1 //===-- ARMISelDAGToDAG.cpp - A dag to dag inst selector for ARM ----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines an instruction selector for the ARM target.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "arm-isel"
16 #include "ARMAddressingModes.h"
17 #include "ARMTargetMachine.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Function.h"
22 #include "llvm/Intrinsics.h"
23 #include "llvm/LLVMContext.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/CodeGen/SelectionDAGISel.h"
29 #include "llvm/Target/TargetLowering.h"
30 #include "llvm/Target/TargetOptions.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Compiler.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
40 DisableShifterOp("disable-shifter-op", cl::Hidden,
41 cl::desc("Disable isel of shifter-op"),
44 //===--------------------------------------------------------------------===//
45 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
46 /// instructions for SelectionDAG operations.
49 class ARMDAGToDAGISel : public SelectionDAGISel {
50 ARMBaseTargetMachine &TM;
52 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
53 /// make the right decision when generating code for different targets.
54 const ARMSubtarget *Subtarget;
57 explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
58 CodeGenOpt::Level OptLevel)
59 : SelectionDAGISel(tm, OptLevel), TM(tm),
60 Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
63 virtual const char *getPassName() const {
64 return "ARM Instruction Selection";
67 /// getI32Imm - Return a target constant of type i32 with the specified
69 inline SDValue getI32Imm(unsigned Imm) {
70 return CurDAG->getTargetConstant(Imm, MVT::i32);
73 SDNode *Select(SDNode *N);
75 bool SelectShifterOperandReg(SDNode *Op, SDValue N, SDValue &A,
76 SDValue &B, SDValue &C);
77 bool SelectAddrMode2(SDNode *Op, SDValue N, SDValue &Base,
78 SDValue &Offset, SDValue &Opc);
79 bool SelectAddrMode2Offset(SDNode *Op, SDValue N,
80 SDValue &Offset, SDValue &Opc);
81 bool SelectAddrMode3(SDNode *Op, SDValue N, SDValue &Base,
82 SDValue &Offset, SDValue &Opc);
83 bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
84 SDValue &Offset, SDValue &Opc);
85 bool SelectAddrMode4(SDNode *Op, SDValue N, SDValue &Addr,
87 bool SelectAddrMode5(SDNode *Op, SDValue N, SDValue &Base,
89 bool SelectAddrMode6(SDNode *Op, SDValue N, SDValue &Addr, SDValue &Align);
91 bool SelectAddrModePC(SDNode *Op, SDValue N, SDValue &Offset,
94 bool SelectThumbAddrModeRR(SDNode *Op, SDValue N, SDValue &Base,
96 bool SelectThumbAddrModeRI5(SDNode *Op, SDValue N, unsigned Scale,
97 SDValue &Base, SDValue &OffImm,
99 bool SelectThumbAddrModeS1(SDNode *Op, SDValue N, SDValue &Base,
100 SDValue &OffImm, SDValue &Offset);
101 bool SelectThumbAddrModeS2(SDNode *Op, SDValue N, SDValue &Base,
102 SDValue &OffImm, SDValue &Offset);
103 bool SelectThumbAddrModeS4(SDNode *Op, SDValue N, SDValue &Base,
104 SDValue &OffImm, SDValue &Offset);
105 bool SelectThumbAddrModeSP(SDNode *Op, SDValue N, SDValue &Base,
108 bool SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
109 SDValue &BaseReg, SDValue &Opc);
110 bool SelectT2AddrModeImm12(SDNode *Op, SDValue N, SDValue &Base,
112 bool SelectT2AddrModeImm8(SDNode *Op, SDValue N, SDValue &Base,
114 bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
116 bool SelectT2AddrModeImm8s4(SDNode *Op, SDValue N, SDValue &Base,
118 bool SelectT2AddrModeSoReg(SDNode *Op, SDValue N, SDValue &Base,
119 SDValue &OffReg, SDValue &ShImm);
121 inline bool Pred_so_imm(SDNode *inN) const {
122 ConstantSDNode *N = cast<ConstantSDNode>(inN);
123 return ARM_AM::getSOImmVal(N->getZExtValue()) != -1;
126 inline bool Pred_t2_so_imm(SDNode *inN) const {
127 ConstantSDNode *N = cast<ConstantSDNode>(inN);
128 return ARM_AM::getT2SOImmVal(N->getZExtValue()) != -1;
131 // Include the pieces autogenerated from the target description.
132 #include "ARMGenDAGISel.inc"
135 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
137 SDNode *SelectARMIndexedLoad(SDNode *N);
138 SDNode *SelectT2IndexedLoad(SDNode *N);
140 /// SelectVLD - Select NEON load intrinsics. NumVecs should be
141 /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for
142 /// loads of D registers and even subregs and odd subregs of Q registers.
143 /// For NumVecs <= 2, QOpcodes1 is not used.
144 SDNode *SelectVLD(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
145 unsigned *QOpcodes0, unsigned *QOpcodes1);
147 /// SelectVST - Select NEON store intrinsics. NumVecs should
148 /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for
149 /// stores of D registers and even subregs and odd subregs of Q registers.
150 /// For NumVecs <= 2, QOpcodes1 is not used.
151 SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
152 unsigned *QOpcodes0, unsigned *QOpcodes1);
154 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
155 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
156 /// load/store of D registers and even subregs and odd subregs of Q registers.
157 SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, unsigned NumVecs,
158 unsigned *DOpcodes, unsigned *QOpcodes0,
159 unsigned *QOpcodes1);
161 /// SelectVTBL - Select NEON VTBL and VTBX intrinsics. NumVecs should be 2,
162 /// 3 or 4. These are custom-selected so that a REG_SEQUENCE can be
163 /// generated to force the table registers to be consecutive.
164 SDNode *SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs, unsigned Opc);
166 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
167 SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
169 /// SelectCMOVOp - Select CMOV instructions for ARM.
170 SDNode *SelectCMOVOp(SDNode *N);
171 SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
172 ARMCC::CondCodes CCVal, SDValue CCR,
174 SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
175 ARMCC::CondCodes CCVal, SDValue CCR,
177 SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
178 ARMCC::CondCodes CCVal, SDValue CCR,
180 SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
181 ARMCC::CondCodes CCVal, SDValue CCR,
184 SDNode *SelectConcatVector(SDNode *N);
186 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
187 /// inline asm expressions.
188 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
190 std::vector<SDValue> &OutOps);
192 // Form pairs of consecutive S, D, or Q registers.
193 SDNode *PairSRegs(EVT VT, SDValue V0, SDValue V1);
194 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
195 SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1);
197 // Form sequences of 4 consecutive S, D, or Q registers.
198 SDNode *QuadSRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
199 SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
200 SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
202 // Form sequences of 8 consecutive D registers.
203 SDNode *OctoDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3,
204 SDValue V4, SDValue V5, SDValue V6, SDValue V7);
208 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
209 /// operand. If so Imm will receive the 32-bit value.
210 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
211 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
212 Imm = cast<ConstantSDNode>(N)->getZExtValue();
218 // isInt32Immediate - This method tests to see if a constant operand.
219 // If so Imm will receive the 32 bit value.
220 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
221 return isInt32Immediate(N.getNode(), Imm);
224 // isOpcWithIntImmediate - This method tests to see if the node is a specific
225 // opcode and that it has a immediate integer right operand.
226 // If so Imm will receive the 32 bit value.
227 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
228 return N->getOpcode() == Opc &&
229 isInt32Immediate(N->getOperand(1).getNode(), Imm);
233 bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
238 if (DisableShifterOp)
241 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
243 // Don't match base register only case. That is matched to a separate
244 // lower complexity pattern with explicit register operand.
245 if (ShOpcVal == ARM_AM::no_shift) return false;
247 BaseReg = N.getOperand(0);
248 unsigned ShImmVal = 0;
249 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
250 ShReg = CurDAG->getRegister(0, MVT::i32);
251 ShImmVal = RHS->getZExtValue() & 31;
253 ShReg = N.getOperand(1);
255 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
260 bool ARMDAGToDAGISel::SelectAddrMode2(SDNode *Op, SDValue N,
261 SDValue &Base, SDValue &Offset,
263 if (N.getOpcode() == ISD::MUL) {
264 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
265 // X * [3,5,9] -> X + X * [2,4,8] etc.
266 int RHSC = (int)RHS->getZExtValue();
269 ARM_AM::AddrOpc AddSub = ARM_AM::add;
271 AddSub = ARM_AM::sub;
274 if (isPowerOf2_32(RHSC)) {
275 unsigned ShAmt = Log2_32(RHSC);
276 Base = Offset = N.getOperand(0);
277 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
286 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
288 if (N.getOpcode() == ISD::FrameIndex) {
289 int FI = cast<FrameIndexSDNode>(N)->getIndex();
290 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
291 } else if (N.getOpcode() == ARMISD::Wrapper &&
292 !(Subtarget->useMovt() &&
293 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
294 Base = N.getOperand(0);
296 Offset = CurDAG->getRegister(0, MVT::i32);
297 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
303 // Match simple R +/- imm12 operands.
304 if (N.getOpcode() == ISD::ADD)
305 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
306 int RHSC = (int)RHS->getZExtValue();
307 if ((RHSC >= 0 && RHSC < 0x1000) ||
308 (RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
309 Base = N.getOperand(0);
310 if (Base.getOpcode() == ISD::FrameIndex) {
311 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
312 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
314 Offset = CurDAG->getRegister(0, MVT::i32);
316 ARM_AM::AddrOpc AddSub = ARM_AM::add;
318 AddSub = ARM_AM::sub;
321 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
328 // Otherwise this is R +/- [possibly shifted] R.
329 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
330 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
333 Base = N.getOperand(0);
334 Offset = N.getOperand(1);
336 if (ShOpcVal != ARM_AM::no_shift) {
337 // Check to see if the RHS of the shift is a constant, if not, we can't fold
339 if (ConstantSDNode *Sh =
340 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
341 ShAmt = Sh->getZExtValue();
342 Offset = N.getOperand(1).getOperand(0);
344 ShOpcVal = ARM_AM::no_shift;
348 // Try matching (R shl C) + (R).
349 if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
350 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
351 if (ShOpcVal != ARM_AM::no_shift) {
352 // Check to see if the RHS of the shift is a constant, if not, we can't
354 if (ConstantSDNode *Sh =
355 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
356 ShAmt = Sh->getZExtValue();
357 Offset = N.getOperand(0).getOperand(0);
358 Base = N.getOperand(1);
360 ShOpcVal = ARM_AM::no_shift;
365 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
370 bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
371 SDValue &Offset, SDValue &Opc) {
372 unsigned Opcode = Op->getOpcode();
373 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
374 ? cast<LoadSDNode>(Op)->getAddressingMode()
375 : cast<StoreSDNode>(Op)->getAddressingMode();
376 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
377 ? ARM_AM::add : ARM_AM::sub;
378 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
379 int Val = (int)C->getZExtValue();
380 if (Val >= 0 && Val < 0x1000) { // 12 bits.
381 Offset = CurDAG->getRegister(0, MVT::i32);
382 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
390 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
392 if (ShOpcVal != ARM_AM::no_shift) {
393 // Check to see if the RHS of the shift is a constant, if not, we can't fold
395 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
396 ShAmt = Sh->getZExtValue();
397 Offset = N.getOperand(0);
399 ShOpcVal = ARM_AM::no_shift;
403 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
409 bool ARMDAGToDAGISel::SelectAddrMode3(SDNode *Op, SDValue N,
410 SDValue &Base, SDValue &Offset,
412 if (N.getOpcode() == ISD::SUB) {
413 // X - C is canonicalize to X + -C, no need to handle it here.
414 Base = N.getOperand(0);
415 Offset = N.getOperand(1);
416 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
420 if (N.getOpcode() != ISD::ADD) {
422 if (N.getOpcode() == ISD::FrameIndex) {
423 int FI = cast<FrameIndexSDNode>(N)->getIndex();
424 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
426 Offset = CurDAG->getRegister(0, MVT::i32);
427 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
431 // If the RHS is +/- imm8, fold into addr mode.
432 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
433 int RHSC = (int)RHS->getZExtValue();
434 if ((RHSC >= 0 && RHSC < 256) ||
435 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
436 Base = N.getOperand(0);
437 if (Base.getOpcode() == ISD::FrameIndex) {
438 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
439 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
441 Offset = CurDAG->getRegister(0, MVT::i32);
443 ARM_AM::AddrOpc AddSub = ARM_AM::add;
445 AddSub = ARM_AM::sub;
448 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
453 Base = N.getOperand(0);
454 Offset = N.getOperand(1);
455 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
459 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
460 SDValue &Offset, SDValue &Opc) {
461 unsigned Opcode = Op->getOpcode();
462 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
463 ? cast<LoadSDNode>(Op)->getAddressingMode()
464 : cast<StoreSDNode>(Op)->getAddressingMode();
465 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
466 ? ARM_AM::add : ARM_AM::sub;
467 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
468 int Val = (int)C->getZExtValue();
469 if (Val >= 0 && Val < 256) {
470 Offset = CurDAG->getRegister(0, MVT::i32);
471 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
477 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
481 bool ARMDAGToDAGISel::SelectAddrMode4(SDNode *Op, SDValue N,
482 SDValue &Addr, SDValue &Mode) {
484 Mode = CurDAG->getTargetConstant(0, MVT::i32);
488 bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N,
489 SDValue &Base, SDValue &Offset) {
490 if (N.getOpcode() != ISD::ADD) {
492 if (N.getOpcode() == ISD::FrameIndex) {
493 int FI = cast<FrameIndexSDNode>(N)->getIndex();
494 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
495 } else if (N.getOpcode() == ARMISD::Wrapper &&
496 !(Subtarget->useMovt() &&
497 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
498 Base = N.getOperand(0);
500 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
505 // If the RHS is +/- imm8, fold into addr mode.
506 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
507 int RHSC = (int)RHS->getZExtValue();
508 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied by 4.
510 if ((RHSC >= 0 && RHSC < 256) ||
511 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
512 Base = N.getOperand(0);
513 if (Base.getOpcode() == ISD::FrameIndex) {
514 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
515 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
518 ARM_AM::AddrOpc AddSub = ARM_AM::add;
520 AddSub = ARM_AM::sub;
523 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
531 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
536 bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N,
537 SDValue &Addr, SDValue &Align) {
539 // Default to no alignment.
540 Align = CurDAG->getTargetConstant(0, MVT::i32);
544 bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
545 SDValue &Offset, SDValue &Label) {
546 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
547 Offset = N.getOperand(0);
548 SDValue N1 = N.getOperand(1);
549 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
556 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
557 SDValue &Base, SDValue &Offset){
558 // FIXME dl should come from the parent load or store, not the address
559 if (N.getOpcode() != ISD::ADD) {
560 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
561 if (!NC || !NC->isNullValue())
568 Base = N.getOperand(0);
569 Offset = N.getOperand(1);
574 ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
575 unsigned Scale, SDValue &Base,
576 SDValue &OffImm, SDValue &Offset) {
578 SDValue TmpBase, TmpOffImm;
579 if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
580 return false; // We want to select tLDRspi / tSTRspi instead.
581 if (N.getOpcode() == ARMISD::Wrapper &&
582 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
583 return false; // We want to select tLDRpci instead.
586 if (N.getOpcode() != ISD::ADD) {
587 if (N.getOpcode() == ARMISD::Wrapper &&
588 !(Subtarget->useMovt() &&
589 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
590 Base = N.getOperand(0);
594 Offset = CurDAG->getRegister(0, MVT::i32);
595 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
599 // Thumb does not have [sp, r] address mode.
600 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
601 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
602 if ((LHSR && LHSR->getReg() == ARM::SP) ||
603 (RHSR && RHSR->getReg() == ARM::SP)) {
605 Offset = CurDAG->getRegister(0, MVT::i32);
606 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
610 // If the RHS is + imm5 * scale, fold into addr mode.
611 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
612 int RHSC = (int)RHS->getZExtValue();
613 if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
615 if (RHSC >= 0 && RHSC < 32) {
616 Base = N.getOperand(0);
617 Offset = CurDAG->getRegister(0, MVT::i32);
618 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
624 Base = N.getOperand(0);
625 Offset = N.getOperand(1);
626 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
630 bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
631 SDValue &Base, SDValue &OffImm,
633 return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
636 bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
637 SDValue &Base, SDValue &OffImm,
639 return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
642 bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
643 SDValue &Base, SDValue &OffImm,
645 return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
648 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
649 SDValue &Base, SDValue &OffImm) {
650 if (N.getOpcode() == ISD::FrameIndex) {
651 int FI = cast<FrameIndexSDNode>(N)->getIndex();
652 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
653 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
657 if (N.getOpcode() != ISD::ADD)
660 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
661 if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
662 (LHSR && LHSR->getReg() == ARM::SP)) {
663 // If the RHS is + imm8 * scale, fold into addr mode.
664 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
665 int RHSC = (int)RHS->getZExtValue();
666 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
668 if (RHSC >= 0 && RHSC < 256) {
669 Base = N.getOperand(0);
670 if (Base.getOpcode() == ISD::FrameIndex) {
671 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
672 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
674 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
684 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
687 if (DisableShifterOp)
690 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
692 // Don't match base register only case. That is matched to a separate
693 // lower complexity pattern with explicit register operand.
694 if (ShOpcVal == ARM_AM::no_shift) return false;
696 BaseReg = N.getOperand(0);
697 unsigned ShImmVal = 0;
698 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
699 ShImmVal = RHS->getZExtValue() & 31;
700 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
707 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
708 SDValue &Base, SDValue &OffImm) {
709 // Match simple R + imm12 operands.
712 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
713 if (N.getOpcode() == ISD::FrameIndex) {
714 // Match frame index...
715 int FI = cast<FrameIndexSDNode>(N)->getIndex();
716 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
717 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
719 } else if (N.getOpcode() == ARMISD::Wrapper &&
720 !(Subtarget->useMovt() &&
721 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
722 Base = N.getOperand(0);
723 if (Base.getOpcode() == ISD::TargetConstantPool)
724 return false; // We want to select t2LDRpci instead.
727 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
731 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
732 if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
733 // Let t2LDRi8 handle (R - imm8).
736 int RHSC = (int)RHS->getZExtValue();
737 if (N.getOpcode() == ISD::SUB)
740 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
741 Base = N.getOperand(0);
742 if (Base.getOpcode() == ISD::FrameIndex) {
743 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
744 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
746 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
753 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
757 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
758 SDValue &Base, SDValue &OffImm) {
759 // Match simple R - imm8 operands.
760 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
761 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
762 int RHSC = (int)RHS->getSExtValue();
763 if (N.getOpcode() == ISD::SUB)
766 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
767 Base = N.getOperand(0);
768 if (Base.getOpcode() == ISD::FrameIndex) {
769 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
770 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
772 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
781 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
783 unsigned Opcode = Op->getOpcode();
784 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
785 ? cast<LoadSDNode>(Op)->getAddressingMode()
786 : cast<StoreSDNode>(Op)->getAddressingMode();
787 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
788 int RHSC = (int)RHS->getZExtValue();
789 if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
790 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
791 ? CurDAG->getTargetConstant(RHSC, MVT::i32)
792 : CurDAG->getTargetConstant(-RHSC, MVT::i32);
800 bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
801 SDValue &Base, SDValue &OffImm) {
802 if (N.getOpcode() == ISD::ADD) {
803 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
804 int RHSC = (int)RHS->getZExtValue();
806 if (((RHSC & 0x3) == 0) &&
807 ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) {
808 Base = N.getOperand(0);
809 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
813 } else if (N.getOpcode() == ISD::SUB) {
814 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
815 int RHSC = (int)RHS->getZExtValue();
817 if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) {
818 Base = N.getOperand(0);
819 OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
828 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDNode *Op, SDValue N,
830 SDValue &OffReg, SDValue &ShImm) {
831 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
832 if (N.getOpcode() != ISD::ADD)
835 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
836 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
837 int RHSC = (int)RHS->getZExtValue();
838 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
840 else if (RHSC < 0 && RHSC >= -255) // 8 bits
844 // Look for (R + R) or (R + (R << [1,2,3])).
846 Base = N.getOperand(0);
847 OffReg = N.getOperand(1);
849 // Swap if it is ((R << c) + R).
850 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
851 if (ShOpcVal != ARM_AM::lsl) {
852 ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
853 if (ShOpcVal == ARM_AM::lsl)
854 std::swap(Base, OffReg);
857 if (ShOpcVal == ARM_AM::lsl) {
858 // Check to see if the RHS of the shift is a constant, if not, we can't fold
860 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
861 ShAmt = Sh->getZExtValue();
864 ShOpcVal = ARM_AM::no_shift;
866 OffReg = OffReg.getOperand(0);
868 ShOpcVal = ARM_AM::no_shift;
872 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
877 //===--------------------------------------------------------------------===//
879 /// getAL - Returns a ARMCC::AL immediate node.
880 static inline SDValue getAL(SelectionDAG *CurDAG) {
881 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
884 SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
885 LoadSDNode *LD = cast<LoadSDNode>(N);
886 ISD::MemIndexedMode AM = LD->getAddressingMode();
887 if (AM == ISD::UNINDEXED)
890 EVT LoadedVT = LD->getMemoryVT();
891 SDValue Offset, AMOpc;
892 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
895 if (LoadedVT == MVT::i32 &&
896 SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
897 Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
899 } else if (LoadedVT == MVT::i16 &&
900 SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
902 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
903 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
904 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
905 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
906 if (LD->getExtensionType() == ISD::SEXTLOAD) {
907 if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
909 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
912 if (SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
914 Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
920 SDValue Chain = LD->getChain();
921 SDValue Base = LD->getBasePtr();
922 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
923 CurDAG->getRegister(0, MVT::i32), Chain };
924 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
931 SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
932 LoadSDNode *LD = cast<LoadSDNode>(N);
933 ISD::MemIndexedMode AM = LD->getAddressingMode();
934 if (AM == ISD::UNINDEXED)
937 EVT LoadedVT = LD->getMemoryVT();
938 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
940 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
943 if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
944 switch (LoadedVT.getSimpleVT().SimpleTy) {
946 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
950 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
952 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
957 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
959 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
968 SDValue Chain = LD->getChain();
969 SDValue Base = LD->getBasePtr();
970 SDValue Ops[]= { Base, Offset, getAL(CurDAG),
971 CurDAG->getRegister(0, MVT::i32), Chain };
972 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
979 /// PairSRegs - Form a D register from a pair of S registers.
981 SDNode *ARMDAGToDAGISel::PairSRegs(EVT VT, SDValue V0, SDValue V1) {
982 DebugLoc dl = V0.getNode()->getDebugLoc();
983 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
984 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
985 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
986 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
989 /// PairDRegs - Form a quad register from a pair of D registers.
991 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
992 DebugLoc dl = V0.getNode()->getDebugLoc();
993 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
994 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
995 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
996 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
999 /// PairQRegs - Form 4 consecutive D registers from a pair of Q registers.
1001 SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) {
1002 DebugLoc dl = V0.getNode()->getDebugLoc();
1003 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
1004 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
1005 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
1006 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
1009 /// QuadSRegs - Form 4 consecutive S registers.
1011 SDNode *ARMDAGToDAGISel::QuadSRegs(EVT VT, SDValue V0, SDValue V1,
1012 SDValue V2, SDValue V3) {
1013 DebugLoc dl = V0.getNode()->getDebugLoc();
1014 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
1015 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
1016 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, MVT::i32);
1017 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, MVT::i32);
1018 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1019 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1022 /// QuadDRegs - Form 4 consecutive D registers.
1024 SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
1025 SDValue V2, SDValue V3) {
1026 DebugLoc dl = V0.getNode()->getDebugLoc();
1027 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1028 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1029 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
1030 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
1031 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1032 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1035 /// QuadQRegs - Form 4 consecutive Q registers.
1037 SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1,
1038 SDValue V2, SDValue V3) {
1039 DebugLoc dl = V0.getNode()->getDebugLoc();
1040 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
1041 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
1042 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, MVT::i32);
1043 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, MVT::i32);
1044 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1045 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1048 /// OctoDRegs - Form 8 consecutive D registers.
1050 SDNode *ARMDAGToDAGISel::OctoDRegs(EVT VT, SDValue V0, SDValue V1,
1051 SDValue V2, SDValue V3,
1052 SDValue V4, SDValue V5,
1053 SDValue V6, SDValue V7) {
1054 DebugLoc dl = V0.getNode()->getDebugLoc();
1055 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1056 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1057 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
1058 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
1059 SDValue SubReg4 = CurDAG->getTargetConstant(ARM::dsub_4, MVT::i32);
1060 SDValue SubReg5 = CurDAG->getTargetConstant(ARM::dsub_5, MVT::i32);
1061 SDValue SubReg6 = CurDAG->getTargetConstant(ARM::dsub_6, MVT::i32);
1062 SDValue SubReg7 = CurDAG->getTargetConstant(ARM::dsub_7, MVT::i32);
1063 const SDValue Ops[] ={ V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3,
1064 V4, SubReg4, V5, SubReg5, V6, SubReg6, V7, SubReg7 };
1065 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 16);
1068 /// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
1069 /// for a 64-bit subregister of the vector.
1070 static EVT GetNEONSubregVT(EVT VT) {
1071 switch (VT.getSimpleVT().SimpleTy) {
1072 default: llvm_unreachable("unhandled NEON type");
1073 case MVT::v16i8: return MVT::v8i8;
1074 case MVT::v8i16: return MVT::v4i16;
1075 case MVT::v4f32: return MVT::v2f32;
1076 case MVT::v4i32: return MVT::v2i32;
1077 case MVT::v2i64: return MVT::v1i64;
1081 SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
1082 unsigned *DOpcodes, unsigned *QOpcodes0,
1083 unsigned *QOpcodes1) {
1084 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1085 DebugLoc dl = N->getDebugLoc();
1087 SDValue MemAddr, Align;
1088 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1091 SDValue Chain = N->getOperand(0);
1092 EVT VT = N->getValueType(0);
1093 bool is64BitVector = VT.is64BitVector();
1095 unsigned OpcodeIndex;
1096 switch (VT.getSimpleVT().SimpleTy) {
1097 default: llvm_unreachable("unhandled vld type");
1098 // Double-register operations:
1099 case MVT::v8i8: OpcodeIndex = 0; break;
1100 case MVT::v4i16: OpcodeIndex = 1; break;
1102 case MVT::v2i32: OpcodeIndex = 2; break;
1103 case MVT::v1i64: OpcodeIndex = 3; break;
1104 // Quad-register operations:
1105 case MVT::v16i8: OpcodeIndex = 0; break;
1106 case MVT::v8i16: OpcodeIndex = 1; break;
1108 case MVT::v4i32: OpcodeIndex = 2; break;
1109 case MVT::v2i64: OpcodeIndex = 3;
1110 assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
1114 SDValue Pred = getAL(CurDAG);
1115 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1116 if (is64BitVector) {
1117 unsigned Opc = DOpcodes[OpcodeIndex];
1118 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1119 std::vector<EVT> ResTys(NumVecs, VT);
1120 ResTys.push_back(MVT::Other);
1121 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1126 SDValue V0 = SDValue(VLd, 0);
1127 SDValue V1 = SDValue(VLd, 1);
1129 // Form a REG_SEQUENCE to force register allocation.
1131 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1133 SDValue V2 = SDValue(VLd, 2);
1134 // If it's a vld3, form a quad D-register but discard the last part.
1135 SDValue V3 = (NumVecs == 3)
1136 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1138 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1141 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1142 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1143 SDValue D = CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec,
1145 ReplaceUses(SDValue(N, Vec), D);
1147 ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, NumVecs));
1151 EVT RegVT = GetNEONSubregVT(VT);
1153 // Quad registers are directly supported for VLD1 and VLD2,
1154 // loading pairs of D regs.
1155 unsigned Opc = QOpcodes0[OpcodeIndex];
1156 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1157 std::vector<EVT> ResTys(2 * NumVecs, RegVT);
1158 ResTys.push_back(MVT::Other);
1159 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1160 Chain = SDValue(VLd, 2 * NumVecs);
1162 // Combine the even and odd subregs to produce the result.
1164 SDNode *Q = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
1165 ReplaceUses(SDValue(N, 0), SDValue(Q, 0));
1167 SDValue QQ = SDValue(QuadDRegs(MVT::v4i64,
1168 SDValue(VLd, 0), SDValue(VLd, 1),
1169 SDValue(VLd, 2), SDValue(VLd, 3)), 0);
1170 SDValue Q0 = CurDAG->getTargetExtractSubreg(ARM::qsub_0, dl, VT, QQ);
1171 SDValue Q1 = CurDAG->getTargetExtractSubreg(ARM::qsub_1, dl, VT, QQ);
1172 ReplaceUses(SDValue(N, 0), Q0);
1173 ReplaceUses(SDValue(N, 1), Q1);
1176 // Otherwise, quad registers are loaded with two separate instructions,
1177 // where one loads the even registers and the other loads the odd registers.
1179 std::vector<EVT> ResTys(NumVecs, RegVT);
1180 ResTys.push_back(MemAddr.getValueType());
1181 ResTys.push_back(MVT::Other);
1183 // Load the even subregs.
1184 unsigned Opc = QOpcodes0[OpcodeIndex];
1185 const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain };
1186 SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6);
1187 Chain = SDValue(VLdA, NumVecs+1);
1189 // Load the odd subregs.
1190 Opc = QOpcodes1[OpcodeIndex];
1191 const SDValue OpsB[] = { SDValue(VLdA, NumVecs),
1192 Align, Reg0, Pred, Reg0, Chain };
1193 SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
1194 Chain = SDValue(VLdB, NumVecs+1);
1196 SDValue V0 = SDValue(VLdA, 0);
1197 SDValue V1 = SDValue(VLdB, 0);
1198 SDValue V2 = SDValue(VLdA, 1);
1199 SDValue V3 = SDValue(VLdB, 1);
1200 SDValue V4 = SDValue(VLdA, 2);
1201 SDValue V5 = SDValue(VLdB, 2);
1202 SDValue V6 = (NumVecs == 3)
1203 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0)
1205 SDValue V7 = (NumVecs == 3)
1206 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0)
1208 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V0, V1, V2, V3,
1209 V4, V5, V6, V7), 0);
1211 // Extract out the 3 / 4 Q registers.
1212 assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1213 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1214 SDValue Q = CurDAG->getTargetExtractSubreg(ARM::qsub_0+Vec,
1216 ReplaceUses(SDValue(N, Vec), Q);
1219 ReplaceUses(SDValue(N, NumVecs), Chain);
1223 SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
1224 unsigned *DOpcodes, unsigned *QOpcodes0,
1225 unsigned *QOpcodes1) {
1226 assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range");
1227 DebugLoc dl = N->getDebugLoc();
1229 SDValue MemAddr, Align;
1230 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1233 SDValue Chain = N->getOperand(0);
1234 EVT VT = N->getOperand(3).getValueType();
1235 bool is64BitVector = VT.is64BitVector();
1237 unsigned OpcodeIndex;
1238 switch (VT.getSimpleVT().SimpleTy) {
1239 default: llvm_unreachable("unhandled vst type");
1240 // Double-register operations:
1241 case MVT::v8i8: OpcodeIndex = 0; break;
1242 case MVT::v4i16: OpcodeIndex = 1; break;
1244 case MVT::v2i32: OpcodeIndex = 2; break;
1245 case MVT::v1i64: OpcodeIndex = 3; break;
1246 // Quad-register operations:
1247 case MVT::v16i8: OpcodeIndex = 0; break;
1248 case MVT::v8i16: OpcodeIndex = 1; break;
1250 case MVT::v4i32: OpcodeIndex = 2; break;
1251 case MVT::v2i64: OpcodeIndex = 3;
1252 assert(NumVecs == 1 && "v2i64 type only supported for VST1");
1256 SDValue Pred = getAL(CurDAG);
1257 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1259 SmallVector<SDValue, 10> Ops;
1260 Ops.push_back(MemAddr);
1261 Ops.push_back(Align);
1263 // FIXME: This is a temporary flag to distinguish VSTs that have been
1264 // converted to pseudo instructions.
1265 bool usePseudoInstrs = (NumVecs >= 3);
1267 if (is64BitVector) {
1270 SDValue V0 = N->getOperand(0+3);
1271 SDValue V1 = N->getOperand(1+3);
1273 // Form a REG_SEQUENCE to force register allocation.
1275 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1277 SDValue V2 = N->getOperand(2+3);
1278 // If it's a vld3, form a quad D-register and leave the last part as
1280 SDValue V3 = (NumVecs == 3)
1281 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1282 : N->getOperand(3+3);
1283 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1285 if (usePseudoInstrs)
1286 Ops.push_back(RegSeq);
1289 // Now extract the D registers back out.
1290 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT,
1292 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT,
1295 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,
1298 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,
1302 Ops.push_back(N->getOperand(3));
1304 Ops.push_back(Pred);
1305 Ops.push_back(Reg0); // predicate register
1306 Ops.push_back(Chain);
1307 unsigned Opc = DOpcodes[OpcodeIndex];
1308 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
1309 usePseudoInstrs ? 6 : NumVecs+5);
1312 EVT RegVT = GetNEONSubregVT(VT);
1314 // Quad registers are directly supported for VST1 and VST2,
1315 // storing pairs of D regs.
1316 unsigned Opc = QOpcodes0[OpcodeIndex];
1318 // First extract the pair of Q registers.
1319 SDValue Q0 = N->getOperand(3);
1320 SDValue Q1 = N->getOperand(4);
1322 // Form a QQ register.
1323 SDValue QQ = SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0);
1325 // Now extract the D registers back out.
1326 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1328 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1330 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, RegVT,
1332 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, RegVT,
1334 Ops.push_back(Pred);
1335 Ops.push_back(Reg0); // predicate register
1336 Ops.push_back(Chain);
1337 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 5 + 4);
1339 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1340 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1341 N->getOperand(Vec+3)));
1342 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1343 N->getOperand(Vec+3)));
1345 Ops.push_back(Pred);
1346 Ops.push_back(Reg0); // predicate register
1347 Ops.push_back(Chain);
1348 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
1353 // Otherwise, quad registers are stored with two separate instructions,
1354 // where one stores the even registers and the other stores the odd registers.
1356 // Form the QQQQ REG_SEQUENCE.
1358 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1359 V[i] = CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1360 N->getOperand(Vec+3));
1361 V[i+1] = CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1362 N->getOperand(Vec+3));
1365 V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1368 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1369 V[4], V[5], V[6], V[7]), 0);
1371 // Store the even D registers.
1372 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1373 Ops.push_back(Reg0); // post-access address offset
1374 if (usePseudoInstrs)
1375 Ops.push_back(RegSeq);
1377 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1378 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec*2, dl,
1380 Ops.push_back(Pred);
1381 Ops.push_back(Reg0); // predicate register
1382 Ops.push_back(Chain);
1383 unsigned Opc = QOpcodes0[OpcodeIndex];
1384 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1385 MVT::Other, Ops.data(),
1386 usePseudoInstrs ? 7 : NumVecs+6);
1387 Chain = SDValue(VStA, 1);
1389 // Store the odd D registers.
1390 Ops[0] = SDValue(VStA, 0); // MemAddr
1391 if (usePseudoInstrs)
1394 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1395 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::dsub_1+Vec*2, dl,
1397 Ops[NumVecs+5] = Chain;
1399 Opc = QOpcodes1[OpcodeIndex];
1400 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1401 MVT::Other, Ops.data(),
1402 usePseudoInstrs ? 7 : NumVecs+6);
1403 Chain = SDValue(VStB, 1);
1404 ReplaceUses(SDValue(N, 0), Chain);
1408 SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1409 unsigned NumVecs, unsigned *DOpcodes,
1410 unsigned *QOpcodes0,
1411 unsigned *QOpcodes1) {
1412 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1413 DebugLoc dl = N->getDebugLoc();
1415 SDValue MemAddr, Align;
1416 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1419 SDValue Chain = N->getOperand(0);
1421 cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1422 EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
1423 bool is64BitVector = VT.is64BitVector();
1425 // Quad registers are handled by load/store of subregs. Find the subreg info.
1426 unsigned NumElts = 0;
1429 if (!is64BitVector) {
1430 RegVT = GetNEONSubregVT(VT);
1431 NumElts = RegVT.getVectorNumElements();
1432 Even = Lane < NumElts;
1435 unsigned OpcodeIndex;
1436 switch (VT.getSimpleVT().SimpleTy) {
1437 default: llvm_unreachable("unhandled vld/vst lane type");
1438 // Double-register operations:
1439 case MVT::v8i8: OpcodeIndex = 0; break;
1440 case MVT::v4i16: OpcodeIndex = 1; break;
1442 case MVT::v2i32: OpcodeIndex = 2; break;
1443 // Quad-register operations:
1444 case MVT::v8i16: OpcodeIndex = 0; break;
1446 case MVT::v4i32: OpcodeIndex = 1; break;
1449 SDValue Pred = getAL(CurDAG);
1450 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1452 SmallVector<SDValue, 10> Ops;
1453 Ops.push_back(MemAddr);
1454 Ops.push_back(Align);
1457 if (is64BitVector) {
1458 Opc = DOpcodes[OpcodeIndex];
1460 SDValue V0 = N->getOperand(0+3);
1461 SDValue V1 = N->getOperand(1+3);
1463 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1465 SDValue V2 = N->getOperand(2+3);
1466 SDValue V3 = (NumVecs == 3)
1467 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1468 : N->getOperand(3+3);
1469 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1472 // Now extract the D registers back out.
1473 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT, RegSeq));
1474 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT, RegSeq));
1476 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,RegSeq));
1478 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,RegSeq));
1480 // Check if this is loading the even or odd subreg of a Q register.
1481 if (Lane < NumElts) {
1482 Opc = QOpcodes0[OpcodeIndex];
1485 Opc = QOpcodes1[OpcodeIndex];
1489 SDValue V0 = N->getOperand(0+3);
1490 SDValue V1 = N->getOperand(1+3);
1492 RegSeq = SDValue(PairQRegs(MVT::v4i64, V0, V1), 0);
1494 SDValue V2 = N->getOperand(2+3);
1495 SDValue V3 = (NumVecs == 3)
1496 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1497 : N->getOperand(3+3);
1498 RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
1501 // Extract the subregs of the input vector.
1502 unsigned SubIdx = Even ? ARM::dsub_0 : ARM::dsub_1;
1503 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1504 Ops.push_back(CurDAG->getTargetExtractSubreg(SubIdx+Vec*2, dl, RegVT,
1507 Ops.push_back(getI32Imm(Lane));
1508 Ops.push_back(Pred);
1509 Ops.push_back(Reg0);
1510 Ops.push_back(Chain);
1513 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6);
1515 std::vector<EVT> ResTys(NumVecs, RegVT);
1516 ResTys.push_back(MVT::Other);
1517 SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(),NumVecs+6);
1519 // Form a REG_SEQUENCE to force register allocation.
1521 if (is64BitVector) {
1522 SDValue V0 = SDValue(VLdLn, 0);
1523 SDValue V1 = SDValue(VLdLn, 1);
1525 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1527 SDValue V2 = SDValue(VLdLn, 2);
1528 // If it's a vld3, form a quad D-register but discard the last part.
1529 SDValue V3 = (NumVecs == 3)
1530 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1531 : SDValue(VLdLn, 3);
1532 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1535 // For 128-bit vectors, take the 64-bit results of the load and insert
1536 // them as subregs into the result.
1538 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1540 V[i] = SDValue(VLdLn, Vec);
1541 V[i+1] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1544 V[i] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1546 V[i+1] = SDValue(VLdLn, Vec);
1550 V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1554 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V[0], V[1], V[2], V[3]), 0);
1556 RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1557 V[4], V[5], V[6], V[7]), 0);
1560 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1561 assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1562 unsigned SubIdx = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
1563 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1564 ReplaceUses(SDValue(N, Vec),
1565 CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, RegSeq));
1566 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, NumVecs));
1570 SDNode *ARMDAGToDAGISel::SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs,
1572 assert(NumVecs >= 2 && NumVecs <= 4 && "VTBL NumVecs out-of-range");
1573 DebugLoc dl = N->getDebugLoc();
1574 EVT VT = N->getValueType(0);
1575 unsigned FirstTblReg = IsExt ? 2 : 1;
1577 // Form a REG_SEQUENCE to force register allocation.
1579 SDValue V0 = N->getOperand(FirstTblReg + 0);
1580 SDValue V1 = N->getOperand(FirstTblReg + 1);
1582 RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0);
1584 SDValue V2 = N->getOperand(FirstTblReg + 2);
1585 // If it's a vtbl3, form a quad D-register and leave the last part as
1587 SDValue V3 = (NumVecs == 3)
1588 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
1589 : N->getOperand(FirstTblReg + 3);
1590 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1593 // Now extract the D registers back out.
1594 SmallVector<SDValue, 6> Ops;
1596 Ops.push_back(N->getOperand(1));
1597 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT, RegSeq));
1598 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT, RegSeq));
1600 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT, RegSeq));
1602 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT, RegSeq));
1604 Ops.push_back(N->getOperand(FirstTblReg + NumVecs));
1605 Ops.push_back(getAL(CurDAG)); // predicate
1606 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // predicate register
1607 return CurDAG->getMachineNode(Opc, dl, VT, Ops.data(), Ops.size());
1610 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
1612 if (!Subtarget->hasV6T2Ops())
1615 unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
1616 : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
1619 // For unsigned extracts, check for a shift right and mask
1620 unsigned And_imm = 0;
1621 if (N->getOpcode() == ISD::AND) {
1622 if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
1624 // The immediate is a mask of the low bits iff imm & (imm+1) == 0
1625 if (And_imm & (And_imm + 1))
1628 unsigned Srl_imm = 0;
1629 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
1631 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1633 unsigned Width = CountTrailingOnes_32(And_imm);
1634 unsigned LSB = Srl_imm;
1635 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1636 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1637 CurDAG->getTargetConstant(LSB, MVT::i32),
1638 CurDAG->getTargetConstant(Width, MVT::i32),
1639 getAL(CurDAG), Reg0 };
1640 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1646 // Otherwise, we're looking for a shift of a shift
1647 unsigned Shl_imm = 0;
1648 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
1649 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1650 unsigned Srl_imm = 0;
1651 if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
1652 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1653 unsigned Width = 32 - Srl_imm;
1654 int LSB = Srl_imm - Shl_imm;
1657 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1658 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1659 CurDAG->getTargetConstant(LSB, MVT::i32),
1660 CurDAG->getTargetConstant(Width, MVT::i32),
1661 getAL(CurDAG), Reg0 };
1662 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1668 SDNode *ARMDAGToDAGISel::
1669 SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1670 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1673 if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
1674 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1675 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1678 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1679 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1680 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1681 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1683 llvm_unreachable("Unknown so_reg opcode!");
1687 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1688 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1689 SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
1690 return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
1695 SDNode *ARMDAGToDAGISel::
1696 SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1697 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1701 if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
1702 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1703 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
1704 return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
1709 SDNode *ARMDAGToDAGISel::
1710 SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1711 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1712 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1716 if (Pred_t2_so_imm(TrueVal.getNode())) {
1717 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1718 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1719 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1720 return CurDAG->SelectNodeTo(N,
1721 ARM::t2MOVCCi, MVT::i32, Ops, 5);
1726 SDNode *ARMDAGToDAGISel::
1727 SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1728 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1729 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1733 if (Pred_so_imm(TrueVal.getNode())) {
1734 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1735 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1736 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1737 return CurDAG->SelectNodeTo(N,
1738 ARM::MOVCCi, MVT::i32, Ops, 5);
1743 SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
1744 EVT VT = N->getValueType(0);
1745 SDValue FalseVal = N->getOperand(0);
1746 SDValue TrueVal = N->getOperand(1);
1747 SDValue CC = N->getOperand(2);
1748 SDValue CCR = N->getOperand(3);
1749 SDValue InFlag = N->getOperand(4);
1750 assert(CC.getOpcode() == ISD::Constant);
1751 assert(CCR.getOpcode() == ISD::Register);
1752 ARMCC::CondCodes CCVal =
1753 (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
1755 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1756 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1757 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1758 // Pattern complexity = 18 cost = 1 size = 0
1762 if (Subtarget->isThumb()) {
1763 SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
1764 CCVal, CCR, InFlag);
1766 Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
1767 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1771 SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
1772 CCVal, CCR, InFlag);
1774 Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
1775 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1780 // Pattern: (ARMcmov:i32 GPR:i32:$false,
1781 // (imm:i32)<<P:Pred_so_imm>>:$true,
1783 // Emits: (MOVCCi:i32 GPR:i32:$false,
1784 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1785 // Pattern complexity = 10 cost = 1 size = 0
1786 if (Subtarget->isThumb()) {
1787 SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
1788 CCVal, CCR, InFlag);
1790 Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
1791 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1795 SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
1796 CCVal, CCR, InFlag);
1798 Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
1799 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1805 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1806 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1807 // Pattern complexity = 6 cost = 1 size = 0
1809 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1810 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1811 // Pattern complexity = 6 cost = 11 size = 0
1813 // Also FCPYScc and FCPYDcc.
1814 SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
1815 SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
1817 switch (VT.getSimpleVT().SimpleTy) {
1818 default: assert(false && "Illegal conditional move type!");
1821 Opc = Subtarget->isThumb()
1822 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1832 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1835 SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
1836 // The only time a CONCAT_VECTORS operation can have legal types is when
1837 // two 64-bit vectors are concatenated to a 128-bit vector.
1838 EVT VT = N->getValueType(0);
1839 if (!VT.is128BitVector() || N->getNumOperands() != 2)
1840 llvm_unreachable("unexpected CONCAT_VECTORS");
1841 DebugLoc dl = N->getDebugLoc();
1842 SDValue V0 = N->getOperand(0);
1843 SDValue V1 = N->getOperand(1);
1844 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1845 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1846 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
1847 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
1850 SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
1851 DebugLoc dl = N->getDebugLoc();
1853 if (N->isMachineOpcode())
1854 return NULL; // Already selected.
1856 switch (N->getOpcode()) {
1858 case ISD::Constant: {
1859 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1861 if (Subtarget->hasThumb2())
1862 // Thumb2-aware targets have the MOVT instruction, so all immediates can
1863 // be done with MOV + MOVT, at worst.
1866 if (Subtarget->isThumb()) {
1867 UseCP = (Val > 255 && // MOV
1868 ~Val > 255 && // MOV + MVN
1869 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
1871 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
1872 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
1873 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
1878 CurDAG->getTargetConstantPool(ConstantInt::get(
1879 Type::getInt32Ty(*CurDAG->getContext()), Val),
1880 TLI.getPointerTy());
1883 if (Subtarget->isThumb1Only()) {
1884 SDValue Pred = getAL(CurDAG);
1885 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1886 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1887 ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1892 CurDAG->getRegister(0, MVT::i32),
1893 CurDAG->getTargetConstant(0, MVT::i32),
1895 CurDAG->getRegister(0, MVT::i32),
1896 CurDAG->getEntryNode()
1898 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1901 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
1905 // Other cases are autogenerated.
1908 case ISD::FrameIndex: {
1909 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1910 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1911 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1912 if (Subtarget->isThumb1Only()) {
1913 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1914 CurDAG->getTargetConstant(0, MVT::i32));
1916 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1917 ARM::t2ADDri : ARM::ADDri);
1918 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1919 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1920 CurDAG->getRegister(0, MVT::i32) };
1921 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1925 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1929 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true))
1933 if (Subtarget->isThumb1Only())
1935 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1936 unsigned RHSV = C->getZExtValue();
1938 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
1939 unsigned ShImm = Log2_32(RHSV-1);
1942 SDValue V = N->getOperand(0);
1943 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1944 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1945 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1946 if (Subtarget->isThumb()) {
1947 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1948 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1950 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1951 return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1954 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
1955 unsigned ShImm = Log2_32(RHSV+1);
1958 SDValue V = N->getOperand(0);
1959 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1960 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1961 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1962 if (Subtarget->isThumb()) {
1963 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1964 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 6);
1966 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1967 return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1973 // Check for unsigned bitfield extract
1974 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1977 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
1978 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
1979 // are entirely contributed by c2 and lower 16-bits are entirely contributed
1980 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
1981 // Select it to: "movt x, ((c1 & 0xffff) >> 16)
1982 EVT VT = N->getValueType(0);
1985 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
1987 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
1990 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
1991 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1994 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
1995 SDValue N2 = N0.getOperand(1);
1996 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1999 unsigned N1CVal = N1C->getZExtValue();
2000 unsigned N2CVal = N2C->getZExtValue();
2001 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
2002 (N1CVal & 0xffffU) == 0xffffU &&
2003 (N2CVal & 0xffffU) == 0x0U) {
2004 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
2006 SDValue Ops[] = { N0.getOperand(0), Imm16,
2007 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2008 return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
2013 case ARMISD::VMOVRRD:
2014 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
2015 N->getOperand(0), getAL(CurDAG),
2016 CurDAG->getRegister(0, MVT::i32));
2017 case ISD::UMUL_LOHI: {
2018 if (Subtarget->isThumb1Only())
2020 if (Subtarget->isThumb()) {
2021 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2022 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2023 CurDAG->getRegister(0, MVT::i32) };
2024 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32,Ops,4);
2026 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2027 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2028 CurDAG->getRegister(0, MVT::i32) };
2029 return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
2032 case ISD::SMUL_LOHI: {
2033 if (Subtarget->isThumb1Only())
2035 if (Subtarget->isThumb()) {
2036 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2037 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2038 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32,Ops,4);
2040 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2041 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2042 CurDAG->getRegister(0, MVT::i32) };
2043 return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
2047 SDNode *ResNode = 0;
2048 if (Subtarget->isThumb() && Subtarget->hasThumb2())
2049 ResNode = SelectT2IndexedLoad(N);
2051 ResNode = SelectARMIndexedLoad(N);
2055 // VLDMQ must be custom-selected for "v2f64 load" to set the AM5Opc value.
2056 if (Subtarget->hasVFP2() &&
2057 N->getValueType(0).getSimpleVT().SimpleTy == MVT::v2f64) {
2058 SDValue Chain = N->getOperand(0);
2060 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
2061 SDValue Pred = getAL(CurDAG);
2062 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2063 SDValue Ops[] = { N->getOperand(1), AM5Opc, Pred, PredReg, Chain };
2064 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2065 MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
2066 SDNode *Ret = CurDAG->getMachineNode(ARM::VLDMQ, dl,
2067 MVT::v2f64, MVT::Other, Ops, 5);
2068 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
2071 // Other cases are autogenerated.
2075 // VSTMQ must be custom-selected for "v2f64 store" to set the AM5Opc value.
2076 if (Subtarget->hasVFP2() &&
2077 N->getOperand(1).getValueType().getSimpleVT().SimpleTy == MVT::v2f64) {
2078 SDValue Chain = N->getOperand(0);
2080 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
2081 SDValue Pred = getAL(CurDAG);
2082 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2083 SDValue Ops[] = { N->getOperand(1), N->getOperand(2),
2084 AM5Opc, Pred, PredReg, Chain };
2085 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2086 MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
2087 SDNode *Ret = CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
2088 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
2091 // Other cases are autogenerated.
2094 case ARMISD::BRCOND: {
2095 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2096 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2097 // Pattern complexity = 6 cost = 1 size = 0
2099 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2100 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
2101 // Pattern complexity = 6 cost = 1 size = 0
2103 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2104 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2105 // Pattern complexity = 6 cost = 1 size = 0
2107 unsigned Opc = Subtarget->isThumb() ?
2108 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
2109 SDValue Chain = N->getOperand(0);
2110 SDValue N1 = N->getOperand(1);
2111 SDValue N2 = N->getOperand(2);
2112 SDValue N3 = N->getOperand(3);
2113 SDValue InFlag = N->getOperand(4);
2114 assert(N1.getOpcode() == ISD::BasicBlock);
2115 assert(N2.getOpcode() == ISD::Constant);
2116 assert(N3.getOpcode() == ISD::Register);
2118 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2119 cast<ConstantSDNode>(N2)->getZExtValue()),
2121 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
2122 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
2124 Chain = SDValue(ResNode, 0);
2125 if (N->getNumValues() == 2) {
2126 InFlag = SDValue(ResNode, 1);
2127 ReplaceUses(SDValue(N, 1), InFlag);
2129 ReplaceUses(SDValue(N, 0),
2130 SDValue(Chain.getNode(), Chain.getResNo()));
2134 return SelectCMOVOp(N);
2135 case ARMISD::CNEG: {
2136 EVT VT = N->getValueType(0);
2137 SDValue N0 = N->getOperand(0);
2138 SDValue N1 = N->getOperand(1);
2139 SDValue N2 = N->getOperand(2);
2140 SDValue N3 = N->getOperand(3);
2141 SDValue InFlag = N->getOperand(4);
2142 assert(N2.getOpcode() == ISD::Constant);
2143 assert(N3.getOpcode() == ISD::Register);
2145 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2146 cast<ConstantSDNode>(N2)->getZExtValue()),
2148 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
2150 switch (VT.getSimpleVT().SimpleTy) {
2151 default: assert(false && "Illegal conditional move type!");
2160 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
2163 case ARMISD::VZIP: {
2165 EVT VT = N->getValueType(0);
2166 switch (VT.getSimpleVT().SimpleTy) {
2167 default: return NULL;
2168 case MVT::v8i8: Opc = ARM::VZIPd8; break;
2169 case MVT::v4i16: Opc = ARM::VZIPd16; break;
2171 case MVT::v2i32: Opc = ARM::VZIPd32; break;
2172 case MVT::v16i8: Opc = ARM::VZIPq8; break;
2173 case MVT::v8i16: Opc = ARM::VZIPq16; break;
2175 case MVT::v4i32: Opc = ARM::VZIPq32; 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::VUZP: {
2184 EVT VT = N->getValueType(0);
2185 switch (VT.getSimpleVT().SimpleTy) {
2186 default: return NULL;
2187 case MVT::v8i8: Opc = ARM::VUZPd8; break;
2188 case MVT::v4i16: Opc = ARM::VUZPd16; break;
2190 case MVT::v2i32: Opc = ARM::VUZPd32; break;
2191 case MVT::v16i8: Opc = ARM::VUZPq8; break;
2192 case MVT::v8i16: Opc = ARM::VUZPq16; break;
2194 case MVT::v4i32: Opc = ARM::VUZPq32; break;
2196 SDValue Pred = getAL(CurDAG);
2197 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2198 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2199 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2201 case ARMISD::VTRN: {
2203 EVT VT = N->getValueType(0);
2204 switch (VT.getSimpleVT().SimpleTy) {
2205 default: return NULL;
2206 case MVT::v8i8: Opc = ARM::VTRNd8; break;
2207 case MVT::v4i16: Opc = ARM::VTRNd16; break;
2209 case MVT::v2i32: Opc = ARM::VTRNd32; break;
2210 case MVT::v16i8: Opc = ARM::VTRNq8; break;
2211 case MVT::v8i16: Opc = ARM::VTRNq16; break;
2213 case MVT::v4i32: Opc = ARM::VTRNq32; break;
2215 SDValue Pred = getAL(CurDAG);
2216 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2217 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2218 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2220 case ARMISD::BUILD_VECTOR: {
2221 EVT VecVT = N->getValueType(0);
2222 EVT EltVT = VecVT.getVectorElementType();
2223 unsigned NumElts = VecVT.getVectorNumElements();
2224 if (EltVT.getSimpleVT() == MVT::f64) {
2225 assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
2226 return PairDRegs(VecVT, N->getOperand(0), N->getOperand(1));
2228 assert(EltVT.getSimpleVT() == MVT::f32 &&
2229 "unexpected type for BUILD_VECTOR");
2231 return PairSRegs(VecVT, N->getOperand(0), N->getOperand(1));
2232 assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
2233 return QuadSRegs(VecVT, N->getOperand(0), N->getOperand(1),
2234 N->getOperand(2), N->getOperand(3));
2237 case ISD::INTRINSIC_VOID:
2238 case ISD::INTRINSIC_W_CHAIN: {
2239 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2244 case Intrinsic::arm_neon_vld1: {
2245 unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
2246 ARM::VLD1d32, ARM::VLD1d64 };
2247 unsigned QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
2248 ARM::VLD1q32, ARM::VLD1q64 };
2249 return SelectVLD(N, 1, DOpcodes, QOpcodes, 0);
2252 case Intrinsic::arm_neon_vld2: {
2253 unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
2254 ARM::VLD2d32, ARM::VLD1q64 };
2255 unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
2256 return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
2259 case Intrinsic::arm_neon_vld3: {
2260 unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
2261 ARM::VLD3d32, ARM::VLD1d64T };
2262 unsigned QOpcodes0[] = { ARM::VLD3q8_UPD,
2265 unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD,
2266 ARM::VLD3q16odd_UPD,
2267 ARM::VLD3q32odd_UPD };
2268 return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2271 case Intrinsic::arm_neon_vld4: {
2272 unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
2273 ARM::VLD4d32, ARM::VLD1d64Q };
2274 unsigned QOpcodes0[] = { ARM::VLD4q8_UPD,
2277 unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD,
2278 ARM::VLD4q16odd_UPD,
2279 ARM::VLD4q32odd_UPD };
2280 return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2283 case Intrinsic::arm_neon_vld2lane: {
2284 unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
2285 unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 };
2286 unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd };
2287 return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
2290 case Intrinsic::arm_neon_vld3lane: {
2291 unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
2292 unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 };
2293 unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd };
2294 return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
2297 case Intrinsic::arm_neon_vld4lane: {
2298 unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
2299 unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 };
2300 unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd };
2301 return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
2304 case Intrinsic::arm_neon_vst1: {
2305 unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
2306 ARM::VST1d32, ARM::VST1d64 };
2307 unsigned QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
2308 ARM::VST1q32, ARM::VST1q64 };
2309 return SelectVST(N, 1, DOpcodes, QOpcodes, 0);
2312 case Intrinsic::arm_neon_vst2: {
2313 unsigned DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
2314 ARM::VST2d32, ARM::VST1q64 };
2315 unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
2316 return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
2319 case Intrinsic::arm_neon_vst3: {
2320 unsigned DOpcodes[] = { ARM::VST3d8Pseudo, ARM::VST3d16Pseudo,
2321 ARM::VST3d32Pseudo, ARM::VST1d64TPseudo };
2322 unsigned QOpcodes0[] = { ARM::VST3q8Pseudo_UPD,
2323 ARM::VST3q16Pseudo_UPD,
2324 ARM::VST3q32Pseudo_UPD };
2325 unsigned QOpcodes1[] = { ARM::VST3q8oddPseudo_UPD,
2326 ARM::VST3q16oddPseudo_UPD,
2327 ARM::VST3q32oddPseudo_UPD };
2328 return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2331 case Intrinsic::arm_neon_vst4: {
2332 unsigned DOpcodes[] = { ARM::VST4d8Pseudo, ARM::VST4d16Pseudo,
2333 ARM::VST4d32Pseudo, ARM::VST1d64QPseudo };
2334 unsigned QOpcodes0[] = { ARM::VST4q8Pseudo_UPD,
2335 ARM::VST4q16Pseudo_UPD,
2336 ARM::VST4q32Pseudo_UPD };
2337 unsigned QOpcodes1[] = { ARM::VST4q8oddPseudo_UPD,
2338 ARM::VST4q16oddPseudo_UPD,
2339 ARM::VST4q32oddPseudo_UPD };
2340 return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2343 case Intrinsic::arm_neon_vst2lane: {
2344 unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
2345 unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 };
2346 unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd };
2347 return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
2350 case Intrinsic::arm_neon_vst3lane: {
2351 unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
2352 unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 };
2353 unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd };
2354 return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
2357 case Intrinsic::arm_neon_vst4lane: {
2358 unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
2359 unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 };
2360 unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd };
2361 return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
2367 case ISD::INTRINSIC_WO_CHAIN: {
2368 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2373 case Intrinsic::arm_neon_vtbl2:
2374 return SelectVTBL(N, false, 2, ARM::VTBL2);
2375 case Intrinsic::arm_neon_vtbl3:
2376 return SelectVTBL(N, false, 3, ARM::VTBL3);
2377 case Intrinsic::arm_neon_vtbl4:
2378 return SelectVTBL(N, false, 4, ARM::VTBL4);
2380 case Intrinsic::arm_neon_vtbx2:
2381 return SelectVTBL(N, true, 2, ARM::VTBX2);
2382 case Intrinsic::arm_neon_vtbx3:
2383 return SelectVTBL(N, true, 3, ARM::VTBX3);
2384 case Intrinsic::arm_neon_vtbx4:
2385 return SelectVTBL(N, true, 4, ARM::VTBX4);
2390 case ISD::CONCAT_VECTORS:
2391 return SelectConcatVector(N);
2394 return SelectCode(N);
2397 bool ARMDAGToDAGISel::
2398 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2399 std::vector<SDValue> &OutOps) {
2400 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
2401 // Require the address to be in a register. That is safe for all ARM
2402 // variants and it is hard to do anything much smarter without knowing
2403 // how the operand is used.
2404 OutOps.push_back(Op);
2408 /// createARMISelDag - This pass converts a legalized DAG into a
2409 /// ARM-specific DAG, ready for instruction scheduling.
2411 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
2412 CodeGenOpt::Level OptLevel) {
2413 return new ARMDAGToDAGISel(TM, OptLevel);