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 /// SelectDYN_ALLOC - Select dynamic alloc for Thumb.
125 SDNode *SelectDYN_ALLOC(SDNode *N);
127 /// SelectVLD - Select NEON load intrinsics. NumVecs should be
128 /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for
129 /// loads of D registers and even subregs and odd subregs of Q registers.
130 /// For NumVecs <= 2, QOpcodes1 is not used.
131 SDNode *SelectVLD(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
132 unsigned *QOpcodes0, unsigned *QOpcodes1);
134 /// SelectVST - Select NEON store intrinsics. NumVecs should
135 /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for
136 /// stores of D registers and even subregs and odd subregs of Q registers.
137 /// For NumVecs <= 2, QOpcodes1 is not used.
138 SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
139 unsigned *QOpcodes0, unsigned *QOpcodes1);
141 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
142 /// be 2, 3 or 4. The opcode arrays specify the instructions used for
143 /// load/store of D registers and even subregs and odd subregs of Q registers.
144 SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, unsigned NumVecs,
145 unsigned *DOpcodes, unsigned *QOpcodes0,
146 unsigned *QOpcodes1);
148 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
149 SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, unsigned Opc);
151 /// SelectCMOVOp - Select CMOV instructions for ARM.
152 SDNode *SelectCMOVOp(SDNode *N);
153 SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
154 ARMCC::CondCodes CCVal, SDValue CCR,
156 SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
157 ARMCC::CondCodes CCVal, SDValue CCR,
159 SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
160 ARMCC::CondCodes CCVal, SDValue CCR,
162 SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
163 ARMCC::CondCodes CCVal, SDValue CCR,
166 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
167 /// inline asm expressions.
168 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
170 std::vector<SDValue> &OutOps);
172 /// PairDRegs - Insert a pair of double registers into an implicit def to
173 /// form a quad register.
174 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
178 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
179 /// operand. If so Imm will receive the 32-bit value.
180 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
181 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
182 Imm = cast<ConstantSDNode>(N)->getZExtValue();
188 // isInt32Immediate - This method tests to see if a constant operand.
189 // If so Imm will receive the 32 bit value.
190 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
191 return isInt32Immediate(N.getNode(), Imm);
194 // isOpcWithIntImmediate - This method tests to see if the node is a specific
195 // opcode and that it has a immediate integer right operand.
196 // If so Imm will receive the 32 bit value.
197 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
198 return N->getOpcode() == Opc &&
199 isInt32Immediate(N->getOperand(1).getNode(), Imm);
203 bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
208 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
210 // Don't match base register only case. That is matched to a separate
211 // lower complexity pattern with explicit register operand.
212 if (ShOpcVal == ARM_AM::no_shift) return false;
214 BaseReg = N.getOperand(0);
215 unsigned ShImmVal = 0;
216 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
217 ShReg = CurDAG->getRegister(0, MVT::i32);
218 ShImmVal = RHS->getZExtValue() & 31;
220 ShReg = N.getOperand(1);
222 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
227 bool ARMDAGToDAGISel::SelectAddrMode2(SDNode *Op, SDValue N,
228 SDValue &Base, SDValue &Offset,
230 if (N.getOpcode() == ISD::MUL) {
231 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
232 // X * [3,5,9] -> X + X * [2,4,8] etc.
233 int RHSC = (int)RHS->getZExtValue();
236 ARM_AM::AddrOpc AddSub = ARM_AM::add;
238 AddSub = ARM_AM::sub;
241 if (isPowerOf2_32(RHSC)) {
242 unsigned ShAmt = Log2_32(RHSC);
243 Base = Offset = N.getOperand(0);
244 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
253 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
255 if (N.getOpcode() == ISD::FrameIndex) {
256 int FI = cast<FrameIndexSDNode>(N)->getIndex();
257 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
258 } else if (N.getOpcode() == ARMISD::Wrapper &&
259 !(Subtarget->useMovt() &&
260 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
261 Base = N.getOperand(0);
263 Offset = CurDAG->getRegister(0, MVT::i32);
264 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
270 // Match simple R +/- imm12 operands.
271 if (N.getOpcode() == ISD::ADD)
272 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
273 int RHSC = (int)RHS->getZExtValue();
274 if ((RHSC >= 0 && RHSC < 0x1000) ||
275 (RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
276 Base = N.getOperand(0);
277 if (Base.getOpcode() == ISD::FrameIndex) {
278 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
279 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
281 Offset = CurDAG->getRegister(0, MVT::i32);
283 ARM_AM::AddrOpc AddSub = ARM_AM::add;
285 AddSub = ARM_AM::sub;
288 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
295 // Otherwise this is R +/- [possibly shifted] R.
296 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
297 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
300 Base = N.getOperand(0);
301 Offset = N.getOperand(1);
303 if (ShOpcVal != ARM_AM::no_shift) {
304 // Check to see if the RHS of the shift is a constant, if not, we can't fold
306 if (ConstantSDNode *Sh =
307 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
308 ShAmt = Sh->getZExtValue();
309 Offset = N.getOperand(1).getOperand(0);
311 ShOpcVal = ARM_AM::no_shift;
315 // Try matching (R shl C) + (R).
316 if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
317 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
318 if (ShOpcVal != ARM_AM::no_shift) {
319 // Check to see if the RHS of the shift is a constant, if not, we can't
321 if (ConstantSDNode *Sh =
322 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
323 ShAmt = Sh->getZExtValue();
324 Offset = N.getOperand(0).getOperand(0);
325 Base = N.getOperand(1);
327 ShOpcVal = ARM_AM::no_shift;
332 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
337 bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
338 SDValue &Offset, SDValue &Opc) {
339 unsigned Opcode = Op->getOpcode();
340 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
341 ? cast<LoadSDNode>(Op)->getAddressingMode()
342 : cast<StoreSDNode>(Op)->getAddressingMode();
343 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
344 ? ARM_AM::add : ARM_AM::sub;
345 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
346 int Val = (int)C->getZExtValue();
347 if (Val >= 0 && Val < 0x1000) { // 12 bits.
348 Offset = CurDAG->getRegister(0, MVT::i32);
349 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
357 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
359 if (ShOpcVal != ARM_AM::no_shift) {
360 // Check to see if the RHS of the shift is a constant, if not, we can't fold
362 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
363 ShAmt = Sh->getZExtValue();
364 Offset = N.getOperand(0);
366 ShOpcVal = ARM_AM::no_shift;
370 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
376 bool ARMDAGToDAGISel::SelectAddrMode3(SDNode *Op, SDValue N,
377 SDValue &Base, SDValue &Offset,
379 if (N.getOpcode() == ISD::SUB) {
380 // X - C is canonicalize to X + -C, no need to handle it here.
381 Base = N.getOperand(0);
382 Offset = N.getOperand(1);
383 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
387 if (N.getOpcode() != ISD::ADD) {
389 if (N.getOpcode() == ISD::FrameIndex) {
390 int FI = cast<FrameIndexSDNode>(N)->getIndex();
391 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
393 Offset = CurDAG->getRegister(0, MVT::i32);
394 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
398 // If the RHS is +/- imm8, fold into addr mode.
399 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
400 int RHSC = (int)RHS->getZExtValue();
401 if ((RHSC >= 0 && RHSC < 256) ||
402 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
403 Base = N.getOperand(0);
404 if (Base.getOpcode() == ISD::FrameIndex) {
405 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
406 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
408 Offset = CurDAG->getRegister(0, MVT::i32);
410 ARM_AM::AddrOpc AddSub = ARM_AM::add;
412 AddSub = ARM_AM::sub;
415 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
420 Base = N.getOperand(0);
421 Offset = N.getOperand(1);
422 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
426 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
427 SDValue &Offset, SDValue &Opc) {
428 unsigned Opcode = Op->getOpcode();
429 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
430 ? cast<LoadSDNode>(Op)->getAddressingMode()
431 : cast<StoreSDNode>(Op)->getAddressingMode();
432 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
433 ? ARM_AM::add : ARM_AM::sub;
434 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
435 int Val = (int)C->getZExtValue();
436 if (Val >= 0 && Val < 256) {
437 Offset = CurDAG->getRegister(0, MVT::i32);
438 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
444 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
448 bool ARMDAGToDAGISel::SelectAddrMode4(SDNode *Op, SDValue N,
449 SDValue &Addr, SDValue &Mode) {
451 Mode = CurDAG->getTargetConstant(0, MVT::i32);
455 bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N,
456 SDValue &Base, SDValue &Offset) {
457 if (N.getOpcode() != ISD::ADD) {
459 if (N.getOpcode() == ISD::FrameIndex) {
460 int FI = cast<FrameIndexSDNode>(N)->getIndex();
461 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
462 } else if (N.getOpcode() == ARMISD::Wrapper &&
463 !(Subtarget->useMovt() &&
464 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
465 Base = N.getOperand(0);
467 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
472 // If the RHS is +/- imm8, fold into addr mode.
473 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
474 int RHSC = (int)RHS->getZExtValue();
475 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied by 4.
477 if ((RHSC >= 0 && RHSC < 256) ||
478 (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
479 Base = N.getOperand(0);
480 if (Base.getOpcode() == ISD::FrameIndex) {
481 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
482 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
485 ARM_AM::AddrOpc AddSub = ARM_AM::add;
487 AddSub = ARM_AM::sub;
490 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
498 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
503 bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N,
504 SDValue &Addr, SDValue &Align) {
506 // Default to no alignment.
507 Align = CurDAG->getTargetConstant(0, MVT::i32);
511 bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
512 SDValue &Offset, SDValue &Label) {
513 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
514 Offset = N.getOperand(0);
515 SDValue N1 = N.getOperand(1);
516 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
523 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
524 SDValue &Base, SDValue &Offset){
525 // FIXME dl should come from the parent load or store, not the address
526 DebugLoc dl = Op->getDebugLoc();
527 if (N.getOpcode() != ISD::ADD) {
528 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
529 if (!NC || NC->getZExtValue() != 0)
536 Base = N.getOperand(0);
537 Offset = N.getOperand(1);
542 ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
543 unsigned Scale, SDValue &Base,
544 SDValue &OffImm, SDValue &Offset) {
546 SDValue TmpBase, TmpOffImm;
547 if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
548 return false; // We want to select tLDRspi / tSTRspi instead.
549 if (N.getOpcode() == ARMISD::Wrapper &&
550 N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
551 return false; // We want to select tLDRpci instead.
554 if (N.getOpcode() != ISD::ADD) {
555 if (N.getOpcode() == ARMISD::Wrapper &&
556 !(Subtarget->useMovt() &&
557 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
558 Base = N.getOperand(0);
562 Offset = CurDAG->getRegister(0, MVT::i32);
563 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
567 // Thumb does not have [sp, r] address mode.
568 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
569 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
570 if ((LHSR && LHSR->getReg() == ARM::SP) ||
571 (RHSR && RHSR->getReg() == ARM::SP)) {
573 Offset = CurDAG->getRegister(0, MVT::i32);
574 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
578 // If the RHS is + imm5 * scale, fold into addr mode.
579 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
580 int RHSC = (int)RHS->getZExtValue();
581 if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
583 if (RHSC >= 0 && RHSC < 32) {
584 Base = N.getOperand(0);
585 Offset = CurDAG->getRegister(0, MVT::i32);
586 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
592 Base = N.getOperand(0);
593 Offset = N.getOperand(1);
594 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
598 bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
599 SDValue &Base, SDValue &OffImm,
601 return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
604 bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
605 SDValue &Base, SDValue &OffImm,
607 return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
610 bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
611 SDValue &Base, SDValue &OffImm,
613 return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
616 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
617 SDValue &Base, SDValue &OffImm) {
618 if (N.getOpcode() == ISD::FrameIndex) {
619 int FI = cast<FrameIndexSDNode>(N)->getIndex();
620 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
621 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
625 if (N.getOpcode() != ISD::ADD)
628 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
629 if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
630 (LHSR && LHSR->getReg() == ARM::SP)) {
631 // If the RHS is + imm8 * scale, fold into addr mode.
632 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
633 int RHSC = (int)RHS->getZExtValue();
634 if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
636 if (RHSC >= 0 && RHSC < 256) {
637 Base = N.getOperand(0);
638 if (Base.getOpcode() == ISD::FrameIndex) {
639 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
640 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
642 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
652 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
655 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
657 // Don't match base register only case. That is matched to a separate
658 // lower complexity pattern with explicit register operand.
659 if (ShOpcVal == ARM_AM::no_shift) return false;
661 BaseReg = N.getOperand(0);
662 unsigned ShImmVal = 0;
663 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
664 ShImmVal = RHS->getZExtValue() & 31;
665 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
672 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
673 SDValue &Base, SDValue &OffImm) {
674 // Match simple R + imm12 operands.
677 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
678 if (N.getOpcode() == ISD::FrameIndex) {
679 // Match frame index...
680 int FI = cast<FrameIndexSDNode>(N)->getIndex();
681 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
682 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
684 } else if (N.getOpcode() == ARMISD::Wrapper &&
685 !(Subtarget->useMovt() &&
686 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
687 Base = N.getOperand(0);
688 if (Base.getOpcode() == ISD::TargetConstantPool)
689 return false; // We want to select t2LDRpci instead.
692 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
696 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
697 if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
698 // Let t2LDRi8 handle (R - imm8).
701 int RHSC = (int)RHS->getZExtValue();
702 if (N.getOpcode() == ISD::SUB)
705 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
706 Base = N.getOperand(0);
707 if (Base.getOpcode() == ISD::FrameIndex) {
708 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
709 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
711 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
718 OffImm = CurDAG->getTargetConstant(0, MVT::i32);
722 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
723 SDValue &Base, SDValue &OffImm) {
724 // Match simple R - imm8 operands.
725 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
726 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
727 int RHSC = (int)RHS->getSExtValue();
728 if (N.getOpcode() == ISD::SUB)
731 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
732 Base = N.getOperand(0);
733 if (Base.getOpcode() == ISD::FrameIndex) {
734 int FI = cast<FrameIndexSDNode>(Base)->getIndex();
735 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
737 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
746 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
748 unsigned Opcode = Op->getOpcode();
749 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
750 ? cast<LoadSDNode>(Op)->getAddressingMode()
751 : cast<StoreSDNode>(Op)->getAddressingMode();
752 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
753 int RHSC = (int)RHS->getZExtValue();
754 if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
755 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
756 ? CurDAG->getTargetConstant(RHSC, MVT::i32)
757 : CurDAG->getTargetConstant(-RHSC, MVT::i32);
765 bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
766 SDValue &Base, SDValue &OffImm) {
767 if (N.getOpcode() == ISD::ADD) {
768 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
769 int RHSC = (int)RHS->getZExtValue();
770 if (((RHSC & 0x3) == 0) &&
771 ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits.
772 Base = N.getOperand(0);
773 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
777 } else if (N.getOpcode() == ISD::SUB) {
778 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
779 int RHSC = (int)RHS->getZExtValue();
780 if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) { // 8 bits.
781 Base = N.getOperand(0);
782 OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
791 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDNode *Op, SDValue N,
793 SDValue &OffReg, SDValue &ShImm) {
794 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
795 if (N.getOpcode() != ISD::ADD)
798 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
799 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
800 int RHSC = (int)RHS->getZExtValue();
801 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
803 else if (RHSC < 0 && RHSC >= -255) // 8 bits
807 // Look for (R + R) or (R + (R << [1,2,3])).
809 Base = N.getOperand(0);
810 OffReg = N.getOperand(1);
812 // Swap if it is ((R << c) + R).
813 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
814 if (ShOpcVal != ARM_AM::lsl) {
815 ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
816 if (ShOpcVal == ARM_AM::lsl)
817 std::swap(Base, OffReg);
820 if (ShOpcVal == ARM_AM::lsl) {
821 // Check to see if the RHS of the shift is a constant, if not, we can't fold
823 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
824 ShAmt = Sh->getZExtValue();
827 ShOpcVal = ARM_AM::no_shift;
829 OffReg = OffReg.getOperand(0);
831 ShOpcVal = ARM_AM::no_shift;
835 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
840 //===--------------------------------------------------------------------===//
842 /// getAL - Returns a ARMCC::AL immediate node.
843 static inline SDValue getAL(SelectionDAG *CurDAG) {
844 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
847 SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
848 LoadSDNode *LD = cast<LoadSDNode>(N);
849 ISD::MemIndexedMode AM = LD->getAddressingMode();
850 if (AM == ISD::UNINDEXED)
853 EVT LoadedVT = LD->getMemoryVT();
854 SDValue Offset, AMOpc;
855 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
858 if (LoadedVT == MVT::i32 &&
859 SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
860 Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
862 } else if (LoadedVT == MVT::i16 &&
863 SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
865 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
866 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
867 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
868 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
869 if (LD->getExtensionType() == ISD::SEXTLOAD) {
870 if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
872 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
875 if (SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
877 Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
883 SDValue Chain = LD->getChain();
884 SDValue Base = LD->getBasePtr();
885 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
886 CurDAG->getRegister(0, MVT::i32), Chain };
887 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
894 SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
895 LoadSDNode *LD = cast<LoadSDNode>(N);
896 ISD::MemIndexedMode AM = LD->getAddressingMode();
897 if (AM == ISD::UNINDEXED)
900 EVT LoadedVT = LD->getMemoryVT();
901 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
903 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
906 if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
907 switch (LoadedVT.getSimpleVT().SimpleTy) {
909 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
913 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
915 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
920 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
922 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
931 SDValue Chain = LD->getChain();
932 SDValue Base = LD->getBasePtr();
933 SDValue Ops[]= { Base, Offset, getAL(CurDAG),
934 CurDAG->getRegister(0, MVT::i32), Chain };
935 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
942 SDNode *ARMDAGToDAGISel::SelectDYN_ALLOC(SDNode *N) {
943 DebugLoc dl = N->getDebugLoc();
944 EVT VT = N->getValueType(0);
945 SDValue Chain = N->getOperand(0);
946 SDValue Size = N->getOperand(1);
947 SDValue Align = N->getOperand(2);
948 SDValue SP = CurDAG->getCopyFromReg(Chain, dl, ARM::SP, MVT::i32);
949 int32_t AlignVal = cast<ConstantSDNode>(Align)->getSExtValue();
951 // We need to align the stack. Use Thumb1 tAND which is the only thumb
952 // instruction that can read and write SP. This matches to a pseudo
953 // instruction that has a chain to ensure the result is written back to
954 // the stack pointer.
955 SP = SDValue(CurDAG->getMachineNode(ARM::tANDsp, dl, VT, SP, Align), 0);
957 bool isC = isa<ConstantSDNode>(Size);
958 uint32_t C = isC ? cast<ConstantSDNode>(Size)->getZExtValue() : ~0UL;
959 // Handle the most common case for both Thumb1 and Thumb2:
960 // tSUBspi - immediate is between 0 ... 508 inclusive.
961 if (C <= 508 && ((C & 3) == 0))
962 // FIXME: tSUBspi encode scale 4 implicitly.
963 return CurDAG->SelectNodeTo(N, ARM::tSUBspi_, VT, MVT::Other, SP,
964 CurDAG->getTargetConstant(C/4, MVT::i32),
967 if (Subtarget->isThumb1Only()) {
968 // Use tADDspr since Thumb1 does not have a sub r, sp, r. ARMISelLowering
969 // should have negated the size operand already. FIXME: We can't insert
970 // new target independent node at this stage so we are forced to negate
971 // it earlier. Is there a better solution?
972 return CurDAG->SelectNodeTo(N, ARM::tADDspr_, VT, MVT::Other, SP, Size,
974 } else if (Subtarget->isThumb2()) {
975 if (isC && Predicate_t2_so_imm(Size.getNode())) {
977 SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
978 return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi_, VT, MVT::Other, Ops, 3);
979 } else if (isC && Predicate_imm0_4095(Size.getNode())) {
981 SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
982 return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi12_, VT, MVT::Other, Ops, 3);
985 SDValue Ops[] = { SP, Size,
986 getI32Imm(ARM_AM::getSORegOpc(ARM_AM::lsl,0)), Chain };
987 return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPs_, VT, MVT::Other, Ops, 4);
991 // FIXME: Add ADD / SUB sp instructions for ARM.
995 /// PairDRegs - Insert a pair of double registers into an implicit def to
996 /// form a quad register.
997 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
998 DebugLoc dl = V0.getNode()->getDebugLoc();
1000 SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0);
1001 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
1002 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
1003 SDNode *Pair = CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
1004 VT, Undef, V0, SubReg0);
1005 return CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
1006 VT, SDValue(Pair, 0), V1, SubReg1);
1009 /// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
1010 /// for a 64-bit subregister of the vector.
1011 static EVT GetNEONSubregVT(EVT VT) {
1012 switch (VT.getSimpleVT().SimpleTy) {
1013 default: llvm_unreachable("unhandled NEON type");
1014 case MVT::v16i8: return MVT::v8i8;
1015 case MVT::v8i16: return MVT::v4i16;
1016 case MVT::v4f32: return MVT::v2f32;
1017 case MVT::v4i32: return MVT::v2i32;
1018 case MVT::v2i64: return MVT::v1i64;
1022 SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
1023 unsigned *DOpcodes, unsigned *QOpcodes0,
1024 unsigned *QOpcodes1) {
1025 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1026 DebugLoc dl = N->getDebugLoc();
1028 SDValue MemAddr, Align;
1029 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1032 SDValue Chain = N->getOperand(0);
1033 EVT VT = N->getValueType(0);
1034 bool is64BitVector = VT.is64BitVector();
1036 unsigned OpcodeIndex;
1037 switch (VT.getSimpleVT().SimpleTy) {
1038 default: llvm_unreachable("unhandled vld type");
1039 // Double-register operations:
1040 case MVT::v8i8: OpcodeIndex = 0; break;
1041 case MVT::v4i16: OpcodeIndex = 1; break;
1043 case MVT::v2i32: OpcodeIndex = 2; break;
1044 case MVT::v1i64: OpcodeIndex = 3; break;
1045 // Quad-register operations:
1046 case MVT::v16i8: OpcodeIndex = 0; break;
1047 case MVT::v8i16: OpcodeIndex = 1; break;
1049 case MVT::v4i32: OpcodeIndex = 2; break;
1050 case MVT::v2i64: OpcodeIndex = 3;
1051 assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
1055 SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1056 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1057 if (is64BitVector) {
1058 unsigned Opc = DOpcodes[OpcodeIndex];
1059 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1060 std::vector<EVT> ResTys(NumVecs, VT);
1061 ResTys.push_back(MVT::Other);
1062 return CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1065 EVT RegVT = GetNEONSubregVT(VT);
1067 // Quad registers are directly supported for VLD1 and VLD2,
1068 // loading pairs of D regs.
1069 unsigned Opc = QOpcodes0[OpcodeIndex];
1070 const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1071 std::vector<EVT> ResTys(2 * NumVecs, RegVT);
1072 ResTys.push_back(MVT::Other);
1073 SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1074 Chain = SDValue(VLd, 2 * NumVecs);
1076 // Combine the even and odd subregs to produce the result.
1077 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1078 SDNode *Q = PairDRegs(VT, SDValue(VLd, 2*Vec), SDValue(VLd, 2*Vec+1));
1079 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1082 // Otherwise, quad registers are loaded with two separate instructions,
1083 // where one loads the even registers and the other loads the odd registers.
1085 std::vector<EVT> ResTys(NumVecs, RegVT);
1086 ResTys.push_back(MemAddr.getValueType());
1087 ResTys.push_back(MVT::Other);
1089 // Load the even subregs.
1090 unsigned Opc = QOpcodes0[OpcodeIndex];
1091 const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain };
1092 SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6);
1093 Chain = SDValue(VLdA, NumVecs+1);
1095 // Load the odd subregs.
1096 Opc = QOpcodes1[OpcodeIndex];
1097 const SDValue OpsB[] = { SDValue(VLdA, NumVecs),
1098 Align, Reg0, Pred, Reg0, Chain };
1099 SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
1100 Chain = SDValue(VLdB, NumVecs+1);
1102 // Combine the even and odd subregs to produce the result.
1103 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1104 SDNode *Q = PairDRegs(VT, SDValue(VLdA, Vec), SDValue(VLdB, Vec));
1105 ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1108 ReplaceUses(SDValue(N, NumVecs), Chain);
1112 SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
1113 unsigned *DOpcodes, unsigned *QOpcodes0,
1114 unsigned *QOpcodes1) {
1115 assert(NumVecs >=1 && NumVecs <= 4 && "VST NumVecs out-of-range");
1116 DebugLoc dl = N->getDebugLoc();
1118 SDValue MemAddr, Align;
1119 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1122 SDValue Chain = N->getOperand(0);
1123 EVT VT = N->getOperand(3).getValueType();
1124 bool is64BitVector = VT.is64BitVector();
1126 unsigned OpcodeIndex;
1127 switch (VT.getSimpleVT().SimpleTy) {
1128 default: llvm_unreachable("unhandled vst type");
1129 // Double-register operations:
1130 case MVT::v8i8: OpcodeIndex = 0; break;
1131 case MVT::v4i16: OpcodeIndex = 1; break;
1133 case MVT::v2i32: OpcodeIndex = 2; break;
1134 case MVT::v1i64: OpcodeIndex = 3; break;
1135 // Quad-register operations:
1136 case MVT::v16i8: OpcodeIndex = 0; break;
1137 case MVT::v8i16: OpcodeIndex = 1; break;
1139 case MVT::v4i32: OpcodeIndex = 2; break;
1140 case MVT::v2i64: OpcodeIndex = 3;
1141 assert(NumVecs == 1 && "v2i64 type only supported for VST1");
1145 SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1146 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1148 SmallVector<SDValue, 10> Ops;
1149 Ops.push_back(MemAddr);
1150 Ops.push_back(Align);
1152 if (is64BitVector) {
1153 unsigned Opc = DOpcodes[OpcodeIndex];
1154 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1155 Ops.push_back(N->getOperand(Vec+3));
1156 Ops.push_back(Pred);
1157 Ops.push_back(Reg0); // predicate register
1158 Ops.push_back(Chain);
1159 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+5);
1162 EVT RegVT = GetNEONSubregVT(VT);
1164 // Quad registers are directly supported for VST1 and VST2,
1165 // storing pairs of D regs.
1166 unsigned Opc = QOpcodes0[OpcodeIndex];
1167 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1168 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1169 N->getOperand(Vec+3)));
1170 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1171 N->getOperand(Vec+3)));
1173 Ops.push_back(Pred);
1174 Ops.push_back(Reg0); // predicate register
1175 Ops.push_back(Chain);
1176 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
1180 // Otherwise, quad registers are stored with two separate instructions,
1181 // where one stores the even registers and the other stores the odd registers.
1183 Ops.push_back(Reg0); // post-access address offset
1185 // Store the even subregs.
1186 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1187 Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1188 N->getOperand(Vec+3)));
1189 Ops.push_back(Pred);
1190 Ops.push_back(Reg0); // predicate register
1191 Ops.push_back(Chain);
1192 unsigned Opc = QOpcodes0[OpcodeIndex];
1193 SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1194 MVT::Other, Ops.data(), NumVecs+6);
1195 Chain = SDValue(VStA, 1);
1197 // Store the odd subregs.
1198 Ops[0] = SDValue(VStA, 0); // MemAddr
1199 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1200 Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1201 N->getOperand(Vec+3));
1202 Ops[NumVecs+5] = Chain;
1203 Opc = QOpcodes1[OpcodeIndex];
1204 SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1205 MVT::Other, Ops.data(), NumVecs+6);
1206 Chain = SDValue(VStB, 1);
1207 ReplaceUses(SDValue(N, 0), Chain);
1211 SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1212 unsigned NumVecs, unsigned *DOpcodes,
1213 unsigned *QOpcodes0,
1214 unsigned *QOpcodes1) {
1215 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1216 DebugLoc dl = N->getDebugLoc();
1218 SDValue MemAddr, Align;
1219 if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1222 SDValue Chain = N->getOperand(0);
1224 cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1225 EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
1226 bool is64BitVector = VT.is64BitVector();
1228 // Quad registers are handled by load/store of subregs. Find the subreg info.
1229 unsigned NumElts = 0;
1232 if (!is64BitVector) {
1233 RegVT = GetNEONSubregVT(VT);
1234 NumElts = RegVT.getVectorNumElements();
1235 SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
1238 unsigned OpcodeIndex;
1239 switch (VT.getSimpleVT().SimpleTy) {
1240 default: llvm_unreachable("unhandled vld/vst lane type");
1241 // Double-register operations:
1242 case MVT::v8i8: OpcodeIndex = 0; break;
1243 case MVT::v4i16: OpcodeIndex = 1; break;
1245 case MVT::v2i32: OpcodeIndex = 2; break;
1246 // Quad-register operations:
1247 case MVT::v8i16: OpcodeIndex = 0; break;
1249 case MVT::v4i32: OpcodeIndex = 1; break;
1252 SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1253 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1255 SmallVector<SDValue, 10> Ops;
1256 Ops.push_back(MemAddr);
1257 Ops.push_back(Align);
1260 if (is64BitVector) {
1261 Opc = DOpcodes[OpcodeIndex];
1262 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1263 Ops.push_back(N->getOperand(Vec+3));
1265 // Check if this is loading the even or odd subreg of a Q register.
1266 if (Lane < NumElts) {
1267 Opc = QOpcodes0[OpcodeIndex];
1270 Opc = QOpcodes1[OpcodeIndex];
1272 // Extract the subregs of the input vector.
1273 for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1274 Ops.push_back(CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1275 N->getOperand(Vec+3)));
1277 Ops.push_back(getI32Imm(Lane));
1278 Ops.push_back(Pred);
1279 Ops.push_back(Reg0);
1280 Ops.push_back(Chain);
1283 return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6);
1285 std::vector<EVT> ResTys(NumVecs, RegVT);
1286 ResTys.push_back(MVT::Other);
1288 CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), NumVecs+6);
1289 // For a 64-bit vector load to D registers, nothing more needs to be done.
1293 // For 128-bit vectors, take the 64-bit results of the load and insert them
1294 // as subregs into the result.
1295 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1296 SDValue QuadVec = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
1297 N->getOperand(Vec+3),
1298 SDValue(VLdLn, Vec));
1299 ReplaceUses(SDValue(N, Vec), QuadVec);
1302 Chain = SDValue(VLdLn, NumVecs);
1303 ReplaceUses(SDValue(N, NumVecs), Chain);
1307 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
1309 if (!Subtarget->hasV6T2Ops())
1312 unsigned Shl_imm = 0;
1313 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
1314 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1315 unsigned Srl_imm = 0;
1316 if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
1317 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1318 unsigned Width = 32 - Srl_imm;
1319 int LSB = Srl_imm - Shl_imm;
1322 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1323 SDValue Ops[] = { N->getOperand(0).getOperand(0),
1324 CurDAG->getTargetConstant(LSB, MVT::i32),
1325 CurDAG->getTargetConstant(Width, MVT::i32),
1326 getAL(CurDAG), Reg0 };
1327 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1333 SDNode *ARMDAGToDAGISel::
1334 SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1335 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1338 if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
1339 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1340 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1343 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1344 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1345 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1346 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1348 llvm_unreachable("Unknown so_reg opcode!");
1352 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1353 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1354 SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
1355 return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
1360 SDNode *ARMDAGToDAGISel::
1361 SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1362 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1366 if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
1367 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1368 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
1369 return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
1374 SDNode *ARMDAGToDAGISel::
1375 SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1376 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1377 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1381 if (Predicate_t2_so_imm(TrueVal.getNode())) {
1382 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1383 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1384 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1385 return CurDAG->SelectNodeTo(N,
1386 ARM::t2MOVCCi, MVT::i32, Ops, 5);
1391 SDNode *ARMDAGToDAGISel::
1392 SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1393 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1394 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1398 if (Predicate_so_imm(TrueVal.getNode())) {
1399 SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1400 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1401 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1402 return CurDAG->SelectNodeTo(N,
1403 ARM::MOVCCi, MVT::i32, Ops, 5);
1408 SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
1409 EVT VT = N->getValueType(0);
1410 SDValue FalseVal = N->getOperand(0);
1411 SDValue TrueVal = N->getOperand(1);
1412 SDValue CC = N->getOperand(2);
1413 SDValue CCR = N->getOperand(3);
1414 SDValue InFlag = N->getOperand(4);
1415 assert(CC.getOpcode() == ISD::Constant);
1416 assert(CCR.getOpcode() == ISD::Register);
1417 ARMCC::CondCodes CCVal =
1418 (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
1420 if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1421 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1422 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1423 // Pattern complexity = 18 cost = 1 size = 0
1427 if (Subtarget->isThumb()) {
1428 SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
1429 CCVal, CCR, InFlag);
1431 Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
1432 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1436 SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
1437 CCVal, CCR, InFlag);
1439 Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
1440 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1445 // Pattern: (ARMcmov:i32 GPR:i32:$false,
1446 // (imm:i32)<<P:Predicate_so_imm>>:$true,
1448 // Emits: (MOVCCi:i32 GPR:i32:$false,
1449 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1450 // Pattern complexity = 10 cost = 1 size = 0
1451 if (Subtarget->isThumb()) {
1452 SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
1453 CCVal, CCR, InFlag);
1455 Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
1456 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1460 SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
1461 CCVal, CCR, InFlag);
1463 Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
1464 ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1470 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1471 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1472 // Pattern complexity = 6 cost = 1 size = 0
1474 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1475 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1476 // Pattern complexity = 6 cost = 11 size = 0
1478 // Also FCPYScc and FCPYDcc.
1479 SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
1480 SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
1482 switch (VT.getSimpleVT().SimpleTy) {
1483 default: assert(false && "Illegal conditional move type!");
1486 Opc = Subtarget->isThumb()
1487 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1497 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1500 SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
1501 DebugLoc dl = N->getDebugLoc();
1503 if (N->isMachineOpcode())
1504 return NULL; // Already selected.
1506 switch (N->getOpcode()) {
1508 case ISD::Constant: {
1509 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1511 if (Subtarget->hasThumb2())
1512 // Thumb2-aware targets have the MOVT instruction, so all immediates can
1513 // be done with MOV + MOVT, at worst.
1516 if (Subtarget->isThumb()) {
1517 UseCP = (Val > 255 && // MOV
1518 ~Val > 255 && // MOV + MVN
1519 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
1521 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
1522 ARM_AM::getSOImmVal(~Val) == -1 && // MVN
1523 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
1528 CurDAG->getTargetConstantPool(ConstantInt::get(
1529 Type::getInt32Ty(*CurDAG->getContext()), Val),
1530 TLI.getPointerTy());
1533 if (Subtarget->isThumb1Only()) {
1534 SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1535 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1536 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1537 ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1542 CurDAG->getRegister(0, MVT::i32),
1543 CurDAG->getTargetConstant(0, MVT::i32),
1545 CurDAG->getRegister(0, MVT::i32),
1546 CurDAG->getEntryNode()
1548 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1551 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
1555 // Other cases are autogenerated.
1558 case ISD::FrameIndex: {
1559 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1560 int FI = cast<FrameIndexSDNode>(N)->getIndex();
1561 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1562 if (Subtarget->isThumb1Only()) {
1563 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1564 CurDAG->getTargetConstant(0, MVT::i32));
1566 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1567 ARM::t2ADDri : ARM::ADDri);
1568 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1569 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1570 CurDAG->getRegister(0, MVT::i32) };
1571 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1574 case ARMISD::DYN_ALLOC:
1575 return SelectDYN_ALLOC(N);
1577 if (SDNode *I = SelectV6T2BitfieldExtractOp(N,
1578 Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX))
1582 if (SDNode *I = SelectV6T2BitfieldExtractOp(N,
1583 Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX))
1587 if (Subtarget->isThumb1Only())
1589 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1590 unsigned RHSV = C->getZExtValue();
1592 if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
1593 unsigned ShImm = Log2_32(RHSV-1);
1596 SDValue V = N->getOperand(0);
1597 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1598 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1599 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1600 if (Subtarget->isThumb()) {
1601 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1602 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1604 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1605 return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1608 if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
1609 unsigned ShImm = Log2_32(RHSV+1);
1612 SDValue V = N->getOperand(0);
1613 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1614 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1615 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1616 if (Subtarget->isThumb()) {
1617 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0 };
1618 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 5);
1620 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1621 return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1627 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
1628 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
1629 // are entirely contributed by c2 and lower 16-bits are entirely contributed
1630 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
1631 // Select it to: "movt x, ((c1 & 0xffff) >> 16)
1632 EVT VT = N->getValueType(0);
1635 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
1637 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
1640 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
1641 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1644 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
1645 SDValue N2 = N0.getOperand(1);
1646 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1649 unsigned N1CVal = N1C->getZExtValue();
1650 unsigned N2CVal = N2C->getZExtValue();
1651 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
1652 (N1CVal & 0xffffU) == 0xffffU &&
1653 (N2CVal & 0xffffU) == 0x0U) {
1654 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
1656 SDValue Ops[] = { N0.getOperand(0), Imm16,
1657 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1658 return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
1663 case ARMISD::VMOVRRD:
1664 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
1665 N->getOperand(0), getAL(CurDAG),
1666 CurDAG->getRegister(0, MVT::i32));
1667 case ISD::UMUL_LOHI: {
1668 if (Subtarget->isThumb1Only())
1670 if (Subtarget->isThumb()) {
1671 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1672 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1673 CurDAG->getRegister(0, MVT::i32) };
1674 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops,4);
1676 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1677 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1678 CurDAG->getRegister(0, MVT::i32) };
1679 return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1682 case ISD::SMUL_LOHI: {
1683 if (Subtarget->isThumb1Only())
1685 if (Subtarget->isThumb()) {
1686 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1687 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1688 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops,4);
1690 SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1691 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1692 CurDAG->getRegister(0, MVT::i32) };
1693 return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1697 SDNode *ResNode = 0;
1698 if (Subtarget->isThumb() && Subtarget->hasThumb2())
1699 ResNode = SelectT2IndexedLoad(N);
1701 ResNode = SelectARMIndexedLoad(N);
1705 // VLDMQ must be custom-selected for "v2f64 load" to set the AM5Opc value.
1706 if (Subtarget->hasVFP2() &&
1707 N->getValueType(0).getSimpleVT().SimpleTy == MVT::v2f64) {
1708 SDValue Chain = N->getOperand(0);
1710 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
1711 SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1712 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1713 SDValue Ops[] = { N->getOperand(1), AM5Opc, Pred, PredReg, Chain };
1714 return CurDAG->getMachineNode(ARM::VLDMQ, dl, MVT::v2f64, MVT::Other,
1717 // Other cases are autogenerated.
1721 // VSTMQ must be custom-selected for "v2f64 store" to set the AM5Opc value.
1722 if (Subtarget->hasVFP2() &&
1723 N->getOperand(1).getValueType().getSimpleVT().SimpleTy == MVT::v2f64) {
1724 SDValue Chain = N->getOperand(0);
1726 CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
1727 SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1728 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1729 SDValue Ops[] = { N->getOperand(1), N->getOperand(2),
1730 AM5Opc, Pred, PredReg, Chain };
1731 return CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
1733 // Other cases are autogenerated.
1736 case ARMISD::BRCOND: {
1737 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1738 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
1739 // Pattern complexity = 6 cost = 1 size = 0
1741 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1742 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
1743 // Pattern complexity = 6 cost = 1 size = 0
1745 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1746 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
1747 // Pattern complexity = 6 cost = 1 size = 0
1749 unsigned Opc = Subtarget->isThumb() ?
1750 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
1751 SDValue Chain = N->getOperand(0);
1752 SDValue N1 = N->getOperand(1);
1753 SDValue N2 = N->getOperand(2);
1754 SDValue N3 = N->getOperand(3);
1755 SDValue InFlag = N->getOperand(4);
1756 assert(N1.getOpcode() == ISD::BasicBlock);
1757 assert(N2.getOpcode() == ISD::Constant);
1758 assert(N3.getOpcode() == ISD::Register);
1760 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1761 cast<ConstantSDNode>(N2)->getZExtValue()),
1763 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
1764 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
1766 Chain = SDValue(ResNode, 0);
1767 if (N->getNumValues() == 2) {
1768 InFlag = SDValue(ResNode, 1);
1769 ReplaceUses(SDValue(N, 1), InFlag);
1771 ReplaceUses(SDValue(N, 0),
1772 SDValue(Chain.getNode(), Chain.getResNo()));
1776 return SelectCMOVOp(N);
1777 case ARMISD::CNEG: {
1778 EVT VT = N->getValueType(0);
1779 SDValue N0 = N->getOperand(0);
1780 SDValue N1 = N->getOperand(1);
1781 SDValue N2 = N->getOperand(2);
1782 SDValue N3 = N->getOperand(3);
1783 SDValue InFlag = N->getOperand(4);
1784 assert(N2.getOpcode() == ISD::Constant);
1785 assert(N3.getOpcode() == ISD::Register);
1787 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1788 cast<ConstantSDNode>(N2)->getZExtValue()),
1790 SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
1792 switch (VT.getSimpleVT().SimpleTy) {
1793 default: assert(false && "Illegal conditional move type!");
1802 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1805 case ARMISD::VZIP: {
1807 EVT VT = N->getValueType(0);
1808 switch (VT.getSimpleVT().SimpleTy) {
1809 default: return NULL;
1810 case MVT::v8i8: Opc = ARM::VZIPd8; break;
1811 case MVT::v4i16: Opc = ARM::VZIPd16; break;
1813 case MVT::v2i32: Opc = ARM::VZIPd32; break;
1814 case MVT::v16i8: Opc = ARM::VZIPq8; break;
1815 case MVT::v8i16: Opc = ARM::VZIPq16; break;
1817 case MVT::v4i32: Opc = ARM::VZIPq32; break;
1819 SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1820 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1821 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
1822 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
1824 case ARMISD::VUZP: {
1826 EVT VT = N->getValueType(0);
1827 switch (VT.getSimpleVT().SimpleTy) {
1828 default: return NULL;
1829 case MVT::v8i8: Opc = ARM::VUZPd8; break;
1830 case MVT::v4i16: Opc = ARM::VUZPd16; break;
1832 case MVT::v2i32: Opc = ARM::VUZPd32; break;
1833 case MVT::v16i8: Opc = ARM::VUZPq8; break;
1834 case MVT::v8i16: Opc = ARM::VUZPq16; break;
1836 case MVT::v4i32: Opc = ARM::VUZPq32; break;
1838 SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1839 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1840 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
1841 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
1843 case ARMISD::VTRN: {
1845 EVT VT = N->getValueType(0);
1846 switch (VT.getSimpleVT().SimpleTy) {
1847 default: return NULL;
1848 case MVT::v8i8: Opc = ARM::VTRNd8; break;
1849 case MVT::v4i16: Opc = ARM::VTRNd16; break;
1851 case MVT::v2i32: Opc = ARM::VTRNd32; break;
1852 case MVT::v16i8: Opc = ARM::VTRNq8; break;
1853 case MVT::v8i16: Opc = ARM::VTRNq16; break;
1855 case MVT::v4i32: Opc = ARM::VTRNq32; break;
1857 SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1858 SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1859 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
1860 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
1863 case ISD::INTRINSIC_VOID:
1864 case ISD::INTRINSIC_W_CHAIN: {
1865 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
1870 case Intrinsic::arm_neon_vld1: {
1871 unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
1872 ARM::VLD1d32, ARM::VLD1d64 };
1873 unsigned QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
1874 ARM::VLD1q32, ARM::VLD1q64 };
1875 return SelectVLD(N, 1, DOpcodes, QOpcodes, 0);
1878 case Intrinsic::arm_neon_vld2: {
1879 unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
1880 ARM::VLD2d32, ARM::VLD1q64 };
1881 unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
1882 return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
1885 case Intrinsic::arm_neon_vld3: {
1886 unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
1887 ARM::VLD3d32, ARM::VLD1d64T };
1888 unsigned QOpcodes0[] = { ARM::VLD3q8_UPD,
1891 unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD,
1892 ARM::VLD3q16odd_UPD,
1893 ARM::VLD3q32odd_UPD };
1894 return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
1897 case Intrinsic::arm_neon_vld4: {
1898 unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
1899 ARM::VLD4d32, ARM::VLD1d64Q };
1900 unsigned QOpcodes0[] = { ARM::VLD4q8_UPD,
1903 unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD,
1904 ARM::VLD4q16odd_UPD,
1905 ARM::VLD4q32odd_UPD };
1906 return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
1909 case Intrinsic::arm_neon_vld2lane: {
1910 unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
1911 unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 };
1912 unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd };
1913 return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
1916 case Intrinsic::arm_neon_vld3lane: {
1917 unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
1918 unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 };
1919 unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd };
1920 return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
1923 case Intrinsic::arm_neon_vld4lane: {
1924 unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
1925 unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 };
1926 unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd };
1927 return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
1930 case Intrinsic::arm_neon_vst1: {
1931 unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
1932 ARM::VST1d32, ARM::VST1d64 };
1933 unsigned QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
1934 ARM::VST1q32, ARM::VST1q64 };
1935 return SelectVST(N, 1, DOpcodes, QOpcodes, 0);
1938 case Intrinsic::arm_neon_vst2: {
1939 unsigned DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
1940 ARM::VST2d32, ARM::VST1q64 };
1941 unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
1942 return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
1945 case Intrinsic::arm_neon_vst3: {
1946 unsigned DOpcodes[] = { ARM::VST3d8, ARM::VST3d16,
1947 ARM::VST3d32, ARM::VST1d64T };
1948 unsigned QOpcodes0[] = { ARM::VST3q8_UPD,
1951 unsigned QOpcodes1[] = { ARM::VST3q8odd_UPD,
1952 ARM::VST3q16odd_UPD,
1953 ARM::VST3q32odd_UPD };
1954 return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
1957 case Intrinsic::arm_neon_vst4: {
1958 unsigned DOpcodes[] = { ARM::VST4d8, ARM::VST4d16,
1959 ARM::VST4d32, ARM::VST1d64Q };
1960 unsigned QOpcodes0[] = { ARM::VST4q8_UPD,
1963 unsigned QOpcodes1[] = { ARM::VST4q8odd_UPD,
1964 ARM::VST4q16odd_UPD,
1965 ARM::VST4q32odd_UPD };
1966 return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
1969 case Intrinsic::arm_neon_vst2lane: {
1970 unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
1971 unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 };
1972 unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd };
1973 return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
1976 case Intrinsic::arm_neon_vst3lane: {
1977 unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
1978 unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 };
1979 unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd };
1980 return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
1983 case Intrinsic::arm_neon_vst4lane: {
1984 unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
1985 unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 };
1986 unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd };
1987 return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
1993 return SelectCode(N);
1996 bool ARMDAGToDAGISel::
1997 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
1998 std::vector<SDValue> &OutOps) {
1999 assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
2000 // Require the address to be in a register. That is safe for all ARM
2001 // variants and it is hard to do anything much smarter without knowing
2002 // how the operand is used.
2003 OutOps.push_back(Op);
2007 /// createARMISelDag - This pass converts a legalized DAG into a
2008 /// ARM-specific DAG, ready for instruction scheduling.
2010 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
2011 CodeGenOpt::Level OptLevel) {
2012 return new ARMDAGToDAGISel(TM, OptLevel);