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 &Update,
84
SDValue &Opc, SDValue &Align);
86
bool SelectAddrModePC(SDNode *Op, SDValue N, SDValue &Offset,
89
bool SelectThumbAddrModeRR(SDNode *Op, SDValue N, SDValue &Base,
91
bool SelectThumbAddrModeRI5(SDNode *Op, SDValue N, unsigned Scale,
92
SDValue &Base, SDValue &OffImm,
94
bool SelectThumbAddrModeS1(SDNode *Op, SDValue N, SDValue &Base,
95
SDValue &OffImm, SDValue &Offset);
96
bool SelectThumbAddrModeS2(SDNode *Op, SDValue N, SDValue &Base,
97
SDValue &OffImm, SDValue &Offset);
98
bool SelectThumbAddrModeS4(SDNode *Op, SDValue N, SDValue &Base,
99
SDValue &OffImm, SDValue &Offset);
100
bool SelectThumbAddrModeSP(SDNode *Op, SDValue N, SDValue &Base,
103
bool SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
104
SDValue &BaseReg, SDValue &Opc);
105
bool SelectT2AddrModeImm12(SDNode *Op, SDValue N, SDValue &Base,
107
bool SelectT2AddrModeImm8(SDNode *Op, SDValue N, SDValue &Base,
109
bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
111
bool SelectT2AddrModeImm8s4(SDNode *Op, SDValue N, SDValue &Base,
113
bool SelectT2AddrModeSoReg(SDNode *Op, SDValue N, SDValue &Base,
114
SDValue &OffReg, SDValue &ShImm);
116
// Include the pieces autogenerated from the target description.
117
#include "ARMGenDAGISel.inc"
120
/// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
122
SDNode *SelectARMIndexedLoad(SDNode *N);
123
SDNode *SelectT2IndexedLoad(SDNode *N);
125
/// SelectDYN_ALLOC - Select dynamic alloc for Thumb.
126
SDNode *SelectDYN_ALLOC(SDNode *N);
128
/// SelectVLD - Select NEON load intrinsics. NumVecs should
129
/// be 2, 3 or 4. The opcode arrays specify the instructions used for
130
/// loads of D registers and even subregs and odd subregs of Q registers.
131
/// For NumVecs == 2, QOpcodes1 is not used.
132
SDNode *SelectVLD(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
133
unsigned *QOpcodes0, unsigned *QOpcodes1);
135
/// SelectVST - Select NEON store intrinsics. NumVecs should
136
/// be 2, 3 or 4. The opcode arrays specify the instructions used for
137
/// stores of D registers and even subregs and odd subregs of Q registers.
138
/// For NumVecs == 2, QOpcodes1 is not used.
139
SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
140
unsigned *QOpcodes0, unsigned *QOpcodes1);
142
/// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
143
/// be 2, 3 or 4. The opcode arrays specify the instructions used for
144
/// load/store of D registers and even subregs and odd subregs of Q registers.
145
SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, unsigned NumVecs,
146
unsigned *DOpcodes, unsigned *QOpcodes0,
147
unsigned *QOpcodes1);
149
/// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
150
SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, unsigned Opc);
152
/// SelectCMOVOp - Select CMOV instructions for ARM.
153
SDNode *SelectCMOVOp(SDNode *N);
154
SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
155
ARMCC::CondCodes CCVal, SDValue CCR,
157
SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
158
ARMCC::CondCodes CCVal, SDValue CCR,
160
SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
161
ARMCC::CondCodes CCVal, SDValue CCR,
163
SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
164
ARMCC::CondCodes CCVal, SDValue CCR,
167
/// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
168
/// inline asm expressions.
169
virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
171
std::vector<SDValue> &OutOps);
173
/// PairDRegs - Insert a pair of double registers into an implicit def to
174
/// form a quad register.
175
SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
179
/// isInt32Immediate - This method tests to see if the node is a 32-bit constant
180
/// operand. If so Imm will receive the 32-bit value.
181
static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
182
if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
183
Imm = cast<ConstantSDNode>(N)->getZExtValue();
189
// isInt32Immediate - This method tests to see if a constant operand.
190
// If so Imm will receive the 32 bit value.
191
static bool isInt32Immediate(SDValue N, unsigned &Imm) {
192
return isInt32Immediate(N.getNode(), Imm);
195
// isOpcWithIntImmediate - This method tests to see if the node is a specific
196
// opcode and that it has a immediate integer right operand.
197
// If so Imm will receive the 32 bit value.
198
static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
199
return N->getOpcode() == Opc &&
200
isInt32Immediate(N->getOperand(1).getNode(), Imm);
204
bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
209
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
211
// Don't match base register only case. That is matched to a separate
212
// lower complexity pattern with explicit register operand.
213
if (ShOpcVal == ARM_AM::no_shift) return false;
215
BaseReg = N.getOperand(0);
216
unsigned ShImmVal = 0;
217
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
218
ShReg = CurDAG->getRegister(0, MVT::i32);
219
ShImmVal = RHS->getZExtValue() & 31;
221
ShReg = N.getOperand(1);
223
Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
228
bool ARMDAGToDAGISel::SelectAddrMode2(SDNode *Op, SDValue N,
229
SDValue &Base, SDValue &Offset,
231
if (N.getOpcode() == ISD::MUL) {
232
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
233
// X * [3,5,9] -> X + X * [2,4,8] etc.
234
int RHSC = (int)RHS->getZExtValue();
237
ARM_AM::AddrOpc AddSub = ARM_AM::add;
239
AddSub = ARM_AM::sub;
242
if (isPowerOf2_32(RHSC)) {
243
unsigned ShAmt = Log2_32(RHSC);
244
Base = Offset = N.getOperand(0);
245
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
254
if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
256
if (N.getOpcode() == ISD::FrameIndex) {
257
int FI = cast<FrameIndexSDNode>(N)->getIndex();
258
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
259
} else if (N.getOpcode() == ARMISD::Wrapper &&
260
!(Subtarget->useMovt() &&
261
N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
262
Base = N.getOperand(0);
264
Offset = CurDAG->getRegister(0, MVT::i32);
265
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
271
// Match simple R +/- imm12 operands.
272
if (N.getOpcode() == ISD::ADD)
273
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
274
int RHSC = (int)RHS->getZExtValue();
275
if ((RHSC >= 0 && RHSC < 0x1000) ||
276
(RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
277
Base = N.getOperand(0);
278
if (Base.getOpcode() == ISD::FrameIndex) {
279
int FI = cast<FrameIndexSDNode>(Base)->getIndex();
280
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
282
Offset = CurDAG->getRegister(0, MVT::i32);
284
ARM_AM::AddrOpc AddSub = ARM_AM::add;
286
AddSub = ARM_AM::sub;
289
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
296
// Otherwise this is R +/- [possibly shifted] R.
297
ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
298
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
301
Base = N.getOperand(0);
302
Offset = N.getOperand(1);
304
if (ShOpcVal != ARM_AM::no_shift) {
305
// Check to see if the RHS of the shift is a constant, if not, we can't fold
307
if (ConstantSDNode *Sh =
308
dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
309
ShAmt = Sh->getZExtValue();
310
Offset = N.getOperand(1).getOperand(0);
312
ShOpcVal = ARM_AM::no_shift;
316
// Try matching (R shl C) + (R).
317
if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
318
ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
319
if (ShOpcVal != ARM_AM::no_shift) {
320
// Check to see if the RHS of the shift is a constant, if not, we can't
322
if (ConstantSDNode *Sh =
323
dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
324
ShAmt = Sh->getZExtValue();
325
Offset = N.getOperand(0).getOperand(0);
326
Base = N.getOperand(1);
328
ShOpcVal = ARM_AM::no_shift;
333
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
338
bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
339
SDValue &Offset, SDValue &Opc) {
340
unsigned Opcode = Op->getOpcode();
341
ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
342
? cast<LoadSDNode>(Op)->getAddressingMode()
343
: cast<StoreSDNode>(Op)->getAddressingMode();
344
ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
345
? ARM_AM::add : ARM_AM::sub;
346
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
347
int Val = (int)C->getZExtValue();
348
if (Val >= 0 && Val < 0x1000) { // 12 bits.
349
Offset = CurDAG->getRegister(0, MVT::i32);
350
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
358
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
360
if (ShOpcVal != ARM_AM::no_shift) {
361
// Check to see if the RHS of the shift is a constant, if not, we can't fold
363
if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
364
ShAmt = Sh->getZExtValue();
365
Offset = N.getOperand(0);
367
ShOpcVal = ARM_AM::no_shift;
371
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
377
bool ARMDAGToDAGISel::SelectAddrMode3(SDNode *Op, SDValue N,
378
SDValue &Base, SDValue &Offset,
380
if (N.getOpcode() == ISD::SUB) {
381
// X - C is canonicalize to X + -C, no need to handle it here.
382
Base = N.getOperand(0);
383
Offset = N.getOperand(1);
384
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
388
if (N.getOpcode() != ISD::ADD) {
390
if (N.getOpcode() == ISD::FrameIndex) {
391
int FI = cast<FrameIndexSDNode>(N)->getIndex();
392
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
394
Offset = CurDAG->getRegister(0, MVT::i32);
395
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
399
// If the RHS is +/- imm8, fold into addr mode.
400
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
401
int RHSC = (int)RHS->getZExtValue();
402
if ((RHSC >= 0 && RHSC < 256) ||
403
(RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
404
Base = N.getOperand(0);
405
if (Base.getOpcode() == ISD::FrameIndex) {
406
int FI = cast<FrameIndexSDNode>(Base)->getIndex();
407
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
409
Offset = CurDAG->getRegister(0, MVT::i32);
411
ARM_AM::AddrOpc AddSub = ARM_AM::add;
413
AddSub = ARM_AM::sub;
416
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
421
Base = N.getOperand(0);
422
Offset = N.getOperand(1);
423
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
427
bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
428
SDValue &Offset, SDValue &Opc) {
429
unsigned Opcode = Op->getOpcode();
430
ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
431
? cast<LoadSDNode>(Op)->getAddressingMode()
432
: cast<StoreSDNode>(Op)->getAddressingMode();
433
ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
434
? ARM_AM::add : ARM_AM::sub;
435
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
436
int Val = (int)C->getZExtValue();
437
if (Val >= 0 && Val < 256) {
438
Offset = CurDAG->getRegister(0, MVT::i32);
439
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
445
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
449
bool ARMDAGToDAGISel::SelectAddrMode4(SDNode *Op, SDValue N,
450
SDValue &Addr, SDValue &Mode) {
452
Mode = CurDAG->getTargetConstant(0, MVT::i32);
456
bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N,
457
SDValue &Base, SDValue &Offset) {
458
if (N.getOpcode() != ISD::ADD) {
460
if (N.getOpcode() == ISD::FrameIndex) {
461
int FI = cast<FrameIndexSDNode>(N)->getIndex();
462
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
463
} else if (N.getOpcode() == ARMISD::Wrapper &&
464
!(Subtarget->useMovt() &&
465
N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
466
Base = N.getOperand(0);
468
Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
473
// If the RHS is +/- imm8, fold into addr mode.
474
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
475
int RHSC = (int)RHS->getZExtValue();
476
if ((RHSC & 3) == 0) { // The constant is implicitly multiplied by 4.
478
if ((RHSC >= 0 && RHSC < 256) ||
479
(RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
480
Base = N.getOperand(0);
481
if (Base.getOpcode() == ISD::FrameIndex) {
482
int FI = cast<FrameIndexSDNode>(Base)->getIndex();
483
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
486
ARM_AM::AddrOpc AddSub = ARM_AM::add;
488
AddSub = ARM_AM::sub;
491
Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
499
Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
504
bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N,
505
SDValue &Addr, SDValue &Update,
506
SDValue &Opc, SDValue &Align) {
508
// Default to no writeback.
509
Update = CurDAG->getRegister(0, MVT::i32);
510
Opc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(false), MVT::i32);
511
// Default to no alignment.
512
Align = CurDAG->getTargetConstant(0, MVT::i32);
516
bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
517
SDValue &Offset, SDValue &Label) {
518
if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
519
Offset = N.getOperand(0);
520
SDValue N1 = N.getOperand(1);
521
Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
528
bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
529
SDValue &Base, SDValue &Offset){
530
// FIXME dl should come from the parent load or store, not the address
531
DebugLoc dl = Op->getDebugLoc();
532
if (N.getOpcode() != ISD::ADD) {
533
ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
534
if (!NC || NC->getZExtValue() != 0)
541
Base = N.getOperand(0);
542
Offset = N.getOperand(1);
547
ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
548
unsigned Scale, SDValue &Base,
549
SDValue &OffImm, SDValue &Offset) {
551
SDValue TmpBase, TmpOffImm;
552
if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
553
return false; // We want to select tLDRspi / tSTRspi instead.
554
if (N.getOpcode() == ARMISD::Wrapper &&
555
N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
556
return false; // We want to select tLDRpci instead.
559
if (N.getOpcode() != ISD::ADD) {
560
if (N.getOpcode() == ARMISD::Wrapper &&
561
!(Subtarget->useMovt() &&
562
N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
563
Base = N.getOperand(0);
567
Offset = CurDAG->getRegister(0, MVT::i32);
568
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
572
// Thumb does not have [sp, r] address mode.
573
RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
574
RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
575
if ((LHSR && LHSR->getReg() == ARM::SP) ||
576
(RHSR && RHSR->getReg() == ARM::SP)) {
578
Offset = CurDAG->getRegister(0, MVT::i32);
579
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
583
// If the RHS is + imm5 * scale, fold into addr mode.
584
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
585
int RHSC = (int)RHS->getZExtValue();
586
if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
588
if (RHSC >= 0 && RHSC < 32) {
589
Base = N.getOperand(0);
590
Offset = CurDAG->getRegister(0, MVT::i32);
591
OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
597
Base = N.getOperand(0);
598
Offset = N.getOperand(1);
599
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
603
bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
604
SDValue &Base, SDValue &OffImm,
606
return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
609
bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
610
SDValue &Base, SDValue &OffImm,
612
return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
615
bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
616
SDValue &Base, SDValue &OffImm,
618
return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
621
bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
622
SDValue &Base, SDValue &OffImm) {
623
if (N.getOpcode() == ISD::FrameIndex) {
624
int FI = cast<FrameIndexSDNode>(N)->getIndex();
625
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
626
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
630
if (N.getOpcode() != ISD::ADD)
633
RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
634
if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
635
(LHSR && LHSR->getReg() == ARM::SP)) {
636
// If the RHS is + imm8 * scale, fold into addr mode.
637
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
638
int RHSC = (int)RHS->getZExtValue();
639
if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
641
if (RHSC >= 0 && RHSC < 256) {
642
Base = N.getOperand(0);
643
if (Base.getOpcode() == ISD::FrameIndex) {
644
int FI = cast<FrameIndexSDNode>(Base)->getIndex();
645
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
647
OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
657
bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
660
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
662
// Don't match base register only case. That is matched to a separate
663
// lower complexity pattern with explicit register operand.
664
if (ShOpcVal == ARM_AM::no_shift) return false;
666
BaseReg = N.getOperand(0);
667
unsigned ShImmVal = 0;
668
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
669
ShImmVal = RHS->getZExtValue() & 31;
670
Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
677
bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
678
SDValue &Base, SDValue &OffImm) {
679
// Match simple R + imm12 operands.
682
if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
683
if (N.getOpcode() == ISD::FrameIndex) {
684
// Match frame index...
685
int FI = cast<FrameIndexSDNode>(N)->getIndex();
686
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
687
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
689
} else if (N.getOpcode() == ARMISD::Wrapper &&
690
!(Subtarget->useMovt() &&
691
N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
692
Base = N.getOperand(0);
693
if (Base.getOpcode() == ISD::TargetConstantPool)
694
return false; // We want to select t2LDRpci instead.
697
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
701
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
702
if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
703
// Let t2LDRi8 handle (R - imm8).
706
int RHSC = (int)RHS->getZExtValue();
707
if (N.getOpcode() == ISD::SUB)
710
if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
711
Base = N.getOperand(0);
712
if (Base.getOpcode() == ISD::FrameIndex) {
713
int FI = cast<FrameIndexSDNode>(Base)->getIndex();
714
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
716
OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
723
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
727
bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
728
SDValue &Base, SDValue &OffImm) {
729
// Match simple R - imm8 operands.
730
if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
731
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
732
int RHSC = (int)RHS->getSExtValue();
733
if (N.getOpcode() == ISD::SUB)
736
if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
737
Base = N.getOperand(0);
738
if (Base.getOpcode() == ISD::FrameIndex) {
739
int FI = cast<FrameIndexSDNode>(Base)->getIndex();
740
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
742
OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
751
bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
753
unsigned Opcode = Op->getOpcode();
754
ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
755
? cast<LoadSDNode>(Op)->getAddressingMode()
756
: cast<StoreSDNode>(Op)->getAddressingMode();
757
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
758
int RHSC = (int)RHS->getZExtValue();
759
if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
760
OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
761
? CurDAG->getTargetConstant(RHSC, MVT::i32)
762
: CurDAG->getTargetConstant(-RHSC, MVT::i32);
770
bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
771
SDValue &Base, SDValue &OffImm) {
772
if (N.getOpcode() == ISD::ADD) {
773
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
774
int RHSC = (int)RHS->getZExtValue();
775
if (((RHSC & 0x3) == 0) &&
776
((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits.
777
Base = N.getOperand(0);
778
OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
782
} else if (N.getOpcode() == ISD::SUB) {
783
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
784
int RHSC = (int)RHS->getZExtValue();
785
if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) { // 8 bits.
786
Base = N.getOperand(0);
787
OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
796
bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDNode *Op, SDValue N,
798
SDValue &OffReg, SDValue &ShImm) {
799
// (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
800
if (N.getOpcode() != ISD::ADD)
803
// Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
804
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
805
int RHSC = (int)RHS->getZExtValue();
806
if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
808
else if (RHSC < 0 && RHSC >= -255) // 8 bits
812
// Look for (R + R) or (R + (R << [1,2,3])).
814
Base = N.getOperand(0);
815
OffReg = N.getOperand(1);
817
// Swap if it is ((R << c) + R).
818
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
819
if (ShOpcVal != ARM_AM::lsl) {
820
ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
821
if (ShOpcVal == ARM_AM::lsl)
822
std::swap(Base, OffReg);
825
if (ShOpcVal == ARM_AM::lsl) {
826
// Check to see if the RHS of the shift is a constant, if not, we can't fold
828
if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
829
ShAmt = Sh->getZExtValue();
832
ShOpcVal = ARM_AM::no_shift;
834
OffReg = OffReg.getOperand(0);
836
ShOpcVal = ARM_AM::no_shift;
840
ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
845
//===--------------------------------------------------------------------===//
847
/// getAL - Returns a ARMCC::AL immediate node.
848
static inline SDValue getAL(SelectionDAG *CurDAG) {
849
return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
852
SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
853
LoadSDNode *LD = cast<LoadSDNode>(N);
854
ISD::MemIndexedMode AM = LD->getAddressingMode();
855
if (AM == ISD::UNINDEXED)
858
EVT LoadedVT = LD->getMemoryVT();
859
SDValue Offset, AMOpc;
860
bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
863
if (LoadedVT == MVT::i32 &&
864
SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
865
Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
867
} else if (LoadedVT == MVT::i16 &&
868
SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
870
Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
871
? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
872
: (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
873
} else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
874
if (LD->getExtensionType() == ISD::SEXTLOAD) {
875
if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
877
Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
880
if (SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
882
Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
888
SDValue Chain = LD->getChain();
889
SDValue Base = LD->getBasePtr();
890
SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
891
CurDAG->getRegister(0, MVT::i32), Chain };
892
return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
899
SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
900
LoadSDNode *LD = cast<LoadSDNode>(N);
901
ISD::MemIndexedMode AM = LD->getAddressingMode();
902
if (AM == ISD::UNINDEXED)
905
EVT LoadedVT = LD->getMemoryVT();
906
bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
908
bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
911
if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
912
switch (LoadedVT.getSimpleVT().SimpleTy) {
914
Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
918
Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
920
Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
925
Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
927
Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
936
SDValue Chain = LD->getChain();
937
SDValue Base = LD->getBasePtr();
938
SDValue Ops[]= { Base, Offset, getAL(CurDAG),
939
CurDAG->getRegister(0, MVT::i32), Chain };
940
return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
947
SDNode *ARMDAGToDAGISel::SelectDYN_ALLOC(SDNode *N) {
948
DebugLoc dl = N->getDebugLoc();
949
EVT VT = N->getValueType(0);
950
SDValue Chain = N->getOperand(0);
951
SDValue Size = N->getOperand(1);
952
SDValue Align = N->getOperand(2);
953
SDValue SP = CurDAG->getRegister(ARM::SP, MVT::i32);
954
int32_t AlignVal = cast<ConstantSDNode>(Align)->getSExtValue();
956
// We need to align the stack. Use Thumb1 tAND which is the only thumb
957
// instruction that can read and write SP. This matches to a pseudo
958
// instruction that has a chain to ensure the result is written back to
959
// the stack pointer.
960
SP = SDValue(CurDAG->getMachineNode(ARM::tANDsp, dl, VT, SP, Align), 0);
962
bool isC = isa<ConstantSDNode>(Size);
963
uint32_t C = isC ? cast<ConstantSDNode>(Size)->getZExtValue() : ~0UL;
964
// Handle the most common case for both Thumb1 and Thumb2:
965
// tSUBspi - immediate is between 0 ... 508 inclusive.
966
if (C <= 508 && ((C & 3) == 0))
967
// FIXME: tSUBspi encode scale 4 implicitly.
968
return CurDAG->SelectNodeTo(N, ARM::tSUBspi_, VT, MVT::Other, SP,
969
CurDAG->getTargetConstant(C/4, MVT::i32),
972
if (Subtarget->isThumb1Only()) {
973
// Use tADDspr since Thumb1 does not have a sub r, sp, r. ARMISelLowering
974
// should have negated the size operand already. FIXME: We can't insert
975
// new target independent node at this stage so we are forced to negate
976
// it earlier. Is there a better solution?
977
return CurDAG->SelectNodeTo(N, ARM::tADDspr_, VT, MVT::Other, SP, Size,
979
} else if (Subtarget->isThumb2()) {
980
if (isC && Predicate_t2_so_imm(Size.getNode())) {
982
SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
983
return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi_, VT, MVT::Other, Ops, 3);
984
} else if (isC && Predicate_imm0_4095(Size.getNode())) {
986
SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
987
return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi12_, VT, MVT::Other, Ops, 3);
990
SDValue Ops[] = { SP, Size,
991
getI32Imm(ARM_AM::getSORegOpc(ARM_AM::lsl,0)), Chain };
992
return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPs_, VT, MVT::Other, Ops, 4);
996
// FIXME: Add ADD / SUB sp instructions for ARM.
1000
/// PairDRegs - Insert a pair of double registers into an implicit def to
1001
/// form a quad register.
1002
SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
1003
DebugLoc dl = V0.getNode()->getDebugLoc();
1005
SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0);
1006
SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
1007
SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
1008
SDNode *Pair = CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
1009
VT, Undef, V0, SubReg0);
1010
return CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
1011
VT, SDValue(Pair, 0), V1, SubReg1);
1014
/// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
1015
/// for a 64-bit subregister of the vector.
1016
static EVT GetNEONSubregVT(EVT VT) {
1017
switch (VT.getSimpleVT().SimpleTy) {
1018
default: llvm_unreachable("unhandled NEON type");
1019
case MVT::v16i8: return MVT::v8i8;
1020
case MVT::v8i16: return MVT::v4i16;
1021
case MVT::v4f32: return MVT::v2f32;
1022
case MVT::v4i32: return MVT::v2i32;
1023
case MVT::v2i64: return MVT::v1i64;
1027
SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
1028
unsigned *DOpcodes, unsigned *QOpcodes0,
1029
unsigned *QOpcodes1) {
1030
assert(NumVecs >=2 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1031
DebugLoc dl = N->getDebugLoc();
1033
SDValue MemAddr, MemUpdate, MemOpc, Align;
1034
if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, MemUpdate, MemOpc, Align))
1037
SDValue Chain = N->getOperand(0);
1038
EVT VT = N->getValueType(0);
1039
bool is64BitVector = VT.is64BitVector();
1041
unsigned OpcodeIndex;
1042
switch (VT.getSimpleVT().SimpleTy) {
1043
default: llvm_unreachable("unhandled vld type");
1044
// Double-register operations:
1045
case MVT::v8i8: OpcodeIndex = 0; break;
1046
case MVT::v4i16: OpcodeIndex = 1; break;
1048
case MVT::v2i32: OpcodeIndex = 2; break;
1049
case MVT::v1i64: OpcodeIndex = 3; break;
1050
// Quad-register operations:
1051
case MVT::v16i8: OpcodeIndex = 0; break;
1052
case MVT::v8i16: OpcodeIndex = 1; break;
1054
case MVT::v4i32: OpcodeIndex = 2; break;
1057
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1058
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1059
if (is64BitVector) {
1060
unsigned Opc = DOpcodes[OpcodeIndex];
1061
const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Align,
1062
Pred, PredReg, Chain };
1063
std::vector<EVT> ResTys(NumVecs, VT);
1064
ResTys.push_back(MVT::Other);
1065
return CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 7);
1068
EVT RegVT = GetNEONSubregVT(VT);
1070
// Quad registers are directly supported for VLD2,
1071
// loading 2 pairs of D regs.
1072
unsigned Opc = QOpcodes0[OpcodeIndex];
1073
const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Align,
1074
Pred, PredReg, Chain };
1075
std::vector<EVT> ResTys(4, VT);
1076
ResTys.push_back(MVT::Other);
1077
SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 7);
1078
Chain = SDValue(VLd, 4);
1080
// Combine the even and odd subregs to produce the result.
1081
for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1082
SDNode *Q = PairDRegs(VT, SDValue(VLd, 2*Vec), SDValue(VLd, 2*Vec+1));
1083
ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1086
// Otherwise, quad registers are loaded with two separate instructions,
1087
// where one loads the even registers and the other loads the odd registers.
1089
// Enable writeback to the address register.
1090
MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
1092
std::vector<EVT> ResTys(NumVecs, RegVT);
1093
ResTys.push_back(MemAddr.getValueType());
1094
ResTys.push_back(MVT::Other);
1096
// Load the even subregs.
1097
unsigned Opc = QOpcodes0[OpcodeIndex];
1098
const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, Align,
1099
Pred, PredReg, Chain };
1100
SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 7);
1101
Chain = SDValue(VLdA, NumVecs+1);
1103
// Load the odd subregs.
1104
Opc = QOpcodes1[OpcodeIndex];
1105
const SDValue OpsB[] = { SDValue(VLdA, NumVecs), MemUpdate, MemOpc,
1106
Align, Pred, PredReg, Chain };
1107
SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 7);
1108
Chain = SDValue(VLdB, NumVecs+1);
1110
// Combine the even and odd subregs to produce the result.
1111
for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1112
SDNode *Q = PairDRegs(VT, SDValue(VLdA, Vec), SDValue(VLdB, Vec));
1113
ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
1116
ReplaceUses(SDValue(N, NumVecs), Chain);
1120
SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
1121
unsigned *DOpcodes, unsigned *QOpcodes0,
1122
unsigned *QOpcodes1) {
1123
assert(NumVecs >=2 && NumVecs <= 4 && "VST NumVecs out-of-range");
1124
DebugLoc dl = N->getDebugLoc();
1126
SDValue MemAddr, MemUpdate, MemOpc, Align;
1127
if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, MemUpdate, MemOpc, Align))
1130
SDValue Chain = N->getOperand(0);
1131
EVT VT = N->getOperand(3).getValueType();
1132
bool is64BitVector = VT.is64BitVector();
1134
unsigned OpcodeIndex;
1135
switch (VT.getSimpleVT().SimpleTy) {
1136
default: llvm_unreachable("unhandled vst type");
1137
// Double-register operations:
1138
case MVT::v8i8: OpcodeIndex = 0; break;
1139
case MVT::v4i16: OpcodeIndex = 1; break;
1141
case MVT::v2i32: OpcodeIndex = 2; break;
1142
case MVT::v1i64: OpcodeIndex = 3; break;
1143
// Quad-register operations:
1144
case MVT::v16i8: OpcodeIndex = 0; break;
1145
case MVT::v8i16: OpcodeIndex = 1; break;
1147
case MVT::v4i32: OpcodeIndex = 2; break;
1150
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1151
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1153
SmallVector<SDValue, 8> Ops;
1154
Ops.push_back(MemAddr);
1155
Ops.push_back(MemUpdate);
1156
Ops.push_back(MemOpc);
1157
Ops.push_back(Align);
1159
if (is64BitVector) {
1160
unsigned Opc = DOpcodes[OpcodeIndex];
1161
for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1162
Ops.push_back(N->getOperand(Vec+3));
1163
Ops.push_back(Pred);
1164
Ops.push_back(PredReg);
1165
Ops.push_back(Chain);
1166
return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+7);
1169
EVT RegVT = GetNEONSubregVT(VT);
1171
// Quad registers are directly supported for VST2,
1172
// storing 2 pairs of D regs.
1173
unsigned Opc = QOpcodes0[OpcodeIndex];
1174
for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1175
Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1176
N->getOperand(Vec+3)));
1177
Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1178
N->getOperand(Vec+3)));
1180
Ops.push_back(Pred);
1181
Ops.push_back(PredReg);
1182
Ops.push_back(Chain);
1183
return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 11);
1186
// Otherwise, quad registers are stored with two separate instructions,
1187
// where one stores the even registers and the other stores the odd registers.
1189
// Enable writeback to the address register.
1190
MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
1192
// Store the even subregs.
1193
for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1194
Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
1195
N->getOperand(Vec+3)));
1196
Ops.push_back(Pred);
1197
Ops.push_back(PredReg);
1198
Ops.push_back(Chain);
1199
unsigned Opc = QOpcodes0[OpcodeIndex];
1200
SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1201
MVT::Other, Ops.data(), NumVecs+7);
1202
Chain = SDValue(VStA, 1);
1204
// Store the odd subregs.
1205
Ops[0] = SDValue(VStA, 0); // MemAddr
1206
for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1207
Ops[Vec+4] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
1208
N->getOperand(Vec+3));
1209
Ops[NumVecs+4] = Pred;
1210
Ops[NumVecs+5] = PredReg;
1211
Ops[NumVecs+6] = Chain;
1212
Opc = QOpcodes1[OpcodeIndex];
1213
SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1214
MVT::Other, Ops.data(), NumVecs+7);
1215
Chain = SDValue(VStB, 1);
1216
ReplaceUses(SDValue(N, 0), Chain);
1220
SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1221
unsigned NumVecs, unsigned *DOpcodes,
1222
unsigned *QOpcodes0,
1223
unsigned *QOpcodes1) {
1224
assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1225
DebugLoc dl = N->getDebugLoc();
1227
SDValue MemAddr, MemUpdate, MemOpc, Align;
1228
if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, MemUpdate, MemOpc, Align))
1231
SDValue Chain = N->getOperand(0);
1233
cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1234
EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
1235
bool is64BitVector = VT.is64BitVector();
1237
// Quad registers are handled by load/store of subregs. Find the subreg info.
1238
unsigned NumElts = 0;
1241
if (!is64BitVector) {
1242
RegVT = GetNEONSubregVT(VT);
1243
NumElts = RegVT.getVectorNumElements();
1244
SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
1247
unsigned OpcodeIndex;
1248
switch (VT.getSimpleVT().SimpleTy) {
1249
default: llvm_unreachable("unhandled vld/vst lane type");
1250
// Double-register operations:
1251
case MVT::v8i8: OpcodeIndex = 0; break;
1252
case MVT::v4i16: OpcodeIndex = 1; break;
1254
case MVT::v2i32: OpcodeIndex = 2; break;
1255
// Quad-register operations:
1256
case MVT::v8i16: OpcodeIndex = 0; break;
1258
case MVT::v4i32: OpcodeIndex = 1; break;
1261
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1262
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1264
SmallVector<SDValue, 9> Ops;
1265
Ops.push_back(MemAddr);
1266
Ops.push_back(MemUpdate);
1267
Ops.push_back(MemOpc);
1268
Ops.push_back(Align);
1271
if (is64BitVector) {
1272
Opc = DOpcodes[OpcodeIndex];
1273
for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1274
Ops.push_back(N->getOperand(Vec+3));
1276
// Check if this is loading the even or odd subreg of a Q register.
1277
if (Lane < NumElts) {
1278
Opc = QOpcodes0[OpcodeIndex];
1281
Opc = QOpcodes1[OpcodeIndex];
1283
// Extract the subregs of the input vector.
1284
for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1285
Ops.push_back(CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
1286
N->getOperand(Vec+3)));
1288
Ops.push_back(getI32Imm(Lane));
1289
Ops.push_back(Pred);
1290
Ops.push_back(PredReg);
1291
Ops.push_back(Chain);
1294
return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+8);
1296
std::vector<EVT> ResTys(NumVecs, RegVT);
1297
ResTys.push_back(MVT::Other);
1299
CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), NumVecs+8);
1300
// For a 64-bit vector load to D registers, nothing more needs to be done.
1304
// For 128-bit vectors, take the 64-bit results of the load and insert them
1305
// as subregs into the result.
1306
for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1307
SDValue QuadVec = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
1308
N->getOperand(Vec+3),
1309
SDValue(VLdLn, Vec));
1310
ReplaceUses(SDValue(N, Vec), QuadVec);
1313
Chain = SDValue(VLdLn, NumVecs);
1314
ReplaceUses(SDValue(N, NumVecs), Chain);
1318
SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
1320
if (!Subtarget->hasV6T2Ops())
1323
unsigned Shl_imm = 0;
1324
if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
1325
assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1326
unsigned Srl_imm = 0;
1327
if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
1328
assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1329
unsigned Width = 32 - Srl_imm;
1330
int LSB = Srl_imm - Shl_imm;
1333
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1334
SDValue Ops[] = { N->getOperand(0).getOperand(0),
1335
CurDAG->getTargetConstant(LSB, MVT::i32),
1336
CurDAG->getTargetConstant(Width, MVT::i32),
1337
getAL(CurDAG), Reg0 };
1338
return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1344
SDNode *ARMDAGToDAGISel::
1345
SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1346
ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1349
if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
1350
unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1351
unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1354
case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1355
case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1356
case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1357
case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1359
llvm_unreachable("Unknown so_reg opcode!");
1363
CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1364
SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1365
SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
1366
return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
1371
SDNode *ARMDAGToDAGISel::
1372
SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1373
ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1377
if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
1378
SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1379
SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
1380
return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
1385
SDNode *ARMDAGToDAGISel::
1386
SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1387
ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1388
ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1392
if (Predicate_t2_so_imm(TrueVal.getNode())) {
1393
SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1394
SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1395
SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1396
return CurDAG->SelectNodeTo(N,
1397
ARM::t2MOVCCi, MVT::i32, Ops, 5);
1402
SDNode *ARMDAGToDAGISel::
1403
SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1404
ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1405
ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1409
if (Predicate_so_imm(TrueVal.getNode())) {
1410
SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1411
SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1412
SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1413
return CurDAG->SelectNodeTo(N,
1414
ARM::MOVCCi, MVT::i32, Ops, 5);
1419
SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
1420
EVT VT = N->getValueType(0);
1421
SDValue FalseVal = N->getOperand(0);
1422
SDValue TrueVal = N->getOperand(1);
1423
SDValue CC = N->getOperand(2);
1424
SDValue CCR = N->getOperand(3);
1425
SDValue InFlag = N->getOperand(4);
1426
assert(CC.getOpcode() == ISD::Constant);
1427
assert(CCR.getOpcode() == ISD::Register);
1428
ARMCC::CondCodes CCVal =
1429
(ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
1431
if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1432
// Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1433
// Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1434
// Pattern complexity = 18 cost = 1 size = 0
1438
if (Subtarget->isThumb()) {
1439
SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
1440
CCVal, CCR, InFlag);
1442
Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
1443
ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1447
SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
1448
CCVal, CCR, InFlag);
1450
Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
1451
ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1456
// Pattern: (ARMcmov:i32 GPR:i32:$false,
1457
// (imm:i32)<<P:Predicate_so_imm>>:$true,
1459
// Emits: (MOVCCi:i32 GPR:i32:$false,
1460
// (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1461
// Pattern complexity = 10 cost = 1 size = 0
1462
if (Subtarget->isThumb()) {
1463
SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
1464
CCVal, CCR, InFlag);
1466
Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
1467
ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1471
SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
1472
CCVal, CCR, InFlag);
1474
Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
1475
ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1481
// Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1482
// Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1483
// Pattern complexity = 6 cost = 1 size = 0
1485
// Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1486
// Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1487
// Pattern complexity = 6 cost = 11 size = 0
1489
// Also FCPYScc and FCPYDcc.
1490
SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
1491
SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
1493
switch (VT.getSimpleVT().SimpleTy) {
1494
default: assert(false && "Illegal conditional move type!");
1497
Opc = Subtarget->isThumb()
1498
? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1508
return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1511
SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
1512
DebugLoc dl = N->getDebugLoc();
1514
if (N->isMachineOpcode())
1515
return NULL; // Already selected.
1517
switch (N->getOpcode()) {
1519
case ISD::Constant: {
1520
unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1522
if (Subtarget->hasThumb2())
1523
// Thumb2-aware targets have the MOVT instruction, so all immediates can
1524
// be done with MOV + MOVT, at worst.
1527
if (Subtarget->isThumb()) {
1528
UseCP = (Val > 255 && // MOV
1529
~Val > 255 && // MOV + MVN
1530
!ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
1532
UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
1533
ARM_AM::getSOImmVal(~Val) == -1 && // MVN
1534
!ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
1539
CurDAG->getTargetConstantPool(ConstantInt::get(
1540
Type::getInt32Ty(*CurDAG->getContext()), Val),
1541
TLI.getPointerTy());
1544
if (Subtarget->isThumb1Only()) {
1545
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1546
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1547
SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1548
ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1553
CurDAG->getRegister(0, MVT::i32),
1554
CurDAG->getTargetConstant(0, MVT::i32),
1556
CurDAG->getRegister(0, MVT::i32),
1557
CurDAG->getEntryNode()
1559
ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1562
ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
1566
// Other cases are autogenerated.
1569
case ISD::FrameIndex: {
1570
// Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1571
int FI = cast<FrameIndexSDNode>(N)->getIndex();
1572
SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1573
if (Subtarget->isThumb1Only()) {
1574
return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1575
CurDAG->getTargetConstant(0, MVT::i32));
1577
unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1578
ARM::t2ADDri : ARM::ADDri);
1579
SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1580
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1581
CurDAG->getRegister(0, MVT::i32) };
1582
return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1585
case ARMISD::DYN_ALLOC:
1586
return SelectDYN_ALLOC(N);
1588
if (SDNode *I = SelectV6T2BitfieldExtractOp(N,
1589
Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX))
1593
if (SDNode *I = SelectV6T2BitfieldExtractOp(N,
1594
Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX))
1598
if (Subtarget->isThumb1Only())
1600
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1601
unsigned RHSV = C->getZExtValue();
1603
if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
1604
unsigned ShImm = Log2_32(RHSV-1);
1607
SDValue V = N->getOperand(0);
1608
ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1609
SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1610
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1611
if (Subtarget->isThumb()) {
1612
SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1613
return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1615
SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1616
return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1619
if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
1620
unsigned ShImm = Log2_32(RHSV+1);
1623
SDValue V = N->getOperand(0);
1624
ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1625
SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1626
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1627
if (Subtarget->isThumb()) {
1628
SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0 };
1629
return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 5);
1631
SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1632
return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1638
// (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
1639
// of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
1640
// are entirely contributed by c2 and lower 16-bits are entirely contributed
1641
// by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
1642
// Select it to: "movt x, ((c1 & 0xffff) >> 16)
1643
EVT VT = N->getValueType(0);
1646
unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
1648
: (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
1651
SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
1652
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1655
if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
1656
SDValue N2 = N0.getOperand(1);
1657
ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1660
unsigned N1CVal = N1C->getZExtValue();
1661
unsigned N2CVal = N2C->getZExtValue();
1662
if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
1663
(N1CVal & 0xffffU) == 0xffffU &&
1664
(N2CVal & 0xffffU) == 0x0U) {
1665
SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
1667
SDValue Ops[] = { N0.getOperand(0), Imm16,
1668
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1669
return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
1674
case ARMISD::VMOVRRD:
1675
return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
1676
N->getOperand(0), getAL(CurDAG),
1677
CurDAG->getRegister(0, MVT::i32));
1678
case ISD::UMUL_LOHI: {
1679
if (Subtarget->isThumb1Only())
1681
if (Subtarget->isThumb()) {
1682
SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1683
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1684
CurDAG->getRegister(0, MVT::i32) };
1685
return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops,4);
1687
SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1688
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1689
CurDAG->getRegister(0, MVT::i32) };
1690
return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1693
case ISD::SMUL_LOHI: {
1694
if (Subtarget->isThumb1Only())
1696
if (Subtarget->isThumb()) {
1697
SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1698
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1699
return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops,4);
1701
SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1702
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1703
CurDAG->getRegister(0, MVT::i32) };
1704
return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1708
SDNode *ResNode = 0;
1709
if (Subtarget->isThumb() && Subtarget->hasThumb2())
1710
ResNode = SelectT2IndexedLoad(N);
1712
ResNode = SelectARMIndexedLoad(N);
1715
// Other cases are autogenerated.
1718
case ARMISD::BRCOND: {
1719
// Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1720
// Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
1721
// Pattern complexity = 6 cost = 1 size = 0
1723
// Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1724
// Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
1725
// Pattern complexity = 6 cost = 1 size = 0
1727
// Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
1728
// Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
1729
// Pattern complexity = 6 cost = 1 size = 0
1731
unsigned Opc = Subtarget->isThumb() ?
1732
((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
1733
SDValue Chain = N->getOperand(0);
1734
SDValue N1 = N->getOperand(1);
1735
SDValue N2 = N->getOperand(2);
1736
SDValue N3 = N->getOperand(3);
1737
SDValue InFlag = N->getOperand(4);
1738
assert(N1.getOpcode() == ISD::BasicBlock);
1739
assert(N2.getOpcode() == ISD::Constant);
1740
assert(N3.getOpcode() == ISD::Register);
1742
SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1743
cast<ConstantSDNode>(N2)->getZExtValue()),
1745
SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
1746
SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
1748
Chain = SDValue(ResNode, 0);
1749
if (N->getNumValues() == 2) {
1750
InFlag = SDValue(ResNode, 1);
1751
ReplaceUses(SDValue(N, 1), InFlag);
1753
ReplaceUses(SDValue(N, 0),
1754
SDValue(Chain.getNode(), Chain.getResNo()));
1758
return SelectCMOVOp(N);
1759
case ARMISD::CNEG: {
1760
EVT VT = N->getValueType(0);
1761
SDValue N0 = N->getOperand(0);
1762
SDValue N1 = N->getOperand(1);
1763
SDValue N2 = N->getOperand(2);
1764
SDValue N3 = N->getOperand(3);
1765
SDValue InFlag = N->getOperand(4);
1766
assert(N2.getOpcode() == ISD::Constant);
1767
assert(N3.getOpcode() == ISD::Register);
1769
SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
1770
cast<ConstantSDNode>(N2)->getZExtValue()),
1772
SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
1774
switch (VT.getSimpleVT().SimpleTy) {
1775
default: assert(false && "Illegal conditional move type!");
1784
return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1787
case ARMISD::VZIP: {
1789
EVT VT = N->getValueType(0);
1790
switch (VT.getSimpleVT().SimpleTy) {
1791
default: return NULL;
1792
case MVT::v8i8: Opc = ARM::VZIPd8; break;
1793
case MVT::v4i16: Opc = ARM::VZIPd16; break;
1795
case MVT::v2i32: Opc = ARM::VZIPd32; break;
1796
case MVT::v16i8: Opc = ARM::VZIPq8; break;
1797
case MVT::v8i16: Opc = ARM::VZIPq16; break;
1799
case MVT::v4i32: Opc = ARM::VZIPq32; break;
1801
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1802
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1803
SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
1804
return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
1806
case ARMISD::VUZP: {
1808
EVT VT = N->getValueType(0);
1809
switch (VT.getSimpleVT().SimpleTy) {
1810
default: return NULL;
1811
case MVT::v8i8: Opc = ARM::VUZPd8; break;
1812
case MVT::v4i16: Opc = ARM::VUZPd16; break;
1814
case MVT::v2i32: Opc = ARM::VUZPd32; break;
1815
case MVT::v16i8: Opc = ARM::VUZPq8; break;
1816
case MVT::v8i16: Opc = ARM::VUZPq16; break;
1818
case MVT::v4i32: Opc = ARM::VUZPq32; break;
1820
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1821
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1822
SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
1823
return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
1825
case ARMISD::VTRN: {
1827
EVT VT = N->getValueType(0);
1828
switch (VT.getSimpleVT().SimpleTy) {
1829
default: return NULL;
1830
case MVT::v8i8: Opc = ARM::VTRNd8; break;
1831
case MVT::v4i16: Opc = ARM::VTRNd16; break;
1833
case MVT::v2i32: Opc = ARM::VTRNd32; break;
1834
case MVT::v16i8: Opc = ARM::VTRNq8; break;
1835
case MVT::v8i16: Opc = ARM::VTRNq16; break;
1837
case MVT::v4i32: Opc = ARM::VTRNq32; break;
1839
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
1840
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1841
SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
1842
return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
1845
case ISD::INTRINSIC_VOID:
1846
case ISD::INTRINSIC_W_CHAIN: {
1847
unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
1852
case Intrinsic::arm_neon_vld2: {
1853
unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
1854
ARM::VLD2d32, ARM::VLD2d64 };
1855
unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
1856
return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
1859
case Intrinsic::arm_neon_vld3: {
1860
unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
1861
ARM::VLD3d32, ARM::VLD3d64 };
1862
unsigned QOpcodes0[] = { ARM::VLD3q8a, ARM::VLD3q16a, ARM::VLD3q32a };
1863
unsigned QOpcodes1[] = { ARM::VLD3q8b, ARM::VLD3q16b, ARM::VLD3q32b };
1864
return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
1867
case Intrinsic::arm_neon_vld4: {
1868
unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
1869
ARM::VLD4d32, ARM::VLD4d64 };
1870
unsigned QOpcodes0[] = { ARM::VLD4q8a, ARM::VLD4q16a, ARM::VLD4q32a };
1871
unsigned QOpcodes1[] = { ARM::VLD4q8b, ARM::VLD4q16b, ARM::VLD4q32b };
1872
return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
1875
case Intrinsic::arm_neon_vld2lane: {
1876
unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
1877
unsigned QOpcodes0[] = { ARM::VLD2LNq16a, ARM::VLD2LNq32a };
1878
unsigned QOpcodes1[] = { ARM::VLD2LNq16b, ARM::VLD2LNq32b };
1879
return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
1882
case Intrinsic::arm_neon_vld3lane: {
1883
unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
1884
unsigned QOpcodes0[] = { ARM::VLD3LNq16a, ARM::VLD3LNq32a };
1885
unsigned QOpcodes1[] = { ARM::VLD3LNq16b, ARM::VLD3LNq32b };
1886
return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
1889
case Intrinsic::arm_neon_vld4lane: {
1890
unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
1891
unsigned QOpcodes0[] = { ARM::VLD4LNq16a, ARM::VLD4LNq32a };
1892
unsigned QOpcodes1[] = { ARM::VLD4LNq16b, ARM::VLD4LNq32b };
1893
return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
1896
case Intrinsic::arm_neon_vst2: {
1897
unsigned DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
1898
ARM::VST2d32, ARM::VST2d64 };
1899
unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
1900
return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
1903
case Intrinsic::arm_neon_vst3: {
1904
unsigned DOpcodes[] = { ARM::VST3d8, ARM::VST3d16,
1905
ARM::VST3d32, ARM::VST3d64 };
1906
unsigned QOpcodes0[] = { ARM::VST3q8a, ARM::VST3q16a, ARM::VST3q32a };
1907
unsigned QOpcodes1[] = { ARM::VST3q8b, ARM::VST3q16b, ARM::VST3q32b };
1908
return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
1911
case Intrinsic::arm_neon_vst4: {
1912
unsigned DOpcodes[] = { ARM::VST4d8, ARM::VST4d16,
1913
ARM::VST4d32, ARM::VST4d64 };
1914
unsigned QOpcodes0[] = { ARM::VST4q8a, ARM::VST4q16a, ARM::VST4q32a };
1915
unsigned QOpcodes1[] = { ARM::VST4q8b, ARM::VST4q16b, ARM::VST4q32b };
1916
return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
1919
case Intrinsic::arm_neon_vst2lane: {
1920
unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
1921
unsigned QOpcodes0[] = { ARM::VST2LNq16a, ARM::VST2LNq32a };
1922
unsigned QOpcodes1[] = { ARM::VST2LNq16b, ARM::VST2LNq32b };
1923
return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
1926
case Intrinsic::arm_neon_vst3lane: {
1927
unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
1928
unsigned QOpcodes0[] = { ARM::VST3LNq16a, ARM::VST3LNq32a };
1929
unsigned QOpcodes1[] = { ARM::VST3LNq16b, ARM::VST3LNq32b };
1930
return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
1933
case Intrinsic::arm_neon_vst4lane: {
1934
unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
1935
unsigned QOpcodes0[] = { ARM::VST4LNq16a, ARM::VST4LNq32a };
1936
unsigned QOpcodes1[] = { ARM::VST4LNq16b, ARM::VST4LNq32b };
1937
return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
1943
return SelectCode(N);
1946
bool ARMDAGToDAGISel::
1947
SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
1948
std::vector<SDValue> &OutOps) {
1949
assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
1950
// Require the address to be in a register. That is safe for all ARM
1951
// variants and it is hard to do anything much smarter without knowing
1952
// how the operand is used.
1953
OutOps.push_back(Op);
1957
/// createARMISelDag - This pass converts a legalized DAG into a
1958
/// ARM-specific DAG, ready for instruction scheduling.
1960
FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
1961
CodeGenOpt::Level OptLevel) {
1962
return new ARMDAGToDAGISel(TM, OptLevel);