1 //===-- MipsSEISelLowering.cpp - MipsSE DAG Lowering Interface --*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Subclass of MipsTargetLowering specialized for mips32/64.
12 //===----------------------------------------------------------------------===//
13 #include "MipsSEISelLowering.h"
14 #include "MipsMachineFunction.h"
15 #include "MipsRegisterInfo.h"
16 #include "MipsTargetMachine.h"
17 #include "llvm/CodeGen/MachineInstrBuilder.h"
18 #include "llvm/CodeGen/MachineRegisterInfo.h"
19 #include "llvm/IR/Intrinsics.h"
20 #include "llvm/Support/CommandLine.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Support/raw_ostream.h"
23 #include "llvm/Target/TargetInstrInfo.h"
27 #define DEBUG_TYPE "mips-isel"
30 EnableMipsTailCalls("enable-mips-tail-calls", cl::Hidden,
31 cl::desc("MIPS: Enable tail calls."), cl::init(false));
33 static cl::opt<bool> NoDPLoadStore("mno-ldc1-sdc1", cl::init(false),
34 cl::desc("Expand double precision loads and "
35 "stores to their single precision "
38 MipsSETargetLowering::MipsSETargetLowering(const MipsTargetMachine &TM,
39 const MipsSubtarget &STI)
40 : MipsTargetLowering(TM, STI) {
41 // Set up the register classes
42 addRegisterClass(MVT::i32, &Mips::GPR32RegClass);
44 if (Subtarget.isGP64bit())
45 addRegisterClass(MVT::i64, &Mips::GPR64RegClass);
47 if (Subtarget.hasDSP() || Subtarget.hasMSA()) {
48 // Expand all truncating stores and extending loads.
49 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 setTargetDAGCombine(ISD::AND);
96 setTargetDAGCombine(ISD::OR);
97 setTargetDAGCombine(ISD::SRA);
98 setTargetDAGCombine(ISD::VSELECT);
99 setTargetDAGCombine(ISD::XOR);
102 if (!Subtarget.useSoftFloat()) {
103 addRegisterClass(MVT::f32, &Mips::FGR32RegClass);
105 // When dealing with single precision only, use libcalls
106 if (!Subtarget.isSingleFloat()) {
107 if (Subtarget.isFP64bit())
108 addRegisterClass(MVT::f64, &Mips::FGR64RegClass);
110 addRegisterClass(MVT::f64, &Mips::AFGR64RegClass);
114 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Custom);
115 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Custom);
116 setOperationAction(ISD::MULHS, MVT::i32, Custom);
117 setOperationAction(ISD::MULHU, MVT::i32, Custom);
119 if (Subtarget.hasCnMips())
120 setOperationAction(ISD::MUL, MVT::i64, Legal);
121 else if (Subtarget.isGP64bit())
122 setOperationAction(ISD::MUL, MVT::i64, Custom);
124 if (Subtarget.isGP64bit()) {
125 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Custom);
126 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Custom);
127 setOperationAction(ISD::MULHS, MVT::i64, Custom);
128 setOperationAction(ISD::MULHU, MVT::i64, Custom);
129 setOperationAction(ISD::SDIVREM, MVT::i64, Custom);
130 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
133 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
134 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
136 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
137 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
138 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
139 setOperationAction(ISD::LOAD, MVT::i32, Custom);
140 setOperationAction(ISD::STORE, MVT::i32, Custom);
142 setTargetDAGCombine(ISD::ADDE);
143 setTargetDAGCombine(ISD::SUBE);
144 setTargetDAGCombine(ISD::MUL);
146 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
147 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
148 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
151 setOperationAction(ISD::LOAD, MVT::f64, Custom);
152 setOperationAction(ISD::STORE, MVT::f64, Custom);
155 if (Subtarget.hasMips32r6()) {
156 // MIPS32r6 replaces the accumulator-based multiplies with a three register
158 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
159 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
160 setOperationAction(ISD::MUL, MVT::i32, Legal);
161 setOperationAction(ISD::MULHS, MVT::i32, Legal);
162 setOperationAction(ISD::MULHU, MVT::i32, Legal);
164 // MIPS32r6 replaces the accumulator-based division/remainder with separate
165 // three register division and remainder instructions.
166 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
167 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
168 setOperationAction(ISD::SDIV, MVT::i32, Legal);
169 setOperationAction(ISD::UDIV, MVT::i32, Legal);
170 setOperationAction(ISD::SREM, MVT::i32, Legal);
171 setOperationAction(ISD::UREM, MVT::i32, Legal);
173 // MIPS32r6 replaces conditional moves with an equivalent that removes the
174 // need for three GPR read ports.
175 setOperationAction(ISD::SETCC, MVT::i32, Legal);
176 setOperationAction(ISD::SELECT, MVT::i32, Legal);
177 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
179 setOperationAction(ISD::SETCC, MVT::f32, Legal);
180 setOperationAction(ISD::SELECT, MVT::f32, Legal);
181 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
183 assert(Subtarget.isFP64bit() && "FR=1 is required for MIPS32r6");
184 setOperationAction(ISD::SETCC, MVT::f64, Legal);
185 setOperationAction(ISD::SELECT, MVT::f64, Legal);
186 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
188 setOperationAction(ISD::BRCOND, MVT::Other, Legal);
190 // Floating point > and >= are supported via < and <=
191 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
192 setCondCodeAction(ISD::SETOGT, MVT::f32, Expand);
193 setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
194 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
196 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
197 setCondCodeAction(ISD::SETOGT, MVT::f64, Expand);
198 setCondCodeAction(ISD::SETUGE, MVT::f64, Expand);
199 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
202 if (Subtarget.hasMips64r6()) {
203 // MIPS64r6 replaces the accumulator-based multiplies with a three register
205 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
206 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
207 setOperationAction(ISD::MUL, MVT::i64, Legal);
208 setOperationAction(ISD::MULHS, MVT::i64, Legal);
209 setOperationAction(ISD::MULHU, MVT::i64, Legal);
211 // MIPS32r6 replaces the accumulator-based division/remainder with separate
212 // three register division and remainder instructions.
213 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
214 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
215 setOperationAction(ISD::SDIV, MVT::i64, Legal);
216 setOperationAction(ISD::UDIV, MVT::i64, Legal);
217 setOperationAction(ISD::SREM, MVT::i64, Legal);
218 setOperationAction(ISD::UREM, MVT::i64, Legal);
220 // MIPS64r6 replaces conditional moves with an equivalent that removes the
221 // need for three GPR read ports.
222 setOperationAction(ISD::SETCC, MVT::i64, Legal);
223 setOperationAction(ISD::SELECT, MVT::i64, Legal);
224 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
227 computeRegisterProperties(Subtarget.getRegisterInfo());
230 const MipsTargetLowering *
231 llvm::createMipsSETargetLowering(const MipsTargetMachine &TM,
232 const MipsSubtarget &STI) {
233 return new MipsSETargetLowering(TM, STI);
236 const TargetRegisterClass *
237 MipsSETargetLowering::getRepRegClassFor(MVT VT) const {
238 if (VT == MVT::Untyped)
239 return Subtarget.hasDSP() ? &Mips::ACC64DSPRegClass : &Mips::ACC64RegClass;
241 return TargetLowering::getRepRegClassFor(VT);
244 // Enable MSA support for the given integer type and Register class.
245 void MipsSETargetLowering::
246 addMSAIntType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
247 addRegisterClass(Ty, RC);
249 // Expand all builtin opcodes.
250 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
251 setOperationAction(Opc, Ty, Expand);
253 setOperationAction(ISD::BITCAST, Ty, Legal);
254 setOperationAction(ISD::LOAD, Ty, Legal);
255 setOperationAction(ISD::STORE, Ty, Legal);
256 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Custom);
257 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
258 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
260 setOperationAction(ISD::ADD, Ty, Legal);
261 setOperationAction(ISD::AND, Ty, Legal);
262 setOperationAction(ISD::CTLZ, Ty, Legal);
263 setOperationAction(ISD::CTPOP, Ty, Legal);
264 setOperationAction(ISD::MUL, Ty, Legal);
265 setOperationAction(ISD::OR, Ty, Legal);
266 setOperationAction(ISD::SDIV, Ty, Legal);
267 setOperationAction(ISD::SREM, Ty, Legal);
268 setOperationAction(ISD::SHL, Ty, Legal);
269 setOperationAction(ISD::SRA, Ty, Legal);
270 setOperationAction(ISD::SRL, Ty, Legal);
271 setOperationAction(ISD::SUB, Ty, Legal);
272 setOperationAction(ISD::UDIV, Ty, Legal);
273 setOperationAction(ISD::UREM, Ty, Legal);
274 setOperationAction(ISD::VECTOR_SHUFFLE, Ty, Custom);
275 setOperationAction(ISD::VSELECT, Ty, Legal);
276 setOperationAction(ISD::XOR, Ty, Legal);
278 if (Ty == MVT::v4i32 || Ty == MVT::v2i64) {
279 setOperationAction(ISD::FP_TO_SINT, Ty, Legal);
280 setOperationAction(ISD::FP_TO_UINT, Ty, Legal);
281 setOperationAction(ISD::SINT_TO_FP, Ty, Legal);
282 setOperationAction(ISD::UINT_TO_FP, Ty, Legal);
285 setOperationAction(ISD::SETCC, Ty, Legal);
286 setCondCodeAction(ISD::SETNE, Ty, Expand);
287 setCondCodeAction(ISD::SETGE, Ty, Expand);
288 setCondCodeAction(ISD::SETGT, Ty, Expand);
289 setCondCodeAction(ISD::SETUGE, Ty, Expand);
290 setCondCodeAction(ISD::SETUGT, Ty, Expand);
293 // Enable MSA support for the given floating-point type and Register class.
294 void MipsSETargetLowering::
295 addMSAFloatType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
296 addRegisterClass(Ty, RC);
298 // Expand all builtin opcodes.
299 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
300 setOperationAction(Opc, Ty, Expand);
302 setOperationAction(ISD::LOAD, Ty, Legal);
303 setOperationAction(ISD::STORE, Ty, Legal);
304 setOperationAction(ISD::BITCAST, Ty, Legal);
305 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Legal);
306 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
307 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
309 if (Ty != MVT::v8f16) {
310 setOperationAction(ISD::FABS, Ty, Legal);
311 setOperationAction(ISD::FADD, Ty, Legal);
312 setOperationAction(ISD::FDIV, Ty, Legal);
313 setOperationAction(ISD::FEXP2, Ty, Legal);
314 setOperationAction(ISD::FLOG2, Ty, Legal);
315 setOperationAction(ISD::FMA, Ty, Legal);
316 setOperationAction(ISD::FMUL, Ty, Legal);
317 setOperationAction(ISD::FRINT, Ty, Legal);
318 setOperationAction(ISD::FSQRT, Ty, Legal);
319 setOperationAction(ISD::FSUB, Ty, Legal);
320 setOperationAction(ISD::VSELECT, Ty, Legal);
322 setOperationAction(ISD::SETCC, Ty, Legal);
323 setCondCodeAction(ISD::SETOGE, Ty, Expand);
324 setCondCodeAction(ISD::SETOGT, Ty, Expand);
325 setCondCodeAction(ISD::SETUGE, Ty, Expand);
326 setCondCodeAction(ISD::SETUGT, Ty, Expand);
327 setCondCodeAction(ISD::SETGE, Ty, Expand);
328 setCondCodeAction(ISD::SETGT, Ty, Expand);
333 MipsSETargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
337 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
339 if (Subtarget.systemSupportsUnalignedAccess()) {
340 // MIPS32r6/MIPS64r6 is required to support unaligned access. It's
341 // implementation defined whether this is handled by hardware, software, or
342 // a hybrid of the two but it's expected that most implementations will
343 // handle the majority of cases in hardware.
360 SDValue MipsSETargetLowering::LowerOperation(SDValue Op,
361 SelectionDAG &DAG) const {
362 switch(Op.getOpcode()) {
363 case ISD::LOAD: return lowerLOAD(Op, DAG);
364 case ISD::STORE: return lowerSTORE(Op, DAG);
365 case ISD::SMUL_LOHI: return lowerMulDiv(Op, MipsISD::Mult, true, true, DAG);
366 case ISD::UMUL_LOHI: return lowerMulDiv(Op, MipsISD::Multu, true, true, DAG);
367 case ISD::MULHS: return lowerMulDiv(Op, MipsISD::Mult, false, true, DAG);
368 case ISD::MULHU: return lowerMulDiv(Op, MipsISD::Multu, false, true, DAG);
369 case ISD::MUL: return lowerMulDiv(Op, MipsISD::Mult, true, false, DAG);
370 case ISD::SDIVREM: return lowerMulDiv(Op, MipsISD::DivRem, true, true, DAG);
371 case ISD::UDIVREM: return lowerMulDiv(Op, MipsISD::DivRemU, true, true,
373 case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG);
374 case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG);
375 case ISD::INTRINSIC_VOID: return lowerINTRINSIC_VOID(Op, DAG);
376 case ISD::EXTRACT_VECTOR_ELT: return lowerEXTRACT_VECTOR_ELT(Op, DAG);
377 case ISD::BUILD_VECTOR: return lowerBUILD_VECTOR(Op, DAG);
378 case ISD::VECTOR_SHUFFLE: return lowerVECTOR_SHUFFLE(Op, DAG);
381 return MipsTargetLowering::LowerOperation(Op, DAG);
385 // Transforms a subgraph in CurDAG if the following pattern is found:
386 // (addc multLo, Lo0), (adde multHi, Hi0),
388 // multHi/Lo: product of multiplication
389 // Lo0: initial value of Lo register
390 // Hi0: initial value of Hi register
391 // Return true if pattern matching was successful.
392 static bool selectMADD(SDNode *ADDENode, SelectionDAG *CurDAG) {
393 // ADDENode's second operand must be a flag output of an ADDC node in order
394 // for the matching to be successful.
395 SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
397 if (ADDCNode->getOpcode() != ISD::ADDC)
400 SDValue MultHi = ADDENode->getOperand(0);
401 SDValue MultLo = ADDCNode->getOperand(0);
402 SDNode *MultNode = MultHi.getNode();
403 unsigned MultOpc = MultHi.getOpcode();
405 // MultHi and MultLo must be generated by the same node,
406 if (MultLo.getNode() != MultNode)
409 // and it must be a multiplication.
410 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
413 // MultLo amd MultHi must be the first and second output of MultNode
415 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
418 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
419 // of the values of MultNode, in which case MultNode will be removed in later
421 // If there exist users other than ADDENode or ADDCNode, this function returns
422 // here, which will result in MultNode being mapped to a single MULT
423 // instruction node rather than a pair of MULT and MADD instructions being
425 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
430 // Initialize accumulator.
431 SDValue ACCIn = CurDAG->getNode(MipsISD::MTLOHI, DL, MVT::Untyped,
432 ADDCNode->getOperand(1),
433 ADDENode->getOperand(1));
435 // create MipsMAdd(u) node
436 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
438 SDValue MAdd = CurDAG->getNode(MultOpc, DL, MVT::Untyped,
439 MultNode->getOperand(0),// Factor 0
440 MultNode->getOperand(1),// Factor 1
443 // replace uses of adde and addc here
444 if (!SDValue(ADDCNode, 0).use_empty()) {
445 SDValue LoOut = CurDAG->getNode(MipsISD::MFLO, DL, MVT::i32, MAdd);
446 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), LoOut);
448 if (!SDValue(ADDENode, 0).use_empty()) {
449 SDValue HiOut = CurDAG->getNode(MipsISD::MFHI, DL, MVT::i32, MAdd);
450 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), HiOut);
457 // Transforms a subgraph in CurDAG if the following pattern is found:
458 // (addc Lo0, multLo), (sube Hi0, multHi),
460 // multHi/Lo: product of multiplication
461 // Lo0: initial value of Lo register
462 // Hi0: initial value of Hi register
463 // Return true if pattern matching was successful.
464 static bool selectMSUB(SDNode *SUBENode, SelectionDAG *CurDAG) {
465 // SUBENode's second operand must be a flag output of an SUBC node in order
466 // for the matching to be successful.
467 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
469 if (SUBCNode->getOpcode() != ISD::SUBC)
472 SDValue MultHi = SUBENode->getOperand(1);
473 SDValue MultLo = SUBCNode->getOperand(1);
474 SDNode *MultNode = MultHi.getNode();
475 unsigned MultOpc = MultHi.getOpcode();
477 // MultHi and MultLo must be generated by the same node,
478 if (MultLo.getNode() != MultNode)
481 // and it must be a multiplication.
482 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
485 // MultLo amd MultHi must be the first and second output of MultNode
487 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
490 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
491 // of the values of MultNode, in which case MultNode will be removed in later
493 // If there exist users other than SUBENode or SUBCNode, this function returns
494 // here, which will result in MultNode being mapped to a single MULT
495 // instruction node rather than a pair of MULT and MSUB instructions being
497 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
502 // Initialize accumulator.
503 SDValue ACCIn = CurDAG->getNode(MipsISD::MTLOHI, DL, MVT::Untyped,
504 SUBCNode->getOperand(0),
505 SUBENode->getOperand(0));
507 // create MipsSub(u) node
508 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
510 SDValue MSub = CurDAG->getNode(MultOpc, DL, MVT::Glue,
511 MultNode->getOperand(0),// Factor 0
512 MultNode->getOperand(1),// Factor 1
515 // replace uses of sube and subc here
516 if (!SDValue(SUBCNode, 0).use_empty()) {
517 SDValue LoOut = CurDAG->getNode(MipsISD::MFLO, DL, MVT::i32, MSub);
518 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), LoOut);
520 if (!SDValue(SUBENode, 0).use_empty()) {
521 SDValue HiOut = CurDAG->getNode(MipsISD::MFHI, DL, MVT::i32, MSub);
522 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), HiOut);
528 static SDValue performADDECombine(SDNode *N, SelectionDAG &DAG,
529 TargetLowering::DAGCombinerInfo &DCI,
530 const MipsSubtarget &Subtarget) {
531 if (DCI.isBeforeLegalize())
534 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() &&
535 N->getValueType(0) == MVT::i32 && selectMADD(N, &DAG))
536 return SDValue(N, 0);
541 // Fold zero extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT
543 // Performs the following transformations:
544 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to zero extension if its
545 // sign/zero-extension is completely overwritten by the new one performed by
547 // - Removes redundant zero extensions performed by an ISD::AND.
548 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
549 TargetLowering::DAGCombinerInfo &DCI,
550 const MipsSubtarget &Subtarget) {
551 if (!Subtarget.hasMSA())
554 SDValue Op0 = N->getOperand(0);
555 SDValue Op1 = N->getOperand(1);
556 unsigned Op0Opcode = Op0->getOpcode();
558 // (and (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d)
559 // where $d + 1 == 2^n and n == 32
560 // or $d + 1 == 2^n and n <= 32 and ZExt
561 // -> (MipsVExtractZExt $a, $b, $c)
562 if (Op0Opcode == MipsISD::VEXTRACT_SEXT_ELT ||
563 Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT) {
564 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(Op1);
569 int32_t Log2IfPositive = (Mask->getAPIntValue() + 1).exactLogBase2();
571 if (Log2IfPositive <= 0)
572 return SDValue(); // Mask+1 is not a power of 2
574 SDValue Op0Op2 = Op0->getOperand(2);
575 EVT ExtendTy = cast<VTSDNode>(Op0Op2)->getVT();
576 unsigned ExtendTySize = ExtendTy.getSizeInBits();
577 unsigned Log2 = Log2IfPositive;
579 if ((Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT && Log2 >= ExtendTySize) ||
580 Log2 == ExtendTySize) {
581 SDValue Ops[] = { Op0->getOperand(0), Op0->getOperand(1), Op0Op2 };
582 return DAG.getNode(MipsISD::VEXTRACT_ZEXT_ELT, SDLoc(Op0),
584 makeArrayRef(Ops, Op0->getNumOperands()));
591 // Determine if the specified node is a constant vector splat.
593 // Returns true and sets Imm if:
594 // * N is a ISD::BUILD_VECTOR representing a constant splat
596 // This function is quite similar to MipsSEDAGToDAGISel::selectVSplat. The
597 // differences are that it assumes the MSA has already been checked and the
598 // arbitrary requirement for a maximum of 32-bit integers isn't applied (and
599 // must not be in order for binsri.d to be selectable).
600 static bool isVSplat(SDValue N, APInt &Imm, bool IsLittleEndian) {
601 BuildVectorSDNode *Node = dyn_cast<BuildVectorSDNode>(N.getNode());
606 APInt SplatValue, SplatUndef;
607 unsigned SplatBitSize;
610 if (!Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
619 // Test whether the given node is an all-ones build_vector.
620 static bool isVectorAllOnes(SDValue N) {
621 // Look through bitcasts. Endianness doesn't matter because we are looking
622 // for an all-ones value.
623 if (N->getOpcode() == ISD::BITCAST)
624 N = N->getOperand(0);
626 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
631 APInt SplatValue, SplatUndef;
632 unsigned SplatBitSize;
635 // Endianness doesn't matter in this context because we are looking for
636 // an all-ones value.
637 if (BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs))
638 return SplatValue.isAllOnesValue();
643 // Test whether N is the bitwise inverse of OfNode.
644 static bool isBitwiseInverse(SDValue N, SDValue OfNode) {
645 if (N->getOpcode() != ISD::XOR)
648 if (isVectorAllOnes(N->getOperand(0)))
649 return N->getOperand(1) == OfNode;
651 if (isVectorAllOnes(N->getOperand(1)))
652 return N->getOperand(0) == OfNode;
657 // Perform combines where ISD::OR is the root node.
659 // Performs the following transformations:
660 // - (or (and $a, $mask), (and $b, $inv_mask)) => (vselect $mask, $a, $b)
661 // where $inv_mask is the bitwise inverse of $mask and the 'or' has a 128-bit
663 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
664 TargetLowering::DAGCombinerInfo &DCI,
665 const MipsSubtarget &Subtarget) {
666 if (!Subtarget.hasMSA())
669 EVT Ty = N->getValueType(0);
671 if (!Ty.is128BitVector())
674 SDValue Op0 = N->getOperand(0);
675 SDValue Op1 = N->getOperand(1);
677 if (Op0->getOpcode() == ISD::AND && Op1->getOpcode() == ISD::AND) {
678 SDValue Op0Op0 = Op0->getOperand(0);
679 SDValue Op0Op1 = Op0->getOperand(1);
680 SDValue Op1Op0 = Op1->getOperand(0);
681 SDValue Op1Op1 = Op1->getOperand(1);
682 bool IsLittleEndian = !Subtarget.isLittle();
684 SDValue IfSet, IfClr, Cond;
685 bool IsConstantMask = false;
688 // If Op0Op0 is an appropriate mask, try to find it's inverse in either
689 // Op1Op0, or Op1Op1. Keep track of the Cond, IfSet, and IfClr nodes, while
691 // IfClr will be set if we find a valid match.
692 if (isVSplat(Op0Op0, Mask, IsLittleEndian)) {
696 if (isVSplat(Op1Op0, InvMask, IsLittleEndian) &&
697 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
699 else if (isVSplat(Op1Op1, InvMask, IsLittleEndian) &&
700 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
703 IsConstantMask = true;
706 // If IfClr is not yet set, and Op0Op1 is an appropriate mask, try the same
707 // thing again using this mask.
708 // IfClr will be set if we find a valid match.
709 if (!IfClr.getNode() && isVSplat(Op0Op1, Mask, IsLittleEndian)) {
713 if (isVSplat(Op1Op0, InvMask, IsLittleEndian) &&
714 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
716 else if (isVSplat(Op1Op1, InvMask, IsLittleEndian) &&
717 Mask.getBitWidth() == InvMask.getBitWidth() && Mask == ~InvMask)
720 IsConstantMask = true;
723 // If IfClr is not yet set, try looking for a non-constant match.
724 // IfClr will be set if we find a valid match amongst the eight
726 if (!IfClr.getNode()) {
727 if (isBitwiseInverse(Op0Op0, Op1Op0)) {
731 } else if (isBitwiseInverse(Op0Op1, Op1Op0)) {
735 } else if (isBitwiseInverse(Op0Op0, Op1Op1)) {
739 } else if (isBitwiseInverse(Op0Op1, Op1Op1)) {
743 } else if (isBitwiseInverse(Op1Op0, Op0Op0)) {
747 } else if (isBitwiseInverse(Op1Op1, Op0Op0)) {
751 } else if (isBitwiseInverse(Op1Op0, Op0Op1)) {
755 } else if (isBitwiseInverse(Op1Op1, Op0Op1)) {
762 // At this point, IfClr will be set if we have a valid match.
763 if (!IfClr.getNode())
766 assert(Cond.getNode() && IfSet.getNode());
768 // Fold degenerate cases.
769 if (IsConstantMask) {
770 if (Mask.isAllOnesValue())
776 // Transform the DAG into an equivalent VSELECT.
777 return DAG.getNode(ISD::VSELECT, SDLoc(N), Ty, Cond, IfSet, IfClr);
783 static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG,
784 TargetLowering::DAGCombinerInfo &DCI,
785 const MipsSubtarget &Subtarget) {
786 if (DCI.isBeforeLegalize())
789 if (Subtarget.hasMips32() && N->getValueType(0) == MVT::i32 &&
791 return SDValue(N, 0);
796 static SDValue genConstMult(SDValue X, uint64_t C, SDLoc DL, EVT VT,
797 EVT ShiftTy, SelectionDAG &DAG) {
798 // Clear the upper (64 - VT.sizeInBits) bits.
799 C &= ((uint64_t)-1) >> (64 - VT.getSizeInBits());
803 return DAG.getConstant(0, DL, VT);
809 // If c is power of 2, return (shl x, log2(c)).
810 if (isPowerOf2_64(C))
811 return DAG.getNode(ISD::SHL, DL, VT, X,
812 DAG.getConstant(Log2_64(C), DL, ShiftTy));
814 unsigned Log2Ceil = Log2_64_Ceil(C);
815 uint64_t Floor = 1LL << Log2_64(C);
816 uint64_t Ceil = Log2Ceil == 64 ? 0LL : 1LL << Log2Ceil;
818 // If |c - floor_c| <= |c - ceil_c|,
819 // where floor_c = pow(2, floor(log2(c))) and ceil_c = pow(2, ceil(log2(c))),
820 // return (add constMult(x, floor_c), constMult(x, c - floor_c)).
821 if (C - Floor <= Ceil - C) {
822 SDValue Op0 = genConstMult(X, Floor, DL, VT, ShiftTy, DAG);
823 SDValue Op1 = genConstMult(X, C - Floor, DL, VT, ShiftTy, DAG);
824 return DAG.getNode(ISD::ADD, DL, VT, Op0, Op1);
827 // If |c - floor_c| > |c - ceil_c|,
828 // return (sub constMult(x, ceil_c), constMult(x, ceil_c - c)).
829 SDValue Op0 = genConstMult(X, Ceil, DL, VT, ShiftTy, DAG);
830 SDValue Op1 = genConstMult(X, Ceil - C, DL, VT, ShiftTy, DAG);
831 return DAG.getNode(ISD::SUB, DL, VT, Op0, Op1);
834 static SDValue performMULCombine(SDNode *N, SelectionDAG &DAG,
835 const TargetLowering::DAGCombinerInfo &DCI,
836 const MipsSETargetLowering *TL) {
837 EVT VT = N->getValueType(0);
839 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
841 return genConstMult(N->getOperand(0), C->getZExtValue(), SDLoc(N), VT,
842 TL->getScalarShiftAmountTy(DAG.getDataLayout(), VT),
845 return SDValue(N, 0);
848 static SDValue performDSPShiftCombine(unsigned Opc, SDNode *N, EVT Ty,
850 const MipsSubtarget &Subtarget) {
851 // See if this is a vector splat immediate node.
852 APInt SplatValue, SplatUndef;
853 unsigned SplatBitSize;
855 unsigned EltSize = Ty.getVectorElementType().getSizeInBits();
856 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
858 if (!Subtarget.hasDSP())
862 !BV->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
863 EltSize, !Subtarget.isLittle()) ||
864 (SplatBitSize != EltSize) ||
865 (SplatValue.getZExtValue() >= EltSize))
869 return DAG.getNode(Opc, DL, Ty, N->getOperand(0),
870 DAG.getConstant(SplatValue.getZExtValue(), DL, MVT::i32));
873 static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
874 TargetLowering::DAGCombinerInfo &DCI,
875 const MipsSubtarget &Subtarget) {
876 EVT Ty = N->getValueType(0);
878 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
881 return performDSPShiftCombine(MipsISD::SHLL_DSP, N, Ty, DAG, Subtarget);
884 // Fold sign-extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT for MSA and fold
885 // constant splats into MipsISD::SHRA_DSP for DSPr2.
887 // Performs the following transformations:
888 // - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to sign extension if its
889 // sign/zero-extension is completely overwritten by the new one performed by
890 // the ISD::SRA and ISD::SHL nodes.
891 // - Removes redundant sign extensions performed by an ISD::SRA and ISD::SHL
894 // See performDSPShiftCombine for more information about the transformation
896 static SDValue performSRACombine(SDNode *N, SelectionDAG &DAG,
897 TargetLowering::DAGCombinerInfo &DCI,
898 const MipsSubtarget &Subtarget) {
899 EVT Ty = N->getValueType(0);
901 if (Subtarget.hasMSA()) {
902 SDValue Op0 = N->getOperand(0);
903 SDValue Op1 = N->getOperand(1);
905 // (sra (shl (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d), imm:$d)
906 // where $d + sizeof($c) == 32
907 // or $d + sizeof($c) <= 32 and SExt
908 // -> (MipsVExtractSExt $a, $b, $c)
909 if (Op0->getOpcode() == ISD::SHL && Op1 == Op0->getOperand(1)) {
910 SDValue Op0Op0 = Op0->getOperand(0);
911 ConstantSDNode *ShAmount = dyn_cast<ConstantSDNode>(Op1);
916 if (Op0Op0->getOpcode() != MipsISD::VEXTRACT_SEXT_ELT &&
917 Op0Op0->getOpcode() != MipsISD::VEXTRACT_ZEXT_ELT)
920 EVT ExtendTy = cast<VTSDNode>(Op0Op0->getOperand(2))->getVT();
921 unsigned TotalBits = ShAmount->getZExtValue() + ExtendTy.getSizeInBits();
923 if (TotalBits == 32 ||
924 (Op0Op0->getOpcode() == MipsISD::VEXTRACT_SEXT_ELT &&
926 SDValue Ops[] = { Op0Op0->getOperand(0), Op0Op0->getOperand(1),
927 Op0Op0->getOperand(2) };
928 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, SDLoc(Op0Op0),
930 makeArrayRef(Ops, Op0Op0->getNumOperands()));
935 if ((Ty != MVT::v2i16) && ((Ty != MVT::v4i8) || !Subtarget.hasDSPR2()))
938 return performDSPShiftCombine(MipsISD::SHRA_DSP, N, Ty, DAG, Subtarget);
942 static SDValue performSRLCombine(SDNode *N, SelectionDAG &DAG,
943 TargetLowering::DAGCombinerInfo &DCI,
944 const MipsSubtarget &Subtarget) {
945 EVT Ty = N->getValueType(0);
947 if (((Ty != MVT::v2i16) || !Subtarget.hasDSPR2()) && (Ty != MVT::v4i8))
950 return performDSPShiftCombine(MipsISD::SHRL_DSP, N, Ty, DAG, Subtarget);
953 static bool isLegalDSPCondCode(EVT Ty, ISD::CondCode CC) {
954 bool IsV216 = (Ty == MVT::v2i16);
958 case ISD::SETNE: return true;
962 case ISD::SETGE: return IsV216;
966 case ISD::SETUGE: return !IsV216;
967 default: return false;
971 static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG) {
972 EVT Ty = N->getValueType(0);
974 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
977 if (!isLegalDSPCondCode(Ty, cast<CondCodeSDNode>(N->getOperand(2))->get()))
980 return DAG.getNode(MipsISD::SETCC_DSP, SDLoc(N), Ty, N->getOperand(0),
981 N->getOperand(1), N->getOperand(2));
984 static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG) {
985 EVT Ty = N->getValueType(0);
987 if (Ty.is128BitVector() && Ty.isInteger()) {
988 // Try the following combines:
989 // (vselect (setcc $a, $b, SETLT), $b, $a)) -> (vsmax $a, $b)
990 // (vselect (setcc $a, $b, SETLE), $b, $a)) -> (vsmax $a, $b)
991 // (vselect (setcc $a, $b, SETLT), $a, $b)) -> (vsmin $a, $b)
992 // (vselect (setcc $a, $b, SETLE), $a, $b)) -> (vsmin $a, $b)
993 // (vselect (setcc $a, $b, SETULT), $b, $a)) -> (vumax $a, $b)
994 // (vselect (setcc $a, $b, SETULE), $b, $a)) -> (vumax $a, $b)
995 // (vselect (setcc $a, $b, SETULT), $a, $b)) -> (vumin $a, $b)
996 // (vselect (setcc $a, $b, SETULE), $a, $b)) -> (vumin $a, $b)
997 // SETGT/SETGE/SETUGT/SETUGE variants of these will show up initially but
998 // will be expanded to equivalent SETLT/SETLE/SETULT/SETULE versions by the
1000 SDValue Op0 = N->getOperand(0);
1002 if (Op0->getOpcode() != ISD::SETCC)
1005 ISD::CondCode CondCode = cast<CondCodeSDNode>(Op0->getOperand(2))->get();
1008 if (CondCode == ISD::SETLT || CondCode == ISD::SETLE)
1010 else if (CondCode == ISD::SETULT || CondCode == ISD::SETULE)
1015 SDValue Op1 = N->getOperand(1);
1016 SDValue Op2 = N->getOperand(2);
1017 SDValue Op0Op0 = Op0->getOperand(0);
1018 SDValue Op0Op1 = Op0->getOperand(1);
1020 if (Op1 == Op0Op0 && Op2 == Op0Op1)
1021 return DAG.getNode(Signed ? MipsISD::VSMIN : MipsISD::VUMIN, SDLoc(N),
1023 else if (Op1 == Op0Op1 && Op2 == Op0Op0)
1024 return DAG.getNode(Signed ? MipsISD::VSMAX : MipsISD::VUMAX, SDLoc(N),
1026 } else if ((Ty == MVT::v2i16) || (Ty == MVT::v4i8)) {
1027 SDValue SetCC = N->getOperand(0);
1029 if (SetCC.getOpcode() != MipsISD::SETCC_DSP)
1032 return DAG.getNode(MipsISD::SELECT_CC_DSP, SDLoc(N), Ty,
1033 SetCC.getOperand(0), SetCC.getOperand(1),
1034 N->getOperand(1), N->getOperand(2), SetCC.getOperand(2));
1040 static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
1041 const MipsSubtarget &Subtarget) {
1042 EVT Ty = N->getValueType(0);
1044 if (Subtarget.hasMSA() && Ty.is128BitVector() && Ty.isInteger()) {
1045 // Try the following combines:
1046 // (xor (or $a, $b), (build_vector allones))
1047 // (xor (or $a, $b), (bitcast (build_vector allones)))
1048 SDValue Op0 = N->getOperand(0);
1049 SDValue Op1 = N->getOperand(1);
1052 if (ISD::isBuildVectorAllOnes(Op0.getNode()))
1054 else if (ISD::isBuildVectorAllOnes(Op1.getNode()))
1059 if (NotOp->getOpcode() == ISD::OR)
1060 return DAG.getNode(MipsISD::VNOR, SDLoc(N), Ty, NotOp->getOperand(0),
1061 NotOp->getOperand(1));
1068 MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1069 SelectionDAG &DAG = DCI.DAG;
1072 switch (N->getOpcode()) {
1074 return performADDECombine(N, DAG, DCI, Subtarget);
1076 Val = performANDCombine(N, DAG, DCI, Subtarget);
1079 Val = performORCombine(N, DAG, DCI, Subtarget);
1082 return performSUBECombine(N, DAG, DCI, Subtarget);
1084 return performMULCombine(N, DAG, DCI, this);
1086 return performSHLCombine(N, DAG, DCI, Subtarget);
1088 return performSRACombine(N, DAG, DCI, Subtarget);
1090 return performSRLCombine(N, DAG, DCI, Subtarget);
1092 return performVSELECTCombine(N, DAG);
1094 Val = performXORCombine(N, DAG, Subtarget);
1097 Val = performSETCCCombine(N, DAG);
1101 if (Val.getNode()) {
1102 DEBUG(dbgs() << "\nMipsSE DAG Combine:\n";
1103 N->printrWithDepth(dbgs(), &DAG);
1104 dbgs() << "\n=> \n";
1105 Val.getNode()->printrWithDepth(dbgs(), &DAG);
1110 return MipsTargetLowering::PerformDAGCombine(N, DCI);
1114 MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1115 MachineBasicBlock *BB) const {
1116 switch (MI->getOpcode()) {
1118 return MipsTargetLowering::EmitInstrWithCustomInserter(MI, BB);
1119 case Mips::BPOSGE32_PSEUDO:
1120 return emitBPOSGE32(MI, BB);
1121 case Mips::SNZ_B_PSEUDO:
1122 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_B);
1123 case Mips::SNZ_H_PSEUDO:
1124 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_H);
1125 case Mips::SNZ_W_PSEUDO:
1126 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_W);
1127 case Mips::SNZ_D_PSEUDO:
1128 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_D);
1129 case Mips::SNZ_V_PSEUDO:
1130 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_V);
1131 case Mips::SZ_B_PSEUDO:
1132 return emitMSACBranchPseudo(MI, BB, Mips::BZ_B);
1133 case Mips::SZ_H_PSEUDO:
1134 return emitMSACBranchPseudo(MI, BB, Mips::BZ_H);
1135 case Mips::SZ_W_PSEUDO:
1136 return emitMSACBranchPseudo(MI, BB, Mips::BZ_W);
1137 case Mips::SZ_D_PSEUDO:
1138 return emitMSACBranchPseudo(MI, BB, Mips::BZ_D);
1139 case Mips::SZ_V_PSEUDO:
1140 return emitMSACBranchPseudo(MI, BB, Mips::BZ_V);
1141 case Mips::COPY_FW_PSEUDO:
1142 return emitCOPY_FW(MI, BB);
1143 case Mips::COPY_FD_PSEUDO:
1144 return emitCOPY_FD(MI, BB);
1145 case Mips::INSERT_FW_PSEUDO:
1146 return emitINSERT_FW(MI, BB);
1147 case Mips::INSERT_FD_PSEUDO:
1148 return emitINSERT_FD(MI, BB);
1149 case Mips::INSERT_B_VIDX_PSEUDO:
1150 case Mips::INSERT_B_VIDX64_PSEUDO:
1151 return emitINSERT_DF_VIDX(MI, BB, 1, false);
1152 case Mips::INSERT_H_VIDX_PSEUDO:
1153 case Mips::INSERT_H_VIDX64_PSEUDO:
1154 return emitINSERT_DF_VIDX(MI, BB, 2, false);
1155 case Mips::INSERT_W_VIDX_PSEUDO:
1156 case Mips::INSERT_W_VIDX64_PSEUDO:
1157 return emitINSERT_DF_VIDX(MI, BB, 4, false);
1158 case Mips::INSERT_D_VIDX_PSEUDO:
1159 case Mips::INSERT_D_VIDX64_PSEUDO:
1160 return emitINSERT_DF_VIDX(MI, BB, 8, false);
1161 case Mips::INSERT_FW_VIDX_PSEUDO:
1162 case Mips::INSERT_FW_VIDX64_PSEUDO:
1163 return emitINSERT_DF_VIDX(MI, BB, 4, true);
1164 case Mips::INSERT_FD_VIDX_PSEUDO:
1165 case Mips::INSERT_FD_VIDX64_PSEUDO:
1166 return emitINSERT_DF_VIDX(MI, BB, 8, true);
1167 case Mips::FILL_FW_PSEUDO:
1168 return emitFILL_FW(MI, BB);
1169 case Mips::FILL_FD_PSEUDO:
1170 return emitFILL_FD(MI, BB);
1171 case Mips::FEXP2_W_1_PSEUDO:
1172 return emitFEXP2_W_1(MI, BB);
1173 case Mips::FEXP2_D_1_PSEUDO:
1174 return emitFEXP2_D_1(MI, BB);
1178 bool MipsSETargetLowering::isEligibleForTailCallOptimization(
1179 const CCState &CCInfo, unsigned NextStackOffset,
1180 const MipsFunctionInfo &FI) const {
1181 if (!EnableMipsTailCalls)
1184 // Return false if either the callee or caller has a byval argument.
1185 if (CCInfo.getInRegsParamsCount() > 0 || FI.hasByvalArg())
1188 // Return true if the callee's argument area is no larger than the
1190 return NextStackOffset <= FI.getIncomingArgSize();
1193 void MipsSETargetLowering::
1194 getOpndList(SmallVectorImpl<SDValue> &Ops,
1195 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
1196 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
1197 bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee,
1198 SDValue Chain) const {
1199 Ops.push_back(Callee);
1200 MipsTargetLowering::getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal,
1201 InternalLinkage, IsCallReloc, CLI, Callee,
1205 SDValue MipsSETargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1206 LoadSDNode &Nd = *cast<LoadSDNode>(Op);
1208 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
1209 return MipsTargetLowering::lowerLOAD(Op, DAG);
1211 // Replace a double precision load with two i32 loads and a buildpair64.
1213 SDValue Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
1214 EVT PtrVT = Ptr.getValueType();
1216 // i32 load from lower address.
1217 SDValue Lo = DAG.getLoad(MVT::i32, DL, Chain, Ptr,
1218 MachinePointerInfo(), Nd.isVolatile(),
1219 Nd.isNonTemporal(), Nd.isInvariant(),
1222 // i32 load from higher address.
1223 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, DL, PtrVT));
1224 SDValue Hi = DAG.getLoad(MVT::i32, DL, Lo.getValue(1), Ptr,
1225 MachinePointerInfo(), Nd.isVolatile(),
1226 Nd.isNonTemporal(), Nd.isInvariant(),
1227 std::min(Nd.getAlignment(), 4U));
1229 if (!Subtarget.isLittle())
1232 SDValue BP = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
1233 SDValue Ops[2] = {BP, Hi.getValue(1)};
1234 return DAG.getMergeValues(Ops, DL);
1237 SDValue MipsSETargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1238 StoreSDNode &Nd = *cast<StoreSDNode>(Op);
1240 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore)
1241 return MipsTargetLowering::lowerSTORE(Op, DAG);
1243 // Replace a double precision store with two extractelement64s and i32 stores.
1245 SDValue Val = Nd.getValue(), Ptr = Nd.getBasePtr(), Chain = Nd.getChain();
1246 EVT PtrVT = Ptr.getValueType();
1247 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1248 Val, DAG.getConstant(0, DL, MVT::i32));
1249 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1250 Val, DAG.getConstant(1, DL, MVT::i32));
1252 if (!Subtarget.isLittle())
1255 // i32 store to lower address.
1256 Chain = DAG.getStore(Chain, DL, Lo, Ptr, MachinePointerInfo(),
1257 Nd.isVolatile(), Nd.isNonTemporal(), Nd.getAlignment(),
1260 // i32 store to higher address.
1261 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, DL, PtrVT));
1262 return DAG.getStore(Chain, DL, Hi, Ptr, MachinePointerInfo(),
1263 Nd.isVolatile(), Nd.isNonTemporal(),
1264 std::min(Nd.getAlignment(), 4U), Nd.getAAInfo());
1267 SDValue MipsSETargetLowering::lowerMulDiv(SDValue Op, unsigned NewOpc,
1268 bool HasLo, bool HasHi,
1269 SelectionDAG &DAG) const {
1270 // MIPS32r6/MIPS64r6 removed accumulator based multiplies.
1271 assert(!Subtarget.hasMips32r6());
1273 EVT Ty = Op.getOperand(0).getValueType();
1275 SDValue Mult = DAG.getNode(NewOpc, DL, MVT::Untyped,
1276 Op.getOperand(0), Op.getOperand(1));
1280 Lo = DAG.getNode(MipsISD::MFLO, DL, Ty, Mult);
1282 Hi = DAG.getNode(MipsISD::MFHI, DL, Ty, Mult);
1284 if (!HasLo || !HasHi)
1285 return HasLo ? Lo : Hi;
1287 SDValue Vals[] = { Lo, Hi };
1288 return DAG.getMergeValues(Vals, DL);
1292 static SDValue initAccumulator(SDValue In, SDLoc DL, SelectionDAG &DAG) {
1293 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
1294 DAG.getConstant(0, DL, MVT::i32));
1295 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In,
1296 DAG.getConstant(1, DL, MVT::i32));
1297 return DAG.getNode(MipsISD::MTLOHI, DL, MVT::Untyped, InLo, InHi);
1300 static SDValue extractLOHI(SDValue Op, SDLoc DL, SelectionDAG &DAG) {
1301 SDValue Lo = DAG.getNode(MipsISD::MFLO, DL, MVT::i32, Op);
1302 SDValue Hi = DAG.getNode(MipsISD::MFHI, DL, MVT::i32, Op);
1303 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi);
1306 // This function expands mips intrinsic nodes which have 64-bit input operands
1307 // or output values.
1309 // out64 = intrinsic-node in64
1311 // lo = copy (extract-element (in64, 0))
1312 // hi = copy (extract-element (in64, 1))
1313 // mips-specific-node
1316 // out64 = merge-values (v0, v1)
1318 static SDValue lowerDSPIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1320 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
1321 SmallVector<SDValue, 3> Ops;
1324 // See if Op has a chain input.
1326 Ops.push_back(Op->getOperand(OpNo++));
1328 // The next operand is the intrinsic opcode.
1329 assert(Op->getOperand(OpNo).getOpcode() == ISD::TargetConstant);
1331 // See if the next operand has type i64.
1332 SDValue Opnd = Op->getOperand(++OpNo), In64;
1334 if (Opnd.getValueType() == MVT::i64)
1335 In64 = initAccumulator(Opnd, DL, DAG);
1337 Ops.push_back(Opnd);
1339 // Push the remaining operands.
1340 for (++OpNo ; OpNo < Op->getNumOperands(); ++OpNo)
1341 Ops.push_back(Op->getOperand(OpNo));
1343 // Add In64 to the end of the list.
1345 Ops.push_back(In64);
1348 SmallVector<EVT, 2> ResTys;
1350 for (SDNode::value_iterator I = Op->value_begin(), E = Op->value_end();
1352 ResTys.push_back((*I == MVT::i64) ? MVT::Untyped : *I);
1355 SDValue Val = DAG.getNode(Opc, DL, ResTys, Ops);
1356 SDValue Out = (ResTys[0] == MVT::Untyped) ? extractLOHI(Val, DL, DAG) : Val;
1361 assert(Val->getValueType(1) == MVT::Other);
1362 SDValue Vals[] = { Out, SDValue(Val.getNode(), 1) };
1363 return DAG.getMergeValues(Vals, DL);
1366 // Lower an MSA copy intrinsic into the specified SelectionDAG node
1367 static SDValue lowerMSACopyIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
1369 SDValue Vec = Op->getOperand(1);
1370 SDValue Idx = Op->getOperand(2);
1371 EVT ResTy = Op->getValueType(0);
1372 EVT EltTy = Vec->getValueType(0).getVectorElementType();
1374 SDValue Result = DAG.getNode(Opc, DL, ResTy, Vec, Idx,
1375 DAG.getValueType(EltTy));
1380 static SDValue lowerMSASplatZExt(SDValue Op, unsigned OpNr, SelectionDAG &DAG) {
1381 EVT ResVecTy = Op->getValueType(0);
1382 EVT ViaVecTy = ResVecTy;
1385 // When ResVecTy == MVT::v2i64, LaneA is the upper 32 bits of the lane and
1386 // LaneB is the lower 32-bits. Otherwise LaneA and LaneB are alternating
1389 SDValue LaneB = Op->getOperand(2);
1391 if (ResVecTy == MVT::v2i64) {
1392 LaneA = DAG.getConstant(0, DL, MVT::i32);
1393 ViaVecTy = MVT::v4i32;
1397 SDValue Ops[16] = { LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB,
1398 LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB };
1400 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy,
1401 makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
1403 if (ViaVecTy != ResVecTy)
1404 Result = DAG.getNode(ISD::BITCAST, DL, ResVecTy, Result);
1409 static SDValue lowerMSASplatImm(SDValue Op, unsigned ImmOp, SelectionDAG &DAG) {
1410 return DAG.getConstant(Op->getConstantOperandVal(ImmOp), SDLoc(Op),
1411 Op->getValueType(0));
1414 static SDValue getBuildVectorSplat(EVT VecTy, SDValue SplatValue,
1415 bool BigEndian, SelectionDAG &DAG) {
1416 EVT ViaVecTy = VecTy;
1417 SDValue SplatValueA = SplatValue;
1418 SDValue SplatValueB = SplatValue;
1419 SDLoc DL(SplatValue);
1421 if (VecTy == MVT::v2i64) {
1422 // v2i64 BUILD_VECTOR must be performed via v4i32 so split into i32's.
1423 ViaVecTy = MVT::v4i32;
1425 SplatValueA = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, SplatValue);
1426 SplatValueB = DAG.getNode(ISD::SRL, DL, MVT::i64, SplatValue,
1427 DAG.getConstant(32, DL, MVT::i32));
1428 SplatValueB = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, SplatValueB);
1431 // We currently hold the parts in little endian order. Swap them if
1434 std::swap(SplatValueA, SplatValueB);
1436 SDValue Ops[16] = { SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1437 SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1438 SplatValueA, SplatValueB, SplatValueA, SplatValueB,
1439 SplatValueA, SplatValueB, SplatValueA, SplatValueB };
1441 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy,
1442 makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
1444 if (VecTy != ViaVecTy)
1445 Result = DAG.getNode(ISD::BITCAST, DL, VecTy, Result);
1450 static SDValue lowerMSABinaryBitImmIntr(SDValue Op, SelectionDAG &DAG,
1451 unsigned Opc, SDValue Imm,
1453 EVT VecTy = Op->getValueType(0);
1457 // The DAG Combiner can't constant fold bitcasted vectors yet so we must do it
1459 if (VecTy == MVT::v2i64) {
1460 if (ConstantSDNode *CImm = dyn_cast<ConstantSDNode>(Imm)) {
1461 APInt BitImm = APInt(64, 1) << CImm->getAPIntValue();
1463 SDValue BitImmHiOp = DAG.getConstant(BitImm.lshr(32).trunc(32), DL,
1465 SDValue BitImmLoOp = DAG.getConstant(BitImm.trunc(32), DL, MVT::i32);
1468 std::swap(BitImmLoOp, BitImmHiOp);
1471 DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
1472 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v4i32, BitImmLoOp,
1473 BitImmHiOp, BitImmLoOp, BitImmHiOp));
1477 if (!Exp2Imm.getNode()) {
1478 // We couldnt constant fold, do a vector shift instead
1480 // Extend i32 to i64 if necessary. Sign or zero extend doesn't matter since
1481 // only values 0-63 are valid.
1482 if (VecTy == MVT::v2i64)
1483 Imm = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Imm);
1485 Exp2Imm = getBuildVectorSplat(VecTy, Imm, BigEndian, DAG);
1487 Exp2Imm = DAG.getNode(ISD::SHL, DL, VecTy, DAG.getConstant(1, DL, VecTy),
1491 return DAG.getNode(Opc, DL, VecTy, Op->getOperand(1), Exp2Imm);
1494 static SDValue lowerMSABitClear(SDValue Op, SelectionDAG &DAG) {
1495 EVT ResTy = Op->getValueType(0);
1497 SDValue One = DAG.getConstant(1, DL, ResTy);
1498 SDValue Bit = DAG.getNode(ISD::SHL, DL, ResTy, One, Op->getOperand(2));
1500 return DAG.getNode(ISD::AND, DL, ResTy, Op->getOperand(1),
1501 DAG.getNOT(DL, Bit, ResTy));
1504 static SDValue lowerMSABitClearImm(SDValue Op, SelectionDAG &DAG) {
1506 EVT ResTy = Op->getValueType(0);
1507 APInt BitImm = APInt(ResTy.getVectorElementType().getSizeInBits(), 1)
1508 << cast<ConstantSDNode>(Op->getOperand(2))->getAPIntValue();
1509 SDValue BitMask = DAG.getConstant(~BitImm, DL, ResTy);
1511 return DAG.getNode(ISD::AND, DL, ResTy, Op->getOperand(1), BitMask);
1514 SDValue MipsSETargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
1515 SelectionDAG &DAG) const {
1518 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) {
1521 case Intrinsic::mips_shilo:
1522 return lowerDSPIntr(Op, DAG, MipsISD::SHILO);
1523 case Intrinsic::mips_dpau_h_qbl:
1524 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL);
1525 case Intrinsic::mips_dpau_h_qbr:
1526 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR);
1527 case Intrinsic::mips_dpsu_h_qbl:
1528 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL);
1529 case Intrinsic::mips_dpsu_h_qbr:
1530 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR);
1531 case Intrinsic::mips_dpa_w_ph:
1532 return lowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH);
1533 case Intrinsic::mips_dps_w_ph:
1534 return lowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH);
1535 case Intrinsic::mips_dpax_w_ph:
1536 return lowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH);
1537 case Intrinsic::mips_dpsx_w_ph:
1538 return lowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH);
1539 case Intrinsic::mips_mulsa_w_ph:
1540 return lowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH);
1541 case Intrinsic::mips_mult:
1542 return lowerDSPIntr(Op, DAG, MipsISD::Mult);
1543 case Intrinsic::mips_multu:
1544 return lowerDSPIntr(Op, DAG, MipsISD::Multu);
1545 case Intrinsic::mips_madd:
1546 return lowerDSPIntr(Op, DAG, MipsISD::MAdd);
1547 case Intrinsic::mips_maddu:
1548 return lowerDSPIntr(Op, DAG, MipsISD::MAddu);
1549 case Intrinsic::mips_msub:
1550 return lowerDSPIntr(Op, DAG, MipsISD::MSub);
1551 case Intrinsic::mips_msubu:
1552 return lowerDSPIntr(Op, DAG, MipsISD::MSubu);
1553 case Intrinsic::mips_addv_b:
1554 case Intrinsic::mips_addv_h:
1555 case Intrinsic::mips_addv_w:
1556 case Intrinsic::mips_addv_d:
1557 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1),
1559 case Intrinsic::mips_addvi_b:
1560 case Intrinsic::mips_addvi_h:
1561 case Intrinsic::mips_addvi_w:
1562 case Intrinsic::mips_addvi_d:
1563 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1),
1564 lowerMSASplatImm(Op, 2, DAG));
1565 case Intrinsic::mips_and_v:
1566 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
1568 case Intrinsic::mips_andi_b:
1569 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
1570 lowerMSASplatImm(Op, 2, DAG));
1571 case Intrinsic::mips_bclr_b:
1572 case Intrinsic::mips_bclr_h:
1573 case Intrinsic::mips_bclr_w:
1574 case Intrinsic::mips_bclr_d:
1575 return lowerMSABitClear(Op, DAG);
1576 case Intrinsic::mips_bclri_b:
1577 case Intrinsic::mips_bclri_h:
1578 case Intrinsic::mips_bclri_w:
1579 case Intrinsic::mips_bclri_d:
1580 return lowerMSABitClearImm(Op, DAG);
1581 case Intrinsic::mips_binsli_b:
1582 case Intrinsic::mips_binsli_h:
1583 case Intrinsic::mips_binsli_w:
1584 case Intrinsic::mips_binsli_d: {
1585 // binsli_x(IfClear, IfSet, nbits) -> (vselect LBitsMask, IfSet, IfClear)
1586 EVT VecTy = Op->getValueType(0);
1587 EVT EltTy = VecTy.getVectorElementType();
1588 APInt Mask = APInt::getHighBitsSet(EltTy.getSizeInBits(),
1589 Op->getConstantOperandVal(3));
1590 return DAG.getNode(ISD::VSELECT, DL, VecTy,
1591 DAG.getConstant(Mask, DL, VecTy, true),
1592 Op->getOperand(2), Op->getOperand(1));
1594 case Intrinsic::mips_binsri_b:
1595 case Intrinsic::mips_binsri_h:
1596 case Intrinsic::mips_binsri_w:
1597 case Intrinsic::mips_binsri_d: {
1598 // binsri_x(IfClear, IfSet, nbits) -> (vselect RBitsMask, IfSet, IfClear)
1599 EVT VecTy = Op->getValueType(0);
1600 EVT EltTy = VecTy.getVectorElementType();
1601 APInt Mask = APInt::getLowBitsSet(EltTy.getSizeInBits(),
1602 Op->getConstantOperandVal(3));
1603 return DAG.getNode(ISD::VSELECT, DL, VecTy,
1604 DAG.getConstant(Mask, DL, VecTy, true),
1605 Op->getOperand(2), Op->getOperand(1));
1607 case Intrinsic::mips_bmnz_v:
1608 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), Op->getOperand(3),
1609 Op->getOperand(2), Op->getOperand(1));
1610 case Intrinsic::mips_bmnzi_b:
1611 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1612 lowerMSASplatImm(Op, 3, DAG), Op->getOperand(2),
1614 case Intrinsic::mips_bmz_v:
1615 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), Op->getOperand(3),
1616 Op->getOperand(1), Op->getOperand(2));
1617 case Intrinsic::mips_bmzi_b:
1618 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1619 lowerMSASplatImm(Op, 3, DAG), Op->getOperand(1),
1621 case Intrinsic::mips_bneg_b:
1622 case Intrinsic::mips_bneg_h:
1623 case Intrinsic::mips_bneg_w:
1624 case Intrinsic::mips_bneg_d: {
1625 EVT VecTy = Op->getValueType(0);
1626 SDValue One = DAG.getConstant(1, DL, VecTy);
1628 return DAG.getNode(ISD::XOR, DL, VecTy, Op->getOperand(1),
1629 DAG.getNode(ISD::SHL, DL, VecTy, One,
1630 Op->getOperand(2)));
1632 case Intrinsic::mips_bnegi_b:
1633 case Intrinsic::mips_bnegi_h:
1634 case Intrinsic::mips_bnegi_w:
1635 case Intrinsic::mips_bnegi_d:
1636 return lowerMSABinaryBitImmIntr(Op, DAG, ISD::XOR, Op->getOperand(2),
1637 !Subtarget.isLittle());
1638 case Intrinsic::mips_bnz_b:
1639 case Intrinsic::mips_bnz_h:
1640 case Intrinsic::mips_bnz_w:
1641 case Intrinsic::mips_bnz_d:
1642 return DAG.getNode(MipsISD::VALL_NONZERO, DL, Op->getValueType(0),
1644 case Intrinsic::mips_bnz_v:
1645 return DAG.getNode(MipsISD::VANY_NONZERO, DL, Op->getValueType(0),
1647 case Intrinsic::mips_bsel_v:
1648 // bsel_v(Mask, IfClear, IfSet) -> (vselect Mask, IfSet, IfClear)
1649 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1650 Op->getOperand(1), Op->getOperand(3),
1652 case Intrinsic::mips_bseli_b:
1653 // bseli_v(Mask, IfClear, IfSet) -> (vselect Mask, IfSet, IfClear)
1654 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0),
1655 Op->getOperand(1), lowerMSASplatImm(Op, 3, DAG),
1657 case Intrinsic::mips_bset_b:
1658 case Intrinsic::mips_bset_h:
1659 case Intrinsic::mips_bset_w:
1660 case Intrinsic::mips_bset_d: {
1661 EVT VecTy = Op->getValueType(0);
1662 SDValue One = DAG.getConstant(1, DL, VecTy);
1664 return DAG.getNode(ISD::OR, DL, VecTy, Op->getOperand(1),
1665 DAG.getNode(ISD::SHL, DL, VecTy, One,
1666 Op->getOperand(2)));
1668 case Intrinsic::mips_bseti_b:
1669 case Intrinsic::mips_bseti_h:
1670 case Intrinsic::mips_bseti_w:
1671 case Intrinsic::mips_bseti_d:
1672 return lowerMSABinaryBitImmIntr(Op, DAG, ISD::OR, Op->getOperand(2),
1673 !Subtarget.isLittle());
1674 case Intrinsic::mips_bz_b:
1675 case Intrinsic::mips_bz_h:
1676 case Intrinsic::mips_bz_w:
1677 case Intrinsic::mips_bz_d:
1678 return DAG.getNode(MipsISD::VALL_ZERO, DL, Op->getValueType(0),
1680 case Intrinsic::mips_bz_v:
1681 return DAG.getNode(MipsISD::VANY_ZERO, DL, Op->getValueType(0),
1683 case Intrinsic::mips_ceq_b:
1684 case Intrinsic::mips_ceq_h:
1685 case Intrinsic::mips_ceq_w:
1686 case Intrinsic::mips_ceq_d:
1687 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1688 Op->getOperand(2), ISD::SETEQ);
1689 case Intrinsic::mips_ceqi_b:
1690 case Intrinsic::mips_ceqi_h:
1691 case Intrinsic::mips_ceqi_w:
1692 case Intrinsic::mips_ceqi_d:
1693 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1694 lowerMSASplatImm(Op, 2, DAG), ISD::SETEQ);
1695 case Intrinsic::mips_cle_s_b:
1696 case Intrinsic::mips_cle_s_h:
1697 case Intrinsic::mips_cle_s_w:
1698 case Intrinsic::mips_cle_s_d:
1699 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1700 Op->getOperand(2), ISD::SETLE);
1701 case Intrinsic::mips_clei_s_b:
1702 case Intrinsic::mips_clei_s_h:
1703 case Intrinsic::mips_clei_s_w:
1704 case Intrinsic::mips_clei_s_d:
1705 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1706 lowerMSASplatImm(Op, 2, DAG), ISD::SETLE);
1707 case Intrinsic::mips_cle_u_b:
1708 case Intrinsic::mips_cle_u_h:
1709 case Intrinsic::mips_cle_u_w:
1710 case Intrinsic::mips_cle_u_d:
1711 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1712 Op->getOperand(2), ISD::SETULE);
1713 case Intrinsic::mips_clei_u_b:
1714 case Intrinsic::mips_clei_u_h:
1715 case Intrinsic::mips_clei_u_w:
1716 case Intrinsic::mips_clei_u_d:
1717 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1718 lowerMSASplatImm(Op, 2, DAG), ISD::SETULE);
1719 case Intrinsic::mips_clt_s_b:
1720 case Intrinsic::mips_clt_s_h:
1721 case Intrinsic::mips_clt_s_w:
1722 case Intrinsic::mips_clt_s_d:
1723 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1724 Op->getOperand(2), ISD::SETLT);
1725 case Intrinsic::mips_clti_s_b:
1726 case Intrinsic::mips_clti_s_h:
1727 case Intrinsic::mips_clti_s_w:
1728 case Intrinsic::mips_clti_s_d:
1729 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1730 lowerMSASplatImm(Op, 2, DAG), ISD::SETLT);
1731 case Intrinsic::mips_clt_u_b:
1732 case Intrinsic::mips_clt_u_h:
1733 case Intrinsic::mips_clt_u_w:
1734 case Intrinsic::mips_clt_u_d:
1735 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1736 Op->getOperand(2), ISD::SETULT);
1737 case Intrinsic::mips_clti_u_b:
1738 case Intrinsic::mips_clti_u_h:
1739 case Intrinsic::mips_clti_u_w:
1740 case Intrinsic::mips_clti_u_d:
1741 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1742 lowerMSASplatImm(Op, 2, DAG), ISD::SETULT);
1743 case Intrinsic::mips_copy_s_b:
1744 case Intrinsic::mips_copy_s_h:
1745 case Intrinsic::mips_copy_s_w:
1746 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
1747 case Intrinsic::mips_copy_s_d:
1748 if (Subtarget.hasMips64())
1749 // Lower directly into VEXTRACT_SEXT_ELT since i64 is legal on Mips64.
1750 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
1752 // Lower into the generic EXTRACT_VECTOR_ELT node and let the type
1753 // legalizer and EXTRACT_VECTOR_ELT lowering sort it out.
1754 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
1755 Op->getValueType(0), Op->getOperand(1),
1758 case Intrinsic::mips_copy_u_b:
1759 case Intrinsic::mips_copy_u_h:
1760 case Intrinsic::mips_copy_u_w:
1761 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
1762 case Intrinsic::mips_copy_u_d:
1763 if (Subtarget.hasMips64())
1764 // Lower directly into VEXTRACT_ZEXT_ELT since i64 is legal on Mips64.
1765 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
1767 // Lower into the generic EXTRACT_VECTOR_ELT node and let the type
1768 // legalizer and EXTRACT_VECTOR_ELT lowering sort it out.
1769 // Note: When i64 is illegal, this results in copy_s.w instructions
1770 // instead of copy_u.w instructions. This makes no difference to the
1771 // behaviour since i64 is only illegal when the register file is 32-bit.
1772 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
1773 Op->getValueType(0), Op->getOperand(1),
1776 case Intrinsic::mips_div_s_b:
1777 case Intrinsic::mips_div_s_h:
1778 case Intrinsic::mips_div_s_w:
1779 case Intrinsic::mips_div_s_d:
1780 return DAG.getNode(ISD::SDIV, DL, Op->getValueType(0), Op->getOperand(1),
1782 case Intrinsic::mips_div_u_b:
1783 case Intrinsic::mips_div_u_h:
1784 case Intrinsic::mips_div_u_w:
1785 case Intrinsic::mips_div_u_d:
1786 return DAG.getNode(ISD::UDIV, DL, Op->getValueType(0), Op->getOperand(1),
1788 case Intrinsic::mips_fadd_w:
1789 case Intrinsic::mips_fadd_d: {
1790 // TODO: If intrinsics have fast-math-flags, propagate them.
1791 return DAG.getNode(ISD::FADD, DL, Op->getValueType(0), Op->getOperand(1),
1794 // Don't lower mips_fcaf_[wd] since LLVM folds SETFALSE condcodes away
1795 case Intrinsic::mips_fceq_w:
1796 case Intrinsic::mips_fceq_d:
1797 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1798 Op->getOperand(2), ISD::SETOEQ);
1799 case Intrinsic::mips_fcle_w:
1800 case Intrinsic::mips_fcle_d:
1801 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1802 Op->getOperand(2), ISD::SETOLE);
1803 case Intrinsic::mips_fclt_w:
1804 case Intrinsic::mips_fclt_d:
1805 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1806 Op->getOperand(2), ISD::SETOLT);
1807 case Intrinsic::mips_fcne_w:
1808 case Intrinsic::mips_fcne_d:
1809 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1810 Op->getOperand(2), ISD::SETONE);
1811 case Intrinsic::mips_fcor_w:
1812 case Intrinsic::mips_fcor_d:
1813 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1814 Op->getOperand(2), ISD::SETO);
1815 case Intrinsic::mips_fcueq_w:
1816 case Intrinsic::mips_fcueq_d:
1817 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1818 Op->getOperand(2), ISD::SETUEQ);
1819 case Intrinsic::mips_fcule_w:
1820 case Intrinsic::mips_fcule_d:
1821 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1822 Op->getOperand(2), ISD::SETULE);
1823 case Intrinsic::mips_fcult_w:
1824 case Intrinsic::mips_fcult_d:
1825 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1826 Op->getOperand(2), ISD::SETULT);
1827 case Intrinsic::mips_fcun_w:
1828 case Intrinsic::mips_fcun_d:
1829 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1830 Op->getOperand(2), ISD::SETUO);
1831 case Intrinsic::mips_fcune_w:
1832 case Intrinsic::mips_fcune_d:
1833 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1),
1834 Op->getOperand(2), ISD::SETUNE);
1835 case Intrinsic::mips_fdiv_w:
1836 case Intrinsic::mips_fdiv_d: {
1837 // TODO: If intrinsics have fast-math-flags, propagate them.
1838 return DAG.getNode(ISD::FDIV, DL, Op->getValueType(0), Op->getOperand(1),
1841 case Intrinsic::mips_ffint_u_w:
1842 case Intrinsic::mips_ffint_u_d:
1843 return DAG.getNode(ISD::UINT_TO_FP, DL, Op->getValueType(0),
1845 case Intrinsic::mips_ffint_s_w:
1846 case Intrinsic::mips_ffint_s_d:
1847 return DAG.getNode(ISD::SINT_TO_FP, DL, Op->getValueType(0),
1849 case Intrinsic::mips_fill_b:
1850 case Intrinsic::mips_fill_h:
1851 case Intrinsic::mips_fill_w:
1852 case Intrinsic::mips_fill_d: {
1853 EVT ResTy = Op->getValueType(0);
1854 SmallVector<SDValue, 16> Ops(ResTy.getVectorNumElements(),
1857 // If ResTy is v2i64 then the type legalizer will break this node down into
1858 // an equivalent v4i32.
1859 return DAG.getNode(ISD::BUILD_VECTOR, DL, ResTy, Ops);
1861 case Intrinsic::mips_fexp2_w:
1862 case Intrinsic::mips_fexp2_d: {
1863 // TODO: If intrinsics have fast-math-flags, propagate them.
1864 EVT ResTy = Op->getValueType(0);
1866 ISD::FMUL, SDLoc(Op), ResTy, Op->getOperand(1),
1867 DAG.getNode(ISD::FEXP2, SDLoc(Op), ResTy, Op->getOperand(2)));
1869 case Intrinsic::mips_flog2_w:
1870 case Intrinsic::mips_flog2_d:
1871 return DAG.getNode(ISD::FLOG2, DL, Op->getValueType(0), Op->getOperand(1));
1872 case Intrinsic::mips_fmadd_w:
1873 case Intrinsic::mips_fmadd_d:
1874 return DAG.getNode(ISD::FMA, SDLoc(Op), Op->getValueType(0),
1875 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3));
1876 case Intrinsic::mips_fmul_w:
1877 case Intrinsic::mips_fmul_d: {
1878 // TODO: If intrinsics have fast-math-flags, propagate them.
1879 return DAG.getNode(ISD::FMUL, DL, Op->getValueType(0), Op->getOperand(1),
1882 case Intrinsic::mips_fmsub_w:
1883 case Intrinsic::mips_fmsub_d: {
1884 // TODO: If intrinsics have fast-math-flags, propagate them.
1885 EVT ResTy = Op->getValueType(0);
1886 return DAG.getNode(ISD::FSUB, SDLoc(Op), ResTy, Op->getOperand(1),
1887 DAG.getNode(ISD::FMUL, SDLoc(Op), ResTy,
1888 Op->getOperand(2), Op->getOperand(3)));
1890 case Intrinsic::mips_frint_w:
1891 case Intrinsic::mips_frint_d:
1892 return DAG.getNode(ISD::FRINT, DL, Op->getValueType(0), Op->getOperand(1));
1893 case Intrinsic::mips_fsqrt_w:
1894 case Intrinsic::mips_fsqrt_d:
1895 return DAG.getNode(ISD::FSQRT, DL, Op->getValueType(0), Op->getOperand(1));
1896 case Intrinsic::mips_fsub_w:
1897 case Intrinsic::mips_fsub_d: {
1898 // TODO: If intrinsics have fast-math-flags, propagate them.
1899 return DAG.getNode(ISD::FSUB, DL, Op->getValueType(0), Op->getOperand(1),
1902 case Intrinsic::mips_ftrunc_u_w:
1903 case Intrinsic::mips_ftrunc_u_d:
1904 return DAG.getNode(ISD::FP_TO_UINT, DL, Op->getValueType(0),
1906 case Intrinsic::mips_ftrunc_s_w:
1907 case Intrinsic::mips_ftrunc_s_d:
1908 return DAG.getNode(ISD::FP_TO_SINT, DL, Op->getValueType(0),
1910 case Intrinsic::mips_ilvev_b:
1911 case Intrinsic::mips_ilvev_h:
1912 case Intrinsic::mips_ilvev_w:
1913 case Intrinsic::mips_ilvev_d:
1914 return DAG.getNode(MipsISD::ILVEV, DL, Op->getValueType(0),
1915 Op->getOperand(1), Op->getOperand(2));
1916 case Intrinsic::mips_ilvl_b:
1917 case Intrinsic::mips_ilvl_h:
1918 case Intrinsic::mips_ilvl_w:
1919 case Intrinsic::mips_ilvl_d:
1920 return DAG.getNode(MipsISD::ILVL, DL, Op->getValueType(0),
1921 Op->getOperand(1), Op->getOperand(2));
1922 case Intrinsic::mips_ilvod_b:
1923 case Intrinsic::mips_ilvod_h:
1924 case Intrinsic::mips_ilvod_w:
1925 case Intrinsic::mips_ilvod_d:
1926 return DAG.getNode(MipsISD::ILVOD, DL, Op->getValueType(0),
1927 Op->getOperand(1), Op->getOperand(2));
1928 case Intrinsic::mips_ilvr_b:
1929 case Intrinsic::mips_ilvr_h:
1930 case Intrinsic::mips_ilvr_w:
1931 case Intrinsic::mips_ilvr_d:
1932 return DAG.getNode(MipsISD::ILVR, DL, Op->getValueType(0),
1933 Op->getOperand(1), Op->getOperand(2));
1934 case Intrinsic::mips_insert_b:
1935 case Intrinsic::mips_insert_h:
1936 case Intrinsic::mips_insert_w:
1937 case Intrinsic::mips_insert_d:
1938 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(Op), Op->getValueType(0),
1939 Op->getOperand(1), Op->getOperand(3), Op->getOperand(2));
1940 case Intrinsic::mips_insve_b:
1941 case Intrinsic::mips_insve_h:
1942 case Intrinsic::mips_insve_w:
1943 case Intrinsic::mips_insve_d:
1944 return DAG.getNode(MipsISD::INSVE, DL, Op->getValueType(0),
1945 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3),
1946 DAG.getConstant(0, DL, MVT::i32));
1947 case Intrinsic::mips_ldi_b:
1948 case Intrinsic::mips_ldi_h:
1949 case Intrinsic::mips_ldi_w:
1950 case Intrinsic::mips_ldi_d:
1951 return lowerMSASplatImm(Op, 1, DAG);
1952 case Intrinsic::mips_lsa:
1953 case Intrinsic::mips_dlsa: {
1954 EVT ResTy = Op->getValueType(0);
1955 return DAG.getNode(ISD::ADD, SDLoc(Op), ResTy, Op->getOperand(1),
1956 DAG.getNode(ISD::SHL, SDLoc(Op), ResTy,
1957 Op->getOperand(2), Op->getOperand(3)));
1959 case Intrinsic::mips_maddv_b:
1960 case Intrinsic::mips_maddv_h:
1961 case Intrinsic::mips_maddv_w:
1962 case Intrinsic::mips_maddv_d: {
1963 EVT ResTy = Op->getValueType(0);
1964 return DAG.getNode(ISD::ADD, SDLoc(Op), ResTy, Op->getOperand(1),
1965 DAG.getNode(ISD::MUL, SDLoc(Op), ResTy,
1966 Op->getOperand(2), Op->getOperand(3)));
1968 case Intrinsic::mips_max_s_b:
1969 case Intrinsic::mips_max_s_h:
1970 case Intrinsic::mips_max_s_w:
1971 case Intrinsic::mips_max_s_d:
1972 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0),
1973 Op->getOperand(1), Op->getOperand(2));
1974 case Intrinsic::mips_max_u_b:
1975 case Intrinsic::mips_max_u_h:
1976 case Intrinsic::mips_max_u_w:
1977 case Intrinsic::mips_max_u_d:
1978 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0),
1979 Op->getOperand(1), Op->getOperand(2));
1980 case Intrinsic::mips_maxi_s_b:
1981 case Intrinsic::mips_maxi_s_h:
1982 case Intrinsic::mips_maxi_s_w:
1983 case Intrinsic::mips_maxi_s_d:
1984 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0),
1985 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1986 case Intrinsic::mips_maxi_u_b:
1987 case Intrinsic::mips_maxi_u_h:
1988 case Intrinsic::mips_maxi_u_w:
1989 case Intrinsic::mips_maxi_u_d:
1990 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0),
1991 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
1992 case Intrinsic::mips_min_s_b:
1993 case Intrinsic::mips_min_s_h:
1994 case Intrinsic::mips_min_s_w:
1995 case Intrinsic::mips_min_s_d:
1996 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0),
1997 Op->getOperand(1), Op->getOperand(2));
1998 case Intrinsic::mips_min_u_b:
1999 case Intrinsic::mips_min_u_h:
2000 case Intrinsic::mips_min_u_w:
2001 case Intrinsic::mips_min_u_d:
2002 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0),
2003 Op->getOperand(1), Op->getOperand(2));
2004 case Intrinsic::mips_mini_s_b:
2005 case Intrinsic::mips_mini_s_h:
2006 case Intrinsic::mips_mini_s_w:
2007 case Intrinsic::mips_mini_s_d:
2008 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0),
2009 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2010 case Intrinsic::mips_mini_u_b:
2011 case Intrinsic::mips_mini_u_h:
2012 case Intrinsic::mips_mini_u_w:
2013 case Intrinsic::mips_mini_u_d:
2014 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0),
2015 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2016 case Intrinsic::mips_mod_s_b:
2017 case Intrinsic::mips_mod_s_h:
2018 case Intrinsic::mips_mod_s_w:
2019 case Intrinsic::mips_mod_s_d:
2020 return DAG.getNode(ISD::SREM, DL, Op->getValueType(0), Op->getOperand(1),
2022 case Intrinsic::mips_mod_u_b:
2023 case Intrinsic::mips_mod_u_h:
2024 case Intrinsic::mips_mod_u_w:
2025 case Intrinsic::mips_mod_u_d:
2026 return DAG.getNode(ISD::UREM, DL, Op->getValueType(0), Op->getOperand(1),
2028 case Intrinsic::mips_mulv_b:
2029 case Intrinsic::mips_mulv_h:
2030 case Intrinsic::mips_mulv_w:
2031 case Intrinsic::mips_mulv_d:
2032 return DAG.getNode(ISD::MUL, DL, Op->getValueType(0), Op->getOperand(1),
2034 case Intrinsic::mips_msubv_b:
2035 case Intrinsic::mips_msubv_h:
2036 case Intrinsic::mips_msubv_w:
2037 case Intrinsic::mips_msubv_d: {
2038 EVT ResTy = Op->getValueType(0);
2039 return DAG.getNode(ISD::SUB, SDLoc(Op), ResTy, Op->getOperand(1),
2040 DAG.getNode(ISD::MUL, SDLoc(Op), ResTy,
2041 Op->getOperand(2), Op->getOperand(3)));
2043 case Intrinsic::mips_nlzc_b:
2044 case Intrinsic::mips_nlzc_h:
2045 case Intrinsic::mips_nlzc_w:
2046 case Intrinsic::mips_nlzc_d:
2047 return DAG.getNode(ISD::CTLZ, DL, Op->getValueType(0), Op->getOperand(1));
2048 case Intrinsic::mips_nor_v: {
2049 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2050 Op->getOperand(1), Op->getOperand(2));
2051 return DAG.getNOT(DL, Res, Res->getValueType(0));
2053 case Intrinsic::mips_nori_b: {
2054 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2056 lowerMSASplatImm(Op, 2, DAG));
2057 return DAG.getNOT(DL, Res, Res->getValueType(0));
2059 case Intrinsic::mips_or_v:
2060 return DAG.getNode(ISD::OR, DL, Op->getValueType(0), Op->getOperand(1),
2062 case Intrinsic::mips_ori_b:
2063 return DAG.getNode(ISD::OR, DL, Op->getValueType(0),
2064 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2065 case Intrinsic::mips_pckev_b:
2066 case Intrinsic::mips_pckev_h:
2067 case Intrinsic::mips_pckev_w:
2068 case Intrinsic::mips_pckev_d:
2069 return DAG.getNode(MipsISD::PCKEV, DL, Op->getValueType(0),
2070 Op->getOperand(1), Op->getOperand(2));
2071 case Intrinsic::mips_pckod_b:
2072 case Intrinsic::mips_pckod_h:
2073 case Intrinsic::mips_pckod_w:
2074 case Intrinsic::mips_pckod_d:
2075 return DAG.getNode(MipsISD::PCKOD, DL, Op->getValueType(0),
2076 Op->getOperand(1), Op->getOperand(2));
2077 case Intrinsic::mips_pcnt_b:
2078 case Intrinsic::mips_pcnt_h:
2079 case Intrinsic::mips_pcnt_w:
2080 case Intrinsic::mips_pcnt_d:
2081 return DAG.getNode(ISD::CTPOP, DL, Op->getValueType(0), Op->getOperand(1));
2082 case Intrinsic::mips_shf_b:
2083 case Intrinsic::mips_shf_h:
2084 case Intrinsic::mips_shf_w:
2085 return DAG.getNode(MipsISD::SHF, DL, Op->getValueType(0),
2086 Op->getOperand(2), Op->getOperand(1));
2087 case Intrinsic::mips_sll_b:
2088 case Intrinsic::mips_sll_h:
2089 case Intrinsic::mips_sll_w:
2090 case Intrinsic::mips_sll_d:
2091 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0), Op->getOperand(1),
2093 case Intrinsic::mips_slli_b:
2094 case Intrinsic::mips_slli_h:
2095 case Intrinsic::mips_slli_w:
2096 case Intrinsic::mips_slli_d:
2097 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0),
2098 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2099 case Intrinsic::mips_splat_b:
2100 case Intrinsic::mips_splat_h:
2101 case Intrinsic::mips_splat_w:
2102 case Intrinsic::mips_splat_d:
2103 // We can't lower via VECTOR_SHUFFLE because it requires constant shuffle
2104 // masks, nor can we lower via BUILD_VECTOR & EXTRACT_VECTOR_ELT because
2105 // EXTRACT_VECTOR_ELT can't extract i64's on MIPS32.
2106 // Instead we lower to MipsISD::VSHF and match from there.
2107 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2108 lowerMSASplatZExt(Op, 2, DAG), Op->getOperand(1),
2110 case Intrinsic::mips_splati_b:
2111 case Intrinsic::mips_splati_h:
2112 case Intrinsic::mips_splati_w:
2113 case Intrinsic::mips_splati_d:
2114 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2115 lowerMSASplatImm(Op, 2, DAG), Op->getOperand(1),
2117 case Intrinsic::mips_sra_b:
2118 case Intrinsic::mips_sra_h:
2119 case Intrinsic::mips_sra_w:
2120 case Intrinsic::mips_sra_d:
2121 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0), Op->getOperand(1),
2123 case Intrinsic::mips_srai_b:
2124 case Intrinsic::mips_srai_h:
2125 case Intrinsic::mips_srai_w:
2126 case Intrinsic::mips_srai_d:
2127 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0),
2128 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2129 case Intrinsic::mips_srl_b:
2130 case Intrinsic::mips_srl_h:
2131 case Intrinsic::mips_srl_w:
2132 case Intrinsic::mips_srl_d:
2133 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0), Op->getOperand(1),
2135 case Intrinsic::mips_srli_b:
2136 case Intrinsic::mips_srli_h:
2137 case Intrinsic::mips_srli_w:
2138 case Intrinsic::mips_srli_d:
2139 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0),
2140 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2141 case Intrinsic::mips_subv_b:
2142 case Intrinsic::mips_subv_h:
2143 case Intrinsic::mips_subv_w:
2144 case Intrinsic::mips_subv_d:
2145 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0), Op->getOperand(1),
2147 case Intrinsic::mips_subvi_b:
2148 case Intrinsic::mips_subvi_h:
2149 case Intrinsic::mips_subvi_w:
2150 case Intrinsic::mips_subvi_d:
2151 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0),
2152 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2153 case Intrinsic::mips_vshf_b:
2154 case Intrinsic::mips_vshf_h:
2155 case Intrinsic::mips_vshf_w:
2156 case Intrinsic::mips_vshf_d:
2157 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0),
2158 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3));
2159 case Intrinsic::mips_xor_v:
2160 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0), Op->getOperand(1),
2162 case Intrinsic::mips_xori_b:
2163 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0),
2164 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG));
2168 static SDValue lowerMSALoadIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
2170 SDValue ChainIn = Op->getOperand(0);
2171 SDValue Address = Op->getOperand(2);
2172 SDValue Offset = Op->getOperand(3);
2173 EVT ResTy = Op->getValueType(0);
2174 EVT PtrTy = Address->getValueType(0);
2176 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
2178 return DAG.getLoad(ResTy, DL, ChainIn, Address, MachinePointerInfo(), false,
2182 SDValue MipsSETargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
2183 SelectionDAG &DAG) const {
2184 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
2188 case Intrinsic::mips_extp:
2189 return lowerDSPIntr(Op, DAG, MipsISD::EXTP);
2190 case Intrinsic::mips_extpdp:
2191 return lowerDSPIntr(Op, DAG, MipsISD::EXTPDP);
2192 case Intrinsic::mips_extr_w:
2193 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_W);
2194 case Intrinsic::mips_extr_r_w:
2195 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W);
2196 case Intrinsic::mips_extr_rs_w:
2197 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W);
2198 case Intrinsic::mips_extr_s_h:
2199 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H);
2200 case Intrinsic::mips_mthlip:
2201 return lowerDSPIntr(Op, DAG, MipsISD::MTHLIP);
2202 case Intrinsic::mips_mulsaq_s_w_ph:
2203 return lowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH);
2204 case Intrinsic::mips_maq_s_w_phl:
2205 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL);
2206 case Intrinsic::mips_maq_s_w_phr:
2207 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR);
2208 case Intrinsic::mips_maq_sa_w_phl:
2209 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL);
2210 case Intrinsic::mips_maq_sa_w_phr:
2211 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR);
2212 case Intrinsic::mips_dpaq_s_w_ph:
2213 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH);
2214 case Intrinsic::mips_dpsq_s_w_ph:
2215 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH);
2216 case Intrinsic::mips_dpaq_sa_l_w:
2217 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W);
2218 case Intrinsic::mips_dpsq_sa_l_w:
2219 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W);
2220 case Intrinsic::mips_dpaqx_s_w_ph:
2221 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH);
2222 case Intrinsic::mips_dpaqx_sa_w_ph:
2223 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH);
2224 case Intrinsic::mips_dpsqx_s_w_ph:
2225 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH);
2226 case Intrinsic::mips_dpsqx_sa_w_ph:
2227 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH);
2228 case Intrinsic::mips_ld_b:
2229 case Intrinsic::mips_ld_h:
2230 case Intrinsic::mips_ld_w:
2231 case Intrinsic::mips_ld_d:
2232 return lowerMSALoadIntr(Op, DAG, Intr);
2236 static SDValue lowerMSAStoreIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) {
2238 SDValue ChainIn = Op->getOperand(0);
2239 SDValue Value = Op->getOperand(2);
2240 SDValue Address = Op->getOperand(3);
2241 SDValue Offset = Op->getOperand(4);
2242 EVT PtrTy = Address->getValueType(0);
2244 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset);
2246 return DAG.getStore(ChainIn, DL, Value, Address, MachinePointerInfo(), false,
2250 SDValue MipsSETargetLowering::lowerINTRINSIC_VOID(SDValue Op,
2251 SelectionDAG &DAG) const {
2252 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
2256 case Intrinsic::mips_st_b:
2257 case Intrinsic::mips_st_h:
2258 case Intrinsic::mips_st_w:
2259 case Intrinsic::mips_st_d:
2260 return lowerMSAStoreIntr(Op, DAG, Intr);
2264 /// \brief Check if the given BuildVectorSDNode is a splat.
2265 /// This method currently relies on DAG nodes being reused when equivalent,
2266 /// so it's possible for this to return false even when isConstantSplat returns
2268 static bool isSplatVector(const BuildVectorSDNode *N) {
2269 unsigned int nOps = N->getNumOperands();
2270 assert(nOps > 1 && "isSplatVector has 0 or 1 sized build vector");
2272 SDValue Operand0 = N->getOperand(0);
2274 for (unsigned int i = 1; i < nOps; ++i) {
2275 if (N->getOperand(i) != Operand0)
2282 // Lower ISD::EXTRACT_VECTOR_ELT into MipsISD::VEXTRACT_SEXT_ELT.
2284 // The non-value bits resulting from ISD::EXTRACT_VECTOR_ELT are undefined. We
2285 // choose to sign-extend but we could have equally chosen zero-extend. The
2286 // DAGCombiner will fold any sign/zero extension of the ISD::EXTRACT_VECTOR_ELT
2287 // result into this node later (possibly changing it to a zero-extend in the
2289 SDValue MipsSETargetLowering::
2290 lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
2292 EVT ResTy = Op->getValueType(0);
2293 SDValue Op0 = Op->getOperand(0);
2294 EVT VecTy = Op0->getValueType(0);
2296 if (!VecTy.is128BitVector())
2299 if (ResTy.isInteger()) {
2300 SDValue Op1 = Op->getOperand(1);
2301 EVT EltTy = VecTy.getVectorElementType();
2302 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, DL, ResTy, Op0, Op1,
2303 DAG.getValueType(EltTy));
2309 static bool isConstantOrUndef(const SDValue Op) {
2310 if (Op->getOpcode() == ISD::UNDEF)
2312 if (isa<ConstantSDNode>(Op))
2314 if (isa<ConstantFPSDNode>(Op))
2319 static bool isConstantOrUndefBUILD_VECTOR(const BuildVectorSDNode *Op) {
2320 for (unsigned i = 0; i < Op->getNumOperands(); ++i)
2321 if (isConstantOrUndef(Op->getOperand(i)))
2326 // Lowers ISD::BUILD_VECTOR into appropriate SelectionDAG nodes for the
2329 // Lowers according to the following rules:
2330 // - Constant splats are legal as-is as long as the SplatBitSize is a power of
2331 // 2 less than or equal to 64 and the value fits into a signed 10-bit
2333 // - Constant splats are lowered to bitconverted BUILD_VECTORs if SplatBitSize
2334 // is a power of 2 less than or equal to 64 and the value does not fit into a
2335 // signed 10-bit immediate
2336 // - Non-constant splats are legal as-is.
2337 // - Non-constant non-splats are lowered to sequences of INSERT_VECTOR_ELT.
2338 // - All others are illegal and must be expanded.
2339 SDValue MipsSETargetLowering::lowerBUILD_VECTOR(SDValue Op,
2340 SelectionDAG &DAG) const {
2341 BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op);
2342 EVT ResTy = Op->getValueType(0);
2344 APInt SplatValue, SplatUndef;
2345 unsigned SplatBitSize;
2348 if (!Subtarget.hasMSA() || !ResTy.is128BitVector())
2351 if (Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
2353 !Subtarget.isLittle()) && SplatBitSize <= 64) {
2354 // We can only cope with 8, 16, 32, or 64-bit elements
2355 if (SplatBitSize != 8 && SplatBitSize != 16 && SplatBitSize != 32 &&
2359 // If the value fits into a simm10 then we can use ldi.[bhwd]
2360 // However, if it isn't an integer type we will have to bitcast from an
2361 // integer type first. Also, if there are any undefs, we must lower them
2362 // to defined values first.
2363 if (ResTy.isInteger() && !HasAnyUndefs && SplatValue.isSignedIntN(10))
2368 switch (SplatBitSize) {
2372 ViaVecTy = MVT::v16i8;
2375 ViaVecTy = MVT::v8i16;
2378 ViaVecTy = MVT::v4i32;
2381 // There's no fill.d to fall back on for 64-bit values
2385 // SelectionDAG::getConstant will promote SplatValue appropriately.
2386 SDValue Result = DAG.getConstant(SplatValue, DL, ViaVecTy);
2388 // Bitcast to the type we originally wanted
2389 if (ViaVecTy != ResTy)
2390 Result = DAG.getNode(ISD::BITCAST, SDLoc(Node), ResTy, Result);
2393 } else if (isSplatVector(Node))
2395 else if (!isConstantOrUndefBUILD_VECTOR(Node)) {
2396 // Use INSERT_VECTOR_ELT operations rather than expand to stores.
2397 // The resulting code is the same length as the expansion, but it doesn't
2398 // use memory operations
2399 EVT ResTy = Node->getValueType(0);
2401 assert(ResTy.isVector());
2403 unsigned NumElts = ResTy.getVectorNumElements();
2404 SDValue Vector = DAG.getUNDEF(ResTy);
2405 for (unsigned i = 0; i < NumElts; ++i) {
2406 Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector,
2407 Node->getOperand(i),
2408 DAG.getConstant(i, DL, MVT::i32));
2416 // Lower VECTOR_SHUFFLE into SHF (if possible).
2418 // SHF splits the vector into blocks of four elements, then shuffles these
2419 // elements according to a <4 x i2> constant (encoded as an integer immediate).
2421 // It is therefore possible to lower into SHF when the mask takes the form:
2422 // <a, b, c, d, a+4, b+4, c+4, d+4, a+8, b+8, c+8, d+8, ...>
2423 // When undef's appear they are treated as if they were whatever value is
2424 // necessary in order to fit the above forms.
2427 // %2 = shufflevector <8 x i16> %0, <8 x i16> undef,
2428 // <8 x i32> <i32 3, i32 2, i32 1, i32 0,
2429 // i32 7, i32 6, i32 5, i32 4>
2431 // (SHF_H $w0, $w1, 27)
2432 // where the 27 comes from:
2433 // 3 + (2 << 2) + (1 << 4) + (0 << 6)
2434 static SDValue lowerVECTOR_SHUFFLE_SHF(SDValue Op, EVT ResTy,
2435 SmallVector<int, 16> Indices,
2436 SelectionDAG &DAG) {
2437 int SHFIndices[4] = { -1, -1, -1, -1 };
2439 if (Indices.size() < 4)
2442 for (unsigned i = 0; i < 4; ++i) {
2443 for (unsigned j = i; j < Indices.size(); j += 4) {
2444 int Idx = Indices[j];
2446 // Convert from vector index to 4-element subvector index
2447 // If an index refers to an element outside of the subvector then give up
2450 if (Idx < 0 || Idx >= 4)
2454 // If the mask has an undef, replace it with the current index.
2455 // Note that it might still be undef if the current index is also undef
2456 if (SHFIndices[i] == -1)
2457 SHFIndices[i] = Idx;
2459 // Check that non-undef values are the same as in the mask. If they
2460 // aren't then give up
2461 if (!(Idx == -1 || Idx == SHFIndices[i]))
2466 // Calculate the immediate. Replace any remaining undefs with zero
2468 for (int i = 3; i >= 0; --i) {
2469 int Idx = SHFIndices[i];
2479 return DAG.getNode(MipsISD::SHF, DL, ResTy,
2480 DAG.getConstant(Imm, DL, MVT::i32), Op->getOperand(0));
2483 /// Determine whether a range fits a regular pattern of values.
2484 /// This function accounts for the possibility of jumping over the End iterator.
2485 template <typename ValType>
2487 fitsRegularPattern(typename SmallVectorImpl<ValType>::const_iterator Begin,
2488 unsigned CheckStride,
2489 typename SmallVectorImpl<ValType>::const_iterator End,
2490 ValType ExpectedIndex, unsigned ExpectedIndexStride) {
2494 if (*I != -1 && *I != ExpectedIndex)
2496 ExpectedIndex += ExpectedIndexStride;
2498 // Incrementing past End is undefined behaviour so we must increment one
2499 // step at a time and check for End at each step.
2500 for (unsigned n = 0; n < CheckStride && I != End; ++n, ++I)
2501 ; // Empty loop body.
2506 // Determine whether VECTOR_SHUFFLE is a SPLATI.
2508 // It is a SPLATI when the mask is:
2510 // where x is any valid index.
2512 // When undef's appear in the mask they are treated as if they were whatever
2513 // value is necessary in order to fit the above form.
2514 static bool isVECTOR_SHUFFLE_SPLATI(SDValue Op, EVT ResTy,
2515 SmallVector<int, 16> Indices,
2516 SelectionDAG &DAG) {
2517 assert((Indices.size() % 2) == 0);
2519 int SplatIndex = -1;
2520 for (const auto &V : Indices) {
2527 return fitsRegularPattern<int>(Indices.begin(), 1, Indices.end(), SplatIndex,
2531 // Lower VECTOR_SHUFFLE into ILVEV (if possible).
2533 // ILVEV interleaves the even elements from each vector.
2535 // It is possible to lower into ILVEV when the mask consists of two of the
2536 // following forms interleaved:
2538 // <n, n+2, n+4, ...>
2539 // where n is the number of elements in the vector.
2541 // <0, 0, 2, 2, 4, 4, ...>
2542 // <0, n, 2, n+2, 4, n+4, ...>
2544 // When undef's appear in the mask they are treated as if they were whatever
2545 // value is necessary in order to fit the above forms.
2546 static SDValue lowerVECTOR_SHUFFLE_ILVEV(SDValue Op, EVT ResTy,
2547 SmallVector<int, 16> Indices,
2548 SelectionDAG &DAG) {
2549 assert((Indices.size() % 2) == 0);
2553 const auto &Begin = Indices.begin();
2554 const auto &End = Indices.end();
2556 // Check even elements are taken from the even elements of one half or the
2557 // other and pick an operand accordingly.
2558 if (fitsRegularPattern<int>(Begin, 2, End, 0, 2))
2559 Wt = Op->getOperand(0);
2560 else if (fitsRegularPattern<int>(Begin, 2, End, Indices.size(), 2))
2561 Wt = Op->getOperand(1);
2565 // Check odd elements are taken from the even elements of one half or the
2566 // other and pick an operand accordingly.
2567 if (fitsRegularPattern<int>(Begin + 1, 2, End, 0, 2))
2568 Ws = Op->getOperand(0);
2569 else if (fitsRegularPattern<int>(Begin + 1, 2, End, Indices.size(), 2))
2570 Ws = Op->getOperand(1);
2574 return DAG.getNode(MipsISD::ILVEV, SDLoc(Op), ResTy, Ws, Wt);
2577 // Lower VECTOR_SHUFFLE into ILVOD (if possible).
2579 // ILVOD interleaves the odd elements from each vector.
2581 // It is possible to lower into ILVOD when the mask consists of two of the
2582 // following forms interleaved:
2584 // <n+1, n+3, n+5, ...>
2585 // where n is the number of elements in the vector.
2587 // <1, 1, 3, 3, 5, 5, ...>
2588 // <1, n+1, 3, n+3, 5, n+5, ...>
2590 // When undef's appear in the mask they are treated as if they were whatever
2591 // value is necessary in order to fit the above forms.
2592 static SDValue lowerVECTOR_SHUFFLE_ILVOD(SDValue Op, EVT ResTy,
2593 SmallVector<int, 16> Indices,
2594 SelectionDAG &DAG) {
2595 assert((Indices.size() % 2) == 0);
2599 const auto &Begin = Indices.begin();
2600 const auto &End = Indices.end();
2602 // Check even elements are taken from the odd elements of one half or the
2603 // other and pick an operand accordingly.
2604 if (fitsRegularPattern<int>(Begin, 2, End, 1, 2))
2605 Wt = Op->getOperand(0);
2606 else if (fitsRegularPattern<int>(Begin, 2, End, Indices.size() + 1, 2))
2607 Wt = Op->getOperand(1);
2611 // Check odd elements are taken from the odd elements of one half or the
2612 // other and pick an operand accordingly.
2613 if (fitsRegularPattern<int>(Begin + 1, 2, End, 1, 2))
2614 Ws = Op->getOperand(0);
2615 else if (fitsRegularPattern<int>(Begin + 1, 2, End, Indices.size() + 1, 2))
2616 Ws = Op->getOperand(1);
2620 return DAG.getNode(MipsISD::ILVOD, SDLoc(Op), ResTy, Wt, Ws);
2623 // Lower VECTOR_SHUFFLE into ILVR (if possible).
2625 // ILVR interleaves consecutive elements from the right (lowest-indexed) half of
2628 // It is possible to lower into ILVR when the mask consists of two of the
2629 // following forms interleaved:
2631 // <n, n+1, n+2, ...>
2632 // where n is the number of elements in the vector.
2634 // <0, 0, 1, 1, 2, 2, ...>
2635 // <0, n, 1, n+1, 2, n+2, ...>
2637 // When undef's appear in the mask they are treated as if they were whatever
2638 // value is necessary in order to fit the above forms.
2639 static SDValue lowerVECTOR_SHUFFLE_ILVR(SDValue Op, EVT ResTy,
2640 SmallVector<int, 16> Indices,
2641 SelectionDAG &DAG) {
2642 assert((Indices.size() % 2) == 0);
2646 const auto &Begin = Indices.begin();
2647 const auto &End = Indices.end();
2649 // Check even elements are taken from the right (lowest-indexed) elements of
2650 // one half or the other and pick an operand accordingly.
2651 if (fitsRegularPattern<int>(Begin, 2, End, 0, 1))
2652 Wt = Op->getOperand(0);
2653 else if (fitsRegularPattern<int>(Begin, 2, End, Indices.size(), 1))
2654 Wt = Op->getOperand(1);
2658 // Check odd elements are taken from the right (lowest-indexed) elements of
2659 // one half or the other and pick an operand accordingly.
2660 if (fitsRegularPattern<int>(Begin + 1, 2, End, 0, 1))
2661 Ws = Op->getOperand(0);
2662 else if (fitsRegularPattern<int>(Begin + 1, 2, End, Indices.size(), 1))
2663 Ws = Op->getOperand(1);
2667 return DAG.getNode(MipsISD::ILVR, SDLoc(Op), ResTy, Ws, Wt);
2670 // Lower VECTOR_SHUFFLE into ILVL (if possible).
2672 // ILVL interleaves consecutive elements from the left (highest-indexed) half
2675 // It is possible to lower into ILVL when the mask consists of two of the
2676 // following forms interleaved:
2677 // <x, x+1, x+2, ...>
2678 // <n+x, n+x+1, n+x+2, ...>
2679 // where n is the number of elements in the vector and x is half n.
2681 // <x, x, x+1, x+1, x+2, x+2, ...>
2682 // <x, n+x, x+1, n+x+1, x+2, n+x+2, ...>
2684 // When undef's appear in the mask they are treated as if they were whatever
2685 // value is necessary in order to fit the above forms.
2686 static SDValue lowerVECTOR_SHUFFLE_ILVL(SDValue Op, EVT ResTy,
2687 SmallVector<int, 16> Indices,
2688 SelectionDAG &DAG) {
2689 assert((Indices.size() % 2) == 0);
2691 unsigned HalfSize = Indices.size() / 2;
2694 const auto &Begin = Indices.begin();
2695 const auto &End = Indices.end();
2697 // Check even elements are taken from the left (highest-indexed) elements of
2698 // one half or the other and pick an operand accordingly.
2699 if (fitsRegularPattern<int>(Begin, 2, End, HalfSize, 1))
2700 Wt = Op->getOperand(0);
2701 else if (fitsRegularPattern<int>(Begin, 2, End, Indices.size() + HalfSize, 1))
2702 Wt = Op->getOperand(1);
2706 // Check odd elements are taken from the left (highest-indexed) elements of
2707 // one half or the other and pick an operand accordingly.
2708 if (fitsRegularPattern<int>(Begin + 1, 2, End, HalfSize, 1))
2709 Ws = Op->getOperand(0);
2710 else if (fitsRegularPattern<int>(Begin + 1, 2, End, Indices.size() + HalfSize,
2712 Ws = Op->getOperand(1);
2716 return DAG.getNode(MipsISD::ILVL, SDLoc(Op), ResTy, Ws, Wt);
2719 // Lower VECTOR_SHUFFLE into PCKEV (if possible).
2721 // PCKEV copies the even elements of each vector into the result vector.
2723 // It is possible to lower into PCKEV when the mask consists of two of the
2724 // following forms concatenated:
2726 // <n, n+2, n+4, ...>
2727 // where n is the number of elements in the vector.
2729 // <0, 2, 4, ..., 0, 2, 4, ...>
2730 // <0, 2, 4, ..., n, n+2, n+4, ...>
2732 // When undef's appear in the mask they are treated as if they were whatever
2733 // value is necessary in order to fit the above forms.
2734 static SDValue lowerVECTOR_SHUFFLE_PCKEV(SDValue Op, EVT ResTy,
2735 SmallVector<int, 16> Indices,
2736 SelectionDAG &DAG) {
2737 assert((Indices.size() % 2) == 0);
2741 const auto &Begin = Indices.begin();
2742 const auto &Mid = Indices.begin() + Indices.size() / 2;
2743 const auto &End = Indices.end();
2745 if (fitsRegularPattern<int>(Begin, 1, Mid, 0, 2))
2746 Wt = Op->getOperand(0);
2747 else if (fitsRegularPattern<int>(Begin, 1, Mid, Indices.size(), 2))
2748 Wt = Op->getOperand(1);
2752 if (fitsRegularPattern<int>(Mid, 1, End, 0, 2))
2753 Ws = Op->getOperand(0);
2754 else if (fitsRegularPattern<int>(Mid, 1, End, Indices.size(), 2))
2755 Ws = Op->getOperand(1);
2759 return DAG.getNode(MipsISD::PCKEV, SDLoc(Op), ResTy, Ws, Wt);
2762 // Lower VECTOR_SHUFFLE into PCKOD (if possible).
2764 // PCKOD copies the odd elements of each vector into the result vector.
2766 // It is possible to lower into PCKOD when the mask consists of two of the
2767 // following forms concatenated:
2769 // <n+1, n+3, n+5, ...>
2770 // where n is the number of elements in the vector.
2772 // <1, 3, 5, ..., 1, 3, 5, ...>
2773 // <1, 3, 5, ..., n+1, n+3, n+5, ...>
2775 // When undef's appear in the mask they are treated as if they were whatever
2776 // value is necessary in order to fit the above forms.
2777 static SDValue lowerVECTOR_SHUFFLE_PCKOD(SDValue Op, EVT ResTy,
2778 SmallVector<int, 16> Indices,
2779 SelectionDAG &DAG) {
2780 assert((Indices.size() % 2) == 0);
2784 const auto &Begin = Indices.begin();
2785 const auto &Mid = Indices.begin() + Indices.size() / 2;
2786 const auto &End = Indices.end();
2788 if (fitsRegularPattern<int>(Begin, 1, Mid, 1, 2))
2789 Wt = Op->getOperand(0);
2790 else if (fitsRegularPattern<int>(Begin, 1, Mid, Indices.size() + 1, 2))
2791 Wt = Op->getOperand(1);
2795 if (fitsRegularPattern<int>(Mid, 1, End, 1, 2))
2796 Ws = Op->getOperand(0);
2797 else if (fitsRegularPattern<int>(Mid, 1, End, Indices.size() + 1, 2))
2798 Ws = Op->getOperand(1);
2802 return DAG.getNode(MipsISD::PCKOD, SDLoc(Op), ResTy, Ws, Wt);
2805 // Lower VECTOR_SHUFFLE into VSHF.
2807 // This mostly consists of converting the shuffle indices in Indices into a
2808 // BUILD_VECTOR and adding it as an operand to the resulting VSHF. There is
2809 // also code to eliminate unused operands of the VECTOR_SHUFFLE. For example,
2810 // if the type is v8i16 and all the indices are less than 8 then the second
2811 // operand is unused and can be replaced with anything. We choose to replace it
2812 // with the used operand since this reduces the number of instructions overall.
2813 static SDValue lowerVECTOR_SHUFFLE_VSHF(SDValue Op, EVT ResTy,
2814 SmallVector<int, 16> Indices,
2815 SelectionDAG &DAG) {
2816 SmallVector<SDValue, 16> Ops;
2819 EVT MaskVecTy = ResTy.changeVectorElementTypeToInteger();
2820 EVT MaskEltTy = MaskVecTy.getVectorElementType();
2821 bool Using1stVec = false;
2822 bool Using2ndVec = false;
2824 int ResTyNumElts = ResTy.getVectorNumElements();
2826 for (int i = 0; i < ResTyNumElts; ++i) {
2827 // Idx == -1 means UNDEF
2828 int Idx = Indices[i];
2830 if (0 <= Idx && Idx < ResTyNumElts)
2832 if (ResTyNumElts <= Idx && Idx < ResTyNumElts * 2)
2836 for (SmallVector<int, 16>::iterator I = Indices.begin(); I != Indices.end();
2838 Ops.push_back(DAG.getTargetConstant(*I, DL, MaskEltTy));
2840 SDValue MaskVec = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecTy, Ops);
2842 if (Using1stVec && Using2ndVec) {
2843 Op0 = Op->getOperand(0);
2844 Op1 = Op->getOperand(1);
2845 } else if (Using1stVec)
2846 Op0 = Op1 = Op->getOperand(0);
2847 else if (Using2ndVec)
2848 Op0 = Op1 = Op->getOperand(1);
2850 llvm_unreachable("shuffle vector mask references neither vector operand?");
2852 // VECTOR_SHUFFLE concatenates the vectors in an vectorwise fashion.
2853 // <0b00, 0b01> + <0b10, 0b11> -> <0b00, 0b01, 0b10, 0b11>
2854 // VSHF concatenates the vectors in a bitwise fashion:
2855 // <0b00, 0b01> + <0b10, 0b11> ->
2856 // 0b0100 + 0b1110 -> 0b01001110
2857 // <0b10, 0b11, 0b00, 0b01>
2858 // We must therefore swap the operands to get the correct result.
2859 return DAG.getNode(MipsISD::VSHF, DL, ResTy, MaskVec, Op1, Op0);
2862 // Lower VECTOR_SHUFFLE into one of a number of instructions depending on the
2863 // indices in the shuffle.
2864 SDValue MipsSETargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
2865 SelectionDAG &DAG) const {
2866 ShuffleVectorSDNode *Node = cast<ShuffleVectorSDNode>(Op);
2867 EVT ResTy = Op->getValueType(0);
2869 if (!ResTy.is128BitVector())
2872 int ResTyNumElts = ResTy.getVectorNumElements();
2873 SmallVector<int, 16> Indices;
2875 for (int i = 0; i < ResTyNumElts; ++i)
2876 Indices.push_back(Node->getMaskElt(i));
2878 // splati.[bhwd] is preferable to the others but is matched from
2880 if (isVECTOR_SHUFFLE_SPLATI(Op, ResTy, Indices, DAG))
2881 return lowerVECTOR_SHUFFLE_VSHF(Op, ResTy, Indices, DAG);
2882 SDValue Result = lowerVECTOR_SHUFFLE_ILVEV(Op, ResTy, Indices, DAG);
2883 if (Result.getNode())
2885 Result = lowerVECTOR_SHUFFLE_ILVOD(Op, ResTy, Indices, DAG);
2886 if (Result.getNode())
2888 Result = lowerVECTOR_SHUFFLE_ILVL(Op, ResTy, Indices, DAG);
2889 if (Result.getNode())
2891 Result = lowerVECTOR_SHUFFLE_ILVR(Op, ResTy, Indices, DAG);
2892 if (Result.getNode())
2894 Result = lowerVECTOR_SHUFFLE_PCKEV(Op, ResTy, Indices, DAG);
2895 if (Result.getNode())
2897 Result = lowerVECTOR_SHUFFLE_PCKOD(Op, ResTy, Indices, DAG);
2898 if (Result.getNode())
2900 Result = lowerVECTOR_SHUFFLE_SHF(Op, ResTy, Indices, DAG);
2901 if (Result.getNode())
2903 return lowerVECTOR_SHUFFLE_VSHF(Op, ResTy, Indices, DAG);
2906 MachineBasicBlock * MipsSETargetLowering::
2907 emitBPOSGE32(MachineInstr *MI, MachineBasicBlock *BB) const{
2909 // bposge32_pseudo $vr0
2919 // $vr0 = phi($vr2, $fbb, $vr1, $tbb)
2921 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2922 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
2923 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
2924 DebugLoc DL = MI->getDebugLoc();
2925 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2926 MachineFunction::iterator It = std::next(MachineFunction::iterator(BB));
2927 MachineFunction *F = BB->getParent();
2928 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
2929 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
2930 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
2933 F->insert(It, Sink);
2935 // Transfer the remainder of BB and its successor edges to Sink.
2936 Sink->splice(Sink->begin(), BB, std::next(MachineBasicBlock::iterator(MI)),
2938 Sink->transferSuccessorsAndUpdatePHIs(BB);
2941 BB->addSuccessor(FBB);
2942 BB->addSuccessor(TBB);
2943 FBB->addSuccessor(Sink);
2944 TBB->addSuccessor(Sink);
2946 // Insert the real bposge32 instruction to $BB.
2947 BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB);
2950 unsigned VR2 = RegInfo.createVirtualRegister(RC);
2951 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2)
2952 .addReg(Mips::ZERO).addImm(0);
2953 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
2956 unsigned VR1 = RegInfo.createVirtualRegister(RC);
2957 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1)
2958 .addReg(Mips::ZERO).addImm(1);
2960 // Insert phi function to $Sink.
2961 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
2962 MI->getOperand(0).getReg())
2963 .addReg(VR2).addMBB(FBB).addReg(VR1).addMBB(TBB);
2965 MI->eraseFromParent(); // The pseudo instruction is gone now.
2969 MachineBasicBlock * MipsSETargetLowering::
2970 emitMSACBranchPseudo(MachineInstr *MI, MachineBasicBlock *BB,
2971 unsigned BranchOp) const{
2973 // vany_nonzero $rd, $ws
2984 // $rd = phi($rd1, $fbb, $rd2, $tbb)
2986 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
2987 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
2988 const TargetRegisterClass *RC = &Mips::GPR32RegClass;
2989 DebugLoc DL = MI->getDebugLoc();
2990 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2991 MachineFunction::iterator It = std::next(MachineFunction::iterator(BB));
2992 MachineFunction *F = BB->getParent();
2993 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
2994 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
2995 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
2998 F->insert(It, Sink);
3000 // Transfer the remainder of BB and its successor edges to Sink.
3001 Sink->splice(Sink->begin(), BB, std::next(MachineBasicBlock::iterator(MI)),
3003 Sink->transferSuccessorsAndUpdatePHIs(BB);
3006 BB->addSuccessor(FBB);
3007 BB->addSuccessor(TBB);
3008 FBB->addSuccessor(Sink);
3009 TBB->addSuccessor(Sink);
3011 // Insert the real bnz.b instruction to $BB.
3012 BuildMI(BB, DL, TII->get(BranchOp))
3013 .addReg(MI->getOperand(1).getReg())
3017 unsigned RD1 = RegInfo.createVirtualRegister(RC);
3018 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), RD1)
3019 .addReg(Mips::ZERO).addImm(0);
3020 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
3023 unsigned RD2 = RegInfo.createVirtualRegister(RC);
3024 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), RD2)
3025 .addReg(Mips::ZERO).addImm(1);
3027 // Insert phi function to $Sink.
3028 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
3029 MI->getOperand(0).getReg())
3030 .addReg(RD1).addMBB(FBB).addReg(RD2).addMBB(TBB);
3032 MI->eraseFromParent(); // The pseudo instruction is gone now.
3036 // Emit the COPY_FW pseudo instruction.
3038 // copy_fw_pseudo $fd, $ws, n
3040 // copy_u_w $rt, $ws, $n
3043 // When n is zero, the equivalent operation can be performed with (potentially)
3044 // zero instructions due to register overlaps. This optimization is never valid
3045 // for lane 1 because it would require FR=0 mode which isn't supported by MSA.
3046 MachineBasicBlock * MipsSETargetLowering::
3047 emitCOPY_FW(MachineInstr *MI, MachineBasicBlock *BB) const{
3048 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3049 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3050 DebugLoc DL = MI->getDebugLoc();
3051 unsigned Fd = MI->getOperand(0).getReg();
3052 unsigned Ws = MI->getOperand(1).getReg();
3053 unsigned Lane = MI->getOperand(2).getImm();
3057 if (!Subtarget.useOddSPReg()) {
3058 // We must copy to an even-numbered MSA register so that the
3059 // single-precision sub-register is also guaranteed to be even-numbered.
3060 Wt = RegInfo.createVirtualRegister(&Mips::MSA128WEvensRegClass);
3062 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Wt).addReg(Ws);
3065 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_lo);
3067 unsigned Wt = RegInfo.createVirtualRegister(
3068 Subtarget.useOddSPReg() ? &Mips::MSA128WRegClass :
3069 &Mips::MSA128WEvensRegClass);
3071 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wt).addReg(Ws).addImm(Lane);
3072 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_lo);
3075 MI->eraseFromParent(); // The pseudo instruction is gone now.
3079 // Emit the COPY_FD pseudo instruction.
3081 // copy_fd_pseudo $fd, $ws, n
3083 // splati.d $wt, $ws, $n
3084 // copy $fd, $wt:sub_64
3086 // When n is zero, the equivalent operation can be performed with (potentially)
3087 // zero instructions due to register overlaps. This optimization is always
3088 // valid because FR=1 mode which is the only supported mode in MSA.
3089 MachineBasicBlock * MipsSETargetLowering::
3090 emitCOPY_FD(MachineInstr *MI, MachineBasicBlock *BB) const{
3091 assert(Subtarget.isFP64bit());
3093 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3094 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3095 unsigned Fd = MI->getOperand(0).getReg();
3096 unsigned Ws = MI->getOperand(1).getReg();
3097 unsigned Lane = MI->getOperand(2).getImm() * 2;
3098 DebugLoc DL = MI->getDebugLoc();
3101 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_64);
3103 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3105 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wt).addReg(Ws).addImm(1);
3106 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_64);
3109 MI->eraseFromParent(); // The pseudo instruction is gone now.
3113 // Emit the INSERT_FW pseudo instruction.
3115 // insert_fw_pseudo $wd, $wd_in, $n, $fs
3117 // subreg_to_reg $wt:sub_lo, $fs
3118 // insve_w $wd[$n], $wd_in, $wt[0]
3120 MipsSETargetLowering::emitINSERT_FW(MachineInstr *MI,
3121 MachineBasicBlock *BB) const {
3122 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3123 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3124 DebugLoc DL = MI->getDebugLoc();
3125 unsigned Wd = MI->getOperand(0).getReg();
3126 unsigned Wd_in = MI->getOperand(1).getReg();
3127 unsigned Lane = MI->getOperand(2).getImm();
3128 unsigned Fs = MI->getOperand(3).getReg();
3129 unsigned Wt = RegInfo.createVirtualRegister(
3130 Subtarget.useOddSPReg() ? &Mips::MSA128WRegClass :
3131 &Mips::MSA128WEvensRegClass);
3133 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
3136 .addImm(Mips::sub_lo);
3137 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_W), Wd)
3143 MI->eraseFromParent(); // The pseudo instruction is gone now.
3147 // Emit the INSERT_FD pseudo instruction.
3149 // insert_fd_pseudo $wd, $fs, n
3151 // subreg_to_reg $wt:sub_64, $fs
3152 // insve_d $wd[$n], $wd_in, $wt[0]
3154 MipsSETargetLowering::emitINSERT_FD(MachineInstr *MI,
3155 MachineBasicBlock *BB) const {
3156 assert(Subtarget.isFP64bit());
3158 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3159 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3160 DebugLoc DL = MI->getDebugLoc();
3161 unsigned Wd = MI->getOperand(0).getReg();
3162 unsigned Wd_in = MI->getOperand(1).getReg();
3163 unsigned Lane = MI->getOperand(2).getImm();
3164 unsigned Fs = MI->getOperand(3).getReg();
3165 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3167 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
3170 .addImm(Mips::sub_64);
3171 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_D), Wd)
3177 MI->eraseFromParent(); // The pseudo instruction is gone now.
3181 // Emit the INSERT_([BHWD]|F[WD])_VIDX pseudo instruction.
3184 // (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $rs)
3186 // (SLL $lanetmp1, $lane, <log2size)
3187 // (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
3188 // (INSERT_[BHWD], $wdtmp2, $wdtmp1, 0, $rs)
3189 // (NEG $lanetmp2, $lanetmp1)
3190 // (SLD_B $wd, $wdtmp2, $wdtmp2, $lanetmp2)
3192 // For floating point:
3193 // (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $fs)
3195 // (SUBREG_TO_REG $wt, $fs, <subreg>)
3196 // (SLL $lanetmp1, $lane, <log2size)
3197 // (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
3198 // (INSVE_[WD], $wdtmp2, 0, $wdtmp1, 0)
3199 // (NEG $lanetmp2, $lanetmp1)
3200 // (SLD_B $wd, $wdtmp2, $wdtmp2, $lanetmp2)
3202 MipsSETargetLowering::emitINSERT_DF_VIDX(MachineInstr *MI,
3203 MachineBasicBlock *BB,
3204 unsigned EltSizeInBytes,
3206 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3207 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3208 DebugLoc DL = MI->getDebugLoc();
3209 unsigned Wd = MI->getOperand(0).getReg();
3210 unsigned SrcVecReg = MI->getOperand(1).getReg();
3211 unsigned LaneReg = MI->getOperand(2).getReg();
3212 unsigned SrcValReg = MI->getOperand(3).getReg();
3214 const TargetRegisterClass *VecRC = nullptr;
3215 const TargetRegisterClass *GPRRC =
3216 Subtarget.isABI_N64() ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
3217 unsigned EltLog2Size;
3218 unsigned InsertOp = 0;
3219 unsigned InsveOp = 0;
3220 switch (EltSizeInBytes) {
3222 llvm_unreachable("Unexpected size");
3225 InsertOp = Mips::INSERT_B;
3226 InsveOp = Mips::INSVE_B;
3227 VecRC = &Mips::MSA128BRegClass;
3231 InsertOp = Mips::INSERT_H;
3232 InsveOp = Mips::INSVE_H;
3233 VecRC = &Mips::MSA128HRegClass;
3237 InsertOp = Mips::INSERT_W;
3238 InsveOp = Mips::INSVE_W;
3239 VecRC = &Mips::MSA128WRegClass;
3243 InsertOp = Mips::INSERT_D;
3244 InsveOp = Mips::INSVE_D;
3245 VecRC = &Mips::MSA128DRegClass;
3250 unsigned Wt = RegInfo.createVirtualRegister(VecRC);
3251 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
3254 .addImm(EltSizeInBytes == 8 ? Mips::sub_64 : Mips::sub_lo);
3258 // Convert the lane index into a byte index
3259 if (EltSizeInBytes != 1) {
3260 unsigned LaneTmp1 = RegInfo.createVirtualRegister(GPRRC);
3261 BuildMI(*BB, MI, DL, TII->get(Mips::SLL), LaneTmp1)
3263 .addImm(EltLog2Size);
3267 // Rotate bytes around so that the desired lane is element zero
3268 unsigned WdTmp1 = RegInfo.createVirtualRegister(VecRC);
3269 BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), WdTmp1)
3274 unsigned WdTmp2 = RegInfo.createVirtualRegister(VecRC);
3276 // Use insve.df to insert to element zero
3277 BuildMI(*BB, MI, DL, TII->get(InsveOp), WdTmp2)
3283 // Use insert.df to insert to element zero
3284 BuildMI(*BB, MI, DL, TII->get(InsertOp), WdTmp2)
3290 // Rotate elements the rest of the way for a full rotation.
3291 // sld.df inteprets $rt modulo the number of columns so we only need to negate
3292 // the lane index to do this.
3293 unsigned LaneTmp2 = RegInfo.createVirtualRegister(GPRRC);
3294 BuildMI(*BB, MI, DL, TII->get(Subtarget.isABI_N64() ? Mips::DSUB : Mips::SUB),
3296 .addReg(Subtarget.isABI_N64() ? Mips::ZERO_64 : Mips::ZERO)
3298 BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), Wd)
3303 MI->eraseFromParent(); // The pseudo instruction is gone now.
3307 // Emit the FILL_FW pseudo instruction.
3309 // fill_fw_pseudo $wd, $fs
3311 // implicit_def $wt1
3312 // insert_subreg $wt2:subreg_lo, $wt1, $fs
3313 // splati.w $wd, $wt2[0]
3315 MipsSETargetLowering::emitFILL_FW(MachineInstr *MI,
3316 MachineBasicBlock *BB) const {
3317 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3318 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3319 DebugLoc DL = MI->getDebugLoc();
3320 unsigned Wd = MI->getOperand(0).getReg();
3321 unsigned Fs = MI->getOperand(1).getReg();
3322 unsigned Wt1 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3323 unsigned Wt2 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
3325 BuildMI(*BB, MI, DL, TII->get(Mips::IMPLICIT_DEF), Wt1);
3326 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2)
3329 .addImm(Mips::sub_lo);
3330 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wd).addReg(Wt2).addImm(0);
3332 MI->eraseFromParent(); // The pseudo instruction is gone now.
3336 // Emit the FILL_FD pseudo instruction.
3338 // fill_fd_pseudo $wd, $fs
3340 // implicit_def $wt1
3341 // insert_subreg $wt2:subreg_64, $wt1, $fs
3342 // splati.d $wd, $wt2[0]
3344 MipsSETargetLowering::emitFILL_FD(MachineInstr *MI,
3345 MachineBasicBlock *BB) const {
3346 assert(Subtarget.isFP64bit());
3348 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3349 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3350 DebugLoc DL = MI->getDebugLoc();
3351 unsigned Wd = MI->getOperand(0).getReg();
3352 unsigned Fs = MI->getOperand(1).getReg();
3353 unsigned Wt1 = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3354 unsigned Wt2 = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
3356 BuildMI(*BB, MI, DL, TII->get(Mips::IMPLICIT_DEF), Wt1);
3357 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2)
3360 .addImm(Mips::sub_64);
3361 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wd).addReg(Wt2).addImm(0);
3363 MI->eraseFromParent(); // The pseudo instruction is gone now.
3367 // Emit the FEXP2_W_1 pseudo instructions.
3369 // fexp2_w_1_pseudo $wd, $wt
3372 // fexp2.w $wd, $ws, $wt
3374 MipsSETargetLowering::emitFEXP2_W_1(MachineInstr *MI,
3375 MachineBasicBlock *BB) const {
3376 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3377 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3378 const TargetRegisterClass *RC = &Mips::MSA128WRegClass;
3379 unsigned Ws1 = RegInfo.createVirtualRegister(RC);
3380 unsigned Ws2 = RegInfo.createVirtualRegister(RC);
3381 DebugLoc DL = MI->getDebugLoc();
3383 // Splat 1.0 into a vector
3384 BuildMI(*BB, MI, DL, TII->get(Mips::LDI_W), Ws1).addImm(1);
3385 BuildMI(*BB, MI, DL, TII->get(Mips::FFINT_U_W), Ws2).addReg(Ws1);
3387 // Emit 1.0 * fexp2(Wt)
3388 BuildMI(*BB, MI, DL, TII->get(Mips::FEXP2_W), MI->getOperand(0).getReg())
3390 .addReg(MI->getOperand(1).getReg());
3392 MI->eraseFromParent(); // The pseudo instruction is gone now.
3396 // Emit the FEXP2_D_1 pseudo instructions.
3398 // fexp2_d_1_pseudo $wd, $wt
3401 // fexp2.d $wd, $ws, $wt
3403 MipsSETargetLowering::emitFEXP2_D_1(MachineInstr *MI,
3404 MachineBasicBlock *BB) const {
3405 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
3406 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
3407 const TargetRegisterClass *RC = &Mips::MSA128DRegClass;
3408 unsigned Ws1 = RegInfo.createVirtualRegister(RC);
3409 unsigned Ws2 = RegInfo.createVirtualRegister(RC);
3410 DebugLoc DL = MI->getDebugLoc();
3412 // Splat 1.0 into a vector
3413 BuildMI(*BB, MI, DL, TII->get(Mips::LDI_D), Ws1).addImm(1);
3414 BuildMI(*BB, MI, DL, TII->get(Mips::FFINT_U_D), Ws2).addReg(Ws1);
3416 // Emit 1.0 * fexp2(Wt)
3417 BuildMI(*BB, MI, DL, TII->get(Mips::FEXP2_D), MI->getOperand(0).getReg())
3419 .addReg(MI->getOperand(1).getReg());
3421 MI->eraseFromParent(); // The pseudo instruction is gone now.