1 //===-- MipsSEISelLowering.cpp - MipsSE DAG Lowering Interface --*- C++ -*-===//
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 // Subclass of MipsTargetLowering specialized for mips32/64.
12 //===----------------------------------------------------------------------===//
13 #include "MipsSEISelLowering.h"
14 #include "MipsRegisterInfo.h"
15 #include "MipsTargetMachine.h"
16 #include "llvm/CodeGen/MachineInstrBuilder.h"
17 #include "llvm/CodeGen/MachineRegisterInfo.h"
18 #include "llvm/IR/Intrinsics.h"
19 #include "llvm/Support/CommandLine.h"
20 #include "llvm/Target/TargetInstrInfo.h"
25 EnableMipsTailCalls("enable-mips-tail-calls", cl::Hidden,
26 cl::desc("MIPS: Enable tail calls."), cl::init(false));
28 static cl::opt<bool> NoDPLoadStore("mno-ldc1-sdc1", cl::init(false),
29 cl::desc("Expand double precision loads and "
30 "stores to their single precision "
33 MipsSETargetLowering::MipsSETargetLowering(MipsTargetMachine &TM)
34 : MipsTargetLowering(TM) {
35 // Set up the register classes
37 clearRegisterClasses();
39 addRegisterClass(MVT::i32, &Mips::GPR32RegClass);
42 addRegisterClass(MVT::i64, &Mips::GPR64RegClass);
44 if (Subtarget->hasDSP()) {
45 MVT::SimpleValueType VecTys[2] = {MVT::v2i16, MVT::v4i8};
47 for (unsigned i = 0; i < array_lengthof(VecTys); ++i) {
48 addRegisterClass(VecTys[i], &Mips::DSPRRegClass);
50 // Expand all builtin opcodes.
51 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
52 setOperationAction(Opc, VecTys[i], Expand);
54 setOperationAction(ISD::ADD, VecTys[i], Legal);
55 setOperationAction(ISD::SUB, VecTys[i], Legal);
56 setOperationAction(ISD::LOAD, VecTys[i], Legal);
57 setOperationAction(ISD::STORE, VecTys[i], Legal);
58 setOperationAction(ISD::BITCAST, VecTys[i], Legal);
61 // Expand all truncating stores and extending loads.
62 unsigned FirstVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
63 unsigned LastVT = (unsigned)MVT::LAST_VECTOR_VALUETYPE;
65 for (unsigned VT0 = FirstVT; VT0 <= LastVT; ++VT0) {
66 for (unsigned VT1 = FirstVT; VT1 <= LastVT; ++VT1)
67 setTruncStoreAction((MVT::SimpleValueType)VT0,
68 (MVT::SimpleValueType)VT1, Expand);
70 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT0, Expand);
71 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT0, Expand);
72 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT0, Expand);
75 setTargetDAGCombine(ISD::SHL);
76 setTargetDAGCombine(ISD::SRA);
77 setTargetDAGCombine(ISD::SRL);
78 setTargetDAGCombine(ISD::SETCC);
79 setTargetDAGCombine(ISD::VSELECT);
82 if (Subtarget->hasDSPR2())
83 setOperationAction(ISD::MUL, MVT::v2i16, Legal);
85 if (Subtarget->hasMSA()) {
86 addMSAIntType(MVT::v16i8, &Mips::MSA128BRegClass);
87 addMSAIntType(MVT::v8i16, &Mips::MSA128HRegClass);
88 addMSAIntType(MVT::v4i32, &Mips::MSA128WRegClass);
89 addMSAIntType(MVT::v2i64, &Mips::MSA128DRegClass);
90 addMSAFloatType(MVT::v8f16, &Mips::MSA128HRegClass);
91 addMSAFloatType(MVT::v4f32, &Mips::MSA128WRegClass);
92 addMSAFloatType(MVT::v2f64, &Mips::MSA128DRegClass);
94 setTargetDAGCombine(ISD::AND);
95 setTargetDAGCombine(ISD::SRA);
96 setTargetDAGCombine(ISD::VSELECT);
97 setTargetDAGCombine(ISD::XOR);
100 if (!Subtarget->mipsSEUsesSoftFloat()) {
101 addRegisterClass(MVT::f32, &Mips::FGR32RegClass);
103 // When dealing with single precision only, use libcalls
104 if (!Subtarget->isSingleFloat()) {
105 if (Subtarget->isFP64bit())
106 addRegisterClass(MVT::f64, &Mips::FGR64RegClass);
108 addRegisterClass(MVT::f64, &Mips::AFGR64RegClass);
112 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Custom);
113 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Custom);
114 setOperationAction(ISD::MULHS, MVT::i32, Custom);
115 setOperationAction(ISD::MULHU, MVT::i32, Custom);
118 setOperationAction(ISD::MULHS, MVT::i64, Custom);
119 setOperationAction(ISD::MULHU, MVT::i64, Custom);
120 setOperationAction(ISD::MUL, MVT::i64, Custom);
123 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
124 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
126 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
127 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
128 setOperationAction(ISD::SDIVREM, MVT::i64, Custom);
129 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
130 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
131 setOperationAction(ISD::LOAD, MVT::i32, Custom);
132 setOperationAction(ISD::STORE, MVT::i32, Custom);
134 setTargetDAGCombine(ISD::ADDE);
135 setTargetDAGCombine(ISD::SUBE);
136 setTargetDAGCombine(ISD::MUL);
138 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
139 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
140 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
143 setOperationAction(ISD::LOAD, MVT::f64, Custom);
144 setOperationAction(ISD::STORE, MVT::f64, Custom);
147 computeRegisterProperties();
150 const MipsTargetLowering *
151 llvm::createMipsSETargetLowering(MipsTargetMachine &TM) {
152 return new MipsSETargetLowering(TM);
155 // Enable MSA support for the given integer type and Register class.
156 void MipsSETargetLowering::
157 addMSAIntType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
158 addRegisterClass(Ty, RC);
160 // Expand all builtin opcodes.
161 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
162 setOperationAction(Opc, Ty, Expand);
164 setOperationAction(ISD::BITCAST, Ty, Legal);
165 setOperationAction(ISD::LOAD, Ty, Legal);
166 setOperationAction(ISD::STORE, Ty, Legal);
167 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Custom);
168 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
169 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
171 setOperationAction(ISD::ADD, Ty, Legal);
172 setOperationAction(ISD::AND, Ty, Legal);
173 setOperationAction(ISD::CTLZ, Ty, Legal);
174 setOperationAction(ISD::CTPOP, Ty, Legal);
175 setOperationAction(ISD::MUL, Ty, Legal);
176 setOperationAction(ISD::OR, Ty, Legal);
177 setOperationAction(ISD::SDIV, Ty, Legal);
178 setOperationAction(ISD::SHL, Ty, Legal);
179 setOperationAction(ISD::SRA, Ty, Legal);
180 setOperationAction(ISD::SRL, Ty, Legal);
181 setOperationAction(ISD::SUB, Ty, Legal);
182 setOperationAction(ISD::UDIV, Ty, Legal);
183 setOperationAction(ISD::VECTOR_SHUFFLE, Ty, Custom);
184 setOperationAction(ISD::VSELECT, Ty, Legal);
185 setOperationAction(ISD::XOR, Ty, Legal);
187 setOperationAction(ISD::SETCC, Ty, Legal);
188 setCondCodeAction(ISD::SETNE, Ty, Expand);
189 setCondCodeAction(ISD::SETGE, Ty, Expand);
190 setCondCodeAction(ISD::SETGT, Ty, Expand);
191 setCondCodeAction(ISD::SETUGE, Ty, Expand);
192 setCondCodeAction(ISD::SETUGT, Ty, Expand);
195 // Enable MSA support for the given floating-point type and Register class.
196 void MipsSETargetLowering::
197 addMSAFloatType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
198 addRegisterClass(Ty, RC);
200 // Expand all builtin opcodes.
201 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
202 setOperationAction(Opc, Ty, Expand);
204 setOperationAction(ISD::LOAD, Ty, Legal);
205 setOperationAction(ISD::STORE, Ty, Legal);
206 setOperationAction(ISD::BITCAST, Ty, Legal);
207 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Legal);
209 if (Ty != MVT::v8f16) {
210 setOperationAction(ISD::FABS, Ty, Legal);
211 setOperationAction(ISD::FADD, Ty, Legal);
212 setOperationAction(ISD::FDIV, Ty, Legal);
213 setOperationAction(ISD::FLOG2, Ty, Legal);
214 setOperationAction(ISD::FMUL, Ty, Legal);
215 setOperationAction(ISD::FRINT, Ty, Legal);
216 setOperationAction(ISD::FSQRT, Ty, Legal);
217 setOperationAction(ISD::FSUB, Ty, Legal);
218 setOperationAction(ISD::VSELECT, Ty, Legal);
220 setOperationAction(ISD::SETCC, Ty, Legal);
221 setCondCodeAction(ISD::SETOGE, Ty, Expand);
222 setCondCodeAction(ISD::SETOGT, Ty, Expand);
223 setCondCodeAction(ISD::SETUGE, Ty, Expand);
224 setCondCodeAction(ISD::SETUGT, Ty, Expand);
225 setCondCodeAction(ISD::SETGE, Ty, Expand);
226 setCondCodeAction(ISD::SETGT, Ty, Expand);
231 MipsSETargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
232 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
245 SDValue MipsSETargetLowering::LowerOperation(SDValue Op,
246 SelectionDAG &DAG) const {
247 switch(Op.getOpcode()) {
248 case ISD::LOAD: return lowerLOAD(Op, DAG);
249 case ISD::STORE: return lowerSTORE(Op, DAG);
250 case ISD::SMUL_LOHI: return lowerMulDiv(Op, MipsISD::Mult, true, true, DAG);
251 case ISD::UMUL_LOHI: return lowerMulDiv(Op, MipsISD::Multu, true, true, DAG);
252 case ISD::MULHS: return lowerMulDiv(Op, MipsISD::Mult, false, true, DAG);
253 case ISD::MULHU: return lowerMulDiv(Op, MipsISD::Multu, false, true, DAG);
254 case ISD::MUL: return lowerMulDiv(Op, MipsISD::Mult, true, false, DAG);
255 case ISD::SDIVREM: return lowerMulDiv(Op, MipsISD::DivRem, true, true, DAG);
256 case ISD::UDIVREM: return lowerMulDiv(Op, MipsISD::DivRemU, true, true,
258 case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG);
259 case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG);
260 case ISD::INTRINSIC_VOID: return lowerINTRINSIC_VOID(Op, DAG);
261 case ISD::EXTRACT_VECTOR_ELT: return lowerEXTRACT_VECTOR_ELT(Op, DAG);
262 case ISD::BUILD_VECTOR: return lowerBUILD_VECTOR(Op, DAG);
263 case ISD::VECTOR_SHUFFLE: return lowerVECTOR_SHUFFLE(Op, DAG);
266 return MipsTargetLowering::LowerOperation(Op, DAG);
270 // Transforms a subgraph in CurDAG if the following pattern is found:
271 // (addc multLo, Lo0), (adde multHi, Hi0),
273 // multHi/Lo: product of multiplication
274 // Lo0: initial value of Lo register
275 // Hi0: initial value of Hi register
276 // Return true if pattern matching was successful.
277 static bool selectMADD(SDNode *ADDENode, SelectionDAG *CurDAG) {
278 // ADDENode's second operand must be a flag output of an ADDC node in order
279 // for the matching to be successful.
280 SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
282 if (ADDCNode->getOpcode() != ISD::ADDC)
285 SDValue MultHi = ADDENode->getOperand(0);
286 SDValue MultLo = ADDCNode->getOperand(0);
287 SDNode *MultNode = MultHi.getNode();
288 unsigned MultOpc = MultHi.getOpcode();
290 // MultHi and MultLo must be generated by the same node,
291 if (MultLo.getNode() != MultNode)
294 // and it must be a multiplication.
295 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
298 // MultLo amd MultHi must be the first and second output of MultNode
300 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
303 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
304 // of the values of MultNode, in which case MultNode will be removed in later
306 // If there exist users other than ADDENode or ADDCNode, this function returns
307 // here, which will result in MultNode being mapped to a single MULT
308 // instruction node rather than a pair of MULT and MADD instructions being
310 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
315 // Initialize accumulator.
316 SDValue ACCIn = CurDAG->getNode(MipsISD::InsertLOHI, DL, MVT::Untyped,
317 ADDCNode->getOperand(1),
318 ADDENode->getOperand(1));
320 // create MipsMAdd(u) node
321 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
323 SDValue MAdd = CurDAG->getNode(MultOpc, DL, MVT::Untyped,
324 MultNode->getOperand(0),// Factor 0
325 MultNode->getOperand(1),// Factor 1
328 // replace uses of adde and addc here
329 if (!SDValue(ADDCNode, 0).use_empty()) {
330 SDValue LoIdx = CurDAG->getConstant(Mips::sub_lo, MVT::i32);
331 SDValue LoOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MAdd,
333 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), LoOut);
335 if (!SDValue(ADDENode, 0).use_empty()) {
336 SDValue HiIdx = CurDAG->getConstant(Mips::sub_hi, MVT::i32);
337 SDValue HiOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MAdd,
339 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), HiOut);
346 // Transforms a subgraph in CurDAG if the following pattern is found:
347 // (addc Lo0, multLo), (sube Hi0, multHi),
349 // multHi/Lo: product of multiplication
350 // Lo0: initial value of Lo register
351 // Hi0: initial value of Hi register
352 // Return true if pattern matching was successful.
353 static bool selectMSUB(SDNode *SUBENode, SelectionDAG *CurDAG) {
354 // SUBENode's second operand must be a flag output of an SUBC node in order
355 // for the matching to be successful.
356 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
358 if (SUBCNode->getOpcode() != ISD::SUBC)
361 SDValue MultHi = SUBENode->getOperand(1);
362 SDValue MultLo = SUBCNode->getOperand(1);
363 SDNode *MultNode = MultHi.getNode();
364 unsigned MultOpc = MultHi.getOpcode();
366 // MultHi and MultLo must be generated by the same node,
367 if (MultLo.getNode() != MultNode)
370 // and it must be a multiplication.
371 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
374 // MultLo amd MultHi must be the first and second output of MultNode
376 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
379 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
380 // of the values of MultNode, in which case MultNode will be removed in later
382 // If there exist users other than SUBENode or SUBCNode, this function returns
383 // here, which will result in MultNode being mapped to a single MULT
384 // instruction node rather than a pair of MULT and MSUB instructions being
386 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
391 // Initialize accumulator.
392 SDValue ACCIn = CurDAG->getNode(MipsISD::InsertLOHI, DL, MVT::Untyped,
393 SUBCNode->getOperand(0),
394 SUBENode->getOperand(0));
396 // create MipsSub(u) node
397 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
399 SDValue MSub = CurDAG->getNode(MultOpc, DL, MVT::Glue,
400 MultNode->getOperand(0),// Factor 0
401 MultNode->getOperand(1),// Factor 1
404 // replace uses of sube and subc here
405 if (!SDValue(SUBCNode, 0).use_empty()) {
406 SDValue LoIdx = CurDAG->getConstant(Mips::sub_lo, MVT::i32);
407 SDValue LoOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MSub,
409 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), LoOut);
411 if (!SDValue(SUBENode, 0).use_empty()) {
412 SDValue HiIdx = CurDAG->getConstant(Mips::sub_hi, MVT::i32);
413 SDValue HiOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MSub,
415 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), HiOut);
421 static SDValue performADDECombine(SDNode *N, SelectionDAG &DAG,
422 TargetLowering::DAGCombinerInfo &DCI,
423 const MipsSubtarget *Subtarget) {
424 if (DCI.isBeforeLegalize())
427 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
429 return SDValue(N, 0);
434 // Fold zero extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT
436 // Performs the following transformations:
437 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to zero extension if its
438 // sign/zero-extension is completely overwritten by the new one performed by
440 // - Removes redundant zero extensions performed by an ISD::AND.
441 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
442 TargetLowering::DAGCombinerInfo &DCI,
443 const MipsSubtarget *Subtarget) {
444 if (!Subtarget->hasMSA())
447 SDValue Op0 = N->getOperand(0);
448 SDValue Op1 = N->getOperand(1);
449 unsigned Op0Opcode = Op0->getOpcode();
451 // (and (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d)
452 // where $d + 1 == 2^n and n == 32
453 // or $d + 1 == 2^n and n <= 32 and ZExt
454 // -> (MipsVExtractZExt $a, $b, $c)
455 if (Op0Opcode == MipsISD::VEXTRACT_SEXT_ELT ||
456 Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT) {
457 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(Op1);
462 int32_t Log2IfPositive = (Mask->getAPIntValue() + 1).exactLogBase2();
464 if (Log2IfPositive <= 0)
465 return SDValue(); // Mask+1 is not a power of 2
467 SDValue Op0Op2 = Op0->getOperand(2);
468 EVT ExtendTy = cast<VTSDNode>(Op0Op2)->getVT();
469 unsigned ExtendTySize = ExtendTy.getSizeInBits();
470 unsigned Log2 = Log2IfPositive;
472 if ((Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT && Log2 >= ExtendTySize) ||
473 Log2 == ExtendTySize) {
474 SDValue Ops[] = { Op0->getOperand(0), Op0->getOperand(1), Op0Op2 };
475 DAG.MorphNodeTo(Op0.getNode(), MipsISD::VEXTRACT_ZEXT_ELT,
476 Op0->getVTList(), Ops, Op0->getNumOperands());
484 static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG,
485 TargetLowering::DAGCombinerInfo &DCI,
486 const MipsSubtarget *Subtarget) {
487 if (DCI.isBeforeLegalize())
490 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
492 return SDValue(N, 0);
497 static SDValue genConstMult(SDValue X, uint64_t C, SDLoc DL, EVT VT,
498 EVT ShiftTy, SelectionDAG &DAG) {
499 // Clear the upper (64 - VT.sizeInBits) bits.
500 C &= ((uint64_t)-1) >> (64 - VT.getSizeInBits());
504 return DAG.getConstant(0, VT);
510 // If c is power of 2, return (shl x, log2(c)).
511 if (isPowerOf2_64(C))
512 return DAG.getNode(ISD::SHL, DL, VT, X,
513 DAG.getConstant(Log2_64(C), ShiftTy));
515 unsigned Log2Ceil = Log2_64_Ceil(C);
516 uint64_t Floor = 1LL << Log2_64(C);
517 uint64_t Ceil = Log2Ceil == 64 ? 0LL : 1LL << Log2Ceil;
519 // If |c - floor_c| <= |c - ceil_c|,
520 // where floor_c = pow(2, floor(log2(c))) and ceil_c = pow(2, ceil(log2(c))),
521 // return (add constMult(x, floor_c), constMult(x, c - floor_c)).
522 if (C - Floor <= Ceil - C) {
523 SDValue Op0 = genConstMult(X, Floor, DL, VT, ShiftTy, DAG);
524 SDValue Op1 = genConstMult(X, C - Floor, DL, VT, ShiftTy, DAG);
525 return DAG.getNode(ISD::ADD, DL, VT, Op0, Op1);
528 // If |c - floor_c| > |c - ceil_c|,
529 // return (sub constMult(x, ceil_c), constMult(x, ceil_c - c)).
530 SDValue Op0 = genConstMult(X, Ceil, DL, VT, ShiftTy, DAG);
531 SDValue Op1 = genConstMult(X, Ceil - C, DL, VT, ShiftTy, DAG);
532 return DAG.getNode(ISD::SUB, DL, VT, Op0, Op1);
535 static SDValue performMULCombine(SDNode *N, SelectionDAG &DAG,
536 const TargetLowering::DAGCombinerInfo &DCI,
537 const MipsSETargetLowering *TL) {
538 EVT VT = N->getValueType(0);
540 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
542 return genConstMult(N->getOperand(0), C->getZExtValue(), SDLoc(N),
543 VT, TL->getScalarShiftAmountTy(VT), DAG);
545 return SDValue(N, 0);
548 static SDValue performDSPShiftCombine(unsigned Opc, SDNode *N, EVT Ty,
550 const MipsSubtarget *Subtarget) {
551 // See if this is a vector splat immediate node.
552 APInt SplatValue, SplatUndef;
553 unsigned SplatBitSize;
555 unsigned EltSize = Ty.getVectorElementType().getSizeInBits();
556 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
559 !BV->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
560 EltSize, !Subtarget->isLittle()) ||
561 (SplatBitSize != EltSize) ||
562 (SplatValue.getZExtValue() >= EltSize))
565 return DAG.getNode(Opc, SDLoc(N), Ty, N->getOperand(0),
566 DAG.getConstant(SplatValue.getZExtValue(), MVT::i32));
569 static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
570 TargetLowering::DAGCombinerInfo &DCI,
571 const MipsSubtarget *Subtarget) {
572 EVT Ty = N->getValueType(0);
574 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
577 return performDSPShiftCombine(MipsISD::SHLL_DSP, N, Ty, DAG, Subtarget);
580 // Fold sign-extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT for MSA and fold
581 // constant splats into MipsISD::SHRA_DSP for DSPr2.
583 // Performs the following transformations:
584 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to sign extension if its
585 // sign/zero-extension is completely overwritten by the new one performed by
586 // the ISD::SRA and ISD::SHL nodes.
587 // - Removes redundant sign extensions performed by an ISD::SRA and ISD::SHL
590 // See performDSPShiftCombine for more information about the transformation
592 static SDValue performSRACombine(SDNode *N, SelectionDAG &DAG,
593 TargetLowering::DAGCombinerInfo &DCI,
594 const MipsSubtarget *Subtarget) {
595 EVT Ty = N->getValueType(0);
597 if (Subtarget->hasMSA()) {
598 SDValue Op0 = N->getOperand(0);
599 SDValue Op1 = N->getOperand(1);
601 // (sra (shl (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d), imm:$d)
602 // where $d + sizeof($c) == 32
603 // or $d + sizeof($c) <= 32 and SExt
604 // -> (MipsVExtractSExt $a, $b, $c)
605 if (Op0->getOpcode() == ISD::SHL && Op1 == Op0->getOperand(1)) {
606 SDValue Op0Op0 = Op0->getOperand(0);
607 ConstantSDNode *ShAmount = dyn_cast<ConstantSDNode>(Op1);
612 if (Op0Op0->getOpcode() != MipsISD::VEXTRACT_SEXT_ELT &&
613 Op0Op0->getOpcode() != MipsISD::VEXTRACT_ZEXT_ELT)
616 EVT ExtendTy = cast<VTSDNode>(Op0Op0->getOperand(2))->getVT();
617 unsigned TotalBits = ShAmount->getZExtValue() + ExtendTy.getSizeInBits();
619 if (TotalBits == 32 ||
620 (Op0Op0->getOpcode() == MipsISD::VEXTRACT_SEXT_ELT &&
622 SDValue Ops[] = { Op0Op0->getOperand(0), Op0Op0->getOperand(1),
623 Op0Op0->getOperand(2) };
624 DAG.MorphNodeTo(Op0Op0.getNode(), MipsISD::VEXTRACT_SEXT_ELT,
625 Op0Op0->getVTList(), Ops, Op0Op0->getNumOperands());
631 if ((Ty != MVT::v2i16) && ((Ty != MVT::v4i8) || !Subtarget->hasDSPR2()))
634 return performDSPShiftCombine(MipsISD::SHRA_DSP, N, Ty, DAG, Subtarget);
638 static SDValue performSRLCombine(SDNode *N, SelectionDAG &DAG,
639 TargetLowering::DAGCombinerInfo &DCI,
640 const MipsSubtarget *Subtarget) {
641 EVT Ty = N->getValueType(0);
643 if (((Ty != MVT::v2i16) || !Subtarget->hasDSPR2()) && (Ty != MVT::v4i8))
646 return performDSPShiftCombine(MipsISD::SHRL_DSP, N, Ty, DAG, Subtarget);
649 static bool isLegalDSPCondCode(EVT Ty, ISD::CondCode CC) {
650 bool IsV216 = (Ty == MVT::v2i16);
654 case ISD::SETNE: return true;
658 case ISD::SETGE: return IsV216;
662 case ISD::SETUGE: return !IsV216;
663 default: return false;
667 static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG) {
668 EVT Ty = N->getValueType(0);
670 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
673 if (!isLegalDSPCondCode(Ty, cast<CondCodeSDNode>(N->getOperand(2))->get()))
676 return DAG.getNode(MipsISD::SETCC_DSP, SDLoc(N), Ty, N->getOperand(0),
677 N->getOperand(1), N->getOperand(2));
680 static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG) {
681 EVT Ty = N->getValueType(0);
683 if (Ty.is128BitVector() && Ty.isInteger()) {
684 // Try the following combines:
685 // (vselect (setcc $a, $b, SETLT), $b, $a)) -> (vsmax $a, $b)
686 // (vselect (setcc $a, $b, SETLE), $b, $a)) -> (vsmax $a, $b)
687 // (vselect (setcc $a, $b, SETLT), $a, $b)) -> (vsmin $a, $b)
688 // (vselect (setcc $a, $b, SETLE), $a, $b)) -> (vsmin $a, $b)
689 // (vselect (setcc $a, $b, SETULT), $b, $a)) -> (vumax $a, $b)
690 // (vselect (setcc $a, $b, SETULE), $b, $a)) -> (vumax $a, $b)
691 // (vselect (setcc $a, $b, SETULT), $a, $b)) -> (vumin $a, $b)
692 // (vselect (setcc $a, $b, SETULE), $a, $b)) -> (vumin $a, $b)
693 // SETGT/SETGE/SETUGT/SETUGE variants of these will show up initially but
694 // will be expanded to equivalent SETLT/SETLE/SETULT/SETULE versions by the
696 SDValue Op0 = N->getOperand(0);
698 if (Op0->getOpcode() != ISD::SETCC)
701 ISD::CondCode CondCode = cast<CondCodeSDNode>(Op0->getOperand(2))->get();
704 if (CondCode == ISD::SETLT || CondCode == ISD::SETLE)
706 else if (CondCode == ISD::SETULT || CondCode == ISD::SETULE)
711 SDValue Op1 = N->getOperand(1);
712 SDValue Op2 = N->getOperand(2);
713 SDValue Op0Op0 = Op0->getOperand(0);
714 SDValue Op0Op1 = Op0->getOperand(1);
716 if (Op1 == Op0Op0 && Op2 == Op0Op1)
717 return DAG.getNode(Signed ? MipsISD::VSMIN : MipsISD::VUMIN, SDLoc(N),
719 else if (Op1 == Op0Op1 && Op2 == Op0Op0)
720 return DAG.getNode(Signed ? MipsISD::VSMAX : MipsISD::VUMAX, SDLoc(N),
722 } else if ((Ty == MVT::v2i16) || (Ty == MVT::v4i8)) {
723 SDValue SetCC = N->getOperand(0);
725 if (SetCC.getOpcode() != MipsISD::SETCC_DSP)
728 return DAG.getNode(MipsISD::SELECT_CC_DSP, SDLoc(N), Ty,
729 SetCC.getOperand(0), SetCC.getOperand(1),
730 N->getOperand(1), N->getOperand(2), SetCC.getOperand(2));
736 static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
737 const MipsSubtarget *Subtarget) {
738 EVT Ty = N->getValueType(0);
740 if (Subtarget->hasMSA() && Ty.is128BitVector() && Ty.isInteger()) {
741 // Try the following combines:
742 // (xor (or $a, $b), (build_vector allones))
743 // (xor (or $a, $b), (bitcast (build_vector allones)))
744 SDValue Op0 = N->getOperand(0);
745 SDValue Op1 = N->getOperand(1);
748 if (ISD::isBuildVectorAllOnes(Op0.getNode()))
750 else if (ISD::isBuildVectorAllOnes(Op1.getNode()))
755 if (NotOp->getOpcode() == ISD::OR)
756 return DAG.getNode(MipsISD::VNOR, SDLoc(N), Ty, NotOp->getOperand(0),
757 NotOp->getOperand(1));
764 MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
765 SelectionDAG &DAG = DCI.DAG;
768 switch (N->getOpcode()) {
770 return performADDECombine(N, DAG, DCI, Subtarget);
772 Val = performANDCombine(N, DAG, DCI, Subtarget);
775 return performSUBECombine(N, DAG, DCI, Subtarget);
777 return performMULCombine(N, DAG, DCI, this);
779 return performSHLCombine(N, DAG, DCI, Subtarget);
781 return performSRACombine(N, DAG, DCI, Subtarget);
783 return performSRLCombine(N, DAG, DCI, Subtarget);
785 return performVSELECTCombine(N, DAG);
787 Val = performXORCombine(N, DAG, Subtarget);
790 Val = performSETCCCombine(N, DAG);
797 return MipsTargetLowering::PerformDAGCombine(N, DCI);
801 MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
802 MachineBasicBlock *BB) const {
803 switch (MI->getOpcode()) {
805 return MipsTargetLowering::EmitInstrWithCustomInserter(MI, BB);
806 case Mips::BPOSGE32_PSEUDO:
807 return emitBPOSGE32(MI, BB);
808 case Mips::SNZ_B_PSEUDO:
809 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_B);
810 case Mips::SNZ_H_PSEUDO:
811 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_H);
812 case Mips::SNZ_W_PSEUDO:
813 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_W);
814 case Mips::SNZ_D_PSEUDO:
815 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_D);
816 case Mips::SNZ_V_PSEUDO:
817 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_V);
818 case Mips::SZ_B_PSEUDO:
819 return emitMSACBranchPseudo(MI, BB, Mips::BZ_B);
820 case Mips::SZ_H_PSEUDO:
821 return emitMSACBranchPseudo(MI, BB, Mips::BZ_H);
822 case Mips::SZ_W_PSEUDO:
823 return emitMSACBranchPseudo(MI, BB, Mips::BZ_W);
824 case Mips::SZ_D_PSEUDO:
825 return emitMSACBranchPseudo(MI, BB, Mips::BZ_D);
826 case Mips::SZ_V_PSEUDO:
827 return emitMSACBranchPseudo(MI, BB, Mips::BZ_V);
831 bool MipsSETargetLowering::
832 isEligibleForTailCallOptimization(const MipsCC &MipsCCInfo,
833 unsigned NextStackOffset,
834 const MipsFunctionInfo& FI) const {
835 if (!EnableMipsTailCalls)
838 // Return false if either the callee or caller has a byval argument.
839 if (MipsCCInfo.hasByValArg() || FI.hasByvalArg())
842 // Return true if the callee's argument area is no larger than the
844 return NextStackOffset <= FI.getIncomingArgSize();
847 void MipsSETargetLowering::
848 getOpndList(SmallVectorImpl<SDValue> &Ops,
849 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
850 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
851 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const {
852 // T9 should contain the address of the callee function if
853 // -reloction-model=pic or it is an indirect call.
854 if (IsPICCall || !GlobalOrExternal) {
855 unsigned T9Reg = IsN64 ? Mips::T9_64 : Mips::T9;
856 RegsToPass.push_front(std::make_pair(T9Reg, Callee));
858 Ops.push_back(Callee);
860 MipsTargetLowering::getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal,
861 InternalLinkage, CLI, Callee, Chain);
864 SDValue MipsSETargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
865 LoadSDNode &Nd = *cast<LoadSDNode>(Op);
867 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
868 return MipsTargetLowering::lowerLOAD(Op, DAG);
870 // Replace a double precision load with two i32 loads and a buildpair64.
872 SDValue Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
873 EVT PtrVT = Ptr.getValueType();
875 // i32 load from lower address.
876 SDValue Lo = DAG.getLoad(MVT::i32, DL, Chain, Ptr,
877 MachinePointerInfo(), Nd.isVolatile(),
878 Nd.isNonTemporal(), Nd.isInvariant(),
881 // i32 load from higher address.
882 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, PtrVT));
883 SDValue Hi = DAG.getLoad(MVT::i32, DL, Lo.getValue(1), Ptr,
884 MachinePointerInfo(), Nd.isVolatile(),
885 Nd.isNonTemporal(), Nd.isInvariant(),
886 std::min(Nd.getAlignment(), 4U));
888 if (!Subtarget->isLittle())
891 SDValue BP = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
892 SDValue Ops[2] = {BP, Hi.getValue(1)};
893 return DAG.getMergeValues(Ops, 2, DL);
896 SDValue MipsSETargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
897 StoreSDNode &Nd = *cast<StoreSDNode>(Op);
899 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
900 return MipsTargetLowering::lowerSTORE(Op, DAG);
902 // Replace a double precision store with two extractelement64s and i32 stores.
904 SDValue Val = Nd.getValue(), Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
905 EVT PtrVT = Ptr.getValueType();
906 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
907 Val, DAG.getConstant(0, MVT::i32));
908 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
909 Val, DAG.getConstant(1, MVT::i32));
911 if (!Subtarget->isLittle())
914 // i32 store to lower address.
915 Chain = DAG.getStore(Chain, DL, Lo, Ptr, MachinePointerInfo(),
916 Nd.isVolatile(), Nd.isNonTemporal(), Nd.getAlignment(),
919 // i32 store to higher address.
920 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, PtrVT));
921 return DAG.getStore(Chain, DL, Hi, Ptr, MachinePointerInfo(),
922 Nd.isVolatile(), Nd.isNonTemporal(),
923 std::min(Nd.getAlignment(), 4U), Nd.getTBAAInfo());
926 SDValue MipsSETargetLowering::lowerMulDiv(SDValue Op, unsigned NewOpc,
927 bool HasLo, bool HasHi,
928 SelectionDAG &DAG) const {
929 EVT Ty = Op.getOperand(0).getValueType();
931 SDValue Mult = DAG.getNode(NewOpc, DL, MVT::Untyped,
932 Op.getOperand(0), Op.getOperand(1));
936 Lo = DAG.getNode(MipsISD::ExtractLOHI, DL, Ty, Mult,
937 DAG.getConstant(Mips::sub_lo, MVT::i32));
939 Hi = DAG.getNode(MipsISD::ExtractLOHI, DL, Ty, Mult,
940 DAG.getConstant(Mips::sub_hi, MVT::i32));
942 if (!HasLo || !HasHi)
943 return HasLo ? Lo : Hi;
945 SDValue Vals[] = { Lo, Hi };
946 return DAG.getMergeValues(Vals, 2, DL);
950 static SDValue initAccumulator(SDValue In, SDLoc DL, SelectionDAG &DAG) {
951 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
952 DAG.getConstant(0, MVT::i32));
953 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
954 DAG.getConstant(1, MVT::i32));
955 return DAG.getNode(MipsISD::InsertLOHI, DL, MVT::Untyped, InLo, InHi);
958 static SDValue extractLOHI(SDValue Op, SDLoc DL, SelectionDAG &DAG) {
959 SDValue Lo = DAG.getNode(MipsISD::ExtractLOHI, DL, MVT::i32, Op,
960 DAG.getConstant(Mips::sub_lo, MVT::i32));
961 SDValue Hi = DAG.getNode(MipsISD::ExtractLOHI, DL, MVT::i32, Op,
962 DAG.getConstant(Mips::sub_hi, MVT::i32));
963 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi);
966 // This function expands mips intrinsic nodes which have 64-bit input operands
969 // out64 = intrinsic-node in64
971 // lo = copy (extract-element (in64, 0))
972 // hi = copy (extract-element (in64, 1))
973 // mips-specific-node
976 // out64 = merge-values (v0, v1)
978 static SDValue lowerDSPIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
980 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
981 SmallVector<SDValue, 3> Ops;
984 // See if Op has a chain input.
986 Ops.push_back(Op->getOperand(OpNo++));
988 // The next operand is the intrinsic opcode.
989 assert(Op->getOperand(OpNo).getOpcode() == ISD::TargetConstant);
991 // See if the next operand has type i64.
992 SDValue Opnd = Op->getOperand(++OpNo), In64;
994 if (Opnd.getValueType() == MVT::i64)
995 In64 = initAccumulator(Opnd, DL, DAG);
999 // Push the remaining operands.
1000 for (++OpNo ; OpNo < Op->getNumOperands(); ++OpNo)
1001 Ops.push_back(Op->getOperand(OpNo));
1003 // Add In64 to the end of the list.
1005 Ops.push_back(In64);
1008 SmallVector<EVT, 2> ResTys;
1010 for (SDNode::value_iterator I = Op->value_begin(), E = Op->value_end();
1012 ResTys.push_back((*I == MVT::i64) ? MVT::Untyped : *I);
1015 SDValue Val = DAG.getNode(Opc, DL, ResTys, &Ops[0], Ops.size());
1016 SDValue Out = (ResTys[0] == MVT::Untyped) ? extractLOHI(Val, DL, DAG) : Val;
1021 assert(Val->getValueType(1) == MVT::Other);
1022 SDValue Vals[] = { Out, SDValue(Val.getNode(), 1) };
1023 return DAG.getMergeValues(Vals, 2, DL);
1026 static SDValue lowerMSABinaryIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1028 SDValue LHS = Op->getOperand(1);
1029 SDValue RHS = Op->getOperand(2);
1030 EVT ResTy = Op->getValueType(0);
1032 SDValue Result = DAG.getNode(Opc, DL, ResTy, LHS, RHS);
1037 static SDValue lowerMSABinaryImmIntr(SDValue Op, SelectionDAG &DAG,
1038 unsigned Opc, SDValue RHS) {
1039 SDValue LHS = Op->getOperand(1);
1040 EVT ResTy = Op->getValueType(0);
1042 return DAG.getNode(Opc, SDLoc(Op), ResTy, LHS, RHS);
1045 static SDValue lowerMSABranchIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1047 SDValue Value = Op->getOperand(1);
1048 EVT ResTy = Op->getValueType(0);
1050 SDValue Result = DAG.getNode(Opc, DL, ResTy, Value);
1055 // Lower an MSA copy intrinsic into the specified SelectionDAG node
1056 static SDValue lowerMSACopyIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1058 SDValue Vec = Op->getOperand(1);
1059 SDValue Idx = Op->getOperand(2);
1060 EVT ResTy = Op->getValueType(0);
1061 EVT EltTy = Vec->getValueType(0).getVectorElementType();
1063 SDValue Result = DAG.getNode(Opc, DL, ResTy, Vec, Idx,
1064 DAG.getValueType(EltTy));
1069 // Lower an MSA insert intrinsic into the specified SelectionDAG node
1070 static SDValue lowerMSAInsertIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1072 SDValue Op0 = Op->getOperand(1);
1073 SDValue Op1 = Op->getOperand(2);
1074 SDValue Op2 = Op->getOperand(3);
1075 EVT ResTy = Op->getValueType(0);
1077 SDValue Result = DAG.getNode(Opc, DL, ResTy, Op0, Op2, Op1);
1082 static SDValue lowerMSASplatImm(SDValue Op, SDValue ImmOp, SelectionDAG &DAG) {
1083 EVT ResTy = Op->getValueType(0);
1084 EVT ViaVecTy = ResTy;
1085 SmallVector<SDValue, 16> Ops;
1089 if (ViaVecTy == MVT::v2i64) {
1090 ImmHiOp = DAG.getNode(ISD::SRA, DL, MVT::i32, ImmOp,
1091 DAG.getConstant(31, MVT::i32));
1092 for (unsigned i = 0; i < ViaVecTy.getVectorNumElements(); ++i) {
1093 Ops.push_back(ImmHiOp);
1094 Ops.push_back(ImmOp);
1096 ViaVecTy = MVT::v4i32;
1098 for (unsigned i = 0; i < ResTy.getVectorNumElements(); ++i)
1099 Ops.push_back(ImmOp);
1102 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy, &Ops[0],
1105 if (ResTy != ViaVecTy)
1106 Result = DAG.getNode(ISD::BITCAST, DL, ResTy, Result);
1112 lowerMSASplatImm(SDValue Op, unsigned ImmOp, SelectionDAG &DAG) {
1113 return lowerMSASplatImm(Op, Op->getOperand(ImmOp), DAG);
1116 static SDValue lowerMSAUnaryIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1118 SDValue Value = Op->getOperand(1);
1119 EVT ResTy = Op->getValueType(0);
1121 SDValue Result = DAG.getNode(Opc, DL, ResTy, Value);
1126 SDValue MipsSETargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
1127 SelectionDAG &DAG) const {
1128 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) {
1131 case Intrinsic::mips_shilo:
1132 return lowerDSPIntr(Op, DAG, MipsISD::SHILO);
1133 case Intrinsic::mips_dpau_h_qbl:
1134 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL);
1135 case Intrinsic::mips_dpau_h_qbr:
1136 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR);
1137 case Intrinsic::mips_dpsu_h_qbl:
1138 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL);
1139 case Intrinsic::mips_dpsu_h_qbr:
1140 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR);
1141 case Intrinsic::mips_dpa_w_ph:
1142 return lowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH);
1143 case Intrinsic::mips_dps_w_ph:
1144 return lowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH);
1145 case Intrinsic::mips_dpax_w_ph:
1146 return lowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH);
1147 case Intrinsic::mips_dpsx_w_ph:
1148 return lowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH);
1149 case Intrinsic::mips_mulsa_w_ph:
1150 return lowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH);
1151 case Intrinsic::mips_mult:
1152 return lowerDSPIntr(Op, DAG, MipsISD::Mult);
1153 case Intrinsic::mips_multu:
1154 return lowerDSPIntr(Op, DAG, MipsISD::Multu);
1155 case Intrinsic::mips_madd:
1156 return lowerDSPIntr(Op, DAG, MipsISD::MAdd);
1157 case Intrinsic::mips_maddu:
1158 return lowerDSPIntr(Op, DAG, MipsISD::MAddu);
1159 case Intrinsic::mips_msub:
1160 return lowerDSPIntr(Op, DAG, MipsISD::MSub);
1161 case Intrinsic::mips_msubu:
1162 return lowerDSPIntr(Op, DAG, MipsISD::MSubu);
1163 case Intrinsic::mips_addv_b:
1164 case Intrinsic::mips_addv_h:
1165 case Intrinsic::mips_addv_w:
1166 case Intrinsic::mips_addv_d:
1167 return lowerMSABinaryIntr(Op, DAG, ISD::ADD);
1168 case Intrinsic::mips_addvi_b:
1169 case Intrinsic::mips_addvi_h:
1170 case Intrinsic::mips_addvi_w:
1171 case Intrinsic::mips_addvi_d:
1172 return lowerMSABinaryImmIntr(Op, DAG, ISD::ADD,
1173 lowerMSASplatImm(Op, 2, DAG));
1174 case Intrinsic::mips_and_v:
1175 return lowerMSABinaryIntr(Op, DAG, ISD::AND);
1176 case Intrinsic::mips_andi_b:
1177 return lowerMSABinaryImmIntr(Op, DAG, ISD::AND,
1178 lowerMSASplatImm(Op, 2, DAG));
1179 case Intrinsic::mips_bnz_b:
1180 case Intrinsic::mips_bnz_h:
1181 case Intrinsic::mips_bnz_w:
1182 case Intrinsic::mips_bnz_d:
1183 return lowerMSABranchIntr(Op, DAG, MipsISD::VALL_NONZERO);
1184 case Intrinsic::mips_bnz_v:
1185 return lowerMSABranchIntr(Op, DAG, MipsISD::VANY_NONZERO);
1186 case Intrinsic::mips_bsel_v:
1187 return DAG.getNode(ISD::VSELECT, SDLoc(Op), Op->getValueType(0),
1188 Op->getOperand(1), Op->getOperand(2),
1190 case Intrinsic::mips_bseli_b:
1191 return DAG.getNode(ISD::VSELECT, SDLoc(Op), Op->getValueType(0),
1192 Op->getOperand(1), Op->getOperand(2),
1193 lowerMSASplatImm(Op, 3, DAG));
1194 case Intrinsic::mips_bz_b:
1195 case Intrinsic::mips_bz_h:
1196 case Intrinsic::mips_bz_w:
1197 case Intrinsic::mips_bz_d:
1198 return lowerMSABranchIntr(Op, DAG, MipsISD::VALL_ZERO);
1199 case Intrinsic::mips_bz_v:
1200 return lowerMSABranchIntr(Op, DAG, MipsISD::VANY_ZERO);
1201 case Intrinsic::mips_ceq_b:
1202 case Intrinsic::mips_ceq_h:
1203 case Intrinsic::mips_ceq_w:
1204 case Intrinsic::mips_ceq_d:
1205 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1206 Op->getOperand(2), ISD::SETEQ);
1207 case Intrinsic::mips_ceqi_b:
1208 case Intrinsic::mips_ceqi_h:
1209 case Intrinsic::mips_ceqi_w:
1210 case Intrinsic::mips_ceqi_d:
1211 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1212 lowerMSASplatImm(Op, 2, DAG), ISD::SETEQ);
1213 case Intrinsic::mips_cle_s_b:
1214 case Intrinsic::mips_cle_s_h:
1215 case Intrinsic::mips_cle_s_w:
1216 case Intrinsic::mips_cle_s_d:
1217 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1218 Op->getOperand(2), ISD::SETLE);
1219 case Intrinsic::mips_clei_s_b:
1220 case Intrinsic::mips_clei_s_h:
1221 case Intrinsic::mips_clei_s_w:
1222 case Intrinsic::mips_clei_s_d:
1223 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1224 lowerMSASplatImm(Op, 2, DAG), ISD::SETLE);
1225 case Intrinsic::mips_cle_u_b:
1226 case Intrinsic::mips_cle_u_h:
1227 case Intrinsic::mips_cle_u_w:
1228 case Intrinsic::mips_cle_u_d:
1229 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1230 Op->getOperand(2), ISD::SETULE);
1231 case Intrinsic::mips_clei_u_b:
1232 case Intrinsic::mips_clei_u_h:
1233 case Intrinsic::mips_clei_u_w:
1234 case Intrinsic::mips_clei_u_d:
1235 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1236 lowerMSASplatImm(Op, 2, DAG), ISD::SETULE);
1237 case Intrinsic::mips_clt_s_b:
1238 case Intrinsic::mips_clt_s_h:
1239 case Intrinsic::mips_clt_s_w:
1240 case Intrinsic::mips_clt_s_d:
1241 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1242 Op->getOperand(2), ISD::SETLT);
1243 case Intrinsic::mips_clti_s_b:
1244 case Intrinsic::mips_clti_s_h:
1245 case Intrinsic::mips_clti_s_w:
1246 case Intrinsic::mips_clti_s_d:
1247 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1248 lowerMSASplatImm(Op, 2, DAG), ISD::SETLT);
1249 case Intrinsic::mips_clt_u_b:
1250 case Intrinsic::mips_clt_u_h:
1251 case Intrinsic::mips_clt_u_w:
1252 case Intrinsic::mips_clt_u_d:
1253 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1254 Op->getOperand(2), ISD::SETULT);
1255 case Intrinsic::mips_clti_u_b:
1256 case Intrinsic::mips_clti_u_h:
1257 case Intrinsic::mips_clti_u_w:
1258 case Intrinsic::mips_clti_u_d:
1259 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1260 lowerMSASplatImm(Op, 2, DAG), ISD::SETULT);
1261 case Intrinsic::mips_copy_s_b:
1262 case Intrinsic::mips_copy_s_h:
1263 case Intrinsic::mips_copy_s_w:
1264 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
1265 case Intrinsic::mips_copy_u_b:
1266 case Intrinsic::mips_copy_u_h:
1267 case Intrinsic::mips_copy_u_w:
1268 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
1269 case Intrinsic::mips_div_s_b:
1270 case Intrinsic::mips_div_s_h:
1271 case Intrinsic::mips_div_s_w:
1272 case Intrinsic::mips_div_s_d:
1273 return lowerMSABinaryIntr(Op, DAG, ISD::SDIV);
1274 case Intrinsic::mips_div_u_b:
1275 case Intrinsic::mips_div_u_h:
1276 case Intrinsic::mips_div_u_w:
1277 case Intrinsic::mips_div_u_d:
1278 return lowerMSABinaryIntr(Op, DAG, ISD::UDIV);
1279 case Intrinsic::mips_fadd_w:
1280 case Intrinsic::mips_fadd_d:
1281 return lowerMSABinaryIntr(Op, DAG, ISD::FADD);
1282 // Don't lower mips_fcaf_[wd] since LLVM folds SETFALSE condcodes away
1283 case Intrinsic::mips_fceq_w:
1284 case Intrinsic::mips_fceq_d:
1285 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1286 Op->getOperand(2), ISD::SETOEQ);
1287 case Intrinsic::mips_fcle_w:
1288 case Intrinsic::mips_fcle_d:
1289 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1290 Op->getOperand(2), ISD::SETOLE);
1291 case Intrinsic::mips_fclt_w:
1292 case Intrinsic::mips_fclt_d:
1293 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1294 Op->getOperand(2), ISD::SETOLT);
1295 case Intrinsic::mips_fcne_w:
1296 case Intrinsic::mips_fcne_d:
1297 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1298 Op->getOperand(2), ISD::SETONE);
1299 case Intrinsic::mips_fcor_w:
1300 case Intrinsic::mips_fcor_d:
1301 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1302 Op->getOperand(2), ISD::SETO);
1303 case Intrinsic::mips_fcueq_w:
1304 case Intrinsic::mips_fcueq_d:
1305 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1306 Op->getOperand(2), ISD::SETUEQ);
1307 case Intrinsic::mips_fcule_w:
1308 case Intrinsic::mips_fcule_d:
1309 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1310 Op->getOperand(2), ISD::SETULE);
1311 case Intrinsic::mips_fcult_w:
1312 case Intrinsic::mips_fcult_d:
1313 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1314 Op->getOperand(2), ISD::SETULT);
1315 case Intrinsic::mips_fcun_w:
1316 case Intrinsic::mips_fcun_d:
1317 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1318 Op->getOperand(2), ISD::SETUO);
1319 case Intrinsic::mips_fcune_w:
1320 case Intrinsic::mips_fcune_d:
1321 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1322 Op->getOperand(2), ISD::SETUNE);
1323 case Intrinsic::mips_fdiv_w:
1324 case Intrinsic::mips_fdiv_d:
1325 return lowerMSABinaryIntr(Op, DAG, ISD::FDIV);
1326 case Intrinsic::mips_fill_b:
1327 case Intrinsic::mips_fill_h:
1328 case Intrinsic::mips_fill_w: {
1329 SmallVector<SDValue, 16> Ops;
1330 EVT ResTy = Op->getValueType(0);
1332 for (unsigned i = 0; i < ResTy.getVectorNumElements(); ++i)
1333 Ops.push_back(Op->getOperand(1));
1335 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), ResTy, &Ops[0],
1338 case Intrinsic::mips_flog2_w:
1339 case Intrinsic::mips_flog2_d:
1340 return lowerMSAUnaryIntr(Op, DAG, ISD::FLOG2);
1341 case Intrinsic::mips_fmul_w:
1342 case Intrinsic::mips_fmul_d:
1343 return lowerMSABinaryIntr(Op, DAG, ISD::FMUL);
1344 case Intrinsic::mips_frint_w:
1345 case Intrinsic::mips_frint_d:
1346 return lowerMSAUnaryIntr(Op, DAG, ISD::FRINT);
1347 case Intrinsic::mips_fsqrt_w:
1348 case Intrinsic::mips_fsqrt_d:
1349 return lowerMSAUnaryIntr(Op, DAG, ISD::FSQRT);
1350 case Intrinsic::mips_fsub_w:
1351 case Intrinsic::mips_fsub_d:
1352 return lowerMSABinaryIntr(Op, DAG, ISD::FSUB);
1353 case Intrinsic::mips_ilvev_b:
1354 case Intrinsic::mips_ilvev_h:
1355 case Intrinsic::mips_ilvev_w:
1356 case Intrinsic::mips_ilvev_d:
1357 return DAG.getNode(MipsISD::ILVEV, SDLoc(Op), Op->getValueType(0),
1358 Op->getOperand(1), Op->getOperand(2));
1359 case Intrinsic::mips_ilvl_b:
1360 case Intrinsic::mips_ilvl_h:
1361 case Intrinsic::mips_ilvl_w:
1362 case Intrinsic::mips_ilvl_d:
1363 return DAG.getNode(MipsISD::ILVL, SDLoc(Op), Op->getValueType(0),
1364 Op->getOperand(1), Op->getOperand(2));
1365 case Intrinsic::mips_ilvod_b:
1366 case Intrinsic::mips_ilvod_h:
1367 case Intrinsic::mips_ilvod_w:
1368 case Intrinsic::mips_ilvod_d:
1369 return DAG.getNode(MipsISD::ILVOD, SDLoc(Op), Op->getValueType(0),
1370 Op->getOperand(1), Op->getOperand(2));
1371 case Intrinsic::mips_ilvr_b:
1372 case Intrinsic::mips_ilvr_h:
1373 case Intrinsic::mips_ilvr_w:
1374 case Intrinsic::mips_ilvr_d:
1375 return DAG.getNode(MipsISD::ILVR, SDLoc(Op), Op->getValueType(0),
1376 Op->getOperand(1), Op->getOperand(2));
1377 case Intrinsic::mips_insert_b:
1378 case Intrinsic::mips_insert_h:
1379 case Intrinsic::mips_insert_w:
1380 return lowerMSAInsertIntr(Op, DAG, ISD::INSERT_VECTOR_ELT);
1381 case Intrinsic::mips_ldi_b:
1382 case Intrinsic::mips_ldi_h:
1383 case Intrinsic::mips_ldi_w:
1384 case Intrinsic::mips_ldi_d:
1385 return lowerMSASplatImm(Op, 1, DAG);
1386 case Intrinsic::mips_max_s_b:
1387 case Intrinsic::mips_max_s_h:
1388 case Intrinsic::mips_max_s_w:
1389 case Intrinsic::mips_max_s_d:
1390 return lowerMSABinaryIntr(Op, DAG, MipsISD::VSMAX);
1391 case Intrinsic::mips_max_u_b:
1392 case Intrinsic::mips_max_u_h:
1393 case Intrinsic::mips_max_u_w:
1394 case Intrinsic::mips_max_u_d:
1395 return lowerMSABinaryIntr(Op, DAG, MipsISD::VUMAX);
1396 case Intrinsic::mips_maxi_s_b:
1397 case Intrinsic::mips_maxi_s_h:
1398 case Intrinsic::mips_maxi_s_w:
1399 case Intrinsic::mips_maxi_s_d:
1400 return lowerMSABinaryImmIntr(Op, DAG, MipsISD::VSMAX,
1401 lowerMSASplatImm(Op, 2, DAG));
1402 case Intrinsic::mips_maxi_u_b:
1403 case Intrinsic::mips_maxi_u_h:
1404 case Intrinsic::mips_maxi_u_w:
1405 case Intrinsic::mips_maxi_u_d:
1406 return lowerMSABinaryImmIntr(Op, DAG, MipsISD::VUMAX,
1407 lowerMSASplatImm(Op, 2, DAG));
1408 case Intrinsic::mips_min_s_b:
1409 case Intrinsic::mips_min_s_h:
1410 case Intrinsic::mips_min_s_w:
1411 case Intrinsic::mips_min_s_d:
1412 return lowerMSABinaryIntr(Op, DAG, MipsISD::VSMIN);
1413 case Intrinsic::mips_min_u_b:
1414 case Intrinsic::mips_min_u_h:
1415 case Intrinsic::mips_min_u_w:
1416 case Intrinsic::mips_min_u_d:
1417 return lowerMSABinaryIntr(Op, DAG, MipsISD::VUMIN);
1418 case Intrinsic::mips_mini_s_b:
1419 case Intrinsic::mips_mini_s_h:
1420 case Intrinsic::mips_mini_s_w:
1421 case Intrinsic::mips_mini_s_d:
1422 return lowerMSABinaryImmIntr(Op, DAG, MipsISD::VSMIN,
1423 lowerMSASplatImm(Op, 2, DAG));
1424 case Intrinsic::mips_mini_u_b:
1425 case Intrinsic::mips_mini_u_h:
1426 case Intrinsic::mips_mini_u_w:
1427 case Intrinsic::mips_mini_u_d:
1428 return lowerMSABinaryImmIntr(Op, DAG, MipsISD::VUMIN,
1429 lowerMSASplatImm(Op, 2, DAG));
1430 case Intrinsic::mips_mulv_b:
1431 case Intrinsic::mips_mulv_h:
1432 case Intrinsic::mips_mulv_w:
1433 case Intrinsic::mips_mulv_d:
1434 return lowerMSABinaryIntr(Op, DAG, ISD::MUL);
1435 case Intrinsic::mips_nlzc_b:
1436 case Intrinsic::mips_nlzc_h:
1437 case Intrinsic::mips_nlzc_w:
1438 case Intrinsic::mips_nlzc_d:
1439 return lowerMSAUnaryIntr(Op, DAG, ISD::CTLZ);
1440 case Intrinsic::mips_nor_v: {
1441 SDValue Res = lowerMSABinaryIntr(Op, DAG, ISD::OR);
1442 return DAG.getNOT(SDLoc(Op), Res, Res->getValueType(0));
1444 case Intrinsic::mips_nori_b: {
1445 SDValue Res = lowerMSABinaryImmIntr(Op, DAG, ISD::OR,
1446 lowerMSASplatImm(Op, 2, DAG));
1447 return DAG.getNOT(SDLoc(Op), Res, Res->getValueType(0));
1449 case Intrinsic::mips_or_v:
1450 return lowerMSABinaryIntr(Op, DAG, ISD::OR);
1451 case Intrinsic::mips_ori_b:
1452 return lowerMSABinaryImmIntr(Op, DAG, ISD::OR,
1453 lowerMSASplatImm(Op, 2, DAG));
1454 case Intrinsic::mips_pckev_b:
1455 case Intrinsic::mips_pckev_h:
1456 case Intrinsic::mips_pckev_w:
1457 case Intrinsic::mips_pckev_d:
1458 return DAG.getNode(MipsISD::PCKEV, SDLoc(Op), Op->getValueType(0),
1459 Op->getOperand(1), Op->getOperand(2));
1460 case Intrinsic::mips_pckod_b:
1461 case Intrinsic::mips_pckod_h:
1462 case Intrinsic::mips_pckod_w:
1463 case Intrinsic::mips_pckod_d:
1464 return DAG.getNode(MipsISD::PCKOD, SDLoc(Op), Op->getValueType(0),
1465 Op->getOperand(1), Op->getOperand(2));
1466 case Intrinsic::mips_pcnt_b:
1467 case Intrinsic::mips_pcnt_h:
1468 case Intrinsic::mips_pcnt_w:
1469 case Intrinsic::mips_pcnt_d:
1470 return lowerMSAUnaryIntr(Op, DAG, ISD::CTPOP);
1471 case Intrinsic::mips_shf_b:
1472 case Intrinsic::mips_shf_h:
1473 case Intrinsic::mips_shf_w:
1474 return DAG.getNode(MipsISD::SHF, SDLoc(Op), Op->getValueType(0),
1475 Op->getOperand(2), Op->getOperand(1));
1476 case Intrinsic::mips_sll_b:
1477 case Intrinsic::mips_sll_h:
1478 case Intrinsic::mips_sll_w:
1479 case Intrinsic::mips_sll_d:
1480 return lowerMSABinaryIntr(Op, DAG, ISD::SHL);
1481 case Intrinsic::mips_slli_b:
1482 case Intrinsic::mips_slli_h:
1483 case Intrinsic::mips_slli_w:
1484 case Intrinsic::mips_slli_d:
1485 return lowerMSABinaryImmIntr(Op, DAG, ISD::SHL,
1486 lowerMSASplatImm(Op, 2, DAG));
1487 case Intrinsic::mips_sra_b:
1488 case Intrinsic::mips_sra_h:
1489 case Intrinsic::mips_sra_w:
1490 case Intrinsic::mips_sra_d:
1491 return lowerMSABinaryIntr(Op, DAG, ISD::SRA);
1492 case Intrinsic::mips_srai_b:
1493 case Intrinsic::mips_srai_h:
1494 case Intrinsic::mips_srai_w:
1495 case Intrinsic::mips_srai_d:
1496 return lowerMSABinaryImmIntr(Op, DAG, ISD::SRA,
1497 lowerMSASplatImm(Op, 2, DAG));
1498 case Intrinsic::mips_srl_b:
1499 case Intrinsic::mips_srl_h:
1500 case Intrinsic::mips_srl_w:
1501 case Intrinsic::mips_srl_d:
1502 return lowerMSABinaryIntr(Op, DAG, ISD::SRL);
1503 case Intrinsic::mips_srli_b:
1504 case Intrinsic::mips_srli_h:
1505 case Intrinsic::mips_srli_w:
1506 case Intrinsic::mips_srli_d:
1507 return lowerMSABinaryImmIntr(Op, DAG, ISD::SRL,
1508 lowerMSASplatImm(Op, 2, DAG));
1509 case Intrinsic::mips_subv_b:
1510 case Intrinsic::mips_subv_h:
1511 case Intrinsic::mips_subv_w:
1512 case Intrinsic::mips_subv_d:
1513 return lowerMSABinaryIntr(Op, DAG, ISD::SUB);
1514 case Intrinsic::mips_subvi_b:
1515 case Intrinsic::mips_subvi_h:
1516 case Intrinsic::mips_subvi_w:
1517 case Intrinsic::mips_subvi_d:
1518 return lowerMSABinaryImmIntr(Op, DAG, ISD::SUB,
1519 lowerMSASplatImm(Op, 2, DAG));
1520 case Intrinsic::mips_vshf_b:
1521 case Intrinsic::mips_vshf_h:
1522 case Intrinsic::mips_vshf_w:
1523 case Intrinsic::mips_vshf_d:
1524 return DAG.getNode(MipsISD::VSHF, SDLoc(Op), Op->getValueType(0),
1525 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3));
1526 case Intrinsic::mips_xor_v:
1527 return lowerMSABinaryIntr(Op, DAG, ISD::XOR);
1528 case Intrinsic::mips_xori_b:
1529 return lowerMSABinaryImmIntr(Op, DAG, ISD::XOR,
1530 lowerMSASplatImm(Op, 2, DAG));
1534 static SDValue lowerMSALoadIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
1536 SDValue ChainIn = Op->getOperand(0);
1537 SDValue Address = Op->getOperand(2);
1538 SDValue Offset = Op->getOperand(3);
1539 EVT ResTy = Op->getValueType(0);
1540 EVT PtrTy = Address->getValueType(0);
1542 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
1544 return DAG.getLoad(ResTy, DL, ChainIn, Address, MachinePointerInfo(), false,
1548 SDValue MipsSETargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
1549 SelectionDAG &DAG) const {
1550 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1554 case Intrinsic::mips_extp:
1555 return lowerDSPIntr(Op, DAG, MipsISD::EXTP);
1556 case Intrinsic::mips_extpdp:
1557 return lowerDSPIntr(Op, DAG, MipsISD::EXTPDP);
1558 case Intrinsic::mips_extr_w:
1559 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_W);
1560 case Intrinsic::mips_extr_r_w:
1561 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W);
1562 case Intrinsic::mips_extr_rs_w:
1563 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W);
1564 case Intrinsic::mips_extr_s_h:
1565 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H);
1566 case Intrinsic::mips_mthlip:
1567 return lowerDSPIntr(Op, DAG, MipsISD::MTHLIP);
1568 case Intrinsic::mips_mulsaq_s_w_ph:
1569 return lowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH);
1570 case Intrinsic::mips_maq_s_w_phl:
1571 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL);
1572 case Intrinsic::mips_maq_s_w_phr:
1573 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR);
1574 case Intrinsic::mips_maq_sa_w_phl:
1575 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL);
1576 case Intrinsic::mips_maq_sa_w_phr:
1577 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR);
1578 case Intrinsic::mips_dpaq_s_w_ph:
1579 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH);
1580 case Intrinsic::mips_dpsq_s_w_ph:
1581 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH);
1582 case Intrinsic::mips_dpaq_sa_l_w:
1583 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W);
1584 case Intrinsic::mips_dpsq_sa_l_w:
1585 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W);
1586 case Intrinsic::mips_dpaqx_s_w_ph:
1587 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH);
1588 case Intrinsic::mips_dpaqx_sa_w_ph:
1589 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH);
1590 case Intrinsic::mips_dpsqx_s_w_ph:
1591 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH);
1592 case Intrinsic::mips_dpsqx_sa_w_ph:
1593 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH);
1594 case Intrinsic::mips_ld_b:
1595 case Intrinsic::mips_ld_h:
1596 case Intrinsic::mips_ld_w:
1597 case Intrinsic::mips_ld_d:
1598 case Intrinsic::mips_ldx_b:
1599 case Intrinsic::mips_ldx_h:
1600 case Intrinsic::mips_ldx_w:
1601 case Intrinsic::mips_ldx_d:
1602 return lowerMSALoadIntr(Op, DAG, Intr);
1606 static SDValue lowerMSAStoreIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
1608 SDValue ChainIn = Op->getOperand(0);
1609 SDValue Value = Op->getOperand(2);
1610 SDValue Address = Op->getOperand(3);
1611 SDValue Offset = Op->getOperand(4);
1612 EVT PtrTy = Address->getValueType(0);
1614 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
1616 return DAG.getStore(ChainIn, DL, Value, Address, MachinePointerInfo(), false,
1620 SDValue MipsSETargetLowering::lowerINTRINSIC_VOID(SDValue Op,
1621 SelectionDAG &DAG) const {
1622 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1626 case Intrinsic::mips_st_b:
1627 case Intrinsic::mips_st_h:
1628 case Intrinsic::mips_st_w:
1629 case Intrinsic::mips_st_d:
1630 case Intrinsic::mips_stx_b:
1631 case Intrinsic::mips_stx_h:
1632 case Intrinsic::mips_stx_w:
1633 case Intrinsic::mips_stx_d:
1634 return lowerMSAStoreIntr(Op, DAG, Intr);
1638 /// \brief Check if the given BuildVectorSDNode is a splat.
1639 /// This method currently relies on DAG nodes being reused when equivalent,
1640 /// so it's possible for this to return false even when isConstantSplat returns
1642 static bool isSplatVector(const BuildVectorSDNode *N) {
1643 unsigned int nOps = N->getNumOperands();
1644 assert(nOps > 1 && "isSplat has 0 or 1 sized build vector");
1646 SDValue Operand0 = N->getOperand(0);
1648 for (unsigned int i = 1; i < nOps; ++i) {
1649 if (N->getOperand(i) != Operand0)
1656 // Lower ISD::EXTRACT_VECTOR_ELT into MipsISD::VEXTRACT_SEXT_ELT.
1658 // The non-value bits resulting from ISD::EXTRACT_VECTOR_ELT are undefined. We
1659 // choose to sign-extend but we could have equally chosen zero-extend. The
1660 // DAGCombiner will fold any sign/zero extension of the ISD::EXTRACT_VECTOR_ELT
1661 // result into this node later (possibly changing it to a zero-extend in the
1663 SDValue MipsSETargetLowering::
1664 lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
1666 EVT ResTy = Op->getValueType(0);
1667 SDValue Op0 = Op->getOperand(0);
1668 SDValue Op1 = Op->getOperand(1);
1669 EVT EltTy = Op0->getValueType(0).getVectorElementType();
1670 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, DL, ResTy, Op0, Op1,
1671 DAG.getValueType(EltTy));
1674 static bool isConstantOrUndef(const SDValue Op) {
1675 if (Op->getOpcode() == ISD::UNDEF)
1677 if (dyn_cast<ConstantSDNode>(Op))
1679 if (dyn_cast<ConstantFPSDNode>(Op))
1684 static bool isConstantOrUndefBUILD_VECTOR(const BuildVectorSDNode *Op) {
1685 for (unsigned i = 0; i < Op->getNumOperands(); ++i)
1686 if (isConstantOrUndef(Op->getOperand(i)))
1691 // Lowers ISD::BUILD_VECTOR into appropriate SelectionDAG nodes for the
1694 // Lowers according to the following rules:
1695 // - Constant splats are legal as-is as long as the SplatBitSize is a power of
1696 // 2 less than or equal to 64 and the value fits into a signed 10-bit
1698 // - Constant splats are lowered to bitconverted BUILD_VECTORs if SplatBitSize
1699 // is a power of 2 less than or equal to 64 and the value does not fit into a
1700 // signed 10-bit immediate
1701 // - Non-constant splats are legal as-is.
1702 // - Non-constant non-splats are lowered to sequences of INSERT_VECTOR_ELT.
1703 // - All others are illegal and must be expanded.
1704 SDValue MipsSETargetLowering::lowerBUILD_VECTOR(SDValue Op,
1705 SelectionDAG &DAG) const {
1706 BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op);
1707 EVT ResTy = Op->getValueType(0);
1709 APInt SplatValue, SplatUndef;
1710 unsigned SplatBitSize;
1713 if (!Subtarget->hasMSA() || !ResTy.is128BitVector())
1716 if (Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
1718 !Subtarget->isLittle()) && SplatBitSize <= 64) {
1719 // We can only cope with 8, 16, 32, or 64-bit elements
1720 if (SplatBitSize != 8 && SplatBitSize != 16 && SplatBitSize != 32 &&
1724 // If the value fits into a simm10 then we can use ldi.[bhwd]
1725 if (SplatValue.isSignedIntN(10))
1730 switch (SplatBitSize) {
1734 ViaVecTy = MVT::v16i8;
1737 ViaVecTy = MVT::v8i16;
1740 ViaVecTy = MVT::v4i32;
1743 // There's no fill.d to fall back on for 64-bit values
1747 SmallVector<SDValue, 16> Ops;
1748 SDValue Constant = DAG.getConstant(SplatValue.sextOrSelf(32), MVT::i32);
1750 for (unsigned i = 0; i < ViaVecTy.getVectorNumElements(); ++i)
1751 Ops.push_back(Constant);
1753 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Node), ViaVecTy,
1754 &Ops[0], Ops.size());
1756 if (ViaVecTy != ResTy)
1757 Result = DAG.getNode(ISD::BITCAST, SDLoc(Node), ResTy, Result);
1760 } else if (isSplatVector(Node))
1762 else if (!isConstantOrUndefBUILD_VECTOR(Node)) {
1763 // Use INSERT_VECTOR_ELT operations rather than expand to stores.
1764 // The resulting code is the same length as the expansion, but it doesn't
1765 // use memory operations
1766 EVT ResTy = Node->getValueType(0);
1768 assert(ResTy.isVector());
1770 unsigned NumElts = ResTy.getVectorNumElements();
1771 SDValue Vector = DAG.getUNDEF(ResTy);
1772 for (unsigned i = 0; i < NumElts; ++i) {
1773 Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector,
1774 Node->getOperand(i),
1775 DAG.getConstant(i, MVT::i32));
1783 // Lower VECTOR_SHUFFLE into SHF (if possible).
1785 // SHF splits the vector into blocks of four elements, then shuffles these
1786 // elements according to a <4 x i2> constant (encoded as an integer immediate).
1788 // It is therefore possible to lower into SHF when the mask takes the form:
1789 // <a, b, c, d, a+4, b+4, c+4, d+4, a+8, b+8, c+8, d+8, ...>
1790 // When undef's appear they are treated as if they were whatever value is
1791 // necessary in order to fit the above form.
1794 // %2 = shufflevector <8 x i16> %0, <8 x i16> undef,
1795 // <8 x i32> <i32 3, i32 2, i32 1, i32 0,
1796 // i32 7, i32 6, i32 5, i32 4>
1798 // (SHF_H $w0, $w1, 27)
1799 // where the 27 comes from:
1800 // 3 + (2 << 2) + (1 << 4) + (0 << 6)
1801 static SDValue lowerVECTOR_SHUFFLE_SHF(SDValue Op, EVT ResTy,
1802 SmallVector<int, 16> Indices,
1803 SelectionDAG &DAG) {
1804 int SHFIndices[4] = { -1, -1, -1, -1 };
1806 if (Indices.size() < 4)
1809 for (unsigned i = 0; i < 4; ++i) {
1810 for (unsigned j = i; j < Indices.size(); j += 4) {
1811 int Idx = Indices[j];
1813 // Convert from vector index to 4-element subvector index
1814 // If an index refers to an element outside of the subvector then give up
1817 if (Idx < 0 || Idx >= 4)
1821 // If the mask has an undef, replace it with the current index.
1822 // Note that it might still be undef if the current index is also undef
1823 if (SHFIndices[i] == -1)
1824 SHFIndices[i] = Idx;
1826 // Check that non-undef values are the same as in the mask. If they
1827 // aren't then give up
1828 if (!(Idx == -1 || Idx == SHFIndices[i]))
1833 // Calculate the immediate. Replace any remaining undefs with zero
1835 for (int i = 3; i >= 0; --i) {
1836 int Idx = SHFIndices[i];
1845 return DAG.getNode(MipsISD::SHF, SDLoc(Op), ResTy,
1846 DAG.getConstant(Imm, MVT::i32), Op->getOperand(0));
1849 // Lower VECTOR_SHUFFLE into ILVEV (if possible).
1851 // ILVEV interleaves the even elements from each vector.
1853 // It is possible to lower into ILVEV when the mask takes the form:
1854 // <0, n, 2, n+2, 4, n+4, ...>
1855 // where n is the number of elements in the vector.
1857 // When undef's appear in the mask they are treated as if they were whatever
1858 // value is necessary in order to fit the above form.
1859 static SDValue lowerVECTOR_SHUFFLE_ILVEV(SDValue Op, EVT ResTy,
1860 SmallVector<int, 16> Indices,
1861 SelectionDAG &DAG) {
1862 assert ((Indices.size() % 2) == 0);
1864 int WtIdx = ResTy.getVectorNumElements();
1866 for (unsigned i = 0; i < Indices.size(); i += 2) {
1867 if (Indices[i] != -1 && Indices[i] != WsIdx)
1869 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
1875 return DAG.getNode(MipsISD::ILVEV, SDLoc(Op), ResTy, Op->getOperand(0),
1879 // Lower VECTOR_SHUFFLE into ILVOD (if possible).
1881 // ILVOD interleaves the odd elements from each vector.
1883 // It is possible to lower into ILVOD when the mask takes the form:
1884 // <1, n+1, 3, n+3, 5, n+5, ...>
1885 // where n is the number of elements in the vector.
1887 // When undef's appear in the mask they are treated as if they were whatever
1888 // value is necessary in order to fit the above form.
1889 static SDValue lowerVECTOR_SHUFFLE_ILVOD(SDValue Op, EVT ResTy,
1890 SmallVector<int, 16> Indices,
1891 SelectionDAG &DAG) {
1892 assert ((Indices.size() % 2) == 0);
1894 int WtIdx = ResTy.getVectorNumElements() + 1;
1896 for (unsigned i = 0; i < Indices.size(); i += 2) {
1897 if (Indices[i] != -1 && Indices[i] != WsIdx)
1899 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
1905 return DAG.getNode(MipsISD::ILVOD, SDLoc(Op), ResTy, Op->getOperand(0),
1909 // Lower VECTOR_SHUFFLE into ILVL (if possible).
1911 // ILVL interleaves consecutive elements from the left half of each vector.
1913 // It is possible to lower into ILVL when the mask takes the form:
1914 // <0, n, 1, n+1, 2, n+2, ...>
1915 // where n is the number of elements in the vector.
1917 // When undef's appear in the mask they are treated as if they were whatever
1918 // value is necessary in order to fit the above form.
1919 static SDValue lowerVECTOR_SHUFFLE_ILVL(SDValue Op, EVT ResTy,
1920 SmallVector<int, 16> Indices,
1921 SelectionDAG &DAG) {
1922 assert ((Indices.size() % 2) == 0);
1924 int WtIdx = ResTy.getVectorNumElements();
1926 for (unsigned i = 0; i < Indices.size(); i += 2) {
1927 if (Indices[i] != -1 && Indices[i] != WsIdx)
1929 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
1935 return DAG.getNode(MipsISD::ILVL, SDLoc(Op), ResTy, Op->getOperand(0),
1939 // Lower VECTOR_SHUFFLE into ILVR (if possible).
1941 // ILVR interleaves consecutive elements from the right half of each vector.
1943 // It is possible to lower into ILVR when the mask takes the form:
1944 // <x, n+x, x+1, n+x+1, x+2, n+x+2, ...>
1945 // where n is the number of elements in the vector and x is half n.
1947 // When undef's appear in the mask they are treated as if they were whatever
1948 // value is necessary in order to fit the above form.
1949 static SDValue lowerVECTOR_SHUFFLE_ILVR(SDValue Op, EVT ResTy,
1950 SmallVector<int, 16> Indices,
1951 SelectionDAG &DAG) {
1952 assert ((Indices.size() % 2) == 0);
1953 unsigned NumElts = ResTy.getVectorNumElements();
1954 int WsIdx = NumElts / 2;
1955 int WtIdx = NumElts + NumElts / 2;
1957 for (unsigned i = 0; i < Indices.size(); i += 2) {
1958 if (Indices[i] != -1 && Indices[i] != WsIdx)
1960 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
1966 return DAG.getNode(MipsISD::ILVR, SDLoc(Op), ResTy, Op->getOperand(0),
1970 // Lower VECTOR_SHUFFLE into PCKEV (if possible).
1972 // PCKEV copies the even elements of each vector into the result vector.
1974 // It is possible to lower into PCKEV when the mask takes the form:
1975 // <0, 2, 4, ..., n, n+2, n+4, ...>
1976 // where n is the number of elements in the vector.
1978 // When undef's appear in the mask they are treated as if they were whatever
1979 // value is necessary in order to fit the above form.
1980 static SDValue lowerVECTOR_SHUFFLE_PCKEV(SDValue Op, EVT ResTy,
1981 SmallVector<int, 16> Indices,
1982 SelectionDAG &DAG) {
1983 assert ((Indices.size() % 2) == 0);
1986 for (unsigned i = 0; i < Indices.size(); ++i) {
1987 if (Indices[i] != -1 && Indices[i] != Idx)
1992 return DAG.getNode(MipsISD::PCKEV, SDLoc(Op), ResTy, Op->getOperand(0),
1996 // Lower VECTOR_SHUFFLE into PCKOD (if possible).
1998 // PCKOD copies the odd elements of each vector into the result vector.
2000 // It is possible to lower into PCKOD when the mask takes the form:
2001 // <1, 3, 5, ..., n+1, n+3, n+5, ...>
2002 // where n is the number of elements in the vector.
2004 // When undef's appear in the mask they are treated as if they were whatever
2005 // value is necessary in order to fit the above form.
2006 static SDValue lowerVECTOR_SHUFFLE_PCKOD(SDValue Op, EVT ResTy,
2007 SmallVector<int, 16> Indices,
2008 SelectionDAG &DAG) {
2009 assert ((Indices.size() % 2) == 0);
2012 for (unsigned i = 0; i < Indices.size(); ++i) {
2013 if (Indices[i] != -1 && Indices[i] != Idx)
2018 return DAG.getNode(MipsISD::PCKOD, SDLoc(Op), ResTy, Op->getOperand(0),
2022 // Lower VECTOR_SHUFFLE into VSHF.
2024 // This mostly consists of converting the shuffle indices in Indices into a
2025 // BUILD_VECTOR and adding it as an operand to the resulting VSHF. There is
2026 // also code to eliminate unused operands of the VECTOR_SHUFFLE. For example,
2027 // if the type is v8i16 and all the indices are less than 8 then the second
2028 // operand is unused and can be replaced with anything. We choose to replace it
2029 // with the used operand since this reduces the number of instructions overall.
2030 static SDValue lowerVECTOR_SHUFFLE_VSHF(SDValue Op, EVT ResTy,
2031 SmallVector<int, 16> Indices,
2032 SelectionDAG &DAG) {
2033 SmallVector<SDValue, 16> Ops;
2036 EVT MaskVecTy = ResTy.changeVectorElementTypeToInteger();
2037 EVT MaskEltTy = MaskVecTy.getVectorElementType();
2038 bool Using1stVec = false;
2039 bool Using2ndVec = false;
2041 int ResTyNumElts = ResTy.getVectorNumElements();
2043 for (int i = 0; i < ResTyNumElts; ++i) {
2044 // Idx == -1 means UNDEF
2045 int Idx = Indices[i];
2047 if (0 <= Idx && Idx < ResTyNumElts)
2049 if (ResTyNumElts <= Idx && Idx < ResTyNumElts * 2)
2053 for (SmallVector<int, 16>::iterator I = Indices.begin(); I != Indices.end();
2055 Ops.push_back(DAG.getTargetConstant(*I, MaskEltTy));
2057 SDValue MaskVec = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecTy, &Ops[0],
2060 if (Using1stVec && Using2ndVec) {
2061 Op0 = Op->getOperand(0);
2062 Op1 = Op->getOperand(1);
2063 } else if (Using1stVec)
2064 Op0 = Op1 = Op->getOperand(0);
2065 else if (Using2ndVec)
2066 Op0 = Op1 = Op->getOperand(1);
2068 llvm_unreachable("shuffle vector mask references neither vector operand?");
2070 return DAG.getNode(MipsISD::VSHF, DL, ResTy, MaskVec, Op0, Op1);
2073 // Lower VECTOR_SHUFFLE into one of a number of instructions depending on the
2074 // indices in the shuffle.
2075 SDValue MipsSETargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
2076 SelectionDAG &DAG) const {
2077 ShuffleVectorSDNode *Node = cast<ShuffleVectorSDNode>(Op);
2078 EVT ResTy = Op->getValueType(0);
2080 if (!ResTy.is128BitVector())
2083 int ResTyNumElts = ResTy.getVectorNumElements();
2084 SmallVector<int, 16> Indices;
2086 for (int i = 0; i < ResTyNumElts; ++i)
2087 Indices.push_back(Node->getMaskElt(i));
2089 SDValue Result = lowerVECTOR_SHUFFLE_SHF(Op, ResTy, Indices, DAG);
2090 if (Result.getNode())
2092 Result = lowerVECTOR_SHUFFLE_ILVEV(Op, ResTy, Indices, DAG);
2093 if (Result.getNode())
2095 Result = lowerVECTOR_SHUFFLE_ILVOD(Op, ResTy, Indices, DAG);
2096 if (Result.getNode())
2098 Result = lowerVECTOR_SHUFFLE_ILVL(Op, ResTy, Indices, DAG);
2099 if (Result.getNode())
2101 Result = lowerVECTOR_SHUFFLE_ILVR(Op, ResTy, Indices, DAG);
2102 if (Result.getNode())
2104 Result = lowerVECTOR_SHUFFLE_PCKEV(Op, ResTy, Indices, DAG);
2105 if (Result.getNode())
2107 Result = lowerVECTOR_SHUFFLE_PCKOD(Op, ResTy, Indices, DAG);
2108 if (Result.getNode())
2110 return lowerVECTOR_SHUFFLE_VSHF(Op, ResTy, Indices, DAG);
2113 MachineBasicBlock * MipsSETargetLowering::
2114 emitBPOSGE32(MachineInstr *MI, MachineBasicBlock *BB) const{
2116 // bposge32_pseudo $vr0
2126 // $vr0 = phi($vr2, $fbb, $vr1, $tbb)
2128 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2129 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2130 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
2131 DebugLoc DL = MI->getDebugLoc();
2132 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2133 MachineFunction::iterator It = llvm::next(MachineFunction::iterator(BB));
2134 MachineFunction *F = BB->getParent();
2135 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
2136 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
2137 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
2140 F->insert(It, Sink);
2142 // Transfer the remainder of BB and its successor edges to Sink.
2143 Sink->splice(Sink->begin(), BB, llvm::next(MachineBasicBlock::iterator(MI)),
2145 Sink->transferSuccessorsAndUpdatePHIs(BB);
2148 BB->addSuccessor(FBB);
2149 BB->addSuccessor(TBB);
2150 FBB->addSuccessor(Sink);
2151 TBB->addSuccessor(Sink);
2153 // Insert the real bposge32 instruction to $BB.
2154 BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB);
2157 unsigned VR2 = RegInfo.createVirtualRegister(RC);
2158 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2)
2159 .addReg(Mips::ZERO).addImm(0);
2160 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
2163 unsigned VR1 = RegInfo.createVirtualRegister(RC);
2164 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1)
2165 .addReg(Mips::ZERO).addImm(1);
2167 // Insert phi function to $Sink.
2168 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
2169 MI->getOperand(0).getReg())
2170 .addReg(VR2).addMBB(FBB).addReg(VR1).addMBB(TBB);
2172 MI->eraseFromParent(); // The pseudo instruction is gone now.
2176 MachineBasicBlock * MipsSETargetLowering::
2177 emitMSACBranchPseudo(MachineInstr *MI, MachineBasicBlock *BB,
2178 unsigned BranchOp) const{
2180 // vany_nonzero $rd, $ws
2191 // $rd = phi($rd1, $fbb, $rd2, $tbb)
2193 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2194 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2195 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
2196 DebugLoc DL = MI->getDebugLoc();
2197 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2198 MachineFunction::iterator It = llvm::next(MachineFunction::iterator(BB));
2199 MachineFunction *F = BB->getParent();
2200 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
2201 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
2202 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
2205 F->insert(It, Sink);
2207 // Transfer the remainder of BB and its successor edges to Sink.
2208 Sink->splice(Sink->begin(), BB, llvm::next(MachineBasicBlock::iterator(MI)),
2210 Sink->transferSuccessorsAndUpdatePHIs(BB);
2213 BB->addSuccessor(FBB);
2214 BB->addSuccessor(TBB);
2215 FBB->addSuccessor(Sink);
2216 TBB->addSuccessor(Sink);
2218 // Insert the real bnz.b instruction to $BB.
2219 BuildMI(BB, DL, TII->get(BranchOp))
2220 .addReg(MI->getOperand(1).getReg())
2224 unsigned RD1 = RegInfo.createVirtualRegister(RC);
2225 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), RD1)
2226 .addReg(Mips::ZERO).addImm(0);
2227 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
2230 unsigned RD2 = RegInfo.createVirtualRegister(RC);
2231 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), RD2)
2232 .addReg(Mips::ZERO).addImm(1);
2234 // Insert phi function to $Sink.
2235 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
2236 MI->getOperand(0).getReg())
2237 .addReg(RD1).addMBB(FBB).addReg(RD2).addMBB(TBB);
2239 MI->eraseFromParent(); // The pseudo instruction is gone now.