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 "MipsMachineFunction.h"
15 #include "MipsRegisterInfo.h"
16 #include "MipsTargetMachine.h"
17 #include "llvm/CodeGen/MachineInstrBuilder.h"
18 #include "llvm/CodeGen/MachineRegisterInfo.h"
19 #include "llvm/IR/Intrinsics.h"
20 #include "llvm/Support/CommandLine.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Support/raw_ostream.h"
23 #include "llvm/Target/TargetInstrInfo.h"
27 #define DEBUG_TYPE "mips-isel"
30 EnableMipsTailCalls("enable-mips-tail-calls", cl::Hidden,
31 cl::desc("MIPS: Enable tail calls."), cl::init(false));
33 static cl::opt<bool> NoDPLoadStore("mno-ldc1-sdc1", cl::init(false),
34 cl::desc("Expand double precision loads and "
35 "stores to their single precision "
38 MipsSETargetLowering::MipsSETargetLowering(const MipsTargetMachine &TM,
39 const MipsSubtarget &STI)
40 : MipsTargetLowering(TM, STI) {
41 // Set up the register classes
42 addRegisterClass(MVT::i32, &Mips::GPR32RegClass);
44 if (Subtarget.isGP64bit())
45 addRegisterClass(MVT::i64, &Mips::GPR64RegClass);
47 if (Subtarget.hasDSP() || Subtarget.hasMSA()) {
48 // Expand all truncating stores and extending loads.
49 unsigned FirstVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
50 unsigned LastVT = (unsigned)MVT::LAST_VECTOR_VALUETYPE;
52 for (unsigned VT0 = FirstVT; VT0 <= LastVT; ++VT0) {
53 for (unsigned VT1 = FirstVT; VT1 <= LastVT; ++VT1)
54 setTruncStoreAction((MVT::SimpleValueType)VT0,
55 (MVT::SimpleValueType)VT1, Expand);
57 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT0, Expand);
58 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT0, Expand);
59 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT0, Expand);
63 if (Subtarget.hasDSP()) {
64 MVT::SimpleValueType VecTys[2] = {MVT::v2i16, MVT::v4i8};
66 for (unsigned i = 0; i < array_lengthof(VecTys); ++i) {
67 addRegisterClass(VecTys[i], &Mips::DSPRRegClass);
69 // Expand all builtin opcodes.
70 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
71 setOperationAction(Opc, VecTys[i], Expand);
73 setOperationAction(ISD::ADD, VecTys[i], Legal);
74 setOperationAction(ISD::SUB, VecTys[i], Legal);
75 setOperationAction(ISD::LOAD, VecTys[i], Legal);
76 setOperationAction(ISD::STORE, VecTys[i], Legal);
77 setOperationAction(ISD::BITCAST, VecTys[i], Legal);
80 setTargetDAGCombine(ISD::SHL);
81 setTargetDAGCombine(ISD::SRA);
82 setTargetDAGCombine(ISD::SRL);
83 setTargetDAGCombine(ISD::SETCC);
84 setTargetDAGCombine(ISD::VSELECT);
87 if (Subtarget.hasDSPR2())
88 setOperationAction(ISD::MUL, MVT::v2i16, Legal);
90 if (Subtarget.hasMSA()) {
91 addMSAIntType(MVT::v16i8, &Mips::MSA128BRegClass);
92 addMSAIntType(MVT::v8i16, &Mips::MSA128HRegClass);
93 addMSAIntType(MVT::v4i32, &Mips::MSA128WRegClass);
94 addMSAIntType(MVT::v2i64, &Mips::MSA128DRegClass);
95 addMSAFloatType(MVT::v8f16, &Mips::MSA128HRegClass);
96 addMSAFloatType(MVT::v4f32, &Mips::MSA128WRegClass);
97 addMSAFloatType(MVT::v2f64, &Mips::MSA128DRegClass);
99 setTargetDAGCombine(ISD::AND);
100 setTargetDAGCombine(ISD::OR);
101 setTargetDAGCombine(ISD::SRA);
102 setTargetDAGCombine(ISD::VSELECT);
103 setTargetDAGCombine(ISD::XOR);
106 if (!Subtarget.abiUsesSoftFloat()) {
107 addRegisterClass(MVT::f32, &Mips::FGR32RegClass);
109 // When dealing with single precision only, use libcalls
110 if (!Subtarget.isSingleFloat()) {
111 if (Subtarget.isFP64bit())
112 addRegisterClass(MVT::f64, &Mips::FGR64RegClass);
114 addRegisterClass(MVT::f64, &Mips::AFGR64RegClass);
118 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Custom);
119 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Custom);
120 setOperationAction(ISD::MULHS, MVT::i32, Custom);
121 setOperationAction(ISD::MULHU, MVT::i32, Custom);
123 if (Subtarget.hasCnMips())
124 setOperationAction(ISD::MUL, MVT::i64, Legal);
125 else if (Subtarget.isGP64bit())
126 setOperationAction(ISD::MUL, MVT::i64, Custom);
128 if (Subtarget.isGP64bit()) {
129 setOperationAction(ISD::MULHS, MVT::i64, Custom);
130 setOperationAction(ISD::MULHU, MVT::i64, Custom);
131 setOperationAction(ISD::SDIVREM, MVT::i64, Custom);
132 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
135 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
136 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
138 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
139 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
140 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
141 setOperationAction(ISD::LOAD, MVT::i32, Custom);
142 setOperationAction(ISD::STORE, MVT::i32, Custom);
144 setTargetDAGCombine(ISD::ADDE);
145 setTargetDAGCombine(ISD::SUBE);
146 setTargetDAGCombine(ISD::MUL);
148 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
149 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
150 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
153 setOperationAction(ISD::LOAD, MVT::f64, Custom);
154 setOperationAction(ISD::STORE, MVT::f64, Custom);
157 if (Subtarget.hasMips32r6()) {
158 // MIPS32r6 replaces the accumulator-based multiplies with a three register
160 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
161 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
162 setOperationAction(ISD::MUL, MVT::i32, Legal);
163 setOperationAction(ISD::MULHS, MVT::i32, Legal);
164 setOperationAction(ISD::MULHU, MVT::i32, Legal);
166 // MIPS32r6 replaces the accumulator-based division/remainder with separate
167 // three register division and remainder instructions.
168 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
169 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
170 setOperationAction(ISD::SDIV, MVT::i32, Legal);
171 setOperationAction(ISD::UDIV, MVT::i32, Legal);
172 setOperationAction(ISD::SREM, MVT::i32, Legal);
173 setOperationAction(ISD::UREM, MVT::i32, Legal);
175 // MIPS32r6 replaces conditional moves with an equivalent that removes the
176 // need for three GPR read ports.
177 setOperationAction(ISD::SETCC, MVT::i32, Legal);
178 setOperationAction(ISD::SELECT, MVT::i32, Legal);
179 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
181 setOperationAction(ISD::SETCC, MVT::f32, Legal);
182 setOperationAction(ISD::SELECT, MVT::f32, Legal);
183 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
185 assert(Subtarget.isFP64bit() && "FR=1 is required for MIPS32r6");
186 setOperationAction(ISD::SETCC, MVT::f64, Legal);
187 setOperationAction(ISD::SELECT, MVT::f64, Legal);
188 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
190 setOperationAction(ISD::BRCOND, MVT::Other, Legal);
192 // Floating point > and >= are supported via < and <=
193 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
194 setCondCodeAction(ISD::SETOGT, MVT::f32, Expand);
195 setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
196 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
198 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
199 setCondCodeAction(ISD::SETOGT, MVT::f64, Expand);
200 setCondCodeAction(ISD::SETUGE, MVT::f64, Expand);
201 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
204 if (Subtarget.hasMips64r6()) {
205 // MIPS64r6 replaces the accumulator-based multiplies with a three register
207 setOperationAction(ISD::MUL, MVT::i64, Legal);
208 setOperationAction(ISD::MULHS, MVT::i64, Legal);
209 setOperationAction(ISD::MULHU, MVT::i64, Legal);
211 // MIPS32r6 replaces the accumulator-based division/remainder with separate
212 // three register division and remainder instructions.
213 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
214 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
215 setOperationAction(ISD::SDIV, MVT::i64, Legal);
216 setOperationAction(ISD::UDIV, MVT::i64, Legal);
217 setOperationAction(ISD::SREM, MVT::i64, Legal);
218 setOperationAction(ISD::UREM, MVT::i64, Legal);
220 // MIPS64r6 replaces conditional moves with an equivalent that removes the
221 // need for three GPR read ports.
222 setOperationAction(ISD::SETCC, MVT::i64, Legal);
223 setOperationAction(ISD::SELECT, MVT::i64, Legal);
224 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
227 computeRegisterProperties();
230 const MipsTargetLowering *
231 llvm::createMipsSETargetLowering(const MipsTargetMachine &TM,
232 const MipsSubtarget &STI) {
233 return new MipsSETargetLowering(TM, STI);
236 const TargetRegisterClass *
237 MipsSETargetLowering::getRepRegClassFor(MVT VT) const {
238 if (VT == MVT::Untyped)
239 return Subtarget.hasDSP() ? &Mips::ACC64DSPRegClass : &Mips::ACC64RegClass;
241 return TargetLowering::getRepRegClassFor(VT);
244 // Enable MSA support for the given integer type and Register class.
245 void MipsSETargetLowering::
246 addMSAIntType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
247 addRegisterClass(Ty, RC);
249 // Expand all builtin opcodes.
250 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
251 setOperationAction(Opc, Ty, Expand);
253 setOperationAction(ISD::BITCAST, Ty, Legal);
254 setOperationAction(ISD::LOAD, Ty, Legal);
255 setOperationAction(ISD::STORE, Ty, Legal);
256 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Custom);
257 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
258 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
260 setOperationAction(ISD::ADD, Ty, Legal);
261 setOperationAction(ISD::AND, Ty, Legal);
262 setOperationAction(ISD::CTLZ, Ty, Legal);
263 setOperationAction(ISD::CTPOP, Ty, Legal);
264 setOperationAction(ISD::MUL, Ty, Legal);
265 setOperationAction(ISD::OR, Ty, Legal);
266 setOperationAction(ISD::SDIV, Ty, Legal);
267 setOperationAction(ISD::SREM, Ty, Legal);
268 setOperationAction(ISD::SHL, Ty, Legal);
269 setOperationAction(ISD::SRA, Ty, Legal);
270 setOperationAction(ISD::SRL, Ty, Legal);
271 setOperationAction(ISD::SUB, Ty, Legal);
272 setOperationAction(ISD::UDIV, Ty, Legal);
273 setOperationAction(ISD::UREM, Ty, Legal);
274 setOperationAction(ISD::VECTOR_SHUFFLE, Ty, Custom);
275 setOperationAction(ISD::VSELECT, Ty, Legal);
276 setOperationAction(ISD::XOR, Ty, Legal);
278 if (Ty == MVT::v4i32 || Ty == MVT::v2i64) {
279 setOperationAction(ISD::FP_TO_SINT, Ty, Legal);
280 setOperationAction(ISD::FP_TO_UINT, Ty, Legal);
281 setOperationAction(ISD::SINT_TO_FP, Ty, Legal);
282 setOperationAction(ISD::UINT_TO_FP, Ty, Legal);
285 setOperationAction(ISD::SETCC, Ty, Legal);
286 setCondCodeAction(ISD::SETNE, Ty, Expand);
287 setCondCodeAction(ISD::SETGE, Ty, Expand);
288 setCondCodeAction(ISD::SETGT, Ty, Expand);
289 setCondCodeAction(ISD::SETUGE, Ty, Expand);
290 setCondCodeAction(ISD::SETUGT, Ty, Expand);
293 // Enable MSA support for the given floating-point type and Register class.
294 void MipsSETargetLowering::
295 addMSAFloatType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
296 addRegisterClass(Ty, RC);
298 // Expand all builtin opcodes.
299 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
300 setOperationAction(Opc, Ty, Expand);
302 setOperationAction(ISD::LOAD, Ty, Legal);
303 setOperationAction(ISD::STORE, Ty, Legal);
304 setOperationAction(ISD::BITCAST, Ty, Legal);
305 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Legal);
306 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
307 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
309 if (Ty != MVT::v8f16) {
310 setOperationAction(ISD::FABS, Ty, Legal);
311 setOperationAction(ISD::FADD, Ty, Legal);
312 setOperationAction(ISD::FDIV, Ty, Legal);
313 setOperationAction(ISD::FEXP2, Ty, Legal);
314 setOperationAction(ISD::FLOG2, Ty, Legal);
315 setOperationAction(ISD::FMA, Ty, Legal);
316 setOperationAction(ISD::FMUL, Ty, Legal);
317 setOperationAction(ISD::FRINT, Ty, Legal);
318 setOperationAction(ISD::FSQRT, Ty, Legal);
319 setOperationAction(ISD::FSUB, Ty, Legal);
320 setOperationAction(ISD::VSELECT, Ty, Legal);
322 setOperationAction(ISD::SETCC, Ty, Legal);
323 setCondCodeAction(ISD::SETOGE, Ty, Expand);
324 setCondCodeAction(ISD::SETOGT, Ty, Expand);
325 setCondCodeAction(ISD::SETUGE, Ty, Expand);
326 setCondCodeAction(ISD::SETUGT, Ty, Expand);
327 setCondCodeAction(ISD::SETGE, Ty, Expand);
328 setCondCodeAction(ISD::SETGT, Ty, Expand);
333 MipsSETargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
337 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
339 if (Subtarget.systemSupportsUnalignedAccess()) {
340 // MIPS32r6/MIPS64r6 is required to support unaligned access. It's
341 // implementation defined whether this is handled by hardware, software, or
342 // a hybrid of the two but it's expected that most implementations will
343 // handle the majority of cases in hardware.
360 SDValue MipsSETargetLowering::LowerOperation(SDValue Op,
361 SelectionDAG &DAG) const {
362 switch(Op.getOpcode()) {
363 case ISD::LOAD: return lowerLOAD(Op, DAG);
364 case ISD::STORE: return lowerSTORE(Op, DAG);
365 case ISD::SMUL_LOHI: return lowerMulDiv(Op, MipsISD::Mult, true, true, DAG);
366 case ISD::UMUL_LOHI: return lowerMulDiv(Op, MipsISD::Multu, true, true, DAG);
367 case ISD::MULHS: return lowerMulDiv(Op, MipsISD::Mult, false, true, DAG);
368 case ISD::MULHU: return lowerMulDiv(Op, MipsISD::Multu, false, true, DAG);
369 case ISD::MUL: return lowerMulDiv(Op, MipsISD::Mult, true, false, DAG);
370 case ISD::SDIVREM: return lowerMulDiv(Op, MipsISD::DivRem, true, true, DAG);
371 case ISD::UDIVREM: return lowerMulDiv(Op, MipsISD::DivRemU, true, true,
373 case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG);
374 case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG);
375 case ISD::INTRINSIC_VOID: return lowerINTRINSIC_VOID(Op, DAG);
376 case ISD::EXTRACT_VECTOR_ELT: return lowerEXTRACT_VECTOR_ELT(Op, DAG);
377 case ISD::BUILD_VECTOR: return lowerBUILD_VECTOR(Op, DAG);
378 case ISD::VECTOR_SHUFFLE: return lowerVECTOR_SHUFFLE(Op, DAG);
381 return MipsTargetLowering::LowerOperation(Op, DAG);
385 // Transforms a subgraph in CurDAG if the following pattern is found:
386 // (addc multLo, Lo0), (adde multHi, Hi0),
388 // multHi/Lo: product of multiplication
389 // Lo0: initial value of Lo register
390 // Hi0: initial value of Hi register
391 // Return true if pattern matching was successful.
392 static bool selectMADD(SDNode *ADDENode, SelectionDAG *CurDAG) {
393 // ADDENode's second operand must be a flag output of an ADDC node in order
394 // for the matching to be successful.
395 SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
397 if (ADDCNode->getOpcode() != ISD::ADDC)
400 SDValue MultHi = ADDENode->getOperand(0);
401 SDValue MultLo = ADDCNode->getOperand(0);
402 SDNode *MultNode = MultHi.getNode();
403 unsigned MultOpc = MultHi.getOpcode();
405 // MultHi and MultLo must be generated by the same node,
406 if (MultLo.getNode() != MultNode)
409 // and it must be a multiplication.
410 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
413 // MultLo amd MultHi must be the first and second output of MultNode
415 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
418 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
419 // of the values of MultNode, in which case MultNode will be removed in later
421 // If there exist users other than ADDENode or ADDCNode, this function returns
422 // here, which will result in MultNode being mapped to a single MULT
423 // instruction node rather than a pair of MULT and MADD instructions being
425 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
430 // Initialize accumulator.
431 SDValue ACCIn = CurDAG->getNode(MipsISD::MTLOHI, DL, MVT::Untyped,
432 ADDCNode->getOperand(1),
433 ADDENode->getOperand(1));
435 // create MipsMAdd(u) node
436 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
438 SDValue MAdd = CurDAG->getNode(MultOpc, DL, MVT::Untyped,
439 MultNode->getOperand(0),// Factor 0
440 MultNode->getOperand(1),// Factor 1
443 // replace uses of adde and addc here
444 if (!SDValue(ADDCNode, 0).use_empty()) {
445 SDValue LoOut = CurDAG->getNode(MipsISD::MFLO, DL, MVT::i32, MAdd);
446 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), LoOut);
448 if (!SDValue(ADDENode, 0).use_empty()) {
449 SDValue HiOut = CurDAG->getNode(MipsISD::MFHI, DL, MVT::i32, MAdd);
450 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), HiOut);
457 // Transforms a subgraph in CurDAG if the following pattern is found:
458 // (addc Lo0, multLo), (sube Hi0, multHi),
460 // multHi/Lo: product of multiplication
461 // Lo0: initial value of Lo register
462 // Hi0: initial value of Hi register
463 // Return true if pattern matching was successful.
464 static bool selectMSUB(SDNode *SUBENode, SelectionDAG *CurDAG) {
465 // SUBENode's second operand must be a flag output of an SUBC node in order
466 // for the matching to be successful.
467 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
469 if (SUBCNode->getOpcode() != ISD::SUBC)
472 SDValue MultHi = SUBENode->getOperand(1);
473 SDValue MultLo = SUBCNode->getOperand(1);
474 SDNode *MultNode = MultHi.getNode();
475 unsigned MultOpc = MultHi.getOpcode();
477 // MultHi and MultLo must be generated by the same node,
478 if (MultLo.getNode() != MultNode)
481 // and it must be a multiplication.
482 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
485 // MultLo amd MultHi must be the first and second output of MultNode
487 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
490 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
491 // of the values of MultNode, in which case MultNode will be removed in later
493 // If there exist users other than SUBENode or SUBCNode, this function returns
494 // here, which will result in MultNode being mapped to a single MULT
495 // instruction node rather than a pair of MULT and MSUB instructions being
497 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
502 // Initialize accumulator.
503 SDValue ACCIn = CurDAG->getNode(MipsISD::MTLOHI, DL, MVT::Untyped,
504 SUBCNode->getOperand(0),
505 SUBENode->getOperand(0));
507 // create MipsSub(u) node
508 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
510 SDValue MSub = CurDAG->getNode(MultOpc, DL, MVT::Glue,
511 MultNode->getOperand(0),// Factor 0
512 MultNode->getOperand(1),// Factor 1
515 // replace uses of sube and subc here
516 if (!SDValue(SUBCNode, 0).use_empty()) {
517 SDValue LoOut = CurDAG->getNode(MipsISD::MFLO, DL, MVT::i32, MSub);
518 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), LoOut);
520 if (!SDValue(SUBENode, 0).use_empty()) {
521 SDValue HiOut = CurDAG->getNode(MipsISD::MFHI, DL, MVT::i32, MSub);
522 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), HiOut);
528 static SDValue performADDECombine(SDNode *N, SelectionDAG &DAG,
529 TargetLowering::DAGCombinerInfo &DCI,
530 const MipsSubtarget &Subtarget) {
531 if (DCI.isBeforeLegalize())
534 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() &&
535 N->getValueType(0) == MVT::i32 && selectMADD(N, &DAG))
536 return SDValue(N, 0);
541 // Fold zero extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT
543 // Performs the following transformations:
544 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to zero extension if its
545 // sign/zero-extension is completely overwritten by the new one performed by
547 // - Removes redundant zero extensions performed by an ISD::AND.
548 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
549 TargetLowering::DAGCombinerInfo &DCI,
550 const MipsSubtarget &Subtarget) {
551 if (!Subtarget.hasMSA())
554 SDValue Op0 = N->getOperand(0);
555 SDValue Op1 = N->getOperand(1);
556 unsigned Op0Opcode = Op0->getOpcode();
558 // (and (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d)
559 // where $d + 1 == 2^n and n == 32
560 // or $d + 1 == 2^n and n <= 32 and ZExt
561 // -> (MipsVExtractZExt $a, $b, $c)
562 if (Op0Opcode == MipsISD::VEXTRACT_SEXT_ELT ||
563 Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT) {
564 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(Op1);
569 int32_t Log2IfPositive = (Mask->getAPIntValue() + 1).exactLogBase2();
571 if (Log2IfPositive <= 0)
572 return SDValue(); // Mask+1 is not a power of 2
574 SDValue Op0Op2 = Op0->getOperand(2);
575 EVT ExtendTy = cast<VTSDNode>(Op0Op2)->getVT();
576 unsigned ExtendTySize = ExtendTy.getSizeInBits();
577 unsigned Log2 = Log2IfPositive;
579 if ((Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT && Log2 >= ExtendTySize) ||
580 Log2 == ExtendTySize) {
581 SDValue Ops[] = { Op0->getOperand(0), Op0->getOperand(1), Op0Op2 };
582 return DAG.getNode(MipsISD::VEXTRACT_ZEXT_ELT, SDLoc(Op0),
584 makeArrayRef(Ops, Op0->getNumOperands()));
591 // Determine if the specified node is a constant vector splat.
593 // Returns true and sets Imm if:
594 // * N is a ISD::BUILD_VECTOR representing a constant splat
596 // This function is quite similar to MipsSEDAGToDAGISel::selectVSplat. The
597 // differences are that it assumes the MSA has already been checked and the
598 // arbitrary requirement for a maximum of 32-bit integers isn't applied (and
599 // must not be in order for binsri.d to be selectable).
600 static bool isVSplat(SDValue N, APInt &Imm, bool IsLittleEndian) {
601 BuildVectorSDNode *Node = dyn_cast<BuildVectorSDNode>(N.getNode());
606 APInt SplatValue, SplatUndef;
607 unsigned SplatBitSize;
610 if (!Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
619 // Test whether the given node is an all-ones build_vector.
620 static bool isVectorAllOnes(SDValue N) {
621 // Look through bitcasts. Endianness doesn't matter because we are looking
622 // for an all-ones value.
623 if (N->getOpcode() == ISD::BITCAST)
624 N = N->getOperand(0);
626 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
631 APInt SplatValue, SplatUndef;
632 unsigned SplatBitSize;
635 // Endianness doesn't matter in this context because we are looking for
636 // an all-ones value.
637 if (BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs))
638 return SplatValue.isAllOnesValue();
643 // Test whether N is the bitwise inverse of OfNode.
644 static bool isBitwiseInverse(SDValue N, SDValue OfNode) {
645 if (N->getOpcode() != ISD::XOR)
648 if (isVectorAllOnes(N->getOperand(0)))
649 return N->getOperand(1) == OfNode;
651 if (isVectorAllOnes(N->getOperand(1)))
652 return N->getOperand(0) == OfNode;
657 // Perform combines where ISD::OR is the root node.
659 // Performs the following transformations:
660 // - (or (and $a, $mask), (and $b, $inv_mask)) => (vselect $mask, $a, $b)
661 // where $inv_mask is the bitwise inverse of $mask and the 'or' has a 128-bit
663 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
664 TargetLowering::DAGCombinerInfo &DCI,
665 const MipsSubtarget &Subtarget) {
666 if (!Subtarget.hasMSA())
669 EVT Ty = N->getValueType(0);
671 if (!Ty.is128BitVector())
674 SDValue Op0 = N->getOperand(0);
675 SDValue Op1 = N->getOperand(1);
677 if (Op0->getOpcode() == ISD::AND && Op1->getOpcode() == ISD::AND) {
678 SDValue Op0Op0 = Op0->getOperand(0);
679 SDValue Op0Op1 = Op0->getOperand(1);
680 SDValue Op1Op0 = Op1->getOperand(0);
681 SDValue Op1Op1 = Op1->getOperand(1);
682 bool IsLittleEndian = !Subtarget.isLittle();
684 SDValue IfSet, IfClr, Cond;
685 bool IsConstantMask = false;
688 // If Op0Op0 is an appropriate mask, try to find it's inverse in either
689 // Op1Op0, or Op1Op1. Keep track of the Cond, IfSet, and IfClr nodes, while
691 // IfClr will be set if we find a valid match.
692 if (isVSplat(Op0Op0, Mask, IsLittleEndian)) {
696 if (isVSplat(Op1Op0, InvMask, IsLittleEndian) &&
697 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
699 else if (isVSplat(Op1Op1, InvMask, IsLittleEndian) &&
700 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
703 IsConstantMask = true;
706 // If IfClr is not yet set, and Op0Op1 is an appropriate mask, try the same
707 // thing again using this mask.
708 // IfClr will be set if we find a valid match.
709 if (!IfClr.getNode() && isVSplat(Op0Op1, Mask, IsLittleEndian)) {
713 if (isVSplat(Op1Op0, InvMask, IsLittleEndian) &&
714 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
716 else if (isVSplat(Op1Op1, InvMask, IsLittleEndian) &&
717 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
720 IsConstantMask = true;
723 // If IfClr is not yet set, try looking for a non-constant match.
724 // IfClr will be set if we find a valid match amongst the eight
726 if (!IfClr.getNode()) {
727 if (isBitwiseInverse(Op0Op0, Op1Op0)) {
731 } else if (isBitwiseInverse(Op0Op1, Op1Op0)) {
735 } else if (isBitwiseInverse(Op0Op0, Op1Op1)) {
739 } else if (isBitwiseInverse(Op0Op1, Op1Op1)) {
743 } else if (isBitwiseInverse(Op1Op0, Op0Op0)) {
747 } else if (isBitwiseInverse(Op1Op1, Op0Op0)) {
751 } else if (isBitwiseInverse(Op1Op0, Op0Op1)) {
755 } else if (isBitwiseInverse(Op1Op1, Op0Op1)) {
762 // At this point, IfClr will be set if we have a valid match.
763 if (!IfClr.getNode())
766 assert(Cond.getNode() && IfSet.getNode());
768 // Fold degenerate cases.
769 if (IsConstantMask) {
770 if (Mask.isAllOnesValue())
776 // Transform the DAG into an equivalent VSELECT.
777 return DAG.getNode(ISD::VSELECT, SDLoc(N), Ty, Cond, IfSet, IfClr);
783 static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG,
784 TargetLowering::DAGCombinerInfo &DCI,
785 const MipsSubtarget &Subtarget) {
786 if (DCI.isBeforeLegalize())
789 if (Subtarget.hasMips32() && N->getValueType(0) == MVT::i32 &&
791 return SDValue(N, 0);
796 static SDValue genConstMult(SDValue X, uint64_t C, SDLoc DL, EVT VT,
797 EVT ShiftTy, SelectionDAG &DAG) {
798 // Clear the upper (64 - VT.sizeInBits) bits.
799 C &= ((uint64_t)-1) >> (64 - VT.getSizeInBits());
803 return DAG.getConstant(0, VT);
809 // If c is power of 2, return (shl x, log2(c)).
810 if (isPowerOf2_64(C))
811 return DAG.getNode(ISD::SHL, DL, VT, X,
812 DAG.getConstant(Log2_64(C), ShiftTy));
814 unsigned Log2Ceil = Log2_64_Ceil(C);
815 uint64_t Floor = 1LL << Log2_64(C);
816 uint64_t Ceil = Log2Ceil == 64 ? 0LL : 1LL << Log2Ceil;
818 // If |c - floor_c| <= |c - ceil_c|,
819 // where floor_c = pow(2, floor(log2(c))) and ceil_c = pow(2, ceil(log2(c))),
820 // return (add constMult(x, floor_c), constMult(x, c - floor_c)).
821 if (C - Floor <= Ceil - C) {
822 SDValue Op0 = genConstMult(X, Floor, DL, VT, ShiftTy, DAG);
823 SDValue Op1 = genConstMult(X, C - Floor, DL, VT, ShiftTy, DAG);
824 return DAG.getNode(ISD::ADD, DL, VT, Op0, Op1);
827 // If |c - floor_c| > |c - ceil_c|,
828 // return (sub constMult(x, ceil_c), constMult(x, ceil_c - c)).
829 SDValue Op0 = genConstMult(X, Ceil, DL, VT, ShiftTy, DAG);
830 SDValue Op1 = genConstMult(X, Ceil - C, DL, VT, ShiftTy, DAG);
831 return DAG.getNode(ISD::SUB, DL, VT, Op0, Op1);
834 static SDValue performMULCombine(SDNode *N, SelectionDAG &DAG,
835 const TargetLowering::DAGCombinerInfo &DCI,
836 const MipsSETargetLowering *TL) {
837 EVT VT = N->getValueType(0);
839 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
841 return genConstMult(N->getOperand(0), C->getZExtValue(), SDLoc(N),
842 VT, TL->getScalarShiftAmountTy(VT), DAG);
844 return SDValue(N, 0);
847 static SDValue performDSPShiftCombine(unsigned Opc, SDNode *N, EVT Ty,
849 const MipsSubtarget &Subtarget) {
850 // See if this is a vector splat immediate node.
851 APInt SplatValue, SplatUndef;
852 unsigned SplatBitSize;
854 unsigned EltSize = Ty.getVectorElementType().getSizeInBits();
855 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
857 if (!Subtarget.hasDSP())
861 !BV->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
862 EltSize, !Subtarget.isLittle()) ||
863 (SplatBitSize != EltSize) ||
864 (SplatValue.getZExtValue() >= EltSize))
867 return DAG.getNode(Opc, SDLoc(N), Ty, N->getOperand(0),
868 DAG.getConstant(SplatValue.getZExtValue(), MVT::i32));
871 static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
872 TargetLowering::DAGCombinerInfo &DCI,
873 const MipsSubtarget &Subtarget) {
874 EVT Ty = N->getValueType(0);
876 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
879 return performDSPShiftCombine(MipsISD::SHLL_DSP, N, Ty, DAG, Subtarget);
882 // Fold sign-extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT for MSA and fold
883 // constant splats into MipsISD::SHRA_DSP for DSPr2.
885 // Performs the following transformations:
886 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to sign extension if its
887 // sign/zero-extension is completely overwritten by the new one performed by
888 // the ISD::SRA and ISD::SHL nodes.
889 // - Removes redundant sign extensions performed by an ISD::SRA and ISD::SHL
892 // See performDSPShiftCombine for more information about the transformation
894 static SDValue performSRACombine(SDNode *N, SelectionDAG &DAG,
895 TargetLowering::DAGCombinerInfo &DCI,
896 const MipsSubtarget &Subtarget) {
897 EVT Ty = N->getValueType(0);
899 if (Subtarget.hasMSA()) {
900 SDValue Op0 = N->getOperand(0);
901 SDValue Op1 = N->getOperand(1);
903 // (sra (shl (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d), imm:$d)
904 // where $d + sizeof($c) == 32
905 // or $d + sizeof($c) <= 32 and SExt
906 // -> (MipsVExtractSExt $a, $b, $c)
907 if (Op0->getOpcode() == ISD::SHL && Op1 == Op0->getOperand(1)) {
908 SDValue Op0Op0 = Op0->getOperand(0);
909 ConstantSDNode *ShAmount = dyn_cast<ConstantSDNode>(Op1);
914 if (Op0Op0->getOpcode() != MipsISD::VEXTRACT_SEXT_ELT &&
915 Op0Op0->getOpcode() != MipsISD::VEXTRACT_ZEXT_ELT)
918 EVT ExtendTy = cast<VTSDNode>(Op0Op0->getOperand(2))->getVT();
919 unsigned TotalBits = ShAmount->getZExtValue() + ExtendTy.getSizeInBits();
921 if (TotalBits == 32 ||
922 (Op0Op0->getOpcode() == MipsISD::VEXTRACT_SEXT_ELT &&
924 SDValue Ops[] = { Op0Op0->getOperand(0), Op0Op0->getOperand(1),
925 Op0Op0->getOperand(2) };
926 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, SDLoc(Op0Op0),
928 makeArrayRef(Ops, Op0Op0->getNumOperands()));
933 if ((Ty != MVT::v2i16) && ((Ty != MVT::v4i8) || !Subtarget.hasDSPR2()))
936 return performDSPShiftCombine(MipsISD::SHRA_DSP, N, Ty, DAG, Subtarget);
940 static SDValue performSRLCombine(SDNode *N, SelectionDAG &DAG,
941 TargetLowering::DAGCombinerInfo &DCI,
942 const MipsSubtarget &Subtarget) {
943 EVT Ty = N->getValueType(0);
945 if (((Ty != MVT::v2i16) || !Subtarget.hasDSPR2()) && (Ty != MVT::v4i8))
948 return performDSPShiftCombine(MipsISD::SHRL_DSP, N, Ty, DAG, Subtarget);
951 static bool isLegalDSPCondCode(EVT Ty, ISD::CondCode CC) {
952 bool IsV216 = (Ty == MVT::v2i16);
956 case ISD::SETNE: return true;
960 case ISD::SETGE: return IsV216;
964 case ISD::SETUGE: return !IsV216;
965 default: return false;
969 static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG) {
970 EVT Ty = N->getValueType(0);
972 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
975 if (!isLegalDSPCondCode(Ty, cast<CondCodeSDNode>(N->getOperand(2))->get()))
978 return DAG.getNode(MipsISD::SETCC_DSP, SDLoc(N), Ty, N->getOperand(0),
979 N->getOperand(1), N->getOperand(2));
982 static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG) {
983 EVT Ty = N->getValueType(0);
985 if (Ty.is128BitVector() && Ty.isInteger()) {
986 // Try the following combines:
987 // (vselect (setcc $a, $b, SETLT), $b, $a)) -> (vsmax $a, $b)
988 // (vselect (setcc $a, $b, SETLE), $b, $a)) -> (vsmax $a, $b)
989 // (vselect (setcc $a, $b, SETLT), $a, $b)) -> (vsmin $a, $b)
990 // (vselect (setcc $a, $b, SETLE), $a, $b)) -> (vsmin $a, $b)
991 // (vselect (setcc $a, $b, SETULT), $b, $a)) -> (vumax $a, $b)
992 // (vselect (setcc $a, $b, SETULE), $b, $a)) -> (vumax $a, $b)
993 // (vselect (setcc $a, $b, SETULT), $a, $b)) -> (vumin $a, $b)
994 // (vselect (setcc $a, $b, SETULE), $a, $b)) -> (vumin $a, $b)
995 // SETGT/SETGE/SETUGT/SETUGE variants of these will show up initially but
996 // will be expanded to equivalent SETLT/SETLE/SETULT/SETULE versions by the
998 SDValue Op0 = N->getOperand(0);
1000 if (Op0->getOpcode() != ISD::SETCC)
1003 ISD::CondCode CondCode = cast<CondCodeSDNode>(Op0->getOperand(2))->get();
1006 if (CondCode == ISD::SETLT || CondCode == ISD::SETLE)
1008 else if (CondCode == ISD::SETULT || CondCode == ISD::SETULE)
1013 SDValue Op1 = N->getOperand(1);
1014 SDValue Op2 = N->getOperand(2);
1015 SDValue Op0Op0 = Op0->getOperand(0);
1016 SDValue Op0Op1 = Op0->getOperand(1);
1018 if (Op1 == Op0Op0 && Op2 == Op0Op1)
1019 return DAG.getNode(Signed ? MipsISD::VSMIN : MipsISD::VUMIN, SDLoc(N),
1021 else if (Op1 == Op0Op1 && Op2 == Op0Op0)
1022 return DAG.getNode(Signed ? MipsISD::VSMAX : MipsISD::VUMAX, SDLoc(N),
1024 } else if ((Ty == MVT::v2i16) || (Ty == MVT::v4i8)) {
1025 SDValue SetCC = N->getOperand(0);
1027 if (SetCC.getOpcode() != MipsISD::SETCC_DSP)
1030 return DAG.getNode(MipsISD::SELECT_CC_DSP, SDLoc(N), Ty,
1031 SetCC.getOperand(0), SetCC.getOperand(1),
1032 N->getOperand(1), N->getOperand(2), SetCC.getOperand(2));
1038 static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
1039 const MipsSubtarget &Subtarget) {
1040 EVT Ty = N->getValueType(0);
1042 if (Subtarget.hasMSA() && Ty.is128BitVector() && Ty.isInteger()) {
1043 // Try the following combines:
1044 // (xor (or $a, $b), (build_vector allones))
1045 // (xor (or $a, $b), (bitcast (build_vector allones)))
1046 SDValue Op0 = N->getOperand(0);
1047 SDValue Op1 = N->getOperand(1);
1050 if (ISD::isBuildVectorAllOnes(Op0.getNode()))
1052 else if (ISD::isBuildVectorAllOnes(Op1.getNode()))
1057 if (NotOp->getOpcode() == ISD::OR)
1058 return DAG.getNode(MipsISD::VNOR, SDLoc(N), Ty, NotOp->getOperand(0),
1059 NotOp->getOperand(1));
1066 MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1067 SelectionDAG &DAG = DCI.DAG;
1070 switch (N->getOpcode()) {
1072 return performADDECombine(N, DAG, DCI, Subtarget);
1074 Val = performANDCombine(N, DAG, DCI, Subtarget);
1077 Val = performORCombine(N, DAG, DCI, Subtarget);
1080 return performSUBECombine(N, DAG, DCI, Subtarget);
1082 return performMULCombine(N, DAG, DCI, this);
1084 return performSHLCombine(N, DAG, DCI, Subtarget);
1086 return performSRACombine(N, DAG, DCI, Subtarget);
1088 return performSRLCombine(N, DAG, DCI, Subtarget);
1090 return performVSELECTCombine(N, DAG);
1092 Val = performXORCombine(N, DAG, Subtarget);
1095 Val = performSETCCCombine(N, DAG);
1099 if (Val.getNode()) {
1100 DEBUG(dbgs() << "\nMipsSE DAG Combine:\n";
1101 N->printrWithDepth(dbgs(), &DAG);
1102 dbgs() << "\n=> \n";
1103 Val.getNode()->printrWithDepth(dbgs(), &DAG);
1108 return MipsTargetLowering::PerformDAGCombine(N, DCI);
1112 MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1113 MachineBasicBlock *BB) const {
1114 switch (MI->getOpcode()) {
1116 return MipsTargetLowering::EmitInstrWithCustomInserter(MI, BB);
1117 case Mips::BPOSGE32_PSEUDO:
1118 return emitBPOSGE32(MI, BB);
1119 case Mips::SNZ_B_PSEUDO:
1120 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_B);
1121 case Mips::SNZ_H_PSEUDO:
1122 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_H);
1123 case Mips::SNZ_W_PSEUDO:
1124 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_W);
1125 case Mips::SNZ_D_PSEUDO:
1126 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_D);
1127 case Mips::SNZ_V_PSEUDO:
1128 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_V);
1129 case Mips::SZ_B_PSEUDO:
1130 return emitMSACBranchPseudo(MI, BB, Mips::BZ_B);
1131 case Mips::SZ_H_PSEUDO:
1132 return emitMSACBranchPseudo(MI, BB, Mips::BZ_H);
1133 case Mips::SZ_W_PSEUDO:
1134 return emitMSACBranchPseudo(MI, BB, Mips::BZ_W);
1135 case Mips::SZ_D_PSEUDO:
1136 return emitMSACBranchPseudo(MI, BB, Mips::BZ_D);
1137 case Mips::SZ_V_PSEUDO:
1138 return emitMSACBranchPseudo(MI, BB, Mips::BZ_V);
1139 case Mips::COPY_FW_PSEUDO:
1140 return emitCOPY_FW(MI, BB);
1141 case Mips::COPY_FD_PSEUDO:
1142 return emitCOPY_FD(MI, BB);
1143 case Mips::INSERT_FW_PSEUDO:
1144 return emitINSERT_FW(MI, BB);
1145 case Mips::INSERT_FD_PSEUDO:
1146 return emitINSERT_FD(MI, BB);
1147 case Mips::INSERT_B_VIDX_PSEUDO:
1148 return emitINSERT_DF_VIDX(MI, BB, 1, false);
1149 case Mips::INSERT_H_VIDX_PSEUDO:
1150 return emitINSERT_DF_VIDX(MI, BB, 2, false);
1151 case Mips::INSERT_W_VIDX_PSEUDO:
1152 return emitINSERT_DF_VIDX(MI, BB, 4, false);
1153 case Mips::INSERT_D_VIDX_PSEUDO:
1154 return emitINSERT_DF_VIDX(MI, BB, 8, false);
1155 case Mips::INSERT_FW_VIDX_PSEUDO:
1156 return emitINSERT_DF_VIDX(MI, BB, 4, true);
1157 case Mips::INSERT_FD_VIDX_PSEUDO:
1158 return emitINSERT_DF_VIDX(MI, BB, 8, true);
1159 case Mips::FILL_FW_PSEUDO:
1160 return emitFILL_FW(MI, BB);
1161 case Mips::FILL_FD_PSEUDO:
1162 return emitFILL_FD(MI, BB);
1163 case Mips::FEXP2_W_1_PSEUDO:
1164 return emitFEXP2_W_1(MI, BB);
1165 case Mips::FEXP2_D_1_PSEUDO:
1166 return emitFEXP2_D_1(MI, BB);
1170 bool MipsSETargetLowering::isEligibleForTailCallOptimization(
1171 const CCState &CCInfo, 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 (CCInfo.getInRegsParamsCount() > 0 || 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 bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee,
1190 SDValue Chain) const {
1191 Ops.push_back(Callee);
1192 MipsTargetLowering::getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal,
1193 InternalLinkage, IsCallReloc, CLI, Callee,
1197 SDValue MipsSETargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1198 LoadSDNode &Nd = *cast<LoadSDNode>(Op);
1200 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
1201 return MipsTargetLowering::lowerLOAD(Op, DAG);
1203 // Replace a double precision load with two i32 loads and a buildpair64.
1205 SDValue Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
1206 EVT PtrVT = Ptr.getValueType();
1208 // i32 load from lower address.
1209 SDValue Lo = DAG.getLoad(MVT::i32, DL, Chain, Ptr,
1210 MachinePointerInfo(), Nd.isVolatile(),
1211 Nd.isNonTemporal(), Nd.isInvariant(),
1214 // i32 load from higher address.
1215 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, PtrVT));
1216 SDValue Hi = DAG.getLoad(MVT::i32, DL, Lo.getValue(1), Ptr,
1217 MachinePointerInfo(), Nd.isVolatile(),
1218 Nd.isNonTemporal(), Nd.isInvariant(),
1219 std::min(Nd.getAlignment(), 4U));
1221 if (!Subtarget.isLittle())
1224 SDValue BP = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
1225 SDValue Ops[2] = {BP, Hi.getValue(1)};
1226 return DAG.getMergeValues(Ops, DL);
1229 SDValue MipsSETargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1230 StoreSDNode &Nd = *cast<StoreSDNode>(Op);
1232 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
1233 return MipsTargetLowering::lowerSTORE(Op, DAG);
1235 // Replace a double precision store with two extractelement64s and i32 stores.
1237 SDValue Val = Nd.getValue(), Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
1238 EVT PtrVT = Ptr.getValueType();
1239 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1240 Val, DAG.getConstant(0, MVT::i32));
1241 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1242 Val, DAG.getConstant(1, MVT::i32));
1244 if (!Subtarget.isLittle())
1247 // i32 store to lower address.
1248 Chain = DAG.getStore(Chain, DL, Lo, Ptr, MachinePointerInfo(),
1249 Nd.isVolatile(), Nd.isNonTemporal(), Nd.getAlignment(),
1252 // i32 store to higher address.
1253 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, PtrVT));
1254 return DAG.getStore(Chain, DL, Hi, Ptr, MachinePointerInfo(),
1255 Nd.isVolatile(), Nd.isNonTemporal(),
1256 std::min(Nd.getAlignment(), 4U), Nd.getAAInfo());
1259 SDValue MipsSETargetLowering::lowerMulDiv(SDValue Op, unsigned NewOpc,
1260 bool HasLo, bool HasHi,
1261 SelectionDAG &DAG) const {
1262 // MIPS32r6/MIPS64r6 removed accumulator based multiplies.
1263 assert(!Subtarget.hasMips32r6());
1265 EVT Ty = Op.getOperand(0).getValueType();
1267 SDValue Mult = DAG.getNode(NewOpc, DL, MVT::Untyped,
1268 Op.getOperand(0), Op.getOperand(1));
1272 Lo = DAG.getNode(MipsISD::MFLO, DL, Ty, Mult);
1274 Hi = DAG.getNode(MipsISD::MFHI, DL, Ty, Mult);
1276 if (!HasLo || !HasHi)
1277 return HasLo ? Lo : Hi;
1279 SDValue Vals[] = { Lo, Hi };
1280 return DAG.getMergeValues(Vals, DL);
1284 static SDValue initAccumulator(SDValue In, SDLoc DL, SelectionDAG &DAG) {
1285 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
1286 DAG.getConstant(0, MVT::i32));
1287 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
1288 DAG.getConstant(1, MVT::i32));
1289 return DAG.getNode(MipsISD::MTLOHI, DL, MVT::Untyped, InLo, InHi);
1292 static SDValue extractLOHI(SDValue Op, SDLoc DL, SelectionDAG &DAG) {
1293 SDValue Lo = DAG.getNode(MipsISD::MFLO, DL, MVT::i32, Op);
1294 SDValue Hi = DAG.getNode(MipsISD::MFHI, DL, MVT::i32, Op);
1295 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi);
1298 // This function expands mips intrinsic nodes which have 64-bit input operands
1299 // or output values.
1301 // out64 = intrinsic-node in64
1303 // lo = copy (extract-element (in64, 0))
1304 // hi = copy (extract-element (in64, 1))
1305 // mips-specific-node
1308 // out64 = merge-values (v0, v1)
1310 static SDValue lowerDSPIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1312 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
1313 SmallVector<SDValue, 3> Ops;
1316 // See if Op has a chain input.
1318 Ops.push_back(Op->getOperand(OpNo++));
1320 // The next operand is the intrinsic opcode.
1321 assert(Op->getOperand(OpNo).getOpcode() == ISD::TargetConstant);
1323 // See if the next operand has type i64.
1324 SDValue Opnd = Op->getOperand(++OpNo), In64;
1326 if (Opnd.getValueType() == MVT::i64)
1327 In64 = initAccumulator(Opnd, DL, DAG);
1329 Ops.push_back(Opnd);
1331 // Push the remaining operands.
1332 for (++OpNo ; OpNo < Op->getNumOperands(); ++OpNo)
1333 Ops.push_back(Op->getOperand(OpNo));
1335 // Add In64 to the end of the list.
1337 Ops.push_back(In64);
1340 SmallVector<EVT, 2> ResTys;
1342 for (SDNode::value_iterator I = Op->value_begin(), E = Op->value_end();
1344 ResTys.push_back((*I == MVT::i64) ? MVT::Untyped : *I);
1347 SDValue Val = DAG.getNode(Opc, DL, ResTys, Ops);
1348 SDValue Out = (ResTys[0] == MVT::Untyped) ? extractLOHI(Val, DL, DAG) : Val;
1353 assert(Val->getValueType(1) == MVT::Other);
1354 SDValue Vals[] = { Out, SDValue(Val.getNode(), 1) };
1355 return DAG.getMergeValues(Vals, DL);
1358 // Lower an MSA copy intrinsic into the specified SelectionDAG node
1359 static SDValue lowerMSACopyIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1361 SDValue Vec = Op->getOperand(1);
1362 SDValue Idx = Op->getOperand(2);
1363 EVT ResTy = Op->getValueType(0);
1364 EVT EltTy = Vec->getValueType(0).getVectorElementType();
1366 SDValue Result = DAG.getNode(Opc, DL, ResTy, Vec, Idx,
1367 DAG.getValueType(EltTy));
1372 static SDValue lowerMSASplatZExt(SDValue Op, unsigned OpNr, SelectionDAG &DAG) {
1373 EVT ResVecTy = Op->getValueType(0);
1374 EVT ViaVecTy = ResVecTy;
1377 // When ResVecTy == MVT::v2i64, LaneA is the upper 32 bits of the lane and
1378 // LaneB is the lower 32-bits. Otherwise LaneA and LaneB are alternating
1381 SDValue LaneB = Op->getOperand(2);
1383 if (ResVecTy == MVT::v2i64) {
1384 LaneA = DAG.getConstant(0, MVT::i32);
1385 ViaVecTy = MVT::v4i32;
1389 SDValue Ops[16] = { LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB,
1390 LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB };
1392 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy,
1393 makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
1395 if (ViaVecTy != ResVecTy)
1396 Result = DAG.getNode(ISD::BITCAST, DL, ResVecTy, Result);
1401 static SDValue lowerMSASplatImm(SDValue Op, unsigned ImmOp, SelectionDAG &DAG) {
1402 return DAG.getConstant(Op->getConstantOperandVal(ImmOp), Op->getValueType(0));
1405 static SDValue getBuildVectorSplat(EVT VecTy, SDValue SplatValue,
1406 bool BigEndian, SelectionDAG &DAG) {
1407 EVT ViaVecTy = VecTy;
1408 SDValue SplatValueA = SplatValue;
1409 SDValue SplatValueB = SplatValue;
1410 SDLoc DL(SplatValue);
1412 if (VecTy == MVT::v2i64) {
1413 // v2i64 BUILD_VECTOR must be performed via v4i32 so split into i32's.
1414 ViaVecTy = MVT::v4i32;
1416 SplatValueA = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, SplatValue);
1417 SplatValueB = DAG.getNode(ISD::SRL, DL, MVT::i64, SplatValue,
1418 DAG.getConstant(32, MVT::i32));
1419 SplatValueB = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, SplatValueB);
1422 // We currently hold the parts in little endian order. Swap them if
1425 std::swap(SplatValueA, SplatValueB);
1427 SDValue Ops[16] = { SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1428 SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1429 SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1430 SplatValueA, SplatValueB, SplatValueA, SplatValueB };
1432 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy,
1433 makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
1435 if (VecTy != ViaVecTy)
1436 Result = DAG.getNode(ISD::BITCAST, DL, VecTy, Result);
1441 static SDValue lowerMSABinaryBitImmIntr(SDValue Op, SelectionDAG &DAG,
1442 unsigned Opc, SDValue Imm,
1444 EVT VecTy = Op->getValueType(0);
1448 // The DAG Combiner can't constant fold bitcasted vectors yet so we must do it
1450 if (VecTy == MVT::v2i64) {
1451 if (ConstantSDNode *CImm = dyn_cast<ConstantSDNode>(Imm)) {
1452 APInt BitImm = APInt(64, 1) << CImm->getAPIntValue();
1454 SDValue BitImmHiOp = DAG.getConstant(BitImm.lshr(32).trunc(32), MVT::i32);
1455 SDValue BitImmLoOp = DAG.getConstant(BitImm.trunc(32), MVT::i32);
1458 std::swap(BitImmLoOp, BitImmHiOp);
1461 DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
1462 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v4i32, BitImmLoOp,
1463 BitImmHiOp, BitImmLoOp, BitImmHiOp));
1467 if (!Exp2Imm.getNode()) {
1468 // We couldnt constant fold, do a vector shift instead
1470 // Extend i32 to i64 if necessary. Sign or zero extend doesn't matter since
1471 // only values 0-63 are valid.
1472 if (VecTy == MVT::v2i64)
1473 Imm = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Imm);
1475 Exp2Imm = getBuildVectorSplat(VecTy, Imm, BigEndian, DAG);
1478 DAG.getNode(ISD::SHL, DL, VecTy, DAG.getConstant(1, VecTy), Exp2Imm);
1481 return DAG.getNode(Opc, DL, VecTy, Op->getOperand(1), Exp2Imm);
1484 static SDValue lowerMSABitClear(SDValue Op, SelectionDAG &DAG) {
1485 EVT ResTy = Op->getValueType(0);
1487 SDValue One = DAG.getConstant(1, ResTy);
1488 SDValue Bit = DAG.getNode(ISD::SHL, DL, ResTy, One, Op->getOperand(2));
1490 return DAG.getNode(ISD::AND, DL, ResTy, Op->getOperand(1),
1491 DAG.getNOT(DL, Bit, ResTy));
1494 static SDValue lowerMSABitClearImm(SDValue Op, SelectionDAG &DAG) {
1496 EVT ResTy = Op->getValueType(0);
1497 APInt BitImm = APInt(ResTy.getVectorElementType().getSizeInBits(), 1)
1498 << cast<ConstantSDNode>(Op->getOperand(2))->getAPIntValue();
1499 SDValue BitMask = DAG.getConstant(~BitImm, ResTy);
1501 return DAG.getNode(ISD::AND, DL, ResTy, Op->getOperand(1), BitMask);
1504 SDValue MipsSETargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
1505 SelectionDAG &DAG) const {
1508 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) {
1511 case Intrinsic::mips_shilo:
1512 return lowerDSPIntr(Op, DAG, MipsISD::SHILO);
1513 case Intrinsic::mips_dpau_h_qbl:
1514 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL);
1515 case Intrinsic::mips_dpau_h_qbr:
1516 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR);
1517 case Intrinsic::mips_dpsu_h_qbl:
1518 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL);
1519 case Intrinsic::mips_dpsu_h_qbr:
1520 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR);
1521 case Intrinsic::mips_dpa_w_ph:
1522 return lowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH);
1523 case Intrinsic::mips_dps_w_ph:
1524 return lowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH);
1525 case Intrinsic::mips_dpax_w_ph:
1526 return lowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH);
1527 case Intrinsic::mips_dpsx_w_ph:
1528 return lowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH);
1529 case Intrinsic::mips_mulsa_w_ph:
1530 return lowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH);
1531 case Intrinsic::mips_mult:
1532 return lowerDSPIntr(Op, DAG, MipsISD::Mult);
1533 case Intrinsic::mips_multu:
1534 return lowerDSPIntr(Op, DAG, MipsISD::Multu);
1535 case Intrinsic::mips_madd:
1536 return lowerDSPIntr(Op, DAG, MipsISD::MAdd);
1537 case Intrinsic::mips_maddu:
1538 return lowerDSPIntr(Op, DAG, MipsISD::MAddu);
1539 case Intrinsic::mips_msub:
1540 return lowerDSPIntr(Op, DAG, MipsISD::MSub);
1541 case Intrinsic::mips_msubu:
1542 return lowerDSPIntr(Op, DAG, MipsISD::MSubu);
1543 case Intrinsic::mips_addv_b:
1544 case Intrinsic::mips_addv_h:
1545 case Intrinsic::mips_addv_w:
1546 case Intrinsic::mips_addv_d:
1547 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1),
1549 case Intrinsic::mips_addvi_b:
1550 case Intrinsic::mips_addvi_h:
1551 case Intrinsic::mips_addvi_w:
1552 case Intrinsic::mips_addvi_d:
1553 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1),
1554 lowerMSASplatImm(Op, 2, DAG));
1555 case Intrinsic::mips_and_v:
1556 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
1558 case Intrinsic::mips_andi_b:
1559 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
1560 lowerMSASplatImm(Op, 2, DAG));
1561 case Intrinsic::mips_bclr_b:
1562 case Intrinsic::mips_bclr_h:
1563 case Intrinsic::mips_bclr_w:
1564 case Intrinsic::mips_bclr_d:
1565 return lowerMSABitClear(Op, DAG);
1566 case Intrinsic::mips_bclri_b:
1567 case Intrinsic::mips_bclri_h:
1568 case Intrinsic::mips_bclri_w:
1569 case Intrinsic::mips_bclri_d:
1570 return lowerMSABitClearImm(Op, DAG);
1571 case Intrinsic::mips_binsli_b:
1572 case Intrinsic::mips_binsli_h:
1573 case Intrinsic::mips_binsli_w:
1574 case Intrinsic::mips_binsli_d: {
1575 // binsli_x(IfClear, IfSet, nbits) -> (vselect LBitsMask, IfSet, IfClear)
1576 EVT VecTy = Op->getValueType(0);
1577 EVT EltTy = VecTy.getVectorElementType();
1578 APInt Mask = APInt::getHighBitsSet(EltTy.getSizeInBits(),
1579 Op->getConstantOperandVal(3));
1580 return DAG.getNode(ISD::VSELECT, DL, VecTy,
1581 DAG.getConstant(Mask, VecTy, true), Op->getOperand(2),
1584 case Intrinsic::mips_binsri_b:
1585 case Intrinsic::mips_binsri_h:
1586 case Intrinsic::mips_binsri_w:
1587 case Intrinsic::mips_binsri_d: {
1588 // binsri_x(IfClear, IfSet, nbits) -> (vselect RBitsMask, IfSet, IfClear)
1589 EVT VecTy = Op->getValueType(0);
1590 EVT EltTy = VecTy.getVectorElementType();
1591 APInt Mask = APInt::getLowBitsSet(EltTy.getSizeInBits(),
1592 Op->getConstantOperandVal(3));
1593 return DAG.getNode(ISD::VSELECT, DL, VecTy,
1594 DAG.getConstant(Mask, VecTy, true), Op->getOperand(2),
1597 case Intrinsic::mips_bmnz_v:
1598 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), Op->getOperand(3),
1599 Op->getOperand(2), Op->getOperand(1));
1600 case Intrinsic::mips_bmnzi_b:
1601 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1602 lowerMSASplatImm(Op, 3, DAG), Op->getOperand(2),
1604 case Intrinsic::mips_bmz_v:
1605 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), Op->getOperand(3),
1606 Op->getOperand(1), Op->getOperand(2));
1607 case Intrinsic::mips_bmzi_b:
1608 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1609 lowerMSASplatImm(Op, 3, DAG), Op->getOperand(1),
1611 case Intrinsic::mips_bneg_b:
1612 case Intrinsic::mips_bneg_h:
1613 case Intrinsic::mips_bneg_w:
1614 case Intrinsic::mips_bneg_d: {
1615 EVT VecTy = Op->getValueType(0);
1616 SDValue One = DAG.getConstant(1, VecTy);
1618 return DAG.getNode(ISD::XOR, DL, VecTy, Op->getOperand(1),
1619 DAG.getNode(ISD::SHL, DL, VecTy, One,
1620 Op->getOperand(2)));
1622 case Intrinsic::mips_bnegi_b:
1623 case Intrinsic::mips_bnegi_h:
1624 case Intrinsic::mips_bnegi_w:
1625 case Intrinsic::mips_bnegi_d:
1626 return lowerMSABinaryBitImmIntr(Op, DAG, ISD::XOR, Op->getOperand(2),
1627 !Subtarget.isLittle());
1628 case Intrinsic::mips_bnz_b:
1629 case Intrinsic::mips_bnz_h:
1630 case Intrinsic::mips_bnz_w:
1631 case Intrinsic::mips_bnz_d:
1632 return DAG.getNode(MipsISD::VALL_NONZERO, DL, Op->getValueType(0),
1634 case Intrinsic::mips_bnz_v:
1635 return DAG.getNode(MipsISD::VANY_NONZERO, DL, Op->getValueType(0),
1637 case Intrinsic::mips_bsel_v:
1638 // bsel_v(Mask, IfClear, IfSet) -> (vselect Mask, IfSet, IfClear)
1639 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1640 Op->getOperand(1), Op->getOperand(3),
1642 case Intrinsic::mips_bseli_b:
1643 // bseli_v(Mask, IfClear, IfSet) -> (vselect Mask, IfSet, IfClear)
1644 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1645 Op->getOperand(1), lowerMSASplatImm(Op, 3, DAG),
1647 case Intrinsic::mips_bset_b:
1648 case Intrinsic::mips_bset_h:
1649 case Intrinsic::mips_bset_w:
1650 case Intrinsic::mips_bset_d: {
1651 EVT VecTy = Op->getValueType(0);
1652 SDValue One = DAG.getConstant(1, VecTy);
1654 return DAG.getNode(ISD::OR, DL, VecTy, Op->getOperand(1),
1655 DAG.getNode(ISD::SHL, DL, VecTy, One,
1656 Op->getOperand(2)));
1658 case Intrinsic::mips_bseti_b:
1659 case Intrinsic::mips_bseti_h:
1660 case Intrinsic::mips_bseti_w:
1661 case Intrinsic::mips_bseti_d:
1662 return lowerMSABinaryBitImmIntr(Op, DAG, ISD::OR, Op->getOperand(2),
1663 !Subtarget.isLittle());
1664 case Intrinsic::mips_bz_b:
1665 case Intrinsic::mips_bz_h:
1666 case Intrinsic::mips_bz_w:
1667 case Intrinsic::mips_bz_d:
1668 return DAG.getNode(MipsISD::VALL_ZERO, DL, Op->getValueType(0),
1670 case Intrinsic::mips_bz_v:
1671 return DAG.getNode(MipsISD::VANY_ZERO, DL, Op->getValueType(0),
1673 case Intrinsic::mips_ceq_b:
1674 case Intrinsic::mips_ceq_h:
1675 case Intrinsic::mips_ceq_w:
1676 case Intrinsic::mips_ceq_d:
1677 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1678 Op->getOperand(2), ISD::SETEQ);
1679 case Intrinsic::mips_ceqi_b:
1680 case Intrinsic::mips_ceqi_h:
1681 case Intrinsic::mips_ceqi_w:
1682 case Intrinsic::mips_ceqi_d:
1683 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1684 lowerMSASplatImm(Op, 2, DAG), ISD::SETEQ);
1685 case Intrinsic::mips_cle_s_b:
1686 case Intrinsic::mips_cle_s_h:
1687 case Intrinsic::mips_cle_s_w:
1688 case Intrinsic::mips_cle_s_d:
1689 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1690 Op->getOperand(2), ISD::SETLE);
1691 case Intrinsic::mips_clei_s_b:
1692 case Intrinsic::mips_clei_s_h:
1693 case Intrinsic::mips_clei_s_w:
1694 case Intrinsic::mips_clei_s_d:
1695 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1696 lowerMSASplatImm(Op, 2, DAG), ISD::SETLE);
1697 case Intrinsic::mips_cle_u_b:
1698 case Intrinsic::mips_cle_u_h:
1699 case Intrinsic::mips_cle_u_w:
1700 case Intrinsic::mips_cle_u_d:
1701 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1702 Op->getOperand(2), ISD::SETULE);
1703 case Intrinsic::mips_clei_u_b:
1704 case Intrinsic::mips_clei_u_h:
1705 case Intrinsic::mips_clei_u_w:
1706 case Intrinsic::mips_clei_u_d:
1707 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1708 lowerMSASplatImm(Op, 2, DAG), ISD::SETULE);
1709 case Intrinsic::mips_clt_s_b:
1710 case Intrinsic::mips_clt_s_h:
1711 case Intrinsic::mips_clt_s_w:
1712 case Intrinsic::mips_clt_s_d:
1713 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1714 Op->getOperand(2), ISD::SETLT);
1715 case Intrinsic::mips_clti_s_b:
1716 case Intrinsic::mips_clti_s_h:
1717 case Intrinsic::mips_clti_s_w:
1718 case Intrinsic::mips_clti_s_d:
1719 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1720 lowerMSASplatImm(Op, 2, DAG), ISD::SETLT);
1721 case Intrinsic::mips_clt_u_b:
1722 case Intrinsic::mips_clt_u_h:
1723 case Intrinsic::mips_clt_u_w:
1724 case Intrinsic::mips_clt_u_d:
1725 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1726 Op->getOperand(2), ISD::SETULT);
1727 case Intrinsic::mips_clti_u_b:
1728 case Intrinsic::mips_clti_u_h:
1729 case Intrinsic::mips_clti_u_w:
1730 case Intrinsic::mips_clti_u_d:
1731 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1732 lowerMSASplatImm(Op, 2, DAG), ISD::SETULT);
1733 case Intrinsic::mips_copy_s_b:
1734 case Intrinsic::mips_copy_s_h:
1735 case Intrinsic::mips_copy_s_w:
1736 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
1737 case Intrinsic::mips_copy_s_d:
1738 if (Subtarget.hasMips64())
1739 // Lower directly into VEXTRACT_SEXT_ELT since i64 is legal on Mips64.
1740 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
1742 // Lower into the generic EXTRACT_VECTOR_ELT node and let the type
1743 // legalizer and EXTRACT_VECTOR_ELT lowering sort it out.
1744 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
1745 Op->getValueType(0), Op->getOperand(1),
1748 case Intrinsic::mips_copy_u_b:
1749 case Intrinsic::mips_copy_u_h:
1750 case Intrinsic::mips_copy_u_w:
1751 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
1752 case Intrinsic::mips_copy_u_d:
1753 if (Subtarget.hasMips64())
1754 // Lower directly into VEXTRACT_ZEXT_ELT since i64 is legal on Mips64.
1755 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
1757 // Lower into the generic EXTRACT_VECTOR_ELT node and let the type
1758 // legalizer and EXTRACT_VECTOR_ELT lowering sort it out.
1759 // Note: When i64 is illegal, this results in copy_s.w instructions
1760 // instead of copy_u.w instructions. This makes no difference to the
1761 // behaviour since i64 is only illegal when the register file is 32-bit.
1762 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
1763 Op->getValueType(0), Op->getOperand(1),
1766 case Intrinsic::mips_div_s_b:
1767 case Intrinsic::mips_div_s_h:
1768 case Intrinsic::mips_div_s_w:
1769 case Intrinsic::mips_div_s_d:
1770 return DAG.getNode(ISD::SDIV, DL, Op->getValueType(0), Op->getOperand(1),
1772 case Intrinsic::mips_div_u_b:
1773 case Intrinsic::mips_div_u_h:
1774 case Intrinsic::mips_div_u_w:
1775 case Intrinsic::mips_div_u_d:
1776 return DAG.getNode(ISD::UDIV, DL, Op->getValueType(0), Op->getOperand(1),
1778 case Intrinsic::mips_fadd_w:
1779 case Intrinsic::mips_fadd_d:
1780 return DAG.getNode(ISD::FADD, DL, Op->getValueType(0), Op->getOperand(1),
1782 // Don't lower mips_fcaf_[wd] since LLVM folds SETFALSE condcodes away
1783 case Intrinsic::mips_fceq_w:
1784 case Intrinsic::mips_fceq_d:
1785 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1786 Op->getOperand(2), ISD::SETOEQ);
1787 case Intrinsic::mips_fcle_w:
1788 case Intrinsic::mips_fcle_d:
1789 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1790 Op->getOperand(2), ISD::SETOLE);
1791 case Intrinsic::mips_fclt_w:
1792 case Intrinsic::mips_fclt_d:
1793 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1794 Op->getOperand(2), ISD::SETOLT);
1795 case Intrinsic::mips_fcne_w:
1796 case Intrinsic::mips_fcne_d:
1797 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1798 Op->getOperand(2), ISD::SETONE);
1799 case Intrinsic::mips_fcor_w:
1800 case Intrinsic::mips_fcor_d:
1801 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1802 Op->getOperand(2), ISD::SETO);
1803 case Intrinsic::mips_fcueq_w:
1804 case Intrinsic::mips_fcueq_d:
1805 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1806 Op->getOperand(2), ISD::SETUEQ);
1807 case Intrinsic::mips_fcule_w:
1808 case Intrinsic::mips_fcule_d:
1809 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1810 Op->getOperand(2), ISD::SETULE);
1811 case Intrinsic::mips_fcult_w:
1812 case Intrinsic::mips_fcult_d:
1813 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1814 Op->getOperand(2), ISD::SETULT);
1815 case Intrinsic::mips_fcun_w:
1816 case Intrinsic::mips_fcun_d:
1817 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1818 Op->getOperand(2), ISD::SETUO);
1819 case Intrinsic::mips_fcune_w:
1820 case Intrinsic::mips_fcune_d:
1821 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1822 Op->getOperand(2), ISD::SETUNE);
1823 case Intrinsic::mips_fdiv_w:
1824 case Intrinsic::mips_fdiv_d:
1825 return DAG.getNode(ISD::FDIV, DL, Op->getValueType(0), Op->getOperand(1),
1827 case Intrinsic::mips_ffint_u_w:
1828 case Intrinsic::mips_ffint_u_d:
1829 return DAG.getNode(ISD::UINT_TO_FP, DL, Op->getValueType(0),
1831 case Intrinsic::mips_ffint_s_w:
1832 case Intrinsic::mips_ffint_s_d:
1833 return DAG.getNode(ISD::SINT_TO_FP, DL, Op->getValueType(0),
1835 case Intrinsic::mips_fill_b:
1836 case Intrinsic::mips_fill_h:
1837 case Intrinsic::mips_fill_w:
1838 case Intrinsic::mips_fill_d: {
1839 SmallVector<SDValue, 16> Ops;
1840 EVT ResTy = Op->getValueType(0);
1842 for (unsigned i = 0; i < ResTy.getVectorNumElements(); ++i)
1843 Ops.push_back(Op->getOperand(1));
1845 // If ResTy is v2i64 then the type legalizer will break this node down into
1846 // an equivalent v4i32.
1847 return DAG.getNode(ISD::BUILD_VECTOR, DL, ResTy, Ops);
1849 case Intrinsic::mips_fexp2_w:
1850 case Intrinsic::mips_fexp2_d: {
1851 EVT ResTy = Op->getValueType(0);
1853 ISD::FMUL, SDLoc(Op), ResTy, Op->getOperand(1),
1854 DAG.getNode(ISD::FEXP2, SDLoc(Op), ResTy, Op->getOperand(2)));
1856 case Intrinsic::mips_flog2_w:
1857 case Intrinsic::mips_flog2_d:
1858 return DAG.getNode(ISD::FLOG2, DL, Op->getValueType(0), Op->getOperand(1));
1859 case Intrinsic::mips_fmadd_w:
1860 case Intrinsic::mips_fmadd_d:
1861 return DAG.getNode(ISD::FMA, SDLoc(Op), Op->getValueType(0),
1862 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3));
1863 case Intrinsic::mips_fmul_w:
1864 case Intrinsic::mips_fmul_d:
1865 return DAG.getNode(ISD::FMUL, DL, Op->getValueType(0), Op->getOperand(1),
1867 case Intrinsic::mips_fmsub_w:
1868 case Intrinsic::mips_fmsub_d: {
1869 EVT ResTy = Op->getValueType(0);
1870 return DAG.getNode(ISD::FSUB, SDLoc(Op), ResTy, Op->getOperand(1),
1871 DAG.getNode(ISD::FMUL, SDLoc(Op), ResTy,
1872 Op->getOperand(2), Op->getOperand(3)));
1874 case Intrinsic::mips_frint_w:
1875 case Intrinsic::mips_frint_d:
1876 return DAG.getNode(ISD::FRINT, DL, Op->getValueType(0), Op->getOperand(1));
1877 case Intrinsic::mips_fsqrt_w:
1878 case Intrinsic::mips_fsqrt_d:
1879 return DAG.getNode(ISD::FSQRT, DL, Op->getValueType(0), Op->getOperand(1));
1880 case Intrinsic::mips_fsub_w:
1881 case Intrinsic::mips_fsub_d:
1882 return DAG.getNode(ISD::FSUB, DL, Op->getValueType(0), Op->getOperand(1),
1884 case Intrinsic::mips_ftrunc_u_w:
1885 case Intrinsic::mips_ftrunc_u_d:
1886 return DAG.getNode(ISD::FP_TO_UINT, DL, Op->getValueType(0),
1888 case Intrinsic::mips_ftrunc_s_w:
1889 case Intrinsic::mips_ftrunc_s_d:
1890 return DAG.getNode(ISD::FP_TO_SINT, DL, Op->getValueType(0),
1892 case Intrinsic::mips_ilvev_b:
1893 case Intrinsic::mips_ilvev_h:
1894 case Intrinsic::mips_ilvev_w:
1895 case Intrinsic::mips_ilvev_d:
1896 return DAG.getNode(MipsISD::ILVEV, DL, Op->getValueType(0),
1897 Op->getOperand(1), Op->getOperand(2));
1898 case Intrinsic::mips_ilvl_b:
1899 case Intrinsic::mips_ilvl_h:
1900 case Intrinsic::mips_ilvl_w:
1901 case Intrinsic::mips_ilvl_d:
1902 return DAG.getNode(MipsISD::ILVL, DL, Op->getValueType(0),
1903 Op->getOperand(1), Op->getOperand(2));
1904 case Intrinsic::mips_ilvod_b:
1905 case Intrinsic::mips_ilvod_h:
1906 case Intrinsic::mips_ilvod_w:
1907 case Intrinsic::mips_ilvod_d:
1908 return DAG.getNode(MipsISD::ILVOD, DL, Op->getValueType(0),
1909 Op->getOperand(1), Op->getOperand(2));
1910 case Intrinsic::mips_ilvr_b:
1911 case Intrinsic::mips_ilvr_h:
1912 case Intrinsic::mips_ilvr_w:
1913 case Intrinsic::mips_ilvr_d:
1914 return DAG.getNode(MipsISD::ILVR, DL, Op->getValueType(0),
1915 Op->getOperand(1), Op->getOperand(2));
1916 case Intrinsic::mips_insert_b:
1917 case Intrinsic::mips_insert_h:
1918 case Intrinsic::mips_insert_w:
1919 case Intrinsic::mips_insert_d:
1920 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(Op), Op->getValueType(0),
1921 Op->getOperand(1), Op->getOperand(3), Op->getOperand(2));
1922 case Intrinsic::mips_insve_b:
1923 case Intrinsic::mips_insve_h:
1924 case Intrinsic::mips_insve_w:
1925 case Intrinsic::mips_insve_d:
1926 return DAG.getNode(MipsISD::INSVE, DL, Op->getValueType(0),
1927 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3),
1928 DAG.getConstant(0, MVT::i32));
1929 case Intrinsic::mips_ldi_b:
1930 case Intrinsic::mips_ldi_h:
1931 case Intrinsic::mips_ldi_w:
1932 case Intrinsic::mips_ldi_d:
1933 return lowerMSASplatImm(Op, 1, DAG);
1934 case Intrinsic::mips_lsa:
1935 case Intrinsic::mips_dlsa: {
1936 EVT ResTy = Op->getValueType(0);
1937 return DAG.getNode(ISD::ADD, SDLoc(Op), ResTy, Op->getOperand(1),
1938 DAG.getNode(ISD::SHL, SDLoc(Op), ResTy,
1939 Op->getOperand(2), Op->getOperand(3)));
1941 case Intrinsic::mips_maddv_b:
1942 case Intrinsic::mips_maddv_h:
1943 case Intrinsic::mips_maddv_w:
1944 case Intrinsic::mips_maddv_d: {
1945 EVT ResTy = Op->getValueType(0);
1946 return DAG.getNode(ISD::ADD, SDLoc(Op), ResTy, Op->getOperand(1),
1947 DAG.getNode(ISD::MUL, SDLoc(Op), ResTy,
1948 Op->getOperand(2), Op->getOperand(3)));
1950 case Intrinsic::mips_max_s_b:
1951 case Intrinsic::mips_max_s_h:
1952 case Intrinsic::mips_max_s_w:
1953 case Intrinsic::mips_max_s_d:
1954 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0),
1955 Op->getOperand(1), Op->getOperand(2));
1956 case Intrinsic::mips_max_u_b:
1957 case Intrinsic::mips_max_u_h:
1958 case Intrinsic::mips_max_u_w:
1959 case Intrinsic::mips_max_u_d:
1960 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0),
1961 Op->getOperand(1), Op->getOperand(2));
1962 case Intrinsic::mips_maxi_s_b:
1963 case Intrinsic::mips_maxi_s_h:
1964 case Intrinsic::mips_maxi_s_w:
1965 case Intrinsic::mips_maxi_s_d:
1966 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0),
1967 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1968 case Intrinsic::mips_maxi_u_b:
1969 case Intrinsic::mips_maxi_u_h:
1970 case Intrinsic::mips_maxi_u_w:
1971 case Intrinsic::mips_maxi_u_d:
1972 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0),
1973 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1974 case Intrinsic::mips_min_s_b:
1975 case Intrinsic::mips_min_s_h:
1976 case Intrinsic::mips_min_s_w:
1977 case Intrinsic::mips_min_s_d:
1978 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0),
1979 Op->getOperand(1), Op->getOperand(2));
1980 case Intrinsic::mips_min_u_b:
1981 case Intrinsic::mips_min_u_h:
1982 case Intrinsic::mips_min_u_w:
1983 case Intrinsic::mips_min_u_d:
1984 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0),
1985 Op->getOperand(1), Op->getOperand(2));
1986 case Intrinsic::mips_mini_s_b:
1987 case Intrinsic::mips_mini_s_h:
1988 case Intrinsic::mips_mini_s_w:
1989 case Intrinsic::mips_mini_s_d:
1990 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0),
1991 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1992 case Intrinsic::mips_mini_u_b:
1993 case Intrinsic::mips_mini_u_h:
1994 case Intrinsic::mips_mini_u_w:
1995 case Intrinsic::mips_mini_u_d:
1996 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0),
1997 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1998 case Intrinsic::mips_mod_s_b:
1999 case Intrinsic::mips_mod_s_h:
2000 case Intrinsic::mips_mod_s_w:
2001 case Intrinsic::mips_mod_s_d:
2002 return DAG.getNode(ISD::SREM, DL, Op->getValueType(0), Op->getOperand(1),
2004 case Intrinsic::mips_mod_u_b:
2005 case Intrinsic::mips_mod_u_h:
2006 case Intrinsic::mips_mod_u_w:
2007 case Intrinsic::mips_mod_u_d:
2008 return DAG.getNode(ISD::UREM, DL, Op->getValueType(0), Op->getOperand(1),
2010 case Intrinsic::mips_mulv_b:
2011 case Intrinsic::mips_mulv_h:
2012 case Intrinsic::mips_mulv_w:
2013 case Intrinsic::mips_mulv_d:
2014 return DAG.getNode(ISD::MUL, DL, Op->getValueType(0), Op->getOperand(1),
2016 case Intrinsic::mips_msubv_b:
2017 case Intrinsic::mips_msubv_h:
2018 case Intrinsic::mips_msubv_w:
2019 case Intrinsic::mips_msubv_d: {
2020 EVT ResTy = Op->getValueType(0);
2021 return DAG.getNode(ISD::SUB, SDLoc(Op), ResTy, Op->getOperand(1),
2022 DAG.getNode(ISD::MUL, SDLoc(Op), ResTy,
2023 Op->getOperand(2), Op->getOperand(3)));
2025 case Intrinsic::mips_nlzc_b:
2026 case Intrinsic::mips_nlzc_h:
2027 case Intrinsic::mips_nlzc_w:
2028 case Intrinsic::mips_nlzc_d:
2029 return DAG.getNode(ISD::CTLZ, DL, Op->getValueType(0), Op->getOperand(1));
2030 case Intrinsic::mips_nor_v: {
2031 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2032 Op->getOperand(1), Op->getOperand(2));
2033 return DAG.getNOT(DL, Res, Res->getValueType(0));
2035 case Intrinsic::mips_nori_b: {
2036 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2038 lowerMSASplatImm(Op, 2, DAG));
2039 return DAG.getNOT(DL, Res, Res->getValueType(0));
2041 case Intrinsic::mips_or_v:
2042 return DAG.getNode(ISD::OR, DL, Op->getValueType(0), Op->getOperand(1),
2044 case Intrinsic::mips_ori_b:
2045 return DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2046 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2047 case Intrinsic::mips_pckev_b:
2048 case Intrinsic::mips_pckev_h:
2049 case Intrinsic::mips_pckev_w:
2050 case Intrinsic::mips_pckev_d:
2051 return DAG.getNode(MipsISD::PCKEV, DL, Op->getValueType(0),
2052 Op->getOperand(1), Op->getOperand(2));
2053 case Intrinsic::mips_pckod_b:
2054 case Intrinsic::mips_pckod_h:
2055 case Intrinsic::mips_pckod_w:
2056 case Intrinsic::mips_pckod_d:
2057 return DAG.getNode(MipsISD::PCKOD, DL, Op->getValueType(0),
2058 Op->getOperand(1), Op->getOperand(2));
2059 case Intrinsic::mips_pcnt_b:
2060 case Intrinsic::mips_pcnt_h:
2061 case Intrinsic::mips_pcnt_w:
2062 case Intrinsic::mips_pcnt_d:
2063 return DAG.getNode(ISD::CTPOP, DL, Op->getValueType(0), Op->getOperand(1));
2064 case Intrinsic::mips_shf_b:
2065 case Intrinsic::mips_shf_h:
2066 case Intrinsic::mips_shf_w:
2067 return DAG.getNode(MipsISD::SHF, DL, Op->getValueType(0),
2068 Op->getOperand(2), Op->getOperand(1));
2069 case Intrinsic::mips_sll_b:
2070 case Intrinsic::mips_sll_h:
2071 case Intrinsic::mips_sll_w:
2072 case Intrinsic::mips_sll_d:
2073 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0), Op->getOperand(1),
2075 case Intrinsic::mips_slli_b:
2076 case Intrinsic::mips_slli_h:
2077 case Intrinsic::mips_slli_w:
2078 case Intrinsic::mips_slli_d:
2079 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0),
2080 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2081 case Intrinsic::mips_splat_b:
2082 case Intrinsic::mips_splat_h:
2083 case Intrinsic::mips_splat_w:
2084 case Intrinsic::mips_splat_d:
2085 // We can't lower via VECTOR_SHUFFLE because it requires constant shuffle
2086 // masks, nor can we lower via BUILD_VECTOR & EXTRACT_VECTOR_ELT because
2087 // EXTRACT_VECTOR_ELT can't extract i64's on MIPS32.
2088 // Instead we lower to MipsISD::VSHF and match from there.
2089 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2090 lowerMSASplatZExt(Op, 2, DAG), Op->getOperand(1),
2092 case Intrinsic::mips_splati_b:
2093 case Intrinsic::mips_splati_h:
2094 case Intrinsic::mips_splati_w:
2095 case Intrinsic::mips_splati_d:
2096 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2097 lowerMSASplatImm(Op, 2, DAG), Op->getOperand(1),
2099 case Intrinsic::mips_sra_b:
2100 case Intrinsic::mips_sra_h:
2101 case Intrinsic::mips_sra_w:
2102 case Intrinsic::mips_sra_d:
2103 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0), Op->getOperand(1),
2105 case Intrinsic::mips_srai_b:
2106 case Intrinsic::mips_srai_h:
2107 case Intrinsic::mips_srai_w:
2108 case Intrinsic::mips_srai_d:
2109 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0),
2110 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2111 case Intrinsic::mips_srl_b:
2112 case Intrinsic::mips_srl_h:
2113 case Intrinsic::mips_srl_w:
2114 case Intrinsic::mips_srl_d:
2115 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0), Op->getOperand(1),
2117 case Intrinsic::mips_srli_b:
2118 case Intrinsic::mips_srli_h:
2119 case Intrinsic::mips_srli_w:
2120 case Intrinsic::mips_srli_d:
2121 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0),
2122 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2123 case Intrinsic::mips_subv_b:
2124 case Intrinsic::mips_subv_h:
2125 case Intrinsic::mips_subv_w:
2126 case Intrinsic::mips_subv_d:
2127 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0), Op->getOperand(1),
2129 case Intrinsic::mips_subvi_b:
2130 case Intrinsic::mips_subvi_h:
2131 case Intrinsic::mips_subvi_w:
2132 case Intrinsic::mips_subvi_d:
2133 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0),
2134 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2135 case Intrinsic::mips_vshf_b:
2136 case Intrinsic::mips_vshf_h:
2137 case Intrinsic::mips_vshf_w:
2138 case Intrinsic::mips_vshf_d:
2139 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2140 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3));
2141 case Intrinsic::mips_xor_v:
2142 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0), Op->getOperand(1),
2144 case Intrinsic::mips_xori_b:
2145 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0),
2146 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2150 static SDValue lowerMSALoadIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
2152 SDValue ChainIn = Op->getOperand(0);
2153 SDValue Address = Op->getOperand(2);
2154 SDValue Offset = Op->getOperand(3);
2155 EVT ResTy = Op->getValueType(0);
2156 EVT PtrTy = Address->getValueType(0);
2158 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
2160 return DAG.getLoad(ResTy, DL, ChainIn, Address, MachinePointerInfo(), false,
2164 SDValue MipsSETargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
2165 SelectionDAG &DAG) const {
2166 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
2170 case Intrinsic::mips_extp:
2171 return lowerDSPIntr(Op, DAG, MipsISD::EXTP);
2172 case Intrinsic::mips_extpdp:
2173 return lowerDSPIntr(Op, DAG, MipsISD::EXTPDP);
2174 case Intrinsic::mips_extr_w:
2175 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_W);
2176 case Intrinsic::mips_extr_r_w:
2177 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W);
2178 case Intrinsic::mips_extr_rs_w:
2179 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W);
2180 case Intrinsic::mips_extr_s_h:
2181 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H);
2182 case Intrinsic::mips_mthlip:
2183 return lowerDSPIntr(Op, DAG, MipsISD::MTHLIP);
2184 case Intrinsic::mips_mulsaq_s_w_ph:
2185 return lowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH);
2186 case Intrinsic::mips_maq_s_w_phl:
2187 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL);
2188 case Intrinsic::mips_maq_s_w_phr:
2189 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR);
2190 case Intrinsic::mips_maq_sa_w_phl:
2191 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL);
2192 case Intrinsic::mips_maq_sa_w_phr:
2193 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR);
2194 case Intrinsic::mips_dpaq_s_w_ph:
2195 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH);
2196 case Intrinsic::mips_dpsq_s_w_ph:
2197 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH);
2198 case Intrinsic::mips_dpaq_sa_l_w:
2199 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W);
2200 case Intrinsic::mips_dpsq_sa_l_w:
2201 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W);
2202 case Intrinsic::mips_dpaqx_s_w_ph:
2203 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH);
2204 case Intrinsic::mips_dpaqx_sa_w_ph:
2205 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH);
2206 case Intrinsic::mips_dpsqx_s_w_ph:
2207 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH);
2208 case Intrinsic::mips_dpsqx_sa_w_ph:
2209 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH);
2210 case Intrinsic::mips_ld_b:
2211 case Intrinsic::mips_ld_h:
2212 case Intrinsic::mips_ld_w:
2213 case Intrinsic::mips_ld_d:
2214 return lowerMSALoadIntr(Op, DAG, Intr);
2218 static SDValue lowerMSAStoreIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
2220 SDValue ChainIn = Op->getOperand(0);
2221 SDValue Value = Op->getOperand(2);
2222 SDValue Address = Op->getOperand(3);
2223 SDValue Offset = Op->getOperand(4);
2224 EVT PtrTy = Address->getValueType(0);
2226 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
2228 return DAG.getStore(ChainIn, DL, Value, Address, MachinePointerInfo(), false,
2232 SDValue MipsSETargetLowering::lowerINTRINSIC_VOID(SDValue Op,
2233 SelectionDAG &DAG) const {
2234 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
2238 case Intrinsic::mips_st_b:
2239 case Intrinsic::mips_st_h:
2240 case Intrinsic::mips_st_w:
2241 case Intrinsic::mips_st_d:
2242 return lowerMSAStoreIntr(Op, DAG, Intr);
2246 /// \brief Check if the given BuildVectorSDNode is a splat.
2247 /// This method currently relies on DAG nodes being reused when equivalent,
2248 /// so it's possible for this to return false even when isConstantSplat returns
2250 static bool isSplatVector(const BuildVectorSDNode *N) {
2251 unsigned int nOps = N->getNumOperands();
2252 assert(nOps > 1 && "isSplatVector has 0 or 1 sized build vector");
2254 SDValue Operand0 = N->getOperand(0);
2256 for (unsigned int i = 1; i < nOps; ++i) {
2257 if (N->getOperand(i) != Operand0)
2264 // Lower ISD::EXTRACT_VECTOR_ELT into MipsISD::VEXTRACT_SEXT_ELT.
2266 // The non-value bits resulting from ISD::EXTRACT_VECTOR_ELT are undefined. We
2267 // choose to sign-extend but we could have equally chosen zero-extend. The
2268 // DAGCombiner will fold any sign/zero extension of the ISD::EXTRACT_VECTOR_ELT
2269 // result into this node later (possibly changing it to a zero-extend in the
2271 SDValue MipsSETargetLowering::
2272 lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
2274 EVT ResTy = Op->getValueType(0);
2275 SDValue Op0 = Op->getOperand(0);
2276 EVT VecTy = Op0->getValueType(0);
2278 if (!VecTy.is128BitVector())
2281 if (ResTy.isInteger()) {
2282 SDValue Op1 = Op->getOperand(1);
2283 EVT EltTy = VecTy.getVectorElementType();
2284 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, DL, ResTy, Op0, Op1,
2285 DAG.getValueType(EltTy));
2291 static bool isConstantOrUndef(const SDValue Op) {
2292 if (Op->getOpcode() == ISD::UNDEF)
2294 if (dyn_cast<ConstantSDNode>(Op))
2296 if (dyn_cast<ConstantFPSDNode>(Op))
2301 static bool isConstantOrUndefBUILD_VECTOR(const BuildVectorSDNode *Op) {
2302 for (unsigned i = 0; i < Op->getNumOperands(); ++i)
2303 if (isConstantOrUndef(Op->getOperand(i)))
2308 // Lowers ISD::BUILD_VECTOR into appropriate SelectionDAG nodes for the
2311 // Lowers according to the following rules:
2312 // - Constant splats are legal as-is as long as the SplatBitSize is a power of
2313 // 2 less than or equal to 64 and the value fits into a signed 10-bit
2315 // - Constant splats are lowered to bitconverted BUILD_VECTORs if SplatBitSize
2316 // is a power of 2 less than or equal to 64 and the value does not fit into a
2317 // signed 10-bit immediate
2318 // - Non-constant splats are legal as-is.
2319 // - Non-constant non-splats are lowered to sequences of INSERT_VECTOR_ELT.
2320 // - All others are illegal and must be expanded.
2321 SDValue MipsSETargetLowering::lowerBUILD_VECTOR(SDValue Op,
2322 SelectionDAG &DAG) const {
2323 BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op);
2324 EVT ResTy = Op->getValueType(0);
2326 APInt SplatValue, SplatUndef;
2327 unsigned SplatBitSize;
2330 if (!Subtarget.hasMSA() || !ResTy.is128BitVector())
2333 if (Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
2335 !Subtarget.isLittle()) && SplatBitSize <= 64) {
2336 // We can only cope with 8, 16, 32, or 64-bit elements
2337 if (SplatBitSize != 8 && SplatBitSize != 16 && SplatBitSize != 32 &&
2341 // If the value fits into a simm10 then we can use ldi.[bhwd]
2342 // However, if it isn't an integer type we will have to bitcast from an
2343 // integer type first. Also, if there are any undefs, we must lower them
2344 // to defined values first.
2345 if (ResTy.isInteger() && !HasAnyUndefs && SplatValue.isSignedIntN(10))
2350 switch (SplatBitSize) {
2354 ViaVecTy = MVT::v16i8;
2357 ViaVecTy = MVT::v8i16;
2360 ViaVecTy = MVT::v4i32;
2363 // There's no fill.d to fall back on for 64-bit values
2367 // SelectionDAG::getConstant will promote SplatValue appropriately.
2368 SDValue Result = DAG.getConstant(SplatValue, ViaVecTy);
2370 // Bitcast to the type we originally wanted
2371 if (ViaVecTy != ResTy)
2372 Result = DAG.getNode(ISD::BITCAST, SDLoc(Node), ResTy, Result);
2375 } else if (isSplatVector(Node))
2377 else if (!isConstantOrUndefBUILD_VECTOR(Node)) {
2378 // Use INSERT_VECTOR_ELT operations rather than expand to stores.
2379 // The resulting code is the same length as the expansion, but it doesn't
2380 // use memory operations
2381 EVT ResTy = Node->getValueType(0);
2383 assert(ResTy.isVector());
2385 unsigned NumElts = ResTy.getVectorNumElements();
2386 SDValue Vector = DAG.getUNDEF(ResTy);
2387 for (unsigned i = 0; i < NumElts; ++i) {
2388 Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector,
2389 Node->getOperand(i),
2390 DAG.getConstant(i, MVT::i32));
2398 // Lower VECTOR_SHUFFLE into SHF (if possible).
2400 // SHF splits the vector into blocks of four elements, then shuffles these
2401 // elements according to a <4 x i2> constant (encoded as an integer immediate).
2403 // It is therefore possible to lower into SHF when the mask takes the form:
2404 // <a, b, c, d, a+4, b+4, c+4, d+4, a+8, b+8, c+8, d+8, ...>
2405 // When undef's appear they are treated as if they were whatever value is
2406 // necessary in order to fit the above form.
2409 // %2 = shufflevector <8 x i16> %0, <8 x i16> undef,
2410 // <8 x i32> <i32 3, i32 2, i32 1, i32 0,
2411 // i32 7, i32 6, i32 5, i32 4>
2413 // (SHF_H $w0, $w1, 27)
2414 // where the 27 comes from:
2415 // 3 + (2 << 2) + (1 << 4) + (0 << 6)
2416 static SDValue lowerVECTOR_SHUFFLE_SHF(SDValue Op, EVT ResTy,
2417 SmallVector<int, 16> Indices,
2418 SelectionDAG &DAG) {
2419 int SHFIndices[4] = { -1, -1, -1, -1 };
2421 if (Indices.size() < 4)
2424 for (unsigned i = 0; i < 4; ++i) {
2425 for (unsigned j = i; j < Indices.size(); j += 4) {
2426 int Idx = Indices[j];
2428 // Convert from vector index to 4-element subvector index
2429 // If an index refers to an element outside of the subvector then give up
2432 if (Idx < 0 || Idx >= 4)
2436 // If the mask has an undef, replace it with the current index.
2437 // Note that it might still be undef if the current index is also undef
2438 if (SHFIndices[i] == -1)
2439 SHFIndices[i] = Idx;
2441 // Check that non-undef values are the same as in the mask. If they
2442 // aren't then give up
2443 if (!(Idx == -1 || Idx == SHFIndices[i]))
2448 // Calculate the immediate. Replace any remaining undefs with zero
2450 for (int i = 3; i >= 0; --i) {
2451 int Idx = SHFIndices[i];
2460 return DAG.getNode(MipsISD::SHF, SDLoc(Op), ResTy,
2461 DAG.getConstant(Imm, MVT::i32), Op->getOperand(0));
2464 // Lower VECTOR_SHUFFLE into ILVEV (if possible).
2466 // ILVEV interleaves the even elements from each vector.
2468 // It is possible to lower into ILVEV when the mask takes the form:
2469 // <0, n, 2, n+2, 4, n+4, ...>
2470 // where n is the number of elements in the vector.
2472 // When undef's appear in the mask they are treated as if they were whatever
2473 // value is necessary in order to fit the above form.
2474 static SDValue lowerVECTOR_SHUFFLE_ILVEV(SDValue Op, EVT ResTy,
2475 SmallVector<int, 16> Indices,
2476 SelectionDAG &DAG) {
2477 assert ((Indices.size() % 2) == 0);
2479 int WtIdx = ResTy.getVectorNumElements();
2481 for (unsigned i = 0; i < Indices.size(); i += 2) {
2482 if (Indices[i] != -1 && Indices[i] != WsIdx)
2484 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
2490 return DAG.getNode(MipsISD::ILVEV, SDLoc(Op), ResTy, Op->getOperand(0),
2494 // Lower VECTOR_SHUFFLE into ILVOD (if possible).
2496 // ILVOD interleaves the odd elements from each vector.
2498 // It is possible to lower into ILVOD when the mask takes the form:
2499 // <1, n+1, 3, n+3, 5, n+5, ...>
2500 // where n is the number of elements in the vector.
2502 // When undef's appear in the mask they are treated as if they were whatever
2503 // value is necessary in order to fit the above form.
2504 static SDValue lowerVECTOR_SHUFFLE_ILVOD(SDValue Op, EVT ResTy,
2505 SmallVector<int, 16> Indices,
2506 SelectionDAG &DAG) {
2507 assert ((Indices.size() % 2) == 0);
2509 int WtIdx = ResTy.getVectorNumElements() + 1;
2511 for (unsigned i = 0; i < Indices.size(); i += 2) {
2512 if (Indices[i] != -1 && Indices[i] != WsIdx)
2514 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
2520 return DAG.getNode(MipsISD::ILVOD, SDLoc(Op), ResTy, Op->getOperand(0),
2524 // Lower VECTOR_SHUFFLE into ILVL (if possible).
2526 // ILVL interleaves consecutive elements from the left half of each vector.
2528 // It is possible to lower into ILVL when the mask takes the form:
2529 // <0, n, 1, n+1, 2, n+2, ...>
2530 // where n is the number of elements in the vector.
2532 // When undef's appear in the mask they are treated as if they were whatever
2533 // value is necessary in order to fit the above form.
2534 static SDValue lowerVECTOR_SHUFFLE_ILVL(SDValue Op, EVT ResTy,
2535 SmallVector<int, 16> Indices,
2536 SelectionDAG &DAG) {
2537 assert ((Indices.size() % 2) == 0);
2539 int WtIdx = ResTy.getVectorNumElements();
2541 for (unsigned i = 0; i < Indices.size(); i += 2) {
2542 if (Indices[i] != -1 && Indices[i] != WsIdx)
2544 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
2550 return DAG.getNode(MipsISD::ILVL, SDLoc(Op), ResTy, Op->getOperand(0),
2554 // Lower VECTOR_SHUFFLE into ILVR (if possible).
2556 // ILVR interleaves consecutive elements from the right half of each vector.
2558 // It is possible to lower into ILVR when the mask takes the form:
2559 // <x, n+x, x+1, n+x+1, x+2, n+x+2, ...>
2560 // where n is the number of elements in the vector and x is half n.
2562 // When undef's appear in the mask they are treated as if they were whatever
2563 // value is necessary in order to fit the above form.
2564 static SDValue lowerVECTOR_SHUFFLE_ILVR(SDValue Op, EVT ResTy,
2565 SmallVector<int, 16> Indices,
2566 SelectionDAG &DAG) {
2567 assert ((Indices.size() % 2) == 0);
2568 unsigned NumElts = ResTy.getVectorNumElements();
2569 int WsIdx = NumElts / 2;
2570 int WtIdx = NumElts + NumElts / 2;
2572 for (unsigned i = 0; i < Indices.size(); i += 2) {
2573 if (Indices[i] != -1 && Indices[i] != WsIdx)
2575 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx)
2581 return DAG.getNode(MipsISD::ILVR, SDLoc(Op), ResTy, Op->getOperand(0),
2585 // Lower VECTOR_SHUFFLE into PCKEV (if possible).
2587 // PCKEV copies the even elements of each vector into the result vector.
2589 // It is possible to lower into PCKEV when the mask takes the form:
2590 // <0, 2, 4, ..., n, n+2, n+4, ...>
2591 // where n is the number of elements in the vector.
2593 // When undef's appear in the mask they are treated as if they were whatever
2594 // value is necessary in order to fit the above form.
2595 static SDValue lowerVECTOR_SHUFFLE_PCKEV(SDValue Op, EVT ResTy,
2596 SmallVector<int, 16> Indices,
2597 SelectionDAG &DAG) {
2598 assert ((Indices.size() % 2) == 0);
2601 for (unsigned i = 0; i < Indices.size(); ++i) {
2602 if (Indices[i] != -1 && Indices[i] != Idx)
2607 return DAG.getNode(MipsISD::PCKEV, SDLoc(Op), ResTy, Op->getOperand(0),
2611 // Lower VECTOR_SHUFFLE into PCKOD (if possible).
2613 // PCKOD copies the odd elements of each vector into the result vector.
2615 // It is possible to lower into PCKOD when the mask takes the form:
2616 // <1, 3, 5, ..., n+1, n+3, n+5, ...>
2617 // where n is the number of elements in the vector.
2619 // When undef's appear in the mask they are treated as if they were whatever
2620 // value is necessary in order to fit the above form.
2621 static SDValue lowerVECTOR_SHUFFLE_PCKOD(SDValue Op, EVT ResTy,
2622 SmallVector<int, 16> Indices,
2623 SelectionDAG &DAG) {
2624 assert ((Indices.size() % 2) == 0);
2627 for (unsigned i = 0; i < Indices.size(); ++i) {
2628 if (Indices[i] != -1 && Indices[i] != Idx)
2633 return DAG.getNode(MipsISD::PCKOD, SDLoc(Op), ResTy, Op->getOperand(0),
2637 // Lower VECTOR_SHUFFLE into VSHF.
2639 // This mostly consists of converting the shuffle indices in Indices into a
2640 // BUILD_VECTOR and adding it as an operand to the resulting VSHF. There is
2641 // also code to eliminate unused operands of the VECTOR_SHUFFLE. For example,
2642 // if the type is v8i16 and all the indices are less than 8 then the second
2643 // operand is unused and can be replaced with anything. We choose to replace it
2644 // with the used operand since this reduces the number of instructions overall.
2645 static SDValue lowerVECTOR_SHUFFLE_VSHF(SDValue Op, EVT ResTy,
2646 SmallVector<int, 16> Indices,
2647 SelectionDAG &DAG) {
2648 SmallVector<SDValue, 16> Ops;
2651 EVT MaskVecTy = ResTy.changeVectorElementTypeToInteger();
2652 EVT MaskEltTy = MaskVecTy.getVectorElementType();
2653 bool Using1stVec = false;
2654 bool Using2ndVec = false;
2656 int ResTyNumElts = ResTy.getVectorNumElements();
2658 for (int i = 0; i < ResTyNumElts; ++i) {
2659 // Idx == -1 means UNDEF
2660 int Idx = Indices[i];
2662 if (0 <= Idx && Idx < ResTyNumElts)
2664 if (ResTyNumElts <= Idx && Idx < ResTyNumElts * 2)
2668 for (SmallVector<int, 16>::iterator I = Indices.begin(); I != Indices.end();
2670 Ops.push_back(DAG.getTargetConstant(*I, MaskEltTy));
2672 SDValue MaskVec = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecTy, Ops);
2674 if (Using1stVec && Using2ndVec) {
2675 Op0 = Op->getOperand(0);
2676 Op1 = Op->getOperand(1);
2677 } else if (Using1stVec)
2678 Op0 = Op1 = Op->getOperand(0);
2679 else if (Using2ndVec)
2680 Op0 = Op1 = Op->getOperand(1);
2682 llvm_unreachable("shuffle vector mask references neither vector operand?");
2684 // VECTOR_SHUFFLE concatenates the vectors in an vectorwise fashion.
2685 // <0b00, 0b01> + <0b10, 0b11> -> <0b00, 0b01, 0b10, 0b11>
2686 // VSHF concatenates the vectors in a bitwise fashion:
2687 // <0b00, 0b01> + <0b10, 0b11> ->
2688 // 0b0100 + 0b1110 -> 0b01001110
2689 // <0b10, 0b11, 0b00, 0b01>
2690 // We must therefore swap the operands to get the correct result.
2691 return DAG.getNode(MipsISD::VSHF, DL, ResTy, MaskVec, Op1, Op0);
2694 // Lower VECTOR_SHUFFLE into one of a number of instructions depending on the
2695 // indices in the shuffle.
2696 SDValue MipsSETargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
2697 SelectionDAG &DAG) const {
2698 ShuffleVectorSDNode *Node = cast<ShuffleVectorSDNode>(Op);
2699 EVT ResTy = Op->getValueType(0);
2701 if (!ResTy.is128BitVector())
2704 int ResTyNumElts = ResTy.getVectorNumElements();
2705 SmallVector<int, 16> Indices;
2707 for (int i = 0; i < ResTyNumElts; ++i)
2708 Indices.push_back(Node->getMaskElt(i));
2710 SDValue Result = lowerVECTOR_SHUFFLE_SHF(Op, ResTy, Indices, DAG);
2711 if (Result.getNode())
2713 Result = lowerVECTOR_SHUFFLE_ILVEV(Op, ResTy, Indices, DAG);
2714 if (Result.getNode())
2716 Result = lowerVECTOR_SHUFFLE_ILVOD(Op, ResTy, Indices, DAG);
2717 if (Result.getNode())
2719 Result = lowerVECTOR_SHUFFLE_ILVL(Op, ResTy, Indices, DAG);
2720 if (Result.getNode())
2722 Result = lowerVECTOR_SHUFFLE_ILVR(Op, ResTy, Indices, DAG);
2723 if (Result.getNode())
2725 Result = lowerVECTOR_SHUFFLE_PCKEV(Op, ResTy, Indices, DAG);
2726 if (Result.getNode())
2728 Result = lowerVECTOR_SHUFFLE_PCKOD(Op, ResTy, Indices, DAG);
2729 if (Result.getNode())
2731 return lowerVECTOR_SHUFFLE_VSHF(Op, ResTy, Indices, DAG);
2734 MachineBasicBlock * MipsSETargetLowering::
2735 emitBPOSGE32(MachineInstr *MI, MachineBasicBlock *BB) const{
2737 // bposge32_pseudo $vr0
2747 // $vr0 = phi($vr2, $fbb, $vr1, $tbb)
2749 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2750 const TargetInstrInfo *TII =
2751 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2752 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
2753 DebugLoc DL = MI->getDebugLoc();
2754 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2755 MachineFunction::iterator It = std::next(MachineFunction::iterator(BB));
2756 MachineFunction *F = BB->getParent();
2757 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
2758 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
2759 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
2762 F->insert(It, Sink);
2764 // Transfer the remainder of BB and its successor edges to Sink.
2765 Sink->splice(Sink->begin(), BB, std::next(MachineBasicBlock::iterator(MI)),
2767 Sink->transferSuccessorsAndUpdatePHIs(BB);
2770 BB->addSuccessor(FBB);
2771 BB->addSuccessor(TBB);
2772 FBB->addSuccessor(Sink);
2773 TBB->addSuccessor(Sink);
2775 // Insert the real bposge32 instruction to $BB.
2776 BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB);
2779 unsigned VR2 = RegInfo.createVirtualRegister(RC);
2780 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2)
2781 .addReg(Mips::ZERO).addImm(0);
2782 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
2785 unsigned VR1 = RegInfo.createVirtualRegister(RC);
2786 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1)
2787 .addReg(Mips::ZERO).addImm(1);
2789 // Insert phi function to $Sink.
2790 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
2791 MI->getOperand(0).getReg())
2792 .addReg(VR2).addMBB(FBB).addReg(VR1).addMBB(TBB);
2794 MI->eraseFromParent(); // The pseudo instruction is gone now.
2798 MachineBasicBlock * MipsSETargetLowering::
2799 emitMSACBranchPseudo(MachineInstr *MI, MachineBasicBlock *BB,
2800 unsigned BranchOp) const{
2802 // vany_nonzero $rd, $ws
2813 // $rd = phi($rd1, $fbb, $rd2, $tbb)
2815 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2816 const TargetInstrInfo *TII =
2817 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2818 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
2819 DebugLoc DL = MI->getDebugLoc();
2820 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2821 MachineFunction::iterator It = std::next(MachineFunction::iterator(BB));
2822 MachineFunction *F = BB->getParent();
2823 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
2824 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
2825 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
2828 F->insert(It, Sink);
2830 // Transfer the remainder of BB and its successor edges to Sink.
2831 Sink->splice(Sink->begin(), BB, std::next(MachineBasicBlock::iterator(MI)),
2833 Sink->transferSuccessorsAndUpdatePHIs(BB);
2836 BB->addSuccessor(FBB);
2837 BB->addSuccessor(TBB);
2838 FBB->addSuccessor(Sink);
2839 TBB->addSuccessor(Sink);
2841 // Insert the real bnz.b instruction to $BB.
2842 BuildMI(BB, DL, TII->get(BranchOp))
2843 .addReg(MI->getOperand(1).getReg())
2847 unsigned RD1 = RegInfo.createVirtualRegister(RC);
2848 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), RD1)
2849 .addReg(Mips::ZERO).addImm(0);
2850 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
2853 unsigned RD2 = RegInfo.createVirtualRegister(RC);
2854 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), RD2)
2855 .addReg(Mips::ZERO).addImm(1);
2857 // Insert phi function to $Sink.
2858 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
2859 MI->getOperand(0).getReg())
2860 .addReg(RD1).addMBB(FBB).addReg(RD2).addMBB(TBB);
2862 MI->eraseFromParent(); // The pseudo instruction is gone now.
2866 // Emit the COPY_FW pseudo instruction.
2868 // copy_fw_pseudo $fd, $ws, n
2870 // copy_u_w $rt, $ws, $n
2873 // When n is zero, the equivalent operation can be performed with (potentially)
2874 // zero instructions due to register overlaps. This optimization is never valid
2875 // for lane 1 because it would require FR=0 mode which isn't supported by MSA.
2876 MachineBasicBlock * MipsSETargetLowering::
2877 emitCOPY_FW(MachineInstr *MI, MachineBasicBlock *BB) const{
2878 const TargetInstrInfo *TII =
2879 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2880 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2881 DebugLoc DL = MI->getDebugLoc();
2882 unsigned Fd = MI->getOperand(0).getReg();
2883 unsigned Ws = MI->getOperand(1).getReg();
2884 unsigned Lane = MI->getOperand(2).getImm();
2887 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_lo);
2889 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
2891 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wt).addReg(Ws).addImm(Lane);
2892 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_lo);
2895 MI->eraseFromParent(); // The pseudo instruction is gone now.
2899 // Emit the COPY_FD pseudo instruction.
2901 // copy_fd_pseudo $fd, $ws, n
2903 // splati.d $wt, $ws, $n
2904 // copy $fd, $wt:sub_64
2906 // When n is zero, the equivalent operation can be performed with (potentially)
2907 // zero instructions due to register overlaps. This optimization is always
2908 // valid because FR=1 mode which is the only supported mode in MSA.
2909 MachineBasicBlock * MipsSETargetLowering::
2910 emitCOPY_FD(MachineInstr *MI, MachineBasicBlock *BB) const{
2911 assert(Subtarget.isFP64bit());
2913 const TargetInstrInfo *TII =
2914 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2915 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2916 unsigned Fd = MI->getOperand(0).getReg();
2917 unsigned Ws = MI->getOperand(1).getReg();
2918 unsigned Lane = MI->getOperand(2).getImm() * 2;
2919 DebugLoc DL = MI->getDebugLoc();
2922 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_64);
2924 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
2926 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wt).addReg(Ws).addImm(1);
2927 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_64);
2930 MI->eraseFromParent(); // The pseudo instruction is gone now.
2934 // Emit the INSERT_FW pseudo instruction.
2936 // insert_fw_pseudo $wd, $wd_in, $n, $fs
2938 // subreg_to_reg $wt:sub_lo, $fs
2939 // insve_w $wd[$n], $wd_in, $wt[0]
2941 MipsSETargetLowering::emitINSERT_FW(MachineInstr *MI,
2942 MachineBasicBlock *BB) const {
2943 const TargetInstrInfo *TII =
2944 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2945 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2946 DebugLoc DL = MI->getDebugLoc();
2947 unsigned Wd = MI->getOperand(0).getReg();
2948 unsigned Wd_in = MI->getOperand(1).getReg();
2949 unsigned Lane = MI->getOperand(2).getImm();
2950 unsigned Fs = MI->getOperand(3).getReg();
2951 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
2953 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
2956 .addImm(Mips::sub_lo);
2957 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_W), Wd)
2963 MI->eraseFromParent(); // The pseudo instruction is gone now.
2967 // Emit the INSERT_FD pseudo instruction.
2969 // insert_fd_pseudo $wd, $fs, n
2971 // subreg_to_reg $wt:sub_64, $fs
2972 // insve_d $wd[$n], $wd_in, $wt[0]
2974 MipsSETargetLowering::emitINSERT_FD(MachineInstr *MI,
2975 MachineBasicBlock *BB) const {
2976 assert(Subtarget.isFP64bit());
2978 const TargetInstrInfo *TII =
2979 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2980 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2981 DebugLoc DL = MI->getDebugLoc();
2982 unsigned Wd = MI->getOperand(0).getReg();
2983 unsigned Wd_in = MI->getOperand(1).getReg();
2984 unsigned Lane = MI->getOperand(2).getImm();
2985 unsigned Fs = MI->getOperand(3).getReg();
2986 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
2988 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
2991 .addImm(Mips::sub_64);
2992 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_D), Wd)
2998 MI->eraseFromParent(); // The pseudo instruction is gone now.
3002 // Emit the INSERT_([BHWD]|F[WD])_VIDX pseudo instruction.
3005 // (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $rs)
3007 // (SLL $lanetmp1, $lane, <log2size)
3008 // (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
3009 // (INSERT_[BHWD], $wdtmp2, $wdtmp1, 0, $rs)
3010 // (NEG $lanetmp2, $lanetmp1)
3011 // (SLD_B $wd, $wdtmp2, $wdtmp2, $lanetmp2)
3013 // For floating point:
3014 // (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $fs)
3016 // (SUBREG_TO_REG $wt, $fs, <subreg>)
3017 // (SLL $lanetmp1, $lane, <log2size)
3018 // (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
3019 // (INSVE_[WD], $wdtmp2, 0, $wdtmp1, 0)
3020 // (NEG $lanetmp2, $lanetmp1)
3021 // (SLD_B $wd, $wdtmp2, $wdtmp2, $lanetmp2)
3023 MipsSETargetLowering::emitINSERT_DF_VIDX(MachineInstr *MI,
3024 MachineBasicBlock *BB,
3025 unsigned EltSizeInBytes,
3027 const TargetInstrInfo *TII =
3028 getTargetMachine().getSubtargetImpl()->getInstrInfo();
3029 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3030 DebugLoc DL = MI->getDebugLoc();
3031 unsigned Wd = MI->getOperand(0).getReg();
3032 unsigned SrcVecReg = MI->getOperand(1).getReg();
3033 unsigned LaneReg = MI->getOperand(2).getReg();
3034 unsigned SrcValReg = MI->getOperand(3).getReg();
3036 const TargetRegisterClass *VecRC = nullptr;
3037 const TargetRegisterClass *GPRRC =
3038 Subtarget.isGP64bit() ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
3039 unsigned EltLog2Size;
3040 unsigned InsertOp = 0;
3041 unsigned InsveOp = 0;
3042 switch (EltSizeInBytes) {
3044 llvm_unreachable("Unexpected size");
3047 InsertOp = Mips::INSERT_B;
3048 InsveOp = Mips::INSVE_B;
3049 VecRC = &Mips::MSA128BRegClass;
3053 InsertOp = Mips::INSERT_H;
3054 InsveOp = Mips::INSVE_H;
3055 VecRC = &Mips::MSA128HRegClass;
3059 InsertOp = Mips::INSERT_W;
3060 InsveOp = Mips::INSVE_W;
3061 VecRC = &Mips::MSA128WRegClass;
3065 InsertOp = Mips::INSERT_D;
3066 InsveOp = Mips::INSVE_D;
3067 VecRC = &Mips::MSA128DRegClass;
3072 unsigned Wt = RegInfo.createVirtualRegister(VecRC);
3073 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
3076 .addImm(EltSizeInBytes == 8 ? Mips::sub_64 : Mips::sub_lo);
3080 // Convert the lane index into a byte index
3081 if (EltSizeInBytes != 1) {
3082 unsigned LaneTmp1 = RegInfo.createVirtualRegister(GPRRC);
3083 BuildMI(*BB, MI, DL, TII->get(Mips::SLL), LaneTmp1)
3085 .addImm(EltLog2Size);
3089 // Rotate bytes around so that the desired lane is element zero
3090 unsigned WdTmp1 = RegInfo.createVirtualRegister(VecRC);
3091 BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), WdTmp1)
3096 unsigned WdTmp2 = RegInfo.createVirtualRegister(VecRC);
3098 // Use insve.df to insert to element zero
3099 BuildMI(*BB, MI, DL, TII->get(InsveOp), WdTmp2)
3105 // Use insert.df to insert to element zero
3106 BuildMI(*BB, MI, DL, TII->get(InsertOp), WdTmp2)
3112 // Rotate elements the rest of the way for a full rotation.
3113 // sld.df inteprets $rt modulo the number of columns so we only need to negate
3114 // the lane index to do this.
3115 unsigned LaneTmp2 = RegInfo.createVirtualRegister(GPRRC);
3116 BuildMI(*BB, MI, DL, TII->get(Mips::SUB), LaneTmp2)
3119 BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), Wd)
3124 MI->eraseFromParent(); // The pseudo instruction is gone now.
3128 // Emit the FILL_FW pseudo instruction.
3130 // fill_fw_pseudo $wd, $fs
3132 // implicit_def $wt1
3133 // insert_subreg $wt2:subreg_lo, $wt1, $fs
3134 // splati.w $wd, $wt2[0]
3136 MipsSETargetLowering::emitFILL_FW(MachineInstr *MI,
3137 MachineBasicBlock *BB) const {
3138 const TargetInstrInfo *TII =
3139 getTargetMachine().getSubtargetImpl()->getInstrInfo();
3140 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3141 DebugLoc DL = MI->getDebugLoc();
3142 unsigned Wd = MI->getOperand(0).getReg();
3143 unsigned Fs = MI->getOperand(1).getReg();
3144 unsigned Wt1 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3145 unsigned Wt2 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3147 BuildMI(*BB, MI, DL, TII->get(Mips::IMPLICIT_DEF), Wt1);
3148 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2)
3151 .addImm(Mips::sub_lo);
3152 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wd).addReg(Wt2).addImm(0);
3154 MI->eraseFromParent(); // The pseudo instruction is gone now.
3158 // Emit the FILL_FD pseudo instruction.
3160 // fill_fd_pseudo $wd, $fs
3162 // implicit_def $wt1
3163 // insert_subreg $wt2:subreg_64, $wt1, $fs
3164 // splati.d $wd, $wt2[0]
3166 MipsSETargetLowering::emitFILL_FD(MachineInstr *MI,
3167 MachineBasicBlock *BB) const {
3168 assert(Subtarget.isFP64bit());
3170 const TargetInstrInfo *TII =
3171 getTargetMachine().getSubtargetImpl()->getInstrInfo();
3172 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3173 DebugLoc DL = MI->getDebugLoc();
3174 unsigned Wd = MI->getOperand(0).getReg();
3175 unsigned Fs = MI->getOperand(1).getReg();
3176 unsigned Wt1 = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3177 unsigned Wt2 = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3179 BuildMI(*BB, MI, DL, TII->get(Mips::IMPLICIT_DEF), Wt1);
3180 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2)
3183 .addImm(Mips::sub_64);
3184 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wd).addReg(Wt2).addImm(0);
3186 MI->eraseFromParent(); // The pseudo instruction is gone now.
3190 // Emit the FEXP2_W_1 pseudo instructions.
3192 // fexp2_w_1_pseudo $wd, $wt
3195 // fexp2.w $wd, $ws, $wt
3197 MipsSETargetLowering::emitFEXP2_W_1(MachineInstr *MI,
3198 MachineBasicBlock *BB) const {
3199 const TargetInstrInfo *TII =
3200 getTargetMachine().getSubtargetImpl()->getInstrInfo();
3201 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3202 const TargetRegisterClass *RC = &Mips::MSA128WRegClass;
3203 unsigned Ws1 = RegInfo.createVirtualRegister(RC);
3204 unsigned Ws2 = RegInfo.createVirtualRegister(RC);
3205 DebugLoc DL = MI->getDebugLoc();
3207 // Splat 1.0 into a vector
3208 BuildMI(*BB, MI, DL, TII->get(Mips::LDI_W), Ws1).addImm(1);
3209 BuildMI(*BB, MI, DL, TII->get(Mips::FFINT_U_W), Ws2).addReg(Ws1);
3211 // Emit 1.0 * fexp2(Wt)
3212 BuildMI(*BB, MI, DL, TII->get(Mips::FEXP2_W), MI->getOperand(0).getReg())
3214 .addReg(MI->getOperand(1).getReg());
3216 MI->eraseFromParent(); // The pseudo instruction is gone now.
3220 // Emit the FEXP2_D_1 pseudo instructions.
3222 // fexp2_d_1_pseudo $wd, $wt
3225 // fexp2.d $wd, $ws, $wt
3227 MipsSETargetLowering::emitFEXP2_D_1(MachineInstr *MI,
3228 MachineBasicBlock *BB) const {
3229 const TargetInstrInfo *TII =
3230 getTargetMachine().getSubtargetImpl()->getInstrInfo();
3231 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3232 const TargetRegisterClass *RC = &Mips::MSA128DRegClass;
3233 unsigned Ws1 = RegInfo.createVirtualRegister(RC);
3234 unsigned Ws2 = RegInfo.createVirtualRegister(RC);
3235 DebugLoc DL = MI->getDebugLoc();
3237 // Splat 1.0 into a vector
3238 BuildMI(*BB, MI, DL, TII->get(Mips::LDI_D), Ws1).addImm(1);
3239 BuildMI(*BB, MI, DL, TII->get(Mips::FFINT_U_D), Ws2).addReg(Ws1);
3241 // Emit 1.0 * fexp2(Wt)
3242 BuildMI(*BB, MI, DL, TII->get(Mips::FEXP2_D), MI->getOperand(0).getReg())
3244 .addReg(MI->getOperand(1).getReg());
3246 MI->eraseFromParent(); // The pseudo instruction is gone now.