1 //===-- ARMISelDAGToDAG.cpp - A dag to dag inst selector for ARM ----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines an instruction selector for the ARM target.
12 //===----------------------------------------------------------------------===//
15 #include "ARMAddressingModes.h"
16 #include "ARMConstantPoolValue.h"
17 #include "ARMISelLowering.h"
18 #include "ARMTargetMachine.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/Intrinsics.h"
24 #include "llvm/LLVMContext.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/SelectionDAGISel.h"
30 #include "llvm/Target/TargetLowering.h"
31 #include "llvm/Target/TargetOptions.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(SDValue Op);
69 virtual void InstructionSelect();
70 bool SelectShifterOperandReg(SDValue Op, SDValue N, SDValue &A,
71 SDValue &B, SDValue &C);
72 bool SelectAddrMode2(SDValue Op, SDValue N, SDValue &Base,
73 SDValue &Offset, SDValue &Opc);
74 bool SelectAddrMode2Offset(SDValue Op, SDValue N,
75 SDValue &Offset, SDValue &Opc);
76 bool SelectAddrMode3(SDValue Op, SDValue N, SDValue &Base,
77 SDValue &Offset, SDValue &Opc);
78 bool SelectAddrMode3Offset(SDValue Op, SDValue N,
79 SDValue &Offset, SDValue &Opc);
80 bool SelectAddrMode4(SDValue Op, SDValue N, SDValue &Addr,
82 bool SelectAddrMode5(SDValue Op, SDValue N, SDValue &Base,
84 bool SelectAddrMode6(SDValue Op, SDValue N, SDValue &Addr, SDValue &Update,
87 bool SelectAddrModePC(SDValue Op, SDValue N, SDValue &Offset,
90 bool SelectThumbAddrModeRR(SDValue Op, SDValue N, SDValue &Base,
92 bool SelectThumbAddrModeRI5(SDValue Op, SDValue N, unsigned Scale,
93 SDValue &Base, SDValue &OffImm,
95 bool SelectThumbAddrModeS1(SDValue Op, SDValue N, SDValue &Base,
96 SDValue &OffImm, SDValue &Offset);
97 bool SelectThumbAddrModeS2(SDValue Op, SDValue N, SDValue &Base,
98 SDValue &OffImm, SDValue &Offset);
99 bool SelectThumbAddrModeS4(SDValue Op, SDValue N, SDValue &Base,
100 SDValue &OffImm, SDValue &Offset);
101 bool SelectThumbAddrModeSP(SDValue Op, SDValue N, SDValue &Base,
104 bool SelectT2ShifterOperandReg(SDValue Op, SDValue N,
105 SDValue &BaseReg, SDValue &Opc);
106 bool SelectT2AddrModeImm12(SDValue Op, SDValue N, SDValue &Base,
108 bool SelectT2AddrModeImm8(SDValue Op, SDValue N, SDValue &Base,
110 bool SelectT2AddrModeImm8Offset(SDValue Op, SDValue N,
112 bool SelectT2AddrModeImm8s4(SDValue Op, SDValue N, SDValue &Base,
114 bool SelectT2AddrModeSoReg(SDValue Op, SDValue N, SDValue &Base,
115 SDValue &OffReg, SDValue &ShImm);
117 // Include the pieces autogenerated from the target description.
118 #include "ARMGenDAGISel.inc"
121 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
123 SDNode *SelectARMIndexedLoad(SDValue Op);
124 SDNode *SelectT2IndexedLoad(SDValue Op);
126 /// SelectDYN_ALLOC - Select dynamic alloc for Thumb.
127 SDNode *SelectDYN_ALLOC(SDValue Op);
129 /// SelectVLD - Select NEON load intrinsics. NumVecs should
130 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
131 /// loads of D registers and even subregs and odd subregs of Q registers.
132 /// For NumVecs == 2, QOpcodes1 is not used.
133 SDNode *SelectVLD(SDValue Op, unsigned NumVecs, unsigned *DOpcodes,
134 unsigned *QOpcodes0, unsigned *QOpcodes1);
136 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
137 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
138 /// load/store of D registers and even subregs and odd subregs of Q registers.
139 SDNode *SelectVLDSTLane(SDValue Op, bool IsLoad, unsigned NumVecs,
140 unsigned *DOpcodes, unsigned *QOpcodes0,
141 unsigned *QOpcodes1);
143 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
144 SDNode *SelectV6T2BitfieldExtractOp(SDValue Op, unsigned Opc);
146 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
147 /// inline asm expressions.
148 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
150 std::vector<SDValue> &OutOps);
152 /// PairDRegs - Insert a pair of double registers into an implicit def to
153 /// form a quad register.
154 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
158 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
159 /// operand. If so Imm will receive the 32-bit value.
160 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
161 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
162 Imm = cast<ConstantSDNode>(N)->getZExtValue();
168 // isInt32Immediate - This method tests to see if a constant operand.
169 // If so Imm will receive the 32 bit value.
170 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
171 return isInt32Immediate(N.getNode(), Imm);
174 // isOpcWithIntImmediate - This method tests to see if the node is a specific
175 // opcode and that it has a immediate integer right operand.
176 // If so Imm will receive the 32 bit value.
177 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
178 return N->getOpcode() == Opc &&
179 isInt32Immediate(N->getOperand(1).getNode(), Imm);
183 void ARMDAGToDAGISel::InstructionSelect() {
187 CurDAG->RemoveDeadNodes();
190 bool ARMDAGToDAGISel::SelectShifterOperandReg(SDValue Op,
195 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
197 // Don't match base register only case. That is matched to a separate
198 // lower complexity pattern with explicit register operand.
199 if (ShOpcVal == ARM_AM::no_shift) return false;
201 BaseReg = N.getOperand(0);
202 unsigned ShImmVal = 0;
203 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
204 ShReg = CurDAG->getRegister(0, MVT::i32);
205 ShImmVal = RHS->getZExtValue() & 31;
207 ShReg = N.getOperand(1);
209 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
214 bool ARMDAGToDAGISel::SelectAddrMode2(SDValue Op, SDValue N,
215 SDValue &Base, SDValue &Offset,
217 if (N.getOpcode() == ISD::MUL) {
218 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
219 // X * [3,5,9] -> X + X * [2,4,8] etc.
220 int RHSC = (int)RHS->getZExtValue();
223 ARM_AM::AddrOpc AddSub = ARM_AM::add;
225 AddSub = ARM_AM::sub;
228 if (isPowerOf2_32(RHSC)) {
229 unsigned ShAmt = Log2_32(RHSC);
230 Base = Offset = N.getOperand(0);
231 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
240 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
242 if (N.getOpcode() == ISD::FrameIndex) {
243 int FI = cast<FrameIndexSDNode>(N)->getIndex();
244 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
245 } else if (N.getOpcode() == ARMISD::Wrapper) {
246 Base = N.getOperand(0);
248 Offset = CurDAG->getRegister(0, MVT::i32);
249 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
255 // Match simple R +/- imm12 operands.
256 if (N.getOpcode() == ISD::ADD)
257 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
258 int RHSC = (int)RHS->getZExtValue();
259 if ((RHSC >= 0 && RHSC < 0x1000) ||
260 (RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
261 Base = N.getOperand(0);
262 if (Base.getOpcode() == ISD::FrameIndex) {
263 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
264 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
266 Offset = CurDAG->getRegister(0, MVT::i32);
268 ARM_AM::AddrOpc AddSub = ARM_AM::add;
270 AddSub = ARM_AM::sub;
273 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
280 // Otherwise this is R +/- [possibly shifted] R
281 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
282 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
285 Base = N.getOperand(0);
286 Offset = N.getOperand(1);
288 if (ShOpcVal != ARM_AM::no_shift) {
289 // Check to see if the RHS of the shift is a constant, if not, we can't fold
291 if (ConstantSDNode *Sh =
292 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
293 ShAmt = Sh->getZExtValue();
294 Offset = N.getOperand(1).getOperand(0);
296 ShOpcVal = ARM_AM::no_shift;
300 // Try matching (R shl C) + (R).
301 if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
302 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
303 if (ShOpcVal != ARM_AM::no_shift) {
304 // Check to see if the RHS of the shift is a constant, if not, we can't
306 if (ConstantSDNode *Sh =
307 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
308 ShAmt = Sh->getZExtValue();
309 Offset = N.getOperand(0).getOperand(0);
310 Base = N.getOperand(1);
312 ShOpcVal = ARM_AM::no_shift;
317 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
322 bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDValue Op, SDValue N,
323 SDValue &Offset, SDValue &Opc) {
324 unsigned Opcode = Op.getOpcode();
325 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
326 ? cast<LoadSDNode>(Op)->getAddressingMode()
327 : cast<StoreSDNode>(Op)->getAddressingMode();
328 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
329 ? ARM_AM::add : ARM_AM::sub;
330 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
331 int Val = (int)C->getZExtValue();
332 if (Val >= 0 && Val < 0x1000) { // 12 bits.
333 Offset = CurDAG->getRegister(0, MVT::i32);
334 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
342 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
344 if (ShOpcVal != ARM_AM::no_shift) {
345 // Check to see if the RHS of the shift is a constant, if not, we can't fold
347 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
348 ShAmt = Sh->getZExtValue();
349 Offset = N.getOperand(0);
351 ShOpcVal = ARM_AM::no_shift;
355 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
361 bool ARMDAGToDAGISel::SelectAddrMode3(SDValue Op, SDValue N,
362 SDValue &Base, SDValue &Offset,
364 if (N.getOpcode() == ISD::SUB) {
365 // X - C is canonicalize to X + -C, no need to handle it here.
366 Base = N.getOperand(0);
367 Offset = N.getOperand(1);
368 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
372 if (N.getOpcode() != ISD::ADD) {
374 if (N.getOpcode() == ISD::FrameIndex) {
375 int FI = cast<FrameIndexSDNode>(N)->getIndex();
376 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
378 Offset = CurDAG->getRegister(0, MVT::i32);
379 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
383 // If the RHS is +/- imm8, fold into addr mode.
384 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
385 int RHSC = (int)RHS->getZExtValue();
386 if ((RHSC >= 0 && RHSC < 256) ||
387 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
388 Base = N.getOperand(0);
389 if (Base.getOpcode() == ISD::FrameIndex) {
390 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
391 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
393 Offset = CurDAG->getRegister(0, MVT::i32);
395 ARM_AM::AddrOpc AddSub = ARM_AM::add;
397 AddSub = ARM_AM::sub;
400 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
405 Base = N.getOperand(0);
406 Offset = N.getOperand(1);
407 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
411 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDValue Op, SDValue N,
412 SDValue &Offset, SDValue &Opc) {
413 unsigned Opcode = Op.getOpcode();
414 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
415 ? cast<LoadSDNode>(Op)->getAddressingMode()
416 : cast<StoreSDNode>(Op)->getAddressingMode();
417 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
418 ? ARM_AM::add : ARM_AM::sub;
419 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
420 int Val = (int)C->getZExtValue();
421 if (Val >= 0 && Val < 256) {
422 Offset = CurDAG->getRegister(0, MVT::i32);
423 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
429 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
433 bool ARMDAGToDAGISel::SelectAddrMode4(SDValue Op, SDValue N,
434 SDValue &Addr, SDValue &Mode) {
436 Mode = CurDAG->getTargetConstant(0, MVT::i32);
440 bool ARMDAGToDAGISel::SelectAddrMode5(SDValue Op, SDValue N,
441 SDValue &Base, SDValue &Offset) {
442 if (N.getOpcode() != ISD::ADD) {
444 if (N.getOpcode() == ISD::FrameIndex) {
445 int FI = cast<FrameIndexSDNode>(N)->getIndex();
446 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
447 } else if (N.getOpcode() == ARMISD::Wrapper) {
448 Base = N.getOperand(0);
450 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
455 // If the RHS is +/- imm8, fold into addr mode.
456 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
457 int RHSC = (int)RHS->getZExtValue();
458 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied by 4.
460 if ((RHSC >= 0 && RHSC < 256) ||
461 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
462 Base = N.getOperand(0);
463 if (Base.getOpcode() == ISD::FrameIndex) {
464 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
465 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
468 ARM_AM::AddrOpc AddSub = ARM_AM::add;
470 AddSub = ARM_AM::sub;
473 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
481 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
486 bool ARMDAGToDAGISel::SelectAddrMode6(SDValue Op, SDValue N,
487 SDValue &Addr, SDValue &Update,
490 // Default to no writeback.
491 Update = CurDAG->getRegister(0, MVT::i32);
492 Opc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(false), MVT::i32);
496 bool ARMDAGToDAGISel::SelectAddrModePC(SDValue Op, SDValue N,
497 SDValue &Offset, SDValue &Label) {
498 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
499 Offset = N.getOperand(0);
500 SDValue N1 = N.getOperand(1);
501 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
508 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue Op, SDValue N,
509 SDValue &Base, SDValue &Offset){
510 // FIXME dl should come from the parent load or store, not the address
511 DebugLoc dl = Op.getDebugLoc();
512 if (N.getOpcode() != ISD::ADD) {
513 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
514 if (!NC || NC->getZExtValue() != 0)
521 Base = N.getOperand(0);
522 Offset = N.getOperand(1);
527 ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDValue Op, SDValue N,
528 unsigned Scale, SDValue &Base,
529 SDValue &OffImm, SDValue &Offset) {
531 SDValue TmpBase, TmpOffImm;
532 if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
533 return false; // We want to select tLDRspi / tSTRspi instead.
534 if (N.getOpcode() == ARMISD::Wrapper &&
535 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
536 return false; // We want to select tLDRpci instead.
539 if (N.getOpcode() != ISD::ADD) {
540 Base = (N.getOpcode() == ARMISD::Wrapper) ? N.getOperand(0) : N;
541 Offset = CurDAG->getRegister(0, MVT::i32);
542 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
546 // Thumb does not have [sp, r] address mode.
547 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
548 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
549 if ((LHSR && LHSR->getReg() == ARM::SP) ||
550 (RHSR && RHSR->getReg() == ARM::SP)) {
552 Offset = CurDAG->getRegister(0, MVT::i32);
553 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
557 // If the RHS is + imm5 * scale, fold into addr mode.
558 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
559 int RHSC = (int)RHS->getZExtValue();
560 if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
562 if (RHSC >= 0 && RHSC < 32) {
563 Base = N.getOperand(0);
564 Offset = CurDAG->getRegister(0, MVT::i32);
565 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
571 Base = N.getOperand(0);
572 Offset = N.getOperand(1);
573 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
577 bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDValue Op, SDValue N,
578 SDValue &Base, SDValue &OffImm,
580 return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
583 bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDValue Op, SDValue N,
584 SDValue &Base, SDValue &OffImm,
586 return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
589 bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDValue Op, SDValue N,
590 SDValue &Base, SDValue &OffImm,
592 return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
595 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDValue Op, SDValue N,
596 SDValue &Base, SDValue &OffImm) {
597 if (N.getOpcode() == ISD::FrameIndex) {
598 int FI = cast<FrameIndexSDNode>(N)->getIndex();
599 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
600 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
604 if (N.getOpcode() != ISD::ADD)
607 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
608 if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
609 (LHSR && LHSR->getReg() == ARM::SP)) {
610 // If the RHS is + imm8 * scale, fold into addr mode.
611 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
612 int RHSC = (int)RHS->getZExtValue();
613 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
615 if (RHSC >= 0 && RHSC < 256) {
616 Base = N.getOperand(0);
617 if (Base.getOpcode() == ISD::FrameIndex) {
618 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
619 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
621 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
631 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDValue Op, SDValue N,
634 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
636 // Don't match base register only case. That is matched to a separate
637 // lower complexity pattern with explicit register operand.
638 if (ShOpcVal == ARM_AM::no_shift) return false;
640 BaseReg = N.getOperand(0);
641 unsigned ShImmVal = 0;
642 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
643 ShImmVal = RHS->getZExtValue() & 31;
644 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
651 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDValue Op, SDValue N,
652 SDValue &Base, SDValue &OffImm) {
653 // Match simple R + imm12 operands.
656 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
657 if (N.getOpcode() == ISD::FrameIndex) {
658 // Match frame index...
659 int FI = cast<FrameIndexSDNode>(N)->getIndex();
660 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
661 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
663 } else if (N.getOpcode() == ARMISD::Wrapper) {
664 Base = N.getOperand(0);
665 if (Base.getOpcode() == ISD::TargetConstantPool)
666 return false; // We want to select t2LDRpci instead.
669 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
673 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
674 if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
675 // Let t2LDRi8 handle (R - imm8).
678 int RHSC = (int)RHS->getZExtValue();
679 if (N.getOpcode() == ISD::SUB)
682 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
683 Base = N.getOperand(0);
684 if (Base.getOpcode() == ISD::FrameIndex) {
685 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
686 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
688 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
695 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
699 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue Op, SDValue N,
700 SDValue &Base, SDValue &OffImm) {
701 // Match simple R - imm8 operands.
702 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
703 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
704 int RHSC = (int)RHS->getSExtValue();
705 if (N.getOpcode() == ISD::SUB)
708 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
709 Base = N.getOperand(0);
710 if (Base.getOpcode() == ISD::FrameIndex) {
711 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
712 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
714 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
723 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDValue Op, SDValue N,
725 unsigned Opcode = Op.getOpcode();
726 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
727 ? cast<LoadSDNode>(Op)->getAddressingMode()
728 : cast<StoreSDNode>(Op)->getAddressingMode();
729 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
730 int RHSC = (int)RHS->getZExtValue();
731 if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
732 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
733 ? CurDAG->getTargetConstant(RHSC, MVT::i32)
734 : CurDAG->getTargetConstant(-RHSC, MVT::i32);
742 bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDValue Op, SDValue N,
743 SDValue &Base, SDValue &OffImm) {
744 if (N.getOpcode() == ISD::ADD) {
745 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
746 int RHSC = (int)RHS->getZExtValue();
747 if (((RHSC & 0x3) == 0) &&
748 ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits.
749 Base = N.getOperand(0);
750 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
754 } else if (N.getOpcode() == ISD::SUB) {
755 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
756 int RHSC = (int)RHS->getZExtValue();
757 if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) { // 8 bits.
758 Base = N.getOperand(0);
759 OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
768 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue Op, SDValue N,
770 SDValue &OffReg, SDValue &ShImm) {
771 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
772 if (N.getOpcode() != ISD::ADD)
775 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
776 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
777 int RHSC = (int)RHS->getZExtValue();
778 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
780 else if (RHSC < 0 && RHSC >= -255) // 8 bits
784 // Look for (R + R) or (R + (R << [1,2,3])).
786 Base = N.getOperand(0);
787 OffReg = N.getOperand(1);
789 // Swap if it is ((R << c) + R).
790 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
791 if (ShOpcVal != ARM_AM::lsl) {
792 ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
793 if (ShOpcVal == ARM_AM::lsl)
794 std::swap(Base, OffReg);
797 if (ShOpcVal == ARM_AM::lsl) {
798 // Check to see if the RHS of the shift is a constant, if not, we can't fold
800 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
801 ShAmt = Sh->getZExtValue();
804 ShOpcVal = ARM_AM::no_shift;
806 OffReg = OffReg.getOperand(0);
808 ShOpcVal = ARM_AM::no_shift;
812 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
817 //===--------------------------------------------------------------------===//
819 /// getAL - Returns a ARMCC::AL immediate node.
820 static inline SDValue getAL(SelectionDAG *CurDAG) {
821 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
824 SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDValue Op) {
825 LoadSDNode *LD = cast<LoadSDNode>(Op);
826 ISD::MemIndexedMode AM = LD->getAddressingMode();
827 if (AM == ISD::UNINDEXED)
830 EVT LoadedVT = LD->getMemoryVT();
831 SDValue Offset, AMOpc;
832 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
835 if (LoadedVT == MVT::i32 &&
836 SelectAddrMode2Offset(Op, LD->getOffset(), Offset, AMOpc)) {
837 Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
839 } else if (LoadedVT == MVT::i16 &&
840 SelectAddrMode3Offset(Op, LD->getOffset(), Offset, AMOpc)) {
842 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
843 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
844 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
845 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
846 if (LD->getExtensionType() == ISD::SEXTLOAD) {
847 if (SelectAddrMode3Offset(Op, LD->getOffset(), Offset, AMOpc)) {
849 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
852 if (SelectAddrMode2Offset(Op, LD->getOffset(), Offset, AMOpc)) {
854 Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
860 SDValue Chain = LD->getChain();
861 SDValue Base = LD->getBasePtr();
862 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
863 CurDAG->getRegister(0, MVT::i32), Chain };
864 return CurDAG->getMachineNode(Opcode, Op.getDebugLoc(), MVT::i32, MVT::i32,
871 SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDValue Op) {
872 LoadSDNode *LD = cast<LoadSDNode>(Op);
873 ISD::MemIndexedMode AM = LD->getAddressingMode();
874 if (AM == ISD::UNINDEXED)
877 EVT LoadedVT = LD->getMemoryVT();
878 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
880 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
883 if (SelectT2AddrModeImm8Offset(Op, LD->getOffset(), Offset)) {
884 switch (LoadedVT.getSimpleVT().SimpleTy) {
886 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
890 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
892 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
897 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
899 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
908 SDValue Chain = LD->getChain();
909 SDValue Base = LD->getBasePtr();
910 SDValue Ops[]= { Base, Offset, getAL(CurDAG),
911 CurDAG->getRegister(0, MVT::i32), Chain };
912 return CurDAG->getMachineNode(Opcode, Op.getDebugLoc(), MVT::i32, MVT::i32,
919 SDNode *ARMDAGToDAGISel::SelectDYN_ALLOC(SDValue Op) {
920 SDNode *N = Op.getNode();
921 DebugLoc dl = N->getDebugLoc();
922 EVT VT = Op.getValueType();
923 SDValue Chain = Op.getOperand(0);
924 SDValue Size = Op.getOperand(1);
925 SDValue Align = Op.getOperand(2);
926 SDValue SP = CurDAG->getRegister(ARM::SP, MVT::i32);
927 int32_t AlignVal = cast<ConstantSDNode>(Align)->getSExtValue();
929 // We need to align the stack. Use Thumb1 tAND which is the only thumb
930 // instruction that can read and write SP. This matches to a pseudo
931 // instruction that has a chain to ensure the result is written back to
932 // the stack pointer.
933 SP = SDValue(CurDAG->getMachineNode(ARM::tANDsp, dl, VT, SP, Align), 0);
935 bool isC = isa<ConstantSDNode>(Size);
936 uint32_t C = isC ? cast<ConstantSDNode>(Size)->getZExtValue() : ~0UL;
937 // Handle the most common case for both Thumb1 and Thumb2:
938 // tSUBspi - immediate is between 0 ... 508 inclusive.
939 if (C <= 508 && ((C & 3) == 0))
940 // FIXME: tSUBspi encode scale 4 implicitly.
941 return CurDAG->SelectNodeTo(N, ARM::tSUBspi_, VT, MVT::Other, SP,
942 CurDAG->getTargetConstant(C/4, MVT::i32),
945 if (Subtarget->isThumb1Only()) {
946 // Use tADDspr since Thumb1 does not have a sub r, sp, r. ARMISelLowering
947 // should have negated the size operand already. FIXME: We can't insert
948 // new target independent node at this stage so we are forced to negate
949 // it earlier. Is there a better solution?
950 return CurDAG->SelectNodeTo(N, ARM::tADDspr_, VT, MVT::Other, SP, Size,
952 } else if (Subtarget->isThumb2()) {
953 if (isC && Predicate_t2_so_imm(Size.getNode())) {
955 SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
956 return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi_, VT, MVT::Other, Ops, 3);
957 } else if (isC && Predicate_imm0_4095(Size.getNode())) {
959 SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
960 return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi12_, VT, MVT::Other, Ops, 3);
963 SDValue Ops[] = { SP, Size,
964 getI32Imm(ARM_AM::getSORegOpc(ARM_AM::lsl,0)), Chain };
965 return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPs_, VT, MVT::Other, Ops, 4);
969 // FIXME: Add ADD / SUB sp instructions for ARM.
973 /// PairDRegs - Insert a pair of double registers into an implicit def to
974 /// form a quad register.
975 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
976 DebugLoc dl = V0.getNode()->getDebugLoc();
978 SDValue(CurDAG->getMachineNode(TargetInstrInfo::IMPLICIT_DEF, dl, VT), 0);
979 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
980 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
981 SDNode *Pair = CurDAG->getMachineNode(TargetInstrInfo::INSERT_SUBREG, dl,
982 VT, Undef, V0, SubReg0);
983 return CurDAG->getMachineNode(TargetInstrInfo::INSERT_SUBREG, dl,
984 VT, SDValue(Pair, 0), V1, SubReg1);
987 /// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
988 /// for a 64-bit subregister of the vector.
989 static EVT GetNEONSubregVT(EVT VT) {
990 switch (VT.getSimpleVT().SimpleTy) {
991 default: llvm_unreachable("unhandled NEON type");
992 case MVT::v16i8: return MVT::v8i8;
993 case MVT::v8i16: return MVT::v4i16;
994 case MVT::v4f32: return MVT::v2f32;
995 case MVT::v4i32: return MVT::v2i32;
996 case MVT::v2i64: return MVT::v1i64;
1000 SDNode *ARMDAGToDAGISel::SelectVLD(SDValue Op, unsigned NumVecs,
1001 unsigned *DOpcodes, unsigned *QOpcodes0,
1002 unsigned *QOpcodes1) {
1003 assert(NumVecs >=2 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1004 SDNode *N = Op.getNode();
1005 DebugLoc dl = N->getDebugLoc();
1007 SDValue MemAddr, MemUpdate, MemOpc;
1008 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1011 SDValue Chain = N->getOperand(0);
1012 EVT VT = N->getValueType(0);
1013 bool is64BitVector = VT.is64BitVector();
1015 unsigned OpcodeIndex;
1016 switch (VT.getSimpleVT().SimpleTy) {
1017 default: llvm_unreachable("unhandled vld type");
1018 // Double-register operations:
1019 case MVT::v8i8: OpcodeIndex = 0; break;
1020 case MVT::v4i16: OpcodeIndex = 1; break;
1022 case MVT::v2i32: OpcodeIndex = 2; break;
1023 case MVT::v1i64: OpcodeIndex = 3; break;
1024 // Quad-register operations:
1025 case MVT::v16i8: OpcodeIndex = 0; break;
1026 case MVT::v8i16: OpcodeIndex = 1; break;
1028 case MVT::v4i32: OpcodeIndex = 2; break;
1031 if (is64BitVector) {
1032 unsigned Opc = DOpcodes[OpcodeIndex];
1033 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
1034 std::vector<EVT> ResTys(NumVecs, VT);
1035 ResTys.push_back(MVT::Other);
1036 return CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 4);
1039 EVT RegVT = GetNEONSubregVT(VT);
1041 // Quad registers are directly supported for VLD2,
1042 // loading 2 pairs of D regs.
1043 unsigned Opc = QOpcodes0[OpcodeIndex];
1044 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
1045 std::vector<EVT> ResTys(4, VT);
1046 ResTys.push_back(MVT::Other);
1047 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 4);
1048 Chain = SDValue(VLd, 4);
1050 // Combine the even and odd subregs to produce the result.
1051 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1052 SDNode *Q = PairDRegs(VT, SDValue(VLd, 2*Vec), SDValue(VLd, 2*Vec+1));
1053 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1056 // Otherwise, quad registers are loaded with two separate instructions,
1057 // where one loads the even registers and the other loads the odd registers.
1059 // Enable writeback to the address register.
1060 MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
1062 std::vector<EVT> ResTys(NumVecs, RegVT);
1063 ResTys.push_back(MemAddr.getValueType());
1064 ResTys.push_back(MVT::Other);
1066 // Load the even subreg.
1067 unsigned Opc = QOpcodes0[OpcodeIndex];
1068 const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, Chain };
1069 SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 4);
1070 Chain = SDValue(VLdA, NumVecs+1);
1072 // Load the odd subreg.
1073 Opc = QOpcodes1[OpcodeIndex];
1074 const SDValue OpsB[] = { SDValue(VLdA, NumVecs), MemUpdate, MemOpc, Chain };
1075 SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 4);
1076 Chain = SDValue(VLdB, NumVecs+1);
1078 // Combine the even and odd subregs to produce the result.
1079 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1080 SDNode *Q = PairDRegs(VT, SDValue(VLdA, Vec), SDValue(VLdB, Vec));
1081 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1084 ReplaceUses(SDValue(N, NumVecs), Chain);
1088 SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDValue Op, bool IsLoad,
1089 unsigned NumVecs, unsigned *DOpcodes,
1090 unsigned *QOpcodes0,
1091 unsigned *QOpcodes1) {
1092 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1093 SDNode *N = Op.getNode();
1094 DebugLoc dl = N->getDebugLoc();
1096 SDValue MemAddr, MemUpdate, MemOpc;
1097 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1100 SDValue Chain = N->getOperand(0);
1102 cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1103 EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
1104 bool is64BitVector = VT.is64BitVector();
1106 // Quad registers are handled by load/store of subregs. Find the subreg info.
1107 unsigned NumElts = 0;
1110 if (!is64BitVector) {
1111 RegVT = GetNEONSubregVT(VT);
1112 NumElts = RegVT.getVectorNumElements();
1113 SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
1116 unsigned OpcodeIndex;
1117 switch (VT.getSimpleVT().SimpleTy) {
1118 default: llvm_unreachable("unhandled vld/vst lane type");
1119 // Double-register operations:
1120 case MVT::v8i8: OpcodeIndex = 0; break;
1121 case MVT::v4i16: OpcodeIndex = 1; break;
1123 case MVT::v2i32: OpcodeIndex = 2; break;
1124 // Quad-register operations:
1125 case MVT::v8i16: OpcodeIndex = 0; break;
1127 case MVT::v4i32: OpcodeIndex = 1; break;
1130 SmallVector<SDValue, 9> Ops;
1131 Ops.push_back(MemAddr);
1132 Ops.push_back(MemUpdate);
1133 Ops.push_back(MemOpc);
1136 if (is64BitVector) {
1137 Opc = DOpcodes[OpcodeIndex];
1138 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1139 Ops.push_back(N->getOperand(Vec+3));
1141 // Check if this is loading the even or odd subreg of a Q register.
1142 if (Lane < NumElts) {
1143 Opc = QOpcodes0[OpcodeIndex];
1146 Opc = QOpcodes1[OpcodeIndex];
1148 // Extract the subregs of the input vector.
1149 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1150 Ops.push_back(CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1151 N->getOperand(Vec+3)));
1153 Ops.push_back(getI32Imm(Lane));
1154 Ops.push_back(Chain);
1157 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+5);
1159 std::vector<EVT> ResTys(NumVecs, RegVT);
1160 ResTys.push_back(MVT::Other);
1162 CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), NumVecs+5);
1163 // For a 64-bit vector load to D registers, nothing more needs to be done.
1167 // For 128-bit vectors, take the 64-bit results of the load and insert them
1168 // as subregs into the result.
1169 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1170 SDValue QuadVec = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
1171 N->getOperand(Vec+3),
1172 SDValue(VLdLn, Vec));
1173 ReplaceUses(SDValue(N, Vec), QuadVec);
1176 Chain = SDValue(VLdLn, NumVecs);
1177 ReplaceUses(SDValue(N, NumVecs), Chain);
1181 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDValue Op,
1183 if (!Subtarget->hasV6T2Ops())
1186 unsigned Shl_imm = 0;
1187 if (isOpcWithIntImmediate(Op.getOperand(0).getNode(), ISD::SHL, Shl_imm)){
1188 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1189 unsigned Srl_imm = 0;
1190 if (isInt32Immediate(Op.getOperand(1), Srl_imm)) {
1191 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1192 unsigned Width = 32 - Srl_imm;
1193 int LSB = Srl_imm - Shl_imm;
1194 if ((LSB + Width) > 32)
1196 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1197 SDValue Ops[] = { Op.getOperand(0).getOperand(0),
1198 CurDAG->getTargetConstant(LSB, MVT::i32),
1199 CurDAG->getTargetConstant(Width, MVT::i32),
1200 getAL(CurDAG), Reg0 };
1201 return CurDAG->SelectNodeTo(Op.getNode(), Opc, MVT::i32, Ops, 5);
1207 SDNode *ARMDAGToDAGISel::Select(SDValue Op) {
1208 SDNode *N = Op.getNode();
1209 DebugLoc dl = N->getDebugLoc();
1211 if (N->isMachineOpcode())
1212 return NULL; // Already selected.
1214 switch (N->getOpcode()) {
1216 case ISD::Constant: {
1217 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1219 if (Subtarget->hasThumb2())
1220 // Thumb2-aware targets have the MOVT instruction, so all immediates can
1221 // be done with MOV + MOVT, at worst.
1224 if (Subtarget->isThumb()) {
1225 UseCP = (Val > 255 && // MOV
1226 ~Val > 255 && // MOV + MVN
1227 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
1229 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
1230 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
1231 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
1236 CurDAG->getTargetConstantPool(ConstantInt::get(
1237 Type::getInt32Ty(*CurDAG->getContext()), Val),
1238 TLI.getPointerTy());
1241 if (Subtarget->isThumb1Only()) {
1242 SDValue Pred = CurDAG->getTargetConstant(0xEULL, MVT::i32);
1243 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1244 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1245 ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1250 CurDAG->getRegister(0, MVT::i32),
1251 CurDAG->getTargetConstant(0, MVT::i32),
1253 CurDAG->getRegister(0, MVT::i32),
1254 CurDAG->getEntryNode()
1256 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1259 ReplaceUses(Op, SDValue(ResNode, 0));
1263 // Other cases are autogenerated.
1266 case ISD::FrameIndex: {
1267 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1268 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1269 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1270 if (Subtarget->isThumb1Only()) {
1271 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1272 CurDAG->getTargetConstant(0, MVT::i32));
1274 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1275 ARM::t2ADDri : ARM::ADDri);
1276 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1277 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1278 CurDAG->getRegister(0, MVT::i32) };
1279 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1282 case ARMISD::DYN_ALLOC:
1283 return SelectDYN_ALLOC(Op);
1285 if (SDNode *I = SelectV6T2BitfieldExtractOp(Op,
1286 Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX))
1290 if (SDNode *I = SelectV6T2BitfieldExtractOp(Op,
1291 Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX))
1295 if (Subtarget->isThumb1Only())
1297 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1298 unsigned RHSV = C->getZExtValue();
1300 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
1301 unsigned ShImm = Log2_32(RHSV-1);
1304 SDValue V = Op.getOperand(0);
1305 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1306 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1307 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1308 if (Subtarget->isThumb()) {
1309 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1310 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1312 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1313 return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1316 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
1317 unsigned ShImm = Log2_32(RHSV+1);
1320 SDValue V = Op.getOperand(0);
1321 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1322 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1323 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1324 if (Subtarget->isThumb()) {
1325 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0 };
1326 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 5);
1328 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1329 return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1335 return CurDAG->getMachineNode(ARM::FMRRD, dl, MVT::i32, MVT::i32,
1336 Op.getOperand(0), getAL(CurDAG),
1337 CurDAG->getRegister(0, MVT::i32));
1338 case ISD::UMUL_LOHI: {
1339 if (Subtarget->isThumb1Only())
1341 if (Subtarget->isThumb()) {
1342 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
1343 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1344 CurDAG->getRegister(0, MVT::i32) };
1345 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops,4);
1347 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
1348 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1349 CurDAG->getRegister(0, MVT::i32) };
1350 return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1353 case ISD::SMUL_LOHI: {
1354 if (Subtarget->isThumb1Only())
1356 if (Subtarget->isThumb()) {
1357 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
1358 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1359 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops,4);
1361 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
1362 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1363 CurDAG->getRegister(0, MVT::i32) };
1364 return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1368 SDNode *ResNode = 0;
1369 if (Subtarget->isThumb() && Subtarget->hasThumb2())
1370 ResNode = SelectT2IndexedLoad(Op);
1372 ResNode = SelectARMIndexedLoad(Op);
1375 // Other cases are autogenerated.
1378 case ARMISD::BRCOND: {
1379 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1380 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
1381 // Pattern complexity = 6 cost = 1 size = 0
1383 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1384 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
1385 // Pattern complexity = 6 cost = 1 size = 0
1387 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1388 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
1389 // Pattern complexity = 6 cost = 1 size = 0
1391 unsigned Opc = Subtarget->isThumb() ?
1392 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
1393 SDValue Chain = Op.getOperand(0);
1394 SDValue N1 = Op.getOperand(1);
1395 SDValue N2 = Op.getOperand(2);
1396 SDValue N3 = Op.getOperand(3);
1397 SDValue InFlag = Op.getOperand(4);
1398 assert(N1.getOpcode() == ISD::BasicBlock);
1399 assert(N2.getOpcode() == ISD::Constant);
1400 assert(N3.getOpcode() == ISD::Register);
1402 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1403 cast<ConstantSDNode>(N2)->getZExtValue()),
1405 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
1406 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
1408 Chain = SDValue(ResNode, 0);
1409 if (Op.getNode()->getNumValues() == 2) {
1410 InFlag = SDValue(ResNode, 1);
1411 ReplaceUses(SDValue(Op.getNode(), 1), InFlag);
1413 ReplaceUses(SDValue(Op.getNode(), 0), SDValue(Chain.getNode(), Chain.getResNo()));
1416 case ARMISD::CMOV: {
1417 EVT VT = Op.getValueType();
1418 SDValue N0 = Op.getOperand(0);
1419 SDValue N1 = Op.getOperand(1);
1420 SDValue N2 = Op.getOperand(2);
1421 SDValue N3 = Op.getOperand(3);
1422 SDValue InFlag = Op.getOperand(4);
1423 assert(N2.getOpcode() == ISD::Constant);
1424 assert(N3.getOpcode() == ISD::Register);
1426 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1427 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1428 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1429 // Pattern complexity = 18 cost = 1 size = 0
1433 if (Subtarget->isThumb()) {
1434 if (SelectT2ShifterOperandReg(Op, N1, CPTmp0, CPTmp1)) {
1435 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1436 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1439 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1440 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1441 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1442 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1444 llvm_unreachable("Unknown so_reg opcode!");
1448 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1449 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1450 cast<ConstantSDNode>(N2)->getZExtValue()),
1452 SDValue Ops[] = { N0, CPTmp0, SOShImm, Tmp2, N3, InFlag };
1453 return CurDAG->SelectNodeTo(Op.getNode(), Opc, MVT::i32,Ops, 6);
1456 if (SelectShifterOperandReg(Op, N1, CPTmp0, CPTmp1, CPTmp2)) {
1457 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1458 cast<ConstantSDNode>(N2)->getZExtValue()),
1460 SDValue Ops[] = { N0, CPTmp0, CPTmp1, CPTmp2, Tmp2, N3, InFlag };
1461 return CurDAG->SelectNodeTo(Op.getNode(),
1462 ARM::MOVCCs, MVT::i32, Ops, 7);
1466 // Pattern: (ARMcmov:i32 GPR:i32:$false,
1467 // (imm:i32)<<P:Predicate_so_imm>>:$true,
1469 // Emits: (MOVCCi:i32 GPR:i32:$false,
1470 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1471 // Pattern complexity = 10 cost = 1 size = 0
1472 if (N3.getOpcode() == ISD::Constant) {
1473 if (Subtarget->isThumb()) {
1474 if (Predicate_t2_so_imm(N3.getNode())) {
1475 SDValue Tmp1 = CurDAG->getTargetConstant(((unsigned)
1476 cast<ConstantSDNode>(N1)->getZExtValue()),
1478 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1479 cast<ConstantSDNode>(N2)->getZExtValue()),
1481 SDValue Ops[] = { N0, Tmp1, Tmp2, N3, InFlag };
1482 return CurDAG->SelectNodeTo(Op.getNode(),
1483 ARM::t2MOVCCi, MVT::i32, Ops, 5);
1486 if (Predicate_so_imm(N3.getNode())) {
1487 SDValue Tmp1 = CurDAG->getTargetConstant(((unsigned)
1488 cast<ConstantSDNode>(N1)->getZExtValue()),
1490 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1491 cast<ConstantSDNode>(N2)->getZExtValue()),
1493 SDValue Ops[] = { N0, Tmp1, Tmp2, N3, InFlag };
1494 return CurDAG->SelectNodeTo(Op.getNode(),
1495 ARM::MOVCCi, MVT::i32, Ops, 5);
1501 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1502 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1503 // Pattern complexity = 6 cost = 1 size = 0
1505 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1506 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1507 // Pattern complexity = 6 cost = 11 size = 0
1509 // Also FCPYScc and FCPYDcc.
1510 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1511 cast<ConstantSDNode>(N2)->getZExtValue()),
1513 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
1515 switch (VT.getSimpleVT().SimpleTy) {
1516 default: assert(false && "Illegal conditional move type!");
1519 Opc = Subtarget->isThumb()
1520 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1530 return CurDAG->SelectNodeTo(Op.getNode(), Opc, VT, Ops, 5);
1532 case ARMISD::CNEG: {
1533 EVT VT = Op.getValueType();
1534 SDValue N0 = Op.getOperand(0);
1535 SDValue N1 = Op.getOperand(1);
1536 SDValue N2 = Op.getOperand(2);
1537 SDValue N3 = Op.getOperand(3);
1538 SDValue InFlag = Op.getOperand(4);
1539 assert(N2.getOpcode() == ISD::Constant);
1540 assert(N3.getOpcode() == ISD::Register);
1542 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1543 cast<ConstantSDNode>(N2)->getZExtValue()),
1545 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
1547 switch (VT.getSimpleVT().SimpleTy) {
1548 default: assert(false && "Illegal conditional move type!");
1557 return CurDAG->SelectNodeTo(Op.getNode(), Opc, VT, Ops, 5);
1560 case ARMISD::VZIP: {
1562 EVT VT = N->getValueType(0);
1563 switch (VT.getSimpleVT().SimpleTy) {
1564 default: return NULL;
1565 case MVT::v8i8: Opc = ARM::VZIPd8; break;
1566 case MVT::v4i16: Opc = ARM::VZIPd16; break;
1568 case MVT::v2i32: Opc = ARM::VZIPd32; break;
1569 case MVT::v16i8: Opc = ARM::VZIPq8; break;
1570 case MVT::v8i16: Opc = ARM::VZIPq16; break;
1572 case MVT::v4i32: Opc = ARM::VZIPq32; break;
1574 return CurDAG->getMachineNode(Opc, dl, VT, VT,
1575 N->getOperand(0), N->getOperand(1));
1577 case ARMISD::VUZP: {
1579 EVT VT = N->getValueType(0);
1580 switch (VT.getSimpleVT().SimpleTy) {
1581 default: return NULL;
1582 case MVT::v8i8: Opc = ARM::VUZPd8; break;
1583 case MVT::v4i16: Opc = ARM::VUZPd16; break;
1585 case MVT::v2i32: Opc = ARM::VUZPd32; break;
1586 case MVT::v16i8: Opc = ARM::VUZPq8; break;
1587 case MVT::v8i16: Opc = ARM::VUZPq16; break;
1589 case MVT::v4i32: Opc = ARM::VUZPq32; break;
1591 return CurDAG->getMachineNode(Opc, dl, VT, VT,
1592 N->getOperand(0), N->getOperand(1));
1594 case ARMISD::VTRN: {
1596 EVT VT = N->getValueType(0);
1597 switch (VT.getSimpleVT().SimpleTy) {
1598 default: return NULL;
1599 case MVT::v8i8: Opc = ARM::VTRNd8; break;
1600 case MVT::v4i16: Opc = ARM::VTRNd16; break;
1602 case MVT::v2i32: Opc = ARM::VTRNd32; break;
1603 case MVT::v16i8: Opc = ARM::VTRNq8; break;
1604 case MVT::v8i16: Opc = ARM::VTRNq16; break;
1606 case MVT::v4i32: Opc = ARM::VTRNq32; break;
1608 return CurDAG->getMachineNode(Opc, dl, VT, VT,
1609 N->getOperand(0), N->getOperand(1));
1612 case ISD::INTRINSIC_VOID:
1613 case ISD::INTRINSIC_W_CHAIN: {
1614 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
1615 EVT VT = N->getValueType(0);
1622 case Intrinsic::arm_neon_vld2: {
1623 unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
1624 ARM::VLD2d32, ARM::VLD2d64 };
1625 unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
1626 return SelectVLD(Op, 2, DOpcodes, QOpcodes, 0);
1629 case Intrinsic::arm_neon_vld3: {
1630 unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
1631 ARM::VLD3d32, ARM::VLD3d64 };
1632 unsigned QOpcodes0[] = { ARM::VLD3q8a, ARM::VLD3q16a, ARM::VLD3q32a };
1633 unsigned QOpcodes1[] = { ARM::VLD3q8b, ARM::VLD3q16b, ARM::VLD3q32b };
1634 return SelectVLD(Op, 3, DOpcodes, QOpcodes0, QOpcodes1);
1637 case Intrinsic::arm_neon_vld4: {
1638 unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
1639 ARM::VLD4d32, ARM::VLD4d64 };
1640 unsigned QOpcodes0[] = { ARM::VLD4q8a, ARM::VLD4q16a, ARM::VLD4q32a };
1641 unsigned QOpcodes1[] = { ARM::VLD4q8b, ARM::VLD4q16b, ARM::VLD4q32b };
1642 return SelectVLD(Op, 4, DOpcodes, QOpcodes0, QOpcodes1);
1645 case Intrinsic::arm_neon_vld2lane: {
1646 unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
1647 unsigned QOpcodes0[] = { ARM::VLD2LNq16a, ARM::VLD2LNq32a };
1648 unsigned QOpcodes1[] = { ARM::VLD2LNq16b, ARM::VLD2LNq32b };
1649 return SelectVLDSTLane(Op, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
1652 case Intrinsic::arm_neon_vld3lane: {
1653 unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
1654 unsigned QOpcodes0[] = { ARM::VLD3LNq16a, ARM::VLD3LNq32a };
1655 unsigned QOpcodes1[] = { ARM::VLD3LNq16b, ARM::VLD3LNq32b };
1656 return SelectVLDSTLane(Op, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
1659 case Intrinsic::arm_neon_vld4lane: {
1660 unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
1661 unsigned QOpcodes0[] = { ARM::VLD4LNq16a, ARM::VLD4LNq32a };
1662 unsigned QOpcodes1[] = { ARM::VLD4LNq16b, ARM::VLD4LNq32b };
1663 return SelectVLDSTLane(Op, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
1666 case Intrinsic::arm_neon_vst2: {
1667 SDValue MemAddr, MemUpdate, MemOpc;
1668 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1670 SDValue Chain = N->getOperand(0);
1671 VT = N->getOperand(3).getValueType();
1672 if (VT.is64BitVector()) {
1673 switch (VT.getSimpleVT().SimpleTy) {
1674 default: llvm_unreachable("unhandled vst2 type");
1675 case MVT::v8i8: Opc = ARM::VST2d8; break;
1676 case MVT::v4i16: Opc = ARM::VST2d16; break;
1678 case MVT::v2i32: Opc = ARM::VST2d32; break;
1679 case MVT::v1i64: Opc = ARM::VST2d64; break;
1681 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
1682 N->getOperand(3), N->getOperand(4), Chain };
1683 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 6);
1685 // Quad registers are stored as pairs of double registers.
1687 switch (VT.getSimpleVT().SimpleTy) {
1688 default: llvm_unreachable("unhandled vst2 type");
1689 case MVT::v16i8: Opc = ARM::VST2q8; RegVT = MVT::v8i8; break;
1690 case MVT::v8i16: Opc = ARM::VST2q16; RegVT = MVT::v4i16; break;
1691 case MVT::v4f32: Opc = ARM::VST2q32; RegVT = MVT::v2f32; break;
1692 case MVT::v4i32: Opc = ARM::VST2q32; RegVT = MVT::v2i32; break;
1694 SDValue D0 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1696 SDValue D1 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1698 SDValue D2 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1700 SDValue D3 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1702 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
1703 D0, D1, D2, D3, Chain };
1704 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
1707 case Intrinsic::arm_neon_vst3: {
1708 SDValue MemAddr, MemUpdate, MemOpc;
1709 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1711 SDValue Chain = N->getOperand(0);
1712 VT = N->getOperand(3).getValueType();
1713 if (VT.is64BitVector()) {
1714 switch (VT.getSimpleVT().SimpleTy) {
1715 default: llvm_unreachable("unhandled vst3 type");
1716 case MVT::v8i8: Opc = ARM::VST3d8; break;
1717 case MVT::v4i16: Opc = ARM::VST3d16; break;
1719 case MVT::v2i32: Opc = ARM::VST3d32; break;
1720 case MVT::v1i64: Opc = ARM::VST3d64; break;
1722 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
1723 N->getOperand(3), N->getOperand(4),
1724 N->getOperand(5), Chain };
1725 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7);
1727 // Quad registers are stored with two separate instructions, where one
1728 // stores the even registers and the other stores the odd registers.
1731 switch (VT.getSimpleVT().SimpleTy) {
1732 default: llvm_unreachable("unhandled vst3 type");
1734 Opc = ARM::VST3q8a; Opc2 = ARM::VST3q8b; RegVT = MVT::v8i8; break;
1736 Opc = ARM::VST3q16a; Opc2 = ARM::VST3q16b; RegVT = MVT::v4i16; break;
1738 Opc = ARM::VST3q32a; Opc2 = ARM::VST3q32b; RegVT = MVT::v2f32; break;
1740 Opc = ARM::VST3q32a; Opc2 = ARM::VST3q32b; RegVT = MVT::v2i32; break;
1742 // Enable writeback to the address register.
1743 MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
1745 SDValue D0 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1747 SDValue D2 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1749 SDValue D4 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1751 const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, D0, D2, D4, Chain };
1752 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1753 MVT::Other, OpsA, 7);
1754 Chain = SDValue(VStA, 1);
1756 SDValue D1 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1758 SDValue D3 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1760 SDValue D5 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1762 MemAddr = SDValue(VStA, 0);
1763 const SDValue OpsB[] = { MemAddr, MemUpdate, MemOpc, D1, D3, D5, Chain };
1764 SDNode *VStB = CurDAG->getMachineNode(Opc2, dl, MemAddr.getValueType(),
1765 MVT::Other, OpsB, 7);
1766 Chain = SDValue(VStB, 1);
1767 ReplaceUses(SDValue(N, 0), Chain);
1771 case Intrinsic::arm_neon_vst4: {
1772 SDValue MemAddr, MemUpdate, MemOpc;
1773 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1775 SDValue Chain = N->getOperand(0);
1776 VT = N->getOperand(3).getValueType();
1777 if (VT.is64BitVector()) {
1778 switch (VT.getSimpleVT().SimpleTy) {
1779 default: llvm_unreachable("unhandled vst4 type");
1780 case MVT::v8i8: Opc = ARM::VST4d8; break;
1781 case MVT::v4i16: Opc = ARM::VST4d16; break;
1783 case MVT::v2i32: Opc = ARM::VST4d32; break;
1784 case MVT::v1i64: Opc = ARM::VST4d64; break;
1786 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
1787 N->getOperand(3), N->getOperand(4),
1788 N->getOperand(5), N->getOperand(6), Chain };
1789 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
1791 // Quad registers are stored with two separate instructions, where one
1792 // stores the even registers and the other stores the odd registers.
1795 switch (VT.getSimpleVT().SimpleTy) {
1796 default: llvm_unreachable("unhandled vst4 type");
1798 Opc = ARM::VST4q8a; Opc2 = ARM::VST4q8b; RegVT = MVT::v8i8; break;
1800 Opc = ARM::VST4q16a; Opc2 = ARM::VST4q16b; RegVT = MVT::v4i16; break;
1802 Opc = ARM::VST4q32a; Opc2 = ARM::VST4q32b; RegVT = MVT::v2f32; break;
1804 Opc = ARM::VST4q32a; Opc2 = ARM::VST4q32b; RegVT = MVT::v2i32; break;
1806 // Enable writeback to the address register.
1807 MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
1809 SDValue D0 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1811 SDValue D2 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1813 SDValue D4 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1815 SDValue D6 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1817 const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc,
1818 D0, D2, D4, D6, Chain };
1819 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1820 MVT::Other, OpsA, 8);
1821 Chain = SDValue(VStA, 1);
1823 SDValue D1 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1825 SDValue D3 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1827 SDValue D5 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1829 SDValue D7 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1831 MemAddr = SDValue(VStA, 0);
1832 const SDValue OpsB[] = { MemAddr, MemUpdate, MemOpc,
1833 D1, D3, D5, D7, Chain };
1834 SDNode *VStB = CurDAG->getMachineNode(Opc2, dl, MemAddr.getValueType(),
1835 MVT::Other, OpsB, 8);
1836 Chain = SDValue(VStB, 1);
1837 ReplaceUses(SDValue(N, 0), Chain);
1841 case Intrinsic::arm_neon_vst2lane: {
1842 unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
1843 unsigned QOpcodes0[] = { ARM::VST2LNq16a, ARM::VST2LNq32a };
1844 unsigned QOpcodes1[] = { ARM::VST2LNq16b, ARM::VST2LNq32b };
1845 return SelectVLDSTLane(Op, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
1848 case Intrinsic::arm_neon_vst3lane: {
1849 unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
1850 unsigned QOpcodes0[] = { ARM::VST3LNq16a, ARM::VST3LNq32a };
1851 unsigned QOpcodes1[] = { ARM::VST3LNq16b, ARM::VST3LNq32b };
1852 return SelectVLDSTLane(Op, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
1855 case Intrinsic::arm_neon_vst4lane: {
1856 unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
1857 unsigned QOpcodes0[] = { ARM::VST4LNq16a, ARM::VST4LNq32a };
1858 unsigned QOpcodes1[] = { ARM::VST4LNq16b, ARM::VST4LNq32b };
1859 return SelectVLDSTLane(Op, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
1865 return SelectCode(Op);
1868 bool ARMDAGToDAGISel::
1869 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
1870 std::vector<SDValue> &OutOps) {
1871 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
1872 // Require the address to be in a register. That is safe for all ARM
1873 // variants and it is hard to do anything much smarter without knowing
1874 // how the operand is used.
1875 OutOps.push_back(Op);
1879 /// createARMISelDag - This pass converts a legalized DAG into a
1880 /// ARM-specific DAG, ready for instruction scheduling.
1882 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
1883 CodeGenOpt::Level OptLevel) {
1884 return new ARMDAGToDAGISel(TM, OptLevel);