1 //===-- ARMISelDAGToDAG.cpp - A dag to dag inst selector for ARM ----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines an instruction selector for the ARM target.
12 //===----------------------------------------------------------------------===//
15 #include "ARMAddressingModes.h"
16 #include "ARMTargetMachine.h"
17 #include "llvm/CallingConv.h"
18 #include "llvm/Constants.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Function.h"
21 #include "llvm/Intrinsics.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/SelectionDAG.h"
27 #include "llvm/CodeGen/SelectionDAGISel.h"
28 #include "llvm/Target/TargetLowering.h"
29 #include "llvm/Target/TargetOptions.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Compiler.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
39 UseRegSeq("neon-reg-sequence", cl::Hidden,
40 cl::desc("Use reg_sequence to model ld / st of multiple neon regs"));
42 //===--------------------------------------------------------------------===//
43 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
44 /// instructions for SelectionDAG operations.
47 class ARMDAGToDAGISel : public SelectionDAGISel {
48 ARMBaseTargetMachine &TM;
50 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
51 /// make the right decision when generating code for different targets.
52 const ARMSubtarget *Subtarget;
55 explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
56 CodeGenOpt::Level OptLevel)
57 : SelectionDAGISel(tm, OptLevel), TM(tm),
58 Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
61 virtual const char *getPassName() const {
62 return "ARM Instruction Selection";
65 /// getI32Imm - Return a target constant of type i32 with the specified
67 inline SDValue getI32Imm(unsigned Imm) {
68 return CurDAG->getTargetConstant(Imm, MVT::i32);
71 SDNode *Select(SDNode *N);
73 bool SelectShifterOperandReg(SDNode *Op, SDValue N, SDValue &A,
74 SDValue &B, SDValue &C);
75 bool SelectAddrMode2(SDNode *Op, SDValue N, SDValue &Base,
76 SDValue &Offset, SDValue &Opc);
77 bool SelectAddrMode2Offset(SDNode *Op, SDValue N,
78 SDValue &Offset, SDValue &Opc);
79 bool SelectAddrMode3(SDNode *Op, SDValue N, SDValue &Base,
80 SDValue &Offset, SDValue &Opc);
81 bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
82 SDValue &Offset, SDValue &Opc);
83 bool SelectAddrMode4(SDNode *Op, SDValue N, SDValue &Addr,
85 bool SelectAddrMode5(SDNode *Op, SDValue N, SDValue &Base,
87 bool SelectAddrMode6(SDNode *Op, SDValue N, SDValue &Addr, SDValue &Align);
89 bool SelectAddrModePC(SDNode *Op, SDValue N, SDValue &Offset,
92 bool SelectThumbAddrModeRR(SDNode *Op, SDValue N, SDValue &Base,
94 bool SelectThumbAddrModeRI5(SDNode *Op, SDValue N, unsigned Scale,
95 SDValue &Base, SDValue &OffImm,
97 bool SelectThumbAddrModeS1(SDNode *Op, SDValue N, SDValue &Base,
98 SDValue &OffImm, SDValue &Offset);
99 bool SelectThumbAddrModeS2(SDNode *Op, SDValue N, SDValue &Base,
100 SDValue &OffImm, SDValue &Offset);
101 bool SelectThumbAddrModeS4(SDNode *Op, SDValue N, SDValue &Base,
102 SDValue &OffImm, SDValue &Offset);
103 bool SelectThumbAddrModeSP(SDNode *Op, SDValue N, SDValue &Base,
106 bool SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
107 SDValue &BaseReg, SDValue &Opc);
108 bool SelectT2AddrModeImm12(SDNode *Op, SDValue N, SDValue &Base,
110 bool SelectT2AddrModeImm8(SDNode *Op, SDValue N, SDValue &Base,
112 bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
114 bool SelectT2AddrModeImm8s4(SDNode *Op, SDValue N, SDValue &Base,
116 bool SelectT2AddrModeSoReg(SDNode *Op, SDValue N, SDValue &Base,
117 SDValue &OffReg, SDValue &ShImm);
119 // Include the pieces autogenerated from the target description.
120 #include "ARMGenDAGISel.inc"
123 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
125 SDNode *SelectARMIndexedLoad(SDNode *N);
126 SDNode *SelectT2IndexedLoad(SDNode *N);
128 /// SelectVLD - Select NEON load intrinsics. NumVecs should be
129 /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for
130 /// loads of D registers and even subregs and odd subregs of Q registers.
131 /// For NumVecs <= 2, QOpcodes1 is not used.
132 SDNode *SelectVLD(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
133 unsigned *QOpcodes0, unsigned *QOpcodes1);
135 /// SelectVST - Select NEON store intrinsics. NumVecs should
136 /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for
137 /// stores of D registers and even subregs and odd subregs of Q registers.
138 /// For NumVecs <= 2, QOpcodes1 is not used.
139 SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
140 unsigned *QOpcodes0, unsigned *QOpcodes1);
142 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
143 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
144 /// load/store of D registers and even subregs and odd subregs of Q registers.
145 SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, unsigned NumVecs,
146 unsigned *DOpcodes, unsigned *QOpcodes0,
147 unsigned *QOpcodes1);
149 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
150 SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
152 /// SelectCMOVOp - Select CMOV instructions for ARM.
153 SDNode *SelectCMOVOp(SDNode *N);
154 SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
155 ARMCC::CondCodes CCVal, SDValue CCR,
157 SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
158 ARMCC::CondCodes CCVal, SDValue CCR,
160 SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
161 ARMCC::CondCodes CCVal, SDValue CCR,
163 SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
164 ARMCC::CondCodes CCVal, SDValue CCR,
167 SDNode *SelectConcatVector(SDNode *N);
169 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
170 /// inline asm expressions.
171 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
173 std::vector<SDValue> &OutOps);
175 /// PairDRegs - Form a quad register from a pair of D registers.
177 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
179 /// PairDRegs - Form a quad register pair from a pair of Q registers.
181 SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1);
183 /// QuadDRegs - Form a quad register pair from a quad of D registers.
185 SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
187 /// QuadQRegs - Form 4 consecutive Q registers.
189 SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
191 /// OctoDRegs - Form 8 consecutive D registers.
193 SDNode *OctoDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3,
194 SDValue V4, SDValue V5, SDValue V6, SDValue V7);
198 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
199 /// operand. If so Imm will receive the 32-bit value.
200 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
201 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
202 Imm = cast<ConstantSDNode>(N)->getZExtValue();
208 // isInt32Immediate - This method tests to see if a constant operand.
209 // If so Imm will receive the 32 bit value.
210 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
211 return isInt32Immediate(N.getNode(), Imm);
214 // isOpcWithIntImmediate - This method tests to see if the node is a specific
215 // opcode and that it has a immediate integer right operand.
216 // If so Imm will receive the 32 bit value.
217 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
218 return N->getOpcode() == Opc &&
219 isInt32Immediate(N->getOperand(1).getNode(), Imm);
223 bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
228 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
230 // Don't match base register only case. That is matched to a separate
231 // lower complexity pattern with explicit register operand.
232 if (ShOpcVal == ARM_AM::no_shift) return false;
234 BaseReg = N.getOperand(0);
235 unsigned ShImmVal = 0;
236 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
237 ShReg = CurDAG->getRegister(0, MVT::i32);
238 ShImmVal = RHS->getZExtValue() & 31;
240 ShReg = N.getOperand(1);
242 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
247 bool ARMDAGToDAGISel::SelectAddrMode2(SDNode *Op, SDValue N,
248 SDValue &Base, SDValue &Offset,
250 if (N.getOpcode() == ISD::MUL) {
251 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
252 // X * [3,5,9] -> X + X * [2,4,8] etc.
253 int RHSC = (int)RHS->getZExtValue();
256 ARM_AM::AddrOpc AddSub = ARM_AM::add;
258 AddSub = ARM_AM::sub;
261 if (isPowerOf2_32(RHSC)) {
262 unsigned ShAmt = Log2_32(RHSC);
263 Base = Offset = N.getOperand(0);
264 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
273 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
275 if (N.getOpcode() == ISD::FrameIndex) {
276 int FI = cast<FrameIndexSDNode>(N)->getIndex();
277 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
278 } else if (N.getOpcode() == ARMISD::Wrapper &&
279 !(Subtarget->useMovt() &&
280 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
281 Base = N.getOperand(0);
283 Offset = CurDAG->getRegister(0, MVT::i32);
284 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
290 // Match simple R +/- imm12 operands.
291 if (N.getOpcode() == ISD::ADD)
292 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
293 int RHSC = (int)RHS->getZExtValue();
294 if ((RHSC >= 0 && RHSC < 0x1000) ||
295 (RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
296 Base = N.getOperand(0);
297 if (Base.getOpcode() == ISD::FrameIndex) {
298 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
299 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
301 Offset = CurDAG->getRegister(0, MVT::i32);
303 ARM_AM::AddrOpc AddSub = ARM_AM::add;
305 AddSub = ARM_AM::sub;
308 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
315 // Otherwise this is R +/- [possibly shifted] R.
316 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
317 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
320 Base = N.getOperand(0);
321 Offset = N.getOperand(1);
323 if (ShOpcVal != ARM_AM::no_shift) {
324 // Check to see if the RHS of the shift is a constant, if not, we can't fold
326 if (ConstantSDNode *Sh =
327 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
328 ShAmt = Sh->getZExtValue();
329 Offset = N.getOperand(1).getOperand(0);
331 ShOpcVal = ARM_AM::no_shift;
335 // Try matching (R shl C) + (R).
336 if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
337 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
338 if (ShOpcVal != ARM_AM::no_shift) {
339 // Check to see if the RHS of the shift is a constant, if not, we can't
341 if (ConstantSDNode *Sh =
342 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
343 ShAmt = Sh->getZExtValue();
344 Offset = N.getOperand(0).getOperand(0);
345 Base = N.getOperand(1);
347 ShOpcVal = ARM_AM::no_shift;
352 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
357 bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
358 SDValue &Offset, SDValue &Opc) {
359 unsigned Opcode = Op->getOpcode();
360 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
361 ? cast<LoadSDNode>(Op)->getAddressingMode()
362 : cast<StoreSDNode>(Op)->getAddressingMode();
363 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
364 ? ARM_AM::add : ARM_AM::sub;
365 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
366 int Val = (int)C->getZExtValue();
367 if (Val >= 0 && Val < 0x1000) { // 12 bits.
368 Offset = CurDAG->getRegister(0, MVT::i32);
369 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
377 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
379 if (ShOpcVal != ARM_AM::no_shift) {
380 // Check to see if the RHS of the shift is a constant, if not, we can't fold
382 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
383 ShAmt = Sh->getZExtValue();
384 Offset = N.getOperand(0);
386 ShOpcVal = ARM_AM::no_shift;
390 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
396 bool ARMDAGToDAGISel::SelectAddrMode3(SDNode *Op, SDValue N,
397 SDValue &Base, SDValue &Offset,
399 if (N.getOpcode() == ISD::SUB) {
400 // X - C is canonicalize to X + -C, no need to handle it here.
401 Base = N.getOperand(0);
402 Offset = N.getOperand(1);
403 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
407 if (N.getOpcode() != ISD::ADD) {
409 if (N.getOpcode() == ISD::FrameIndex) {
410 int FI = cast<FrameIndexSDNode>(N)->getIndex();
411 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
413 Offset = CurDAG->getRegister(0, MVT::i32);
414 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
418 // If the RHS is +/- imm8, fold into addr mode.
419 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
420 int RHSC = (int)RHS->getZExtValue();
421 if ((RHSC >= 0 && RHSC < 256) ||
422 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
423 Base = N.getOperand(0);
424 if (Base.getOpcode() == ISD::FrameIndex) {
425 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
426 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
428 Offset = CurDAG->getRegister(0, MVT::i32);
430 ARM_AM::AddrOpc AddSub = ARM_AM::add;
432 AddSub = ARM_AM::sub;
435 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
440 Base = N.getOperand(0);
441 Offset = N.getOperand(1);
442 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
446 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
447 SDValue &Offset, SDValue &Opc) {
448 unsigned Opcode = Op->getOpcode();
449 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
450 ? cast<LoadSDNode>(Op)->getAddressingMode()
451 : cast<StoreSDNode>(Op)->getAddressingMode();
452 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
453 ? ARM_AM::add : ARM_AM::sub;
454 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
455 int Val = (int)C->getZExtValue();
456 if (Val >= 0 && Val < 256) {
457 Offset = CurDAG->getRegister(0, MVT::i32);
458 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
464 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
468 bool ARMDAGToDAGISel::SelectAddrMode4(SDNode *Op, SDValue N,
469 SDValue &Addr, SDValue &Mode) {
471 Mode = CurDAG->getTargetConstant(0, MVT::i32);
475 bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N,
476 SDValue &Base, SDValue &Offset) {
477 if (N.getOpcode() != ISD::ADD) {
479 if (N.getOpcode() == ISD::FrameIndex) {
480 int FI = cast<FrameIndexSDNode>(N)->getIndex();
481 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
482 } else if (N.getOpcode() == ARMISD::Wrapper &&
483 !(Subtarget->useMovt() &&
484 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
485 Base = N.getOperand(0);
487 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
492 // If the RHS is +/- imm8, fold into addr mode.
493 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
494 int RHSC = (int)RHS->getZExtValue();
495 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied by 4.
497 if ((RHSC >= 0 && RHSC < 256) ||
498 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
499 Base = N.getOperand(0);
500 if (Base.getOpcode() == ISD::FrameIndex) {
501 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
502 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
505 ARM_AM::AddrOpc AddSub = ARM_AM::add;
507 AddSub = ARM_AM::sub;
510 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
518 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
523 bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N,
524 SDValue &Addr, SDValue &Align) {
526 // Default to no alignment.
527 Align = CurDAG->getTargetConstant(0, MVT::i32);
531 bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
532 SDValue &Offset, SDValue &Label) {
533 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
534 Offset = N.getOperand(0);
535 SDValue N1 = N.getOperand(1);
536 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
543 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
544 SDValue &Base, SDValue &Offset){
545 // FIXME dl should come from the parent load or store, not the address
546 DebugLoc dl = Op->getDebugLoc();
547 if (N.getOpcode() != ISD::ADD) {
548 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
549 if (!NC || NC->getZExtValue() != 0)
556 Base = N.getOperand(0);
557 Offset = N.getOperand(1);
562 ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
563 unsigned Scale, SDValue &Base,
564 SDValue &OffImm, SDValue &Offset) {
566 SDValue TmpBase, TmpOffImm;
567 if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
568 return false; // We want to select tLDRspi / tSTRspi instead.
569 if (N.getOpcode() == ARMISD::Wrapper &&
570 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
571 return false; // We want to select tLDRpci instead.
574 if (N.getOpcode() != ISD::ADD) {
575 if (N.getOpcode() == ARMISD::Wrapper &&
576 !(Subtarget->useMovt() &&
577 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
578 Base = N.getOperand(0);
582 Offset = CurDAG->getRegister(0, MVT::i32);
583 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
587 // Thumb does not have [sp, r] address mode.
588 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
589 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
590 if ((LHSR && LHSR->getReg() == ARM::SP) ||
591 (RHSR && RHSR->getReg() == ARM::SP)) {
593 Offset = CurDAG->getRegister(0, MVT::i32);
594 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
598 // If the RHS is + imm5 * scale, fold into addr mode.
599 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
600 int RHSC = (int)RHS->getZExtValue();
601 if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
603 if (RHSC >= 0 && RHSC < 32) {
604 Base = N.getOperand(0);
605 Offset = CurDAG->getRegister(0, MVT::i32);
606 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
612 Base = N.getOperand(0);
613 Offset = N.getOperand(1);
614 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
618 bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
619 SDValue &Base, SDValue &OffImm,
621 return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
624 bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
625 SDValue &Base, SDValue &OffImm,
627 return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
630 bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
631 SDValue &Base, SDValue &OffImm,
633 return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
636 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
637 SDValue &Base, SDValue &OffImm) {
638 if (N.getOpcode() == ISD::FrameIndex) {
639 int FI = cast<FrameIndexSDNode>(N)->getIndex();
640 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
641 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
645 if (N.getOpcode() != ISD::ADD)
648 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
649 if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
650 (LHSR && LHSR->getReg() == ARM::SP)) {
651 // If the RHS is + imm8 * scale, fold into addr mode.
652 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
653 int RHSC = (int)RHS->getZExtValue();
654 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
656 if (RHSC >= 0 && RHSC < 256) {
657 Base = N.getOperand(0);
658 if (Base.getOpcode() == ISD::FrameIndex) {
659 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
660 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
662 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
672 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
675 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
677 // Don't match base register only case. That is matched to a separate
678 // lower complexity pattern with explicit register operand.
679 if (ShOpcVal == ARM_AM::no_shift) return false;
681 BaseReg = N.getOperand(0);
682 unsigned ShImmVal = 0;
683 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
684 ShImmVal = RHS->getZExtValue() & 31;
685 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
692 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
693 SDValue &Base, SDValue &OffImm) {
694 // Match simple R + imm12 operands.
697 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
698 if (N.getOpcode() == ISD::FrameIndex) {
699 // Match frame index...
700 int FI = cast<FrameIndexSDNode>(N)->getIndex();
701 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
702 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
704 } else if (N.getOpcode() == ARMISD::Wrapper &&
705 !(Subtarget->useMovt() &&
706 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
707 Base = N.getOperand(0);
708 if (Base.getOpcode() == ISD::TargetConstantPool)
709 return false; // We want to select t2LDRpci instead.
712 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
716 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
717 if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
718 // Let t2LDRi8 handle (R - imm8).
721 int RHSC = (int)RHS->getZExtValue();
722 if (N.getOpcode() == ISD::SUB)
725 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
726 Base = N.getOperand(0);
727 if (Base.getOpcode() == ISD::FrameIndex) {
728 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
729 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
731 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
738 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
742 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
743 SDValue &Base, SDValue &OffImm) {
744 // Match simple R - imm8 operands.
745 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
746 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
747 int RHSC = (int)RHS->getSExtValue();
748 if (N.getOpcode() == ISD::SUB)
751 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
752 Base = N.getOperand(0);
753 if (Base.getOpcode() == ISD::FrameIndex) {
754 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
755 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
757 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
766 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
768 unsigned Opcode = Op->getOpcode();
769 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
770 ? cast<LoadSDNode>(Op)->getAddressingMode()
771 : cast<StoreSDNode>(Op)->getAddressingMode();
772 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
773 int RHSC = (int)RHS->getZExtValue();
774 if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
775 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
776 ? CurDAG->getTargetConstant(RHSC, MVT::i32)
777 : CurDAG->getTargetConstant(-RHSC, MVT::i32);
785 bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
786 SDValue &Base, SDValue &OffImm) {
787 if (N.getOpcode() == ISD::ADD) {
788 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
789 int RHSC = (int)RHS->getZExtValue();
790 if (((RHSC & 0x3) == 0) &&
791 ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits.
792 Base = N.getOperand(0);
793 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
797 } else if (N.getOpcode() == ISD::SUB) {
798 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
799 int RHSC = (int)RHS->getZExtValue();
800 if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) { // 8 bits.
801 Base = N.getOperand(0);
802 OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
811 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDNode *Op, SDValue N,
813 SDValue &OffReg, SDValue &ShImm) {
814 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
815 if (N.getOpcode() != ISD::ADD)
818 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
819 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
820 int RHSC = (int)RHS->getZExtValue();
821 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
823 else if (RHSC < 0 && RHSC >= -255) // 8 bits
827 // Look for (R + R) or (R + (R << [1,2,3])).
829 Base = N.getOperand(0);
830 OffReg = N.getOperand(1);
832 // Swap if it is ((R << c) + R).
833 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
834 if (ShOpcVal != ARM_AM::lsl) {
835 ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
836 if (ShOpcVal == ARM_AM::lsl)
837 std::swap(Base, OffReg);
840 if (ShOpcVal == ARM_AM::lsl) {
841 // Check to see if the RHS of the shift is a constant, if not, we can't fold
843 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
844 ShAmt = Sh->getZExtValue();
847 ShOpcVal = ARM_AM::no_shift;
849 OffReg = OffReg.getOperand(0);
851 ShOpcVal = ARM_AM::no_shift;
855 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
860 //===--------------------------------------------------------------------===//
862 /// getAL - Returns a ARMCC::AL immediate node.
863 static inline SDValue getAL(SelectionDAG *CurDAG) {
864 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
867 SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
868 LoadSDNode *LD = cast<LoadSDNode>(N);
869 ISD::MemIndexedMode AM = LD->getAddressingMode();
870 if (AM == ISD::UNINDEXED)
873 EVT LoadedVT = LD->getMemoryVT();
874 SDValue Offset, AMOpc;
875 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
878 if (LoadedVT == MVT::i32 &&
879 SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
880 Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
882 } else if (LoadedVT == MVT::i16 &&
883 SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
885 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
886 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
887 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
888 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
889 if (LD->getExtensionType() == ISD::SEXTLOAD) {
890 if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
892 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
895 if (SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
897 Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
903 SDValue Chain = LD->getChain();
904 SDValue Base = LD->getBasePtr();
905 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
906 CurDAG->getRegister(0, MVT::i32), Chain };
907 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
914 SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
915 LoadSDNode *LD = cast<LoadSDNode>(N);
916 ISD::MemIndexedMode AM = LD->getAddressingMode();
917 if (AM == ISD::UNINDEXED)
920 EVT LoadedVT = LD->getMemoryVT();
921 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
923 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
926 if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
927 switch (LoadedVT.getSimpleVT().SimpleTy) {
929 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
933 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
935 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
940 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
942 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
951 SDValue Chain = LD->getChain();
952 SDValue Base = LD->getBasePtr();
953 SDValue Ops[]= { Base, Offset, getAL(CurDAG),
954 CurDAG->getRegister(0, MVT::i32), Chain };
955 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
962 /// PairDRegs - Form a quad register from a pair of D registers.
964 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
965 DebugLoc dl = V0.getNode()->getDebugLoc();
966 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
967 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
968 if (llvm::ModelWithRegSequence()) {
969 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
970 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
973 SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0);
974 SDNode *Pair = CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
975 VT, Undef, V0, SubReg0);
976 return CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
977 VT, SDValue(Pair, 0), V1, SubReg1);
980 /// PairQRegs - Form 4 consecutive D registers from a pair of Q registers.
982 SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) {
983 DebugLoc dl = V0.getNode()->getDebugLoc();
984 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::QSUBREG_0, MVT::i32);
985 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::QSUBREG_1, MVT::i32);
986 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
987 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
990 /// QuadDRegs - Form 4 consecutive D registers.
992 SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
993 SDValue V2, SDValue V3) {
994 DebugLoc dl = V0.getNode()->getDebugLoc();
995 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
996 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
997 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::DSUBREG_2, MVT::i32);
998 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::DSUBREG_3, MVT::i32);
999 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1000 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1003 /// QuadQRegs - Form 4 consecutive Q registers.
1005 SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1,
1006 SDValue V2, SDValue V3) {
1007 DebugLoc dl = V0.getNode()->getDebugLoc();
1008 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::QSUBREG_0, MVT::i32);
1009 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::QSUBREG_1, MVT::i32);
1010 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::QSUBREG_2, MVT::i32);
1011 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::QSUBREG_3, MVT::i32);
1012 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1013 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1016 /// OctoDRegs - Form 8 consecutive D registers.
1018 SDNode *ARMDAGToDAGISel::OctoDRegs(EVT VT, SDValue V0, SDValue V1,
1019 SDValue V2, SDValue V3,
1020 SDValue V4, SDValue V5,
1021 SDValue V6, SDValue V7) {
1022 DebugLoc dl = V0.getNode()->getDebugLoc();
1023 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
1024 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
1025 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::DSUBREG_2, MVT::i32);
1026 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::DSUBREG_3, MVT::i32);
1027 SDValue SubReg4 = CurDAG->getTargetConstant(ARM::DSUBREG_4, MVT::i32);
1028 SDValue SubReg5 = CurDAG->getTargetConstant(ARM::DSUBREG_5, MVT::i32);
1029 SDValue SubReg6 = CurDAG->getTargetConstant(ARM::DSUBREG_6, MVT::i32);
1030 SDValue SubReg7 = CurDAG->getTargetConstant(ARM::DSUBREG_7, MVT::i32);
1031 const SDValue Ops[] ={ V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3,
1032 V4, SubReg4, V5, SubReg5, V6, SubReg6, V7, SubReg7 };
1033 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 16);
1036 /// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
1037 /// for a 64-bit subregister of the vector.
1038 static EVT GetNEONSubregVT(EVT VT) {
1039 switch (VT.getSimpleVT().SimpleTy) {
1040 default: llvm_unreachable("unhandled NEON type");
1041 case MVT::v16i8: return MVT::v8i8;
1042 case MVT::v8i16: return MVT::v4i16;
1043 case MVT::v4f32: return MVT::v2f32;
1044 case MVT::v4i32: return MVT::v2i32;
1045 case MVT::v2i64: return MVT::v1i64;
1049 SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
1050 unsigned *DOpcodes, unsigned *QOpcodes0,
1051 unsigned *QOpcodes1) {
1052 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1053 DebugLoc dl = N->getDebugLoc();
1055 SDValue MemAddr, Align;
1056 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1059 SDValue Chain = N->getOperand(0);
1060 EVT VT = N->getValueType(0);
1061 bool is64BitVector = VT.is64BitVector();
1063 unsigned OpcodeIndex;
1064 switch (VT.getSimpleVT().SimpleTy) {
1065 default: llvm_unreachable("unhandled vld type");
1066 // Double-register operations:
1067 case MVT::v8i8: OpcodeIndex = 0; break;
1068 case MVT::v4i16: OpcodeIndex = 1; break;
1070 case MVT::v2i32: OpcodeIndex = 2; break;
1071 case MVT::v1i64: OpcodeIndex = 3; break;
1072 // Quad-register operations:
1073 case MVT::v16i8: OpcodeIndex = 0; break;
1074 case MVT::v8i16: OpcodeIndex = 1; break;
1076 case MVT::v4i32: OpcodeIndex = 2; break;
1077 case MVT::v2i64: OpcodeIndex = 3;
1078 assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
1082 SDValue Pred = getAL(CurDAG);
1083 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1084 if (is64BitVector) {
1085 unsigned Opc = DOpcodes[OpcodeIndex];
1086 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1087 std::vector<EVT> ResTys(NumVecs, VT);
1088 ResTys.push_back(MVT::Other);
1089 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1090 if (!llvm::ModelWithRegSequence() || NumVecs < 2)
1094 SDValue V0 = SDValue(VLd, 0);
1095 SDValue V1 = SDValue(VLd, 1);
1097 // Form a REG_SEQUENCE to force register allocation.
1099 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1101 SDValue V2 = SDValue(VLd, 2);
1102 // If it's a vld3, form a quad D-register but discard the last part.
1103 SDValue V3 = (NumVecs == 3)
1104 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1106 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1109 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1110 SDValue D = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0+Vec,
1112 ReplaceUses(SDValue(N, Vec), D);
1114 ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, NumVecs));
1118 EVT RegVT = GetNEONSubregVT(VT);
1120 // Quad registers are directly supported for VLD1 and VLD2,
1121 // loading pairs of D regs.
1122 unsigned Opc = QOpcodes0[OpcodeIndex];
1123 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1124 std::vector<EVT> ResTys(2 * NumVecs, RegVT);
1125 ResTys.push_back(MVT::Other);
1126 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1127 Chain = SDValue(VLd, 2 * NumVecs);
1129 // Combine the even and odd subregs to produce the result.
1130 if (llvm::ModelWithRegSequence()) {
1132 SDNode *Q = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
1133 ReplaceUses(SDValue(N, 0), SDValue(Q, 0));
1135 SDValue QQ = SDValue(QuadDRegs(MVT::v4i64,
1136 SDValue(VLd, 0), SDValue(VLd, 1),
1137 SDValue(VLd, 2), SDValue(VLd, 3)), 0);
1138 SDValue Q0 = CurDAG->getTargetExtractSubreg(ARM::QSUBREG_0, dl, VT, QQ);
1139 SDValue Q1 = CurDAG->getTargetExtractSubreg(ARM::QSUBREG_1, dl, VT, QQ);
1140 ReplaceUses(SDValue(N, 0), Q0);
1141 ReplaceUses(SDValue(N, 1), Q1);
1144 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1145 SDNode *Q = PairDRegs(VT, SDValue(VLd, 2*Vec), SDValue(VLd, 2*Vec+1));
1146 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1150 // Otherwise, quad registers are loaded with two separate instructions,
1151 // where one loads the even registers and the other loads the odd registers.
1153 std::vector<EVT> ResTys(NumVecs, RegVT);
1154 ResTys.push_back(MemAddr.getValueType());
1155 ResTys.push_back(MVT::Other);
1157 // Load the even subregs.
1158 unsigned Opc = QOpcodes0[OpcodeIndex];
1159 const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain };
1160 SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6);
1161 Chain = SDValue(VLdA, NumVecs+1);
1163 // Load the odd subregs.
1164 Opc = QOpcodes1[OpcodeIndex];
1165 const SDValue OpsB[] = { SDValue(VLdA, NumVecs),
1166 Align, Reg0, Pred, Reg0, Chain };
1167 SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
1168 Chain = SDValue(VLdB, NumVecs+1);
1170 if (llvm::ModelWithRegSequence()) {
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),
1181 SDValue V7 = (NumVecs == 3)
1182 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT),
1185 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V0, V1, V2, V3,
1186 V4, V5, V6, V7), 0);
1188 // Extract out the 3 / 4 Q registers.
1189 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1190 SDValue Q = CurDAG->getTargetExtractSubreg(ARM::QSUBREG_0+Vec,
1192 ReplaceUses(SDValue(N, Vec), Q);
1195 // Combine the even and odd subregs to produce the result.
1196 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1197 SDNode *Q = PairDRegs(VT, SDValue(VLdA, Vec), SDValue(VLdB, Vec));
1198 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1202 ReplaceUses(SDValue(N, NumVecs), Chain);
1206 SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
1207 unsigned *DOpcodes, unsigned *QOpcodes0,
1208 unsigned *QOpcodes1) {
1209 assert(NumVecs >=1 && NumVecs <= 4 && "VST NumVecs out-of-range");
1210 DebugLoc dl = N->getDebugLoc();
1212 SDValue MemAddr, Align;
1213 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1216 SDValue Chain = N->getOperand(0);
1217 EVT VT = N->getOperand(3).getValueType();
1218 bool is64BitVector = VT.is64BitVector();
1220 unsigned OpcodeIndex;
1221 switch (VT.getSimpleVT().SimpleTy) {
1222 default: llvm_unreachable("unhandled vst type");
1223 // Double-register operations:
1224 case MVT::v8i8: OpcodeIndex = 0; break;
1225 case MVT::v4i16: OpcodeIndex = 1; break;
1227 case MVT::v2i32: OpcodeIndex = 2; break;
1228 case MVT::v1i64: OpcodeIndex = 3; break;
1229 // Quad-register operations:
1230 case MVT::v16i8: OpcodeIndex = 0; break;
1231 case MVT::v8i16: OpcodeIndex = 1; break;
1233 case MVT::v4i32: OpcodeIndex = 2; break;
1234 case MVT::v2i64: OpcodeIndex = 3;
1235 assert(NumVecs == 1 && "v2i64 type only supported for VST1");
1239 SDValue Pred = getAL(CurDAG);
1240 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1242 SmallVector<SDValue, 10> Ops;
1243 Ops.push_back(MemAddr);
1244 Ops.push_back(Align);
1246 if (is64BitVector) {
1247 if (llvm::ModelWithRegSequence() && NumVecs >= 2) {
1249 SDValue V0 = N->getOperand(0+3);
1250 SDValue V1 = N->getOperand(1+3);
1252 // Form a REG_SEQUENCE to force register allocation.
1254 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1256 SDValue V2 = N->getOperand(2+3);
1257 // If it's a vld3, form a quad D-register and leave the last part as
1259 SDValue V3 = (NumVecs == 3)
1260 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1261 : N->getOperand(3+3);
1262 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1265 // Now extract the D registers back out.
1266 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, VT,
1268 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, VT,
1271 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_2, dl, VT,
1274 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_3, dl, VT,
1277 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1278 Ops.push_back(N->getOperand(Vec+3));
1280 Ops.push_back(Pred);
1281 Ops.push_back(Reg0); // predicate register
1282 Ops.push_back(Chain);
1283 unsigned Opc = DOpcodes[OpcodeIndex];
1284 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+5);
1287 EVT RegVT = GetNEONSubregVT(VT);
1289 // Quad registers are directly supported for VST1 and VST2,
1290 // storing pairs of D regs.
1291 unsigned Opc = QOpcodes0[OpcodeIndex];
1292 if (llvm::ModelWithRegSequence() && NumVecs == 2) {
1293 // First extract the pair of Q registers.
1294 SDValue Q0 = N->getOperand(3);
1295 SDValue Q1 = N->getOperand(4);
1297 // Form a QQ register.
1298 SDValue QQ = SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0);
1300 // Now extract the D registers back out.
1301 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1303 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1305 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_2, dl, RegVT,
1307 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_3, dl, RegVT,
1309 Ops.push_back(Pred);
1310 Ops.push_back(Reg0); // predicate register
1311 Ops.push_back(Chain);
1312 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 5 + 4);
1314 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1315 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1316 N->getOperand(Vec+3)));
1317 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1318 N->getOperand(Vec+3)));
1320 Ops.push_back(Pred);
1321 Ops.push_back(Reg0); // predicate register
1322 Ops.push_back(Chain);
1323 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
1328 // Otherwise, quad registers are stored with two separate instructions,
1329 // where one stores the even registers and the other stores the odd registers.
1330 if (llvm::ModelWithRegSequence()) {
1331 // Form the QQQQ REG_SEQUENCE.
1333 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1334 V[i] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1335 N->getOperand(Vec+3));
1336 V[i+1] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1337 N->getOperand(Vec+3));
1340 V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1343 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1344 V[4], V[5], V[6], V[7]), 0);
1346 // Store the even D registers.
1347 Ops.push_back(Reg0); // post-access address offset
1348 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1349 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0+Vec*2, dl,
1351 Ops.push_back(Pred);
1352 Ops.push_back(Reg0); // predicate register
1353 Ops.push_back(Chain);
1354 unsigned Opc = QOpcodes0[OpcodeIndex];
1355 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1356 MVT::Other, Ops.data(), NumVecs+6);
1357 Chain = SDValue(VStA, 1);
1359 // Store the odd D registers.
1360 Ops[0] = SDValue(VStA, 0); // MemAddr
1361 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1362 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1+Vec*2, dl,
1364 Ops[NumVecs+5] = Chain;
1365 Opc = QOpcodes1[OpcodeIndex];
1366 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1367 MVT::Other, Ops.data(), NumVecs+6);
1368 Chain = SDValue(VStB, 1);
1369 ReplaceUses(SDValue(N, 0), Chain);
1372 Ops.push_back(Reg0); // post-access address offset
1374 // Store the even subregs.
1375 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1376 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1377 N->getOperand(Vec+3)));
1378 Ops.push_back(Pred);
1379 Ops.push_back(Reg0); // predicate register
1380 Ops.push_back(Chain);
1381 unsigned Opc = QOpcodes0[OpcodeIndex];
1382 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1383 MVT::Other, Ops.data(), NumVecs+6);
1384 Chain = SDValue(VStA, 1);
1386 // Store the odd subregs.
1387 Ops[0] = SDValue(VStA, 0); // MemAddr
1388 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1389 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1390 N->getOperand(Vec+3));
1391 Ops[NumVecs+5] = Chain;
1392 Opc = QOpcodes1[OpcodeIndex];
1393 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1394 MVT::Other, Ops.data(), NumVecs+6);
1395 Chain = SDValue(VStB, 1);
1396 ReplaceUses(SDValue(N, 0), Chain);
1401 SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1402 unsigned NumVecs, unsigned *DOpcodes,
1403 unsigned *QOpcodes0,
1404 unsigned *QOpcodes1) {
1405 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1406 DebugLoc dl = N->getDebugLoc();
1408 SDValue MemAddr, Align;
1409 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1412 SDValue Chain = N->getOperand(0);
1414 cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1415 EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
1416 bool is64BitVector = VT.is64BitVector();
1418 // Quad registers are handled by load/store of subregs. Find the subreg info.
1419 unsigned NumElts = 0;
1423 if (!is64BitVector) {
1424 RegVT = GetNEONSubregVT(VT);
1425 NumElts = RegVT.getVectorNumElements();
1426 SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
1427 Even = Lane < NumElts;
1430 unsigned OpcodeIndex;
1431 switch (VT.getSimpleVT().SimpleTy) {
1432 default: llvm_unreachable("unhandled vld/vst lane type");
1433 // Double-register operations:
1434 case MVT::v8i8: OpcodeIndex = 0; break;
1435 case MVT::v4i16: OpcodeIndex = 1; break;
1437 case MVT::v2i32: OpcodeIndex = 2; break;
1438 // Quad-register operations:
1439 case MVT::v8i16: OpcodeIndex = 0; break;
1441 case MVT::v4i32: OpcodeIndex = 1; break;
1444 SDValue Pred = getAL(CurDAG);
1445 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1447 SmallVector<SDValue, 10> Ops;
1448 Ops.push_back(MemAddr);
1449 Ops.push_back(Align);
1452 if (is64BitVector) {
1453 Opc = DOpcodes[OpcodeIndex];
1454 if (llvm::ModelWithRegSequence()) {
1456 SDValue V0 = N->getOperand(0+3);
1457 SDValue V1 = N->getOperand(1+3);
1459 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1461 SDValue V2 = N->getOperand(2+3);
1462 SDValue V3 = (NumVecs == 3)
1463 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1464 : N->getOperand(3+3);
1465 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1468 // Now extract the D registers back out.
1469 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, VT,
1471 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, VT,
1474 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_2, dl, VT,
1477 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_3, dl, VT,
1480 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1481 Ops.push_back(N->getOperand(Vec+3));
1484 // Check if this is loading the even or odd subreg of a Q register.
1485 if (Lane < NumElts) {
1486 Opc = QOpcodes0[OpcodeIndex];
1489 Opc = QOpcodes1[OpcodeIndex];
1492 if (llvm::ModelWithRegSequence()) {
1494 SDValue V0 = N->getOperand(0+3);
1495 SDValue V1 = N->getOperand(1+3);
1497 RegSeq = SDValue(PairQRegs(MVT::v4i64, V0, V1), 0);
1499 SDValue V2 = N->getOperand(2+3);
1500 SDValue V3 = (NumVecs == 3)
1501 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1502 : N->getOperand(3+3);
1503 RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
1506 // Extract the subregs of the input vector.
1507 unsigned SubIdx = Even ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
1508 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1509 Ops.push_back(CurDAG->getTargetExtractSubreg(SubIdx+Vec*2, dl, RegVT,
1512 // Extract the subregs of the input vector.
1513 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1514 Ops.push_back(CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1515 N->getOperand(Vec+3)));
1518 Ops.push_back(getI32Imm(Lane));
1519 Ops.push_back(Pred);
1520 Ops.push_back(Reg0);
1521 Ops.push_back(Chain);
1524 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6);
1526 std::vector<EVT> ResTys(NumVecs, RegVT);
1527 ResTys.push_back(MVT::Other);
1528 SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(),NumVecs+6);
1530 if (llvm::ModelWithRegSequence()) {
1531 // Form a REG_SEQUENCE to force register allocation.
1533 if (is64BitVector) {
1534 SDValue V0 = SDValue(VLdLn, 0);
1535 SDValue V1 = SDValue(VLdLn, 1);
1537 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1539 SDValue V2 = SDValue(VLdLn, 2);
1540 // If it's a vld3, form a quad D-register but discard the last part.
1541 SDValue V3 = (NumVecs == 3)
1542 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1543 : SDValue(VLdLn, 3);
1544 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1547 // For 128-bit vectors, take the 64-bit results of the load and insert them
1548 // as subregs into the result.
1550 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1552 V[i] = SDValue(VLdLn, Vec);
1553 V[i+1] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1556 V[i] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1558 V[i+1] = SDValue(VLdLn, Vec);
1562 V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1566 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V[0], V[1], V[2], V[3]), 0);
1568 RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1569 V[4], V[5], V[6], V[7]), 0);
1572 unsigned SubIdx = is64BitVector ? ARM::DSUBREG_0 : ARM::QSUBREG_0;
1573 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1574 ReplaceUses(SDValue(N, Vec),
1575 CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, RegSeq));
1576 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, NumVecs));
1580 // For a 64-bit vector load to D registers, nothing more needs to be done.
1584 // For 128-bit vectors, take the 64-bit results of the load and insert them
1585 // as subregs into the result.
1586 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1587 SDValue QuadVec = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
1588 N->getOperand(Vec+3),
1589 SDValue(VLdLn, Vec));
1590 ReplaceUses(SDValue(N, Vec), QuadVec);
1593 Chain = SDValue(VLdLn, NumVecs);
1594 ReplaceUses(SDValue(N, NumVecs), Chain);
1598 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
1600 if (!Subtarget->hasV6T2Ops())
1603 unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
1604 : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
1607 // For unsigned extracts, check for a shift right and mask
1608 unsigned And_imm = 0;
1609 if (N->getOpcode() == ISD::AND) {
1610 if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
1612 // The immediate is a mask of the low bits iff imm & (imm+1) == 0
1613 if (And_imm & (And_imm + 1))
1616 unsigned Srl_imm = 0;
1617 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
1619 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1621 unsigned Width = CountTrailingOnes_32(And_imm);
1622 unsigned LSB = Srl_imm;
1623 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1624 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1625 CurDAG->getTargetConstant(LSB, MVT::i32),
1626 CurDAG->getTargetConstant(Width, MVT::i32),
1627 getAL(CurDAG), Reg0 };
1628 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1634 // Otherwise, we're looking for a shift of a shift
1635 unsigned Shl_imm = 0;
1636 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
1637 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1638 unsigned Srl_imm = 0;
1639 if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
1640 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1641 unsigned Width = 32 - Srl_imm;
1642 int LSB = Srl_imm - Shl_imm;
1645 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1646 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1647 CurDAG->getTargetConstant(LSB, MVT::i32),
1648 CurDAG->getTargetConstant(Width, MVT::i32),
1649 getAL(CurDAG), Reg0 };
1650 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1656 SDNode *ARMDAGToDAGISel::
1657 SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1658 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1661 if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
1662 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1663 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1666 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1667 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1668 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1669 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1671 llvm_unreachable("Unknown so_reg opcode!");
1675 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1676 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1677 SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
1678 return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
1683 SDNode *ARMDAGToDAGISel::
1684 SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1685 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1689 if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
1690 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1691 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
1692 return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
1697 SDNode *ARMDAGToDAGISel::
1698 SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1699 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1700 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1704 if (Predicate_t2_so_imm(TrueVal.getNode())) {
1705 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1706 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1707 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1708 return CurDAG->SelectNodeTo(N,
1709 ARM::t2MOVCCi, MVT::i32, Ops, 5);
1714 SDNode *ARMDAGToDAGISel::
1715 SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1716 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1717 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1721 if (Predicate_so_imm(TrueVal.getNode())) {
1722 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1723 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1724 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1725 return CurDAG->SelectNodeTo(N,
1726 ARM::MOVCCi, MVT::i32, Ops, 5);
1731 SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
1732 EVT VT = N->getValueType(0);
1733 SDValue FalseVal = N->getOperand(0);
1734 SDValue TrueVal = N->getOperand(1);
1735 SDValue CC = N->getOperand(2);
1736 SDValue CCR = N->getOperand(3);
1737 SDValue InFlag = N->getOperand(4);
1738 assert(CC.getOpcode() == ISD::Constant);
1739 assert(CCR.getOpcode() == ISD::Register);
1740 ARMCC::CondCodes CCVal =
1741 (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
1743 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1744 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1745 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1746 // Pattern complexity = 18 cost = 1 size = 0
1750 if (Subtarget->isThumb()) {
1751 SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
1752 CCVal, CCR, InFlag);
1754 Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
1755 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1759 SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
1760 CCVal, CCR, InFlag);
1762 Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
1763 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1768 // Pattern: (ARMcmov:i32 GPR:i32:$false,
1769 // (imm:i32)<<P:Predicate_so_imm>>:$true,
1771 // Emits: (MOVCCi:i32 GPR:i32:$false,
1772 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1773 // Pattern complexity = 10 cost = 1 size = 0
1774 if (Subtarget->isThumb()) {
1775 SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
1776 CCVal, CCR, InFlag);
1778 Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
1779 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1783 SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
1784 CCVal, CCR, InFlag);
1786 Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
1787 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1793 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1794 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1795 // Pattern complexity = 6 cost = 1 size = 0
1797 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1798 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1799 // Pattern complexity = 6 cost = 11 size = 0
1801 // Also FCPYScc and FCPYDcc.
1802 SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
1803 SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
1805 switch (VT.getSimpleVT().SimpleTy) {
1806 default: assert(false && "Illegal conditional move type!");
1809 Opc = Subtarget->isThumb()
1810 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1820 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1823 SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
1824 // The only time a CONCAT_VECTORS operation can have legal types is when
1825 // two 64-bit vectors are concatenated to a 128-bit vector.
1826 EVT VT = N->getValueType(0);
1827 if (!VT.is128BitVector() || N->getNumOperands() != 2)
1828 llvm_unreachable("unexpected CONCAT_VECTORS");
1829 DebugLoc dl = N->getDebugLoc();
1830 SDValue V0 = N->getOperand(0);
1831 SDValue V1 = N->getOperand(1);
1832 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
1833 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
1834 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
1835 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
1838 SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
1839 DebugLoc dl = N->getDebugLoc();
1841 if (N->isMachineOpcode())
1842 return NULL; // Already selected.
1844 switch (N->getOpcode()) {
1846 case ISD::Constant: {
1847 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1849 if (Subtarget->hasThumb2())
1850 // Thumb2-aware targets have the MOVT instruction, so all immediates can
1851 // be done with MOV + MOVT, at worst.
1854 if (Subtarget->isThumb()) {
1855 UseCP = (Val > 255 && // MOV
1856 ~Val > 255 && // MOV + MVN
1857 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
1859 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
1860 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
1861 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
1866 CurDAG->getTargetConstantPool(ConstantInt::get(
1867 Type::getInt32Ty(*CurDAG->getContext()), Val),
1868 TLI.getPointerTy());
1871 if (Subtarget->isThumb1Only()) {
1872 SDValue Pred = getAL(CurDAG);
1873 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1874 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1875 ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1880 CurDAG->getRegister(0, MVT::i32),
1881 CurDAG->getTargetConstant(0, MVT::i32),
1883 CurDAG->getRegister(0, MVT::i32),
1884 CurDAG->getEntryNode()
1886 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1889 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
1893 // Other cases are autogenerated.
1896 case ISD::FrameIndex: {
1897 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1898 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1899 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1900 if (Subtarget->isThumb1Only()) {
1901 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1902 CurDAG->getTargetConstant(0, MVT::i32));
1904 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1905 ARM::t2ADDri : ARM::ADDri);
1906 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1907 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1908 CurDAG->getRegister(0, MVT::i32) };
1909 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1913 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1917 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true))
1921 if (Subtarget->isThumb1Only())
1923 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1924 unsigned RHSV = C->getZExtValue();
1926 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
1927 unsigned ShImm = Log2_32(RHSV-1);
1930 SDValue V = N->getOperand(0);
1931 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1932 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1933 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1934 if (Subtarget->isThumb()) {
1935 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1936 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1938 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1939 return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1942 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
1943 unsigned ShImm = Log2_32(RHSV+1);
1946 SDValue V = N->getOperand(0);
1947 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1948 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1949 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1950 if (Subtarget->isThumb()) {
1951 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0 };
1952 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 5);
1954 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1955 return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1961 // Check for unsigned bitfield extract
1962 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1965 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
1966 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
1967 // are entirely contributed by c2 and lower 16-bits are entirely contributed
1968 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
1969 // Select it to: "movt x, ((c1 & 0xffff) >> 16)
1970 EVT VT = N->getValueType(0);
1973 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
1975 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
1978 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
1979 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1982 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
1983 SDValue N2 = N0.getOperand(1);
1984 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1987 unsigned N1CVal = N1C->getZExtValue();
1988 unsigned N2CVal = N2C->getZExtValue();
1989 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
1990 (N1CVal & 0xffffU) == 0xffffU &&
1991 (N2CVal & 0xffffU) == 0x0U) {
1992 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
1994 SDValue Ops[] = { N0.getOperand(0), Imm16,
1995 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1996 return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
2001 case ARMISD::VMOVRRD:
2002 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
2003 N->getOperand(0), getAL(CurDAG),
2004 CurDAG->getRegister(0, MVT::i32));
2005 case ISD::UMUL_LOHI: {
2006 if (Subtarget->isThumb1Only())
2008 if (Subtarget->isThumb()) {
2009 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2010 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2011 CurDAG->getRegister(0, MVT::i32) };
2012 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops,4);
2014 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2015 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2016 CurDAG->getRegister(0, MVT::i32) };
2017 return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
2020 case ISD::SMUL_LOHI: {
2021 if (Subtarget->isThumb1Only())
2023 if (Subtarget->isThumb()) {
2024 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2025 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2026 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops,4);
2028 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2029 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2030 CurDAG->getRegister(0, MVT::i32) };
2031 return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
2035 SDNode *ResNode = 0;
2036 if (Subtarget->isThumb() && Subtarget->hasThumb2())
2037 ResNode = SelectT2IndexedLoad(N);
2039 ResNode = SelectARMIndexedLoad(N);
2043 // VLDMQ must be custom-selected for "v2f64 load" to set the AM5Opc value.
2044 if (Subtarget->hasVFP2() &&
2045 N->getValueType(0).getSimpleVT().SimpleTy == MVT::v2f64) {
2046 SDValue Chain = N->getOperand(0);
2048 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
2049 SDValue Pred = getAL(CurDAG);
2050 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2051 SDValue Ops[] = { N->getOperand(1), AM5Opc, Pred, PredReg, Chain };
2052 return CurDAG->getMachineNode(ARM::VLDMQ, dl, MVT::v2f64, MVT::Other,
2055 // Other cases are autogenerated.
2059 // VSTMQ must be custom-selected for "v2f64 store" to set the AM5Opc value.
2060 if (Subtarget->hasVFP2() &&
2061 N->getOperand(1).getValueType().getSimpleVT().SimpleTy == MVT::v2f64) {
2062 SDValue Chain = N->getOperand(0);
2064 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
2065 SDValue Pred = getAL(CurDAG);
2066 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2067 SDValue Ops[] = { N->getOperand(1), N->getOperand(2),
2068 AM5Opc, Pred, PredReg, Chain };
2069 return CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
2071 // Other cases are autogenerated.
2074 case ARMISD::BRCOND: {
2075 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2076 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2077 // Pattern complexity = 6 cost = 1 size = 0
2079 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2080 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
2081 // Pattern complexity = 6 cost = 1 size = 0
2083 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2084 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2085 // Pattern complexity = 6 cost = 1 size = 0
2087 unsigned Opc = Subtarget->isThumb() ?
2088 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
2089 SDValue Chain = N->getOperand(0);
2090 SDValue N1 = N->getOperand(1);
2091 SDValue N2 = N->getOperand(2);
2092 SDValue N3 = N->getOperand(3);
2093 SDValue InFlag = N->getOperand(4);
2094 assert(N1.getOpcode() == ISD::BasicBlock);
2095 assert(N2.getOpcode() == ISD::Constant);
2096 assert(N3.getOpcode() == ISD::Register);
2098 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2099 cast<ConstantSDNode>(N2)->getZExtValue()),
2101 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
2102 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
2104 Chain = SDValue(ResNode, 0);
2105 if (N->getNumValues() == 2) {
2106 InFlag = SDValue(ResNode, 1);
2107 ReplaceUses(SDValue(N, 1), InFlag);
2109 ReplaceUses(SDValue(N, 0),
2110 SDValue(Chain.getNode(), Chain.getResNo()));
2114 return SelectCMOVOp(N);
2115 case ARMISD::CNEG: {
2116 EVT VT = N->getValueType(0);
2117 SDValue N0 = N->getOperand(0);
2118 SDValue N1 = N->getOperand(1);
2119 SDValue N2 = N->getOperand(2);
2120 SDValue N3 = N->getOperand(3);
2121 SDValue InFlag = N->getOperand(4);
2122 assert(N2.getOpcode() == ISD::Constant);
2123 assert(N3.getOpcode() == ISD::Register);
2125 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2126 cast<ConstantSDNode>(N2)->getZExtValue()),
2128 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
2130 switch (VT.getSimpleVT().SimpleTy) {
2131 default: assert(false && "Illegal conditional move type!");
2140 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
2143 case ARMISD::VZIP: {
2145 EVT VT = N->getValueType(0);
2146 switch (VT.getSimpleVT().SimpleTy) {
2147 default: return NULL;
2148 case MVT::v8i8: Opc = ARM::VZIPd8; break;
2149 case MVT::v4i16: Opc = ARM::VZIPd16; break;
2151 case MVT::v2i32: Opc = ARM::VZIPd32; break;
2152 case MVT::v16i8: Opc = ARM::VZIPq8; break;
2153 case MVT::v8i16: Opc = ARM::VZIPq16; break;
2155 case MVT::v4i32: Opc = ARM::VZIPq32; break;
2157 SDValue Pred = getAL(CurDAG);
2158 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2159 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2160 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2162 case ARMISD::VUZP: {
2164 EVT VT = N->getValueType(0);
2165 switch (VT.getSimpleVT().SimpleTy) {
2166 default: return NULL;
2167 case MVT::v8i8: Opc = ARM::VUZPd8; break;
2168 case MVT::v4i16: Opc = ARM::VUZPd16; break;
2170 case MVT::v2i32: Opc = ARM::VUZPd32; break;
2171 case MVT::v16i8: Opc = ARM::VUZPq8; break;
2172 case MVT::v8i16: Opc = ARM::VUZPq16; break;
2174 case MVT::v4i32: Opc = ARM::VUZPq32; break;
2176 SDValue Pred = getAL(CurDAG);
2177 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2178 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2179 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2181 case ARMISD::VTRN: {
2183 EVT VT = N->getValueType(0);
2184 switch (VT.getSimpleVT().SimpleTy) {
2185 default: return NULL;
2186 case MVT::v8i8: Opc = ARM::VTRNd8; break;
2187 case MVT::v4i16: Opc = ARM::VTRNd16; break;
2189 case MVT::v2i32: Opc = ARM::VTRNd32; break;
2190 case MVT::v16i8: Opc = ARM::VTRNq8; break;
2191 case MVT::v8i16: Opc = ARM::VTRNq16; break;
2193 case MVT::v4i32: Opc = ARM::VTRNq32; break;
2195 SDValue Pred = getAL(CurDAG);
2196 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2197 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2198 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2201 case ISD::INTRINSIC_VOID:
2202 case ISD::INTRINSIC_W_CHAIN: {
2203 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2208 case Intrinsic::arm_neon_vld1: {
2209 unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
2210 ARM::VLD1d32, ARM::VLD1d64 };
2211 unsigned QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
2212 ARM::VLD1q32, ARM::VLD1q64 };
2213 return SelectVLD(N, 1, DOpcodes, QOpcodes, 0);
2216 case Intrinsic::arm_neon_vld2: {
2217 unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
2218 ARM::VLD2d32, ARM::VLD1q64 };
2219 unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
2220 return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
2223 case Intrinsic::arm_neon_vld3: {
2224 unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
2225 ARM::VLD3d32, ARM::VLD1d64T };
2226 unsigned QOpcodes0[] = { ARM::VLD3q8_UPD,
2229 unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD,
2230 ARM::VLD3q16odd_UPD,
2231 ARM::VLD3q32odd_UPD };
2232 return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2235 case Intrinsic::arm_neon_vld4: {
2236 unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
2237 ARM::VLD4d32, ARM::VLD1d64Q };
2238 unsigned QOpcodes0[] = { ARM::VLD4q8_UPD,
2241 unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD,
2242 ARM::VLD4q16odd_UPD,
2243 ARM::VLD4q32odd_UPD };
2244 return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2247 case Intrinsic::arm_neon_vld2lane: {
2248 unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
2249 unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 };
2250 unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd };
2251 return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
2254 case Intrinsic::arm_neon_vld3lane: {
2255 unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
2256 unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 };
2257 unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd };
2258 return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
2261 case Intrinsic::arm_neon_vld4lane: {
2262 unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
2263 unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 };
2264 unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd };
2265 return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
2268 case Intrinsic::arm_neon_vst1: {
2269 unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
2270 ARM::VST1d32, ARM::VST1d64 };
2271 unsigned QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
2272 ARM::VST1q32, ARM::VST1q64 };
2273 return SelectVST(N, 1, DOpcodes, QOpcodes, 0);
2276 case Intrinsic::arm_neon_vst2: {
2277 unsigned DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
2278 ARM::VST2d32, ARM::VST1q64 };
2279 unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
2280 return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
2283 case Intrinsic::arm_neon_vst3: {
2284 unsigned DOpcodes[] = { ARM::VST3d8, ARM::VST3d16,
2285 ARM::VST3d32, ARM::VST1d64T };
2286 unsigned QOpcodes0[] = { ARM::VST3q8_UPD,
2289 unsigned QOpcodes1[] = { ARM::VST3q8odd_UPD,
2290 ARM::VST3q16odd_UPD,
2291 ARM::VST3q32odd_UPD };
2292 return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2295 case Intrinsic::arm_neon_vst4: {
2296 unsigned DOpcodes[] = { ARM::VST4d8, ARM::VST4d16,
2297 ARM::VST4d32, ARM::VST1d64Q };
2298 unsigned QOpcodes0[] = { ARM::VST4q8_UPD,
2301 unsigned QOpcodes1[] = { ARM::VST4q8odd_UPD,
2302 ARM::VST4q16odd_UPD,
2303 ARM::VST4q32odd_UPD };
2304 return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2307 case Intrinsic::arm_neon_vst2lane: {
2308 unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
2309 unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 };
2310 unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd };
2311 return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
2314 case Intrinsic::arm_neon_vst3lane: {
2315 unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
2316 unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 };
2317 unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd };
2318 return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
2321 case Intrinsic::arm_neon_vst4lane: {
2322 unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
2323 unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 };
2324 unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd };
2325 return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
2331 case ISD::CONCAT_VECTORS:
2332 return SelectConcatVector(N);
2335 return SelectCode(N);
2338 bool ARMDAGToDAGISel::
2339 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2340 std::vector<SDValue> &OutOps) {
2341 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
2342 // Require the address to be in a register. That is safe for all ARM
2343 // variants and it is hard to do anything much smarter without knowing
2344 // how the operand is used.
2345 OutOps.push_back(Op);
2349 /// createARMISelDag - This pass converts a legalized DAG into a
2350 /// ARM-specific DAG, ready for instruction scheduling.
2352 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
2353 CodeGenOpt::Level OptLevel) {
2354 return new ARMDAGToDAGISel(TM, OptLevel);
2357 /// ModelWithRegSequence - Return true if isel should use REG_SEQUENCE to model
2358 /// operations involving sub-registers.
2359 bool llvm::ModelWithRegSequence() {