1 //===-- SIISelLowering.cpp - SI DAG Lowering Implementation ---------------===//
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 //===----------------------------------------------------------------------===//
11 /// \brief Custom DAG lowering for SI
13 //===----------------------------------------------------------------------===//
17 #define _USE_MATH_DEFINES
21 #include "SIISelLowering.h"
23 #include "AMDGPUIntrinsicInfo.h"
24 #include "AMDGPUSubtarget.h"
25 #include "SIInstrInfo.h"
26 #include "SIMachineFunctionInfo.h"
27 #include "SIRegisterInfo.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/SelectionDAG.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/ADT/SmallString.h"
37 SITargetLowering::SITargetLowering(TargetMachine &TM) :
38 AMDGPUTargetLowering(TM) {
39 addRegisterClass(MVT::i1, &AMDGPU::VReg_1RegClass);
40 addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
42 addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
43 addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);
45 addRegisterClass(MVT::i32, &AMDGPU::SReg_32RegClass);
46 addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
48 addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
49 addRegisterClass(MVT::v2i32, &AMDGPU::SReg_64RegClass);
50 addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
52 addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
53 addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
55 addRegisterClass(MVT::v8i32, &AMDGPU::VReg_256RegClass);
56 addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
58 addRegisterClass(MVT::v16i32, &AMDGPU::VReg_512RegClass);
59 addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);
61 computeRegisterProperties();
64 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
65 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
66 setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
67 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
68 setCondCodeAction(ISD::SETULE, MVT::f32, Expand);
69 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
71 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
72 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
73 setCondCodeAction(ISD::SETUGE, MVT::f64, Expand);
74 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
75 setCondCodeAction(ISD::SETULE, MVT::f64, Expand);
76 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
78 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
79 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
80 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
81 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);
83 setOperationAction(ISD::ADD, MVT::i32, Legal);
84 setOperationAction(ISD::ADDC, MVT::i32, Legal);
85 setOperationAction(ISD::ADDE, MVT::i32, Legal);
86 setOperationAction(ISD::SUBC, MVT::i32, Legal);
87 setOperationAction(ISD::SUBE, MVT::i32, Legal);
89 setOperationAction(ISD::FSIN, MVT::f32, Custom);
90 setOperationAction(ISD::FCOS, MVT::f32, Custom);
92 // We need to custom lower vector stores from local memory
93 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
94 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
95 setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
97 setOperationAction(ISD::STORE, MVT::v8i32, Custom);
98 setOperationAction(ISD::STORE, MVT::v16i32, Custom);
100 setOperationAction(ISD::STORE, MVT::i1, Custom);
101 setOperationAction(ISD::STORE, MVT::i32, Custom);
102 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
103 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
105 setOperationAction(ISD::SELECT, MVT::f32, Promote);
106 AddPromotedToType(ISD::SELECT, MVT::f32, MVT::i32);
107 setOperationAction(ISD::SELECT, MVT::i64, Custom);
108 setOperationAction(ISD::SELECT, MVT::f64, Promote);
109 AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
111 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
112 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
113 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
114 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
116 setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
117 setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
119 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
120 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
121 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
123 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
124 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
125 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
127 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
128 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
129 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
131 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Custom);
133 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
135 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
136 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
137 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom);
138 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
140 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
141 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
143 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
144 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Custom);
145 setLoadExtAction(ISD::SEXTLOAD, MVT::i16, Custom);
146 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Expand);
147 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, Expand);
148 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, Expand);
150 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
151 setLoadExtAction(ISD::ZEXTLOAD, MVT::i8, Custom);
152 setLoadExtAction(ISD::ZEXTLOAD, MVT::i16, Custom);
153 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Expand);
155 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
156 setLoadExtAction(ISD::EXTLOAD, MVT::i8, Custom);
157 setLoadExtAction(ISD::EXTLOAD, MVT::i16, Custom);
158 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Expand);
159 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
161 setTruncStoreAction(MVT::i32, MVT::i8, Custom);
162 setTruncStoreAction(MVT::i32, MVT::i16, Custom);
163 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
164 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
165 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Expand);
166 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Expand);
168 setOperationAction(ISD::LOAD, MVT::i1, Custom);
170 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Expand);
171 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
173 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
174 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
175 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
177 // These should use UDIVREM, so set them to expand
178 setOperationAction(ISD::UDIV, MVT::i64, Expand);
179 setOperationAction(ISD::UREM, MVT::i64, Expand);
181 // We only support LOAD/STORE and vector manipulation ops for vectors
182 // with > 4 elements.
184 MVT::v8i32, MVT::v8f32, MVT::v16i32, MVT::v16f32
187 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
188 setOperationAction(ISD::SELECT, MVT::i1, Promote);
190 for (MVT VT : VecTypes) {
191 for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
195 case ISD::BUILD_VECTOR:
197 case ISD::EXTRACT_VECTOR_ELT:
198 case ISD::INSERT_VECTOR_ELT:
199 case ISD::CONCAT_VECTORS:
200 case ISD::INSERT_SUBVECTOR:
201 case ISD::EXTRACT_SUBVECTOR:
204 setOperationAction(Op, VT, Expand);
210 for (int I = MVT::v1f64; I <= MVT::v8f64; ++I) {
211 MVT::SimpleValueType VT = static_cast<MVT::SimpleValueType>(I);
212 setOperationAction(ISD::FTRUNC, VT, Expand);
213 setOperationAction(ISD::FCEIL, VT, Expand);
214 setOperationAction(ISD::FFLOOR, VT, Expand);
217 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
218 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
219 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
220 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
221 setOperationAction(ISD::FRINT, MVT::f64, Legal);
224 // FIXME: These should be removed and handled the same was as f32 fneg. Source
225 // modifiers also work for the double instructions.
226 setOperationAction(ISD::FNEG, MVT::f64, Expand);
227 setOperationAction(ISD::FABS, MVT::f64, Expand);
229 setOperationAction(ISD::FDIV, MVT::f32, Custom);
231 setTargetDAGCombine(ISD::SELECT_CC);
232 setTargetDAGCombine(ISD::SETCC);
234 setTargetDAGCombine(ISD::UINT_TO_FP);
236 setSchedulingPreference(Sched::RegPressure);
239 //===----------------------------------------------------------------------===//
240 // TargetLowering queries
241 //===----------------------------------------------------------------------===//
243 bool SITargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
246 bool *IsFast) const {
250 // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
251 // which isn't a simple VT.
252 if (!VT.isSimple() || VT == MVT::Other)
255 // XXX - CI changes say "Support for unaligned memory accesses" but I don't
256 // see what for specifically. The wording everywhere else seems to be the
259 // XXX - The only mention I see of this in the ISA manual is for LDS direct
260 // reads the "byte address and must be dword aligned". Is it also true for the
261 // normal loads and stores?
262 if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS) {
263 // ds_read/write_b64 require 8-byte alignment, but we can do a 4 byte
264 // aligned, 8 byte access in a single operation using ds_read2/write2_b32
265 // with adjacent offsets.
266 return Align % 4 == 0;
269 // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
270 // byte-address are ignored, thus forcing Dword alignment.
271 // This applies to private, global, and constant memory.
274 return VT.bitsGT(MVT::i32);
277 EVT SITargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
278 unsigned SrcAlign, bool IsMemset,
281 MachineFunction &MF) const {
282 // FIXME: Should account for address space here.
284 // The default fallback uses the private pointer size as a guess for a type to
285 // use. Make sure we switch these to 64-bit accesses.
287 if (Size >= 16 && DstAlign >= 4) // XXX: Should only do for global
290 if (Size >= 8 && DstAlign >= 4)
297 TargetLoweringBase::LegalizeTypeAction
298 SITargetLowering::getPreferredVectorAction(EVT VT) const {
299 if (VT.getVectorNumElements() != 1 && VT.getScalarType().bitsLE(MVT::i16))
300 return TypeSplitVector;
302 return TargetLoweringBase::getPreferredVectorAction(VT);
305 bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
307 const SIInstrInfo *TII =
308 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
309 return TII->isInlineConstant(Imm);
312 SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
313 SDLoc SL, SDValue Chain,
314 unsigned Offset, bool Signed) const {
315 const DataLayout *DL = getDataLayout();
317 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
319 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
320 PointerType *PtrTy = PointerType::get(Ty, AMDGPUAS::CONSTANT_ADDRESS);
321 SDValue BasePtr = DAG.getCopyFromReg(Chain, SL,
322 MRI.getLiveInVirtReg(AMDGPU::SGPR0_SGPR1), MVT::i64);
323 SDValue Ptr = DAG.getNode(ISD::ADD, SL, MVT::i64, BasePtr,
324 DAG.getConstant(Offset, MVT::i64));
325 SDValue PtrOffset = DAG.getUNDEF(getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
326 MachinePointerInfo PtrInfo(UndefValue::get(PtrTy));
328 return DAG.getLoad(ISD::UNINDEXED, Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD,
329 VT, SL, Chain, Ptr, PtrOffset, PtrInfo, MemVT,
331 true, // isNonTemporal
333 DL->getABITypeAlignment(Ty)); // Alignment
336 SDValue SITargetLowering::LowerFormalArguments(
338 CallingConv::ID CallConv,
340 const SmallVectorImpl<ISD::InputArg> &Ins,
341 SDLoc DL, SelectionDAG &DAG,
342 SmallVectorImpl<SDValue> &InVals) const {
344 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
346 MachineFunction &MF = DAG.getMachineFunction();
347 FunctionType *FType = MF.getFunction()->getFunctionType();
348 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
350 assert(CallConv == CallingConv::C);
352 SmallVector<ISD::InputArg, 16> Splits;
353 uint32_t Skipped = 0;
355 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
356 const ISD::InputArg &Arg = Ins[i];
358 // First check if it's a PS input addr
359 if (Info->getShaderType() == ShaderType::PIXEL && !Arg.Flags.isInReg() &&
360 !Arg.Flags.isByVal()) {
362 assert((PSInputNum <= 15) && "Too many PS inputs!");
365 // We can savely skip PS inputs
371 Info->PSInputAddr |= 1 << PSInputNum++;
374 // Second split vertices into their elements
375 if (Info->getShaderType() != ShaderType::COMPUTE && Arg.VT.isVector()) {
376 ISD::InputArg NewArg = Arg;
377 NewArg.Flags.setSplit();
378 NewArg.VT = Arg.VT.getVectorElementType();
380 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
381 // three or five element vertex only needs three or five registers,
382 // NOT four or eigth.
383 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
384 unsigned NumElements = ParamType->getVectorNumElements();
386 for (unsigned j = 0; j != NumElements; ++j) {
387 Splits.push_back(NewArg);
388 NewArg.PartOffset += NewArg.VT.getStoreSize();
391 } else if (Info->getShaderType() != ShaderType::COMPUTE) {
392 Splits.push_back(Arg);
396 SmallVector<CCValAssign, 16> ArgLocs;
397 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
398 getTargetMachine(), ArgLocs, *DAG.getContext());
400 // At least one interpolation mode must be enabled or else the GPU will hang.
401 if (Info->getShaderType() == ShaderType::PIXEL &&
402 (Info->PSInputAddr & 0x7F) == 0) {
403 Info->PSInputAddr |= 1;
404 CCInfo.AllocateReg(AMDGPU::VGPR0);
405 CCInfo.AllocateReg(AMDGPU::VGPR1);
408 // The pointer to the list of arguments is stored in SGPR0, SGPR1
409 // The pointer to the scratch buffer is stored in SGPR2, SGPR3
410 if (Info->getShaderType() == ShaderType::COMPUTE) {
411 Info->NumUserSGPRs = 4;
412 CCInfo.AllocateReg(AMDGPU::SGPR0);
413 CCInfo.AllocateReg(AMDGPU::SGPR1);
414 CCInfo.AllocateReg(AMDGPU::SGPR2);
415 CCInfo.AllocateReg(AMDGPU::SGPR3);
416 MF.addLiveIn(AMDGPU::SGPR0_SGPR1, &AMDGPU::SReg_64RegClass);
417 MF.addLiveIn(AMDGPU::SGPR2_SGPR3, &AMDGPU::SReg_64RegClass);
420 if (Info->getShaderType() == ShaderType::COMPUTE) {
421 getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
425 AnalyzeFormalArguments(CCInfo, Splits);
427 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
429 const ISD::InputArg &Arg = Ins[i];
430 if (Skipped & (1 << i)) {
431 InVals.push_back(DAG.getUNDEF(Arg.VT));
435 CCValAssign &VA = ArgLocs[ArgIdx++];
436 EVT VT = VA.getLocVT();
440 EVT MemVT = Splits[i].VT;
441 // The first 36 bytes of the input buffer contains information about
442 // thread group and global sizes.
443 SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, DAG.getRoot(),
444 36 + VA.getLocMemOffset(),
445 Ins[i].Flags.isSExt());
446 InVals.push_back(Arg);
449 assert(VA.isRegLoc() && "Parameter must be in a register!");
451 unsigned Reg = VA.getLocReg();
453 if (VT == MVT::i64) {
454 // For now assume it is a pointer
455 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
456 &AMDGPU::SReg_64RegClass);
457 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
458 InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
462 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
464 Reg = MF.addLiveIn(Reg, RC);
465 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
467 if (Arg.VT.isVector()) {
469 // Build a vector from the registers
470 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
471 unsigned NumElements = ParamType->getVectorNumElements();
473 SmallVector<SDValue, 4> Regs;
475 for (unsigned j = 1; j != NumElements; ++j) {
476 Reg = ArgLocs[ArgIdx++].getLocReg();
477 Reg = MF.addLiveIn(Reg, RC);
478 Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
481 // Fill up the missing vector elements
482 NumElements = Arg.VT.getVectorNumElements() - NumElements;
483 for (unsigned j = 0; j != NumElements; ++j)
484 Regs.push_back(DAG.getUNDEF(VT));
486 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT, Regs));
490 InVals.push_back(Val);
495 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
496 MachineInstr * MI, MachineBasicBlock * BB) const {
498 MachineBasicBlock::iterator I = *MI;
499 const SIInstrInfo *TII =
500 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
501 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
503 switch (MI->getOpcode()) {
505 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
506 case AMDGPU::BRANCH: return BB;
507 case AMDGPU::SI_ADDR64_RSRC: {
508 unsigned SuperReg = MI->getOperand(0).getReg();
509 unsigned SubRegLo = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
510 unsigned SubRegHi = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
511 unsigned SubRegHiHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
512 unsigned SubRegHiLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
513 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B64), SubRegLo)
514 .addOperand(MI->getOperand(1));
515 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiLo)
517 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiHi)
518 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
519 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SubRegHi)
521 .addImm(AMDGPU::sub0)
523 .addImm(AMDGPU::sub1);
524 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SuperReg)
526 .addImm(AMDGPU::sub0_sub1)
528 .addImm(AMDGPU::sub2_sub3);
529 MI->eraseFromParent();
532 case AMDGPU::SI_BUFFER_RSRC: {
533 unsigned SuperReg = MI->getOperand(0).getReg();
535 for (unsigned i = 0, e = 4; i < e; ++i) {
536 MachineOperand &Arg = MI->getOperand(i + 1);
539 Args[i] = Arg.getReg();
544 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
545 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), Reg)
546 .addImm(Arg.getImm());
549 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE),
552 .addImm(AMDGPU::sub0)
554 .addImm(AMDGPU::sub1)
556 .addImm(AMDGPU::sub2)
558 .addImm(AMDGPU::sub3);
559 MI->eraseFromParent();
562 case AMDGPU::V_SUB_F64: {
563 unsigned DestReg = MI->getOperand(0).getReg();
564 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F64), DestReg)
565 .addImm(0) // SRC0 modifiers
566 .addReg(MI->getOperand(1).getReg())
567 .addImm(1) // SRC1 modifiers
568 .addReg(MI->getOperand(2).getReg())
571 MI->eraseFromParent();
574 case AMDGPU::SI_RegisterStorePseudo: {
575 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
576 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
577 MachineInstrBuilder MIB =
578 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::SI_RegisterStore),
580 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i)
581 MIB.addOperand(MI->getOperand(i));
583 MI->eraseFromParent();
586 case AMDGPU::FABS_SI: {
587 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
588 const SIInstrInfo *TII =
589 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
590 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
591 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
594 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_AND_B32_e32),
595 MI->getOperand(0).getReg())
596 .addReg(MI->getOperand(1).getReg())
598 MI->eraseFromParent();
601 case AMDGPU::FNEG_SI: {
602 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
603 const SIInstrInfo *TII =
604 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
605 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
606 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
609 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_XOR_B32_e32),
610 MI->getOperand(0).getReg())
611 .addReg(MI->getOperand(1).getReg())
613 MI->eraseFromParent();
616 case AMDGPU::FCLAMP_SI: {
617 const SIInstrInfo *TII =
618 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
619 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F32_e64),
620 MI->getOperand(0).getReg())
621 .addImm(0) // SRC0 modifiers
622 .addOperand(MI->getOperand(1))
623 .addImm(0) // SRC1 modifiers
627 MI->eraseFromParent();
633 EVT SITargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
634 if (!VT.isVector()) {
637 return MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
640 MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
644 bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
645 VT = VT.getScalarType();
650 switch (VT.getSimpleVT().SimpleTy) {
652 return false; /* There is V_MAD_F32 for f32 */
662 //===----------------------------------------------------------------------===//
663 // Custom DAG Lowering Operations
664 //===----------------------------------------------------------------------===//
666 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
667 switch (Op.getOpcode()) {
668 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
669 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
670 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
672 SDValue Result = LowerLOAD(Op, DAG);
673 assert((!Result.getNode() ||
674 Result.getNode()->getNumValues() == 2) &&
675 "Load should return a value and a chain");
681 return LowerTrig(Op, DAG);
682 case ISD::SELECT: return LowerSELECT(Op, DAG);
683 case ISD::FDIV: return LowerFDIV(Op, DAG);
684 case ISD::STORE: return LowerSTORE(Op, DAG);
685 case ISD::GlobalAddress: {
686 MachineFunction &MF = DAG.getMachineFunction();
687 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
688 return LowerGlobalAddress(MFI, Op, DAG);
690 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
691 case ISD::INTRINSIC_VOID: return LowerINTRINSIC_VOID(Op, DAG);
696 /// \brief Helper function for LowerBRCOND
697 static SDNode *findUser(SDValue Value, unsigned Opcode) {
699 SDNode *Parent = Value.getNode();
700 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
703 if (I.getUse().get() != Value)
706 if (I->getOpcode() == Opcode)
712 SDValue SITargetLowering::LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const {
714 MachineFunction &MF = DAG.getMachineFunction();
715 const SIInstrInfo *TII =
716 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
717 const SIRegisterInfo &TRI = TII->getRegisterInfo();
718 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Op);
719 unsigned FrameIndex = FINode->getIndex();
721 CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
722 TRI.getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET), MVT::i32);
724 return DAG.getTargetFrameIndex(FrameIndex, MVT::i32);
727 /// This transforms the control flow intrinsics to get the branch destination as
728 /// last parameter, also switches branch target with BR if the need arise
729 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
730 SelectionDAG &DAG) const {
734 SDNode *Intr = BRCOND.getOperand(1).getNode();
735 SDValue Target = BRCOND.getOperand(2);
736 SDNode *BR = nullptr;
738 if (Intr->getOpcode() == ISD::SETCC) {
739 // As long as we negate the condition everything is fine
740 SDNode *SetCC = Intr;
741 assert(SetCC->getConstantOperandVal(1) == 1);
742 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
744 Intr = SetCC->getOperand(0).getNode();
747 // Get the target from BR if we don't negate the condition
748 BR = findUser(BRCOND, ISD::BR);
749 Target = BR->getOperand(1);
752 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
754 // Build the result and
755 SmallVector<EVT, 4> Res;
756 for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
757 Res.push_back(Intr->getValueType(i));
759 // operands of the new intrinsic call
760 SmallVector<SDValue, 4> Ops;
761 Ops.push_back(BRCOND.getOperand(0));
762 for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
763 Ops.push_back(Intr->getOperand(i));
764 Ops.push_back(Target);
766 // build the new intrinsic call
767 SDNode *Result = DAG.getNode(
768 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
769 DAG.getVTList(Res), Ops).getNode();
772 // Give the branch instruction our target
777 SDValue NewBR = DAG.getNode(ISD::BR, DL, BR->getVTList(), Ops);
778 DAG.ReplaceAllUsesWith(BR, NewBR.getNode());
779 BR = NewBR.getNode();
782 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
784 // Copy the intrinsic results to registers
785 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
786 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
790 Chain = DAG.getCopyToReg(
792 CopyToReg->getOperand(1),
793 SDValue(Result, i - 1),
796 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
799 // Remove the old intrinsic from the chain
800 DAG.ReplaceAllUsesOfValueWith(
801 SDValue(Intr, Intr->getNumValues() - 1),
802 Intr->getOperand(0));
807 SDValue SITargetLowering::LowerGlobalAddress(AMDGPUMachineFunction *MFI,
809 SelectionDAG &DAG) const {
810 GlobalAddressSDNode *GSD = cast<GlobalAddressSDNode>(Op);
812 if (GSD->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS)
813 return AMDGPUTargetLowering::LowerGlobalAddress(MFI, Op, DAG);
816 const GlobalValue *GV = GSD->getGlobal();
817 MVT PtrVT = getPointerTy(GSD->getAddressSpace());
819 SDValue Ptr = DAG.getNode(AMDGPUISD::CONST_DATA_PTR, DL, PtrVT);
820 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32);
822 SDValue PtrLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Ptr,
823 DAG.getConstant(0, MVT::i32));
824 SDValue PtrHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Ptr,
825 DAG.getConstant(1, MVT::i32));
827 SDValue Lo = DAG.getNode(ISD::ADDC, DL, DAG.getVTList(MVT::i32, MVT::Glue),
829 SDValue Hi = DAG.getNode(ISD::ADDE, DL, DAG.getVTList(MVT::i32, MVT::Glue),
830 PtrHi, DAG.getConstant(0, MVT::i32),
831 SDValue(Lo.getNode(), 1));
832 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi);
835 SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
836 SelectionDAG &DAG) const {
837 MachineFunction &MF = DAG.getMachineFunction();
838 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
840 EVT VT = Op.getValueType();
842 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
844 switch (IntrinsicID) {
845 case Intrinsic::r600_read_ngroups_x:
846 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 0, false);
847 case Intrinsic::r600_read_ngroups_y:
848 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 4, false);
849 case Intrinsic::r600_read_ngroups_z:
850 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 8, false);
851 case Intrinsic::r600_read_global_size_x:
852 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 12, false);
853 case Intrinsic::r600_read_global_size_y:
854 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 16, false);
855 case Intrinsic::r600_read_global_size_z:
856 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 20, false);
857 case Intrinsic::r600_read_local_size_x:
858 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 24, false);
859 case Intrinsic::r600_read_local_size_y:
860 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 28, false);
861 case Intrinsic::r600_read_local_size_z:
862 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 32, false);
863 case Intrinsic::r600_read_tgid_x:
864 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
865 AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 0), VT);
866 case Intrinsic::r600_read_tgid_y:
867 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
868 AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 1), VT);
869 case Intrinsic::r600_read_tgid_z:
870 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
871 AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 2), VT);
872 case Intrinsic::r600_read_tidig_x:
873 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
875 case Intrinsic::r600_read_tidig_y:
876 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
878 case Intrinsic::r600_read_tidig_z:
879 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
881 case AMDGPUIntrinsic::SI_load_const: {
887 MachineMemOperand *MMO = MF.getMachineMemOperand(
888 MachinePointerInfo(),
889 MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
890 VT.getStoreSize(), 4);
891 return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
892 Op->getVTList(), Ops, VT, MMO);
894 case AMDGPUIntrinsic::SI_sample:
895 return LowerSampleIntrinsic(AMDGPUISD::SAMPLE, Op, DAG);
896 case AMDGPUIntrinsic::SI_sampleb:
897 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEB, Op, DAG);
898 case AMDGPUIntrinsic::SI_sampled:
899 return LowerSampleIntrinsic(AMDGPUISD::SAMPLED, Op, DAG);
900 case AMDGPUIntrinsic::SI_samplel:
901 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEL, Op, DAG);
902 case AMDGPUIntrinsic::SI_vs_load_input:
903 return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
908 return AMDGPUTargetLowering::LowerOperation(Op, DAG);
912 SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
913 SelectionDAG &DAG) const {
914 MachineFunction &MF = DAG.getMachineFunction();
915 SDValue Chain = Op.getOperand(0);
916 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
918 switch (IntrinsicID) {
919 case AMDGPUIntrinsic::SI_tbuffer_store: {
938 EVT VT = Op.getOperand(3).getValueType();
940 MachineMemOperand *MMO = MF.getMachineMemOperand(
941 MachinePointerInfo(),
942 MachineMemOperand::MOStore,
943 VT.getStoreSize(), 4);
944 return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL,
945 Op->getVTList(), Ops, VT, MMO);
952 SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
954 LoadSDNode *Load = cast<LoadSDNode>(Op);
956 if (Op.getValueType().isVector()) {
957 assert(Op.getValueType().getVectorElementType() == MVT::i32 &&
958 "Custom lowering for non-i32 vectors hasn't been implemented.");
959 unsigned NumElements = Op.getValueType().getVectorNumElements();
960 assert(NumElements != 2 && "v2 loads are supported for all address spaces.");
961 switch (Load->getAddressSpace()) {
963 case AMDGPUAS::GLOBAL_ADDRESS:
964 case AMDGPUAS::PRIVATE_ADDRESS:
965 // v4 loads are supported for private and global memory.
966 if (NumElements <= 4)
969 case AMDGPUAS::LOCAL_ADDRESS:
970 return ScalarizeVectorLoad(Op, DAG);
974 return AMDGPUTargetLowering::LowerLOAD(Op, DAG);
977 SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
979 SelectionDAG &DAG) const {
980 return DAG.getNode(Opcode, SDLoc(Op), Op.getValueType(), Op.getOperand(1),
986 SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
987 if (Op.getValueType() != MVT::i64)
991 SDValue Cond = Op.getOperand(0);
993 SDValue Zero = DAG.getConstant(0, MVT::i32);
994 SDValue One = DAG.getConstant(1, MVT::i32);
996 SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1));
997 SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2));
999 SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero);
1000 SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero);
1002 SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1);
1004 SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One);
1005 SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One);
1007 SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1);
1009 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i32, Lo, Hi);
1010 return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
1013 // Catch division cases where we can use shortcuts with rcp and rsq
1015 SDValue SITargetLowering::LowerFastFDIV(SDValue Op, SelectionDAG &DAG) const {
1017 SDValue LHS = Op.getOperand(0);
1018 SDValue RHS = Op.getOperand(1);
1019 EVT VT = Op.getValueType();
1020 bool Unsafe = DAG.getTarget().Options.UnsafeFPMath;
1022 if (const ConstantFPSDNode *CLHS = dyn_cast<ConstantFPSDNode>(LHS)) {
1023 if ((Unsafe || (VT == MVT::f32 && !Subtarget->hasFP32Denormals())) &&
1024 CLHS->isExactlyValue(1.0)) {
1025 // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to
1026 // the CI documentation has a worst case error of 1 ulp.
1027 // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to
1028 // use it as long as we aren't trying to use denormals.
1030 // 1.0 / sqrt(x) -> rsq(x)
1032 // XXX - Is UnsafeFPMath sufficient to do this for f64? The maximum ULP
1033 // error seems really high at 2^29 ULP.
1034 if (RHS.getOpcode() == ISD::FSQRT)
1035 return DAG.getNode(AMDGPUISD::RSQ, SL, VT, RHS.getOperand(0));
1037 // 1.0 / x -> rcp(x)
1038 return DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
1043 // Turn into multiply by the reciprocal.
1044 // x / y -> x * (1.0 / y)
1045 SDValue Recip = DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
1046 return DAG.getNode(ISD::FMUL, SL, VT, LHS, Recip);
1052 SDValue SITargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const {
1053 SDValue FastLowered = LowerFastFDIV(Op, DAG);
1054 if (FastLowered.getNode())
1057 // This uses v_rcp_f32 which does not handle denormals. Let this hit a
1058 // selection error for now rather than do something incorrect.
1059 if (Subtarget->hasFP32Denormals())
1063 SDValue LHS = Op.getOperand(0);
1064 SDValue RHS = Op.getOperand(1);
1066 SDValue r1 = DAG.getNode(ISD::FABS, SL, MVT::f32, RHS);
1068 const APFloat K0Val(BitsToFloat(0x6f800000));
1069 const SDValue K0 = DAG.getConstantFP(K0Val, MVT::f32);
1071 const APFloat K1Val(BitsToFloat(0x2f800000));
1072 const SDValue K1 = DAG.getConstantFP(K1Val, MVT::f32);
1074 const SDValue One = DAG.getTargetConstantFP(1.0, MVT::f32);
1076 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f32);
1078 SDValue r2 = DAG.getSetCC(SL, SetCCVT, r1, K0, ISD::SETOGT);
1080 SDValue r3 = DAG.getNode(ISD::SELECT, SL, MVT::f32, r2, K1, One);
1082 r1 = DAG.getNode(ISD::FMUL, SL, MVT::f32, RHS, r3);
1084 SDValue r0 = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f32, r1);
1086 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f32, LHS, r0);
1088 return DAG.getNode(ISD::FMUL, SL, MVT::f32, r3, Mul);
1091 SDValue SITargetLowering::LowerFDIV64(SDValue Op, SelectionDAG &DAG) const {
1095 SDValue SITargetLowering::LowerFDIV(SDValue Op, SelectionDAG &DAG) const {
1096 EVT VT = Op.getValueType();
1099 return LowerFDIV32(Op, DAG);
1102 return LowerFDIV64(Op, DAG);
1104 llvm_unreachable("Unexpected type for fdiv");
1107 SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1109 StoreSDNode *Store = cast<StoreSDNode>(Op);
1110 EVT VT = Store->getMemoryVT();
1112 // These stores are legal.
1113 if (Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS &&
1114 VT.isVector() && VT.getVectorNumElements() == 2 &&
1115 VT.getVectorElementType() == MVT::i32)
1118 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) {
1119 if (VT.isVector() && VT.getVectorNumElements() > 4)
1120 return ScalarizeVectorStore(Op, DAG);
1124 SDValue Ret = AMDGPUTargetLowering::LowerSTORE(Op, DAG);
1128 if (VT.isVector() && VT.getVectorNumElements() >= 8)
1129 return ScalarizeVectorStore(Op, DAG);
1132 return DAG.getTruncStore(Store->getChain(), DL,
1133 DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32),
1134 Store->getBasePtr(), MVT::i1, Store->getMemOperand());
1139 SDValue SITargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const {
1140 EVT VT = Op.getValueType();
1141 SDValue Arg = Op.getOperand(0);
1142 SDValue FractPart = DAG.getNode(AMDGPUISD::FRACT, SDLoc(Op), VT,
1143 DAG.getNode(ISD::FMUL, SDLoc(Op), VT, Arg,
1144 DAG.getConstantFP(0.5 / M_PI, VT)));
1146 switch (Op.getOpcode()) {
1148 return DAG.getNode(AMDGPUISD::COS_HW, SDLoc(Op), VT, FractPart);
1150 return DAG.getNode(AMDGPUISD::SIN_HW, SDLoc(Op), VT, FractPart);
1152 llvm_unreachable("Wrong trig opcode");
1156 //===----------------------------------------------------------------------===//
1157 // Custom DAG optimizations
1158 //===----------------------------------------------------------------------===//
1160 SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
1161 DAGCombinerInfo &DCI) {
1162 EVT VT = N->getValueType(0);
1163 EVT ScalarVT = VT.getScalarType();
1164 if (ScalarVT != MVT::f32)
1167 SelectionDAG &DAG = DCI.DAG;
1170 SDValue Src = N->getOperand(0);
1171 EVT SrcVT = Src.getValueType();
1173 // TODO: We could try to match extracting the higher bytes, which would be
1174 // easier if i8 vectors weren't promoted to i32 vectors, particularly after
1175 // types are legalized. v4i8 -> v4f32 is probably the only case to worry
1176 // about in practice.
1177 if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) {
1178 if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) {
1179 SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src);
1180 DCI.AddToWorklist(Cvt.getNode());
1185 // We are primarily trying to catch operations on illegal vector types
1186 // before they are expanded.
1187 // For scalars, we can use the more flexible method of checking masked bits
1188 // after legalization.
1189 if (!DCI.isBeforeLegalize() ||
1190 !SrcVT.isVector() ||
1191 SrcVT.getVectorElementType() != MVT::i8) {
1195 assert(DCI.isBeforeLegalize() && "Unexpected legal type");
1197 // Weird sized vectors are a pain to handle, but we know 3 is really the same
1199 unsigned NElts = SrcVT.getVectorNumElements();
1200 if (!SrcVT.isSimple() && NElts != 3)
1203 // Handle v4i8 -> v4f32 extload. Replace the v4i8 with a legal i32 load to
1204 // prevent a mess from expanding to v4i32 and repacking.
1205 if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse()) {
1206 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), SrcVT);
1207 EVT RegVT = getEquivalentLoadRegType(*DAG.getContext(), SrcVT);
1208 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f32, NElts);
1210 LoadSDNode *Load = cast<LoadSDNode>(Src);
1211 SDValue NewLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegVT,
1215 Load->getMemOperand());
1217 // Make sure successors of the original load stay after it by updating
1218 // them to use the new Chain.
1219 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), NewLoad.getValue(1));
1221 SmallVector<SDValue, 4> Elts;
1222 if (RegVT.isVector())
1223 DAG.ExtractVectorElements(NewLoad, Elts);
1225 Elts.push_back(NewLoad);
1227 SmallVector<SDValue, 4> Ops;
1229 unsigned EltIdx = 0;
1230 for (SDValue Elt : Elts) {
1231 unsigned ComponentsInElt = std::min(4u, NElts - 4 * EltIdx);
1232 for (unsigned I = 0; I < ComponentsInElt; ++I) {
1233 unsigned Opc = AMDGPUISD::CVT_F32_UBYTE0 + I;
1234 SDValue Cvt = DAG.getNode(Opc, DL, MVT::f32, Elt);
1235 DCI.AddToWorklist(Cvt.getNode());
1242 assert(Ops.size() == NElts);
1244 return DAG.getNode(ISD::BUILD_VECTOR, DL, FloatVT, Ops);
1250 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
1251 DAGCombinerInfo &DCI) const {
1252 SelectionDAG &DAG = DCI.DAG;
1254 EVT VT = N->getValueType(0);
1256 switch (N->getOpcode()) {
1257 default: return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1259 SDValue Arg0 = N->getOperand(0);
1260 SDValue Arg1 = N->getOperand(1);
1261 SDValue CC = N->getOperand(2);
1262 ConstantSDNode * C = nullptr;
1263 ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
1265 // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
1267 && Arg0.getOpcode() == ISD::SIGN_EXTEND
1268 && Arg0.getOperand(0).getValueType() == MVT::i1
1269 && (C = dyn_cast<ConstantSDNode>(Arg1))
1271 && CCOp == ISD::SETNE) {
1272 return SimplifySetCC(VT, Arg0.getOperand(0),
1273 DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
1278 case AMDGPUISD::CVT_F32_UBYTE0:
1279 case AMDGPUISD::CVT_F32_UBYTE1:
1280 case AMDGPUISD::CVT_F32_UBYTE2:
1281 case AMDGPUISD::CVT_F32_UBYTE3: {
1282 unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
1284 SDValue Src = N->getOperand(0);
1285 APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);
1287 APInt KnownZero, KnownOne;
1288 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1289 !DCI.isBeforeLegalizeOps());
1290 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1291 if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
1292 TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {
1293 DCI.CommitTargetLoweringOpt(TLO);
1299 case ISD::UINT_TO_FP: {
1300 return performUCharToFloatCombine(N, DCI);
1304 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1307 /// \brief Test if RegClass is one of the VSrc classes
1308 static bool isVSrc(unsigned RegClass) {
1309 return AMDGPU::VSrc_32RegClassID == RegClass ||
1310 AMDGPU::VSrc_64RegClassID == RegClass;
1313 /// \brief Test if RegClass is one of the SSrc classes
1314 static bool isSSrc(unsigned RegClass) {
1315 return AMDGPU::SSrc_32RegClassID == RegClass ||
1316 AMDGPU::SSrc_64RegClassID == RegClass;
1319 /// \brief Analyze the possible immediate value Op
1321 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
1322 /// and the immediate value if it's a literal immediate
1323 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
1330 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
1331 if (Node->getZExtValue() >> 32) {
1334 Imm.I = Node->getSExtValue();
1335 } else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
1336 if (N->getValueType(0) != MVT::f32)
1338 Imm.F = Node->getValueAPF().convertToFloat();
1340 return -1; // It isn't an immediate
1342 if ((Imm.I >= -16 && Imm.I <= 64) ||
1343 Imm.F == 0.5f || Imm.F == -0.5f ||
1344 Imm.F == 1.0f || Imm.F == -1.0f ||
1345 Imm.F == 2.0f || Imm.F == -2.0f ||
1346 Imm.F == 4.0f || Imm.F == -4.0f)
1347 return 0; // It's an inline immediate
1349 return Imm.I; // It's a literal immediate
1352 /// \brief Try to fold an immediate directly into an instruction
1353 bool SITargetLowering::foldImm(SDValue &Operand, int32_t &Immediate,
1354 bool &ScalarSlotUsed) const {
1356 MachineSDNode *Mov = dyn_cast<MachineSDNode>(Operand);
1357 const SIInstrInfo *TII =
1358 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1359 if (!Mov || !TII->isMov(Mov->getMachineOpcode()))
1362 const SDValue &Op = Mov->getOperand(0);
1363 int32_t Value = analyzeImmediate(Op.getNode());
1365 // Not an immediate at all
1368 } else if (Value == 0) {
1369 // Inline immediates can always be fold
1373 } else if (Value == Immediate) {
1374 // Already fold literal immediate
1378 } else if (!ScalarSlotUsed && !Immediate) {
1379 // Fold this literal immediate
1380 ScalarSlotUsed = true;
1390 const TargetRegisterClass *SITargetLowering::getRegClassForNode(
1391 SelectionDAG &DAG, const SDValue &Op) const {
1392 const SIInstrInfo *TII =
1393 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1394 const SIRegisterInfo &TRI = TII->getRegisterInfo();
1396 if (!Op->isMachineOpcode()) {
1397 switch(Op->getOpcode()) {
1398 case ISD::CopyFromReg: {
1399 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1400 unsigned Reg = cast<RegisterSDNode>(Op->getOperand(1))->getReg();
1401 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
1402 return MRI.getRegClass(Reg);
1404 return TRI.getPhysRegClass(Reg);
1406 default: return nullptr;
1409 const MCInstrDesc &Desc = TII->get(Op->getMachineOpcode());
1410 int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
1411 if (OpClassID != -1) {
1412 return TRI.getRegClass(OpClassID);
1414 switch(Op.getMachineOpcode()) {
1415 case AMDGPU::COPY_TO_REGCLASS:
1416 // Operand 1 is the register class id for COPY_TO_REGCLASS instructions.
1417 OpClassID = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1419 // If the COPY_TO_REGCLASS instruction is copying to a VSrc register
1420 // class, then the register class for the value could be either a
1421 // VReg or and SReg. In order to get a more accurate
1422 if (OpClassID == AMDGPU::VSrc_32RegClassID ||
1423 OpClassID == AMDGPU::VSrc_64RegClassID) {
1424 return getRegClassForNode(DAG, Op.getOperand(0));
1426 return TRI.getRegClass(OpClassID);
1427 case AMDGPU::EXTRACT_SUBREG: {
1428 int SubIdx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1429 const TargetRegisterClass *SuperClass =
1430 getRegClassForNode(DAG, Op.getOperand(0));
1431 return TRI.getSubClassWithSubReg(SuperClass, SubIdx);
1433 case AMDGPU::REG_SEQUENCE:
1434 // Operand 0 is the register class id for REG_SEQUENCE instructions.
1435 return TRI.getRegClass(
1436 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue());
1438 return getRegClassFor(Op.getSimpleValueType());
1442 /// \brief Does "Op" fit into register class "RegClass" ?
1443 bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, const SDValue &Op,
1444 unsigned RegClass) const {
1445 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1446 const TargetRegisterClass *RC = getRegClassForNode(DAG, Op);
1450 return TRI->getRegClass(RegClass)->hasSubClassEq(RC);
1453 /// \brief Make sure that we don't exeed the number of allowed scalars
1454 void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
1456 bool &ScalarSlotUsed) const {
1458 // First map the operands register class to a destination class
1459 if (RegClass == AMDGPU::VSrc_32RegClassID)
1460 RegClass = AMDGPU::VReg_32RegClassID;
1461 else if (RegClass == AMDGPU::VSrc_64RegClassID)
1462 RegClass = AMDGPU::VReg_64RegClassID;
1466 // Nothing to do if they fit naturally
1467 if (fitsRegClass(DAG, Operand, RegClass))
1470 // If the scalar slot isn't used yet use it now
1471 if (!ScalarSlotUsed) {
1472 ScalarSlotUsed = true;
1476 // This is a conservative aproach. It is possible that we can't determine the
1477 // correct register class and copy too often, but better safe than sorry.
1480 // We can't use COPY_TO_REGCLASS with FrameIndex arguments.
1481 if (isa<FrameIndexSDNode>(Operand)) {
1482 unsigned Opcode = Operand.getValueType() == MVT::i32 ?
1483 AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64;
1484 Node = DAG.getMachineNode(Opcode, SDLoc(), Operand.getValueType(),
1487 SDValue RC = DAG.getTargetConstant(RegClass, MVT::i32);
1488 Node = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, SDLoc(),
1489 Operand.getValueType(), Operand, RC);
1491 Operand = SDValue(Node, 0);
1494 /// \returns true if \p Node's operands are different from the SDValue list
1496 static bool isNodeChanged(const SDNode *Node, const std::vector<SDValue> &Ops) {
1497 for (unsigned i = 0, e = Node->getNumOperands(); i < e; ++i) {
1498 if (Ops[i].getNode() != Node->getOperand(i).getNode()) {
1505 /// \brief Try to fold the Nodes operands into the Node
1506 SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
1507 SelectionDAG &DAG) const {
1509 // Original encoding (either e32 or e64)
1510 int Opcode = Node->getMachineOpcode();
1511 const SIInstrInfo *TII =
1512 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1513 const MCInstrDesc *Desc = &TII->get(Opcode);
1515 unsigned NumDefs = Desc->getNumDefs();
1516 unsigned NumOps = Desc->getNumOperands();
1518 // Commuted opcode if available
1519 int OpcodeRev = Desc->isCommutable() ? TII->commuteOpcode(Opcode) : -1;
1520 const MCInstrDesc *DescRev = OpcodeRev == -1 ? nullptr : &TII->get(OpcodeRev);
1522 assert(!DescRev || DescRev->getNumDefs() == NumDefs);
1523 assert(!DescRev || DescRev->getNumOperands() == NumOps);
1525 // e64 version if available, -1 otherwise
1526 int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
1527 const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? nullptr : &TII->get(OpcodeE64);
1528 int InputModifiers[3] = {0};
1530 assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
1532 int32_t Immediate = Desc->getSize() == 4 ? 0 : -1;
1533 bool HaveVSrc = false, HaveSSrc = false;
1535 // First figure out what we already have in this instruction.
1536 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1537 i != e && Op < NumOps; ++i, ++Op) {
1539 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1540 if (isVSrc(RegClass))
1542 else if (isSSrc(RegClass))
1547 int32_t Imm = analyzeImmediate(Node->getOperand(i).getNode());
1548 if (Imm != -1 && Imm != 0) {
1549 // Literal immediate
1554 // If we neither have VSrc nor SSrc, it makes no sense to continue.
1555 if (!HaveVSrc && !HaveSSrc)
1558 // No scalar allowed when we have both VSrc and SSrc
1559 bool ScalarSlotUsed = HaveVSrc && HaveSSrc;
1561 // Second go over the operands and try to fold them
1562 std::vector<SDValue> Ops;
1563 bool Promote2e64 = false;
1564 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1565 i != e && Op < NumOps; ++i, ++Op) {
1567 const SDValue &Operand = Node->getOperand(i);
1568 Ops.push_back(Operand);
1570 // Already folded immediate?
1571 if (isa<ConstantSDNode>(Operand.getNode()) ||
1572 isa<ConstantFPSDNode>(Operand.getNode()))
1575 // Is this a VSrc or SSrc operand?
1576 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1577 if (isVSrc(RegClass) || isSSrc(RegClass)) {
1578 // Try to fold the immediates
1579 if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
1580 // Folding didn't work, make sure we don't hit the SReg limit.
1581 ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
1585 // If it's not a VSrc or SSrc operand check if we have a GlobalAddress.
1586 // These will be lowered to immediates, so we will need to insert a MOV.
1587 if (isa<GlobalAddressSDNode>(Ops[i])) {
1588 SDNode *Node = DAG.getMachineNode(AMDGPU::V_MOV_B32_e32, SDLoc(),
1589 Operand.getValueType(), Operand);
1590 Ops[i] = SDValue(Node, 0);
1594 if (i == 1 && DescRev && fitsRegClass(DAG, Ops[0], RegClass)) {
1596 unsigned OtherRegClass = Desc->OpInfo[NumDefs].RegClass;
1597 assert(isVSrc(OtherRegClass) || isSSrc(OtherRegClass));
1599 // Test if it makes sense to swap operands
1600 if (foldImm(Ops[1], Immediate, ScalarSlotUsed) ||
1601 (!fitsRegClass(DAG, Ops[1], RegClass) &&
1602 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1604 // Swap commutable operands
1605 std::swap(Ops[0], Ops[1]);
1617 // Test if it makes sense to switch to e64 encoding
1618 unsigned OtherRegClass = DescE64->OpInfo[Op].RegClass;
1619 if (!isVSrc(OtherRegClass) && !isSSrc(OtherRegClass))
1622 int32_t TmpImm = -1;
1623 if (foldImm(Ops[i], TmpImm, ScalarSlotUsed) ||
1624 (!fitsRegClass(DAG, Ops[i], RegClass) &&
1625 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1627 // Switch to e64 encoding
1635 if (!DescE64 && !Promote2e64)
1637 if (!Operand.isMachineOpcode())
1642 std::vector<SDValue> OldOps(Ops);
1644 bool HasModifiers = TII->hasModifiers(Desc->Opcode);
1645 for (unsigned i = 0; i < OldOps.size(); ++i) {
1648 Ops.push_back(DAG.getTargetConstant(InputModifiers[i], MVT::i32));
1649 Ops.push_back(OldOps[i]);
1651 // Add the modifier flags while promoting
1653 for (unsigned i = 0; i < 2; ++i)
1654 Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
1658 // Add optional chain and glue
1659 for (unsigned i = NumOps - NumDefs, e = Node->getNumOperands(); i < e; ++i)
1660 Ops.push_back(Node->getOperand(i));
1662 // Nodes that have a glue result are not CSE'd by getMachineNode(), so in
1663 // this case a brand new node is always be created, even if the operands
1664 // are the same as before. So, manually check if anything has been changed.
1665 if (Desc->Opcode == Opcode && !isNodeChanged(Node, Ops)) {
1669 // Create a complete new instruction
1670 return DAG.getMachineNode(Desc->Opcode, SDLoc(Node), Node->getVTList(), Ops);
1673 /// \brief Helper function for adjustWritemask
1674 static unsigned SubIdx2Lane(unsigned Idx) {
1677 case AMDGPU::sub0: return 0;
1678 case AMDGPU::sub1: return 1;
1679 case AMDGPU::sub2: return 2;
1680 case AMDGPU::sub3: return 3;
1684 /// \brief Adjust the writemask of MIMG instructions
1685 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
1686 SelectionDAG &DAG) const {
1687 SDNode *Users[4] = { };
1689 unsigned OldDmask = Node->getConstantOperandVal(0);
1690 unsigned NewDmask = 0;
1692 // Try to figure out the used register components
1693 for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
1696 // Abort if we can't understand the usage
1697 if (!I->isMachineOpcode() ||
1698 I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
1701 // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used.
1702 // Note that subregs are packed, i.e. Lane==0 is the first bit set
1703 // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit
1705 Lane = SubIdx2Lane(I->getConstantOperandVal(1));
1707 // Set which texture component corresponds to the lane.
1709 for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) {
1711 Comp = countTrailingZeros(Dmask);
1712 Dmask &= ~(1 << Comp);
1715 // Abort if we have more than one user per component
1720 NewDmask |= 1 << Comp;
1723 // Abort if there's no change
1724 if (NewDmask == OldDmask)
1727 // Adjust the writemask in the node
1728 std::vector<SDValue> Ops;
1729 Ops.push_back(DAG.getTargetConstant(NewDmask, MVT::i32));
1730 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
1731 Ops.push_back(Node->getOperand(i));
1732 Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops);
1734 // If we only got one lane, replace it with a copy
1735 // (if NewDmask has only one bit set...)
1736 if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
1737 SDValue RC = DAG.getTargetConstant(AMDGPU::VReg_32RegClassID, MVT::i32);
1738 SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1739 SDLoc(), Users[Lane]->getValueType(0),
1740 SDValue(Node, 0), RC);
1741 DAG.ReplaceAllUsesWith(Users[Lane], Copy);
1745 // Update the users of the node with the new indices
1746 for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
1748 SDNode *User = Users[i];
1752 SDValue Op = DAG.getTargetConstant(Idx, MVT::i32);
1753 DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
1757 case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
1758 case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
1759 case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
1764 /// \brief Fold the instructions after selecting them.
1765 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
1766 SelectionDAG &DAG) const {
1767 const SIInstrInfo *TII =
1768 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1769 Node = AdjustRegClass(Node, DAG);
1771 if (TII->isMIMG(Node->getMachineOpcode()))
1772 adjustWritemask(Node, DAG);
1774 return foldOperands(Node, DAG);
1777 /// \brief Assign the register class depending on the number of
1778 /// bits set in the writemask
1779 void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
1780 SDNode *Node) const {
1781 const SIInstrInfo *TII =
1782 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1783 if (!TII->isMIMG(MI->getOpcode()))
1786 unsigned VReg = MI->getOperand(0).getReg();
1787 unsigned Writemask = MI->getOperand(1).getImm();
1788 unsigned BitsSet = 0;
1789 for (unsigned i = 0; i < 4; ++i)
1790 BitsSet += Writemask & (1 << i) ? 1 : 0;
1792 const TargetRegisterClass *RC;
1795 case 1: RC = &AMDGPU::VReg_32RegClass; break;
1796 case 2: RC = &AMDGPU::VReg_64RegClass; break;
1797 case 3: RC = &AMDGPU::VReg_96RegClass; break;
1800 unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
1801 MI->setDesc(TII->get(NewOpcode));
1802 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1803 MRI.setRegClass(VReg, RC);
1806 MachineSDNode *SITargetLowering::AdjustRegClass(MachineSDNode *N,
1807 SelectionDAG &DAG) const {
1810 unsigned NewOpcode = N->getMachineOpcode();
1812 switch (N->getMachineOpcode()) {
1814 case AMDGPU::S_LOAD_DWORD_IMM:
1815 NewOpcode = AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1817 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1818 if (NewOpcode == N->getMachineOpcode()) {
1819 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1822 case AMDGPU::S_LOAD_DWORDX4_IMM:
1823 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1824 if (NewOpcode == N->getMachineOpcode()) {
1825 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1827 if (fitsRegClass(DAG, N->getOperand(0), AMDGPU::SReg_64RegClassID)) {
1830 ConstantSDNode *Offset = cast<ConstantSDNode>(N->getOperand(1));
1832 SDValue(DAG.getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::i128,
1833 DAG.getConstant(0, MVT::i64)), 0),
1835 DAG.getConstant(Offset->getSExtValue() << 2, MVT::i32)
1837 return DAG.getMachineNode(NewOpcode, DL, N->getVTList(), Ops);
1842 SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1843 const TargetRegisterClass *RC,
1844 unsigned Reg, EVT VT) const {
1845 SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT);
1847 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()),
1848 cast<RegisterSDNode>(VReg)->getReg(), VT);