154 EVT VT = N->getValueType(0);
3156 // Form a REG_SEQUENCE to force register allocation.
3157 SDValue V0 = N->getOperand(0);
3158 SDValue V1 = N->getOperand(1);
3159 SDValue RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0);
3161 SmallVector<SDValue, 6> Ops;
3162 Ops.push_back(RegSeq);
3163 Ops.push_back(N->getOperand(2));
3164 Ops.push_back(getAL(CurDAG)); // Predicate
3165 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // Predicate Register
3166 return CurDAG->getMachineNode(ARM::VTBL2Pseudo, dl, VT,
3167 Ops.data(), Ops.size());
3170 case ISD::CONCAT_VECTORS:
3171 return SelectConcatVector(N);
3173 case ARMISD::ATOMOR64_DAG:
3174 return SelectAtomic64(N, ARM::ATOMOR6432);
3175 case ARMISD::ATOMXOR64_DAG:
3176 return SelectAtomic64(N, ARM::ATOMXOR6432);
3177 case ARMISD::ATOMADD64_DAG:
3178 return SelectAtomic64(N, ARM::ATOMADD6432);
3179 case ARMISD::ATOMSUB64_DAG:
3180 return SelectAtomic64(N, ARM::ATOMSUB6432);
3181 case ARMISD::ATOMNAND64_DAG:
3182 return SelectAtomic64(N, ARM::ATOMNAND6432);
3183 case ARMISD::ATOMAND64_DAG:
3184 return SelectAtomic64(N, ARM::ATOMAND6432);
3185 case ARMISD::ATOMSWAP64_DAG:
3186 return SelectAtomic64(N, ARM::ATOMSWAP6432);
3187 case ARMISD::ATOMCMPXCHG64_DAG:
3188 return SelectAtomic64(N, ARM::ATOMCMPXCHG6432);
3191 return SelectCode(N);
3194 bool ARMDAGToDAGISel::
3195 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
3196 std::vector<SDValue> &OutOps) {
3197 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
3198 // Require the address to be in a register. That is safe for all ARM
3199 // variants and it is hard to do anything much smarter without knowing
3200 // how the operand is used.
3201 OutOps.push_back(Op);
3205 /// createARMISelDag - This pass converts a legalized DAG into a
3206 /// ARM-specific DAG, ready for instruction scheduling.
3208 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
3209 CodeGenOpt::Level OptLevel) {
3210 return new ARMDAGToDAGISel(TM, OptLevel);