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"
39 //===--------------------------------------------------------------------===//
40 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
41 /// instructions for SelectionDAG operations.
44 class ARMDAGToDAGISel : public SelectionDAGISel {
45 ARMBaseTargetMachine &TM;
47 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
48 /// make the right decision when generating code for different targets.
49 const ARMSubtarget *Subtarget;
52 explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
53 CodeGenOpt::Level OptLevel)
54 : SelectionDAGISel(tm, OptLevel), TM(tm),
55 Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
58 virtual const char *getPassName() const {
59 return "ARM Instruction Selection";
62 /// getI32Imm - Return a target constant of type i32 with the specified
64 inline SDValue getI32Imm(unsigned Imm) {
65 return CurDAG->getTargetConstant(Imm, MVT::i32);
68 SDNode *Select(SDNode *N);
70 bool SelectShifterOperandReg(SDNode *Op, SDValue N, SDValue &A,
71 SDValue &B, SDValue &C);
72 bool SelectAddrMode2(SDNode *Op, SDValue N, SDValue &Base,
73 SDValue &Offset, SDValue &Opc);
74 bool SelectAddrMode2Offset(SDNode *Op, SDValue N,
75 SDValue &Offset, SDValue &Opc);
76 bool SelectAddrMode3(SDNode *Op, SDValue N, SDValue &Base,
77 SDValue &Offset, SDValue &Opc);
78 bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
79 SDValue &Offset, SDValue &Opc);
80 bool SelectAddrMode4(SDNode *Op, SDValue N, SDValue &Addr,
82 bool SelectAddrMode5(SDNode *Op, SDValue N, SDValue &Base,
84 bool SelectAddrMode6(SDNode *Op, SDValue N, SDValue &Addr, SDValue &Align);
86 bool SelectAddrModePC(SDNode *Op, SDValue N, SDValue &Offset,
89 bool SelectThumbAddrModeRR(SDNode *Op, SDValue N, SDValue &Base,
91 bool SelectThumbAddrModeRI5(SDNode *Op, SDValue N, unsigned Scale,
92 SDValue &Base, SDValue &OffImm,
94 bool SelectThumbAddrModeS1(SDNode *Op, SDValue N, SDValue &Base,
95 SDValue &OffImm, SDValue &Offset);
96 bool SelectThumbAddrModeS2(SDNode *Op, SDValue N, SDValue &Base,
97 SDValue &OffImm, SDValue &Offset);
98 bool SelectThumbAddrModeS4(SDNode *Op, SDValue N, SDValue &Base,
99 SDValue &OffImm, SDValue &Offset);
100 bool SelectThumbAddrModeSP(SDNode *Op, SDValue N, SDValue &Base,
103 bool SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
104 SDValue &BaseReg, SDValue &Opc);
105 bool SelectT2AddrModeImm12(SDNode *Op, SDValue N, SDValue &Base,
107 bool SelectT2AddrModeImm8(SDNode *Op, SDValue N, SDValue &Base,
109 bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
111 bool SelectT2AddrModeImm8s4(SDNode *Op, SDValue N, SDValue &Base,
113 bool SelectT2AddrModeSoReg(SDNode *Op, SDValue N, SDValue &Base,
114 SDValue &OffReg, SDValue &ShImm);
116 // Include the pieces autogenerated from the target description.
117 #include "ARMGenDAGISel.inc"
120 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
122 SDNode *SelectARMIndexedLoad(SDNode *N);
123 SDNode *SelectT2IndexedLoad(SDNode *N);
125 /// SelectVLD - Select NEON load intrinsics. NumVecs should be
126 /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for
127 /// loads of D registers and even subregs and odd subregs of Q registers.
128 /// For NumVecs <= 2, QOpcodes1 is not used.
129 SDNode *SelectVLD(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
130 unsigned *QOpcodes0, unsigned *QOpcodes1);
132 /// SelectVST - Select NEON store intrinsics. NumVecs should
133 /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for
134 /// stores of D registers and even subregs and odd subregs of Q registers.
135 /// For NumVecs <= 2, QOpcodes1 is not used.
136 SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
137 unsigned *QOpcodes0, unsigned *QOpcodes1);
139 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
140 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
141 /// load/store of D registers and even subregs and odd subregs of Q registers.
142 SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, unsigned NumVecs,
143 unsigned *DOpcodes, unsigned *QOpcodes0,
144 unsigned *QOpcodes1);
146 /// SelectVTBL - Select NEON VTBL and VTBX intrinsics. NumVecs should be 2,
147 /// 3 or 4. These are custom-selected so that a REG_SEQUENCE can be
148 /// generated to force the table registers to be consecutive.
149 SDNode *SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs, unsigned Opc);
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 if (N.getOpcode() != ISD::ADD) {
542 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
543 if (!NC || !NC->isNullValue())
550 Base = N.getOperand(0);
551 Offset = N.getOperand(1);
556 ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
557 unsigned Scale, SDValue &Base,
558 SDValue &OffImm, SDValue &Offset) {
560 SDValue TmpBase, TmpOffImm;
561 if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
562 return false; // We want to select tLDRspi / tSTRspi instead.
563 if (N.getOpcode() == ARMISD::Wrapper &&
564 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
565 return false; // We want to select tLDRpci instead.
568 if (N.getOpcode() != ISD::ADD) {
569 if (N.getOpcode() == ARMISD::Wrapper &&
570 !(Subtarget->useMovt() &&
571 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
572 Base = N.getOperand(0);
576 Offset = CurDAG->getRegister(0, MVT::i32);
577 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
581 // Thumb does not have [sp, r] address mode.
582 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
583 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
584 if ((LHSR && LHSR->getReg() == ARM::SP) ||
585 (RHSR && RHSR->getReg() == ARM::SP)) {
587 Offset = CurDAG->getRegister(0, MVT::i32);
588 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
592 // If the RHS is + imm5 * scale, fold into addr mode.
593 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
594 int RHSC = (int)RHS->getZExtValue();
595 if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
597 if (RHSC >= 0 && RHSC < 32) {
598 Base = N.getOperand(0);
599 Offset = CurDAG->getRegister(0, MVT::i32);
600 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
606 Base = N.getOperand(0);
607 Offset = N.getOperand(1);
608 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
612 bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
613 SDValue &Base, SDValue &OffImm,
615 return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
618 bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
619 SDValue &Base, SDValue &OffImm,
621 return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
624 bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
625 SDValue &Base, SDValue &OffImm,
627 return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
630 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
631 SDValue &Base, SDValue &OffImm) {
632 if (N.getOpcode() == ISD::FrameIndex) {
633 int FI = cast<FrameIndexSDNode>(N)->getIndex();
634 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
635 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
639 if (N.getOpcode() != ISD::ADD)
642 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
643 if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
644 (LHSR && LHSR->getReg() == ARM::SP)) {
645 // If the RHS is + imm8 * scale, fold into addr mode.
646 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
647 int RHSC = (int)RHS->getZExtValue();
648 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
650 if (RHSC >= 0 && RHSC < 256) {
651 Base = N.getOperand(0);
652 if (Base.getOpcode() == ISD::FrameIndex) {
653 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
654 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
656 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
666 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
669 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
671 // Don't match base register only case. That is matched to a separate
672 // lower complexity pattern with explicit register operand.
673 if (ShOpcVal == ARM_AM::no_shift) return false;
675 BaseReg = N.getOperand(0);
676 unsigned ShImmVal = 0;
677 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
678 ShImmVal = RHS->getZExtValue() & 31;
679 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
686 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
687 SDValue &Base, SDValue &OffImm) {
688 // Match simple R + imm12 operands.
691 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
692 if (N.getOpcode() == ISD::FrameIndex) {
693 // Match frame index...
694 int FI = cast<FrameIndexSDNode>(N)->getIndex();
695 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
696 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
698 } else if (N.getOpcode() == ARMISD::Wrapper &&
699 !(Subtarget->useMovt() &&
700 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
701 Base = N.getOperand(0);
702 if (Base.getOpcode() == ISD::TargetConstantPool)
703 return false; // We want to select t2LDRpci instead.
706 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
710 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
711 if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
712 // Let t2LDRi8 handle (R - imm8).
715 int RHSC = (int)RHS->getZExtValue();
716 if (N.getOpcode() == ISD::SUB)
719 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
720 Base = N.getOperand(0);
721 if (Base.getOpcode() == ISD::FrameIndex) {
722 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
723 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
725 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
732 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
736 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
737 SDValue &Base, SDValue &OffImm) {
738 // Match simple R - imm8 operands.
739 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
740 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
741 int RHSC = (int)RHS->getSExtValue();
742 if (N.getOpcode() == ISD::SUB)
745 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
746 Base = N.getOperand(0);
747 if (Base.getOpcode() == ISD::FrameIndex) {
748 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
749 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
751 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
760 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
762 unsigned Opcode = Op->getOpcode();
763 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
764 ? cast<LoadSDNode>(Op)->getAddressingMode()
765 : cast<StoreSDNode>(Op)->getAddressingMode();
766 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
767 int RHSC = (int)RHS->getZExtValue();
768 if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
769 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
770 ? CurDAG->getTargetConstant(RHSC, MVT::i32)
771 : CurDAG->getTargetConstant(-RHSC, MVT::i32);
779 bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
780 SDValue &Base, SDValue &OffImm) {
781 if (N.getOpcode() == ISD::ADD) {
782 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
783 int RHSC = (int)RHS->getZExtValue();
785 if (((RHSC & 0x3) == 0) &&
786 ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) {
787 Base = N.getOperand(0);
788 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
792 } else if (N.getOpcode() == ISD::SUB) {
793 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
794 int RHSC = (int)RHS->getZExtValue();
796 if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) {
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 /// PairSRegs - Form a D register from a pair of S registers.
960 SDNode *ARMDAGToDAGISel::PairSRegs(EVT VT, SDValue V0, SDValue V1) {
961 DebugLoc dl = V0.getNode()->getDebugLoc();
962 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
963 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
964 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
965 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
968 /// PairDRegs - Form a quad register from a pair of D registers.
970 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
971 DebugLoc dl = V0.getNode()->getDebugLoc();
972 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
973 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
974 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
975 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
978 /// PairQRegs - Form 4 consecutive D registers from a pair of Q registers.
980 SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) {
981 DebugLoc dl = V0.getNode()->getDebugLoc();
982 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
983 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
984 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
985 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
988 /// QuadSRegs - Form 4 consecutive S registers.
990 SDNode *ARMDAGToDAGISel::QuadSRegs(EVT VT, SDValue V0, SDValue V1,
991 SDValue V2, SDValue V3) {
992 DebugLoc dl = V0.getNode()->getDebugLoc();
993 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
994 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
995 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, MVT::i32);
996 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, MVT::i32);
997 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
998 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1001 /// QuadDRegs - Form 4 consecutive D registers.
1003 SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
1004 SDValue V2, SDValue V3) {
1005 DebugLoc dl = V0.getNode()->getDebugLoc();
1006 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1007 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1008 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
1009 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
1010 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1011 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1014 /// QuadQRegs - Form 4 consecutive Q registers.
1016 SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1,
1017 SDValue V2, SDValue V3) {
1018 DebugLoc dl = V0.getNode()->getDebugLoc();
1019 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
1020 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
1021 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, MVT::i32);
1022 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, MVT::i32);
1023 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1024 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1027 /// OctoDRegs - Form 8 consecutive D registers.
1029 SDNode *ARMDAGToDAGISel::OctoDRegs(EVT VT, SDValue V0, SDValue V1,
1030 SDValue V2, SDValue V3,
1031 SDValue V4, SDValue V5,
1032 SDValue V6, SDValue V7) {
1033 DebugLoc dl = V0.getNode()->getDebugLoc();
1034 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1035 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1036 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
1037 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
1038 SDValue SubReg4 = CurDAG->getTargetConstant(ARM::dsub_4, MVT::i32);
1039 SDValue SubReg5 = CurDAG->getTargetConstant(ARM::dsub_5, MVT::i32);
1040 SDValue SubReg6 = CurDAG->getTargetConstant(ARM::dsub_6, MVT::i32);
1041 SDValue SubReg7 = CurDAG->getTargetConstant(ARM::dsub_7, MVT::i32);
1042 const SDValue Ops[] ={ V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3,
1043 V4, SubReg4, V5, SubReg5, V6, SubReg6, V7, SubReg7 };
1044 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 16);
1047 /// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
1048 /// for a 64-bit subregister of the vector.
1049 static EVT GetNEONSubregVT(EVT VT) {
1050 switch (VT.getSimpleVT().SimpleTy) {
1051 default: llvm_unreachable("unhandled NEON type");
1052 case MVT::v16i8: return MVT::v8i8;
1053 case MVT::v8i16: return MVT::v4i16;
1054 case MVT::v4f32: return MVT::v2f32;
1055 case MVT::v4i32: return MVT::v2i32;
1056 case MVT::v2i64: return MVT::v1i64;
1060 SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
1061 unsigned *DOpcodes, unsigned *QOpcodes0,
1062 unsigned *QOpcodes1) {
1063 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1064 DebugLoc dl = N->getDebugLoc();
1066 SDValue MemAddr, Align;
1067 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1070 SDValue Chain = N->getOperand(0);
1071 EVT VT = N->getValueType(0);
1072 bool is64BitVector = VT.is64BitVector();
1074 unsigned OpcodeIndex;
1075 switch (VT.getSimpleVT().SimpleTy) {
1076 default: llvm_unreachable("unhandled vld type");
1077 // Double-register operations:
1078 case MVT::v8i8: OpcodeIndex = 0; break;
1079 case MVT::v4i16: OpcodeIndex = 1; break;
1081 case MVT::v2i32: OpcodeIndex = 2; break;
1082 case MVT::v1i64: OpcodeIndex = 3; break;
1083 // Quad-register operations:
1084 case MVT::v16i8: OpcodeIndex = 0; break;
1085 case MVT::v8i16: OpcodeIndex = 1; break;
1087 case MVT::v4i32: OpcodeIndex = 2; break;
1088 case MVT::v2i64: OpcodeIndex = 3;
1089 assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
1093 SDValue Pred = getAL(CurDAG);
1094 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1095 if (is64BitVector) {
1096 unsigned Opc = DOpcodes[OpcodeIndex];
1097 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1098 std::vector<EVT> ResTys(NumVecs, VT);
1099 ResTys.push_back(MVT::Other);
1100 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1105 SDValue V0 = SDValue(VLd, 0);
1106 SDValue V1 = SDValue(VLd, 1);
1108 // Form a REG_SEQUENCE to force register allocation.
1110 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1112 SDValue V2 = SDValue(VLd, 2);
1113 // If it's a vld3, form a quad D-register but discard the last part.
1114 SDValue V3 = (NumVecs == 3)
1115 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1117 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1120 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1121 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1122 SDValue D = CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec,
1124 ReplaceUses(SDValue(N, Vec), D);
1126 ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, NumVecs));
1130 EVT RegVT = GetNEONSubregVT(VT);
1132 // Quad registers are directly supported for VLD1 and VLD2,
1133 // loading pairs of D regs.
1134 unsigned Opc = QOpcodes0[OpcodeIndex];
1135 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1136 std::vector<EVT> ResTys(2 * NumVecs, RegVT);
1137 ResTys.push_back(MVT::Other);
1138 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1139 Chain = SDValue(VLd, 2 * NumVecs);
1141 // Combine the even and odd subregs to produce the result.
1143 SDNode *Q = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
1144 ReplaceUses(SDValue(N, 0), SDValue(Q, 0));
1146 SDValue QQ = SDValue(QuadDRegs(MVT::v4i64,
1147 SDValue(VLd, 0), SDValue(VLd, 1),
1148 SDValue(VLd, 2), SDValue(VLd, 3)), 0);
1149 SDValue Q0 = CurDAG->getTargetExtractSubreg(ARM::qsub_0, dl, VT, QQ);
1150 SDValue Q1 = CurDAG->getTargetExtractSubreg(ARM::qsub_1, dl, VT, QQ);
1151 ReplaceUses(SDValue(N, 0), Q0);
1152 ReplaceUses(SDValue(N, 1), Q1);
1155 // Otherwise, quad registers are loaded with two separate instructions,
1156 // where one loads the even registers and the other loads the odd registers.
1158 std::vector<EVT> ResTys(NumVecs, RegVT);
1159 ResTys.push_back(MemAddr.getValueType());
1160 ResTys.push_back(MVT::Other);
1162 // Load the even subregs.
1163 unsigned Opc = QOpcodes0[OpcodeIndex];
1164 const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain };
1165 SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6);
1166 Chain = SDValue(VLdA, NumVecs+1);
1168 // Load the odd subregs.
1169 Opc = QOpcodes1[OpcodeIndex];
1170 const SDValue OpsB[] = { SDValue(VLdA, NumVecs),
1171 Align, Reg0, Pred, Reg0, Chain };
1172 SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
1173 Chain = SDValue(VLdB, NumVecs+1);
1175 SDValue V0 = SDValue(VLdA, 0);
1176 SDValue V1 = SDValue(VLdB, 0);
1177 SDValue V2 = SDValue(VLdA, 1);
1178 SDValue V3 = SDValue(VLdB, 1);
1179 SDValue V4 = SDValue(VLdA, 2);
1180 SDValue V5 = SDValue(VLdB, 2);
1181 SDValue V6 = (NumVecs == 3)
1182 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0)
1184 SDValue V7 = (NumVecs == 3)
1185 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0)
1187 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V0, V1, V2, V3,
1188 V4, V5, V6, V7), 0);
1190 // Extract out the 3 / 4 Q registers.
1191 assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1192 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1193 SDValue Q = CurDAG->getTargetExtractSubreg(ARM::qsub_0+Vec,
1195 ReplaceUses(SDValue(N, Vec), Q);
1198 ReplaceUses(SDValue(N, NumVecs), Chain);
1202 SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
1203 unsigned *DOpcodes, unsigned *QOpcodes0,
1204 unsigned *QOpcodes1) {
1205 assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range");
1206 DebugLoc dl = N->getDebugLoc();
1208 SDValue MemAddr, Align;
1209 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1212 SDValue Chain = N->getOperand(0);
1213 EVT VT = N->getOperand(3).getValueType();
1214 bool is64BitVector = VT.is64BitVector();
1216 unsigned OpcodeIndex;
1217 switch (VT.getSimpleVT().SimpleTy) {
1218 default: llvm_unreachable("unhandled vst type");
1219 // Double-register operations:
1220 case MVT::v8i8: OpcodeIndex = 0; break;
1221 case MVT::v4i16: OpcodeIndex = 1; break;
1223 case MVT::v2i32: OpcodeIndex = 2; break;
1224 case MVT::v1i64: OpcodeIndex = 3; break;
1225 // Quad-register operations:
1226 case MVT::v16i8: OpcodeIndex = 0; break;
1227 case MVT::v8i16: OpcodeIndex = 1; break;
1229 case MVT::v4i32: OpcodeIndex = 2; break;
1230 case MVT::v2i64: OpcodeIndex = 3;
1231 assert(NumVecs == 1 && "v2i64 type only supported for VST1");
1235 SDValue Pred = getAL(CurDAG);
1236 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1238 SmallVector<SDValue, 10> Ops;
1239 Ops.push_back(MemAddr);
1240 Ops.push_back(Align);
1242 if (is64BitVector) {
1245 SDValue V0 = N->getOperand(0+3);
1246 SDValue V1 = N->getOperand(1+3);
1248 // Form a REG_SEQUENCE to force register allocation.
1250 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1252 SDValue V2 = N->getOperand(2+3);
1253 // If it's a vld3, form a quad D-register and leave the last part as
1255 SDValue V3 = (NumVecs == 3)
1256 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1257 : N->getOperand(3+3);
1258 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1261 // Now extract the D registers back out.
1262 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT,
1264 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT,
1267 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,
1270 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,
1273 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1274 Ops.push_back(N->getOperand(Vec+3));
1276 Ops.push_back(Pred);
1277 Ops.push_back(Reg0); // predicate register
1278 Ops.push_back(Chain);
1279 unsigned Opc = DOpcodes[OpcodeIndex];
1280 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+5);
1283 EVT RegVT = GetNEONSubregVT(VT);
1285 // Quad registers are directly supported for VST1 and VST2,
1286 // storing pairs of D regs.
1287 unsigned Opc = QOpcodes0[OpcodeIndex];
1289 // First extract the pair of Q registers.
1290 SDValue Q0 = N->getOperand(3);
1291 SDValue Q1 = N->getOperand(4);
1293 // Form a QQ register.
1294 SDValue QQ = SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0);
1296 // Now extract the D registers back out.
1297 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1299 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1301 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, RegVT,
1303 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, RegVT,
1305 Ops.push_back(Pred);
1306 Ops.push_back(Reg0); // predicate register
1307 Ops.push_back(Chain);
1308 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 5 + 4);
1310 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1311 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1312 N->getOperand(Vec+3)));
1313 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1314 N->getOperand(Vec+3)));
1316 Ops.push_back(Pred);
1317 Ops.push_back(Reg0); // predicate register
1318 Ops.push_back(Chain);
1319 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
1324 // Otherwise, quad registers are stored with two separate instructions,
1325 // where one stores the even registers and the other stores the odd registers.
1327 // Form the QQQQ REG_SEQUENCE.
1329 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1330 V[i] = CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1331 N->getOperand(Vec+3));
1332 V[i+1] = CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1333 N->getOperand(Vec+3));
1336 V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1339 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1340 V[4], V[5], V[6], V[7]), 0);
1342 // Store the even D registers.
1343 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1344 Ops.push_back(Reg0); // post-access address offset
1345 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1346 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec*2, dl,
1348 Ops.push_back(Pred);
1349 Ops.push_back(Reg0); // predicate register
1350 Ops.push_back(Chain);
1351 unsigned Opc = QOpcodes0[OpcodeIndex];
1352 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1353 MVT::Other, Ops.data(), NumVecs+6);
1354 Chain = SDValue(VStA, 1);
1356 // Store the odd D registers.
1357 Ops[0] = SDValue(VStA, 0); // MemAddr
1358 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1359 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::dsub_1+Vec*2, dl,
1361 Ops[NumVecs+5] = Chain;
1362 Opc = QOpcodes1[OpcodeIndex];
1363 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1364 MVT::Other, Ops.data(), NumVecs+6);
1365 Chain = SDValue(VStB, 1);
1366 ReplaceUses(SDValue(N, 0), Chain);
1370 SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1371 unsigned NumVecs, unsigned *DOpcodes,
1372 unsigned *QOpcodes0,
1373 unsigned *QOpcodes1) {
1374 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1375 DebugLoc dl = N->getDebugLoc();
1377 SDValue MemAddr, Align;
1378 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1381 SDValue Chain = N->getOperand(0);
1383 cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1384 EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
1385 bool is64BitVector = VT.is64BitVector();
1387 // Quad registers are handled by load/store of subregs. Find the subreg info.
1388 unsigned NumElts = 0;
1391 if (!is64BitVector) {
1392 RegVT = GetNEONSubregVT(VT);
1393 NumElts = RegVT.getVectorNumElements();
1394 Even = Lane < NumElts;
1397 unsigned OpcodeIndex;
1398 switch (VT.getSimpleVT().SimpleTy) {
1399 default: llvm_unreachable("unhandled vld/vst lane type");
1400 // Double-register operations:
1401 case MVT::v8i8: OpcodeIndex = 0; break;
1402 case MVT::v4i16: OpcodeIndex = 1; break;
1404 case MVT::v2i32: OpcodeIndex = 2; break;
1405 // Quad-register operations:
1406 case MVT::v8i16: OpcodeIndex = 0; break;
1408 case MVT::v4i32: OpcodeIndex = 1; break;
1411 SDValue Pred = getAL(CurDAG);
1412 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1414 SmallVector<SDValue, 10> Ops;
1415 Ops.push_back(MemAddr);
1416 Ops.push_back(Align);
1419 if (is64BitVector) {
1420 Opc = DOpcodes[OpcodeIndex];
1422 SDValue V0 = N->getOperand(0+3);
1423 SDValue V1 = N->getOperand(1+3);
1425 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1427 SDValue V2 = N->getOperand(2+3);
1428 SDValue V3 = (NumVecs == 3)
1429 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1430 : N->getOperand(3+3);
1431 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1434 // Now extract the D registers back out.
1435 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT, RegSeq));
1436 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT, RegSeq));
1438 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,RegSeq));
1440 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,RegSeq));
1442 // Check if this is loading the even or odd subreg of a Q register.
1443 if (Lane < NumElts) {
1444 Opc = QOpcodes0[OpcodeIndex];
1447 Opc = QOpcodes1[OpcodeIndex];
1451 SDValue V0 = N->getOperand(0+3);
1452 SDValue V1 = N->getOperand(1+3);
1454 RegSeq = SDValue(PairQRegs(MVT::v4i64, V0, V1), 0);
1456 SDValue V2 = N->getOperand(2+3);
1457 SDValue V3 = (NumVecs == 3)
1458 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1459 : N->getOperand(3+3);
1460 RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
1463 // Extract the subregs of the input vector.
1464 unsigned SubIdx = Even ? ARM::dsub_0 : ARM::dsub_1;
1465 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1466 Ops.push_back(CurDAG->getTargetExtractSubreg(SubIdx+Vec*2, dl, RegVT,
1469 Ops.push_back(getI32Imm(Lane));
1470 Ops.push_back(Pred);
1471 Ops.push_back(Reg0);
1472 Ops.push_back(Chain);
1475 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6);
1477 std::vector<EVT> ResTys(NumVecs, RegVT);
1478 ResTys.push_back(MVT::Other);
1479 SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(),NumVecs+6);
1481 // Form a REG_SEQUENCE to force register allocation.
1483 if (is64BitVector) {
1484 SDValue V0 = SDValue(VLdLn, 0);
1485 SDValue V1 = SDValue(VLdLn, 1);
1487 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1489 SDValue V2 = SDValue(VLdLn, 2);
1490 // If it's a vld3, form a quad D-register but discard the last part.
1491 SDValue V3 = (NumVecs == 3)
1492 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1493 : SDValue(VLdLn, 3);
1494 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1497 // For 128-bit vectors, take the 64-bit results of the load and insert
1498 // them as subregs into the result.
1500 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1502 V[i] = SDValue(VLdLn, Vec);
1503 V[i+1] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1506 V[i] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1508 V[i+1] = SDValue(VLdLn, Vec);
1512 V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1516 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V[0], V[1], V[2], V[3]), 0);
1518 RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1519 V[4], V[5], V[6], V[7]), 0);
1522 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1523 assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1524 unsigned SubIdx = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
1525 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1526 ReplaceUses(SDValue(N, Vec),
1527 CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, RegSeq));
1528 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, NumVecs));
1532 SDNode *ARMDAGToDAGISel::SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs,
1534 assert(NumVecs >= 2 && NumVecs <= 4 && "VTBL NumVecs out-of-range");
1535 DebugLoc dl = N->getDebugLoc();
1536 EVT VT = N->getValueType(0);
1537 unsigned FirstTblReg = IsExt ? 2 : 1;
1539 // Form a REG_SEQUENCE to force register allocation.
1541 SDValue V0 = N->getOperand(FirstTblReg + 0);
1542 SDValue V1 = N->getOperand(FirstTblReg + 1);
1544 RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0);
1546 SDValue V2 = N->getOperand(FirstTblReg + 2);
1547 // If it's a vtbl3, form a quad D-register and leave the last part as
1549 SDValue V3 = (NumVecs == 3)
1550 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
1551 : N->getOperand(FirstTblReg + 3);
1552 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1555 // Now extract the D registers back out.
1556 SmallVector<SDValue, 6> Ops;
1558 Ops.push_back(N->getOperand(1));
1559 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT, RegSeq));
1560 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT, RegSeq));
1562 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT, RegSeq));
1564 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT, RegSeq));
1566 Ops.push_back(N->getOperand(FirstTblReg + NumVecs));
1567 Ops.push_back(getAL(CurDAG)); // predicate
1568 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // predicate register
1569 return CurDAG->getMachineNode(Opc, dl, VT, Ops.data(), Ops.size());
1572 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
1574 if (!Subtarget->hasV6T2Ops())
1577 unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
1578 : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
1581 // For unsigned extracts, check for a shift right and mask
1582 unsigned And_imm = 0;
1583 if (N->getOpcode() == ISD::AND) {
1584 if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
1586 // The immediate is a mask of the low bits iff imm & (imm+1) == 0
1587 if (And_imm & (And_imm + 1))
1590 unsigned Srl_imm = 0;
1591 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
1593 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1595 unsigned Width = CountTrailingOnes_32(And_imm);
1596 unsigned LSB = Srl_imm;
1597 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1598 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1599 CurDAG->getTargetConstant(LSB, MVT::i32),
1600 CurDAG->getTargetConstant(Width, MVT::i32),
1601 getAL(CurDAG), Reg0 };
1602 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1608 // Otherwise, we're looking for a shift of a shift
1609 unsigned Shl_imm = 0;
1610 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
1611 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1612 unsigned Srl_imm = 0;
1613 if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
1614 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1615 unsigned Width = 32 - Srl_imm;
1616 int LSB = Srl_imm - Shl_imm;
1619 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1620 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1621 CurDAG->getTargetConstant(LSB, MVT::i32),
1622 CurDAG->getTargetConstant(Width, MVT::i32),
1623 getAL(CurDAG), Reg0 };
1624 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1630 SDNode *ARMDAGToDAGISel::
1631 SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1632 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1635 if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
1636 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1637 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1640 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1641 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1642 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1643 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1645 llvm_unreachable("Unknown so_reg opcode!");
1649 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1650 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1651 SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
1652 return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
1657 SDNode *ARMDAGToDAGISel::
1658 SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1659 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1663 if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
1664 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1665 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
1666 return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
1671 SDNode *ARMDAGToDAGISel::
1672 SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1673 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1674 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1678 if (Predicate_t2_so_imm(TrueVal.getNode())) {
1679 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1680 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1681 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1682 return CurDAG->SelectNodeTo(N,
1683 ARM::t2MOVCCi, MVT::i32, Ops, 5);
1688 SDNode *ARMDAGToDAGISel::
1689 SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1690 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1691 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1695 if (Predicate_so_imm(TrueVal.getNode())) {
1696 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1697 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1698 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1699 return CurDAG->SelectNodeTo(N,
1700 ARM::MOVCCi, MVT::i32, Ops, 5);
1705 SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
1706 EVT VT = N->getValueType(0);
1707 SDValue FalseVal = N->getOperand(0);
1708 SDValue TrueVal = N->getOperand(1);
1709 SDValue CC = N->getOperand(2);
1710 SDValue CCR = N->getOperand(3);
1711 SDValue InFlag = N->getOperand(4);
1712 assert(CC.getOpcode() == ISD::Constant);
1713 assert(CCR.getOpcode() == ISD::Register);
1714 ARMCC::CondCodes CCVal =
1715 (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
1717 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1718 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1719 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1720 // Pattern complexity = 18 cost = 1 size = 0
1724 if (Subtarget->isThumb()) {
1725 SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
1726 CCVal, CCR, InFlag);
1728 Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
1729 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1733 SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
1734 CCVal, CCR, InFlag);
1736 Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
1737 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1742 // Pattern: (ARMcmov:i32 GPR:i32:$false,
1743 // (imm:i32)<<P:Predicate_so_imm>>:$true,
1745 // Emits: (MOVCCi:i32 GPR:i32:$false,
1746 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1747 // Pattern complexity = 10 cost = 1 size = 0
1748 if (Subtarget->isThumb()) {
1749 SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
1750 CCVal, CCR, InFlag);
1752 Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
1753 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1757 SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
1758 CCVal, CCR, InFlag);
1760 Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
1761 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1767 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1768 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1769 // Pattern complexity = 6 cost = 1 size = 0
1771 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1772 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1773 // Pattern complexity = 6 cost = 11 size = 0
1775 // Also FCPYScc and FCPYDcc.
1776 SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
1777 SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
1779 switch (VT.getSimpleVT().SimpleTy) {
1780 default: assert(false && "Illegal conditional move type!");
1783 Opc = Subtarget->isThumb()
1784 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1794 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1797 SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
1798 // The only time a CONCAT_VECTORS operation can have legal types is when
1799 // two 64-bit vectors are concatenated to a 128-bit vector.
1800 EVT VT = N->getValueType(0);
1801 if (!VT.is128BitVector() || N->getNumOperands() != 2)
1802 llvm_unreachable("unexpected CONCAT_VECTORS");
1803 DebugLoc dl = N->getDebugLoc();
1804 SDValue V0 = N->getOperand(0);
1805 SDValue V1 = N->getOperand(1);
1806 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1807 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1808 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
1809 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
1812 SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
1813 DebugLoc dl = N->getDebugLoc();
1815 if (N->isMachineOpcode())
1816 return NULL; // Already selected.
1818 switch (N->getOpcode()) {
1820 case ISD::Constant: {
1821 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1823 if (Subtarget->hasThumb2())
1824 // Thumb2-aware targets have the MOVT instruction, so all immediates can
1825 // be done with MOV + MOVT, at worst.
1828 if (Subtarget->isThumb()) {
1829 UseCP = (Val > 255 && // MOV
1830 ~Val > 255 && // MOV + MVN
1831 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
1833 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
1834 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
1835 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
1840 CurDAG->getTargetConstantPool(ConstantInt::get(
1841 Type::getInt32Ty(*CurDAG->getContext()), Val),
1842 TLI.getPointerTy());
1845 if (Subtarget->isThumb1Only()) {
1846 SDValue Pred = getAL(CurDAG);
1847 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1848 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1849 ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1854 CurDAG->getRegister(0, MVT::i32),
1855 CurDAG->getTargetConstant(0, MVT::i32),
1857 CurDAG->getRegister(0, MVT::i32),
1858 CurDAG->getEntryNode()
1860 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1863 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
1867 // Other cases are autogenerated.
1870 case ISD::FrameIndex: {
1871 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1872 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1873 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1874 if (Subtarget->isThumb1Only()) {
1875 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1876 CurDAG->getTargetConstant(0, MVT::i32));
1878 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1879 ARM::t2ADDri : ARM::ADDri);
1880 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1881 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1882 CurDAG->getRegister(0, MVT::i32) };
1883 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1887 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1891 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true))
1895 if (Subtarget->isThumb1Only())
1897 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1898 unsigned RHSV = C->getZExtValue();
1900 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
1901 unsigned ShImm = Log2_32(RHSV-1);
1904 SDValue V = N->getOperand(0);
1905 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1906 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1907 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1908 if (Subtarget->isThumb()) {
1909 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1910 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1912 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1913 return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1916 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
1917 unsigned ShImm = Log2_32(RHSV+1);
1920 SDValue V = N->getOperand(0);
1921 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1922 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1923 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1924 if (Subtarget->isThumb()) {
1925 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1926 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 6);
1928 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1929 return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1935 // Check for unsigned bitfield extract
1936 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1939 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
1940 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
1941 // are entirely contributed by c2 and lower 16-bits are entirely contributed
1942 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
1943 // Select it to: "movt x, ((c1 & 0xffff) >> 16)
1944 EVT VT = N->getValueType(0);
1947 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
1949 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
1952 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
1953 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1956 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
1957 SDValue N2 = N0.getOperand(1);
1958 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1961 unsigned N1CVal = N1C->getZExtValue();
1962 unsigned N2CVal = N2C->getZExtValue();
1963 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
1964 (N1CVal & 0xffffU) == 0xffffU &&
1965 (N2CVal & 0xffffU) == 0x0U) {
1966 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
1968 SDValue Ops[] = { N0.getOperand(0), Imm16,
1969 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1970 return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
1975 case ARMISD::VMOVRRD:
1976 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
1977 N->getOperand(0), getAL(CurDAG),
1978 CurDAG->getRegister(0, MVT::i32));
1979 case ISD::UMUL_LOHI: {
1980 if (Subtarget->isThumb1Only())
1982 if (Subtarget->isThumb()) {
1983 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1984 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1985 CurDAG->getRegister(0, MVT::i32) };
1986 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32,Ops,4);
1988 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1989 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1990 CurDAG->getRegister(0, MVT::i32) };
1991 return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1994 case ISD::SMUL_LOHI: {
1995 if (Subtarget->isThumb1Only())
1997 if (Subtarget->isThumb()) {
1998 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1999 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2000 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32,Ops,4);
2002 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2003 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2004 CurDAG->getRegister(0, MVT::i32) };
2005 return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
2009 SDNode *ResNode = 0;
2010 if (Subtarget->isThumb() && Subtarget->hasThumb2())
2011 ResNode = SelectT2IndexedLoad(N);
2013 ResNode = SelectARMIndexedLoad(N);
2017 // VLDMQ must be custom-selected for "v2f64 load" to set the AM5Opc value.
2018 if (Subtarget->hasVFP2() &&
2019 N->getValueType(0).getSimpleVT().SimpleTy == MVT::v2f64) {
2020 SDValue Chain = N->getOperand(0);
2022 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
2023 SDValue Pred = getAL(CurDAG);
2024 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2025 SDValue Ops[] = { N->getOperand(1), AM5Opc, Pred, PredReg, Chain };
2026 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2027 MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
2028 SDNode *Ret = CurDAG->getMachineNode(ARM::VLDMQ, dl,
2029 MVT::v2f64, MVT::Other, Ops, 5);
2030 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
2033 // Other cases are autogenerated.
2037 // VSTMQ must be custom-selected for "v2f64 store" to set the AM5Opc value.
2038 if (Subtarget->hasVFP2() &&
2039 N->getOperand(1).getValueType().getSimpleVT().SimpleTy == MVT::v2f64) {
2040 SDValue Chain = N->getOperand(0);
2042 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
2043 SDValue Pred = getAL(CurDAG);
2044 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2045 SDValue Ops[] = { N->getOperand(1), N->getOperand(2),
2046 AM5Opc, Pred, PredReg, Chain };
2047 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2048 MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
2049 SDNode *Ret = CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
2050 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
2053 // Other cases are autogenerated.
2056 case ARMISD::BRCOND: {
2057 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2058 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2059 // Pattern complexity = 6 cost = 1 size = 0
2061 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2062 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
2063 // Pattern complexity = 6 cost = 1 size = 0
2065 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2066 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2067 // Pattern complexity = 6 cost = 1 size = 0
2069 unsigned Opc = Subtarget->isThumb() ?
2070 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
2071 SDValue Chain = N->getOperand(0);
2072 SDValue N1 = N->getOperand(1);
2073 SDValue N2 = N->getOperand(2);
2074 SDValue N3 = N->getOperand(3);
2075 SDValue InFlag = N->getOperand(4);
2076 assert(N1.getOpcode() == ISD::BasicBlock);
2077 assert(N2.getOpcode() == ISD::Constant);
2078 assert(N3.getOpcode() == ISD::Register);
2080 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2081 cast<ConstantSDNode>(N2)->getZExtValue()),
2083 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
2084 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
2086 Chain = SDValue(ResNode, 0);
2087 if (N->getNumValues() == 2) {
2088 InFlag = SDValue(ResNode, 1);
2089 ReplaceUses(SDValue(N, 1), InFlag);
2091 ReplaceUses(SDValue(N, 0),
2092 SDValue(Chain.getNode(), Chain.getResNo()));
2096 return SelectCMOVOp(N);
2097 case ARMISD::CNEG: {
2098 EVT VT = N->getValueType(0);
2099 SDValue N0 = N->getOperand(0);
2100 SDValue N1 = N->getOperand(1);
2101 SDValue N2 = N->getOperand(2);
2102 SDValue N3 = N->getOperand(3);
2103 SDValue InFlag = N->getOperand(4);
2104 assert(N2.getOpcode() == ISD::Constant);
2105 assert(N3.getOpcode() == ISD::Register);
2107 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2108 cast<ConstantSDNode>(N2)->getZExtValue()),
2110 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
2112 switch (VT.getSimpleVT().SimpleTy) {
2113 default: assert(false && "Illegal conditional move type!");
2122 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
2125 case ARMISD::VZIP: {
2127 EVT VT = N->getValueType(0);
2128 switch (VT.getSimpleVT().SimpleTy) {
2129 default: return NULL;
2130 case MVT::v8i8: Opc = ARM::VZIPd8; break;
2131 case MVT::v4i16: Opc = ARM::VZIPd16; break;
2133 case MVT::v2i32: Opc = ARM::VZIPd32; break;
2134 case MVT::v16i8: Opc = ARM::VZIPq8; break;
2135 case MVT::v8i16: Opc = ARM::VZIPq16; break;
2137 case MVT::v4i32: Opc = ARM::VZIPq32; break;
2139 SDValue Pred = getAL(CurDAG);
2140 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2141 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2142 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2144 case ARMISD::VUZP: {
2146 EVT VT = N->getValueType(0);
2147 switch (VT.getSimpleVT().SimpleTy) {
2148 default: return NULL;
2149 case MVT::v8i8: Opc = ARM::VUZPd8; break;
2150 case MVT::v4i16: Opc = ARM::VUZPd16; break;
2152 case MVT::v2i32: Opc = ARM::VUZPd32; break;
2153 case MVT::v16i8: Opc = ARM::VUZPq8; break;
2154 case MVT::v8i16: Opc = ARM::VUZPq16; break;
2156 case MVT::v4i32: Opc = ARM::VUZPq32; break;
2158 SDValue Pred = getAL(CurDAG);
2159 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2160 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2161 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2163 case ARMISD::VTRN: {
2165 EVT VT = N->getValueType(0);
2166 switch (VT.getSimpleVT().SimpleTy) {
2167 default: return NULL;
2168 case MVT::v8i8: Opc = ARM::VTRNd8; break;
2169 case MVT::v4i16: Opc = ARM::VTRNd16; break;
2171 case MVT::v2i32: Opc = ARM::VTRNd32; break;
2172 case MVT::v16i8: Opc = ARM::VTRNq8; break;
2173 case MVT::v8i16: Opc = ARM::VTRNq16; break;
2175 case MVT::v4i32: Opc = ARM::VTRNq32; break;
2177 SDValue Pred = getAL(CurDAG);
2178 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2179 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2180 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2182 case ARMISD::BUILD_VECTOR: {
2183 EVT VecVT = N->getValueType(0);
2184 EVT EltVT = VecVT.getVectorElementType();
2185 unsigned NumElts = VecVT.getVectorNumElements();
2186 if (EltVT.getSimpleVT() == MVT::f64) {
2187 assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
2188 return PairDRegs(VecVT, N->getOperand(0), N->getOperand(1));
2190 assert(EltVT.getSimpleVT() == MVT::f32 &&
2191 "unexpected type for BUILD_VECTOR");
2193 return PairSRegs(VecVT, N->getOperand(0), N->getOperand(1));
2194 assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
2195 return QuadSRegs(VecVT, N->getOperand(0), N->getOperand(1),
2196 N->getOperand(2), N->getOperand(3));
2199 case ISD::INTRINSIC_VOID:
2200 case ISD::INTRINSIC_W_CHAIN: {
2201 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2206 case Intrinsic::arm_neon_vld1: {
2207 unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
2208 ARM::VLD1d32, ARM::VLD1d64 };
2209 unsigned QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
2210 ARM::VLD1q32, ARM::VLD1q64 };
2211 return SelectVLD(N, 1, DOpcodes, QOpcodes, 0);
2214 case Intrinsic::arm_neon_vld2: {
2215 unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
2216 ARM::VLD2d32, ARM::VLD1q64 };
2217 unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
2218 return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
2221 case Intrinsic::arm_neon_vld3: {
2222 unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
2223 ARM::VLD3d32, ARM::VLD1d64T };
2224 unsigned QOpcodes0[] = { ARM::VLD3q8_UPD,
2227 unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD,
2228 ARM::VLD3q16odd_UPD,
2229 ARM::VLD3q32odd_UPD };
2230 return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2233 case Intrinsic::arm_neon_vld4: {
2234 unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
2235 ARM::VLD4d32, ARM::VLD1d64Q };
2236 unsigned QOpcodes0[] = { ARM::VLD4q8_UPD,
2239 unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD,
2240 ARM::VLD4q16odd_UPD,
2241 ARM::VLD4q32odd_UPD };
2242 return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2245 case Intrinsic::arm_neon_vld2lane: {
2246 unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
2247 unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 };
2248 unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd };
2249 return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
2252 case Intrinsic::arm_neon_vld3lane: {
2253 unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
2254 unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 };
2255 unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd };
2256 return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
2259 case Intrinsic::arm_neon_vld4lane: {
2260 unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
2261 unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 };
2262 unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd };
2263 return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
2266 case Intrinsic::arm_neon_vst1: {
2267 unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
2268 ARM::VST1d32, ARM::VST1d64 };
2269 unsigned QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
2270 ARM::VST1q32, ARM::VST1q64 };
2271 return SelectVST(N, 1, DOpcodes, QOpcodes, 0);
2274 case Intrinsic::arm_neon_vst2: {
2275 unsigned DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
2276 ARM::VST2d32, ARM::VST1q64 };
2277 unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
2278 return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
2281 case Intrinsic::arm_neon_vst3: {
2282 unsigned DOpcodes[] = { ARM::VST3d8, ARM::VST3d16,
2283 ARM::VST3d32, ARM::VST1d64T };
2284 unsigned QOpcodes0[] = { ARM::VST3q8_UPD,
2287 unsigned QOpcodes1[] = { ARM::VST3q8odd_UPD,
2288 ARM::VST3q16odd_UPD,
2289 ARM::VST3q32odd_UPD };
2290 return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2293 case Intrinsic::arm_neon_vst4: {
2294 unsigned DOpcodes[] = { ARM::VST4d8, ARM::VST4d16,
2295 ARM::VST4d32, ARM::VST1d64Q };
2296 unsigned QOpcodes0[] = { ARM::VST4q8_UPD,
2299 unsigned QOpcodes1[] = { ARM::VST4q8odd_UPD,
2300 ARM::VST4q16odd_UPD,
2301 ARM::VST4q32odd_UPD };
2302 return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2305 case Intrinsic::arm_neon_vst2lane: {
2306 unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
2307 unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 };
2308 unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd };
2309 return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
2312 case Intrinsic::arm_neon_vst3lane: {
2313 unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
2314 unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 };
2315 unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd };
2316 return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
2319 case Intrinsic::arm_neon_vst4lane: {
2320 unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
2321 unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 };
2322 unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd };
2323 return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
2329 case ISD::INTRINSIC_WO_CHAIN: {
2330 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2335 case Intrinsic::arm_neon_vtbl2:
2336 return SelectVTBL(N, false, 2, ARM::VTBL2);
2337 case Intrinsic::arm_neon_vtbl3:
2338 return SelectVTBL(N, false, 3, ARM::VTBL3);
2339 case Intrinsic::arm_neon_vtbl4:
2340 return SelectVTBL(N, false, 4, ARM::VTBL4);
2342 case Intrinsic::arm_neon_vtbx2:
2343 return SelectVTBL(N, true, 2, ARM::VTBX2);
2344 case Intrinsic::arm_neon_vtbx3:
2345 return SelectVTBL(N, true, 3, ARM::VTBX3);
2346 case Intrinsic::arm_neon_vtbx4:
2347 return SelectVTBL(N, true, 4, ARM::VTBX4);
2352 case ISD::CONCAT_VECTORS:
2353 return SelectConcatVector(N);
2356 return SelectCode(N);
2359 bool ARMDAGToDAGISel::
2360 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2361 std::vector<SDValue> &OutOps) {
2362 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
2363 // Require the address to be in a register. That is safe for all ARM
2364 // variants and it is hard to do anything much smarter without knowing
2365 // how the operand is used.
2366 OutOps.push_back(Op);
2370 /// createARMISelDag - This pass converts a legalized DAG into a
2371 /// ARM-specific DAG, ready for instruction scheduling.
2373 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
2374 CodeGenOpt::Level OptLevel) {
2375 return new ARMDAGToDAGISel(TM, OptLevel);