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"),
43 //===--------------------------------------------------------------------===//
44 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
45 /// instructions for SelectionDAG operations.
48 class ARMDAGToDAGISel : public SelectionDAGISel {
49 ARMBaseTargetMachine &TM;
51 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
52 /// make the right decision when generating code for different targets.
53 const ARMSubtarget *Subtarget;
56 explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
57 CodeGenOpt::Level OptLevel)
58 : SelectionDAGISel(tm, OptLevel), TM(tm),
59 Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
62 virtual const char *getPassName() const {
63 return "ARM Instruction Selection";
66 /// getI32Imm - Return a target constant of type i32 with the specified
68 inline SDValue getI32Imm(unsigned Imm) {
69 return CurDAG->getTargetConstant(Imm, MVT::i32);
72 SDNode *Select(SDNode *N);
74 bool SelectShifterOperandReg(SDNode *Op, SDValue N, SDValue &A,
75 SDValue &B, SDValue &C);
76 bool SelectAddrMode2(SDNode *Op, SDValue N, SDValue &Base,
77 SDValue &Offset, SDValue &Opc);
78 bool SelectAddrMode2Offset(SDNode *Op, SDValue N,
79 SDValue &Offset, SDValue &Opc);
80 bool SelectAddrMode3(SDNode *Op, SDValue N, SDValue &Base,
81 SDValue &Offset, SDValue &Opc);
82 bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
83 SDValue &Offset, SDValue &Opc);
84 bool SelectAddrMode4(SDNode *Op, SDValue N, SDValue &Addr,
86 bool SelectAddrMode5(SDNode *Op, SDValue N, SDValue &Base,
88 bool SelectAddrMode6(SDNode *Op, SDValue N, SDValue &Addr, SDValue &Align);
90 bool SelectAddrModePC(SDNode *Op, SDValue N, SDValue &Offset,
93 bool SelectThumbAddrModeRR(SDNode *Op, SDValue N, SDValue &Base,
95 bool SelectThumbAddrModeRI5(SDNode *Op, SDValue N, unsigned Scale,
96 SDValue &Base, SDValue &OffImm,
98 bool SelectThumbAddrModeS1(SDNode *Op, SDValue N, SDValue &Base,
99 SDValue &OffImm, SDValue &Offset);
100 bool SelectThumbAddrModeS2(SDNode *Op, SDValue N, SDValue &Base,
101 SDValue &OffImm, SDValue &Offset);
102 bool SelectThumbAddrModeS4(SDNode *Op, SDValue N, SDValue &Base,
103 SDValue &OffImm, SDValue &Offset);
104 bool SelectThumbAddrModeSP(SDNode *Op, SDValue N, SDValue &Base,
107 bool SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
108 SDValue &BaseReg, SDValue &Opc);
109 bool SelectT2AddrModeImm12(SDNode *Op, SDValue N, SDValue &Base,
111 bool SelectT2AddrModeImm8(SDNode *Op, SDValue N, SDValue &Base,
113 bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
115 bool SelectT2AddrModeImm8s4(SDNode *Op, SDValue N, SDValue &Base,
117 bool SelectT2AddrModeSoReg(SDNode *Op, SDValue N, SDValue &Base,
118 SDValue &OffReg, SDValue &ShImm);
120 // Include the pieces autogenerated from the target description.
121 #include "ARMGenDAGISel.inc"
124 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
126 SDNode *SelectARMIndexedLoad(SDNode *N);
127 SDNode *SelectT2IndexedLoad(SDNode *N);
129 /// SelectVLD - Select NEON load intrinsics. NumVecs should be
130 /// 1, 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(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
134 unsigned *QOpcodes0, unsigned *QOpcodes1);
136 /// SelectVST - Select NEON store intrinsics. NumVecs should
137 /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for
138 /// stores of D registers and even subregs and odd subregs of Q registers.
139 /// For NumVecs <= 2, QOpcodes1 is not used.
140 SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
141 unsigned *QOpcodes0, unsigned *QOpcodes1);
143 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
144 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
145 /// load/store of D registers and even subregs and odd subregs of Q registers.
146 SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, unsigned NumVecs,
147 unsigned *DOpcodes, unsigned *QOpcodes0,
148 unsigned *QOpcodes1);
150 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
151 SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
153 /// SelectCMOVOp - Select CMOV instructions for ARM.
154 SDNode *SelectCMOVOp(SDNode *N);
155 SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
156 ARMCC::CondCodes CCVal, SDValue CCR,
158 SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
159 ARMCC::CondCodes CCVal, SDValue CCR,
161 SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
162 ARMCC::CondCodes CCVal, SDValue CCR,
164 SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
165 ARMCC::CondCodes CCVal, SDValue CCR,
168 SDNode *SelectConcatVector(SDNode *N);
170 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
171 /// inline asm expressions.
172 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
174 std::vector<SDValue> &OutOps);
176 /// PairDRegs - Form a quad register from a pair of D registers.
178 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
180 /// PairDRegs - Form a quad register pair from a pair of Q registers.
182 SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1);
184 /// QuadDRegs - Form a quad register pair from a quad of D registers.
186 SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
188 /// QuadQRegs - Form 4 consecutive Q registers.
190 SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
192 /// OctoDRegs - Form 8 consecutive D registers.
194 SDNode *OctoDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3,
195 SDValue V4, SDValue V5, SDValue V6, SDValue V7);
199 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
200 /// operand. If so Imm will receive the 32-bit value.
201 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
202 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
203 Imm = cast<ConstantSDNode>(N)->getZExtValue();
209 // isInt32Immediate - This method tests to see if a constant operand.
210 // If so Imm will receive the 32 bit value.
211 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
212 return isInt32Immediate(N.getNode(), Imm);
215 // isOpcWithIntImmediate - This method tests to see if the node is a specific
216 // opcode and that it has a immediate integer right operand.
217 // If so Imm will receive the 32 bit value.
218 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
219 return N->getOpcode() == Opc &&
220 isInt32Immediate(N->getOperand(1).getNode(), Imm);
224 bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
229 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
231 // Don't match base register only case. That is matched to a separate
232 // lower complexity pattern with explicit register operand.
233 if (ShOpcVal == ARM_AM::no_shift) return false;
235 BaseReg = N.getOperand(0);
236 unsigned ShImmVal = 0;
237 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
238 ShReg = CurDAG->getRegister(0, MVT::i32);
239 ShImmVal = RHS->getZExtValue() & 31;
241 ShReg = N.getOperand(1);
243 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
248 bool ARMDAGToDAGISel::SelectAddrMode2(SDNode *Op, SDValue N,
249 SDValue &Base, SDValue &Offset,
251 if (N.getOpcode() == ISD::MUL) {
252 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
253 // X * [3,5,9] -> X + X * [2,4,8] etc.
254 int RHSC = (int)RHS->getZExtValue();
257 ARM_AM::AddrOpc AddSub = ARM_AM::add;
259 AddSub = ARM_AM::sub;
262 if (isPowerOf2_32(RHSC)) {
263 unsigned ShAmt = Log2_32(RHSC);
264 Base = Offset = N.getOperand(0);
265 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
274 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
276 if (N.getOpcode() == ISD::FrameIndex) {
277 int FI = cast<FrameIndexSDNode>(N)->getIndex();
278 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
279 } else if (N.getOpcode() == ARMISD::Wrapper &&
280 !(Subtarget->useMovt() &&
281 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
282 Base = N.getOperand(0);
284 Offset = CurDAG->getRegister(0, MVT::i32);
285 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
291 // Match simple R +/- imm12 operands.
292 if (N.getOpcode() == ISD::ADD)
293 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
294 int RHSC = (int)RHS->getZExtValue();
295 if ((RHSC >= 0 && RHSC < 0x1000) ||
296 (RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
297 Base = N.getOperand(0);
298 if (Base.getOpcode() == ISD::FrameIndex) {
299 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
300 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
302 Offset = CurDAG->getRegister(0, MVT::i32);
304 ARM_AM::AddrOpc AddSub = ARM_AM::add;
306 AddSub = ARM_AM::sub;
309 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
316 // Otherwise this is R +/- [possibly shifted] R.
317 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
318 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
321 Base = N.getOperand(0);
322 Offset = N.getOperand(1);
324 if (ShOpcVal != ARM_AM::no_shift) {
325 // Check to see if the RHS of the shift is a constant, if not, we can't fold
327 if (ConstantSDNode *Sh =
328 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
329 ShAmt = Sh->getZExtValue();
330 Offset = N.getOperand(1).getOperand(0);
332 ShOpcVal = ARM_AM::no_shift;
336 // Try matching (R shl C) + (R).
337 if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
338 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
339 if (ShOpcVal != ARM_AM::no_shift) {
340 // Check to see if the RHS of the shift is a constant, if not, we can't
342 if (ConstantSDNode *Sh =
343 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
344 ShAmt = Sh->getZExtValue();
345 Offset = N.getOperand(0).getOperand(0);
346 Base = N.getOperand(1);
348 ShOpcVal = ARM_AM::no_shift;
353 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
358 bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
359 SDValue &Offset, SDValue &Opc) {
360 unsigned Opcode = Op->getOpcode();
361 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
362 ? cast<LoadSDNode>(Op)->getAddressingMode()
363 : cast<StoreSDNode>(Op)->getAddressingMode();
364 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
365 ? ARM_AM::add : ARM_AM::sub;
366 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
367 int Val = (int)C->getZExtValue();
368 if (Val >= 0 && Val < 0x1000) { // 12 bits.
369 Offset = CurDAG->getRegister(0, MVT::i32);
370 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
378 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
380 if (ShOpcVal != ARM_AM::no_shift) {
381 // Check to see if the RHS of the shift is a constant, if not, we can't fold
383 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
384 ShAmt = Sh->getZExtValue();
385 Offset = N.getOperand(0);
387 ShOpcVal = ARM_AM::no_shift;
391 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
397 bool ARMDAGToDAGISel::SelectAddrMode3(SDNode *Op, SDValue N,
398 SDValue &Base, SDValue &Offset,
400 if (N.getOpcode() == ISD::SUB) {
401 // X - C is canonicalize to X + -C, no need to handle it here.
402 Base = N.getOperand(0);
403 Offset = N.getOperand(1);
404 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
408 if (N.getOpcode() != ISD::ADD) {
410 if (N.getOpcode() == ISD::FrameIndex) {
411 int FI = cast<FrameIndexSDNode>(N)->getIndex();
412 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
414 Offset = CurDAG->getRegister(0, MVT::i32);
415 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
419 // If the RHS is +/- imm8, fold into addr mode.
420 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
421 int RHSC = (int)RHS->getZExtValue();
422 if ((RHSC >= 0 && RHSC < 256) ||
423 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
424 Base = N.getOperand(0);
425 if (Base.getOpcode() == ISD::FrameIndex) {
426 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
427 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
429 Offset = CurDAG->getRegister(0, MVT::i32);
431 ARM_AM::AddrOpc AddSub = ARM_AM::add;
433 AddSub = ARM_AM::sub;
436 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
441 Base = N.getOperand(0);
442 Offset = N.getOperand(1);
443 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
447 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
448 SDValue &Offset, SDValue &Opc) {
449 unsigned Opcode = Op->getOpcode();
450 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
451 ? cast<LoadSDNode>(Op)->getAddressingMode()
452 : cast<StoreSDNode>(Op)->getAddressingMode();
453 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
454 ? ARM_AM::add : ARM_AM::sub;
455 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
456 int Val = (int)C->getZExtValue();
457 if (Val >= 0 && Val < 256) {
458 Offset = CurDAG->getRegister(0, MVT::i32);
459 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
465 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
469 bool ARMDAGToDAGISel::SelectAddrMode4(SDNode *Op, SDValue N,
470 SDValue &Addr, SDValue &Mode) {
472 Mode = CurDAG->getTargetConstant(0, MVT::i32);
476 bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N,
477 SDValue &Base, SDValue &Offset) {
478 if (N.getOpcode() != ISD::ADD) {
480 if (N.getOpcode() == ISD::FrameIndex) {
481 int FI = cast<FrameIndexSDNode>(N)->getIndex();
482 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
483 } else if (N.getOpcode() == ARMISD::Wrapper &&
484 !(Subtarget->useMovt() &&
485 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
486 Base = N.getOperand(0);
488 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
493 // If the RHS is +/- imm8, fold into addr mode.
494 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
495 int RHSC = (int)RHS->getZExtValue();
496 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied by 4.
498 if ((RHSC >= 0 && RHSC < 256) ||
499 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
500 Base = N.getOperand(0);
501 if (Base.getOpcode() == ISD::FrameIndex) {
502 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
503 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
506 ARM_AM::AddrOpc AddSub = ARM_AM::add;
508 AddSub = ARM_AM::sub;
511 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
519 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
524 bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N,
525 SDValue &Addr, SDValue &Align) {
527 // Default to no alignment.
528 Align = CurDAG->getTargetConstant(0, MVT::i32);
532 bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
533 SDValue &Offset, SDValue &Label) {
534 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
535 Offset = N.getOperand(0);
536 SDValue N1 = N.getOperand(1);
537 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
544 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
545 SDValue &Base, SDValue &Offset){
546 // FIXME dl should come from the parent load or store, not the address
547 DebugLoc dl = Op->getDebugLoc();
548 if (N.getOpcode() != ISD::ADD) {
549 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
550 if (!NC || NC->getZExtValue() != 0)
557 Base = N.getOperand(0);
558 Offset = N.getOperand(1);
563 ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
564 unsigned Scale, SDValue &Base,
565 SDValue &OffImm, SDValue &Offset) {
567 SDValue TmpBase, TmpOffImm;
568 if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
569 return false; // We want to select tLDRspi / tSTRspi instead.
570 if (N.getOpcode() == ARMISD::Wrapper &&
571 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
572 return false; // We want to select tLDRpci instead.
575 if (N.getOpcode() != ISD::ADD) {
576 if (N.getOpcode() == ARMISD::Wrapper &&
577 !(Subtarget->useMovt() &&
578 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
579 Base = N.getOperand(0);
583 Offset = CurDAG->getRegister(0, MVT::i32);
584 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
588 // Thumb does not have [sp, r] address mode.
589 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
590 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
591 if ((LHSR && LHSR->getReg() == ARM::SP) ||
592 (RHSR && RHSR->getReg() == ARM::SP)) {
594 Offset = CurDAG->getRegister(0, MVT::i32);
595 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
599 // If the RHS is + imm5 * scale, fold into addr mode.
600 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
601 int RHSC = (int)RHS->getZExtValue();
602 if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
604 if (RHSC >= 0 && RHSC < 32) {
605 Base = N.getOperand(0);
606 Offset = CurDAG->getRegister(0, MVT::i32);
607 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
613 Base = N.getOperand(0);
614 Offset = N.getOperand(1);
615 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
619 bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
620 SDValue &Base, SDValue &OffImm,
622 return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
625 bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
626 SDValue &Base, SDValue &OffImm,
628 return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
631 bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
632 SDValue &Base, SDValue &OffImm,
634 return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
637 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
638 SDValue &Base, SDValue &OffImm) {
639 if (N.getOpcode() == ISD::FrameIndex) {
640 int FI = cast<FrameIndexSDNode>(N)->getIndex();
641 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
642 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
646 if (N.getOpcode() != ISD::ADD)
649 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
650 if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
651 (LHSR && LHSR->getReg() == ARM::SP)) {
652 // If the RHS is + imm8 * scale, fold into addr mode.
653 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
654 int RHSC = (int)RHS->getZExtValue();
655 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
657 if (RHSC >= 0 && RHSC < 256) {
658 Base = N.getOperand(0);
659 if (Base.getOpcode() == ISD::FrameIndex) {
660 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
661 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
663 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
673 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
676 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
678 // Don't match base register only case. That is matched to a separate
679 // lower complexity pattern with explicit register operand.
680 if (ShOpcVal == ARM_AM::no_shift) return false;
682 BaseReg = N.getOperand(0);
683 unsigned ShImmVal = 0;
684 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
685 ShImmVal = RHS->getZExtValue() & 31;
686 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
693 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
694 SDValue &Base, SDValue &OffImm) {
695 // Match simple R + imm12 operands.
698 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
699 if (N.getOpcode() == ISD::FrameIndex) {
700 // Match frame index...
701 int FI = cast<FrameIndexSDNode>(N)->getIndex();
702 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
703 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
705 } else if (N.getOpcode() == ARMISD::Wrapper &&
706 !(Subtarget->useMovt() &&
707 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
708 Base = N.getOperand(0);
709 if (Base.getOpcode() == ISD::TargetConstantPool)
710 return false; // We want to select t2LDRpci instead.
713 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
717 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
718 if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
719 // Let t2LDRi8 handle (R - imm8).
722 int RHSC = (int)RHS->getZExtValue();
723 if (N.getOpcode() == ISD::SUB)
726 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
727 Base = N.getOperand(0);
728 if (Base.getOpcode() == ISD::FrameIndex) {
729 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
730 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
732 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
739 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
743 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
744 SDValue &Base, SDValue &OffImm) {
745 // Match simple R - imm8 operands.
746 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
747 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
748 int RHSC = (int)RHS->getSExtValue();
749 if (N.getOpcode() == ISD::SUB)
752 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
753 Base = N.getOperand(0);
754 if (Base.getOpcode() == ISD::FrameIndex) {
755 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
756 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
758 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
767 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
769 unsigned Opcode = Op->getOpcode();
770 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
771 ? cast<LoadSDNode>(Op)->getAddressingMode()
772 : cast<StoreSDNode>(Op)->getAddressingMode();
773 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
774 int RHSC = (int)RHS->getZExtValue();
775 if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
776 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
777 ? CurDAG->getTargetConstant(RHSC, MVT::i32)
778 : CurDAG->getTargetConstant(-RHSC, MVT::i32);
786 bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
787 SDValue &Base, SDValue &OffImm) {
788 if (N.getOpcode() == ISD::ADD) {
789 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
790 int RHSC = (int)RHS->getZExtValue();
791 if (((RHSC & 0x3) == 0) &&
792 ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits.
793 Base = N.getOperand(0);
794 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
798 } else if (N.getOpcode() == ISD::SUB) {
799 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
800 int RHSC = (int)RHS->getZExtValue();
801 if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) { // 8 bits.
802 Base = N.getOperand(0);
803 OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
812 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDNode *Op, SDValue N,
814 SDValue &OffReg, SDValue &ShImm) {
815 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
816 if (N.getOpcode() != ISD::ADD)
819 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
820 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
821 int RHSC = (int)RHS->getZExtValue();
822 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
824 else if (RHSC < 0 && RHSC >= -255) // 8 bits
828 // Look for (R + R) or (R + (R << [1,2,3])).
830 Base = N.getOperand(0);
831 OffReg = N.getOperand(1);
833 // Swap if it is ((R << c) + R).
834 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
835 if (ShOpcVal != ARM_AM::lsl) {
836 ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
837 if (ShOpcVal == ARM_AM::lsl)
838 std::swap(Base, OffReg);
841 if (ShOpcVal == ARM_AM::lsl) {
842 // Check to see if the RHS of the shift is a constant, if not, we can't fold
844 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
845 ShAmt = Sh->getZExtValue();
848 ShOpcVal = ARM_AM::no_shift;
850 OffReg = OffReg.getOperand(0);
852 ShOpcVal = ARM_AM::no_shift;
856 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
861 //===--------------------------------------------------------------------===//
863 /// getAL - Returns a ARMCC::AL immediate node.
864 static inline SDValue getAL(SelectionDAG *CurDAG) {
865 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
868 SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
869 LoadSDNode *LD = cast<LoadSDNode>(N);
870 ISD::MemIndexedMode AM = LD->getAddressingMode();
871 if (AM == ISD::UNINDEXED)
874 EVT LoadedVT = LD->getMemoryVT();
875 SDValue Offset, AMOpc;
876 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
879 if (LoadedVT == MVT::i32 &&
880 SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
881 Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
883 } else if (LoadedVT == MVT::i16 &&
884 SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
886 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
887 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
888 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
889 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
890 if (LD->getExtensionType() == ISD::SEXTLOAD) {
891 if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
893 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
896 if (SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
898 Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
904 SDValue Chain = LD->getChain();
905 SDValue Base = LD->getBasePtr();
906 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
907 CurDAG->getRegister(0, MVT::i32), Chain };
908 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
915 SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
916 LoadSDNode *LD = cast<LoadSDNode>(N);
917 ISD::MemIndexedMode AM = LD->getAddressingMode();
918 if (AM == ISD::UNINDEXED)
921 EVT LoadedVT = LD->getMemoryVT();
922 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
924 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
927 if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
928 switch (LoadedVT.getSimpleVT().SimpleTy) {
930 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
934 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
936 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
941 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
943 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
952 SDValue Chain = LD->getChain();
953 SDValue Base = LD->getBasePtr();
954 SDValue Ops[]= { Base, Offset, getAL(CurDAG),
955 CurDAG->getRegister(0, MVT::i32), Chain };
956 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
963 /// PairDRegs - Form a quad register from a pair of D registers.
965 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
966 DebugLoc dl = V0.getNode()->getDebugLoc();
967 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
968 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
969 if (llvm::ModelWithRegSequence()) {
970 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
971 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
974 SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0);
975 SDNode *Pair = CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
976 VT, Undef, V0, SubReg0);
977 return CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
978 VT, SDValue(Pair, 0), V1, SubReg1);
981 /// PairQRegs - Form 4 consecutive D registers from a pair of Q registers.
983 SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) {
984 DebugLoc dl = V0.getNode()->getDebugLoc();
985 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::QSUBREG_0, MVT::i32);
986 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::QSUBREG_1, MVT::i32);
987 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
988 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
991 /// QuadDRegs - Form 4 consecutive D registers.
993 SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
994 SDValue V2, SDValue V3) {
995 DebugLoc dl = V0.getNode()->getDebugLoc();
996 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
997 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
998 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::DSUBREG_2, MVT::i32);
999 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::DSUBREG_3, MVT::i32);
1000 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1001 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1004 /// QuadQRegs - Form 4 consecutive Q registers.
1006 SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1,
1007 SDValue V2, SDValue V3) {
1008 DebugLoc dl = V0.getNode()->getDebugLoc();
1009 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::QSUBREG_0, MVT::i32);
1010 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::QSUBREG_1, MVT::i32);
1011 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::QSUBREG_2, MVT::i32);
1012 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::QSUBREG_3, MVT::i32);
1013 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1014 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1017 /// OctoDRegs - Form 8 consecutive D registers.
1019 SDNode *ARMDAGToDAGISel::OctoDRegs(EVT VT, SDValue V0, SDValue V1,
1020 SDValue V2, SDValue V3,
1021 SDValue V4, SDValue V5,
1022 SDValue V6, SDValue V7) {
1023 DebugLoc dl = V0.getNode()->getDebugLoc();
1024 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
1025 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
1026 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::DSUBREG_2, MVT::i32);
1027 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::DSUBREG_3, MVT::i32);
1028 SDValue SubReg4 = CurDAG->getTargetConstant(ARM::DSUBREG_4, MVT::i32);
1029 SDValue SubReg5 = CurDAG->getTargetConstant(ARM::DSUBREG_5, MVT::i32);
1030 SDValue SubReg6 = CurDAG->getTargetConstant(ARM::DSUBREG_6, MVT::i32);
1031 SDValue SubReg7 = CurDAG->getTargetConstant(ARM::DSUBREG_7, MVT::i32);
1032 const SDValue Ops[] ={ V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3,
1033 V4, SubReg4, V5, SubReg5, V6, SubReg6, V7, SubReg7 };
1034 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 16);
1037 /// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
1038 /// for a 64-bit subregister of the vector.
1039 static EVT GetNEONSubregVT(EVT VT) {
1040 switch (VT.getSimpleVT().SimpleTy) {
1041 default: llvm_unreachable("unhandled NEON type");
1042 case MVT::v16i8: return MVT::v8i8;
1043 case MVT::v8i16: return MVT::v4i16;
1044 case MVT::v4f32: return MVT::v2f32;
1045 case MVT::v4i32: return MVT::v2i32;
1046 case MVT::v2i64: return MVT::v1i64;
1050 SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
1051 unsigned *DOpcodes, unsigned *QOpcodes0,
1052 unsigned *QOpcodes1) {
1053 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1054 DebugLoc dl = N->getDebugLoc();
1056 SDValue MemAddr, Align;
1057 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1060 SDValue Chain = N->getOperand(0);
1061 EVT VT = N->getValueType(0);
1062 bool is64BitVector = VT.is64BitVector();
1064 unsigned OpcodeIndex;
1065 switch (VT.getSimpleVT().SimpleTy) {
1066 default: llvm_unreachable("unhandled vld type");
1067 // Double-register operations:
1068 case MVT::v8i8: OpcodeIndex = 0; break;
1069 case MVT::v4i16: OpcodeIndex = 1; break;
1071 case MVT::v2i32: OpcodeIndex = 2; break;
1072 case MVT::v1i64: OpcodeIndex = 3; break;
1073 // Quad-register operations:
1074 case MVT::v16i8: OpcodeIndex = 0; break;
1075 case MVT::v8i16: OpcodeIndex = 1; break;
1077 case MVT::v4i32: OpcodeIndex = 2; break;
1078 case MVT::v2i64: OpcodeIndex = 3;
1079 assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
1083 SDValue Pred = getAL(CurDAG);
1084 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1085 if (is64BitVector) {
1086 unsigned Opc = DOpcodes[OpcodeIndex];
1087 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1088 std::vector<EVT> ResTys(NumVecs, VT);
1089 ResTys.push_back(MVT::Other);
1090 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1091 if (!llvm::ModelWithRegSequence() || NumVecs < 2)
1095 SDValue V0 = SDValue(VLd, 0);
1096 SDValue V1 = SDValue(VLd, 1);
1098 // Form a REG_SEQUENCE to force register allocation.
1100 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1102 SDValue V2 = SDValue(VLd, 2);
1103 // If it's a vld3, form a quad D-register but discard the last part.
1104 SDValue V3 = (NumVecs == 3)
1105 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1107 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1110 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1111 SDValue D = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0+Vec,
1113 ReplaceUses(SDValue(N, Vec), D);
1115 ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, NumVecs));
1119 EVT RegVT = GetNEONSubregVT(VT);
1121 // Quad registers are directly supported for VLD1 and VLD2,
1122 // loading pairs of D regs.
1123 unsigned Opc = QOpcodes0[OpcodeIndex];
1124 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1125 std::vector<EVT> ResTys(2 * NumVecs, RegVT);
1126 ResTys.push_back(MVT::Other);
1127 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1128 Chain = SDValue(VLd, 2 * NumVecs);
1130 // Combine the even and odd subregs to produce the result.
1131 if (llvm::ModelWithRegSequence()) {
1133 SDNode *Q = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
1134 ReplaceUses(SDValue(N, 0), SDValue(Q, 0));
1136 SDValue QQ = SDValue(QuadDRegs(MVT::v4i64,
1137 SDValue(VLd, 0), SDValue(VLd, 1),
1138 SDValue(VLd, 2), SDValue(VLd, 3)), 0);
1139 SDValue Q0 = CurDAG->getTargetExtractSubreg(ARM::QSUBREG_0, dl, VT, QQ);
1140 SDValue Q1 = CurDAG->getTargetExtractSubreg(ARM::QSUBREG_1, dl, VT, QQ);
1141 ReplaceUses(SDValue(N, 0), Q0);
1142 ReplaceUses(SDValue(N, 1), Q1);
1145 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1146 SDNode *Q = PairDRegs(VT, SDValue(VLd, 2*Vec), SDValue(VLd, 2*Vec+1));
1147 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1151 // Otherwise, quad registers are loaded with two separate instructions,
1152 // where one loads the even registers and the other loads the odd registers.
1154 std::vector<EVT> ResTys(NumVecs, RegVT);
1155 ResTys.push_back(MemAddr.getValueType());
1156 ResTys.push_back(MVT::Other);
1158 // Load the even subregs.
1159 unsigned Opc = QOpcodes0[OpcodeIndex];
1160 const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain };
1161 SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6);
1162 Chain = SDValue(VLdA, NumVecs+1);
1164 // Load the odd subregs.
1165 Opc = QOpcodes1[OpcodeIndex];
1166 const SDValue OpsB[] = { SDValue(VLdA, NumVecs),
1167 Align, Reg0, Pred, Reg0, Chain };
1168 SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
1169 Chain = SDValue(VLdB, NumVecs+1);
1171 if (llvm::ModelWithRegSequence()) {
1172 SDValue V0 = SDValue(VLdA, 0);
1173 SDValue V1 = SDValue(VLdB, 0);
1174 SDValue V2 = SDValue(VLdA, 1);
1175 SDValue V3 = SDValue(VLdB, 1);
1176 SDValue V4 = SDValue(VLdA, 2);
1177 SDValue V5 = SDValue(VLdB, 2);
1178 SDValue V6 = (NumVecs == 3)
1179 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT),
1182 SDValue V7 = (NumVecs == 3)
1183 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT),
1186 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V0, V1, V2, V3,
1187 V4, V5, V6, V7), 0);
1189 // Extract out the 3 / 4 Q registers.
1190 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1191 SDValue Q = CurDAG->getTargetExtractSubreg(ARM::QSUBREG_0+Vec,
1193 ReplaceUses(SDValue(N, Vec), Q);
1196 // Combine the even and odd subregs to produce the result.
1197 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1198 SDNode *Q = PairDRegs(VT, SDValue(VLdA, Vec), SDValue(VLdB, Vec));
1199 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1203 ReplaceUses(SDValue(N, NumVecs), Chain);
1207 SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
1208 unsigned *DOpcodes, unsigned *QOpcodes0,
1209 unsigned *QOpcodes1) {
1210 assert(NumVecs >=1 && NumVecs <= 4 && "VST NumVecs out-of-range");
1211 DebugLoc dl = N->getDebugLoc();
1213 SDValue MemAddr, Align;
1214 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1217 SDValue Chain = N->getOperand(0);
1218 EVT VT = N->getOperand(3).getValueType();
1219 bool is64BitVector = VT.is64BitVector();
1221 unsigned OpcodeIndex;
1222 switch (VT.getSimpleVT().SimpleTy) {
1223 default: llvm_unreachable("unhandled vst type");
1224 // Double-register operations:
1225 case MVT::v8i8: OpcodeIndex = 0; break;
1226 case MVT::v4i16: OpcodeIndex = 1; break;
1228 case MVT::v2i32: OpcodeIndex = 2; break;
1229 case MVT::v1i64: OpcodeIndex = 3; break;
1230 // Quad-register operations:
1231 case MVT::v16i8: OpcodeIndex = 0; break;
1232 case MVT::v8i16: OpcodeIndex = 1; break;
1234 case MVT::v4i32: OpcodeIndex = 2; break;
1235 case MVT::v2i64: OpcodeIndex = 3;
1236 assert(NumVecs == 1 && "v2i64 type only supported for VST1");
1240 SDValue Pred = getAL(CurDAG);
1241 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1243 SmallVector<SDValue, 10> Ops;
1244 Ops.push_back(MemAddr);
1245 Ops.push_back(Align);
1247 if (is64BitVector) {
1248 if (llvm::ModelWithRegSequence() && NumVecs >= 2) {
1250 SDValue V0 = N->getOperand(0+3);
1251 SDValue V1 = N->getOperand(1+3);
1253 // Form a REG_SEQUENCE to force register allocation.
1255 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1257 SDValue V2 = N->getOperand(2+3);
1258 // If it's a vld3, form a quad D-register and leave the last part as
1260 SDValue V3 = (NumVecs == 3)
1261 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1262 : N->getOperand(3+3);
1263 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1266 // Now extract the D registers back out.
1267 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, VT,
1269 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, VT,
1272 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_2, dl, VT,
1275 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_3, dl, VT,
1278 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1279 Ops.push_back(N->getOperand(Vec+3));
1281 Ops.push_back(Pred);
1282 Ops.push_back(Reg0); // predicate register
1283 Ops.push_back(Chain);
1284 unsigned Opc = DOpcodes[OpcodeIndex];
1285 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+5);
1288 EVT RegVT = GetNEONSubregVT(VT);
1290 // Quad registers are directly supported for VST1 and VST2,
1291 // storing pairs of D regs.
1292 unsigned Opc = QOpcodes0[OpcodeIndex];
1293 if (llvm::ModelWithRegSequence() && NumVecs == 2) {
1294 // First extract the pair of Q registers.
1295 SDValue Q0 = N->getOperand(3);
1296 SDValue Q1 = N->getOperand(4);
1298 // Form a QQ register.
1299 SDValue QQ = SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0);
1301 // Now extract the D registers back out.
1302 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1304 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1306 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_2, dl, RegVT,
1308 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_3, dl, RegVT,
1310 Ops.push_back(Pred);
1311 Ops.push_back(Reg0); // predicate register
1312 Ops.push_back(Chain);
1313 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 5 + 4);
1315 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1316 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1317 N->getOperand(Vec+3)));
1318 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1319 N->getOperand(Vec+3)));
1321 Ops.push_back(Pred);
1322 Ops.push_back(Reg0); // predicate register
1323 Ops.push_back(Chain);
1324 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
1329 // Otherwise, quad registers are stored with two separate instructions,
1330 // where one stores the even registers and the other stores the odd registers.
1331 if (llvm::ModelWithRegSequence()) {
1332 // Form the QQQQ REG_SEQUENCE.
1334 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1335 V[i] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1336 N->getOperand(Vec+3));
1337 V[i+1] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1338 N->getOperand(Vec+3));
1341 V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1344 SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1345 V[4], V[5], V[6], V[7]), 0);
1347 // Store the even D registers.
1348 Ops.push_back(Reg0); // post-access address offset
1349 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1350 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0+Vec*2, dl,
1352 Ops.push_back(Pred);
1353 Ops.push_back(Reg0); // predicate register
1354 Ops.push_back(Chain);
1355 unsigned Opc = QOpcodes0[OpcodeIndex];
1356 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1357 MVT::Other, Ops.data(), NumVecs+6);
1358 Chain = SDValue(VStA, 1);
1360 // Store the odd D registers.
1361 Ops[0] = SDValue(VStA, 0); // MemAddr
1362 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1363 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1+Vec*2, dl,
1365 Ops[NumVecs+5] = Chain;
1366 Opc = QOpcodes1[OpcodeIndex];
1367 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1368 MVT::Other, Ops.data(), NumVecs+6);
1369 Chain = SDValue(VStB, 1);
1370 ReplaceUses(SDValue(N, 0), Chain);
1373 Ops.push_back(Reg0); // post-access address offset
1375 // Store the even subregs.
1376 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1377 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1378 N->getOperand(Vec+3)));
1379 Ops.push_back(Pred);
1380 Ops.push_back(Reg0); // predicate register
1381 Ops.push_back(Chain);
1382 unsigned Opc = QOpcodes0[OpcodeIndex];
1383 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1384 MVT::Other, Ops.data(), NumVecs+6);
1385 Chain = SDValue(VStA, 1);
1387 // Store the odd subregs.
1388 Ops[0] = SDValue(VStA, 0); // MemAddr
1389 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1390 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1391 N->getOperand(Vec+3));
1392 Ops[NumVecs+5] = Chain;
1393 Opc = QOpcodes1[OpcodeIndex];
1394 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1395 MVT::Other, Ops.data(), NumVecs+6);
1396 Chain = SDValue(VStB, 1);
1397 ReplaceUses(SDValue(N, 0), Chain);
1402 SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1403 unsigned NumVecs, unsigned *DOpcodes,
1404 unsigned *QOpcodes0,
1405 unsigned *QOpcodes1) {
1406 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1407 DebugLoc dl = N->getDebugLoc();
1409 SDValue MemAddr, Align;
1410 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1413 SDValue Chain = N->getOperand(0);
1415 cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1416 EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
1417 bool is64BitVector = VT.is64BitVector();
1419 // Quad registers are handled by load/store of subregs. Find the subreg info.
1420 unsigned NumElts = 0;
1424 if (!is64BitVector) {
1425 RegVT = GetNEONSubregVT(VT);
1426 NumElts = RegVT.getVectorNumElements();
1427 SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
1428 Even = Lane < NumElts;
1431 unsigned OpcodeIndex;
1432 switch (VT.getSimpleVT().SimpleTy) {
1433 default: llvm_unreachable("unhandled vld/vst lane type");
1434 // Double-register operations:
1435 case MVT::v8i8: OpcodeIndex = 0; break;
1436 case MVT::v4i16: OpcodeIndex = 1; break;
1438 case MVT::v2i32: OpcodeIndex = 2; break;
1439 // Quad-register operations:
1440 case MVT::v8i16: OpcodeIndex = 0; break;
1442 case MVT::v4i32: OpcodeIndex = 1; break;
1445 SDValue Pred = getAL(CurDAG);
1446 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1448 SmallVector<SDValue, 10> Ops;
1449 Ops.push_back(MemAddr);
1450 Ops.push_back(Align);
1453 if (is64BitVector) {
1454 Opc = DOpcodes[OpcodeIndex];
1455 if (llvm::ModelWithRegSequence()) {
1457 SDValue V0 = N->getOperand(0+3);
1458 SDValue V1 = N->getOperand(1+3);
1460 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1462 SDValue V2 = N->getOperand(2+3);
1463 SDValue V3 = (NumVecs == 3)
1464 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1465 : N->getOperand(3+3);
1466 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1469 // Now extract the D registers back out.
1470 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, VT,
1472 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, VT,
1475 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_2, dl, VT,
1478 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_3, dl, VT,
1481 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1482 Ops.push_back(N->getOperand(Vec+3));
1485 // Check if this is loading the even or odd subreg of a Q register.
1486 if (Lane < NumElts) {
1487 Opc = QOpcodes0[OpcodeIndex];
1490 Opc = QOpcodes1[OpcodeIndex];
1493 if (llvm::ModelWithRegSequence()) {
1495 SDValue V0 = N->getOperand(0+3);
1496 SDValue V1 = N->getOperand(1+3);
1498 RegSeq = SDValue(PairQRegs(MVT::v4i64, V0, V1), 0);
1500 SDValue V2 = N->getOperand(2+3);
1501 SDValue V3 = (NumVecs == 3)
1502 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1503 : N->getOperand(3+3);
1504 RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
1507 // Extract the subregs of the input vector.
1508 unsigned SubIdx = Even ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
1509 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1510 Ops.push_back(CurDAG->getTargetExtractSubreg(SubIdx+Vec*2, dl, RegVT,
1513 // Extract the subregs of the input vector.
1514 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1515 Ops.push_back(CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1516 N->getOperand(Vec+3)));
1519 Ops.push_back(getI32Imm(Lane));
1520 Ops.push_back(Pred);
1521 Ops.push_back(Reg0);
1522 Ops.push_back(Chain);
1525 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6);
1527 std::vector<EVT> ResTys(NumVecs, RegVT);
1528 ResTys.push_back(MVT::Other);
1529 SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(),NumVecs+6);
1531 if (llvm::ModelWithRegSequence()) {
1532 // Form a REG_SEQUENCE to force register allocation.
1534 if (is64BitVector) {
1535 SDValue V0 = SDValue(VLdLn, 0);
1536 SDValue V1 = SDValue(VLdLn, 1);
1538 RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1540 SDValue V2 = SDValue(VLdLn, 2);
1541 // If it's a vld3, form a quad D-register but discard the last part.
1542 SDValue V3 = (NumVecs == 3)
1543 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1544 : SDValue(VLdLn, 3);
1545 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1548 // For 128-bit vectors, take the 64-bit results of the load and insert them
1549 // as subregs into the result.
1551 for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1553 V[i] = SDValue(VLdLn, Vec);
1554 V[i+1] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1557 V[i] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1559 V[i+1] = SDValue(VLdLn, Vec);
1563 V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1567 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V[0], V[1], V[2], V[3]), 0);
1569 RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1570 V[4], V[5], V[6], V[7]), 0);
1573 unsigned SubIdx = is64BitVector ? ARM::DSUBREG_0 : ARM::QSUBREG_0;
1574 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1575 ReplaceUses(SDValue(N, Vec),
1576 CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, RegSeq));
1577 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, NumVecs));
1581 // For a 64-bit vector load to D registers, nothing more needs to be done.
1585 // For 128-bit vectors, take the 64-bit results of the load and insert them
1586 // as subregs into the result.
1587 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1588 SDValue QuadVec = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
1589 N->getOperand(Vec+3),
1590 SDValue(VLdLn, Vec));
1591 ReplaceUses(SDValue(N, Vec), QuadVec);
1594 Chain = SDValue(VLdLn, NumVecs);
1595 ReplaceUses(SDValue(N, NumVecs), Chain);
1599 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
1601 if (!Subtarget->hasV6T2Ops())
1604 unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
1605 : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
1608 // For unsigned extracts, check for a shift right and mask
1609 unsigned And_imm = 0;
1610 if (N->getOpcode() == ISD::AND) {
1611 if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
1613 // The immediate is a mask of the low bits iff imm & (imm+1) == 0
1614 if (And_imm & (And_imm + 1))
1617 unsigned Srl_imm = 0;
1618 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
1620 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1622 unsigned Width = CountTrailingOnes_32(And_imm);
1623 unsigned LSB = Srl_imm;
1624 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1625 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1626 CurDAG->getTargetConstant(LSB, MVT::i32),
1627 CurDAG->getTargetConstant(Width, MVT::i32),
1628 getAL(CurDAG), Reg0 };
1629 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1635 // Otherwise, we're looking for a shift of a shift
1636 unsigned Shl_imm = 0;
1637 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
1638 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1639 unsigned Srl_imm = 0;
1640 if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
1641 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1642 unsigned Width = 32 - Srl_imm;
1643 int LSB = Srl_imm - Shl_imm;
1646 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1647 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1648 CurDAG->getTargetConstant(LSB, MVT::i32),
1649 CurDAG->getTargetConstant(Width, MVT::i32),
1650 getAL(CurDAG), Reg0 };
1651 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1657 SDNode *ARMDAGToDAGISel::
1658 SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1659 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1662 if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
1663 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1664 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1667 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1668 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1669 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1670 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1672 llvm_unreachable("Unknown so_reg opcode!");
1676 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1677 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1678 SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
1679 return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
1684 SDNode *ARMDAGToDAGISel::
1685 SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1686 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1690 if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
1691 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1692 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
1693 return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
1698 SDNode *ARMDAGToDAGISel::
1699 SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1700 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1701 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1705 if (Predicate_t2_so_imm(TrueVal.getNode())) {
1706 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1707 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1708 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1709 return CurDAG->SelectNodeTo(N,
1710 ARM::t2MOVCCi, MVT::i32, Ops, 5);
1715 SDNode *ARMDAGToDAGISel::
1716 SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1717 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1718 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1722 if (Predicate_so_imm(TrueVal.getNode())) {
1723 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1724 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1725 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1726 return CurDAG->SelectNodeTo(N,
1727 ARM::MOVCCi, MVT::i32, Ops, 5);
1732 SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
1733 EVT VT = N->getValueType(0);
1734 SDValue FalseVal = N->getOperand(0);
1735 SDValue TrueVal = N->getOperand(1);
1736 SDValue CC = N->getOperand(2);
1737 SDValue CCR = N->getOperand(3);
1738 SDValue InFlag = N->getOperand(4);
1739 assert(CC.getOpcode() == ISD::Constant);
1740 assert(CCR.getOpcode() == ISD::Register);
1741 ARMCC::CondCodes CCVal =
1742 (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
1744 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1745 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1746 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1747 // Pattern complexity = 18 cost = 1 size = 0
1751 if (Subtarget->isThumb()) {
1752 SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
1753 CCVal, CCR, InFlag);
1755 Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
1756 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1760 SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
1761 CCVal, CCR, InFlag);
1763 Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
1764 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1769 // Pattern: (ARMcmov:i32 GPR:i32:$false,
1770 // (imm:i32)<<P:Predicate_so_imm>>:$true,
1772 // Emits: (MOVCCi:i32 GPR:i32:$false,
1773 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1774 // Pattern complexity = 10 cost = 1 size = 0
1775 if (Subtarget->isThumb()) {
1776 SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
1777 CCVal, CCR, InFlag);
1779 Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
1780 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1784 SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
1785 CCVal, CCR, InFlag);
1787 Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
1788 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1794 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1795 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1796 // Pattern complexity = 6 cost = 1 size = 0
1798 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1799 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1800 // Pattern complexity = 6 cost = 11 size = 0
1802 // Also FCPYScc and FCPYDcc.
1803 SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
1804 SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
1806 switch (VT.getSimpleVT().SimpleTy) {
1807 default: assert(false && "Illegal conditional move type!");
1810 Opc = Subtarget->isThumb()
1811 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1821 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1824 SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
1825 // The only time a CONCAT_VECTORS operation can have legal types is when
1826 // two 64-bit vectors are concatenated to a 128-bit vector.
1827 EVT VT = N->getValueType(0);
1828 if (!VT.is128BitVector() || N->getNumOperands() != 2)
1829 llvm_unreachable("unexpected CONCAT_VECTORS");
1830 DebugLoc dl = N->getDebugLoc();
1831 SDValue V0 = N->getOperand(0);
1832 SDValue V1 = N->getOperand(1);
1833 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
1834 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
1835 const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
1836 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
1839 SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
1840 DebugLoc dl = N->getDebugLoc();
1842 if (N->isMachineOpcode())
1843 return NULL; // Already selected.
1845 switch (N->getOpcode()) {
1847 case ISD::Constant: {
1848 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1850 if (Subtarget->hasThumb2())
1851 // Thumb2-aware targets have the MOVT instruction, so all immediates can
1852 // be done with MOV + MOVT, at worst.
1855 if (Subtarget->isThumb()) {
1856 UseCP = (Val > 255 && // MOV
1857 ~Val > 255 && // MOV + MVN
1858 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
1860 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
1861 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
1862 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
1867 CurDAG->getTargetConstantPool(ConstantInt::get(
1868 Type::getInt32Ty(*CurDAG->getContext()), Val),
1869 TLI.getPointerTy());
1872 if (Subtarget->isThumb1Only()) {
1873 SDValue Pred = getAL(CurDAG);
1874 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1875 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1876 ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1881 CurDAG->getRegister(0, MVT::i32),
1882 CurDAG->getTargetConstant(0, MVT::i32),
1884 CurDAG->getRegister(0, MVT::i32),
1885 CurDAG->getEntryNode()
1887 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1890 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
1894 // Other cases are autogenerated.
1897 case ISD::FrameIndex: {
1898 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1899 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1900 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1901 if (Subtarget->isThumb1Only()) {
1902 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1903 CurDAG->getTargetConstant(0, MVT::i32));
1905 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1906 ARM::t2ADDri : ARM::ADDri);
1907 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1908 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1909 CurDAG->getRegister(0, MVT::i32) };
1910 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1914 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1918 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true))
1922 if (Subtarget->isThumb1Only())
1924 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1925 unsigned RHSV = C->getZExtValue();
1927 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
1928 unsigned ShImm = Log2_32(RHSV-1);
1931 SDValue V = N->getOperand(0);
1932 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1933 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1934 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1935 if (Subtarget->isThumb()) {
1936 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1937 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1939 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1940 return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1943 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
1944 unsigned ShImm = Log2_32(RHSV+1);
1947 SDValue V = N->getOperand(0);
1948 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1949 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1950 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1951 if (Subtarget->isThumb()) {
1952 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0 };
1953 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 5);
1955 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1956 return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1962 // Check for unsigned bitfield extract
1963 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1966 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
1967 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
1968 // are entirely contributed by c2 and lower 16-bits are entirely contributed
1969 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
1970 // Select it to: "movt x, ((c1 & 0xffff) >> 16)
1971 EVT VT = N->getValueType(0);
1974 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
1976 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
1979 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
1980 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1983 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
1984 SDValue N2 = N0.getOperand(1);
1985 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1988 unsigned N1CVal = N1C->getZExtValue();
1989 unsigned N2CVal = N2C->getZExtValue();
1990 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
1991 (N1CVal & 0xffffU) == 0xffffU &&
1992 (N2CVal & 0xffffU) == 0x0U) {
1993 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
1995 SDValue Ops[] = { N0.getOperand(0), Imm16,
1996 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1997 return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
2002 case ARMISD::VMOVRRD:
2003 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
2004 N->getOperand(0), getAL(CurDAG),
2005 CurDAG->getRegister(0, MVT::i32));
2006 case ISD::UMUL_LOHI: {
2007 if (Subtarget->isThumb1Only())
2009 if (Subtarget->isThumb()) {
2010 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2011 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2012 CurDAG->getRegister(0, MVT::i32) };
2013 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops,4);
2015 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2016 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2017 CurDAG->getRegister(0, MVT::i32) };
2018 return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
2021 case ISD::SMUL_LOHI: {
2022 if (Subtarget->isThumb1Only())
2024 if (Subtarget->isThumb()) {
2025 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2026 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
2027 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops,4);
2029 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
2030 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
2031 CurDAG->getRegister(0, MVT::i32) };
2032 return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
2036 SDNode *ResNode = 0;
2037 if (Subtarget->isThumb() && Subtarget->hasThumb2())
2038 ResNode = SelectT2IndexedLoad(N);
2040 ResNode = SelectARMIndexedLoad(N);
2044 // VLDMQ must be custom-selected for "v2f64 load" to set the AM5Opc value.
2045 if (Subtarget->hasVFP2() &&
2046 N->getValueType(0).getSimpleVT().SimpleTy == MVT::v2f64) {
2047 SDValue Chain = N->getOperand(0);
2049 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
2050 SDValue Pred = getAL(CurDAG);
2051 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2052 SDValue Ops[] = { N->getOperand(1), AM5Opc, Pred, PredReg, Chain };
2053 return CurDAG->getMachineNode(ARM::VLDMQ, dl, MVT::v2f64, MVT::Other,
2056 // Other cases are autogenerated.
2060 // VSTMQ must be custom-selected for "v2f64 store" to set the AM5Opc value.
2061 if (Subtarget->hasVFP2() &&
2062 N->getOperand(1).getValueType().getSimpleVT().SimpleTy == MVT::v2f64) {
2063 SDValue Chain = N->getOperand(0);
2065 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
2066 SDValue Pred = getAL(CurDAG);
2067 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2068 SDValue Ops[] = { N->getOperand(1), N->getOperand(2),
2069 AM5Opc, Pred, PredReg, Chain };
2070 return CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
2072 // Other cases are autogenerated.
2075 case ARMISD::BRCOND: {
2076 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2077 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2078 // Pattern complexity = 6 cost = 1 size = 0
2080 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2081 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
2082 // Pattern complexity = 6 cost = 1 size = 0
2084 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2085 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2086 // Pattern complexity = 6 cost = 1 size = 0
2088 unsigned Opc = Subtarget->isThumb() ?
2089 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
2090 SDValue Chain = N->getOperand(0);
2091 SDValue N1 = N->getOperand(1);
2092 SDValue N2 = N->getOperand(2);
2093 SDValue N3 = N->getOperand(3);
2094 SDValue InFlag = N->getOperand(4);
2095 assert(N1.getOpcode() == ISD::BasicBlock);
2096 assert(N2.getOpcode() == ISD::Constant);
2097 assert(N3.getOpcode() == ISD::Register);
2099 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2100 cast<ConstantSDNode>(N2)->getZExtValue()),
2102 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
2103 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
2105 Chain = SDValue(ResNode, 0);
2106 if (N->getNumValues() == 2) {
2107 InFlag = SDValue(ResNode, 1);
2108 ReplaceUses(SDValue(N, 1), InFlag);
2110 ReplaceUses(SDValue(N, 0),
2111 SDValue(Chain.getNode(), Chain.getResNo()));
2115 return SelectCMOVOp(N);
2116 case ARMISD::CNEG: {
2117 EVT VT = N->getValueType(0);
2118 SDValue N0 = N->getOperand(0);
2119 SDValue N1 = N->getOperand(1);
2120 SDValue N2 = N->getOperand(2);
2121 SDValue N3 = N->getOperand(3);
2122 SDValue InFlag = N->getOperand(4);
2123 assert(N2.getOpcode() == ISD::Constant);
2124 assert(N3.getOpcode() == ISD::Register);
2126 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2127 cast<ConstantSDNode>(N2)->getZExtValue()),
2129 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
2131 switch (VT.getSimpleVT().SimpleTy) {
2132 default: assert(false && "Illegal conditional move type!");
2141 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
2144 case ARMISD::VZIP: {
2146 EVT VT = N->getValueType(0);
2147 switch (VT.getSimpleVT().SimpleTy) {
2148 default: return NULL;
2149 case MVT::v8i8: Opc = ARM::VZIPd8; break;
2150 case MVT::v4i16: Opc = ARM::VZIPd16; break;
2152 case MVT::v2i32: Opc = ARM::VZIPd32; break;
2153 case MVT::v16i8: Opc = ARM::VZIPq8; break;
2154 case MVT::v8i16: Opc = ARM::VZIPq16; break;
2156 case MVT::v4i32: Opc = ARM::VZIPq32; break;
2158 SDValue Pred = getAL(CurDAG);
2159 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2160 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2161 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2163 case ARMISD::VUZP: {
2165 EVT VT = N->getValueType(0);
2166 switch (VT.getSimpleVT().SimpleTy) {
2167 default: return NULL;
2168 case MVT::v8i8: Opc = ARM::VUZPd8; break;
2169 case MVT::v4i16: Opc = ARM::VUZPd16; break;
2171 case MVT::v2i32: Opc = ARM::VUZPd32; break;
2172 case MVT::v16i8: Opc = ARM::VUZPq8; break;
2173 case MVT::v8i16: Opc = ARM::VUZPq16; break;
2175 case MVT::v4i32: Opc = ARM::VUZPq32; break;
2177 SDValue Pred = getAL(CurDAG);
2178 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2179 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2180 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2182 case ARMISD::VTRN: {
2184 EVT VT = N->getValueType(0);
2185 switch (VT.getSimpleVT().SimpleTy) {
2186 default: return NULL;
2187 case MVT::v8i8: Opc = ARM::VTRNd8; break;
2188 case MVT::v4i16: Opc = ARM::VTRNd16; break;
2190 case MVT::v2i32: Opc = ARM::VTRNd32; break;
2191 case MVT::v16i8: Opc = ARM::VTRNq8; break;
2192 case MVT::v8i16: Opc = ARM::VTRNq16; break;
2194 case MVT::v4i32: Opc = ARM::VTRNq32; break;
2196 SDValue Pred = getAL(CurDAG);
2197 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2198 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2199 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2202 case ISD::INTRINSIC_VOID:
2203 case ISD::INTRINSIC_W_CHAIN: {
2204 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2209 case Intrinsic::arm_neon_vld1: {
2210 unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
2211 ARM::VLD1d32, ARM::VLD1d64 };
2212 unsigned QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
2213 ARM::VLD1q32, ARM::VLD1q64 };
2214 return SelectVLD(N, 1, DOpcodes, QOpcodes, 0);
2217 case Intrinsic::arm_neon_vld2: {
2218 unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
2219 ARM::VLD2d32, ARM::VLD1q64 };
2220 unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
2221 return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
2224 case Intrinsic::arm_neon_vld3: {
2225 unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
2226 ARM::VLD3d32, ARM::VLD1d64T };
2227 unsigned QOpcodes0[] = { ARM::VLD3q8_UPD,
2230 unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD,
2231 ARM::VLD3q16odd_UPD,
2232 ARM::VLD3q32odd_UPD };
2233 return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2236 case Intrinsic::arm_neon_vld4: {
2237 unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
2238 ARM::VLD4d32, ARM::VLD1d64Q };
2239 unsigned QOpcodes0[] = { ARM::VLD4q8_UPD,
2242 unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD,
2243 ARM::VLD4q16odd_UPD,
2244 ARM::VLD4q32odd_UPD };
2245 return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2248 case Intrinsic::arm_neon_vld2lane: {
2249 unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
2250 unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 };
2251 unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd };
2252 return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
2255 case Intrinsic::arm_neon_vld3lane: {
2256 unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
2257 unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 };
2258 unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd };
2259 return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
2262 case Intrinsic::arm_neon_vld4lane: {
2263 unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
2264 unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 };
2265 unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd };
2266 return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
2269 case Intrinsic::arm_neon_vst1: {
2270 unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
2271 ARM::VST1d32, ARM::VST1d64 };
2272 unsigned QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
2273 ARM::VST1q32, ARM::VST1q64 };
2274 return SelectVST(N, 1, DOpcodes, QOpcodes, 0);
2277 case Intrinsic::arm_neon_vst2: {
2278 unsigned DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
2279 ARM::VST2d32, ARM::VST1q64 };
2280 unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
2281 return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
2284 case Intrinsic::arm_neon_vst3: {
2285 unsigned DOpcodes[] = { ARM::VST3d8, ARM::VST3d16,
2286 ARM::VST3d32, ARM::VST1d64T };
2287 unsigned QOpcodes0[] = { ARM::VST3q8_UPD,
2290 unsigned QOpcodes1[] = { ARM::VST3q8odd_UPD,
2291 ARM::VST3q16odd_UPD,
2292 ARM::VST3q32odd_UPD };
2293 return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2296 case Intrinsic::arm_neon_vst4: {
2297 unsigned DOpcodes[] = { ARM::VST4d8, ARM::VST4d16,
2298 ARM::VST4d32, ARM::VST1d64Q };
2299 unsigned QOpcodes0[] = { ARM::VST4q8_UPD,
2302 unsigned QOpcodes1[] = { ARM::VST4q8odd_UPD,
2303 ARM::VST4q16odd_UPD,
2304 ARM::VST4q32odd_UPD };
2305 return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2308 case Intrinsic::arm_neon_vst2lane: {
2309 unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
2310 unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 };
2311 unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd };
2312 return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
2315 case Intrinsic::arm_neon_vst3lane: {
2316 unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
2317 unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 };
2318 unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd };
2319 return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
2322 case Intrinsic::arm_neon_vst4lane: {
2323 unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
2324 unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 };
2325 unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd };
2326 return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
2332 case ISD::CONCAT_VECTORS:
2333 return SelectConcatVector(N);
2336 return SelectCode(N);
2339 bool ARMDAGToDAGISel::
2340 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2341 std::vector<SDValue> &OutOps) {
2342 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
2343 // Require the address to be in a register. That is safe for all ARM
2344 // variants and it is hard to do anything much smarter without knowing
2345 // how the operand is used.
2346 OutOps.push_back(Op);
2350 /// createARMISelDag - This pass converts a legalized DAG into a
2351 /// ARM-specific DAG, ready for instruction scheduling.
2353 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
2354 CodeGenOpt::Level OptLevel) {
2355 return new ARMDAGToDAGISel(TM, OptLevel);
2358 /// ModelWithRegSequence - Return true if isel should use REG_SEQUENCE to model
2359 /// operations involving sub-registers.
2360 bool llvm::ModelWithRegSequence() {