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::VSELECT, Ty, Legal);
184 setOperationAction(ISD::XOR, Ty, Legal);
186 setOperationAction(ISD::SETCC, Ty, Legal);
187 setCondCodeAction(ISD::SETNE, Ty, Expand);
188 setCondCodeAction(ISD::SETGE, Ty, Expand);
189 setCondCodeAction(ISD::SETGT, Ty, Expand);
190 setCondCodeAction(ISD::SETUGE, Ty, Expand);
191 setCondCodeAction(ISD::SETUGT, Ty, Expand);
194 // Enable MSA support for the given floating-point type and Register class.
195 void MipsSETargetLowering::
196 addMSAFloatType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
197 addRegisterClass(Ty, RC);
199 // Expand all builtin opcodes.
200 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
201 setOperationAction(Opc, Ty, Expand);
203 setOperationAction(ISD::LOAD, Ty, Legal);
204 setOperationAction(ISD::STORE, Ty, Legal);
205 setOperationAction(ISD::BITCAST, Ty, Legal);
206 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Legal);
208 if (Ty != MVT::v8f16) {
209 setOperationAction(ISD::FADD, Ty, Legal);
210 setOperationAction(ISD::FDIV, Ty, Legal);
211 setOperationAction(ISD::FLOG2, Ty, Legal);
212 setOperationAction(ISD::FMUL, Ty, Legal);
213 setOperationAction(ISD::FRINT, Ty, Legal);
214 setOperationAction(ISD::FSQRT, Ty, Legal);
215 setOperationAction(ISD::FSUB, Ty, Legal);
216 setOperationAction(ISD::VSELECT, Ty, Legal);
218 setOperationAction(ISD::SETCC, Ty, Legal);
219 setCondCodeAction(ISD::SETOGE, Ty, Expand);
220 setCondCodeAction(ISD::SETOGT, Ty, Expand);
221 setCondCodeAction(ISD::SETUGE, Ty, Expand);
222 setCondCodeAction(ISD::SETUGT, Ty, Expand);
223 setCondCodeAction(ISD::SETGE, Ty, Expand);
224 setCondCodeAction(ISD::SETGT, Ty, Expand);
229 MipsSETargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
230 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
243 SDValue MipsSETargetLowering::LowerOperation(SDValue Op,
244 SelectionDAG &DAG) const {
245 switch(Op.getOpcode()) {
246 case ISD::LOAD: return lowerLOAD(Op, DAG);
247 case ISD::STORE: return lowerSTORE(Op, DAG);
248 case ISD::SMUL_LOHI: return lowerMulDiv(Op, MipsISD::Mult, true, true, DAG);
249 case ISD::UMUL_LOHI: return lowerMulDiv(Op, MipsISD::Multu, true, true, DAG);
250 case ISD::MULHS: return lowerMulDiv(Op, MipsISD::Mult, false, true, DAG);
251 case ISD::MULHU: return lowerMulDiv(Op, MipsISD::Multu, false, true, DAG);
252 case ISD::MUL: return lowerMulDiv(Op, MipsISD::Mult, true, false, DAG);
253 case ISD::SDIVREM: return lowerMulDiv(Op, MipsISD::DivRem, true, true, DAG);
254 case ISD::UDIVREM: return lowerMulDiv(Op, MipsISD::DivRemU, true, true,
256 case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG);
257 case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG);
258 case ISD::INTRINSIC_VOID: return lowerINTRINSIC_VOID(Op, DAG);
259 case ISD::EXTRACT_VECTOR_ELT: return lowerEXTRACT_VECTOR_ELT(Op, DAG);
260 case ISD::BUILD_VECTOR: return lowerBUILD_VECTOR(Op, DAG);
263 return MipsTargetLowering::LowerOperation(Op, DAG);
267 // Transforms a subgraph in CurDAG if the following pattern is found:
268 // (addc multLo, Lo0), (adde multHi, Hi0),
270 // multHi/Lo: product of multiplication
271 // Lo0: initial value of Lo register
272 // Hi0: initial value of Hi register
273 // Return true if pattern matching was successful.
274 static bool selectMADD(SDNode *ADDENode, SelectionDAG *CurDAG) {
275 // ADDENode's second operand must be a flag output of an ADDC node in order
276 // for the matching to be successful.
277 SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
279 if (ADDCNode->getOpcode() != ISD::ADDC)
282 SDValue MultHi = ADDENode->getOperand(0);
283 SDValue MultLo = ADDCNode->getOperand(0);
284 SDNode *MultNode = MultHi.getNode();
285 unsigned MultOpc = MultHi.getOpcode();
287 // MultHi and MultLo must be generated by the same node,
288 if (MultLo.getNode() != MultNode)
291 // and it must be a multiplication.
292 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
295 // MultLo amd MultHi must be the first and second output of MultNode
297 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
300 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
301 // of the values of MultNode, in which case MultNode will be removed in later
303 // If there exist users other than ADDENode or ADDCNode, this function returns
304 // here, which will result in MultNode being mapped to a single MULT
305 // instruction node rather than a pair of MULT and MADD instructions being
307 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
312 // Initialize accumulator.
313 SDValue ACCIn = CurDAG->getNode(MipsISD::InsertLOHI, DL, MVT::Untyped,
314 ADDCNode->getOperand(1),
315 ADDENode->getOperand(1));
317 // create MipsMAdd(u) node
318 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
320 SDValue MAdd = CurDAG->getNode(MultOpc, DL, MVT::Untyped,
321 MultNode->getOperand(0),// Factor 0
322 MultNode->getOperand(1),// Factor 1
325 // replace uses of adde and addc here
326 if (!SDValue(ADDCNode, 0).use_empty()) {
327 SDValue LoIdx = CurDAG->getConstant(Mips::sub_lo, MVT::i32);
328 SDValue LoOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MAdd,
330 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), LoOut);
332 if (!SDValue(ADDENode, 0).use_empty()) {
333 SDValue HiIdx = CurDAG->getConstant(Mips::sub_hi, MVT::i32);
334 SDValue HiOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MAdd,
336 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), HiOut);
343 // Transforms a subgraph in CurDAG if the following pattern is found:
344 // (addc Lo0, multLo), (sube Hi0, multHi),
346 // multHi/Lo: product of multiplication
347 // Lo0: initial value of Lo register
348 // Hi0: initial value of Hi register
349 // Return true if pattern matching was successful.
350 static bool selectMSUB(SDNode *SUBENode, SelectionDAG *CurDAG) {
351 // SUBENode's second operand must be a flag output of an SUBC node in order
352 // for the matching to be successful.
353 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
355 if (SUBCNode->getOpcode() != ISD::SUBC)
358 SDValue MultHi = SUBENode->getOperand(1);
359 SDValue MultLo = SUBCNode->getOperand(1);
360 SDNode *MultNode = MultHi.getNode();
361 unsigned MultOpc = MultHi.getOpcode();
363 // MultHi and MultLo must be generated by the same node,
364 if (MultLo.getNode() != MultNode)
367 // and it must be a multiplication.
368 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
371 // MultLo amd MultHi must be the first and second output of MultNode
373 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
376 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
377 // of the values of MultNode, in which case MultNode will be removed in later
379 // If there exist users other than SUBENode or SUBCNode, this function returns
380 // here, which will result in MultNode being mapped to a single MULT
381 // instruction node rather than a pair of MULT and MSUB instructions being
383 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
388 // Initialize accumulator.
389 SDValue ACCIn = CurDAG->getNode(MipsISD::InsertLOHI, DL, MVT::Untyped,
390 SUBCNode->getOperand(0),
391 SUBENode->getOperand(0));
393 // create MipsSub(u) node
394 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
396 SDValue MSub = CurDAG->getNode(MultOpc, DL, MVT::Glue,
397 MultNode->getOperand(0),// Factor 0
398 MultNode->getOperand(1),// Factor 1
401 // replace uses of sube and subc here
402 if (!SDValue(SUBCNode, 0).use_empty()) {
403 SDValue LoIdx = CurDAG->getConstant(Mips::sub_lo, MVT::i32);
404 SDValue LoOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MSub,
406 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), LoOut);
408 if (!SDValue(SUBENode, 0).use_empty()) {
409 SDValue HiIdx = CurDAG->getConstant(Mips::sub_hi, MVT::i32);
410 SDValue HiOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MSub,
412 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), HiOut);
418 static SDValue performADDECombine(SDNode *N, SelectionDAG &DAG,
419 TargetLowering::DAGCombinerInfo &DCI,
420 const MipsSubtarget *Subtarget) {
421 if (DCI.isBeforeLegalize())
424 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
426 return SDValue(N, 0);
431 // Fold zero extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT
433 // Performs the following transformations:
434 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to zero extension if its
435 // sign/zero-extension is completely overwritten by the new one performed by
437 // - Removes redundant zero extensions performed by an ISD::AND.
438 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
439 TargetLowering::DAGCombinerInfo &DCI,
440 const MipsSubtarget *Subtarget) {
441 if (!Subtarget->hasMSA())
444 SDValue Op0 = N->getOperand(0);
445 SDValue Op1 = N->getOperand(1);
446 unsigned Op0Opcode = Op0->getOpcode();
448 // (and (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d)
449 // where $d + 1 == 2^n and n == 32
450 // or $d + 1 == 2^n and n <= 32 and ZExt
451 // -> (MipsVExtractZExt $a, $b, $c)
452 if (Op0Opcode == MipsISD::VEXTRACT_SEXT_ELT ||
453 Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT) {
454 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(Op1);
459 int32_t Log2IfPositive = (Mask->getAPIntValue() + 1).exactLogBase2();
461 if (Log2IfPositive <= 0)
462 return SDValue(); // Mask+1 is not a power of 2
464 SDValue Op0Op2 = Op0->getOperand(2);
465 EVT ExtendTy = cast<VTSDNode>(Op0Op2)->getVT();
466 unsigned ExtendTySize = ExtendTy.getSizeInBits();
467 unsigned Log2 = Log2IfPositive;
469 if ((Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT && Log2 >= ExtendTySize) ||
470 Log2 == ExtendTySize) {
471 SDValue Ops[] = { Op0->getOperand(0), Op0->getOperand(1), Op0Op2 };
472 DAG.MorphNodeTo(Op0.getNode(), MipsISD::VEXTRACT_ZEXT_ELT,
473 Op0->getVTList(), Ops, Op0->getNumOperands());
481 static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG,
482 TargetLowering::DAGCombinerInfo &DCI,
483 const MipsSubtarget *Subtarget) {
484 if (DCI.isBeforeLegalize())
487 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
489 return SDValue(N, 0);
494 static SDValue genConstMult(SDValue X, uint64_t C, SDLoc DL, EVT VT,
495 EVT ShiftTy, SelectionDAG &DAG) {
496 // Clear the upper (64 - VT.sizeInBits) bits.
497 C &= ((uint64_t)-1) >> (64 - VT.getSizeInBits());
501 return DAG.getConstant(0, VT);
507 // If c is power of 2, return (shl x, log2(c)).
508 if (isPowerOf2_64(C))
509 return DAG.getNode(ISD::SHL, DL, VT, X,
510 DAG.getConstant(Log2_64(C), ShiftTy));
512 unsigned Log2Ceil = Log2_64_Ceil(C);
513 uint64_t Floor = 1LL << Log2_64(C);
514 uint64_t Ceil = Log2Ceil == 64 ? 0LL : 1LL << Log2Ceil;
516 // If |c - floor_c| <= |c - ceil_c|,
517 // where floor_c = pow(2, floor(log2(c))) and ceil_c = pow(2, ceil(log2(c))),
518 // return (add constMult(x, floor_c), constMult(x, c - floor_c)).
519 if (C - Floor <= Ceil - C) {
520 SDValue Op0 = genConstMult(X, Floor, DL, VT, ShiftTy, DAG);
521 SDValue Op1 = genConstMult(X, C - Floor, DL, VT, ShiftTy, DAG);
522 return DAG.getNode(ISD::ADD, DL, VT, Op0, Op1);
525 // If |c - floor_c| > |c - ceil_c|,
526 // return (sub constMult(x, ceil_c), constMult(x, ceil_c - c)).
527 SDValue Op0 = genConstMult(X, Ceil, DL, VT, ShiftTy, DAG);
528 SDValue Op1 = genConstMult(X, Ceil - C, DL, VT, ShiftTy, DAG);
529 return DAG.getNode(ISD::SUB, DL, VT, Op0, Op1);
532 static SDValue performMULCombine(SDNode *N, SelectionDAG &DAG,
533 const TargetLowering::DAGCombinerInfo &DCI,
534 const MipsSETargetLowering *TL) {
535 EVT VT = N->getValueType(0);
537 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
539 return genConstMult(N->getOperand(0), C->getZExtValue(), SDLoc(N),
540 VT, TL->getScalarShiftAmountTy(VT), DAG);
542 return SDValue(N, 0);
545 static SDValue performDSPShiftCombine(unsigned Opc, SDNode *N, EVT Ty,
547 const MipsSubtarget *Subtarget) {
548 // See if this is a vector splat immediate node.
549 APInt SplatValue, SplatUndef;
550 unsigned SplatBitSize;
552 unsigned EltSize = Ty.getVectorElementType().getSizeInBits();
553 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
556 !BV->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
557 EltSize, !Subtarget->isLittle()) ||
558 (SplatBitSize != EltSize) ||
559 (SplatValue.getZExtValue() >= EltSize))
562 return DAG.getNode(Opc, SDLoc(N), Ty, N->getOperand(0),
563 DAG.getConstant(SplatValue.getZExtValue(), MVT::i32));
566 static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
567 TargetLowering::DAGCombinerInfo &DCI,
568 const MipsSubtarget *Subtarget) {
569 EVT Ty = N->getValueType(0);
571 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
574 return performDSPShiftCombine(MipsISD::SHLL_DSP, N, Ty, DAG, Subtarget);
577 // Fold sign-extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT for MSA and fold
578 // constant splats into MipsISD::SHRA_DSP for DSPr2.
580 // Performs the following transformations:
581 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to sign extension if its
582 // sign/zero-extension is completely overwritten by the new one performed by
583 // the ISD::SRA and ISD::SHL nodes.
584 // - Removes redundant sign extensions performed by an ISD::SRA and ISD::SHL
587 // See performDSPShiftCombine for more information about the transformation
589 static SDValue performSRACombine(SDNode *N, SelectionDAG &DAG,
590 TargetLowering::DAGCombinerInfo &DCI,
591 const MipsSubtarget *Subtarget) {
592 EVT Ty = N->getValueType(0);
594 if (Subtarget->hasMSA()) {
595 SDValue Op0 = N->getOperand(0);
596 SDValue Op1 = N->getOperand(1);
598 // (sra (shl (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d), imm:$d)
599 // where $d + sizeof($c) == 32
600 // or $d + sizeof($c) <= 32 and SExt
601 // -> (MipsVExtractSExt $a, $b, $c)
602 if (Op0->getOpcode() == ISD::SHL && Op1 == Op0->getOperand(1)) {
603 SDValue Op0Op0 = Op0->getOperand(0);
604 ConstantSDNode *ShAmount = dyn_cast<ConstantSDNode>(Op1);
609 EVT ExtendTy = cast<VTSDNode>(Op0Op0->getOperand(2))->getVT();
610 unsigned TotalBits = ShAmount->getZExtValue() + ExtendTy.getSizeInBits();
612 if (TotalBits == 32 ||
613 (Op0Op0->getOpcode() == MipsISD::VEXTRACT_SEXT_ELT &&
615 SDValue Ops[] = { Op0Op0->getOperand(0), Op0Op0->getOperand(1),
616 Op0Op0->getOperand(2) };
617 DAG.MorphNodeTo(Op0Op0.getNode(), MipsISD::VEXTRACT_SEXT_ELT,
618 Op0Op0->getVTList(), Ops, Op0Op0->getNumOperands());
624 if ((Ty != MVT::v2i16) && ((Ty != MVT::v4i8) || !Subtarget->hasDSPR2()))
627 return performDSPShiftCombine(MipsISD::SHRA_DSP, N, Ty, DAG, Subtarget);
631 static SDValue performSRLCombine(SDNode *N, SelectionDAG &DAG,
632 TargetLowering::DAGCombinerInfo &DCI,
633 const MipsSubtarget *Subtarget) {
634 EVT Ty = N->getValueType(0);
636 if (((Ty != MVT::v2i16) || !Subtarget->hasDSPR2()) && (Ty != MVT::v4i8))
639 return performDSPShiftCombine(MipsISD::SHRL_DSP, N, Ty, DAG, Subtarget);
642 static bool isLegalDSPCondCode(EVT Ty, ISD::CondCode CC) {
643 bool IsV216 = (Ty == MVT::v2i16);
647 case ISD::SETNE: return true;
651 case ISD::SETGE: return IsV216;
655 case ISD::SETUGE: return !IsV216;
656 default: return false;
660 static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG) {
661 EVT Ty = N->getValueType(0);
663 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
666 if (!isLegalDSPCondCode(Ty, cast<CondCodeSDNode>(N->getOperand(2))->get()))
669 return DAG.getNode(MipsISD::SETCC_DSP, SDLoc(N), Ty, N->getOperand(0),
670 N->getOperand(1), N->getOperand(2));
673 static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG) {
674 EVT Ty = N->getValueType(0);
676 if (Ty.is128BitVector() && Ty.isInteger()) {
677 // Try the following combines:
678 // (vselect (setcc $a, $b, SETLT), $b, $a)) -> (vsmax $a, $b)
679 // (vselect (setcc $a, $b, SETLE), $b, $a)) -> (vsmax $a, $b)
680 // (vselect (setcc $a, $b, SETLT), $a, $b)) -> (vsmin $a, $b)
681 // (vselect (setcc $a, $b, SETLE), $a, $b)) -> (vsmin $a, $b)
682 // (vselect (setcc $a, $b, SETULT), $b, $a)) -> (vumax $a, $b)
683 // (vselect (setcc $a, $b, SETULE), $b, $a)) -> (vumax $a, $b)
684 // (vselect (setcc $a, $b, SETULT), $a, $b)) -> (vumin $a, $b)
685 // (vselect (setcc $a, $b, SETULE), $a, $b)) -> (vumin $a, $b)
686 // SETGT/SETGE/SETUGT/SETUGE variants of these will show up initially but
687 // will be expanded to equivalent SETLT/SETLE/SETULT/SETULE versions by the
689 SDValue Op0 = N->getOperand(0);
691 if (Op0->getOpcode() != ISD::SETCC)
694 ISD::CondCode CondCode = cast<CondCodeSDNode>(Op0->getOperand(2))->get();
697 if (CondCode == ISD::SETLT || CondCode == ISD::SETLE)
699 else if (CondCode == ISD::SETULT || CondCode == ISD::SETULE)
704 SDValue Op1 = N->getOperand(1);
705 SDValue Op2 = N->getOperand(2);
706 SDValue Op0Op0 = Op0->getOperand(0);
707 SDValue Op0Op1 = Op0->getOperand(1);
709 if (Op1 == Op0Op0 && Op2 == Op0Op1)
710 return DAG.getNode(Signed ? MipsISD::VSMIN : MipsISD::VUMIN, SDLoc(N),
712 else if (Op1 == Op0Op1 && Op2 == Op0Op0)
713 return DAG.getNode(Signed ? MipsISD::VSMAX : MipsISD::VUMAX, SDLoc(N),
715 } else if ((Ty == MVT::v2i16) || (Ty == MVT::v4i8)) {
716 SDValue SetCC = N->getOperand(0);
718 if (SetCC.getOpcode() != MipsISD::SETCC_DSP)
721 return DAG.getNode(MipsISD::SELECT_CC_DSP, SDLoc(N), Ty,
722 SetCC.getOperand(0), SetCC.getOperand(1),
723 N->getOperand(1), N->getOperand(2), SetCC.getOperand(2));
729 static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
730 const MipsSubtarget *Subtarget) {
731 EVT Ty = N->getValueType(0);
733 if (Subtarget->hasMSA() && Ty.is128BitVector() && Ty.isInteger()) {
734 // Try the following combines:
735 // (xor (or $a, $b), (build_vector allones))
736 // (xor (or $a, $b), (bitcast (build_vector allones)))
737 SDValue Op0 = N->getOperand(0);
738 SDValue Op1 = N->getOperand(1);
740 ConstantSDNode *Const;
742 if (ISD::isBuildVectorAllOnes(Op0.getNode()))
744 else if (ISD::isBuildVectorAllOnes(Op1.getNode()))
746 else if ((Op0->getOpcode() == MipsISD::VSPLAT ||
747 Op0->getOpcode() == MipsISD::VSPLATD) &&
748 (Const = dyn_cast<ConstantSDNode>(Op0->getOperand(0))) &&
749 Const->isAllOnesValue())
751 else if ((Op1->getOpcode() == MipsISD::VSPLAT ||
752 Op1->getOpcode() == MipsISD::VSPLATD) &&
753 (Const = dyn_cast<ConstantSDNode>(Op1->getOperand(0))) &&
754 Const->isAllOnesValue())
759 if (NotOp->getOpcode() == ISD::OR)
760 return DAG.getNode(MipsISD::VNOR, SDLoc(N), Ty, NotOp->getOperand(0),
761 NotOp->getOperand(1));
768 MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
769 SelectionDAG &DAG = DCI.DAG;
772 switch (N->getOpcode()) {
774 return performADDECombine(N, DAG, DCI, Subtarget);
776 Val = performANDCombine(N, DAG, DCI, Subtarget);
779 return performSUBECombine(N, DAG, DCI, Subtarget);
781 return performMULCombine(N, DAG, DCI, this);
783 return performSHLCombine(N, DAG, DCI, Subtarget);
785 return performSRACombine(N, DAG, DCI, Subtarget);
787 return performSRLCombine(N, DAG, DCI, Subtarget);
789 return performVSELECTCombine(N, DAG);
791 Val = performXORCombine(N, DAG, Subtarget);
794 Val = performSETCCCombine(N, DAG);
801 return MipsTargetLowering::PerformDAGCombine(N, DCI);
805 MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
806 MachineBasicBlock *BB) const {
807 switch (MI->getOpcode()) {
809 return MipsTargetLowering::EmitInstrWithCustomInserter(MI, BB);
810 case Mips::BPOSGE32_PSEUDO:
811 return emitBPOSGE32(MI, BB);
812 case Mips::SNZ_B_PSEUDO:
813 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_B);
814 case Mips::SNZ_H_PSEUDO:
815 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_H);
816 case Mips::SNZ_W_PSEUDO:
817 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_W);
818 case Mips::SNZ_D_PSEUDO:
819 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_D);
820 case Mips::SNZ_V_PSEUDO:
821 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_V);
822 case Mips::SZ_B_PSEUDO:
823 return emitMSACBranchPseudo(MI, BB, Mips::BZ_B);
824 case Mips::SZ_H_PSEUDO:
825 return emitMSACBranchPseudo(MI, BB, Mips::BZ_H);
826 case Mips::SZ_W_PSEUDO:
827 return emitMSACBranchPseudo(MI, BB, Mips::BZ_W);
828 case Mips::SZ_D_PSEUDO:
829 return emitMSACBranchPseudo(MI, BB, Mips::BZ_D);
830 case Mips::SZ_V_PSEUDO:
831 return emitMSACBranchPseudo(MI, BB, Mips::BZ_V);
835 bool MipsSETargetLowering::
836 isEligibleForTailCallOptimization(const MipsCC &MipsCCInfo,
837 unsigned NextStackOffset,
838 const MipsFunctionInfo& FI) const {
839 if (!EnableMipsTailCalls)
842 // Return false if either the callee or caller has a byval argument.
843 if (MipsCCInfo.hasByValArg() || FI.hasByvalArg())
846 // Return true if the callee's argument area is no larger than the
848 return NextStackOffset <= FI.getIncomingArgSize();
851 void MipsSETargetLowering::
852 getOpndList(SmallVectorImpl<SDValue> &Ops,
853 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
854 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
855 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const {
856 // T9 should contain the address of the callee function if
857 // -reloction-model=pic or it is an indirect call.
858 if (IsPICCall || !GlobalOrExternal) {
859 unsigned T9Reg = IsN64 ? Mips::T9_64 : Mips::T9;
860 RegsToPass.push_front(std::make_pair(T9Reg, Callee));
862 Ops.push_back(Callee);
864 MipsTargetLowering::getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal,
865 InternalLinkage, CLI, Callee, Chain);
868 SDValue MipsSETargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
869 LoadSDNode &Nd = *cast<LoadSDNode>(Op);
871 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
872 return MipsTargetLowering::lowerLOAD(Op, DAG);
874 // Replace a double precision load with two i32 loads and a buildpair64.
876 SDValue Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
877 EVT PtrVT = Ptr.getValueType();
879 // i32 load from lower address.
880 SDValue Lo = DAG.getLoad(MVT::i32, DL, Chain, Ptr,
881 MachinePointerInfo(), Nd.isVolatile(),
882 Nd.isNonTemporal(), Nd.isInvariant(),
885 // i32 load from higher address.
886 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, PtrVT));
887 SDValue Hi = DAG.getLoad(MVT::i32, DL, Lo.getValue(1), Ptr,
888 MachinePointerInfo(), Nd.isVolatile(),
889 Nd.isNonTemporal(), Nd.isInvariant(),
890 std::min(Nd.getAlignment(), 4U));
892 if (!Subtarget->isLittle())
895 SDValue BP = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
896 SDValue Ops[2] = {BP, Hi.getValue(1)};
897 return DAG.getMergeValues(Ops, 2, DL);
900 SDValue MipsSETargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
901 StoreSDNode &Nd = *cast<StoreSDNode>(Op);
903 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
904 return MipsTargetLowering::lowerSTORE(Op, DAG);
906 // Replace a double precision store with two extractelement64s and i32 stores.
908 SDValue Val = Nd.getValue(), Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
909 EVT PtrVT = Ptr.getValueType();
910 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
911 Val, DAG.getConstant(0, MVT::i32));
912 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
913 Val, DAG.getConstant(1, MVT::i32));
915 if (!Subtarget->isLittle())
918 // i32 store to lower address.
919 Chain = DAG.getStore(Chain, DL, Lo, Ptr, MachinePointerInfo(),
920 Nd.isVolatile(), Nd.isNonTemporal(), Nd.getAlignment(),
923 // i32 store to higher address.
924 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, PtrVT));
925 return DAG.getStore(Chain, DL, Hi, Ptr, MachinePointerInfo(),
926 Nd.isVolatile(), Nd.isNonTemporal(),
927 std::min(Nd.getAlignment(), 4U), Nd.getTBAAInfo());
930 SDValue MipsSETargetLowering::lowerMulDiv(SDValue Op, unsigned NewOpc,
931 bool HasLo, bool HasHi,
932 SelectionDAG &DAG) const {
933 EVT Ty = Op.getOperand(0).getValueType();
935 SDValue Mult = DAG.getNode(NewOpc, DL, MVT::Untyped,
936 Op.getOperand(0), Op.getOperand(1));
940 Lo = DAG.getNode(MipsISD::ExtractLOHI, DL, Ty, Mult,
941 DAG.getConstant(Mips::sub_lo, MVT::i32));
943 Hi = DAG.getNode(MipsISD::ExtractLOHI, DL, Ty, Mult,
944 DAG.getConstant(Mips::sub_hi, MVT::i32));
946 if (!HasLo || !HasHi)
947 return HasLo ? Lo : Hi;
949 SDValue Vals[] = { Lo, Hi };
950 return DAG.getMergeValues(Vals, 2, DL);
954 static SDValue initAccumulator(SDValue In, SDLoc DL, SelectionDAG &DAG) {
955 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
956 DAG.getConstant(0, MVT::i32));
957 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
958 DAG.getConstant(1, MVT::i32));
959 return DAG.getNode(MipsISD::InsertLOHI, DL, MVT::Untyped, InLo, InHi);
962 static SDValue extractLOHI(SDValue Op, SDLoc DL, SelectionDAG &DAG) {
963 SDValue Lo = DAG.getNode(MipsISD::ExtractLOHI, DL, MVT::i32, Op,
964 DAG.getConstant(Mips::sub_lo, MVT::i32));
965 SDValue Hi = DAG.getNode(MipsISD::ExtractLOHI, DL, MVT::i32, Op,
966 DAG.getConstant(Mips::sub_hi, MVT::i32));
967 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi);
970 // This function expands mips intrinsic nodes which have 64-bit input operands
973 // out64 = intrinsic-node in64
975 // lo = copy (extract-element (in64, 0))
976 // hi = copy (extract-element (in64, 1))
977 // mips-specific-node
980 // out64 = merge-values (v0, v1)
982 static SDValue lowerDSPIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
984 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
985 SmallVector<SDValue, 3> Ops;
988 // See if Op has a chain input.
990 Ops.push_back(Op->getOperand(OpNo++));
992 // The next operand is the intrinsic opcode.
993 assert(Op->getOperand(OpNo).getOpcode() == ISD::TargetConstant);
995 // See if the next operand has type i64.
996 SDValue Opnd = Op->getOperand(++OpNo), In64;
998 if (Opnd.getValueType() == MVT::i64)
999 In64 = initAccumulator(Opnd, DL, DAG);
1001 Ops.push_back(Opnd);
1003 // Push the remaining operands.
1004 for (++OpNo ; OpNo < Op->getNumOperands(); ++OpNo)
1005 Ops.push_back(Op->getOperand(OpNo));
1007 // Add In64 to the end of the list.
1009 Ops.push_back(In64);
1012 SmallVector<EVT, 2> ResTys;
1014 for (SDNode::value_iterator I = Op->value_begin(), E = Op->value_end();
1016 ResTys.push_back((*I == MVT::i64) ? MVT::Untyped : *I);
1019 SDValue Val = DAG.getNode(Opc, DL, ResTys, &Ops[0], Ops.size());
1020 SDValue Out = (ResTys[0] == MVT::Untyped) ? extractLOHI(Val, DL, DAG) : Val;
1025 assert(Val->getValueType(1) == MVT::Other);
1026 SDValue Vals[] = { Out, SDValue(Val.getNode(), 1) };
1027 return DAG.getMergeValues(Vals, 2, DL);
1030 static SDValue lowerMSABinaryIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1032 SDValue LHS = Op->getOperand(1);
1033 SDValue RHS = Op->getOperand(2);
1034 EVT ResTy = Op->getValueType(0);
1036 SDValue Result = DAG.getNode(Opc, DL, ResTy, LHS, RHS);
1041 static SDValue lowerMSABinaryImmIntr(SDValue Op, SelectionDAG &DAG,
1042 unsigned Opc, SDValue RHS) {
1043 SDValue LHS = Op->getOperand(1);
1044 EVT ResTy = Op->getValueType(0);
1046 return DAG.getNode(Opc, SDLoc(Op), ResTy, LHS, RHS);
1049 static SDValue lowerMSABranchIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1051 SDValue Value = Op->getOperand(1);
1052 EVT ResTy = Op->getValueType(0);
1054 SDValue Result = DAG.getNode(Opc, DL, ResTy, Value);
1059 // Lower an MSA copy intrinsic into the specified SelectionDAG node
1060 static SDValue lowerMSACopyIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1062 SDValue Vec = Op->getOperand(1);
1063 SDValue Idx = Op->getOperand(2);
1064 EVT ResTy = Op->getValueType(0);
1065 EVT EltTy = Vec->getValueType(0).getVectorElementType();
1067 SDValue Result = DAG.getNode(Opc, DL, ResTy, Vec, Idx,
1068 DAG.getValueType(EltTy));
1073 // Lower an MSA insert intrinsic into the specified SelectionDAG node
1074 static SDValue lowerMSAInsertIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1076 SDValue Op0 = Op->getOperand(1);
1077 SDValue Op1 = Op->getOperand(2);
1078 SDValue Op2 = Op->getOperand(3);
1079 EVT ResTy = Op->getValueType(0);
1081 SDValue Result = DAG.getNode(Opc, DL, ResTy, Op0, Op2, Op1);
1086 static SDValue lowerMSASplatImm(SDValue Op, unsigned ImmOp, SelectionDAG &DAG) {
1087 EVT ResTy = Op->getValueType(0);
1089 unsigned SplatOp = MipsISD::VSPLAT;
1090 if (ResTy == MVT::v2i64)
1091 SplatOp = MipsISD::VSPLATD;
1093 return DAG.getNode(SplatOp, SDLoc(Op), ResTy, Op->getOperand(ImmOp));
1096 static SDValue lowerMSAUnaryIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1098 SDValue Value = Op->getOperand(1);
1099 EVT ResTy = Op->getValueType(0);
1101 SDValue Result = DAG.getNode(Opc, DL, ResTy, Value);
1106 SDValue MipsSETargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
1107 SelectionDAG &DAG) const {
1108 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) {
1111 case Intrinsic::mips_shilo:
1112 return lowerDSPIntr(Op, DAG, MipsISD::SHILO);
1113 case Intrinsic::mips_dpau_h_qbl:
1114 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL);
1115 case Intrinsic::mips_dpau_h_qbr:
1116 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR);
1117 case Intrinsic::mips_dpsu_h_qbl:
1118 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL);
1119 case Intrinsic::mips_dpsu_h_qbr:
1120 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR);
1121 case Intrinsic::mips_dpa_w_ph:
1122 return lowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH);
1123 case Intrinsic::mips_dps_w_ph:
1124 return lowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH);
1125 case Intrinsic::mips_dpax_w_ph:
1126 return lowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH);
1127 case Intrinsic::mips_dpsx_w_ph:
1128 return lowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH);
1129 case Intrinsic::mips_mulsa_w_ph:
1130 return lowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH);
1131 case Intrinsic::mips_mult:
1132 return lowerDSPIntr(Op, DAG, MipsISD::Mult);
1133 case Intrinsic::mips_multu:
1134 return lowerDSPIntr(Op, DAG, MipsISD::Multu);
1135 case Intrinsic::mips_madd:
1136 return lowerDSPIntr(Op, DAG, MipsISD::MAdd);
1137 case Intrinsic::mips_maddu:
1138 return lowerDSPIntr(Op, DAG, MipsISD::MAddu);
1139 case Intrinsic::mips_msub:
1140 return lowerDSPIntr(Op, DAG, MipsISD::MSub);
1141 case Intrinsic::mips_msubu:
1142 return lowerDSPIntr(Op, DAG, MipsISD::MSubu);
1143 case Intrinsic::mips_addv_b:
1144 case Intrinsic::mips_addv_h:
1145 case Intrinsic::mips_addv_w:
1146 case Intrinsic::mips_addv_d:
1147 return lowerMSABinaryIntr(Op, DAG, ISD::ADD);
1148 case Intrinsic::mips_addvi_b:
1149 case Intrinsic::mips_addvi_h:
1150 case Intrinsic::mips_addvi_w:
1151 case Intrinsic::mips_addvi_d:
1152 return lowerMSABinaryImmIntr(Op, DAG, ISD::ADD,
1153 lowerMSASplatImm(Op, 2, DAG));
1154 case Intrinsic::mips_and_v:
1155 return lowerMSABinaryIntr(Op, DAG, ISD::AND);
1156 case Intrinsic::mips_bnz_b:
1157 case Intrinsic::mips_bnz_h:
1158 case Intrinsic::mips_bnz_w:
1159 case Intrinsic::mips_bnz_d:
1160 return lowerMSABranchIntr(Op, DAG, MipsISD::VALL_NONZERO);
1161 case Intrinsic::mips_bnz_v:
1162 return lowerMSABranchIntr(Op, DAG, MipsISD::VANY_NONZERO);
1163 case Intrinsic::mips_bsel_v:
1164 return DAG.getNode(ISD::VSELECT, SDLoc(Op), Op->getValueType(0),
1165 Op->getOperand(1), Op->getOperand(2),
1167 case Intrinsic::mips_bseli_b:
1168 return DAG.getNode(ISD::VSELECT, SDLoc(Op), Op->getValueType(0),
1169 Op->getOperand(1), Op->getOperand(2),
1170 lowerMSASplatImm(Op, 3, DAG));
1171 case Intrinsic::mips_bz_b:
1172 case Intrinsic::mips_bz_h:
1173 case Intrinsic::mips_bz_w:
1174 case Intrinsic::mips_bz_d:
1175 return lowerMSABranchIntr(Op, DAG, MipsISD::VALL_ZERO);
1176 case Intrinsic::mips_bz_v:
1177 return lowerMSABranchIntr(Op, DAG, MipsISD::VANY_ZERO);
1178 case Intrinsic::mips_ceq_b:
1179 case Intrinsic::mips_ceq_h:
1180 case Intrinsic::mips_ceq_w:
1181 case Intrinsic::mips_ceq_d:
1182 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1183 Op->getOperand(2), ISD::SETEQ);
1184 case Intrinsic::mips_ceqi_b:
1185 case Intrinsic::mips_ceqi_h:
1186 case Intrinsic::mips_ceqi_w:
1187 case Intrinsic::mips_ceqi_d:
1188 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1189 lowerMSASplatImm(Op, 2, DAG), ISD::SETEQ);
1190 case Intrinsic::mips_cle_s_b:
1191 case Intrinsic::mips_cle_s_h:
1192 case Intrinsic::mips_cle_s_w:
1193 case Intrinsic::mips_cle_s_d:
1194 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1195 Op->getOperand(2), ISD::SETLE);
1196 case Intrinsic::mips_clei_s_b:
1197 case Intrinsic::mips_clei_s_h:
1198 case Intrinsic::mips_clei_s_w:
1199 case Intrinsic::mips_clei_s_d:
1200 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1201 lowerMSASplatImm(Op, 2, DAG), ISD::SETLE);
1202 case Intrinsic::mips_cle_u_b:
1203 case Intrinsic::mips_cle_u_h:
1204 case Intrinsic::mips_cle_u_w:
1205 case Intrinsic::mips_cle_u_d:
1206 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1207 Op->getOperand(2), ISD::SETULE);
1208 case Intrinsic::mips_clei_u_b:
1209 case Intrinsic::mips_clei_u_h:
1210 case Intrinsic::mips_clei_u_w:
1211 case Intrinsic::mips_clei_u_d:
1212 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1213 lowerMSASplatImm(Op, 2, DAG), ISD::SETULE);
1214 case Intrinsic::mips_clt_s_b:
1215 case Intrinsic::mips_clt_s_h:
1216 case Intrinsic::mips_clt_s_w:
1217 case Intrinsic::mips_clt_s_d:
1218 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1219 Op->getOperand(2), ISD::SETLT);
1220 case Intrinsic::mips_clti_s_b:
1221 case Intrinsic::mips_clti_s_h:
1222 case Intrinsic::mips_clti_s_w:
1223 case Intrinsic::mips_clti_s_d:
1224 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1225 lowerMSASplatImm(Op, 2, DAG), ISD::SETLT);
1226 case Intrinsic::mips_clt_u_b:
1227 case Intrinsic::mips_clt_u_h:
1228 case Intrinsic::mips_clt_u_w:
1229 case Intrinsic::mips_clt_u_d:
1230 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1231 Op->getOperand(2), ISD::SETULT);
1232 case Intrinsic::mips_clti_u_b:
1233 case Intrinsic::mips_clti_u_h:
1234 case Intrinsic::mips_clti_u_w:
1235 case Intrinsic::mips_clti_u_d:
1236 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1237 lowerMSASplatImm(Op, 2, DAG), ISD::SETULT);
1238 case Intrinsic::mips_copy_s_b:
1239 case Intrinsic::mips_copy_s_h:
1240 case Intrinsic::mips_copy_s_w:
1241 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
1242 case Intrinsic::mips_copy_u_b:
1243 case Intrinsic::mips_copy_u_h:
1244 case Intrinsic::mips_copy_u_w:
1245 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
1246 case Intrinsic::mips_div_s_b:
1247 case Intrinsic::mips_div_s_h:
1248 case Intrinsic::mips_div_s_w:
1249 case Intrinsic::mips_div_s_d:
1250 return lowerMSABinaryIntr(Op, DAG, ISD::SDIV);
1251 case Intrinsic::mips_div_u_b:
1252 case Intrinsic::mips_div_u_h:
1253 case Intrinsic::mips_div_u_w:
1254 case Intrinsic::mips_div_u_d:
1255 return lowerMSABinaryIntr(Op, DAG, ISD::UDIV);
1256 case Intrinsic::mips_fadd_w:
1257 case Intrinsic::mips_fadd_d:
1258 return lowerMSABinaryIntr(Op, DAG, ISD::FADD);
1259 // Don't lower mips_fcaf_[wd] since LLVM folds SETFALSE condcodes away
1260 case Intrinsic::mips_fceq_w:
1261 case Intrinsic::mips_fceq_d:
1262 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1263 Op->getOperand(2), ISD::SETOEQ);
1264 case Intrinsic::mips_fcle_w:
1265 case Intrinsic::mips_fcle_d:
1266 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1267 Op->getOperand(2), ISD::SETOLE);
1268 case Intrinsic::mips_fclt_w:
1269 case Intrinsic::mips_fclt_d:
1270 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1271 Op->getOperand(2), ISD::SETOLT);
1272 case Intrinsic::mips_fcne_w:
1273 case Intrinsic::mips_fcne_d:
1274 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1275 Op->getOperand(2), ISD::SETONE);
1276 case Intrinsic::mips_fcor_w:
1277 case Intrinsic::mips_fcor_d:
1278 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1279 Op->getOperand(2), ISD::SETO);
1280 case Intrinsic::mips_fcueq_w:
1281 case Intrinsic::mips_fcueq_d:
1282 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1283 Op->getOperand(2), ISD::SETUEQ);
1284 case Intrinsic::mips_fcule_w:
1285 case Intrinsic::mips_fcule_d:
1286 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1287 Op->getOperand(2), ISD::SETULE);
1288 case Intrinsic::mips_fcult_w:
1289 case Intrinsic::mips_fcult_d:
1290 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1291 Op->getOperand(2), ISD::SETULT);
1292 case Intrinsic::mips_fcun_w:
1293 case Intrinsic::mips_fcun_d:
1294 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1295 Op->getOperand(2), ISD::SETUO);
1296 case Intrinsic::mips_fcune_w:
1297 case Intrinsic::mips_fcune_d:
1298 return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
1299 Op->getOperand(2), ISD::SETUNE);
1300 case Intrinsic::mips_fdiv_w:
1301 case Intrinsic::mips_fdiv_d:
1302 return lowerMSABinaryIntr(Op, DAG, ISD::FDIV);
1303 case Intrinsic::mips_fill_b:
1304 case Intrinsic::mips_fill_h:
1305 case Intrinsic::mips_fill_w:
1306 return lowerMSAUnaryIntr(Op, DAG, MipsISD::VSPLAT);
1307 case Intrinsic::mips_flog2_w:
1308 case Intrinsic::mips_flog2_d:
1309 return lowerMSAUnaryIntr(Op, DAG, ISD::FLOG2);
1310 case Intrinsic::mips_fmul_w:
1311 case Intrinsic::mips_fmul_d:
1312 return lowerMSABinaryIntr(Op, DAG, ISD::FMUL);
1313 case Intrinsic::mips_frint_w:
1314 case Intrinsic::mips_frint_d:
1315 return lowerMSAUnaryIntr(Op, DAG, ISD::FRINT);
1316 case Intrinsic::mips_fsqrt_w:
1317 case Intrinsic::mips_fsqrt_d:
1318 return lowerMSAUnaryIntr(Op, DAG, ISD::FSQRT);
1319 case Intrinsic::mips_fsub_w:
1320 case Intrinsic::mips_fsub_d:
1321 return lowerMSABinaryIntr(Op, DAG, ISD::FSUB);
1322 case Intrinsic::mips_insert_b:
1323 case Intrinsic::mips_insert_h:
1324 case Intrinsic::mips_insert_w:
1325 return lowerMSAInsertIntr(Op, DAG, ISD::INSERT_VECTOR_ELT);
1326 case Intrinsic::mips_ldi_b:
1327 case Intrinsic::mips_ldi_h:
1328 case Intrinsic::mips_ldi_w:
1329 case Intrinsic::mips_ldi_d:
1330 return lowerMSAUnaryIntr(Op, DAG, MipsISD::VSPLAT);
1331 case Intrinsic::mips_max_s_b:
1332 case Intrinsic::mips_max_s_h:
1333 case Intrinsic::mips_max_s_w:
1334 case Intrinsic::mips_max_s_d:
1335 return lowerMSABinaryIntr(Op, DAG, MipsISD::VSMAX);
1336 case Intrinsic::mips_max_u_b:
1337 case Intrinsic::mips_max_u_h:
1338 case Intrinsic::mips_max_u_w:
1339 case Intrinsic::mips_max_u_d:
1340 return lowerMSABinaryIntr(Op, DAG, MipsISD::VUMAX);
1341 case Intrinsic::mips_maxi_s_b:
1342 case Intrinsic::mips_maxi_s_h:
1343 case Intrinsic::mips_maxi_s_w:
1344 case Intrinsic::mips_maxi_s_d:
1345 return lowerMSABinaryImmIntr(Op, DAG, MipsISD::VSMAX,
1346 lowerMSASplatImm(Op, 2, DAG));
1347 case Intrinsic::mips_maxi_u_b:
1348 case Intrinsic::mips_maxi_u_h:
1349 case Intrinsic::mips_maxi_u_w:
1350 case Intrinsic::mips_maxi_u_d:
1351 return lowerMSABinaryImmIntr(Op, DAG, MipsISD::VUMAX,
1352 lowerMSASplatImm(Op, 2, DAG));
1353 case Intrinsic::mips_min_s_b:
1354 case Intrinsic::mips_min_s_h:
1355 case Intrinsic::mips_min_s_w:
1356 case Intrinsic::mips_min_s_d:
1357 return lowerMSABinaryIntr(Op, DAG, MipsISD::VSMIN);
1358 case Intrinsic::mips_min_u_b:
1359 case Intrinsic::mips_min_u_h:
1360 case Intrinsic::mips_min_u_w:
1361 case Intrinsic::mips_min_u_d:
1362 return lowerMSABinaryIntr(Op, DAG, MipsISD::VUMIN);
1363 case Intrinsic::mips_mini_s_b:
1364 case Intrinsic::mips_mini_s_h:
1365 case Intrinsic::mips_mini_s_w:
1366 case Intrinsic::mips_mini_s_d:
1367 return lowerMSABinaryImmIntr(Op, DAG, MipsISD::VSMIN,
1368 lowerMSASplatImm(Op, 2, DAG));
1369 case Intrinsic::mips_mini_u_b:
1370 case Intrinsic::mips_mini_u_h:
1371 case Intrinsic::mips_mini_u_w:
1372 case Intrinsic::mips_mini_u_d:
1373 return lowerMSABinaryImmIntr(Op, DAG, MipsISD::VUMIN,
1374 lowerMSASplatImm(Op, 2, DAG));
1375 case Intrinsic::mips_mulv_b:
1376 case Intrinsic::mips_mulv_h:
1377 case Intrinsic::mips_mulv_w:
1378 case Intrinsic::mips_mulv_d:
1379 return lowerMSABinaryIntr(Op, DAG, ISD::MUL);
1380 case Intrinsic::mips_nlzc_b:
1381 case Intrinsic::mips_nlzc_h:
1382 case Intrinsic::mips_nlzc_w:
1383 case Intrinsic::mips_nlzc_d:
1384 return lowerMSAUnaryIntr(Op, DAG, ISD::CTLZ);
1385 case Intrinsic::mips_nor_v: {
1386 SDValue Res = lowerMSABinaryIntr(Op, DAG, ISD::OR);
1387 return DAG.getNOT(SDLoc(Op), Res, Res->getValueType(0));
1389 case Intrinsic::mips_or_v:
1390 return lowerMSABinaryIntr(Op, DAG, ISD::OR);
1391 case Intrinsic::mips_pcnt_b:
1392 case Intrinsic::mips_pcnt_h:
1393 case Intrinsic::mips_pcnt_w:
1394 case Intrinsic::mips_pcnt_d:
1395 return lowerMSAUnaryIntr(Op, DAG, ISD::CTPOP);
1396 case Intrinsic::mips_sll_b:
1397 case Intrinsic::mips_sll_h:
1398 case Intrinsic::mips_sll_w:
1399 case Intrinsic::mips_sll_d:
1400 return lowerMSABinaryIntr(Op, DAG, ISD::SHL);
1401 case Intrinsic::mips_slli_b:
1402 case Intrinsic::mips_slli_h:
1403 case Intrinsic::mips_slli_w:
1404 case Intrinsic::mips_slli_d:
1405 return lowerMSABinaryImmIntr(Op, DAG, ISD::SHL,
1406 lowerMSASplatImm(Op, 2, DAG));
1407 case Intrinsic::mips_sra_b:
1408 case Intrinsic::mips_sra_h:
1409 case Intrinsic::mips_sra_w:
1410 case Intrinsic::mips_sra_d:
1411 return lowerMSABinaryIntr(Op, DAG, ISD::SRA);
1412 case Intrinsic::mips_srai_b:
1413 case Intrinsic::mips_srai_h:
1414 case Intrinsic::mips_srai_w:
1415 case Intrinsic::mips_srai_d:
1416 return lowerMSABinaryImmIntr(Op, DAG, ISD::SRA,
1417 lowerMSASplatImm(Op, 2, DAG));
1418 case Intrinsic::mips_srl_b:
1419 case Intrinsic::mips_srl_h:
1420 case Intrinsic::mips_srl_w:
1421 case Intrinsic::mips_srl_d:
1422 return lowerMSABinaryIntr(Op, DAG, ISD::SRL);
1423 case Intrinsic::mips_srli_b:
1424 case Intrinsic::mips_srli_h:
1425 case Intrinsic::mips_srli_w:
1426 case Intrinsic::mips_srli_d:
1427 return lowerMSABinaryImmIntr(Op, DAG, ISD::SRL,
1428 lowerMSASplatImm(Op, 2, DAG));
1429 case Intrinsic::mips_subv_b:
1430 case Intrinsic::mips_subv_h:
1431 case Intrinsic::mips_subv_w:
1432 case Intrinsic::mips_subv_d:
1433 return lowerMSABinaryIntr(Op, DAG, ISD::SUB);
1434 case Intrinsic::mips_subvi_b:
1435 case Intrinsic::mips_subvi_h:
1436 case Intrinsic::mips_subvi_w:
1437 case Intrinsic::mips_subvi_d:
1438 return lowerMSABinaryImmIntr(Op, DAG, ISD::SUB,
1439 lowerMSASplatImm(Op, 2, DAG));
1440 case Intrinsic::mips_xor_v:
1441 return lowerMSABinaryIntr(Op, DAG, ISD::XOR);
1445 static SDValue lowerMSALoadIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
1447 SDValue ChainIn = Op->getOperand(0);
1448 SDValue Address = Op->getOperand(2);
1449 SDValue Offset = Op->getOperand(3);
1450 EVT ResTy = Op->getValueType(0);
1451 EVT PtrTy = Address->getValueType(0);
1453 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
1455 return DAG.getLoad(ResTy, DL, ChainIn, Address, MachinePointerInfo(), false,
1459 SDValue MipsSETargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
1460 SelectionDAG &DAG) const {
1461 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1465 case Intrinsic::mips_extp:
1466 return lowerDSPIntr(Op, DAG, MipsISD::EXTP);
1467 case Intrinsic::mips_extpdp:
1468 return lowerDSPIntr(Op, DAG, MipsISD::EXTPDP);
1469 case Intrinsic::mips_extr_w:
1470 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_W);
1471 case Intrinsic::mips_extr_r_w:
1472 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W);
1473 case Intrinsic::mips_extr_rs_w:
1474 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W);
1475 case Intrinsic::mips_extr_s_h:
1476 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H);
1477 case Intrinsic::mips_mthlip:
1478 return lowerDSPIntr(Op, DAG, MipsISD::MTHLIP);
1479 case Intrinsic::mips_mulsaq_s_w_ph:
1480 return lowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH);
1481 case Intrinsic::mips_maq_s_w_phl:
1482 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL);
1483 case Intrinsic::mips_maq_s_w_phr:
1484 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR);
1485 case Intrinsic::mips_maq_sa_w_phl:
1486 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL);
1487 case Intrinsic::mips_maq_sa_w_phr:
1488 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR);
1489 case Intrinsic::mips_dpaq_s_w_ph:
1490 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH);
1491 case Intrinsic::mips_dpsq_s_w_ph:
1492 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH);
1493 case Intrinsic::mips_dpaq_sa_l_w:
1494 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W);
1495 case Intrinsic::mips_dpsq_sa_l_w:
1496 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W);
1497 case Intrinsic::mips_dpaqx_s_w_ph:
1498 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH);
1499 case Intrinsic::mips_dpaqx_sa_w_ph:
1500 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH);
1501 case Intrinsic::mips_dpsqx_s_w_ph:
1502 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH);
1503 case Intrinsic::mips_dpsqx_sa_w_ph:
1504 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH);
1505 case Intrinsic::mips_ld_b:
1506 case Intrinsic::mips_ld_h:
1507 case Intrinsic::mips_ld_w:
1508 case Intrinsic::mips_ld_d:
1509 case Intrinsic::mips_ldx_b:
1510 case Intrinsic::mips_ldx_h:
1511 case Intrinsic::mips_ldx_w:
1512 case Intrinsic::mips_ldx_d:
1513 return lowerMSALoadIntr(Op, DAG, Intr);
1517 static SDValue lowerMSAStoreIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
1519 SDValue ChainIn = Op->getOperand(0);
1520 SDValue Value = Op->getOperand(2);
1521 SDValue Address = Op->getOperand(3);
1522 SDValue Offset = Op->getOperand(4);
1523 EVT PtrTy = Address->getValueType(0);
1525 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
1527 return DAG.getStore(ChainIn, DL, Value, Address, MachinePointerInfo(), false,
1531 SDValue MipsSETargetLowering::lowerINTRINSIC_VOID(SDValue Op,
1532 SelectionDAG &DAG) const {
1533 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1537 case Intrinsic::mips_st_b:
1538 case Intrinsic::mips_st_h:
1539 case Intrinsic::mips_st_w:
1540 case Intrinsic::mips_st_d:
1541 case Intrinsic::mips_stx_b:
1542 case Intrinsic::mips_stx_h:
1543 case Intrinsic::mips_stx_w:
1544 case Intrinsic::mips_stx_d:
1545 return lowerMSAStoreIntr(Op, DAG, Intr);
1549 /// \brief Check if the given BuildVectorSDNode is a splat.
1550 /// This method currently relies on DAG nodes being reused when equivalent,
1551 /// so it's possible for this to return false even when isConstantSplat returns
1553 static bool isSplatVector(const BuildVectorSDNode *N) {
1554 unsigned int nOps = N->getNumOperands();
1555 assert(nOps > 1 && "isSplat has 0 or 1 sized build vector");
1557 SDValue Operand0 = N->getOperand(0);
1559 for (unsigned int i = 1; i < nOps; ++i) {
1560 if (N->getOperand(i) != Operand0)
1567 // Lower ISD::EXTRACT_VECTOR_ELT into MipsISD::VEXTRACT_SEXT_ELT.
1569 // The non-value bits resulting from ISD::EXTRACT_VECTOR_ELT are undefined. We
1570 // choose to sign-extend but we could have equally chosen zero-extend. The
1571 // DAGCombiner will fold any sign/zero extension of the ISD::EXTRACT_VECTOR_ELT
1572 // result into this node later (possibly changing it to a zero-extend in the
1574 SDValue MipsSETargetLowering::
1575 lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
1577 EVT ResTy = Op->getValueType(0);
1578 SDValue Op0 = Op->getOperand(0);
1579 SDValue Op1 = Op->getOperand(1);
1580 EVT EltTy = Op0->getValueType(0).getVectorElementType();
1581 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, DL, ResTy, Op0, Op1,
1582 DAG.getValueType(EltTy));
1585 // Lowers ISD::BUILD_VECTOR into appropriate SelectionDAG nodes for the
1588 // Lowers according to the following rules:
1589 // - Vectors of 128-bits may be legal subject to the other rules. Other sizes
1591 // - Non-constant splats are legal and are lowered to MipsISD::VSPLAT.
1592 // - Constant splats with an element size of 32-bits or less are legal and are
1593 // lowered to MipsISD::VSPLAT.
1594 // - Constant splats with an element size of 64-bits but whose value would fit
1595 // within a 10 bit immediate are legal and are lowered to MipsISD::VSPLATD.
1596 // - All other ISD::BUILD_VECTORS are not legal
1597 SDValue MipsSETargetLowering::lowerBUILD_VECTOR(SDValue Op,
1598 SelectionDAG &DAG) const {
1599 BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op);
1600 EVT ResTy = Op->getValueType(0);
1602 APInt SplatValue, SplatUndef;
1603 unsigned SplatBitSize;
1606 if (!Subtarget->hasMSA() || !ResTy.is128BitVector())
1609 if (Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
1611 !Subtarget->isLittle())) {
1614 EVT ConstTy = MVT::i32;
1615 unsigned SplatOp = MipsISD::VSPLAT;
1617 switch (SplatBitSize) {
1621 TmpVecTy = MVT::v2i64;
1623 // i64 is an illegal type on Mips32, but if it the constant fits into a
1624 // signed 10-bit value then we can still handle it using VSPLATD and an
1628 else if (isInt<10>(SplatValue.getSExtValue())) {
1629 SplatValue = SplatValue.trunc(32);
1630 SplatOp = MipsISD::VSPLATD;
1635 TmpVecTy = MVT::v4i32;
1638 TmpVecTy = MVT::v8i16;
1639 SplatValue = SplatValue.sext(32);
1642 TmpVecTy = MVT::v16i8;
1643 SplatValue = SplatValue.sext(32);
1647 Result = DAG.getNode(SplatOp, DL, TmpVecTy,
1648 DAG.getConstant(SplatValue, ConstTy));
1649 if (ResTy != Result.getValueType())
1650 Result = DAG.getNode(ISD::BITCAST, DL, ResTy, Result);
1654 else if (isSplatVector(Node))
1655 return DAG.getNode(MipsISD::VSPLAT, DL, ResTy, Op->getOperand(0));
1660 MachineBasicBlock * MipsSETargetLowering::
1661 emitBPOSGE32(MachineInstr *MI, MachineBasicBlock *BB) const{
1663 // bposge32_pseudo $vr0
1673 // $vr0 = phi($vr2, $fbb, $vr1, $tbb)
1675 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
1676 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1677 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
1678 DebugLoc DL = MI->getDebugLoc();
1679 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1680 MachineFunction::iterator It = llvm::next(MachineFunction::iterator(BB));
1681 MachineFunction *F = BB->getParent();
1682 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
1683 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
1684 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
1687 F->insert(It, Sink);
1689 // Transfer the remainder of BB and its successor edges to Sink.
1690 Sink->splice(Sink->begin(), BB, llvm::next(MachineBasicBlock::iterator(MI)),
1692 Sink->transferSuccessorsAndUpdatePHIs(BB);
1695 BB->addSuccessor(FBB);
1696 BB->addSuccessor(TBB);
1697 FBB->addSuccessor(Sink);
1698 TBB->addSuccessor(Sink);
1700 // Insert the real bposge32 instruction to $BB.
1701 BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB);
1704 unsigned VR2 = RegInfo.createVirtualRegister(RC);
1705 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2)
1706 .addReg(Mips::ZERO).addImm(0);
1707 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
1710 unsigned VR1 = RegInfo.createVirtualRegister(RC);
1711 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1)
1712 .addReg(Mips::ZERO).addImm(1);
1714 // Insert phi function to $Sink.
1715 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
1716 MI->getOperand(0).getReg())
1717 .addReg(VR2).addMBB(FBB).addReg(VR1).addMBB(TBB);
1719 MI->eraseFromParent(); // The pseudo instruction is gone now.
1723 MachineBasicBlock * MipsSETargetLowering::
1724 emitMSACBranchPseudo(MachineInstr *MI, MachineBasicBlock *BB,
1725 unsigned BranchOp) const{
1727 // vany_nonzero $rd, $ws
1738 // $rd = phi($rd1, $fbb, $rd2, $tbb)
1740 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
1741 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1742 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
1743 DebugLoc DL = MI->getDebugLoc();
1744 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1745 MachineFunction::iterator It = llvm::next(MachineFunction::iterator(BB));
1746 MachineFunction *F = BB->getParent();
1747 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
1748 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
1749 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
1752 F->insert(It, Sink);
1754 // Transfer the remainder of BB and its successor edges to Sink.
1755 Sink->splice(Sink->begin(), BB, llvm::next(MachineBasicBlock::iterator(MI)),
1757 Sink->transferSuccessorsAndUpdatePHIs(BB);
1760 BB->addSuccessor(FBB);
1761 BB->addSuccessor(TBB);
1762 FBB->addSuccessor(Sink);
1763 TBB->addSuccessor(Sink);
1765 // Insert the real bnz.b instruction to $BB.
1766 BuildMI(BB, DL, TII->get(BranchOp))
1767 .addReg(MI->getOperand(1).getReg())
1771 unsigned RD1 = RegInfo.createVirtualRegister(RC);
1772 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), RD1)
1773 .addReg(Mips::ZERO).addImm(0);
1774 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
1777 unsigned RD2 = RegInfo.createVirtualRegister(RC);
1778 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), RD2)
1779 .addReg(Mips::ZERO).addImm(1);
1781 // Insert phi function to $Sink.
1782 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
1783 MI->getOperand(0).getReg())
1784 .addReg(RD1).addMBB(FBB).addReg(RD2).addMBB(TBB);
1786 MI->eraseFromParent(); // The pseudo instruction is gone now.