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::v2i32, Custom);
94 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
95 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
96 setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
98 setOperationAction(ISD::STORE, MVT::v8i32, Custom);
99 setOperationAction(ISD::STORE, MVT::v16i32, Custom);
101 // We need to custom lower loads/stores from private memory
102 setOperationAction(ISD::LOAD, MVT::i32, Custom);
103 setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
104 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
105 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
107 setOperationAction(ISD::STORE, MVT::i1, Custom);
108 setOperationAction(ISD::STORE, MVT::i32, Custom);
109 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
110 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
112 setOperationAction(ISD::SELECT, MVT::f32, Promote);
113 AddPromotedToType(ISD::SELECT, MVT::f32, MVT::i32);
114 setOperationAction(ISD::SELECT, MVT::i64, Custom);
115 setOperationAction(ISD::SELECT, MVT::f64, Promote);
116 AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
118 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
119 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
120 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
121 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
123 setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
124 setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
127 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
128 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
130 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
131 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
132 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
134 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
135 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
136 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
138 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Custom);
140 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
142 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
143 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
144 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom);
145 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
147 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
148 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
150 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
151 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Custom);
152 setLoadExtAction(ISD::SEXTLOAD, MVT::i16, Custom);
153 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Expand);
154 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, Expand);
155 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, Expand);
157 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
158 setLoadExtAction(ISD::ZEXTLOAD, MVT::i8, Custom);
159 setLoadExtAction(ISD::ZEXTLOAD, MVT::i16, Custom);
160 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Expand);
162 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
163 setLoadExtAction(ISD::EXTLOAD, MVT::i8, Custom);
164 setLoadExtAction(ISD::EXTLOAD, MVT::i16, Custom);
165 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Expand);
166 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
168 setTruncStoreAction(MVT::i32, MVT::i8, Custom);
169 setTruncStoreAction(MVT::i32, MVT::i16, Custom);
170 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
171 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
172 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Expand);
173 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Expand);
175 setOperationAction(ISD::LOAD, MVT::i1, Custom);
177 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Expand);
178 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
180 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
181 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
182 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
184 // These should use UDIVREM, so set them to expand
185 setOperationAction(ISD::UDIV, MVT::i64, Expand);
186 setOperationAction(ISD::UREM, MVT::i64, Expand);
188 // We only support LOAD/STORE and vector manipulation ops for vectors
189 // with > 4 elements.
191 MVT::v8i32, MVT::v8f32, MVT::v16i32, MVT::v16f32
194 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
195 setOperationAction(ISD::SELECT, MVT::i1, Promote);
197 for (MVT VT : VecTypes) {
198 for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
202 case ISD::BUILD_VECTOR:
204 case ISD::EXTRACT_VECTOR_ELT:
205 case ISD::INSERT_VECTOR_ELT:
206 case ISD::CONCAT_VECTORS:
207 case ISD::INSERT_SUBVECTOR:
208 case ISD::EXTRACT_SUBVECTOR:
211 setOperationAction(Op, VT, Expand);
217 for (int I = MVT::v1f64; I <= MVT::v8f64; ++I) {
218 MVT::SimpleValueType VT = static_cast<MVT::SimpleValueType>(I);
219 setOperationAction(ISD::FTRUNC, VT, Expand);
220 setOperationAction(ISD::FCEIL, VT, Expand);
221 setOperationAction(ISD::FFLOOR, VT, Expand);
224 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
225 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
226 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
227 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
228 setOperationAction(ISD::FRINT, MVT::f64, Legal);
231 // FIXME: These should be removed and handled the same was as f32 fneg. Source
232 // modifiers also work for the double instructions.
233 setOperationAction(ISD::FNEG, MVT::f64, Expand);
234 setOperationAction(ISD::FABS, MVT::f64, Expand);
236 setOperationAction(ISD::FDIV, MVT::f32, Custom);
238 setTargetDAGCombine(ISD::SELECT_CC);
239 setTargetDAGCombine(ISD::SETCC);
241 setTargetDAGCombine(ISD::UINT_TO_FP);
243 setSchedulingPreference(Sched::RegPressure);
246 //===----------------------------------------------------------------------===//
247 // TargetLowering queries
248 //===----------------------------------------------------------------------===//
250 bool SITargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
252 bool *IsFast) const {
256 // XXX: This depends on the address space and also we may want to revist
257 // the alignment values we specify in the DataLayout.
259 // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
260 // which isn't a simple VT.
261 if (!VT.isSimple() || VT == MVT::Other)
264 // XXX - CI changes say "Support for unaligned memory accesses" but I don't
265 // see what for specifically. The wording everywhere else seems to be the
268 // 3.6.4 - Operations using pairs of VGPRs (for example: double-floats) have
269 // no alignment restrictions.
270 if (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS) {
271 // Using any pair of GPRs should be the same as any other pair.
274 return VT.bitsGE(MVT::i64);
277 // XXX - The only mention I see of this in the ISA manual is for LDS direct
278 // reads the "byte address and must be dword aligned". Is it also true for the
279 // normal loads and stores?
280 if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS)
283 // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
284 // byte-address are ignored, thus forcing Dword alignment.
287 return VT.bitsGT(MVT::i32);
290 TargetLoweringBase::LegalizeTypeAction
291 SITargetLowering::getPreferredVectorAction(EVT VT) const {
292 if (VT.getVectorNumElements() != 1 && VT.getScalarType().bitsLE(MVT::i16))
293 return TypeSplitVector;
295 return TargetLoweringBase::getPreferredVectorAction(VT);
298 bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
300 const SIInstrInfo *TII =
301 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
302 return TII->isInlineConstant(Imm);
305 SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
306 SDLoc DL, SDValue Chain,
307 unsigned Offset, bool Signed) const {
308 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
309 PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
310 AMDGPUAS::CONSTANT_ADDRESS);
311 SDValue BasePtr = DAG.getCopyFromReg(Chain, DL,
312 MRI.getLiveInVirtReg(AMDGPU::SGPR0_SGPR1), MVT::i64);
313 SDValue Ptr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
314 DAG.getConstant(Offset, MVT::i64));
315 return DAG.getExtLoad(Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD, DL, VT, Chain, Ptr,
316 MachinePointerInfo(UndefValue::get(PtrTy)), MemVT,
317 false, false, MemVT.getSizeInBits() >> 3);
321 SDValue SITargetLowering::LowerFormalArguments(
323 CallingConv::ID CallConv,
325 const SmallVectorImpl<ISD::InputArg> &Ins,
326 SDLoc DL, SelectionDAG &DAG,
327 SmallVectorImpl<SDValue> &InVals) const {
329 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
331 MachineFunction &MF = DAG.getMachineFunction();
332 FunctionType *FType = MF.getFunction()->getFunctionType();
333 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
335 assert(CallConv == CallingConv::C);
337 SmallVector<ISD::InputArg, 16> Splits;
338 uint32_t Skipped = 0;
340 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
341 const ISD::InputArg &Arg = Ins[i];
343 // First check if it's a PS input addr
344 if (Info->getShaderType() == ShaderType::PIXEL && !Arg.Flags.isInReg() &&
345 !Arg.Flags.isByVal()) {
347 assert((PSInputNum <= 15) && "Too many PS inputs!");
350 // We can savely skip PS inputs
356 Info->PSInputAddr |= 1 << PSInputNum++;
359 // Second split vertices into their elements
360 if (Info->getShaderType() != ShaderType::COMPUTE && Arg.VT.isVector()) {
361 ISD::InputArg NewArg = Arg;
362 NewArg.Flags.setSplit();
363 NewArg.VT = Arg.VT.getVectorElementType();
365 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
366 // three or five element vertex only needs three or five registers,
367 // NOT four or eigth.
368 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
369 unsigned NumElements = ParamType->getVectorNumElements();
371 for (unsigned j = 0; j != NumElements; ++j) {
372 Splits.push_back(NewArg);
373 NewArg.PartOffset += NewArg.VT.getStoreSize();
376 } else if (Info->getShaderType() != ShaderType::COMPUTE) {
377 Splits.push_back(Arg);
381 SmallVector<CCValAssign, 16> ArgLocs;
382 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
383 getTargetMachine(), ArgLocs, *DAG.getContext());
385 // At least one interpolation mode must be enabled or else the GPU will hang.
386 if (Info->getShaderType() == ShaderType::PIXEL &&
387 (Info->PSInputAddr & 0x7F) == 0) {
388 Info->PSInputAddr |= 1;
389 CCInfo.AllocateReg(AMDGPU::VGPR0);
390 CCInfo.AllocateReg(AMDGPU::VGPR1);
393 // The pointer to the list of arguments is stored in SGPR0, SGPR1
394 if (Info->getShaderType() == ShaderType::COMPUTE) {
395 CCInfo.AllocateReg(AMDGPU::SGPR0);
396 CCInfo.AllocateReg(AMDGPU::SGPR1);
397 MF.addLiveIn(AMDGPU::SGPR0_SGPR1, &AMDGPU::SReg_64RegClass);
400 if (Info->getShaderType() == ShaderType::COMPUTE) {
401 getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
405 AnalyzeFormalArguments(CCInfo, Splits);
407 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
409 const ISD::InputArg &Arg = Ins[i];
410 if (Skipped & (1 << i)) {
411 InVals.push_back(DAG.getUNDEF(Arg.VT));
415 CCValAssign &VA = ArgLocs[ArgIdx++];
416 EVT VT = VA.getLocVT();
420 EVT MemVT = Splits[i].VT;
421 // The first 36 bytes of the input buffer contains information about
422 // thread group and global sizes.
423 SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, DAG.getRoot(),
424 36 + VA.getLocMemOffset(),
425 Ins[i].Flags.isSExt());
426 InVals.push_back(Arg);
429 assert(VA.isRegLoc() && "Parameter must be in a register!");
431 unsigned Reg = VA.getLocReg();
433 if (VT == MVT::i64) {
434 // For now assume it is a pointer
435 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
436 &AMDGPU::SReg_64RegClass);
437 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
438 InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
442 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
444 Reg = MF.addLiveIn(Reg, RC);
445 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
447 if (Arg.VT.isVector()) {
449 // Build a vector from the registers
450 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
451 unsigned NumElements = ParamType->getVectorNumElements();
453 SmallVector<SDValue, 4> Regs;
455 for (unsigned j = 1; j != NumElements; ++j) {
456 Reg = ArgLocs[ArgIdx++].getLocReg();
457 Reg = MF.addLiveIn(Reg, RC);
458 Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
461 // Fill up the missing vector elements
462 NumElements = Arg.VT.getVectorNumElements() - NumElements;
463 for (unsigned j = 0; j != NumElements; ++j)
464 Regs.push_back(DAG.getUNDEF(VT));
466 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT, Regs));
470 InVals.push_back(Val);
475 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
476 MachineInstr * MI, MachineBasicBlock * BB) const {
478 MachineBasicBlock::iterator I = *MI;
479 const SIInstrInfo *TII =
480 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
481 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
483 switch (MI->getOpcode()) {
485 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
486 case AMDGPU::BRANCH: return BB;
487 case AMDGPU::SI_ADDR64_RSRC: {
488 unsigned SuperReg = MI->getOperand(0).getReg();
489 unsigned SubRegLo = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
490 unsigned SubRegHi = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
491 unsigned SubRegHiHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
492 unsigned SubRegHiLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
493 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B64), SubRegLo)
494 .addOperand(MI->getOperand(1));
495 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiLo)
497 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiHi)
498 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
499 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SubRegHi)
501 .addImm(AMDGPU::sub0)
503 .addImm(AMDGPU::sub1);
504 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SuperReg)
506 .addImm(AMDGPU::sub0_sub1)
508 .addImm(AMDGPU::sub2_sub3);
509 MI->eraseFromParent();
512 case AMDGPU::V_SUB_F64: {
513 unsigned DestReg = MI->getOperand(0).getReg();
514 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F64), DestReg)
515 .addImm(0) // SRC0 modifiers
516 .addReg(MI->getOperand(1).getReg())
517 .addImm(1) // SRC1 modifiers
518 .addReg(MI->getOperand(2).getReg())
519 .addImm(0) // SRC2 modifiers
523 MI->eraseFromParent();
526 case AMDGPU::SI_RegisterStorePseudo: {
527 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
528 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
529 MachineInstrBuilder MIB =
530 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::SI_RegisterStore),
532 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i)
533 MIB.addOperand(MI->getOperand(i));
535 MI->eraseFromParent();
538 case AMDGPU::FABS_SI: {
539 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
540 const SIInstrInfo *TII =
541 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
542 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
543 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
546 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_AND_B32_e32),
547 MI->getOperand(0).getReg())
548 .addReg(MI->getOperand(1).getReg())
550 MI->eraseFromParent();
553 case AMDGPU::FNEG_SI: {
554 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
555 const SIInstrInfo *TII =
556 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
557 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
558 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
561 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_XOR_B32_e32),
562 MI->getOperand(0).getReg())
563 .addReg(MI->getOperand(1).getReg())
565 MI->eraseFromParent();
568 case AMDGPU::FCLAMP_SI: {
569 const SIInstrInfo *TII =
570 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
571 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F32_e64),
572 MI->getOperand(0).getReg())
573 .addImm(0) // SRC0 modifiers
574 .addOperand(MI->getOperand(1))
575 .addImm(0) // SRC1 modifiers
579 MI->eraseFromParent();
585 EVT SITargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
586 if (!VT.isVector()) {
589 return MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
592 MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
596 bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
597 VT = VT.getScalarType();
602 switch (VT.getSimpleVT().SimpleTy) {
604 return false; /* There is V_MAD_F32 for f32 */
614 //===----------------------------------------------------------------------===//
615 // Custom DAG Lowering Operations
616 //===----------------------------------------------------------------------===//
618 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
619 MachineFunction &MF = DAG.getMachineFunction();
620 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
621 switch (Op.getOpcode()) {
622 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
623 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
625 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
626 EVT VT = Op.getValueType();
628 // These loads are legal.
629 if (Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS &&
630 VT.isVector() && VT.getVectorNumElements() == 2 &&
631 VT.getVectorElementType() == MVT::i32)
634 if (Op.getValueType().isVector() &&
635 (Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
636 Load->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS ||
637 (Load->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
638 Op.getValueType().getVectorNumElements() > 4))) {
639 return SplitVectorLoad(Op, DAG);
641 SDValue Result = LowerLOAD(Op, DAG);
642 assert((!Result.getNode() ||
643 Result.getNode()->getNumValues() == 2) &&
644 "Load should return a value and a chain");
651 return LowerTrig(Op, DAG);
652 case ISD::SELECT: return LowerSELECT(Op, DAG);
653 case ISD::FDIV: return LowerFDIV(Op, DAG);
654 case ISD::STORE: return LowerSTORE(Op, DAG);
655 case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
656 case ISD::INTRINSIC_WO_CHAIN: {
657 unsigned IntrinsicID =
658 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
659 EVT VT = Op.getValueType();
661 //XXX: Hardcoded we only use two to store the pointer to the parameters.
662 unsigned NumUserSGPRs = 2;
663 switch (IntrinsicID) {
664 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
665 case Intrinsic::r600_read_ngroups_x:
666 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 0, false);
667 case Intrinsic::r600_read_ngroups_y:
668 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 4, false);
669 case Intrinsic::r600_read_ngroups_z:
670 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 8, false);
671 case Intrinsic::r600_read_global_size_x:
672 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 12, false);
673 case Intrinsic::r600_read_global_size_y:
674 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 16, false);
675 case Intrinsic::r600_read_global_size_z:
676 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 20, false);
677 case Intrinsic::r600_read_local_size_x:
678 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 24, false);
679 case Intrinsic::r600_read_local_size_y:
680 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 28, false);
681 case Intrinsic::r600_read_local_size_z:
682 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 32, false);
683 case Intrinsic::r600_read_tgid_x:
684 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
685 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 0), VT);
686 case Intrinsic::r600_read_tgid_y:
687 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
688 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 1), VT);
689 case Intrinsic::r600_read_tgid_z:
690 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
691 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 2), VT);
692 case Intrinsic::r600_read_tidig_x:
693 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
695 case Intrinsic::r600_read_tidig_y:
696 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
698 case Intrinsic::r600_read_tidig_z:
699 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
701 case AMDGPUIntrinsic::SI_load_const: {
707 MachineMemOperand *MMO = MF.getMachineMemOperand(
708 MachinePointerInfo(),
709 MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
710 VT.getSizeInBits() / 8, 4);
711 return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
712 Op->getVTList(), Ops, VT, MMO);
714 case AMDGPUIntrinsic::SI_sample:
715 return LowerSampleIntrinsic(AMDGPUISD::SAMPLE, Op, DAG);
716 case AMDGPUIntrinsic::SI_sampleb:
717 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEB, Op, DAG);
718 case AMDGPUIntrinsic::SI_sampled:
719 return LowerSampleIntrinsic(AMDGPUISD::SAMPLED, Op, DAG);
720 case AMDGPUIntrinsic::SI_samplel:
721 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEL, Op, DAG);
722 case AMDGPUIntrinsic::SI_vs_load_input:
723 return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
730 case ISD::INTRINSIC_VOID:
731 SDValue Chain = Op.getOperand(0);
732 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
734 switch (IntrinsicID) {
735 case AMDGPUIntrinsic::SI_tbuffer_store: {
753 EVT VT = Op.getOperand(3).getValueType();
755 MachineMemOperand *MMO = MF.getMachineMemOperand(
756 MachinePointerInfo(),
757 MachineMemOperand::MOStore,
758 VT.getSizeInBits() / 8, 4);
759 return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL,
760 Op->getVTList(), Ops, VT, MMO);
769 /// \brief Helper function for LowerBRCOND
770 static SDNode *findUser(SDValue Value, unsigned Opcode) {
772 SDNode *Parent = Value.getNode();
773 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
776 if (I.getUse().get() != Value)
779 if (I->getOpcode() == Opcode)
785 /// This transforms the control flow intrinsics to get the branch destination as
786 /// last parameter, also switches branch target with BR if the need arise
787 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
788 SelectionDAG &DAG) const {
792 SDNode *Intr = BRCOND.getOperand(1).getNode();
793 SDValue Target = BRCOND.getOperand(2);
794 SDNode *BR = nullptr;
796 if (Intr->getOpcode() == ISD::SETCC) {
797 // As long as we negate the condition everything is fine
798 SDNode *SetCC = Intr;
799 assert(SetCC->getConstantOperandVal(1) == 1);
800 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
802 Intr = SetCC->getOperand(0).getNode();
805 // Get the target from BR if we don't negate the condition
806 BR = findUser(BRCOND, ISD::BR);
807 Target = BR->getOperand(1);
810 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
812 // Build the result and
813 SmallVector<EVT, 4> Res;
814 for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
815 Res.push_back(Intr->getValueType(i));
817 // operands of the new intrinsic call
818 SmallVector<SDValue, 4> Ops;
819 Ops.push_back(BRCOND.getOperand(0));
820 for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
821 Ops.push_back(Intr->getOperand(i));
822 Ops.push_back(Target);
824 // build the new intrinsic call
825 SDNode *Result = DAG.getNode(
826 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
827 DAG.getVTList(Res), Ops).getNode();
830 // Give the branch instruction our target
835 DAG.MorphNodeTo(BR, ISD::BR, BR->getVTList(), Ops);
838 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
840 // Copy the intrinsic results to registers
841 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
842 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
846 Chain = DAG.getCopyToReg(
848 CopyToReg->getOperand(1),
849 SDValue(Result, i - 1),
852 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
855 // Remove the old intrinsic from the chain
856 DAG.ReplaceAllUsesOfValueWith(
857 SDValue(Intr, Intr->getNumValues() - 1),
858 Intr->getOperand(0));
863 SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
865 LoadSDNode *Load = cast<LoadSDNode>(Op);
866 SDValue Lowered = AMDGPUTargetLowering::LowerLOAD(Op, DAG);
867 if (Lowered.getNode())
870 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
874 EVT MemVT = Load->getMemoryVT();
876 assert(!MemVT.isVector() && "Private loads should be scalarized");
877 assert(!MemVT.isFloatingPoint() && "FP loads should be promoted to int");
879 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
880 DAG.getConstant(2, MVT::i32));
882 // FIXME: REGISTER_LOAD should probably have a chain result.
883 SDValue Chain = Load->getChain();
884 SDValue LoLoad = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
886 DAG.getTargetConstant(0, MVT::i32),
889 SDValue Ret = LoLoad.getValue(0);
890 if (MemVT.getSizeInBits() == 64) {
891 // TODO: This needs a test to make sure the right thing is happening with
892 // the chain. That is hard without general function support.
894 SDValue IncPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
895 DAG.getConstant(1, MVT::i32));
897 SDValue HiLoad = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
899 DAG.getTargetConstant(0, MVT::i32),
902 Ret = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, LoLoad, HiLoad);
903 // Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
904 // LoLoad.getValue(1), HiLoad.getValue(1));
912 return DAG.getMergeValues(Ops, DL);
915 SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
917 SelectionDAG &DAG) const {
918 return DAG.getNode(Opcode, SDLoc(Op), Op.getValueType(), Op.getOperand(1),
924 SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
925 if (Op.getValueType() != MVT::i64)
929 SDValue Cond = Op.getOperand(0);
931 SDValue Zero = DAG.getConstant(0, MVT::i32);
932 SDValue One = DAG.getConstant(1, MVT::i32);
934 SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1));
935 SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2));
937 SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero);
938 SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero);
940 SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1);
942 SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One);
943 SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One);
945 SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1);
947 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i32, Lo, Hi);
948 return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
951 // Catch division cases where we can use shortcuts with rcp and rsq
953 SDValue SITargetLowering::LowerFastFDIV(SDValue Op, SelectionDAG &DAG) const {
955 SDValue LHS = Op.getOperand(0);
956 SDValue RHS = Op.getOperand(1);
957 EVT VT = Op.getValueType();
958 bool Unsafe = DAG.getTarget().Options.UnsafeFPMath;
960 if (const ConstantFPSDNode *CLHS = dyn_cast<ConstantFPSDNode>(LHS)) {
961 if ((Unsafe || (VT == MVT::f32 && !Subtarget->hasFP32Denormals())) &&
962 CLHS->isExactlyValue(1.0)) {
963 // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to
964 // the CI documentation has a worst case error of 1 ulp.
965 // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to
966 // use it as long as we aren't trying to use denormals.
968 // 1.0 / sqrt(x) -> rsq(x)
970 // XXX - Is UnsafeFPMath sufficient to do this for f64? The maximum ULP
971 // error seems really high at 2^29 ULP.
972 if (RHS.getOpcode() == ISD::FSQRT)
973 return DAG.getNode(AMDGPUISD::RSQ, SL, VT, RHS.getOperand(0));
976 return DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
981 // Turn into multiply by the reciprocal.
982 // x / y -> x * (1.0 / y)
983 SDValue Recip = DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
984 return DAG.getNode(ISD::FMUL, SL, VT, LHS, Recip);
990 SDValue SITargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const {
991 SDValue FastLowered = LowerFastFDIV(Op, DAG);
992 if (FastLowered.getNode())
995 // This uses v_rcp_f32 which does not handle denormals. Let this hit a
996 // selection error for now rather than do something incorrect.
997 if (Subtarget->hasFP32Denormals())
1001 SDValue LHS = Op.getOperand(0);
1002 SDValue RHS = Op.getOperand(1);
1004 SDValue r1 = DAG.getNode(ISD::FABS, SL, MVT::f32, RHS);
1006 const APFloat K0Val(BitsToFloat(0x6f800000));
1007 const SDValue K0 = DAG.getConstantFP(K0Val, MVT::f32);
1009 const APFloat K1Val(BitsToFloat(0x2f800000));
1010 const SDValue K1 = DAG.getConstantFP(K1Val, MVT::f32);
1012 const SDValue One = DAG.getTargetConstantFP(1.0, MVT::f32);
1014 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f32);
1016 SDValue r2 = DAG.getSetCC(SL, SetCCVT, r1, K0, ISD::SETOGT);
1018 SDValue r3 = DAG.getNode(ISD::SELECT, SL, MVT::f32, r2, K1, One);
1020 r1 = DAG.getNode(ISD::FMUL, SL, MVT::f32, RHS, r3);
1022 SDValue r0 = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f32, r1);
1024 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f32, LHS, r0);
1026 return DAG.getNode(ISD::FMUL, SL, MVT::f32, r3, Mul);
1029 SDValue SITargetLowering::LowerFDIV64(SDValue Op, SelectionDAG &DAG) const {
1033 SDValue SITargetLowering::LowerFDIV(SDValue Op, SelectionDAG &DAG) const {
1034 EVT VT = Op.getValueType();
1037 return LowerFDIV32(Op, DAG);
1040 return LowerFDIV64(Op, DAG);
1042 llvm_unreachable("Unexpected type for fdiv");
1045 SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1047 StoreSDNode *Store = cast<StoreSDNode>(Op);
1048 EVT VT = Store->getMemoryVT();
1050 // These stores are legal.
1051 if (Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS &&
1052 VT.isVector() && VT.getVectorNumElements() == 2 &&
1053 VT.getVectorElementType() == MVT::i32)
1056 SDValue Ret = AMDGPUTargetLowering::LowerSTORE(Op, DAG);
1060 if (VT.isVector() && VT.getVectorNumElements() >= 8)
1061 return SplitVectorStore(Op, DAG);
1064 return DAG.getTruncStore(Store->getChain(), DL,
1065 DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32),
1066 Store->getBasePtr(), MVT::i1, Store->getMemOperand());
1068 if (Store->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
1071 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Store->getBasePtr(),
1072 DAG.getConstant(2, MVT::i32));
1073 SDValue Chain = Store->getChain();
1074 SmallVector<SDValue, 8> Values;
1076 if (Store->isTruncatingStore()) {
1078 if (Store->getMemoryVT() == MVT::i8) {
1080 } else if (Store->getMemoryVT() == MVT::i16) {
1083 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1084 Chain, Store->getBasePtr(),
1085 DAG.getConstant(0, MVT::i32));
1086 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, Store->getBasePtr(),
1087 DAG.getConstant(0x3, MVT::i32));
1088 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1089 DAG.getConstant(3, MVT::i32));
1090 SDValue MaskedValue = DAG.getNode(ISD::AND, DL, MVT::i32, Store->getValue(),
1091 DAG.getConstant(Mask, MVT::i32));
1092 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1093 MaskedValue, ShiftAmt);
1094 SDValue RotrAmt = DAG.getNode(ISD::SUB, DL, MVT::i32,
1095 DAG.getConstant(32, MVT::i32), ShiftAmt);
1096 SDValue DstMask = DAG.getNode(ISD::ROTR, DL, MVT::i32,
1097 DAG.getConstant(Mask, MVT::i32),
1099 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1100 Dst = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1102 Values.push_back(Dst);
1103 } else if (VT == MVT::i64) {
1104 for (unsigned i = 0; i < 2; ++i) {
1105 Values.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
1106 Store->getValue(), DAG.getConstant(i, MVT::i32)));
1108 } else if (VT == MVT::i128) {
1109 for (unsigned i = 0; i < 2; ++i) {
1110 for (unsigned j = 0; j < 2; ++j) {
1111 Values.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
1112 DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
1113 Store->getValue(), DAG.getConstant(i, MVT::i32)),
1114 DAG.getConstant(j, MVT::i32)));
1118 Values.push_back(Store->getValue());
1121 for (unsigned i = 0; i < Values.size(); ++i) {
1122 SDValue PartPtr = DAG.getNode(ISD::ADD, DL, MVT::i32,
1123 Ptr, DAG.getConstant(i, MVT::i32));
1124 Chain = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1125 Chain, Values[i], PartPtr,
1126 DAG.getTargetConstant(0, MVT::i32));
1131 SDValue SITargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const {
1132 EVT VT = Op.getValueType();
1133 SDValue Arg = Op.getOperand(0);
1134 SDValue FractPart = DAG.getNode(AMDGPUISD::FRACT, SDLoc(Op), VT,
1135 DAG.getNode(ISD::FMUL, SDLoc(Op), VT, Arg,
1136 DAG.getConstantFP(0.5 / M_PI, VT)));
1138 switch (Op.getOpcode()) {
1140 return DAG.getNode(AMDGPUISD::COS_HW, SDLoc(Op), VT, FractPart);
1142 return DAG.getNode(AMDGPUISD::SIN_HW, SDLoc(Op), VT, FractPart);
1144 llvm_unreachable("Wrong trig opcode");
1148 //===----------------------------------------------------------------------===//
1149 // Custom DAG optimizations
1150 //===----------------------------------------------------------------------===//
1152 SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
1153 DAGCombinerInfo &DCI) {
1154 EVT VT = N->getValueType(0);
1155 EVT ScalarVT = VT.getScalarType();
1156 if (ScalarVT != MVT::f32)
1159 SelectionDAG &DAG = DCI.DAG;
1162 SDValue Src = N->getOperand(0);
1163 EVT SrcVT = Src.getValueType();
1165 // TODO: We could try to match extracting the higher bytes, which would be
1166 // easier if i8 vectors weren't promoted to i32 vectors, particularly after
1167 // types are legalized. v4i8 -> v4f32 is probably the only case to worry
1168 // about in practice.
1169 if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) {
1170 if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) {
1171 SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src);
1172 DCI.AddToWorklist(Cvt.getNode());
1177 // We are primarily trying to catch operations on illegal vector types
1178 // before they are expanded.
1179 // For scalars, we can use the more flexible method of checking masked bits
1180 // after legalization.
1181 if (!DCI.isBeforeLegalize() ||
1182 !SrcVT.isVector() ||
1183 SrcVT.getVectorElementType() != MVT::i8) {
1187 assert(DCI.isBeforeLegalize() && "Unexpected legal type");
1189 // Weird sized vectors are a pain to handle, but we know 3 is really the same
1191 unsigned NElts = SrcVT.getVectorNumElements();
1192 if (!SrcVT.isSimple() && NElts != 3)
1195 // Handle v4i8 -> v4f32 extload. Replace the v4i8 with a legal i32 load to
1196 // prevent a mess from expanding to v4i32 and repacking.
1197 if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse()) {
1198 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), SrcVT);
1199 EVT RegVT = getEquivalentLoadRegType(*DAG.getContext(), SrcVT);
1200 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f32, NElts);
1202 LoadSDNode *Load = cast<LoadSDNode>(Src);
1203 SDValue NewLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegVT,
1207 Load->getMemOperand());
1209 // Make sure successors of the original load stay after it by updating
1210 // them to use the new Chain.
1211 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), NewLoad.getValue(1));
1213 SmallVector<SDValue, 4> Elts;
1214 if (RegVT.isVector())
1215 DAG.ExtractVectorElements(NewLoad, Elts);
1217 Elts.push_back(NewLoad);
1219 SmallVector<SDValue, 4> Ops;
1221 unsigned EltIdx = 0;
1222 for (SDValue Elt : Elts) {
1223 unsigned ComponentsInElt = std::min(4u, NElts - 4 * EltIdx);
1224 for (unsigned I = 0; I < ComponentsInElt; ++I) {
1225 unsigned Opc = AMDGPUISD::CVT_F32_UBYTE0 + I;
1226 SDValue Cvt = DAG.getNode(Opc, DL, MVT::f32, Elt);
1227 DCI.AddToWorklist(Cvt.getNode());
1234 assert(Ops.size() == NElts);
1236 return DAG.getNode(ISD::BUILD_VECTOR, DL, FloatVT, Ops);
1242 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
1243 DAGCombinerInfo &DCI) const {
1244 SelectionDAG &DAG = DCI.DAG;
1246 EVT VT = N->getValueType(0);
1248 switch (N->getOpcode()) {
1249 default: return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1251 SDValue Arg0 = N->getOperand(0);
1252 SDValue Arg1 = N->getOperand(1);
1253 SDValue CC = N->getOperand(2);
1254 ConstantSDNode * C = nullptr;
1255 ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
1257 // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
1259 && Arg0.getOpcode() == ISD::SIGN_EXTEND
1260 && Arg0.getOperand(0).getValueType() == MVT::i1
1261 && (C = dyn_cast<ConstantSDNode>(Arg1))
1263 && CCOp == ISD::SETNE) {
1264 return SimplifySetCC(VT, Arg0.getOperand(0),
1265 DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
1270 case AMDGPUISD::CVT_F32_UBYTE0:
1271 case AMDGPUISD::CVT_F32_UBYTE1:
1272 case AMDGPUISD::CVT_F32_UBYTE2:
1273 case AMDGPUISD::CVT_F32_UBYTE3: {
1274 unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
1276 SDValue Src = N->getOperand(0);
1277 APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);
1279 APInt KnownZero, KnownOne;
1280 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1281 !DCI.isBeforeLegalizeOps());
1282 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1283 if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
1284 TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {
1285 DCI.CommitTargetLoweringOpt(TLO);
1291 case ISD::UINT_TO_FP: {
1292 return performUCharToFloatCombine(N, DCI);
1296 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1299 /// \brief Test if RegClass is one of the VSrc classes
1300 static bool isVSrc(unsigned RegClass) {
1301 return AMDGPU::VSrc_32RegClassID == RegClass ||
1302 AMDGPU::VSrc_64RegClassID == RegClass;
1305 /// \brief Test if RegClass is one of the SSrc classes
1306 static bool isSSrc(unsigned RegClass) {
1307 return AMDGPU::SSrc_32RegClassID == RegClass ||
1308 AMDGPU::SSrc_64RegClassID == RegClass;
1311 /// \brief Analyze the possible immediate value Op
1313 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
1314 /// and the immediate value if it's a literal immediate
1315 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
1322 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
1323 if (Node->getZExtValue() >> 32) {
1326 Imm.I = Node->getSExtValue();
1327 } else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
1328 if (N->getValueType(0) != MVT::f32)
1330 Imm.F = Node->getValueAPF().convertToFloat();
1332 return -1; // It isn't an immediate
1334 if ((Imm.I >= -16 && Imm.I <= 64) ||
1335 Imm.F == 0.5f || Imm.F == -0.5f ||
1336 Imm.F == 1.0f || Imm.F == -1.0f ||
1337 Imm.F == 2.0f || Imm.F == -2.0f ||
1338 Imm.F == 4.0f || Imm.F == -4.0f)
1339 return 0; // It's an inline immediate
1341 return Imm.I; // It's a literal immediate
1344 /// \brief Try to fold an immediate directly into an instruction
1345 bool SITargetLowering::foldImm(SDValue &Operand, int32_t &Immediate,
1346 bool &ScalarSlotUsed) const {
1348 MachineSDNode *Mov = dyn_cast<MachineSDNode>(Operand);
1349 const SIInstrInfo *TII =
1350 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1351 if (!Mov || !TII->isMov(Mov->getMachineOpcode()))
1354 const SDValue &Op = Mov->getOperand(0);
1355 int32_t Value = analyzeImmediate(Op.getNode());
1357 // Not an immediate at all
1360 } else if (Value == 0) {
1361 // Inline immediates can always be fold
1365 } else if (Value == Immediate) {
1366 // Already fold literal immediate
1370 } else if (!ScalarSlotUsed && !Immediate) {
1371 // Fold this literal immediate
1372 ScalarSlotUsed = true;
1382 const TargetRegisterClass *SITargetLowering::getRegClassForNode(
1383 SelectionDAG &DAG, const SDValue &Op) const {
1384 const SIInstrInfo *TII =
1385 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1386 const SIRegisterInfo &TRI = TII->getRegisterInfo();
1388 if (!Op->isMachineOpcode()) {
1389 switch(Op->getOpcode()) {
1390 case ISD::CopyFromReg: {
1391 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1392 unsigned Reg = cast<RegisterSDNode>(Op->getOperand(1))->getReg();
1393 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
1394 return MRI.getRegClass(Reg);
1396 return TRI.getPhysRegClass(Reg);
1398 default: return nullptr;
1401 const MCInstrDesc &Desc = TII->get(Op->getMachineOpcode());
1402 int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
1403 if (OpClassID != -1) {
1404 return TRI.getRegClass(OpClassID);
1406 switch(Op.getMachineOpcode()) {
1407 case AMDGPU::COPY_TO_REGCLASS:
1408 // Operand 1 is the register class id for COPY_TO_REGCLASS instructions.
1409 OpClassID = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1411 // If the COPY_TO_REGCLASS instruction is copying to a VSrc register
1412 // class, then the register class for the value could be either a
1413 // VReg or and SReg. In order to get a more accurate
1414 if (OpClassID == AMDGPU::VSrc_32RegClassID ||
1415 OpClassID == AMDGPU::VSrc_64RegClassID) {
1416 return getRegClassForNode(DAG, Op.getOperand(0));
1418 return TRI.getRegClass(OpClassID);
1419 case AMDGPU::EXTRACT_SUBREG: {
1420 int SubIdx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1421 const TargetRegisterClass *SuperClass =
1422 getRegClassForNode(DAG, Op.getOperand(0));
1423 return TRI.getSubClassWithSubReg(SuperClass, SubIdx);
1425 case AMDGPU::REG_SEQUENCE:
1426 // Operand 0 is the register class id for REG_SEQUENCE instructions.
1427 return TRI.getRegClass(
1428 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue());
1430 return getRegClassFor(Op.getSimpleValueType());
1434 /// \brief Does "Op" fit into register class "RegClass" ?
1435 bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, const SDValue &Op,
1436 unsigned RegClass) const {
1437 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1438 const TargetRegisterClass *RC = getRegClassForNode(DAG, Op);
1442 return TRI->getRegClass(RegClass)->hasSubClassEq(RC);
1445 /// \brief Make sure that we don't exeed the number of allowed scalars
1446 void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
1448 bool &ScalarSlotUsed) const {
1450 // First map the operands register class to a destination class
1451 if (RegClass == AMDGPU::VSrc_32RegClassID)
1452 RegClass = AMDGPU::VReg_32RegClassID;
1453 else if (RegClass == AMDGPU::VSrc_64RegClassID)
1454 RegClass = AMDGPU::VReg_64RegClassID;
1458 // Nothing to do if they fit naturally
1459 if (fitsRegClass(DAG, Operand, RegClass))
1462 // If the scalar slot isn't used yet use it now
1463 if (!ScalarSlotUsed) {
1464 ScalarSlotUsed = true;
1468 // This is a conservative aproach. It is possible that we can't determine the
1469 // correct register class and copy too often, but better safe than sorry.
1470 SDValue RC = DAG.getTargetConstant(RegClass, MVT::i32);
1471 SDNode *Node = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, SDLoc(),
1472 Operand.getValueType(), Operand, RC);
1473 Operand = SDValue(Node, 0);
1476 /// \returns true if \p Node's operands are different from the SDValue list
1478 static bool isNodeChanged(const SDNode *Node, const std::vector<SDValue> &Ops) {
1479 for (unsigned i = 0, e = Node->getNumOperands(); i < e; ++i) {
1480 if (Ops[i].getNode() != Node->getOperand(i).getNode()) {
1487 /// \brief Try to fold the Nodes operands into the Node
1488 SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
1489 SelectionDAG &DAG) const {
1491 // Original encoding (either e32 or e64)
1492 int Opcode = Node->getMachineOpcode();
1493 const SIInstrInfo *TII =
1494 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1495 const MCInstrDesc *Desc = &TII->get(Opcode);
1497 unsigned NumDefs = Desc->getNumDefs();
1498 unsigned NumOps = Desc->getNumOperands();
1500 // Commuted opcode if available
1501 int OpcodeRev = Desc->isCommutable() ? TII->commuteOpcode(Opcode) : -1;
1502 const MCInstrDesc *DescRev = OpcodeRev == -1 ? nullptr : &TII->get(OpcodeRev);
1504 assert(!DescRev || DescRev->getNumDefs() == NumDefs);
1505 assert(!DescRev || DescRev->getNumOperands() == NumOps);
1507 // e64 version if available, -1 otherwise
1508 int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
1509 const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? nullptr : &TII->get(OpcodeE64);
1510 int InputModifiers[3] = {0};
1512 assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
1514 int32_t Immediate = Desc->getSize() == 4 ? 0 : -1;
1515 bool HaveVSrc = false, HaveSSrc = false;
1517 // First figure out what we already have in this instruction.
1518 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1519 i != e && Op < NumOps; ++i, ++Op) {
1521 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1522 if (isVSrc(RegClass))
1524 else if (isSSrc(RegClass))
1529 int32_t Imm = analyzeImmediate(Node->getOperand(i).getNode());
1530 if (Imm != -1 && Imm != 0) {
1531 // Literal immediate
1536 // If we neither have VSrc nor SSrc, it makes no sense to continue.
1537 if (!HaveVSrc && !HaveSSrc)
1540 // No scalar allowed when we have both VSrc and SSrc
1541 bool ScalarSlotUsed = HaveVSrc && HaveSSrc;
1543 // Second go over the operands and try to fold them
1544 std::vector<SDValue> Ops;
1545 bool Promote2e64 = false;
1546 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1547 i != e && Op < NumOps; ++i, ++Op) {
1549 const SDValue &Operand = Node->getOperand(i);
1550 Ops.push_back(Operand);
1552 // Already folded immediate?
1553 if (isa<ConstantSDNode>(Operand.getNode()) ||
1554 isa<ConstantFPSDNode>(Operand.getNode()))
1557 // Is this a VSrc or SSrc operand?
1558 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1559 if (isVSrc(RegClass) || isSSrc(RegClass)) {
1560 // Try to fold the immediates
1561 if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
1562 // Folding didn't work, make sure we don't hit the SReg limit.
1563 ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
1568 if (i == 1 && DescRev && fitsRegClass(DAG, Ops[0], RegClass)) {
1570 unsigned OtherRegClass = Desc->OpInfo[NumDefs].RegClass;
1571 assert(isVSrc(OtherRegClass) || isSSrc(OtherRegClass));
1573 // Test if it makes sense to swap operands
1574 if (foldImm(Ops[1], Immediate, ScalarSlotUsed) ||
1575 (!fitsRegClass(DAG, Ops[1], RegClass) &&
1576 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1578 // Swap commutable operands
1579 std::swap(Ops[0], Ops[1]);
1591 // Test if it makes sense to switch to e64 encoding
1592 unsigned OtherRegClass = DescE64->OpInfo[Op].RegClass;
1593 if (!isVSrc(OtherRegClass) && !isSSrc(OtherRegClass))
1596 int32_t TmpImm = -1;
1597 if (foldImm(Ops[i], TmpImm, ScalarSlotUsed) ||
1598 (!fitsRegClass(DAG, Ops[i], RegClass) &&
1599 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1601 // Switch to e64 encoding
1609 if (!DescE64 && !Promote2e64)
1611 if (!Operand.isMachineOpcode())
1613 if (Operand.getMachineOpcode() == AMDGPU::FNEG_SI) {
1615 Ops.push_back(Operand.getOperand(0));
1616 InputModifiers[i] = 1;
1623 else if (Operand.getMachineOpcode() == AMDGPU::FABS_SI) {
1625 Ops.push_back(Operand.getOperand(0));
1626 InputModifiers[i] = 2;
1636 std::vector<SDValue> OldOps(Ops);
1638 for (unsigned i = 0; i < OldOps.size(); ++i) {
1640 Ops.push_back(DAG.getTargetConstant(InputModifiers[i], MVT::i32));
1641 Ops.push_back(OldOps[i]);
1643 // Add the modifier flags while promoting
1644 for (unsigned i = 0; i < 2; ++i)
1645 Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
1648 // Add optional chain and glue
1649 for (unsigned i = NumOps - NumDefs, e = Node->getNumOperands(); i < e; ++i)
1650 Ops.push_back(Node->getOperand(i));
1652 // Nodes that have a glue result are not CSE'd by getMachineNode(), so in
1653 // this case a brand new node is always be created, even if the operands
1654 // are the same as before. So, manually check if anything has been changed.
1655 if (Desc->Opcode == Opcode && !isNodeChanged(Node, Ops)) {
1659 // Create a complete new instruction
1660 return DAG.getMachineNode(Desc->Opcode, SDLoc(Node), Node->getVTList(), Ops);
1663 /// \brief Helper function for adjustWritemask
1664 static unsigned SubIdx2Lane(unsigned Idx) {
1667 case AMDGPU::sub0: return 0;
1668 case AMDGPU::sub1: return 1;
1669 case AMDGPU::sub2: return 2;
1670 case AMDGPU::sub3: return 3;
1674 /// \brief Adjust the writemask of MIMG instructions
1675 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
1676 SelectionDAG &DAG) const {
1677 SDNode *Users[4] = { };
1679 unsigned OldDmask = Node->getConstantOperandVal(0);
1680 unsigned NewDmask = 0;
1682 // Try to figure out the used register components
1683 for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
1686 // Abort if we can't understand the usage
1687 if (!I->isMachineOpcode() ||
1688 I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
1691 // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used.
1692 // Note that subregs are packed, i.e. Lane==0 is the first bit set
1693 // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit
1695 Lane = SubIdx2Lane(I->getConstantOperandVal(1));
1697 // Set which texture component corresponds to the lane.
1699 for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) {
1701 Comp = countTrailingZeros(Dmask);
1702 Dmask &= ~(1 << Comp);
1705 // Abort if we have more than one user per component
1710 NewDmask |= 1 << Comp;
1713 // Abort if there's no change
1714 if (NewDmask == OldDmask)
1717 // Adjust the writemask in the node
1718 std::vector<SDValue> Ops;
1719 Ops.push_back(DAG.getTargetConstant(NewDmask, MVT::i32));
1720 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
1721 Ops.push_back(Node->getOperand(i));
1722 Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops);
1724 // If we only got one lane, replace it with a copy
1725 // (if NewDmask has only one bit set...)
1726 if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
1727 SDValue RC = DAG.getTargetConstant(AMDGPU::VReg_32RegClassID, MVT::i32);
1728 SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1729 SDLoc(), Users[Lane]->getValueType(0),
1730 SDValue(Node, 0), RC);
1731 DAG.ReplaceAllUsesWith(Users[Lane], Copy);
1735 // Update the users of the node with the new indices
1736 for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
1738 SDNode *User = Users[i];
1742 SDValue Op = DAG.getTargetConstant(Idx, MVT::i32);
1743 DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
1747 case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
1748 case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
1749 case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
1754 /// \brief Fold the instructions after selecting them.
1755 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
1756 SelectionDAG &DAG) const {
1757 const SIInstrInfo *TII =
1758 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1759 Node = AdjustRegClass(Node, DAG);
1761 if (TII->isMIMG(Node->getMachineOpcode()))
1762 adjustWritemask(Node, DAG);
1764 return foldOperands(Node, DAG);
1767 /// \brief Assign the register class depending on the number of
1768 /// bits set in the writemask
1769 void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
1770 SDNode *Node) const {
1771 const SIInstrInfo *TII =
1772 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1773 if (!TII->isMIMG(MI->getOpcode()))
1776 unsigned VReg = MI->getOperand(0).getReg();
1777 unsigned Writemask = MI->getOperand(1).getImm();
1778 unsigned BitsSet = 0;
1779 for (unsigned i = 0; i < 4; ++i)
1780 BitsSet += Writemask & (1 << i) ? 1 : 0;
1782 const TargetRegisterClass *RC;
1785 case 1: RC = &AMDGPU::VReg_32RegClass; break;
1786 case 2: RC = &AMDGPU::VReg_64RegClass; break;
1787 case 3: RC = &AMDGPU::VReg_96RegClass; break;
1790 unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
1791 MI->setDesc(TII->get(NewOpcode));
1792 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1793 MRI.setRegClass(VReg, RC);
1796 MachineSDNode *SITargetLowering::AdjustRegClass(MachineSDNode *N,
1797 SelectionDAG &DAG) const {
1800 unsigned NewOpcode = N->getMachineOpcode();
1802 switch (N->getMachineOpcode()) {
1804 case AMDGPU::S_LOAD_DWORD_IMM:
1805 NewOpcode = AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1807 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1808 if (NewOpcode == N->getMachineOpcode()) {
1809 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1812 case AMDGPU::S_LOAD_DWORDX4_IMM:
1813 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1814 if (NewOpcode == N->getMachineOpcode()) {
1815 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1817 if (fitsRegClass(DAG, N->getOperand(0), AMDGPU::SReg_64RegClassID)) {
1820 ConstantSDNode *Offset = cast<ConstantSDNode>(N->getOperand(1));
1822 SDValue(DAG.getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::i128,
1823 DAG.getConstant(0, MVT::i64)), 0),
1825 DAG.getConstant(Offset->getSExtValue() << 2, MVT::i32)
1827 return DAG.getMachineNode(NewOpcode, DL, N->getVTList(), Ops);
1832 SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1833 const TargetRegisterClass *RC,
1834 unsigned Reg, EVT VT) const {
1835 SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT);
1837 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()),
1838 cast<RegisterSDNode>(VReg)->getReg(), VT);