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 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
147 SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
149 /// SelectCMOVOp - Select CMOV instructions for ARM.
150 SDNode *SelectCMOVOp(SDNode *N);
151 SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
152 ARMCC::CondCodes CCVal, SDValue CCR,
154 SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
155 ARMCC::CondCodes CCVal, SDValue CCR,
157 SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
158 ARMCC::CondCodes CCVal, SDValue CCR,
160 SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
161 ARMCC::CondCodes CCVal, SDValue CCR,
164 SDNode *SelectConcatVector(SDNode *N);
166 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
167 /// inline asm expressions.
168 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
170 std::vector<SDValue> &OutOps);
172 // Form pairs of consecutive S, D, or Q registers.
173 SDNode *PairSRegs(EVT VT, SDValue V0, SDValue V1);
174 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
175 SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1);
177 // Form sequences of 4 consecutive S, D, or Q registers.
178 SDNode *QuadSRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
179 SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
180 SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
182 // Form sequences of 8 consecutive D registers.
183 SDNode *OctoDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3,
184 SDValue V4, SDValue V5, SDValue V6, SDValue V7);
188 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
189 /// operand. If so Imm will receive the 32-bit value.
190 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
191 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
192 Imm = cast<ConstantSDNode>(N)->getZExtValue();
198 // isInt32Immediate - This method tests to see if a constant operand.
199 // If so Imm will receive the 32 bit value.
200 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
201 return isInt32Immediate(N.getNode(), Imm);
204 // isOpcWithIntImmediate - This method tests to see if the node is a specific
205 // opcode and that it has a immediate integer right operand.
206 // If so Imm will receive the 32 bit value.
207 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
208 return N->getOpcode() == Opc &&
209 isInt32Immediate(N->getOperand(1).getNode(), Imm);
213 bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
218 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
220 // Don't match base register only case. That is matched to a separate
221 // lower complexity pattern with explicit register operand.
222 if (ShOpcVal == ARM_AM::no_shift) return false;
224 BaseReg = N.getOperand(0);
225 unsigned ShImmVal = 0;
226 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
227 ShReg = CurDAG->getRegister(0, MVT::i32);
228 ShImmVal = RHS->getZExtValue() & 31;
230 ShReg = N.getOperand(1);
232 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
237 bool ARMDAGToDAGISel::SelectAddrMode2(SDNode *Op, SDValue N,
238 SDValue &Base, SDValue &Offset,
240 if (N.getOpcode() == ISD::MUL) {
241 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
242 // X * [3,5,9] -> X + X * [2,4,8] etc.
243 int RHSC = (int)RHS->getZExtValue();
246 ARM_AM::AddrOpc AddSub = ARM_AM::add;
248 AddSub = ARM_AM::sub;
251 if (isPowerOf2_32(RHSC)) {
252 unsigned ShAmt = Log2_32(RHSC);
253 Base = Offset = N.getOperand(0);
254 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
263 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
265 if (N.getOpcode() == ISD::FrameIndex) {
266 int FI = cast<FrameIndexSDNode>(N)->getIndex();
267 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
268 } else if (N.getOpcode() == ARMISD::Wrapper &&
269 !(Subtarget->useMovt() &&
270 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
271 Base = N.getOperand(0);
273 Offset = CurDAG->getRegister(0, MVT::i32);
274 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
280 // Match simple R +/- imm12 operands.
281 if (N.getOpcode() == ISD::ADD)
282 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
283 int RHSC = (int)RHS->getZExtValue();
284 if ((RHSC >= 0 && RHSC < 0x1000) ||
285 (RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
286 Base = N.getOperand(0);
287 if (Base.getOpcode() == ISD::FrameIndex) {
288 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
289 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
291 Offset = CurDAG->getRegister(0, MVT::i32);
293 ARM_AM::AddrOpc AddSub = ARM_AM::add;
295 AddSub = ARM_AM::sub;
298 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
305 // Otherwise this is R +/- [possibly shifted] R.
306 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
307 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
310 Base = N.getOperand(0);
311 Offset = N.getOperand(1);
313 if (ShOpcVal != ARM_AM::no_shift) {
314 // Check to see if the RHS of the shift is a constant, if not, we can't fold
316 if (ConstantSDNode *Sh =
317 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
318 ShAmt = Sh->getZExtValue();
319 Offset = N.getOperand(1).getOperand(0);
321 ShOpcVal = ARM_AM::no_shift;
325 // Try matching (R shl C) + (R).
326 if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
327 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
328 if (ShOpcVal != ARM_AM::no_shift) {
329 // Check to see if the RHS of the shift is a constant, if not, we can't
331 if (ConstantSDNode *Sh =
332 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
333 ShAmt = Sh->getZExtValue();
334 Offset = N.getOperand(0).getOperand(0);
335 Base = N.getOperand(1);
337 ShOpcVal = ARM_AM::no_shift;
342 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
347 bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
348 SDValue &Offset, SDValue &Opc) {
349 unsigned Opcode = Op->getOpcode();
350 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
351 ? cast<LoadSDNode>(Op)->getAddressingMode()
352 : cast<StoreSDNode>(Op)->getAddressingMode();
353 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
354 ? ARM_AM::add : ARM_AM::sub;
355 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
356 int Val = (int)C->getZExtValue();
357 if (Val >= 0 && Val < 0x1000) { // 12 bits.
358 Offset = CurDAG->getRegister(0, MVT::i32);
359 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
367 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
369 if (ShOpcVal != ARM_AM::no_shift) {
370 // Check to see if the RHS of the shift is a constant, if not, we can't fold
372 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
373 ShAmt = Sh->getZExtValue();
374 Offset = N.getOperand(0);
376 ShOpcVal = ARM_AM::no_shift;
380 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
386 bool ARMDAGToDAGISel::SelectAddrMode3(SDNode *Op, SDValue N,
387 SDValue &Base, SDValue &Offset,
389 if (N.getOpcode() == ISD::SUB) {
390 // X - C is canonicalize to X + -C, no need to handle it here.
391 Base = N.getOperand(0);
392 Offset = N.getOperand(1);
393 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
397 if (N.getOpcode() != ISD::ADD) {
399 if (N.getOpcode() == ISD::FrameIndex) {
400 int FI = cast<FrameIndexSDNode>(N)->getIndex();
401 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
403 Offset = CurDAG->getRegister(0, MVT::i32);
404 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
408 // If the RHS is +/- imm8, fold into addr mode.
409 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
410 int RHSC = (int)RHS->getZExtValue();
411 if ((RHSC >= 0 && RHSC < 256) ||
412 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
413 Base = N.getOperand(0);
414 if (Base.getOpcode() == ISD::FrameIndex) {
415 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
416 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
418 Offset = CurDAG->getRegister(0, MVT::i32);
420 ARM_AM::AddrOpc AddSub = ARM_AM::add;
422 AddSub = ARM_AM::sub;
425 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
430 Base = N.getOperand(0);
431 Offset = N.getOperand(1);
432 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
436 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
437 SDValue &Offset, SDValue &Opc) {
438 unsigned Opcode = Op->getOpcode();
439 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
440 ? cast<LoadSDNode>(Op)->getAddressingMode()
441 : cast<StoreSDNode>(Op)->getAddressingMode();
442 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
443 ? ARM_AM::add : ARM_AM::sub;
444 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
445 int Val = (int)C->getZExtValue();
446 if (Val >= 0 && Val < 256) {
447 Offset = CurDAG->getRegister(0, MVT::i32);
448 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
454 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
458 bool ARMDAGToDAGISel::SelectAddrMode4(SDNode *Op, SDValue N,
459 SDValue &Addr, SDValue &Mode) {
461 Mode = CurDAG->getTargetConstant(0, MVT::i32);
465 bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N,
466 SDValue &Base, SDValue &Offset) {
467 if (N.getOpcode() != ISD::ADD) {
469 if (N.getOpcode() == ISD::FrameIndex) {
470 int FI = cast<FrameIndexSDNode>(N)->getIndex();
471 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
472 } else if (N.getOpcode() == ARMISD::Wrapper &&
473 !(Subtarget->useMovt() &&
474 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
475 Base = N.getOperand(0);
477 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
482 // If the RHS is +/- imm8, fold into addr mode.
483 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
484 int RHSC = (int)RHS->getZExtValue();
485 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied by 4.
487 if ((RHSC >= 0 && RHSC < 256) ||
488 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
489 Base = N.getOperand(0);
490 if (Base.getOpcode() == ISD::FrameIndex) {
491 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
492 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
495 ARM_AM::AddrOpc AddSub = ARM_AM::add;
497 AddSub = ARM_AM::sub;
500 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
508 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
513 bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N,
514 SDValue &Addr, SDValue &Align) {
516 // Default to no alignment.
517 Align = CurDAG->getTargetConstant(0, MVT::i32);
521 bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
522 SDValue &Offset, SDValue &Label) {
523 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
524 Offset = N.getOperand(0);
525 SDValue N1 = N.getOperand(1);
526 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
533 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
534 SDValue &Base, SDValue &Offset){
535 // FIXME dl should come from the parent load or store, not the address
536 DebugLoc dl = Op->getDebugLoc();
537 if (N.getOpcode() != ISD::ADD) {
538 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
539 if (!NC || !NC->isNullValue())
546 Base = N.getOperand(0);
547 Offset = N.getOperand(1);
552 ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
553 unsigned Scale, SDValue &Base,
554 SDValue &OffImm, SDValue &Offset) {
556 SDValue TmpBase, TmpOffImm;
557 if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
558 return false; // We want to select tLDRspi / tSTRspi instead.
559 if (N.getOpcode() == ARMISD::Wrapper &&
560 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
561 return false; // We want to select tLDRpci instead.
564 if (N.getOpcode() != ISD::ADD) {
565 if (N.getOpcode() == ARMISD::Wrapper &&
566 !(Subtarget->useMovt() &&
567 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
568 Base = N.getOperand(0);
572 Offset = CurDAG->getRegister(0, MVT::i32);
573 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
577 // Thumb does not have [sp, r] address mode.
578 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
579 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
580 if ((LHSR && LHSR->getReg() == ARM::SP) ||
581 (RHSR && RHSR->getReg() == ARM::SP)) {
583 Offset = CurDAG->getRegister(0, MVT::i32);
584 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
588 // If the RHS is + imm5 * scale, fold into addr mode.
589 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
590 int RHSC = (int)RHS->getZExtValue();
591 if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
593 if (RHSC >= 0 && RHSC < 32) {
594 Base = N.getOperand(0);
595 Offset = CurDAG->getRegister(0, MVT::i32);
596 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
602 Base = N.getOperand(0);
603 Offset = N.getOperand(1);
604 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
608 bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
609 SDValue &Base, SDValue &OffImm,
611 return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
614 bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
615 SDValue &Base, SDValue &OffImm,
617 return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
620 bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
621 SDValue &Base, SDValue &OffImm,
623 return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
626 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
627 SDValue &Base, SDValue &OffImm) {
628 if (N.getOpcode() == ISD::FrameIndex) {
629 int FI = cast<FrameIndexSDNode>(N)->getIndex();
630 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
631 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
635 if (N.getOpcode() != ISD::ADD)
638 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
639 if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
640 (LHSR && LHSR->getReg() == ARM::SP)) {
641 // If the RHS is + imm8 * scale, fold into addr mode.
642 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
643 int RHSC = (int)RHS->getZExtValue();
644 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
646 if (RHSC >= 0 && RHSC < 256) {
647 Base = N.getOperand(0);
648 if (Base.getOpcode() == ISD::FrameIndex) {
649 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
650 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
652 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
662 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
665 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
667 // Don't match base register only case. That is matched to a separate
668 // lower complexity pattern with explicit register operand.
669 if (ShOpcVal == ARM_AM::no_shift) return false;
671 BaseReg = N.getOperand(0);
672 unsigned ShImmVal = 0;
673 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
674 ShImmVal = RHS->getZExtValue() & 31;
675 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
682 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
683 SDValue &Base, SDValue &OffImm) {
684 // Match simple R + imm12 operands.
687 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
688 if (N.getOpcode() == ISD::FrameIndex) {
689 // Match frame index...
690 int FI = cast<FrameIndexSDNode>(N)->getIndex();
691 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
692 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
694 } else if (N.getOpcode() == ARMISD::Wrapper &&
695 !(Subtarget->useMovt() &&
696 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
697 Base = N.getOperand(0);
698 if (Base.getOpcode() == ISD::TargetConstantPool)
699 return false; // We want to select t2LDRpci instead.
702 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
706 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
707 if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
708 // Let t2LDRi8 handle (R - imm8).
711 int RHSC = (int)RHS->getZExtValue();
712 if (N.getOpcode() == ISD::SUB)
715 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
716 Base = N.getOperand(0);
717 if (Base.getOpcode() == ISD::FrameIndex) {
718 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
719 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
721 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
728 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
732 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
733 SDValue &Base, SDValue &OffImm) {
734 // Match simple R - imm8 operands.
735 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
736 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
737 int RHSC = (int)RHS->getSExtValue();
738 if (N.getOpcode() == ISD::SUB)
741 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
742 Base = N.getOperand(0);
743 if (Base.getOpcode() == ISD::FrameIndex) {
744 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
745 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
747 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
756 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
758 unsigned Opcode = Op->getOpcode();
759 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
760 ? cast<LoadSDNode>(Op)->getAddressingMode()
761 : cast<StoreSDNode>(Op)->getAddressingMode();
762 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
763 int RHSC = (int)RHS->getZExtValue();
764 if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
765 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
766 ? CurDAG->getTargetConstant(RHSC, MVT::i32)
767 : CurDAG->getTargetConstant(-RHSC, MVT::i32);
775 bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
776 SDValue &Base, SDValue &OffImm) {
777 if (N.getOpcode() == ISD::ADD) {
778 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
779 int RHSC = (int)RHS->getZExtValue();
781 if (((RHSC & 0x3) == 0) &&
782 ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) {
783 Base = N.getOperand(0);
784 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
788 } else if (N.getOpcode() == ISD::SUB) {
789 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
790 int RHSC = (int)RHS->getZExtValue();
792 if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) {
793 Base = N.getOperand(0);
794 OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
803 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDNode *Op, SDValue N,
805 SDValue &OffReg, SDValue &ShImm) {
806 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
807 if (N.getOpcode() != ISD::ADD)
810 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
811 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
812 int RHSC = (int)RHS->getZExtValue();
813 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
815 else if (RHSC < 0 && RHSC >= -255) // 8 bits
819 // Look for (R + R) or (R + (R << [1,2,3])).
821 Base = N.getOperand(0);
822 OffReg = N.getOperand(1);
824 // Swap if it is ((R << c) + R).
825 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
826 if (ShOpcVal != ARM_AM::lsl) {
827 ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
828 if (ShOpcVal == ARM_AM::lsl)
829 std::swap(Base, OffReg);
832 if (ShOpcVal == ARM_AM::lsl) {
833 // Check to see if the RHS of the shift is a constant, if not, we can't fold
835 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
836 ShAmt = Sh->getZExtValue();
839 ShOpcVal = ARM_AM::no_shift;
841 OffReg = OffReg.getOperand(0);
843 ShOpcVal = ARM_AM::no_shift;
847 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
852 //===--------------------------------------------------------------------===//
854 /// getAL - Returns a ARMCC::AL immediate node.
855 static inline SDValue getAL(SelectionDAG *CurDAG) {
856 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
859 SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
860 LoadSDNode *LD = cast<LoadSDNode>(N);
861 ISD::MemIndexedMode AM = LD->getAddressingMode();
862 if (AM == ISD::UNINDEXED)
865 EVT LoadedVT = LD->getMemoryVT();
866 SDValue Offset, AMOpc;
867 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
870 if (LoadedVT == MVT::i32 &&
871 SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
872 Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
874 } else if (LoadedVT == MVT::i16 &&
875 SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
877 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
878 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
879 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
880 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
881 if (LD->getExtensionType() == ISD::SEXTLOAD) {
882 if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
884 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
887 if (SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
889 Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
895 SDValue Chain = LD->getChain();
896 SDValue Base = LD->getBasePtr();
897 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
898 CurDAG->getRegister(0, MVT::i32), Chain };
899 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
906 SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
907 LoadSDNode *LD = cast<LoadSDNode>(N);
908 ISD::MemIndexedMode AM = LD->getAddressingMode();
909 if (AM == ISD::UNINDEXED)
912 EVT LoadedVT = LD->getMemoryVT();
913 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
915 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
918 if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
919 switch (LoadedVT.getSimpleVT().SimpleTy) {
921 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
925 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
927 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
932 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
934 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
943 SDValue Chain = LD->getChain();
944 SDValue Base = LD->getBasePtr();
945 SDValue Ops[]= { Base, Offset, getAL(CurDAG),
946 CurDAG->getRegister(0, MVT::i32), Chain };
947 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
954 /// PairSRegs - Form a D register from a pair of S registers.
956 SDNode *ARMDAGToDAGISel::PairSRegs(EVT VT, SDValue V0, SDValue V1) {
957 DebugLoc dl = V0.getNode()->getDebugLoc();
958 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
959 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
960 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
961 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
964 /// PairDRegs - Form a quad register from a pair of D registers.
966 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
967 DebugLoc dl = V0.getNode()->getDebugLoc();
968 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
969 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
970 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
971 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
974 /// PairQRegs - Form 4 consecutive D registers from a pair of Q registers.
976 SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) {
977 DebugLoc dl = V0.getNode()->getDebugLoc();
978 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
979 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
980 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
981 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
984 /// QuadSRegs - Form 4 consecutive S registers.
986 SDNode *ARMDAGToDAGISel::QuadSRegs(EVT VT, SDValue V0, SDValue V1,
987 SDValue V2, SDValue V3) {
988 DebugLoc dl = V0.getNode()->getDebugLoc();
989 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
990 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
991 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, MVT::i32);
992 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, MVT::i32);
993 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
994 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
997 /// QuadDRegs - Form 4 consecutive D registers.
999 SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
1000 SDValue V2, SDValue V3) {
1001 DebugLoc dl = V0.getNode()->getDebugLoc();
1002 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1003 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1004 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
1005 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
1006 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1007 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1010 /// QuadQRegs - Form 4 consecutive Q registers.
1012 SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1,
1013 SDValue V2, SDValue V3) {
1014 DebugLoc dl = V0.getNode()->getDebugLoc();
1015 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
1016 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
1017 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, MVT::i32);
1018 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, MVT::i32);
1019 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1020 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1023 /// OctoDRegs - Form 8 consecutive D registers.
1025 SDNode *ARMDAGToDAGISel::OctoDRegs(EVT VT, SDValue V0, SDValue V1,
1026 SDValue V2, SDValue V3,
1027 SDValue V4, SDValue V5,
1028 SDValue V6, SDValue V7) {
1029 DebugLoc dl = V0.getNode()->getDebugLoc();
1030 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1031 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1032 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
1033 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
1034 SDValue SubReg4 = CurDAG->getTargetConstant(ARM::dsub_4, MVT::i32);
1035 SDValue SubReg5 = CurDAG->getTargetConstant(ARM::dsub_5, MVT::i32);
1036 SDValue SubReg6 = CurDAG->getTargetConstant(ARM::dsub_6, MVT::i32);
1037 SDValue SubReg7 = CurDAG->getTargetConstant(ARM::dsub_7, MVT::i32);
1038 const SDValue Ops[] ={ V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3,
1039 V4, SubReg4, V5, SubReg5, V6, SubReg6, V7, SubReg7 };
1040 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 16);
1043 /// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
1044 /// for a 64-bit subregister of the vector.
1045 static EVT GetNEONSubregVT(EVT VT) {
1046 switch (VT.getSimpleVT().SimpleTy) {
1047 default: llvm_unreachable("unhandled NEON type");
1048 case MVT::v16i8: return MVT::v8i8;
1049 case MVT::v8i16: return MVT::v4i16;
1050 case MVT::v4f32: return MVT::v2f32;
1051 case MVT::v4i32: return MVT::v2i32;
1052 case MVT::v2i64: return MVT::v1i64;
1056 SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
1057 unsigned *DOpcodes, unsigned *QOpcodes0,
1058 unsigned *QOpcodes1) {
1059 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1060 DebugLoc dl = N->getDebugLoc();
1062 SDValue MemAddr, Align;
1063 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1066 SDValue Chain = N->getOperand(0);
1067 EVT VT = N->getValueType(0);
1068 bool is64BitVector = VT.is64BitVector();
1070 unsigned OpcodeIndex;
1071 switch (VT.getSimpleVT().SimpleTy) {
1072 default: llvm_unreachable("unhandled vld type");
1073 // Double-register operations:
1074 case MVT::v8i8: OpcodeIndex = 0; break;
1075 case MVT::v4i16: OpcodeIndex = 1; break;
1077 case MVT::v2i32: OpcodeIndex = 2; break;
1078 case MVT::v1i64: OpcodeIndex = 3; break;
1079 // Quad-register operations:
1080 case MVT::v16i8: OpcodeIndex = 0; break;
1081 case MVT::v8i16: OpcodeIndex = 1; break;
1083 case MVT::v4i32: OpcodeIndex = 2; break;
1084 case MVT::v2i64: OpcodeIndex = 3;
1085 assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
1089 SDValue Pred = getAL(CurDAG);
1090 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1091 if (is64BitVector) {
1092 unsigned Opc = DOpcodes[OpcodeIndex];
1093 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1094 std::vector<EVT> ResTys(NumVecs, VT);
1095 ResTys.push_back(MVT::Other);
1096 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1101 SDValue V0 = SDValue(VLd, 0);
1102 SDValue V1 = SDValue(VLd, 1);
1104 // Form a REG_SEQUENCE to force register allocation.
1106 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1108 SDValue V2 = SDValue(VLd, 2);
1109 // If it's a vld3, form a quad D-register but discard the last part.
1110 SDValue V3 = (NumVecs == 3)
1111 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1113 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1116 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1117 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1118 SDValue D = CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec,
1120 ReplaceUses(SDValue(N, Vec), D);
1122 ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, NumVecs));
1126 EVT RegVT = GetNEONSubregVT(VT);
1128 // Quad registers are directly supported for VLD1 and VLD2,
1129 // loading pairs of D regs.
1130 unsigned Opc = QOpcodes0[OpcodeIndex];
1131 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1132 std::vector<EVT> ResTys(2 * NumVecs, RegVT);
1133 ResTys.push_back(MVT::Other);
1134 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1135 Chain = SDValue(VLd, 2 * NumVecs);
1137 // Combine the even and odd subregs to produce the result.
1139 SDNode *Q = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
1140 ReplaceUses(SDValue(N, 0), SDValue(Q, 0));
1142 SDValue QQ = SDValue(QuadDRegs(MVT::v4i64,
1143 SDValue(VLd, 0), SDValue(VLd, 1),
1144 SDValue(VLd, 2), SDValue(VLd, 3)), 0);
1145 SDValue Q0 = CurDAG->getTargetExtractSubreg(ARM::qsub_0, dl, VT, QQ);
1146 SDValue Q1 = CurDAG->getTargetExtractSubreg(ARM::qsub_1, dl, VT, QQ);
1147 ReplaceUses(SDValue(N, 0), Q0);
1148 ReplaceUses(SDValue(N, 1), Q1);
1151 // Otherwise, quad registers are loaded with two separate instructions,
1152 // where one loads the even registers and the other loads the odd registers.
1154 std::vector<EVT> ResTys(NumVecs, RegVT);
1155 ResTys.push_back(MemAddr.getValueType());
1156 ResTys.push_back(MVT::Other);
1158 // Load the even subregs.
1159 unsigned Opc = QOpcodes0[OpcodeIndex];
1160 const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain };
1161 SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6);
1162 Chain = SDValue(VLdA, NumVecs+1);
1164 // Load the odd subregs.
1165 Opc = QOpcodes1[OpcodeIndex];
1166 const SDValue OpsB[] = { SDValue(VLdA, NumVecs),
1167 Align, Reg0, Pred, Reg0, Chain };
1168 SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
1169 Chain = SDValue(VLdB, NumVecs+1);
1171 SDValue V0 = SDValue(VLdA, 0);
1172 SDValue V1 = SDValue(VLdB, 0);
1173 SDValue V2 = SDValue(VLdA, 1);
1174 SDValue V3 = SDValue(VLdB, 1);
1175 SDValue V4 = SDValue(VLdA, 2);
1176 SDValue V5 = SDValue(VLdB, 2);
1177 SDValue V6 = (NumVecs == 3)
1178 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0)
1180 SDValue V7 = (NumVecs == 3)
1181 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0)
1183 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V0, V1, V2, V3,
1184 V4, V5, V6, V7), 0);
1186 // Extract out the 3 / 4 Q registers.
1187 assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1188 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1189 SDValue Q = CurDAG->getTargetExtractSubreg(ARM::qsub_0+Vec,
1191 ReplaceUses(SDValue(N, Vec), Q);
1194 ReplaceUses(SDValue(N, NumVecs), Chain);
1198 SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
1199 unsigned *DOpcodes, unsigned *QOpcodes0,
1200 unsigned *QOpcodes1) {
1201 assert(NumVecs >=1 && NumVecs <= 4 && "VST NumVecs out-of-range");
1202 DebugLoc dl = N->getDebugLoc();
1204 SDValue MemAddr, Align;
1205 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1208 SDValue Chain = N->getOperand(0);
1209 EVT VT = N->getOperand(3).getValueType();
1210 bool is64BitVector = VT.is64BitVector();
1212 unsigned OpcodeIndex;
1213 switch (VT.getSimpleVT().SimpleTy) {
1214 default: llvm_unreachable("unhandled vst type");
1215 // Double-register operations:
1216 case MVT::v8i8: OpcodeIndex = 0; break;
1217 case MVT::v4i16: OpcodeIndex = 1; break;
1219 case MVT::v2i32: OpcodeIndex = 2; break;
1220 case MVT::v1i64: OpcodeIndex = 3; break;
1221 // Quad-register operations:
1222 case MVT::v16i8: OpcodeIndex = 0; break;
1223 case MVT::v8i16: OpcodeIndex = 1; break;
1225 case MVT::v4i32: OpcodeIndex = 2; break;
1226 case MVT::v2i64: OpcodeIndex = 3;
1227 assert(NumVecs == 1 && "v2i64 type only supported for VST1");
1231 SDValue Pred = getAL(CurDAG);
1232 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1234 SmallVector<SDValue, 10> Ops;
1235 Ops.push_back(MemAddr);
1236 Ops.push_back(Align);
1238 if (is64BitVector) {
1241 SDValue V0 = N->getOperand(0+3);
1242 SDValue V1 = N->getOperand(1+3);
1244 // Form a REG_SEQUENCE to force register allocation.
1246 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1248 SDValue V2 = N->getOperand(2+3);
1249 // If it's a vld3, form a quad D-register and leave the last part as
1251 SDValue V3 = (NumVecs == 3)
1252 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1253 : N->getOperand(3+3);
1254 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1257 // Now extract the D registers back out.
1258 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT,
1260 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT,
1263 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,
1266 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,
1269 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1270 Ops.push_back(N->getOperand(Vec+3));
1272 Ops.push_back(Pred);
1273 Ops.push_back(Reg0); // predicate register
1274 Ops.push_back(Chain);
1275 unsigned Opc = DOpcodes[OpcodeIndex];
1276 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+5);
1279 EVT RegVT = GetNEONSubregVT(VT);
1281 // Quad registers are directly supported for VST1 and VST2,
1282 // storing pairs of D regs.
1283 unsigned Opc = QOpcodes0[OpcodeIndex];
1285 // First extract the pair of Q registers.
1286 SDValue Q0 = N->getOperand(3);
1287 SDValue Q1 = N->getOperand(4);
1289 // Form a QQ register.
1290 SDValue QQ = SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0);
1292 // Now extract the D registers back out.
1293 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1295 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1297 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, RegVT,
1299 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, RegVT,
1301 Ops.push_back(Pred);
1302 Ops.push_back(Reg0); // predicate register
1303 Ops.push_back(Chain);
1304 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 5 + 4);
1306 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1307 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1308 N->getOperand(Vec+3)));
1309 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1310 N->getOperand(Vec+3)));
1312 Ops.push_back(Pred);
1313 Ops.push_back(Reg0); // predicate register
1314 Ops.push_back(Chain);
1315 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
1320 // Otherwise, quad registers are stored with two separate instructions,
1321 // where one stores the even registers and the other stores the odd registers.
1323 // Form the QQQQ REG_SEQUENCE.
1325 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1326 V[i] = CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
1327 N->getOperand(Vec+3));
1328 V[i+1] = CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
1329 N->getOperand(Vec+3));
1332 V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1335 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1336 V[4], V[5], V[6], V[7]), 0);
1338 // Store the even D registers.
1339 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1340 Ops.push_back(Reg0); // post-access address offset
1341 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1342 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec*2, dl,
1344 Ops.push_back(Pred);
1345 Ops.push_back(Reg0); // predicate register
1346 Ops.push_back(Chain);
1347 unsigned Opc = QOpcodes0[OpcodeIndex];
1348 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1349 MVT::Other, Ops.data(), NumVecs+6);
1350 Chain = SDValue(VStA, 1);
1352 // Store the odd D registers.
1353 Ops[0] = SDValue(VStA, 0); // MemAddr
1354 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1355 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::dsub_1+Vec*2, dl,
1357 Ops[NumVecs+5] = Chain;
1358 Opc = QOpcodes1[OpcodeIndex];
1359 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1360 MVT::Other, Ops.data(), NumVecs+6);
1361 Chain = SDValue(VStB, 1);
1362 ReplaceUses(SDValue(N, 0), Chain);
1366 SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1367 unsigned NumVecs, unsigned *DOpcodes,
1368 unsigned *QOpcodes0,
1369 unsigned *QOpcodes1) {
1370 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1371 DebugLoc dl = N->getDebugLoc();
1373 SDValue MemAddr, Align;
1374 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1377 SDValue Chain = N->getOperand(0);
1379 cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1380 EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
1381 bool is64BitVector = VT.is64BitVector();
1383 // Quad registers are handled by load/store of subregs. Find the subreg info.
1384 unsigned NumElts = 0;
1388 if (!is64BitVector) {
1389 RegVT = GetNEONSubregVT(VT);
1390 NumElts = RegVT.getVectorNumElements();
1391 SubregIdx = (Lane < NumElts) ? ARM::dsub_0 : ARM::dsub_1;
1392 Even = Lane < NumElts;
1395 unsigned OpcodeIndex;
1396 switch (VT.getSimpleVT().SimpleTy) {
1397 default: llvm_unreachable("unhandled vld/vst lane type");
1398 // Double-register operations:
1399 case MVT::v8i8: OpcodeIndex = 0; break;
1400 case MVT::v4i16: OpcodeIndex = 1; break;
1402 case MVT::v2i32: OpcodeIndex = 2; break;
1403 // Quad-register operations:
1404 case MVT::v8i16: OpcodeIndex = 0; break;
1406 case MVT::v4i32: OpcodeIndex = 1; break;
1409 SDValue Pred = getAL(CurDAG);
1410 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1412 SmallVector<SDValue, 10> Ops;
1413 Ops.push_back(MemAddr);
1414 Ops.push_back(Align);
1417 if (is64BitVector) {
1418 Opc = DOpcodes[OpcodeIndex];
1420 SDValue V0 = N->getOperand(0+3);
1421 SDValue V1 = N->getOperand(1+3);
1423 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1425 SDValue V2 = N->getOperand(2+3);
1426 SDValue V3 = (NumVecs == 3)
1427 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1428 : N->getOperand(3+3);
1429 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1432 // Now extract the D registers back out.
1433 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT, RegSeq));
1434 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT, RegSeq));
1436 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,RegSeq));
1438 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,RegSeq));
1440 // Check if this is loading the even or odd subreg of a Q register.
1441 if (Lane < NumElts) {
1442 Opc = QOpcodes0[OpcodeIndex];
1445 Opc = QOpcodes1[OpcodeIndex];
1449 SDValue V0 = N->getOperand(0+3);
1450 SDValue V1 = N->getOperand(1+3);
1452 RegSeq = SDValue(PairQRegs(MVT::v4i64, V0, V1), 0);
1454 SDValue V2 = N->getOperand(2+3);
1455 SDValue V3 = (NumVecs == 3)
1456 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1457 : N->getOperand(3+3);
1458 RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
1461 // Extract the subregs of the input vector.
1462 unsigned SubIdx = Even ? ARM::dsub_0 : ARM::dsub_1;
1463 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1464 Ops.push_back(CurDAG->getTargetExtractSubreg(SubIdx+Vec*2, dl, RegVT,
1467 Ops.push_back(getI32Imm(Lane));
1468 Ops.push_back(Pred);
1469 Ops.push_back(Reg0);
1470 Ops.push_back(Chain);
1473 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6);
1475 std::vector<EVT> ResTys(NumVecs, RegVT);
1476 ResTys.push_back(MVT::Other);
1477 SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(),NumVecs+6);
1479 // Form a REG_SEQUENCE to force register allocation.
1481 if (is64BitVector) {
1482 SDValue V0 = SDValue(VLdLn, 0);
1483 SDValue V1 = SDValue(VLdLn, 1);
1485 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1487 SDValue V2 = SDValue(VLdLn, 2);
1488 // If it's a vld3, form a quad D-register but discard the last part.
1489 SDValue V3 = (NumVecs == 3)
1490 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1491 : SDValue(VLdLn, 3);
1492 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1495 // For 128-bit vectors, take the 64-bit results of the load and insert
1496 // them as subregs into the result.
1498 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1500 V[i] = SDValue(VLdLn, Vec);
1501 V[i+1] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1504 V[i] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1506 V[i+1] = SDValue(VLdLn, Vec);
1510 V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1514 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V[0], V[1], V[2], V[3]), 0);
1516 RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1517 V[4], V[5], V[6], V[7]), 0);
1520 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1521 assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1522 unsigned SubIdx = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
1523 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1524 ReplaceUses(SDValue(N, Vec),
1525 CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, RegSeq));
1526 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, NumVecs));
1530 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
1532 if (!Subtarget->hasV6T2Ops())
1535 unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
1536 : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
1539 // For unsigned extracts, check for a shift right and mask
1540 unsigned And_imm = 0;
1541 if (N->getOpcode() == ISD::AND) {
1542 if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
1544 // The immediate is a mask of the low bits iff imm & (imm+1) == 0
1545 if (And_imm & (And_imm + 1))
1548 unsigned Srl_imm = 0;
1549 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
1551 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1553 unsigned Width = CountTrailingOnes_32(And_imm);
1554 unsigned LSB = Srl_imm;
1555 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1556 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1557 CurDAG->getTargetConstant(LSB, MVT::i32),
1558 CurDAG->getTargetConstant(Width, MVT::i32),
1559 getAL(CurDAG), Reg0 };
1560 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1566 // Otherwise, we're looking for a shift of a shift
1567 unsigned Shl_imm = 0;
1568 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
1569 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1570 unsigned Srl_imm = 0;
1571 if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
1572 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1573 unsigned Width = 32 - Srl_imm;
1574 int LSB = Srl_imm - Shl_imm;
1577 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1578 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1579 CurDAG->getTargetConstant(LSB, MVT::i32),
1580 CurDAG->getTargetConstant(Width, MVT::i32),
1581 getAL(CurDAG), Reg0 };
1582 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1588 SDNode *ARMDAGToDAGISel::
1589 SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1590 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1593 if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
1594 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1595 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1598 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1599 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1600 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1601 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1603 llvm_unreachable("Unknown so_reg opcode!");
1607 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1608 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1609 SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
1610 return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
1615 SDNode *ARMDAGToDAGISel::
1616 SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1617 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1621 if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
1622 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1623 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
1624 return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
1629 SDNode *ARMDAGToDAGISel::
1630 SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1631 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1632 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1636 if (Predicate_t2_so_imm(TrueVal.getNode())) {
1637 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1638 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1639 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1640 return CurDAG->SelectNodeTo(N,
1641 ARM::t2MOVCCi, MVT::i32, Ops, 5);
1646 SDNode *ARMDAGToDAGISel::
1647 SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1648 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1649 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1653 if (Predicate_so_imm(TrueVal.getNode())) {
1654 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1655 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1656 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1657 return CurDAG->SelectNodeTo(N,
1658 ARM::MOVCCi, MVT::i32, Ops, 5);
1663 SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
1664 EVT VT = N->getValueType(0);
1665 SDValue FalseVal = N->getOperand(0);
1666 SDValue TrueVal = N->getOperand(1);
1667 SDValue CC = N->getOperand(2);
1668 SDValue CCR = N->getOperand(3);
1669 SDValue InFlag = N->getOperand(4);
1670 assert(CC.getOpcode() == ISD::Constant);
1671 assert(CCR.getOpcode() == ISD::Register);
1672 ARMCC::CondCodes CCVal =
1673 (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
1675 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1676 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1677 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1678 // Pattern complexity = 18 cost = 1 size = 0
1682 if (Subtarget->isThumb()) {
1683 SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
1684 CCVal, CCR, InFlag);
1686 Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
1687 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1691 SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
1692 CCVal, CCR, InFlag);
1694 Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
1695 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1700 // Pattern: (ARMcmov:i32 GPR:i32:$false,
1701 // (imm:i32)<<P:Predicate_so_imm>>:$true,
1703 // Emits: (MOVCCi:i32 GPR:i32:$false,
1704 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1705 // Pattern complexity = 10 cost = 1 size = 0
1706 if (Subtarget->isThumb()) {
1707 SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
1708 CCVal, CCR, InFlag);
1710 Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
1711 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1715 SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
1716 CCVal, CCR, InFlag);
1718 Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
1719 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1725 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1726 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1727 // Pattern complexity = 6 cost = 1 size = 0
1729 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1730 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1731 // Pattern complexity = 6 cost = 11 size = 0
1733 // Also FCPYScc and FCPYDcc.
1734 SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
1735 SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
1737 switch (VT.getSimpleVT().SimpleTy) {
1738 default: assert(false && "Illegal conditional move type!");
1741 Opc = Subtarget->isThumb()
1742 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1752 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1755 SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
1756 // The only time a CONCAT_VECTORS operation can have legal types is when
1757 // two 64-bit vectors are concatenated to a 128-bit vector.
1758 EVT VT = N->getValueType(0);
1759 if (!VT.is128BitVector() || N->getNumOperands() != 2)
1760 llvm_unreachable("unexpected CONCAT_VECTORS");
1761 DebugLoc dl = N->getDebugLoc();
1762 SDValue V0 = N->getOperand(0);
1763 SDValue V1 = N->getOperand(1);
1764 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1765 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1766 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
1767 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
1770 SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
1771 DebugLoc dl = N->getDebugLoc();
1773 if (N->isMachineOpcode())
1774 return NULL; // Already selected.
1776 switch (N->getOpcode()) {
1778 case ISD::Constant: {
1779 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1781 if (Subtarget->hasThumb2())
1782 // Thumb2-aware targets have the MOVT instruction, so all immediates can
1783 // be done with MOV + MOVT, at worst.
1786 if (Subtarget->isThumb()) {
1787 UseCP = (Val > 255 && // MOV
1788 ~Val > 255 && // MOV + MVN
1789 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
1791 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
1792 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
1793 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
1798 CurDAG->getTargetConstantPool(ConstantInt::get(
1799 Type::getInt32Ty(*CurDAG->getContext()), Val),
1800 TLI.getPointerTy());
1803 if (Subtarget->isThumb1Only()) {
1804 SDValue Pred = getAL(CurDAG);
1805 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1806 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1807 ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1812 CurDAG->getRegister(0, MVT::i32),
1813 CurDAG->getTargetConstant(0, MVT::i32),
1815 CurDAG->getRegister(0, MVT::i32),
1816 CurDAG->getEntryNode()
1818 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1821 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
1825 // Other cases are autogenerated.
1828 case ISD::FrameIndex: {
1829 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1830 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1831 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1832 if (Subtarget->isThumb1Only()) {
1833 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1834 CurDAG->getTargetConstant(0, MVT::i32));
1836 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1837 ARM::t2ADDri : ARM::ADDri);
1838 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1839 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1840 CurDAG->getRegister(0, MVT::i32) };
1841 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1845 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1849 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true))
1853 if (Subtarget->isThumb1Only())
1855 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1856 unsigned RHSV = C->getZExtValue();
1858 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
1859 unsigned ShImm = Log2_32(RHSV-1);
1862 SDValue V = N->getOperand(0);
1863 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1864 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1865 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1866 if (Subtarget->isThumb()) {
1867 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1868 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1870 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1871 return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1874 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
1875 unsigned ShImm = Log2_32(RHSV+1);
1878 SDValue V = N->getOperand(0);
1879 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1880 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1881 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1882 if (Subtarget->isThumb()) {
1883 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1884 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 6);
1886 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1887 return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1893 // Check for unsigned bitfield extract
1894 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1897 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
1898 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
1899 // are entirely contributed by c2 and lower 16-bits are entirely contributed
1900 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
1901 // Select it to: "movt x, ((c1 & 0xffff) >> 16)
1902 EVT VT = N->getValueType(0);
1905 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
1907 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
1910 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
1911 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1914 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
1915 SDValue N2 = N0.getOperand(1);
1916 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1919 unsigned N1CVal = N1C->getZExtValue();
1920 unsigned N2CVal = N2C->getZExtValue();
1921 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
1922 (N1CVal & 0xffffU) == 0xffffU &&
1923 (N2CVal & 0xffffU) == 0x0U) {
1924 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
1926 SDValue Ops[] = { N0.getOperand(0), Imm16,
1927 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1928 return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
1933 case ARMISD::VMOVRRD:
1934 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
1935 N->getOperand(0), getAL(CurDAG),
1936 CurDAG->getRegister(0, MVT::i32));
1937 case ISD::UMUL_LOHI: {
1938 if (Subtarget->isThumb1Only())
1940 if (Subtarget->isThumb()) {
1941 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1942 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1943 CurDAG->getRegister(0, MVT::i32) };
1944 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32,Ops,4);
1946 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1947 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1948 CurDAG->getRegister(0, MVT::i32) };
1949 return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1952 case ISD::SMUL_LOHI: {
1953 if (Subtarget->isThumb1Only())
1955 if (Subtarget->isThumb()) {
1956 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1957 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1958 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32,Ops,4);
1960 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1961 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1962 CurDAG->getRegister(0, MVT::i32) };
1963 return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1967 SDNode *ResNode = 0;
1968 if (Subtarget->isThumb() && Subtarget->hasThumb2())
1969 ResNode = SelectT2IndexedLoad(N);
1971 ResNode = SelectARMIndexedLoad(N);
1975 // VLDMQ must be custom-selected for "v2f64 load" to set the AM5Opc value.
1976 if (Subtarget->hasVFP2() &&
1977 N->getValueType(0).getSimpleVT().SimpleTy == MVT::v2f64) {
1978 SDValue Chain = N->getOperand(0);
1980 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
1981 SDValue Pred = getAL(CurDAG);
1982 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1983 SDValue Ops[] = { N->getOperand(1), AM5Opc, Pred, PredReg, Chain };
1984 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1985 MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
1986 SDNode *Ret = CurDAG->getMachineNode(ARM::VLDMQ, dl,
1987 MVT::v2f64, MVT::Other, Ops, 5);
1988 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1991 // Other cases are autogenerated.
1995 // VSTMQ must be custom-selected for "v2f64 store" to set the AM5Opc value.
1996 if (Subtarget->hasVFP2() &&
1997 N->getOperand(1).getValueType().getSimpleVT().SimpleTy == MVT::v2f64) {
1998 SDValue Chain = N->getOperand(0);
2000 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
2001 SDValue Pred = getAL(CurDAG);
2002 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2003 SDValue Ops[] = { N->getOperand(1), N->getOperand(2),
2004 AM5Opc, Pred, PredReg, Chain };
2005 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
2006 MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
2007 SDNode *Ret = CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
2008 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
2011 // Other cases are autogenerated.
2014 case ARMISD::BRCOND: {
2015 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2016 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2017 // Pattern complexity = 6 cost = 1 size = 0
2019 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2020 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
2021 // Pattern complexity = 6 cost = 1 size = 0
2023 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2024 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2025 // Pattern complexity = 6 cost = 1 size = 0
2027 unsigned Opc = Subtarget->isThumb() ?
2028 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
2029 SDValue Chain = N->getOperand(0);
2030 SDValue N1 = N->getOperand(1);
2031 SDValue N2 = N->getOperand(2);
2032 SDValue N3 = N->getOperand(3);
2033 SDValue InFlag = N->getOperand(4);
2034 assert(N1.getOpcode() == ISD::BasicBlock);
2035 assert(N2.getOpcode() == ISD::Constant);
2036 assert(N3.getOpcode() == ISD::Register);
2038 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2039 cast<ConstantSDNode>(N2)->getZExtValue()),
2041 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
2042 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
2044 Chain = SDValue(ResNode, 0);
2045 if (N->getNumValues() == 2) {
2046 InFlag = SDValue(ResNode, 1);
2047 ReplaceUses(SDValue(N, 1), InFlag);
2049 ReplaceUses(SDValue(N, 0),
2050 SDValue(Chain.getNode(), Chain.getResNo()));
2054 return SelectCMOVOp(N);
2055 case ARMISD::CNEG: {
2056 EVT VT = N->getValueType(0);
2057 SDValue N0 = N->getOperand(0);
2058 SDValue N1 = N->getOperand(1);
2059 SDValue N2 = N->getOperand(2);
2060 SDValue N3 = N->getOperand(3);
2061 SDValue InFlag = N->getOperand(4);
2062 assert(N2.getOpcode() == ISD::Constant);
2063 assert(N3.getOpcode() == ISD::Register);
2065 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2066 cast<ConstantSDNode>(N2)->getZExtValue()),
2068 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
2070 switch (VT.getSimpleVT().SimpleTy) {
2071 default: assert(false && "Illegal conditional move type!");
2080 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
2083 case ARMISD::VZIP: {
2085 EVT VT = N->getValueType(0);
2086 switch (VT.getSimpleVT().SimpleTy) {
2087 default: return NULL;
2088 case MVT::v8i8: Opc = ARM::VZIPd8; break;
2089 case MVT::v4i16: Opc = ARM::VZIPd16; break;
2091 case MVT::v2i32: Opc = ARM::VZIPd32; break;
2092 case MVT::v16i8: Opc = ARM::VZIPq8; break;
2093 case MVT::v8i16: Opc = ARM::VZIPq16; break;
2095 case MVT::v4i32: Opc = ARM::VZIPq32; break;
2097 SDValue Pred = getAL(CurDAG);
2098 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2099 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2100 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2102 case ARMISD::VUZP: {
2104 EVT VT = N->getValueType(0);
2105 switch (VT.getSimpleVT().SimpleTy) {
2106 default: return NULL;
2107 case MVT::v8i8: Opc = ARM::VUZPd8; break;
2108 case MVT::v4i16: Opc = ARM::VUZPd16; break;
2110 case MVT::v2i32: Opc = ARM::VUZPd32; break;
2111 case MVT::v16i8: Opc = ARM::VUZPq8; break;
2112 case MVT::v8i16: Opc = ARM::VUZPq16; break;
2114 case MVT::v4i32: Opc = ARM::VUZPq32; break;
2116 SDValue Pred = getAL(CurDAG);
2117 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2118 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2119 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2121 case ARMISD::VTRN: {
2123 EVT VT = N->getValueType(0);
2124 switch (VT.getSimpleVT().SimpleTy) {
2125 default: return NULL;
2126 case MVT::v8i8: Opc = ARM::VTRNd8; break;
2127 case MVT::v4i16: Opc = ARM::VTRNd16; break;
2129 case MVT::v2i32: Opc = ARM::VTRNd32; break;
2130 case MVT::v16i8: Opc = ARM::VTRNq8; break;
2131 case MVT::v8i16: Opc = ARM::VTRNq16; break;
2133 case MVT::v4i32: Opc = ARM::VTRNq32; break;
2135 SDValue Pred = getAL(CurDAG);
2136 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2137 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2138 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2140 case ARMISD::BUILD_VECTOR: {
2141 EVT VecVT = N->getValueType(0);
2142 EVT EltVT = VecVT.getVectorElementType();
2143 unsigned NumElts = VecVT.getVectorNumElements();
2144 if (EltVT.getSimpleVT() == MVT::f64) {
2145 assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
2146 return PairDRegs(VecVT, N->getOperand(0), N->getOperand(1));
2148 assert(EltVT.getSimpleVT() == MVT::f32 &&
2149 "unexpected type for BUILD_VECTOR");
2151 return PairSRegs(VecVT, N->getOperand(0), N->getOperand(1));
2152 assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
2153 return QuadSRegs(VecVT, N->getOperand(0), N->getOperand(1),
2154 N->getOperand(2), N->getOperand(3));
2157 case ISD::INTRINSIC_VOID:
2158 case ISD::INTRINSIC_W_CHAIN: {
2159 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2164 case Intrinsic::arm_neon_vld1: {
2165 unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
2166 ARM::VLD1d32, ARM::VLD1d64 };
2167 unsigned QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
2168 ARM::VLD1q32, ARM::VLD1q64 };
2169 return SelectVLD(N, 1, DOpcodes, QOpcodes, 0);
2172 case Intrinsic::arm_neon_vld2: {
2173 unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
2174 ARM::VLD2d32, ARM::VLD1q64 };
2175 unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
2176 return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
2179 case Intrinsic::arm_neon_vld3: {
2180 unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
2181 ARM::VLD3d32, ARM::VLD1d64T };
2182 unsigned QOpcodes0[] = { ARM::VLD3q8_UPD,
2185 unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD,
2186 ARM::VLD3q16odd_UPD,
2187 ARM::VLD3q32odd_UPD };
2188 return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2191 case Intrinsic::arm_neon_vld4: {
2192 unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
2193 ARM::VLD4d32, ARM::VLD1d64Q };
2194 unsigned QOpcodes0[] = { ARM::VLD4q8_UPD,
2197 unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD,
2198 ARM::VLD4q16odd_UPD,
2199 ARM::VLD4q32odd_UPD };
2200 return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2203 case Intrinsic::arm_neon_vld2lane: {
2204 unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
2205 unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 };
2206 unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd };
2207 return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
2210 case Intrinsic::arm_neon_vld3lane: {
2211 unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
2212 unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 };
2213 unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd };
2214 return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
2217 case Intrinsic::arm_neon_vld4lane: {
2218 unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
2219 unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 };
2220 unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd };
2221 return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
2224 case Intrinsic::arm_neon_vst1: {
2225 unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
2226 ARM::VST1d32, ARM::VST1d64 };
2227 unsigned QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
2228 ARM::VST1q32, ARM::VST1q64 };
2229 return SelectVST(N, 1, DOpcodes, QOpcodes, 0);
2232 case Intrinsic::arm_neon_vst2: {
2233 unsigned DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
2234 ARM::VST2d32, ARM::VST1q64 };
2235 unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
2236 return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
2239 case Intrinsic::arm_neon_vst3: {
2240 unsigned DOpcodes[] = { ARM::VST3d8, ARM::VST3d16,
2241 ARM::VST3d32, ARM::VST1d64T };
2242 unsigned QOpcodes0[] = { ARM::VST3q8_UPD,
2245 unsigned QOpcodes1[] = { ARM::VST3q8odd_UPD,
2246 ARM::VST3q16odd_UPD,
2247 ARM::VST3q32odd_UPD };
2248 return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2251 case Intrinsic::arm_neon_vst4: {
2252 unsigned DOpcodes[] = { ARM::VST4d8, ARM::VST4d16,
2253 ARM::VST4d32, ARM::VST1d64Q };
2254 unsigned QOpcodes0[] = { ARM::VST4q8_UPD,
2257 unsigned QOpcodes1[] = { ARM::VST4q8odd_UPD,
2258 ARM::VST4q16odd_UPD,
2259 ARM::VST4q32odd_UPD };
2260 return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2263 case Intrinsic::arm_neon_vst2lane: {
2264 unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
2265 unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 };
2266 unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd };
2267 return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
2270 case Intrinsic::arm_neon_vst3lane: {
2271 unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
2272 unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 };
2273 unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd };
2274 return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
2277 case Intrinsic::arm_neon_vst4lane: {
2278 unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
2279 unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 };
2280 unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd };
2281 return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
2287 case ISD::CONCAT_VECTORS:
2288 return SelectConcatVector(N);
2291 return SelectCode(N);
2294 bool ARMDAGToDAGISel::
2295 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2296 std::vector<SDValue> &OutOps) {
2297 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
2298 // Require the address to be in a register. That is safe for all ARM
2299 // variants and it is hard to do anything much smarter without knowing
2300 // how the operand is used.
2301 OutOps.push_back(Op);
2305 /// createARMISelDag - This pass converts a legalized DAG into a
2306 /// ARM-specific DAG, ready for instruction scheduling.
2308 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
2309 CodeGenOpt::Level OptLevel) {
2310 return new ARMDAGToDAGISel(TM, OptLevel);