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/Support/Debug.h"
21 #include "llvm/Support/raw_ostream.h"
22 #include "llvm/Target/TargetInstrInfo.h"
26 #define DEBUG_TYPE "mips-isel"
29 EnableMipsTailCalls("enable-mips-tail-calls", cl::Hidden,
30 cl::desc("MIPS: Enable tail calls."), cl::init(false));
32 static cl::opt<bool> NoDPLoadStore("mno-ldc1-sdc1", cl::init(false),
33 cl::desc("Expand double precision loads and "
34 "stores to their single precision "
37 MipsSETargetLowering::MipsSETargetLowering(MipsTargetMachine &TM,
38 const MipsSubtarget &STI)
39 : MipsTargetLowering(TM, STI) {
40 // Set up the register classes
41 addRegisterClass(MVT::i32, &Mips::GPR32RegClass);
43 if (Subtarget.isGP64bit())
44 addRegisterClass(MVT::i64, &Mips::GPR64RegClass);
46 if (Subtarget.hasDSP() || Subtarget.hasMSA()) {
47 // Expand all truncating stores and extending loads.
48 unsigned FirstVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
49 unsigned LastVT = (unsigned)MVT::LAST_VECTOR_VALUETYPE;
51 for (unsigned VT0 = FirstVT; VT0 <= LastVT; ++VT0) {
52 for (unsigned VT1 = FirstVT; VT1 <= LastVT; ++VT1)
53 setTruncStoreAction((MVT::SimpleValueType)VT0,
54 (MVT::SimpleValueType)VT1, Expand);
56 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT0, Expand);
57 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT0, Expand);
58 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT0, Expand);
62 if (Subtarget.hasDSP()) {
63 MVT::SimpleValueType VecTys[2] = {MVT::v2i16, MVT::v4i8};
65 for (unsigned i = 0; i < array_lengthof(VecTys); ++i) {
66 addRegisterClass(VecTys[i], &Mips::DSPRRegClass);
68 // Expand all builtin opcodes.
69 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
70 setOperationAction(Opc, VecTys[i], Expand);
72 setOperationAction(ISD::ADD, VecTys[i], Legal);
73 setOperationAction(ISD::SUB, VecTys[i], Legal);
74 setOperationAction(ISD::LOAD, VecTys[i], Legal);
75 setOperationAction(ISD::STORE, VecTys[i], Legal);
76 setOperationAction(ISD::BITCAST, VecTys[i], Legal);
79 setTargetDAGCombine(ISD::SHL);
80 setTargetDAGCombine(ISD::SRA);
81 setTargetDAGCombine(ISD::SRL);
82 setTargetDAGCombine(ISD::SETCC);
83 setTargetDAGCombine(ISD::VSELECT);
86 if (Subtarget.hasDSPR2())
87 setOperationAction(ISD::MUL, MVT::v2i16, Legal);
89 if (Subtarget.hasMSA()) {
90 addMSAIntType(MVT::v16i8, &Mips::MSA128BRegClass);
91 addMSAIntType(MVT::v8i16, &Mips::MSA128HRegClass);
92 addMSAIntType(MVT::v4i32, &Mips::MSA128WRegClass);
93 addMSAIntType(MVT::v2i64, &Mips::MSA128DRegClass);
94 addMSAFloatType(MVT::v8f16, &Mips::MSA128HRegClass);
95 addMSAFloatType(MVT::v4f32, &Mips::MSA128WRegClass);
96 addMSAFloatType(MVT::v2f64, &Mips::MSA128DRegClass);
98 setTargetDAGCombine(ISD::AND);
99 setTargetDAGCombine(ISD::OR);
100 setTargetDAGCombine(ISD::SRA);
101 setTargetDAGCombine(ISD::VSELECT);
102 setTargetDAGCombine(ISD::XOR);
105 if (!Subtarget.abiUsesSoftFloat()) {
106 addRegisterClass(MVT::f32, &Mips::FGR32RegClass);
108 // When dealing with single precision only, use libcalls
109 if (!Subtarget.isSingleFloat()) {
110 if (Subtarget.isFP64bit())
111 addRegisterClass(MVT::f64, &Mips::FGR64RegClass);
113 addRegisterClass(MVT::f64, &Mips::AFGR64RegClass);
117 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Custom);
118 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Custom);
119 setOperationAction(ISD::MULHS, MVT::i32, Custom);
120 setOperationAction(ISD::MULHU, MVT::i32, Custom);
122 if (Subtarget.hasCnMips())
123 setOperationAction(ISD::MUL, MVT::i64, Legal);
124 else if (Subtarget.isGP64bit())
125 setOperationAction(ISD::MUL, MVT::i64, Custom);
127 if (Subtarget.isGP64bit()) {
128 setOperationAction(ISD::MULHS, MVT::i64, Custom);
129 setOperationAction(ISD::MULHU, MVT::i64, Custom);
132 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
133 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
135 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
136 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
137 setOperationAction(ISD::SDIVREM, MVT::i64, Custom);
138 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
139 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
140 setOperationAction(ISD::LOAD, MVT::i32, Custom);
141 setOperationAction(ISD::STORE, MVT::i32, Custom);
143 setTargetDAGCombine(ISD::ADDE);
144 setTargetDAGCombine(ISD::SUBE);
145 setTargetDAGCombine(ISD::MUL);
147 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
148 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
149 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
152 setOperationAction(ISD::LOAD, MVT::f64, Custom);
153 setOperationAction(ISD::STORE, MVT::f64, Custom);
156 if (Subtarget.hasMips32r6()) {
157 // MIPS32r6 replaces the accumulator-based multiplies with a three register
159 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
160 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
161 setOperationAction(ISD::MUL, MVT::i32, Legal);
162 setOperationAction(ISD::MULHS, MVT::i32, Legal);
163 setOperationAction(ISD::MULHU, MVT::i32, Legal);
165 // MIPS32r6 replaces the accumulator-based division/remainder with separate
166 // three register division and remainder instructions.
167 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
168 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
169 setOperationAction(ISD::SDIV, MVT::i32, Legal);
170 setOperationAction(ISD::UDIV, MVT::i32, Legal);
171 setOperationAction(ISD::SREM, MVT::i32, Legal);
172 setOperationAction(ISD::UREM, MVT::i32, Legal);
174 // MIPS32r6 replaces conditional moves with an equivalent that removes the
175 // need for three GPR read ports.
176 setOperationAction(ISD::SETCC, MVT::i32, Legal);
177 setOperationAction(ISD::SELECT, MVT::i32, Legal);
178 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
180 setOperationAction(ISD::SETCC, MVT::f32, Legal);
181 setOperationAction(ISD::SELECT, MVT::f32, Legal);
182 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
184 assert(Subtarget.isFP64bit() && "FR=1 is required for MIPS32r6");
185 setOperationAction(ISD::SETCC, MVT::f64, Legal);
186 setOperationAction(ISD::SELECT, MVT::f64, Legal);
187 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
189 setOperationAction(ISD::BRCOND, MVT::Other, Legal);
191 // Floating point > and >= are supported via < and <=
192 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
193 setCondCodeAction(ISD::SETOGT, MVT::f32, Expand);
194 setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
195 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
197 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
198 setCondCodeAction(ISD::SETOGT, MVT::f64, Expand);
199 setCondCodeAction(ISD::SETUGE, MVT::f64, Expand);
200 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
203 if (Subtarget.hasMips64r6()) {
204 // MIPS64r6 replaces the accumulator-based multiplies with a three register
206 setOperationAction(ISD::MUL, MVT::i64, Legal);
207 setOperationAction(ISD::MULHS, MVT::i64, Legal);
208 setOperationAction(ISD::MULHU, MVT::i64, Legal);
210 // MIPS32r6 replaces the accumulator-based division/remainder with separate
211 // three register division and remainder instructions.
212 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
213 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
214 setOperationAction(ISD::SDIV, MVT::i64, Legal);
215 setOperationAction(ISD::UDIV, MVT::i64, Legal);
216 setOperationAction(ISD::SREM, MVT::i64, Legal);
217 setOperationAction(ISD::UREM, MVT::i64, Legal);
219 // MIPS64r6 replaces conditional moves with an equivalent that removes the
220 // need for three GPR read ports.
221 setOperationAction(ISD::SETCC, MVT::i64, Legal);
222 setOperationAction(ISD::SELECT, MVT::i64, Legal);
223 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
226 computeRegisterProperties();
229 const MipsTargetLowering *
230 llvm::createMipsSETargetLowering(MipsTargetMachine &TM,
231 const MipsSubtarget &STI) {
232 return new MipsSETargetLowering(TM, STI);
235 const TargetRegisterClass *
236 MipsSETargetLowering::getRepRegClassFor(MVT VT) const {
237 if (VT == MVT::Untyped)
238 return Subtarget.hasDSP() ? &Mips::ACC64DSPRegClass : &Mips::ACC64RegClass;
240 return TargetLowering::getRepRegClassFor(VT);
243 // Enable MSA support for the given integer type and Register class.
244 void MipsSETargetLowering::
245 addMSAIntType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
246 addRegisterClass(Ty, RC);
248 // Expand all builtin opcodes.
249 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
250 setOperationAction(Opc, Ty, Expand);
252 setOperationAction(ISD::BITCAST, Ty, Legal);
253 setOperationAction(ISD::LOAD, Ty, Legal);
254 setOperationAction(ISD::STORE, Ty, Legal);
255 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Custom);
256 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
257 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
259 setOperationAction(ISD::ADD, Ty, Legal);
260 setOperationAction(ISD::AND, Ty, Legal);
261 setOperationAction(ISD::CTLZ, Ty, Legal);
262 setOperationAction(ISD::CTPOP, Ty, Legal);
263 setOperationAction(ISD::MUL, Ty, Legal);
264 setOperationAction(ISD::OR, Ty, Legal);
265 setOperationAction(ISD::SDIV, Ty, Legal);
266 setOperationAction(ISD::SREM, Ty, Legal);
267 setOperationAction(ISD::SHL, Ty, Legal);
268 setOperationAction(ISD::SRA, Ty, Legal);
269 setOperationAction(ISD::SRL, Ty, Legal);
270 setOperationAction(ISD::SUB, Ty, Legal);
271 setOperationAction(ISD::UDIV, Ty, Legal);
272 setOperationAction(ISD::UREM, Ty, Legal);
273 setOperationAction(ISD::VECTOR_SHUFFLE, Ty, Custom);
274 setOperationAction(ISD::VSELECT, Ty, Legal);
275 setOperationAction(ISD::XOR, Ty, Legal);
277 if (Ty == MVT::v4i32 || Ty == MVT::v2i64) {
278 setOperationAction(ISD::FP_TO_SINT, Ty, Legal);
279 setOperationAction(ISD::FP_TO_UINT, Ty, Legal);
280 setOperationAction(ISD::SINT_TO_FP, Ty, Legal);
281 setOperationAction(ISD::UINT_TO_FP, Ty, Legal);
284 setOperationAction(ISD::SETCC, Ty, Legal);
285 setCondCodeAction(ISD::SETNE, Ty, Expand);
286 setCondCodeAction(ISD::SETGE, Ty, Expand);
287 setCondCodeAction(ISD::SETGT, Ty, Expand);
288 setCondCodeAction(ISD::SETUGE, Ty, Expand);
289 setCondCodeAction(ISD::SETUGT, Ty, Expand);
292 // Enable MSA support for the given floating-point type and Register class.
293 void MipsSETargetLowering::
294 addMSAFloatType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
295 addRegisterClass(Ty, RC);
297 // Expand all builtin opcodes.
298 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
299 setOperationAction(Opc, Ty, Expand);
301 setOperationAction(ISD::LOAD, Ty, Legal);
302 setOperationAction(ISD::STORE, Ty, Legal);
303 setOperationAction(ISD::BITCAST, Ty, Legal);
304 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Legal);
305 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
306 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
308 if (Ty != MVT::v8f16) {
309 setOperationAction(ISD::FABS, Ty, Legal);
310 setOperationAction(ISD::FADD, Ty, Legal);
311 setOperationAction(ISD::FDIV, Ty, Legal);
312 setOperationAction(ISD::FEXP2, Ty, Legal);
313 setOperationAction(ISD::FLOG2, Ty, Legal);
314 setOperationAction(ISD::FMA, Ty, Legal);
315 setOperationAction(ISD::FMUL, Ty, Legal);
316 setOperationAction(ISD::FRINT, Ty, Legal);
317 setOperationAction(ISD::FSQRT, Ty, Legal);
318 setOperationAction(ISD::FSUB, Ty, Legal);
319 setOperationAction(ISD::VSELECT, Ty, Legal);
321 setOperationAction(ISD::SETCC, Ty, Legal);
322 setCondCodeAction(ISD::SETOGE, Ty, Expand);
323 setCondCodeAction(ISD::SETOGT, Ty, Expand);
324 setCondCodeAction(ISD::SETUGE, Ty, Expand);
325 setCondCodeAction(ISD::SETUGT, Ty, Expand);
326 setCondCodeAction(ISD::SETGE, Ty, Expand);
327 setCondCodeAction(ISD::SETGT, Ty, Expand);
332 MipsSETargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
336 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
338 if (Subtarget.systemSupportsUnalignedAccess()) {
339 // MIPS32r6/MIPS64r6 is required to support unaligned access. It's
340 // implementation defined whether this is handled by hardware, software, or
341 // a hybrid of the two but it's expected that most implementations will
342 // handle the majority of cases in hardware.
359 SDValue MipsSETargetLowering::LowerOperation(SDValue Op,
360 SelectionDAG &DAG) const {
361 switch(Op.getOpcode()) {
362 case ISD::LOAD: return lowerLOAD(Op, DAG);
363 case ISD::STORE: return lowerSTORE(Op, DAG);
364 case ISD::SMUL_LOHI: return lowerMulDiv(Op, MipsISD::Mult, true, true, DAG);
365 case ISD::UMUL_LOHI: return lowerMulDiv(Op, MipsISD::Multu, true, true, DAG);
366 case ISD::MULHS: return lowerMulDiv(Op, MipsISD::Mult, false, true, DAG);
367 case ISD::MULHU: return lowerMulDiv(Op, MipsISD::Multu, false, true, DAG);
368 case ISD::MUL: return lowerMulDiv(Op, MipsISD::Mult, true, false, DAG);
369 case ISD::SDIVREM: return lowerMulDiv(Op, MipsISD::DivRem, true, true, DAG);
370 case ISD::UDIVREM: return lowerMulDiv(Op, MipsISD::DivRemU, true, true,
372 case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG);
373 case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG);
374 case ISD::INTRINSIC_VOID: return lowerINTRINSIC_VOID(Op, DAG);
375 case ISD::EXTRACT_VECTOR_ELT: return lowerEXTRACT_VECTOR_ELT(Op, DAG);
376 case ISD::BUILD_VECTOR: return lowerBUILD_VECTOR(Op, DAG);
377 case ISD::VECTOR_SHUFFLE: return lowerVECTOR_SHUFFLE(Op, DAG);
380 return MipsTargetLowering::LowerOperation(Op, DAG);
384 // Transforms a subgraph in CurDAG if the following pattern is found:
385 // (addc multLo, Lo0), (adde multHi, Hi0),
387 // multHi/Lo: product of multiplication
388 // Lo0: initial value of Lo register
389 // Hi0: initial value of Hi register
390 // Return true if pattern matching was successful.
391 static bool selectMADD(SDNode *ADDENode, SelectionDAG *CurDAG) {
392 // ADDENode's second operand must be a flag output of an ADDC node in order
393 // for the matching to be successful.
394 SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
396 if (ADDCNode->getOpcode() != ISD::ADDC)
399 SDValue MultHi = ADDENode->getOperand(0);
400 SDValue MultLo = ADDCNode->getOperand(0);
401 SDNode *MultNode = MultHi.getNode();
402 unsigned MultOpc = MultHi.getOpcode();
404 // MultHi and MultLo must be generated by the same node,
405 if (MultLo.getNode() != MultNode)
408 // and it must be a multiplication.
409 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
412 // MultLo amd MultHi must be the first and second output of MultNode
414 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
417 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
418 // of the values of MultNode, in which case MultNode will be removed in later
420 // If there exist users other than ADDENode or ADDCNode, this function returns
421 // here, which will result in MultNode being mapped to a single MULT
422 // instruction node rather than a pair of MULT and MADD instructions being
424 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
429 // Initialize accumulator.
430 SDValue ACCIn = CurDAG->getNode(MipsISD::MTLOHI, DL, MVT::Untyped,
431 ADDCNode->getOperand(1),
432 ADDENode->getOperand(1));
434 // create MipsMAdd(u) node
435 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
437 SDValue MAdd = CurDAG->getNode(MultOpc, DL, MVT::Untyped,
438 MultNode->getOperand(0),// Factor 0
439 MultNode->getOperand(1),// Factor 1
442 // replace uses of adde and addc here
443 if (!SDValue(ADDCNode, 0).use_empty()) {
444 SDValue LoOut = CurDAG->getNode(MipsISD::MFLO, DL, MVT::i32, MAdd);
445 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), LoOut);
447 if (!SDValue(ADDENode, 0).use_empty()) {
448 SDValue HiOut = CurDAG->getNode(MipsISD::MFHI, DL, MVT::i32, MAdd);
449 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), HiOut);
456 // Transforms a subgraph in CurDAG if the following pattern is found:
457 // (addc Lo0, multLo), (sube Hi0, multHi),
459 // multHi/Lo: product of multiplication
460 // Lo0: initial value of Lo register
461 // Hi0: initial value of Hi register
462 // Return true if pattern matching was successful.
463 static bool selectMSUB(SDNode *SUBENode, SelectionDAG *CurDAG) {
464 // SUBENode's second operand must be a flag output of an SUBC node in order
465 // for the matching to be successful.
466 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
468 if (SUBCNode->getOpcode() != ISD::SUBC)
471 SDValue MultHi = SUBENode->getOperand(1);
472 SDValue MultLo = SUBCNode->getOperand(1);
473 SDNode *MultNode = MultHi.getNode();
474 unsigned MultOpc = MultHi.getOpcode();
476 // MultHi and MultLo must be generated by the same node,
477 if (MultLo.getNode() != MultNode)
480 // and it must be a multiplication.
481 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
484 // MultLo amd MultHi must be the first and second output of MultNode
486 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
489 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
490 // of the values of MultNode, in which case MultNode will be removed in later
492 // If there exist users other than SUBENode or SUBCNode, this function returns
493 // here, which will result in MultNode being mapped to a single MULT
494 // instruction node rather than a pair of MULT and MSUB instructions being
496 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
501 // Initialize accumulator.
502 SDValue ACCIn = CurDAG->getNode(MipsISD::MTLOHI, DL, MVT::Untyped,
503 SUBCNode->getOperand(0),
504 SUBENode->getOperand(0));
506 // create MipsSub(u) node
507 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
509 SDValue MSub = CurDAG->getNode(MultOpc, DL, MVT::Glue,
510 MultNode->getOperand(0),// Factor 0
511 MultNode->getOperand(1),// Factor 1
514 // replace uses of sube and subc here
515 if (!SDValue(SUBCNode, 0).use_empty()) {
516 SDValue LoOut = CurDAG->getNode(MipsISD::MFLO, DL, MVT::i32, MSub);
517 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), LoOut);
519 if (!SDValue(SUBENode, 0).use_empty()) {
520 SDValue HiOut = CurDAG->getNode(MipsISD::MFHI, DL, MVT::i32, MSub);
521 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), HiOut);
527 static SDValue performADDECombine(SDNode *N, SelectionDAG &DAG,
528 TargetLowering::DAGCombinerInfo &DCI,
529 const MipsSubtarget &Subtarget) {
530 if (DCI.isBeforeLegalize())
533 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() &&
534 N->getValueType(0) == MVT::i32 && selectMADD(N, &DAG))
535 return SDValue(N, 0);
540 // Fold zero extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT
542 // Performs the following transformations:
543 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to zero extension if its
544 // sign/zero-extension is completely overwritten by the new one performed by
546 // - Removes redundant zero extensions performed by an ISD::AND.
547 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
548 TargetLowering::DAGCombinerInfo &DCI,
549 const MipsSubtarget &Subtarget) {
550 if (!Subtarget.hasMSA())
553 SDValue Op0 = N->getOperand(0);
554 SDValue Op1 = N->getOperand(1);
555 unsigned Op0Opcode = Op0->getOpcode();
557 // (and (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d)
558 // where $d + 1 == 2^n and n == 32
559 // or $d + 1 == 2^n and n <= 32 and ZExt
560 // -> (MipsVExtractZExt $a, $b, $c)
561 if (Op0Opcode == MipsISD::VEXTRACT_SEXT_ELT ||
562 Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT) {
563 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(Op1);
568 int32_t Log2IfPositive = (Mask->getAPIntValue() + 1).exactLogBase2();
570 if (Log2IfPositive <= 0)
571 return SDValue(); // Mask+1 is not a power of 2
573 SDValue Op0Op2 = Op0->getOperand(2);
574 EVT ExtendTy = cast<VTSDNode>(Op0Op2)->getVT();
575 unsigned ExtendTySize = ExtendTy.getSizeInBits();
576 unsigned Log2 = Log2IfPositive;
578 if ((Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT && Log2 >= ExtendTySize) ||
579 Log2 == ExtendTySize) {
580 SDValue Ops[] = { Op0->getOperand(0), Op0->getOperand(1), Op0Op2 };
581 return DAG.getNode(MipsISD::VEXTRACT_ZEXT_ELT, SDLoc(Op0),
583 makeArrayRef(Ops, Op0->getNumOperands()));
590 // Determine if the specified node is a constant vector splat.
592 // Returns true and sets Imm if:
593 // * N is a ISD::BUILD_VECTOR representing a constant splat
595 // This function is quite similar to MipsSEDAGToDAGISel::selectVSplat. The
596 // differences are that it assumes the MSA has already been checked and the
597 // arbitrary requirement for a maximum of 32-bit integers isn't applied (and
598 // must not be in order for binsri.d to be selectable).
599 static bool isVSplat(SDValue N, APInt &Imm, bool IsLittleEndian) {
600 BuildVectorSDNode *Node = dyn_cast<BuildVectorSDNode>(N.getNode());
605 APInt SplatValue, SplatUndef;
606 unsigned SplatBitSize;
609 if (!Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
618 // Test whether the given node is an all-ones build_vector.
619 static bool isVectorAllOnes(SDValue N) {
620 // Look through bitcasts. Endianness doesn't matter because we are looking
621 // for an all-ones value.
622 if (N->getOpcode() == ISD::BITCAST)
623 N = N->getOperand(0);
625 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
630 APInt SplatValue, SplatUndef;
631 unsigned SplatBitSize;
634 // Endianness doesn't matter in this context because we are looking for
635 // an all-ones value.
636 if (BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs))
637 return SplatValue.isAllOnesValue();
642 // Test whether N is the bitwise inverse of OfNode.
643 static bool isBitwiseInverse(SDValue N, SDValue OfNode) {
644 if (N->getOpcode() != ISD::XOR)
647 if (isVectorAllOnes(N->getOperand(0)))
648 return N->getOperand(1) == OfNode;
650 if (isVectorAllOnes(N->getOperand(1)))
651 return N->getOperand(0) == OfNode;
656 // Perform combines where ISD::OR is the root node.
658 // Performs the following transformations:
659 // - (or (and $a, $mask), (and $b, $inv_mask)) => (vselect $mask, $a, $b)
660 // where $inv_mask is the bitwise inverse of $mask and the 'or' has a 128-bit
662 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
663 TargetLowering::DAGCombinerInfo &DCI,
664 const MipsSubtarget &Subtarget) {
665 if (!Subtarget.hasMSA())
668 EVT Ty = N->getValueType(0);
670 if (!Ty.is128BitVector())
673 SDValue Op0 = N->getOperand(0);
674 SDValue Op1 = N->getOperand(1);
676 if (Op0->getOpcode() == ISD::AND && Op1->getOpcode() == ISD::AND) {
677 SDValue Op0Op0 = Op0->getOperand(0);
678 SDValue Op0Op1 = Op0->getOperand(1);
679 SDValue Op1Op0 = Op1->getOperand(0);
680 SDValue Op1Op1 = Op1->getOperand(1);
681 bool IsLittleEndian = !Subtarget.isLittle();
683 SDValue IfSet, IfClr, Cond;
684 bool IsConstantMask = false;
687 // If Op0Op0 is an appropriate mask, try to find it's inverse in either
688 // Op1Op0, or Op1Op1. Keep track of the Cond, IfSet, and IfClr nodes, while
690 // IfClr will be set if we find a valid match.
691 if (isVSplat(Op0Op0, Mask, IsLittleEndian)) {
695 if (isVSplat(Op1Op0, InvMask, IsLittleEndian) &&
696 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
698 else if (isVSplat(Op1Op1, InvMask, IsLittleEndian) &&
699 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
702 IsConstantMask = true;
705 // If IfClr is not yet set, and Op0Op1 is an appropriate mask, try the same
706 // thing again using this mask.
707 // IfClr will be set if we find a valid match.
708 if (!IfClr.getNode() && isVSplat(Op0Op1, Mask, IsLittleEndian)) {
712 if (isVSplat(Op1Op0, InvMask, IsLittleEndian) &&
713 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
715 else if (isVSplat(Op1Op1, InvMask, IsLittleEndian) &&
716 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
719 IsConstantMask = true;
722 // If IfClr is not yet set, try looking for a non-constant match.
723 // IfClr will be set if we find a valid match amongst the eight
725 if (!IfClr.getNode()) {
726 if (isBitwiseInverse(Op0Op0, Op1Op0)) {
730 } else if (isBitwiseInverse(Op0Op1, Op1Op0)) {
734 } else if (isBitwiseInverse(Op0Op0, Op1Op1)) {
738 } else if (isBitwiseInverse(Op0Op1, Op1Op1)) {
742 } else if (isBitwiseInverse(Op1Op0, Op0Op0)) {
746 } else if (isBitwiseInverse(Op1Op1, Op0Op0)) {
750 } else if (isBitwiseInverse(Op1Op0, Op0Op1)) {
754 } else if (isBitwiseInverse(Op1Op1, Op0Op1)) {
761 // At this point, IfClr will be set if we have a valid match.
762 if (!IfClr.getNode())
765 assert(Cond.getNode() && IfSet.getNode());
767 // Fold degenerate cases.
768 if (IsConstantMask) {
769 if (Mask.isAllOnesValue())
775 // Transform the DAG into an equivalent VSELECT.
776 return DAG.getNode(ISD::VSELECT, SDLoc(N), Ty, Cond, IfSet, IfClr);
782 static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG,
783 TargetLowering::DAGCombinerInfo &DCI,
784 const MipsSubtarget &Subtarget) {
785 if (DCI.isBeforeLegalize())
788 if (Subtarget.hasMips32() && N->getValueType(0) == MVT::i32 &&
790 return SDValue(N, 0);
795 static SDValue genConstMult(SDValue X, uint64_t C, SDLoc DL, EVT VT,
796 EVT ShiftTy, SelectionDAG &DAG) {
797 // Clear the upper (64 - VT.sizeInBits) bits.
798 C &= ((uint64_t)-1) >> (64 - VT.getSizeInBits());
802 return DAG.getConstant(0, VT);
808 // If c is power of 2, return (shl x, log2(c)).
809 if (isPowerOf2_64(C))
810 return DAG.getNode(ISD::SHL, DL, VT, X,
811 DAG.getConstant(Log2_64(C), ShiftTy));
813 unsigned Log2Ceil = Log2_64_Ceil(C);
814 uint64_t Floor = 1LL << Log2_64(C);
815 uint64_t Ceil = Log2Ceil == 64 ? 0LL : 1LL << Log2Ceil;
817 // If |c - floor_c| <= |c - ceil_c|,
818 // where floor_c = pow(2, floor(log2(c))) and ceil_c = pow(2, ceil(log2(c))),
819 // return (add constMult(x, floor_c), constMult(x, c - floor_c)).
820 if (C - Floor <= Ceil - C) {
821 SDValue Op0 = genConstMult(X, Floor, DL, VT, ShiftTy, DAG);
822 SDValue Op1 = genConstMult(X, C - Floor, DL, VT, ShiftTy, DAG);
823 return DAG.getNode(ISD::ADD, DL, VT, Op0, Op1);
826 // If |c - floor_c| > |c - ceil_c|,
827 // return (sub constMult(x, ceil_c), constMult(x, ceil_c - c)).
828 SDValue Op0 = genConstMult(X, Ceil, DL, VT, ShiftTy, DAG);
829 SDValue Op1 = genConstMult(X, Ceil - C, DL, VT, ShiftTy, DAG);
830 return DAG.getNode(ISD::SUB, DL, VT, Op0, Op1);
833 static SDValue performMULCombine(SDNode *N, SelectionDAG &DAG,
834 const TargetLowering::DAGCombinerInfo &DCI,
835 const MipsSETargetLowering *TL) {
836 EVT VT = N->getValueType(0);
838 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
840 return genConstMult(N->getOperand(0), C->getZExtValue(), SDLoc(N),
841 VT, TL->getScalarShiftAmountTy(VT), DAG);
843 return SDValue(N, 0);
846 static SDValue performDSPShiftCombine(unsigned Opc, SDNode *N, EVT Ty,
848 const MipsSubtarget &Subtarget) {
849 // See if this is a vector splat immediate node.
850 APInt SplatValue, SplatUndef;
851 unsigned SplatBitSize;
853 unsigned EltSize = Ty.getVectorElementType().getSizeInBits();
854 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
856 if (!Subtarget.hasDSP())
860 !BV->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
861 EltSize, !Subtarget.isLittle()) ||
862 (SplatBitSize != EltSize) ||
863 (SplatValue.getZExtValue() >= EltSize))
866 return DAG.getNode(Opc, SDLoc(N), Ty, N->getOperand(0),
867 DAG.getConstant(SplatValue.getZExtValue(), MVT::i32));
870 static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
871 TargetLowering::DAGCombinerInfo &DCI,
872 const MipsSubtarget &Subtarget) {
873 EVT Ty = N->getValueType(0);
875 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
878 return performDSPShiftCombine(MipsISD::SHLL_DSP, N, Ty, DAG, Subtarget);
881 // Fold sign-extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT for MSA and fold
882 // constant splats into MipsISD::SHRA_DSP for DSPr2.
884 // Performs the following transformations:
885 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to sign extension if its
886 // sign/zero-extension is completely overwritten by the new one performed by
887 // the ISD::SRA and ISD::SHL nodes.
888 // - Removes redundant sign extensions performed by an ISD::SRA and ISD::SHL
891 // See performDSPShiftCombine for more information about the transformation
893 static SDValue performSRACombine(SDNode *N, SelectionDAG &DAG,
894 TargetLowering::DAGCombinerInfo &DCI,
895 const MipsSubtarget &Subtarget) {
896 EVT Ty = N->getValueType(0);
898 if (Subtarget.hasMSA()) {
899 SDValue Op0 = N->getOperand(0);
900 SDValue Op1 = N->getOperand(1);
902 // (sra (shl (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d), imm:$d)
903 // where $d + sizeof($c) == 32
904 // or $d + sizeof($c) <= 32 and SExt
905 // -> (MipsVExtractSExt $a, $b, $c)
906 if (Op0->getOpcode() == ISD::SHL && Op1 == Op0->getOperand(1)) {
907 SDValue Op0Op0 = Op0->getOperand(0);
908 ConstantSDNode *ShAmount = dyn_cast<ConstantSDNode>(Op1);
913 if (Op0Op0->getOpcode() != MipsISD::VEXTRACT_SEXT_ELT &&
914 Op0Op0->getOpcode() != MipsISD::VEXTRACT_ZEXT_ELT)
917 EVT ExtendTy = cast<VTSDNode>(Op0Op0->getOperand(2))->getVT();
918 unsigned TotalBits = ShAmount->getZExtValue() + ExtendTy.getSizeInBits();
920 if (TotalBits == 32 ||
921 (Op0Op0->getOpcode() == MipsISD::VEXTRACT_SEXT_ELT &&
923 SDValue Ops[] = { Op0Op0->getOperand(0), Op0Op0->getOperand(1),
924 Op0Op0->getOperand(2) };
925 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, SDLoc(Op0Op0),
927 makeArrayRef(Ops, Op0Op0->getNumOperands()));
932 if ((Ty != MVT::v2i16) && ((Ty != MVT::v4i8) || !Subtarget.hasDSPR2()))
935 return performDSPShiftCombine(MipsISD::SHRA_DSP, N, Ty, DAG, Subtarget);
939 static SDValue performSRLCombine(SDNode *N, SelectionDAG &DAG,
940 TargetLowering::DAGCombinerInfo &DCI,
941 const MipsSubtarget &Subtarget) {
942 EVT Ty = N->getValueType(0);
944 if (((Ty != MVT::v2i16) || !Subtarget.hasDSPR2()) && (Ty != MVT::v4i8))
947 return performDSPShiftCombine(MipsISD::SHRL_DSP, N, Ty, DAG, Subtarget);
950 static bool isLegalDSPCondCode(EVT Ty, ISD::CondCode CC) {
951 bool IsV216 = (Ty == MVT::v2i16);
955 case ISD::SETNE: return true;
959 case ISD::SETGE: return IsV216;
963 case ISD::SETUGE: return !IsV216;
964 default: return false;
968 static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG) {
969 EVT Ty = N->getValueType(0);
971 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
974 if (!isLegalDSPCondCode(Ty, cast<CondCodeSDNode>(N->getOperand(2))->get()))
977 return DAG.getNode(MipsISD::SETCC_DSP, SDLoc(N), Ty, N->getOperand(0),
978 N->getOperand(1), N->getOperand(2));
981 static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG) {
982 EVT Ty = N->getValueType(0);
984 if (Ty.is128BitVector() && Ty.isInteger()) {
985 // Try the following combines:
986 // (vselect (setcc $a, $b, SETLT), $b, $a)) -> (vsmax $a, $b)
987 // (vselect (setcc $a, $b, SETLE), $b, $a)) -> (vsmax $a, $b)
988 // (vselect (setcc $a, $b, SETLT), $a, $b)) -> (vsmin $a, $b)
989 // (vselect (setcc $a, $b, SETLE), $a, $b)) -> (vsmin $a, $b)
990 // (vselect (setcc $a, $b, SETULT), $b, $a)) -> (vumax $a, $b)
991 // (vselect (setcc $a, $b, SETULE), $b, $a)) -> (vumax $a, $b)
992 // (vselect (setcc $a, $b, SETULT), $a, $b)) -> (vumin $a, $b)
993 // (vselect (setcc $a, $b, SETULE), $a, $b)) -> (vumin $a, $b)
994 // SETGT/SETGE/SETUGT/SETUGE variants of these will show up initially but
995 // will be expanded to equivalent SETLT/SETLE/SETULT/SETULE versions by the
997 SDValue Op0 = N->getOperand(0);
999 if (Op0->getOpcode() != ISD::SETCC)
1002 ISD::CondCode CondCode = cast<CondCodeSDNode>(Op0->getOperand(2))->get();
1005 if (CondCode == ISD::SETLT || CondCode == ISD::SETLE)
1007 else if (CondCode == ISD::SETULT || CondCode == ISD::SETULE)
1012 SDValue Op1 = N->getOperand(1);
1013 SDValue Op2 = N->getOperand(2);
1014 SDValue Op0Op0 = Op0->getOperand(0);
1015 SDValue Op0Op1 = Op0->getOperand(1);
1017 if (Op1 == Op0Op0 && Op2 == Op0Op1)
1018 return DAG.getNode(Signed ? MipsISD::VSMIN : MipsISD::VUMIN, SDLoc(N),
1020 else if (Op1 == Op0Op1 && Op2 == Op0Op0)
1021 return DAG.getNode(Signed ? MipsISD::VSMAX : MipsISD::VUMAX, SDLoc(N),
1023 } else if ((Ty == MVT::v2i16) || (Ty == MVT::v4i8)) {
1024 SDValue SetCC = N->getOperand(0);
1026 if (SetCC.getOpcode() != MipsISD::SETCC_DSP)
1029 return DAG.getNode(MipsISD::SELECT_CC_DSP, SDLoc(N), Ty,
1030 SetCC.getOperand(0), SetCC.getOperand(1),
1031 N->getOperand(1), N->getOperand(2), SetCC.getOperand(2));
1037 static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
1038 const MipsSubtarget &Subtarget) {
1039 EVT Ty = N->getValueType(0);
1041 if (Subtarget.hasMSA() && Ty.is128BitVector() && Ty.isInteger()) {
1042 // Try the following combines:
1043 // (xor (or $a, $b), (build_vector allones))
1044 // (xor (or $a, $b), (bitcast (build_vector allones)))
1045 SDValue Op0 = N->getOperand(0);
1046 SDValue Op1 = N->getOperand(1);
1049 if (ISD::isBuildVectorAllOnes(Op0.getNode()))
1051 else if (ISD::isBuildVectorAllOnes(Op1.getNode()))
1056 if (NotOp->getOpcode() == ISD::OR)
1057 return DAG.getNode(MipsISD::VNOR, SDLoc(N), Ty, NotOp->getOperand(0),
1058 NotOp->getOperand(1));
1065 MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1066 SelectionDAG &DAG = DCI.DAG;
1069 switch (N->getOpcode()) {
1071 return performADDECombine(N, DAG, DCI, Subtarget);
1073 Val = performANDCombine(N, DAG, DCI, Subtarget);
1076 Val = performORCombine(N, DAG, DCI, Subtarget);
1079 return performSUBECombine(N, DAG, DCI, Subtarget);
1081 return performMULCombine(N, DAG, DCI, this);
1083 return performSHLCombine(N, DAG, DCI, Subtarget);
1085 return performSRACombine(N, DAG, DCI, Subtarget);
1087 return performSRLCombine(N, DAG, DCI, Subtarget);
1089 return performVSELECTCombine(N, DAG);
1091 Val = performXORCombine(N, DAG, Subtarget);
1094 Val = performSETCCCombine(N, DAG);
1098 if (Val.getNode()) {
1099 DEBUG(dbgs() << "\nMipsSE DAG Combine:\n";
1100 N->printrWithDepth(dbgs(), &DAG);
1101 dbgs() << "\n=> \n";
1102 Val.getNode()->printrWithDepth(dbgs(), &DAG);
1107 return MipsTargetLowering::PerformDAGCombine(N, DCI);
1111 MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1112 MachineBasicBlock *BB) const {
1113 switch (MI->getOpcode()) {
1115 return MipsTargetLowering::EmitInstrWithCustomInserter(MI, BB);
1116 case Mips::BPOSGE32_PSEUDO:
1117 return emitBPOSGE32(MI, BB);
1118 case Mips::SNZ_B_PSEUDO:
1119 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_B);
1120 case Mips::SNZ_H_PSEUDO:
1121 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_H);
1122 case Mips::SNZ_W_PSEUDO:
1123 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_W);
1124 case Mips::SNZ_D_PSEUDO:
1125 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_D);
1126 case Mips::SNZ_V_PSEUDO:
1127 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_V);
1128 case Mips::SZ_B_PSEUDO:
1129 return emitMSACBranchPseudo(MI, BB, Mips::BZ_B);
1130 case Mips::SZ_H_PSEUDO:
1131 return emitMSACBranchPseudo(MI, BB, Mips::BZ_H);
1132 case Mips::SZ_W_PSEUDO:
1133 return emitMSACBranchPseudo(MI, BB, Mips::BZ_W);
1134 case Mips::SZ_D_PSEUDO:
1135 return emitMSACBranchPseudo(MI, BB, Mips::BZ_D);
1136 case Mips::SZ_V_PSEUDO:
1137 return emitMSACBranchPseudo(MI, BB, Mips::BZ_V);
1138 case Mips::COPY_FW_PSEUDO:
1139 return emitCOPY_FW(MI, BB);
1140 case Mips::COPY_FD_PSEUDO:
1141 return emitCOPY_FD(MI, BB);
1142 case Mips::INSERT_FW_PSEUDO:
1143 return emitINSERT_FW(MI, BB);
1144 case Mips::INSERT_FD_PSEUDO:
1145 return emitINSERT_FD(MI, BB);
1146 case Mips::INSERT_B_VIDX_PSEUDO:
1147 return emitINSERT_DF_VIDX(MI, BB, 1, false);
1148 case Mips::INSERT_H_VIDX_PSEUDO:
1149 return emitINSERT_DF_VIDX(MI, BB, 2, false);
1150 case Mips::INSERT_W_VIDX_PSEUDO:
1151 return emitINSERT_DF_VIDX(MI, BB, 4, false);
1152 case Mips::INSERT_D_VIDX_PSEUDO:
1153 return emitINSERT_DF_VIDX(MI, BB, 8, false);
1154 case Mips::INSERT_FW_VIDX_PSEUDO:
1155 return emitINSERT_DF_VIDX(MI, BB, 4, true);
1156 case Mips::INSERT_FD_VIDX_PSEUDO:
1157 return emitINSERT_DF_VIDX(MI, BB, 8, true);
1158 case Mips::FILL_FW_PSEUDO:
1159 return emitFILL_FW(MI, BB);
1160 case Mips::FILL_FD_PSEUDO:
1161 return emitFILL_FD(MI, BB);
1162 case Mips::FEXP2_W_1_PSEUDO:
1163 return emitFEXP2_W_1(MI, BB);
1164 case Mips::FEXP2_D_1_PSEUDO:
1165 return emitFEXP2_D_1(MI, BB);
1169 bool MipsSETargetLowering::
1170 isEligibleForTailCallOptimization(const MipsCC &MipsCCInfo,
1171 unsigned NextStackOffset,
1172 const MipsFunctionInfo& FI) const {
1173 if (!EnableMipsTailCalls)
1176 // Return false if either the callee or caller has a byval argument.
1177 if (MipsCCInfo.hasByValArg() || FI.hasByvalArg())
1180 // Return true if the callee's argument area is no larger than the
1182 return NextStackOffset <= FI.getIncomingArgSize();
1185 void MipsSETargetLowering::
1186 getOpndList(SmallVectorImpl<SDValue> &Ops,
1187 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
1188 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
1189 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const {
1190 Ops.push_back(Callee);
1191 MipsTargetLowering::getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal,
1192 InternalLinkage, CLI, Callee, Chain);
1195 SDValue MipsSETargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1196 LoadSDNode &Nd = *cast<LoadSDNode>(Op);
1198 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
1199 return MipsTargetLowering::lowerLOAD(Op, DAG);
1201 // Replace a double precision load with two i32 loads and a buildpair64.
1203 SDValue Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
1204 EVT PtrVT = Ptr.getValueType();
1206 // i32 load from lower address.
1207 SDValue Lo = DAG.getLoad(MVT::i32, DL, Chain, Ptr,
1208 MachinePointerInfo(), Nd.isVolatile(),
1209 Nd.isNonTemporal(), Nd.isInvariant(),
1212 // i32 load from higher address.
1213 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, PtrVT));
1214 SDValue Hi = DAG.getLoad(MVT::i32, DL, Lo.getValue(1), Ptr,
1215 MachinePointerInfo(), Nd.isVolatile(),
1216 Nd.isNonTemporal(), Nd.isInvariant(),
1217 std::min(Nd.getAlignment(), 4U));
1219 if (!Subtarget.isLittle())
1222 SDValue BP = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
1223 SDValue Ops[2] = {BP, Hi.getValue(1)};
1224 return DAG.getMergeValues(Ops, DL);
1227 SDValue MipsSETargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1228 StoreSDNode &Nd = *cast<StoreSDNode>(Op);
1230 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
1231 return MipsTargetLowering::lowerSTORE(Op, DAG);
1233 // Replace a double precision store with two extractelement64s and i32 stores.
1235 SDValue Val = Nd.getValue(), Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
1236 EVT PtrVT = Ptr.getValueType();
1237 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1238 Val, DAG.getConstant(0, MVT::i32));
1239 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1240 Val, DAG.getConstant(1, MVT::i32));
1242 if (!Subtarget.isLittle())
1245 // i32 store to lower address.
1246 Chain = DAG.getStore(Chain, DL, Lo, Ptr, MachinePointerInfo(),
1247 Nd.isVolatile(), Nd.isNonTemporal(), Nd.getAlignment(),
1250 // i32 store to higher address.
1251 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, PtrVT));
1252 return DAG.getStore(Chain, DL, Hi, Ptr, MachinePointerInfo(),
1253 Nd.isVolatile(), Nd.isNonTemporal(),
1254 std::min(Nd.getAlignment(), 4U), Nd.getAAInfo());
1257 SDValue MipsSETargetLowering::lowerMulDiv(SDValue Op, unsigned NewOpc,
1258 bool HasLo, bool HasHi,
1259 SelectionDAG &DAG) const {
1260 // MIPS32r6/MIPS64r6 removed accumulator based multiplies.
1261 assert(!Subtarget.hasMips32r6());
1263 EVT Ty = Op.getOperand(0).getValueType();
1265 SDValue Mult = DAG.getNode(NewOpc, DL, MVT::Untyped,
1266 Op.getOperand(0), Op.getOperand(1));
1270 Lo = DAG.getNode(MipsISD::MFLO, DL, Ty, Mult);
1272 Hi = DAG.getNode(MipsISD::MFHI, DL, Ty, Mult);
1274 if (!HasLo || !HasHi)
1275 return HasLo ? Lo : Hi;
1277 SDValue Vals[] = { Lo, Hi };
1278 return DAG.getMergeValues(Vals, DL);
1282 static SDValue initAccumulator(SDValue In, SDLoc DL, SelectionDAG &DAG) {
1283 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
1284 DAG.getConstant(0, MVT::i32));
1285 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
1286 DAG.getConstant(1, MVT::i32));
1287 return DAG.getNode(MipsISD::MTLOHI, DL, MVT::Untyped, InLo, InHi);
1290 static SDValue extractLOHI(SDValue Op, SDLoc DL, SelectionDAG &DAG) {
1291 SDValue Lo = DAG.getNode(MipsISD::MFLO, DL, MVT::i32, Op);
1292 SDValue Hi = DAG.getNode(MipsISD::MFHI, DL, MVT::i32, Op);
1293 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi);
1296 // This function expands mips intrinsic nodes which have 64-bit input operands
1297 // or output values.
1299 // out64 = intrinsic-node in64
1301 // lo = copy (extract-element (in64, 0))
1302 // hi = copy (extract-element (in64, 1))
1303 // mips-specific-node
1306 // out64 = merge-values (v0, v1)
1308 static SDValue lowerDSPIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1310 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
1311 SmallVector<SDValue, 3> Ops;
1314 // See if Op has a chain input.
1316 Ops.push_back(Op->getOperand(OpNo++));
1318 // The next operand is the intrinsic opcode.
1319 assert(Op->getOperand(OpNo).getOpcode() == ISD::TargetConstant);
1321 // See if the next operand has type i64.
1322 SDValue Opnd = Op->getOperand(++OpNo), In64;
1324 if (Opnd.getValueType() == MVT::i64)
1325 In64 = initAccumulator(Opnd, DL, DAG);
1327 Ops.push_back(Opnd);
1329 // Push the remaining operands.
1330 for (++OpNo ; OpNo < Op->getNumOperands(); ++OpNo)
1331 Ops.push_back(Op->getOperand(OpNo));
1333 // Add In64 to the end of the list.
1335 Ops.push_back(In64);
1338 SmallVector<EVT, 2> ResTys;
1340 for (SDNode::value_iterator I = Op->value_begin(), E = Op->value_end();
1342 ResTys.push_back((*I == MVT::i64) ? MVT::Untyped : *I);
1345 SDValue Val = DAG.getNode(Opc, DL, ResTys, Ops);
1346 SDValue Out = (ResTys[0] == MVT::Untyped) ? extractLOHI(Val, DL, DAG) : Val;
1351 assert(Val->getValueType(1) == MVT::Other);
1352 SDValue Vals[] = { Out, SDValue(Val.getNode(), 1) };
1353 return DAG.getMergeValues(Vals, DL);
1356 // Lower an MSA copy intrinsic into the specified SelectionDAG node
1357 static SDValue lowerMSACopyIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1359 SDValue Vec = Op->getOperand(1);
1360 SDValue Idx = Op->getOperand(2);
1361 EVT ResTy = Op->getValueType(0);
1362 EVT EltTy = Vec->getValueType(0).getVectorElementType();
1364 SDValue Result = DAG.getNode(Opc, DL, ResTy, Vec, Idx,
1365 DAG.getValueType(EltTy));
1370 static SDValue lowerMSASplatZExt(SDValue Op, unsigned OpNr, SelectionDAG &DAG) {
1371 EVT ResVecTy = Op->getValueType(0);
1372 EVT ViaVecTy = ResVecTy;
1375 // When ResVecTy == MVT::v2i64, LaneA is the upper 32 bits of the lane and
1376 // LaneB is the lower 32-bits. Otherwise LaneA and LaneB are alternating
1379 SDValue LaneB = Op->getOperand(2);
1381 if (ResVecTy == MVT::v2i64) {
1382 LaneA = DAG.getConstant(0, MVT::i32);
1383 ViaVecTy = MVT::v4i32;
1387 SDValue Ops[16] = { LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB,
1388 LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB };
1390 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy,
1391 makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
1393 if (ViaVecTy != ResVecTy)
1394 Result = DAG.getNode(ISD::BITCAST, DL, ResVecTy, Result);
1399 static SDValue lowerMSASplatImm(SDValue Op, unsigned ImmOp, SelectionDAG &DAG) {
1400 return DAG.getConstant(Op->getConstantOperandVal(ImmOp), Op->getValueType(0));
1403 static SDValue getBuildVectorSplat(EVT VecTy, SDValue SplatValue,
1404 bool BigEndian, SelectionDAG &DAG) {
1405 EVT ViaVecTy = VecTy;
1406 SDValue SplatValueA = SplatValue;
1407 SDValue SplatValueB = SplatValue;
1408 SDLoc DL(SplatValue);
1410 if (VecTy == MVT::v2i64) {
1411 // v2i64 BUILD_VECTOR must be performed via v4i32 so split into i32's.
1412 ViaVecTy = MVT::v4i32;
1414 SplatValueA = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, SplatValue);
1415 SplatValueB = DAG.getNode(ISD::SRL, DL, MVT::i64, SplatValue,
1416 DAG.getConstant(32, MVT::i32));
1417 SplatValueB = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, SplatValueB);
1420 // We currently hold the parts in little endian order. Swap them if
1423 std::swap(SplatValueA, SplatValueB);
1425 SDValue Ops[16] = { SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1426 SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1427 SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1428 SplatValueA, SplatValueB, SplatValueA, SplatValueB };
1430 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy,
1431 makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
1433 if (VecTy != ViaVecTy)
1434 Result = DAG.getNode(ISD::BITCAST, DL, VecTy, Result);
1439 static SDValue lowerMSABinaryBitImmIntr(SDValue Op, SelectionDAG &DAG,
1440 unsigned Opc, SDValue Imm,
1442 EVT VecTy = Op->getValueType(0);
1446 // The DAG Combiner can't constant fold bitcasted vectors yet so we must do it
1448 if (VecTy == MVT::v2i64) {
1449 if (ConstantSDNode *CImm = dyn_cast<ConstantSDNode>(Imm)) {
1450 APInt BitImm = APInt(64, 1) << CImm->getAPIntValue();
1452 SDValue BitImmHiOp = DAG.getConstant(BitImm.lshr(32).trunc(32), MVT::i32);
1453 SDValue BitImmLoOp = DAG.getConstant(BitImm.trunc(32), MVT::i32);
1456 std::swap(BitImmLoOp, BitImmHiOp);
1459 DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
1460 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v4i32, BitImmLoOp,
1461 BitImmHiOp, BitImmLoOp, BitImmHiOp));
1465 if (!Exp2Imm.getNode()) {
1466 // We couldnt constant fold, do a vector shift instead
1468 // Extend i32 to i64 if necessary. Sign or zero extend doesn't matter since
1469 // only values 0-63 are valid.
1470 if (VecTy == MVT::v2i64)
1471 Imm = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Imm);
1473 Exp2Imm = getBuildVectorSplat(VecTy, Imm, BigEndian, DAG);
1476 DAG.getNode(ISD::SHL, DL, VecTy, DAG.getConstant(1, VecTy), Exp2Imm);
1479 return DAG.getNode(Opc, DL, VecTy, Op->getOperand(1), Exp2Imm);
1482 static SDValue lowerMSABitClear(SDValue Op, SelectionDAG &DAG) {
1483 EVT ResTy = Op->getValueType(0);
1485 SDValue One = DAG.getConstant(1, ResTy);
1486 SDValue Bit = DAG.getNode(ISD::SHL, DL, ResTy, One, Op->getOperand(2));
1488 return DAG.getNode(ISD::AND, DL, ResTy, Op->getOperand(1),
1489 DAG.getNOT(DL, Bit, ResTy));
1492 static SDValue lowerMSABitClearImm(SDValue Op, SelectionDAG &DAG) {
1494 EVT ResTy = Op->getValueType(0);
1495 APInt BitImm = APInt(ResTy.getVectorElementType().getSizeInBits(), 1)
1496 << cast<ConstantSDNode>(Op->getOperand(2))->getAPIntValue();
1497 SDValue BitMask = DAG.getConstant(~BitImm, ResTy);
1499 return DAG.getNode(ISD::AND, DL, ResTy, Op->getOperand(1), BitMask);
1502 SDValue MipsSETargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
1503 SelectionDAG &DAG) const {
1506 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) {
1509 case Intrinsic::mips_shilo:
1510 return lowerDSPIntr(Op, DAG, MipsISD::SHILO);
1511 case Intrinsic::mips_dpau_h_qbl:
1512 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL);
1513 case Intrinsic::mips_dpau_h_qbr:
1514 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR);
1515 case Intrinsic::mips_dpsu_h_qbl:
1516 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL);
1517 case Intrinsic::mips_dpsu_h_qbr:
1518 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR);
1519 case Intrinsic::mips_dpa_w_ph:
1520 return lowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH);
1521 case Intrinsic::mips_dps_w_ph:
1522 return lowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH);
1523 case Intrinsic::mips_dpax_w_ph:
1524 return lowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH);
1525 case Intrinsic::mips_dpsx_w_ph:
1526 return lowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH);
1527 case Intrinsic::mips_mulsa_w_ph:
1528 return lowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH);
1529 case Intrinsic::mips_mult:
1530 return lowerDSPIntr(Op, DAG, MipsISD::Mult);
1531 case Intrinsic::mips_multu:
1532 return lowerDSPIntr(Op, DAG, MipsISD::Multu);
1533 case Intrinsic::mips_madd:
1534 return lowerDSPIntr(Op, DAG, MipsISD::MAdd);
1535 case Intrinsic::mips_maddu:
1536 return lowerDSPIntr(Op, DAG, MipsISD::MAddu);
1537 case Intrinsic::mips_msub:
1538 return lowerDSPIntr(Op, DAG, MipsISD::MSub);
1539 case Intrinsic::mips_msubu:
1540 return lowerDSPIntr(Op, DAG, MipsISD::MSubu);
1541 case Intrinsic::mips_addv_b:
1542 case Intrinsic::mips_addv_h:
1543 case Intrinsic::mips_addv_w:
1544 case Intrinsic::mips_addv_d:
1545 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1),
1547 case Intrinsic::mips_addvi_b:
1548 case Intrinsic::mips_addvi_h:
1549 case Intrinsic::mips_addvi_w:
1550 case Intrinsic::mips_addvi_d:
1551 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1),
1552 lowerMSASplatImm(Op, 2, DAG));
1553 case Intrinsic::mips_and_v:
1554 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
1556 case Intrinsic::mips_andi_b:
1557 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
1558 lowerMSASplatImm(Op, 2, DAG));
1559 case Intrinsic::mips_bclr_b:
1560 case Intrinsic::mips_bclr_h:
1561 case Intrinsic::mips_bclr_w:
1562 case Intrinsic::mips_bclr_d:
1563 return lowerMSABitClear(Op, DAG);
1564 case Intrinsic::mips_bclri_b:
1565 case Intrinsic::mips_bclri_h:
1566 case Intrinsic::mips_bclri_w:
1567 case Intrinsic::mips_bclri_d:
1568 return lowerMSABitClearImm(Op, DAG);
1569 case Intrinsic::mips_binsli_b:
1570 case Intrinsic::mips_binsli_h:
1571 case Intrinsic::mips_binsli_w:
1572 case Intrinsic::mips_binsli_d: {
1573 // binsli_x(IfClear, IfSet, nbits) -> (vselect LBitsMask, IfSet, IfClear)
1574 EVT VecTy = Op->getValueType(0);
1575 EVT EltTy = VecTy.getVectorElementType();
1576 APInt Mask = APInt::getHighBitsSet(EltTy.getSizeInBits(),
1577 Op->getConstantOperandVal(3));
1578 return DAG.getNode(ISD::VSELECT, DL, VecTy,
1579 DAG.getConstant(Mask, VecTy, true), Op->getOperand(2),
1582 case Intrinsic::mips_binsri_b:
1583 case Intrinsic::mips_binsri_h:
1584 case Intrinsic::mips_binsri_w:
1585 case Intrinsic::mips_binsri_d: {
1586 // binsri_x(IfClear, IfSet, nbits) -> (vselect RBitsMask, IfSet, IfClear)
1587 EVT VecTy = Op->getValueType(0);
1588 EVT EltTy = VecTy.getVectorElementType();
1589 APInt Mask = APInt::getLowBitsSet(EltTy.getSizeInBits(),
1590 Op->getConstantOperandVal(3));
1591 return DAG.getNode(ISD::VSELECT, DL, VecTy,
1592 DAG.getConstant(Mask, VecTy, true), Op->getOperand(2),
1595 case Intrinsic::mips_bmnz_v:
1596 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), Op->getOperand(3),
1597 Op->getOperand(2), Op->getOperand(1));
1598 case Intrinsic::mips_bmnzi_b:
1599 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1600 lowerMSASplatImm(Op, 3, DAG), Op->getOperand(2),
1602 case Intrinsic::mips_bmz_v:
1603 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), Op->getOperand(3),
1604 Op->getOperand(1), Op->getOperand(2));
1605 case Intrinsic::mips_bmzi_b:
1606 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1607 lowerMSASplatImm(Op, 3, DAG), Op->getOperand(1),
1609 case Intrinsic::mips_bneg_b:
1610 case Intrinsic::mips_bneg_h:
1611 case Intrinsic::mips_bneg_w:
1612 case Intrinsic::mips_bneg_d: {
1613 EVT VecTy = Op->getValueType(0);
1614 SDValue One = DAG.getConstant(1, VecTy);
1616 return DAG.getNode(ISD::XOR, DL, VecTy, Op->getOperand(1),
1617 DAG.getNode(ISD::SHL, DL, VecTy, One,
1618 Op->getOperand(2)));
1620 case Intrinsic::mips_bnegi_b:
1621 case Intrinsic::mips_bnegi_h:
1622 case Intrinsic::mips_bnegi_w:
1623 case Intrinsic::mips_bnegi_d:
1624 return lowerMSABinaryBitImmIntr(Op, DAG, ISD::XOR, Op->getOperand(2),
1625 !Subtarget.isLittle());
1626 case Intrinsic::mips_bnz_b:
1627 case Intrinsic::mips_bnz_h:
1628 case Intrinsic::mips_bnz_w:
1629 case Intrinsic::mips_bnz_d:
1630 return DAG.getNode(MipsISD::VALL_NONZERO, DL, Op->getValueType(0),
1632 case Intrinsic::mips_bnz_v:
1633 return DAG.getNode(MipsISD::VANY_NONZERO, DL, Op->getValueType(0),
1635 case Intrinsic::mips_bsel_v:
1636 // bsel_v(Mask, IfClear, IfSet) -> (vselect Mask, IfSet, IfClear)
1637 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1638 Op->getOperand(1), Op->getOperand(3),
1640 case Intrinsic::mips_bseli_b:
1641 // bseli_v(Mask, IfClear, IfSet) -> (vselect Mask, IfSet, IfClear)
1642 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1643 Op->getOperand(1), lowerMSASplatImm(Op, 3, DAG),
1645 case Intrinsic::mips_bset_b:
1646 case Intrinsic::mips_bset_h:
1647 case Intrinsic::mips_bset_w:
1648 case Intrinsic::mips_bset_d: {
1649 EVT VecTy = Op->getValueType(0);
1650 SDValue One = DAG.getConstant(1, VecTy);
1652 return DAG.getNode(ISD::OR, DL, VecTy, Op->getOperand(1),
1653 DAG.getNode(ISD::SHL, DL, VecTy, One,
1654 Op->getOperand(2)));
1656 case Intrinsic::mips_bseti_b:
1657 case Intrinsic::mips_bseti_h:
1658 case Intrinsic::mips_bseti_w:
1659 case Intrinsic::mips_bseti_d:
1660 return lowerMSABinaryBitImmIntr(Op, DAG, ISD::OR, Op->getOperand(2),
1661 !Subtarget.isLittle());
1662 case Intrinsic::mips_bz_b:
1663 case Intrinsic::mips_bz_h:
1664 case Intrinsic::mips_bz_w:
1665 case Intrinsic::mips_bz_d:
1666 return DAG.getNode(MipsISD::VALL_ZERO, DL, Op->getValueType(0),
1668 case Intrinsic::mips_bz_v:
1669 return DAG.getNode(MipsISD::VANY_ZERO, DL, Op->getValueType(0),
1671 case Intrinsic::mips_ceq_b:
1672 case Intrinsic::mips_ceq_h:
1673 case Intrinsic::mips_ceq_w:
1674 case Intrinsic::mips_ceq_d:
1675 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1676 Op->getOperand(2), ISD::SETEQ);
1677 case Intrinsic::mips_ceqi_b:
1678 case Intrinsic::mips_ceqi_h:
1679 case Intrinsic::mips_ceqi_w:
1680 case Intrinsic::mips_ceqi_d:
1681 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1682 lowerMSASplatImm(Op, 2, DAG), ISD::SETEQ);
1683 case Intrinsic::mips_cle_s_b:
1684 case Intrinsic::mips_cle_s_h:
1685 case Intrinsic::mips_cle_s_w:
1686 case Intrinsic::mips_cle_s_d:
1687 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1688 Op->getOperand(2), ISD::SETLE);
1689 case Intrinsic::mips_clei_s_b:
1690 case Intrinsic::mips_clei_s_h:
1691 case Intrinsic::mips_clei_s_w:
1692 case Intrinsic::mips_clei_s_d:
1693 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1694 lowerMSASplatImm(Op, 2, DAG), ISD::SETLE);
1695 case Intrinsic::mips_cle_u_b:
1696 case Intrinsic::mips_cle_u_h:
1697 case Intrinsic::mips_cle_u_w:
1698 case Intrinsic::mips_cle_u_d:
1699 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1700 Op->getOperand(2), ISD::SETULE);
1701 case Intrinsic::mips_clei_u_b:
1702 case Intrinsic::mips_clei_u_h:
1703 case Intrinsic::mips_clei_u_w:
1704 case Intrinsic::mips_clei_u_d:
1705 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1706 lowerMSASplatImm(Op, 2, DAG), ISD::SETULE);
1707 case Intrinsic::mips_clt_s_b:
1708 case Intrinsic::mips_clt_s_h:
1709 case Intrinsic::mips_clt_s_w:
1710 case Intrinsic::mips_clt_s_d:
1711 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1712 Op->getOperand(2), ISD::SETLT);
1713 case Intrinsic::mips_clti_s_b:
1714 case Intrinsic::mips_clti_s_h:
1715 case Intrinsic::mips_clti_s_w:
1716 case Intrinsic::mips_clti_s_d:
1717 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1718 lowerMSASplatImm(Op, 2, DAG), ISD::SETLT);
1719 case Intrinsic::mips_clt_u_b:
1720 case Intrinsic::mips_clt_u_h:
1721 case Intrinsic::mips_clt_u_w:
1722 case Intrinsic::mips_clt_u_d:
1723 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1724 Op->getOperand(2), ISD::SETULT);
1725 case Intrinsic::mips_clti_u_b:
1726 case Intrinsic::mips_clti_u_h:
1727 case Intrinsic::mips_clti_u_w:
1728 case Intrinsic::mips_clti_u_d:
1729 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1730 lowerMSASplatImm(Op, 2, DAG), ISD::SETULT);
1731 case Intrinsic::mips_copy_s_b:
1732 case Intrinsic::mips_copy_s_h:
1733 case Intrinsic::mips_copy_s_w:
1734 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
1735 case Intrinsic::mips_copy_s_d:
1736 if (Subtarget.hasMips64())
1737 // Lower directly into VEXTRACT_SEXT_ELT since i64 is legal on Mips64.
1738 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
1740 // Lower into the generic EXTRACT_VECTOR_ELT node and let the type
1741 // legalizer and EXTRACT_VECTOR_ELT lowering sort it out.
1742 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
1743 Op->getValueType(0), Op->getOperand(1),
1746 case Intrinsic::mips_copy_u_b:
1747 case Intrinsic::mips_copy_u_h:
1748 case Intrinsic::mips_copy_u_w:
1749 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
1750 case Intrinsic::mips_copy_u_d:
1751 if (Subtarget.hasMips64())
1752 // Lower directly into VEXTRACT_ZEXT_ELT since i64 is legal on Mips64.
1753 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
1755 // Lower into the generic EXTRACT_VECTOR_ELT node and let the type
1756 // legalizer and EXTRACT_VECTOR_ELT lowering sort it out.
1757 // Note: When i64 is illegal, this results in copy_s.w instructions
1758 // instead of copy_u.w instructions. This makes no difference to the
1759 // behaviour since i64 is only illegal when the register file is 32-bit.
1760 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
1761 Op->getValueType(0), Op->getOperand(1),
1764 case Intrinsic::mips_div_s_b:
1765 case Intrinsic::mips_div_s_h:
1766 case Intrinsic::mips_div_s_w:
1767 case Intrinsic::mips_div_s_d:
1768 return DAG.getNode(ISD::SDIV, DL, Op->getValueType(0), Op->getOperand(1),
1770 case Intrinsic::mips_div_u_b:
1771 case Intrinsic::mips_div_u_h:
1772 case Intrinsic::mips_div_u_w:
1773 case Intrinsic::mips_div_u_d:
1774 return DAG.getNode(ISD::UDIV, DL, Op->getValueType(0), Op->getOperand(1),
1776 case Intrinsic::mips_fadd_w:
1777 case Intrinsic::mips_fadd_d:
1778 return DAG.getNode(ISD::FADD, DL, Op->getValueType(0), Op->getOperand(1),
1780 // Don't lower mips_fcaf_[wd] since LLVM folds SETFALSE condcodes away
1781 case Intrinsic::mips_fceq_w:
1782 case Intrinsic::mips_fceq_d:
1783 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1784 Op->getOperand(2), ISD::SETOEQ);
1785 case Intrinsic::mips_fcle_w:
1786 case Intrinsic::mips_fcle_d:
1787 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1788 Op->getOperand(2), ISD::SETOLE);
1789 case Intrinsic::mips_fclt_w:
1790 case Intrinsic::mips_fclt_d:
1791 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1792 Op->getOperand(2), ISD::SETOLT);
1793 case Intrinsic::mips_fcne_w:
1794 case Intrinsic::mips_fcne_d:
1795 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1796 Op->getOperand(2), ISD::SETONE);
1797 case Intrinsic::mips_fcor_w:
1798 case Intrinsic::mips_fcor_d:
1799 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1800 Op->getOperand(2), ISD::SETO);
1801 case Intrinsic::mips_fcueq_w:
1802 case Intrinsic::mips_fcueq_d:
1803 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1804 Op->getOperand(2), ISD::SETUEQ);
1805 case Intrinsic::mips_fcule_w:
1806 case Intrinsic::mips_fcule_d:
1807 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1808 Op->getOperand(2), ISD::SETULE);
1809 case Intrinsic::mips_fcult_w:
1810 case Intrinsic::mips_fcult_d:
1811 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1812 Op->getOperand(2), ISD::SETULT);
1813 case Intrinsic::mips_fcun_w:
1814 case Intrinsic::mips_fcun_d:
1815 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1816 Op->getOperand(2), ISD::SETUO);
1817 case Intrinsic::mips_fcune_w:
1818 case Intrinsic::mips_fcune_d:
1819 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1820 Op->getOperand(2), ISD::SETUNE);
1821 case Intrinsic::mips_fdiv_w:
1822 case Intrinsic::mips_fdiv_d:
1823 return DAG.getNode(ISD::FDIV, DL, Op->getValueType(0), Op->getOperand(1),
1825 case Intrinsic::mips_ffint_u_w:
1826 case Intrinsic::mips_ffint_u_d:
1827 return DAG.getNode(ISD::UINT_TO_FP, DL, Op->getValueType(0),
1829 case Intrinsic::mips_ffint_s_w:
1830 case Intrinsic::mips_ffint_s_d:
1831 return DAG.getNode(ISD::SINT_TO_FP, DL, Op->getValueType(0),
1833 case Intrinsic::mips_fill_b:
1834 case Intrinsic::mips_fill_h:
1835 case Intrinsic::mips_fill_w:
1836 case Intrinsic::mips_fill_d: {
1837 SmallVector<SDValue, 16> Ops;
1838 EVT ResTy = Op->getValueType(0);
1840 for (unsigned i = 0; i < ResTy.getVectorNumElements(); ++i)
1841 Ops.push_back(Op->getOperand(1));
1843 // If ResTy is v2i64 then the type legalizer will break this node down into
1844 // an equivalent v4i32.
1845 return DAG.getNode(ISD::BUILD_VECTOR, DL, ResTy, Ops);
1847 case Intrinsic::mips_fexp2_w:
1848 case Intrinsic::mips_fexp2_d: {
1849 EVT ResTy = Op->getValueType(0);
1851 ISD::FMUL, SDLoc(Op), ResTy, Op->getOperand(1),
1852 DAG.getNode(ISD::FEXP2, SDLoc(Op), ResTy, Op->getOperand(2)));
1854 case Intrinsic::mips_flog2_w:
1855 case Intrinsic::mips_flog2_d:
1856 return DAG.getNode(ISD::FLOG2, DL, Op->getValueType(0), Op->getOperand(1));
1857 case Intrinsic::mips_fmadd_w:
1858 case Intrinsic::mips_fmadd_d:
1859 return DAG.getNode(ISD::FMA, SDLoc(Op), Op->getValueType(0),
1860 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3));
1861 case Intrinsic::mips_fmul_w:
1862 case Intrinsic::mips_fmul_d:
1863 return DAG.getNode(ISD::FMUL, DL, Op->getValueType(0), Op->getOperand(1),
1865 case Intrinsic::mips_fmsub_w:
1866 case Intrinsic::mips_fmsub_d: {
1867 EVT ResTy = Op->getValueType(0);
1868 return DAG.getNode(ISD::FSUB, SDLoc(Op), ResTy, Op->getOperand(1),
1869 DAG.getNode(ISD::FMUL, SDLoc(Op), ResTy,
1870 Op->getOperand(2), Op->getOperand(3)));
1872 case Intrinsic::mips_frint_w:
1873 case Intrinsic::mips_frint_d:
1874 return DAG.getNode(ISD::FRINT, DL, Op->getValueType(0), Op->getOperand(1));
1875 case Intrinsic::mips_fsqrt_w:
1876 case Intrinsic::mips_fsqrt_d:
1877 return DAG.getNode(ISD::FSQRT, DL, Op->getValueType(0), Op->getOperand(1));
1878 case Intrinsic::mips_fsub_w:
1879 case Intrinsic::mips_fsub_d:
1880 return DAG.getNode(ISD::FSUB, DL, Op->getValueType(0), Op->getOperand(1),
1882 case Intrinsic::mips_ftrunc_u_w:
1883 case Intrinsic::mips_ftrunc_u_d:
1884 return DAG.getNode(ISD::FP_TO_UINT, DL, Op->getValueType(0),
1886 case Intrinsic::mips_ftrunc_s_w:
1887 case Intrinsic::mips_ftrunc_s_d:
1888 return DAG.getNode(ISD::FP_TO_SINT, DL, Op->getValueType(0),
1890 case Intrinsic::mips_ilvev_b:
1891 case Intrinsic::mips_ilvev_h:
1892 case Intrinsic::mips_ilvev_w:
1893 case Intrinsic::mips_ilvev_d:
1894 return DAG.getNode(MipsISD::ILVEV, DL, Op->getValueType(0),
1895 Op->getOperand(1), Op->getOperand(2));
1896 case Intrinsic::mips_ilvl_b:
1897 case Intrinsic::mips_ilvl_h:
1898 case Intrinsic::mips_ilvl_w:
1899 case Intrinsic::mips_ilvl_d:
1900 return DAG.getNode(MipsISD::ILVL, DL, Op->getValueType(0),
1901 Op->getOperand(1), Op->getOperand(2));
1902 case Intrinsic::mips_ilvod_b:
1903 case Intrinsic::mips_ilvod_h:
1904 case Intrinsic::mips_ilvod_w:
1905 case Intrinsic::mips_ilvod_d:
1906 return DAG.getNode(MipsISD::ILVOD, DL, Op->getValueType(0),
1907 Op->getOperand(1), Op->getOperand(2));
1908 case Intrinsic::mips_ilvr_b:
1909 case Intrinsic::mips_ilvr_h:
1910 case Intrinsic::mips_ilvr_w:
1911 case Intrinsic::mips_ilvr_d:
1912 return DAG.getNode(MipsISD::ILVR, DL, Op->getValueType(0),
1913 Op->getOperand(1), Op->getOperand(2));
1914 case Intrinsic::mips_insert_b:
1915 case Intrinsic::mips_insert_h:
1916 case Intrinsic::mips_insert_w:
1917 case Intrinsic::mips_insert_d:
1918 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(Op), Op->getValueType(0),
1919 Op->getOperand(1), Op->getOperand(3), Op->getOperand(2));
1920 case Intrinsic::mips_insve_b:
1921 case Intrinsic::mips_insve_h:
1922 case Intrinsic::mips_insve_w:
1923 case Intrinsic::mips_insve_d:
1924 return DAG.getNode(MipsISD::INSVE, DL, Op->getValueType(0),
1925 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3),
1926 DAG.getConstant(0, MVT::i32));
1927 case Intrinsic::mips_ldi_b:
1928 case Intrinsic::mips_ldi_h:
1929 case Intrinsic::mips_ldi_w:
1930 case Intrinsic::mips_ldi_d:
1931 return lowerMSASplatImm(Op, 1, DAG);
1932 case Intrinsic::mips_lsa:
1933 case Intrinsic::mips_dlsa: {
1934 EVT ResTy = Op->getValueType(0);
1935 return DAG.getNode(ISD::ADD, SDLoc(Op), ResTy, Op->getOperand(1),
1936 DAG.getNode(ISD::SHL, SDLoc(Op), ResTy,
1937 Op->getOperand(2), Op->getOperand(3)));
1939 case Intrinsic::mips_maddv_b:
1940 case Intrinsic::mips_maddv_h:
1941 case Intrinsic::mips_maddv_w:
1942 case Intrinsic::mips_maddv_d: {
1943 EVT ResTy = Op->getValueType(0);
1944 return DAG.getNode(ISD::ADD, SDLoc(Op), ResTy, Op->getOperand(1),
1945 DAG.getNode(ISD::MUL, SDLoc(Op), ResTy,
1946 Op->getOperand(2), Op->getOperand(3)));
1948 case Intrinsic::mips_max_s_b:
1949 case Intrinsic::mips_max_s_h:
1950 case Intrinsic::mips_max_s_w:
1951 case Intrinsic::mips_max_s_d:
1952 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0),
1953 Op->getOperand(1), Op->getOperand(2));
1954 case Intrinsic::mips_max_u_b:
1955 case Intrinsic::mips_max_u_h:
1956 case Intrinsic::mips_max_u_w:
1957 case Intrinsic::mips_max_u_d:
1958 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0),
1959 Op->getOperand(1), Op->getOperand(2));
1960 case Intrinsic::mips_maxi_s_b:
1961 case Intrinsic::mips_maxi_s_h:
1962 case Intrinsic::mips_maxi_s_w:
1963 case Intrinsic::mips_maxi_s_d:
1964 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0),
1965 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1966 case Intrinsic::mips_maxi_u_b:
1967 case Intrinsic::mips_maxi_u_h:
1968 case Intrinsic::mips_maxi_u_w:
1969 case Intrinsic::mips_maxi_u_d:
1970 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0),
1971 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1972 case Intrinsic::mips_min_s_b:
1973 case Intrinsic::mips_min_s_h:
1974 case Intrinsic::mips_min_s_w:
1975 case Intrinsic::mips_min_s_d:
1976 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0),
1977 Op->getOperand(1), Op->getOperand(2));
1978 case Intrinsic::mips_min_u_b:
1979 case Intrinsic::mips_min_u_h:
1980 case Intrinsic::mips_min_u_w:
1981 case Intrinsic::mips_min_u_d:
1982 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0),
1983 Op->getOperand(1), Op->getOperand(2));
1984 case Intrinsic::mips_mini_s_b:
1985 case Intrinsic::mips_mini_s_h:
1986 case Intrinsic::mips_mini_s_w:
1987 case Intrinsic::mips_mini_s_d:
1988 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0),
1989 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1990 case Intrinsic::mips_mini_u_b:
1991 case Intrinsic::mips_mini_u_h:
1992 case Intrinsic::mips_mini_u_w:
1993 case Intrinsic::mips_mini_u_d:
1994 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0),
1995 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1996 case Intrinsic::mips_mod_s_b:
1997 case Intrinsic::mips_mod_s_h:
1998 case Intrinsic::mips_mod_s_w:
1999 case Intrinsic::mips_mod_s_d:
2000 return DAG.getNode(ISD::SREM, DL, Op->getValueType(0), Op->getOperand(1),
2002 case Intrinsic::mips_mod_u_b:
2003 case Intrinsic::mips_mod_u_h:
2004 case Intrinsic::mips_mod_u_w:
2005 case Intrinsic::mips_mod_u_d:
2006 return DAG.getNode(ISD::UREM, DL, Op->getValueType(0), Op->getOperand(1),
2008 case Intrinsic::mips_mulv_b:
2009 case Intrinsic::mips_mulv_h:
2010 case Intrinsic::mips_mulv_w:
2011 case Intrinsic::mips_mulv_d:
2012 return DAG.getNode(ISD::MUL, DL, Op->getValueType(0), Op->getOperand(1),
2014 case Intrinsic::mips_msubv_b:
2015 case Intrinsic::mips_msubv_h:
2016 case Intrinsic::mips_msubv_w:
2017 case Intrinsic::mips_msubv_d: {
2018 EVT ResTy = Op->getValueType(0);
2019 return DAG.getNode(ISD::SUB, SDLoc(Op), ResTy, Op->getOperand(1),
2020 DAG.getNode(ISD::MUL, SDLoc(Op), ResTy,
2021 Op->getOperand(2), Op->getOperand(3)));
2023 case Intrinsic::mips_nlzc_b:
2024 case Intrinsic::mips_nlzc_h:
2025 case Intrinsic::mips_nlzc_w:
2026 case Intrinsic::mips_nlzc_d:
2027 return DAG.getNode(ISD::CTLZ, DL, Op->getValueType(0), Op->getOperand(1));
2028 case Intrinsic::mips_nor_v: {
2029 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2030 Op->getOperand(1), Op->getOperand(2));
2031 return DAG.getNOT(DL, Res, Res->getValueType(0));
2033 case Intrinsic::mips_nori_b: {
2034 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2036 lowerMSASplatImm(Op, 2, DAG));
2037 return DAG.getNOT(DL, Res, Res->getValueType(0));
2039 case Intrinsic::mips_or_v:
2040 return DAG.getNode(ISD::OR, DL, Op->getValueType(0), Op->getOperand(1),
2042 case Intrinsic::mips_ori_b:
2043 return DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2044 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2045 case Intrinsic::mips_pckev_b:
2046 case Intrinsic::mips_pckev_h:
2047 case Intrinsic::mips_pckev_w:
2048 case Intrinsic::mips_pckev_d:
2049 return DAG.getNode(MipsISD::PCKEV, DL, Op->getValueType(0),
2050 Op->getOperand(1), Op->getOperand(2));
2051 case Intrinsic::mips_pckod_b:
2052 case Intrinsic::mips_pckod_h:
2053 case Intrinsic::mips_pckod_w:
2054 case Intrinsic::mips_pckod_d:
2055 return DAG.getNode(MipsISD::PCKOD, DL, Op->getValueType(0),
2056 Op->getOperand(1), Op->getOperand(2));
2057 case Intrinsic::mips_pcnt_b:
2058 case Intrinsic::mips_pcnt_h:
2059 case Intrinsic::mips_pcnt_w:
2060 case Intrinsic::mips_pcnt_d:
2061 return DAG.getNode(ISD::CTPOP, DL, Op->getValueType(0), Op->getOperand(1));
2062 case Intrinsic::mips_shf_b:
2063 case Intrinsic::mips_shf_h:
2064 case Intrinsic::mips_shf_w:
2065 return DAG.getNode(MipsISD::SHF, DL, Op->getValueType(0),
2066 Op->getOperand(2), Op->getOperand(1));
2067 case Intrinsic::mips_sll_b:
2068 case Intrinsic::mips_sll_h:
2069 case Intrinsic::mips_sll_w:
2070 case Intrinsic::mips_sll_d:
2071 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0), Op->getOperand(1),
2073 case Intrinsic::mips_slli_b:
2074 case Intrinsic::mips_slli_h:
2075 case Intrinsic::mips_slli_w:
2076 case Intrinsic::mips_slli_d:
2077 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0),
2078 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2079 case Intrinsic::mips_splat_b:
2080 case Intrinsic::mips_splat_h:
2081 case Intrinsic::mips_splat_w:
2082 case Intrinsic::mips_splat_d:
2083 // We can't lower via VECTOR_SHUFFLE because it requires constant shuffle
2084 // masks, nor can we lower via BUILD_VECTOR & EXTRACT_VECTOR_ELT because
2085 // EXTRACT_VECTOR_ELT can't extract i64's on MIPS32.
2086 // Instead we lower to MipsISD::VSHF and match from there.
2087 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2088 lowerMSASplatZExt(Op, 2, DAG), Op->getOperand(1),
2090 case Intrinsic::mips_splati_b:
2091 case Intrinsic::mips_splati_h:
2092 case Intrinsic::mips_splati_w:
2093 case Intrinsic::mips_splati_d:
2094 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2095 lowerMSASplatImm(Op, 2, DAG), Op->getOperand(1),
2097 case Intrinsic::mips_sra_b:
2098 case Intrinsic::mips_sra_h:
2099 case Intrinsic::mips_sra_w:
2100 case Intrinsic::mips_sra_d:
2101 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0), Op->getOperand(1),
2103 case Intrinsic::mips_srai_b:
2104 case Intrinsic::mips_srai_h:
2105 case Intrinsic::mips_srai_w:
2106 case Intrinsic::mips_srai_d:
2107 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0),
2108 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2109 case Intrinsic::mips_srl_b:
2110 case Intrinsic::mips_srl_h:
2111 case Intrinsic::mips_srl_w:
2112 case Intrinsic::mips_srl_d:
2113 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0), Op->getOperand(1),
2115 case Intrinsic::mips_srli_b:
2116 case Intrinsic::mips_srli_h:
2117 case Intrinsic::mips_srli_w:
2118 case Intrinsic::mips_srli_d:
2119 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0),
2120 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2121 case Intrinsic::mips_subv_b:
2122 case Intrinsic::mips_subv_h:
2123 case Intrinsic::mips_subv_w:
2124 case Intrinsic::mips_subv_d:
2125 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0), Op->getOperand(1),
2127 case Intrinsic::mips_subvi_b:
2128 case Intrinsic::mips_subvi_h:
2129 case Intrinsic::mips_subvi_w:
2130 case Intrinsic::mips_subvi_d:
2131 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0),
2132 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2133 case Intrinsic::mips_vshf_b:
2134 case Intrinsic::mips_vshf_h:
2135 case Intrinsic::mips_vshf_w:
2136 case Intrinsic::mips_vshf_d:
2137 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2138 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3));
2139 case Intrinsic::mips_xor_v:
2140 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0), Op->getOperand(1),
2142 case Intrinsic::mips_xori_b:
2143 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0),
2144 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2148 static SDValue lowerMSALoadIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
2150 SDValue ChainIn = Op->getOperand(0);
2151 SDValue Address = Op->getOperand(2);
2152 SDValue Offset = Op->getOperand(3);
2153 EVT ResTy = Op->getValueType(0);
2154 EVT PtrTy = Address->getValueType(0);
2156 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
2158 return DAG.getLoad(ResTy, DL, ChainIn, Address, MachinePointerInfo(), false,
2162 SDValue MipsSETargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
2163 SelectionDAG &DAG) const {
2164 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
2168 case Intrinsic::mips_extp:
2169 return lowerDSPIntr(Op, DAG, MipsISD::EXTP);
2170 case Intrinsic::mips_extpdp:
2171 return lowerDSPIntr(Op, DAG, MipsISD::EXTPDP);
2172 case Intrinsic::mips_extr_w:
2173 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_W);
2174 case Intrinsic::mips_extr_r_w:
2175 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W);
2176 case Intrinsic::mips_extr_rs_w:
2177 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W);
2178 case Intrinsic::mips_extr_s_h:
2179 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H);
2180 case Intrinsic::mips_mthlip:
2181 return lowerDSPIntr(Op, DAG, MipsISD::MTHLIP);
2182 case Intrinsic::mips_mulsaq_s_w_ph:
2183 return lowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH);
2184 case Intrinsic::mips_maq_s_w_phl:
2185 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL);
2186 case Intrinsic::mips_maq_s_w_phr:
2187 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR);
2188 case Intrinsic::mips_maq_sa_w_phl:
2189 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL);
2190 case Intrinsic::mips_maq_sa_w_phr:
2191 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR);
2192 case Intrinsic::mips_dpaq_s_w_ph:
2193 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH);
2194 case Intrinsic::mips_dpsq_s_w_ph:
2195 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH);
2196 case Intrinsic::mips_dpaq_sa_l_w:
2197 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W);
2198 case Intrinsic::mips_dpsq_sa_l_w:
2199 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W);
2200 case Intrinsic::mips_dpaqx_s_w_ph:
2201 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH);
2202 case Intrinsic::mips_dpaqx_sa_w_ph:
2203 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH);
2204 case Intrinsic::mips_dpsqx_s_w_ph:
2205 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH);
2206 case Intrinsic::mips_dpsqx_sa_w_ph:
2207 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH);
2208 case Intrinsic::mips_ld_b:
2209 case Intrinsic::mips_ld_h:
2210 case Intrinsic::mips_ld_w:
2211 case Intrinsic::mips_ld_d:
2212 return lowerMSALoadIntr(Op, DAG, Intr);
2216 static SDValue lowerMSAStoreIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
2218 SDValue ChainIn = Op->getOperand(0);
2219 SDValue Value = Op->getOperand(2);
2220 SDValue Address = Op->getOperand(3);
2221 SDValue Offset = Op->getOperand(4);
2222 EVT PtrTy = Address->getValueType(0);
2224 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
2226 return DAG.getStore(ChainIn, DL, Value, Address, MachinePointerInfo(), false,
2230 SDValue MipsSETargetLowering::lowerINTRINSIC_VOID(SDValue Op,
2231 SelectionDAG &DAG) const {
2232 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
2236 case Intrinsic::mips_st_b:
2237 case Intrinsic::mips_st_h:
2238 case Intrinsic::mips_st_w:
2239 case Intrinsic::mips_st_d:
2240 return lowerMSAStoreIntr(Op, DAG, Intr);
2244 /// \brief Check if the given BuildVectorSDNode is a splat.
2245 /// This method currently relies on DAG nodes being reused when equivalent,
2246 /// so it's possible for this to return false even when isConstantSplat returns
2248 static bool isSplatVector(const BuildVectorSDNode *N) {
2249 unsigned int nOps = N->getNumOperands();
2250 assert(nOps > 1 && "isSplatVector has 0 or 1 sized build vector");
2252 SDValue Operand0 = N->getOperand(0);
2254 for (unsigned int i = 1; i < nOps; ++i) {
2255 if (N->getOperand(i) != Operand0)
2262 // Lower ISD::EXTRACT_VECTOR_ELT into MipsISD::VEXTRACT_SEXT_ELT.
2264 // The non-value bits resulting from ISD::EXTRACT_VECTOR_ELT are undefined. We
2265 // choose to sign-extend but we could have equally chosen zero-extend. The
2266 // DAGCombiner will fold any sign/zero extension of the ISD::EXTRACT_VECTOR_ELT
2267 // result into this node later (possibly changing it to a zero-extend in the
2269 SDValue MipsSETargetLowering::
2270 lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
2272 EVT ResTy = Op->getValueType(0);
2273 SDValue Op0 = Op->getOperand(0);
2274 EVT VecTy = Op0->getValueType(0);
2276 if (!VecTy.is128BitVector())
2279 if (ResTy.isInteger()) {
2280 SDValue Op1 = Op->getOperand(1);
2281 EVT EltTy = VecTy.getVectorElementType();
2282 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, DL, ResTy, Op0, Op1,
2283 DAG.getValueType(EltTy));
2289 static bool isConstantOrUndef(const SDValue Op) {
2290 if (Op->getOpcode() == ISD::UNDEF)
2292 if (dyn_cast<ConstantSDNode>(Op))
2294 if (dyn_cast<ConstantFPSDNode>(Op))
2299 static bool isConstantOrUndefBUILD_VECTOR(const BuildVectorSDNode *Op) {
2300 for (unsigned i = 0; i < Op->getNumOperands(); ++i)
2301 if (isConstantOrUndef(Op->getOperand(i)))
2306 // Lowers ISD::BUILD_VECTOR into appropriate SelectionDAG nodes for the
2309 // Lowers according to the following rules:
2310 // - Constant splats are legal as-is as long as the SplatBitSize is a power of
2311 // 2 less than or equal to 64 and the value fits into a signed 10-bit
2313 // - Constant splats are lowered to bitconverted BUILD_VECTORs if SplatBitSize
2314 // is a power of 2 less than or equal to 64 and the value does not fit into a
2315 // signed 10-bit immediate
2316 // - Non-constant splats are legal as-is.
2317 // - Non-constant non-splats are lowered to sequences of INSERT_VECTOR_ELT.
2318 // - All others are illegal and must be expanded.
2319 SDValue MipsSETargetLowering::lowerBUILD_VECTOR(SDValue Op,
2320 SelectionDAG &DAG) const {
2321 BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op);
2322 EVT ResTy = Op->getValueType(0);
2324 APInt SplatValue, SplatUndef;
2325 unsigned SplatBitSize;
2328 if (!Subtarget.hasMSA() || !ResTy.is128BitVector())
2331 if (Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
2333 !Subtarget.isLittle()) && SplatBitSize <= 64) {
2334 // We can only cope with 8, 16, 32, or 64-bit elements
2335 if (SplatBitSize != 8 && SplatBitSize != 16 && SplatBitSize != 32 &&
2339 // If the value fits into a simm10 then we can use ldi.[bhwd]
2340 // However, if it isn't an integer type we will have to bitcast from an
2341 // integer type first. Also, if there are any undefs, we must lower them
2342 // to defined values first.
2343 if (ResTy.isInteger() && !HasAnyUndefs && SplatValue.isSignedIntN(10))
2348 switch (SplatBitSize) {
2352 ViaVecTy = MVT::v16i8;
2355 ViaVecTy = MVT::v8i16;
2358 ViaVecTy = MVT::v4i32;
2361 // There's no fill.d to fall back on for 64-bit values
2365 // SelectionDAG::getConstant will promote SplatValue appropriately.
2366 SDValue Result = DAG.getConstant(SplatValue, ViaVecTy);
2368 // Bitcast to the type we originally wanted
2369 if (ViaVecTy != ResTy)
2370 Result = DAG.getNode(ISD::BITCAST, SDLoc(Node), ResTy, Result);
2373 } else if (isSplatVector(Node))
2375 else if (!isConstantOrUndefBUILD_VECTOR(Node)) {
2376 // Use INSERT_VECTOR_ELT operations rather than expand to stores.
2377 // The resulting code is the same length as the expansion, but it doesn't
2378 // use memory operations
2379 EVT ResTy = Node->getValueType(0);
2381 assert(ResTy.isVector());
2383 unsigned NumElts = ResTy.getVectorNumElements();
2384 SDValue Vector = DAG.getUNDEF(ResTy);
2385 for (unsigned i = 0; i < NumElts; ++i) {
2386 Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector,
2387 Node->getOperand(i),
2388 DAG.getConstant(i, MVT::i32));
2396 // Lower VECTOR_SHUFFLE into SHF (if possible).
2398 // SHF splits the vector into blocks of four elements, then shuffles these
2399 // elements according to a <4 x i2> constant (encoded as an integer immediate).
2401 // It is therefore possible to lower into SHF when the mask takes the form:
2402 // <a, b, c, d, a+4, b+4, c+4, d+4, a+8, b+8, c+8, d+8, ...>
2403 // When undef's appear they are treated as if they were whatever value is
2404 // necessary in order to fit the above form.
2407 // %2 = shufflevector <8 x i16> %0, <8 x i16> undef,
2408 // <8 x i32> <i32 3, i32 2, i32 1, i32 0,
2409 // i32 7, i32 6, i32 5, i32 4>
2411 // (SHF_H $w0, $w1, 27)
2412 // where the 27 comes from:
2413 // 3 + (2 << 2) + (1 << 4) + (0 << 6)
2414 static SDValue lowerVECTOR_SHUFFLE_SHF(SDValue Op, EVT ResTy,
2415 SmallVector<int, 16> Indices,
2416 SelectionDAG &DAG) {
2417 int SHFIndices[4] = { -1, -1, -1, -1 };
2419 if (Indices.size() < 4)
2422 for (unsigned i = 0; i < 4; ++i) {
2423 for (unsigned j = i; j < Indices.size(); j += 4) {
2424 int Idx = Indices[j];
2426 // Convert from vector index to 4-element subvector index
2427 // If an index refers to an element outside of the subvector then give up
2430 if (Idx < 0 || Idx >= 4)
2434 // If the mask has an undef, replace it with the current index.
2435 // Note that it might still be undef if the current index is also undef
2436 if (SHFIndices[i] == -1)
2437 SHFIndices[i] = Idx;
2439 // Check that non-undef values are the same as in the mask. If they
2440 // aren't then give up
2441 if (!(Idx == -1 || Idx == SHFIndices[i]))
2446 // Calculate the immediate. Replace any remaining undefs with zero
2448 for (int i = 3; i >= 0; --i) {
2449 int Idx = SHFIndices[i];
2458 return DAG.getNode(MipsISD::SHF, SDLoc(Op), ResTy,
2459 DAG.getConstant(Imm, MVT::i32), Op->getOperand(0));
2462 // Lower VECTOR_SHUFFLE into ILVEV (if possible).
2464 // ILVEV interleaves the even elements from each vector.
2466 // It is possible to lower into ILVEV when the mask takes the form:
2467 // <0, n, 2, n+2, 4, n+4, ...>
2468 // where n is the number of elements in the vector.
2470 // When undef's appear in the mask they are treated as if they were whatever
2471 // value is necessary in order to fit the above form.
2472 static SDValue lowerVECTOR_SHUFFLE_ILVEV(SDValue Op, EVT ResTy,
2473 SmallVector<int, 16> Indices,
2474 SelectionDAG &DAG) {
2475 assert ((Indices.size() % 2) == 0);
2477 int WtIdx = ResTy.getVectorNumElements();
2479 for (unsigned i = 0; i < Indices.size(); i += 2) {
2480 if (Indices[i] != -1 && Indices[i] != WsIdx)
2482 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
2488 return DAG.getNode(MipsISD::ILVEV, SDLoc(Op), ResTy, Op->getOperand(0),
2492 // Lower VECTOR_SHUFFLE into ILVOD (if possible).
2494 // ILVOD interleaves the odd elements from each vector.
2496 // It is possible to lower into ILVOD when the mask takes the form:
2497 // <1, n+1, 3, n+3, 5, n+5, ...>
2498 // where n is the number of elements in the vector.
2500 // When undef's appear in the mask they are treated as if they were whatever
2501 // value is necessary in order to fit the above form.
2502 static SDValue lowerVECTOR_SHUFFLE_ILVOD(SDValue Op, EVT ResTy,
2503 SmallVector<int, 16> Indices,
2504 SelectionDAG &DAG) {
2505 assert ((Indices.size() % 2) == 0);
2507 int WtIdx = ResTy.getVectorNumElements() + 1;
2509 for (unsigned i = 0; i < Indices.size(); i += 2) {
2510 if (Indices[i] != -1 && Indices[i] != WsIdx)
2512 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
2518 return DAG.getNode(MipsISD::ILVOD, SDLoc(Op), ResTy, Op->getOperand(0),
2522 // Lower VECTOR_SHUFFLE into ILVL (if possible).
2524 // ILVL interleaves consecutive elements from the left half of each vector.
2526 // It is possible to lower into ILVL when the mask takes the form:
2527 // <0, n, 1, n+1, 2, n+2, ...>
2528 // where n is the number of elements in the vector.
2530 // When undef's appear in the mask they are treated as if they were whatever
2531 // value is necessary in order to fit the above form.
2532 static SDValue lowerVECTOR_SHUFFLE_ILVL(SDValue Op, EVT ResTy,
2533 SmallVector<int, 16> Indices,
2534 SelectionDAG &DAG) {
2535 assert ((Indices.size() % 2) == 0);
2537 int WtIdx = ResTy.getVectorNumElements();
2539 for (unsigned i = 0; i < Indices.size(); i += 2) {
2540 if (Indices[i] != -1 && Indices[i] != WsIdx)
2542 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
2548 return DAG.getNode(MipsISD::ILVL, SDLoc(Op), ResTy, Op->getOperand(0),
2552 // Lower VECTOR_SHUFFLE into ILVR (if possible).
2554 // ILVR interleaves consecutive elements from the right half of each vector.
2556 // It is possible to lower into ILVR when the mask takes the form:
2557 // <x, n+x, x+1, n+x+1, x+2, n+x+2, ...>
2558 // where n is the number of elements in the vector and x is half n.
2560 // When undef's appear in the mask they are treated as if they were whatever
2561 // value is necessary in order to fit the above form.
2562 static SDValue lowerVECTOR_SHUFFLE_ILVR(SDValue Op, EVT ResTy,
2563 SmallVector<int, 16> Indices,
2564 SelectionDAG &DAG) {
2565 assert ((Indices.size() % 2) == 0);
2566 unsigned NumElts = ResTy.getVectorNumElements();
2567 int WsIdx = NumElts / 2;
2568 int WtIdx = NumElts + NumElts / 2;
2570 for (unsigned i = 0; i < Indices.size(); i += 2) {
2571 if (Indices[i] != -1 && Indices[i] != WsIdx)
2573 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
2579 return DAG.getNode(MipsISD::ILVR, SDLoc(Op), ResTy, Op->getOperand(0),
2583 // Lower VECTOR_SHUFFLE into PCKEV (if possible).
2585 // PCKEV copies the even elements of each vector into the result vector.
2587 // It is possible to lower into PCKEV when the mask takes the form:
2588 // <0, 2, 4, ..., n, n+2, n+4, ...>
2589 // where n is the number of elements in the vector.
2591 // When undef's appear in the mask they are treated as if they were whatever
2592 // value is necessary in order to fit the above form.
2593 static SDValue lowerVECTOR_SHUFFLE_PCKEV(SDValue Op, EVT ResTy,
2594 SmallVector<int, 16> Indices,
2595 SelectionDAG &DAG) {
2596 assert ((Indices.size() % 2) == 0);
2599 for (unsigned i = 0; i < Indices.size(); ++i) {
2600 if (Indices[i] != -1 && Indices[i] != Idx)
2605 return DAG.getNode(MipsISD::PCKEV, SDLoc(Op), ResTy, Op->getOperand(0),
2609 // Lower VECTOR_SHUFFLE into PCKOD (if possible).
2611 // PCKOD copies the odd elements of each vector into the result vector.
2613 // It is possible to lower into PCKOD when the mask takes the form:
2614 // <1, 3, 5, ..., n+1, n+3, n+5, ...>
2615 // where n is the number of elements in the vector.
2617 // When undef's appear in the mask they are treated as if they were whatever
2618 // value is necessary in order to fit the above form.
2619 static SDValue lowerVECTOR_SHUFFLE_PCKOD(SDValue Op, EVT ResTy,
2620 SmallVector<int, 16> Indices,
2621 SelectionDAG &DAG) {
2622 assert ((Indices.size() % 2) == 0);
2625 for (unsigned i = 0; i < Indices.size(); ++i) {
2626 if (Indices[i] != -1 && Indices[i] != Idx)
2631 return DAG.getNode(MipsISD::PCKOD, SDLoc(Op), ResTy, Op->getOperand(0),
2635 // Lower VECTOR_SHUFFLE into VSHF.
2637 // This mostly consists of converting the shuffle indices in Indices into a
2638 // BUILD_VECTOR and adding it as an operand to the resulting VSHF. There is
2639 // also code to eliminate unused operands of the VECTOR_SHUFFLE. For example,
2640 // if the type is v8i16 and all the indices are less than 8 then the second
2641 // operand is unused and can be replaced with anything. We choose to replace it
2642 // with the used operand since this reduces the number of instructions overall.
2643 static SDValue lowerVECTOR_SHUFFLE_VSHF(SDValue Op, EVT ResTy,
2644 SmallVector<int, 16> Indices,
2645 SelectionDAG &DAG) {
2646 SmallVector<SDValue, 16> Ops;
2649 EVT MaskVecTy = ResTy.changeVectorElementTypeToInteger();
2650 EVT MaskEltTy = MaskVecTy.getVectorElementType();
2651 bool Using1stVec = false;
2652 bool Using2ndVec = false;
2654 int ResTyNumElts = ResTy.getVectorNumElements();
2656 for (int i = 0; i < ResTyNumElts; ++i) {
2657 // Idx == -1 means UNDEF
2658 int Idx = Indices[i];
2660 if (0 <= Idx && Idx < ResTyNumElts)
2662 if (ResTyNumElts <= Idx && Idx < ResTyNumElts * 2)
2666 for (SmallVector<int, 16>::iterator I = Indices.begin(); I != Indices.end();
2668 Ops.push_back(DAG.getTargetConstant(*I, MaskEltTy));
2670 SDValue MaskVec = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecTy, Ops);
2672 if (Using1stVec && Using2ndVec) {
2673 Op0 = Op->getOperand(0);
2674 Op1 = Op->getOperand(1);
2675 } else if (Using1stVec)
2676 Op0 = Op1 = Op->getOperand(0);
2677 else if (Using2ndVec)
2678 Op0 = Op1 = Op->getOperand(1);
2680 llvm_unreachable("shuffle vector mask references neither vector operand?");
2682 // VECTOR_SHUFFLE concatenates the vectors in an vectorwise fashion.
2683 // <0b00, 0b01> + <0b10, 0b11> -> <0b00, 0b01, 0b10, 0b11>
2684 // VSHF concatenates the vectors in a bitwise fashion:
2685 // <0b00, 0b01> + <0b10, 0b11> ->
2686 // 0b0100 + 0b1110 -> 0b01001110
2687 // <0b10, 0b11, 0b00, 0b01>
2688 // We must therefore swap the operands to get the correct result.
2689 return DAG.getNode(MipsISD::VSHF, DL, ResTy, MaskVec, Op1, Op0);
2692 // Lower VECTOR_SHUFFLE into one of a number of instructions depending on the
2693 // indices in the shuffle.
2694 SDValue MipsSETargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
2695 SelectionDAG &DAG) const {
2696 ShuffleVectorSDNode *Node = cast<ShuffleVectorSDNode>(Op);
2697 EVT ResTy = Op->getValueType(0);
2699 if (!ResTy.is128BitVector())
2702 int ResTyNumElts = ResTy.getVectorNumElements();
2703 SmallVector<int, 16> Indices;
2705 for (int i = 0; i < ResTyNumElts; ++i)
2706 Indices.push_back(Node->getMaskElt(i));
2708 SDValue Result = lowerVECTOR_SHUFFLE_SHF(Op, ResTy, Indices, DAG);
2709 if (Result.getNode())
2711 Result = lowerVECTOR_SHUFFLE_ILVEV(Op, ResTy, Indices, DAG);
2712 if (Result.getNode())
2714 Result = lowerVECTOR_SHUFFLE_ILVOD(Op, ResTy, Indices, DAG);
2715 if (Result.getNode())
2717 Result = lowerVECTOR_SHUFFLE_ILVL(Op, ResTy, Indices, DAG);
2718 if (Result.getNode())
2720 Result = lowerVECTOR_SHUFFLE_ILVR(Op, ResTy, Indices, DAG);
2721 if (Result.getNode())
2723 Result = lowerVECTOR_SHUFFLE_PCKEV(Op, ResTy, Indices, DAG);
2724 if (Result.getNode())
2726 Result = lowerVECTOR_SHUFFLE_PCKOD(Op, ResTy, Indices, DAG);
2727 if (Result.getNode())
2729 return lowerVECTOR_SHUFFLE_VSHF(Op, ResTy, Indices, DAG);
2732 MachineBasicBlock * MipsSETargetLowering::
2733 emitBPOSGE32(MachineInstr *MI, MachineBasicBlock *BB) const{
2735 // bposge32_pseudo $vr0
2745 // $vr0 = phi($vr2, $fbb, $vr1, $tbb)
2747 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2748 const TargetInstrInfo *TII =
2749 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2750 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
2751 DebugLoc DL = MI->getDebugLoc();
2752 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2753 MachineFunction::iterator It = std::next(MachineFunction::iterator(BB));
2754 MachineFunction *F = BB->getParent();
2755 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
2756 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
2757 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
2760 F->insert(It, Sink);
2762 // Transfer the remainder of BB and its successor edges to Sink.
2763 Sink->splice(Sink->begin(), BB, std::next(MachineBasicBlock::iterator(MI)),
2765 Sink->transferSuccessorsAndUpdatePHIs(BB);
2768 BB->addSuccessor(FBB);
2769 BB->addSuccessor(TBB);
2770 FBB->addSuccessor(Sink);
2771 TBB->addSuccessor(Sink);
2773 // Insert the real bposge32 instruction to $BB.
2774 BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB);
2777 unsigned VR2 = RegInfo.createVirtualRegister(RC);
2778 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2)
2779 .addReg(Mips::ZERO).addImm(0);
2780 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
2783 unsigned VR1 = RegInfo.createVirtualRegister(RC);
2784 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1)
2785 .addReg(Mips::ZERO).addImm(1);
2787 // Insert phi function to $Sink.
2788 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
2789 MI->getOperand(0).getReg())
2790 .addReg(VR2).addMBB(FBB).addReg(VR1).addMBB(TBB);
2792 MI->eraseFromParent(); // The pseudo instruction is gone now.
2796 MachineBasicBlock * MipsSETargetLowering::
2797 emitMSACBranchPseudo(MachineInstr *MI, MachineBasicBlock *BB,
2798 unsigned BranchOp) const{
2800 // vany_nonzero $rd, $ws
2811 // $rd = phi($rd1, $fbb, $rd2, $tbb)
2813 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2814 const TargetInstrInfo *TII =
2815 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2816 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
2817 DebugLoc DL = MI->getDebugLoc();
2818 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2819 MachineFunction::iterator It = std::next(MachineFunction::iterator(BB));
2820 MachineFunction *F = BB->getParent();
2821 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
2822 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
2823 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
2826 F->insert(It, Sink);
2828 // Transfer the remainder of BB and its successor edges to Sink.
2829 Sink->splice(Sink->begin(), BB, std::next(MachineBasicBlock::iterator(MI)),
2831 Sink->transferSuccessorsAndUpdatePHIs(BB);
2834 BB->addSuccessor(FBB);
2835 BB->addSuccessor(TBB);
2836 FBB->addSuccessor(Sink);
2837 TBB->addSuccessor(Sink);
2839 // Insert the real bnz.b instruction to $BB.
2840 BuildMI(BB, DL, TII->get(BranchOp))
2841 .addReg(MI->getOperand(1).getReg())
2845 unsigned RD1 = RegInfo.createVirtualRegister(RC);
2846 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), RD1)
2847 .addReg(Mips::ZERO).addImm(0);
2848 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
2851 unsigned RD2 = RegInfo.createVirtualRegister(RC);
2852 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), RD2)
2853 .addReg(Mips::ZERO).addImm(1);
2855 // Insert phi function to $Sink.
2856 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
2857 MI->getOperand(0).getReg())
2858 .addReg(RD1).addMBB(FBB).addReg(RD2).addMBB(TBB);
2860 MI->eraseFromParent(); // The pseudo instruction is gone now.
2864 // Emit the COPY_FW pseudo instruction.
2866 // copy_fw_pseudo $fd, $ws, n
2868 // copy_u_w $rt, $ws, $n
2871 // When n is zero, the equivalent operation can be performed with (potentially)
2872 // zero instructions due to register overlaps. This optimization is never valid
2873 // for lane 1 because it would require FR=0 mode which isn't supported by MSA.
2874 MachineBasicBlock * MipsSETargetLowering::
2875 emitCOPY_FW(MachineInstr *MI, MachineBasicBlock *BB) const{
2876 const TargetInstrInfo *TII =
2877 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2878 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2879 DebugLoc DL = MI->getDebugLoc();
2880 unsigned Fd = MI->getOperand(0).getReg();
2881 unsigned Ws = MI->getOperand(1).getReg();
2882 unsigned Lane = MI->getOperand(2).getImm();
2885 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_lo);
2887 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
2889 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wt).addReg(Ws).addImm(Lane);
2890 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_lo);
2893 MI->eraseFromParent(); // The pseudo instruction is gone now.
2897 // Emit the COPY_FD pseudo instruction.
2899 // copy_fd_pseudo $fd, $ws, n
2901 // splati.d $wt, $ws, $n
2902 // copy $fd, $wt:sub_64
2904 // When n is zero, the equivalent operation can be performed with (potentially)
2905 // zero instructions due to register overlaps. This optimization is always
2906 // valid because FR=1 mode which is the only supported mode in MSA.
2907 MachineBasicBlock * MipsSETargetLowering::
2908 emitCOPY_FD(MachineInstr *MI, MachineBasicBlock *BB) const{
2909 assert(Subtarget.isFP64bit());
2911 const TargetInstrInfo *TII =
2912 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2913 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2914 unsigned Fd = MI->getOperand(0).getReg();
2915 unsigned Ws = MI->getOperand(1).getReg();
2916 unsigned Lane = MI->getOperand(2).getImm() * 2;
2917 DebugLoc DL = MI->getDebugLoc();
2920 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_64);
2922 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
2924 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wt).addReg(Ws).addImm(1);
2925 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_64);
2928 MI->eraseFromParent(); // The pseudo instruction is gone now.
2932 // Emit the INSERT_FW pseudo instruction.
2934 // insert_fw_pseudo $wd, $wd_in, $n, $fs
2936 // subreg_to_reg $wt:sub_lo, $fs
2937 // insve_w $wd[$n], $wd_in, $wt[0]
2939 MipsSETargetLowering::emitINSERT_FW(MachineInstr *MI,
2940 MachineBasicBlock *BB) const {
2941 const TargetInstrInfo *TII =
2942 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2943 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2944 DebugLoc DL = MI->getDebugLoc();
2945 unsigned Wd = MI->getOperand(0).getReg();
2946 unsigned Wd_in = MI->getOperand(1).getReg();
2947 unsigned Lane = MI->getOperand(2).getImm();
2948 unsigned Fs = MI->getOperand(3).getReg();
2949 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
2951 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
2954 .addImm(Mips::sub_lo);
2955 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_W), Wd)
2961 MI->eraseFromParent(); // The pseudo instruction is gone now.
2965 // Emit the INSERT_FD pseudo instruction.
2967 // insert_fd_pseudo $wd, $fs, n
2969 // subreg_to_reg $wt:sub_64, $fs
2970 // insve_d $wd[$n], $wd_in, $wt[0]
2972 MipsSETargetLowering::emitINSERT_FD(MachineInstr *MI,
2973 MachineBasicBlock *BB) const {
2974 assert(Subtarget.isFP64bit());
2976 const TargetInstrInfo *TII =
2977 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2978 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2979 DebugLoc DL = MI->getDebugLoc();
2980 unsigned Wd = MI->getOperand(0).getReg();
2981 unsigned Wd_in = MI->getOperand(1).getReg();
2982 unsigned Lane = MI->getOperand(2).getImm();
2983 unsigned Fs = MI->getOperand(3).getReg();
2984 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
2986 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
2989 .addImm(Mips::sub_64);
2990 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_D), Wd)
2996 MI->eraseFromParent(); // The pseudo instruction is gone now.
3000 // Emit the INSERT_([BHWD]|F[WD])_VIDX pseudo instruction.
3003 // (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $rs)
3005 // (SLL $lanetmp1, $lane, <log2size)
3006 // (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
3007 // (INSERT_[BHWD], $wdtmp2, $wdtmp1, 0, $rs)
3008 // (NEG $lanetmp2, $lanetmp1)
3009 // (SLD_B $wd, $wdtmp2, $wdtmp2, $lanetmp2)
3011 // For floating point:
3012 // (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $fs)
3014 // (SUBREG_TO_REG $wt, $fs, <subreg>)
3015 // (SLL $lanetmp1, $lane, <log2size)
3016 // (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
3017 // (INSVE_[WD], $wdtmp2, 0, $wdtmp1, 0)
3018 // (NEG $lanetmp2, $lanetmp1)
3019 // (SLD_B $wd, $wdtmp2, $wdtmp2, $lanetmp2)
3021 MipsSETargetLowering::emitINSERT_DF_VIDX(MachineInstr *MI,
3022 MachineBasicBlock *BB,
3023 unsigned EltSizeInBytes,
3025 const TargetInstrInfo *TII =
3026 getTargetMachine().getSubtargetImpl()->getInstrInfo();
3027 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3028 DebugLoc DL = MI->getDebugLoc();
3029 unsigned Wd = MI->getOperand(0).getReg();
3030 unsigned SrcVecReg = MI->getOperand(1).getReg();
3031 unsigned LaneReg = MI->getOperand(2).getReg();
3032 unsigned SrcValReg = MI->getOperand(3).getReg();
3034 const TargetRegisterClass *VecRC = nullptr;
3035 const TargetRegisterClass *GPRRC =
3036 Subtarget.isGP64bit() ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
3037 unsigned EltLog2Size;
3038 unsigned InsertOp = 0;
3039 unsigned InsveOp = 0;
3040 switch (EltSizeInBytes) {
3042 llvm_unreachable("Unexpected size");
3045 InsertOp = Mips::INSERT_B;
3046 InsveOp = Mips::INSVE_B;
3047 VecRC = &Mips::MSA128BRegClass;
3051 InsertOp = Mips::INSERT_H;
3052 InsveOp = Mips::INSVE_H;
3053 VecRC = &Mips::MSA128HRegClass;
3057 InsertOp = Mips::INSERT_W;
3058 InsveOp = Mips::INSVE_W;
3059 VecRC = &Mips::MSA128WRegClass;
3063 InsertOp = Mips::INSERT_D;
3064 InsveOp = Mips::INSVE_D;
3065 VecRC = &Mips::MSA128DRegClass;
3070 unsigned Wt = RegInfo.createVirtualRegister(VecRC);
3071 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
3074 .addImm(EltSizeInBytes == 8 ? Mips::sub_64 : Mips::sub_lo);
3078 // Convert the lane index into a byte index
3079 if (EltSizeInBytes != 1) {
3080 unsigned LaneTmp1 = RegInfo.createVirtualRegister(GPRRC);
3081 BuildMI(*BB, MI, DL, TII->get(Mips::SLL), LaneTmp1)
3083 .addImm(EltLog2Size);
3087 // Rotate bytes around so that the desired lane is element zero
3088 unsigned WdTmp1 = RegInfo.createVirtualRegister(VecRC);
3089 BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), WdTmp1)
3094 unsigned WdTmp2 = RegInfo.createVirtualRegister(VecRC);
3096 // Use insve.df to insert to element zero
3097 BuildMI(*BB, MI, DL, TII->get(InsveOp), WdTmp2)
3103 // Use insert.df to insert to element zero
3104 BuildMI(*BB, MI, DL, TII->get(InsertOp), WdTmp2)
3110 // Rotate elements the rest of the way for a full rotation.
3111 // sld.df inteprets $rt modulo the number of columns so we only need to negate
3112 // the lane index to do this.
3113 unsigned LaneTmp2 = RegInfo.createVirtualRegister(GPRRC);
3114 BuildMI(*BB, MI, DL, TII->get(Mips::SUB), LaneTmp2)
3117 BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), Wd)
3122 MI->eraseFromParent(); // The pseudo instruction is gone now.
3126 // Emit the FILL_FW pseudo instruction.
3128 // fill_fw_pseudo $wd, $fs
3130 // implicit_def $wt1
3131 // insert_subreg $wt2:subreg_lo, $wt1, $fs
3132 // splati.w $wd, $wt2[0]
3134 MipsSETargetLowering::emitFILL_FW(MachineInstr *MI,
3135 MachineBasicBlock *BB) const {
3136 const TargetInstrInfo *TII =
3137 getTargetMachine().getSubtargetImpl()->getInstrInfo();
3138 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3139 DebugLoc DL = MI->getDebugLoc();
3140 unsigned Wd = MI->getOperand(0).getReg();
3141 unsigned Fs = MI->getOperand(1).getReg();
3142 unsigned Wt1 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3143 unsigned Wt2 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3145 BuildMI(*BB, MI, DL, TII->get(Mips::IMPLICIT_DEF), Wt1);
3146 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2)
3149 .addImm(Mips::sub_lo);
3150 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wd).addReg(Wt2).addImm(0);
3152 MI->eraseFromParent(); // The pseudo instruction is gone now.
3156 // Emit the FILL_FD pseudo instruction.
3158 // fill_fd_pseudo $wd, $fs
3160 // implicit_def $wt1
3161 // insert_subreg $wt2:subreg_64, $wt1, $fs
3162 // splati.d $wd, $wt2[0]
3164 MipsSETargetLowering::emitFILL_FD(MachineInstr *MI,
3165 MachineBasicBlock *BB) const {
3166 assert(Subtarget.isFP64bit());
3168 const TargetInstrInfo *TII =
3169 getTargetMachine().getSubtargetImpl()->getInstrInfo();
3170 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3171 DebugLoc DL = MI->getDebugLoc();
3172 unsigned Wd = MI->getOperand(0).getReg();
3173 unsigned Fs = MI->getOperand(1).getReg();
3174 unsigned Wt1 = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3175 unsigned Wt2 = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3177 BuildMI(*BB, MI, DL, TII->get(Mips::IMPLICIT_DEF), Wt1);
3178 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2)
3181 .addImm(Mips::sub_64);
3182 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wd).addReg(Wt2).addImm(0);
3184 MI->eraseFromParent(); // The pseudo instruction is gone now.
3188 // Emit the FEXP2_W_1 pseudo instructions.
3190 // fexp2_w_1_pseudo $wd, $wt
3193 // fexp2.w $wd, $ws, $wt
3195 MipsSETargetLowering::emitFEXP2_W_1(MachineInstr *MI,
3196 MachineBasicBlock *BB) const {
3197 const TargetInstrInfo *TII =
3198 getTargetMachine().getSubtargetImpl()->getInstrInfo();
3199 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3200 const TargetRegisterClass *RC = &Mips::MSA128WRegClass;
3201 unsigned Ws1 = RegInfo.createVirtualRegister(RC);
3202 unsigned Ws2 = RegInfo.createVirtualRegister(RC);
3203 DebugLoc DL = MI->getDebugLoc();
3205 // Splat 1.0 into a vector
3206 BuildMI(*BB, MI, DL, TII->get(Mips::LDI_W), Ws1).addImm(1);
3207 BuildMI(*BB, MI, DL, TII->get(Mips::FFINT_U_W), Ws2).addReg(Ws1);
3209 // Emit 1.0 * fexp2(Wt)
3210 BuildMI(*BB, MI, DL, TII->get(Mips::FEXP2_W), MI->getOperand(0).getReg())
3212 .addReg(MI->getOperand(1).getReg());
3214 MI->eraseFromParent(); // The pseudo instruction is gone now.
3218 // Emit the FEXP2_D_1 pseudo instructions.
3220 // fexp2_d_1_pseudo $wd, $wt
3223 // fexp2.d $wd, $ws, $wt
3225 MipsSETargetLowering::emitFEXP2_D_1(MachineInstr *MI,
3226 MachineBasicBlock *BB) const {
3227 const TargetInstrInfo *TII =
3228 getTargetMachine().getSubtargetImpl()->getInstrInfo();
3229 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3230 const TargetRegisterClass *RC = &Mips::MSA128DRegClass;
3231 unsigned Ws1 = RegInfo.createVirtualRegister(RC);
3232 unsigned Ws2 = RegInfo.createVirtualRegister(RC);
3233 DebugLoc DL = MI->getDebugLoc();
3235 // Splat 1.0 into a vector
3236 BuildMI(*BB, MI, DL, TII->get(Mips::LDI_D), Ws1).addImm(1);
3237 BuildMI(*BB, MI, DL, TII->get(Mips::FFINT_U_D), Ws2).addReg(Ws1);
3239 // Emit 1.0 * fexp2(Wt)
3240 BuildMI(*BB, MI, DL, TII->get(Mips::FEXP2_D), MI->getOperand(0).getReg())
3242 .addReg(MI->getOperand(1).getReg());
3244 MI->eraseFromParent(); // The pseudo instruction is gone now.