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 "ARMISelLowering.h"
17 #include "ARMTargetMachine.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Function.h"
22 #include "llvm/Intrinsics.h"
23 #include "llvm/LLVMContext.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/CodeGen/SelectionDAGISel.h"
29 #include "llvm/Target/TargetLowering.h"
30 #include "llvm/Target/TargetOptions.h"
31 #include "llvm/Support/Compiler.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
38 //===--------------------------------------------------------------------===//
39 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
40 /// instructions for SelectionDAG operations.
43 class ARMDAGToDAGISel : public SelectionDAGISel {
44 ARMBaseTargetMachine &TM;
46 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
47 /// make the right decision when generating code for different targets.
48 const ARMSubtarget *Subtarget;
51 explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
52 CodeGenOpt::Level OptLevel)
53 : SelectionDAGISel(tm, OptLevel), TM(tm),
54 Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
57 virtual const char *getPassName() const {
58 return "ARM Instruction Selection";
61 /// getI32Imm - Return a target constant of type i32 with the specified
63 inline SDValue getI32Imm(unsigned Imm) {
64 return CurDAG->getTargetConstant(Imm, MVT::i32);
67 SDNode *Select(SDNode *N);
69 bool SelectShifterOperandReg(SDNode *Op, SDValue N, SDValue &A,
70 SDValue &B, SDValue &C);
71 bool SelectAddrMode2(SDNode *Op, SDValue N, SDValue &Base,
72 SDValue &Offset, SDValue &Opc);
73 bool SelectAddrMode2Offset(SDNode *Op, SDValue N,
74 SDValue &Offset, SDValue &Opc);
75 bool SelectAddrMode3(SDNode *Op, SDValue N, SDValue &Base,
76 SDValue &Offset, SDValue &Opc);
77 bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
78 SDValue &Offset, SDValue &Opc);
79 bool SelectAddrMode4(SDNode *Op, SDValue N, SDValue &Addr,
81 bool SelectAddrMode5(SDNode *Op, SDValue N, SDValue &Base,
83 bool SelectAddrMode6(SDNode *Op, SDValue N, SDValue &Addr, SDValue &Align);
85 bool SelectAddrModePC(SDNode *Op, SDValue N, SDValue &Offset,
88 bool SelectThumbAddrModeRR(SDNode *Op, SDValue N, SDValue &Base,
90 bool SelectThumbAddrModeRI5(SDNode *Op, SDValue N, unsigned Scale,
91 SDValue &Base, SDValue &OffImm,
93 bool SelectThumbAddrModeS1(SDNode *Op, SDValue N, SDValue &Base,
94 SDValue &OffImm, SDValue &Offset);
95 bool SelectThumbAddrModeS2(SDNode *Op, SDValue N, SDValue &Base,
96 SDValue &OffImm, SDValue &Offset);
97 bool SelectThumbAddrModeS4(SDNode *Op, SDValue N, SDValue &Base,
98 SDValue &OffImm, SDValue &Offset);
99 bool SelectThumbAddrModeSP(SDNode *Op, SDValue N, SDValue &Base,
102 bool SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
103 SDValue &BaseReg, SDValue &Opc);
104 bool SelectT2AddrModeImm12(SDNode *Op, SDValue N, SDValue &Base,
106 bool SelectT2AddrModeImm8(SDNode *Op, SDValue N, SDValue &Base,
108 bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
110 bool SelectT2AddrModeImm8s4(SDNode *Op, SDValue N, SDValue &Base,
112 bool SelectT2AddrModeSoReg(SDNode *Op, SDValue N, SDValue &Base,
113 SDValue &OffReg, SDValue &ShImm);
115 // Include the pieces autogenerated from the target description.
116 #include "ARMGenDAGISel.inc"
119 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
121 SDNode *SelectARMIndexedLoad(SDNode *N);
122 SDNode *SelectT2IndexedLoad(SDNode *N);
124 /// SelectVLD - Select NEON load intrinsics. NumVecs should be
125 /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for
126 /// loads of D registers and even subregs and odd subregs of Q registers.
127 /// For NumVecs <= 2, QOpcodes1 is not used.
128 SDNode *SelectVLD(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
129 unsigned *QOpcodes0, unsigned *QOpcodes1);
131 /// SelectVST - Select NEON store intrinsics. NumVecs should
132 /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for
133 /// stores of D registers and even subregs and odd subregs of Q registers.
134 /// For NumVecs <= 2, QOpcodes1 is not used.
135 SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
136 unsigned *QOpcodes0, unsigned *QOpcodes1);
138 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
139 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
140 /// load/store of D registers and even subregs and odd subregs of Q registers.
141 SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, unsigned NumVecs,
142 unsigned *DOpcodes, unsigned *QOpcodes0,
143 unsigned *QOpcodes1);
145 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
146 SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, unsigned Opc);
148 /// SelectCMOVOp - Select CMOV instructions for ARM.
149 SDNode *SelectCMOVOp(SDNode *N);
150 SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
151 ARMCC::CondCodes CCVal, SDValue CCR,
153 SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
154 ARMCC::CondCodes CCVal, SDValue CCR,
156 SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
157 ARMCC::CondCodes CCVal, SDValue CCR,
159 SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
160 ARMCC::CondCodes CCVal, SDValue CCR,
163 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
164 /// inline asm expressions.
165 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
167 std::vector<SDValue> &OutOps);
169 /// PairDRegs - Insert a pair of double registers into an implicit def to
170 /// form a quad register.
171 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
175 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
176 /// operand. If so Imm will receive the 32-bit value.
177 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
178 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
179 Imm = cast<ConstantSDNode>(N)->getZExtValue();
185 // isInt32Immediate - This method tests to see if a constant operand.
186 // If so Imm will receive the 32 bit value.
187 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
188 return isInt32Immediate(N.getNode(), Imm);
191 // isOpcWithIntImmediate - This method tests to see if the node is a specific
192 // opcode and that it has a immediate integer right operand.
193 // If so Imm will receive the 32 bit value.
194 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
195 return N->getOpcode() == Opc &&
196 isInt32Immediate(N->getOperand(1).getNode(), Imm);
200 bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
205 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
207 // Don't match base register only case. That is matched to a separate
208 // lower complexity pattern with explicit register operand.
209 if (ShOpcVal == ARM_AM::no_shift) return false;
211 BaseReg = N.getOperand(0);
212 unsigned ShImmVal = 0;
213 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
214 ShReg = CurDAG->getRegister(0, MVT::i32);
215 ShImmVal = RHS->getZExtValue() & 31;
217 ShReg = N.getOperand(1);
219 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
224 bool ARMDAGToDAGISel::SelectAddrMode2(SDNode *Op, SDValue N,
225 SDValue &Base, SDValue &Offset,
227 if (N.getOpcode() == ISD::MUL) {
228 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
229 // X * [3,5,9] -> X + X * [2,4,8] etc.
230 int RHSC = (int)RHS->getZExtValue();
233 ARM_AM::AddrOpc AddSub = ARM_AM::add;
235 AddSub = ARM_AM::sub;
238 if (isPowerOf2_32(RHSC)) {
239 unsigned ShAmt = Log2_32(RHSC);
240 Base = Offset = N.getOperand(0);
241 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
250 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
252 if (N.getOpcode() == ISD::FrameIndex) {
253 int FI = cast<FrameIndexSDNode>(N)->getIndex();
254 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
255 } else if (N.getOpcode() == ARMISD::Wrapper &&
256 !(Subtarget->useMovt() &&
257 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
258 Base = N.getOperand(0);
260 Offset = CurDAG->getRegister(0, MVT::i32);
261 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
267 // Match simple R +/- imm12 operands.
268 if (N.getOpcode() == ISD::ADD)
269 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
270 int RHSC = (int)RHS->getZExtValue();
271 if ((RHSC >= 0 && RHSC < 0x1000) ||
272 (RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
273 Base = N.getOperand(0);
274 if (Base.getOpcode() == ISD::FrameIndex) {
275 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
276 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
278 Offset = CurDAG->getRegister(0, MVT::i32);
280 ARM_AM::AddrOpc AddSub = ARM_AM::add;
282 AddSub = ARM_AM::sub;
285 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
292 // Otherwise this is R +/- [possibly shifted] R.
293 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
294 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
297 Base = N.getOperand(0);
298 Offset = N.getOperand(1);
300 if (ShOpcVal != ARM_AM::no_shift) {
301 // Check to see if the RHS of the shift is a constant, if not, we can't fold
303 if (ConstantSDNode *Sh =
304 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
305 ShAmt = Sh->getZExtValue();
306 Offset = N.getOperand(1).getOperand(0);
308 ShOpcVal = ARM_AM::no_shift;
312 // Try matching (R shl C) + (R).
313 if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
314 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
315 if (ShOpcVal != ARM_AM::no_shift) {
316 // Check to see if the RHS of the shift is a constant, if not, we can't
318 if (ConstantSDNode *Sh =
319 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
320 ShAmt = Sh->getZExtValue();
321 Offset = N.getOperand(0).getOperand(0);
322 Base = N.getOperand(1);
324 ShOpcVal = ARM_AM::no_shift;
329 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
334 bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
335 SDValue &Offset, SDValue &Opc) {
336 unsigned Opcode = Op->getOpcode();
337 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
338 ? cast<LoadSDNode>(Op)->getAddressingMode()
339 : cast<StoreSDNode>(Op)->getAddressingMode();
340 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
341 ? ARM_AM::add : ARM_AM::sub;
342 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
343 int Val = (int)C->getZExtValue();
344 if (Val >= 0 && Val < 0x1000) { // 12 bits.
345 Offset = CurDAG->getRegister(0, MVT::i32);
346 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
354 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
356 if (ShOpcVal != ARM_AM::no_shift) {
357 // Check to see if the RHS of the shift is a constant, if not, we can't fold
359 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
360 ShAmt = Sh->getZExtValue();
361 Offset = N.getOperand(0);
363 ShOpcVal = ARM_AM::no_shift;
367 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
373 bool ARMDAGToDAGISel::SelectAddrMode3(SDNode *Op, SDValue N,
374 SDValue &Base, SDValue &Offset,
376 if (N.getOpcode() == ISD::SUB) {
377 // X - C is canonicalize to X + -C, no need to handle it here.
378 Base = N.getOperand(0);
379 Offset = N.getOperand(1);
380 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
384 if (N.getOpcode() != ISD::ADD) {
386 if (N.getOpcode() == ISD::FrameIndex) {
387 int FI = cast<FrameIndexSDNode>(N)->getIndex();
388 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
390 Offset = CurDAG->getRegister(0, MVT::i32);
391 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
395 // If the RHS is +/- imm8, fold into addr mode.
396 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
397 int RHSC = (int)RHS->getZExtValue();
398 if ((RHSC >= 0 && RHSC < 256) ||
399 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
400 Base = N.getOperand(0);
401 if (Base.getOpcode() == ISD::FrameIndex) {
402 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
403 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
405 Offset = CurDAG->getRegister(0, MVT::i32);
407 ARM_AM::AddrOpc AddSub = ARM_AM::add;
409 AddSub = ARM_AM::sub;
412 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
417 Base = N.getOperand(0);
418 Offset = N.getOperand(1);
419 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
423 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
424 SDValue &Offset, SDValue &Opc) {
425 unsigned Opcode = Op->getOpcode();
426 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
427 ? cast<LoadSDNode>(Op)->getAddressingMode()
428 : cast<StoreSDNode>(Op)->getAddressingMode();
429 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
430 ? ARM_AM::add : ARM_AM::sub;
431 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
432 int Val = (int)C->getZExtValue();
433 if (Val >= 0 && Val < 256) {
434 Offset = CurDAG->getRegister(0, MVT::i32);
435 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
441 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
445 bool ARMDAGToDAGISel::SelectAddrMode4(SDNode *Op, SDValue N,
446 SDValue &Addr, SDValue &Mode) {
448 Mode = CurDAG->getTargetConstant(0, MVT::i32);
452 bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N,
453 SDValue &Base, SDValue &Offset) {
454 if (N.getOpcode() != ISD::ADD) {
456 if (N.getOpcode() == ISD::FrameIndex) {
457 int FI = cast<FrameIndexSDNode>(N)->getIndex();
458 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
459 } else if (N.getOpcode() == ARMISD::Wrapper &&
460 !(Subtarget->useMovt() &&
461 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
462 Base = N.getOperand(0);
464 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
469 // If the RHS is +/- imm8, fold into addr mode.
470 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
471 int RHSC = (int)RHS->getZExtValue();
472 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied by 4.
474 if ((RHSC >= 0 && RHSC < 256) ||
475 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
476 Base = N.getOperand(0);
477 if (Base.getOpcode() == ISD::FrameIndex) {
478 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
479 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
482 ARM_AM::AddrOpc AddSub = ARM_AM::add;
484 AddSub = ARM_AM::sub;
487 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
495 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
500 bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N,
501 SDValue &Addr, SDValue &Align) {
503 // Default to no alignment.
504 Align = CurDAG->getTargetConstant(0, MVT::i32);
508 bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
509 SDValue &Offset, SDValue &Label) {
510 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
511 Offset = N.getOperand(0);
512 SDValue N1 = N.getOperand(1);
513 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
520 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
521 SDValue &Base, SDValue &Offset){
522 // FIXME dl should come from the parent load or store, not the address
523 DebugLoc dl = Op->getDebugLoc();
524 if (N.getOpcode() != ISD::ADD) {
525 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
526 if (!NC || NC->getZExtValue() != 0)
533 Base = N.getOperand(0);
534 Offset = N.getOperand(1);
539 ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
540 unsigned Scale, SDValue &Base,
541 SDValue &OffImm, SDValue &Offset) {
543 SDValue TmpBase, TmpOffImm;
544 if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
545 return false; // We want to select tLDRspi / tSTRspi instead.
546 if (N.getOpcode() == ARMISD::Wrapper &&
547 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
548 return false; // We want to select tLDRpci instead.
551 if (N.getOpcode() != ISD::ADD) {
552 if (N.getOpcode() == ARMISD::Wrapper &&
553 !(Subtarget->useMovt() &&
554 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
555 Base = N.getOperand(0);
559 Offset = CurDAG->getRegister(0, MVT::i32);
560 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
564 // Thumb does not have [sp, r] address mode.
565 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
566 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
567 if ((LHSR && LHSR->getReg() == ARM::SP) ||
568 (RHSR && RHSR->getReg() == ARM::SP)) {
570 Offset = CurDAG->getRegister(0, MVT::i32);
571 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
575 // If the RHS is + imm5 * scale, fold into addr mode.
576 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
577 int RHSC = (int)RHS->getZExtValue();
578 if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
580 if (RHSC >= 0 && RHSC < 32) {
581 Base = N.getOperand(0);
582 Offset = CurDAG->getRegister(0, MVT::i32);
583 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
589 Base = N.getOperand(0);
590 Offset = N.getOperand(1);
591 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
595 bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
596 SDValue &Base, SDValue &OffImm,
598 return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
601 bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
602 SDValue &Base, SDValue &OffImm,
604 return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
607 bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
608 SDValue &Base, SDValue &OffImm,
610 return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
613 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
614 SDValue &Base, SDValue &OffImm) {
615 if (N.getOpcode() == ISD::FrameIndex) {
616 int FI = cast<FrameIndexSDNode>(N)->getIndex();
617 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
618 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
622 if (N.getOpcode() != ISD::ADD)
625 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
626 if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
627 (LHSR && LHSR->getReg() == ARM::SP)) {
628 // If the RHS is + imm8 * scale, fold into addr mode.
629 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
630 int RHSC = (int)RHS->getZExtValue();
631 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
633 if (RHSC >= 0 && RHSC < 256) {
634 Base = N.getOperand(0);
635 if (Base.getOpcode() == ISD::FrameIndex) {
636 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
637 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
639 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
649 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
652 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
654 // Don't match base register only case. That is matched to a separate
655 // lower complexity pattern with explicit register operand.
656 if (ShOpcVal == ARM_AM::no_shift) return false;
658 BaseReg = N.getOperand(0);
659 unsigned ShImmVal = 0;
660 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
661 ShImmVal = RHS->getZExtValue() & 31;
662 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
669 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
670 SDValue &Base, SDValue &OffImm) {
671 // Match simple R + imm12 operands.
674 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
675 if (N.getOpcode() == ISD::FrameIndex) {
676 // Match frame index...
677 int FI = cast<FrameIndexSDNode>(N)->getIndex();
678 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
679 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
681 } else if (N.getOpcode() == ARMISD::Wrapper &&
682 !(Subtarget->useMovt() &&
683 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
684 Base = N.getOperand(0);
685 if (Base.getOpcode() == ISD::TargetConstantPool)
686 return false; // We want to select t2LDRpci instead.
689 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
693 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
694 if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
695 // Let t2LDRi8 handle (R - imm8).
698 int RHSC = (int)RHS->getZExtValue();
699 if (N.getOpcode() == ISD::SUB)
702 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
703 Base = N.getOperand(0);
704 if (Base.getOpcode() == ISD::FrameIndex) {
705 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
706 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
708 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
715 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
719 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
720 SDValue &Base, SDValue &OffImm) {
721 // Match simple R - imm8 operands.
722 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
723 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
724 int RHSC = (int)RHS->getSExtValue();
725 if (N.getOpcode() == ISD::SUB)
728 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
729 Base = N.getOperand(0);
730 if (Base.getOpcode() == ISD::FrameIndex) {
731 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
732 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
734 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
743 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
745 unsigned Opcode = Op->getOpcode();
746 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
747 ? cast<LoadSDNode>(Op)->getAddressingMode()
748 : cast<StoreSDNode>(Op)->getAddressingMode();
749 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
750 int RHSC = (int)RHS->getZExtValue();
751 if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
752 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
753 ? CurDAG->getTargetConstant(RHSC, MVT::i32)
754 : CurDAG->getTargetConstant(-RHSC, MVT::i32);
762 bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
763 SDValue &Base, SDValue &OffImm) {
764 if (N.getOpcode() == ISD::ADD) {
765 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
766 int RHSC = (int)RHS->getZExtValue();
767 if (((RHSC & 0x3) == 0) &&
768 ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits.
769 Base = N.getOperand(0);
770 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
774 } else if (N.getOpcode() == ISD::SUB) {
775 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
776 int RHSC = (int)RHS->getZExtValue();
777 if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) { // 8 bits.
778 Base = N.getOperand(0);
779 OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
788 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDNode *Op, SDValue N,
790 SDValue &OffReg, SDValue &ShImm) {
791 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
792 if (N.getOpcode() != ISD::ADD)
795 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
796 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
797 int RHSC = (int)RHS->getZExtValue();
798 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
800 else if (RHSC < 0 && RHSC >= -255) // 8 bits
804 // Look for (R + R) or (R + (R << [1,2,3])).
806 Base = N.getOperand(0);
807 OffReg = N.getOperand(1);
809 // Swap if it is ((R << c) + R).
810 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
811 if (ShOpcVal != ARM_AM::lsl) {
812 ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
813 if (ShOpcVal == ARM_AM::lsl)
814 std::swap(Base, OffReg);
817 if (ShOpcVal == ARM_AM::lsl) {
818 // Check to see if the RHS of the shift is a constant, if not, we can't fold
820 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
821 ShAmt = Sh->getZExtValue();
824 ShOpcVal = ARM_AM::no_shift;
826 OffReg = OffReg.getOperand(0);
828 ShOpcVal = ARM_AM::no_shift;
832 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
837 //===--------------------------------------------------------------------===//
839 /// getAL - Returns a ARMCC::AL immediate node.
840 static inline SDValue getAL(SelectionDAG *CurDAG) {
841 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
844 SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
845 LoadSDNode *LD = cast<LoadSDNode>(N);
846 ISD::MemIndexedMode AM = LD->getAddressingMode();
847 if (AM == ISD::UNINDEXED)
850 EVT LoadedVT = LD->getMemoryVT();
851 SDValue Offset, AMOpc;
852 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
855 if (LoadedVT == MVT::i32 &&
856 SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
857 Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
859 } else if (LoadedVT == MVT::i16 &&
860 SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
862 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
863 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
864 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
865 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
866 if (LD->getExtensionType() == ISD::SEXTLOAD) {
867 if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
869 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
872 if (SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
874 Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
880 SDValue Chain = LD->getChain();
881 SDValue Base = LD->getBasePtr();
882 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
883 CurDAG->getRegister(0, MVT::i32), Chain };
884 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
891 SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
892 LoadSDNode *LD = cast<LoadSDNode>(N);
893 ISD::MemIndexedMode AM = LD->getAddressingMode();
894 if (AM == ISD::UNINDEXED)
897 EVT LoadedVT = LD->getMemoryVT();
898 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
900 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
903 if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
904 switch (LoadedVT.getSimpleVT().SimpleTy) {
906 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
910 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
912 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
917 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
919 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
928 SDValue Chain = LD->getChain();
929 SDValue Base = LD->getBasePtr();
930 SDValue Ops[]= { Base, Offset, getAL(CurDAG),
931 CurDAG->getRegister(0, MVT::i32), Chain };
932 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
939 /// PairDRegs - Insert a pair of double registers into an implicit def to
940 /// form a quad register.
941 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
942 DebugLoc dl = V0.getNode()->getDebugLoc();
944 SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0);
945 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
946 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
947 SDNode *Pair = CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
948 VT, Undef, V0, SubReg0);
949 return CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
950 VT, SDValue(Pair, 0), V1, SubReg1);
953 /// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
954 /// for a 64-bit subregister of the vector.
955 static EVT GetNEONSubregVT(EVT VT) {
956 switch (VT.getSimpleVT().SimpleTy) {
957 default: llvm_unreachable("unhandled NEON type");
958 case MVT::v16i8: return MVT::v8i8;
959 case MVT::v8i16: return MVT::v4i16;
960 case MVT::v4f32: return MVT::v2f32;
961 case MVT::v4i32: return MVT::v2i32;
962 case MVT::v2i64: return MVT::v1i64;
966 SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
967 unsigned *DOpcodes, unsigned *QOpcodes0,
968 unsigned *QOpcodes1) {
969 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
970 DebugLoc dl = N->getDebugLoc();
972 SDValue MemAddr, Align;
973 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
976 SDValue Chain = N->getOperand(0);
977 EVT VT = N->getValueType(0);
978 bool is64BitVector = VT.is64BitVector();
980 unsigned OpcodeIndex;
981 switch (VT.getSimpleVT().SimpleTy) {
982 default: llvm_unreachable("unhandled vld type");
983 // Double-register operations:
984 case MVT::v8i8: OpcodeIndex = 0; break;
985 case MVT::v4i16: OpcodeIndex = 1; break;
987 case MVT::v2i32: OpcodeIndex = 2; break;
988 case MVT::v1i64: OpcodeIndex = 3; break;
989 // Quad-register operations:
990 case MVT::v16i8: OpcodeIndex = 0; break;
991 case MVT::v8i16: OpcodeIndex = 1; break;
993 case MVT::v4i32: OpcodeIndex = 2; break;
994 case MVT::v2i64: OpcodeIndex = 3;
995 assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
999 SDValue Pred = getAL(CurDAG);
1000 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1001 if (is64BitVector) {
1002 unsigned Opc = DOpcodes[OpcodeIndex];
1003 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1004 std::vector<EVT> ResTys(NumVecs, VT);
1005 ResTys.push_back(MVT::Other);
1006 return CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1009 EVT RegVT = GetNEONSubregVT(VT);
1011 // Quad registers are directly supported for VLD1 and VLD2,
1012 // loading pairs of D regs.
1013 unsigned Opc = QOpcodes0[OpcodeIndex];
1014 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1015 std::vector<EVT> ResTys(2 * NumVecs, RegVT);
1016 ResTys.push_back(MVT::Other);
1017 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1018 Chain = SDValue(VLd, 2 * NumVecs);
1020 // Combine the even and odd subregs to produce the result.
1021 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1022 SDNode *Q = PairDRegs(VT, SDValue(VLd, 2*Vec), SDValue(VLd, 2*Vec+1));
1023 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1026 // Otherwise, quad registers are loaded with two separate instructions,
1027 // where one loads the even registers and the other loads the odd registers.
1029 std::vector<EVT> ResTys(NumVecs, RegVT);
1030 ResTys.push_back(MemAddr.getValueType());
1031 ResTys.push_back(MVT::Other);
1033 // Load the even subregs.
1034 unsigned Opc = QOpcodes0[OpcodeIndex];
1035 const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain };
1036 SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6);
1037 Chain = SDValue(VLdA, NumVecs+1);
1039 // Load the odd subregs.
1040 Opc = QOpcodes1[OpcodeIndex];
1041 const SDValue OpsB[] = { SDValue(VLdA, NumVecs),
1042 Align, Reg0, Pred, Reg0, Chain };
1043 SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
1044 Chain = SDValue(VLdB, NumVecs+1);
1046 // Combine the even and odd subregs to produce the result.
1047 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1048 SDNode *Q = PairDRegs(VT, SDValue(VLdA, Vec), SDValue(VLdB, Vec));
1049 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1052 ReplaceUses(SDValue(N, NumVecs), Chain);
1056 SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
1057 unsigned *DOpcodes, unsigned *QOpcodes0,
1058 unsigned *QOpcodes1) {
1059 assert(NumVecs >=1 && NumVecs <= 4 && "VST NumVecs out-of-range");
1060 DebugLoc dl = N->getDebugLoc();
1062 SDValue MemAddr, Align;
1063 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1066 SDValue Chain = N->getOperand(0);
1067 EVT VT = N->getOperand(3).getValueType();
1068 bool is64BitVector = VT.is64BitVector();
1070 unsigned OpcodeIndex;
1071 switch (VT.getSimpleVT().SimpleTy) {
1072 default: llvm_unreachable("unhandled vst type");
1073 // Double-register operations:
1074 case MVT::v8i8: OpcodeIndex = 0; break;
1075 case MVT::v4i16: OpcodeIndex = 1; break;
1077 case MVT::v2i32: OpcodeIndex = 2; break;
1078 case MVT::v1i64: OpcodeIndex = 3; break;
1079 // Quad-register operations:
1080 case MVT::v16i8: OpcodeIndex = 0; break;
1081 case MVT::v8i16: OpcodeIndex = 1; break;
1083 case MVT::v4i32: OpcodeIndex = 2; break;
1084 case MVT::v2i64: OpcodeIndex = 3;
1085 assert(NumVecs == 1 && "v2i64 type only supported for VST1");
1089 SDValue Pred = getAL(CurDAG);
1090 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1092 SmallVector<SDValue, 10> Ops;
1093 Ops.push_back(MemAddr);
1094 Ops.push_back(Align);
1096 if (is64BitVector) {
1097 unsigned Opc = DOpcodes[OpcodeIndex];
1098 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1099 Ops.push_back(N->getOperand(Vec+3));
1100 Ops.push_back(Pred);
1101 Ops.push_back(Reg0); // predicate register
1102 Ops.push_back(Chain);
1103 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+5);
1106 EVT RegVT = GetNEONSubregVT(VT);
1108 // Quad registers are directly supported for VST1 and VST2,
1109 // storing pairs of D regs.
1110 unsigned Opc = QOpcodes0[OpcodeIndex];
1111 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1112 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1113 N->getOperand(Vec+3)));
1114 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1115 N->getOperand(Vec+3)));
1117 Ops.push_back(Pred);
1118 Ops.push_back(Reg0); // predicate register
1119 Ops.push_back(Chain);
1120 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
1124 // Otherwise, quad registers are stored with two separate instructions,
1125 // where one stores the even registers and the other stores the odd registers.
1127 Ops.push_back(Reg0); // post-access address offset
1129 // Store the even subregs.
1130 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1131 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1132 N->getOperand(Vec+3)));
1133 Ops.push_back(Pred);
1134 Ops.push_back(Reg0); // predicate register
1135 Ops.push_back(Chain);
1136 unsigned Opc = QOpcodes0[OpcodeIndex];
1137 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1138 MVT::Other, Ops.data(), NumVecs+6);
1139 Chain = SDValue(VStA, 1);
1141 // Store the odd subregs.
1142 Ops[0] = SDValue(VStA, 0); // MemAddr
1143 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1144 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1145 N->getOperand(Vec+3));
1146 Ops[NumVecs+5] = Chain;
1147 Opc = QOpcodes1[OpcodeIndex];
1148 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1149 MVT::Other, Ops.data(), NumVecs+6);
1150 Chain = SDValue(VStB, 1);
1151 ReplaceUses(SDValue(N, 0), Chain);
1155 SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1156 unsigned NumVecs, unsigned *DOpcodes,
1157 unsigned *QOpcodes0,
1158 unsigned *QOpcodes1) {
1159 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1160 DebugLoc dl = N->getDebugLoc();
1162 SDValue MemAddr, Align;
1163 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1166 SDValue Chain = N->getOperand(0);
1168 cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1169 EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
1170 bool is64BitVector = VT.is64BitVector();
1172 // Quad registers are handled by load/store of subregs. Find the subreg info.
1173 unsigned NumElts = 0;
1176 if (!is64BitVector) {
1177 RegVT = GetNEONSubregVT(VT);
1178 NumElts = RegVT.getVectorNumElements();
1179 SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
1182 unsigned OpcodeIndex;
1183 switch (VT.getSimpleVT().SimpleTy) {
1184 default: llvm_unreachable("unhandled vld/vst lane type");
1185 // Double-register operations:
1186 case MVT::v8i8: OpcodeIndex = 0; break;
1187 case MVT::v4i16: OpcodeIndex = 1; break;
1189 case MVT::v2i32: OpcodeIndex = 2; break;
1190 // Quad-register operations:
1191 case MVT::v8i16: OpcodeIndex = 0; break;
1193 case MVT::v4i32: OpcodeIndex = 1; break;
1196 SDValue Pred = getAL(CurDAG);
1197 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1199 SmallVector<SDValue, 10> Ops;
1200 Ops.push_back(MemAddr);
1201 Ops.push_back(Align);
1204 if (is64BitVector) {
1205 Opc = DOpcodes[OpcodeIndex];
1206 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1207 Ops.push_back(N->getOperand(Vec+3));
1209 // Check if this is loading the even or odd subreg of a Q register.
1210 if (Lane < NumElts) {
1211 Opc = QOpcodes0[OpcodeIndex];
1214 Opc = QOpcodes1[OpcodeIndex];
1216 // Extract the subregs of the input vector.
1217 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1218 Ops.push_back(CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1219 N->getOperand(Vec+3)));
1221 Ops.push_back(getI32Imm(Lane));
1222 Ops.push_back(Pred);
1223 Ops.push_back(Reg0);
1224 Ops.push_back(Chain);
1227 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6);
1229 std::vector<EVT> ResTys(NumVecs, RegVT);
1230 ResTys.push_back(MVT::Other);
1232 CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), NumVecs+6);
1233 // For a 64-bit vector load to D registers, nothing more needs to be done.
1237 // For 128-bit vectors, take the 64-bit results of the load and insert them
1238 // as subregs into the result.
1239 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1240 SDValue QuadVec = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
1241 N->getOperand(Vec+3),
1242 SDValue(VLdLn, Vec));
1243 ReplaceUses(SDValue(N, Vec), QuadVec);
1246 Chain = SDValue(VLdLn, NumVecs);
1247 ReplaceUses(SDValue(N, NumVecs), Chain);
1251 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
1253 if (!Subtarget->hasV6T2Ops())
1256 unsigned Shl_imm = 0;
1257 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
1258 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1259 unsigned Srl_imm = 0;
1260 if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
1261 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1262 unsigned Width = 32 - Srl_imm;
1263 int LSB = Srl_imm - Shl_imm;
1266 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1267 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1268 CurDAG->getTargetConstant(LSB, MVT::i32),
1269 CurDAG->getTargetConstant(Width, MVT::i32),
1270 getAL(CurDAG), Reg0 };
1271 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1277 SDNode *ARMDAGToDAGISel::
1278 SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1279 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1282 if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
1283 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1284 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1287 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1288 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1289 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1290 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1292 llvm_unreachable("Unknown so_reg opcode!");
1296 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1297 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1298 SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
1299 return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
1304 SDNode *ARMDAGToDAGISel::
1305 SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1306 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1310 if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
1311 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1312 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
1313 return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
1318 SDNode *ARMDAGToDAGISel::
1319 SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1320 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1321 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1325 if (Predicate_t2_so_imm(TrueVal.getNode())) {
1326 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1327 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1328 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1329 return CurDAG->SelectNodeTo(N,
1330 ARM::t2MOVCCi, MVT::i32, Ops, 5);
1335 SDNode *ARMDAGToDAGISel::
1336 SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1337 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1338 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1342 if (Predicate_so_imm(TrueVal.getNode())) {
1343 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1344 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1345 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1346 return CurDAG->SelectNodeTo(N,
1347 ARM::MOVCCi, MVT::i32, Ops, 5);
1352 SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
1353 EVT VT = N->getValueType(0);
1354 SDValue FalseVal = N->getOperand(0);
1355 SDValue TrueVal = N->getOperand(1);
1356 SDValue CC = N->getOperand(2);
1357 SDValue CCR = N->getOperand(3);
1358 SDValue InFlag = N->getOperand(4);
1359 assert(CC.getOpcode() == ISD::Constant);
1360 assert(CCR.getOpcode() == ISD::Register);
1361 ARMCC::CondCodes CCVal =
1362 (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
1364 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1365 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1366 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1367 // Pattern complexity = 18 cost = 1 size = 0
1371 if (Subtarget->isThumb()) {
1372 SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
1373 CCVal, CCR, InFlag);
1375 Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
1376 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1380 SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
1381 CCVal, CCR, InFlag);
1383 Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
1384 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1389 // Pattern: (ARMcmov:i32 GPR:i32:$false,
1390 // (imm:i32)<<P:Predicate_so_imm>>:$true,
1392 // Emits: (MOVCCi:i32 GPR:i32:$false,
1393 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1394 // Pattern complexity = 10 cost = 1 size = 0
1395 if (Subtarget->isThumb()) {
1396 SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
1397 CCVal, CCR, InFlag);
1399 Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
1400 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1404 SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
1405 CCVal, CCR, InFlag);
1407 Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
1408 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1414 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1415 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1416 // Pattern complexity = 6 cost = 1 size = 0
1418 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1419 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1420 // Pattern complexity = 6 cost = 11 size = 0
1422 // Also FCPYScc and FCPYDcc.
1423 SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
1424 SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
1426 switch (VT.getSimpleVT().SimpleTy) {
1427 default: assert(false && "Illegal conditional move type!");
1430 Opc = Subtarget->isThumb()
1431 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1441 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1444 SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
1445 DebugLoc dl = N->getDebugLoc();
1447 if (N->isMachineOpcode())
1448 return NULL; // Already selected.
1450 switch (N->getOpcode()) {
1452 case ISD::Constant: {
1453 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1455 if (Subtarget->hasThumb2())
1456 // Thumb2-aware targets have the MOVT instruction, so all immediates can
1457 // be done with MOV + MOVT, at worst.
1460 if (Subtarget->isThumb()) {
1461 UseCP = (Val > 255 && // MOV
1462 ~Val > 255 && // MOV + MVN
1463 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
1465 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
1466 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
1467 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
1472 CurDAG->getTargetConstantPool(ConstantInt::get(
1473 Type::getInt32Ty(*CurDAG->getContext()), Val),
1474 TLI.getPointerTy());
1477 if (Subtarget->isThumb1Only()) {
1478 SDValue Pred = getAL(CurDAG);
1479 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1480 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1481 ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1486 CurDAG->getRegister(0, MVT::i32),
1487 CurDAG->getTargetConstant(0, MVT::i32),
1489 CurDAG->getRegister(0, MVT::i32),
1490 CurDAG->getEntryNode()
1492 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1495 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
1499 // Other cases are autogenerated.
1502 case ISD::FrameIndex: {
1503 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1504 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1505 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1506 if (Subtarget->isThumb1Only()) {
1507 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1508 CurDAG->getTargetConstant(0, MVT::i32));
1510 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1511 ARM::t2ADDri : ARM::ADDri);
1512 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1513 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1514 CurDAG->getRegister(0, MVT::i32) };
1515 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1519 if (SDNode *I = SelectV6T2BitfieldExtractOp(N,
1520 Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX))
1524 if (SDNode *I = SelectV6T2BitfieldExtractOp(N,
1525 Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX))
1529 if (Subtarget->isThumb1Only())
1531 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1532 unsigned RHSV = C->getZExtValue();
1534 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
1535 unsigned ShImm = Log2_32(RHSV-1);
1538 SDValue V = N->getOperand(0);
1539 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1540 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1541 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1542 if (Subtarget->isThumb()) {
1543 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1544 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1546 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1547 return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1550 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
1551 unsigned ShImm = Log2_32(RHSV+1);
1554 SDValue V = N->getOperand(0);
1555 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1556 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1557 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1558 if (Subtarget->isThumb()) {
1559 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0 };
1560 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 5);
1562 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1563 return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1569 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
1570 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
1571 // are entirely contributed by c2 and lower 16-bits are entirely contributed
1572 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
1573 // Select it to: "movt x, ((c1 & 0xffff) >> 16)
1574 EVT VT = N->getValueType(0);
1577 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
1579 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
1582 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
1583 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1586 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
1587 SDValue N2 = N0.getOperand(1);
1588 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1591 unsigned N1CVal = N1C->getZExtValue();
1592 unsigned N2CVal = N2C->getZExtValue();
1593 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
1594 (N1CVal & 0xffffU) == 0xffffU &&
1595 (N2CVal & 0xffffU) == 0x0U) {
1596 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
1598 SDValue Ops[] = { N0.getOperand(0), Imm16,
1599 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1600 return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
1605 case ARMISD::VMOVRRD:
1606 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
1607 N->getOperand(0), getAL(CurDAG),
1608 CurDAG->getRegister(0, MVT::i32));
1609 case ISD::UMUL_LOHI: {
1610 if (Subtarget->isThumb1Only())
1612 if (Subtarget->isThumb()) {
1613 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1614 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1615 CurDAG->getRegister(0, MVT::i32) };
1616 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops,4);
1618 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1619 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1620 CurDAG->getRegister(0, MVT::i32) };
1621 return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1624 case ISD::SMUL_LOHI: {
1625 if (Subtarget->isThumb1Only())
1627 if (Subtarget->isThumb()) {
1628 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1629 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1630 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops,4);
1632 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1633 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1634 CurDAG->getRegister(0, MVT::i32) };
1635 return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1639 SDNode *ResNode = 0;
1640 if (Subtarget->isThumb() && Subtarget->hasThumb2())
1641 ResNode = SelectT2IndexedLoad(N);
1643 ResNode = SelectARMIndexedLoad(N);
1647 // VLDMQ must be custom-selected for "v2f64 load" to set the AM5Opc value.
1648 if (Subtarget->hasVFP2() &&
1649 N->getValueType(0).getSimpleVT().SimpleTy == MVT::v2f64) {
1650 SDValue Chain = N->getOperand(0);
1652 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
1653 SDValue Pred = getAL(CurDAG);
1654 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1655 SDValue Ops[] = { N->getOperand(1), AM5Opc, Pred, PredReg, Chain };
1656 return CurDAG->getMachineNode(ARM::VLDMQ, dl, MVT::v2f64, MVT::Other,
1659 // Other cases are autogenerated.
1663 // VSTMQ must be custom-selected for "v2f64 store" to set the AM5Opc value.
1664 if (Subtarget->hasVFP2() &&
1665 N->getOperand(1).getValueType().getSimpleVT().SimpleTy == MVT::v2f64) {
1666 SDValue Chain = N->getOperand(0);
1668 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
1669 SDValue Pred = getAL(CurDAG);
1670 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1671 SDValue Ops[] = { N->getOperand(1), N->getOperand(2),
1672 AM5Opc, Pred, PredReg, Chain };
1673 return CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
1675 // Other cases are autogenerated.
1678 case ARMISD::BRCOND: {
1679 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1680 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
1681 // Pattern complexity = 6 cost = 1 size = 0
1683 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1684 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
1685 // Pattern complexity = 6 cost = 1 size = 0
1687 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1688 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
1689 // Pattern complexity = 6 cost = 1 size = 0
1691 unsigned Opc = Subtarget->isThumb() ?
1692 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
1693 SDValue Chain = N->getOperand(0);
1694 SDValue N1 = N->getOperand(1);
1695 SDValue N2 = N->getOperand(2);
1696 SDValue N3 = N->getOperand(3);
1697 SDValue InFlag = N->getOperand(4);
1698 assert(N1.getOpcode() == ISD::BasicBlock);
1699 assert(N2.getOpcode() == ISD::Constant);
1700 assert(N3.getOpcode() == ISD::Register);
1702 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1703 cast<ConstantSDNode>(N2)->getZExtValue()),
1705 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
1706 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
1708 Chain = SDValue(ResNode, 0);
1709 if (N->getNumValues() == 2) {
1710 InFlag = SDValue(ResNode, 1);
1711 ReplaceUses(SDValue(N, 1), InFlag);
1713 ReplaceUses(SDValue(N, 0),
1714 SDValue(Chain.getNode(), Chain.getResNo()));
1718 return SelectCMOVOp(N);
1719 case ARMISD::CNEG: {
1720 EVT VT = N->getValueType(0);
1721 SDValue N0 = N->getOperand(0);
1722 SDValue N1 = N->getOperand(1);
1723 SDValue N2 = N->getOperand(2);
1724 SDValue N3 = N->getOperand(3);
1725 SDValue InFlag = N->getOperand(4);
1726 assert(N2.getOpcode() == ISD::Constant);
1727 assert(N3.getOpcode() == ISD::Register);
1729 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1730 cast<ConstantSDNode>(N2)->getZExtValue()),
1732 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
1734 switch (VT.getSimpleVT().SimpleTy) {
1735 default: assert(false && "Illegal conditional move type!");
1744 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1747 case ARMISD::VZIP: {
1749 EVT VT = N->getValueType(0);
1750 switch (VT.getSimpleVT().SimpleTy) {
1751 default: return NULL;
1752 case MVT::v8i8: Opc = ARM::VZIPd8; break;
1753 case MVT::v4i16: Opc = ARM::VZIPd16; break;
1755 case MVT::v2i32: Opc = ARM::VZIPd32; break;
1756 case MVT::v16i8: Opc = ARM::VZIPq8; break;
1757 case MVT::v8i16: Opc = ARM::VZIPq16; break;
1759 case MVT::v4i32: Opc = ARM::VZIPq32; break;
1761 SDValue Pred = getAL(CurDAG);
1762 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1763 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
1764 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
1766 case ARMISD::VUZP: {
1768 EVT VT = N->getValueType(0);
1769 switch (VT.getSimpleVT().SimpleTy) {
1770 default: return NULL;
1771 case MVT::v8i8: Opc = ARM::VUZPd8; break;
1772 case MVT::v4i16: Opc = ARM::VUZPd16; break;
1774 case MVT::v2i32: Opc = ARM::VUZPd32; break;
1775 case MVT::v16i8: Opc = ARM::VUZPq8; break;
1776 case MVT::v8i16: Opc = ARM::VUZPq16; break;
1778 case MVT::v4i32: Opc = ARM::VUZPq32; break;
1780 SDValue Pred = getAL(CurDAG);
1781 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1782 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
1783 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
1785 case ARMISD::VTRN: {
1787 EVT VT = N->getValueType(0);
1788 switch (VT.getSimpleVT().SimpleTy) {
1789 default: return NULL;
1790 case MVT::v8i8: Opc = ARM::VTRNd8; break;
1791 case MVT::v4i16: Opc = ARM::VTRNd16; break;
1793 case MVT::v2i32: Opc = ARM::VTRNd32; break;
1794 case MVT::v16i8: Opc = ARM::VTRNq8; break;
1795 case MVT::v8i16: Opc = ARM::VTRNq16; break;
1797 case MVT::v4i32: Opc = ARM::VTRNq32; break;
1799 SDValue Pred = getAL(CurDAG);
1800 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1801 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
1802 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
1805 case ISD::INTRINSIC_VOID:
1806 case ISD::INTRINSIC_W_CHAIN: {
1807 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
1812 case Intrinsic::arm_neon_vld1: {
1813 unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
1814 ARM::VLD1d32, ARM::VLD1d64 };
1815 unsigned QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
1816 ARM::VLD1q32, ARM::VLD1q64 };
1817 return SelectVLD(N, 1, DOpcodes, QOpcodes, 0);
1820 case Intrinsic::arm_neon_vld2: {
1821 unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
1822 ARM::VLD2d32, ARM::VLD1q64 };
1823 unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
1824 return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
1827 case Intrinsic::arm_neon_vld3: {
1828 unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
1829 ARM::VLD3d32, ARM::VLD1d64T };
1830 unsigned QOpcodes0[] = { ARM::VLD3q8_UPD,
1833 unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD,
1834 ARM::VLD3q16odd_UPD,
1835 ARM::VLD3q32odd_UPD };
1836 return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
1839 case Intrinsic::arm_neon_vld4: {
1840 unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
1841 ARM::VLD4d32, ARM::VLD1d64Q };
1842 unsigned QOpcodes0[] = { ARM::VLD4q8_UPD,
1845 unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD,
1846 ARM::VLD4q16odd_UPD,
1847 ARM::VLD4q32odd_UPD };
1848 return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
1851 case Intrinsic::arm_neon_vld2lane: {
1852 unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
1853 unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 };
1854 unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd };
1855 return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
1858 case Intrinsic::arm_neon_vld3lane: {
1859 unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
1860 unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 };
1861 unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd };
1862 return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
1865 case Intrinsic::arm_neon_vld4lane: {
1866 unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
1867 unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 };
1868 unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd };
1869 return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
1872 case Intrinsic::arm_neon_vst1: {
1873 unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
1874 ARM::VST1d32, ARM::VST1d64 };
1875 unsigned QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
1876 ARM::VST1q32, ARM::VST1q64 };
1877 return SelectVST(N, 1, DOpcodes, QOpcodes, 0);
1880 case Intrinsic::arm_neon_vst2: {
1881 unsigned DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
1882 ARM::VST2d32, ARM::VST1q64 };
1883 unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
1884 return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
1887 case Intrinsic::arm_neon_vst3: {
1888 unsigned DOpcodes[] = { ARM::VST3d8, ARM::VST3d16,
1889 ARM::VST3d32, ARM::VST1d64T };
1890 unsigned QOpcodes0[] = { ARM::VST3q8_UPD,
1893 unsigned QOpcodes1[] = { ARM::VST3q8odd_UPD,
1894 ARM::VST3q16odd_UPD,
1895 ARM::VST3q32odd_UPD };
1896 return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
1899 case Intrinsic::arm_neon_vst4: {
1900 unsigned DOpcodes[] = { ARM::VST4d8, ARM::VST4d16,
1901 ARM::VST4d32, ARM::VST1d64Q };
1902 unsigned QOpcodes0[] = { ARM::VST4q8_UPD,
1905 unsigned QOpcodes1[] = { ARM::VST4q8odd_UPD,
1906 ARM::VST4q16odd_UPD,
1907 ARM::VST4q32odd_UPD };
1908 return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
1911 case Intrinsic::arm_neon_vst2lane: {
1912 unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
1913 unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 };
1914 unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd };
1915 return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
1918 case Intrinsic::arm_neon_vst3lane: {
1919 unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
1920 unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 };
1921 unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd };
1922 return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
1925 case Intrinsic::arm_neon_vst4lane: {
1926 unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
1927 unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 };
1928 unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd };
1929 return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
1935 return SelectCode(N);
1938 bool ARMDAGToDAGISel::
1939 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
1940 std::vector<SDValue> &OutOps) {
1941 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
1942 // Require the address to be in a register. That is safe for all ARM
1943 // variants and it is hard to do anything much smarter without knowing
1944 // how the operand is used.
1945 OutOps.push_back(Op);
1949 /// createARMISelDag - This pass converts a legalized DAG into a
1950 /// ARM-specific DAG, ready for instruction scheduling.
1952 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
1953 CodeGenOpt::Level OptLevel) {
1954 return new ARMDAGToDAGISel(TM, OptLevel);