1 //===-- ARMISelDAGToDAG.cpp - A dag to dag inst selector for ARM ----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines an instruction selector for the ARM target.
12 //===----------------------------------------------------------------------===//
15 #include "ARMAddressingModes.h"
16 #include "ARMConstantPoolValue.h"
17 #include "ARMISelLowering.h"
18 #include "ARMTargetMachine.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/Intrinsics.h"
24 #include "llvm/LLVMContext.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/SelectionDAGISel.h"
30 #include "llvm/Target/TargetLowering.h"
31 #include "llvm/Target/TargetOptions.h"
32 #include "llvm/Support/Compiler.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
39 //===--------------------------------------------------------------------===//
40 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
41 /// instructions for SelectionDAG operations.
44 class ARMDAGToDAGISel : public SelectionDAGISel {
45 ARMBaseTargetMachine &TM;
47 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
48 /// make the right decision when generating code for different targets.
49 const ARMSubtarget *Subtarget;
52 explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
53 CodeGenOpt::Level OptLevel)
54 : SelectionDAGISel(tm, OptLevel), TM(tm),
55 Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
58 virtual const char *getPassName() const {
59 return "ARM Instruction Selection";
62 /// getI32Imm - Return a target constant of type i32 with the specified
64 inline SDValue getI32Imm(unsigned Imm) {
65 return CurDAG->getTargetConstant(Imm, MVT::i32);
68 SDNode *Select(SDValue Op);
69 virtual void InstructionSelect();
70 bool SelectShifterOperandReg(SDValue Op, SDValue N, SDValue &A,
71 SDValue &B, SDValue &C);
72 bool SelectAddrMode2(SDValue Op, SDValue N, SDValue &Base,
73 SDValue &Offset, SDValue &Opc);
74 bool SelectAddrMode2Offset(SDValue Op, SDValue N,
75 SDValue &Offset, SDValue &Opc);
76 bool SelectAddrMode3(SDValue Op, SDValue N, SDValue &Base,
77 SDValue &Offset, SDValue &Opc);
78 bool SelectAddrMode3Offset(SDValue Op, SDValue N,
79 SDValue &Offset, SDValue &Opc);
80 bool SelectAddrMode4(SDValue Op, SDValue N, SDValue &Addr,
82 bool SelectAddrMode5(SDValue Op, SDValue N, SDValue &Base,
84 bool SelectAddrMode6(SDValue Op, SDValue N, SDValue &Addr, SDValue &Update,
87 bool SelectAddrModePC(SDValue Op, SDValue N, SDValue &Offset,
90 bool SelectThumbAddrModeRR(SDValue Op, SDValue N, SDValue &Base,
92 bool SelectThumbAddrModeRI5(SDValue Op, SDValue N, unsigned Scale,
93 SDValue &Base, SDValue &OffImm,
95 bool SelectThumbAddrModeS1(SDValue Op, SDValue N, SDValue &Base,
96 SDValue &OffImm, SDValue &Offset);
97 bool SelectThumbAddrModeS2(SDValue Op, SDValue N, SDValue &Base,
98 SDValue &OffImm, SDValue &Offset);
99 bool SelectThumbAddrModeS4(SDValue Op, SDValue N, SDValue &Base,
100 SDValue &OffImm, SDValue &Offset);
101 bool SelectThumbAddrModeSP(SDValue Op, SDValue N, SDValue &Base,
104 bool SelectT2ShifterOperandReg(SDValue Op, SDValue N,
105 SDValue &BaseReg, SDValue &Opc);
106 bool SelectT2AddrModeImm12(SDValue Op, SDValue N, SDValue &Base,
108 bool SelectT2AddrModeImm8(SDValue Op, SDValue N, SDValue &Base,
110 bool SelectT2AddrModeImm8Offset(SDValue Op, SDValue N,
112 bool SelectT2AddrModeImm8s4(SDValue Op, SDValue N, SDValue &Base,
114 bool SelectT2AddrModeSoReg(SDValue Op, SDValue N, SDValue &Base,
115 SDValue &OffReg, SDValue &ShImm);
117 // Include the pieces autogenerated from the target description.
118 #include "ARMGenDAGISel.inc"
121 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
123 SDNode *SelectARMIndexedLoad(SDValue Op);
124 SDNode *SelectT2IndexedLoad(SDValue Op);
126 /// SelectDYN_ALLOC - Select dynamic alloc for Thumb.
127 SDNode *SelectDYN_ALLOC(SDValue Op);
129 /// SelectVLDLane - Select NEON load structure to one lane. NumVecs should
130 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
131 /// loading D registers and even subregs and odd subregs of Q registers.
132 SDNode *SelectVLDLane(SDValue Op, unsigned NumVecs,
133 unsigned *DOpcodes, unsigned *QOpcodes0,
134 unsigned *QOpcodes1);
136 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
137 SDNode *SelectV6T2BitfieldExtractOp(SDValue Op, unsigned Opc);
139 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
140 /// inline asm expressions.
141 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
143 std::vector<SDValue> &OutOps);
145 /// PairDRegs - Insert a pair of double registers into an implicit def to
146 /// form a quad register.
147 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
151 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
152 /// operand. If so Imm will receive the 32-bit value.
153 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
154 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
155 Imm = cast<ConstantSDNode>(N)->getZExtValue();
161 // isInt32Immediate - This method tests to see if a constant operand.
162 // If so Imm will receive the 32 bit value.
163 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
164 return isInt32Immediate(N.getNode(), Imm);
167 // isOpcWithIntImmediate - This method tests to see if the node is a specific
168 // opcode and that it has a immediate integer right operand.
169 // If so Imm will receive the 32 bit value.
170 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
171 return N->getOpcode() == Opc &&
172 isInt32Immediate(N->getOperand(1).getNode(), Imm);
176 void ARMDAGToDAGISel::InstructionSelect() {
180 CurDAG->RemoveDeadNodes();
183 bool ARMDAGToDAGISel::SelectShifterOperandReg(SDValue Op,
188 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
190 // Don't match base register only case. That is matched to a separate
191 // lower complexity pattern with explicit register operand.
192 if (ShOpcVal == ARM_AM::no_shift) return false;
194 BaseReg = N.getOperand(0);
195 unsigned ShImmVal = 0;
196 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
197 ShReg = CurDAG->getRegister(0, MVT::i32);
198 ShImmVal = RHS->getZExtValue() & 31;
200 ShReg = N.getOperand(1);
202 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
207 bool ARMDAGToDAGISel::SelectAddrMode2(SDValue Op, SDValue N,
208 SDValue &Base, SDValue &Offset,
210 if (N.getOpcode() == ISD::MUL) {
211 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
212 // X * [3,5,9] -> X + X * [2,4,8] etc.
213 int RHSC = (int)RHS->getZExtValue();
216 ARM_AM::AddrOpc AddSub = ARM_AM::add;
218 AddSub = ARM_AM::sub;
221 if (isPowerOf2_32(RHSC)) {
222 unsigned ShAmt = Log2_32(RHSC);
223 Base = Offset = N.getOperand(0);
224 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
233 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
235 if (N.getOpcode() == ISD::FrameIndex) {
236 int FI = cast<FrameIndexSDNode>(N)->getIndex();
237 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
238 } else if (N.getOpcode() == ARMISD::Wrapper) {
239 Base = N.getOperand(0);
241 Offset = CurDAG->getRegister(0, MVT::i32);
242 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
248 // Match simple R +/- imm12 operands.
249 if (N.getOpcode() == ISD::ADD)
250 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
251 int RHSC = (int)RHS->getZExtValue();
252 if ((RHSC >= 0 && RHSC < 0x1000) ||
253 (RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
254 Base = N.getOperand(0);
255 if (Base.getOpcode() == ISD::FrameIndex) {
256 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
257 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
259 Offset = CurDAG->getRegister(0, MVT::i32);
261 ARM_AM::AddrOpc AddSub = ARM_AM::add;
263 AddSub = ARM_AM::sub;
266 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
273 // Otherwise this is R +/- [possibly shifted] R
274 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
275 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
278 Base = N.getOperand(0);
279 Offset = N.getOperand(1);
281 if (ShOpcVal != ARM_AM::no_shift) {
282 // Check to see if the RHS of the shift is a constant, if not, we can't fold
284 if (ConstantSDNode *Sh =
285 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
286 ShAmt = Sh->getZExtValue();
287 Offset = N.getOperand(1).getOperand(0);
289 ShOpcVal = ARM_AM::no_shift;
293 // Try matching (R shl C) + (R).
294 if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
295 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
296 if (ShOpcVal != ARM_AM::no_shift) {
297 // Check to see if the RHS of the shift is a constant, if not, we can't
299 if (ConstantSDNode *Sh =
300 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
301 ShAmt = Sh->getZExtValue();
302 Offset = N.getOperand(0).getOperand(0);
303 Base = N.getOperand(1);
305 ShOpcVal = ARM_AM::no_shift;
310 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
315 bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDValue Op, SDValue N,
316 SDValue &Offset, SDValue &Opc) {
317 unsigned Opcode = Op.getOpcode();
318 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
319 ? cast<LoadSDNode>(Op)->getAddressingMode()
320 : cast<StoreSDNode>(Op)->getAddressingMode();
321 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
322 ? ARM_AM::add : ARM_AM::sub;
323 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
324 int Val = (int)C->getZExtValue();
325 if (Val >= 0 && Val < 0x1000) { // 12 bits.
326 Offset = CurDAG->getRegister(0, MVT::i32);
327 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
335 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
337 if (ShOpcVal != ARM_AM::no_shift) {
338 // Check to see if the RHS of the shift is a constant, if not, we can't fold
340 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
341 ShAmt = Sh->getZExtValue();
342 Offset = N.getOperand(0);
344 ShOpcVal = ARM_AM::no_shift;
348 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
354 bool ARMDAGToDAGISel::SelectAddrMode3(SDValue Op, SDValue N,
355 SDValue &Base, SDValue &Offset,
357 if (N.getOpcode() == ISD::SUB) {
358 // X - C is canonicalize to X + -C, no need to handle it here.
359 Base = N.getOperand(0);
360 Offset = N.getOperand(1);
361 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
365 if (N.getOpcode() != ISD::ADD) {
367 if (N.getOpcode() == ISD::FrameIndex) {
368 int FI = cast<FrameIndexSDNode>(N)->getIndex();
369 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
371 Offset = CurDAG->getRegister(0, MVT::i32);
372 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
376 // If the RHS is +/- imm8, fold into addr mode.
377 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
378 int RHSC = (int)RHS->getZExtValue();
379 if ((RHSC >= 0 && RHSC < 256) ||
380 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
381 Base = N.getOperand(0);
382 if (Base.getOpcode() == ISD::FrameIndex) {
383 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
384 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
386 Offset = CurDAG->getRegister(0, MVT::i32);
388 ARM_AM::AddrOpc AddSub = ARM_AM::add;
390 AddSub = ARM_AM::sub;
393 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
398 Base = N.getOperand(0);
399 Offset = N.getOperand(1);
400 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
404 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDValue Op, SDValue N,
405 SDValue &Offset, SDValue &Opc) {
406 unsigned Opcode = Op.getOpcode();
407 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
408 ? cast<LoadSDNode>(Op)->getAddressingMode()
409 : cast<StoreSDNode>(Op)->getAddressingMode();
410 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
411 ? ARM_AM::add : ARM_AM::sub;
412 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
413 int Val = (int)C->getZExtValue();
414 if (Val >= 0 && Val < 256) {
415 Offset = CurDAG->getRegister(0, MVT::i32);
416 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
422 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
426 bool ARMDAGToDAGISel::SelectAddrMode4(SDValue Op, SDValue N,
427 SDValue &Addr, SDValue &Mode) {
429 Mode = CurDAG->getTargetConstant(0, MVT::i32);
433 bool ARMDAGToDAGISel::SelectAddrMode5(SDValue Op, SDValue N,
434 SDValue &Base, SDValue &Offset) {
435 if (N.getOpcode() != ISD::ADD) {
437 if (N.getOpcode() == ISD::FrameIndex) {
438 int FI = cast<FrameIndexSDNode>(N)->getIndex();
439 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
440 } else if (N.getOpcode() == ARMISD::Wrapper) {
441 Base = N.getOperand(0);
443 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
448 // If the RHS is +/- imm8, fold into addr mode.
449 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
450 int RHSC = (int)RHS->getZExtValue();
451 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied by 4.
453 if ((RHSC >= 0 && RHSC < 256) ||
454 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
455 Base = N.getOperand(0);
456 if (Base.getOpcode() == ISD::FrameIndex) {
457 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
458 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
461 ARM_AM::AddrOpc AddSub = ARM_AM::add;
463 AddSub = ARM_AM::sub;
466 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
474 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
479 bool ARMDAGToDAGISel::SelectAddrMode6(SDValue Op, SDValue N,
480 SDValue &Addr, SDValue &Update,
483 // Default to no writeback.
484 Update = CurDAG->getRegister(0, MVT::i32);
485 Opc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(false), MVT::i32);
489 bool ARMDAGToDAGISel::SelectAddrModePC(SDValue Op, SDValue N,
490 SDValue &Offset, SDValue &Label) {
491 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
492 Offset = N.getOperand(0);
493 SDValue N1 = N.getOperand(1);
494 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
501 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue Op, SDValue N,
502 SDValue &Base, SDValue &Offset){
503 // FIXME dl should come from the parent load or store, not the address
504 DebugLoc dl = Op.getDebugLoc();
505 if (N.getOpcode() != ISD::ADD) {
506 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
507 if (!NC || NC->getZExtValue() != 0)
514 Base = N.getOperand(0);
515 Offset = N.getOperand(1);
520 ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDValue Op, SDValue N,
521 unsigned Scale, SDValue &Base,
522 SDValue &OffImm, SDValue &Offset) {
524 SDValue TmpBase, TmpOffImm;
525 if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
526 return false; // We want to select tLDRspi / tSTRspi instead.
527 if (N.getOpcode() == ARMISD::Wrapper &&
528 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
529 return false; // We want to select tLDRpci instead.
532 if (N.getOpcode() != ISD::ADD) {
533 Base = (N.getOpcode() == ARMISD::Wrapper) ? N.getOperand(0) : N;
534 Offset = CurDAG->getRegister(0, MVT::i32);
535 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
539 // Thumb does not have [sp, r] address mode.
540 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
541 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
542 if ((LHSR && LHSR->getReg() == ARM::SP) ||
543 (RHSR && RHSR->getReg() == ARM::SP)) {
545 Offset = CurDAG->getRegister(0, MVT::i32);
546 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
550 // If the RHS is + imm5 * scale, fold into addr mode.
551 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
552 int RHSC = (int)RHS->getZExtValue();
553 if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
555 if (RHSC >= 0 && RHSC < 32) {
556 Base = N.getOperand(0);
557 Offset = CurDAG->getRegister(0, MVT::i32);
558 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
564 Base = N.getOperand(0);
565 Offset = N.getOperand(1);
566 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
570 bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDValue Op, SDValue N,
571 SDValue &Base, SDValue &OffImm,
573 return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
576 bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDValue Op, SDValue N,
577 SDValue &Base, SDValue &OffImm,
579 return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
582 bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDValue Op, SDValue N,
583 SDValue &Base, SDValue &OffImm,
585 return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
588 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDValue Op, SDValue N,
589 SDValue &Base, SDValue &OffImm) {
590 if (N.getOpcode() == ISD::FrameIndex) {
591 int FI = cast<FrameIndexSDNode>(N)->getIndex();
592 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
593 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
597 if (N.getOpcode() != ISD::ADD)
600 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
601 if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
602 (LHSR && LHSR->getReg() == ARM::SP)) {
603 // If the RHS is + imm8 * scale, fold into addr mode.
604 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
605 int RHSC = (int)RHS->getZExtValue();
606 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
608 if (RHSC >= 0 && RHSC < 256) {
609 Base = N.getOperand(0);
610 if (Base.getOpcode() == ISD::FrameIndex) {
611 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
612 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
614 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
624 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDValue Op, SDValue N,
627 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
629 // Don't match base register only case. That is matched to a separate
630 // lower complexity pattern with explicit register operand.
631 if (ShOpcVal == ARM_AM::no_shift) return false;
633 BaseReg = N.getOperand(0);
634 unsigned ShImmVal = 0;
635 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
636 ShImmVal = RHS->getZExtValue() & 31;
637 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
644 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDValue Op, SDValue N,
645 SDValue &Base, SDValue &OffImm) {
646 // Match simple R + imm12 operands.
649 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
650 if (N.getOpcode() == ISD::FrameIndex) {
651 // Match frame index...
652 int FI = cast<FrameIndexSDNode>(N)->getIndex();
653 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
654 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
656 } else if (N.getOpcode() == ARMISD::Wrapper) {
657 Base = N.getOperand(0);
658 if (Base.getOpcode() == ISD::TargetConstantPool)
659 return false; // We want to select t2LDRpci instead.
662 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
666 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
667 if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
668 // Let t2LDRi8 handle (R - imm8).
671 int RHSC = (int)RHS->getZExtValue();
672 if (N.getOpcode() == ISD::SUB)
675 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
676 Base = N.getOperand(0);
677 if (Base.getOpcode() == ISD::FrameIndex) {
678 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
679 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
681 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
688 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
692 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue Op, SDValue N,
693 SDValue &Base, SDValue &OffImm) {
694 // Match simple R - imm8 operands.
695 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
696 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
697 int RHSC = (int)RHS->getSExtValue();
698 if (N.getOpcode() == ISD::SUB)
701 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
702 Base = N.getOperand(0);
703 if (Base.getOpcode() == ISD::FrameIndex) {
704 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
705 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
707 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
716 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDValue Op, SDValue N,
718 unsigned Opcode = Op.getOpcode();
719 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
720 ? cast<LoadSDNode>(Op)->getAddressingMode()
721 : cast<StoreSDNode>(Op)->getAddressingMode();
722 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
723 int RHSC = (int)RHS->getZExtValue();
724 if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
725 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
726 ? CurDAG->getTargetConstant(RHSC, MVT::i32)
727 : CurDAG->getTargetConstant(-RHSC, MVT::i32);
735 bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDValue Op, SDValue N,
736 SDValue &Base, SDValue &OffImm) {
737 if (N.getOpcode() == ISD::ADD) {
738 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
739 int RHSC = (int)RHS->getZExtValue();
740 if (((RHSC & 0x3) == 0) &&
741 ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits.
742 Base = N.getOperand(0);
743 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
747 } else if (N.getOpcode() == ISD::SUB) {
748 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
749 int RHSC = (int)RHS->getZExtValue();
750 if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) { // 8 bits.
751 Base = N.getOperand(0);
752 OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
761 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue Op, SDValue N,
763 SDValue &OffReg, SDValue &ShImm) {
764 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
765 if (N.getOpcode() != ISD::ADD)
768 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
769 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
770 int RHSC = (int)RHS->getZExtValue();
771 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
773 else if (RHSC < 0 && RHSC >= -255) // 8 bits
777 // Look for (R + R) or (R + (R << [1,2,3])).
779 Base = N.getOperand(0);
780 OffReg = N.getOperand(1);
782 // Swap if it is ((R << c) + R).
783 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
784 if (ShOpcVal != ARM_AM::lsl) {
785 ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
786 if (ShOpcVal == ARM_AM::lsl)
787 std::swap(Base, OffReg);
790 if (ShOpcVal == ARM_AM::lsl) {
791 // Check to see if the RHS of the shift is a constant, if not, we can't fold
793 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
794 ShAmt = Sh->getZExtValue();
797 ShOpcVal = ARM_AM::no_shift;
799 OffReg = OffReg.getOperand(0);
801 ShOpcVal = ARM_AM::no_shift;
805 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
810 //===--------------------------------------------------------------------===//
812 /// getAL - Returns a ARMCC::AL immediate node.
813 static inline SDValue getAL(SelectionDAG *CurDAG) {
814 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
817 SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDValue Op) {
818 LoadSDNode *LD = cast<LoadSDNode>(Op);
819 ISD::MemIndexedMode AM = LD->getAddressingMode();
820 if (AM == ISD::UNINDEXED)
823 EVT LoadedVT = LD->getMemoryVT();
824 SDValue Offset, AMOpc;
825 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
828 if (LoadedVT == MVT::i32 &&
829 SelectAddrMode2Offset(Op, LD->getOffset(), Offset, AMOpc)) {
830 Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
832 } else if (LoadedVT == MVT::i16 &&
833 SelectAddrMode3Offset(Op, LD->getOffset(), Offset, AMOpc)) {
835 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
836 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
837 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
838 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
839 if (LD->getExtensionType() == ISD::SEXTLOAD) {
840 if (SelectAddrMode3Offset(Op, LD->getOffset(), Offset, AMOpc)) {
842 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
845 if (SelectAddrMode2Offset(Op, LD->getOffset(), Offset, AMOpc)) {
847 Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
853 SDValue Chain = LD->getChain();
854 SDValue Base = LD->getBasePtr();
855 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
856 CurDAG->getRegister(0, MVT::i32), Chain };
857 return CurDAG->getMachineNode(Opcode, Op.getDebugLoc(), MVT::i32, MVT::i32,
864 SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDValue Op) {
865 LoadSDNode *LD = cast<LoadSDNode>(Op);
866 ISD::MemIndexedMode AM = LD->getAddressingMode();
867 if (AM == ISD::UNINDEXED)
870 EVT LoadedVT = LD->getMemoryVT();
871 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
873 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
876 if (SelectT2AddrModeImm8Offset(Op, LD->getOffset(), Offset)) {
877 switch (LoadedVT.getSimpleVT().SimpleTy) {
879 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
883 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
885 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
890 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
892 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
901 SDValue Chain = LD->getChain();
902 SDValue Base = LD->getBasePtr();
903 SDValue Ops[]= { Base, Offset, getAL(CurDAG),
904 CurDAG->getRegister(0, MVT::i32), Chain };
905 return CurDAG->getMachineNode(Opcode, Op.getDebugLoc(), MVT::i32, MVT::i32,
912 SDNode *ARMDAGToDAGISel::SelectDYN_ALLOC(SDValue Op) {
913 SDNode *N = Op.getNode();
914 DebugLoc dl = N->getDebugLoc();
915 EVT VT = Op.getValueType();
916 SDValue Chain = Op.getOperand(0);
917 SDValue Size = Op.getOperand(1);
918 SDValue Align = Op.getOperand(2);
919 SDValue SP = CurDAG->getRegister(ARM::SP, MVT::i32);
920 int32_t AlignVal = cast<ConstantSDNode>(Align)->getSExtValue();
922 // We need to align the stack. Use Thumb1 tAND which is the only thumb
923 // instruction that can read and write SP. This matches to a pseudo
924 // instruction that has a chain to ensure the result is written back to
925 // the stack pointer.
926 SP = SDValue(CurDAG->getMachineNode(ARM::tANDsp, dl, VT, SP, Align), 0);
928 bool isC = isa<ConstantSDNode>(Size);
929 uint32_t C = isC ? cast<ConstantSDNode>(Size)->getZExtValue() : ~0UL;
930 // Handle the most common case for both Thumb1 and Thumb2:
931 // tSUBspi - immediate is between 0 ... 508 inclusive.
932 if (C <= 508 && ((C & 3) == 0))
933 // FIXME: tSUBspi encode scale 4 implicitly.
934 return CurDAG->SelectNodeTo(N, ARM::tSUBspi_, VT, MVT::Other, SP,
935 CurDAG->getTargetConstant(C/4, MVT::i32),
938 if (Subtarget->isThumb1Only()) {
939 // Use tADDspr since Thumb1 does not have a sub r, sp, r. ARMISelLowering
940 // should have negated the size operand already. FIXME: We can't insert
941 // new target independent node at this stage so we are forced to negate
942 // it earlier. Is there a better solution?
943 return CurDAG->SelectNodeTo(N, ARM::tADDspr_, VT, MVT::Other, SP, Size,
945 } else if (Subtarget->isThumb2()) {
946 if (isC && Predicate_t2_so_imm(Size.getNode())) {
948 SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
949 return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi_, VT, MVT::Other, Ops, 3);
950 } else if (isC && Predicate_imm0_4095(Size.getNode())) {
952 SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
953 return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi12_, VT, MVT::Other, Ops, 3);
956 SDValue Ops[] = { SP, Size,
957 getI32Imm(ARM_AM::getSORegOpc(ARM_AM::lsl,0)), Chain };
958 return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPs_, VT, MVT::Other, Ops, 4);
962 // FIXME: Add ADD / SUB sp instructions for ARM.
966 /// PairDRegs - Insert a pair of double registers into an implicit def to
967 /// form a quad register.
968 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
969 DebugLoc dl = V0.getNode()->getDebugLoc();
971 SDValue(CurDAG->getMachineNode(TargetInstrInfo::IMPLICIT_DEF, dl, VT), 0);
972 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
973 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
974 SDNode *Pair = CurDAG->getMachineNode(TargetInstrInfo::INSERT_SUBREG, dl,
975 VT, Undef, V0, SubReg0);
976 return CurDAG->getMachineNode(TargetInstrInfo::INSERT_SUBREG, dl,
977 VT, SDValue(Pair, 0), V1, SubReg1);
980 /// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
981 /// for a 64-bit subregister of the vector.
982 static EVT GetNEONSubregVT(EVT VT) {
983 switch (VT.getSimpleVT().SimpleTy) {
984 default: llvm_unreachable("unhandled NEON type");
985 case MVT::v16i8: return MVT::v8i8;
986 case MVT::v8i16: return MVT::v4i16;
987 case MVT::v4f32: return MVT::v2f32;
988 case MVT::v4i32: return MVT::v2i32;
989 case MVT::v2i64: return MVT::v1i64;
993 SDNode *ARMDAGToDAGISel::SelectVLDLane(SDValue Op, unsigned NumVecs,
994 unsigned *DOpcodes, unsigned *QOpcodes0,
995 unsigned *QOpcodes1) {
996 assert(NumVecs >=2 && NumVecs <= 4 && "VLDLane NumVecs out-of-range");
997 SDNode *N = Op.getNode();
998 DebugLoc dl = N->getDebugLoc();
1000 SDValue MemAddr, MemUpdate, MemOpc;
1001 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1004 SDValue Chain = N->getOperand(0);
1006 cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1007 EVT VT = N->getValueType(0);
1008 bool is64BitVector = VT.is64BitVector();
1010 // Quad registers are handled by extracting subregs, doing the load,
1011 // and then inserting the results as subregs. Find the subreg info.
1012 unsigned NumElts = 0;
1015 if (!is64BitVector) {
1016 RegVT = GetNEONSubregVT(VT);
1017 NumElts = RegVT.getVectorNumElements();
1018 SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
1021 unsigned OpcodeIndex;
1022 switch (VT.getSimpleVT().SimpleTy) {
1023 default: llvm_unreachable("unhandled vld lane type");
1024 // Double-register operations:
1025 case MVT::v8i8: OpcodeIndex = 0; break;
1026 case MVT::v4i16: OpcodeIndex = 1; break;
1028 case MVT::v2i32: OpcodeIndex = 2; break;
1029 // Quad-register operations:
1030 case MVT::v8i16: OpcodeIndex = 0; break;
1032 case MVT::v4i32: OpcodeIndex = 1; break;
1035 SmallVector<SDValue, 9> Ops;
1036 Ops.push_back(MemAddr);
1037 Ops.push_back(MemUpdate);
1038 Ops.push_back(MemOpc);
1041 if (is64BitVector) {
1042 Opc = DOpcodes[OpcodeIndex];
1043 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1044 Ops.push_back(N->getOperand(Vec+3));
1046 // Check if this is loading the even or odd subreg of a Q register.
1047 if (Lane < NumElts) {
1048 Opc = QOpcodes0[OpcodeIndex];
1051 Opc = QOpcodes1[OpcodeIndex];
1053 // Extract the subregs of the input vector.
1054 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1055 Ops.push_back(CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1056 N->getOperand(Vec+3)));
1058 Ops.push_back(getI32Imm(Lane));
1059 Ops.push_back(Chain);
1061 std::vector<EVT> ResTys(NumVecs, RegVT);
1062 ResTys.push_back(MVT::Other);
1064 CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), NumVecs+5);
1065 // For a 64-bit vector load to D registers, nothing more needs to be done.
1069 // For 128-bit vectors, take the 64-bit results of the load and insert them
1070 // as subregs into the result.
1071 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1072 SDValue QuadVec = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
1073 N->getOperand(Vec+3),
1074 SDValue(VLdLn, Vec));
1075 ReplaceUses(SDValue(N, Vec), QuadVec);
1078 Chain = SDValue(VLdLn, NumVecs);
1079 ReplaceUses(SDValue(N, NumVecs), Chain);
1083 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDValue Op,
1085 if (!Subtarget->hasV6T2Ops())
1088 unsigned Shl_imm = 0;
1089 if (isOpcWithIntImmediate(Op.getOperand(0).getNode(), ISD::SHL, Shl_imm)){
1090 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1091 unsigned Srl_imm = 0;
1092 if (isInt32Immediate(Op.getOperand(1), Srl_imm)) {
1093 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1094 unsigned Width = 32 - Srl_imm;
1095 int LSB = Srl_imm - Shl_imm;
1096 if ((LSB + Width) > 32)
1098 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1099 SDValue Ops[] = { Op.getOperand(0).getOperand(0),
1100 CurDAG->getTargetConstant(LSB, MVT::i32),
1101 CurDAG->getTargetConstant(Width, MVT::i32),
1102 getAL(CurDAG), Reg0 };
1103 return CurDAG->SelectNodeTo(Op.getNode(), Opc, MVT::i32, Ops, 5);
1109 SDNode *ARMDAGToDAGISel::Select(SDValue Op) {
1110 SDNode *N = Op.getNode();
1111 DebugLoc dl = N->getDebugLoc();
1113 if (N->isMachineOpcode())
1114 return NULL; // Already selected.
1116 switch (N->getOpcode()) {
1118 case ISD::Constant: {
1119 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1121 if (Subtarget->hasThumb2())
1122 // Thumb2-aware targets have the MOVT instruction, so all immediates can
1123 // be done with MOV + MOVT, at worst.
1126 if (Subtarget->isThumb()) {
1127 UseCP = (Val > 255 && // MOV
1128 ~Val > 255 && // MOV + MVN
1129 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
1131 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
1132 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
1133 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
1138 CurDAG->getTargetConstantPool(ConstantInt::get(
1139 Type::getInt32Ty(*CurDAG->getContext()), Val),
1140 TLI.getPointerTy());
1143 if (Subtarget->isThumb1Only()) {
1144 SDValue Pred = CurDAG->getTargetConstant(0xEULL, MVT::i32);
1145 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1146 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1147 ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1152 CurDAG->getRegister(0, MVT::i32),
1153 CurDAG->getTargetConstant(0, MVT::i32),
1155 CurDAG->getRegister(0, MVT::i32),
1156 CurDAG->getEntryNode()
1158 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1161 ReplaceUses(Op, SDValue(ResNode, 0));
1165 // Other cases are autogenerated.
1168 case ISD::FrameIndex: {
1169 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1170 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1171 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1172 if (Subtarget->isThumb1Only()) {
1173 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1174 CurDAG->getTargetConstant(0, MVT::i32));
1176 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1177 ARM::t2ADDri : ARM::ADDri);
1178 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1179 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1180 CurDAG->getRegister(0, MVT::i32) };
1181 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1184 case ARMISD::DYN_ALLOC:
1185 return SelectDYN_ALLOC(Op);
1187 if (SDNode *I = SelectV6T2BitfieldExtractOp(Op,
1188 Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX))
1192 if (SDNode *I = SelectV6T2BitfieldExtractOp(Op,
1193 Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX))
1197 if (Subtarget->isThumb1Only())
1199 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1200 unsigned RHSV = C->getZExtValue();
1202 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
1203 unsigned ShImm = Log2_32(RHSV-1);
1206 SDValue V = Op.getOperand(0);
1207 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1208 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1209 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1210 if (Subtarget->isThumb()) {
1211 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1212 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1214 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1215 return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1218 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
1219 unsigned ShImm = Log2_32(RHSV+1);
1222 SDValue V = Op.getOperand(0);
1223 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1224 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1225 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1226 if (Subtarget->isThumb()) {
1227 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0 };
1228 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 5);
1230 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1231 return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1237 return CurDAG->getMachineNode(ARM::FMRRD, dl, MVT::i32, MVT::i32,
1238 Op.getOperand(0), getAL(CurDAG),
1239 CurDAG->getRegister(0, MVT::i32));
1240 case ISD::UMUL_LOHI: {
1241 if (Subtarget->isThumb1Only())
1243 if (Subtarget->isThumb()) {
1244 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
1245 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1246 CurDAG->getRegister(0, MVT::i32) };
1247 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops,4);
1249 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
1250 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1251 CurDAG->getRegister(0, MVT::i32) };
1252 return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1255 case ISD::SMUL_LOHI: {
1256 if (Subtarget->isThumb1Only())
1258 if (Subtarget->isThumb()) {
1259 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
1260 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1261 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops,4);
1263 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
1264 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1265 CurDAG->getRegister(0, MVT::i32) };
1266 return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1270 SDNode *ResNode = 0;
1271 if (Subtarget->isThumb() && Subtarget->hasThumb2())
1272 ResNode = SelectT2IndexedLoad(Op);
1274 ResNode = SelectARMIndexedLoad(Op);
1277 // Other cases are autogenerated.
1280 case ARMISD::BRCOND: {
1281 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1282 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
1283 // Pattern complexity = 6 cost = 1 size = 0
1285 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1286 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
1287 // Pattern complexity = 6 cost = 1 size = 0
1289 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1290 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
1291 // Pattern complexity = 6 cost = 1 size = 0
1293 unsigned Opc = Subtarget->isThumb() ?
1294 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
1295 SDValue Chain = Op.getOperand(0);
1296 SDValue N1 = Op.getOperand(1);
1297 SDValue N2 = Op.getOperand(2);
1298 SDValue N3 = Op.getOperand(3);
1299 SDValue InFlag = Op.getOperand(4);
1300 assert(N1.getOpcode() == ISD::BasicBlock);
1301 assert(N2.getOpcode() == ISD::Constant);
1302 assert(N3.getOpcode() == ISD::Register);
1304 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1305 cast<ConstantSDNode>(N2)->getZExtValue()),
1307 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
1308 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
1310 Chain = SDValue(ResNode, 0);
1311 if (Op.getNode()->getNumValues() == 2) {
1312 InFlag = SDValue(ResNode, 1);
1313 ReplaceUses(SDValue(Op.getNode(), 1), InFlag);
1315 ReplaceUses(SDValue(Op.getNode(), 0), SDValue(Chain.getNode(), Chain.getResNo()));
1318 case ARMISD::CMOV: {
1319 EVT VT = Op.getValueType();
1320 SDValue N0 = Op.getOperand(0);
1321 SDValue N1 = Op.getOperand(1);
1322 SDValue N2 = Op.getOperand(2);
1323 SDValue N3 = Op.getOperand(3);
1324 SDValue InFlag = Op.getOperand(4);
1325 assert(N2.getOpcode() == ISD::Constant);
1326 assert(N3.getOpcode() == ISD::Register);
1328 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1329 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1330 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1331 // Pattern complexity = 18 cost = 1 size = 0
1335 if (Subtarget->isThumb()) {
1336 if (SelectT2ShifterOperandReg(Op, N1, CPTmp0, CPTmp1)) {
1337 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1338 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1341 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1342 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1343 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1344 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1346 llvm_unreachable("Unknown so_reg opcode!");
1350 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1351 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1352 cast<ConstantSDNode>(N2)->getZExtValue()),
1354 SDValue Ops[] = { N0, CPTmp0, SOShImm, Tmp2, N3, InFlag };
1355 return CurDAG->SelectNodeTo(Op.getNode(), Opc, MVT::i32,Ops, 6);
1358 if (SelectShifterOperandReg(Op, N1, CPTmp0, CPTmp1, CPTmp2)) {
1359 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1360 cast<ConstantSDNode>(N2)->getZExtValue()),
1362 SDValue Ops[] = { N0, CPTmp0, CPTmp1, CPTmp2, Tmp2, N3, InFlag };
1363 return CurDAG->SelectNodeTo(Op.getNode(),
1364 ARM::MOVCCs, MVT::i32, Ops, 7);
1368 // Pattern: (ARMcmov:i32 GPR:i32:$false,
1369 // (imm:i32)<<P:Predicate_so_imm>>:$true,
1371 // Emits: (MOVCCi:i32 GPR:i32:$false,
1372 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1373 // Pattern complexity = 10 cost = 1 size = 0
1374 if (N3.getOpcode() == ISD::Constant) {
1375 if (Subtarget->isThumb()) {
1376 if (Predicate_t2_so_imm(N3.getNode())) {
1377 SDValue Tmp1 = CurDAG->getTargetConstant(((unsigned)
1378 cast<ConstantSDNode>(N1)->getZExtValue()),
1380 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1381 cast<ConstantSDNode>(N2)->getZExtValue()),
1383 SDValue Ops[] = { N0, Tmp1, Tmp2, N3, InFlag };
1384 return CurDAG->SelectNodeTo(Op.getNode(),
1385 ARM::t2MOVCCi, MVT::i32, Ops, 5);
1388 if (Predicate_so_imm(N3.getNode())) {
1389 SDValue Tmp1 = CurDAG->getTargetConstant(((unsigned)
1390 cast<ConstantSDNode>(N1)->getZExtValue()),
1392 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1393 cast<ConstantSDNode>(N2)->getZExtValue()),
1395 SDValue Ops[] = { N0, Tmp1, Tmp2, N3, InFlag };
1396 return CurDAG->SelectNodeTo(Op.getNode(),
1397 ARM::MOVCCi, MVT::i32, Ops, 5);
1403 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1404 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1405 // Pattern complexity = 6 cost = 1 size = 0
1407 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1408 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1409 // Pattern complexity = 6 cost = 11 size = 0
1411 // Also FCPYScc and FCPYDcc.
1412 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1413 cast<ConstantSDNode>(N2)->getZExtValue()),
1415 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
1417 switch (VT.getSimpleVT().SimpleTy) {
1418 default: assert(false && "Illegal conditional move type!");
1421 Opc = Subtarget->isThumb()
1422 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1432 return CurDAG->SelectNodeTo(Op.getNode(), Opc, VT, Ops, 5);
1434 case ARMISD::CNEG: {
1435 EVT VT = Op.getValueType();
1436 SDValue N0 = Op.getOperand(0);
1437 SDValue N1 = Op.getOperand(1);
1438 SDValue N2 = Op.getOperand(2);
1439 SDValue N3 = Op.getOperand(3);
1440 SDValue InFlag = Op.getOperand(4);
1441 assert(N2.getOpcode() == ISD::Constant);
1442 assert(N3.getOpcode() == ISD::Register);
1444 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1445 cast<ConstantSDNode>(N2)->getZExtValue()),
1447 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
1449 switch (VT.getSimpleVT().SimpleTy) {
1450 default: assert(false && "Illegal conditional move type!");
1459 return CurDAG->SelectNodeTo(Op.getNode(), Opc, VT, Ops, 5);
1462 case ARMISD::VZIP: {
1464 EVT VT = N->getValueType(0);
1465 switch (VT.getSimpleVT().SimpleTy) {
1466 default: return NULL;
1467 case MVT::v8i8: Opc = ARM::VZIPd8; break;
1468 case MVT::v4i16: Opc = ARM::VZIPd16; break;
1470 case MVT::v2i32: Opc = ARM::VZIPd32; break;
1471 case MVT::v16i8: Opc = ARM::VZIPq8; break;
1472 case MVT::v8i16: Opc = ARM::VZIPq16; break;
1474 case MVT::v4i32: Opc = ARM::VZIPq32; break;
1476 return CurDAG->getMachineNode(Opc, dl, VT, VT,
1477 N->getOperand(0), N->getOperand(1));
1479 case ARMISD::VUZP: {
1481 EVT VT = N->getValueType(0);
1482 switch (VT.getSimpleVT().SimpleTy) {
1483 default: return NULL;
1484 case MVT::v8i8: Opc = ARM::VUZPd8; break;
1485 case MVT::v4i16: Opc = ARM::VUZPd16; break;
1487 case MVT::v2i32: Opc = ARM::VUZPd32; break;
1488 case MVT::v16i8: Opc = ARM::VUZPq8; break;
1489 case MVT::v8i16: Opc = ARM::VUZPq16; break;
1491 case MVT::v4i32: Opc = ARM::VUZPq32; break;
1493 return CurDAG->getMachineNode(Opc, dl, VT, VT,
1494 N->getOperand(0), N->getOperand(1));
1496 case ARMISD::VTRN: {
1498 EVT VT = N->getValueType(0);
1499 switch (VT.getSimpleVT().SimpleTy) {
1500 default: return NULL;
1501 case MVT::v8i8: Opc = ARM::VTRNd8; break;
1502 case MVT::v4i16: Opc = ARM::VTRNd16; break;
1504 case MVT::v2i32: Opc = ARM::VTRNd32; break;
1505 case MVT::v16i8: Opc = ARM::VTRNq8; break;
1506 case MVT::v8i16: Opc = ARM::VTRNq16; break;
1508 case MVT::v4i32: Opc = ARM::VTRNq32; break;
1510 return CurDAG->getMachineNode(Opc, dl, VT, VT,
1511 N->getOperand(0), N->getOperand(1));
1514 case ISD::INTRINSIC_VOID:
1515 case ISD::INTRINSIC_W_CHAIN: {
1516 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
1517 EVT VT = N->getValueType(0);
1524 case Intrinsic::arm_neon_vld2: {
1525 SDValue MemAddr, MemUpdate, MemOpc;
1526 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1528 SDValue Chain = N->getOperand(0);
1529 if (VT.is64BitVector()) {
1530 switch (VT.getSimpleVT().SimpleTy) {
1531 default: llvm_unreachable("unhandled vld2 type");
1532 case MVT::v8i8: Opc = ARM::VLD2d8; break;
1533 case MVT::v4i16: Opc = ARM::VLD2d16; break;
1535 case MVT::v2i32: Opc = ARM::VLD2d32; break;
1536 case MVT::v1i64: Opc = ARM::VLD2d64; break;
1538 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
1539 return CurDAG->getMachineNode(Opc, dl, VT, VT, MVT::Other, Ops, 4);
1541 // Quad registers are loaded as pairs of double registers.
1543 switch (VT.getSimpleVT().SimpleTy) {
1544 default: llvm_unreachable("unhandled vld2 type");
1545 case MVT::v16i8: Opc = ARM::VLD2q8; RegVT = MVT::v8i8; break;
1546 case MVT::v8i16: Opc = ARM::VLD2q16; RegVT = MVT::v4i16; break;
1547 case MVT::v4f32: Opc = ARM::VLD2q32; RegVT = MVT::v2f32; break;
1548 case MVT::v4i32: Opc = ARM::VLD2q32; RegVT = MVT::v2i32; break;
1550 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
1551 std::vector<EVT> ResTys(4, RegVT);
1552 ResTys.push_back(MVT::Other);
1553 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 4);
1554 SDNode *Q0 = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
1555 SDNode *Q1 = PairDRegs(VT, SDValue(VLd, 2), SDValue(VLd, 3));
1556 ReplaceUses(SDValue(N, 0), SDValue(Q0, 0));
1557 ReplaceUses(SDValue(N, 1), SDValue(Q1, 0));
1558 ReplaceUses(SDValue(N, 2), SDValue(VLd, 4));
1562 case Intrinsic::arm_neon_vld3: {
1563 SDValue MemAddr, MemUpdate, MemOpc;
1564 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1566 SDValue Chain = N->getOperand(0);
1567 if (VT.is64BitVector()) {
1568 switch (VT.getSimpleVT().SimpleTy) {
1569 default: llvm_unreachable("unhandled vld3 type");
1570 case MVT::v8i8: Opc = ARM::VLD3d8; break;
1571 case MVT::v4i16: Opc = ARM::VLD3d16; break;
1573 case MVT::v2i32: Opc = ARM::VLD3d32; break;
1574 case MVT::v1i64: Opc = ARM::VLD3d64; break;
1576 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
1577 return CurDAG->getMachineNode(Opc, dl, VT, VT, VT, MVT::Other, Ops, 4);
1579 // Quad registers are loaded with two separate instructions, where one
1580 // loads the even registers and the other loads the odd registers.
1583 switch (VT.getSimpleVT().SimpleTy) {
1584 default: llvm_unreachable("unhandled vld3 type");
1586 Opc = ARM::VLD3q8a; Opc2 = ARM::VLD3q8b; RegVT = MVT::v8i8; break;
1588 Opc = ARM::VLD3q16a; Opc2 = ARM::VLD3q16b; RegVT = MVT::v4i16; break;
1590 Opc = ARM::VLD3q32a; Opc2 = ARM::VLD3q32b; RegVT = MVT::v2f32; break;
1592 Opc = ARM::VLD3q32a; Opc2 = ARM::VLD3q32b; RegVT = MVT::v2i32; break;
1594 // Enable writeback to the address register.
1595 MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
1597 std::vector<EVT> ResTys(3, RegVT);
1598 ResTys.push_back(MemAddr.getValueType());
1599 ResTys.push_back(MVT::Other);
1601 const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, Chain };
1602 SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 4);
1603 Chain = SDValue(VLdA, 4);
1605 const SDValue OpsB[] = { SDValue(VLdA, 3), MemUpdate, MemOpc, Chain };
1606 SDNode *VLdB = CurDAG->getMachineNode(Opc2, dl, ResTys, OpsB, 4);
1607 Chain = SDValue(VLdB, 4);
1609 SDNode *Q0 = PairDRegs(VT, SDValue(VLdA, 0), SDValue(VLdB, 0));
1610 SDNode *Q1 = PairDRegs(VT, SDValue(VLdA, 1), SDValue(VLdB, 1));
1611 SDNode *Q2 = PairDRegs(VT, SDValue(VLdA, 2), SDValue(VLdB, 2));
1612 ReplaceUses(SDValue(N, 0), SDValue(Q0, 0));
1613 ReplaceUses(SDValue(N, 1), SDValue(Q1, 0));
1614 ReplaceUses(SDValue(N, 2), SDValue(Q2, 0));
1615 ReplaceUses(SDValue(N, 3), Chain);
1619 case Intrinsic::arm_neon_vld4: {
1620 SDValue MemAddr, MemUpdate, MemOpc;
1621 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1623 SDValue Chain = N->getOperand(0);
1624 if (VT.is64BitVector()) {
1625 switch (VT.getSimpleVT().SimpleTy) {
1626 default: llvm_unreachable("unhandled vld4 type");
1627 case MVT::v8i8: Opc = ARM::VLD4d8; break;
1628 case MVT::v4i16: Opc = ARM::VLD4d16; break;
1630 case MVT::v2i32: Opc = ARM::VLD4d32; break;
1631 case MVT::v1i64: Opc = ARM::VLD4d64; break;
1633 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
1634 std::vector<EVT> ResTys(4, VT);
1635 ResTys.push_back(MVT::Other);
1636 return CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 4);
1638 // Quad registers are loaded with two separate instructions, where one
1639 // loads the even registers and the other loads the odd registers.
1642 switch (VT.getSimpleVT().SimpleTy) {
1643 default: llvm_unreachable("unhandled vld4 type");
1645 Opc = ARM::VLD4q8a; Opc2 = ARM::VLD4q8b; RegVT = MVT::v8i8; break;
1647 Opc = ARM::VLD4q16a; Opc2 = ARM::VLD4q16b; RegVT = MVT::v4i16; break;
1649 Opc = ARM::VLD4q32a; Opc2 = ARM::VLD4q32b; RegVT = MVT::v2f32; break;
1651 Opc = ARM::VLD4q32a; Opc2 = ARM::VLD4q32b; RegVT = MVT::v2i32; break;
1653 // Enable writeback to the address register.
1654 MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
1656 std::vector<EVT> ResTys(4, RegVT);
1657 ResTys.push_back(MemAddr.getValueType());
1658 ResTys.push_back(MVT::Other);
1660 const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, Chain };
1661 SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 4);
1662 Chain = SDValue(VLdA, 5);
1664 const SDValue OpsB[] = { SDValue(VLdA, 4), MemUpdate, MemOpc, Chain };
1665 SDNode *VLdB = CurDAG->getMachineNode(Opc2, dl, ResTys, OpsB, 4);
1666 Chain = SDValue(VLdB, 5);
1668 SDNode *Q0 = PairDRegs(VT, SDValue(VLdA, 0), SDValue(VLdB, 0));
1669 SDNode *Q1 = PairDRegs(VT, SDValue(VLdA, 1), SDValue(VLdB, 1));
1670 SDNode *Q2 = PairDRegs(VT, SDValue(VLdA, 2), SDValue(VLdB, 2));
1671 SDNode *Q3 = PairDRegs(VT, SDValue(VLdA, 3), SDValue(VLdB, 3));
1672 ReplaceUses(SDValue(N, 0), SDValue(Q0, 0));
1673 ReplaceUses(SDValue(N, 1), SDValue(Q1, 0));
1674 ReplaceUses(SDValue(N, 2), SDValue(Q2, 0));
1675 ReplaceUses(SDValue(N, 3), SDValue(Q3, 0));
1676 ReplaceUses(SDValue(N, 4), Chain);
1680 case Intrinsic::arm_neon_vld2lane: {
1681 unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
1682 unsigned QOpcodes0[] = { ARM::VLD2LNq16a, ARM::VLD2LNq32a };
1683 unsigned QOpcodes1[] = { ARM::VLD2LNq16b, ARM::VLD2LNq32b };
1684 return SelectVLDLane(Op, 2, DOpcodes, QOpcodes0, QOpcodes1);
1687 case Intrinsic::arm_neon_vld3lane: {
1688 unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
1689 unsigned QOpcodes0[] = { ARM::VLD3LNq16a, ARM::VLD3LNq32a };
1690 unsigned QOpcodes1[] = { ARM::VLD3LNq16b, ARM::VLD3LNq32b };
1691 return SelectVLDLane(Op, 3, DOpcodes, QOpcodes0, QOpcodes1);
1694 case Intrinsic::arm_neon_vld4lane: {
1695 unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
1696 unsigned QOpcodes0[] = { ARM::VLD4LNq16a, ARM::VLD4LNq32a };
1697 unsigned QOpcodes1[] = { ARM::VLD4LNq16b, ARM::VLD4LNq32b };
1698 return SelectVLDLane(Op, 4, DOpcodes, QOpcodes0, QOpcodes1);
1701 case Intrinsic::arm_neon_vst2: {
1702 SDValue MemAddr, MemUpdate, MemOpc;
1703 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1705 SDValue Chain = N->getOperand(0);
1706 VT = N->getOperand(3).getValueType();
1707 if (VT.is64BitVector()) {
1708 switch (VT.getSimpleVT().SimpleTy) {
1709 default: llvm_unreachable("unhandled vst2 type");
1710 case MVT::v8i8: Opc = ARM::VST2d8; break;
1711 case MVT::v4i16: Opc = ARM::VST2d16; break;
1713 case MVT::v2i32: Opc = ARM::VST2d32; break;
1714 case MVT::v1i64: Opc = ARM::VST2d64; break;
1716 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
1717 N->getOperand(3), N->getOperand(4), Chain };
1718 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 6);
1720 // Quad registers are stored as pairs of double registers.
1722 switch (VT.getSimpleVT().SimpleTy) {
1723 default: llvm_unreachable("unhandled vst2 type");
1724 case MVT::v16i8: Opc = ARM::VST2q8; RegVT = MVT::v8i8; break;
1725 case MVT::v8i16: Opc = ARM::VST2q16; RegVT = MVT::v4i16; break;
1726 case MVT::v4f32: Opc = ARM::VST2q32; RegVT = MVT::v2f32; break;
1727 case MVT::v4i32: Opc = ARM::VST2q32; RegVT = MVT::v2i32; break;
1729 SDValue D0 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1731 SDValue D1 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1733 SDValue D2 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1735 SDValue D3 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1737 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
1738 D0, D1, D2, D3, Chain };
1739 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
1742 case Intrinsic::arm_neon_vst3: {
1743 SDValue MemAddr, MemUpdate, MemOpc;
1744 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1746 SDValue Chain = N->getOperand(0);
1747 VT = N->getOperand(3).getValueType();
1748 if (VT.is64BitVector()) {
1749 switch (VT.getSimpleVT().SimpleTy) {
1750 default: llvm_unreachable("unhandled vst3 type");
1751 case MVT::v8i8: Opc = ARM::VST3d8; break;
1752 case MVT::v4i16: Opc = ARM::VST3d16; break;
1754 case MVT::v2i32: Opc = ARM::VST3d32; break;
1755 case MVT::v1i64: Opc = ARM::VST3d64; break;
1757 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
1758 N->getOperand(3), N->getOperand(4),
1759 N->getOperand(5), Chain };
1760 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7);
1762 // Quad registers are stored with two separate instructions, where one
1763 // stores the even registers and the other stores the odd registers.
1766 switch (VT.getSimpleVT().SimpleTy) {
1767 default: llvm_unreachable("unhandled vst3 type");
1769 Opc = ARM::VST3q8a; Opc2 = ARM::VST3q8b; RegVT = MVT::v8i8; break;
1771 Opc = ARM::VST3q16a; Opc2 = ARM::VST3q16b; RegVT = MVT::v4i16; break;
1773 Opc = ARM::VST3q32a; Opc2 = ARM::VST3q32b; RegVT = MVT::v2f32; break;
1775 Opc = ARM::VST3q32a; Opc2 = ARM::VST3q32b; RegVT = MVT::v2i32; break;
1777 // Enable writeback to the address register.
1778 MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
1780 SDValue D0 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1782 SDValue D2 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1784 SDValue D4 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1786 const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, D0, D2, D4, Chain };
1787 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1788 MVT::Other, OpsA, 7);
1789 Chain = SDValue(VStA, 1);
1791 SDValue D1 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1793 SDValue D3 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1795 SDValue D5 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1797 MemAddr = SDValue(VStA, 0);
1798 const SDValue OpsB[] = { MemAddr, MemUpdate, MemOpc, D1, D3, D5, Chain };
1799 SDNode *VStB = CurDAG->getMachineNode(Opc2, dl, MemAddr.getValueType(),
1800 MVT::Other, OpsB, 7);
1801 Chain = SDValue(VStB, 1);
1802 ReplaceUses(SDValue(N, 0), Chain);
1806 case Intrinsic::arm_neon_vst4: {
1807 SDValue MemAddr, MemUpdate, MemOpc;
1808 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1810 SDValue Chain = N->getOperand(0);
1811 VT = N->getOperand(3).getValueType();
1812 if (VT.is64BitVector()) {
1813 switch (VT.getSimpleVT().SimpleTy) {
1814 default: llvm_unreachable("unhandled vst4 type");
1815 case MVT::v8i8: Opc = ARM::VST4d8; break;
1816 case MVT::v4i16: Opc = ARM::VST4d16; break;
1818 case MVT::v2i32: Opc = ARM::VST4d32; break;
1819 case MVT::v1i64: Opc = ARM::VST4d64; break;
1821 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
1822 N->getOperand(3), N->getOperand(4),
1823 N->getOperand(5), N->getOperand(6), Chain };
1824 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
1826 // Quad registers are stored with two separate instructions, where one
1827 // stores the even registers and the other stores the odd registers.
1830 switch (VT.getSimpleVT().SimpleTy) {
1831 default: llvm_unreachable("unhandled vst4 type");
1833 Opc = ARM::VST4q8a; Opc2 = ARM::VST4q8b; RegVT = MVT::v8i8; break;
1835 Opc = ARM::VST4q16a; Opc2 = ARM::VST4q16b; RegVT = MVT::v4i16; break;
1837 Opc = ARM::VST4q32a; Opc2 = ARM::VST4q32b; RegVT = MVT::v2f32; break;
1839 Opc = ARM::VST4q32a; Opc2 = ARM::VST4q32b; RegVT = MVT::v2i32; break;
1841 // Enable writeback to the address register.
1842 MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
1844 SDValue D0 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1846 SDValue D2 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1848 SDValue D4 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1850 SDValue D6 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1852 const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc,
1853 D0, D2, D4, D6, Chain };
1854 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1855 MVT::Other, OpsA, 8);
1856 Chain = SDValue(VStA, 1);
1858 SDValue D1 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1860 SDValue D3 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1862 SDValue D5 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1864 SDValue D7 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1866 MemAddr = SDValue(VStA, 0);
1867 const SDValue OpsB[] = { MemAddr, MemUpdate, MemOpc,
1868 D1, D3, D5, D7, Chain };
1869 SDNode *VStB = CurDAG->getMachineNode(Opc2, dl, MemAddr.getValueType(),
1870 MVT::Other, OpsB, 8);
1871 Chain = SDValue(VStB, 1);
1872 ReplaceUses(SDValue(N, 0), Chain);
1876 case Intrinsic::arm_neon_vst2lane: {
1877 SDValue MemAddr, MemUpdate, MemOpc;
1878 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1880 SDValue Chain = N->getOperand(0);
1881 unsigned Lane = cast<ConstantSDNode>(N->getOperand(5))->getZExtValue();
1882 VT = N->getOperand(3).getValueType();
1883 if (VT.is64BitVector()) {
1884 switch (VT.getSimpleVT().SimpleTy) {
1885 default: llvm_unreachable("unhandled vst2lane type");
1886 case MVT::v8i8: Opc = ARM::VST2LNd8; break;
1887 case MVT::v4i16: Opc = ARM::VST2LNd16; break;
1889 case MVT::v2i32: Opc = ARM::VST2LNd32; break;
1891 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
1892 N->getOperand(3), N->getOperand(4),
1893 getI32Imm(Lane), Chain };
1894 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7);
1896 // Quad registers are handled by extracting subregs and then doing
1900 switch (VT.getSimpleVT().SimpleTy) {
1901 default: llvm_unreachable("unhandled vst2lane type");
1903 Opc = ARM::VST2LNq16a;
1904 Opc2 = ARM::VST2LNq16b;
1908 Opc = ARM::VST2LNq32a;
1909 Opc2 = ARM::VST2LNq32b;
1913 Opc = ARM::VST2LNq32a;
1914 Opc2 = ARM::VST2LNq32b;
1918 unsigned NumElts = RegVT.getVectorNumElements();
1919 int SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
1921 SDValue D0 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1923 SDValue D1 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1925 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1,
1926 getI32Imm(Lane % NumElts), Chain };
1927 return CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
1928 dl, MVT::Other, Ops, 7);
1931 case Intrinsic::arm_neon_vst3lane: {
1932 SDValue MemAddr, MemUpdate, MemOpc;
1933 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1935 SDValue Chain = N->getOperand(0);
1936 unsigned Lane = cast<ConstantSDNode>(N->getOperand(6))->getZExtValue();
1937 VT = N->getOperand(3).getValueType();
1938 if (VT.is64BitVector()) {
1939 switch (VT.getSimpleVT().SimpleTy) {
1940 default: llvm_unreachable("unhandled vst3lane type");
1941 case MVT::v8i8: Opc = ARM::VST3LNd8; break;
1942 case MVT::v4i16: Opc = ARM::VST3LNd16; break;
1944 case MVT::v2i32: Opc = ARM::VST3LNd32; break;
1946 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
1947 N->getOperand(3), N->getOperand(4),
1948 N->getOperand(5), getI32Imm(Lane), Chain };
1949 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
1951 // Quad registers are handled by extracting subregs and then doing
1955 switch (VT.getSimpleVT().SimpleTy) {
1956 default: llvm_unreachable("unhandled vst3lane type");
1958 Opc = ARM::VST3LNq16a;
1959 Opc2 = ARM::VST3LNq16b;
1963 Opc = ARM::VST3LNq32a;
1964 Opc2 = ARM::VST3LNq32b;
1968 Opc = ARM::VST3LNq32a;
1969 Opc2 = ARM::VST3LNq32b;
1973 unsigned NumElts = RegVT.getVectorNumElements();
1974 int SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
1976 SDValue D0 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1978 SDValue D1 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1980 SDValue D2 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1982 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1, D2,
1983 getI32Imm(Lane % NumElts), Chain };
1984 return CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
1985 dl, MVT::Other, Ops, 8);
1988 case Intrinsic::arm_neon_vst4lane: {
1989 SDValue MemAddr, MemUpdate, MemOpc;
1990 if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
1992 SDValue Chain = N->getOperand(0);
1993 unsigned Lane = cast<ConstantSDNode>(N->getOperand(7))->getZExtValue();
1994 VT = N->getOperand(3).getValueType();
1995 if (VT.is64BitVector()) {
1996 switch (VT.getSimpleVT().SimpleTy) {
1997 default: llvm_unreachable("unhandled vst4lane type");
1998 case MVT::v8i8: Opc = ARM::VST4LNd8; break;
1999 case MVT::v4i16: Opc = ARM::VST4LNd16; break;
2001 case MVT::v2i32: Opc = ARM::VST4LNd32; break;
2003 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
2004 N->getOperand(3), N->getOperand(4),
2005 N->getOperand(5), N->getOperand(6),
2006 getI32Imm(Lane), Chain };
2007 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 9);
2009 // Quad registers are handled by extracting subregs and then doing
2013 switch (VT.getSimpleVT().SimpleTy) {
2014 default: llvm_unreachable("unhandled vst4lane type");
2016 Opc = ARM::VST4LNq16a;
2017 Opc2 = ARM::VST4LNq16b;
2021 Opc = ARM::VST4LNq32a;
2022 Opc2 = ARM::VST4LNq32b;
2026 Opc = ARM::VST4LNq32a;
2027 Opc2 = ARM::VST4LNq32b;
2031 unsigned NumElts = RegVT.getVectorNumElements();
2032 int SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
2034 SDValue D0 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
2036 SDValue D1 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
2038 SDValue D2 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
2040 SDValue D3 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
2042 const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1, D2, D3,
2043 getI32Imm(Lane % NumElts), Chain };
2044 return CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
2045 dl, MVT::Other, Ops, 9);
2051 return SelectCode(Op);
2054 bool ARMDAGToDAGISel::
2055 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2056 std::vector<SDValue> &OutOps) {
2057 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
2058 // Require the address to be in a register. That is safe for all ARM
2059 // variants and it is hard to do anything much smarter without knowing
2060 // how the operand is used.
2061 OutOps.push_back(Op);
2065 /// createARMISelDag - This pass converts a legalized DAG into a
2066 /// ARM-specific DAG, ready for instruction scheduling.
2068 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
2069 CodeGenOpt::Level OptLevel) {
2070 return new ARMDAGToDAGISel(TM, OptLevel);