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 //===----------------------------------------------------------------------===//
15 #include "SIISelLowering.h"
17 #include "AMDGPUSubtarget.h"
18 #include "AMDILIntrinsicInfo.h"
19 #include "SIInstrInfo.h"
20 #include "SIMachineFunctionInfo.h"
21 #include "SIRegisterInfo.h"
22 #include "llvm/CodeGen/CallingConvLower.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/SelectionDAG.h"
26 #include "llvm/IR/Function.h"
30 SITargetLowering::SITargetLowering(TargetMachine &TM) :
31 AMDGPUTargetLowering(TM) {
32 addRegisterClass(MVT::i1, &AMDGPU::VReg_1RegClass);
33 addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
35 addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
36 addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);
38 addRegisterClass(MVT::i32, &AMDGPU::SReg_32RegClass);
39 addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
41 addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
42 addRegisterClass(MVT::v2i32, &AMDGPU::SReg_64RegClass);
43 addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
45 addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
46 addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
48 addRegisterClass(MVT::v8i32, &AMDGPU::VReg_256RegClass);
49 addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
51 addRegisterClass(MVT::v16i32, &AMDGPU::VReg_512RegClass);
52 addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);
54 computeRegisterProperties();
57 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
58 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
59 setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
60 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
61 setCondCodeAction(ISD::SETULE, MVT::f32, Expand);
62 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
64 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
65 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
66 setCondCodeAction(ISD::SETUGE, MVT::f64, Expand);
67 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
68 setCondCodeAction(ISD::SETULE, MVT::f64, Expand);
69 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
71 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
72 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
73 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
74 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);
76 setOperationAction(ISD::ADD, MVT::i32, Legal);
77 setOperationAction(ISD::ADDC, MVT::i32, Legal);
78 setOperationAction(ISD::ADDE, MVT::i32, Legal);
80 // We need to custom lower vector stores from local memory
81 setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
82 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
83 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
84 setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
86 setOperationAction(ISD::STORE, MVT::v8i32, Custom);
87 setOperationAction(ISD::STORE, MVT::v16i32, Custom);
89 // We need to custom lower loads/stores from private memory
90 setOperationAction(ISD::LOAD, MVT::i32, Custom);
91 setOperationAction(ISD::LOAD, MVT::i64, Custom);
92 setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
93 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
94 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
96 setOperationAction(ISD::STORE, MVT::i1, Custom);
97 setOperationAction(ISD::STORE, MVT::i32, Custom);
98 setOperationAction(ISD::STORE, MVT::i64, Custom);
99 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
100 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
102 setOperationAction(ISD::SELECT, MVT::i64, Custom);
103 setOperationAction(ISD::SELECT, MVT::f64, Promote);
104 AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
106 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
107 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
109 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
111 setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
112 setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
114 setOperationAction(ISD::ANY_EXTEND, MVT::i64, Custom);
115 setOperationAction(ISD::SIGN_EXTEND, MVT::i64, Custom);
116 setOperationAction(ISD::ZERO_EXTEND, MVT::i64, Custom);
118 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
119 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
120 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
122 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
123 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
124 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
127 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
128 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
130 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Custom);
132 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
134 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
135 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
136 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom);
137 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
139 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
141 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
142 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Custom);
143 setLoadExtAction(ISD::SEXTLOAD, MVT::i16, Custom);
144 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Expand);
145 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, Expand);
146 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, Expand);
148 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
149 setLoadExtAction(ISD::ZEXTLOAD, MVT::i8, Custom);
150 setLoadExtAction(ISD::ZEXTLOAD, MVT::i16, Custom);
151 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Expand);
153 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
154 setLoadExtAction(ISD::EXTLOAD, MVT::i8, Custom);
155 setLoadExtAction(ISD::EXTLOAD, MVT::i16, Custom);
156 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Expand);
157 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
159 setTruncStoreAction(MVT::i32, MVT::i8, Custom);
160 setTruncStoreAction(MVT::i32, MVT::i16, Custom);
161 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
162 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
163 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Expand);
164 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Expand);
166 setOperationAction(ISD::LOAD, MVT::i1, Custom);
168 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
169 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
170 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
172 // These should use UDIVREM, so set them to expand
173 setOperationAction(ISD::UDIV, MVT::i64, Expand);
174 setOperationAction(ISD::UREM, MVT::i64, Expand);
176 // We only support LOAD/STORE and vector manipulation ops for vectors
177 // with > 4 elements.
179 MVT::v8i32, MVT::v8f32, MVT::v16i32, MVT::v16f32
182 const size_t NumVecTypes = array_lengthof(VecTypes);
183 for (unsigned Type = 0; Type < NumVecTypes; ++Type) {
184 for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
188 case ISD::BUILD_VECTOR:
190 case ISD::EXTRACT_VECTOR_ELT:
191 case ISD::INSERT_VECTOR_ELT:
192 case ISD::CONCAT_VECTORS:
193 case ISD::INSERT_SUBVECTOR:
194 case ISD::EXTRACT_SUBVECTOR:
197 setOperationAction(Op, VecTypes[Type], Expand);
203 for (int I = MVT::v1f64; I <= MVT::v8f64; ++I) {
204 MVT::SimpleValueType VT = static_cast<MVT::SimpleValueType>(I);
205 setOperationAction(ISD::FTRUNC, VT, Expand);
206 setOperationAction(ISD::FCEIL, VT, Expand);
207 setOperationAction(ISD::FFLOOR, VT, Expand);
210 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
211 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
212 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
213 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
214 setOperationAction(ISD::FRINT, MVT::f64, Legal);
217 setTargetDAGCombine(ISD::SELECT_CC);
218 setTargetDAGCombine(ISD::SETCC);
220 setSchedulingPreference(Sched::RegPressure);
223 //===----------------------------------------------------------------------===//
224 // TargetLowering queries
225 //===----------------------------------------------------------------------===//
227 bool SITargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
229 bool *IsFast) const {
233 // XXX: This depends on the address space and also we may want to revist
234 // the alignment values we specify in the DataLayout.
236 // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
237 // which isn't a simple VT.
238 if (!VT.isSimple() || VT == MVT::Other)
241 // XXX - CI changes say "Support for unaligned memory accesses" but I don't
242 // see what for specifically. The wording everywhere else seems to be the
245 // 3.6.4 - Operations using pairs of VGPRs (for example: double-floats) have
246 // no alignment restrictions.
247 if (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS) {
248 // Using any pair of GPRs should be the same as any other pair.
251 return VT.bitsGE(MVT::i64);
254 // XXX - The only mention I see of this in the ISA manual is for LDS direct
255 // reads the "byte address and must be dword aligned". Is it also true for the
256 // normal loads and stores?
257 if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS)
260 // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
261 // byte-address are ignored, thus forcing Dword alignment.
264 return VT.bitsGT(MVT::i32);
267 bool SITargetLowering::shouldSplitVectorType(EVT VT) const {
268 return VT.getScalarType().bitsLE(MVT::i16);
271 bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
273 const SIInstrInfo *TII =
274 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
275 return TII->isInlineConstant(Imm);
278 SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
279 SDLoc DL, SDValue Chain,
280 unsigned Offset, bool Signed) const {
281 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
282 PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
283 AMDGPUAS::CONSTANT_ADDRESS);
284 SDValue BasePtr = DAG.getCopyFromReg(Chain, DL,
285 MRI.getLiveInVirtReg(AMDGPU::SGPR0_SGPR1), MVT::i64);
286 SDValue Ptr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
287 DAG.getConstant(Offset, MVT::i64));
288 return DAG.getExtLoad(Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD, DL, VT, Chain, Ptr,
289 MachinePointerInfo(UndefValue::get(PtrTy)), MemVT,
290 false, false, MemVT.getSizeInBits() >> 3);
294 SDValue SITargetLowering::LowerFormalArguments(
296 CallingConv::ID CallConv,
298 const SmallVectorImpl<ISD::InputArg> &Ins,
299 SDLoc DL, SelectionDAG &DAG,
300 SmallVectorImpl<SDValue> &InVals) const {
302 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
304 MachineFunction &MF = DAG.getMachineFunction();
305 FunctionType *FType = MF.getFunction()->getFunctionType();
306 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
308 assert(CallConv == CallingConv::C);
310 SmallVector<ISD::InputArg, 16> Splits;
311 uint32_t Skipped = 0;
313 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
314 const ISD::InputArg &Arg = Ins[i];
316 // First check if it's a PS input addr
317 if (Info->ShaderType == ShaderType::PIXEL && !Arg.Flags.isInReg() &&
318 !Arg.Flags.isByVal()) {
320 assert((PSInputNum <= 15) && "Too many PS inputs!");
323 // We can savely skip PS inputs
329 Info->PSInputAddr |= 1 << PSInputNum++;
332 // Second split vertices into their elements
333 if (Info->ShaderType != ShaderType::COMPUTE && Arg.VT.isVector()) {
334 ISD::InputArg NewArg = Arg;
335 NewArg.Flags.setSplit();
336 NewArg.VT = Arg.VT.getVectorElementType();
338 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
339 // three or five element vertex only needs three or five registers,
340 // NOT four or eigth.
341 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
342 unsigned NumElements = ParamType->getVectorNumElements();
344 for (unsigned j = 0; j != NumElements; ++j) {
345 Splits.push_back(NewArg);
346 NewArg.PartOffset += NewArg.VT.getStoreSize();
349 } else if (Info->ShaderType != ShaderType::COMPUTE) {
350 Splits.push_back(Arg);
354 SmallVector<CCValAssign, 16> ArgLocs;
355 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
356 getTargetMachine(), ArgLocs, *DAG.getContext());
358 // At least one interpolation mode must be enabled or else the GPU will hang.
359 if (Info->ShaderType == ShaderType::PIXEL && (Info->PSInputAddr & 0x7F) == 0) {
360 Info->PSInputAddr |= 1;
361 CCInfo.AllocateReg(AMDGPU::VGPR0);
362 CCInfo.AllocateReg(AMDGPU::VGPR1);
365 // The pointer to the list of arguments is stored in SGPR0, SGPR1
366 if (Info->ShaderType == ShaderType::COMPUTE) {
367 CCInfo.AllocateReg(AMDGPU::SGPR0);
368 CCInfo.AllocateReg(AMDGPU::SGPR1);
369 MF.addLiveIn(AMDGPU::SGPR0_SGPR1, &AMDGPU::SReg_64RegClass);
372 if (Info->ShaderType == ShaderType::COMPUTE) {
373 getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
377 AnalyzeFormalArguments(CCInfo, Splits);
379 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
381 const ISD::InputArg &Arg = Ins[i];
382 if (Skipped & (1 << i)) {
383 InVals.push_back(DAG.getUNDEF(Arg.VT));
387 CCValAssign &VA = ArgLocs[ArgIdx++];
388 EVT VT = VA.getLocVT();
392 EVT MemVT = Splits[i].VT;
393 // The first 36 bytes of the input buffer contains information about
394 // thread group and global sizes.
395 SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, DAG.getRoot(),
396 36 + VA.getLocMemOffset(),
397 Ins[i].Flags.isSExt());
398 InVals.push_back(Arg);
401 assert(VA.isRegLoc() && "Parameter must be in a register!");
403 unsigned Reg = VA.getLocReg();
405 if (VT == MVT::i64) {
406 // For now assume it is a pointer
407 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
408 &AMDGPU::SReg_64RegClass);
409 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
410 InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
414 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
416 Reg = MF.addLiveIn(Reg, RC);
417 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
419 if (Arg.VT.isVector()) {
421 // Build a vector from the registers
422 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
423 unsigned NumElements = ParamType->getVectorNumElements();
425 SmallVector<SDValue, 4> Regs;
427 for (unsigned j = 1; j != NumElements; ++j) {
428 Reg = ArgLocs[ArgIdx++].getLocReg();
429 Reg = MF.addLiveIn(Reg, RC);
430 Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
433 // Fill up the missing vector elements
434 NumElements = Arg.VT.getVectorNumElements() - NumElements;
435 for (unsigned j = 0; j != NumElements; ++j)
436 Regs.push_back(DAG.getUNDEF(VT));
438 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT, Regs));
442 InVals.push_back(Val);
447 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
448 MachineInstr * MI, MachineBasicBlock * BB) const {
450 MachineBasicBlock::iterator I = *MI;
451 const SIInstrInfo *TII =
452 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
453 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
455 switch (MI->getOpcode()) {
457 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
458 case AMDGPU::BRANCH: return BB;
459 case AMDGPU::SI_ADDR64_RSRC: {
460 unsigned SuperReg = MI->getOperand(0).getReg();
461 unsigned SubRegLo = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
462 unsigned SubRegHi = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
463 unsigned SubRegHiHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
464 unsigned SubRegHiLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
465 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B64), SubRegLo)
466 .addOperand(MI->getOperand(1));
467 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiLo)
469 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiHi)
470 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
471 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SubRegHi)
473 .addImm(AMDGPU::sub0)
475 .addImm(AMDGPU::sub1);
476 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SuperReg)
478 .addImm(AMDGPU::sub0_sub1)
480 .addImm(AMDGPU::sub2_sub3);
481 MI->eraseFromParent();
484 case AMDGPU::V_SUB_F64:
485 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F64),
486 MI->getOperand(0).getReg())
487 .addReg(MI->getOperand(1).getReg())
488 .addReg(MI->getOperand(2).getReg())
489 .addImm(0) /* src2 */
491 .addImm(0) /* CLAMP */
492 .addImm(0) /* OMOD */
493 .addImm(2); /* NEG */
494 MI->eraseFromParent();
497 case AMDGPU::SI_RegisterStorePseudo: {
498 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
499 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
500 MachineInstrBuilder MIB =
501 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::SI_RegisterStore),
503 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i)
504 MIB.addOperand(MI->getOperand(i));
506 MI->eraseFromParent();
509 case AMDGPU::FABS_SI: {
510 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
511 const SIInstrInfo *TII =
512 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
513 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
514 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
517 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_AND_B32_e32),
518 MI->getOperand(0).getReg())
519 .addReg(MI->getOperand(1).getReg())
521 MI->eraseFromParent();
524 case AMDGPU::FNEG_SI: {
525 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
526 const SIInstrInfo *TII =
527 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
528 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
529 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
532 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_XOR_B32_e32),
533 MI->getOperand(0).getReg())
534 .addReg(MI->getOperand(1).getReg())
536 MI->eraseFromParent();
539 case AMDGPU::FCLAMP_SI: {
540 const SIInstrInfo *TII =
541 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
542 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F32_e64),
543 MI->getOperand(0).getReg())
544 .addImm(0) // SRC0 modifiers
545 .addOperand(MI->getOperand(1))
546 .addImm(0) // SRC1 modifiers
550 MI->eraseFromParent();
556 EVT SITargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
557 if (!VT.isVector()) {
560 return MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
563 MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
567 bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
568 VT = VT.getScalarType();
573 switch (VT.getSimpleVT().SimpleTy) {
575 return false; /* There is V_MAD_F32 for f32 */
585 //===----------------------------------------------------------------------===//
586 // Custom DAG Lowering Operations
587 //===----------------------------------------------------------------------===//
589 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
590 MachineFunction &MF = DAG.getMachineFunction();
591 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
592 switch (Op.getOpcode()) {
593 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
594 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
596 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
597 if (Op.getValueType().isVector() &&
598 (Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
599 Load->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS ||
600 (Load->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
601 Op.getValueType().getVectorNumElements() > 4))) {
602 SDValue MergedValues[2] = {
603 SplitVectorLoad(Op, DAG),
606 return DAG.getMergeValues(MergedValues, SDLoc(Op));
608 return LowerLOAD(Op, DAG);
612 case ISD::SELECT: return LowerSELECT(Op, DAG);
613 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
614 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, DAG);
615 case ISD::STORE: return LowerSTORE(Op, DAG);
616 case ISD::ANY_EXTEND: // Fall-through
617 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, DAG);
618 case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
619 case ISD::INTRINSIC_WO_CHAIN: {
620 unsigned IntrinsicID =
621 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
622 EVT VT = Op.getValueType();
624 //XXX: Hardcoded we only use two to store the pointer to the parameters.
625 unsigned NumUserSGPRs = 2;
626 switch (IntrinsicID) {
627 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
628 case Intrinsic::r600_read_ngroups_x:
629 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 0, false);
630 case Intrinsic::r600_read_ngroups_y:
631 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 4, false);
632 case Intrinsic::r600_read_ngroups_z:
633 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 8, false);
634 case Intrinsic::r600_read_global_size_x:
635 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 12, false);
636 case Intrinsic::r600_read_global_size_y:
637 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 16, false);
638 case Intrinsic::r600_read_global_size_z:
639 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 20, false);
640 case Intrinsic::r600_read_local_size_x:
641 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 24, false);
642 case Intrinsic::r600_read_local_size_y:
643 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 28, false);
644 case Intrinsic::r600_read_local_size_z:
645 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 32, false);
646 case Intrinsic::r600_read_tgid_x:
647 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
648 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 0), VT);
649 case Intrinsic::r600_read_tgid_y:
650 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
651 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 1), VT);
652 case Intrinsic::r600_read_tgid_z:
653 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
654 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 2), VT);
655 case Intrinsic::r600_read_tidig_x:
656 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
658 case Intrinsic::r600_read_tidig_y:
659 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
661 case Intrinsic::r600_read_tidig_z:
662 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
664 case AMDGPUIntrinsic::SI_load_const: {
670 MachineMemOperand *MMO = MF.getMachineMemOperand(
671 MachinePointerInfo(),
672 MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
673 VT.getSizeInBits() / 8, 4);
674 return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
675 Op->getVTList(), Ops, VT, MMO);
677 case AMDGPUIntrinsic::SI_sample:
678 return LowerSampleIntrinsic(AMDGPUISD::SAMPLE, Op, DAG);
679 case AMDGPUIntrinsic::SI_sampleb:
680 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEB, Op, DAG);
681 case AMDGPUIntrinsic::SI_sampled:
682 return LowerSampleIntrinsic(AMDGPUISD::SAMPLED, Op, DAG);
683 case AMDGPUIntrinsic::SI_samplel:
684 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEL, Op, DAG);
685 case AMDGPUIntrinsic::SI_vs_load_input:
686 return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
693 case ISD::INTRINSIC_VOID:
694 SDValue Chain = Op.getOperand(0);
695 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
697 switch (IntrinsicID) {
698 case AMDGPUIntrinsic::SI_tbuffer_store: {
716 EVT VT = Op.getOperand(3).getValueType();
718 MachineMemOperand *MMO = MF.getMachineMemOperand(
719 MachinePointerInfo(),
720 MachineMemOperand::MOStore,
721 VT.getSizeInBits() / 8, 4);
722 return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL,
723 Op->getVTList(), Ops, VT, MMO);
732 /// \brief Helper function for LowerBRCOND
733 static SDNode *findUser(SDValue Value, unsigned Opcode) {
735 SDNode *Parent = Value.getNode();
736 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
739 if (I.getUse().get() != Value)
742 if (I->getOpcode() == Opcode)
748 /// This transforms the control flow intrinsics to get the branch destination as
749 /// last parameter, also switches branch target with BR if the need arise
750 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
751 SelectionDAG &DAG) const {
755 SDNode *Intr = BRCOND.getOperand(1).getNode();
756 SDValue Target = BRCOND.getOperand(2);
757 SDNode *BR = nullptr;
759 if (Intr->getOpcode() == ISD::SETCC) {
760 // As long as we negate the condition everything is fine
761 SDNode *SetCC = Intr;
762 assert(SetCC->getConstantOperandVal(1) == 1);
763 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
765 Intr = SetCC->getOperand(0).getNode();
768 // Get the target from BR if we don't negate the condition
769 BR = findUser(BRCOND, ISD::BR);
770 Target = BR->getOperand(1);
773 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
775 // Build the result and
776 SmallVector<EVT, 4> Res;
777 for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
778 Res.push_back(Intr->getValueType(i));
780 // operands of the new intrinsic call
781 SmallVector<SDValue, 4> Ops;
782 Ops.push_back(BRCOND.getOperand(0));
783 for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
784 Ops.push_back(Intr->getOperand(i));
785 Ops.push_back(Target);
787 // build the new intrinsic call
788 SDNode *Result = DAG.getNode(
789 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
790 DAG.getVTList(Res), Ops).getNode();
793 // Give the branch instruction our target
798 DAG.MorphNodeTo(BR, ISD::BR, BR->getVTList(), Ops);
801 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
803 // Copy the intrinsic results to registers
804 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
805 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
809 Chain = DAG.getCopyToReg(
811 CopyToReg->getOperand(1),
812 SDValue(Result, i - 1),
815 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
818 // Remove the old intrinsic from the chain
819 DAG.ReplaceAllUsesOfValueWith(
820 SDValue(Intr, Intr->getNumValues() - 1),
821 Intr->getOperand(0));
826 SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
828 LoadSDNode *Load = cast<LoadSDNode>(Op);
829 SDValue Ret = AMDGPUTargetLowering::LowerLOAD(Op, DAG);
830 SDValue MergedValues[2];
831 MergedValues[1] = Load->getChain();
833 MergedValues[0] = Ret;
834 return DAG.getMergeValues(MergedValues, DL);
837 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
841 EVT MemVT = Load->getMemoryVT();
843 assert(!MemVT.isVector() && "Private loads should be scalarized");
844 assert(!MemVT.isFloatingPoint() && "FP loads should be promoted to int");
846 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
847 DAG.getConstant(2, MVT::i32));
848 Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
849 Load->getChain(), Ptr,
850 DAG.getTargetConstant(0, MVT::i32),
852 if (MemVT.getSizeInBits() == 64) {
853 SDValue IncPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
854 DAG.getConstant(1, MVT::i32));
856 SDValue LoadUpper = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
857 Load->getChain(), IncPtr,
858 DAG.getTargetConstant(0, MVT::i32),
861 Ret = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ret, LoadUpper);
864 MergedValues[0] = Ret;
865 return DAG.getMergeValues(MergedValues, DL);
869 SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
871 SelectionDAG &DAG) const {
872 return DAG.getNode(Opcode, SDLoc(Op), Op.getValueType(), Op.getOperand(1),
878 SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
879 if (Op.getValueType() != MVT::i64)
883 SDValue Cond = Op.getOperand(0);
885 SDValue Zero = DAG.getConstant(0, MVT::i32);
886 SDValue One = DAG.getConstant(1, MVT::i32);
888 SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1));
889 SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2));
891 SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero);
892 SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero);
894 SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1);
896 SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One);
897 SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One);
899 SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1);
901 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i32, Lo, Hi);
902 return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
905 SDValue SITargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
906 SDValue LHS = Op.getOperand(0);
907 SDValue RHS = Op.getOperand(1);
908 SDValue True = Op.getOperand(2);
909 SDValue False = Op.getOperand(3);
910 SDValue CC = Op.getOperand(4);
911 EVT VT = Op.getValueType();
914 SDValue Cond = DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS, CC);
915 return DAG.getNode(ISD::SELECT, DL, VT, Cond, True, False);
918 SDValue SITargetLowering::LowerSIGN_EXTEND(SDValue Op,
919 SelectionDAG &DAG) const {
920 EVT VT = Op.getValueType();
923 if (VT != MVT::i64) {
927 SDValue Hi = DAG.getNode(ISD::SRA, DL, MVT::i32, Op.getOperand(0),
928 DAG.getConstant(31, MVT::i32));
930 return DAG.getNode(ISD::BUILD_PAIR, DL, VT, Op.getOperand(0), Hi);
933 SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
935 StoreSDNode *Store = cast<StoreSDNode>(Op);
936 EVT VT = Store->getMemoryVT();
938 SDValue Ret = AMDGPUTargetLowering::LowerSTORE(Op, DAG);
942 if (VT.isVector() && VT.getVectorNumElements() >= 8)
943 return SplitVectorStore(Op, DAG);
946 return DAG.getTruncStore(Store->getChain(), DL,
947 DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32),
948 Store->getBasePtr(), MVT::i1, Store->getMemOperand());
950 if (Store->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
953 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Store->getBasePtr(),
954 DAG.getConstant(2, MVT::i32));
955 SDValue Chain = Store->getChain();
956 SmallVector<SDValue, 8> Values;
958 if (Store->isTruncatingStore()) {
960 if (Store->getMemoryVT() == MVT::i8) {
962 } else if (Store->getMemoryVT() == MVT::i16) {
965 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
966 Chain, Store->getBasePtr(),
967 DAG.getConstant(0, MVT::i32));
968 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, Store->getBasePtr(),
969 DAG.getConstant(0x3, MVT::i32));
970 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
971 DAG.getConstant(3, MVT::i32));
972 SDValue MaskedValue = DAG.getNode(ISD::AND, DL, MVT::i32, Store->getValue(),
973 DAG.getConstant(Mask, MVT::i32));
974 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
975 MaskedValue, ShiftAmt);
976 SDValue RotrAmt = DAG.getNode(ISD::SUB, DL, MVT::i32,
977 DAG.getConstant(32, MVT::i32), ShiftAmt);
978 SDValue DstMask = DAG.getNode(ISD::ROTR, DL, MVT::i32,
979 DAG.getConstant(Mask, MVT::i32),
981 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
982 Dst = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
984 Values.push_back(Dst);
985 } else if (VT == MVT::i64) {
986 for (unsigned i = 0; i < 2; ++i) {
987 Values.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
988 Store->getValue(), DAG.getConstant(i, MVT::i32)));
990 } else if (VT == MVT::i128) {
991 for (unsigned i = 0; i < 2; ++i) {
992 for (unsigned j = 0; j < 2; ++j) {
993 Values.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
994 DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
995 Store->getValue(), DAG.getConstant(i, MVT::i32)),
996 DAG.getConstant(j, MVT::i32)));
1000 Values.push_back(Store->getValue());
1003 for (unsigned i = 0; i < Values.size(); ++i) {
1004 SDValue PartPtr = DAG.getNode(ISD::ADD, DL, MVT::i32,
1005 Ptr, DAG.getConstant(i, MVT::i32));
1006 Chain = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1007 Chain, Values[i], PartPtr,
1008 DAG.getTargetConstant(0, MVT::i32));
1014 SDValue SITargetLowering::LowerZERO_EXTEND(SDValue Op,
1015 SelectionDAG &DAG) const {
1016 EVT VT = Op.getValueType();
1019 if (VT != MVT::i64) {
1023 SDValue Src = Op.getOperand(0);
1024 if (Src.getValueType() != MVT::i32)
1025 Src = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Src);
1027 SDValue Zero = DAG.getConstant(0, MVT::i32);
1028 return DAG.getNode(ISD::BUILD_PAIR, DL, VT, Src, Zero);
1031 //===----------------------------------------------------------------------===//
1032 // Custom DAG optimizations
1033 //===----------------------------------------------------------------------===//
1035 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
1036 DAGCombinerInfo &DCI) const {
1037 SelectionDAG &DAG = DCI.DAG;
1039 EVT VT = N->getValueType(0);
1041 switch (N->getOpcode()) {
1042 default: return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1043 case ISD::SELECT_CC: {
1044 ConstantSDNode *True, *False;
1045 // i1 selectcc(l, r, -1, 0, cc) -> i1 setcc(l, r, cc)
1046 if ((True = dyn_cast<ConstantSDNode>(N->getOperand(2)))
1047 && (False = dyn_cast<ConstantSDNode>(N->getOperand(3)))
1048 && True->isAllOnesValue()
1049 && False->isNullValue()
1051 return DAG.getNode(ISD::SETCC, DL, VT, N->getOperand(0),
1052 N->getOperand(1), N->getOperand(4));
1058 SDValue Arg0 = N->getOperand(0);
1059 SDValue Arg1 = N->getOperand(1);
1060 SDValue CC = N->getOperand(2);
1061 ConstantSDNode * C = nullptr;
1062 ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
1064 // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
1066 && Arg0.getOpcode() == ISD::SIGN_EXTEND
1067 && Arg0.getOperand(0).getValueType() == MVT::i1
1068 && (C = dyn_cast<ConstantSDNode>(Arg1))
1070 && CCOp == ISD::SETNE) {
1071 return SimplifySetCC(VT, Arg0.getOperand(0),
1072 DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
1080 /// \brief Test if RegClass is one of the VSrc classes
1081 static bool isVSrc(unsigned RegClass) {
1082 return AMDGPU::VSrc_32RegClassID == RegClass ||
1083 AMDGPU::VSrc_64RegClassID == RegClass;
1086 /// \brief Test if RegClass is one of the SSrc classes
1087 static bool isSSrc(unsigned RegClass) {
1088 return AMDGPU::SSrc_32RegClassID == RegClass ||
1089 AMDGPU::SSrc_64RegClassID == RegClass;
1092 /// \brief Analyze the possible immediate value Op
1094 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
1095 /// and the immediate value if it's a literal immediate
1096 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
1103 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
1104 if (Node->getZExtValue() >> 32) {
1107 Imm.I = Node->getSExtValue();
1108 } else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
1109 if (N->getValueType(0) != MVT::f32)
1111 Imm.F = Node->getValueAPF().convertToFloat();
1113 return -1; // It isn't an immediate
1115 if ((Imm.I >= -16 && Imm.I <= 64) ||
1116 Imm.F == 0.5f || Imm.F == -0.5f ||
1117 Imm.F == 1.0f || Imm.F == -1.0f ||
1118 Imm.F == 2.0f || Imm.F == -2.0f ||
1119 Imm.F == 4.0f || Imm.F == -4.0f)
1120 return 0; // It's an inline immediate
1122 return Imm.I; // It's a literal immediate
1125 /// \brief Try to fold an immediate directly into an instruction
1126 bool SITargetLowering::foldImm(SDValue &Operand, int32_t &Immediate,
1127 bool &ScalarSlotUsed) const {
1129 MachineSDNode *Mov = dyn_cast<MachineSDNode>(Operand);
1130 const SIInstrInfo *TII =
1131 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1132 if (!Mov || !TII->isMov(Mov->getMachineOpcode()))
1135 const SDValue &Op = Mov->getOperand(0);
1136 int32_t Value = analyzeImmediate(Op.getNode());
1138 // Not an immediate at all
1141 } else if (Value == 0) {
1142 // Inline immediates can always be fold
1146 } else if (Value == Immediate) {
1147 // Already fold literal immediate
1151 } else if (!ScalarSlotUsed && !Immediate) {
1152 // Fold this literal immediate
1153 ScalarSlotUsed = true;
1163 const TargetRegisterClass *SITargetLowering::getRegClassForNode(
1164 SelectionDAG &DAG, const SDValue &Op) const {
1165 const SIInstrInfo *TII =
1166 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1167 const SIRegisterInfo &TRI = TII->getRegisterInfo();
1169 if (!Op->isMachineOpcode()) {
1170 switch(Op->getOpcode()) {
1171 case ISD::CopyFromReg: {
1172 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1173 unsigned Reg = cast<RegisterSDNode>(Op->getOperand(1))->getReg();
1174 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
1175 return MRI.getRegClass(Reg);
1177 return TRI.getPhysRegClass(Reg);
1179 default: return nullptr;
1182 const MCInstrDesc &Desc = TII->get(Op->getMachineOpcode());
1183 int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
1184 if (OpClassID != -1) {
1185 return TRI.getRegClass(OpClassID);
1187 switch(Op.getMachineOpcode()) {
1188 case AMDGPU::COPY_TO_REGCLASS:
1189 // Operand 1 is the register class id for COPY_TO_REGCLASS instructions.
1190 OpClassID = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1192 // If the COPY_TO_REGCLASS instruction is copying to a VSrc register
1193 // class, then the register class for the value could be either a
1194 // VReg or and SReg. In order to get a more accurate
1195 if (OpClassID == AMDGPU::VSrc_32RegClassID ||
1196 OpClassID == AMDGPU::VSrc_64RegClassID) {
1197 return getRegClassForNode(DAG, Op.getOperand(0));
1199 return TRI.getRegClass(OpClassID);
1200 case AMDGPU::EXTRACT_SUBREG: {
1201 int SubIdx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1202 const TargetRegisterClass *SuperClass =
1203 getRegClassForNode(DAG, Op.getOperand(0));
1204 return TRI.getSubClassWithSubReg(SuperClass, SubIdx);
1206 case AMDGPU::REG_SEQUENCE:
1207 // Operand 0 is the register class id for REG_SEQUENCE instructions.
1208 return TRI.getRegClass(
1209 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue());
1211 return getRegClassFor(Op.getSimpleValueType());
1215 /// \brief Does "Op" fit into register class "RegClass" ?
1216 bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, const SDValue &Op,
1217 unsigned RegClass) const {
1218 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1219 const TargetRegisterClass *RC = getRegClassForNode(DAG, Op);
1223 return TRI->getRegClass(RegClass)->hasSubClassEq(RC);
1226 /// \brief Make sure that we don't exeed the number of allowed scalars
1227 void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
1229 bool &ScalarSlotUsed) const {
1231 // First map the operands register class to a destination class
1232 if (RegClass == AMDGPU::VSrc_32RegClassID)
1233 RegClass = AMDGPU::VReg_32RegClassID;
1234 else if (RegClass == AMDGPU::VSrc_64RegClassID)
1235 RegClass = AMDGPU::VReg_64RegClassID;
1239 // Nothing to do if they fit naturally
1240 if (fitsRegClass(DAG, Operand, RegClass))
1243 // If the scalar slot isn't used yet use it now
1244 if (!ScalarSlotUsed) {
1245 ScalarSlotUsed = true;
1249 // This is a conservative aproach. It is possible that we can't determine the
1250 // correct register class and copy too often, but better safe than sorry.
1251 SDValue RC = DAG.getTargetConstant(RegClass, MVT::i32);
1252 SDNode *Node = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, SDLoc(),
1253 Operand.getValueType(), Operand, RC);
1254 Operand = SDValue(Node, 0);
1257 /// \returns true if \p Node's operands are different from the SDValue list
1259 static bool isNodeChanged(const SDNode *Node, const std::vector<SDValue> &Ops) {
1260 for (unsigned i = 0, e = Node->getNumOperands(); i < e; ++i) {
1261 if (Ops[i].getNode() != Node->getOperand(i).getNode()) {
1268 /// \brief Try to fold the Nodes operands into the Node
1269 SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
1270 SelectionDAG &DAG) const {
1272 // Original encoding (either e32 or e64)
1273 int Opcode = Node->getMachineOpcode();
1274 const SIInstrInfo *TII =
1275 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1276 const MCInstrDesc *Desc = &TII->get(Opcode);
1278 unsigned NumDefs = Desc->getNumDefs();
1279 unsigned NumOps = Desc->getNumOperands();
1281 // Commuted opcode if available
1282 int OpcodeRev = Desc->isCommutable() ? TII->commuteOpcode(Opcode) : -1;
1283 const MCInstrDesc *DescRev = OpcodeRev == -1 ? nullptr : &TII->get(OpcodeRev);
1285 assert(!DescRev || DescRev->getNumDefs() == NumDefs);
1286 assert(!DescRev || DescRev->getNumOperands() == NumOps);
1288 // e64 version if available, -1 otherwise
1289 int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
1290 const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? nullptr : &TII->get(OpcodeE64);
1291 int InputModifiers[3] = {0};
1293 assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
1295 int32_t Immediate = Desc->getSize() == 4 ? 0 : -1;
1296 bool HaveVSrc = false, HaveSSrc = false;
1298 // First figure out what we alread have in this instruction
1299 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1300 i != e && Op < NumOps; ++i, ++Op) {
1302 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1303 if (isVSrc(RegClass))
1305 else if (isSSrc(RegClass))
1310 int32_t Imm = analyzeImmediate(Node->getOperand(i).getNode());
1311 if (Imm != -1 && Imm != 0) {
1312 // Literal immediate
1317 // If we neither have VSrc nor SSrc it makes no sense to continue
1318 if (!HaveVSrc && !HaveSSrc)
1321 // No scalar allowed when we have both VSrc and SSrc
1322 bool ScalarSlotUsed = HaveVSrc && HaveSSrc;
1324 // Second go over the operands and try to fold them
1325 std::vector<SDValue> Ops;
1326 bool Promote2e64 = false;
1327 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1328 i != e && Op < NumOps; ++i, ++Op) {
1330 const SDValue &Operand = Node->getOperand(i);
1331 Ops.push_back(Operand);
1333 // Already folded immediate ?
1334 if (isa<ConstantSDNode>(Operand.getNode()) ||
1335 isa<ConstantFPSDNode>(Operand.getNode()))
1338 // Is this a VSrc or SSrc operand ?
1339 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1340 if (isVSrc(RegClass) || isSSrc(RegClass)) {
1341 // Try to fold the immediates
1342 if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
1343 // Folding didn't worked, make sure we don't hit the SReg limit
1344 ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
1349 if (i == 1 && DescRev && fitsRegClass(DAG, Ops[0], RegClass)) {
1351 unsigned OtherRegClass = Desc->OpInfo[NumDefs].RegClass;
1352 assert(isVSrc(OtherRegClass) || isSSrc(OtherRegClass));
1354 // Test if it makes sense to swap operands
1355 if (foldImm(Ops[1], Immediate, ScalarSlotUsed) ||
1356 (!fitsRegClass(DAG, Ops[1], RegClass) &&
1357 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1359 // Swap commutable operands
1360 std::swap(Ops[0], Ops[1]);
1373 // Test if it makes sense to switch to e64 encoding
1374 unsigned OtherRegClass = DescE64->OpInfo[Op].RegClass;
1375 if (!isVSrc(OtherRegClass) && !isSSrc(OtherRegClass))
1378 int32_t TmpImm = -1;
1379 if (foldImm(Ops[i], TmpImm, ScalarSlotUsed) ||
1380 (!fitsRegClass(DAG, Ops[i], RegClass) &&
1381 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1383 // Switch to e64 encoding
1391 if (!DescE64 && !Promote2e64)
1393 if (!Operand.isMachineOpcode())
1395 if (Operand.getMachineOpcode() == AMDGPU::FNEG_SI) {
1397 Ops.push_back(Operand.getOperand(0));
1398 InputModifiers[i] = 1;
1405 else if (Operand.getMachineOpcode() == AMDGPU::FABS_SI) {
1407 Ops.push_back(Operand.getOperand(0));
1408 InputModifiers[i] = 2;
1418 std::vector<SDValue> OldOps(Ops);
1420 for (unsigned i = 0; i < OldOps.size(); ++i) {
1422 Ops.push_back(DAG.getTargetConstant(InputModifiers[i], MVT::i32));
1423 Ops.push_back(OldOps[i]);
1425 // Add the modifier flags while promoting
1426 for (unsigned i = 0; i < 2; ++i)
1427 Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
1430 // Add optional chain and glue
1431 for (unsigned i = NumOps - NumDefs, e = Node->getNumOperands(); i < e; ++i)
1432 Ops.push_back(Node->getOperand(i));
1434 // Nodes that have a glue result are not CSE'd by getMachineNode(), so in
1435 // this case a brand new node is always be created, even if the operands
1436 // are the same as before. So, manually check if anything has been changed.
1437 if (Desc->Opcode == Opcode && !isNodeChanged(Node, Ops)) {
1441 // Create a complete new instruction
1442 return DAG.getMachineNode(Desc->Opcode, SDLoc(Node), Node->getVTList(), Ops);
1445 /// \brief Helper function for adjustWritemask
1446 static unsigned SubIdx2Lane(unsigned Idx) {
1449 case AMDGPU::sub0: return 0;
1450 case AMDGPU::sub1: return 1;
1451 case AMDGPU::sub2: return 2;
1452 case AMDGPU::sub3: return 3;
1456 /// \brief Adjust the writemask of MIMG instructions
1457 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
1458 SelectionDAG &DAG) const {
1459 SDNode *Users[4] = { };
1461 unsigned OldDmask = Node->getConstantOperandVal(0);
1462 unsigned NewDmask = 0;
1464 // Try to figure out the used register components
1465 for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
1468 // Abort if we can't understand the usage
1469 if (!I->isMachineOpcode() ||
1470 I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
1473 // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used.
1474 // Note that subregs are packed, i.e. Lane==0 is the first bit set
1475 // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit
1477 Lane = SubIdx2Lane(I->getConstantOperandVal(1));
1479 // Set which texture component corresponds to the lane.
1481 for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) {
1483 Comp = countTrailingZeros(Dmask);
1484 Dmask &= ~(1 << Comp);
1487 // Abort if we have more than one user per component
1492 NewDmask |= 1 << Comp;
1495 // Abort if there's no change
1496 if (NewDmask == OldDmask)
1499 // Adjust the writemask in the node
1500 std::vector<SDValue> Ops;
1501 Ops.push_back(DAG.getTargetConstant(NewDmask, MVT::i32));
1502 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
1503 Ops.push_back(Node->getOperand(i));
1504 Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops);
1506 // If we only got one lane, replace it with a copy
1507 // (if NewDmask has only one bit set...)
1508 if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
1509 SDValue RC = DAG.getTargetConstant(AMDGPU::VReg_32RegClassID, MVT::i32);
1510 SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1511 SDLoc(), Users[Lane]->getValueType(0),
1512 SDValue(Node, 0), RC);
1513 DAG.ReplaceAllUsesWith(Users[Lane], Copy);
1517 // Update the users of the node with the new indices
1518 for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
1520 SDNode *User = Users[i];
1524 SDValue Op = DAG.getTargetConstant(Idx, MVT::i32);
1525 DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
1529 case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
1530 case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
1531 case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
1536 /// \brief Fold the instructions after slecting them
1537 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
1538 SelectionDAG &DAG) const {
1539 const SIInstrInfo *TII =
1540 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1541 Node = AdjustRegClass(Node, DAG);
1543 if (TII->isMIMG(Node->getMachineOpcode()))
1544 adjustWritemask(Node, DAG);
1546 return foldOperands(Node, DAG);
1549 /// \brief Assign the register class depending on the number of
1550 /// bits set in the writemask
1551 void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
1552 SDNode *Node) const {
1553 const SIInstrInfo *TII =
1554 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1555 if (!TII->isMIMG(MI->getOpcode()))
1558 unsigned VReg = MI->getOperand(0).getReg();
1559 unsigned Writemask = MI->getOperand(1).getImm();
1560 unsigned BitsSet = 0;
1561 for (unsigned i = 0; i < 4; ++i)
1562 BitsSet += Writemask & (1 << i) ? 1 : 0;
1564 const TargetRegisterClass *RC;
1567 case 1: RC = &AMDGPU::VReg_32RegClass; break;
1568 case 2: RC = &AMDGPU::VReg_64RegClass; break;
1569 case 3: RC = &AMDGPU::VReg_96RegClass; break;
1572 unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
1573 MI->setDesc(TII->get(NewOpcode));
1574 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1575 MRI.setRegClass(VReg, RC);
1578 MachineSDNode *SITargetLowering::AdjustRegClass(MachineSDNode *N,
1579 SelectionDAG &DAG) const {
1582 unsigned NewOpcode = N->getMachineOpcode();
1584 switch (N->getMachineOpcode()) {
1586 case AMDGPU::S_LOAD_DWORD_IMM:
1587 NewOpcode = AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1589 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1590 if (NewOpcode == N->getMachineOpcode()) {
1591 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1594 case AMDGPU::S_LOAD_DWORDX4_IMM:
1595 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1596 if (NewOpcode == N->getMachineOpcode()) {
1597 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1599 if (fitsRegClass(DAG, N->getOperand(0), AMDGPU::SReg_64RegClassID)) {
1602 ConstantSDNode *Offset = cast<ConstantSDNode>(N->getOperand(1));
1604 SDValue(DAG.getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::i128,
1605 DAG.getConstant(0, MVT::i64)), 0),
1607 DAG.getConstant(Offset->getSExtValue() << 2, MVT::i32)
1609 return DAG.getMachineNode(NewOpcode, DL, N->getVTList(), Ops);
1614 SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1615 const TargetRegisterClass *RC,
1616 unsigned Reg, EVT VT) const {
1617 SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT);
1619 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()),
1620 cast<RegisterSDNode>(VReg)->getReg(), VT);