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 UseMipsTailCalls("mips-tail-calls", cl::Hidden,
31 cl::desc("MIPS: permit 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 for (MVT VT0 : MVT::vector_valuetypes()) {
50 for (MVT VT1 : MVT::vector_valuetypes()) {
51 setTruncStoreAction(VT0, VT1, Expand);
52 setLoadExtAction(ISD::SEXTLOAD, VT0, VT1, Expand);
53 setLoadExtAction(ISD::ZEXTLOAD, VT0, VT1, Expand);
54 setLoadExtAction(ISD::EXTLOAD, VT0, VT1, Expand);
59 if (Subtarget.hasDSP()) {
60 MVT::SimpleValueType VecTys[2] = {MVT::v2i16, MVT::v4i8};
62 for (unsigned i = 0; i < array_lengthof(VecTys); ++i) {
63 addRegisterClass(VecTys[i], &Mips::DSPRRegClass);
65 // Expand all builtin opcodes.
66 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
67 setOperationAction(Opc, VecTys[i], Expand);
69 setOperationAction(ISD::ADD, VecTys[i], Legal);
70 setOperationAction(ISD::SUB, VecTys[i], Legal);
71 setOperationAction(ISD::LOAD, VecTys[i], Legal);
72 setOperationAction(ISD::STORE, VecTys[i], Legal);
73 setOperationAction(ISD::BITCAST, VecTys[i], Legal);
76 setTargetDAGCombine(ISD::SHL);
77 setTargetDAGCombine(ISD::SRA);
78 setTargetDAGCombine(ISD::SRL);
79 setTargetDAGCombine(ISD::SETCC);
80 setTargetDAGCombine(ISD::VSELECT);
83 if (Subtarget.hasDSPR2())
84 setOperationAction(ISD::MUL, MVT::v2i16, Legal);
86 if (Subtarget.hasMSA()) {
87 addMSAIntType(MVT::v16i8, &Mips::MSA128BRegClass);
88 addMSAIntType(MVT::v8i16, &Mips::MSA128HRegClass);
89 addMSAIntType(MVT::v4i32, &Mips::MSA128WRegClass);
90 addMSAIntType(MVT::v2i64, &Mips::MSA128DRegClass);
91 addMSAFloatType(MVT::v8f16, &Mips::MSA128HRegClass);
92 addMSAFloatType(MVT::v4f32, &Mips::MSA128WRegClass);
93 addMSAFloatType(MVT::v2f64, &Mips::MSA128DRegClass);
95 // f16 is a storage-only type, always promote it to f32.
96 addRegisterClass(MVT::f16, &Mips::MSA128HRegClass);
97 setOperationAction(ISD::SETCC, MVT::f16, Promote);
98 setOperationAction(ISD::BR_CC, MVT::f16, Promote);
99 setOperationAction(ISD::SELECT_CC, MVT::f16, Promote);
100 setOperationAction(ISD::SELECT, MVT::f16, Promote);
101 setOperationAction(ISD::FADD, MVT::f16, Promote);
102 setOperationAction(ISD::FSUB, MVT::f16, Promote);
103 setOperationAction(ISD::FMUL, MVT::f16, Promote);
104 setOperationAction(ISD::FDIV, MVT::f16, Promote);
105 setOperationAction(ISD::FREM, MVT::f16, Promote);
106 setOperationAction(ISD::FMA, MVT::f16, Promote);
107 setOperationAction(ISD::FNEG, MVT::f16, Promote);
108 setOperationAction(ISD::FABS, MVT::f16, Promote);
109 setOperationAction(ISD::FCEIL, MVT::f16, Promote);
110 setOperationAction(ISD::FCOPYSIGN, MVT::f16, Promote);
111 setOperationAction(ISD::FCOS, MVT::f16, Promote);
112 setOperationAction(ISD::FP_EXTEND, MVT::f16, Promote);
113 setOperationAction(ISD::FFLOOR, MVT::f16, Promote);
114 setOperationAction(ISD::FNEARBYINT, MVT::f16, Promote);
115 setOperationAction(ISD::FPOW, MVT::f16, Promote);
116 setOperationAction(ISD::FPOWI, MVT::f16, Promote);
117 setOperationAction(ISD::FRINT, MVT::f16, Promote);
118 setOperationAction(ISD::FSIN, MVT::f16, Promote);
119 setOperationAction(ISD::FSINCOS, MVT::f16, Promote);
120 setOperationAction(ISD::FSQRT, MVT::f16, Promote);
121 setOperationAction(ISD::FEXP, MVT::f16, Promote);
122 setOperationAction(ISD::FEXP2, MVT::f16, Promote);
123 setOperationAction(ISD::FLOG, MVT::f16, Promote);
124 setOperationAction(ISD::FLOG2, MVT::f16, Promote);
125 setOperationAction(ISD::FLOG10, MVT::f16, Promote);
126 setOperationAction(ISD::FROUND, MVT::f16, Promote);
127 setOperationAction(ISD::FTRUNC, MVT::f16, Promote);
128 setOperationAction(ISD::FMINNUM, MVT::f16, Promote);
129 setOperationAction(ISD::FMAXNUM, MVT::f16, Promote);
130 setOperationAction(ISD::FMINNAN, MVT::f16, Promote);
131 setOperationAction(ISD::FMAXNAN, MVT::f16, Promote);
133 setTargetDAGCombine(ISD::AND);
134 setTargetDAGCombine(ISD::OR);
135 setTargetDAGCombine(ISD::SRA);
136 setTargetDAGCombine(ISD::VSELECT);
137 setTargetDAGCombine(ISD::XOR);
140 if (!Subtarget.useSoftFloat()) {
141 addRegisterClass(MVT::f32, &Mips::FGR32RegClass);
143 // When dealing with single precision only, use libcalls
144 if (!Subtarget.isSingleFloat()) {
145 if (Subtarget.isFP64bit())
146 addRegisterClass(MVT::f64, &Mips::FGR64RegClass);
148 addRegisterClass(MVT::f64, &Mips::AFGR64RegClass);
152 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Custom);
153 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Custom);
154 setOperationAction(ISD::MULHS, MVT::i32, Custom);
155 setOperationAction(ISD::MULHU, MVT::i32, Custom);
157 if (Subtarget.hasCnMips())
158 setOperationAction(ISD::MUL, MVT::i64, Legal);
159 else if (Subtarget.isGP64bit())
160 setOperationAction(ISD::MUL, MVT::i64, Custom);
162 if (Subtarget.isGP64bit()) {
163 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Custom);
164 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Custom);
165 setOperationAction(ISD::MULHS, MVT::i64, Custom);
166 setOperationAction(ISD::MULHU, MVT::i64, Custom);
167 setOperationAction(ISD::SDIVREM, MVT::i64, Custom);
168 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
171 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
172 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
174 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
175 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
176 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
177 setOperationAction(ISD::LOAD, MVT::i32, Custom);
178 setOperationAction(ISD::STORE, MVT::i32, Custom);
180 setTargetDAGCombine(ISD::ADDE);
181 setTargetDAGCombine(ISD::SUBE);
182 setTargetDAGCombine(ISD::MUL);
184 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
185 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
186 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
189 setOperationAction(ISD::LOAD, MVT::f64, Custom);
190 setOperationAction(ISD::STORE, MVT::f64, Custom);
193 if (Subtarget.hasMips32r6()) {
194 // MIPS32r6 replaces the accumulator-based multiplies with a three register
196 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
197 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
198 setOperationAction(ISD::MUL, MVT::i32, Legal);
199 setOperationAction(ISD::MULHS, MVT::i32, Legal);
200 setOperationAction(ISD::MULHU, MVT::i32, Legal);
202 // MIPS32r6 replaces the accumulator-based division/remainder with separate
203 // three register division and remainder instructions.
204 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
205 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
206 setOperationAction(ISD::SDIV, MVT::i32, Legal);
207 setOperationAction(ISD::UDIV, MVT::i32, Legal);
208 setOperationAction(ISD::SREM, MVT::i32, Legal);
209 setOperationAction(ISD::UREM, MVT::i32, Legal);
211 // MIPS32r6 replaces conditional moves with an equivalent that removes the
212 // need for three GPR read ports.
213 setOperationAction(ISD::SETCC, MVT::i32, Legal);
214 setOperationAction(ISD::SELECT, MVT::i32, Legal);
215 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
217 setOperationAction(ISD::SETCC, MVT::f32, Legal);
218 setOperationAction(ISD::SELECT, MVT::f32, Legal);
219 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
221 assert(Subtarget.isFP64bit() && "FR=1 is required for MIPS32r6");
222 setOperationAction(ISD::SETCC, MVT::f64, Legal);
223 setOperationAction(ISD::SELECT, MVT::f64, Legal);
224 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
226 setOperationAction(ISD::BRCOND, MVT::Other, Legal);
228 // Floating point > and >= are supported via < and <=
229 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
230 setCondCodeAction(ISD::SETOGT, MVT::f32, Expand);
231 setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
232 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
234 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
235 setCondCodeAction(ISD::SETOGT, MVT::f64, Expand);
236 setCondCodeAction(ISD::SETUGE, MVT::f64, Expand);
237 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
240 if (Subtarget.hasMips64r6()) {
241 // MIPS64r6 replaces the accumulator-based multiplies with a three register
243 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
244 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
245 setOperationAction(ISD::MUL, MVT::i64, Legal);
246 setOperationAction(ISD::MULHS, MVT::i64, Legal);
247 setOperationAction(ISD::MULHU, MVT::i64, Legal);
249 // MIPS32r6 replaces the accumulator-based division/remainder with separate
250 // three register division and remainder instructions.
251 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
252 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
253 setOperationAction(ISD::SDIV, MVT::i64, Legal);
254 setOperationAction(ISD::UDIV, MVT::i64, Legal);
255 setOperationAction(ISD::SREM, MVT::i64, Legal);
256 setOperationAction(ISD::UREM, MVT::i64, Legal);
258 // MIPS64r6 replaces conditional moves with an equivalent that removes the
259 // need for three GPR read ports.
260 setOperationAction(ISD::SETCC, MVT::i64, Legal);
261 setOperationAction(ISD::SELECT, MVT::i64, Legal);
262 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
265 computeRegisterProperties(Subtarget.getRegisterInfo());
268 const MipsTargetLowering *
269 llvm::createMipsSETargetLowering(const MipsTargetMachine &TM,
270 const MipsSubtarget &STI) {
271 return new MipsSETargetLowering(TM, STI);
274 const TargetRegisterClass *
275 MipsSETargetLowering::getRepRegClassFor(MVT VT) const {
276 if (VT == MVT::Untyped)
277 return Subtarget.hasDSP() ? &Mips::ACC64DSPRegClass : &Mips::ACC64RegClass;
279 return TargetLowering::getRepRegClassFor(VT);
282 // Enable MSA support for the given integer type and Register class.
283 void MipsSETargetLowering::
284 addMSAIntType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
285 addRegisterClass(Ty, RC);
287 // Expand all builtin opcodes.
288 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
289 setOperationAction(Opc, Ty, Expand);
291 setOperationAction(ISD::BITCAST, Ty, Legal);
292 setOperationAction(ISD::LOAD, Ty, Legal);
293 setOperationAction(ISD::STORE, Ty, Legal);
294 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Custom);
295 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
296 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
298 setOperationAction(ISD::ADD, Ty, Legal);
299 setOperationAction(ISD::AND, Ty, Legal);
300 setOperationAction(ISD::CTLZ, Ty, Legal);
301 setOperationAction(ISD::CTPOP, Ty, Legal);
302 setOperationAction(ISD::MUL, Ty, Legal);
303 setOperationAction(ISD::OR, Ty, Legal);
304 setOperationAction(ISD::SDIV, Ty, Legal);
305 setOperationAction(ISD::SREM, Ty, Legal);
306 setOperationAction(ISD::SHL, Ty, Legal);
307 setOperationAction(ISD::SRA, Ty, Legal);
308 setOperationAction(ISD::SRL, Ty, Legal);
309 setOperationAction(ISD::SUB, Ty, Legal);
310 setOperationAction(ISD::UDIV, Ty, Legal);
311 setOperationAction(ISD::UREM, Ty, Legal);
312 setOperationAction(ISD::VECTOR_SHUFFLE, Ty, Custom);
313 setOperationAction(ISD::VSELECT, Ty, Legal);
314 setOperationAction(ISD::XOR, Ty, Legal);
316 if (Ty == MVT::v4i32 || Ty == MVT::v2i64) {
317 setOperationAction(ISD::FP_TO_SINT, Ty, Legal);
318 setOperationAction(ISD::FP_TO_UINT, Ty, Legal);
319 setOperationAction(ISD::SINT_TO_FP, Ty, Legal);
320 setOperationAction(ISD::UINT_TO_FP, Ty, Legal);
323 setOperationAction(ISD::SETCC, Ty, Legal);
324 setCondCodeAction(ISD::SETNE, Ty, Expand);
325 setCondCodeAction(ISD::SETGE, Ty, Expand);
326 setCondCodeAction(ISD::SETGT, Ty, Expand);
327 setCondCodeAction(ISD::SETUGE, Ty, Expand);
328 setCondCodeAction(ISD::SETUGT, Ty, Expand);
331 // Enable MSA support for the given floating-point type and Register class.
332 void MipsSETargetLowering::
333 addMSAFloatType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
334 addRegisterClass(Ty, RC);
336 // Expand all builtin opcodes.
337 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
338 setOperationAction(Opc, Ty, Expand);
340 setOperationAction(ISD::LOAD, Ty, Legal);
341 setOperationAction(ISD::STORE, Ty, Legal);
342 setOperationAction(ISD::BITCAST, Ty, Legal);
343 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Legal);
344 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
345 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
347 if (Ty != MVT::v8f16) {
348 setOperationAction(ISD::FABS, Ty, Legal);
349 setOperationAction(ISD::FADD, Ty, Legal);
350 setOperationAction(ISD::FDIV, Ty, Legal);
351 setOperationAction(ISD::FEXP2, Ty, Legal);
352 setOperationAction(ISD::FLOG2, Ty, Legal);
353 setOperationAction(ISD::FMA, Ty, Legal);
354 setOperationAction(ISD::FMUL, Ty, Legal);
355 setOperationAction(ISD::FRINT, Ty, Legal);
356 setOperationAction(ISD::FSQRT, Ty, Legal);
357 setOperationAction(ISD::FSUB, Ty, Legal);
358 setOperationAction(ISD::VSELECT, Ty, Legal);
360 setOperationAction(ISD::SETCC, Ty, Legal);
361 setCondCodeAction(ISD::SETOGE, Ty, Expand);
362 setCondCodeAction(ISD::SETOGT, Ty, Expand);
363 setCondCodeAction(ISD::SETUGE, Ty, Expand);
364 setCondCodeAction(ISD::SETUGT, Ty, Expand);
365 setCondCodeAction(ISD::SETGE, Ty, Expand);
366 setCondCodeAction(ISD::SETGT, Ty, Expand);
371 MipsSETargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
375 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
377 if (Subtarget.systemSupportsUnalignedAccess()) {
378 // MIPS32r6/MIPS64r6 is required to support unaligned access. It's
379 // implementation defined whether this is handled by hardware, software, or
380 // a hybrid of the two but it's expected that most implementations will
381 // handle the majority of cases in hardware.
398 SDValue MipsSETargetLowering::LowerOperation(SDValue Op,
399 SelectionDAG &DAG) const {
400 switch(Op.getOpcode()) {
401 case ISD::LOAD: return lowerLOAD(Op, DAG);
402 case ISD::STORE: return lowerSTORE(Op, DAG);
403 case ISD::SMUL_LOHI: return lowerMulDiv(Op, MipsISD::Mult, true, true, DAG);
404 case ISD::UMUL_LOHI: return lowerMulDiv(Op, MipsISD::Multu, true, true, DAG);
405 case ISD::MULHS: return lowerMulDiv(Op, MipsISD::Mult, false, true, DAG);
406 case ISD::MULHU: return lowerMulDiv(Op, MipsISD::Multu, false, true, DAG);
407 case ISD::MUL: return lowerMulDiv(Op, MipsISD::Mult, true, false, DAG);
408 case ISD::SDIVREM: return lowerMulDiv(Op, MipsISD::DivRem, true, true, DAG);
409 case ISD::UDIVREM: return lowerMulDiv(Op, MipsISD::DivRemU, true, true,
411 case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG);
412 case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG);
413 case ISD::INTRINSIC_VOID: return lowerINTRINSIC_VOID(Op, DAG);
414 case ISD::EXTRACT_VECTOR_ELT: return lowerEXTRACT_VECTOR_ELT(Op, DAG);
415 case ISD::BUILD_VECTOR: return lowerBUILD_VECTOR(Op, DAG);
416 case ISD::VECTOR_SHUFFLE: return lowerVECTOR_SHUFFLE(Op, DAG);
419 return MipsTargetLowering::LowerOperation(Op, DAG);
423 // Transforms a subgraph in CurDAG if the following pattern is found:
424 // (addc multLo, Lo0), (adde multHi, Hi0),
426 // multHi/Lo: product of multiplication
427 // Lo0: initial value of Lo register
428 // Hi0: initial value of Hi register
429 // Return true if pattern matching was successful.
430 static bool selectMADD(SDNode *ADDENode, SelectionDAG *CurDAG) {
431 // ADDENode's second operand must be a flag output of an ADDC node in order
432 // for the matching to be successful.
433 SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
435 if (ADDCNode->getOpcode() != ISD::ADDC)
438 SDValue MultHi = ADDENode->getOperand(0);
439 SDValue MultLo = ADDCNode->getOperand(0);
440 SDNode *MultNode = MultHi.getNode();
441 unsigned MultOpc = MultHi.getOpcode();
443 // MultHi and MultLo must be generated by the same node,
444 if (MultLo.getNode() != MultNode)
447 // and it must be a multiplication.
448 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
451 // MultLo amd MultHi must be the first and second output of MultNode
453 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
456 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
457 // of the values of MultNode, in which case MultNode will be removed in later
459 // If there exist users other than ADDENode or ADDCNode, this function returns
460 // here, which will result in MultNode being mapped to a single MULT
461 // instruction node rather than a pair of MULT and MADD instructions being
463 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
468 // Initialize accumulator.
469 SDValue ACCIn = CurDAG->getNode(MipsISD::MTLOHI, DL, MVT::Untyped,
470 ADDCNode->getOperand(1),
471 ADDENode->getOperand(1));
473 // create MipsMAdd(u) node
474 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
476 SDValue MAdd = CurDAG->getNode(MultOpc, DL, MVT::Untyped,
477 MultNode->getOperand(0),// Factor 0
478 MultNode->getOperand(1),// Factor 1
481 // replace uses of adde and addc here
482 if (!SDValue(ADDCNode, 0).use_empty()) {
483 SDValue LoOut = CurDAG->getNode(MipsISD::MFLO, DL, MVT::i32, MAdd);
484 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), LoOut);
486 if (!SDValue(ADDENode, 0).use_empty()) {
487 SDValue HiOut = CurDAG->getNode(MipsISD::MFHI, DL, MVT::i32, MAdd);
488 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), HiOut);
495 // Transforms a subgraph in CurDAG if the following pattern is found:
496 // (addc Lo0, multLo), (sube Hi0, multHi),
498 // multHi/Lo: product of multiplication
499 // Lo0: initial value of Lo register
500 // Hi0: initial value of Hi register
501 // Return true if pattern matching was successful.
502 static bool selectMSUB(SDNode *SUBENode, SelectionDAG *CurDAG) {
503 // SUBENode's second operand must be a flag output of an SUBC node in order
504 // for the matching to be successful.
505 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
507 if (SUBCNode->getOpcode() != ISD::SUBC)
510 SDValue MultHi = SUBENode->getOperand(1);
511 SDValue MultLo = SUBCNode->getOperand(1);
512 SDNode *MultNode = MultHi.getNode();
513 unsigned MultOpc = MultHi.getOpcode();
515 // MultHi and MultLo must be generated by the same node,
516 if (MultLo.getNode() != MultNode)
519 // and it must be a multiplication.
520 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
523 // MultLo amd MultHi must be the first and second output of MultNode
525 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
528 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
529 // of the values of MultNode, in which case MultNode will be removed in later
531 // If there exist users other than SUBENode or SUBCNode, this function returns
532 // here, which will result in MultNode being mapped to a single MULT
533 // instruction node rather than a pair of MULT and MSUB instructions being
535 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
540 // Initialize accumulator.
541 SDValue ACCIn = CurDAG->getNode(MipsISD::MTLOHI, DL, MVT::Untyped,
542 SUBCNode->getOperand(0),
543 SUBENode->getOperand(0));
545 // create MipsSub(u) node
546 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
548 SDValue MSub = CurDAG->getNode(MultOpc, DL, MVT::Glue,
549 MultNode->getOperand(0),// Factor 0
550 MultNode->getOperand(1),// Factor 1
553 // replace uses of sube and subc here
554 if (!SDValue(SUBCNode, 0).use_empty()) {
555 SDValue LoOut = CurDAG->getNode(MipsISD::MFLO, DL, MVT::i32, MSub);
556 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), LoOut);
558 if (!SDValue(SUBENode, 0).use_empty()) {
559 SDValue HiOut = CurDAG->getNode(MipsISD::MFHI, DL, MVT::i32, MSub);
560 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), HiOut);
566 static SDValue performADDECombine(SDNode *N, SelectionDAG &DAG,
567 TargetLowering::DAGCombinerInfo &DCI,
568 const MipsSubtarget &Subtarget) {
569 if (DCI.isBeforeLegalize())
572 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() &&
573 N->getValueType(0) == MVT::i32 && selectMADD(N, &DAG))
574 return SDValue(N, 0);
579 // Fold zero extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT
581 // Performs the following transformations:
582 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to zero extension if its
583 // sign/zero-extension is completely overwritten by the new one performed by
585 // - Removes redundant zero extensions performed by an ISD::AND.
586 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
587 TargetLowering::DAGCombinerInfo &DCI,
588 const MipsSubtarget &Subtarget) {
589 if (!Subtarget.hasMSA())
592 SDValue Op0 = N->getOperand(0);
593 SDValue Op1 = N->getOperand(1);
594 unsigned Op0Opcode = Op0->getOpcode();
596 // (and (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d)
597 // where $d + 1 == 2^n and n == 32
598 // or $d + 1 == 2^n and n <= 32 and ZExt
599 // -> (MipsVExtractZExt $a, $b, $c)
600 if (Op0Opcode == MipsISD::VEXTRACT_SEXT_ELT ||
601 Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT) {
602 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(Op1);
607 int32_t Log2IfPositive = (Mask->getAPIntValue() + 1).exactLogBase2();
609 if (Log2IfPositive <= 0)
610 return SDValue(); // Mask+1 is not a power of 2
612 SDValue Op0Op2 = Op0->getOperand(2);
613 EVT ExtendTy = cast<VTSDNode>(Op0Op2)->getVT();
614 unsigned ExtendTySize = ExtendTy.getSizeInBits();
615 unsigned Log2 = Log2IfPositive;
617 if ((Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT && Log2 >= ExtendTySize) ||
618 Log2 == ExtendTySize) {
619 SDValue Ops[] = { Op0->getOperand(0), Op0->getOperand(1), Op0Op2 };
620 return DAG.getNode(MipsISD::VEXTRACT_ZEXT_ELT, SDLoc(Op0),
622 makeArrayRef(Ops, Op0->getNumOperands()));
629 // Determine if the specified node is a constant vector splat.
631 // Returns true and sets Imm if:
632 // * N is a ISD::BUILD_VECTOR representing a constant splat
634 // This function is quite similar to MipsSEDAGToDAGISel::selectVSplat. The
635 // differences are that it assumes the MSA has already been checked and the
636 // arbitrary requirement for a maximum of 32-bit integers isn't applied (and
637 // must not be in order for binsri.d to be selectable).
638 static bool isVSplat(SDValue N, APInt &Imm, bool IsLittleEndian) {
639 BuildVectorSDNode *Node = dyn_cast<BuildVectorSDNode>(N.getNode());
644 APInt SplatValue, SplatUndef;
645 unsigned SplatBitSize;
648 if (!Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
657 // Test whether the given node is an all-ones build_vector.
658 static bool isVectorAllOnes(SDValue N) {
659 // Look through bitcasts. Endianness doesn't matter because we are looking
660 // for an all-ones value.
661 if (N->getOpcode() == ISD::BITCAST)
662 N = N->getOperand(0);
664 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
669 APInt SplatValue, SplatUndef;
670 unsigned SplatBitSize;
673 // Endianness doesn't matter in this context because we are looking for
674 // an all-ones value.
675 if (BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs))
676 return SplatValue.isAllOnesValue();
681 // Test whether N is the bitwise inverse of OfNode.
682 static bool isBitwiseInverse(SDValue N, SDValue OfNode) {
683 if (N->getOpcode() != ISD::XOR)
686 if (isVectorAllOnes(N->getOperand(0)))
687 return N->getOperand(1) == OfNode;
689 if (isVectorAllOnes(N->getOperand(1)))
690 return N->getOperand(0) == OfNode;
695 // Perform combines where ISD::OR is the root node.
697 // Performs the following transformations:
698 // - (or (and $a, $mask), (and $b, $inv_mask)) => (vselect $mask, $a, $b)
699 // where $inv_mask is the bitwise inverse of $mask and the 'or' has a 128-bit
701 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
702 TargetLowering::DAGCombinerInfo &DCI,
703 const MipsSubtarget &Subtarget) {
704 if (!Subtarget.hasMSA())
707 EVT Ty = N->getValueType(0);
709 if (!Ty.is128BitVector())
712 SDValue Op0 = N->getOperand(0);
713 SDValue Op1 = N->getOperand(1);
715 if (Op0->getOpcode() == ISD::AND && Op1->getOpcode() == ISD::AND) {
716 SDValue Op0Op0 = Op0->getOperand(0);
717 SDValue Op0Op1 = Op0->getOperand(1);
718 SDValue Op1Op0 = Op1->getOperand(0);
719 SDValue Op1Op1 = Op1->getOperand(1);
720 bool IsLittleEndian = !Subtarget.isLittle();
722 SDValue IfSet, IfClr, Cond;
723 bool IsConstantMask = false;
726 // If Op0Op0 is an appropriate mask, try to find it's inverse in either
727 // Op1Op0, or Op1Op1. Keep track of the Cond, IfSet, and IfClr nodes, while
729 // IfClr will be set if we find a valid match.
730 if (isVSplat(Op0Op0, Mask, IsLittleEndian)) {
734 if (isVSplat(Op1Op0, InvMask, IsLittleEndian) &&
735 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
737 else if (isVSplat(Op1Op1, InvMask, IsLittleEndian) &&
738 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
741 IsConstantMask = true;
744 // If IfClr is not yet set, and Op0Op1 is an appropriate mask, try the same
745 // thing again using this mask.
746 // IfClr will be set if we find a valid match.
747 if (!IfClr.getNode() && isVSplat(Op0Op1, Mask, IsLittleEndian)) {
751 if (isVSplat(Op1Op0, InvMask, IsLittleEndian) &&
752 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
754 else if (isVSplat(Op1Op1, InvMask, IsLittleEndian) &&
755 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
758 IsConstantMask = true;
761 // If IfClr is not yet set, try looking for a non-constant match.
762 // IfClr will be set if we find a valid match amongst the eight
764 if (!IfClr.getNode()) {
765 if (isBitwiseInverse(Op0Op0, Op1Op0)) {
769 } else if (isBitwiseInverse(Op0Op1, Op1Op0)) {
773 } else if (isBitwiseInverse(Op0Op0, Op1Op1)) {
777 } else if (isBitwiseInverse(Op0Op1, Op1Op1)) {
781 } else if (isBitwiseInverse(Op1Op0, Op0Op0)) {
785 } else if (isBitwiseInverse(Op1Op1, Op0Op0)) {
789 } else if (isBitwiseInverse(Op1Op0, Op0Op1)) {
793 } else if (isBitwiseInverse(Op1Op1, Op0Op1)) {
800 // At this point, IfClr will be set if we have a valid match.
801 if (!IfClr.getNode())
804 assert(Cond.getNode() && IfSet.getNode());
806 // Fold degenerate cases.
807 if (IsConstantMask) {
808 if (Mask.isAllOnesValue())
814 // Transform the DAG into an equivalent VSELECT.
815 return DAG.getNode(ISD::VSELECT, SDLoc(N), Ty, Cond, IfSet, IfClr);
821 static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG,
822 TargetLowering::DAGCombinerInfo &DCI,
823 const MipsSubtarget &Subtarget) {
824 if (DCI.isBeforeLegalize())
827 if (Subtarget.hasMips32() && N->getValueType(0) == MVT::i32 &&
829 return SDValue(N, 0);
834 static SDValue genConstMult(SDValue X, uint64_t C, const SDLoc &DL, EVT VT,
835 EVT ShiftTy, SelectionDAG &DAG) {
836 // Clear the upper (64 - VT.sizeInBits) bits.
837 C &= ((uint64_t)-1) >> (64 - VT.getSizeInBits());
841 return DAG.getConstant(0, DL, VT);
847 // If c is power of 2, return (shl x, log2(c)).
848 if (isPowerOf2_64(C))
849 return DAG.getNode(ISD::SHL, DL, VT, X,
850 DAG.getConstant(Log2_64(C), DL, ShiftTy));
852 unsigned Log2Ceil = Log2_64_Ceil(C);
853 uint64_t Floor = 1LL << Log2_64(C);
854 uint64_t Ceil = Log2Ceil == 64 ? 0LL : 1LL << Log2Ceil;
856 // If |c - floor_c| <= |c - ceil_c|,
857 // where floor_c = pow(2, floor(log2(c))) and ceil_c = pow(2, ceil(log2(c))),
858 // return (add constMult(x, floor_c), constMult(x, c - floor_c)).
859 if (C - Floor <= Ceil - C) {
860 SDValue Op0 = genConstMult(X, Floor, DL, VT, ShiftTy, DAG);
861 SDValue Op1 = genConstMult(X, C - Floor, DL, VT, ShiftTy, DAG);
862 return DAG.getNode(ISD::ADD, DL, VT, Op0, Op1);
865 // If |c - floor_c| > |c - ceil_c|,
866 // return (sub constMult(x, ceil_c), constMult(x, ceil_c - c)).
867 SDValue Op0 = genConstMult(X, Ceil, DL, VT, ShiftTy, DAG);
868 SDValue Op1 = genConstMult(X, Ceil - C, DL, VT, ShiftTy, DAG);
869 return DAG.getNode(ISD::SUB, DL, VT, Op0, Op1);
872 static SDValue performMULCombine(SDNode *N, SelectionDAG &DAG,
873 const TargetLowering::DAGCombinerInfo &DCI,
874 const MipsSETargetLowering *TL) {
875 EVT VT = N->getValueType(0);
877 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
879 return genConstMult(N->getOperand(0), C->getZExtValue(), SDLoc(N), VT,
880 TL->getScalarShiftAmountTy(DAG.getDataLayout(), VT),
883 return SDValue(N, 0);
886 static SDValue performDSPShiftCombine(unsigned Opc, SDNode *N, EVT Ty,
888 const MipsSubtarget &Subtarget) {
889 // See if this is a vector splat immediate node.
890 APInt SplatValue, SplatUndef;
891 unsigned SplatBitSize;
893 unsigned EltSize = Ty.getScalarSizeInBits();
894 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
896 if (!Subtarget.hasDSP())
900 !BV->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
901 EltSize, !Subtarget.isLittle()) ||
902 (SplatBitSize != EltSize) ||
903 (SplatValue.getZExtValue() >= EltSize))
907 return DAG.getNode(Opc, DL, Ty, N->getOperand(0),
908 DAG.getConstant(SplatValue.getZExtValue(), DL, MVT::i32));
911 static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
912 TargetLowering::DAGCombinerInfo &DCI,
913 const MipsSubtarget &Subtarget) {
914 EVT Ty = N->getValueType(0);
916 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
919 return performDSPShiftCombine(MipsISD::SHLL_DSP, N, Ty, DAG, Subtarget);
922 // Fold sign-extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT for MSA and fold
923 // constant splats into MipsISD::SHRA_DSP for DSPr2.
925 // Performs the following transformations:
926 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to sign extension if its
927 // sign/zero-extension is completely overwritten by the new one performed by
928 // the ISD::SRA and ISD::SHL nodes.
929 // - Removes redundant sign extensions performed by an ISD::SRA and ISD::SHL
932 // See performDSPShiftCombine for more information about the transformation
934 static SDValue performSRACombine(SDNode *N, SelectionDAG &DAG,
935 TargetLowering::DAGCombinerInfo &DCI,
936 const MipsSubtarget &Subtarget) {
937 EVT Ty = N->getValueType(0);
939 if (Subtarget.hasMSA()) {
940 SDValue Op0 = N->getOperand(0);
941 SDValue Op1 = N->getOperand(1);
943 // (sra (shl (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d), imm:$d)
944 // where $d + sizeof($c) == 32
945 // or $d + sizeof($c) <= 32 and SExt
946 // -> (MipsVExtractSExt $a, $b, $c)
947 if (Op0->getOpcode() == ISD::SHL && Op1 == Op0->getOperand(1)) {
948 SDValue Op0Op0 = Op0->getOperand(0);
949 ConstantSDNode *ShAmount = dyn_cast<ConstantSDNode>(Op1);
954 if (Op0Op0->getOpcode() != MipsISD::VEXTRACT_SEXT_ELT &&
955 Op0Op0->getOpcode() != MipsISD::VEXTRACT_ZEXT_ELT)
958 EVT ExtendTy = cast<VTSDNode>(Op0Op0->getOperand(2))->getVT();
959 unsigned TotalBits = ShAmount->getZExtValue() + ExtendTy.getSizeInBits();
961 if (TotalBits == 32 ||
962 (Op0Op0->getOpcode() == MipsISD::VEXTRACT_SEXT_ELT &&
964 SDValue Ops[] = { Op0Op0->getOperand(0), Op0Op0->getOperand(1),
965 Op0Op0->getOperand(2) };
966 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, SDLoc(Op0Op0),
968 makeArrayRef(Ops, Op0Op0->getNumOperands()));
973 if ((Ty != MVT::v2i16) && ((Ty != MVT::v4i8) || !Subtarget.hasDSPR2()))
976 return performDSPShiftCombine(MipsISD::SHRA_DSP, N, Ty, DAG, Subtarget);
980 static SDValue performSRLCombine(SDNode *N, SelectionDAG &DAG,
981 TargetLowering::DAGCombinerInfo &DCI,
982 const MipsSubtarget &Subtarget) {
983 EVT Ty = N->getValueType(0);
985 if (((Ty != MVT::v2i16) || !Subtarget.hasDSPR2()) && (Ty != MVT::v4i8))
988 return performDSPShiftCombine(MipsISD::SHRL_DSP, N, Ty, DAG, Subtarget);
991 static bool isLegalDSPCondCode(EVT Ty, ISD::CondCode CC) {
992 bool IsV216 = (Ty == MVT::v2i16);
996 case ISD::SETNE: return true;
1000 case ISD::SETGE: return IsV216;
1004 case ISD::SETUGE: return !IsV216;
1005 default: return false;
1009 static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG) {
1010 EVT Ty = N->getValueType(0);
1012 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
1015 if (!isLegalDSPCondCode(Ty, cast<CondCodeSDNode>(N->getOperand(2))->get()))
1018 return DAG.getNode(MipsISD::SETCC_DSP, SDLoc(N), Ty, N->getOperand(0),
1019 N->getOperand(1), N->getOperand(2));
1022 static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG) {
1023 EVT Ty = N->getValueType(0);
1025 if (Ty.is128BitVector() && Ty.isInteger()) {
1026 // Try the following combines:
1027 // (vselect (setcc $a, $b, SETLT), $b, $a)) -> (vsmax $a, $b)
1028 // (vselect (setcc $a, $b, SETLE), $b, $a)) -> (vsmax $a, $b)
1029 // (vselect (setcc $a, $b, SETLT), $a, $b)) -> (vsmin $a, $b)
1030 // (vselect (setcc $a, $b, SETLE), $a, $b)) -> (vsmin $a, $b)
1031 // (vselect (setcc $a, $b, SETULT), $b, $a)) -> (vumax $a, $b)
1032 // (vselect (setcc $a, $b, SETULE), $b, $a)) -> (vumax $a, $b)
1033 // (vselect (setcc $a, $b, SETULT), $a, $b)) -> (vumin $a, $b)
1034 // (vselect (setcc $a, $b, SETULE), $a, $b)) -> (vumin $a, $b)
1035 // SETGT/SETGE/SETUGT/SETUGE variants of these will show up initially but
1036 // will be expanded to equivalent SETLT/SETLE/SETULT/SETULE versions by the
1038 SDValue Op0 = N->getOperand(0);
1040 if (Op0->getOpcode() != ISD::SETCC)
1043 ISD::CondCode CondCode = cast<CondCodeSDNode>(Op0->getOperand(2))->get();
1046 if (CondCode == ISD::SETLT || CondCode == ISD::SETLE)
1048 else if (CondCode == ISD::SETULT || CondCode == ISD::SETULE)
1053 SDValue Op1 = N->getOperand(1);
1054 SDValue Op2 = N->getOperand(2);
1055 SDValue Op0Op0 = Op0->getOperand(0);
1056 SDValue Op0Op1 = Op0->getOperand(1);
1058 if (Op1 == Op0Op0 && Op2 == Op0Op1)
1059 return DAG.getNode(Signed ? MipsISD::VSMIN : MipsISD::VUMIN, SDLoc(N),
1061 else if (Op1 == Op0Op1 && Op2 == Op0Op0)
1062 return DAG.getNode(Signed ? MipsISD::VSMAX : MipsISD::VUMAX, SDLoc(N),
1064 } else if ((Ty == MVT::v2i16) || (Ty == MVT::v4i8)) {
1065 SDValue SetCC = N->getOperand(0);
1067 if (SetCC.getOpcode() != MipsISD::SETCC_DSP)
1070 return DAG.getNode(MipsISD::SELECT_CC_DSP, SDLoc(N), Ty,
1071 SetCC.getOperand(0), SetCC.getOperand(1),
1072 N->getOperand(1), N->getOperand(2), SetCC.getOperand(2));
1078 static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
1079 const MipsSubtarget &Subtarget) {
1080 EVT Ty = N->getValueType(0);
1082 if (Subtarget.hasMSA() && Ty.is128BitVector() && Ty.isInteger()) {
1083 // Try the following combines:
1084 // (xor (or $a, $b), (build_vector allones))
1085 // (xor (or $a, $b), (bitcast (build_vector allones)))
1086 SDValue Op0 = N->getOperand(0);
1087 SDValue Op1 = N->getOperand(1);
1090 if (ISD::isBuildVectorAllOnes(Op0.getNode()))
1092 else if (ISD::isBuildVectorAllOnes(Op1.getNode()))
1097 if (NotOp->getOpcode() == ISD::OR)
1098 return DAG.getNode(MipsISD::VNOR, SDLoc(N), Ty, NotOp->getOperand(0),
1099 NotOp->getOperand(1));
1106 MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1107 SelectionDAG &DAG = DCI.DAG;
1110 switch (N->getOpcode()) {
1112 return performADDECombine(N, DAG, DCI, Subtarget);
1114 Val = performANDCombine(N, DAG, DCI, Subtarget);
1117 Val = performORCombine(N, DAG, DCI, Subtarget);
1120 return performSUBECombine(N, DAG, DCI, Subtarget);
1122 return performMULCombine(N, DAG, DCI, this);
1124 return performSHLCombine(N, DAG, DCI, Subtarget);
1126 return performSRACombine(N, DAG, DCI, Subtarget);
1128 return performSRLCombine(N, DAG, DCI, Subtarget);
1130 return performVSELECTCombine(N, DAG);
1132 Val = performXORCombine(N, DAG, Subtarget);
1135 Val = performSETCCCombine(N, DAG);
1139 if (Val.getNode()) {
1140 DEBUG(dbgs() << "\nMipsSE DAG Combine:\n";
1141 N->printrWithDepth(dbgs(), &DAG);
1142 dbgs() << "\n=> \n";
1143 Val.getNode()->printrWithDepth(dbgs(), &DAG);
1148 return MipsTargetLowering::PerformDAGCombine(N, DCI);
1152 MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
1153 MachineBasicBlock *BB) const {
1154 switch (MI.getOpcode()) {
1156 return MipsTargetLowering::EmitInstrWithCustomInserter(MI, BB);
1157 case Mips::BPOSGE32_PSEUDO:
1158 return emitBPOSGE32(MI, BB);
1159 case Mips::SNZ_B_PSEUDO:
1160 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_B);
1161 case Mips::SNZ_H_PSEUDO:
1162 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_H);
1163 case Mips::SNZ_W_PSEUDO:
1164 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_W);
1165 case Mips::SNZ_D_PSEUDO:
1166 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_D);
1167 case Mips::SNZ_V_PSEUDO:
1168 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_V);
1169 case Mips::SZ_B_PSEUDO:
1170 return emitMSACBranchPseudo(MI, BB, Mips::BZ_B);
1171 case Mips::SZ_H_PSEUDO:
1172 return emitMSACBranchPseudo(MI, BB, Mips::BZ_H);
1173 case Mips::SZ_W_PSEUDO:
1174 return emitMSACBranchPseudo(MI, BB, Mips::BZ_W);
1175 case Mips::SZ_D_PSEUDO:
1176 return emitMSACBranchPseudo(MI, BB, Mips::BZ_D);
1177 case Mips::SZ_V_PSEUDO:
1178 return emitMSACBranchPseudo(MI, BB, Mips::BZ_V);
1179 case Mips::COPY_FW_PSEUDO:
1180 return emitCOPY_FW(MI, BB);
1181 case Mips::COPY_FD_PSEUDO:
1182 return emitCOPY_FD(MI, BB);
1183 case Mips::INSERT_FW_PSEUDO:
1184 return emitINSERT_FW(MI, BB);
1185 case Mips::INSERT_FD_PSEUDO:
1186 return emitINSERT_FD(MI, BB);
1187 case Mips::INSERT_B_VIDX_PSEUDO:
1188 case Mips::INSERT_B_VIDX64_PSEUDO:
1189 return emitINSERT_DF_VIDX(MI, BB, 1, false);
1190 case Mips::INSERT_H_VIDX_PSEUDO:
1191 case Mips::INSERT_H_VIDX64_PSEUDO:
1192 return emitINSERT_DF_VIDX(MI, BB, 2, false);
1193 case Mips::INSERT_W_VIDX_PSEUDO:
1194 case Mips::INSERT_W_VIDX64_PSEUDO:
1195 return emitINSERT_DF_VIDX(MI, BB, 4, false);
1196 case Mips::INSERT_D_VIDX_PSEUDO:
1197 case Mips::INSERT_D_VIDX64_PSEUDO:
1198 return emitINSERT_DF_VIDX(MI, BB, 8, false);
1199 case Mips::INSERT_FW_VIDX_PSEUDO:
1200 case Mips::INSERT_FW_VIDX64_PSEUDO:
1201 return emitINSERT_DF_VIDX(MI, BB, 4, true);
1202 case Mips::INSERT_FD_VIDX_PSEUDO:
1203 case Mips::INSERT_FD_VIDX64_PSEUDO:
1204 return emitINSERT_DF_VIDX(MI, BB, 8, true);
1205 case Mips::FILL_FW_PSEUDO:
1206 return emitFILL_FW(MI, BB);
1207 case Mips::FILL_FD_PSEUDO:
1208 return emitFILL_FD(MI, BB);
1209 case Mips::FEXP2_W_1_PSEUDO:
1210 return emitFEXP2_W_1(MI, BB);
1211 case Mips::FEXP2_D_1_PSEUDO:
1212 return emitFEXP2_D_1(MI, BB);
1214 return emitST_F16_PSEUDO(MI, BB);
1216 return emitLD_F16_PSEUDO(MI, BB);
1217 case Mips::MSA_FP_EXTEND_W_PSEUDO:
1218 return emitFPEXTEND_PSEUDO(MI, BB, false);
1219 case Mips::MSA_FP_ROUND_W_PSEUDO:
1220 return emitFPROUND_PSEUDO(MI, BB, false);
1221 case Mips::MSA_FP_EXTEND_D_PSEUDO:
1222 return emitFPEXTEND_PSEUDO(MI, BB, true);
1223 case Mips::MSA_FP_ROUND_D_PSEUDO:
1224 return emitFPROUND_PSEUDO(MI, BB, true);
1228 bool MipsSETargetLowering::isEligibleForTailCallOptimization(
1229 const CCState &CCInfo, unsigned NextStackOffset,
1230 const MipsFunctionInfo &FI) const {
1231 if (!UseMipsTailCalls)
1234 // Exception has to be cleared with eret.
1238 // Return false if either the callee or caller has a byval argument.
1239 if (CCInfo.getInRegsParamsCount() > 0 || FI.hasByvalArg())
1242 // Return true if the callee's argument area is no larger than the
1244 return NextStackOffset <= FI.getIncomingArgSize();
1247 void MipsSETargetLowering::
1248 getOpndList(SmallVectorImpl<SDValue> &Ops,
1249 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
1250 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
1251 bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee,
1252 SDValue Chain) const {
1253 Ops.push_back(Callee);
1254 MipsTargetLowering::getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal,
1255 InternalLinkage, IsCallReloc, CLI, Callee,
1259 SDValue MipsSETargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1260 LoadSDNode &Nd = *cast<LoadSDNode>(Op);
1262 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
1263 return MipsTargetLowering::lowerLOAD(Op, DAG);
1265 // Replace a double precision load with two i32 loads and a buildpair64.
1267 SDValue Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
1268 EVT PtrVT = Ptr.getValueType();
1270 // i32 load from lower address.
1271 SDValue Lo = DAG.getLoad(MVT::i32, DL, Chain, Ptr, MachinePointerInfo(),
1272 Nd.getAlignment(), Nd.getMemOperand()->getFlags());
1274 // i32 load from higher address.
1275 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, DL, PtrVT));
1276 SDValue Hi = DAG.getLoad(
1277 MVT::i32, DL, Lo.getValue(1), Ptr, MachinePointerInfo(),
1278 std::min(Nd.getAlignment(), 4U), Nd.getMemOperand()->getFlags());
1280 if (!Subtarget.isLittle())
1283 SDValue BP = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
1284 SDValue Ops[2] = {BP, Hi.getValue(1)};
1285 return DAG.getMergeValues(Ops, DL);
1288 SDValue MipsSETargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1289 StoreSDNode &Nd = *cast<StoreSDNode>(Op);
1291 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
1292 return MipsTargetLowering::lowerSTORE(Op, DAG);
1294 // Replace a double precision store with two extractelement64s and i32 stores.
1296 SDValue Val = Nd.getValue(), Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
1297 EVT PtrVT = Ptr.getValueType();
1298 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1299 Val, DAG.getConstant(0, DL, MVT::i32));
1300 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1301 Val, DAG.getConstant(1, DL, MVT::i32));
1303 if (!Subtarget.isLittle())
1306 // i32 store to lower address.
1308 DAG.getStore(Chain, DL, Lo, Ptr, MachinePointerInfo(), Nd.getAlignment(),
1309 Nd.getMemOperand()->getFlags(), Nd.getAAInfo());
1311 // i32 store to higher address.
1312 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, DL, PtrVT));
1313 return DAG.getStore(Chain, DL, Hi, Ptr, MachinePointerInfo(),
1314 std::min(Nd.getAlignment(), 4U),
1315 Nd.getMemOperand()->getFlags(), Nd.getAAInfo());
1318 SDValue MipsSETargetLowering::lowerMulDiv(SDValue Op, unsigned NewOpc,
1319 bool HasLo, bool HasHi,
1320 SelectionDAG &DAG) const {
1321 // MIPS32r6/MIPS64r6 removed accumulator based multiplies.
1322 assert(!Subtarget.hasMips32r6());
1324 EVT Ty = Op.getOperand(0).getValueType();
1326 SDValue Mult = DAG.getNode(NewOpc, DL, MVT::Untyped,
1327 Op.getOperand(0), Op.getOperand(1));
1331 Lo = DAG.getNode(MipsISD::MFLO, DL, Ty, Mult);
1333 Hi = DAG.getNode(MipsISD::MFHI, DL, Ty, Mult);
1335 if (!HasLo || !HasHi)
1336 return HasLo ? Lo : Hi;
1338 SDValue Vals[] = { Lo, Hi };
1339 return DAG.getMergeValues(Vals, DL);
1342 static SDValue initAccumulator(SDValue In, const SDLoc &DL, SelectionDAG &DAG) {
1343 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
1344 DAG.getConstant(0, DL, MVT::i32));
1345 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
1346 DAG.getConstant(1, DL, MVT::i32));
1347 return DAG.getNode(MipsISD::MTLOHI, DL, MVT::Untyped, InLo, InHi);
1350 static SDValue extractLOHI(SDValue Op, const SDLoc &DL, SelectionDAG &DAG) {
1351 SDValue Lo = DAG.getNode(MipsISD::MFLO, DL, MVT::i32, Op);
1352 SDValue Hi = DAG.getNode(MipsISD::MFHI, DL, MVT::i32, Op);
1353 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi);
1356 // This function expands mips intrinsic nodes which have 64-bit input operands
1357 // or output values.
1359 // out64 = intrinsic-node in64
1361 // lo = copy (extract-element (in64, 0))
1362 // hi = copy (extract-element (in64, 1))
1363 // mips-specific-node
1366 // out64 = merge-values (v0, v1)
1368 static SDValue lowerDSPIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1370 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
1371 SmallVector<SDValue, 3> Ops;
1374 // See if Op has a chain input.
1376 Ops.push_back(Op->getOperand(OpNo++));
1378 // The next operand is the intrinsic opcode.
1379 assert(Op->getOperand(OpNo).getOpcode() == ISD::TargetConstant);
1381 // See if the next operand has type i64.
1382 SDValue Opnd = Op->getOperand(++OpNo), In64;
1384 if (Opnd.getValueType() == MVT::i64)
1385 In64 = initAccumulator(Opnd, DL, DAG);
1387 Ops.push_back(Opnd);
1389 // Push the remaining operands.
1390 for (++OpNo ; OpNo < Op->getNumOperands(); ++OpNo)
1391 Ops.push_back(Op->getOperand(OpNo));
1393 // Add In64 to the end of the list.
1395 Ops.push_back(In64);
1398 SmallVector<EVT, 2> ResTys;
1400 for (SDNode::value_iterator I = Op->value_begin(), E = Op->value_end();
1402 ResTys.push_back((*I == MVT::i64) ? MVT::Untyped : *I);
1405 SDValue Val = DAG.getNode(Opc, DL, ResTys, Ops);
1406 SDValue Out = (ResTys[0] == MVT::Untyped) ? extractLOHI(Val, DL, DAG) : Val;
1411 assert(Val->getValueType(1) == MVT::Other);
1412 SDValue Vals[] = { Out, SDValue(Val.getNode(), 1) };
1413 return DAG.getMergeValues(Vals, DL);
1416 // Lower an MSA copy intrinsic into the specified SelectionDAG node
1417 static SDValue lowerMSACopyIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1419 SDValue Vec = Op->getOperand(1);
1420 SDValue Idx = Op->getOperand(2);
1421 EVT ResTy = Op->getValueType(0);
1422 EVT EltTy = Vec->getValueType(0).getVectorElementType();
1424 SDValue Result = DAG.getNode(Opc, DL, ResTy, Vec, Idx,
1425 DAG.getValueType(EltTy));
1430 static SDValue lowerMSASplatZExt(SDValue Op, unsigned OpNr, SelectionDAG &DAG) {
1431 EVT ResVecTy = Op->getValueType(0);
1432 EVT ViaVecTy = ResVecTy;
1435 // When ResVecTy == MVT::v2i64, LaneA is the upper 32 bits of the lane and
1436 // LaneB is the lower 32-bits. Otherwise LaneA and LaneB are alternating
1439 SDValue LaneB = Op->getOperand(2);
1441 if (ResVecTy == MVT::v2i64) {
1442 LaneA = DAG.getConstant(0, DL, MVT::i32);
1443 ViaVecTy = MVT::v4i32;
1447 SDValue Ops[16] = { LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB,
1448 LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB };
1450 SDValue Result = DAG.getBuildVector(
1451 ViaVecTy, DL, makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
1453 if (ViaVecTy != ResVecTy)
1454 Result = DAG.getNode(ISD::BITCAST, DL, ResVecTy, Result);
1459 static SDValue lowerMSASplatImm(SDValue Op, unsigned ImmOp, SelectionDAG &DAG) {
1460 return DAG.getConstant(Op->getConstantOperandVal(ImmOp), SDLoc(Op),
1461 Op->getValueType(0));
1464 static SDValue getBuildVectorSplat(EVT VecTy, SDValue SplatValue,
1465 bool BigEndian, SelectionDAG &DAG) {
1466 EVT ViaVecTy = VecTy;
1467 SDValue SplatValueA = SplatValue;
1468 SDValue SplatValueB = SplatValue;
1469 SDLoc DL(SplatValue);
1471 if (VecTy == MVT::v2i64) {
1472 // v2i64 BUILD_VECTOR must be performed via v4i32 so split into i32's.
1473 ViaVecTy = MVT::v4i32;
1475 SplatValueA = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, SplatValue);
1476 SplatValueB = DAG.getNode(ISD::SRL, DL, MVT::i64, SplatValue,
1477 DAG.getConstant(32, DL, MVT::i32));
1478 SplatValueB = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, SplatValueB);
1481 // We currently hold the parts in little endian order. Swap them if
1484 std::swap(SplatValueA, SplatValueB);
1486 SDValue Ops[16] = { SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1487 SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1488 SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1489 SplatValueA, SplatValueB, SplatValueA, SplatValueB };
1491 SDValue Result = DAG.getBuildVector(
1492 ViaVecTy, DL, makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
1494 if (VecTy != ViaVecTy)
1495 Result = DAG.getNode(ISD::BITCAST, DL, VecTy, Result);
1500 static SDValue lowerMSABinaryBitImmIntr(SDValue Op, SelectionDAG &DAG,
1501 unsigned Opc, SDValue Imm,
1503 EVT VecTy = Op->getValueType(0);
1507 // The DAG Combiner can't constant fold bitcasted vectors yet so we must do it
1509 if (VecTy == MVT::v2i64) {
1510 if (ConstantSDNode *CImm = dyn_cast<ConstantSDNode>(Imm)) {
1511 APInt BitImm = APInt(64, 1) << CImm->getAPIntValue();
1513 SDValue BitImmHiOp = DAG.getConstant(BitImm.lshr(32).trunc(32), DL,
1515 SDValue BitImmLoOp = DAG.getConstant(BitImm.trunc(32), DL, MVT::i32);
1518 std::swap(BitImmLoOp, BitImmHiOp);
1520 Exp2Imm = DAG.getNode(
1521 ISD::BITCAST, DL, MVT::v2i64,
1522 DAG.getBuildVector(MVT::v4i32, DL,
1523 {BitImmLoOp, BitImmHiOp, BitImmLoOp, BitImmHiOp}));
1527 if (!Exp2Imm.getNode()) {
1528 // We couldnt constant fold, do a vector shift instead
1530 // Extend i32 to i64 if necessary. Sign or zero extend doesn't matter since
1531 // only values 0-63 are valid.
1532 if (VecTy == MVT::v2i64)
1533 Imm = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Imm);
1535 Exp2Imm = getBuildVectorSplat(VecTy, Imm, BigEndian, DAG);
1537 Exp2Imm = DAG.getNode(ISD::SHL, DL, VecTy, DAG.getConstant(1, DL, VecTy),
1541 return DAG.getNode(Opc, DL, VecTy, Op->getOperand(1), Exp2Imm);
1544 static SDValue lowerMSABitClear(SDValue Op, SelectionDAG &DAG) {
1545 EVT ResTy = Op->getValueType(0);
1547 SDValue One = DAG.getConstant(1, DL, ResTy);
1548 SDValue Bit = DAG.getNode(ISD::SHL, DL, ResTy, One, Op->getOperand(2));
1550 return DAG.getNode(ISD::AND, DL, ResTy, Op->getOperand(1),
1551 DAG.getNOT(DL, Bit, ResTy));
1554 static SDValue lowerMSABitClearImm(SDValue Op, SelectionDAG &DAG) {
1556 EVT ResTy = Op->getValueType(0);
1557 APInt BitImm = APInt(ResTy.getScalarSizeInBits(), 1)
1558 << cast<ConstantSDNode>(Op->getOperand(2))->getAPIntValue();
1559 SDValue BitMask = DAG.getConstant(~BitImm, DL, ResTy);
1561 return DAG.getNode(ISD::AND, DL, ResTy, Op->getOperand(1), BitMask);
1564 SDValue MipsSETargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
1565 SelectionDAG &DAG) const {
1568 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) {
1571 case Intrinsic::mips_shilo:
1572 return lowerDSPIntr(Op, DAG, MipsISD::SHILO);
1573 case Intrinsic::mips_dpau_h_qbl:
1574 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL);
1575 case Intrinsic::mips_dpau_h_qbr:
1576 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR);
1577 case Intrinsic::mips_dpsu_h_qbl:
1578 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL);
1579 case Intrinsic::mips_dpsu_h_qbr:
1580 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR);
1581 case Intrinsic::mips_dpa_w_ph:
1582 return lowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH);
1583 case Intrinsic::mips_dps_w_ph:
1584 return lowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH);
1585 case Intrinsic::mips_dpax_w_ph:
1586 return lowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH);
1587 case Intrinsic::mips_dpsx_w_ph:
1588 return lowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH);
1589 case Intrinsic::mips_mulsa_w_ph:
1590 return lowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH);
1591 case Intrinsic::mips_mult:
1592 return lowerDSPIntr(Op, DAG, MipsISD::Mult);
1593 case Intrinsic::mips_multu:
1594 return lowerDSPIntr(Op, DAG, MipsISD::Multu);
1595 case Intrinsic::mips_madd:
1596 return lowerDSPIntr(Op, DAG, MipsISD::MAdd);
1597 case Intrinsic::mips_maddu:
1598 return lowerDSPIntr(Op, DAG, MipsISD::MAddu);
1599 case Intrinsic::mips_msub:
1600 return lowerDSPIntr(Op, DAG, MipsISD::MSub);
1601 case Intrinsic::mips_msubu:
1602 return lowerDSPIntr(Op, DAG, MipsISD::MSubu);
1603 case Intrinsic::mips_addv_b:
1604 case Intrinsic::mips_addv_h:
1605 case Intrinsic::mips_addv_w:
1606 case Intrinsic::mips_addv_d:
1607 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1),
1609 case Intrinsic::mips_addvi_b:
1610 case Intrinsic::mips_addvi_h:
1611 case Intrinsic::mips_addvi_w:
1612 case Intrinsic::mips_addvi_d:
1613 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1),
1614 lowerMSASplatImm(Op, 2, DAG));
1615 case Intrinsic::mips_and_v:
1616 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
1618 case Intrinsic::mips_andi_b:
1619 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
1620 lowerMSASplatImm(Op, 2, DAG));
1621 case Intrinsic::mips_bclr_b:
1622 case Intrinsic::mips_bclr_h:
1623 case Intrinsic::mips_bclr_w:
1624 case Intrinsic::mips_bclr_d:
1625 return lowerMSABitClear(Op, DAG);
1626 case Intrinsic::mips_bclri_b:
1627 case Intrinsic::mips_bclri_h:
1628 case Intrinsic::mips_bclri_w:
1629 case Intrinsic::mips_bclri_d:
1630 return lowerMSABitClearImm(Op, DAG);
1631 case Intrinsic::mips_binsli_b:
1632 case Intrinsic::mips_binsli_h:
1633 case Intrinsic::mips_binsli_w:
1634 case Intrinsic::mips_binsli_d: {
1635 // binsli_x(IfClear, IfSet, nbits) -> (vselect LBitsMask, IfSet, IfClear)
1636 EVT VecTy = Op->getValueType(0);
1637 EVT EltTy = VecTy.getVectorElementType();
1638 APInt Mask = APInt::getHighBitsSet(EltTy.getSizeInBits(),
1639 Op->getConstantOperandVal(3));
1640 return DAG.getNode(ISD::VSELECT, DL, VecTy,
1641 DAG.getConstant(Mask, DL, VecTy, true),
1642 Op->getOperand(2), Op->getOperand(1));
1644 case Intrinsic::mips_binsri_b:
1645 case Intrinsic::mips_binsri_h:
1646 case Intrinsic::mips_binsri_w:
1647 case Intrinsic::mips_binsri_d: {
1648 // binsri_x(IfClear, IfSet, nbits) -> (vselect RBitsMask, IfSet, IfClear)
1649 EVT VecTy = Op->getValueType(0);
1650 EVT EltTy = VecTy.getVectorElementType();
1651 APInt Mask = APInt::getLowBitsSet(EltTy.getSizeInBits(),
1652 Op->getConstantOperandVal(3));
1653 return DAG.getNode(ISD::VSELECT, DL, VecTy,
1654 DAG.getConstant(Mask, DL, VecTy, true),
1655 Op->getOperand(2), Op->getOperand(1));
1657 case Intrinsic::mips_bmnz_v:
1658 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), Op->getOperand(3),
1659 Op->getOperand(2), Op->getOperand(1));
1660 case Intrinsic::mips_bmnzi_b:
1661 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1662 lowerMSASplatImm(Op, 3, DAG), Op->getOperand(2),
1664 case Intrinsic::mips_bmz_v:
1665 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), Op->getOperand(3),
1666 Op->getOperand(1), Op->getOperand(2));
1667 case Intrinsic::mips_bmzi_b:
1668 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1669 lowerMSASplatImm(Op, 3, DAG), Op->getOperand(1),
1671 case Intrinsic::mips_bneg_b:
1672 case Intrinsic::mips_bneg_h:
1673 case Intrinsic::mips_bneg_w:
1674 case Intrinsic::mips_bneg_d: {
1675 EVT VecTy = Op->getValueType(0);
1676 SDValue One = DAG.getConstant(1, DL, VecTy);
1678 return DAG.getNode(ISD::XOR, DL, VecTy, Op->getOperand(1),
1679 DAG.getNode(ISD::SHL, DL, VecTy, One,
1680 Op->getOperand(2)));
1682 case Intrinsic::mips_bnegi_b:
1683 case Intrinsic::mips_bnegi_h:
1684 case Intrinsic::mips_bnegi_w:
1685 case Intrinsic::mips_bnegi_d:
1686 return lowerMSABinaryBitImmIntr(Op, DAG, ISD::XOR, Op->getOperand(2),
1687 !Subtarget.isLittle());
1688 case Intrinsic::mips_bnz_b:
1689 case Intrinsic::mips_bnz_h:
1690 case Intrinsic::mips_bnz_w:
1691 case Intrinsic::mips_bnz_d:
1692 return DAG.getNode(MipsISD::VALL_NONZERO, DL, Op->getValueType(0),
1694 case Intrinsic::mips_bnz_v:
1695 return DAG.getNode(MipsISD::VANY_NONZERO, DL, Op->getValueType(0),
1697 case Intrinsic::mips_bsel_v:
1698 // bsel_v(Mask, IfClear, IfSet) -> (vselect Mask, IfSet, IfClear)
1699 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1700 Op->getOperand(1), Op->getOperand(3),
1702 case Intrinsic::mips_bseli_b:
1703 // bseli_v(Mask, IfClear, IfSet) -> (vselect Mask, IfSet, IfClear)
1704 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1705 Op->getOperand(1), lowerMSASplatImm(Op, 3, DAG),
1707 case Intrinsic::mips_bset_b:
1708 case Intrinsic::mips_bset_h:
1709 case Intrinsic::mips_bset_w:
1710 case Intrinsic::mips_bset_d: {
1711 EVT VecTy = Op->getValueType(0);
1712 SDValue One = DAG.getConstant(1, DL, VecTy);
1714 return DAG.getNode(ISD::OR, DL, VecTy, Op->getOperand(1),
1715 DAG.getNode(ISD::SHL, DL, VecTy, One,
1716 Op->getOperand(2)));
1718 case Intrinsic::mips_bseti_b:
1719 case Intrinsic::mips_bseti_h:
1720 case Intrinsic::mips_bseti_w:
1721 case Intrinsic::mips_bseti_d:
1722 return lowerMSABinaryBitImmIntr(Op, DAG, ISD::OR, Op->getOperand(2),
1723 !Subtarget.isLittle());
1724 case Intrinsic::mips_bz_b:
1725 case Intrinsic::mips_bz_h:
1726 case Intrinsic::mips_bz_w:
1727 case Intrinsic::mips_bz_d:
1728 return DAG.getNode(MipsISD::VALL_ZERO, DL, Op->getValueType(0),
1730 case Intrinsic::mips_bz_v:
1731 return DAG.getNode(MipsISD::VANY_ZERO, DL, Op->getValueType(0),
1733 case Intrinsic::mips_ceq_b:
1734 case Intrinsic::mips_ceq_h:
1735 case Intrinsic::mips_ceq_w:
1736 case Intrinsic::mips_ceq_d:
1737 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1738 Op->getOperand(2), ISD::SETEQ);
1739 case Intrinsic::mips_ceqi_b:
1740 case Intrinsic::mips_ceqi_h:
1741 case Intrinsic::mips_ceqi_w:
1742 case Intrinsic::mips_ceqi_d:
1743 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1744 lowerMSASplatImm(Op, 2, DAG), ISD::SETEQ);
1745 case Intrinsic::mips_cle_s_b:
1746 case Intrinsic::mips_cle_s_h:
1747 case Intrinsic::mips_cle_s_w:
1748 case Intrinsic::mips_cle_s_d:
1749 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1750 Op->getOperand(2), ISD::SETLE);
1751 case Intrinsic::mips_clei_s_b:
1752 case Intrinsic::mips_clei_s_h:
1753 case Intrinsic::mips_clei_s_w:
1754 case Intrinsic::mips_clei_s_d:
1755 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1756 lowerMSASplatImm(Op, 2, DAG), ISD::SETLE);
1757 case Intrinsic::mips_cle_u_b:
1758 case Intrinsic::mips_cle_u_h:
1759 case Intrinsic::mips_cle_u_w:
1760 case Intrinsic::mips_cle_u_d:
1761 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1762 Op->getOperand(2), ISD::SETULE);
1763 case Intrinsic::mips_clei_u_b:
1764 case Intrinsic::mips_clei_u_h:
1765 case Intrinsic::mips_clei_u_w:
1766 case Intrinsic::mips_clei_u_d:
1767 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1768 lowerMSASplatImm(Op, 2, DAG), ISD::SETULE);
1769 case Intrinsic::mips_clt_s_b:
1770 case Intrinsic::mips_clt_s_h:
1771 case Intrinsic::mips_clt_s_w:
1772 case Intrinsic::mips_clt_s_d:
1773 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1774 Op->getOperand(2), ISD::SETLT);
1775 case Intrinsic::mips_clti_s_b:
1776 case Intrinsic::mips_clti_s_h:
1777 case Intrinsic::mips_clti_s_w:
1778 case Intrinsic::mips_clti_s_d:
1779 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1780 lowerMSASplatImm(Op, 2, DAG), ISD::SETLT);
1781 case Intrinsic::mips_clt_u_b:
1782 case Intrinsic::mips_clt_u_h:
1783 case Intrinsic::mips_clt_u_w:
1784 case Intrinsic::mips_clt_u_d:
1785 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1786 Op->getOperand(2), ISD::SETULT);
1787 case Intrinsic::mips_clti_u_b:
1788 case Intrinsic::mips_clti_u_h:
1789 case Intrinsic::mips_clti_u_w:
1790 case Intrinsic::mips_clti_u_d:
1791 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1792 lowerMSASplatImm(Op, 2, DAG), ISD::SETULT);
1793 case Intrinsic::mips_copy_s_b:
1794 case Intrinsic::mips_copy_s_h:
1795 case Intrinsic::mips_copy_s_w:
1796 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
1797 case Intrinsic::mips_copy_s_d:
1798 if (Subtarget.hasMips64())
1799 // Lower directly into VEXTRACT_SEXT_ELT since i64 is legal on Mips64.
1800 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
1802 // Lower into the generic EXTRACT_VECTOR_ELT node and let the type
1803 // legalizer and EXTRACT_VECTOR_ELT lowering sort it out.
1804 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
1805 Op->getValueType(0), Op->getOperand(1),
1808 case Intrinsic::mips_copy_u_b:
1809 case Intrinsic::mips_copy_u_h:
1810 case Intrinsic::mips_copy_u_w:
1811 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
1812 case Intrinsic::mips_copy_u_d:
1813 if (Subtarget.hasMips64())
1814 // Lower directly into VEXTRACT_ZEXT_ELT since i64 is legal on Mips64.
1815 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
1817 // Lower into the generic EXTRACT_VECTOR_ELT node and let the type
1818 // legalizer and EXTRACT_VECTOR_ELT lowering sort it out.
1819 // Note: When i64 is illegal, this results in copy_s.w instructions
1820 // instead of copy_u.w instructions. This makes no difference to the
1821 // behaviour since i64 is only illegal when the register file is 32-bit.
1822 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
1823 Op->getValueType(0), Op->getOperand(1),
1826 case Intrinsic::mips_div_s_b:
1827 case Intrinsic::mips_div_s_h:
1828 case Intrinsic::mips_div_s_w:
1829 case Intrinsic::mips_div_s_d:
1830 return DAG.getNode(ISD::SDIV, DL, Op->getValueType(0), Op->getOperand(1),
1832 case Intrinsic::mips_div_u_b:
1833 case Intrinsic::mips_div_u_h:
1834 case Intrinsic::mips_div_u_w:
1835 case Intrinsic::mips_div_u_d:
1836 return DAG.getNode(ISD::UDIV, DL, Op->getValueType(0), Op->getOperand(1),
1838 case Intrinsic::mips_fadd_w:
1839 case Intrinsic::mips_fadd_d: {
1840 // TODO: If intrinsics have fast-math-flags, propagate them.
1841 return DAG.getNode(ISD::FADD, DL, Op->getValueType(0), Op->getOperand(1),
1844 // Don't lower mips_fcaf_[wd] since LLVM folds SETFALSE condcodes away
1845 case Intrinsic::mips_fceq_w:
1846 case Intrinsic::mips_fceq_d:
1847 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1848 Op->getOperand(2), ISD::SETOEQ);
1849 case Intrinsic::mips_fcle_w:
1850 case Intrinsic::mips_fcle_d:
1851 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1852 Op->getOperand(2), ISD::SETOLE);
1853 case Intrinsic::mips_fclt_w:
1854 case Intrinsic::mips_fclt_d:
1855 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1856 Op->getOperand(2), ISD::SETOLT);
1857 case Intrinsic::mips_fcne_w:
1858 case Intrinsic::mips_fcne_d:
1859 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1860 Op->getOperand(2), ISD::SETONE);
1861 case Intrinsic::mips_fcor_w:
1862 case Intrinsic::mips_fcor_d:
1863 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1864 Op->getOperand(2), ISD::SETO);
1865 case Intrinsic::mips_fcueq_w:
1866 case Intrinsic::mips_fcueq_d:
1867 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1868 Op->getOperand(2), ISD::SETUEQ);
1869 case Intrinsic::mips_fcule_w:
1870 case Intrinsic::mips_fcule_d:
1871 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1872 Op->getOperand(2), ISD::SETULE);
1873 case Intrinsic::mips_fcult_w:
1874 case Intrinsic::mips_fcult_d:
1875 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1876 Op->getOperand(2), ISD::SETULT);
1877 case Intrinsic::mips_fcun_w:
1878 case Intrinsic::mips_fcun_d:
1879 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1880 Op->getOperand(2), ISD::SETUO);
1881 case Intrinsic::mips_fcune_w:
1882 case Intrinsic::mips_fcune_d:
1883 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1884 Op->getOperand(2), ISD::SETUNE);
1885 case Intrinsic::mips_fdiv_w:
1886 case Intrinsic::mips_fdiv_d: {
1887 // TODO: If intrinsics have fast-math-flags, propagate them.
1888 return DAG.getNode(ISD::FDIV, DL, Op->getValueType(0), Op->getOperand(1),
1891 case Intrinsic::mips_ffint_u_w:
1892 case Intrinsic::mips_ffint_u_d:
1893 return DAG.getNode(ISD::UINT_TO_FP, DL, Op->getValueType(0),
1895 case Intrinsic::mips_ffint_s_w:
1896 case Intrinsic::mips_ffint_s_d:
1897 return DAG.getNode(ISD::SINT_TO_FP, DL, Op->getValueType(0),
1899 case Intrinsic::mips_fill_b:
1900 case Intrinsic::mips_fill_h:
1901 case Intrinsic::mips_fill_w:
1902 case Intrinsic::mips_fill_d: {
1903 EVT ResTy = Op->getValueType(0);
1904 SmallVector<SDValue, 16> Ops(ResTy.getVectorNumElements(),
1907 // If ResTy is v2i64 then the type legalizer will break this node down into
1908 // an equivalent v4i32.
1909 return DAG.getBuildVector(ResTy, DL, Ops);
1911 case Intrinsic::mips_fexp2_w:
1912 case Intrinsic::mips_fexp2_d: {
1913 // TODO: If intrinsics have fast-math-flags, propagate them.
1914 EVT ResTy = Op->getValueType(0);
1916 ISD::FMUL, SDLoc(Op), ResTy, Op->getOperand(1),
1917 DAG.getNode(ISD::FEXP2, SDLoc(Op), ResTy, Op->getOperand(2)));
1919 case Intrinsic::mips_flog2_w:
1920 case Intrinsic::mips_flog2_d:
1921 return DAG.getNode(ISD::FLOG2, DL, Op->getValueType(0), Op->getOperand(1));
1922 case Intrinsic::mips_fmadd_w:
1923 case Intrinsic::mips_fmadd_d:
1924 return DAG.getNode(ISD::FMA, SDLoc(Op), Op->getValueType(0),
1925 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3));
1926 case Intrinsic::mips_fmul_w:
1927 case Intrinsic::mips_fmul_d: {
1928 // TODO: If intrinsics have fast-math-flags, propagate them.
1929 return DAG.getNode(ISD::FMUL, DL, Op->getValueType(0), Op->getOperand(1),
1932 case Intrinsic::mips_fmsub_w:
1933 case Intrinsic::mips_fmsub_d: {
1934 // TODO: If intrinsics have fast-math-flags, propagate them.
1935 EVT ResTy = Op->getValueType(0);
1936 return DAG.getNode(ISD::FSUB, SDLoc(Op), ResTy, Op->getOperand(1),
1937 DAG.getNode(ISD::FMUL, SDLoc(Op), ResTy,
1938 Op->getOperand(2), Op->getOperand(3)));
1940 case Intrinsic::mips_frint_w:
1941 case Intrinsic::mips_frint_d:
1942 return DAG.getNode(ISD::FRINT, DL, Op->getValueType(0), Op->getOperand(1));
1943 case Intrinsic::mips_fsqrt_w:
1944 case Intrinsic::mips_fsqrt_d:
1945 return DAG.getNode(ISD::FSQRT, DL, Op->getValueType(0), Op->getOperand(1));
1946 case Intrinsic::mips_fsub_w:
1947 case Intrinsic::mips_fsub_d: {
1948 // TODO: If intrinsics have fast-math-flags, propagate them.
1949 return DAG.getNode(ISD::FSUB, DL, Op->getValueType(0), Op->getOperand(1),
1952 case Intrinsic::mips_ftrunc_u_w:
1953 case Intrinsic::mips_ftrunc_u_d:
1954 return DAG.getNode(ISD::FP_TO_UINT, DL, Op->getValueType(0),
1956 case Intrinsic::mips_ftrunc_s_w:
1957 case Intrinsic::mips_ftrunc_s_d:
1958 return DAG.getNode(ISD::FP_TO_SINT, DL, Op->getValueType(0),
1960 case Intrinsic::mips_ilvev_b:
1961 case Intrinsic::mips_ilvev_h:
1962 case Intrinsic::mips_ilvev_w:
1963 case Intrinsic::mips_ilvev_d:
1964 return DAG.getNode(MipsISD::ILVEV, DL, Op->getValueType(0),
1965 Op->getOperand(1), Op->getOperand(2));
1966 case Intrinsic::mips_ilvl_b:
1967 case Intrinsic::mips_ilvl_h:
1968 case Intrinsic::mips_ilvl_w:
1969 case Intrinsic::mips_ilvl_d:
1970 return DAG.getNode(MipsISD::ILVL, DL, Op->getValueType(0),
1971 Op->getOperand(1), Op->getOperand(2));
1972 case Intrinsic::mips_ilvod_b:
1973 case Intrinsic::mips_ilvod_h:
1974 case Intrinsic::mips_ilvod_w:
1975 case Intrinsic::mips_ilvod_d:
1976 return DAG.getNode(MipsISD::ILVOD, DL, Op->getValueType(0),
1977 Op->getOperand(1), Op->getOperand(2));
1978 case Intrinsic::mips_ilvr_b:
1979 case Intrinsic::mips_ilvr_h:
1980 case Intrinsic::mips_ilvr_w:
1981 case Intrinsic::mips_ilvr_d:
1982 return DAG.getNode(MipsISD::ILVR, DL, Op->getValueType(0),
1983 Op->getOperand(1), Op->getOperand(2));
1984 case Intrinsic::mips_insert_b:
1985 case Intrinsic::mips_insert_h:
1986 case Intrinsic::mips_insert_w:
1987 case Intrinsic::mips_insert_d:
1988 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(Op), Op->getValueType(0),
1989 Op->getOperand(1), Op->getOperand(3), Op->getOperand(2));
1990 case Intrinsic::mips_insve_b:
1991 case Intrinsic::mips_insve_h:
1992 case Intrinsic::mips_insve_w:
1993 case Intrinsic::mips_insve_d:
1994 return DAG.getNode(MipsISD::INSVE, DL, Op->getValueType(0),
1995 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3),
1996 DAG.getConstant(0, DL, MVT::i32));
1997 case Intrinsic::mips_ldi_b:
1998 case Intrinsic::mips_ldi_h:
1999 case Intrinsic::mips_ldi_w:
2000 case Intrinsic::mips_ldi_d:
2001 return lowerMSASplatImm(Op, 1, DAG);
2002 case Intrinsic::mips_lsa:
2003 case Intrinsic::mips_dlsa: {
2004 EVT ResTy = Op->getValueType(0);
2005 return DAG.getNode(ISD::ADD, SDLoc(Op), ResTy, Op->getOperand(1),
2006 DAG.getNode(ISD::SHL, SDLoc(Op), ResTy,
2007 Op->getOperand(2), Op->getOperand(3)));
2009 case Intrinsic::mips_maddv_b:
2010 case Intrinsic::mips_maddv_h:
2011 case Intrinsic::mips_maddv_w:
2012 case Intrinsic::mips_maddv_d: {
2013 EVT ResTy = Op->getValueType(0);
2014 return DAG.getNode(ISD::ADD, SDLoc(Op), ResTy, Op->getOperand(1),
2015 DAG.getNode(ISD::MUL, SDLoc(Op), ResTy,
2016 Op->getOperand(2), Op->getOperand(3)));
2018 case Intrinsic::mips_max_s_b:
2019 case Intrinsic::mips_max_s_h:
2020 case Intrinsic::mips_max_s_w:
2021 case Intrinsic::mips_max_s_d:
2022 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0),
2023 Op->getOperand(1), Op->getOperand(2));
2024 case Intrinsic::mips_max_u_b:
2025 case Intrinsic::mips_max_u_h:
2026 case Intrinsic::mips_max_u_w:
2027 case Intrinsic::mips_max_u_d:
2028 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0),
2029 Op->getOperand(1), Op->getOperand(2));
2030 case Intrinsic::mips_maxi_s_b:
2031 case Intrinsic::mips_maxi_s_h:
2032 case Intrinsic::mips_maxi_s_w:
2033 case Intrinsic::mips_maxi_s_d:
2034 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0),
2035 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2036 case Intrinsic::mips_maxi_u_b:
2037 case Intrinsic::mips_maxi_u_h:
2038 case Intrinsic::mips_maxi_u_w:
2039 case Intrinsic::mips_maxi_u_d:
2040 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0),
2041 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2042 case Intrinsic::mips_min_s_b:
2043 case Intrinsic::mips_min_s_h:
2044 case Intrinsic::mips_min_s_w:
2045 case Intrinsic::mips_min_s_d:
2046 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0),
2047 Op->getOperand(1), Op->getOperand(2));
2048 case Intrinsic::mips_min_u_b:
2049 case Intrinsic::mips_min_u_h:
2050 case Intrinsic::mips_min_u_w:
2051 case Intrinsic::mips_min_u_d:
2052 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0),
2053 Op->getOperand(1), Op->getOperand(2));
2054 case Intrinsic::mips_mini_s_b:
2055 case Intrinsic::mips_mini_s_h:
2056 case Intrinsic::mips_mini_s_w:
2057 case Intrinsic::mips_mini_s_d:
2058 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0),
2059 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2060 case Intrinsic::mips_mini_u_b:
2061 case Intrinsic::mips_mini_u_h:
2062 case Intrinsic::mips_mini_u_w:
2063 case Intrinsic::mips_mini_u_d:
2064 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0),
2065 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2066 case Intrinsic::mips_mod_s_b:
2067 case Intrinsic::mips_mod_s_h:
2068 case Intrinsic::mips_mod_s_w:
2069 case Intrinsic::mips_mod_s_d:
2070 return DAG.getNode(ISD::SREM, DL, Op->getValueType(0), Op->getOperand(1),
2072 case Intrinsic::mips_mod_u_b:
2073 case Intrinsic::mips_mod_u_h:
2074 case Intrinsic::mips_mod_u_w:
2075 case Intrinsic::mips_mod_u_d:
2076 return DAG.getNode(ISD::UREM, DL, Op->getValueType(0), Op->getOperand(1),
2078 case Intrinsic::mips_mulv_b:
2079 case Intrinsic::mips_mulv_h:
2080 case Intrinsic::mips_mulv_w:
2081 case Intrinsic::mips_mulv_d:
2082 return DAG.getNode(ISD::MUL, DL, Op->getValueType(0), Op->getOperand(1),
2084 case Intrinsic::mips_msubv_b:
2085 case Intrinsic::mips_msubv_h:
2086 case Intrinsic::mips_msubv_w:
2087 case Intrinsic::mips_msubv_d: {
2088 EVT ResTy = Op->getValueType(0);
2089 return DAG.getNode(ISD::SUB, SDLoc(Op), ResTy, Op->getOperand(1),
2090 DAG.getNode(ISD::MUL, SDLoc(Op), ResTy,
2091 Op->getOperand(2), Op->getOperand(3)));
2093 case Intrinsic::mips_nlzc_b:
2094 case Intrinsic::mips_nlzc_h:
2095 case Intrinsic::mips_nlzc_w:
2096 case Intrinsic::mips_nlzc_d:
2097 return DAG.getNode(ISD::CTLZ, DL, Op->getValueType(0), Op->getOperand(1));
2098 case Intrinsic::mips_nor_v: {
2099 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2100 Op->getOperand(1), Op->getOperand(2));
2101 return DAG.getNOT(DL, Res, Res->getValueType(0));
2103 case Intrinsic::mips_nori_b: {
2104 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2106 lowerMSASplatImm(Op, 2, DAG));
2107 return DAG.getNOT(DL, Res, Res->getValueType(0));
2109 case Intrinsic::mips_or_v:
2110 return DAG.getNode(ISD::OR, DL, Op->getValueType(0), Op->getOperand(1),
2112 case Intrinsic::mips_ori_b:
2113 return DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2114 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2115 case Intrinsic::mips_pckev_b:
2116 case Intrinsic::mips_pckev_h:
2117 case Intrinsic::mips_pckev_w:
2118 case Intrinsic::mips_pckev_d:
2119 return DAG.getNode(MipsISD::PCKEV, DL, Op->getValueType(0),
2120 Op->getOperand(1), Op->getOperand(2));
2121 case Intrinsic::mips_pckod_b:
2122 case Intrinsic::mips_pckod_h:
2123 case Intrinsic::mips_pckod_w:
2124 case Intrinsic::mips_pckod_d:
2125 return DAG.getNode(MipsISD::PCKOD, DL, Op->getValueType(0),
2126 Op->getOperand(1), Op->getOperand(2));
2127 case Intrinsic::mips_pcnt_b:
2128 case Intrinsic::mips_pcnt_h:
2129 case Intrinsic::mips_pcnt_w:
2130 case Intrinsic::mips_pcnt_d:
2131 return DAG.getNode(ISD::CTPOP, DL, Op->getValueType(0), Op->getOperand(1));
2132 case Intrinsic::mips_shf_b:
2133 case Intrinsic::mips_shf_h:
2134 case Intrinsic::mips_shf_w:
2135 return DAG.getNode(MipsISD::SHF, DL, Op->getValueType(0),
2136 Op->getOperand(2), Op->getOperand(1));
2137 case Intrinsic::mips_sll_b:
2138 case Intrinsic::mips_sll_h:
2139 case Intrinsic::mips_sll_w:
2140 case Intrinsic::mips_sll_d:
2141 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0), Op->getOperand(1),
2143 case Intrinsic::mips_slli_b:
2144 case Intrinsic::mips_slli_h:
2145 case Intrinsic::mips_slli_w:
2146 case Intrinsic::mips_slli_d:
2147 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0),
2148 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2149 case Intrinsic::mips_splat_b:
2150 case Intrinsic::mips_splat_h:
2151 case Intrinsic::mips_splat_w:
2152 case Intrinsic::mips_splat_d:
2153 // We can't lower via VECTOR_SHUFFLE because it requires constant shuffle
2154 // masks, nor can we lower via BUILD_VECTOR & EXTRACT_VECTOR_ELT because
2155 // EXTRACT_VECTOR_ELT can't extract i64's on MIPS32.
2156 // Instead we lower to MipsISD::VSHF and match from there.
2157 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2158 lowerMSASplatZExt(Op, 2, DAG), Op->getOperand(1),
2160 case Intrinsic::mips_splati_b:
2161 case Intrinsic::mips_splati_h:
2162 case Intrinsic::mips_splati_w:
2163 case Intrinsic::mips_splati_d:
2164 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2165 lowerMSASplatImm(Op, 2, DAG), Op->getOperand(1),
2167 case Intrinsic::mips_sra_b:
2168 case Intrinsic::mips_sra_h:
2169 case Intrinsic::mips_sra_w:
2170 case Intrinsic::mips_sra_d:
2171 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0), Op->getOperand(1),
2173 case Intrinsic::mips_srai_b:
2174 case Intrinsic::mips_srai_h:
2175 case Intrinsic::mips_srai_w:
2176 case Intrinsic::mips_srai_d:
2177 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0),
2178 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2179 case Intrinsic::mips_srl_b:
2180 case Intrinsic::mips_srl_h:
2181 case Intrinsic::mips_srl_w:
2182 case Intrinsic::mips_srl_d:
2183 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0), Op->getOperand(1),
2185 case Intrinsic::mips_srli_b:
2186 case Intrinsic::mips_srli_h:
2187 case Intrinsic::mips_srli_w:
2188 case Intrinsic::mips_srli_d:
2189 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0),
2190 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2191 case Intrinsic::mips_subv_b:
2192 case Intrinsic::mips_subv_h:
2193 case Intrinsic::mips_subv_w:
2194 case Intrinsic::mips_subv_d:
2195 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0), Op->getOperand(1),
2197 case Intrinsic::mips_subvi_b:
2198 case Intrinsic::mips_subvi_h:
2199 case Intrinsic::mips_subvi_w:
2200 case Intrinsic::mips_subvi_d:
2201 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0),
2202 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2203 case Intrinsic::mips_vshf_b:
2204 case Intrinsic::mips_vshf_h:
2205 case Intrinsic::mips_vshf_w:
2206 case Intrinsic::mips_vshf_d:
2207 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2208 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3));
2209 case Intrinsic::mips_xor_v:
2210 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0), Op->getOperand(1),
2212 case Intrinsic::mips_xori_b:
2213 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0),
2214 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2215 case Intrinsic::thread_pointer: {
2216 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2217 return DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
2222 static SDValue lowerMSALoadIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
2224 SDValue ChainIn = Op->getOperand(0);
2225 SDValue Address = Op->getOperand(2);
2226 SDValue Offset = Op->getOperand(3);
2227 EVT ResTy = Op->getValueType(0);
2228 EVT PtrTy = Address->getValueType(0);
2230 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
2231 return DAG.getLoad(ResTy, DL, ChainIn, Address, MachinePointerInfo(),
2232 /* Alignment = */ 16);
2235 SDValue MipsSETargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
2236 SelectionDAG &DAG) const {
2237 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
2241 case Intrinsic::mips_extp:
2242 return lowerDSPIntr(Op, DAG, MipsISD::EXTP);
2243 case Intrinsic::mips_extpdp:
2244 return lowerDSPIntr(Op, DAG, MipsISD::EXTPDP);
2245 case Intrinsic::mips_extr_w:
2246 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_W);
2247 case Intrinsic::mips_extr_r_w:
2248 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W);
2249 case Intrinsic::mips_extr_rs_w:
2250 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W);
2251 case Intrinsic::mips_extr_s_h:
2252 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H);
2253 case Intrinsic::mips_mthlip:
2254 return lowerDSPIntr(Op, DAG, MipsISD::MTHLIP);
2255 case Intrinsic::mips_mulsaq_s_w_ph:
2256 return lowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH);
2257 case Intrinsic::mips_maq_s_w_phl:
2258 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL);
2259 case Intrinsic::mips_maq_s_w_phr:
2260 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR);
2261 case Intrinsic::mips_maq_sa_w_phl:
2262 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL);
2263 case Intrinsic::mips_maq_sa_w_phr:
2264 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR);
2265 case Intrinsic::mips_dpaq_s_w_ph:
2266 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH);
2267 case Intrinsic::mips_dpsq_s_w_ph:
2268 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH);
2269 case Intrinsic::mips_dpaq_sa_l_w:
2270 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W);
2271 case Intrinsic::mips_dpsq_sa_l_w:
2272 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W);
2273 case Intrinsic::mips_dpaqx_s_w_ph:
2274 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH);
2275 case Intrinsic::mips_dpaqx_sa_w_ph:
2276 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH);
2277 case Intrinsic::mips_dpsqx_s_w_ph:
2278 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH);
2279 case Intrinsic::mips_dpsqx_sa_w_ph:
2280 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH);
2281 case Intrinsic::mips_ld_b:
2282 case Intrinsic::mips_ld_h:
2283 case Intrinsic::mips_ld_w:
2284 case Intrinsic::mips_ld_d:
2285 return lowerMSALoadIntr(Op, DAG, Intr);
2289 static SDValue lowerMSAStoreIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
2291 SDValue ChainIn = Op->getOperand(0);
2292 SDValue Value = Op->getOperand(2);
2293 SDValue Address = Op->getOperand(3);
2294 SDValue Offset = Op->getOperand(4);
2295 EVT PtrTy = Address->getValueType(0);
2297 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
2299 return DAG.getStore(ChainIn, DL, Value, Address, MachinePointerInfo(),
2300 /* Alignment = */ 16);
2303 SDValue MipsSETargetLowering::lowerINTRINSIC_VOID(SDValue Op,
2304 SelectionDAG &DAG) const {
2305 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
2309 case Intrinsic::mips_st_b:
2310 case Intrinsic::mips_st_h:
2311 case Intrinsic::mips_st_w:
2312 case Intrinsic::mips_st_d:
2313 return lowerMSAStoreIntr(Op, DAG, Intr);
2317 /// \brief Check if the given BuildVectorSDNode is a splat.
2318 /// This method currently relies on DAG nodes being reused when equivalent,
2319 /// so it's possible for this to return false even when isConstantSplat returns
2321 static bool isSplatVector(const BuildVectorSDNode *N) {
2322 unsigned int nOps = N->getNumOperands();
2323 assert(nOps > 1 && "isSplatVector has 0 or 1 sized build vector");
2325 SDValue Operand0 = N->getOperand(0);
2327 for (unsigned int i = 1; i < nOps; ++i) {
2328 if (N->getOperand(i) != Operand0)
2335 // Lower ISD::EXTRACT_VECTOR_ELT into MipsISD::VEXTRACT_SEXT_ELT.
2337 // The non-value bits resulting from ISD::EXTRACT_VECTOR_ELT are undefined. We
2338 // choose to sign-extend but we could have equally chosen zero-extend. The
2339 // DAGCombiner will fold any sign/zero extension of the ISD::EXTRACT_VECTOR_ELT
2340 // result into this node later (possibly changing it to a zero-extend in the
2342 SDValue MipsSETargetLowering::
2343 lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
2345 EVT ResTy = Op->getValueType(0);
2346 SDValue Op0 = Op->getOperand(0);
2347 EVT VecTy = Op0->getValueType(0);
2349 if (!VecTy.is128BitVector())
2352 if (ResTy.isInteger()) {
2353 SDValue Op1 = Op->getOperand(1);
2354 EVT EltTy = VecTy.getVectorElementType();
2355 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, DL, ResTy, Op0, Op1,
2356 DAG.getValueType(EltTy));
2362 static bool isConstantOrUndef(const SDValue Op) {
2365 if (isa<ConstantSDNode>(Op))
2367 if (isa<ConstantFPSDNode>(Op))
2372 static bool isConstantOrUndefBUILD_VECTOR(const BuildVectorSDNode *Op) {
2373 for (unsigned i = 0; i < Op->getNumOperands(); ++i)
2374 if (isConstantOrUndef(Op->getOperand(i)))
2379 // Lowers ISD::BUILD_VECTOR into appropriate SelectionDAG nodes for the
2382 // Lowers according to the following rules:
2383 // - Constant splats are legal as-is as long as the SplatBitSize is a power of
2384 // 2 less than or equal to 64 and the value fits into a signed 10-bit
2386 // - Constant splats are lowered to bitconverted BUILD_VECTORs if SplatBitSize
2387 // is a power of 2 less than or equal to 64 and the value does not fit into a
2388 // signed 10-bit immediate
2389 // - Non-constant splats are legal as-is.
2390 // - Non-constant non-splats are lowered to sequences of INSERT_VECTOR_ELT.
2391 // - All others are illegal and must be expanded.
2392 SDValue MipsSETargetLowering::lowerBUILD_VECTOR(SDValue Op,
2393 SelectionDAG &DAG) const {
2394 BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op);
2395 EVT ResTy = Op->getValueType(0);
2397 APInt SplatValue, SplatUndef;
2398 unsigned SplatBitSize;
2401 if (!Subtarget.hasMSA() || !ResTy.is128BitVector())
2404 if (Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
2406 !Subtarget.isLittle()) && SplatBitSize <= 64) {
2407 // We can only cope with 8, 16, 32, or 64-bit elements
2408 if (SplatBitSize != 8 && SplatBitSize != 16 && SplatBitSize != 32 &&
2412 // If the value fits into a simm10 then we can use ldi.[bhwd]
2413 // However, if it isn't an integer type we will have to bitcast from an
2414 // integer type first. Also, if there are any undefs, we must lower them
2415 // to defined values first.
2416 if (ResTy.isInteger() && !HasAnyUndefs && SplatValue.isSignedIntN(10))
2421 switch (SplatBitSize) {
2425 ViaVecTy = MVT::v16i8;
2428 ViaVecTy = MVT::v8i16;
2431 ViaVecTy = MVT::v4i32;
2434 // There's no fill.d to fall back on for 64-bit values
2438 // SelectionDAG::getConstant will promote SplatValue appropriately.
2439 SDValue Result = DAG.getConstant(SplatValue, DL, ViaVecTy);
2441 // Bitcast to the type we originally wanted
2442 if (ViaVecTy != ResTy)
2443 Result = DAG.getNode(ISD::BITCAST, SDLoc(Node), ResTy, Result);
2446 } else if (isSplatVector(Node))
2448 else if (!isConstantOrUndefBUILD_VECTOR(Node)) {
2449 // Use INSERT_VECTOR_ELT operations rather than expand to stores.
2450 // The resulting code is the same length as the expansion, but it doesn't
2451 // use memory operations
2452 EVT ResTy = Node->getValueType(0);
2454 assert(ResTy.isVector());
2456 unsigned NumElts = ResTy.getVectorNumElements();
2457 SDValue Vector = DAG.getUNDEF(ResTy);
2458 for (unsigned i = 0; i < NumElts; ++i) {
2459 Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector,
2460 Node->getOperand(i),
2461 DAG.getConstant(i, DL, MVT::i32));
2469 // Lower VECTOR_SHUFFLE into SHF (if possible).
2471 // SHF splits the vector into blocks of four elements, then shuffles these
2472 // elements according to a <4 x i2> constant (encoded as an integer immediate).
2474 // It is therefore possible to lower into SHF when the mask takes the form:
2475 // <a, b, c, d, a+4, b+4, c+4, d+4, a+8, b+8, c+8, d+8, ...>
2476 // When undef's appear they are treated as if they were whatever value is
2477 // necessary in order to fit the above forms.
2480 // %2 = shufflevector <8 x i16> %0, <8 x i16> undef,
2481 // <8 x i32> <i32 3, i32 2, i32 1, i32 0,
2482 // i32 7, i32 6, i32 5, i32 4>
2484 // (SHF_H $w0, $w1, 27)
2485 // where the 27 comes from:
2486 // 3 + (2 << 2) + (1 << 4) + (0 << 6)
2487 static SDValue lowerVECTOR_SHUFFLE_SHF(SDValue Op, EVT ResTy,
2488 SmallVector<int, 16> Indices,
2489 SelectionDAG &DAG) {
2490 int SHFIndices[4] = { -1, -1, -1, -1 };
2492 if (Indices.size() < 4)
2495 for (unsigned i = 0; i < 4; ++i) {
2496 for (unsigned j = i; j < Indices.size(); j += 4) {
2497 int Idx = Indices[j];
2499 // Convert from vector index to 4-element subvector index
2500 // If an index refers to an element outside of the subvector then give up
2503 if (Idx < 0 || Idx >= 4)
2507 // If the mask has an undef, replace it with the current index.
2508 // Note that it might still be undef if the current index is also undef
2509 if (SHFIndices[i] == -1)
2510 SHFIndices[i] = Idx;
2512 // Check that non-undef values are the same as in the mask. If they
2513 // aren't then give up
2514 if (!(Idx == -1 || Idx == SHFIndices[i]))
2519 // Calculate the immediate. Replace any remaining undefs with zero
2521 for (int i = 3; i >= 0; --i) {
2522 int Idx = SHFIndices[i];
2532 return DAG.getNode(MipsISD::SHF, DL, ResTy,
2533 DAG.getConstant(Imm, DL, MVT::i32), Op->getOperand(0));
2536 /// Determine whether a range fits a regular pattern of values.
2537 /// This function accounts for the possibility of jumping over the End iterator.
2538 template <typename ValType>
2540 fitsRegularPattern(typename SmallVectorImpl<ValType>::const_iterator Begin,
2541 unsigned CheckStride,
2542 typename SmallVectorImpl<ValType>::const_iterator End,
2543 ValType ExpectedIndex, unsigned ExpectedIndexStride) {
2547 if (*I != -1 && *I != ExpectedIndex)
2549 ExpectedIndex += ExpectedIndexStride;
2551 // Incrementing past End is undefined behaviour so we must increment one
2552 // step at a time and check for End at each step.
2553 for (unsigned n = 0; n < CheckStride && I != End; ++n, ++I)
2554 ; // Empty loop body.
2559 // Determine whether VECTOR_SHUFFLE is a SPLATI.
2561 // It is a SPLATI when the mask is:
2563 // where x is any valid index.
2565 // When undef's appear in the mask they are treated as if they were whatever
2566 // value is necessary in order to fit the above form.
2567 static bool isVECTOR_SHUFFLE_SPLATI(SDValue Op, EVT ResTy,
2568 SmallVector<int, 16> Indices,
2569 SelectionDAG &DAG) {
2570 assert((Indices.size() % 2) == 0);
2572 int SplatIndex = -1;
2573 for (const auto &V : Indices) {
2580 return fitsRegularPattern<int>(Indices.begin(), 1, Indices.end(), SplatIndex,
2584 // Lower VECTOR_SHUFFLE into ILVEV (if possible).
2586 // ILVEV interleaves the even elements from each vector.
2588 // It is possible to lower into ILVEV when the mask consists of two of the
2589 // following forms interleaved:
2591 // <n, n+2, n+4, ...>
2592 // where n is the number of elements in the vector.
2594 // <0, 0, 2, 2, 4, 4, ...>
2595 // <0, n, 2, n+2, 4, n+4, ...>
2597 // When undef's appear in the mask they are treated as if they were whatever
2598 // value is necessary in order to fit the above forms.
2599 static SDValue lowerVECTOR_SHUFFLE_ILVEV(SDValue Op, EVT ResTy,
2600 SmallVector<int, 16> Indices,
2601 SelectionDAG &DAG) {
2602 assert((Indices.size() % 2) == 0);
2606 const auto &Begin = Indices.begin();
2607 const auto &End = Indices.end();
2609 // Check even elements are taken from the even elements of one half or the
2610 // other and pick an operand accordingly.
2611 if (fitsRegularPattern<int>(Begin, 2, End, 0, 2))
2612 Wt = Op->getOperand(0);
2613 else if (fitsRegularPattern<int>(Begin, 2, End, Indices.size(), 2))
2614 Wt = Op->getOperand(1);
2618 // Check odd elements are taken from the even elements of one half or the
2619 // other and pick an operand accordingly.
2620 if (fitsRegularPattern<int>(Begin + 1, 2, End, 0, 2))
2621 Ws = Op->getOperand(0);
2622 else if (fitsRegularPattern<int>(Begin + 1, 2, End, Indices.size(), 2))
2623 Ws = Op->getOperand(1);
2627 return DAG.getNode(MipsISD::ILVEV, SDLoc(Op), ResTy, Ws, Wt);
2630 // Lower VECTOR_SHUFFLE into ILVOD (if possible).
2632 // ILVOD interleaves the odd elements from each vector.
2634 // It is possible to lower into ILVOD when the mask consists of two of the
2635 // following forms interleaved:
2637 // <n+1, n+3, n+5, ...>
2638 // where n is the number of elements in the vector.
2640 // <1, 1, 3, 3, 5, 5, ...>
2641 // <1, n+1, 3, n+3, 5, n+5, ...>
2643 // When undef's appear in the mask they are treated as if they were whatever
2644 // value is necessary in order to fit the above forms.
2645 static SDValue lowerVECTOR_SHUFFLE_ILVOD(SDValue Op, EVT ResTy,
2646 SmallVector<int, 16> Indices,
2647 SelectionDAG &DAG) {
2648 assert((Indices.size() % 2) == 0);
2652 const auto &Begin = Indices.begin();
2653 const auto &End = Indices.end();
2655 // Check even elements are taken from the odd elements of one half or the
2656 // other and pick an operand accordingly.
2657 if (fitsRegularPattern<int>(Begin, 2, End, 1, 2))
2658 Wt = Op->getOperand(0);
2659 else if (fitsRegularPattern<int>(Begin, 2, End, Indices.size() + 1, 2))
2660 Wt = Op->getOperand(1);
2664 // Check odd elements are taken from the odd elements of one half or the
2665 // other and pick an operand accordingly.
2666 if (fitsRegularPattern<int>(Begin + 1, 2, End, 1, 2))
2667 Ws = Op->getOperand(0);
2668 else if (fitsRegularPattern<int>(Begin + 1, 2, End, Indices.size() + 1, 2))
2669 Ws = Op->getOperand(1);
2673 return DAG.getNode(MipsISD::ILVOD, SDLoc(Op), ResTy, Wt, Ws);
2676 // Lower VECTOR_SHUFFLE into ILVR (if possible).
2678 // ILVR interleaves consecutive elements from the right (lowest-indexed) half of
2681 // It is possible to lower into ILVR when the mask consists of two of the
2682 // following forms interleaved:
2684 // <n, n+1, n+2, ...>
2685 // where n is the number of elements in the vector.
2687 // <0, 0, 1, 1, 2, 2, ...>
2688 // <0, n, 1, n+1, 2, n+2, ...>
2690 // When undef's appear in the mask they are treated as if they were whatever
2691 // value is necessary in order to fit the above forms.
2692 static SDValue lowerVECTOR_SHUFFLE_ILVR(SDValue Op, EVT ResTy,
2693 SmallVector<int, 16> Indices,
2694 SelectionDAG &DAG) {
2695 assert((Indices.size() % 2) == 0);
2699 const auto &Begin = Indices.begin();
2700 const auto &End = Indices.end();
2702 // Check even elements are taken from the right (lowest-indexed) elements of
2703 // one half or the other and pick an operand accordingly.
2704 if (fitsRegularPattern<int>(Begin, 2, End, 0, 1))
2705 Wt = Op->getOperand(0);
2706 else if (fitsRegularPattern<int>(Begin, 2, End, Indices.size(), 1))
2707 Wt = Op->getOperand(1);
2711 // Check odd elements are taken from the right (lowest-indexed) elements of
2712 // one half or the other and pick an operand accordingly.
2713 if (fitsRegularPattern<int>(Begin + 1, 2, End, 0, 1))
2714 Ws = Op->getOperand(0);
2715 else if (fitsRegularPattern<int>(Begin + 1, 2, End, Indices.size(), 1))
2716 Ws = Op->getOperand(1);
2720 return DAG.getNode(MipsISD::ILVR, SDLoc(Op), ResTy, Ws, Wt);
2723 // Lower VECTOR_SHUFFLE into ILVL (if possible).
2725 // ILVL interleaves consecutive elements from the left (highest-indexed) half
2728 // It is possible to lower into ILVL when the mask consists of two of the
2729 // following forms interleaved:
2730 // <x, x+1, x+2, ...>
2731 // <n+x, n+x+1, n+x+2, ...>
2732 // where n is the number of elements in the vector and x is half n.
2734 // <x, x, x+1, x+1, x+2, x+2, ...>
2735 // <x, n+x, x+1, n+x+1, x+2, n+x+2, ...>
2737 // When undef's appear in the mask they are treated as if they were whatever
2738 // value is necessary in order to fit the above forms.
2739 static SDValue lowerVECTOR_SHUFFLE_ILVL(SDValue Op, EVT ResTy,
2740 SmallVector<int, 16> Indices,
2741 SelectionDAG &DAG) {
2742 assert((Indices.size() % 2) == 0);
2744 unsigned HalfSize = Indices.size() / 2;
2747 const auto &Begin = Indices.begin();
2748 const auto &End = Indices.end();
2750 // Check even elements are taken from the left (highest-indexed) elements of
2751 // one half or the other and pick an operand accordingly.
2752 if (fitsRegularPattern<int>(Begin, 2, End, HalfSize, 1))
2753 Wt = Op->getOperand(0);
2754 else if (fitsRegularPattern<int>(Begin, 2, End, Indices.size() + HalfSize, 1))
2755 Wt = Op->getOperand(1);
2759 // Check odd elements are taken from the left (highest-indexed) elements of
2760 // one half or the other and pick an operand accordingly.
2761 if (fitsRegularPattern<int>(Begin + 1, 2, End, HalfSize, 1))
2762 Ws = Op->getOperand(0);
2763 else if (fitsRegularPattern<int>(Begin + 1, 2, End, Indices.size() + HalfSize,
2765 Ws = Op->getOperand(1);
2769 return DAG.getNode(MipsISD::ILVL, SDLoc(Op), ResTy, Ws, Wt);
2772 // Lower VECTOR_SHUFFLE into PCKEV (if possible).
2774 // PCKEV copies the even elements of each vector into the result vector.
2776 // It is possible to lower into PCKEV when the mask consists of two of the
2777 // following forms concatenated:
2779 // <n, n+2, n+4, ...>
2780 // where n is the number of elements in the vector.
2782 // <0, 2, 4, ..., 0, 2, 4, ...>
2783 // <0, 2, 4, ..., n, n+2, n+4, ...>
2785 // When undef's appear in the mask they are treated as if they were whatever
2786 // value is necessary in order to fit the above forms.
2787 static SDValue lowerVECTOR_SHUFFLE_PCKEV(SDValue Op, EVT ResTy,
2788 SmallVector<int, 16> Indices,
2789 SelectionDAG &DAG) {
2790 assert((Indices.size() % 2) == 0);
2794 const auto &Begin = Indices.begin();
2795 const auto &Mid = Indices.begin() + Indices.size() / 2;
2796 const auto &End = Indices.end();
2798 if (fitsRegularPattern<int>(Begin, 1, Mid, 0, 2))
2799 Wt = Op->getOperand(0);
2800 else if (fitsRegularPattern<int>(Begin, 1, Mid, Indices.size(), 2))
2801 Wt = Op->getOperand(1);
2805 if (fitsRegularPattern<int>(Mid, 1, End, 0, 2))
2806 Ws = Op->getOperand(0);
2807 else if (fitsRegularPattern<int>(Mid, 1, End, Indices.size(), 2))
2808 Ws = Op->getOperand(1);
2812 return DAG.getNode(MipsISD::PCKEV, SDLoc(Op), ResTy, Ws, Wt);
2815 // Lower VECTOR_SHUFFLE into PCKOD (if possible).
2817 // PCKOD copies the odd elements of each vector into the result vector.
2819 // It is possible to lower into PCKOD when the mask consists of two of the
2820 // following forms concatenated:
2822 // <n+1, n+3, n+5, ...>
2823 // where n is the number of elements in the vector.
2825 // <1, 3, 5, ..., 1, 3, 5, ...>
2826 // <1, 3, 5, ..., n+1, n+3, n+5, ...>
2828 // When undef's appear in the mask they are treated as if they were whatever
2829 // value is necessary in order to fit the above forms.
2830 static SDValue lowerVECTOR_SHUFFLE_PCKOD(SDValue Op, EVT ResTy,
2831 SmallVector<int, 16> Indices,
2832 SelectionDAG &DAG) {
2833 assert((Indices.size() % 2) == 0);
2837 const auto &Begin = Indices.begin();
2838 const auto &Mid = Indices.begin() + Indices.size() / 2;
2839 const auto &End = Indices.end();
2841 if (fitsRegularPattern<int>(Begin, 1, Mid, 1, 2))
2842 Wt = Op->getOperand(0);
2843 else if (fitsRegularPattern<int>(Begin, 1, Mid, Indices.size() + 1, 2))
2844 Wt = Op->getOperand(1);
2848 if (fitsRegularPattern<int>(Mid, 1, End, 1, 2))
2849 Ws = Op->getOperand(0);
2850 else if (fitsRegularPattern<int>(Mid, 1, End, Indices.size() + 1, 2))
2851 Ws = Op->getOperand(1);
2855 return DAG.getNode(MipsISD::PCKOD, SDLoc(Op), ResTy, Ws, Wt);
2858 // Lower VECTOR_SHUFFLE into VSHF.
2860 // This mostly consists of converting the shuffle indices in Indices into a
2861 // BUILD_VECTOR and adding it as an operand to the resulting VSHF. There is
2862 // also code to eliminate unused operands of the VECTOR_SHUFFLE. For example,
2863 // if the type is v8i16 and all the indices are less than 8 then the second
2864 // operand is unused and can be replaced with anything. We choose to replace it
2865 // with the used operand since this reduces the number of instructions overall.
2866 static SDValue lowerVECTOR_SHUFFLE_VSHF(SDValue Op, EVT ResTy,
2867 SmallVector<int, 16> Indices,
2868 SelectionDAG &DAG) {
2869 SmallVector<SDValue, 16> Ops;
2872 EVT MaskVecTy = ResTy.changeVectorElementTypeToInteger();
2873 EVT MaskEltTy = MaskVecTy.getVectorElementType();
2874 bool Using1stVec = false;
2875 bool Using2ndVec = false;
2877 int ResTyNumElts = ResTy.getVectorNumElements();
2879 for (int i = 0; i < ResTyNumElts; ++i) {
2880 // Idx == -1 means UNDEF
2881 int Idx = Indices[i];
2883 if (0 <= Idx && Idx < ResTyNumElts)
2885 if (ResTyNumElts <= Idx && Idx < ResTyNumElts * 2)
2889 for (SmallVector<int, 16>::iterator I = Indices.begin(); I != Indices.end();
2891 Ops.push_back(DAG.getTargetConstant(*I, DL, MaskEltTy));
2893 SDValue MaskVec = DAG.getBuildVector(MaskVecTy, DL, Ops);
2895 if (Using1stVec && Using2ndVec) {
2896 Op0 = Op->getOperand(0);
2897 Op1 = Op->getOperand(1);
2898 } else if (Using1stVec)
2899 Op0 = Op1 = Op->getOperand(0);
2900 else if (Using2ndVec)
2901 Op0 = Op1 = Op->getOperand(1);
2903 llvm_unreachable("shuffle vector mask references neither vector operand?");
2905 // VECTOR_SHUFFLE concatenates the vectors in an vectorwise fashion.
2906 // <0b00, 0b01> + <0b10, 0b11> -> <0b00, 0b01, 0b10, 0b11>
2907 // VSHF concatenates the vectors in a bitwise fashion:
2908 // <0b00, 0b01> + <0b10, 0b11> ->
2909 // 0b0100 + 0b1110 -> 0b01001110
2910 // <0b10, 0b11, 0b00, 0b01>
2911 // We must therefore swap the operands to get the correct result.
2912 return DAG.getNode(MipsISD::VSHF, DL, ResTy, MaskVec, Op1, Op0);
2915 // Lower VECTOR_SHUFFLE into one of a number of instructions depending on the
2916 // indices in the shuffle.
2917 SDValue MipsSETargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
2918 SelectionDAG &DAG) const {
2919 ShuffleVectorSDNode *Node = cast<ShuffleVectorSDNode>(Op);
2920 EVT ResTy = Op->getValueType(0);
2922 if (!ResTy.is128BitVector())
2925 int ResTyNumElts = ResTy.getVectorNumElements();
2926 SmallVector<int, 16> Indices;
2928 for (int i = 0; i < ResTyNumElts; ++i)
2929 Indices.push_back(Node->getMaskElt(i));
2931 // splati.[bhwd] is preferable to the others but is matched from
2933 if (isVECTOR_SHUFFLE_SPLATI(Op, ResTy, Indices, DAG))
2934 return lowerVECTOR_SHUFFLE_VSHF(Op, ResTy, Indices, DAG);
2936 if ((Result = lowerVECTOR_SHUFFLE_ILVEV(Op, ResTy, Indices, DAG)))
2938 if ((Result = lowerVECTOR_SHUFFLE_ILVOD(Op, ResTy, Indices, DAG)))
2940 if ((Result = lowerVECTOR_SHUFFLE_ILVL(Op, ResTy, Indices, DAG)))
2942 if ((Result = lowerVECTOR_SHUFFLE_ILVR(Op, ResTy, Indices, DAG)))
2944 if ((Result = lowerVECTOR_SHUFFLE_PCKEV(Op, ResTy, Indices, DAG)))
2946 if ((Result = lowerVECTOR_SHUFFLE_PCKOD(Op, ResTy, Indices, DAG)))
2948 if ((Result = lowerVECTOR_SHUFFLE_SHF(Op, ResTy, Indices, DAG)))
2950 return lowerVECTOR_SHUFFLE_VSHF(Op, ResTy, Indices, DAG);
2954 MipsSETargetLowering::emitBPOSGE32(MachineInstr &MI,
2955 MachineBasicBlock *BB) const {
2957 // bposge32_pseudo $vr0
2967 // $vr0 = phi($vr2, $fbb, $vr1, $tbb)
2969 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2970 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
2971 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
2972 DebugLoc DL = MI.getDebugLoc();
2973 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2974 MachineFunction::iterator It = std::next(MachineFunction::iterator(BB));
2975 MachineFunction *F = BB->getParent();
2976 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
2977 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
2978 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
2981 F->insert(It, Sink);
2983 // Transfer the remainder of BB and its successor edges to Sink.
2984 Sink->splice(Sink->begin(), BB, std::next(MachineBasicBlock::iterator(MI)),
2986 Sink->transferSuccessorsAndUpdatePHIs(BB);
2989 BB->addSuccessor(FBB);
2990 BB->addSuccessor(TBB);
2991 FBB->addSuccessor(Sink);
2992 TBB->addSuccessor(Sink);
2994 // Insert the real bposge32 instruction to $BB.
2995 BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB);
2996 // Insert the real bposge32c instruction to $BB.
2997 BuildMI(BB, DL, TII->get(Mips::BPOSGE32C_MMR3)).addMBB(TBB);
3000 unsigned VR2 = RegInfo.createVirtualRegister(RC);
3001 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2)
3002 .addReg(Mips::ZERO).addImm(0);
3003 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
3006 unsigned VR1 = RegInfo.createVirtualRegister(RC);
3007 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1)
3008 .addReg(Mips::ZERO).addImm(1);
3010 // Insert phi function to $Sink.
3011 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
3012 MI.getOperand(0).getReg())
3018 MI.eraseFromParent(); // The pseudo instruction is gone now.
3022 MachineBasicBlock *MipsSETargetLowering::emitMSACBranchPseudo(
3023 MachineInstr &MI, MachineBasicBlock *BB, unsigned BranchOp) const {
3025 // vany_nonzero $rd, $ws
3036 // $rd = phi($rd1, $fbb, $rd2, $tbb)
3038 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3039 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3040 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
3041 DebugLoc DL = MI.getDebugLoc();
3042 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3043 MachineFunction::iterator It = std::next(MachineFunction::iterator(BB));
3044 MachineFunction *F = BB->getParent();
3045 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
3046 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
3047 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
3050 F->insert(It, Sink);
3052 // Transfer the remainder of BB and its successor edges to Sink.
3053 Sink->splice(Sink->begin(), BB, std::next(MachineBasicBlock::iterator(MI)),
3055 Sink->transferSuccessorsAndUpdatePHIs(BB);
3058 BB->addSuccessor(FBB);
3059 BB->addSuccessor(TBB);
3060 FBB->addSuccessor(Sink);
3061 TBB->addSuccessor(Sink);
3063 // Insert the real bnz.b instruction to $BB.
3064 BuildMI(BB, DL, TII->get(BranchOp))
3065 .addReg(MI.getOperand(1).getReg())
3069 unsigned RD1 = RegInfo.createVirtualRegister(RC);
3070 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), RD1)
3071 .addReg(Mips::ZERO).addImm(0);
3072 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
3075 unsigned RD2 = RegInfo.createVirtualRegister(RC);
3076 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), RD2)
3077 .addReg(Mips::ZERO).addImm(1);
3079 // Insert phi function to $Sink.
3080 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
3081 MI.getOperand(0).getReg())
3087 MI.eraseFromParent(); // The pseudo instruction is gone now.
3091 // Emit the COPY_FW pseudo instruction.
3093 // copy_fw_pseudo $fd, $ws, n
3095 // copy_u_w $rt, $ws, $n
3098 // When n is zero, the equivalent operation can be performed with (potentially)
3099 // zero instructions due to register overlaps. This optimization is never valid
3100 // for lane 1 because it would require FR=0 mode which isn't supported by MSA.
3102 MipsSETargetLowering::emitCOPY_FW(MachineInstr &MI,
3103 MachineBasicBlock *BB) const {
3104 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3105 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3106 DebugLoc DL = MI.getDebugLoc();
3107 unsigned Fd = MI.getOperand(0).getReg();
3108 unsigned Ws = MI.getOperand(1).getReg();
3109 unsigned Lane = MI.getOperand(2).getImm();
3113 if (!Subtarget.useOddSPReg()) {
3114 // We must copy to an even-numbered MSA register so that the
3115 // single-precision sub-register is also guaranteed to be even-numbered.
3116 Wt = RegInfo.createVirtualRegister(&Mips::MSA128WEvensRegClass);
3118 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Wt).addReg(Ws);
3121 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_lo);
3123 unsigned Wt = RegInfo.createVirtualRegister(
3124 Subtarget.useOddSPReg() ? &Mips::MSA128WRegClass :
3125 &Mips::MSA128WEvensRegClass);
3127 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wt).addReg(Ws).addImm(Lane);
3128 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_lo);
3131 MI.eraseFromParent(); // The pseudo instruction is gone now.
3135 // Emit the COPY_FD pseudo instruction.
3137 // copy_fd_pseudo $fd, $ws, n
3139 // splati.d $wt, $ws, $n
3140 // copy $fd, $wt:sub_64
3142 // When n is zero, the equivalent operation can be performed with (potentially)
3143 // zero instructions due to register overlaps. This optimization is always
3144 // valid because FR=1 mode which is the only supported mode in MSA.
3146 MipsSETargetLowering::emitCOPY_FD(MachineInstr &MI,
3147 MachineBasicBlock *BB) const {
3148 assert(Subtarget.isFP64bit());
3150 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3151 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3152 unsigned Fd = MI.getOperand(0).getReg();
3153 unsigned Ws = MI.getOperand(1).getReg();
3154 unsigned Lane = MI.getOperand(2).getImm() * 2;
3155 DebugLoc DL = MI.getDebugLoc();
3158 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_64);
3160 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3162 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wt).addReg(Ws).addImm(1);
3163 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_64);
3166 MI.eraseFromParent(); // The pseudo instruction is gone now.
3170 // Emit the INSERT_FW pseudo instruction.
3172 // insert_fw_pseudo $wd, $wd_in, $n, $fs
3174 // subreg_to_reg $wt:sub_lo, $fs
3175 // insve_w $wd[$n], $wd_in, $wt[0]
3177 MipsSETargetLowering::emitINSERT_FW(MachineInstr &MI,
3178 MachineBasicBlock *BB) const {
3179 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3180 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3181 DebugLoc DL = MI.getDebugLoc();
3182 unsigned Wd = MI.getOperand(0).getReg();
3183 unsigned Wd_in = MI.getOperand(1).getReg();
3184 unsigned Lane = MI.getOperand(2).getImm();
3185 unsigned Fs = MI.getOperand(3).getReg();
3186 unsigned Wt = RegInfo.createVirtualRegister(
3187 Subtarget.useOddSPReg() ? &Mips::MSA128WRegClass :
3188 &Mips::MSA128WEvensRegClass);
3190 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
3193 .addImm(Mips::sub_lo);
3194 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_W), Wd)
3200 MI.eraseFromParent(); // The pseudo instruction is gone now.
3204 // Emit the INSERT_FD pseudo instruction.
3206 // insert_fd_pseudo $wd, $fs, n
3208 // subreg_to_reg $wt:sub_64, $fs
3209 // insve_d $wd[$n], $wd_in, $wt[0]
3211 MipsSETargetLowering::emitINSERT_FD(MachineInstr &MI,
3212 MachineBasicBlock *BB) const {
3213 assert(Subtarget.isFP64bit());
3215 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3216 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3217 DebugLoc DL = MI.getDebugLoc();
3218 unsigned Wd = MI.getOperand(0).getReg();
3219 unsigned Wd_in = MI.getOperand(1).getReg();
3220 unsigned Lane = MI.getOperand(2).getImm();
3221 unsigned Fs = MI.getOperand(3).getReg();
3222 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3224 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
3227 .addImm(Mips::sub_64);
3228 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_D), Wd)
3234 MI.eraseFromParent(); // The pseudo instruction is gone now.
3238 // Emit the INSERT_([BHWD]|F[WD])_VIDX pseudo instruction.
3241 // (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $rs)
3243 // (SLL $lanetmp1, $lane, <log2size)
3244 // (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
3245 // (INSERT_[BHWD], $wdtmp2, $wdtmp1, 0, $rs)
3246 // (NEG $lanetmp2, $lanetmp1)
3247 // (SLD_B $wd, $wdtmp2, $wdtmp2, $lanetmp2)
3249 // For floating point:
3250 // (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $fs)
3252 // (SUBREG_TO_REG $wt, $fs, <subreg>)
3253 // (SLL $lanetmp1, $lane, <log2size)
3254 // (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
3255 // (INSVE_[WD], $wdtmp2, 0, $wdtmp1, 0)
3256 // (NEG $lanetmp2, $lanetmp1)
3257 // (SLD_B $wd, $wdtmp2, $wdtmp2, $lanetmp2)
3258 MachineBasicBlock *MipsSETargetLowering::emitINSERT_DF_VIDX(
3259 MachineInstr &MI, MachineBasicBlock *BB, unsigned EltSizeInBytes,
3261 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3262 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3263 DebugLoc DL = MI.getDebugLoc();
3264 unsigned Wd = MI.getOperand(0).getReg();
3265 unsigned SrcVecReg = MI.getOperand(1).getReg();
3266 unsigned LaneReg = MI.getOperand(2).getReg();
3267 unsigned SrcValReg = MI.getOperand(3).getReg();
3269 const TargetRegisterClass *VecRC = nullptr;
3270 // FIXME: This should be true for N32 too.
3271 const TargetRegisterClass *GPRRC =
3272 Subtarget.isABI_N64() ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
3273 unsigned SubRegIdx = Subtarget.isABI_N64() ? Mips::sub_32 : 0;
3274 unsigned ShiftOp = Subtarget.isABI_N64() ? Mips::DSLL : Mips::SLL;
3275 unsigned EltLog2Size;
3276 unsigned InsertOp = 0;
3277 unsigned InsveOp = 0;
3278 switch (EltSizeInBytes) {
3280 llvm_unreachable("Unexpected size");
3283 InsertOp = Mips::INSERT_B;
3284 InsveOp = Mips::INSVE_B;
3285 VecRC = &Mips::MSA128BRegClass;
3289 InsertOp = Mips::INSERT_H;
3290 InsveOp = Mips::INSVE_H;
3291 VecRC = &Mips::MSA128HRegClass;
3295 InsertOp = Mips::INSERT_W;
3296 InsveOp = Mips::INSVE_W;
3297 VecRC = &Mips::MSA128WRegClass;
3301 InsertOp = Mips::INSERT_D;
3302 InsveOp = Mips::INSVE_D;
3303 VecRC = &Mips::MSA128DRegClass;
3308 unsigned Wt = RegInfo.createVirtualRegister(VecRC);
3309 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
3312 .addImm(EltSizeInBytes == 8 ? Mips::sub_64 : Mips::sub_lo);
3316 // Convert the lane index into a byte index
3317 if (EltSizeInBytes != 1) {
3318 unsigned LaneTmp1 = RegInfo.createVirtualRegister(GPRRC);
3319 BuildMI(*BB, MI, DL, TII->get(ShiftOp), LaneTmp1)
3321 .addImm(EltLog2Size);
3325 // Rotate bytes around so that the desired lane is element zero
3326 unsigned WdTmp1 = RegInfo.createVirtualRegister(VecRC);
3327 BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), WdTmp1)
3330 .addReg(LaneReg, 0, SubRegIdx);
3332 unsigned WdTmp2 = RegInfo.createVirtualRegister(VecRC);
3334 // Use insve.df to insert to element zero
3335 BuildMI(*BB, MI, DL, TII->get(InsveOp), WdTmp2)
3341 // Use insert.df to insert to element zero
3342 BuildMI(*BB, MI, DL, TII->get(InsertOp), WdTmp2)
3348 // Rotate elements the rest of the way for a full rotation.
3349 // sld.df inteprets $rt modulo the number of columns so we only need to negate
3350 // the lane index to do this.
3351 unsigned LaneTmp2 = RegInfo.createVirtualRegister(GPRRC);
3352 BuildMI(*BB, MI, DL, TII->get(Subtarget.isABI_N64() ? Mips::DSUB : Mips::SUB),
3354 .addReg(Subtarget.isABI_N64() ? Mips::ZERO_64 : Mips::ZERO)
3356 BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), Wd)
3359 .addReg(LaneTmp2, 0, SubRegIdx);
3361 MI.eraseFromParent(); // The pseudo instruction is gone now.
3365 // Emit the FILL_FW pseudo instruction.
3367 // fill_fw_pseudo $wd, $fs
3369 // implicit_def $wt1
3370 // insert_subreg $wt2:subreg_lo, $wt1, $fs
3371 // splati.w $wd, $wt2[0]
3373 MipsSETargetLowering::emitFILL_FW(MachineInstr &MI,
3374 MachineBasicBlock *BB) const {
3375 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3376 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3377 DebugLoc DL = MI.getDebugLoc();
3378 unsigned Wd = MI.getOperand(0).getReg();
3379 unsigned Fs = MI.getOperand(1).getReg();
3380 unsigned Wt1 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3381 unsigned Wt2 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3383 BuildMI(*BB, MI, DL, TII->get(Mips::IMPLICIT_DEF), Wt1);
3384 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2)
3387 .addImm(Mips::sub_lo);
3388 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wd).addReg(Wt2).addImm(0);
3390 MI.eraseFromParent(); // The pseudo instruction is gone now.
3394 // Emit the FILL_FD pseudo instruction.
3396 // fill_fd_pseudo $wd, $fs
3398 // implicit_def $wt1
3399 // insert_subreg $wt2:subreg_64, $wt1, $fs
3400 // splati.d $wd, $wt2[0]
3402 MipsSETargetLowering::emitFILL_FD(MachineInstr &MI,
3403 MachineBasicBlock *BB) const {
3404 assert(Subtarget.isFP64bit());
3406 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3407 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3408 DebugLoc DL = MI.getDebugLoc();
3409 unsigned Wd = MI.getOperand(0).getReg();
3410 unsigned Fs = MI.getOperand(1).getReg();
3411 unsigned Wt1 = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3412 unsigned Wt2 = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3414 BuildMI(*BB, MI, DL, TII->get(Mips::IMPLICIT_DEF), Wt1);
3415 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2)
3418 .addImm(Mips::sub_64);
3419 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wd).addReg(Wt2).addImm(0);
3421 MI.eraseFromParent(); // The pseudo instruction is gone now.
3425 // Emit the ST_F16_PSEDUO instruction to store a f16 value from an MSA
3428 // STF16 MSA128F16:$wd, mem_simm10:$addr
3430 // copy_u.h $rtemp,$wd[0]
3433 // Safety: We can't use st.h & co as they would over write the memory after
3434 // the destination. It would require half floats be allocated 16 bytes(!) of
3437 MipsSETargetLowering::emitST_F16_PSEUDO(MachineInstr &MI,
3438 MachineBasicBlock *BB) const {
3440 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3441 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3442 DebugLoc DL = MI.getDebugLoc();
3443 unsigned Ws = MI.getOperand(0).getReg();
3444 unsigned Rt = MI.getOperand(1).getReg();
3445 const MachineMemOperand &MMO = **MI.memoperands_begin();
3446 unsigned Imm = MMO.getOffset();
3448 // Caution: A load via the GOT can expand to a GPR32 operand, a load via
3449 // spill and reload can expand as a GPR64 operand. Examine the
3450 // operand in detail and default to ABI.
3451 const TargetRegisterClass *RC =
3452 MI.getOperand(1).isReg() ? RegInfo.getRegClass(MI.getOperand(1).getReg())
3453 : (Subtarget.isABI_O32() ? &Mips::GPR32RegClass
3454 : &Mips::GPR64RegClass);
3455 const bool UsingMips32 = RC == &Mips::GPR32RegClass;
3456 unsigned Rs = RegInfo.createVirtualRegister(RC);
3458 BuildMI(*BB, MI, DL, TII->get(Mips::COPY_U_H), Rs).addReg(Ws).addImm(0);
3459 BuildMI(*BB, MI, DL, TII->get(UsingMips32 ? Mips::SH : Mips::SH64))
3463 .addMemOperand(BB->getParent()->getMachineMemOperand(
3464 &MMO, MMO.getOffset(), MMO.getSize()));
3466 MI.eraseFromParent();
3470 // Emit the LD_F16_PSEDUO instruction to load a f16 value into an MSA register.
3472 // LD_F16 MSA128F16:$wd, mem_simm10:$addr
3475 // fill.h $wd, $rtemp
3477 // Safety: We can't use ld.h & co as they over-read from the source.
3478 // Additionally, if the address is not modulo 16, 2 cases can occur:
3479 // a) Segmentation fault as the load instruction reads from a memory page
3480 // memory it's not supposed to.
3481 // b) The load crosses an implementation specific boundary, requiring OS
3485 MipsSETargetLowering::emitLD_F16_PSEUDO(MachineInstr &MI,
3486 MachineBasicBlock *BB) const {
3488 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3489 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3490 DebugLoc DL = MI.getDebugLoc();
3491 unsigned Wd = MI.getOperand(0).getReg();
3493 // Caution: A load via the GOT can expand to a GPR32 operand, a load via
3494 // spill and reload can expand as a GPR64 operand. Examine the
3495 // operand in detail and default to ABI.
3496 const TargetRegisterClass *RC =
3497 MI.getOperand(1).isReg() ? RegInfo.getRegClass(MI.getOperand(1).getReg())
3498 : (Subtarget.isABI_O32() ? &Mips::GPR32RegClass
3499 : &Mips::GPR64RegClass);
3501 const bool UsingMips32 = RC == &Mips::GPR32RegClass;
3502 unsigned Rt = RegInfo.createVirtualRegister(RC);
3504 MachineInstrBuilder MIB =
3505 BuildMI(*BB, MI, DL, TII->get(UsingMips32 ? Mips::LH : Mips::LH64), Rt);
3506 for (unsigned i = 1; i < MI.getNumOperands(); i++)
3507 MIB.addOperand(MI.getOperand(i));
3509 BuildMI(*BB, MI, DL, TII->get(Mips::FILL_H), Wd).addReg(Rt);
3511 MI.eraseFromParent();
3515 // Emit the FPROUND_PSEUDO instruction.
3517 // Round an FGR64Opnd, FGR32Opnd to an f16.
3519 // Safety: Cycle the operand through the GPRs so the result always ends up
3520 // the correct MSA register.
3522 // FIXME: This copying is strictly unnecessary. If we could tie FGR32Opnd:$Fs
3523 // / FGR64Opnd:$Fs and MSA128F16:$Wd to the same physical register
3524 // (which they can be, as the MSA registers are defined to alias the
3525 // FPU's 64 bit and 32 bit registers) the result can be accessed using
3526 // the correct register class. That requires operands be tie-able across
3527 // register classes which have a sub/super register class relationship.
3531 // FPROUND MSA128F16:$wd, FGR32Opnd:$fs
3534 // fill.w $rtemp, $wtemp
3535 // fexdo.w $wd, $wtemp, $wtemp
3537 // For FPG64Opnd on mips32r2+:
3539 // FPROUND MSA128F16:$wd, FGR64Opnd:$fs
3542 // fill.w $rtemp, $wtemp
3543 // mfhc1 $rtemp2, $fs
3544 // insert.w $wtemp[1], $rtemp2
3545 // insert.w $wtemp[3], $rtemp2
3546 // fexdo.w $wtemp2, $wtemp, $wtemp
3547 // fexdo.h $wd, $temp2, $temp2
3549 // For FGR64Opnd on mips64r2+:
3551 // FPROUND MSA128F16:$wd, FGR64Opnd:$fs
3553 // dmfc1 $rtemp, $fs
3554 // fill.d $rtemp, $wtemp
3555 // fexdo.w $wtemp2, $wtemp, $wtemp
3556 // fexdo.h $wd, $wtemp2, $wtemp2
3558 // Safety note: As $wtemp is UNDEF, we may provoke a spurious exception if the
3559 // undef bits are "just right" and the exception enable bits are
3560 // set. By using fill.w to replicate $fs into all elements over
3561 // insert.w for one element, we avoid that potiential case. If
3562 // fexdo.[hw] causes an exception in, the exception is valid and it
3563 // occurs for all elements.
3566 MipsSETargetLowering::emitFPROUND_PSEUDO(MachineInstr &MI,
3567 MachineBasicBlock *BB,
3568 bool IsFGR64) const {
3570 // Strictly speaking, we need MIPS32R5 to support MSA. We'll be generous
3571 // here. It's technically doable to support MIPS32 here, but the ISA forbids
3573 assert(Subtarget.hasMSA() && Subtarget.hasMips32r2());
3575 bool IsFGR64onMips64 = Subtarget.hasMips64() && IsFGR64;
3577 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3578 DebugLoc DL = MI.getDebugLoc();
3579 unsigned Wd = MI.getOperand(0).getReg();
3580 unsigned Fs = MI.getOperand(1).getReg();
3582 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3583 unsigned Wtemp = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3584 const TargetRegisterClass *GPRRC =
3585 IsFGR64onMips64 ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
3586 unsigned MFC1Opc = IsFGR64onMips64 ? Mips::DMFC1 : Mips::MFC1;
3587 unsigned FILLOpc = IsFGR64onMips64 ? Mips::FILL_D : Mips::FILL_W;
3589 // Perform the register class copy as mentioned above.
3590 unsigned Rtemp = RegInfo.createVirtualRegister(GPRRC);
3591 BuildMI(*BB, MI, DL, TII->get(MFC1Opc), Rtemp).addReg(Fs);
3592 BuildMI(*BB, MI, DL, TII->get(FILLOpc), Wtemp).addReg(Rtemp);
3593 unsigned WPHI = Wtemp;
3595 if (!Subtarget.hasMips64() && IsFGR64) {
3596 unsigned Rtemp2 = RegInfo.createVirtualRegister(GPRRC);
3597 BuildMI(*BB, MI, DL, TII->get(Mips::MFHC1_D64), Rtemp2).addReg(Fs);
3598 unsigned Wtemp2 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3599 unsigned Wtemp3 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3600 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_W), Wtemp2)
3604 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_W), Wtemp3)
3612 unsigned Wtemp2 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3613 BuildMI(*BB, MI, DL, TII->get(Mips::FEXDO_W), Wtemp2)
3619 BuildMI(*BB, MI, DL, TII->get(Mips::FEXDO_H), Wd).addReg(WPHI).addReg(WPHI);
3621 MI.eraseFromParent();
3625 // Emit the FPEXTEND_PSEUDO instruction.
3627 // Expand an f16 to either a FGR32Opnd or FGR64Opnd.
3629 // Safety: Cycle the result through the GPRs so the result always ends up
3630 // the correct floating point register.
3632 // FIXME: This copying is strictly unnecessary. If we could tie FGR32Opnd:$Fd
3633 // / FGR64Opnd:$Fd and MSA128F16:$Ws to the same physical register
3634 // (which they can be, as the MSA registers are defined to alias the
3635 // FPU's 64 bit and 32 bit registers) the result can be accessed using
3636 // the correct register class. That requires operands be tie-able across
3637 // register classes which have a sub/super register class relationship. I
3642 // FPEXTEND FGR32Opnd:$fd, MSA128F16:$ws
3644 // fexupr.w $wtemp, $ws
3645 // copy_s.w $rtemp, $ws[0]
3648 // For FGR64Opnd on Mips64:
3650 // FPEXTEND FGR64Opnd:$fd, MSA128F16:$ws
3652 // fexupr.w $wtemp, $ws
3653 // fexupr.d $wtemp2, $wtemp
3654 // copy_s.d $rtemp, $wtemp2s[0]
3655 // dmtc1 $rtemp, $fd
3657 // For FGR64Opnd on Mips32:
3659 // FPEXTEND FGR64Opnd:$fd, MSA128F16:$ws
3661 // fexupr.w $wtemp, $ws
3662 // fexupr.d $wtemp2, $wtemp
3663 // copy_s.w $rtemp, $wtemp2[0]
3664 // mtc1 $rtemp, $ftemp
3665 // copy_s.w $rtemp2, $wtemp2[1]
3666 // $fd = mthc1 $rtemp2, $ftemp
3669 MipsSETargetLowering::emitFPEXTEND_PSEUDO(MachineInstr &MI,
3670 MachineBasicBlock *BB,
3671 bool IsFGR64) const {
3673 // Strictly speaking, we need MIPS32R5 to support MSA. We'll be generous
3674 // here. It's technically doable to support MIPS32 here, but the ISA forbids
3676 assert(Subtarget.hasMSA() && Subtarget.hasMips32r2());
3678 bool IsFGR64onMips64 = Subtarget.hasMips64() && IsFGR64;
3679 bool IsFGR64onMips32 = !Subtarget.hasMips64() && IsFGR64;
3681 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3682 DebugLoc DL = MI.getDebugLoc();
3683 unsigned Fd = MI.getOperand(0).getReg();
3684 unsigned Ws = MI.getOperand(1).getReg();
3686 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3687 const TargetRegisterClass *GPRRC =
3688 IsFGR64onMips64 ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
3689 unsigned MTC1Opc = IsFGR64onMips64 ? Mips::DMTC1 : Mips::MTC1;
3690 unsigned COPYOpc = IsFGR64onMips64 ? Mips::COPY_S_D : Mips::COPY_S_W;
3692 unsigned Wtemp = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3693 unsigned WPHI = Wtemp;
3695 BuildMI(*BB, MI, DL, TII->get(Mips::FEXUPR_W), Wtemp).addReg(Ws);
3697 WPHI = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3698 BuildMI(*BB, MI, DL, TII->get(Mips::FEXUPR_D), WPHI).addReg(Wtemp);
3701 // Perform the safety regclass copy mentioned above.
3702 unsigned Rtemp = RegInfo.createVirtualRegister(GPRRC);
3703 unsigned FPRPHI = IsFGR64onMips32
3704 ? RegInfo.createVirtualRegister(&Mips::FGR64RegClass)
3706 BuildMI(*BB, MI, DL, TII->get(COPYOpc), Rtemp).addReg(WPHI).addImm(0);
3707 BuildMI(*BB, MI, DL, TII->get(MTC1Opc), FPRPHI).addReg(Rtemp);
3709 if (IsFGR64onMips32) {
3710 unsigned Rtemp2 = RegInfo.createVirtualRegister(GPRRC);
3711 BuildMI(*BB, MI, DL, TII->get(Mips::COPY_S_W), Rtemp2)
3714 BuildMI(*BB, MI, DL, TII->get(Mips::MTHC1_D64), Fd)
3719 MI.eraseFromParent();
3723 // Emit the FEXP2_W_1 pseudo instructions.
3725 // fexp2_w_1_pseudo $wd, $wt
3728 // fexp2.w $wd, $ws, $wt
3730 MipsSETargetLowering::emitFEXP2_W_1(MachineInstr &MI,
3731 MachineBasicBlock *BB) const {
3732 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3733 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3734 const TargetRegisterClass *RC = &Mips::MSA128WRegClass;
3735 unsigned Ws1 = RegInfo.createVirtualRegister(RC);
3736 unsigned Ws2 = RegInfo.createVirtualRegister(RC);
3737 DebugLoc DL = MI.getDebugLoc();
3739 // Splat 1.0 into a vector
3740 BuildMI(*BB, MI, DL, TII->get(Mips::LDI_W), Ws1).addImm(1);
3741 BuildMI(*BB, MI, DL, TII->get(Mips::FFINT_U_W), Ws2).addReg(Ws1);
3743 // Emit 1.0 * fexp2(Wt)
3744 BuildMI(*BB, MI, DL, TII->get(Mips::FEXP2_W), MI.getOperand(0).getReg())
3746 .addReg(MI.getOperand(1).getReg());
3748 MI.eraseFromParent(); // The pseudo instruction is gone now.
3752 // Emit the FEXP2_D_1 pseudo instructions.
3754 // fexp2_d_1_pseudo $wd, $wt
3757 // fexp2.d $wd, $ws, $wt
3759 MipsSETargetLowering::emitFEXP2_D_1(MachineInstr &MI,
3760 MachineBasicBlock *BB) const {
3761 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3762 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3763 const TargetRegisterClass *RC = &Mips::MSA128DRegClass;
3764 unsigned Ws1 = RegInfo.createVirtualRegister(RC);
3765 unsigned Ws2 = RegInfo.createVirtualRegister(RC);
3766 DebugLoc DL = MI.getDebugLoc();
3768 // Splat 1.0 into a vector
3769 BuildMI(*BB, MI, DL, TII->get(Mips::LDI_D), Ws1).addImm(1);
3770 BuildMI(*BB, MI, DL, TII->get(Mips::FFINT_U_D), Ws2).addReg(Ws1);
3772 // Emit 1.0 * fexp2(Wt)
3773 BuildMI(*BB, MI, DL, TII->get(Mips::FEXP2_D), MI.getOperand(0).getReg())
3775 .addReg(MI.getOperand(1).getReg());
3777 MI.eraseFromParent(); // The pseudo instruction is gone now.