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() || Subtarget->hasMSA()) {
45 // Expand all truncating stores and extending loads.
46 unsigned FirstVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
47 unsigned LastVT = (unsigned)MVT::LAST_VECTOR_VALUETYPE;
49 for (unsigned VT0 = FirstVT; VT0 <= LastVT; ++VT0) {
50 for (unsigned VT1 = FirstVT; VT1 <= LastVT; ++VT1)
51 setTruncStoreAction((MVT::SimpleValueType)VT0,
52 (MVT::SimpleValueType)VT1, Expand);
54 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT0, Expand);
55 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT0, Expand);
56 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT0, Expand);
60 if (Subtarget->hasDSP()) {
61 MVT::SimpleValueType VecTys[2] = {MVT::v2i16, MVT::v4i8};
63 for (unsigned i = 0; i < array_lengthof(VecTys); ++i) {
64 addRegisterClass(VecTys[i], &Mips::DSPRRegClass);
66 // Expand all builtin opcodes.
67 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
68 setOperationAction(Opc, VecTys[i], Expand);
70 setOperationAction(ISD::ADD, VecTys[i], Legal);
71 setOperationAction(ISD::SUB, VecTys[i], Legal);
72 setOperationAction(ISD::LOAD, VecTys[i], Legal);
73 setOperationAction(ISD::STORE, VecTys[i], Legal);
74 setOperationAction(ISD::BITCAST, VecTys[i], Legal);
77 setTargetDAGCombine(ISD::SHL);
78 setTargetDAGCombine(ISD::SRA);
79 setTargetDAGCombine(ISD::SRL);
80 setTargetDAGCombine(ISD::SETCC);
81 setTargetDAGCombine(ISD::VSELECT);
84 if (Subtarget->hasDSPR2())
85 setOperationAction(ISD::MUL, MVT::v2i16, Legal);
87 if (Subtarget->hasMSA()) {
88 addMSAIntType(MVT::v16i8, &Mips::MSA128BRegClass);
89 addMSAIntType(MVT::v8i16, &Mips::MSA128HRegClass);
90 addMSAIntType(MVT::v4i32, &Mips::MSA128WRegClass);
91 addMSAIntType(MVT::v2i64, &Mips::MSA128DRegClass);
92 addMSAFloatType(MVT::v8f16, &Mips::MSA128HRegClass);
93 addMSAFloatType(MVT::v4f32, &Mips::MSA128WRegClass);
94 addMSAFloatType(MVT::v2f64, &Mips::MSA128DRegClass);
96 setTargetDAGCombine(ISD::AND);
97 setTargetDAGCombine(ISD::SRA);
98 setTargetDAGCombine(ISD::VSELECT);
99 setTargetDAGCombine(ISD::XOR);
102 if (!Subtarget->mipsSEUsesSoftFloat()) {
103 addRegisterClass(MVT::f32, &Mips::FGR32RegClass);
105 // When dealing with single precision only, use libcalls
106 if (!Subtarget->isSingleFloat()) {
107 if (Subtarget->isFP64bit())
108 addRegisterClass(MVT::f64, &Mips::FGR64RegClass);
110 addRegisterClass(MVT::f64, &Mips::AFGR64RegClass);
114 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Custom);
115 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Custom);
116 setOperationAction(ISD::MULHS, MVT::i32, Custom);
117 setOperationAction(ISD::MULHU, MVT::i32, Custom);
120 setOperationAction(ISD::MULHS, MVT::i64, Custom);
121 setOperationAction(ISD::MULHU, MVT::i64, Custom);
122 setOperationAction(ISD::MUL, MVT::i64, Custom);
125 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
126 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
128 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
129 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
130 setOperationAction(ISD::SDIVREM, MVT::i64, Custom);
131 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
132 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
133 setOperationAction(ISD::LOAD, MVT::i32, Custom);
134 setOperationAction(ISD::STORE, MVT::i32, Custom);
136 setTargetDAGCombine(ISD::ADDE);
137 setTargetDAGCombine(ISD::SUBE);
138 setTargetDAGCombine(ISD::MUL);
140 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
141 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
142 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
145 setOperationAction(ISD::LOAD, MVT::f64, Custom);
146 setOperationAction(ISD::STORE, MVT::f64, Custom);
149 computeRegisterProperties();
152 const MipsTargetLowering *
153 llvm::createMipsSETargetLowering(MipsTargetMachine &TM) {
154 return new MipsSETargetLowering(TM);
157 // Enable MSA support for the given integer type and Register class.
158 void MipsSETargetLowering::
159 addMSAIntType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
160 addRegisterClass(Ty, RC);
162 // Expand all builtin opcodes.
163 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
164 setOperationAction(Opc, Ty, Expand);
166 setOperationAction(ISD::BITCAST, Ty, Legal);
167 setOperationAction(ISD::LOAD, Ty, Legal);
168 setOperationAction(ISD::STORE, Ty, Legal);
169 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Custom);
170 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
171 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
173 setOperationAction(ISD::ADD, Ty, Legal);
174 setOperationAction(ISD::AND, Ty, Legal);
175 setOperationAction(ISD::CTLZ, Ty, Legal);
176 setOperationAction(ISD::CTPOP, Ty, Legal);
177 setOperationAction(ISD::MUL, Ty, Legal);
178 setOperationAction(ISD::OR, Ty, Legal);
179 setOperationAction(ISD::SDIV, Ty, Legal);
180 setOperationAction(ISD::SHL, Ty, Legal);
181 setOperationAction(ISD::SRA, Ty, Legal);
182 setOperationAction(ISD::SRL, Ty, Legal);
183 setOperationAction(ISD::SUB, Ty, Legal);
184 setOperationAction(ISD::UDIV, Ty, Legal);
185 setOperationAction(ISD::VECTOR_SHUFFLE, Ty, Custom);
186 setOperationAction(ISD::VSELECT, Ty, Legal);
187 setOperationAction(ISD::XOR, Ty, Legal);
189 setOperationAction(ISD::SETCC, Ty, Legal);
190 setCondCodeAction(ISD::SETNE, Ty, Expand);
191 setCondCodeAction(ISD::SETGE, Ty, Expand);
192 setCondCodeAction(ISD::SETGT, Ty, Expand);
193 setCondCodeAction(ISD::SETUGE, Ty, Expand);
194 setCondCodeAction(ISD::SETUGT, Ty, Expand);
197 // Enable MSA support for the given floating-point type and Register class.
198 void MipsSETargetLowering::
199 addMSAFloatType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
200 addRegisterClass(Ty, RC);
202 // Expand all builtin opcodes.
203 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
204 setOperationAction(Opc, Ty, Expand);
206 setOperationAction(ISD::LOAD, Ty, Legal);
207 setOperationAction(ISD::STORE, Ty, Legal);
208 setOperationAction(ISD::BITCAST, Ty, Legal);
209 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Legal);
210 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
212 if (Ty != MVT::v8f16) {
213 setOperationAction(ISD::FABS, Ty, Legal);
214 setOperationAction(ISD::FADD, Ty, Legal);
215 setOperationAction(ISD::FDIV, Ty, Legal);
216 setOperationAction(ISD::FLOG2, Ty, Legal);
217 setOperationAction(ISD::FMUL, Ty, Legal);
218 setOperationAction(ISD::FRINT, Ty, Legal);
219 setOperationAction(ISD::FSQRT, Ty, Legal);
220 setOperationAction(ISD::FSUB, Ty, Legal);
221 setOperationAction(ISD::VSELECT, Ty, Legal);
223 setOperationAction(ISD::SETCC, Ty, Legal);
224 setCondCodeAction(ISD::SETOGE, Ty, Expand);
225 setCondCodeAction(ISD::SETOGT, Ty, Expand);
226 setCondCodeAction(ISD::SETUGE, Ty, Expand);
227 setCondCodeAction(ISD::SETUGT, Ty, Expand);
228 setCondCodeAction(ISD::SETGE, Ty, Expand);
229 setCondCodeAction(ISD::SETGT, Ty, Expand);
234 MipsSETargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
235 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
248 SDValue MipsSETargetLowering::LowerOperation(SDValue Op,
249 SelectionDAG &DAG) const {
250 switch(Op.getOpcode()) {
251 case ISD::LOAD: return lowerLOAD(Op, DAG);
252 case ISD::STORE: return lowerSTORE(Op, DAG);
253 case ISD::SMUL_LOHI: return lowerMulDiv(Op, MipsISD::Mult, true, true, DAG);
254 case ISD::UMUL_LOHI: return lowerMulDiv(Op, MipsISD::Multu, true, true, DAG);
255 case ISD::MULHS: return lowerMulDiv(Op, MipsISD::Mult, false, true, DAG);
256 case ISD::MULHU: return lowerMulDiv(Op, MipsISD::Multu, false, true, DAG);
257 case ISD::MUL: return lowerMulDiv(Op, MipsISD::Mult, true, false, DAG);
258 case ISD::SDIVREM: return lowerMulDiv(Op, MipsISD::DivRem, true, true, DAG);
259 case ISD::UDIVREM: return lowerMulDiv(Op, MipsISD::DivRemU, true, true,
261 case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG);
262 case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG);
263 case ISD::INTRINSIC_VOID: return lowerINTRINSIC_VOID(Op, DAG);
264 case ISD::EXTRACT_VECTOR_ELT: return lowerEXTRACT_VECTOR_ELT(Op, DAG);
265 case ISD::BUILD_VECTOR: return lowerBUILD_VECTOR(Op, DAG);
266 case ISD::VECTOR_SHUFFLE: return lowerVECTOR_SHUFFLE(Op, DAG);
269 return MipsTargetLowering::LowerOperation(Op, DAG);
273 // Transforms a subgraph in CurDAG if the following pattern is found:
274 // (addc multLo, Lo0), (adde multHi, Hi0),
276 // multHi/Lo: product of multiplication
277 // Lo0: initial value of Lo register
278 // Hi0: initial value of Hi register
279 // Return true if pattern matching was successful.
280 static bool selectMADD(SDNode *ADDENode, SelectionDAG *CurDAG) {
281 // ADDENode's second operand must be a flag output of an ADDC node in order
282 // for the matching to be successful.
283 SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
285 if (ADDCNode->getOpcode() != ISD::ADDC)
288 SDValue MultHi = ADDENode->getOperand(0);
289 SDValue MultLo = ADDCNode->getOperand(0);
290 SDNode *MultNode = MultHi.getNode();
291 unsigned MultOpc = MultHi.getOpcode();
293 // MultHi and MultLo must be generated by the same node,
294 if (MultLo.getNode() != MultNode)
297 // and it must be a multiplication.
298 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
301 // MultLo amd MultHi must be the first and second output of MultNode
303 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
306 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
307 // of the values of MultNode, in which case MultNode will be removed in later
309 // If there exist users other than ADDENode or ADDCNode, this function returns
310 // here, which will result in MultNode being mapped to a single MULT
311 // instruction node rather than a pair of MULT and MADD instructions being
313 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
318 // Initialize accumulator.
319 SDValue ACCIn = CurDAG->getNode(MipsISD::InsertLOHI, DL, MVT::Untyped,
320 ADDCNode->getOperand(1),
321 ADDENode->getOperand(1));
323 // create MipsMAdd(u) node
324 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
326 SDValue MAdd = CurDAG->getNode(MultOpc, DL, MVT::Untyped,
327 MultNode->getOperand(0),// Factor 0
328 MultNode->getOperand(1),// Factor 1
331 // replace uses of adde and addc here
332 if (!SDValue(ADDCNode, 0).use_empty()) {
333 SDValue LoIdx = CurDAG->getConstant(Mips::sub_lo, MVT::i32);
334 SDValue LoOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MAdd,
336 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), LoOut);
338 if (!SDValue(ADDENode, 0).use_empty()) {
339 SDValue HiIdx = CurDAG->getConstant(Mips::sub_hi, MVT::i32);
340 SDValue HiOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MAdd,
342 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), HiOut);
349 // Transforms a subgraph in CurDAG if the following pattern is found:
350 // (addc Lo0, multLo), (sube Hi0, multHi),
352 // multHi/Lo: product of multiplication
353 // Lo0: initial value of Lo register
354 // Hi0: initial value of Hi register
355 // Return true if pattern matching was successful.
356 static bool selectMSUB(SDNode *SUBENode, SelectionDAG *CurDAG) {
357 // SUBENode's second operand must be a flag output of an SUBC node in order
358 // for the matching to be successful.
359 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
361 if (SUBCNode->getOpcode() != ISD::SUBC)
364 SDValue MultHi = SUBENode->getOperand(1);
365 SDValue MultLo = SUBCNode->getOperand(1);
366 SDNode *MultNode = MultHi.getNode();
367 unsigned MultOpc = MultHi.getOpcode();
369 // MultHi and MultLo must be generated by the same node,
370 if (MultLo.getNode() != MultNode)
373 // and it must be a multiplication.
374 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
377 // MultLo amd MultHi must be the first and second output of MultNode
379 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
382 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
383 // of the values of MultNode, in which case MultNode will be removed in later
385 // If there exist users other than SUBENode or SUBCNode, this function returns
386 // here, which will result in MultNode being mapped to a single MULT
387 // instruction node rather than a pair of MULT and MSUB instructions being
389 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
394 // Initialize accumulator.
395 SDValue ACCIn = CurDAG->getNode(MipsISD::InsertLOHI, DL, MVT::Untyped,
396 SUBCNode->getOperand(0),
397 SUBENode->getOperand(0));
399 // create MipsSub(u) node
400 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
402 SDValue MSub = CurDAG->getNode(MultOpc, DL, MVT::Glue,
403 MultNode->getOperand(0),// Factor 0
404 MultNode->getOperand(1),// Factor 1
407 // replace uses of sube and subc here
408 if (!SDValue(SUBCNode, 0).use_empty()) {
409 SDValue LoIdx = CurDAG->getConstant(Mips::sub_lo, MVT::i32);
410 SDValue LoOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MSub,
412 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), LoOut);
414 if (!SDValue(SUBENode, 0).use_empty()) {
415 SDValue HiIdx = CurDAG->getConstant(Mips::sub_hi, MVT::i32);
416 SDValue HiOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MSub,
418 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), HiOut);
424 static SDValue performADDECombine(SDNode *N, SelectionDAG &DAG,
425 TargetLowering::DAGCombinerInfo &DCI,
426 const MipsSubtarget *Subtarget) {
427 if (DCI.isBeforeLegalize())
430 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
432 return SDValue(N, 0);
437 // Fold zero extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT
439 // Performs the following transformations:
440 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to zero extension if its
441 // sign/zero-extension is completely overwritten by the new one performed by
443 // - Removes redundant zero extensions performed by an ISD::AND.
444 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
445 TargetLowering::DAGCombinerInfo &DCI,
446 const MipsSubtarget *Subtarget) {
447 if (!Subtarget->hasMSA())
450 SDValue Op0 = N->getOperand(0);
451 SDValue Op1 = N->getOperand(1);
452 unsigned Op0Opcode = Op0->getOpcode();
454 // (and (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d)
455 // where $d + 1 == 2^n and n == 32
456 // or $d + 1 == 2^n and n <= 32 and ZExt
457 // -> (MipsVExtractZExt $a, $b, $c)
458 if (Op0Opcode == MipsISD::VEXTRACT_SEXT_ELT ||
459 Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT) {
460 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(Op1);
465 int32_t Log2IfPositive = (Mask->getAPIntValue() + 1).exactLogBase2();
467 if (Log2IfPositive <= 0)
468 return SDValue(); // Mask+1 is not a power of 2
470 SDValue Op0Op2 = Op0->getOperand(2);
471 EVT ExtendTy = cast<VTSDNode>(Op0Op2)->getVT();
472 unsigned ExtendTySize = ExtendTy.getSizeInBits();
473 unsigned Log2 = Log2IfPositive;
475 if ((Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT && Log2 >= ExtendTySize) ||
476 Log2 == ExtendTySize) {
477 SDValue Ops[] = { Op0->getOperand(0), Op0->getOperand(1), Op0Op2 };
478 DAG.MorphNodeTo(Op0.getNode(), MipsISD::VEXTRACT_ZEXT_ELT,
479 Op0->getVTList(), Ops, Op0->getNumOperands());
487 static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG,
488 TargetLowering::DAGCombinerInfo &DCI,
489 const MipsSubtarget *Subtarget) {
490 if (DCI.isBeforeLegalize())
493 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
495 return SDValue(N, 0);
500 static SDValue genConstMult(SDValue X, uint64_t C, SDLoc DL, EVT VT,
501 EVT ShiftTy, SelectionDAG &DAG) {
502 // Clear the upper (64 - VT.sizeInBits) bits.
503 C &= ((uint64_t)-1) >> (64 - VT.getSizeInBits());
507 return DAG.getConstant(0, VT);
513 // If c is power of 2, return (shl x, log2(c)).
514 if (isPowerOf2_64(C))
515 return DAG.getNode(ISD::SHL, DL, VT, X,
516 DAG.getConstant(Log2_64(C), ShiftTy));
518 unsigned Log2Ceil = Log2_64_Ceil(C);
519 uint64_t Floor = 1LL << Log2_64(C);
520 uint64_t Ceil = Log2Ceil == 64 ? 0LL : 1LL << Log2Ceil;
522 // If |c - floor_c| <= |c - ceil_c|,
523 // where floor_c = pow(2, floor(log2(c))) and ceil_c = pow(2, ceil(log2(c))),
524 // return (add constMult(x, floor_c), constMult(x, c - floor_c)).
525 if (C - Floor <= Ceil - C) {
526 SDValue Op0 = genConstMult(X, Floor, DL, VT, ShiftTy, DAG);
527 SDValue Op1 = genConstMult(X, C - Floor, DL, VT, ShiftTy, DAG);
528 return DAG.getNode(ISD::ADD, DL, VT, Op0, Op1);
531 // If |c - floor_c| > |c - ceil_c|,
532 // return (sub constMult(x, ceil_c), constMult(x, ceil_c - c)).
533 SDValue Op0 = genConstMult(X, Ceil, DL, VT, ShiftTy, DAG);
534 SDValue Op1 = genConstMult(X, Ceil - C, DL, VT, ShiftTy, DAG);
535 return DAG.getNode(ISD::SUB, DL, VT, Op0, Op1);
538 static SDValue performMULCombine(SDNode *N, SelectionDAG &DAG,
539 const TargetLowering::DAGCombinerInfo &DCI,
540 const MipsSETargetLowering *TL) {
541 EVT VT = N->getValueType(0);
543 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
545 return genConstMult(N->getOperand(0), C->getZExtValue(), SDLoc(N),
546 VT, TL->getScalarShiftAmountTy(VT), DAG);
548 return SDValue(N, 0);
551 static SDValue performDSPShiftCombine(unsigned Opc, SDNode *N, EVT Ty,
553 const MipsSubtarget *Subtarget) {
554 // See if this is a vector splat immediate node.
555 APInt SplatValue, SplatUndef;
556 unsigned SplatBitSize;
558 unsigned EltSize = Ty.getVectorElementType().getSizeInBits();
559 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
562 !BV->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
563 EltSize, !Subtarget->isLittle()) ||
564 (SplatBitSize != EltSize) ||
565 (SplatValue.getZExtValue() >= EltSize))
568 return DAG.getNode(Opc, SDLoc(N), Ty, N->getOperand(0),
569 DAG.getConstant(SplatValue.getZExtValue(), MVT::i32));
572 static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
573 TargetLowering::DAGCombinerInfo &DCI,
574 const MipsSubtarget *Subtarget) {
575 EVT Ty = N->getValueType(0);
577 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
580 return performDSPShiftCombine(MipsISD::SHLL_DSP, N, Ty, DAG, Subtarget);
583 // Fold sign-extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT for MSA and fold
584 // constant splats into MipsISD::SHRA_DSP for DSPr2.
586 // Performs the following transformations:
587 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to sign extension if its
588 // sign/zero-extension is completely overwritten by the new one performed by
589 // the ISD::SRA and ISD::SHL nodes.
590 // - Removes redundant sign extensions performed by an ISD::SRA and ISD::SHL
593 // See performDSPShiftCombine for more information about the transformation
595 static SDValue performSRACombine(SDNode *N, SelectionDAG &DAG,
596 TargetLowering::DAGCombinerInfo &DCI,
597 const MipsSubtarget *Subtarget) {
598 EVT Ty = N->getValueType(0);
600 if (Subtarget->hasMSA()) {
601 SDValue Op0 = N->getOperand(0);
602 SDValue Op1 = N->getOperand(1);
604 // (sra (shl (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d), imm:$d)
605 // where $d + sizeof($c) == 32
606 // or $d + sizeof($c) <= 32 and SExt
607 // -> (MipsVExtractSExt $a, $b, $c)
608 if (Op0->getOpcode() == ISD::SHL && Op1 == Op0->getOperand(1)) {
609 SDValue Op0Op0 = Op0->getOperand(0);
610 ConstantSDNode *ShAmount = dyn_cast<ConstantSDNode>(Op1);
615 if (Op0Op0->getOpcode() != MipsISD::VEXTRACT_SEXT_ELT &&
616 Op0Op0->getOpcode() != MipsISD::VEXTRACT_ZEXT_ELT)
619 EVT ExtendTy = cast<VTSDNode>(Op0Op0->getOperand(2))->getVT();
620 unsigned TotalBits = ShAmount->getZExtValue() + ExtendTy.getSizeInBits();
622 if (TotalBits == 32 ||
623 (Op0Op0->getOpcode() == MipsISD::VEXTRACT_SEXT_ELT &&
625 SDValue Ops[] = { Op0Op0->getOperand(0), Op0Op0->getOperand(1),
626 Op0Op0->getOperand(2) };
627 DAG.MorphNodeTo(Op0Op0.getNode(), MipsISD::VEXTRACT_SEXT_ELT,
628 Op0Op0->getVTList(), Ops, Op0Op0->getNumOperands());
634 if ((Ty != MVT::v2i16) && ((Ty != MVT::v4i8) || !Subtarget->hasDSPR2()))
637 return performDSPShiftCombine(MipsISD::SHRA_DSP, N, Ty, DAG, Subtarget);
641 static SDValue performSRLCombine(SDNode *N, SelectionDAG &DAG,
642 TargetLowering::DAGCombinerInfo &DCI,
643 const MipsSubtarget *Subtarget) {
644 EVT Ty = N->getValueType(0);
646 if (((Ty != MVT::v2i16) || !Subtarget->hasDSPR2()) && (Ty != MVT::v4i8))
649 return performDSPShiftCombine(MipsISD::SHRL_DSP, N, Ty, DAG, Subtarget);
652 static bool isLegalDSPCondCode(EVT Ty, ISD::CondCode CC) {
653 bool IsV216 = (Ty == MVT::v2i16);
657 case ISD::SETNE: return true;
661 case ISD::SETGE: return IsV216;
665 case ISD::SETUGE: return !IsV216;
666 default: return false;
670 static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG) {
671 EVT Ty = N->getValueType(0);
673 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
676 if (!isLegalDSPCondCode(Ty, cast<CondCodeSDNode>(N->getOperand(2))->get()))
679 return DAG.getNode(MipsISD::SETCC_DSP, SDLoc(N), Ty, N->getOperand(0),
680 N->getOperand(1), N->getOperand(2));
683 static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG) {
684 EVT Ty = N->getValueType(0);
686 if (Ty.is128BitVector() && Ty.isInteger()) {
687 // Try the following combines:
688 // (vselect (setcc $a, $b, SETLT), $b, $a)) -> (vsmax $a, $b)
689 // (vselect (setcc $a, $b, SETLE), $b, $a)) -> (vsmax $a, $b)
690 // (vselect (setcc $a, $b, SETLT), $a, $b)) -> (vsmin $a, $b)
691 // (vselect (setcc $a, $b, SETLE), $a, $b)) -> (vsmin $a, $b)
692 // (vselect (setcc $a, $b, SETULT), $b, $a)) -> (vumax $a, $b)
693 // (vselect (setcc $a, $b, SETULE), $b, $a)) -> (vumax $a, $b)
694 // (vselect (setcc $a, $b, SETULT), $a, $b)) -> (vumin $a, $b)
695 // (vselect (setcc $a, $b, SETULE), $a, $b)) -> (vumin $a, $b)
696 // SETGT/SETGE/SETUGT/SETUGE variants of these will show up initially but
697 // will be expanded to equivalent SETLT/SETLE/SETULT/SETULE versions by the
699 SDValue Op0 = N->getOperand(0);
701 if (Op0->getOpcode() != ISD::SETCC)
704 ISD::CondCode CondCode = cast<CondCodeSDNode>(Op0->getOperand(2))->get();
707 if (CondCode == ISD::SETLT || CondCode == ISD::SETLE)
709 else if (CondCode == ISD::SETULT || CondCode == ISD::SETULE)
714 SDValue Op1 = N->getOperand(1);
715 SDValue Op2 = N->getOperand(2);
716 SDValue Op0Op0 = Op0->getOperand(0);
717 SDValue Op0Op1 = Op0->getOperand(1);
719 if (Op1 == Op0Op0 && Op2 == Op0Op1)
720 return DAG.getNode(Signed ? MipsISD::VSMIN : MipsISD::VUMIN, SDLoc(N),
722 else if (Op1 == Op0Op1 && Op2 == Op0Op0)
723 return DAG.getNode(Signed ? MipsISD::VSMAX : MipsISD::VUMAX, SDLoc(N),
725 } else if ((Ty == MVT::v2i16) || (Ty == MVT::v4i8)) {
726 SDValue SetCC = N->getOperand(0);
728 if (SetCC.getOpcode() != MipsISD::SETCC_DSP)
731 return DAG.getNode(MipsISD::SELECT_CC_DSP, SDLoc(N), Ty,
732 SetCC.getOperand(0), SetCC.getOperand(1),
733 N->getOperand(1), N->getOperand(2), SetCC.getOperand(2));
739 static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
740 const MipsSubtarget *Subtarget) {
741 EVT Ty = N->getValueType(0);
743 if (Subtarget->hasMSA() && Ty.is128BitVector() && Ty.isInteger()) {
744 // Try the following combines:
745 // (xor (or $a, $b), (build_vector allones))
746 // (xor (or $a, $b), (bitcast (build_vector allones)))
747 SDValue Op0 = N->getOperand(0);
748 SDValue Op1 = N->getOperand(1);
751 if (ISD::isBuildVectorAllOnes(Op0.getNode()))
753 else if (ISD::isBuildVectorAllOnes(Op1.getNode()))
758 if (NotOp->getOpcode() == ISD::OR)
759 return DAG.getNode(MipsISD::VNOR, SDLoc(N), Ty, NotOp->getOperand(0),
760 NotOp->getOperand(1));
767 MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
768 SelectionDAG &DAG = DCI.DAG;
771 switch (N->getOpcode()) {
773 return performADDECombine(N, DAG, DCI, Subtarget);
775 Val = performANDCombine(N, DAG, DCI, Subtarget);
778 return performSUBECombine(N, DAG, DCI, Subtarget);
780 return performMULCombine(N, DAG, DCI, this);
782 return performSHLCombine(N, DAG, DCI, Subtarget);
784 return performSRACombine(N, DAG, DCI, Subtarget);
786 return performSRLCombine(N, DAG, DCI, Subtarget);
788 return performVSELECTCombine(N, DAG);
790 Val = performXORCombine(N, DAG, Subtarget);
793 Val = performSETCCCombine(N, DAG);
800 return MipsTargetLowering::PerformDAGCombine(N, DCI);
804 MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
805 MachineBasicBlock *BB) const {
806 switch (MI->getOpcode()) {
808 return MipsTargetLowering::EmitInstrWithCustomInserter(MI, BB);
809 case Mips::BPOSGE32_PSEUDO:
810 return emitBPOSGE32(MI, BB);
811 case Mips::SNZ_B_PSEUDO:
812 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_B);
813 case Mips::SNZ_H_PSEUDO:
814 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_H);
815 case Mips::SNZ_W_PSEUDO:
816 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_W);
817 case Mips::SNZ_D_PSEUDO:
818 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_D);
819 case Mips::SNZ_V_PSEUDO:
820 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_V);
821 case Mips::SZ_B_PSEUDO:
822 return emitMSACBranchPseudo(MI, BB, Mips::BZ_B);
823 case Mips::SZ_H_PSEUDO:
824 return emitMSACBranchPseudo(MI, BB, Mips::BZ_H);
825 case Mips::SZ_W_PSEUDO:
826 return emitMSACBranchPseudo(MI, BB, Mips::BZ_W);
827 case Mips::SZ_D_PSEUDO:
828 return emitMSACBranchPseudo(MI, BB, Mips::BZ_D);
829 case Mips::SZ_V_PSEUDO:
830 return emitMSACBranchPseudo(MI, BB, Mips::BZ_V);
831 case Mips::COPY_FW_PSEUDO:
832 return emitCOPY_FW(MI, BB);
833 case Mips::COPY_FD_PSEUDO:
834 return emitCOPY_FD(MI, BB);
835 case Mips::INSERT_FW_PSEUDO:
836 return emitINSERT_FW(MI, BB);
837 case Mips::INSERT_FD_PSEUDO:
838 return emitINSERT_FD(MI, BB);
842 bool MipsSETargetLowering::
843 isEligibleForTailCallOptimization(const MipsCC &MipsCCInfo,
844 unsigned NextStackOffset,
845 const MipsFunctionInfo& FI) const {
846 if (!EnableMipsTailCalls)
849 // Return false if either the callee or caller has a byval argument.
850 if (MipsCCInfo.hasByValArg() || FI.hasByvalArg())
853 // Return true if the callee's argument area is no larger than the
855 return NextStackOffset <= FI.getIncomingArgSize();
858 void MipsSETargetLowering::
859 getOpndList(SmallVectorImpl<SDValue> &Ops,
860 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
861 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
862 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const {
863 // T9 should contain the address of the callee function if
864 // -reloction-model=pic or it is an indirect call.
865 if (IsPICCall || !GlobalOrExternal) {
866 unsigned T9Reg = IsN64 ? Mips::T9_64 : Mips::T9;
867 RegsToPass.push_front(std::make_pair(T9Reg, Callee));
869 Ops.push_back(Callee);
871 MipsTargetLowering::getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal,
872 InternalLinkage, CLI, Callee, Chain);
875 SDValue MipsSETargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
876 LoadSDNode &Nd = *cast<LoadSDNode>(Op);
878 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
879 return MipsTargetLowering::lowerLOAD(Op, DAG);
881 // Replace a double precision load with two i32 loads and a buildpair64.
883 SDValue Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
884 EVT PtrVT = Ptr.getValueType();
886 // i32 load from lower address.
887 SDValue Lo = DAG.getLoad(MVT::i32, DL, Chain, Ptr,
888 MachinePointerInfo(), Nd.isVolatile(),
889 Nd.isNonTemporal(), Nd.isInvariant(),
892 // i32 load from higher address.
893 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, PtrVT));
894 SDValue Hi = DAG.getLoad(MVT::i32, DL, Lo.getValue(1), Ptr,
895 MachinePointerInfo(), Nd.isVolatile(),
896 Nd.isNonTemporal(), Nd.isInvariant(),
897 std::min(Nd.getAlignment(), 4U));
899 if (!Subtarget->isLittle())
902 SDValue BP = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
903 SDValue Ops[2] = {BP, Hi.getValue(1)};
904 return DAG.getMergeValues(Ops, 2, DL);
907 SDValue MipsSETargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
908 StoreSDNode &Nd = *cast<StoreSDNode>(Op);
910 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
911 return MipsTargetLowering::lowerSTORE(Op, DAG);
913 // Replace a double precision store with two extractelement64s and i32 stores.
915 SDValue Val = Nd.getValue(), Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
916 EVT PtrVT = Ptr.getValueType();
917 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
918 Val, DAG.getConstant(0, MVT::i32));
919 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
920 Val, DAG.getConstant(1, MVT::i32));
922 if (!Subtarget->isLittle())
925 // i32 store to lower address.
926 Chain = DAG.getStore(Chain, DL, Lo, Ptr, MachinePointerInfo(),
927 Nd.isVolatile(), Nd.isNonTemporal(), Nd.getAlignment(),
930 // i32 store to higher address.
931 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, PtrVT));
932 return DAG.getStore(Chain, DL, Hi, Ptr, MachinePointerInfo(),
933 Nd.isVolatile(), Nd.isNonTemporal(),
934 std::min(Nd.getAlignment(), 4U), Nd.getTBAAInfo());
937 SDValue MipsSETargetLowering::lowerMulDiv(SDValue Op, unsigned NewOpc,
938 bool HasLo, bool HasHi,
939 SelectionDAG &DAG) const {
940 EVT Ty = Op.getOperand(0).getValueType();
942 SDValue Mult = DAG.getNode(NewOpc, DL, MVT::Untyped,
943 Op.getOperand(0), Op.getOperand(1));
947 Lo = DAG.getNode(MipsISD::ExtractLOHI, DL, Ty, Mult,
948 DAG.getConstant(Mips::sub_lo, MVT::i32));
950 Hi = DAG.getNode(MipsISD::ExtractLOHI, DL, Ty, Mult,
951 DAG.getConstant(Mips::sub_hi, MVT::i32));
953 if (!HasLo || !HasHi)
954 return HasLo ? Lo : Hi;
956 SDValue Vals[] = { Lo, Hi };
957 return DAG.getMergeValues(Vals, 2, DL);
961 static SDValue initAccumulator(SDValue In, SDLoc DL, SelectionDAG &DAG) {
962 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
963 DAG.getConstant(0, MVT::i32));
964 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
965 DAG.getConstant(1, MVT::i32));
966 return DAG.getNode(MipsISD::InsertLOHI, DL, MVT::Untyped, InLo, InHi);
969 static SDValue extractLOHI(SDValue Op, SDLoc DL, SelectionDAG &DAG) {
970 SDValue Lo = DAG.getNode(MipsISD::ExtractLOHI, DL, MVT::i32, Op,
971 DAG.getConstant(Mips::sub_lo, MVT::i32));
972 SDValue Hi = DAG.getNode(MipsISD::ExtractLOHI, DL, MVT::i32, Op,
973 DAG.getConstant(Mips::sub_hi, MVT::i32));
974 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi);
977 // This function expands mips intrinsic nodes which have 64-bit input operands
980 // out64 = intrinsic-node in64
982 // lo = copy (extract-element (in64, 0))
983 // hi = copy (extract-element (in64, 1))
984 // mips-specific-node
987 // out64 = merge-values (v0, v1)
989 static SDValue lowerDSPIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
991 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
992 SmallVector<SDValue, 3> Ops;
995 // See if Op has a chain input.
997 Ops.push_back(Op->getOperand(OpNo++));
999 // The next operand is the intrinsic opcode.
1000 assert(Op->getOperand(OpNo).getOpcode() == ISD::TargetConstant);
1002 // See if the next operand has type i64.
1003 SDValue Opnd = Op->getOperand(++OpNo), In64;
1005 if (Opnd.getValueType() == MVT::i64)
1006 In64 = initAccumulator(Opnd, DL, DAG);
1008 Ops.push_back(Opnd);
1010 // Push the remaining operands.
1011 for (++OpNo ; OpNo < Op->getNumOperands(); ++OpNo)
1012 Ops.push_back(Op->getOperand(OpNo));
1014 // Add In64 to the end of the list.
1016 Ops.push_back(In64);
1019 SmallVector<EVT, 2> ResTys;
1021 for (SDNode::value_iterator I = Op->value_begin(), E = Op->value_end();
1023 ResTys.push_back((*I == MVT::i64) ? MVT::Untyped : *I);
1026 SDValue Val = DAG.getNode(Opc, DL, ResTys, &Ops[0], Ops.size());
1027 SDValue Out = (ResTys[0] == MVT::Untyped) ? extractLOHI(Val, DL, DAG) : Val;
1032 assert(Val->getValueType(1) == MVT::Other);
1033 SDValue Vals[] = { Out, SDValue(Val.getNode(), 1) };
1034 return DAG.getMergeValues(Vals, 2, DL);
1037 // Lower an MSA copy intrinsic into the specified SelectionDAG node
1038 static SDValue lowerMSACopyIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1040 SDValue Vec = Op->getOperand(1);
1041 SDValue Idx = Op->getOperand(2);
1042 EVT ResTy = Op->getValueType(0);
1043 EVT EltTy = Vec->getValueType(0).getVectorElementType();
1045 SDValue Result = DAG.getNode(Opc, DL, ResTy, Vec, Idx,
1046 DAG.getValueType(EltTy));
1051 // Lower an MSA insert intrinsic into the specified SelectionDAG node
1052 static SDValue lowerMSAInsertIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1054 SDValue Op0 = Op->getOperand(1);
1055 SDValue Op1 = Op->getOperand(2);
1056 SDValue Op2 = Op->getOperand(3);
1057 EVT ResTy = Op->getValueType(0);
1059 SDValue Result = DAG.getNode(Opc, DL, ResTy, Op0, Op2, Op1);
1065 lowerMSASplatImm(SDLoc DL, EVT ResTy, SDValue ImmOp, SelectionDAG &DAG) {
1066 EVT ViaVecTy = ResTy;
1067 SmallVector<SDValue, 16> Ops;
1070 if (ViaVecTy == MVT::v2i64) {
1071 ImmHiOp = DAG.getNode(ISD::SRA, DL, MVT::i32, ImmOp,
1072 DAG.getConstant(31, MVT::i32));
1073 for (unsigned i = 0; i < ViaVecTy.getVectorNumElements(); ++i) {
1074 Ops.push_back(ImmHiOp);
1075 Ops.push_back(ImmOp);
1077 ViaVecTy = MVT::v4i32;
1079 for (unsigned i = 0; i < ResTy.getVectorNumElements(); ++i)
1080 Ops.push_back(ImmOp);
1083 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy, &Ops[0],
1086 if (ResTy != ViaVecTy)
1087 Result = DAG.getNode(ISD::BITCAST, DL, ResTy, Result);
1093 lowerMSASplatImm(SDValue Op, unsigned ImmOp, SelectionDAG &DAG) {
1094 return lowerMSASplatImm(SDLoc(Op), Op->getValueType(0),
1095 Op->getOperand(ImmOp), DAG);
1098 SDValue MipsSETargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
1099 SelectionDAG &DAG) const {
1102 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) {
1105 case Intrinsic::mips_shilo:
1106 return lowerDSPIntr(Op, DAG, MipsISD::SHILO);
1107 case Intrinsic::mips_dpau_h_qbl:
1108 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL);
1109 case Intrinsic::mips_dpau_h_qbr:
1110 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR);
1111 case Intrinsic::mips_dpsu_h_qbl:
1112 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL);
1113 case Intrinsic::mips_dpsu_h_qbr:
1114 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR);
1115 case Intrinsic::mips_dpa_w_ph:
1116 return lowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH);
1117 case Intrinsic::mips_dps_w_ph:
1118 return lowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH);
1119 case Intrinsic::mips_dpax_w_ph:
1120 return lowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH);
1121 case Intrinsic::mips_dpsx_w_ph:
1122 return lowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH);
1123 case Intrinsic::mips_mulsa_w_ph:
1124 return lowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH);
1125 case Intrinsic::mips_mult:
1126 return lowerDSPIntr(Op, DAG, MipsISD::Mult);
1127 case Intrinsic::mips_multu:
1128 return lowerDSPIntr(Op, DAG, MipsISD::Multu);
1129 case Intrinsic::mips_madd:
1130 return lowerDSPIntr(Op, DAG, MipsISD::MAdd);
1131 case Intrinsic::mips_maddu:
1132 return lowerDSPIntr(Op, DAG, MipsISD::MAddu);
1133 case Intrinsic::mips_msub:
1134 return lowerDSPIntr(Op, DAG, MipsISD::MSub);
1135 case Intrinsic::mips_msubu:
1136 return lowerDSPIntr(Op, DAG, MipsISD::MSubu);
1137 case Intrinsic::mips_addv_b:
1138 case Intrinsic::mips_addv_h:
1139 case Intrinsic::mips_addv_w:
1140 case Intrinsic::mips_addv_d:
1141 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1),
1143 case Intrinsic::mips_addvi_b:
1144 case Intrinsic::mips_addvi_h:
1145 case Intrinsic::mips_addvi_w:
1146 case Intrinsic::mips_addvi_d:
1147 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1),
1148 lowerMSASplatImm(Op, 2, DAG));
1149 case Intrinsic::mips_and_v:
1150 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
1152 case Intrinsic::mips_andi_b:
1153 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
1154 lowerMSASplatImm(Op, 2, DAG));
1155 case Intrinsic::mips_bnz_b:
1156 case Intrinsic::mips_bnz_h:
1157 case Intrinsic::mips_bnz_w:
1158 case Intrinsic::mips_bnz_d:
1159 return DAG.getNode(MipsISD::VALL_NONZERO, DL, Op->getValueType(0),
1161 case Intrinsic::mips_bnz_v:
1162 return DAG.getNode(MipsISD::VANY_NONZERO, DL, Op->getValueType(0),
1164 case Intrinsic::mips_bsel_v:
1165 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1166 Op->getOperand(1), Op->getOperand(2),
1168 case Intrinsic::mips_bseli_b:
1169 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1170 Op->getOperand(1), Op->getOperand(2),
1171 lowerMSASplatImm(Op, 3, DAG));
1172 case Intrinsic::mips_bz_b:
1173 case Intrinsic::mips_bz_h:
1174 case Intrinsic::mips_bz_w:
1175 case Intrinsic::mips_bz_d:
1176 return DAG.getNode(MipsISD::VALL_ZERO, DL, Op->getValueType(0),
1178 case Intrinsic::mips_bz_v:
1179 return DAG.getNode(MipsISD::VANY_ZERO, DL, Op->getValueType(0),
1181 case Intrinsic::mips_ceq_b:
1182 case Intrinsic::mips_ceq_h:
1183 case Intrinsic::mips_ceq_w:
1184 case Intrinsic::mips_ceq_d:
1185 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1186 Op->getOperand(2), ISD::SETEQ);
1187 case Intrinsic::mips_ceqi_b:
1188 case Intrinsic::mips_ceqi_h:
1189 case Intrinsic::mips_ceqi_w:
1190 case Intrinsic::mips_ceqi_d:
1191 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1192 lowerMSASplatImm(Op, 2, DAG), ISD::SETEQ);
1193 case Intrinsic::mips_cle_s_b:
1194 case Intrinsic::mips_cle_s_h:
1195 case Intrinsic::mips_cle_s_w:
1196 case Intrinsic::mips_cle_s_d:
1197 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1198 Op->getOperand(2), ISD::SETLE);
1199 case Intrinsic::mips_clei_s_b:
1200 case Intrinsic::mips_clei_s_h:
1201 case Intrinsic::mips_clei_s_w:
1202 case Intrinsic::mips_clei_s_d:
1203 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1204 lowerMSASplatImm(Op, 2, DAG), ISD::SETLE);
1205 case Intrinsic::mips_cle_u_b:
1206 case Intrinsic::mips_cle_u_h:
1207 case Intrinsic::mips_cle_u_w:
1208 case Intrinsic::mips_cle_u_d:
1209 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1210 Op->getOperand(2), ISD::SETULE);
1211 case Intrinsic::mips_clei_u_b:
1212 case Intrinsic::mips_clei_u_h:
1213 case Intrinsic::mips_clei_u_w:
1214 case Intrinsic::mips_clei_u_d:
1215 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1216 lowerMSASplatImm(Op, 2, DAG), ISD::SETULE);
1217 case Intrinsic::mips_clt_s_b:
1218 case Intrinsic::mips_clt_s_h:
1219 case Intrinsic::mips_clt_s_w:
1220 case Intrinsic::mips_clt_s_d:
1221 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1222 Op->getOperand(2), ISD::SETLT);
1223 case Intrinsic::mips_clti_s_b:
1224 case Intrinsic::mips_clti_s_h:
1225 case Intrinsic::mips_clti_s_w:
1226 case Intrinsic::mips_clti_s_d:
1227 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1228 lowerMSASplatImm(Op, 2, DAG), ISD::SETLT);
1229 case Intrinsic::mips_clt_u_b:
1230 case Intrinsic::mips_clt_u_h:
1231 case Intrinsic::mips_clt_u_w:
1232 case Intrinsic::mips_clt_u_d:
1233 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1234 Op->getOperand(2), ISD::SETULT);
1235 case Intrinsic::mips_clti_u_b:
1236 case Intrinsic::mips_clti_u_h:
1237 case Intrinsic::mips_clti_u_w:
1238 case Intrinsic::mips_clti_u_d:
1239 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1240 lowerMSASplatImm(Op, 2, DAG), ISD::SETULT);
1241 case Intrinsic::mips_copy_s_b:
1242 case Intrinsic::mips_copy_s_h:
1243 case Intrinsic::mips_copy_s_w:
1244 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
1245 case Intrinsic::mips_copy_u_b:
1246 case Intrinsic::mips_copy_u_h:
1247 case Intrinsic::mips_copy_u_w:
1248 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
1249 case Intrinsic::mips_div_s_b:
1250 case Intrinsic::mips_div_s_h:
1251 case Intrinsic::mips_div_s_w:
1252 case Intrinsic::mips_div_s_d:
1253 return DAG.getNode(ISD::SDIV, DL, Op->getValueType(0), Op->getOperand(1),
1255 case Intrinsic::mips_div_u_b:
1256 case Intrinsic::mips_div_u_h:
1257 case Intrinsic::mips_div_u_w:
1258 case Intrinsic::mips_div_u_d:
1259 return DAG.getNode(ISD::UDIV, DL, Op->getValueType(0), Op->getOperand(1),
1261 case Intrinsic::mips_fadd_w:
1262 case Intrinsic::mips_fadd_d:
1263 return DAG.getNode(ISD::FADD, DL, Op->getValueType(0), Op->getOperand(1),
1265 // Don't lower mips_fcaf_[wd] since LLVM folds SETFALSE condcodes away
1266 case Intrinsic::mips_fceq_w:
1267 case Intrinsic::mips_fceq_d:
1268 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1269 Op->getOperand(2), ISD::SETOEQ);
1270 case Intrinsic::mips_fcle_w:
1271 case Intrinsic::mips_fcle_d:
1272 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1273 Op->getOperand(2), ISD::SETOLE);
1274 case Intrinsic::mips_fclt_w:
1275 case Intrinsic::mips_fclt_d:
1276 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1277 Op->getOperand(2), ISD::SETOLT);
1278 case Intrinsic::mips_fcne_w:
1279 case Intrinsic::mips_fcne_d:
1280 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1281 Op->getOperand(2), ISD::SETONE);
1282 case Intrinsic::mips_fcor_w:
1283 case Intrinsic::mips_fcor_d:
1284 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1285 Op->getOperand(2), ISD::SETO);
1286 case Intrinsic::mips_fcueq_w:
1287 case Intrinsic::mips_fcueq_d:
1288 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1289 Op->getOperand(2), ISD::SETUEQ);
1290 case Intrinsic::mips_fcule_w:
1291 case Intrinsic::mips_fcule_d:
1292 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1293 Op->getOperand(2), ISD::SETULE);
1294 case Intrinsic::mips_fcult_w:
1295 case Intrinsic::mips_fcult_d:
1296 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1297 Op->getOperand(2), ISD::SETULT);
1298 case Intrinsic::mips_fcun_w:
1299 case Intrinsic::mips_fcun_d:
1300 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1301 Op->getOperand(2), ISD::SETUO);
1302 case Intrinsic::mips_fcune_w:
1303 case Intrinsic::mips_fcune_d:
1304 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1305 Op->getOperand(2), ISD::SETUNE);
1306 case Intrinsic::mips_fdiv_w:
1307 case Intrinsic::mips_fdiv_d:
1308 return DAG.getNode(ISD::FDIV, DL, Op->getValueType(0), Op->getOperand(1),
1310 case Intrinsic::mips_fill_b:
1311 case Intrinsic::mips_fill_h:
1312 case Intrinsic::mips_fill_w: {
1313 SmallVector<SDValue, 16> Ops;
1314 EVT ResTy = Op->getValueType(0);
1316 for (unsigned i = 0; i < ResTy.getVectorNumElements(); ++i)
1317 Ops.push_back(Op->getOperand(1));
1319 return DAG.getNode(ISD::BUILD_VECTOR, DL, ResTy, &Ops[0],
1322 case Intrinsic::mips_flog2_w:
1323 case Intrinsic::mips_flog2_d:
1324 return DAG.getNode(ISD::FLOG2, DL, Op->getValueType(0), Op->getOperand(1));
1325 case Intrinsic::mips_fmul_w:
1326 case Intrinsic::mips_fmul_d:
1327 return DAG.getNode(ISD::FMUL, DL, Op->getValueType(0), Op->getOperand(1),
1329 case Intrinsic::mips_frint_w:
1330 case Intrinsic::mips_frint_d:
1331 return DAG.getNode(ISD::FRINT, DL, Op->getValueType(0), Op->getOperand(1));
1332 case Intrinsic::mips_fsqrt_w:
1333 case Intrinsic::mips_fsqrt_d:
1334 return DAG.getNode(ISD::FSQRT, DL, Op->getValueType(0), Op->getOperand(1));
1335 case Intrinsic::mips_fsub_w:
1336 case Intrinsic::mips_fsub_d:
1337 return DAG.getNode(ISD::FSUB, DL, Op->getValueType(0), Op->getOperand(1),
1339 case Intrinsic::mips_ilvev_b:
1340 case Intrinsic::mips_ilvev_h:
1341 case Intrinsic::mips_ilvev_w:
1342 case Intrinsic::mips_ilvev_d:
1343 return DAG.getNode(MipsISD::ILVEV, DL, Op->getValueType(0),
1344 Op->getOperand(1), Op->getOperand(2));
1345 case Intrinsic::mips_ilvl_b:
1346 case Intrinsic::mips_ilvl_h:
1347 case Intrinsic::mips_ilvl_w:
1348 case Intrinsic::mips_ilvl_d:
1349 return DAG.getNode(MipsISD::ILVL, DL, Op->getValueType(0),
1350 Op->getOperand(1), Op->getOperand(2));
1351 case Intrinsic::mips_ilvod_b:
1352 case Intrinsic::mips_ilvod_h:
1353 case Intrinsic::mips_ilvod_w:
1354 case Intrinsic::mips_ilvod_d:
1355 return DAG.getNode(MipsISD::ILVOD, DL, Op->getValueType(0),
1356 Op->getOperand(1), Op->getOperand(2));
1357 case Intrinsic::mips_ilvr_b:
1358 case Intrinsic::mips_ilvr_h:
1359 case Intrinsic::mips_ilvr_w:
1360 case Intrinsic::mips_ilvr_d:
1361 return DAG.getNode(MipsISD::ILVR, DL, Op->getValueType(0),
1362 Op->getOperand(1), Op->getOperand(2));
1363 case Intrinsic::mips_insert_b:
1364 case Intrinsic::mips_insert_h:
1365 case Intrinsic::mips_insert_w:
1366 return lowerMSAInsertIntr(Op, DAG, ISD::INSERT_VECTOR_ELT);
1367 case Intrinsic::mips_ldi_b:
1368 case Intrinsic::mips_ldi_h:
1369 case Intrinsic::mips_ldi_w:
1370 case Intrinsic::mips_ldi_d:
1371 return lowerMSASplatImm(Op, 1, DAG);
1372 case Intrinsic::mips_max_s_b:
1373 case Intrinsic::mips_max_s_h:
1374 case Intrinsic::mips_max_s_w:
1375 case Intrinsic::mips_max_s_d:
1376 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0),
1377 Op->getOperand(1), Op->getOperand(2));
1378 case Intrinsic::mips_max_u_b:
1379 case Intrinsic::mips_max_u_h:
1380 case Intrinsic::mips_max_u_w:
1381 case Intrinsic::mips_max_u_d:
1382 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0),
1383 Op->getOperand(1), Op->getOperand(2));
1384 case Intrinsic::mips_maxi_s_b:
1385 case Intrinsic::mips_maxi_s_h:
1386 case Intrinsic::mips_maxi_s_w:
1387 case Intrinsic::mips_maxi_s_d:
1388 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0),
1389 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1390 case Intrinsic::mips_maxi_u_b:
1391 case Intrinsic::mips_maxi_u_h:
1392 case Intrinsic::mips_maxi_u_w:
1393 case Intrinsic::mips_maxi_u_d:
1394 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0),
1395 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1396 case Intrinsic::mips_min_s_b:
1397 case Intrinsic::mips_min_s_h:
1398 case Intrinsic::mips_min_s_w:
1399 case Intrinsic::mips_min_s_d:
1400 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0),
1401 Op->getOperand(1), Op->getOperand(2));
1402 case Intrinsic::mips_min_u_b:
1403 case Intrinsic::mips_min_u_h:
1404 case Intrinsic::mips_min_u_w:
1405 case Intrinsic::mips_min_u_d:
1406 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0),
1407 Op->getOperand(1), Op->getOperand(2));
1408 case Intrinsic::mips_mini_s_b:
1409 case Intrinsic::mips_mini_s_h:
1410 case Intrinsic::mips_mini_s_w:
1411 case Intrinsic::mips_mini_s_d:
1412 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0),
1413 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1414 case Intrinsic::mips_mini_u_b:
1415 case Intrinsic::mips_mini_u_h:
1416 case Intrinsic::mips_mini_u_w:
1417 case Intrinsic::mips_mini_u_d:
1418 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0),
1419 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1420 case Intrinsic::mips_mulv_b:
1421 case Intrinsic::mips_mulv_h:
1422 case Intrinsic::mips_mulv_w:
1423 case Intrinsic::mips_mulv_d:
1424 return DAG.getNode(ISD::MUL, DL, Op->getValueType(0), Op->getOperand(1),
1426 case Intrinsic::mips_nlzc_b:
1427 case Intrinsic::mips_nlzc_h:
1428 case Intrinsic::mips_nlzc_w:
1429 case Intrinsic::mips_nlzc_d:
1430 return DAG.getNode(ISD::CTLZ, DL, Op->getValueType(0), Op->getOperand(1));
1431 case Intrinsic::mips_nor_v: {
1432 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0),
1433 Op->getOperand(1), Op->getOperand(2));
1434 return DAG.getNOT(DL, Res, Res->getValueType(0));
1436 case Intrinsic::mips_nori_b: {
1437 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0),
1439 lowerMSASplatImm(Op, 2, DAG));
1440 return DAG.getNOT(DL, Res, Res->getValueType(0));
1442 case Intrinsic::mips_or_v:
1443 return DAG.getNode(ISD::OR, DL, Op->getValueType(0), Op->getOperand(1),
1445 case Intrinsic::mips_ori_b:
1446 return DAG.getNode(ISD::OR, DL, Op->getValueType(0),
1447 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1448 case Intrinsic::mips_pckev_b:
1449 case Intrinsic::mips_pckev_h:
1450 case Intrinsic::mips_pckev_w:
1451 case Intrinsic::mips_pckev_d:
1452 return DAG.getNode(MipsISD::PCKEV, DL, Op->getValueType(0),
1453 Op->getOperand(1), Op->getOperand(2));
1454 case Intrinsic::mips_pckod_b:
1455 case Intrinsic::mips_pckod_h:
1456 case Intrinsic::mips_pckod_w:
1457 case Intrinsic::mips_pckod_d:
1458 return DAG.getNode(MipsISD::PCKOD, DL, Op->getValueType(0),
1459 Op->getOperand(1), Op->getOperand(2));
1460 case Intrinsic::mips_pcnt_b:
1461 case Intrinsic::mips_pcnt_h:
1462 case Intrinsic::mips_pcnt_w:
1463 case Intrinsic::mips_pcnt_d:
1464 return DAG.getNode(ISD::CTPOP, DL, Op->getValueType(0), Op->getOperand(1));
1465 case Intrinsic::mips_shf_b:
1466 case Intrinsic::mips_shf_h:
1467 case Intrinsic::mips_shf_w:
1468 return DAG.getNode(MipsISD::SHF, DL, Op->getValueType(0),
1469 Op->getOperand(2), Op->getOperand(1));
1470 case Intrinsic::mips_sll_b:
1471 case Intrinsic::mips_sll_h:
1472 case Intrinsic::mips_sll_w:
1473 case Intrinsic::mips_sll_d:
1474 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0), Op->getOperand(1),
1476 case Intrinsic::mips_slli_b:
1477 case Intrinsic::mips_slli_h:
1478 case Intrinsic::mips_slli_w:
1479 case Intrinsic::mips_slli_d:
1480 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0),
1481 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1482 case Intrinsic::mips_splati_b:
1483 case Intrinsic::mips_splati_h:
1484 case Intrinsic::mips_splati_w:
1485 case Intrinsic::mips_splati_d:
1486 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
1487 lowerMSASplatImm(Op, 2, DAG), Op->getOperand(1),
1489 case Intrinsic::mips_sra_b:
1490 case Intrinsic::mips_sra_h:
1491 case Intrinsic::mips_sra_w:
1492 case Intrinsic::mips_sra_d:
1493 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0), Op->getOperand(1),
1495 case Intrinsic::mips_srai_b:
1496 case Intrinsic::mips_srai_h:
1497 case Intrinsic::mips_srai_w:
1498 case Intrinsic::mips_srai_d:
1499 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0),
1500 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1501 case Intrinsic::mips_srl_b:
1502 case Intrinsic::mips_srl_h:
1503 case Intrinsic::mips_srl_w:
1504 case Intrinsic::mips_srl_d:
1505 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0), Op->getOperand(1),
1507 case Intrinsic::mips_srli_b:
1508 case Intrinsic::mips_srli_h:
1509 case Intrinsic::mips_srli_w:
1510 case Intrinsic::mips_srli_d:
1511 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0),
1512 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1513 case Intrinsic::mips_subv_b:
1514 case Intrinsic::mips_subv_h:
1515 case Intrinsic::mips_subv_w:
1516 case Intrinsic::mips_subv_d:
1517 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0), Op->getOperand(1),
1519 case Intrinsic::mips_subvi_b:
1520 case Intrinsic::mips_subvi_h:
1521 case Intrinsic::mips_subvi_w:
1522 case Intrinsic::mips_subvi_d:
1523 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0),
1524 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1525 case Intrinsic::mips_vshf_b:
1526 case Intrinsic::mips_vshf_h:
1527 case Intrinsic::mips_vshf_w:
1528 case Intrinsic::mips_vshf_d:
1529 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
1530 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3));
1531 case Intrinsic::mips_xor_v:
1532 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0), Op->getOperand(1),
1534 case Intrinsic::mips_xori_b:
1535 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0),
1536 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1540 static SDValue lowerMSALoadIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
1542 SDValue ChainIn = Op->getOperand(0);
1543 SDValue Address = Op->getOperand(2);
1544 SDValue Offset = Op->getOperand(3);
1545 EVT ResTy = Op->getValueType(0);
1546 EVT PtrTy = Address->getValueType(0);
1548 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
1550 return DAG.getLoad(ResTy, DL, ChainIn, Address, MachinePointerInfo(), false,
1554 SDValue MipsSETargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
1555 SelectionDAG &DAG) const {
1556 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1560 case Intrinsic::mips_extp:
1561 return lowerDSPIntr(Op, DAG, MipsISD::EXTP);
1562 case Intrinsic::mips_extpdp:
1563 return lowerDSPIntr(Op, DAG, MipsISD::EXTPDP);
1564 case Intrinsic::mips_extr_w:
1565 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_W);
1566 case Intrinsic::mips_extr_r_w:
1567 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W);
1568 case Intrinsic::mips_extr_rs_w:
1569 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W);
1570 case Intrinsic::mips_extr_s_h:
1571 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H);
1572 case Intrinsic::mips_mthlip:
1573 return lowerDSPIntr(Op, DAG, MipsISD::MTHLIP);
1574 case Intrinsic::mips_mulsaq_s_w_ph:
1575 return lowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH);
1576 case Intrinsic::mips_maq_s_w_phl:
1577 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL);
1578 case Intrinsic::mips_maq_s_w_phr:
1579 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR);
1580 case Intrinsic::mips_maq_sa_w_phl:
1581 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL);
1582 case Intrinsic::mips_maq_sa_w_phr:
1583 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR);
1584 case Intrinsic::mips_dpaq_s_w_ph:
1585 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH);
1586 case Intrinsic::mips_dpsq_s_w_ph:
1587 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH);
1588 case Intrinsic::mips_dpaq_sa_l_w:
1589 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W);
1590 case Intrinsic::mips_dpsq_sa_l_w:
1591 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W);
1592 case Intrinsic::mips_dpaqx_s_w_ph:
1593 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH);
1594 case Intrinsic::mips_dpaqx_sa_w_ph:
1595 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH);
1596 case Intrinsic::mips_dpsqx_s_w_ph:
1597 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH);
1598 case Intrinsic::mips_dpsqx_sa_w_ph:
1599 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH);
1600 case Intrinsic::mips_ld_b:
1601 case Intrinsic::mips_ld_h:
1602 case Intrinsic::mips_ld_w:
1603 case Intrinsic::mips_ld_d:
1604 case Intrinsic::mips_ldx_b:
1605 case Intrinsic::mips_ldx_h:
1606 case Intrinsic::mips_ldx_w:
1607 case Intrinsic::mips_ldx_d:
1608 return lowerMSALoadIntr(Op, DAG, Intr);
1612 static SDValue lowerMSAStoreIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
1614 SDValue ChainIn = Op->getOperand(0);
1615 SDValue Value = Op->getOperand(2);
1616 SDValue Address = Op->getOperand(3);
1617 SDValue Offset = Op->getOperand(4);
1618 EVT PtrTy = Address->getValueType(0);
1620 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
1622 return DAG.getStore(ChainIn, DL, Value, Address, MachinePointerInfo(), false,
1626 SDValue MipsSETargetLowering::lowerINTRINSIC_VOID(SDValue Op,
1627 SelectionDAG &DAG) const {
1628 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1632 case Intrinsic::mips_st_b:
1633 case Intrinsic::mips_st_h:
1634 case Intrinsic::mips_st_w:
1635 case Intrinsic::mips_st_d:
1636 case Intrinsic::mips_stx_b:
1637 case Intrinsic::mips_stx_h:
1638 case Intrinsic::mips_stx_w:
1639 case Intrinsic::mips_stx_d:
1640 return lowerMSAStoreIntr(Op, DAG, Intr);
1644 /// \brief Check if the given BuildVectorSDNode is a splat.
1645 /// This method currently relies on DAG nodes being reused when equivalent,
1646 /// so it's possible for this to return false even when isConstantSplat returns
1648 static bool isSplatVector(const BuildVectorSDNode *N) {
1649 unsigned int nOps = N->getNumOperands();
1650 assert(nOps > 1 && "isSplat has 0 or 1 sized build vector");
1652 SDValue Operand0 = N->getOperand(0);
1654 for (unsigned int i = 1; i < nOps; ++i) {
1655 if (N->getOperand(i) != Operand0)
1662 // Lower ISD::EXTRACT_VECTOR_ELT into MipsISD::VEXTRACT_SEXT_ELT.
1664 // The non-value bits resulting from ISD::EXTRACT_VECTOR_ELT are undefined. We
1665 // choose to sign-extend but we could have equally chosen zero-extend. The
1666 // DAGCombiner will fold any sign/zero extension of the ISD::EXTRACT_VECTOR_ELT
1667 // result into this node later (possibly changing it to a zero-extend in the
1669 SDValue MipsSETargetLowering::
1670 lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
1672 EVT ResTy = Op->getValueType(0);
1673 SDValue Op0 = Op->getOperand(0);
1674 EVT VecTy = Op0->getValueType(0);
1676 if (!VecTy.is128BitVector())
1679 if (ResTy.isInteger()) {
1680 SDValue Op1 = Op->getOperand(1);
1681 EVT EltTy = VecTy.getVectorElementType();
1682 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, DL, ResTy, Op0, Op1,
1683 DAG.getValueType(EltTy));
1689 static bool isConstantOrUndef(const SDValue Op) {
1690 if (Op->getOpcode() == ISD::UNDEF)
1692 if (dyn_cast<ConstantSDNode>(Op))
1694 if (dyn_cast<ConstantFPSDNode>(Op))
1699 static bool isConstantOrUndefBUILD_VECTOR(const BuildVectorSDNode *Op) {
1700 for (unsigned i = 0; i < Op->getNumOperands(); ++i)
1701 if (isConstantOrUndef(Op->getOperand(i)))
1706 // Lowers ISD::BUILD_VECTOR into appropriate SelectionDAG nodes for the
1709 // Lowers according to the following rules:
1710 // - Constant splats are legal as-is as long as the SplatBitSize is a power of
1711 // 2 less than or equal to 64 and the value fits into a signed 10-bit
1713 // - Constant splats are lowered to bitconverted BUILD_VECTORs if SplatBitSize
1714 // is a power of 2 less than or equal to 64 and the value does not fit into a
1715 // signed 10-bit immediate
1716 // - Non-constant splats are legal as-is.
1717 // - Non-constant non-splats are lowered to sequences of INSERT_VECTOR_ELT.
1718 // - All others are illegal and must be expanded.
1719 SDValue MipsSETargetLowering::lowerBUILD_VECTOR(SDValue Op,
1720 SelectionDAG &DAG) const {
1721 BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op);
1722 EVT ResTy = Op->getValueType(0);
1724 APInt SplatValue, SplatUndef;
1725 unsigned SplatBitSize;
1728 if (!Subtarget->hasMSA() || !ResTy.is128BitVector())
1731 if (Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
1733 !Subtarget->isLittle()) && SplatBitSize <= 64) {
1734 // We can only cope with 8, 16, 32, or 64-bit elements
1735 if (SplatBitSize != 8 && SplatBitSize != 16 && SplatBitSize != 32 &&
1739 // If the value fits into a simm10 then we can use ldi.[bhwd]
1740 if (SplatValue.isSignedIntN(10))
1745 switch (SplatBitSize) {
1749 ViaVecTy = MVT::v16i8;
1752 ViaVecTy = MVT::v8i16;
1755 ViaVecTy = MVT::v4i32;
1758 // There's no fill.d to fall back on for 64-bit values
1762 SmallVector<SDValue, 16> Ops;
1763 SDValue Constant = DAG.getConstant(SplatValue.sextOrSelf(32), MVT::i32);
1765 for (unsigned i = 0; i < ViaVecTy.getVectorNumElements(); ++i)
1766 Ops.push_back(Constant);
1768 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Node), ViaVecTy,
1769 &Ops[0], Ops.size());
1771 if (ViaVecTy != ResTy)
1772 Result = DAG.getNode(ISD::BITCAST, SDLoc(Node), ResTy, Result);
1775 } else if (isSplatVector(Node))
1777 else if (!isConstantOrUndefBUILD_VECTOR(Node)) {
1778 // Use INSERT_VECTOR_ELT operations rather than expand to stores.
1779 // The resulting code is the same length as the expansion, but it doesn't
1780 // use memory operations
1781 EVT ResTy = Node->getValueType(0);
1783 assert(ResTy.isVector());
1785 unsigned NumElts = ResTy.getVectorNumElements();
1786 SDValue Vector = DAG.getUNDEF(ResTy);
1787 for (unsigned i = 0; i < NumElts; ++i) {
1788 Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector,
1789 Node->getOperand(i),
1790 DAG.getConstant(i, MVT::i32));
1798 // Lower VECTOR_SHUFFLE into SHF (if possible).
1800 // SHF splits the vector into blocks of four elements, then shuffles these
1801 // elements according to a <4 x i2> constant (encoded as an integer immediate).
1803 // It is therefore possible to lower into SHF when the mask takes the form:
1804 // <a, b, c, d, a+4, b+4, c+4, d+4, a+8, b+8, c+8, d+8, ...>
1805 // When undef's appear they are treated as if they were whatever value is
1806 // necessary in order to fit the above form.
1809 // %2 = shufflevector <8 x i16> %0, <8 x i16> undef,
1810 // <8 x i32> <i32 3, i32 2, i32 1, i32 0,
1811 // i32 7, i32 6, i32 5, i32 4>
1813 // (SHF_H $w0, $w1, 27)
1814 // where the 27 comes from:
1815 // 3 + (2 << 2) + (1 << 4) + (0 << 6)
1816 static SDValue lowerVECTOR_SHUFFLE_SHF(SDValue Op, EVT ResTy,
1817 SmallVector<int, 16> Indices,
1818 SelectionDAG &DAG) {
1819 int SHFIndices[4] = { -1, -1, -1, -1 };
1821 if (Indices.size() < 4)
1824 for (unsigned i = 0; i < 4; ++i) {
1825 for (unsigned j = i; j < Indices.size(); j += 4) {
1826 int Idx = Indices[j];
1828 // Convert from vector index to 4-element subvector index
1829 // If an index refers to an element outside of the subvector then give up
1832 if (Idx < 0 || Idx >= 4)
1836 // If the mask has an undef, replace it with the current index.
1837 // Note that it might still be undef if the current index is also undef
1838 if (SHFIndices[i] == -1)
1839 SHFIndices[i] = Idx;
1841 // Check that non-undef values are the same as in the mask. If they
1842 // aren't then give up
1843 if (!(Idx == -1 || Idx == SHFIndices[i]))
1848 // Calculate the immediate. Replace any remaining undefs with zero
1850 for (int i = 3; i >= 0; --i) {
1851 int Idx = SHFIndices[i];
1860 return DAG.getNode(MipsISD::SHF, SDLoc(Op), ResTy,
1861 DAG.getConstant(Imm, MVT::i32), Op->getOperand(0));
1864 // Lower VECTOR_SHUFFLE into ILVEV (if possible).
1866 // ILVEV interleaves the even elements from each vector.
1868 // It is possible to lower into ILVEV when the mask takes the form:
1869 // <0, n, 2, n+2, 4, n+4, ...>
1870 // where n is the number of elements in the vector.
1872 // When undef's appear in the mask they are treated as if they were whatever
1873 // value is necessary in order to fit the above form.
1874 static SDValue lowerVECTOR_SHUFFLE_ILVEV(SDValue Op, EVT ResTy,
1875 SmallVector<int, 16> Indices,
1876 SelectionDAG &DAG) {
1877 assert ((Indices.size() % 2) == 0);
1879 int WtIdx = ResTy.getVectorNumElements();
1881 for (unsigned i = 0; i < Indices.size(); i += 2) {
1882 if (Indices[i] != -1 && Indices[i] != WsIdx)
1884 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
1890 return DAG.getNode(MipsISD::ILVEV, SDLoc(Op), ResTy, Op->getOperand(0),
1894 // Lower VECTOR_SHUFFLE into ILVOD (if possible).
1896 // ILVOD interleaves the odd elements from each vector.
1898 // It is possible to lower into ILVOD when the mask takes the form:
1899 // <1, n+1, 3, n+3, 5, n+5, ...>
1900 // where n is the number of elements in the vector.
1902 // When undef's appear in the mask they are treated as if they were whatever
1903 // value is necessary in order to fit the above form.
1904 static SDValue lowerVECTOR_SHUFFLE_ILVOD(SDValue Op, EVT ResTy,
1905 SmallVector<int, 16> Indices,
1906 SelectionDAG &DAG) {
1907 assert ((Indices.size() % 2) == 0);
1909 int WtIdx = ResTy.getVectorNumElements() + 1;
1911 for (unsigned i = 0; i < Indices.size(); i += 2) {
1912 if (Indices[i] != -1 && Indices[i] != WsIdx)
1914 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
1920 return DAG.getNode(MipsISD::ILVOD, SDLoc(Op), ResTy, Op->getOperand(0),
1924 // Lower VECTOR_SHUFFLE into ILVL (if possible).
1926 // ILVL interleaves consecutive elements from the left half of each vector.
1928 // It is possible to lower into ILVL when the mask takes the form:
1929 // <0, n, 1, n+1, 2, n+2, ...>
1930 // where n is the number of elements in the vector.
1932 // When undef's appear in the mask they are treated as if they were whatever
1933 // value is necessary in order to fit the above form.
1934 static SDValue lowerVECTOR_SHUFFLE_ILVL(SDValue Op, EVT ResTy,
1935 SmallVector<int, 16> Indices,
1936 SelectionDAG &DAG) {
1937 assert ((Indices.size() % 2) == 0);
1939 int WtIdx = ResTy.getVectorNumElements();
1941 for (unsigned i = 0; i < Indices.size(); i += 2) {
1942 if (Indices[i] != -1 && Indices[i] != WsIdx)
1944 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
1950 return DAG.getNode(MipsISD::ILVL, SDLoc(Op), ResTy, Op->getOperand(0),
1954 // Lower VECTOR_SHUFFLE into ILVR (if possible).
1956 // ILVR interleaves consecutive elements from the right half of each vector.
1958 // It is possible to lower into ILVR when the mask takes the form:
1959 // <x, n+x, x+1, n+x+1, x+2, n+x+2, ...>
1960 // where n is the number of elements in the vector and x is half n.
1962 // When undef's appear in the mask they are treated as if they were whatever
1963 // value is necessary in order to fit the above form.
1964 static SDValue lowerVECTOR_SHUFFLE_ILVR(SDValue Op, EVT ResTy,
1965 SmallVector<int, 16> Indices,
1966 SelectionDAG &DAG) {
1967 assert ((Indices.size() % 2) == 0);
1968 unsigned NumElts = ResTy.getVectorNumElements();
1969 int WsIdx = NumElts / 2;
1970 int WtIdx = NumElts + NumElts / 2;
1972 for (unsigned i = 0; i < Indices.size(); i += 2) {
1973 if (Indices[i] != -1 && Indices[i] != WsIdx)
1975 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
1981 return DAG.getNode(MipsISD::ILVR, SDLoc(Op), ResTy, Op->getOperand(0),
1985 // Lower VECTOR_SHUFFLE into PCKEV (if possible).
1987 // PCKEV copies the even elements of each vector into the result vector.
1989 // It is possible to lower into PCKEV when the mask takes the form:
1990 // <0, 2, 4, ..., n, n+2, n+4, ...>
1991 // where n is the number of elements in the vector.
1993 // When undef's appear in the mask they are treated as if they were whatever
1994 // value is necessary in order to fit the above form.
1995 static SDValue lowerVECTOR_SHUFFLE_PCKEV(SDValue Op, EVT ResTy,
1996 SmallVector<int, 16> Indices,
1997 SelectionDAG &DAG) {
1998 assert ((Indices.size() % 2) == 0);
2001 for (unsigned i = 0; i < Indices.size(); ++i) {
2002 if (Indices[i] != -1 && Indices[i] != Idx)
2007 return DAG.getNode(MipsISD::PCKEV, SDLoc(Op), ResTy, Op->getOperand(0),
2011 // Lower VECTOR_SHUFFLE into PCKOD (if possible).
2013 // PCKOD copies the odd elements of each vector into the result vector.
2015 // It is possible to lower into PCKOD when the mask takes the form:
2016 // <1, 3, 5, ..., n+1, n+3, n+5, ...>
2017 // where n is the number of elements in the vector.
2019 // When undef's appear in the mask they are treated as if they were whatever
2020 // value is necessary in order to fit the above form.
2021 static SDValue lowerVECTOR_SHUFFLE_PCKOD(SDValue Op, EVT ResTy,
2022 SmallVector<int, 16> Indices,
2023 SelectionDAG &DAG) {
2024 assert ((Indices.size() % 2) == 0);
2027 for (unsigned i = 0; i < Indices.size(); ++i) {
2028 if (Indices[i] != -1 && Indices[i] != Idx)
2033 return DAG.getNode(MipsISD::PCKOD, SDLoc(Op), ResTy, Op->getOperand(0),
2037 // Lower VECTOR_SHUFFLE into VSHF.
2039 // This mostly consists of converting the shuffle indices in Indices into a
2040 // BUILD_VECTOR and adding it as an operand to the resulting VSHF. There is
2041 // also code to eliminate unused operands of the VECTOR_SHUFFLE. For example,
2042 // if the type is v8i16 and all the indices are less than 8 then the second
2043 // operand is unused and can be replaced with anything. We choose to replace it
2044 // with the used operand since this reduces the number of instructions overall.
2045 static SDValue lowerVECTOR_SHUFFLE_VSHF(SDValue Op, EVT ResTy,
2046 SmallVector<int, 16> Indices,
2047 SelectionDAG &DAG) {
2048 SmallVector<SDValue, 16> Ops;
2051 EVT MaskVecTy = ResTy.changeVectorElementTypeToInteger();
2052 EVT MaskEltTy = MaskVecTy.getVectorElementType();
2053 bool Using1stVec = false;
2054 bool Using2ndVec = false;
2056 int ResTyNumElts = ResTy.getVectorNumElements();
2058 for (int i = 0; i < ResTyNumElts; ++i) {
2059 // Idx == -1 means UNDEF
2060 int Idx = Indices[i];
2062 if (0 <= Idx && Idx < ResTyNumElts)
2064 if (ResTyNumElts <= Idx && Idx < ResTyNumElts * 2)
2068 for (SmallVector<int, 16>::iterator I = Indices.begin(); I != Indices.end();
2070 Ops.push_back(DAG.getTargetConstant(*I, MaskEltTy));
2072 SDValue MaskVec = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecTy, &Ops[0],
2075 if (Using1stVec && Using2ndVec) {
2076 Op0 = Op->getOperand(0);
2077 Op1 = Op->getOperand(1);
2078 } else if (Using1stVec)
2079 Op0 = Op1 = Op->getOperand(0);
2080 else if (Using2ndVec)
2081 Op0 = Op1 = Op->getOperand(1);
2083 llvm_unreachable("shuffle vector mask references neither vector operand?");
2085 return DAG.getNode(MipsISD::VSHF, DL, ResTy, MaskVec, Op0, Op1);
2088 // Lower VECTOR_SHUFFLE into one of a number of instructions depending on the
2089 // indices in the shuffle.
2090 SDValue MipsSETargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
2091 SelectionDAG &DAG) const {
2092 ShuffleVectorSDNode *Node = cast<ShuffleVectorSDNode>(Op);
2093 EVT ResTy = Op->getValueType(0);
2095 if (!ResTy.is128BitVector())
2098 int ResTyNumElts = ResTy.getVectorNumElements();
2099 SmallVector<int, 16> Indices;
2101 for (int i = 0; i < ResTyNumElts; ++i)
2102 Indices.push_back(Node->getMaskElt(i));
2104 SDValue Result = lowerVECTOR_SHUFFLE_SHF(Op, ResTy, Indices, DAG);
2105 if (Result.getNode())
2107 Result = lowerVECTOR_SHUFFLE_ILVEV(Op, ResTy, Indices, DAG);
2108 if (Result.getNode())
2110 Result = lowerVECTOR_SHUFFLE_ILVOD(Op, ResTy, Indices, DAG);
2111 if (Result.getNode())
2113 Result = lowerVECTOR_SHUFFLE_ILVL(Op, ResTy, Indices, DAG);
2114 if (Result.getNode())
2116 Result = lowerVECTOR_SHUFFLE_ILVR(Op, ResTy, Indices, DAG);
2117 if (Result.getNode())
2119 Result = lowerVECTOR_SHUFFLE_PCKEV(Op, ResTy, Indices, DAG);
2120 if (Result.getNode())
2122 Result = lowerVECTOR_SHUFFLE_PCKOD(Op, ResTy, Indices, DAG);
2123 if (Result.getNode())
2125 return lowerVECTOR_SHUFFLE_VSHF(Op, ResTy, Indices, DAG);
2128 MachineBasicBlock * MipsSETargetLowering::
2129 emitBPOSGE32(MachineInstr *MI, MachineBasicBlock *BB) const{
2131 // bposge32_pseudo $vr0
2141 // $vr0 = phi($vr2, $fbb, $vr1, $tbb)
2143 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2144 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2145 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
2146 DebugLoc DL = MI->getDebugLoc();
2147 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2148 MachineFunction::iterator It = llvm::next(MachineFunction::iterator(BB));
2149 MachineFunction *F = BB->getParent();
2150 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
2151 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
2152 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
2155 F->insert(It, Sink);
2157 // Transfer the remainder of BB and its successor edges to Sink.
2158 Sink->splice(Sink->begin(), BB, llvm::next(MachineBasicBlock::iterator(MI)),
2160 Sink->transferSuccessorsAndUpdatePHIs(BB);
2163 BB->addSuccessor(FBB);
2164 BB->addSuccessor(TBB);
2165 FBB->addSuccessor(Sink);
2166 TBB->addSuccessor(Sink);
2168 // Insert the real bposge32 instruction to $BB.
2169 BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB);
2172 unsigned VR2 = RegInfo.createVirtualRegister(RC);
2173 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2)
2174 .addReg(Mips::ZERO).addImm(0);
2175 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
2178 unsigned VR1 = RegInfo.createVirtualRegister(RC);
2179 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1)
2180 .addReg(Mips::ZERO).addImm(1);
2182 // Insert phi function to $Sink.
2183 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
2184 MI->getOperand(0).getReg())
2185 .addReg(VR2).addMBB(FBB).addReg(VR1).addMBB(TBB);
2187 MI->eraseFromParent(); // The pseudo instruction is gone now.
2191 MachineBasicBlock * MipsSETargetLowering::
2192 emitMSACBranchPseudo(MachineInstr *MI, MachineBasicBlock *BB,
2193 unsigned BranchOp) const{
2195 // vany_nonzero $rd, $ws
2206 // $rd = phi($rd1, $fbb, $rd2, $tbb)
2208 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2209 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2210 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
2211 DebugLoc DL = MI->getDebugLoc();
2212 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2213 MachineFunction::iterator It = llvm::next(MachineFunction::iterator(BB));
2214 MachineFunction *F = BB->getParent();
2215 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
2216 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
2217 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
2220 F->insert(It, Sink);
2222 // Transfer the remainder of BB and its successor edges to Sink.
2223 Sink->splice(Sink->begin(), BB, llvm::next(MachineBasicBlock::iterator(MI)),
2225 Sink->transferSuccessorsAndUpdatePHIs(BB);
2228 BB->addSuccessor(FBB);
2229 BB->addSuccessor(TBB);
2230 FBB->addSuccessor(Sink);
2231 TBB->addSuccessor(Sink);
2233 // Insert the real bnz.b instruction to $BB.
2234 BuildMI(BB, DL, TII->get(BranchOp))
2235 .addReg(MI->getOperand(1).getReg())
2239 unsigned RD1 = RegInfo.createVirtualRegister(RC);
2240 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), RD1)
2241 .addReg(Mips::ZERO).addImm(0);
2242 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
2245 unsigned RD2 = RegInfo.createVirtualRegister(RC);
2246 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), RD2)
2247 .addReg(Mips::ZERO).addImm(1);
2249 // Insert phi function to $Sink.
2250 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
2251 MI->getOperand(0).getReg())
2252 .addReg(RD1).addMBB(FBB).addReg(RD2).addMBB(TBB);
2254 MI->eraseFromParent(); // The pseudo instruction is gone now.
2258 // Emit the COPY_FW pseudo instruction.
2260 // copy_fw_pseudo $fd, $ws, n
2262 // copy_u_w $rt, $ws, $n
2265 // When n is zero, the equivalent operation can be performed with (potentially)
2266 // zero instructions due to register overlaps. This optimization is never valid
2267 // for lane 1 because it would require FR=0 mode which isn't supported by MSA.
2268 MachineBasicBlock * MipsSETargetLowering::
2269 emitCOPY_FW(MachineInstr *MI, MachineBasicBlock *BB) const{
2270 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2271 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2272 DebugLoc DL = MI->getDebugLoc();
2273 unsigned Fd = MI->getOperand(0).getReg();
2274 unsigned Ws = MI->getOperand(1).getReg();
2275 unsigned Lane = MI->getOperand(2).getImm();
2278 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_lo);
2280 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
2282 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wt).addReg(Ws).addImm(1);
2283 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_lo);
2286 MI->eraseFromParent(); // The pseudo instruction is gone now.
2290 // Emit the COPY_FD pseudo instruction.
2292 // copy_fd_pseudo $fd, $ws, n
2294 // splati.d $wt, $ws, $n
2295 // copy $fd, $wt:sub_64
2297 // When n is zero, the equivalent operation can be performed with (potentially)
2298 // zero instructions due to register overlaps. This optimization is always
2299 // valid because FR=1 mode which is the only supported mode in MSA.
2300 MachineBasicBlock * MipsSETargetLowering::
2301 emitCOPY_FD(MachineInstr *MI, MachineBasicBlock *BB) const{
2302 assert(Subtarget->isFP64bit());
2304 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2305 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2306 unsigned Fd = MI->getOperand(0).getReg();
2307 unsigned Ws = MI->getOperand(1).getReg();
2308 unsigned Lane = MI->getOperand(2).getImm() * 2;
2309 DebugLoc DL = MI->getDebugLoc();
2312 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_64);
2314 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
2316 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wt).addReg(Ws).addImm(1);
2317 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_64);
2320 MI->eraseFromParent(); // The pseudo instruction is gone now.
2324 // Emit the INSERT_FW pseudo instruction.
2326 // insert_fw_pseudo $wd, $wd_in, $n, $fs
2328 // subreg_to_reg $wt:sub_lo, $fs
2329 // insve_w $wd[$n], $wd_in, $wt[0]
2330 MachineBasicBlock * MipsSETargetLowering::
2331 emitINSERT_FW(MachineInstr *MI, MachineBasicBlock *BB) const{
2332 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2333 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2334 DebugLoc DL = MI->getDebugLoc();
2335 unsigned Wd = MI->getOperand(0).getReg();
2336 unsigned Wd_in = MI->getOperand(1).getReg();
2337 unsigned Lane = MI->getOperand(2).getImm();
2338 unsigned Fs = MI->getOperand(3).getReg();
2339 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
2341 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
2342 .addImm(0).addReg(Fs).addImm(Mips::sub_lo);
2343 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_W), Wd)
2344 .addReg(Wd_in).addImm(Lane).addReg(Wt);
2346 MI->eraseFromParent(); // The pseudo instruction is gone now.
2350 // Emit the INSERT_FD pseudo instruction.
2352 // insert_fd_pseudo $wd, $fs, n
2354 // subreg_to_reg $wt:sub_64, $fs
2355 // insve_d $wd[$n], $wd_in, $wt[0]
2356 MachineBasicBlock * MipsSETargetLowering::
2357 emitINSERT_FD(MachineInstr *MI, MachineBasicBlock *BB) const{
2358 assert(Subtarget->isFP64bit());
2360 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2361 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2362 DebugLoc DL = MI->getDebugLoc();
2363 unsigned Wd = MI->getOperand(0).getReg();
2364 unsigned Wd_in = MI->getOperand(1).getReg();
2365 unsigned Lane = MI->getOperand(2).getImm();
2366 unsigned Fs = MI->getOperand(3).getReg();
2367 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
2369 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
2370 .addImm(0).addReg(Fs).addImm(Mips::sub_64);
2371 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_D), Wd)
2372 .addReg(Wd_in).addImm(Lane).addReg(Wt);
2374 MI->eraseFromParent(); // The pseudo instruction is gone now.