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 //===----------------------------------------------------------------------===//
15 #include "ARMAddressingModes.h"
16 #include "ARMTargetMachine.h"
17 #include "llvm/CallingConv.h"
18 #include "llvm/Constants.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Function.h"
21 #include "llvm/Intrinsics.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/SelectionDAG.h"
27 #include "llvm/CodeGen/SelectionDAGISel.h"
28 #include "llvm/Target/TargetLowering.h"
29 #include "llvm/Target/TargetOptions.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Compiler.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
39 UseRegSeq("neon-reg-sequence", cl::Hidden,
40 cl::desc("Use reg_sequence to model ld / st of multiple neon regs"));
42 //===--------------------------------------------------------------------===//
43 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
44 /// instructions for SelectionDAG operations.
47 class ARMDAGToDAGISel : public SelectionDAGISel {
48 ARMBaseTargetMachine &TM;
50 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
51 /// make the right decision when generating code for different targets.
52 const ARMSubtarget *Subtarget;
55 explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
56 CodeGenOpt::Level OptLevel)
57 : SelectionDAGISel(tm, OptLevel), TM(tm),
58 Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
61 virtual const char *getPassName() const {
62 return "ARM Instruction Selection";
65 /// getI32Imm - Return a target constant of type i32 with the specified
67 inline SDValue getI32Imm(unsigned Imm) {
68 return CurDAG->getTargetConstant(Imm, MVT::i32);
71 SDNode *Select(SDNode *N);
73 bool SelectShifterOperandReg(SDNode *Op, SDValue N, SDValue &A,
74 SDValue &B, SDValue &C);
75 bool SelectAddrMode2(SDNode *Op, SDValue N, SDValue &Base,
76 SDValue &Offset, SDValue &Opc);
77 bool SelectAddrMode2Offset(SDNode *Op, SDValue N,
78 SDValue &Offset, SDValue &Opc);
79 bool SelectAddrMode3(SDNode *Op, SDValue N, SDValue &Base,
80 SDValue &Offset, SDValue &Opc);
81 bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
82 SDValue &Offset, SDValue &Opc);
83 bool SelectAddrMode4(SDNode *Op, SDValue N, SDValue &Addr,
85 bool SelectAddrMode5(SDNode *Op, SDValue N, SDValue &Base,
87 bool SelectAddrMode6(SDNode *Op, SDValue N, SDValue &Addr, SDValue &Align);
89 bool SelectAddrModePC(SDNode *Op, SDValue N, SDValue &Offset,
92 bool SelectThumbAddrModeRR(SDNode *Op, SDValue N, SDValue &Base,
94 bool SelectThumbAddrModeRI5(SDNode *Op, SDValue N, unsigned Scale,
95 SDValue &Base, SDValue &OffImm,
97 bool SelectThumbAddrModeS1(SDNode *Op, SDValue N, SDValue &Base,
98 SDValue &OffImm, SDValue &Offset);
99 bool SelectThumbAddrModeS2(SDNode *Op, SDValue N, SDValue &Base,
100 SDValue &OffImm, SDValue &Offset);
101 bool SelectThumbAddrModeS4(SDNode *Op, SDValue N, SDValue &Base,
102 SDValue &OffImm, SDValue &Offset);
103 bool SelectThumbAddrModeSP(SDNode *Op, SDValue N, SDValue &Base,
106 bool SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
107 SDValue &BaseReg, SDValue &Opc);
108 bool SelectT2AddrModeImm12(SDNode *Op, SDValue N, SDValue &Base,
110 bool SelectT2AddrModeImm8(SDNode *Op, SDValue N, SDValue &Base,
112 bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
114 bool SelectT2AddrModeImm8s4(SDNode *Op, SDValue N, SDValue &Base,
116 bool SelectT2AddrModeSoReg(SDNode *Op, SDValue N, SDValue &Base,
117 SDValue &OffReg, SDValue &ShImm);
119 // Include the pieces autogenerated from the target description.
120 #include "ARMGenDAGISel.inc"
123 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
125 SDNode *SelectARMIndexedLoad(SDNode *N);
126 SDNode *SelectT2IndexedLoad(SDNode *N);
128 /// SelectVLD - Select NEON load intrinsics. NumVecs should be
129 /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for
130 /// loads of D registers and even subregs and odd subregs of Q registers.
131 /// For NumVecs <= 2, QOpcodes1 is not used.
132 SDNode *SelectVLD(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
133 unsigned *QOpcodes0, unsigned *QOpcodes1);
135 /// SelectVST - Select NEON store intrinsics. NumVecs should
136 /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for
137 /// stores of D registers and even subregs and odd subregs of Q registers.
138 /// For NumVecs <= 2, QOpcodes1 is not used.
139 SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
140 unsigned *QOpcodes0, unsigned *QOpcodes1);
142 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
143 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
144 /// load/store of D registers and even subregs and odd subregs of Q registers.
145 SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, unsigned NumVecs,
146 unsigned *DOpcodes, unsigned *QOpcodes0,
147 unsigned *QOpcodes1);
149 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
150 SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
152 /// SelectCMOVOp - Select CMOV instructions for ARM.
153 SDNode *SelectCMOVOp(SDNode *N);
154 SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
155 ARMCC::CondCodes CCVal, SDValue CCR,
157 SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
158 ARMCC::CondCodes CCVal, SDValue CCR,
160 SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
161 ARMCC::CondCodes CCVal, SDValue CCR,
163 SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
164 ARMCC::CondCodes CCVal, SDValue CCR,
167 SDNode *SelectConcatVector(SDNode *N);
169 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
170 /// inline asm expressions.
171 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
173 std::vector<SDValue> &OutOps);
175 /// PairDRegs - Form a quad register from a pair of D registers.
177 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
179 /// PairDRegs - Form a quad register pair from a pair of Q registers.
181 SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1);
183 /// QuadDRegs - Form a quad register pair from a quad of D registers.
185 SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
187 /// OctoDRegs - Form 8 consecutive D registers.
189 SDNode *OctoDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3,
190 SDValue V4, SDValue V5, SDValue V6, SDValue V7);
194 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
195 /// operand. If so Imm will receive the 32-bit value.
196 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
197 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
198 Imm = cast<ConstantSDNode>(N)->getZExtValue();
204 // isInt32Immediate - This method tests to see if a constant operand.
205 // If so Imm will receive the 32 bit value.
206 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
207 return isInt32Immediate(N.getNode(), Imm);
210 // isOpcWithIntImmediate - This method tests to see if the node is a specific
211 // opcode and that it has a immediate integer right operand.
212 // If so Imm will receive the 32 bit value.
213 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
214 return N->getOpcode() == Opc &&
215 isInt32Immediate(N->getOperand(1).getNode(), Imm);
219 bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
224 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
226 // Don't match base register only case. That is matched to a separate
227 // lower complexity pattern with explicit register operand.
228 if (ShOpcVal == ARM_AM::no_shift) return false;
230 BaseReg = N.getOperand(0);
231 unsigned ShImmVal = 0;
232 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
233 ShReg = CurDAG->getRegister(0, MVT::i32);
234 ShImmVal = RHS->getZExtValue() & 31;
236 ShReg = N.getOperand(1);
238 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
243 bool ARMDAGToDAGISel::SelectAddrMode2(SDNode *Op, SDValue N,
244 SDValue &Base, SDValue &Offset,
246 if (N.getOpcode() == ISD::MUL) {
247 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
248 // X * [3,5,9] -> X + X * [2,4,8] etc.
249 int RHSC = (int)RHS->getZExtValue();
252 ARM_AM::AddrOpc AddSub = ARM_AM::add;
254 AddSub = ARM_AM::sub;
257 if (isPowerOf2_32(RHSC)) {
258 unsigned ShAmt = Log2_32(RHSC);
259 Base = Offset = N.getOperand(0);
260 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
269 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
271 if (N.getOpcode() == ISD::FrameIndex) {
272 int FI = cast<FrameIndexSDNode>(N)->getIndex();
273 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
274 } else if (N.getOpcode() == ARMISD::Wrapper &&
275 !(Subtarget->useMovt() &&
276 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
277 Base = N.getOperand(0);
279 Offset = CurDAG->getRegister(0, MVT::i32);
280 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
286 // Match simple R +/- imm12 operands.
287 if (N.getOpcode() == ISD::ADD)
288 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
289 int RHSC = (int)RHS->getZExtValue();
290 if ((RHSC >= 0 && RHSC < 0x1000) ||
291 (RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
292 Base = N.getOperand(0);
293 if (Base.getOpcode() == ISD::FrameIndex) {
294 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
295 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
297 Offset = CurDAG->getRegister(0, MVT::i32);
299 ARM_AM::AddrOpc AddSub = ARM_AM::add;
301 AddSub = ARM_AM::sub;
304 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
311 // Otherwise this is R +/- [possibly shifted] R.
312 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
313 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
316 Base = N.getOperand(0);
317 Offset = N.getOperand(1);
319 if (ShOpcVal != ARM_AM::no_shift) {
320 // Check to see if the RHS of the shift is a constant, if not, we can't fold
322 if (ConstantSDNode *Sh =
323 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
324 ShAmt = Sh->getZExtValue();
325 Offset = N.getOperand(1).getOperand(0);
327 ShOpcVal = ARM_AM::no_shift;
331 // Try matching (R shl C) + (R).
332 if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
333 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
334 if (ShOpcVal != ARM_AM::no_shift) {
335 // Check to see if the RHS of the shift is a constant, if not, we can't
337 if (ConstantSDNode *Sh =
338 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
339 ShAmt = Sh->getZExtValue();
340 Offset = N.getOperand(0).getOperand(0);
341 Base = N.getOperand(1);
343 ShOpcVal = ARM_AM::no_shift;
348 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
353 bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
354 SDValue &Offset, SDValue &Opc) {
355 unsigned Opcode = Op->getOpcode();
356 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
357 ? cast<LoadSDNode>(Op)->getAddressingMode()
358 : cast<StoreSDNode>(Op)->getAddressingMode();
359 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
360 ? ARM_AM::add : ARM_AM::sub;
361 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
362 int Val = (int)C->getZExtValue();
363 if (Val >= 0 && Val < 0x1000) { // 12 bits.
364 Offset = CurDAG->getRegister(0, MVT::i32);
365 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
373 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
375 if (ShOpcVal != ARM_AM::no_shift) {
376 // Check to see if the RHS of the shift is a constant, if not, we can't fold
378 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
379 ShAmt = Sh->getZExtValue();
380 Offset = N.getOperand(0);
382 ShOpcVal = ARM_AM::no_shift;
386 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
392 bool ARMDAGToDAGISel::SelectAddrMode3(SDNode *Op, SDValue N,
393 SDValue &Base, SDValue &Offset,
395 if (N.getOpcode() == ISD::SUB) {
396 // X - C is canonicalize to X + -C, no need to handle it here.
397 Base = N.getOperand(0);
398 Offset = N.getOperand(1);
399 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
403 if (N.getOpcode() != ISD::ADD) {
405 if (N.getOpcode() == ISD::FrameIndex) {
406 int FI = cast<FrameIndexSDNode>(N)->getIndex();
407 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
409 Offset = CurDAG->getRegister(0, MVT::i32);
410 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
414 // If the RHS is +/- imm8, fold into addr mode.
415 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
416 int RHSC = (int)RHS->getZExtValue();
417 if ((RHSC >= 0 && RHSC < 256) ||
418 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
419 Base = N.getOperand(0);
420 if (Base.getOpcode() == ISD::FrameIndex) {
421 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
422 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
424 Offset = CurDAG->getRegister(0, MVT::i32);
426 ARM_AM::AddrOpc AddSub = ARM_AM::add;
428 AddSub = ARM_AM::sub;
431 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
436 Base = N.getOperand(0);
437 Offset = N.getOperand(1);
438 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
442 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
443 SDValue &Offset, SDValue &Opc) {
444 unsigned Opcode = Op->getOpcode();
445 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
446 ? cast<LoadSDNode>(Op)->getAddressingMode()
447 : cast<StoreSDNode>(Op)->getAddressingMode();
448 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
449 ? ARM_AM::add : ARM_AM::sub;
450 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
451 int Val = (int)C->getZExtValue();
452 if (Val >= 0 && Val < 256) {
453 Offset = CurDAG->getRegister(0, MVT::i32);
454 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
460 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
464 bool ARMDAGToDAGISel::SelectAddrMode4(SDNode *Op, SDValue N,
465 SDValue &Addr, SDValue &Mode) {
467 Mode = CurDAG->getTargetConstant(0, MVT::i32);
471 bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N,
472 SDValue &Base, SDValue &Offset) {
473 if (N.getOpcode() != ISD::ADD) {
475 if (N.getOpcode() == ISD::FrameIndex) {
476 int FI = cast<FrameIndexSDNode>(N)->getIndex();
477 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
478 } else if (N.getOpcode() == ARMISD::Wrapper &&
479 !(Subtarget->useMovt() &&
480 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
481 Base = N.getOperand(0);
483 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
488 // If the RHS is +/- imm8, fold into addr mode.
489 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
490 int RHSC = (int)RHS->getZExtValue();
491 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied by 4.
493 if ((RHSC >= 0 && RHSC < 256) ||
494 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
495 Base = N.getOperand(0);
496 if (Base.getOpcode() == ISD::FrameIndex) {
497 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
498 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
501 ARM_AM::AddrOpc AddSub = ARM_AM::add;
503 AddSub = ARM_AM::sub;
506 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
514 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
519 bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N,
520 SDValue &Addr, SDValue &Align) {
522 // Default to no alignment.
523 Align = CurDAG->getTargetConstant(0, MVT::i32);
527 bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
528 SDValue &Offset, SDValue &Label) {
529 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
530 Offset = N.getOperand(0);
531 SDValue N1 = N.getOperand(1);
532 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
539 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
540 SDValue &Base, SDValue &Offset){
541 // FIXME dl should come from the parent load or store, not the address
542 DebugLoc dl = Op->getDebugLoc();
543 if (N.getOpcode() != ISD::ADD) {
544 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
545 if (!NC || NC->getZExtValue() != 0)
552 Base = N.getOperand(0);
553 Offset = N.getOperand(1);
558 ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
559 unsigned Scale, SDValue &Base,
560 SDValue &OffImm, SDValue &Offset) {
562 SDValue TmpBase, TmpOffImm;
563 if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
564 return false; // We want to select tLDRspi / tSTRspi instead.
565 if (N.getOpcode() == ARMISD::Wrapper &&
566 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
567 return false; // We want to select tLDRpci instead.
570 if (N.getOpcode() != ISD::ADD) {
571 if (N.getOpcode() == ARMISD::Wrapper &&
572 !(Subtarget->useMovt() &&
573 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
574 Base = N.getOperand(0);
578 Offset = CurDAG->getRegister(0, MVT::i32);
579 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
583 // Thumb does not have [sp, r] address mode.
584 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
585 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
586 if ((LHSR && LHSR->getReg() == ARM::SP) ||
587 (RHSR && RHSR->getReg() == ARM::SP)) {
589 Offset = CurDAG->getRegister(0, MVT::i32);
590 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
594 // If the RHS is + imm5 * scale, fold into addr mode.
595 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
596 int RHSC = (int)RHS->getZExtValue();
597 if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
599 if (RHSC >= 0 && RHSC < 32) {
600 Base = N.getOperand(0);
601 Offset = CurDAG->getRegister(0, MVT::i32);
602 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
608 Base = N.getOperand(0);
609 Offset = N.getOperand(1);
610 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
614 bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
615 SDValue &Base, SDValue &OffImm,
617 return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
620 bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
621 SDValue &Base, SDValue &OffImm,
623 return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
626 bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
627 SDValue &Base, SDValue &OffImm,
629 return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
632 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
633 SDValue &Base, SDValue &OffImm) {
634 if (N.getOpcode() == ISD::FrameIndex) {
635 int FI = cast<FrameIndexSDNode>(N)->getIndex();
636 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
637 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
641 if (N.getOpcode() != ISD::ADD)
644 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
645 if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
646 (LHSR && LHSR->getReg() == ARM::SP)) {
647 // If the RHS is + imm8 * scale, fold into addr mode.
648 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
649 int RHSC = (int)RHS->getZExtValue();
650 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
652 if (RHSC >= 0 && RHSC < 256) {
653 Base = N.getOperand(0);
654 if (Base.getOpcode() == ISD::FrameIndex) {
655 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
656 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
658 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
668 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
671 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
673 // Don't match base register only case. That is matched to a separate
674 // lower complexity pattern with explicit register operand.
675 if (ShOpcVal == ARM_AM::no_shift) return false;
677 BaseReg = N.getOperand(0);
678 unsigned ShImmVal = 0;
679 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
680 ShImmVal = RHS->getZExtValue() & 31;
681 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
688 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
689 SDValue &Base, SDValue &OffImm) {
690 // Match simple R + imm12 operands.
693 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
694 if (N.getOpcode() == ISD::FrameIndex) {
695 // Match frame index...
696 int FI = cast<FrameIndexSDNode>(N)->getIndex();
697 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
698 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
700 } else if (N.getOpcode() == ARMISD::Wrapper &&
701 !(Subtarget->useMovt() &&
702 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
703 Base = N.getOperand(0);
704 if (Base.getOpcode() == ISD::TargetConstantPool)
705 return false; // We want to select t2LDRpci instead.
708 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
712 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
713 if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
714 // Let t2LDRi8 handle (R - imm8).
717 int RHSC = (int)RHS->getZExtValue();
718 if (N.getOpcode() == ISD::SUB)
721 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
722 Base = N.getOperand(0);
723 if (Base.getOpcode() == ISD::FrameIndex) {
724 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
725 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
727 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
734 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
738 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
739 SDValue &Base, SDValue &OffImm) {
740 // Match simple R - imm8 operands.
741 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
742 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
743 int RHSC = (int)RHS->getSExtValue();
744 if (N.getOpcode() == ISD::SUB)
747 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
748 Base = N.getOperand(0);
749 if (Base.getOpcode() == ISD::FrameIndex) {
750 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
751 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
753 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
762 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
764 unsigned Opcode = Op->getOpcode();
765 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
766 ? cast<LoadSDNode>(Op)->getAddressingMode()
767 : cast<StoreSDNode>(Op)->getAddressingMode();
768 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
769 int RHSC = (int)RHS->getZExtValue();
770 if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
771 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
772 ? CurDAG->getTargetConstant(RHSC, MVT::i32)
773 : CurDAG->getTargetConstant(-RHSC, MVT::i32);
781 bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
782 SDValue &Base, SDValue &OffImm) {
783 if (N.getOpcode() == ISD::ADD) {
784 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
785 int RHSC = (int)RHS->getZExtValue();
786 if (((RHSC & 0x3) == 0) &&
787 ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits.
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();
796 if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) { // 8 bits.
797 Base = N.getOperand(0);
798 OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
807 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDNode *Op, SDValue N,
809 SDValue &OffReg, SDValue &ShImm) {
810 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
811 if (N.getOpcode() != ISD::ADD)
814 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
815 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
816 int RHSC = (int)RHS->getZExtValue();
817 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
819 else if (RHSC < 0 && RHSC >= -255) // 8 bits
823 // Look for (R + R) or (R + (R << [1,2,3])).
825 Base = N.getOperand(0);
826 OffReg = N.getOperand(1);
828 // Swap if it is ((R << c) + R).
829 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
830 if (ShOpcVal != ARM_AM::lsl) {
831 ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
832 if (ShOpcVal == ARM_AM::lsl)
833 std::swap(Base, OffReg);
836 if (ShOpcVal == ARM_AM::lsl) {
837 // Check to see if the RHS of the shift is a constant, if not, we can't fold
839 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
840 ShAmt = Sh->getZExtValue();
843 ShOpcVal = ARM_AM::no_shift;
845 OffReg = OffReg.getOperand(0);
847 ShOpcVal = ARM_AM::no_shift;
851 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
856 //===--------------------------------------------------------------------===//
858 /// getAL - Returns a ARMCC::AL immediate node.
859 static inline SDValue getAL(SelectionDAG *CurDAG) {
860 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
863 SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
864 LoadSDNode *LD = cast<LoadSDNode>(N);
865 ISD::MemIndexedMode AM = LD->getAddressingMode();
866 if (AM == ISD::UNINDEXED)
869 EVT LoadedVT = LD->getMemoryVT();
870 SDValue Offset, AMOpc;
871 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
874 if (LoadedVT == MVT::i32 &&
875 SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
876 Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
878 } else if (LoadedVT == MVT::i16 &&
879 SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
881 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
882 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
883 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
884 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
885 if (LD->getExtensionType() == ISD::SEXTLOAD) {
886 if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
888 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
891 if (SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
893 Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
899 SDValue Chain = LD->getChain();
900 SDValue Base = LD->getBasePtr();
901 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
902 CurDAG->getRegister(0, MVT::i32), Chain };
903 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
910 SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
911 LoadSDNode *LD = cast<LoadSDNode>(N);
912 ISD::MemIndexedMode AM = LD->getAddressingMode();
913 if (AM == ISD::UNINDEXED)
916 EVT LoadedVT = LD->getMemoryVT();
917 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
919 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
922 if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
923 switch (LoadedVT.getSimpleVT().SimpleTy) {
925 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
929 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
931 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
936 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
938 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
947 SDValue Chain = LD->getChain();
948 SDValue Base = LD->getBasePtr();
949 SDValue Ops[]= { Base, Offset, getAL(CurDAG),
950 CurDAG->getRegister(0, MVT::i32), Chain };
951 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
958 /// PairDRegs - Form a quad register from a pair of D registers.
960 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
961 DebugLoc dl = V0.getNode()->getDebugLoc();
962 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
963 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
964 if (llvm::ModelWithRegSequence()) {
965 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
966 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
969 SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0);
970 SDNode *Pair = CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
971 VT, Undef, V0, SubReg0);
972 return CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
973 VT, SDValue(Pair, 0), V1, SubReg1);
976 /// PairQRegs - Form 4 consecutive D registers from a pair of Q registers.
978 SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) {
979 DebugLoc dl = V0.getNode()->getDebugLoc();
980 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::QSUBREG_0, MVT::i32);
981 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::QSUBREG_1, MVT::i32);
982 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
983 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
986 /// QuadDRegs - Form 4 consecutive D registers.
988 SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
989 SDValue V2, SDValue V3) {
990 DebugLoc dl = V0.getNode()->getDebugLoc();
991 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
992 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
993 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::DSUBREG_2, MVT::i32);
994 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::DSUBREG_3, MVT::i32);
995 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
996 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
999 /// OctoDRegs - Form 8 consecutive D registers.
1001 SDNode *ARMDAGToDAGISel::OctoDRegs(EVT VT, SDValue V0, SDValue V1,
1002 SDValue V2, SDValue V3,
1003 SDValue V4, SDValue V5,
1004 SDValue V6, SDValue V7) {
1005 DebugLoc dl = V0.getNode()->getDebugLoc();
1006 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
1007 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
1008 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::DSUBREG_2, MVT::i32);
1009 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::DSUBREG_3, MVT::i32);
1010 SDValue SubReg4 = CurDAG->getTargetConstant(ARM::DSUBREG_4, MVT::i32);
1011 SDValue SubReg5 = CurDAG->getTargetConstant(ARM::DSUBREG_5, MVT::i32);
1012 SDValue SubReg6 = CurDAG->getTargetConstant(ARM::DSUBREG_6, MVT::i32);
1013 SDValue SubReg7 = CurDAG->getTargetConstant(ARM::DSUBREG_7, MVT::i32);
1014 const SDValue Ops[] ={ V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3,
1015 V4, SubReg4, V5, SubReg5, V6, SubReg6, V7, SubReg7 };
1016 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 16);
1019 /// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
1020 /// for a 64-bit subregister of the vector.
1021 static EVT GetNEONSubregVT(EVT VT) {
1022 switch (VT.getSimpleVT().SimpleTy) {
1023 default: llvm_unreachable("unhandled NEON type");
1024 case MVT::v16i8: return MVT::v8i8;
1025 case MVT::v8i16: return MVT::v4i16;
1026 case MVT::v4f32: return MVT::v2f32;
1027 case MVT::v4i32: return MVT::v2i32;
1028 case MVT::v2i64: return MVT::v1i64;
1032 SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
1033 unsigned *DOpcodes, unsigned *QOpcodes0,
1034 unsigned *QOpcodes1) {
1035 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1036 DebugLoc dl = N->getDebugLoc();
1038 SDValue MemAddr, Align;
1039 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1042 SDValue Chain = N->getOperand(0);
1043 EVT VT = N->getValueType(0);
1044 bool is64BitVector = VT.is64BitVector();
1046 unsigned OpcodeIndex;
1047 switch (VT.getSimpleVT().SimpleTy) {
1048 default: llvm_unreachable("unhandled vld type");
1049 // Double-register operations:
1050 case MVT::v8i8: OpcodeIndex = 0; break;
1051 case MVT::v4i16: OpcodeIndex = 1; break;
1053 case MVT::v2i32: OpcodeIndex = 2; break;
1054 case MVT::v1i64: OpcodeIndex = 3; break;
1055 // Quad-register operations:
1056 case MVT::v16i8: OpcodeIndex = 0; break;
1057 case MVT::v8i16: OpcodeIndex = 1; break;
1059 case MVT::v4i32: OpcodeIndex = 2; break;
1060 case MVT::v2i64: OpcodeIndex = 3;
1061 assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
1065 SDValue Pred = getAL(CurDAG);
1066 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1067 if (is64BitVector) {
1068 unsigned Opc = DOpcodes[OpcodeIndex];
1069 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1070 std::vector<EVT> ResTys(NumVecs, VT);
1071 ResTys.push_back(MVT::Other);
1072 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1073 if (!llvm::ModelWithRegSequence() || NumVecs < 2)
1077 SDValue V0 = SDValue(VLd, 0);
1078 SDValue V1 = SDValue(VLd, 1);
1080 // Form a REG_SEQUENCE to force register allocation.
1082 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1084 SDValue V2 = SDValue(VLd, 2);
1085 // If it's a vld3, form a quad D-register but discard the last part.
1086 SDValue V3 = (NumVecs == 3)
1087 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1089 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1092 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1093 SDValue D = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0+Vec,
1095 ReplaceUses(SDValue(N, Vec), D);
1097 ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, NumVecs));
1101 EVT RegVT = GetNEONSubregVT(VT);
1103 // Quad registers are directly supported for VLD1 and VLD2,
1104 // loading pairs of D regs.
1105 unsigned Opc = QOpcodes0[OpcodeIndex];
1106 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1107 std::vector<EVT> ResTys(2 * NumVecs, RegVT);
1108 ResTys.push_back(MVT::Other);
1109 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1110 Chain = SDValue(VLd, 2 * NumVecs);
1112 // Combine the even and odd subregs to produce the result.
1113 if (llvm::ModelWithRegSequence()) {
1115 SDNode *Q = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
1116 ReplaceUses(SDValue(N, 0), SDValue(Q, 0));
1118 SDValue QQ = SDValue(QuadDRegs(MVT::v4i64,
1119 SDValue(VLd, 0), SDValue(VLd, 1),
1120 SDValue(VLd, 2), SDValue(VLd, 3)), 0);
1121 SDValue Q0 = CurDAG->getTargetExtractSubreg(ARM::QSUBREG_0, dl, VT, QQ);
1122 SDValue Q1 = CurDAG->getTargetExtractSubreg(ARM::QSUBREG_1, dl, VT, QQ);
1123 ReplaceUses(SDValue(N, 0), Q0);
1124 ReplaceUses(SDValue(N, 1), Q1);
1127 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1128 SDNode *Q = PairDRegs(VT, SDValue(VLd, 2*Vec), SDValue(VLd, 2*Vec+1));
1129 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1133 // Otherwise, quad registers are loaded with two separate instructions,
1134 // where one loads the even registers and the other loads the odd registers.
1136 std::vector<EVT> ResTys(NumVecs, RegVT);
1137 ResTys.push_back(MemAddr.getValueType());
1138 ResTys.push_back(MVT::Other);
1140 // Load the even subregs.
1141 unsigned Opc = QOpcodes0[OpcodeIndex];
1142 const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain };
1143 SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6);
1144 Chain = SDValue(VLdA, NumVecs+1);
1146 // Load the odd subregs.
1147 Opc = QOpcodes1[OpcodeIndex];
1148 const SDValue OpsB[] = { SDValue(VLdA, NumVecs),
1149 Align, Reg0, Pred, Reg0, Chain };
1150 SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
1151 Chain = SDValue(VLdB, NumVecs+1);
1153 if (llvm::ModelWithRegSequence()) {
1154 SDValue V0 = SDValue(VLdA, 0);
1155 SDValue V1 = SDValue(VLdB, 0);
1156 SDValue V2 = SDValue(VLdA, 1);
1157 SDValue V3 = SDValue(VLdB, 1);
1158 SDValue V4 = SDValue(VLdA, 2);
1159 SDValue V5 = SDValue(VLdB, 2);
1160 SDValue V6 = (NumVecs == 3)
1161 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT),
1164 SDValue V7 = (NumVecs == 3)
1165 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT),
1168 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V0, V1, V2, V3,
1169 V4, V5, V6, V7), 0);
1171 // Extract out the 3 / 4 Q registers.
1172 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1173 SDValue Q = CurDAG->getTargetExtractSubreg(ARM::QSUBREG_0+Vec,
1175 ReplaceUses(SDValue(N, Vec), Q);
1178 // Combine the even and odd subregs to produce the result.
1179 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1180 SDNode *Q = PairDRegs(VT, SDValue(VLdA, Vec), SDValue(VLdB, Vec));
1181 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1185 ReplaceUses(SDValue(N, NumVecs), Chain);
1189 SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
1190 unsigned *DOpcodes, unsigned *QOpcodes0,
1191 unsigned *QOpcodes1) {
1192 assert(NumVecs >=1 && NumVecs <= 4 && "VST NumVecs out-of-range");
1193 DebugLoc dl = N->getDebugLoc();
1195 SDValue MemAddr, Align;
1196 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1199 SDValue Chain = N->getOperand(0);
1200 EVT VT = N->getOperand(3).getValueType();
1201 bool is64BitVector = VT.is64BitVector();
1203 unsigned OpcodeIndex;
1204 switch (VT.getSimpleVT().SimpleTy) {
1205 default: llvm_unreachable("unhandled vst type");
1206 // Double-register operations:
1207 case MVT::v8i8: OpcodeIndex = 0; break;
1208 case MVT::v4i16: OpcodeIndex = 1; break;
1210 case MVT::v2i32: OpcodeIndex = 2; break;
1211 case MVT::v1i64: OpcodeIndex = 3; break;
1212 // Quad-register operations:
1213 case MVT::v16i8: OpcodeIndex = 0; break;
1214 case MVT::v8i16: OpcodeIndex = 1; break;
1216 case MVT::v4i32: OpcodeIndex = 2; break;
1217 case MVT::v2i64: OpcodeIndex = 3;
1218 assert(NumVecs == 1 && "v2i64 type only supported for VST1");
1222 SDValue Pred = getAL(CurDAG);
1223 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1225 SmallVector<SDValue, 10> Ops;
1226 Ops.push_back(MemAddr);
1227 Ops.push_back(Align);
1229 if (is64BitVector) {
1230 if (llvm::ModelWithRegSequence() && NumVecs >= 2) {
1232 SDValue V0 = N->getOperand(0+3);
1233 SDValue V1 = N->getOperand(1+3);
1235 // Form a REG_SEQUENCE to force register allocation.
1237 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1239 SDValue V2 = N->getOperand(2+3);
1240 // If it's a vld3, form a quad D-register and leave the last part as
1242 SDValue V3 = (NumVecs == 3)
1243 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1244 : N->getOperand(3+3);
1245 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1248 // Now extract the D registers back out.
1249 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, VT,
1251 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, VT,
1254 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_2, dl, VT,
1257 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_3, dl, VT,
1260 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1261 Ops.push_back(N->getOperand(Vec+3));
1263 Ops.push_back(Pred);
1264 Ops.push_back(Reg0); // predicate register
1265 Ops.push_back(Chain);
1266 unsigned Opc = DOpcodes[OpcodeIndex];
1267 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+5);
1270 EVT RegVT = GetNEONSubregVT(VT);
1272 // Quad registers are directly supported for VST1 and VST2,
1273 // storing pairs of D regs.
1274 unsigned Opc = QOpcodes0[OpcodeIndex];
1275 if (llvm::ModelWithRegSequence() && NumVecs == 2) {
1276 // First extract the pair of Q registers.
1277 SDValue Q0 = N->getOperand(3);
1278 SDValue Q1 = N->getOperand(4);
1280 // Form a QQ register.
1281 SDValue QQ = SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0);
1283 // Now extract the D registers back out.
1284 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1286 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1288 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_2, dl, RegVT,
1290 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_3, dl, RegVT,
1292 Ops.push_back(Pred);
1293 Ops.push_back(Reg0); // predicate register
1294 Ops.push_back(Chain);
1295 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 5 + 4);
1297 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1298 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1299 N->getOperand(Vec+3)));
1300 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1301 N->getOperand(Vec+3)));
1303 Ops.push_back(Pred);
1304 Ops.push_back(Reg0); // predicate register
1305 Ops.push_back(Chain);
1306 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
1311 // Otherwise, quad registers are stored with two separate instructions,
1312 // where one stores the even registers and the other stores the odd registers.
1313 if (llvm::ModelWithRegSequence()) {
1314 // Form the QQQQ REG_SEQUENCE.
1316 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1317 V[i] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1318 N->getOperand(Vec+3));
1319 V[i+1] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1320 N->getOperand(Vec+3));
1324 SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0);
1325 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1326 V[4], V[5], V[6], V[7]), 0);
1328 // Store the even D registers.
1329 Ops.push_back(Reg0); // post-access address offset
1330 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1331 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0+Vec*2, dl,
1333 Ops.push_back(Pred);
1334 Ops.push_back(Reg0); // predicate register
1335 Ops.push_back(Chain);
1336 unsigned Opc = QOpcodes0[OpcodeIndex];
1337 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1338 MVT::Other, Ops.data(), NumVecs+6);
1339 Chain = SDValue(VStA, 1);
1341 // Store the odd D registers.
1342 Ops[0] = SDValue(VStA, 0); // MemAddr
1343 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1344 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1+Vec*2, dl,
1346 Ops[NumVecs+5] = Chain;
1347 Opc = QOpcodes1[OpcodeIndex];
1348 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1349 MVT::Other, Ops.data(), NumVecs+6);
1350 Chain = SDValue(VStB, 1);
1351 ReplaceUses(SDValue(N, 0), Chain);
1354 Ops.push_back(Reg0); // post-access address offset
1356 // Store the even subregs.
1357 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1358 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1359 N->getOperand(Vec+3)));
1360 Ops.push_back(Pred);
1361 Ops.push_back(Reg0); // predicate register
1362 Ops.push_back(Chain);
1363 unsigned Opc = QOpcodes0[OpcodeIndex];
1364 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1365 MVT::Other, Ops.data(), NumVecs+6);
1366 Chain = SDValue(VStA, 1);
1368 // Store the odd subregs.
1369 Ops[0] = SDValue(VStA, 0); // MemAddr
1370 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1371 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1372 N->getOperand(Vec+3));
1373 Ops[NumVecs+5] = Chain;
1374 Opc = QOpcodes1[OpcodeIndex];
1375 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1376 MVT::Other, Ops.data(), NumVecs+6);
1377 Chain = SDValue(VStB, 1);
1378 ReplaceUses(SDValue(N, 0), Chain);
1383 SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1384 unsigned NumVecs, unsigned *DOpcodes,
1385 unsigned *QOpcodes0,
1386 unsigned *QOpcodes1) {
1387 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1388 DebugLoc dl = N->getDebugLoc();
1390 SDValue MemAddr, Align;
1391 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1394 SDValue Chain = N->getOperand(0);
1396 cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1397 EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
1398 bool is64BitVector = VT.is64BitVector();
1400 // Quad registers are handled by load/store of subregs. Find the subreg info.
1401 unsigned NumElts = 0;
1404 if (!is64BitVector) {
1405 RegVT = GetNEONSubregVT(VT);
1406 NumElts = RegVT.getVectorNumElements();
1407 SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
1410 unsigned OpcodeIndex;
1411 switch (VT.getSimpleVT().SimpleTy) {
1412 default: llvm_unreachable("unhandled vld/vst lane type");
1413 // Double-register operations:
1414 case MVT::v8i8: OpcodeIndex = 0; break;
1415 case MVT::v4i16: OpcodeIndex = 1; break;
1417 case MVT::v2i32: OpcodeIndex = 2; break;
1418 // Quad-register operations:
1419 case MVT::v8i16: OpcodeIndex = 0; break;
1421 case MVT::v4i32: OpcodeIndex = 1; break;
1424 SDValue Pred = getAL(CurDAG);
1425 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1427 SmallVector<SDValue, 10> Ops;
1428 Ops.push_back(MemAddr);
1429 Ops.push_back(Align);
1432 if (is64BitVector) {
1433 Opc = DOpcodes[OpcodeIndex];
1434 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1435 Ops.push_back(N->getOperand(Vec+3));
1437 // Check if this is loading the even or odd subreg of a Q register.
1438 if (Lane < NumElts) {
1439 Opc = QOpcodes0[OpcodeIndex];
1442 Opc = QOpcodes1[OpcodeIndex];
1444 // Extract the subregs of the input vector.
1445 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1446 Ops.push_back(CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1447 N->getOperand(Vec+3)));
1449 Ops.push_back(getI32Imm(Lane));
1450 Ops.push_back(Pred);
1451 Ops.push_back(Reg0);
1452 Ops.push_back(Chain);
1455 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6);
1457 std::vector<EVT> ResTys(NumVecs, RegVT);
1458 ResTys.push_back(MVT::Other);
1459 SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(),NumVecs+6);
1461 if (llvm::ModelWithRegSequence() && is64BitVector) {
1463 SDValue V0 = SDValue(VLdLn, 0);
1464 SDValue V1 = SDValue(VLdLn, 1);
1466 // Form a REG_SEQUENCE to force register allocation.
1468 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1470 SDValue V2 = SDValue(VLdLn, 2);
1471 // If it's a vld3, form a quad D-register but discard the last part.
1472 SDValue V3 = (NumVecs == 3)
1473 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1474 : SDValue(VLdLn, 3);
1475 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1478 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1479 SDValue D = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0+Vec, dl, VT,
1481 ReplaceUses(SDValue(N, Vec), D);
1483 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, NumVecs));
1487 // For a 64-bit vector load to D registers, nothing more needs to be done.
1491 // For 128-bit vectors, take the 64-bit results of the load and insert them
1492 // as subregs into the result.
1493 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1494 SDValue QuadVec = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
1495 N->getOperand(Vec+3),
1496 SDValue(VLdLn, Vec));
1497 ReplaceUses(SDValue(N, Vec), QuadVec);
1500 Chain = SDValue(VLdLn, NumVecs);
1501 ReplaceUses(SDValue(N, NumVecs), Chain);
1505 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
1507 if (!Subtarget->hasV6T2Ops())
1510 unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
1511 : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
1514 // For unsigned extracts, check for a shift right and mask
1515 unsigned And_imm = 0;
1516 if (N->getOpcode() == ISD::AND) {
1517 if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
1519 // The immediate is a mask of the low bits iff imm & (imm+1) == 0
1520 if (And_imm & (And_imm + 1))
1523 unsigned Srl_imm = 0;
1524 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
1526 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1528 unsigned Width = CountTrailingOnes_32(And_imm);
1529 unsigned LSB = Srl_imm;
1530 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1531 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1532 CurDAG->getTargetConstant(LSB, MVT::i32),
1533 CurDAG->getTargetConstant(Width, MVT::i32),
1534 getAL(CurDAG), Reg0 };
1535 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1541 // Otherwise, we're looking for a shift of a shift
1542 unsigned Shl_imm = 0;
1543 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
1544 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1545 unsigned Srl_imm = 0;
1546 if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
1547 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1548 unsigned Width = 32 - Srl_imm;
1549 int LSB = Srl_imm - Shl_imm;
1552 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1553 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1554 CurDAG->getTargetConstant(LSB, MVT::i32),
1555 CurDAG->getTargetConstant(Width, MVT::i32),
1556 getAL(CurDAG), Reg0 };
1557 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1563 SDNode *ARMDAGToDAGISel::
1564 SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1565 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1568 if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
1569 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1570 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1573 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1574 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1575 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1576 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1578 llvm_unreachable("Unknown so_reg opcode!");
1582 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1583 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1584 SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
1585 return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
1590 SDNode *ARMDAGToDAGISel::
1591 SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1592 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1596 if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
1597 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1598 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
1599 return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
1604 SDNode *ARMDAGToDAGISel::
1605 SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1606 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1607 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1611 if (Predicate_t2_so_imm(TrueVal.getNode())) {
1612 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1613 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1614 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1615 return CurDAG->SelectNodeTo(N,
1616 ARM::t2MOVCCi, MVT::i32, Ops, 5);
1621 SDNode *ARMDAGToDAGISel::
1622 SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1623 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1624 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1628 if (Predicate_so_imm(TrueVal.getNode())) {
1629 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1630 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1631 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1632 return CurDAG->SelectNodeTo(N,
1633 ARM::MOVCCi, MVT::i32, Ops, 5);
1638 SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
1639 EVT VT = N->getValueType(0);
1640 SDValue FalseVal = N->getOperand(0);
1641 SDValue TrueVal = N->getOperand(1);
1642 SDValue CC = N->getOperand(2);
1643 SDValue CCR = N->getOperand(3);
1644 SDValue InFlag = N->getOperand(4);
1645 assert(CC.getOpcode() == ISD::Constant);
1646 assert(CCR.getOpcode() == ISD::Register);
1647 ARMCC::CondCodes CCVal =
1648 (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
1650 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1651 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1652 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1653 // Pattern complexity = 18 cost = 1 size = 0
1657 if (Subtarget->isThumb()) {
1658 SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
1659 CCVal, CCR, InFlag);
1661 Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
1662 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1666 SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
1667 CCVal, CCR, InFlag);
1669 Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
1670 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1675 // Pattern: (ARMcmov:i32 GPR:i32:$false,
1676 // (imm:i32)<<P:Predicate_so_imm>>:$true,
1678 // Emits: (MOVCCi:i32 GPR:i32:$false,
1679 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1680 // Pattern complexity = 10 cost = 1 size = 0
1681 if (Subtarget->isThumb()) {
1682 SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
1683 CCVal, CCR, InFlag);
1685 Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
1686 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1690 SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
1691 CCVal, CCR, InFlag);
1693 Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
1694 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1700 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1701 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1702 // Pattern complexity = 6 cost = 1 size = 0
1704 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1705 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1706 // Pattern complexity = 6 cost = 11 size = 0
1708 // Also FCPYScc and FCPYDcc.
1709 SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
1710 SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
1712 switch (VT.getSimpleVT().SimpleTy) {
1713 default: assert(false && "Illegal conditional move type!");
1716 Opc = Subtarget->isThumb()
1717 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1727 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1730 SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
1731 // The only time a CONCAT_VECTORS operation can have legal types is when
1732 // two 64-bit vectors are concatenated to a 128-bit vector.
1733 EVT VT = N->getValueType(0);
1734 if (!VT.is128BitVector() || N->getNumOperands() != 2)
1735 llvm_unreachable("unexpected CONCAT_VECTORS");
1736 DebugLoc dl = N->getDebugLoc();
1737 SDValue V0 = N->getOperand(0);
1738 SDValue V1 = N->getOperand(1);
1739 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
1740 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
1741 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
1742 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
1745 SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
1746 DebugLoc dl = N->getDebugLoc();
1748 if (N->isMachineOpcode())
1749 return NULL; // Already selected.
1751 switch (N->getOpcode()) {
1753 case ISD::Constant: {
1754 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1756 if (Subtarget->hasThumb2())
1757 // Thumb2-aware targets have the MOVT instruction, so all immediates can
1758 // be done with MOV + MOVT, at worst.
1761 if (Subtarget->isThumb()) {
1762 UseCP = (Val > 255 && // MOV
1763 ~Val > 255 && // MOV + MVN
1764 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
1766 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
1767 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
1768 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
1773 CurDAG->getTargetConstantPool(ConstantInt::get(
1774 Type::getInt32Ty(*CurDAG->getContext()), Val),
1775 TLI.getPointerTy());
1778 if (Subtarget->isThumb1Only()) {
1779 SDValue Pred = getAL(CurDAG);
1780 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1781 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1782 ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1787 CurDAG->getRegister(0, MVT::i32),
1788 CurDAG->getTargetConstant(0, MVT::i32),
1790 CurDAG->getRegister(0, MVT::i32),
1791 CurDAG->getEntryNode()
1793 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1796 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
1800 // Other cases are autogenerated.
1803 case ISD::FrameIndex: {
1804 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1805 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1806 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1807 if (Subtarget->isThumb1Only()) {
1808 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1809 CurDAG->getTargetConstant(0, MVT::i32));
1811 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1812 ARM::t2ADDri : ARM::ADDri);
1813 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1814 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1815 CurDAG->getRegister(0, MVT::i32) };
1816 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1820 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1824 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true))
1828 if (Subtarget->isThumb1Only())
1830 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1831 unsigned RHSV = C->getZExtValue();
1833 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
1834 unsigned ShImm = Log2_32(RHSV-1);
1837 SDValue V = N->getOperand(0);
1838 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1839 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1840 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1841 if (Subtarget->isThumb()) {
1842 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1843 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1845 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1846 return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1849 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
1850 unsigned ShImm = Log2_32(RHSV+1);
1853 SDValue V = N->getOperand(0);
1854 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1855 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1856 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1857 if (Subtarget->isThumb()) {
1858 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0 };
1859 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 5);
1861 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1862 return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1868 // Check for unsigned bitfield extract
1869 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1872 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
1873 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
1874 // are entirely contributed by c2 and lower 16-bits are entirely contributed
1875 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
1876 // Select it to: "movt x, ((c1 & 0xffff) >> 16)
1877 EVT VT = N->getValueType(0);
1880 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
1882 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
1885 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
1886 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1889 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
1890 SDValue N2 = N0.getOperand(1);
1891 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1894 unsigned N1CVal = N1C->getZExtValue();
1895 unsigned N2CVal = N2C->getZExtValue();
1896 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
1897 (N1CVal & 0xffffU) == 0xffffU &&
1898 (N2CVal & 0xffffU) == 0x0U) {
1899 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
1901 SDValue Ops[] = { N0.getOperand(0), Imm16,
1902 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1903 return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
1908 case ARMISD::VMOVRRD:
1909 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
1910 N->getOperand(0), getAL(CurDAG),
1911 CurDAG->getRegister(0, MVT::i32));
1912 case ISD::UMUL_LOHI: {
1913 if (Subtarget->isThumb1Only())
1915 if (Subtarget->isThumb()) {
1916 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1917 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1918 CurDAG->getRegister(0, MVT::i32) };
1919 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops,4);
1921 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1922 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1923 CurDAG->getRegister(0, MVT::i32) };
1924 return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1927 case ISD::SMUL_LOHI: {
1928 if (Subtarget->isThumb1Only())
1930 if (Subtarget->isThumb()) {
1931 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1932 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1933 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops,4);
1935 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1936 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1937 CurDAG->getRegister(0, MVT::i32) };
1938 return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1942 SDNode *ResNode = 0;
1943 if (Subtarget->isThumb() && Subtarget->hasThumb2())
1944 ResNode = SelectT2IndexedLoad(N);
1946 ResNode = SelectARMIndexedLoad(N);
1950 // VLDMQ must be custom-selected for "v2f64 load" to set the AM5Opc value.
1951 if (Subtarget->hasVFP2() &&
1952 N->getValueType(0).getSimpleVT().SimpleTy == MVT::v2f64) {
1953 SDValue Chain = N->getOperand(0);
1955 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
1956 SDValue Pred = getAL(CurDAG);
1957 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1958 SDValue Ops[] = { N->getOperand(1), AM5Opc, Pred, PredReg, Chain };
1959 return CurDAG->getMachineNode(ARM::VLDMQ, dl, MVT::v2f64, MVT::Other,
1962 // Other cases are autogenerated.
1966 // VSTMQ must be custom-selected for "v2f64 store" to set the AM5Opc value.
1967 if (Subtarget->hasVFP2() &&
1968 N->getOperand(1).getValueType().getSimpleVT().SimpleTy == MVT::v2f64) {
1969 SDValue Chain = N->getOperand(0);
1971 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
1972 SDValue Pred = getAL(CurDAG);
1973 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1974 SDValue Ops[] = { N->getOperand(1), N->getOperand(2),
1975 AM5Opc, Pred, PredReg, Chain };
1976 return CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
1978 // Other cases are autogenerated.
1981 case ARMISD::BRCOND: {
1982 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1983 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
1984 // Pattern complexity = 6 cost = 1 size = 0
1986 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1987 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
1988 // Pattern complexity = 6 cost = 1 size = 0
1990 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1991 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
1992 // Pattern complexity = 6 cost = 1 size = 0
1994 unsigned Opc = Subtarget->isThumb() ?
1995 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
1996 SDValue Chain = N->getOperand(0);
1997 SDValue N1 = N->getOperand(1);
1998 SDValue N2 = N->getOperand(2);
1999 SDValue N3 = N->getOperand(3);
2000 SDValue InFlag = N->getOperand(4);
2001 assert(N1.getOpcode() == ISD::BasicBlock);
2002 assert(N2.getOpcode() == ISD::Constant);
2003 assert(N3.getOpcode() == ISD::Register);
2005 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2006 cast<ConstantSDNode>(N2)->getZExtValue()),
2008 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
2009 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
2011 Chain = SDValue(ResNode, 0);
2012 if (N->getNumValues() == 2) {
2013 InFlag = SDValue(ResNode, 1);
2014 ReplaceUses(SDValue(N, 1), InFlag);
2016 ReplaceUses(SDValue(N, 0),
2017 SDValue(Chain.getNode(), Chain.getResNo()));
2021 return SelectCMOVOp(N);
2022 case ARMISD::CNEG: {
2023 EVT VT = N->getValueType(0);
2024 SDValue N0 = N->getOperand(0);
2025 SDValue N1 = N->getOperand(1);
2026 SDValue N2 = N->getOperand(2);
2027 SDValue N3 = N->getOperand(3);
2028 SDValue InFlag = N->getOperand(4);
2029 assert(N2.getOpcode() == ISD::Constant);
2030 assert(N3.getOpcode() == ISD::Register);
2032 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2033 cast<ConstantSDNode>(N2)->getZExtValue()),
2035 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
2037 switch (VT.getSimpleVT().SimpleTy) {
2038 default: assert(false && "Illegal conditional move type!");
2047 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
2050 case ARMISD::VZIP: {
2052 EVT VT = N->getValueType(0);
2053 switch (VT.getSimpleVT().SimpleTy) {
2054 default: return NULL;
2055 case MVT::v8i8: Opc = ARM::VZIPd8; break;
2056 case MVT::v4i16: Opc = ARM::VZIPd16; break;
2058 case MVT::v2i32: Opc = ARM::VZIPd32; break;
2059 case MVT::v16i8: Opc = ARM::VZIPq8; break;
2060 case MVT::v8i16: Opc = ARM::VZIPq16; break;
2062 case MVT::v4i32: Opc = ARM::VZIPq32; break;
2064 SDValue Pred = getAL(CurDAG);
2065 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2066 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2067 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2069 case ARMISD::VUZP: {
2071 EVT VT = N->getValueType(0);
2072 switch (VT.getSimpleVT().SimpleTy) {
2073 default: return NULL;
2074 case MVT::v8i8: Opc = ARM::VUZPd8; break;
2075 case MVT::v4i16: Opc = ARM::VUZPd16; break;
2077 case MVT::v2i32: Opc = ARM::VUZPd32; break;
2078 case MVT::v16i8: Opc = ARM::VUZPq8; break;
2079 case MVT::v8i16: Opc = ARM::VUZPq16; break;
2081 case MVT::v4i32: Opc = ARM::VUZPq32; break;
2083 SDValue Pred = getAL(CurDAG);
2084 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2085 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2086 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2088 case ARMISD::VTRN: {
2090 EVT VT = N->getValueType(0);
2091 switch (VT.getSimpleVT().SimpleTy) {
2092 default: return NULL;
2093 case MVT::v8i8: Opc = ARM::VTRNd8; break;
2094 case MVT::v4i16: Opc = ARM::VTRNd16; break;
2096 case MVT::v2i32: Opc = ARM::VTRNd32; break;
2097 case MVT::v16i8: Opc = ARM::VTRNq8; break;
2098 case MVT::v8i16: Opc = ARM::VTRNq16; break;
2100 case MVT::v4i32: Opc = ARM::VTRNq32; break;
2102 SDValue Pred = getAL(CurDAG);
2103 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2104 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2105 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2108 case ISD::INTRINSIC_VOID:
2109 case ISD::INTRINSIC_W_CHAIN: {
2110 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2115 case Intrinsic::arm_neon_vld1: {
2116 unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
2117 ARM::VLD1d32, ARM::VLD1d64 };
2118 unsigned QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
2119 ARM::VLD1q32, ARM::VLD1q64 };
2120 return SelectVLD(N, 1, DOpcodes, QOpcodes, 0);
2123 case Intrinsic::arm_neon_vld2: {
2124 unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
2125 ARM::VLD2d32, ARM::VLD1q64 };
2126 unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
2127 return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
2130 case Intrinsic::arm_neon_vld3: {
2131 unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
2132 ARM::VLD3d32, ARM::VLD1d64T };
2133 unsigned QOpcodes0[] = { ARM::VLD3q8_UPD,
2136 unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD,
2137 ARM::VLD3q16odd_UPD,
2138 ARM::VLD3q32odd_UPD };
2139 return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2142 case Intrinsic::arm_neon_vld4: {
2143 unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
2144 ARM::VLD4d32, ARM::VLD1d64Q };
2145 unsigned QOpcodes0[] = { ARM::VLD4q8_UPD,
2148 unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD,
2149 ARM::VLD4q16odd_UPD,
2150 ARM::VLD4q32odd_UPD };
2151 return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2154 case Intrinsic::arm_neon_vld2lane: {
2155 unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
2156 unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 };
2157 unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd };
2158 return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
2161 case Intrinsic::arm_neon_vld3lane: {
2162 unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
2163 unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 };
2164 unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd };
2165 return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
2168 case Intrinsic::arm_neon_vld4lane: {
2169 unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
2170 unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 };
2171 unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd };
2172 return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
2175 case Intrinsic::arm_neon_vst1: {
2176 unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
2177 ARM::VST1d32, ARM::VST1d64 };
2178 unsigned QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
2179 ARM::VST1q32, ARM::VST1q64 };
2180 return SelectVST(N, 1, DOpcodes, QOpcodes, 0);
2183 case Intrinsic::arm_neon_vst2: {
2184 unsigned DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
2185 ARM::VST2d32, ARM::VST1q64 };
2186 unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
2187 return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
2190 case Intrinsic::arm_neon_vst3: {
2191 unsigned DOpcodes[] = { ARM::VST3d8, ARM::VST3d16,
2192 ARM::VST3d32, ARM::VST1d64T };
2193 unsigned QOpcodes0[] = { ARM::VST3q8_UPD,
2196 unsigned QOpcodes1[] = { ARM::VST3q8odd_UPD,
2197 ARM::VST3q16odd_UPD,
2198 ARM::VST3q32odd_UPD };
2199 return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2202 case Intrinsic::arm_neon_vst4: {
2203 unsigned DOpcodes[] = { ARM::VST4d8, ARM::VST4d16,
2204 ARM::VST4d32, ARM::VST1d64Q };
2205 unsigned QOpcodes0[] = { ARM::VST4q8_UPD,
2208 unsigned QOpcodes1[] = { ARM::VST4q8odd_UPD,
2209 ARM::VST4q16odd_UPD,
2210 ARM::VST4q32odd_UPD };
2211 return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2214 case Intrinsic::arm_neon_vst2lane: {
2215 unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
2216 unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 };
2217 unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd };
2218 return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
2221 case Intrinsic::arm_neon_vst3lane: {
2222 unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
2223 unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 };
2224 unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd };
2225 return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
2228 case Intrinsic::arm_neon_vst4lane: {
2229 unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
2230 unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 };
2231 unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd };
2232 return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
2238 case ISD::CONCAT_VECTORS:
2239 return SelectConcatVector(N);
2242 return SelectCode(N);
2245 bool ARMDAGToDAGISel::
2246 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2247 std::vector<SDValue> &OutOps) {
2248 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
2249 // Require the address to be in a register. That is safe for all ARM
2250 // variants and it is hard to do anything much smarter without knowing
2251 // how the operand is used.
2252 OutOps.push_back(Op);
2256 /// createARMISelDag - This pass converts a legalized DAG into a
2257 /// ARM-specific DAG, ready for instruction scheduling.
2259 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
2260 CodeGenOpt::Level OptLevel) {
2261 return new ARMDAGToDAGISel(TM, OptLevel);
2264 /// ModelWithRegSequence - Return true if isel should use REG_SEQUENCE to model
2265 /// operations involving sub-registers.
2266 bool llvm::ModelWithRegSequence() {