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 "ARMAddressingModes.h"
17 #include "ARMTargetMachine.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Function.h"
22 #include "llvm/Intrinsics.h"
23 #include "llvm/LLVMContext.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/CodeGen/SelectionDAGISel.h"
29 #include "llvm/Target/TargetLowering.h"
30 #include "llvm/Target/TargetOptions.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Compiler.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
40 UseRegSeq("neon-reg-sequence", cl::Hidden,
41 cl::desc("Use reg_sequence to model ld / st of multiple neon regs"),
44 //===--------------------------------------------------------------------===//
45 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
46 /// instructions for SelectionDAG operations.
49 class ARMDAGToDAGISel : public SelectionDAGISel {
50 ARMBaseTargetMachine &TM;
52 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
53 /// make the right decision when generating code for different targets.
54 const ARMSubtarget *Subtarget;
57 explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
58 CodeGenOpt::Level OptLevel)
59 : SelectionDAGISel(tm, OptLevel), TM(tm),
60 Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
63 virtual const char *getPassName() const {
64 return "ARM Instruction Selection";
67 /// getI32Imm - Return a target constant of type i32 with the specified
69 inline SDValue getI32Imm(unsigned Imm) {
70 return CurDAG->getTargetConstant(Imm, MVT::i32);
73 SDNode *Select(SDNode *N);
75 bool SelectShifterOperandReg(SDNode *Op, SDValue N, SDValue &A,
76 SDValue &B, SDValue &C);
77 bool SelectAddrMode2(SDNode *Op, SDValue N, SDValue &Base,
78 SDValue &Offset, SDValue &Opc);
79 bool SelectAddrMode2Offset(SDNode *Op, SDValue N,
80 SDValue &Offset, SDValue &Opc);
81 bool SelectAddrMode3(SDNode *Op, SDValue N, SDValue &Base,
82 SDValue &Offset, SDValue &Opc);
83 bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
84 SDValue &Offset, SDValue &Opc);
85 bool SelectAddrMode4(SDNode *Op, SDValue N, SDValue &Addr,
87 bool SelectAddrMode5(SDNode *Op, SDValue N, SDValue &Base,
89 bool SelectAddrMode6(SDNode *Op, SDValue N, SDValue &Addr, SDValue &Align);
91 bool SelectAddrModePC(SDNode *Op, SDValue N, SDValue &Offset,
94 bool SelectThumbAddrModeRR(SDNode *Op, SDValue N, SDValue &Base,
96 bool SelectThumbAddrModeRI5(SDNode *Op, SDValue N, unsigned Scale,
97 SDValue &Base, SDValue &OffImm,
99 bool SelectThumbAddrModeS1(SDNode *Op, SDValue N, SDValue &Base,
100 SDValue &OffImm, SDValue &Offset);
101 bool SelectThumbAddrModeS2(SDNode *Op, SDValue N, SDValue &Base,
102 SDValue &OffImm, SDValue &Offset);
103 bool SelectThumbAddrModeS4(SDNode *Op, SDValue N, SDValue &Base,
104 SDValue &OffImm, SDValue &Offset);
105 bool SelectThumbAddrModeSP(SDNode *Op, SDValue N, SDValue &Base,
108 bool SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
109 SDValue &BaseReg, SDValue &Opc);
110 bool SelectT2AddrModeImm12(SDNode *Op, SDValue N, SDValue &Base,
112 bool SelectT2AddrModeImm8(SDNode *Op, SDValue N, SDValue &Base,
114 bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
116 bool SelectT2AddrModeImm8s4(SDNode *Op, SDValue N, SDValue &Base,
118 bool SelectT2AddrModeSoReg(SDNode *Op, SDValue N, SDValue &Base,
119 SDValue &OffReg, SDValue &ShImm);
121 // Include the pieces autogenerated from the target description.
122 #include "ARMGenDAGISel.inc"
125 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
127 SDNode *SelectARMIndexedLoad(SDNode *N);
128 SDNode *SelectT2IndexedLoad(SDNode *N);
130 /// SelectVLD - Select NEON load intrinsics. NumVecs should be
131 /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for
132 /// loads of D registers and even subregs and odd subregs of Q registers.
133 /// For NumVecs <= 2, QOpcodes1 is not used.
134 SDNode *SelectVLD(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
135 unsigned *QOpcodes0, unsigned *QOpcodes1);
137 /// SelectVST - Select NEON store intrinsics. NumVecs should
138 /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for
139 /// stores of D registers and even subregs and odd subregs of Q registers.
140 /// For NumVecs <= 2, QOpcodes1 is not used.
141 SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
142 unsigned *QOpcodes0, unsigned *QOpcodes1);
144 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
145 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
146 /// load/store of D registers and even subregs and odd subregs of Q registers.
147 SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, unsigned NumVecs,
148 unsigned *DOpcodes, unsigned *QOpcodes0,
149 unsigned *QOpcodes1);
151 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
152 SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
154 /// SelectCMOVOp - Select CMOV instructions for ARM.
155 SDNode *SelectCMOVOp(SDNode *N);
156 SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
157 ARMCC::CondCodes CCVal, SDValue CCR,
159 SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
160 ARMCC::CondCodes CCVal, SDValue CCR,
162 SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
163 ARMCC::CondCodes CCVal, SDValue CCR,
165 SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
166 ARMCC::CondCodes CCVal, SDValue CCR,
169 SDNode *SelectConcatVector(SDNode *N);
171 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
172 /// inline asm expressions.
173 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
175 std::vector<SDValue> &OutOps);
177 // Form pairs of consecutive S, D, or Q registers.
178 SDNode *PairSRegs(EVT VT, SDValue V0, SDValue V1);
179 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
180 SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1);
182 // Form sequences of 4 consecutive S, D, or Q registers.
183 SDNode *QuadSRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
184 SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
185 SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
187 // Form sequences of 8 consecutive D registers.
188 SDNode *OctoDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3,
189 SDValue V4, SDValue V5, SDValue V6, SDValue V7);
193 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
194 /// operand. If so Imm will receive the 32-bit value.
195 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
196 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
197 Imm = cast<ConstantSDNode>(N)->getZExtValue();
203 // isInt32Immediate - This method tests to see if a constant operand.
204 // If so Imm will receive the 32 bit value.
205 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
206 return isInt32Immediate(N.getNode(), Imm);
209 // isOpcWithIntImmediate - This method tests to see if the node is a specific
210 // opcode and that it has a immediate integer right operand.
211 // If so Imm will receive the 32 bit value.
212 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
213 return N->getOpcode() == Opc &&
214 isInt32Immediate(N->getOperand(1).getNode(), Imm);
218 bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
223 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
225 // Don't match base register only case. That is matched to a separate
226 // lower complexity pattern with explicit register operand.
227 if (ShOpcVal == ARM_AM::no_shift) return false;
229 BaseReg = N.getOperand(0);
230 unsigned ShImmVal = 0;
231 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
232 ShReg = CurDAG->getRegister(0, MVT::i32);
233 ShImmVal = RHS->getZExtValue() & 31;
235 ShReg = N.getOperand(1);
237 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
242 bool ARMDAGToDAGISel::SelectAddrMode2(SDNode *Op, SDValue N,
243 SDValue &Base, SDValue &Offset,
245 if (N.getOpcode() == ISD::MUL) {
246 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
247 // X * [3,5,9] -> X + X * [2,4,8] etc.
248 int RHSC = (int)RHS->getZExtValue();
251 ARM_AM::AddrOpc AddSub = ARM_AM::add;
253 AddSub = ARM_AM::sub;
256 if (isPowerOf2_32(RHSC)) {
257 unsigned ShAmt = Log2_32(RHSC);
258 Base = Offset = N.getOperand(0);
259 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
268 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
270 if (N.getOpcode() == ISD::FrameIndex) {
271 int FI = cast<FrameIndexSDNode>(N)->getIndex();
272 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
273 } else if (N.getOpcode() == ARMISD::Wrapper &&
274 !(Subtarget->useMovt() &&
275 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
276 Base = N.getOperand(0);
278 Offset = CurDAG->getRegister(0, MVT::i32);
279 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
285 // Match simple R +/- imm12 operands.
286 if (N.getOpcode() == ISD::ADD)
287 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
288 int RHSC = (int)RHS->getZExtValue();
289 if ((RHSC >= 0 && RHSC < 0x1000) ||
290 (RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
291 Base = N.getOperand(0);
292 if (Base.getOpcode() == ISD::FrameIndex) {
293 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
294 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
296 Offset = CurDAG->getRegister(0, MVT::i32);
298 ARM_AM::AddrOpc AddSub = ARM_AM::add;
300 AddSub = ARM_AM::sub;
303 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
310 // Otherwise this is R +/- [possibly shifted] R.
311 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
312 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
315 Base = N.getOperand(0);
316 Offset = N.getOperand(1);
318 if (ShOpcVal != ARM_AM::no_shift) {
319 // Check to see if the RHS of the shift is a constant, if not, we can't fold
321 if (ConstantSDNode *Sh =
322 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
323 ShAmt = Sh->getZExtValue();
324 Offset = N.getOperand(1).getOperand(0);
326 ShOpcVal = ARM_AM::no_shift;
330 // Try matching (R shl C) + (R).
331 if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
332 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
333 if (ShOpcVal != ARM_AM::no_shift) {
334 // Check to see if the RHS of the shift is a constant, if not, we can't
336 if (ConstantSDNode *Sh =
337 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
338 ShAmt = Sh->getZExtValue();
339 Offset = N.getOperand(0).getOperand(0);
340 Base = N.getOperand(1);
342 ShOpcVal = ARM_AM::no_shift;
347 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
352 bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
353 SDValue &Offset, SDValue &Opc) {
354 unsigned Opcode = Op->getOpcode();
355 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
356 ? cast<LoadSDNode>(Op)->getAddressingMode()
357 : cast<StoreSDNode>(Op)->getAddressingMode();
358 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
359 ? ARM_AM::add : ARM_AM::sub;
360 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
361 int Val = (int)C->getZExtValue();
362 if (Val >= 0 && Val < 0x1000) { // 12 bits.
363 Offset = CurDAG->getRegister(0, MVT::i32);
364 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
372 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
374 if (ShOpcVal != ARM_AM::no_shift) {
375 // Check to see if the RHS of the shift is a constant, if not, we can't fold
377 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
378 ShAmt = Sh->getZExtValue();
379 Offset = N.getOperand(0);
381 ShOpcVal = ARM_AM::no_shift;
385 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
391 bool ARMDAGToDAGISel::SelectAddrMode3(SDNode *Op, SDValue N,
392 SDValue &Base, SDValue &Offset,
394 if (N.getOpcode() == ISD::SUB) {
395 // X - C is canonicalize to X + -C, no need to handle it here.
396 Base = N.getOperand(0);
397 Offset = N.getOperand(1);
398 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
402 if (N.getOpcode() != ISD::ADD) {
404 if (N.getOpcode() == ISD::FrameIndex) {
405 int FI = cast<FrameIndexSDNode>(N)->getIndex();
406 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
408 Offset = CurDAG->getRegister(0, MVT::i32);
409 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
413 // If the RHS is +/- imm8, fold into addr mode.
414 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
415 int RHSC = (int)RHS->getZExtValue();
416 if ((RHSC >= 0 && RHSC < 256) ||
417 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
418 Base = N.getOperand(0);
419 if (Base.getOpcode() == ISD::FrameIndex) {
420 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
421 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
423 Offset = CurDAG->getRegister(0, MVT::i32);
425 ARM_AM::AddrOpc AddSub = ARM_AM::add;
427 AddSub = ARM_AM::sub;
430 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
435 Base = N.getOperand(0);
436 Offset = N.getOperand(1);
437 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
441 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
442 SDValue &Offset, SDValue &Opc) {
443 unsigned Opcode = Op->getOpcode();
444 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
445 ? cast<LoadSDNode>(Op)->getAddressingMode()
446 : cast<StoreSDNode>(Op)->getAddressingMode();
447 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
448 ? ARM_AM::add : ARM_AM::sub;
449 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
450 int Val = (int)C->getZExtValue();
451 if (Val >= 0 && Val < 256) {
452 Offset = CurDAG->getRegister(0, MVT::i32);
453 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
459 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
463 bool ARMDAGToDAGISel::SelectAddrMode4(SDNode *Op, SDValue N,
464 SDValue &Addr, SDValue &Mode) {
466 Mode = CurDAG->getTargetConstant(0, MVT::i32);
470 bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N,
471 SDValue &Base, SDValue &Offset) {
472 if (N.getOpcode() != ISD::ADD) {
474 if (N.getOpcode() == ISD::FrameIndex) {
475 int FI = cast<FrameIndexSDNode>(N)->getIndex();
476 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
477 } else if (N.getOpcode() == ARMISD::Wrapper &&
478 !(Subtarget->useMovt() &&
479 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
480 Base = N.getOperand(0);
482 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
487 // If the RHS is +/- imm8, fold into addr mode.
488 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
489 int RHSC = (int)RHS->getZExtValue();
490 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied by 4.
492 if ((RHSC >= 0 && RHSC < 256) ||
493 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
494 Base = N.getOperand(0);
495 if (Base.getOpcode() == ISD::FrameIndex) {
496 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
497 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
500 ARM_AM::AddrOpc AddSub = ARM_AM::add;
502 AddSub = ARM_AM::sub;
505 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
513 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
518 bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N,
519 SDValue &Addr, SDValue &Align) {
521 // Default to no alignment.
522 Align = CurDAG->getTargetConstant(0, MVT::i32);
526 bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
527 SDValue &Offset, SDValue &Label) {
528 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
529 Offset = N.getOperand(0);
530 SDValue N1 = N.getOperand(1);
531 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
538 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
539 SDValue &Base, SDValue &Offset){
540 // FIXME dl should come from the parent load or store, not the address
541 DebugLoc dl = Op->getDebugLoc();
542 if (N.getOpcode() != ISD::ADD) {
543 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
544 if (!NC || NC->getZExtValue() != 0)
551 Base = N.getOperand(0);
552 Offset = N.getOperand(1);
557 ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
558 unsigned Scale, SDValue &Base,
559 SDValue &OffImm, SDValue &Offset) {
561 SDValue TmpBase, TmpOffImm;
562 if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
563 return false; // We want to select tLDRspi / tSTRspi instead.
564 if (N.getOpcode() == ARMISD::Wrapper &&
565 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
566 return false; // We want to select tLDRpci instead.
569 if (N.getOpcode() != ISD::ADD) {
570 if (N.getOpcode() == ARMISD::Wrapper &&
571 !(Subtarget->useMovt() &&
572 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
573 Base = N.getOperand(0);
577 Offset = CurDAG->getRegister(0, MVT::i32);
578 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
582 // Thumb does not have [sp, r] address mode.
583 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
584 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
585 if ((LHSR && LHSR->getReg() == ARM::SP) ||
586 (RHSR && RHSR->getReg() == ARM::SP)) {
588 Offset = CurDAG->getRegister(0, MVT::i32);
589 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
593 // If the RHS is + imm5 * scale, fold into addr mode.
594 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
595 int RHSC = (int)RHS->getZExtValue();
596 if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
598 if (RHSC >= 0 && RHSC < 32) {
599 Base = N.getOperand(0);
600 Offset = CurDAG->getRegister(0, MVT::i32);
601 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
607 Base = N.getOperand(0);
608 Offset = N.getOperand(1);
609 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
613 bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
614 SDValue &Base, SDValue &OffImm,
616 return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
619 bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
620 SDValue &Base, SDValue &OffImm,
622 return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
625 bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
626 SDValue &Base, SDValue &OffImm,
628 return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
631 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
632 SDValue &Base, SDValue &OffImm) {
633 if (N.getOpcode() == ISD::FrameIndex) {
634 int FI = cast<FrameIndexSDNode>(N)->getIndex();
635 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
636 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
640 if (N.getOpcode() != ISD::ADD)
643 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
644 if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
645 (LHSR && LHSR->getReg() == ARM::SP)) {
646 // If the RHS is + imm8 * scale, fold into addr mode.
647 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
648 int RHSC = (int)RHS->getZExtValue();
649 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
651 if (RHSC >= 0 && RHSC < 256) {
652 Base = N.getOperand(0);
653 if (Base.getOpcode() == ISD::FrameIndex) {
654 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
655 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
657 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
667 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
670 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
672 // Don't match base register only case. That is matched to a separate
673 // lower complexity pattern with explicit register operand.
674 if (ShOpcVal == ARM_AM::no_shift) return false;
676 BaseReg = N.getOperand(0);
677 unsigned ShImmVal = 0;
678 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
679 ShImmVal = RHS->getZExtValue() & 31;
680 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
687 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
688 SDValue &Base, SDValue &OffImm) {
689 // Match simple R + imm12 operands.
692 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
693 if (N.getOpcode() == ISD::FrameIndex) {
694 // Match frame index...
695 int FI = cast<FrameIndexSDNode>(N)->getIndex();
696 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
697 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
699 } else if (N.getOpcode() == ARMISD::Wrapper &&
700 !(Subtarget->useMovt() &&
701 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
702 Base = N.getOperand(0);
703 if (Base.getOpcode() == ISD::TargetConstantPool)
704 return false; // We want to select t2LDRpci instead.
707 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
711 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
712 if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
713 // Let t2LDRi8 handle (R - imm8).
716 int RHSC = (int)RHS->getZExtValue();
717 if (N.getOpcode() == ISD::SUB)
720 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
721 Base = N.getOperand(0);
722 if (Base.getOpcode() == ISD::FrameIndex) {
723 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
724 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
726 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
733 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
737 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
738 SDValue &Base, SDValue &OffImm) {
739 // Match simple R - imm8 operands.
740 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
741 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
742 int RHSC = (int)RHS->getSExtValue();
743 if (N.getOpcode() == ISD::SUB)
746 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
747 Base = N.getOperand(0);
748 if (Base.getOpcode() == ISD::FrameIndex) {
749 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
750 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
752 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
761 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
763 unsigned Opcode = Op->getOpcode();
764 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
765 ? cast<LoadSDNode>(Op)->getAddressingMode()
766 : cast<StoreSDNode>(Op)->getAddressingMode();
767 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
768 int RHSC = (int)RHS->getZExtValue();
769 if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
770 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
771 ? CurDAG->getTargetConstant(RHSC, MVT::i32)
772 : CurDAG->getTargetConstant(-RHSC, MVT::i32);
780 bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
781 SDValue &Base, SDValue &OffImm) {
782 if (N.getOpcode() == ISD::ADD) {
783 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
784 int RHSC = (int)RHS->getZExtValue();
786 if (((RHSC & 0x3) == 0) &&
787 ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) {
788 Base = N.getOperand(0);
789 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
793 } else if (N.getOpcode() == ISD::SUB) {
794 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
795 int RHSC = (int)RHS->getZExtValue();
797 if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) {
798 Base = N.getOperand(0);
799 OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
808 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDNode *Op, SDValue N,
810 SDValue &OffReg, SDValue &ShImm) {
811 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
812 if (N.getOpcode() != ISD::ADD)
815 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
816 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
817 int RHSC = (int)RHS->getZExtValue();
818 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
820 else if (RHSC < 0 && RHSC >= -255) // 8 bits
824 // Look for (R + R) or (R + (R << [1,2,3])).
826 Base = N.getOperand(0);
827 OffReg = N.getOperand(1);
829 // Swap if it is ((R << c) + R).
830 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
831 if (ShOpcVal != ARM_AM::lsl) {
832 ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
833 if (ShOpcVal == ARM_AM::lsl)
834 std::swap(Base, OffReg);
837 if (ShOpcVal == ARM_AM::lsl) {
838 // Check to see if the RHS of the shift is a constant, if not, we can't fold
840 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
841 ShAmt = Sh->getZExtValue();
844 ShOpcVal = ARM_AM::no_shift;
846 OffReg = OffReg.getOperand(0);
848 ShOpcVal = ARM_AM::no_shift;
852 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
857 //===--------------------------------------------------------------------===//
859 /// getAL - Returns a ARMCC::AL immediate node.
860 static inline SDValue getAL(SelectionDAG *CurDAG) {
861 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
864 SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
865 LoadSDNode *LD = cast<LoadSDNode>(N);
866 ISD::MemIndexedMode AM = LD->getAddressingMode();
867 if (AM == ISD::UNINDEXED)
870 EVT LoadedVT = LD->getMemoryVT();
871 SDValue Offset, AMOpc;
872 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
875 if (LoadedVT == MVT::i32 &&
876 SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
877 Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
879 } else if (LoadedVT == MVT::i16 &&
880 SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
882 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
883 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
884 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
885 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
886 if (LD->getExtensionType() == ISD::SEXTLOAD) {
887 if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
889 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
892 if (SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
894 Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
900 SDValue Chain = LD->getChain();
901 SDValue Base = LD->getBasePtr();
902 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
903 CurDAG->getRegister(0, MVT::i32), Chain };
904 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
911 SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
912 LoadSDNode *LD = cast<LoadSDNode>(N);
913 ISD::MemIndexedMode AM = LD->getAddressingMode();
914 if (AM == ISD::UNINDEXED)
917 EVT LoadedVT = LD->getMemoryVT();
918 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
920 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
923 if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
924 switch (LoadedVT.getSimpleVT().SimpleTy) {
926 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
930 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
932 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
937 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
939 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
948 SDValue Chain = LD->getChain();
949 SDValue Base = LD->getBasePtr();
950 SDValue Ops[]= { Base, Offset, getAL(CurDAG),
951 CurDAG->getRegister(0, MVT::i32), Chain };
952 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
959 /// PairSRegs - Form a D register from a pair of S registers.
961 SDNode *ARMDAGToDAGISel::PairSRegs(EVT VT, SDValue V0, SDValue V1) {
962 DebugLoc dl = V0.getNode()->getDebugLoc();
963 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
964 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
965 if (llvm::ModelWithRegSequence()) {
966 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
967 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
970 SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0);
971 SDNode *Pair = CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
972 VT, Undef, V0, SubReg0);
973 return CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
974 VT, SDValue(Pair, 0), V1, SubReg1);
977 /// PairDRegs - Form a quad register from a pair of D registers.
979 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
980 DebugLoc dl = V0.getNode()->getDebugLoc();
981 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
982 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
983 if (llvm::ModelWithRegSequence()) {
984 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
985 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
988 SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0);
989 SDNode *Pair = CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
990 VT, Undef, V0, SubReg0);
991 return CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
992 VT, SDValue(Pair, 0), V1, SubReg1);
995 /// PairQRegs - Form 4 consecutive D registers from a pair of Q registers.
997 SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) {
998 DebugLoc dl = V0.getNode()->getDebugLoc();
999 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
1000 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
1001 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
1002 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
1005 /// QuadSRegs - Form 4 consecutive S registers.
1007 SDNode *ARMDAGToDAGISel::QuadSRegs(EVT VT, SDValue V0, SDValue V1,
1008 SDValue V2, SDValue V3) {
1009 DebugLoc dl = V0.getNode()->getDebugLoc();
1010 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
1011 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
1012 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, MVT::i32);
1013 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, MVT::i32);
1014 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1015 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1018 /// QuadDRegs - Form 4 consecutive D registers.
1020 SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
1021 SDValue V2, SDValue V3) {
1022 DebugLoc dl = V0.getNode()->getDebugLoc();
1023 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1024 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1025 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
1026 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
1027 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1028 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1031 /// QuadQRegs - Form 4 consecutive Q registers.
1033 SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1,
1034 SDValue V2, SDValue V3) {
1035 DebugLoc dl = V0.getNode()->getDebugLoc();
1036 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
1037 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
1038 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, MVT::i32);
1039 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, MVT::i32);
1040 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1041 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1044 /// OctoDRegs - Form 8 consecutive D registers.
1046 SDNode *ARMDAGToDAGISel::OctoDRegs(EVT VT, SDValue V0, SDValue V1,
1047 SDValue V2, SDValue V3,
1048 SDValue V4, SDValue V5,
1049 SDValue V6, SDValue V7) {
1050 DebugLoc dl = V0.getNode()->getDebugLoc();
1051 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1052 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1053 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
1054 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
1055 SDValue SubReg4 = CurDAG->getTargetConstant(ARM::dsub_4, MVT::i32);
1056 SDValue SubReg5 = CurDAG->getTargetConstant(ARM::dsub_5, MVT::i32);
1057 SDValue SubReg6 = CurDAG->getTargetConstant(ARM::dsub_6, MVT::i32);
1058 SDValue SubReg7 = CurDAG->getTargetConstant(ARM::dsub_7, MVT::i32);
1059 const SDValue Ops[] ={ V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3,
1060 V4, SubReg4, V5, SubReg5, V6, SubReg6, V7, SubReg7 };
1061 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 16);
1064 /// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
1065 /// for a 64-bit subregister of the vector.
1066 static EVT GetNEONSubregVT(EVT VT) {
1067 switch (VT.getSimpleVT().SimpleTy) {
1068 default: llvm_unreachable("unhandled NEON type");
1069 case MVT::v16i8: return MVT::v8i8;
1070 case MVT::v8i16: return MVT::v4i16;
1071 case MVT::v4f32: return MVT::v2f32;
1072 case MVT::v4i32: return MVT::v2i32;
1073 case MVT::v2i64: return MVT::v1i64;
1077 SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
1078 unsigned *DOpcodes, unsigned *QOpcodes0,
1079 unsigned *QOpcodes1) {
1080 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1081 DebugLoc dl = N->getDebugLoc();
1083 SDValue MemAddr, Align;
1084 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1087 SDValue Chain = N->getOperand(0);
1088 EVT VT = N->getValueType(0);
1089 bool is64BitVector = VT.is64BitVector();
1091 unsigned OpcodeIndex;
1092 switch (VT.getSimpleVT().SimpleTy) {
1093 default: llvm_unreachable("unhandled vld type");
1094 // Double-register operations:
1095 case MVT::v8i8: OpcodeIndex = 0; break;
1096 case MVT::v4i16: OpcodeIndex = 1; break;
1098 case MVT::v2i32: OpcodeIndex = 2; break;
1099 case MVT::v1i64: OpcodeIndex = 3; break;
1100 // Quad-register operations:
1101 case MVT::v16i8: OpcodeIndex = 0; break;
1102 case MVT::v8i16: OpcodeIndex = 1; break;
1104 case MVT::v4i32: OpcodeIndex = 2; break;
1105 case MVT::v2i64: OpcodeIndex = 3;
1106 assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
1110 SDValue Pred = getAL(CurDAG);
1111 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1112 if (is64BitVector) {
1113 unsigned Opc = DOpcodes[OpcodeIndex];
1114 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1115 std::vector<EVT> ResTys(NumVecs, VT);
1116 ResTys.push_back(MVT::Other);
1117 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1118 if (!llvm::ModelWithRegSequence() || NumVecs < 2)
1122 SDValue V0 = SDValue(VLd, 0);
1123 SDValue V1 = SDValue(VLd, 1);
1125 // Form a REG_SEQUENCE to force register allocation.
1127 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1129 SDValue V2 = SDValue(VLd, 2);
1130 // If it's a vld3, form a quad D-register but discard the last part.
1131 SDValue V3 = (NumVecs == 3)
1132 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1134 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1137 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1138 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1139 SDValue D = CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec,
1141 ReplaceUses(SDValue(N, Vec), D);
1143 ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, NumVecs));
1147 EVT RegVT = GetNEONSubregVT(VT);
1149 // Quad registers are directly supported for VLD1 and VLD2,
1150 // loading pairs of D regs.
1151 unsigned Opc = QOpcodes0[OpcodeIndex];
1152 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1153 std::vector<EVT> ResTys(2 * NumVecs, RegVT);
1154 ResTys.push_back(MVT::Other);
1155 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1156 Chain = SDValue(VLd, 2 * NumVecs);
1158 // Combine the even and odd subregs to produce the result.
1159 if (llvm::ModelWithRegSequence()) {
1161 SDNode *Q = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
1162 ReplaceUses(SDValue(N, 0), SDValue(Q, 0));
1164 SDValue QQ = SDValue(QuadDRegs(MVT::v4i64,
1165 SDValue(VLd, 0), SDValue(VLd, 1),
1166 SDValue(VLd, 2), SDValue(VLd, 3)), 0);
1167 SDValue Q0 = CurDAG->getTargetExtractSubreg(ARM::qsub_0, dl, VT, QQ);
1168 SDValue Q1 = CurDAG->getTargetExtractSubreg(ARM::qsub_1, dl, VT, QQ);
1169 ReplaceUses(SDValue(N, 0), Q0);
1170 ReplaceUses(SDValue(N, 1), Q1);
1173 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1174 SDNode *Q = PairDRegs(VT, SDValue(VLd, 2*Vec), SDValue(VLd, 2*Vec+1));
1175 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1179 // Otherwise, quad registers are loaded with two separate instructions,
1180 // where one loads the even registers and the other loads the odd registers.
1182 std::vector<EVT> ResTys(NumVecs, RegVT);
1183 ResTys.push_back(MemAddr.getValueType());
1184 ResTys.push_back(MVT::Other);
1186 // Load the even subregs.
1187 unsigned Opc = QOpcodes0[OpcodeIndex];
1188 const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain };
1189 SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6);
1190 Chain = SDValue(VLdA, NumVecs+1);
1192 // Load the odd subregs.
1193 Opc = QOpcodes1[OpcodeIndex];
1194 const SDValue OpsB[] = { SDValue(VLdA, NumVecs),
1195 Align, Reg0, Pred, Reg0, Chain };
1196 SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
1197 Chain = SDValue(VLdB, NumVecs+1);
1199 if (llvm::ModelWithRegSequence()) {
1200 SDValue V0 = SDValue(VLdA, 0);
1201 SDValue V1 = SDValue(VLdB, 0);
1202 SDValue V2 = SDValue(VLdA, 1);
1203 SDValue V3 = SDValue(VLdB, 1);
1204 SDValue V4 = SDValue(VLdA, 2);
1205 SDValue V5 = SDValue(VLdB, 2);
1206 SDValue V6 = (NumVecs == 3)
1207 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT),
1210 SDValue V7 = (NumVecs == 3)
1211 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT),
1214 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V0, V1, V2, V3,
1215 V4, V5, V6, V7), 0);
1217 // Extract out the 3 / 4 Q registers.
1218 assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1219 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1220 SDValue Q = CurDAG->getTargetExtractSubreg(ARM::qsub_0+Vec,
1222 ReplaceUses(SDValue(N, Vec), Q);
1225 // Combine the even and odd subregs to produce the result.
1226 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1227 SDNode *Q = PairDRegs(VT, SDValue(VLdA, Vec), SDValue(VLdB, Vec));
1228 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1232 ReplaceUses(SDValue(N, NumVecs), Chain);
1236 SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
1237 unsigned *DOpcodes, unsigned *QOpcodes0,
1238 unsigned *QOpcodes1) {
1239 assert(NumVecs >=1 && NumVecs <= 4 && "VST NumVecs out-of-range");
1240 DebugLoc dl = N->getDebugLoc();
1242 SDValue MemAddr, Align;
1243 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1246 SDValue Chain = N->getOperand(0);
1247 EVT VT = N->getOperand(3).getValueType();
1248 bool is64BitVector = VT.is64BitVector();
1250 unsigned OpcodeIndex;
1251 switch (VT.getSimpleVT().SimpleTy) {
1252 default: llvm_unreachable("unhandled vst type");
1253 // Double-register operations:
1254 case MVT::v8i8: OpcodeIndex = 0; break;
1255 case MVT::v4i16: OpcodeIndex = 1; break;
1257 case MVT::v2i32: OpcodeIndex = 2; break;
1258 case MVT::v1i64: OpcodeIndex = 3; break;
1259 // Quad-register operations:
1260 case MVT::v16i8: OpcodeIndex = 0; break;
1261 case MVT::v8i16: OpcodeIndex = 1; break;
1263 case MVT::v4i32: OpcodeIndex = 2; break;
1264 case MVT::v2i64: OpcodeIndex = 3;
1265 assert(NumVecs == 1 && "v2i64 type only supported for VST1");
1269 SDValue Pred = getAL(CurDAG);
1270 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1272 SmallVector<SDValue, 10> Ops;
1273 Ops.push_back(MemAddr);
1274 Ops.push_back(Align);
1276 if (is64BitVector) {
1277 if (llvm::ModelWithRegSequence() && NumVecs >= 2) {
1279 SDValue V0 = N->getOperand(0+3);
1280 SDValue V1 = N->getOperand(1+3);
1282 // Form a REG_SEQUENCE to force register allocation.
1284 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1286 SDValue V2 = N->getOperand(2+3);
1287 // If it's a vld3, form a quad D-register and leave the last part as
1289 SDValue V3 = (NumVecs == 3)
1290 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1291 : N->getOperand(3+3);
1292 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1295 // Now extract the D registers back out.
1296 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT,
1298 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT,
1301 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,
1304 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,
1307 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1308 Ops.push_back(N->getOperand(Vec+3));
1310 Ops.push_back(Pred);
1311 Ops.push_back(Reg0); // predicate register
1312 Ops.push_back(Chain);
1313 unsigned Opc = DOpcodes[OpcodeIndex];
1314 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+5);
1317 EVT RegVT = GetNEONSubregVT(VT);
1319 // Quad registers are directly supported for VST1 and VST2,
1320 // storing pairs of D regs.
1321 unsigned Opc = QOpcodes0[OpcodeIndex];
1322 if (llvm::ModelWithRegSequence() && NumVecs == 2) {
1323 // First extract the pair of Q registers.
1324 SDValue Q0 = N->getOperand(3);
1325 SDValue Q1 = N->getOperand(4);
1327 // Form a QQ register.
1328 SDValue QQ = SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0);
1330 // Now extract the D registers back out.
1331 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1333 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1335 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, RegVT,
1337 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, RegVT,
1339 Ops.push_back(Pred);
1340 Ops.push_back(Reg0); // predicate register
1341 Ops.push_back(Chain);
1342 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 5 + 4);
1344 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1345 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1346 N->getOperand(Vec+3)));
1347 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1348 N->getOperand(Vec+3)));
1350 Ops.push_back(Pred);
1351 Ops.push_back(Reg0); // predicate register
1352 Ops.push_back(Chain);
1353 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
1358 // Otherwise, quad registers are stored with two separate instructions,
1359 // where one stores the even registers and the other stores the odd registers.
1360 if (llvm::ModelWithRegSequence()) {
1361 // Form the QQQQ REG_SEQUENCE.
1363 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1364 V[i] = CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1365 N->getOperand(Vec+3));
1366 V[i+1] = CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1367 N->getOperand(Vec+3));
1370 V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1373 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1374 V[4], V[5], V[6], V[7]), 0);
1376 // Store the even D registers.
1377 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1378 Ops.push_back(Reg0); // post-access address offset
1379 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1380 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec*2, dl,
1382 Ops.push_back(Pred);
1383 Ops.push_back(Reg0); // predicate register
1384 Ops.push_back(Chain);
1385 unsigned Opc = QOpcodes0[OpcodeIndex];
1386 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1387 MVT::Other, Ops.data(), NumVecs+6);
1388 Chain = SDValue(VStA, 1);
1390 // Store the odd D registers.
1391 Ops[0] = SDValue(VStA, 0); // MemAddr
1392 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1393 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::dsub_1+Vec*2, dl,
1395 Ops[NumVecs+5] = Chain;
1396 Opc = QOpcodes1[OpcodeIndex];
1397 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1398 MVT::Other, Ops.data(), NumVecs+6);
1399 Chain = SDValue(VStB, 1);
1400 ReplaceUses(SDValue(N, 0), Chain);
1403 Ops.push_back(Reg0); // post-access address offset
1405 // Store the even subregs.
1406 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1407 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1408 N->getOperand(Vec+3)));
1409 Ops.push_back(Pred);
1410 Ops.push_back(Reg0); // predicate register
1411 Ops.push_back(Chain);
1412 unsigned Opc = QOpcodes0[OpcodeIndex];
1413 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1414 MVT::Other, Ops.data(), NumVecs+6);
1415 Chain = SDValue(VStA, 1);
1417 // Store the odd subregs.
1418 Ops[0] = SDValue(VStA, 0); // MemAddr
1419 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1420 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1421 N->getOperand(Vec+3));
1422 Ops[NumVecs+5] = Chain;
1423 Opc = QOpcodes1[OpcodeIndex];
1424 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1425 MVT::Other, Ops.data(), NumVecs+6);
1426 Chain = SDValue(VStB, 1);
1427 ReplaceUses(SDValue(N, 0), Chain);
1432 SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1433 unsigned NumVecs, unsigned *DOpcodes,
1434 unsigned *QOpcodes0,
1435 unsigned *QOpcodes1) {
1436 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1437 DebugLoc dl = N->getDebugLoc();
1439 SDValue MemAddr, Align;
1440 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1443 SDValue Chain = N->getOperand(0);
1445 cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1446 EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
1447 bool is64BitVector = VT.is64BitVector();
1449 // Quad registers are handled by load/store of subregs. Find the subreg info.
1450 unsigned NumElts = 0;
1454 if (!is64BitVector) {
1455 RegVT = GetNEONSubregVT(VT);
1456 NumElts = RegVT.getVectorNumElements();
1457 SubregIdx = (Lane < NumElts) ? ARM::dsub_0 : ARM::dsub_1;
1458 Even = Lane < NumElts;
1461 unsigned OpcodeIndex;
1462 switch (VT.getSimpleVT().SimpleTy) {
1463 default: llvm_unreachable("unhandled vld/vst lane type");
1464 // Double-register operations:
1465 case MVT::v8i8: OpcodeIndex = 0; break;
1466 case MVT::v4i16: OpcodeIndex = 1; break;
1468 case MVT::v2i32: OpcodeIndex = 2; break;
1469 // Quad-register operations:
1470 case MVT::v8i16: OpcodeIndex = 0; break;
1472 case MVT::v4i32: OpcodeIndex = 1; break;
1475 SDValue Pred = getAL(CurDAG);
1476 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1478 SmallVector<SDValue, 10> Ops;
1479 Ops.push_back(MemAddr);
1480 Ops.push_back(Align);
1483 if (is64BitVector) {
1484 Opc = DOpcodes[OpcodeIndex];
1485 if (llvm::ModelWithRegSequence()) {
1487 SDValue V0 = N->getOperand(0+3);
1488 SDValue V1 = N->getOperand(1+3);
1490 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1492 SDValue V2 = N->getOperand(2+3);
1493 SDValue V3 = (NumVecs == 3)
1494 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1495 : N->getOperand(3+3);
1496 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1499 // Now extract the D registers back out.
1500 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT,
1502 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT,
1505 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,
1508 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,
1511 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1512 Ops.push_back(N->getOperand(Vec+3));
1515 // Check if this is loading the even or odd subreg of a Q register.
1516 if (Lane < NumElts) {
1517 Opc = QOpcodes0[OpcodeIndex];
1520 Opc = QOpcodes1[OpcodeIndex];
1523 if (llvm::ModelWithRegSequence()) {
1525 SDValue V0 = N->getOperand(0+3);
1526 SDValue V1 = N->getOperand(1+3);
1528 RegSeq = SDValue(PairQRegs(MVT::v4i64, V0, V1), 0);
1530 SDValue V2 = N->getOperand(2+3);
1531 SDValue V3 = (NumVecs == 3)
1532 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1533 : N->getOperand(3+3);
1534 RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
1537 // Extract the subregs of the input vector.
1538 unsigned SubIdx = Even ? ARM::dsub_0 : ARM::dsub_1;
1539 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1540 Ops.push_back(CurDAG->getTargetExtractSubreg(SubIdx+Vec*2, dl, RegVT,
1543 // Extract the subregs of the input vector.
1544 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1545 Ops.push_back(CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1546 N->getOperand(Vec+3)));
1549 Ops.push_back(getI32Imm(Lane));
1550 Ops.push_back(Pred);
1551 Ops.push_back(Reg0);
1552 Ops.push_back(Chain);
1555 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6);
1557 std::vector<EVT> ResTys(NumVecs, RegVT);
1558 ResTys.push_back(MVT::Other);
1559 SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(),NumVecs+6);
1561 if (llvm::ModelWithRegSequence()) {
1562 // Form a REG_SEQUENCE to force register allocation.
1564 if (is64BitVector) {
1565 SDValue V0 = SDValue(VLdLn, 0);
1566 SDValue V1 = SDValue(VLdLn, 1);
1568 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1570 SDValue V2 = SDValue(VLdLn, 2);
1571 // If it's a vld3, form a quad D-register but discard the last part.
1572 SDValue V3 = (NumVecs == 3)
1573 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1574 : SDValue(VLdLn, 3);
1575 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1578 // For 128-bit vectors, take the 64-bit results of the load and insert
1579 // them as subregs into the result.
1581 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1583 V[i] = SDValue(VLdLn, Vec);
1584 V[i+1] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1587 V[i] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1589 V[i+1] = SDValue(VLdLn, Vec);
1593 V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1597 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V[0], V[1], V[2], V[3]), 0);
1599 RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1600 V[4], V[5], V[6], V[7]), 0);
1603 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1604 assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1605 unsigned SubIdx = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
1606 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1607 ReplaceUses(SDValue(N, Vec),
1608 CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, RegSeq));
1609 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, NumVecs));
1613 // For a 64-bit vector load to D registers, nothing more needs to be done.
1617 // For 128-bit vectors, take the 64-bit results of the load and insert them
1618 // as subregs into the result.
1619 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1620 SDValue QuadVec = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
1621 N->getOperand(Vec+3),
1622 SDValue(VLdLn, Vec));
1623 ReplaceUses(SDValue(N, Vec), QuadVec);
1626 Chain = SDValue(VLdLn, NumVecs);
1627 ReplaceUses(SDValue(N, NumVecs), Chain);
1631 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
1633 if (!Subtarget->hasV6T2Ops())
1636 unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
1637 : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
1640 // For unsigned extracts, check for a shift right and mask
1641 unsigned And_imm = 0;
1642 if (N->getOpcode() == ISD::AND) {
1643 if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
1645 // The immediate is a mask of the low bits iff imm & (imm+1) == 0
1646 if (And_imm & (And_imm + 1))
1649 unsigned Srl_imm = 0;
1650 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
1652 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1654 unsigned Width = CountTrailingOnes_32(And_imm);
1655 unsigned LSB = Srl_imm;
1656 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1657 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1658 CurDAG->getTargetConstant(LSB, MVT::i32),
1659 CurDAG->getTargetConstant(Width, MVT::i32),
1660 getAL(CurDAG), Reg0 };
1661 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1667 // Otherwise, we're looking for a shift of a shift
1668 unsigned Shl_imm = 0;
1669 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
1670 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1671 unsigned Srl_imm = 0;
1672 if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
1673 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1674 unsigned Width = 32 - Srl_imm;
1675 int LSB = Srl_imm - Shl_imm;
1678 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1679 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1680 CurDAG->getTargetConstant(LSB, MVT::i32),
1681 CurDAG->getTargetConstant(Width, MVT::i32),
1682 getAL(CurDAG), Reg0 };
1683 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1689 SDNode *ARMDAGToDAGISel::
1690 SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1691 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1694 if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
1695 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1696 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1699 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1700 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1701 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1702 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1704 llvm_unreachable("Unknown so_reg opcode!");
1708 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1709 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1710 SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
1711 return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
1716 SDNode *ARMDAGToDAGISel::
1717 SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1718 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1722 if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
1723 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1724 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
1725 return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
1730 SDNode *ARMDAGToDAGISel::
1731 SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1732 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1733 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1737 if (Predicate_t2_so_imm(TrueVal.getNode())) {
1738 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1739 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1740 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1741 return CurDAG->SelectNodeTo(N,
1742 ARM::t2MOVCCi, MVT::i32, Ops, 5);
1747 SDNode *ARMDAGToDAGISel::
1748 SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1749 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1750 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1754 if (Predicate_so_imm(TrueVal.getNode())) {
1755 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1756 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1757 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1758 return CurDAG->SelectNodeTo(N,
1759 ARM::MOVCCi, MVT::i32, Ops, 5);
1764 SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
1765 EVT VT = N->getValueType(0);
1766 SDValue FalseVal = N->getOperand(0);
1767 SDValue TrueVal = N->getOperand(1);
1768 SDValue CC = N->getOperand(2);
1769 SDValue CCR = N->getOperand(3);
1770 SDValue InFlag = N->getOperand(4);
1771 assert(CC.getOpcode() == ISD::Constant);
1772 assert(CCR.getOpcode() == ISD::Register);
1773 ARMCC::CondCodes CCVal =
1774 (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
1776 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1777 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1778 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1779 // Pattern complexity = 18 cost = 1 size = 0
1783 if (Subtarget->isThumb()) {
1784 SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
1785 CCVal, CCR, InFlag);
1787 Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
1788 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1792 SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
1793 CCVal, CCR, InFlag);
1795 Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
1796 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1801 // Pattern: (ARMcmov:i32 GPR:i32:$false,
1802 // (imm:i32)<<P:Predicate_so_imm>>:$true,
1804 // Emits: (MOVCCi:i32 GPR:i32:$false,
1805 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1806 // Pattern complexity = 10 cost = 1 size = 0
1807 if (Subtarget->isThumb()) {
1808 SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
1809 CCVal, CCR, InFlag);
1811 Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
1812 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1816 SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
1817 CCVal, CCR, InFlag);
1819 Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
1820 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1826 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1827 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1828 // Pattern complexity = 6 cost = 1 size = 0
1830 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1831 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1832 // Pattern complexity = 6 cost = 11 size = 0
1834 // Also FCPYScc and FCPYDcc.
1835 SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
1836 SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
1838 switch (VT.getSimpleVT().SimpleTy) {
1839 default: assert(false && "Illegal conditional move type!");
1842 Opc = Subtarget->isThumb()
1843 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1853 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1856 SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
1857 // The only time a CONCAT_VECTORS operation can have legal types is when
1858 // two 64-bit vectors are concatenated to a 128-bit vector.
1859 EVT VT = N->getValueType(0);
1860 if (!VT.is128BitVector() || N->getNumOperands() != 2)
1861 llvm_unreachable("unexpected CONCAT_VECTORS");
1862 DebugLoc dl = N->getDebugLoc();
1863 SDValue V0 = N->getOperand(0);
1864 SDValue V1 = N->getOperand(1);
1865 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1866 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1867 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
1868 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
1871 SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
1872 DebugLoc dl = N->getDebugLoc();
1874 if (N->isMachineOpcode())
1875 return NULL; // Already selected.
1877 switch (N->getOpcode()) {
1879 case ISD::Constant: {
1880 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1882 if (Subtarget->hasThumb2())
1883 // Thumb2-aware targets have the MOVT instruction, so all immediates can
1884 // be done with MOV + MOVT, at worst.
1887 if (Subtarget->isThumb()) {
1888 UseCP = (Val > 255 && // MOV
1889 ~Val > 255 && // MOV + MVN
1890 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
1892 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
1893 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
1894 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
1899 CurDAG->getTargetConstantPool(ConstantInt::get(
1900 Type::getInt32Ty(*CurDAG->getContext()), Val),
1901 TLI.getPointerTy());
1904 if (Subtarget->isThumb1Only()) {
1905 SDValue Pred = getAL(CurDAG);
1906 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1907 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1908 ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1913 CurDAG->getRegister(0, MVT::i32),
1914 CurDAG->getTargetConstant(0, MVT::i32),
1916 CurDAG->getRegister(0, MVT::i32),
1917 CurDAG->getEntryNode()
1919 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1922 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
1926 // Other cases are autogenerated.
1929 case ISD::FrameIndex: {
1930 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1931 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1932 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1933 if (Subtarget->isThumb1Only()) {
1934 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1935 CurDAG->getTargetConstant(0, MVT::i32));
1937 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1938 ARM::t2ADDri : ARM::ADDri);
1939 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1940 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1941 CurDAG->getRegister(0, MVT::i32) };
1942 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1946 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1950 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true))
1954 if (Subtarget->isThumb1Only())
1956 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1957 unsigned RHSV = C->getZExtValue();
1959 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
1960 unsigned ShImm = Log2_32(RHSV-1);
1963 SDValue V = N->getOperand(0);
1964 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1965 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1966 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1967 if (Subtarget->isThumb()) {
1968 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1969 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1971 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1972 return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1975 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
1976 unsigned ShImm = Log2_32(RHSV+1);
1979 SDValue V = N->getOperand(0);
1980 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1981 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1982 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1983 if (Subtarget->isThumb()) {
1984 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1985 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 6);
1987 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1988 return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1994 // Check for unsigned bitfield extract
1995 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1998 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
1999 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
2000 // are entirely contributed by c2 and lower 16-bits are entirely contributed
2001 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
2002 // Select it to: "movt x, ((c1 & 0xffff) >> 16)
2003 EVT VT = N->getValueType(0);
2006 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
2008 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
2011 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2012 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2015 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
2016 SDValue N2 = N0.getOperand(1);
2017 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
2020 unsigned N1CVal = N1C->getZExtValue();
2021 unsigned N2CVal = N2C->getZExtValue();
2022 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
2023 (N1CVal & 0xffffU) == 0xffffU &&
2024 (N2CVal & 0xffffU) == 0x0U) {
2025 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
2027 SDValue Ops[] = { N0.getOperand(0), Imm16,
2028 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2029 return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
2034 case ARMISD::VMOVRRD:
2035 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
2036 N->getOperand(0), getAL(CurDAG),
2037 CurDAG->getRegister(0, MVT::i32));
2038 case ISD::UMUL_LOHI: {
2039 if (Subtarget->isThumb1Only())
2041 if (Subtarget->isThumb()) {
2042 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2043 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2044 CurDAG->getRegister(0, MVT::i32) };
2045 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32,Ops,4);
2047 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2048 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2049 CurDAG->getRegister(0, MVT::i32) };
2050 return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
2053 case ISD::SMUL_LOHI: {
2054 if (Subtarget->isThumb1Only())
2056 if (Subtarget->isThumb()) {
2057 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2058 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2059 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32,Ops,4);
2061 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2062 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2063 CurDAG->getRegister(0, MVT::i32) };
2064 return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
2068 SDNode *ResNode = 0;
2069 if (Subtarget->isThumb() && Subtarget->hasThumb2())
2070 ResNode = SelectT2IndexedLoad(N);
2072 ResNode = SelectARMIndexedLoad(N);
2076 // VLDMQ must be custom-selected for "v2f64 load" to set the AM5Opc value.
2077 if (Subtarget->hasVFP2() &&
2078 N->getValueType(0).getSimpleVT().SimpleTy == MVT::v2f64) {
2079 SDValue Chain = N->getOperand(0);
2081 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
2082 SDValue Pred = getAL(CurDAG);
2083 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2084 SDValue Ops[] = { N->getOperand(1), AM5Opc, Pred, PredReg, Chain };
2085 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2086 MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
2087 SDNode *Ret = CurDAG->getMachineNode(ARM::VLDMQ, dl,
2088 MVT::v2f64, MVT::Other, Ops, 5);
2089 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
2092 // Other cases are autogenerated.
2096 // VSTMQ must be custom-selected for "v2f64 store" to set the AM5Opc value.
2097 if (Subtarget->hasVFP2() &&
2098 N->getOperand(1).getValueType().getSimpleVT().SimpleTy == MVT::v2f64) {
2099 SDValue Chain = N->getOperand(0);
2101 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
2102 SDValue Pred = getAL(CurDAG);
2103 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2104 SDValue Ops[] = { N->getOperand(1), N->getOperand(2),
2105 AM5Opc, Pred, PredReg, Chain };
2106 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2107 MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
2108 SDNode *Ret = CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
2109 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
2112 // Other cases are autogenerated.
2115 case ARMISD::BRCOND: {
2116 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2117 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2118 // Pattern complexity = 6 cost = 1 size = 0
2120 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2121 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
2122 // Pattern complexity = 6 cost = 1 size = 0
2124 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2125 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2126 // Pattern complexity = 6 cost = 1 size = 0
2128 unsigned Opc = Subtarget->isThumb() ?
2129 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
2130 SDValue Chain = N->getOperand(0);
2131 SDValue N1 = N->getOperand(1);
2132 SDValue N2 = N->getOperand(2);
2133 SDValue N3 = N->getOperand(3);
2134 SDValue InFlag = N->getOperand(4);
2135 assert(N1.getOpcode() == ISD::BasicBlock);
2136 assert(N2.getOpcode() == ISD::Constant);
2137 assert(N3.getOpcode() == ISD::Register);
2139 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2140 cast<ConstantSDNode>(N2)->getZExtValue()),
2142 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
2143 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
2145 Chain = SDValue(ResNode, 0);
2146 if (N->getNumValues() == 2) {
2147 InFlag = SDValue(ResNode, 1);
2148 ReplaceUses(SDValue(N, 1), InFlag);
2150 ReplaceUses(SDValue(N, 0),
2151 SDValue(Chain.getNode(), Chain.getResNo()));
2155 return SelectCMOVOp(N);
2156 case ARMISD::CNEG: {
2157 EVT VT = N->getValueType(0);
2158 SDValue N0 = N->getOperand(0);
2159 SDValue N1 = N->getOperand(1);
2160 SDValue N2 = N->getOperand(2);
2161 SDValue N3 = N->getOperand(3);
2162 SDValue InFlag = N->getOperand(4);
2163 assert(N2.getOpcode() == ISD::Constant);
2164 assert(N3.getOpcode() == ISD::Register);
2166 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2167 cast<ConstantSDNode>(N2)->getZExtValue()),
2169 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
2171 switch (VT.getSimpleVT().SimpleTy) {
2172 default: assert(false && "Illegal conditional move type!");
2181 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
2184 case ARMISD::VZIP: {
2186 EVT VT = N->getValueType(0);
2187 switch (VT.getSimpleVT().SimpleTy) {
2188 default: return NULL;
2189 case MVT::v8i8: Opc = ARM::VZIPd8; break;
2190 case MVT::v4i16: Opc = ARM::VZIPd16; break;
2192 case MVT::v2i32: Opc = ARM::VZIPd32; break;
2193 case MVT::v16i8: Opc = ARM::VZIPq8; break;
2194 case MVT::v8i16: Opc = ARM::VZIPq16; break;
2196 case MVT::v4i32: Opc = ARM::VZIPq32; break;
2198 SDValue Pred = getAL(CurDAG);
2199 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2200 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2201 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2203 case ARMISD::VUZP: {
2205 EVT VT = N->getValueType(0);
2206 switch (VT.getSimpleVT().SimpleTy) {
2207 default: return NULL;
2208 case MVT::v8i8: Opc = ARM::VUZPd8; break;
2209 case MVT::v4i16: Opc = ARM::VUZPd16; break;
2211 case MVT::v2i32: Opc = ARM::VUZPd32; break;
2212 case MVT::v16i8: Opc = ARM::VUZPq8; break;
2213 case MVT::v8i16: Opc = ARM::VUZPq16; break;
2215 case MVT::v4i32: Opc = ARM::VUZPq32; break;
2217 SDValue Pred = getAL(CurDAG);
2218 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2219 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2220 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2222 case ARMISD::VTRN: {
2224 EVT VT = N->getValueType(0);
2225 switch (VT.getSimpleVT().SimpleTy) {
2226 default: return NULL;
2227 case MVT::v8i8: Opc = ARM::VTRNd8; break;
2228 case MVT::v4i16: Opc = ARM::VTRNd16; break;
2230 case MVT::v2i32: Opc = ARM::VTRNd32; break;
2231 case MVT::v16i8: Opc = ARM::VTRNq8; break;
2232 case MVT::v8i16: Opc = ARM::VTRNq16; break;
2234 case MVT::v4i32: Opc = ARM::VTRNq32; break;
2236 SDValue Pred = getAL(CurDAG);
2237 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2238 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2239 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2241 case ARMISD::BUILD_VECTOR: {
2242 EVT VecVT = N->getValueType(0);
2243 EVT EltVT = VecVT.getVectorElementType();
2244 unsigned NumElts = VecVT.getVectorNumElements();
2245 if (EltVT.getSimpleVT() == MVT::f64) {
2246 assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
2247 return PairDRegs(VecVT, N->getOperand(0), N->getOperand(1));
2249 assert(EltVT.getSimpleVT() == MVT::f32 &&
2250 "unexpected type for BUILD_VECTOR");
2252 return PairSRegs(VecVT, N->getOperand(0), N->getOperand(1));
2253 assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
2254 return QuadSRegs(VecVT, N->getOperand(0), N->getOperand(1),
2255 N->getOperand(2), N->getOperand(3));
2258 case ISD::INTRINSIC_VOID:
2259 case ISD::INTRINSIC_W_CHAIN: {
2260 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2265 case Intrinsic::arm_neon_vld1: {
2266 unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
2267 ARM::VLD1d32, ARM::VLD1d64 };
2268 unsigned QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
2269 ARM::VLD1q32, ARM::VLD1q64 };
2270 return SelectVLD(N, 1, DOpcodes, QOpcodes, 0);
2273 case Intrinsic::arm_neon_vld2: {
2274 unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
2275 ARM::VLD2d32, ARM::VLD1q64 };
2276 unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
2277 return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
2280 case Intrinsic::arm_neon_vld3: {
2281 unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
2282 ARM::VLD3d32, ARM::VLD1d64T };
2283 unsigned QOpcodes0[] = { ARM::VLD3q8_UPD,
2286 unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD,
2287 ARM::VLD3q16odd_UPD,
2288 ARM::VLD3q32odd_UPD };
2289 return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2292 case Intrinsic::arm_neon_vld4: {
2293 unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
2294 ARM::VLD4d32, ARM::VLD1d64Q };
2295 unsigned QOpcodes0[] = { ARM::VLD4q8_UPD,
2298 unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD,
2299 ARM::VLD4q16odd_UPD,
2300 ARM::VLD4q32odd_UPD };
2301 return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2304 case Intrinsic::arm_neon_vld2lane: {
2305 unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
2306 unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 };
2307 unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd };
2308 return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
2311 case Intrinsic::arm_neon_vld3lane: {
2312 unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
2313 unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 };
2314 unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd };
2315 return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
2318 case Intrinsic::arm_neon_vld4lane: {
2319 unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
2320 unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 };
2321 unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd };
2322 return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
2325 case Intrinsic::arm_neon_vst1: {
2326 unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
2327 ARM::VST1d32, ARM::VST1d64 };
2328 unsigned QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
2329 ARM::VST1q32, ARM::VST1q64 };
2330 return SelectVST(N, 1, DOpcodes, QOpcodes, 0);
2333 case Intrinsic::arm_neon_vst2: {
2334 unsigned DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
2335 ARM::VST2d32, ARM::VST1q64 };
2336 unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
2337 return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
2340 case Intrinsic::arm_neon_vst3: {
2341 unsigned DOpcodes[] = { ARM::VST3d8, ARM::VST3d16,
2342 ARM::VST3d32, ARM::VST1d64T };
2343 unsigned QOpcodes0[] = { ARM::VST3q8_UPD,
2346 unsigned QOpcodes1[] = { ARM::VST3q8odd_UPD,
2347 ARM::VST3q16odd_UPD,
2348 ARM::VST3q32odd_UPD };
2349 return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2352 case Intrinsic::arm_neon_vst4: {
2353 unsigned DOpcodes[] = { ARM::VST4d8, ARM::VST4d16,
2354 ARM::VST4d32, ARM::VST1d64Q };
2355 unsigned QOpcodes0[] = { ARM::VST4q8_UPD,
2358 unsigned QOpcodes1[] = { ARM::VST4q8odd_UPD,
2359 ARM::VST4q16odd_UPD,
2360 ARM::VST4q32odd_UPD };
2361 return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2364 case Intrinsic::arm_neon_vst2lane: {
2365 unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
2366 unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 };
2367 unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd };
2368 return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
2371 case Intrinsic::arm_neon_vst3lane: {
2372 unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
2373 unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 };
2374 unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd };
2375 return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
2378 case Intrinsic::arm_neon_vst4lane: {
2379 unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
2380 unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 };
2381 unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd };
2382 return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
2388 case ISD::CONCAT_VECTORS:
2389 return SelectConcatVector(N);
2392 return SelectCode(N);
2395 bool ARMDAGToDAGISel::
2396 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2397 std::vector<SDValue> &OutOps) {
2398 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
2399 // Require the address to be in a register. That is safe for all ARM
2400 // variants and it is hard to do anything much smarter without knowing
2401 // how the operand is used.
2402 OutOps.push_back(Op);
2406 /// createARMISelDag - This pass converts a legalized DAG into a
2407 /// ARM-specific DAG, ready for instruction scheduling.
2409 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
2410 CodeGenOpt::Level OptLevel) {
2411 return new ARMDAGToDAGISel(TM, OptLevel);
2414 /// ModelWithRegSequence - Return true if isel should use REG_SEQUENCE to model
2415 /// operations involving sub-registers.
2416 bool llvm::ModelWithRegSequence() {