1 //===-- HexagonISelLoweringHVX.cpp --- Lowering HVX operations ------------===//
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 #include "HexagonISelLowering.h"
11 #include "HexagonRegisterInfo.h"
12 #include "HexagonSubtarget.h"
13 #include "llvm/Support/CommandLine.h"
17 static const MVT LegalV64[] = { MVT::v64i8, MVT::v32i16, MVT::v16i32 };
18 static const MVT LegalW64[] = { MVT::v128i8, MVT::v64i16, MVT::v32i32 };
19 static const MVT LegalV128[] = { MVT::v128i8, MVT::v64i16, MVT::v32i32 };
20 static const MVT LegalW128[] = { MVT::v256i8, MVT::v128i16, MVT::v64i32 };
24 HexagonTargetLowering::initializeHVXLowering() {
25 if (Subtarget.useHVX64BOps()) {
26 addRegisterClass(MVT::v64i8, &Hexagon::HvxVRRegClass);
27 addRegisterClass(MVT::v32i16, &Hexagon::HvxVRRegClass);
28 addRegisterClass(MVT::v16i32, &Hexagon::HvxVRRegClass);
29 addRegisterClass(MVT::v128i8, &Hexagon::HvxWRRegClass);
30 addRegisterClass(MVT::v64i16, &Hexagon::HvxWRRegClass);
31 addRegisterClass(MVT::v32i32, &Hexagon::HvxWRRegClass);
32 // These "short" boolean vector types should be legal because
33 // they will appear as results of vector compares. If they were
34 // not legal, type legalization would try to make them legal
35 // and that would require using operations that do not use or
36 // produce such types. That, in turn, would imply using custom
37 // nodes, which would be unoptimizable by the DAG combiner.
38 // The idea is to rely on target-independent operations as much
40 addRegisterClass(MVT::v16i1, &Hexagon::HvxQRRegClass);
41 addRegisterClass(MVT::v32i1, &Hexagon::HvxQRRegClass);
42 addRegisterClass(MVT::v64i1, &Hexagon::HvxQRRegClass);
43 addRegisterClass(MVT::v512i1, &Hexagon::HvxQRRegClass);
44 } else if (Subtarget.useHVX128BOps()) {
45 addRegisterClass(MVT::v128i8, &Hexagon::HvxVRRegClass);
46 addRegisterClass(MVT::v64i16, &Hexagon::HvxVRRegClass);
47 addRegisterClass(MVT::v32i32, &Hexagon::HvxVRRegClass);
48 addRegisterClass(MVT::v256i8, &Hexagon::HvxWRRegClass);
49 addRegisterClass(MVT::v128i16, &Hexagon::HvxWRRegClass);
50 addRegisterClass(MVT::v64i32, &Hexagon::HvxWRRegClass);
51 addRegisterClass(MVT::v32i1, &Hexagon::HvxQRRegClass);
52 addRegisterClass(MVT::v64i1, &Hexagon::HvxQRRegClass);
53 addRegisterClass(MVT::v128i1, &Hexagon::HvxQRRegClass);
54 addRegisterClass(MVT::v1024i1, &Hexagon::HvxQRRegClass);
57 // Set up operation actions.
59 bool Use64b = Subtarget.useHVX64BOps();
60 ArrayRef<MVT> LegalV = Use64b ? LegalV64 : LegalV128;
61 ArrayRef<MVT> LegalW = Use64b ? LegalW64 : LegalW128;
62 MVT ByteV = Use64b ? MVT::v64i8 : MVT::v128i8;
63 MVT ByteW = Use64b ? MVT::v128i8 : MVT::v256i8;
65 auto setPromoteTo = [this] (unsigned Opc, MVT FromTy, MVT ToTy) {
66 setOperationAction(Opc, FromTy, Promote);
67 AddPromotedToType(Opc, FromTy, ToTy);
70 setOperationAction(ISD::VECTOR_SHUFFLE, ByteV, Legal);
71 setOperationAction(ISD::VECTOR_SHUFFLE, ByteW, Legal);
73 for (MVT T : LegalV) {
74 setIndexedLoadAction(ISD::POST_INC, T, Legal);
75 setIndexedStoreAction(ISD::POST_INC, T, Legal);
77 setOperationAction(ISD::AND, T, Legal);
78 setOperationAction(ISD::OR, T, Legal);
79 setOperationAction(ISD::XOR, T, Legal);
80 setOperationAction(ISD::ADD, T, Legal);
81 setOperationAction(ISD::SUB, T, Legal);
82 setOperationAction(ISD::CTPOP, T, Legal);
83 setOperationAction(ISD::CTLZ, T, Legal);
85 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, T, Legal);
86 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, T, Legal);
87 setOperationAction(ISD::BSWAP, T, Legal);
90 setOperationAction(ISD::CTTZ, T, Custom);
91 setOperationAction(ISD::LOAD, T, Custom);
92 setOperationAction(ISD::MUL, T, Custom);
93 setOperationAction(ISD::MULHS, T, Custom);
94 setOperationAction(ISD::MULHU, T, Custom);
95 setOperationAction(ISD::BUILD_VECTOR, T, Custom);
96 // Make concat-vectors custom to handle concats of more than 2 vectors.
97 setOperationAction(ISD::CONCAT_VECTORS, T, Custom);
98 setOperationAction(ISD::INSERT_SUBVECTOR, T, Custom);
99 setOperationAction(ISD::INSERT_VECTOR_ELT, T, Custom);
100 setOperationAction(ISD::EXTRACT_SUBVECTOR, T, Custom);
101 setOperationAction(ISD::EXTRACT_VECTOR_ELT, T, Custom);
102 setOperationAction(ISD::ANY_EXTEND, T, Custom);
103 setOperationAction(ISD::SIGN_EXTEND, T, Custom);
104 setOperationAction(ISD::ZERO_EXTEND, T, Custom);
106 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, T, Custom);
107 // HVX only has shifts of words and halfwords.
108 setOperationAction(ISD::SRA, T, Custom);
109 setOperationAction(ISD::SHL, T, Custom);
110 setOperationAction(ISD::SRL, T, Custom);
112 // Promote all shuffles to operate on vectors of bytes.
113 setPromoteTo(ISD::VECTOR_SHUFFLE, T, ByteV);
116 setCondCodeAction(ISD::SETNE, T, Expand);
117 setCondCodeAction(ISD::SETLE, T, Expand);
118 setCondCodeAction(ISD::SETGE, T, Expand);
119 setCondCodeAction(ISD::SETLT, T, Expand);
120 setCondCodeAction(ISD::SETULE, T, Expand);
121 setCondCodeAction(ISD::SETUGE, T, Expand);
122 setCondCodeAction(ISD::SETULT, T, Expand);
125 for (MVT T : LegalW) {
126 // Custom-lower BUILD_VECTOR for vector pairs. The standard (target-
127 // independent) handling of it would convert it to a load, which is
128 // not always the optimal choice.
129 setOperationAction(ISD::BUILD_VECTOR, T, Custom);
130 // Make concat-vectors custom to handle concats of more than 2 vectors.
131 setOperationAction(ISD::CONCAT_VECTORS, T, Custom);
133 // Custom-lower these operations for pairs. Expand them into a concat
134 // of the corresponding operations on individual vectors.
135 setOperationAction(ISD::ANY_EXTEND, T, Custom);
136 setOperationAction(ISD::SIGN_EXTEND, T, Custom);
137 setOperationAction(ISD::ZERO_EXTEND, T, Custom);
138 setOperationAction(ISD::SIGN_EXTEND_INREG, T, Custom);
139 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, T, Custom);
140 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, T, Legal);
141 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, T, Legal);
143 setOperationAction(ISD::LOAD, T, Custom);
144 setOperationAction(ISD::STORE, T, Custom);
145 setOperationAction(ISD::CTLZ, T, Custom);
146 setOperationAction(ISD::CTTZ, T, Custom);
147 setOperationAction(ISD::CTPOP, T, Custom);
149 setOperationAction(ISD::ADD, T, Legal);
150 setOperationAction(ISD::SUB, T, Legal);
151 setOperationAction(ISD::MUL, T, Custom);
152 setOperationAction(ISD::MULHS, T, Custom);
153 setOperationAction(ISD::MULHU, T, Custom);
154 setOperationAction(ISD::AND, T, Custom);
155 setOperationAction(ISD::OR, T, Custom);
156 setOperationAction(ISD::XOR, T, Custom);
157 setOperationAction(ISD::SETCC, T, Custom);
158 setOperationAction(ISD::VSELECT, T, Custom);
160 setOperationAction(ISD::SRA, T, Custom);
161 setOperationAction(ISD::SHL, T, Custom);
162 setOperationAction(ISD::SRL, T, Custom);
164 // Promote all shuffles to operate on vectors of bytes.
165 setPromoteTo(ISD::VECTOR_SHUFFLE, T, ByteW);
171 for (MVT T : LegalW) {
172 // Boolean types for vector pairs will overlap with the boolean
173 // types for single vectors, e.g.
174 // v64i8 -> v64i1 (single)
175 // v64i16 -> v64i1 (pair)
176 // Set these actions first, and allow the single actions to overwrite
178 MVT BoolW = MVT::getVectorVT(MVT::i1, T.getVectorNumElements());
179 setOperationAction(ISD::SETCC, BoolW, Custom);
180 setOperationAction(ISD::AND, BoolW, Custom);
181 setOperationAction(ISD::OR, BoolW, Custom);
182 setOperationAction(ISD::XOR, BoolW, Custom);
185 for (MVT T : LegalV) {
186 MVT BoolV = MVT::getVectorVT(MVT::i1, T.getVectorNumElements());
187 setOperationAction(ISD::BUILD_VECTOR, BoolV, Custom);
188 setOperationAction(ISD::CONCAT_VECTORS, BoolV, Custom);
189 setOperationAction(ISD::INSERT_SUBVECTOR, BoolV, Custom);
190 setOperationAction(ISD::INSERT_VECTOR_ELT, BoolV, Custom);
191 setOperationAction(ISD::EXTRACT_SUBVECTOR, BoolV, Custom);
192 setOperationAction(ISD::EXTRACT_VECTOR_ELT, BoolV, Custom);
193 setOperationAction(ISD::AND, BoolV, Legal);
194 setOperationAction(ISD::OR, BoolV, Legal);
195 setOperationAction(ISD::XOR, BoolV, Legal);
200 HexagonTargetLowering::getInt(unsigned IntId, MVT ResTy, ArrayRef<SDValue> Ops,
201 const SDLoc &dl, SelectionDAG &DAG) const {
202 SmallVector<SDValue,4> IntOps;
203 IntOps.push_back(DAG.getConstant(IntId, dl, MVT::i32));
204 for (const SDValue &Op : Ops)
205 IntOps.push_back(Op);
206 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, ResTy, IntOps);
210 HexagonTargetLowering::typeJoin(const TypePair &Tys) const {
211 assert(Tys.first.getVectorElementType() == Tys.second.getVectorElementType());
213 MVT ElemTy = Tys.first.getVectorElementType();
214 return MVT::getVectorVT(ElemTy, Tys.first.getVectorNumElements() +
215 Tys.second.getVectorNumElements());
218 HexagonTargetLowering::TypePair
219 HexagonTargetLowering::typeSplit(MVT VecTy) const {
220 assert(VecTy.isVector());
221 unsigned NumElem = VecTy.getVectorNumElements();
222 assert((NumElem % 2) == 0 && "Expecting even-sized vector type");
223 MVT HalfTy = MVT::getVectorVT(VecTy.getVectorElementType(), NumElem/2);
224 return { HalfTy, HalfTy };
228 HexagonTargetLowering::typeExtElem(MVT VecTy, unsigned Factor) const {
229 MVT ElemTy = VecTy.getVectorElementType();
230 MVT NewElemTy = MVT::getIntegerVT(ElemTy.getSizeInBits() * Factor);
231 return MVT::getVectorVT(NewElemTy, VecTy.getVectorNumElements());
235 HexagonTargetLowering::typeTruncElem(MVT VecTy, unsigned Factor) const {
236 MVT ElemTy = VecTy.getVectorElementType();
237 MVT NewElemTy = MVT::getIntegerVT(ElemTy.getSizeInBits() / Factor);
238 return MVT::getVectorVT(NewElemTy, VecTy.getVectorNumElements());
242 HexagonTargetLowering::opCastElem(SDValue Vec, MVT ElemTy,
243 SelectionDAG &DAG) const {
244 if (ty(Vec).getVectorElementType() == ElemTy)
246 MVT CastTy = tyVector(Vec.getValueType().getSimpleVT(), ElemTy);
247 return DAG.getBitcast(CastTy, Vec);
251 HexagonTargetLowering::opJoin(const VectorPair &Ops, const SDLoc &dl,
252 SelectionDAG &DAG) const {
253 return DAG.getNode(ISD::CONCAT_VECTORS, dl, typeJoin(ty(Ops)),
254 Ops.second, Ops.first);
257 HexagonTargetLowering::VectorPair
258 HexagonTargetLowering::opSplit(SDValue Vec, const SDLoc &dl,
259 SelectionDAG &DAG) const {
260 TypePair Tys = typeSplit(ty(Vec));
261 if (Vec.getOpcode() == HexagonISD::QCAT)
262 return VectorPair(Vec.getOperand(0), Vec.getOperand(1));
263 return DAG.SplitVector(Vec, dl, Tys.first, Tys.second);
267 HexagonTargetLowering::isHvxSingleTy(MVT Ty) const {
268 return Subtarget.isHVXVectorType(Ty) &&
269 Ty.getSizeInBits() == 8 * Subtarget.getVectorLength();
273 HexagonTargetLowering::isHvxPairTy(MVT Ty) const {
274 return Subtarget.isHVXVectorType(Ty) &&
275 Ty.getSizeInBits() == 16 * Subtarget.getVectorLength();
279 HexagonTargetLowering::convertToByteIndex(SDValue ElemIdx, MVT ElemTy,
280 SelectionDAG &DAG) const {
281 if (ElemIdx.getValueType().getSimpleVT() != MVT::i32)
282 ElemIdx = DAG.getBitcast(MVT::i32, ElemIdx);
284 unsigned ElemWidth = ElemTy.getSizeInBits();
288 unsigned L = Log2_32(ElemWidth/8);
289 const SDLoc &dl(ElemIdx);
290 return DAG.getNode(ISD::SHL, dl, MVT::i32,
291 {ElemIdx, DAG.getConstant(L, dl, MVT::i32)});
295 HexagonTargetLowering::getIndexInWord32(SDValue Idx, MVT ElemTy,
296 SelectionDAG &DAG) const {
297 unsigned ElemWidth = ElemTy.getSizeInBits();
298 assert(ElemWidth >= 8 && ElemWidth <= 32);
302 if (ty(Idx) != MVT::i32)
303 Idx = DAG.getBitcast(MVT::i32, Idx);
304 const SDLoc &dl(Idx);
305 SDValue Mask = DAG.getConstant(32/ElemWidth - 1, dl, MVT::i32);
306 SDValue SubIdx = DAG.getNode(ISD::AND, dl, MVT::i32, {Idx, Mask});
311 HexagonTargetLowering::getByteShuffle(const SDLoc &dl, SDValue Op0,
312 SDValue Op1, ArrayRef<int> Mask,
313 SelectionDAG &DAG) const {
315 assert(OpTy == ty(Op1));
317 MVT ElemTy = OpTy.getVectorElementType();
318 if (ElemTy == MVT::i8)
319 return DAG.getVectorShuffle(OpTy, dl, Op0, Op1, Mask);
320 assert(ElemTy.getSizeInBits() >= 8);
322 MVT ResTy = tyVector(OpTy, MVT::i8);
323 unsigned ElemSize = ElemTy.getSizeInBits() / 8;
325 SmallVector<int,128> ByteMask;
328 for (unsigned I = 0; I != ElemSize; ++I)
329 ByteMask.push_back(-1);
331 int NewM = M*ElemSize;
332 for (unsigned I = 0; I != ElemSize; ++I)
333 ByteMask.push_back(NewM+I);
336 assert(ResTy.getVectorNumElements() == ByteMask.size());
337 return DAG.getVectorShuffle(ResTy, dl, opCastElem(Op0, MVT::i8, DAG),
338 opCastElem(Op1, MVT::i8, DAG), ByteMask);
342 HexagonTargetLowering::buildHvxVectorReg(ArrayRef<SDValue> Values,
343 const SDLoc &dl, MVT VecTy,
344 SelectionDAG &DAG) const {
345 unsigned VecLen = Values.size();
346 MachineFunction &MF = DAG.getMachineFunction();
347 MVT ElemTy = VecTy.getVectorElementType();
348 unsigned ElemWidth = ElemTy.getSizeInBits();
349 unsigned HwLen = Subtarget.getVectorLength();
351 unsigned ElemSize = ElemWidth / 8;
352 assert(ElemSize*VecLen == HwLen);
353 SmallVector<SDValue,32> Words;
355 if (VecTy.getVectorElementType() != MVT::i32) {
356 assert((ElemSize == 1 || ElemSize == 2) && "Invalid element size");
357 unsigned OpsPerWord = (ElemSize == 1) ? 4 : 2;
358 MVT PartVT = MVT::getVectorVT(VecTy.getVectorElementType(), OpsPerWord);
359 for (unsigned i = 0; i != VecLen; i += OpsPerWord) {
360 SDValue W = buildVector32(Values.slice(i, OpsPerWord), dl, PartVT, DAG);
361 Words.push_back(DAG.getBitcast(MVT::i32, W));
364 Words.assign(Values.begin(), Values.end());
367 unsigned NumWords = Words.size();
368 bool IsSplat = true, IsUndef = true;
370 for (unsigned i = 0; i != NumWords && IsSplat; ++i) {
371 if (isUndef(Words[i]))
374 if (!SplatV.getNode())
376 else if (SplatV != Words[i])
380 return DAG.getUNDEF(VecTy);
382 assert(SplatV.getNode());
383 auto *IdxN = dyn_cast<ConstantSDNode>(SplatV.getNode());
384 if (IdxN && IdxN->isNullValue())
385 return getZero(dl, VecTy, DAG);
386 return DAG.getNode(HexagonISD::VSPLATW, dl, VecTy, SplatV);
389 // Delay recognizing constant vectors until here, so that we can generate
391 SmallVector<ConstantInt*, 128> Consts(VecLen);
392 bool AllConst = getBuildVectorConstInts(Values, VecTy, DAG, Consts);
394 ArrayRef<Constant*> Tmp((Constant**)Consts.begin(),
395 (Constant**)Consts.end());
396 Constant *CV = ConstantVector::get(Tmp);
397 unsigned Align = HwLen;
398 SDValue CP = LowerConstantPool(DAG.getConstantPool(CV, VecTy, Align), DAG);
399 return DAG.getLoad(VecTy, dl, DAG.getEntryNode(), CP,
400 MachinePointerInfo::getConstantPool(MF), Align);
403 // A special case is a situation where the vector is built entirely from
404 // elements extracted from another vector. This could be done via a shuffle
405 // more efficiently, but typically, the size of the source vector will not
406 // match the size of the vector being built (which precludes the use of a
407 // shuffle directly).
408 // This only handles a single source vector, and the vector being built
409 // should be of a sub-vector type of the source vector type.
410 auto IsBuildFromExtracts = [this,&Values] (SDValue &SrcVec,
411 SmallVectorImpl<int> &SrcIdx) {
413 for (SDValue V : Values) {
415 SrcIdx.push_back(-1);
418 if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
420 // All extracts should come from the same vector.
421 SDValue T = V.getOperand(0);
422 if (Vec.getNode() != nullptr && T.getNode() != Vec.getNode())
425 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
428 int I = C->getSExtValue();
429 assert(I >= 0 && "Negative element index");
436 SmallVector<int,128> ExtIdx;
438 if (IsBuildFromExtracts(ExtVec, ExtIdx)) {
439 MVT ExtTy = ty(ExtVec);
440 unsigned ExtLen = ExtTy.getVectorNumElements();
441 if (ExtLen == VecLen || ExtLen == 2*VecLen) {
442 // Construct a new shuffle mask that will produce a vector with the same
443 // number of elements as the input vector, and such that the vector we
444 // want will be the initial subvector of it.
445 SmallVector<int,128> Mask;
446 BitVector Used(ExtLen);
448 for (int M : ExtIdx) {
453 // Fill the rest of the mask with the unused elements of ExtVec in hopes
454 // that it will result in a permutation of ExtVec's elements. It's still
455 // fine if it doesn't (e.g. if undefs are present, or elements are
456 // repeated), but permutations can always be done efficiently via vdelta
458 for (unsigned I = 0; I != ExtLen; ++I) {
459 if (Mask.size() == ExtLen)
465 SDValue S = DAG.getVectorShuffle(ExtTy, dl, ExtVec,
466 DAG.getUNDEF(ExtTy), Mask);
467 if (ExtLen == VecLen)
469 return DAG.getTargetExtractSubreg(Hexagon::vsub_lo, dl, VecTy, S);
473 // Construct two halves in parallel, then or them together.
474 assert(4*Words.size() == Subtarget.getVectorLength());
475 SDValue HalfV0 = getInstr(Hexagon::V6_vd0, dl, VecTy, {}, DAG);
476 SDValue HalfV1 = getInstr(Hexagon::V6_vd0, dl, VecTy, {}, DAG);
477 SDValue S = DAG.getConstant(4, dl, MVT::i32);
478 for (unsigned i = 0; i != NumWords/2; ++i) {
479 SDValue N = DAG.getNode(HexagonISD::VINSERTW0, dl, VecTy,
481 SDValue M = DAG.getNode(HexagonISD::VINSERTW0, dl, VecTy,
482 {HalfV1, Words[i+NumWords/2]});
483 HalfV0 = DAG.getNode(HexagonISD::VROR, dl, VecTy, {N, S});
484 HalfV1 = DAG.getNode(HexagonISD::VROR, dl, VecTy, {M, S});
487 HalfV0 = DAG.getNode(HexagonISD::VROR, dl, VecTy,
488 {HalfV0, DAG.getConstant(HwLen/2, dl, MVT::i32)});
489 SDValue DstV = DAG.getNode(ISD::OR, dl, VecTy, {HalfV0, HalfV1});
494 HexagonTargetLowering::createHvxPrefixPred(SDValue PredV, const SDLoc &dl,
495 unsigned BitBytes, bool ZeroFill, SelectionDAG &DAG) const {
496 MVT PredTy = ty(PredV);
497 unsigned HwLen = Subtarget.getVectorLength();
498 MVT ByteTy = MVT::getVectorVT(MVT::i8, HwLen);
500 if (Subtarget.isHVXVectorType(PredTy, true)) {
501 // Move the vector predicate SubV to a vector register, and scale it
502 // down to match the representation (bytes per type element) that VecV
503 // uses. The scaling down will pick every 2nd or 4th (every Scale-th
504 // in general) element and put them at the front of the resulting
505 // vector. This subvector will then be inserted into the Q2V of VecV.
506 // To avoid having an operation that generates an illegal type (short
507 // vector), generate a full size vector.
509 SDValue T = DAG.getNode(HexagonISD::Q2V, dl, ByteTy, PredV);
510 SmallVector<int,128> Mask(HwLen);
511 // Scale = BitBytes(PredV) / Given BitBytes.
512 unsigned Scale = HwLen / (PredTy.getVectorNumElements() * BitBytes);
513 unsigned BlockLen = PredTy.getVectorNumElements() * BitBytes;
515 for (unsigned i = 0; i != HwLen; ++i) {
516 unsigned Num = i % Scale;
517 unsigned Off = i / Scale;
518 Mask[BlockLen*Num + Off] = i;
520 SDValue S = DAG.getVectorShuffle(ByteTy, dl, T, DAG.getUNDEF(ByteTy), Mask);
523 // Fill the bytes beyond BlockLen with 0s.
524 MVT BoolTy = MVT::getVectorVT(MVT::i1, HwLen);
525 SDValue Q = getInstr(Hexagon::V6_pred_scalar2, dl, BoolTy,
526 {DAG.getConstant(BlockLen, dl, MVT::i32)}, DAG);
527 SDValue M = DAG.getNode(HexagonISD::Q2V, dl, ByteTy, Q);
528 return DAG.getNode(ISD::AND, dl, ByteTy, S, M);
531 // Make sure that this is a valid scalar predicate.
532 assert(PredTy == MVT::v2i1 || PredTy == MVT::v4i1 || PredTy == MVT::v8i1);
534 unsigned Bytes = 8 / PredTy.getVectorNumElements();
535 SmallVector<SDValue,4> Words[2];
538 auto Lo32 = [&DAG, &dl] (SDValue P) {
539 return DAG.getTargetExtractSubreg(Hexagon::isub_lo, dl, MVT::i32, P);
541 auto Hi32 = [&DAG, &dl] (SDValue P) {
542 return DAG.getTargetExtractSubreg(Hexagon::isub_hi, dl, MVT::i32, P);
545 SDValue W0 = isUndef(PredV)
546 ? DAG.getUNDEF(MVT::i64)
547 : DAG.getNode(HexagonISD::P2D, dl, MVT::i64, PredV);
548 Words[IdxW].push_back(Hi32(W0));
549 Words[IdxW].push_back(Lo32(W0));
551 while (Bytes < BitBytes) {
556 for (const SDValue &W : Words[IdxW ^ 1]) {
557 SDValue T = expandPredicate(W, dl, DAG);
558 Words[IdxW].push_back(Hi32(T));
559 Words[IdxW].push_back(Lo32(T));
562 for (const SDValue &W : Words[IdxW ^ 1]) {
563 Words[IdxW].push_back(W);
564 Words[IdxW].push_back(W);
570 assert(Bytes == BitBytes);
572 SDValue Vec = ZeroFill ? getZero(dl, ByteTy, DAG) : DAG.getUNDEF(ByteTy);
573 SDValue S4 = DAG.getConstant(HwLen-4, dl, MVT::i32);
574 for (const SDValue &W : Words[IdxW]) {
575 Vec = DAG.getNode(HexagonISD::VROR, dl, ByteTy, Vec, S4);
576 Vec = DAG.getNode(HexagonISD::VINSERTW0, dl, ByteTy, Vec, W);
583 HexagonTargetLowering::buildHvxVectorPred(ArrayRef<SDValue> Values,
584 const SDLoc &dl, MVT VecTy,
585 SelectionDAG &DAG) const {
586 // Construct a vector V of bytes, such that a comparison V >u 0 would
587 // produce the required vector predicate.
588 unsigned VecLen = Values.size();
589 unsigned HwLen = Subtarget.getVectorLength();
590 assert(VecLen <= HwLen || VecLen == 8*HwLen);
591 SmallVector<SDValue,128> Bytes;
592 bool AllT = true, AllF = true;
594 auto IsTrue = [] (SDValue V) {
595 if (const auto *N = dyn_cast<ConstantSDNode>(V.getNode()))
596 return !N->isNullValue();
599 auto IsFalse = [] (SDValue V) {
600 if (const auto *N = dyn_cast<ConstantSDNode>(V.getNode()))
601 return N->isNullValue();
605 if (VecLen <= HwLen) {
606 // In the hardware, each bit of a vector predicate corresponds to a byte
607 // of a vector register. Calculate how many bytes does a bit of VecTy
609 assert(HwLen % VecLen == 0);
610 unsigned BitBytes = HwLen / VecLen;
611 for (SDValue V : Values) {
615 SDValue Ext = !V.isUndef() ? DAG.getZExtOrTrunc(V, dl, MVT::i8)
616 : DAG.getUNDEF(MVT::i8);
617 for (unsigned B = 0; B != BitBytes; ++B)
618 Bytes.push_back(Ext);
621 // There are as many i1 values, as there are bits in a vector register.
622 // Divide the values into groups of 8 and check that each group consists
623 // of the same value (ignoring undefs).
624 for (unsigned I = 0; I != VecLen; I += 8) {
626 // Find the first non-undef value in this group.
627 for (; B != 8; ++B) {
628 if (!Values[I+B].isUndef())
631 SDValue F = Values[I+B];
635 SDValue Ext = (B < 8) ? DAG.getZExtOrTrunc(F, dl, MVT::i8)
636 : DAG.getUNDEF(MVT::i8);
637 Bytes.push_back(Ext);
638 // Verify that the rest of values in the group are the same as the
641 assert(Values[I+B].isUndef() || Values[I+B] == F);
646 return DAG.getNode(HexagonISD::QTRUE, dl, VecTy);
648 return DAG.getNode(HexagonISD::QFALSE, dl, VecTy);
650 MVT ByteTy = MVT::getVectorVT(MVT::i8, HwLen);
651 SDValue ByteVec = buildHvxVectorReg(Bytes, dl, ByteTy, DAG);
652 return DAG.getNode(HexagonISD::V2Q, dl, VecTy, ByteVec);
656 HexagonTargetLowering::extractHvxElementReg(SDValue VecV, SDValue IdxV,
657 const SDLoc &dl, MVT ResTy, SelectionDAG &DAG) const {
658 MVT ElemTy = ty(VecV).getVectorElementType();
660 unsigned ElemWidth = ElemTy.getSizeInBits();
661 assert(ElemWidth >= 8 && ElemWidth <= 32);
664 SDValue ByteIdx = convertToByteIndex(IdxV, ElemTy, DAG);
665 SDValue ExWord = DAG.getNode(HexagonISD::VEXTRACTW, dl, MVT::i32,
667 if (ElemTy == MVT::i32)
670 // Have an extracted word, need to extract the smaller element out of it.
671 // 1. Extract the bits of (the original) IdxV that correspond to the index
672 // of the desired element in the 32-bit word.
673 SDValue SubIdx = getIndexInWord32(IdxV, ElemTy, DAG);
674 // 2. Extract the element from the word.
675 SDValue ExVec = DAG.getBitcast(tyVector(ty(ExWord), ElemTy), ExWord);
676 return extractVector(ExVec, SubIdx, dl, ElemTy, MVT::i32, DAG);
680 HexagonTargetLowering::extractHvxElementPred(SDValue VecV, SDValue IdxV,
681 const SDLoc &dl, MVT ResTy, SelectionDAG &DAG) const {
682 // Implement other return types if necessary.
683 assert(ResTy == MVT::i1);
685 unsigned HwLen = Subtarget.getVectorLength();
686 MVT ByteTy = MVT::getVectorVT(MVT::i8, HwLen);
687 SDValue ByteVec = DAG.getNode(HexagonISD::Q2V, dl, ByteTy, VecV);
689 unsigned Scale = HwLen / ty(VecV).getVectorNumElements();
690 SDValue ScV = DAG.getConstant(Scale, dl, MVT::i32);
691 IdxV = DAG.getNode(ISD::MUL, dl, MVT::i32, IdxV, ScV);
693 SDValue ExtB = extractHvxElementReg(ByteVec, IdxV, dl, MVT::i32, DAG);
694 SDValue Zero = DAG.getTargetConstant(0, dl, MVT::i32);
695 return getInstr(Hexagon::C2_cmpgtui, dl, MVT::i1, {ExtB, Zero}, DAG);
699 HexagonTargetLowering::insertHvxElementReg(SDValue VecV, SDValue IdxV,
700 SDValue ValV, const SDLoc &dl, SelectionDAG &DAG) const {
701 MVT ElemTy = ty(VecV).getVectorElementType();
703 unsigned ElemWidth = ElemTy.getSizeInBits();
704 assert(ElemWidth >= 8 && ElemWidth <= 32);
707 auto InsertWord = [&DAG,&dl,this] (SDValue VecV, SDValue ValV,
709 MVT VecTy = ty(VecV);
710 unsigned HwLen = Subtarget.getVectorLength();
711 SDValue MaskV = DAG.getNode(ISD::AND, dl, MVT::i32,
712 {ByteIdxV, DAG.getConstant(-4, dl, MVT::i32)});
713 SDValue RotV = DAG.getNode(HexagonISD::VROR, dl, VecTy, {VecV, MaskV});
714 SDValue InsV = DAG.getNode(HexagonISD::VINSERTW0, dl, VecTy, {RotV, ValV});
715 SDValue SubV = DAG.getNode(ISD::SUB, dl, MVT::i32,
716 {DAG.getConstant(HwLen, dl, MVT::i32), MaskV});
717 SDValue TorV = DAG.getNode(HexagonISD::VROR, dl, VecTy, {InsV, SubV});
721 SDValue ByteIdx = convertToByteIndex(IdxV, ElemTy, DAG);
722 if (ElemTy == MVT::i32)
723 return InsertWord(VecV, ValV, ByteIdx);
725 // If this is not inserting a 32-bit word, convert it into such a thing.
726 // 1. Extract the existing word from the target vector.
727 SDValue WordIdx = DAG.getNode(ISD::SRL, dl, MVT::i32,
728 {ByteIdx, DAG.getConstant(2, dl, MVT::i32)});
729 SDValue Ext = extractHvxElementReg(opCastElem(VecV, MVT::i32, DAG), WordIdx,
732 // 2. Treating the extracted word as a 32-bit vector, insert the given
734 SDValue SubIdx = getIndexInWord32(IdxV, ElemTy, DAG);
735 MVT SubVecTy = tyVector(ty(Ext), ElemTy);
736 SDValue Ins = insertVector(DAG.getBitcast(SubVecTy, Ext),
737 ValV, SubIdx, dl, ElemTy, DAG);
739 // 3. Insert the 32-bit word back into the original vector.
740 return InsertWord(VecV, Ins, ByteIdx);
744 HexagonTargetLowering::insertHvxElementPred(SDValue VecV, SDValue IdxV,
745 SDValue ValV, const SDLoc &dl, SelectionDAG &DAG) const {
746 unsigned HwLen = Subtarget.getVectorLength();
747 MVT ByteTy = MVT::getVectorVT(MVT::i8, HwLen);
748 SDValue ByteVec = DAG.getNode(HexagonISD::Q2V, dl, ByteTy, VecV);
750 unsigned Scale = HwLen / ty(VecV).getVectorNumElements();
751 SDValue ScV = DAG.getConstant(Scale, dl, MVT::i32);
752 IdxV = DAG.getNode(ISD::MUL, dl, MVT::i32, IdxV, ScV);
753 ValV = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, ValV);
755 SDValue InsV = insertHvxElementReg(ByteVec, IdxV, ValV, dl, DAG);
756 return DAG.getNode(HexagonISD::V2Q, dl, ty(VecV), InsV);
760 HexagonTargetLowering::extractHvxSubvectorReg(SDValue VecV, SDValue IdxV,
761 const SDLoc &dl, MVT ResTy, SelectionDAG &DAG) const {
762 MVT VecTy = ty(VecV);
763 unsigned HwLen = Subtarget.getVectorLength();
764 unsigned Idx = cast<ConstantSDNode>(IdxV.getNode())->getZExtValue();
765 MVT ElemTy = VecTy.getVectorElementType();
766 unsigned ElemWidth = ElemTy.getSizeInBits();
768 // If the source vector is a vector pair, get the single vector containing
769 // the subvector of interest. The subvector will never overlap two single
771 if (isHvxPairTy(VecTy)) {
773 if (Idx * ElemWidth >= 8*HwLen) {
774 SubIdx = Hexagon::vsub_hi;
775 Idx -= VecTy.getVectorNumElements() / 2;
777 SubIdx = Hexagon::vsub_lo;
779 VecTy = typeSplit(VecTy).first;
780 VecV = DAG.getTargetExtractSubreg(SubIdx, dl, VecTy, VecV);
785 // The only meaningful subvectors of a single HVX vector are those that
786 // fit in a scalar register.
787 assert(ResTy.getSizeInBits() == 32 || ResTy.getSizeInBits() == 64);
789 MVT WordTy = tyVector(VecTy, MVT::i32);
790 SDValue WordVec = DAG.getBitcast(WordTy, VecV);
791 unsigned WordIdx = (Idx*ElemWidth) / 32;
793 SDValue W0Idx = DAG.getConstant(WordIdx, dl, MVT::i32);
794 SDValue W0 = extractHvxElementReg(WordVec, W0Idx, dl, MVT::i32, DAG);
795 if (ResTy.getSizeInBits() == 32)
796 return DAG.getBitcast(ResTy, W0);
798 SDValue W1Idx = DAG.getConstant(WordIdx+1, dl, MVT::i32);
799 SDValue W1 = extractHvxElementReg(WordVec, W1Idx, dl, MVT::i32, DAG);
800 SDValue WW = DAG.getNode(HexagonISD::COMBINE, dl, MVT::i64, {W1, W0});
801 return DAG.getBitcast(ResTy, WW);
805 HexagonTargetLowering::extractHvxSubvectorPred(SDValue VecV, SDValue IdxV,
806 const SDLoc &dl, MVT ResTy, SelectionDAG &DAG) const {
807 MVT VecTy = ty(VecV);
808 unsigned HwLen = Subtarget.getVectorLength();
809 MVT ByteTy = MVT::getVectorVT(MVT::i8, HwLen);
810 SDValue ByteVec = DAG.getNode(HexagonISD::Q2V, dl, ByteTy, VecV);
811 // IdxV is required to be a constant.
812 unsigned Idx = cast<ConstantSDNode>(IdxV.getNode())->getZExtValue();
814 unsigned ResLen = ResTy.getVectorNumElements();
815 unsigned BitBytes = HwLen / VecTy.getVectorNumElements();
816 unsigned Offset = Idx * BitBytes;
817 SDValue Undef = DAG.getUNDEF(ByteTy);
818 SmallVector<int,128> Mask;
820 if (Subtarget.isHVXVectorType(ResTy, true)) {
821 // Converting between two vector predicates. Since the result is shorter
822 // than the source, it will correspond to a vector predicate with the
823 // relevant bits replicated. The replication count is the ratio of the
824 // source and target vector lengths.
825 unsigned Rep = VecTy.getVectorNumElements() / ResLen;
826 assert(isPowerOf2_32(Rep) && HwLen % Rep == 0);
827 for (unsigned i = 0; i != HwLen/Rep; ++i) {
828 for (unsigned j = 0; j != Rep; ++j)
829 Mask.push_back(i + Offset);
831 SDValue ShuffV = DAG.getVectorShuffle(ByteTy, dl, ByteVec, Undef, Mask);
832 return DAG.getNode(HexagonISD::V2Q, dl, ResTy, ShuffV);
835 // Converting between a vector predicate and a scalar predicate. In the
836 // vector predicate, a group of BitBytes bits will correspond to a single
837 // i1 element of the source vector type. Those bits will all have the same
838 // value. The same will be true for ByteVec, where each byte corresponds
839 // to a bit in the vector predicate.
840 // The algorithm is to traverse the ByteVec, going over the i1 values from
841 // the source vector, and generate the corresponding representation in an
842 // 8-byte vector. To avoid repeated extracts from ByteVec, shuffle the
843 // elements so that the interesting 8 bytes will be in the low end of the
845 unsigned Rep = 8 / ResLen;
846 // Make sure the output fill the entire vector register, so repeat the
847 // 8-byte groups as many times as necessary.
848 for (unsigned r = 0; r != HwLen/ResLen; ++r) {
849 // This will generate the indexes of the 8 interesting bytes.
850 for (unsigned i = 0; i != ResLen; ++i) {
851 for (unsigned j = 0; j != Rep; ++j)
852 Mask.push_back(Offset + i*BitBytes);
856 SDValue Zero = getZero(dl, MVT::i32, DAG);
857 SDValue ShuffV = DAG.getVectorShuffle(ByteTy, dl, ByteVec, Undef, Mask);
858 // Combine the two low words from ShuffV into a v8i8, and byte-compare
860 SDValue W0 = DAG.getNode(HexagonISD::VEXTRACTW, dl, MVT::i32, {ShuffV, Zero});
861 SDValue W1 = DAG.getNode(HexagonISD::VEXTRACTW, dl, MVT::i32,
862 {ShuffV, DAG.getConstant(4, dl, MVT::i32)});
863 SDValue Vec64 = DAG.getNode(HexagonISD::COMBINE, dl, MVT::v8i8, {W1, W0});
864 return getInstr(Hexagon::A4_vcmpbgtui, dl, ResTy,
865 {Vec64, DAG.getTargetConstant(0, dl, MVT::i32)}, DAG);
869 HexagonTargetLowering::insertHvxSubvectorReg(SDValue VecV, SDValue SubV,
870 SDValue IdxV, const SDLoc &dl, SelectionDAG &DAG) const {
871 MVT VecTy = ty(VecV);
872 MVT SubTy = ty(SubV);
873 unsigned HwLen = Subtarget.getVectorLength();
874 MVT ElemTy = VecTy.getVectorElementType();
875 unsigned ElemWidth = ElemTy.getSizeInBits();
877 bool IsPair = isHvxPairTy(VecTy);
878 MVT SingleTy = MVT::getVectorVT(ElemTy, (8*HwLen)/ElemWidth);
879 // The two single vectors that VecV consists of, if it's a pair.
881 SDValue SingleV = VecV;
885 V0 = DAG.getTargetExtractSubreg(Hexagon::vsub_lo, dl, SingleTy, VecV);
886 V1 = DAG.getTargetExtractSubreg(Hexagon::vsub_hi, dl, SingleTy, VecV);
888 SDValue HalfV = DAG.getConstant(SingleTy.getVectorNumElements(),
890 PickHi = DAG.getSetCC(dl, MVT::i1, IdxV, HalfV, ISD::SETUGT);
891 if (isHvxSingleTy(SubTy)) {
892 if (const auto *CN = dyn_cast<const ConstantSDNode>(IdxV.getNode())) {
893 unsigned Idx = CN->getZExtValue();
894 assert(Idx == 0 || Idx == VecTy.getVectorNumElements()/2);
895 unsigned SubIdx = (Idx == 0) ? Hexagon::vsub_lo : Hexagon::vsub_hi;
896 return DAG.getTargetInsertSubreg(SubIdx, dl, VecTy, VecV, SubV);
898 // If IdxV is not a constant, generate the two variants: with the
899 // SubV as the high and as the low subregister, and select the right
900 // pair based on the IdxV.
901 SDValue InLo = DAG.getNode(ISD::CONCAT_VECTORS, dl, VecTy, {SubV, V1});
902 SDValue InHi = DAG.getNode(ISD::CONCAT_VECTORS, dl, VecTy, {V0, SubV});
903 return DAG.getNode(ISD::SELECT, dl, VecTy, PickHi, InHi, InLo);
905 // The subvector being inserted must be entirely contained in one of
906 // the vectors V0 or V1. Set SingleV to the correct one, and update
907 // IdxV to be the index relative to the beginning of that vector.
908 SDValue S = DAG.getNode(ISD::SUB, dl, MVT::i32, IdxV, HalfV);
909 IdxV = DAG.getNode(ISD::SELECT, dl, MVT::i32, PickHi, S, IdxV);
910 SingleV = DAG.getNode(ISD::SELECT, dl, SingleTy, PickHi, V1, V0);
913 // The only meaningful subvectors of a single HVX vector are those that
914 // fit in a scalar register.
915 assert(SubTy.getSizeInBits() == 32 || SubTy.getSizeInBits() == 64);
916 // Convert IdxV to be index in bytes.
917 auto *IdxN = dyn_cast<ConstantSDNode>(IdxV.getNode());
918 if (!IdxN || !IdxN->isNullValue()) {
919 IdxV = DAG.getNode(ISD::MUL, dl, MVT::i32, IdxV,
920 DAG.getConstant(ElemWidth/8, dl, MVT::i32));
921 SingleV = DAG.getNode(HexagonISD::VROR, dl, SingleTy, SingleV, IdxV);
923 // When inserting a single word, the rotation back to the original position
924 // would be by HwLen-Idx, but if two words are inserted, it will need to be
926 unsigned RolBase = HwLen;
927 if (VecTy.getSizeInBits() == 32) {
928 SDValue V = DAG.getBitcast(MVT::i32, SubV);
929 SingleV = DAG.getNode(HexagonISD::VINSERTW0, dl, SingleTy, V);
931 SDValue V = DAG.getBitcast(MVT::i64, SubV);
932 SDValue R0 = DAG.getTargetExtractSubreg(Hexagon::isub_lo, dl, MVT::i32, V);
933 SDValue R1 = DAG.getTargetExtractSubreg(Hexagon::isub_hi, dl, MVT::i32, V);
934 SingleV = DAG.getNode(HexagonISD::VINSERTW0, dl, SingleTy, SingleV, R0);
935 SingleV = DAG.getNode(HexagonISD::VROR, dl, SingleTy, SingleV,
936 DAG.getConstant(4, dl, MVT::i32));
937 SingleV = DAG.getNode(HexagonISD::VINSERTW0, dl, SingleTy, SingleV, R1);
940 // If the vector wasn't ror'ed, don't ror it back.
941 if (RolBase != 4 || !IdxN || !IdxN->isNullValue()) {
942 SDValue RolV = DAG.getNode(ISD::SUB, dl, MVT::i32,
943 DAG.getConstant(RolBase, dl, MVT::i32), IdxV);
944 SingleV = DAG.getNode(HexagonISD::VROR, dl, SingleTy, SingleV, RolV);
948 SDValue InLo = DAG.getNode(ISD::CONCAT_VECTORS, dl, VecTy, {SingleV, V1});
949 SDValue InHi = DAG.getNode(ISD::CONCAT_VECTORS, dl, VecTy, {V0, SingleV});
950 return DAG.getNode(ISD::SELECT, dl, VecTy, PickHi, InHi, InLo);
956 HexagonTargetLowering::insertHvxSubvectorPred(SDValue VecV, SDValue SubV,
957 SDValue IdxV, const SDLoc &dl, SelectionDAG &DAG) const {
958 MVT VecTy = ty(VecV);
959 MVT SubTy = ty(SubV);
960 assert(Subtarget.isHVXVectorType(VecTy, true));
961 // VecV is an HVX vector predicate. SubV may be either an HVX vector
962 // predicate as well, or it can be a scalar predicate.
964 unsigned VecLen = VecTy.getVectorNumElements();
965 unsigned HwLen = Subtarget.getVectorLength();
966 assert(HwLen % VecLen == 0 && "Unexpected vector type");
968 unsigned Scale = VecLen / SubTy.getVectorNumElements();
969 unsigned BitBytes = HwLen / VecLen;
970 unsigned BlockLen = HwLen / Scale;
972 MVT ByteTy = MVT::getVectorVT(MVT::i8, HwLen);
973 SDValue ByteVec = DAG.getNode(HexagonISD::Q2V, dl, ByteTy, VecV);
974 SDValue ByteSub = createHvxPrefixPred(SubV, dl, BitBytes, false, DAG);
977 auto *IdxN = dyn_cast<ConstantSDNode>(IdxV.getNode());
978 if (!IdxN || !IdxN->isNullValue()) {
979 ByteIdx = DAG.getNode(ISD::MUL, dl, MVT::i32, IdxV,
980 DAG.getConstant(BitBytes, dl, MVT::i32));
981 ByteVec = DAG.getNode(HexagonISD::VROR, dl, ByteTy, ByteVec, ByteIdx);
984 // ByteVec is the target vector VecV rotated in such a way that the
985 // subvector should be inserted at index 0. Generate a predicate mask
986 // and use vmux to do the insertion.
987 MVT BoolTy = MVT::getVectorVT(MVT::i1, HwLen);
988 SDValue Q = getInstr(Hexagon::V6_pred_scalar2, dl, BoolTy,
989 {DAG.getConstant(BlockLen, dl, MVT::i32)}, DAG);
990 ByteVec = getInstr(Hexagon::V6_vmux, dl, ByteTy, {Q, ByteSub, ByteVec}, DAG);
991 // Rotate ByteVec back, and convert to a vector predicate.
992 if (!IdxN || !IdxN->isNullValue()) {
993 SDValue HwLenV = DAG.getConstant(HwLen, dl, MVT::i32);
994 SDValue ByteXdi = DAG.getNode(ISD::SUB, dl, MVT::i32, HwLenV, ByteIdx);
995 ByteVec = DAG.getNode(HexagonISD::VROR, dl, ByteTy, ByteVec, ByteXdi);
997 return DAG.getNode(HexagonISD::V2Q, dl, VecTy, ByteVec);
1001 HexagonTargetLowering::extendHvxVectorPred(SDValue VecV, const SDLoc &dl,
1002 MVT ResTy, bool ZeroExt, SelectionDAG &DAG) const {
1003 // Sign- and any-extending of a vector predicate to a vector register is
1004 // equivalent to Q2V. For zero-extensions, generate a vmux between 0 and
1005 // a vector of 1s (where the 1s are of type matching the vector type).
1006 assert(Subtarget.isHVXVectorType(ResTy));
1008 return DAG.getNode(HexagonISD::Q2V, dl, ResTy, VecV);
1010 assert(ty(VecV).getVectorNumElements() == ResTy.getVectorNumElements());
1011 SDValue True = DAG.getNode(HexagonISD::VSPLAT, dl, ResTy,
1012 DAG.getConstant(1, dl, MVT::i32));
1013 SDValue False = getZero(dl, ResTy, DAG);
1014 return DAG.getSelect(dl, ResTy, VecV, True, False);
1018 HexagonTargetLowering::LowerHvxBuildVector(SDValue Op, SelectionDAG &DAG)
1020 const SDLoc &dl(Op);
1023 unsigned Size = Op.getNumOperands();
1024 SmallVector<SDValue,128> Ops;
1025 for (unsigned i = 0; i != Size; ++i)
1026 Ops.push_back(Op.getOperand(i));
1028 if (VecTy.getVectorElementType() == MVT::i1)
1029 return buildHvxVectorPred(Ops, dl, VecTy, DAG);
1031 if (VecTy.getSizeInBits() == 16*Subtarget.getVectorLength()) {
1032 ArrayRef<SDValue> A(Ops);
1033 MVT SingleTy = typeSplit(VecTy).first;
1034 SDValue V0 = buildHvxVectorReg(A.take_front(Size/2), dl, SingleTy, DAG);
1035 SDValue V1 = buildHvxVectorReg(A.drop_front(Size/2), dl, SingleTy, DAG);
1036 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VecTy, V0, V1);
1039 return buildHvxVectorReg(Ops, dl, VecTy, DAG);
1043 HexagonTargetLowering::LowerHvxConcatVectors(SDValue Op, SelectionDAG &DAG)
1045 // Vector concatenation of two integer (non-bool) vectors does not need
1046 // special lowering. Custom-lower concats of bool vectors and expand
1047 // concats of more than 2 vectors.
1049 const SDLoc &dl(Op);
1050 unsigned NumOp = Op.getNumOperands();
1051 if (VecTy.getVectorElementType() != MVT::i1) {
1054 // Expand the other cases into a build-vector.
1055 SmallVector<SDValue,8> Elems;
1056 for (SDValue V : Op.getNode()->ops())
1057 DAG.ExtractVectorElements(V, Elems);
1058 // A vector of i16 will be broken up into a build_vector of i16's.
1059 // This is a problem, since at the time of operation legalization,
1060 // all operations are expected to be type-legalized, and i16 is not
1061 // a legal type. If any of the extracted elements is not of a valid
1062 // type, sign-extend it to a valid one.
1063 for (unsigned i = 0, e = Elems.size(); i != e; ++i) {
1064 SDValue V = Elems[i];
1066 if (!isTypeLegal(Ty)) {
1067 EVT NTy = getTypeToTransformTo(*DAG.getContext(), Ty);
1068 if (V.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
1069 Elems[i] = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NTy,
1070 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NTy,
1071 V.getOperand(0), V.getOperand(1)),
1072 DAG.getValueType(Ty));
1075 // A few less complicated cases.
1076 if (V.getOpcode() == ISD::Constant)
1077 Elems[i] = DAG.getSExtOrTrunc(V, dl, NTy);
1078 else if (V.isUndef())
1079 Elems[i] = DAG.getUNDEF(NTy);
1081 llvm_unreachable("Unexpected vector element");
1084 return DAG.getBuildVector(VecTy, dl, Elems);
1087 assert(VecTy.getVectorElementType() == MVT::i1);
1088 unsigned HwLen = Subtarget.getVectorLength();
1089 assert(isPowerOf2_32(NumOp) && HwLen % NumOp == 0);
1091 SDValue Op0 = Op.getOperand(0);
1093 // If the operands are HVX types (i.e. not scalar predicates), then
1094 // defer the concatenation, and create QCAT instead.
1095 if (Subtarget.isHVXVectorType(ty(Op0), true)) {
1097 return DAG.getNode(HexagonISD::QCAT, dl, VecTy, Op0, Op.getOperand(1));
1099 ArrayRef<SDUse> U(Op.getNode()->ops());
1100 SmallVector<SDValue,4> SV(U.begin(), U.end());
1101 ArrayRef<SDValue> Ops(SV);
1103 MVT HalfTy = typeSplit(VecTy).first;
1104 SDValue V0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfTy,
1105 Ops.take_front(NumOp/2));
1106 SDValue V1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfTy,
1107 Ops.take_back(NumOp/2));
1108 return DAG.getNode(HexagonISD::QCAT, dl, VecTy, V0, V1);
1111 // Count how many bytes (in a vector register) each bit in VecTy
1113 unsigned BitBytes = HwLen / VecTy.getVectorNumElements();
1115 SmallVector<SDValue,8> Prefixes;
1116 for (SDValue V : Op.getNode()->op_values()) {
1117 SDValue P = createHvxPrefixPred(V, dl, BitBytes, true, DAG);
1118 Prefixes.push_back(P);
1121 unsigned InpLen = ty(Op.getOperand(0)).getVectorNumElements();
1122 MVT ByteTy = MVT::getVectorVT(MVT::i8, HwLen);
1123 SDValue S = DAG.getConstant(InpLen*BitBytes, dl, MVT::i32);
1124 SDValue Res = getZero(dl, ByteTy, DAG);
1125 for (unsigned i = 0, e = Prefixes.size(); i != e; ++i) {
1126 Res = DAG.getNode(HexagonISD::VROR, dl, ByteTy, Res, S);
1127 Res = DAG.getNode(ISD::OR, dl, ByteTy, Res, Prefixes[e-i-1]);
1129 return DAG.getNode(HexagonISD::V2Q, dl, VecTy, Res);
1133 HexagonTargetLowering::LowerHvxExtractElement(SDValue Op, SelectionDAG &DAG)
1135 // Change the type of the extracted element to i32.
1136 SDValue VecV = Op.getOperand(0);
1137 MVT ElemTy = ty(VecV).getVectorElementType();
1138 const SDLoc &dl(Op);
1139 SDValue IdxV = Op.getOperand(1);
1140 if (ElemTy == MVT::i1)
1141 return extractHvxElementPred(VecV, IdxV, dl, ty(Op), DAG);
1143 return extractHvxElementReg(VecV, IdxV, dl, ty(Op), DAG);
1147 HexagonTargetLowering::LowerHvxInsertElement(SDValue Op, SelectionDAG &DAG)
1149 const SDLoc &dl(Op);
1150 SDValue VecV = Op.getOperand(0);
1151 SDValue ValV = Op.getOperand(1);
1152 SDValue IdxV = Op.getOperand(2);
1153 MVT ElemTy = ty(VecV).getVectorElementType();
1154 if (ElemTy == MVT::i1)
1155 return insertHvxElementPred(VecV, IdxV, ValV, dl, DAG);
1157 return insertHvxElementReg(VecV, IdxV, ValV, dl, DAG);
1161 HexagonTargetLowering::LowerHvxExtractSubvector(SDValue Op, SelectionDAG &DAG)
1163 SDValue SrcV = Op.getOperand(0);
1164 MVT SrcTy = ty(SrcV);
1166 SDValue IdxV = Op.getOperand(1);
1167 unsigned Idx = cast<ConstantSDNode>(IdxV.getNode())->getZExtValue();
1168 assert(Idx % DstTy.getVectorNumElements() == 0);
1170 const SDLoc &dl(Op);
1172 MVT ElemTy = SrcTy.getVectorElementType();
1173 if (ElemTy == MVT::i1)
1174 return extractHvxSubvectorPred(SrcV, IdxV, dl, DstTy, DAG);
1176 return extractHvxSubvectorReg(SrcV, IdxV, dl, DstTy, DAG);
1180 HexagonTargetLowering::LowerHvxInsertSubvector(SDValue Op, SelectionDAG &DAG)
1182 // Idx does not need to be a constant.
1183 SDValue VecV = Op.getOperand(0);
1184 SDValue ValV = Op.getOperand(1);
1185 SDValue IdxV = Op.getOperand(2);
1187 const SDLoc &dl(Op);
1188 MVT VecTy = ty(VecV);
1189 MVT ElemTy = VecTy.getVectorElementType();
1190 if (ElemTy == MVT::i1)
1191 return insertHvxSubvectorPred(VecV, ValV, IdxV, dl, DAG);
1193 return insertHvxSubvectorReg(VecV, ValV, IdxV, dl, DAG);
1197 HexagonTargetLowering::LowerHvxAnyExt(SDValue Op, SelectionDAG &DAG) const {
1198 // Lower any-extends of boolean vectors to sign-extends, since they
1199 // translate directly to Q2V. Zero-extending could also be done equally
1200 // fast, but Q2V is used/recognized in more places.
1201 // For all other vectors, use zero-extend.
1203 SDValue InpV = Op.getOperand(0);
1204 MVT ElemTy = ty(InpV).getVectorElementType();
1205 if (ElemTy == MVT::i1 && Subtarget.isHVXVectorType(ResTy))
1206 return LowerHvxSignExt(Op, DAG);
1207 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(Op), ResTy, InpV);
1211 HexagonTargetLowering::LowerHvxSignExt(SDValue Op, SelectionDAG &DAG) const {
1213 SDValue InpV = Op.getOperand(0);
1214 MVT ElemTy = ty(InpV).getVectorElementType();
1215 if (ElemTy == MVT::i1 && Subtarget.isHVXVectorType(ResTy))
1216 return extendHvxVectorPred(InpV, SDLoc(Op), ty(Op), false, DAG);
1221 HexagonTargetLowering::LowerHvxZeroExt(SDValue Op, SelectionDAG &DAG) const {
1223 SDValue InpV = Op.getOperand(0);
1224 MVT ElemTy = ty(InpV).getVectorElementType();
1225 if (ElemTy == MVT::i1 && Subtarget.isHVXVectorType(ResTy))
1226 return extendHvxVectorPred(InpV, SDLoc(Op), ty(Op), true, DAG);
1231 HexagonTargetLowering::LowerHvxCttz(SDValue Op, SelectionDAG &DAG) const {
1232 // Lower vector CTTZ into a computation using CTLZ (Hacker's Delight):
1233 // cttz(x) = bitwidth(x) - ctlz(~x & (x-1))
1234 const SDLoc &dl(Op);
1236 SDValue InpV = Op.getOperand(0);
1237 assert(ResTy == ty(InpV));
1239 // Calculate the vectors of 1 and bitwidth(x).
1240 MVT ElemTy = ty(InpV).getVectorElementType();
1241 unsigned ElemWidth = ElemTy.getSizeInBits();
1242 // Using uint64_t because a shift by 32 can happen.
1243 uint64_t Splat1 = 0, SplatW = 0;
1244 assert(isPowerOf2_32(ElemWidth) && ElemWidth <= 32);
1245 for (unsigned i = 0; i != 32/ElemWidth; ++i) {
1246 Splat1 = (Splat1 << ElemWidth) | 1;
1247 SplatW = (SplatW << ElemWidth) | ElemWidth;
1249 SDValue Vec1 = DAG.getNode(HexagonISD::VSPLATW, dl, ResTy,
1250 DAG.getConstant(uint32_t(Splat1), dl, MVT::i32));
1251 SDValue VecW = DAG.getNode(HexagonISD::VSPLATW, dl, ResTy,
1252 DAG.getConstant(uint32_t(SplatW), dl, MVT::i32));
1253 SDValue VecN1 = DAG.getNode(HexagonISD::VSPLATW, dl, ResTy,
1254 DAG.getConstant(-1, dl, MVT::i32));
1255 // Do not use DAG.getNOT, because that would create BUILD_VECTOR with
1256 // a BITCAST. Here we can skip the BITCAST (so we don't have to handle
1257 // it separately in custom combine or selection).
1258 SDValue A = DAG.getNode(ISD::AND, dl, ResTy,
1259 {DAG.getNode(ISD::XOR, dl, ResTy, {InpV, VecN1}),
1260 DAG.getNode(ISD::SUB, dl, ResTy, {InpV, Vec1})});
1261 return DAG.getNode(ISD::SUB, dl, ResTy,
1262 {VecW, DAG.getNode(ISD::CTLZ, dl, ResTy, A)});
1266 HexagonTargetLowering::LowerHvxMul(SDValue Op, SelectionDAG &DAG) const {
1268 assert(ResTy.isVector() && isHvxSingleTy(ResTy));
1269 const SDLoc &dl(Op);
1270 SmallVector<int,256> ShuffMask;
1272 MVT ElemTy = ResTy.getVectorElementType();
1273 unsigned VecLen = ResTy.getVectorNumElements();
1274 SDValue Vs = Op.getOperand(0);
1275 SDValue Vt = Op.getOperand(1);
1277 switch (ElemTy.SimpleTy) {
1279 // For i8 vectors Vs = (a0, a1, ...), Vt = (b0, b1, ...),
1280 // V6_vmpybv Vs, Vt produces a pair of i16 vectors Hi:Lo,
1281 // where Lo = (a0*b0, a2*b2, ...), Hi = (a1*b1, a3*b3, ...).
1282 MVT ExtTy = typeExtElem(ResTy, 2);
1283 unsigned MpyOpc = ElemTy == MVT::i8 ? Hexagon::V6_vmpybv
1284 : Hexagon::V6_vmpyhv;
1285 SDValue M = getInstr(MpyOpc, dl, ExtTy, {Vs, Vt}, DAG);
1287 // Discard high halves of the resulting values, collect the low halves.
1288 for (unsigned I = 0; I < VecLen; I += 2) {
1289 ShuffMask.push_back(I); // Pick even element.
1290 ShuffMask.push_back(I+VecLen); // Pick odd element.
1292 VectorPair P = opSplit(opCastElem(M, ElemTy, DAG), dl, DAG);
1293 SDValue BS = getByteShuffle(dl, P.first, P.second, ShuffMask, DAG);
1294 return DAG.getBitcast(ResTy, BS);
1297 // For i16 there is V6_vmpyih, which acts exactly like the MUL opcode.
1298 // (There is also V6_vmpyhv, which behaves in an analogous way to
1300 return getInstr(Hexagon::V6_vmpyih, dl, ResTy, {Vs, Vt}, DAG);
1302 // Use the following sequence for signed word multiply:
1303 // T0 = V6_vmpyiowh Vs, Vt
1304 // T1 = V6_vaslw T0, 16
1305 // T2 = V6_vmpyiewuh_acc T1, Vs, Vt
1306 SDValue S16 = DAG.getConstant(16, dl, MVT::i32);
1307 SDValue T0 = getInstr(Hexagon::V6_vmpyiowh, dl, ResTy, {Vs, Vt}, DAG);
1308 SDValue T1 = getInstr(Hexagon::V6_vaslw, dl, ResTy, {T0, S16}, DAG);
1309 SDValue T2 = getInstr(Hexagon::V6_vmpyiewuh_acc, dl, ResTy,
1320 HexagonTargetLowering::LowerHvxMulh(SDValue Op, SelectionDAG &DAG) const {
1322 assert(ResTy.isVector());
1323 const SDLoc &dl(Op);
1324 SmallVector<int,256> ShuffMask;
1326 MVT ElemTy = ResTy.getVectorElementType();
1327 unsigned VecLen = ResTy.getVectorNumElements();
1328 SDValue Vs = Op.getOperand(0);
1329 SDValue Vt = Op.getOperand(1);
1330 bool IsSigned = Op.getOpcode() == ISD::MULHS;
1332 if (ElemTy == MVT::i8 || ElemTy == MVT::i16) {
1333 // For i8 vectors Vs = (a0, a1, ...), Vt = (b0, b1, ...),
1334 // V6_vmpybv Vs, Vt produces a pair of i16 vectors Hi:Lo,
1335 // where Lo = (a0*b0, a2*b2, ...), Hi = (a1*b1, a3*b3, ...).
1336 // For i16, use V6_vmpyhv, which behaves in an analogous way to
1337 // V6_vmpybv: results Lo and Hi are products of even/odd elements
1339 MVT ExtTy = typeExtElem(ResTy, 2);
1340 unsigned MpyOpc = ElemTy == MVT::i8
1341 ? (IsSigned ? Hexagon::V6_vmpybv : Hexagon::V6_vmpyubv)
1342 : (IsSigned ? Hexagon::V6_vmpyhv : Hexagon::V6_vmpyuhv);
1343 SDValue M = getInstr(MpyOpc, dl, ExtTy, {Vs, Vt}, DAG);
1345 // Discard low halves of the resulting values, collect the high halves.
1346 for (unsigned I = 0; I < VecLen; I += 2) {
1347 ShuffMask.push_back(I+1); // Pick even element.
1348 ShuffMask.push_back(I+VecLen+1); // Pick odd element.
1350 VectorPair P = opSplit(opCastElem(M, ElemTy, DAG), dl, DAG);
1351 SDValue BS = getByteShuffle(dl, P.first, P.second, ShuffMask, DAG);
1352 return DAG.getBitcast(ResTy, BS);
1355 assert(ElemTy == MVT::i32);
1356 SDValue S16 = DAG.getConstant(16, dl, MVT::i32);
1360 // = [(Hi(Vs)*2^16 + Lo(Vs)) *s (Hi(Vt)*2^16 + Lo(Vt))] >> 32
1361 // = [Hi(Vs)*2^16 *s Hi(Vt)*2^16 + Hi(Vs) *su Lo(Vt)*2^16
1362 // + Lo(Vs) *us (Hi(Vt)*2^16 + Lo(Vt))] >> 32
1363 // = [Hi(Vs) *s Hi(Vt)*2^32 + Hi(Vs) *su Lo(Vt)*2^16
1364 // + Lo(Vs) *us Vt] >> 32
1365 // The low half of Lo(Vs)*Lo(Vt) will be discarded (it's not added to
1366 // anything, so it cannot produce any carry over to higher bits),
1367 // so everything in [] can be shifted by 16 without loss of precision.
1368 // = [Hi(Vs) *s Hi(Vt)*2^16 + Hi(Vs)*su Lo(Vt) + Lo(Vs)*Vt >> 16] >> 16
1369 // = [Hi(Vs) *s Hi(Vt)*2^16 + Hi(Vs)*su Lo(Vt) + V6_vmpyewuh(Vs,Vt)] >> 16
1370 // Denote Hi(Vs) = Vs':
1371 // = [Vs'*s Hi(Vt)*2^16 + Vs' *su Lo(Vt) + V6_vmpyewuh(Vt,Vs)] >> 16
1372 // = Vs'*s Hi(Vt) + (V6_vmpyiewuh(Vs',Vt) + V6_vmpyewuh(Vt,Vs)) >> 16
1373 SDValue T0 = getInstr(Hexagon::V6_vmpyewuh, dl, ResTy, {Vt, Vs}, DAG);
1375 SDValue S0 = getInstr(Hexagon::V6_vasrw, dl, ResTy, {Vs, S16}, DAG);
1376 SDValue T1 = getInstr(Hexagon::V6_vmpyiewuh_acc, dl, ResTy,
1379 SDValue S2 = getInstr(Hexagon::V6_vasrw, dl, ResTy, {T1, S16}, DAG);
1381 SDValue T2 = getInstr(Hexagon::V6_vmpyiowh, dl, ResTy, {S0, Vt}, DAG);
1383 SDValue T3 = DAG.getNode(ISD::ADD, dl, ResTy, {S2, T2});
1387 // Unsigned mulhw. (Would expansion using signed mulhw be better?)
1389 auto LoVec = [&DAG,ResTy,dl] (SDValue Pair) {
1390 return DAG.getTargetExtractSubreg(Hexagon::vsub_lo, dl, ResTy, Pair);
1392 auto HiVec = [&DAG,ResTy,dl] (SDValue Pair) {
1393 return DAG.getTargetExtractSubreg(Hexagon::vsub_hi, dl, ResTy, Pair);
1396 MVT PairTy = typeJoin({ResTy, ResTy});
1397 SDValue P = getInstr(Hexagon::V6_lvsplatw, dl, ResTy,
1398 {DAG.getConstant(0x02020202, dl, MVT::i32)}, DAG);
1399 // Multiply-unsigned halfwords:
1400 // LoVec = Vs.uh[2i] * Vt.uh[2i],
1401 // HiVec = Vs.uh[2i+1] * Vt.uh[2i+1]
1402 SDValue T0 = getInstr(Hexagon::V6_vmpyuhv, dl, PairTy, {Vs, Vt}, DAG);
1403 // The low halves in the LoVec of the pair can be discarded. They are
1404 // not added to anything (in the full-precision product), so they cannot
1405 // produce a carry into the higher bits.
1406 SDValue T1 = getInstr(Hexagon::V6_vlsrw, dl, ResTy, {LoVec(T0), S16}, DAG);
1407 // Swap low and high halves in Vt, and do the halfword multiplication
1408 // to get products Vs.uh[2i] * Vt.uh[2i+1] and Vs.uh[2i+1] * Vt.uh[2i].
1409 SDValue D0 = getInstr(Hexagon::V6_vdelta, dl, ResTy, {Vt, P}, DAG);
1410 SDValue T2 = getInstr(Hexagon::V6_vmpyuhv, dl, PairTy, {Vs, D0}, DAG);
1411 // T2 has mixed products of halfwords: Lo(Vt)*Hi(Vs) and Hi(Vt)*Lo(Vs).
1412 // These products are words, but cannot be added directly because the
1413 // sums could overflow. Add these products, by halfwords, where each sum
1414 // of a pair of halfwords gives a word.
1415 SDValue T3 = getInstr(Hexagon::V6_vadduhw, dl, PairTy,
1416 {LoVec(T2), HiVec(T2)}, DAG);
1417 // Add the high halfwords from the products of the low halfwords.
1418 SDValue T4 = DAG.getNode(ISD::ADD, dl, ResTy, {T1, LoVec(T3)});
1419 SDValue T5 = getInstr(Hexagon::V6_vlsrw, dl, ResTy, {T4, S16}, DAG);
1420 SDValue T6 = DAG.getNode(ISD::ADD, dl, ResTy, {HiVec(T0), HiVec(T3)});
1421 SDValue T7 = DAG.getNode(ISD::ADD, dl, ResTy, {T5, T6});
1426 HexagonTargetLowering::LowerHvxExtend(SDValue Op, SelectionDAG &DAG) const {
1427 // Sign- and zero-extends are legal.
1428 assert(Op.getOpcode() == ISD::ANY_EXTEND_VECTOR_INREG);
1429 return DAG.getNode(ISD::ZERO_EXTEND_VECTOR_INREG, SDLoc(Op), ty(Op),
1434 HexagonTargetLowering::LowerHvxShift(SDValue Op, SelectionDAG &DAG) const {
1435 if (SDValue S = getVectorShiftByInt(Op, DAG))
1441 HexagonTargetLowering::SplitHvxPairOp(SDValue Op, SelectionDAG &DAG) const {
1442 assert(!Op.isMachineOpcode());
1443 SmallVector<SDValue,2> OpsL, OpsH;
1444 const SDLoc &dl(Op);
1446 auto SplitVTNode = [&DAG,this] (const VTSDNode *N) {
1447 MVT Ty = typeSplit(N->getVT().getSimpleVT()).first;
1448 SDValue TV = DAG.getValueType(Ty);
1449 return std::make_pair(TV, TV);
1452 for (SDValue A : Op.getNode()->ops()) {
1453 VectorPair P = Subtarget.isHVXVectorType(ty(A), true)
1454 ? opSplit(A, dl, DAG)
1455 : std::make_pair(A, A);
1456 // Special case for type operand.
1457 if (Op.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1458 if (const auto *N = dyn_cast<const VTSDNode>(A.getNode()))
1461 OpsL.push_back(P.first);
1462 OpsH.push_back(P.second);
1466 MVT HalfTy = typeSplit(ResTy).first;
1467 SDValue L = DAG.getNode(Op.getOpcode(), dl, HalfTy, OpsL);
1468 SDValue H = DAG.getNode(Op.getOpcode(), dl, HalfTy, OpsH);
1469 SDValue S = DAG.getNode(ISD::CONCAT_VECTORS, dl, ResTy, L, H);
1474 HexagonTargetLowering::SplitHvxMemOp(SDValue Op, SelectionDAG &DAG) const {
1475 LSBaseSDNode *BN = cast<LSBaseSDNode>(Op.getNode());
1476 assert(BN->isUnindexed());
1477 MVT MemTy = BN->getMemoryVT().getSimpleVT();
1478 if (!isHvxPairTy(MemTy))
1481 const SDLoc &dl(Op);
1482 unsigned HwLen = Subtarget.getVectorLength();
1483 MVT SingleTy = typeSplit(MemTy).first;
1484 SDValue Chain = BN->getChain();
1485 SDValue Base0 = BN->getBasePtr();
1486 SDValue Base1 = DAG.getMemBasePlusOffset(Base0, HwLen, dl);
1488 MachineMemOperand *MOp0 = nullptr, *MOp1 = nullptr;
1489 if (MachineMemOperand *MMO = BN->getMemOperand()) {
1490 MachineFunction &MF = DAG.getMachineFunction();
1491 MOp0 = MF.getMachineMemOperand(MMO, 0, HwLen);
1492 MOp1 = MF.getMachineMemOperand(MMO, HwLen, HwLen);
1495 unsigned MemOpc = BN->getOpcode();
1498 if (MemOpc == ISD::LOAD) {
1499 SDValue Load0 = DAG.getLoad(SingleTy, dl, Chain, Base0, MOp0);
1500 SDValue Load1 = DAG.getLoad(SingleTy, dl, Chain, Base1, MOp1);
1501 NewOp = DAG.getMergeValues(
1502 { DAG.getNode(ISD::CONCAT_VECTORS, dl, MemTy, Load0, Load1),
1503 DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1504 Load0.getValue(1), Load1.getValue(1)) }, dl);
1506 assert(MemOpc == ISD::STORE);
1507 VectorPair Vals = opSplit(cast<StoreSDNode>(Op)->getValue(), dl, DAG);
1508 SDValue Store0 = DAG.getStore(Chain, dl, Vals.first, Base0, MOp0);
1509 SDValue Store1 = DAG.getStore(Chain, dl, Vals.second, Base1, MOp1);
1510 NewOp = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store0, Store1);
1517 HexagonTargetLowering::LowerHvxOperation(SDValue Op, SelectionDAG &DAG) const {
1518 unsigned Opc = Op.getOpcode();
1519 bool IsPairOp = isHvxPairTy(ty(Op)) ||
1520 llvm::any_of(Op.getNode()->ops(), [this] (SDValue V) {
1521 return isHvxPairTy(ty(V));
1530 return SplitHvxMemOp(Op, DAG);
1545 case ISD::SIGN_EXTEND_INREG:
1546 return SplitHvxPairOp(Op, DAG);
1553 case ISD::BUILD_VECTOR: return LowerHvxBuildVector(Op, DAG);
1554 case ISD::CONCAT_VECTORS: return LowerHvxConcatVectors(Op, DAG);
1555 case ISD::INSERT_SUBVECTOR: return LowerHvxInsertSubvector(Op, DAG);
1556 case ISD::INSERT_VECTOR_ELT: return LowerHvxInsertElement(Op, DAG);
1557 case ISD::EXTRACT_SUBVECTOR: return LowerHvxExtractSubvector(Op, DAG);
1558 case ISD::EXTRACT_VECTOR_ELT: return LowerHvxExtractElement(Op, DAG);
1560 case ISD::ANY_EXTEND: return LowerHvxAnyExt(Op, DAG);
1561 case ISD::SIGN_EXTEND: return LowerHvxSignExt(Op, DAG);
1562 case ISD::ZERO_EXTEND: return LowerHvxZeroExt(Op, DAG);
1563 case ISD::CTTZ: return LowerHvxCttz(Op, DAG);
1566 case ISD::SRL: return LowerHvxShift(Op, DAG);
1567 case ISD::MUL: return LowerHvxMul(Op, DAG);
1569 case ISD::MULHU: return LowerHvxMulh(Op, DAG);
1570 case ISD::ANY_EXTEND_VECTOR_INREG: return LowerHvxExtend(Op, DAG);
1572 case ISD::INTRINSIC_VOID: return Op;
1573 // Unaligned loads will be handled by the default lowering.
1574 case ISD::LOAD: return SDValue();
1579 llvm_unreachable("Unhandled HVX operation");
1583 HexagonTargetLowering::isHvxOperation(SDValue Op) const {
1584 // If the type of the result, or any operand type are HVX vector types,
1585 // this is an HVX operation.
1586 return Subtarget.isHVXVectorType(ty(Op), true) ||
1587 llvm::any_of(Op.getNode()->ops(),
1588 [this] (SDValue V) {
1589 return Subtarget.isHVXVectorType(ty(V), true);