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 "AMDGPUIntrinsicInfo.h"
18 #include "AMDGPUSubtarget.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"
27 #include "llvm/ADT/SmallString.h"
31 SITargetLowering::SITargetLowering(TargetMachine &TM) :
32 AMDGPUTargetLowering(TM) {
33 addRegisterClass(MVT::i1, &AMDGPU::VReg_1RegClass);
34 addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
36 addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
37 addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);
39 addRegisterClass(MVT::i32, &AMDGPU::SReg_32RegClass);
40 addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
42 addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
43 addRegisterClass(MVT::v2i32, &AMDGPU::SReg_64RegClass);
44 addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
46 addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
47 addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
49 addRegisterClass(MVT::v8i32, &AMDGPU::VReg_256RegClass);
50 addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
52 addRegisterClass(MVT::v16i32, &AMDGPU::VReg_512RegClass);
53 addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);
55 computeRegisterProperties();
58 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
59 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
60 setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
61 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
62 setCondCodeAction(ISD::SETULE, MVT::f32, Expand);
63 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
65 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
66 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
67 setCondCodeAction(ISD::SETUGE, MVT::f64, Expand);
68 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
69 setCondCodeAction(ISD::SETULE, MVT::f64, Expand);
70 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
72 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
73 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
74 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
75 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);
77 setOperationAction(ISD::ADD, MVT::i32, Legal);
78 setOperationAction(ISD::ADDC, MVT::i32, Legal);
79 setOperationAction(ISD::ADDE, MVT::i32, Legal);
80 setOperationAction(ISD::SUBC, MVT::i32, Legal);
81 setOperationAction(ISD::SUBE, MVT::i32, Legal);
83 // We need to custom lower vector stores from local memory
84 setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
85 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
86 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
87 setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
89 setOperationAction(ISD::STORE, MVT::v8i32, Custom);
90 setOperationAction(ISD::STORE, MVT::v16i32, Custom);
92 // We need to custom lower loads/stores from private memory
93 setOperationAction(ISD::LOAD, MVT::i32, Custom);
94 setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
95 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
96 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
98 setOperationAction(ISD::STORE, MVT::i1, Custom);
99 setOperationAction(ISD::STORE, MVT::i32, Custom);
100 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
101 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
103 setOperationAction(ISD::SELECT, MVT::f32, Promote);
104 AddPromotedToType(ISD::SELECT, MVT::f32, MVT::i32);
105 setOperationAction(ISD::SELECT, MVT::i64, Custom);
106 setOperationAction(ISD::SELECT, MVT::f64, Promote);
107 AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
109 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
110 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
111 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
112 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
114 setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
115 setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
117 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
118 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
119 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
121 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
122 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
123 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
125 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
127 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
129 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Custom);
131 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
133 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
134 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
135 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom);
136 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
138 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
139 setOperationAction(ISD::BRCOND, 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 for (MVT VT : VecTypes) {
183 for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
187 case ISD::BUILD_VECTOR:
189 case ISD::EXTRACT_VECTOR_ELT:
190 case ISD::INSERT_VECTOR_ELT:
191 case ISD::CONCAT_VECTORS:
192 case ISD::INSERT_SUBVECTOR:
193 case ISD::EXTRACT_SUBVECTOR:
196 setOperationAction(Op, VT, Expand);
202 for (int I = MVT::v1f64; I <= MVT::v8f64; ++I) {
203 MVT::SimpleValueType VT = static_cast<MVT::SimpleValueType>(I);
204 setOperationAction(ISD::FTRUNC, VT, Expand);
205 setOperationAction(ISD::FCEIL, VT, Expand);
206 setOperationAction(ISD::FFLOOR, VT, Expand);
209 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
210 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
211 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
212 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
213 setOperationAction(ISD::FRINT, MVT::f64, Legal);
216 // FIXME: These should be removed and handled the same was as f32 fneg. Source
217 // modifiers also work for the double instructions.
218 setOperationAction(ISD::FNEG, MVT::f64, Expand);
219 setOperationAction(ISD::FABS, MVT::f64, Expand);
221 setTargetDAGCombine(ISD::SELECT_CC);
222 setTargetDAGCombine(ISD::SETCC);
224 setTargetDAGCombine(ISD::UINT_TO_FP);
226 setSchedulingPreference(Sched::RegPressure);
229 //===----------------------------------------------------------------------===//
230 // TargetLowering queries
231 //===----------------------------------------------------------------------===//
233 bool SITargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
235 bool *IsFast) const {
239 // XXX: This depends on the address space and also we may want to revist
240 // the alignment values we specify in the DataLayout.
242 // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
243 // which isn't a simple VT.
244 if (!VT.isSimple() || VT == MVT::Other)
247 // XXX - CI changes say "Support for unaligned memory accesses" but I don't
248 // see what for specifically. The wording everywhere else seems to be the
251 // 3.6.4 - Operations using pairs of VGPRs (for example: double-floats) have
252 // no alignment restrictions.
253 if (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS) {
254 // Using any pair of GPRs should be the same as any other pair.
257 return VT.bitsGE(MVT::i64);
260 // XXX - The only mention I see of this in the ISA manual is for LDS direct
261 // reads the "byte address and must be dword aligned". Is it also true for the
262 // normal loads and stores?
263 if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS)
266 // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
267 // byte-address are ignored, thus forcing Dword alignment.
270 return VT.bitsGT(MVT::i32);
273 TargetLoweringBase::LegalizeTypeAction
274 SITargetLowering::getPreferredVectorAction(EVT VT) const {
275 if (VT.getVectorNumElements() != 1 && VT.getScalarType().bitsLE(MVT::i16))
276 return TypeSplitVector;
278 return TargetLoweringBase::getPreferredVectorAction(VT);
281 bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
283 const SIInstrInfo *TII =
284 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
285 return TII->isInlineConstant(Imm);
288 SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
289 SDLoc DL, SDValue Chain,
290 unsigned Offset, bool Signed) const {
291 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
292 PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
293 AMDGPUAS::CONSTANT_ADDRESS);
294 SDValue BasePtr = DAG.getCopyFromReg(Chain, DL,
295 MRI.getLiveInVirtReg(AMDGPU::SGPR0_SGPR1), MVT::i64);
296 SDValue Ptr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
297 DAG.getConstant(Offset, MVT::i64));
298 return DAG.getExtLoad(Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD, DL, VT, Chain, Ptr,
299 MachinePointerInfo(UndefValue::get(PtrTy)), MemVT,
300 false, false, MemVT.getSizeInBits() >> 3);
304 SDValue SITargetLowering::LowerFormalArguments(
306 CallingConv::ID CallConv,
308 const SmallVectorImpl<ISD::InputArg> &Ins,
309 SDLoc DL, SelectionDAG &DAG,
310 SmallVectorImpl<SDValue> &InVals) const {
312 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
314 MachineFunction &MF = DAG.getMachineFunction();
315 FunctionType *FType = MF.getFunction()->getFunctionType();
316 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
318 assert(CallConv == CallingConv::C);
320 SmallVector<ISD::InputArg, 16> Splits;
321 uint32_t Skipped = 0;
323 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
324 const ISD::InputArg &Arg = Ins[i];
326 // First check if it's a PS input addr
327 if (Info->ShaderType == ShaderType::PIXEL && !Arg.Flags.isInReg() &&
328 !Arg.Flags.isByVal()) {
330 assert((PSInputNum <= 15) && "Too many PS inputs!");
333 // We can savely skip PS inputs
339 Info->PSInputAddr |= 1 << PSInputNum++;
342 // Second split vertices into their elements
343 if (Info->ShaderType != ShaderType::COMPUTE && Arg.VT.isVector()) {
344 ISD::InputArg NewArg = Arg;
345 NewArg.Flags.setSplit();
346 NewArg.VT = Arg.VT.getVectorElementType();
348 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
349 // three or five element vertex only needs three or five registers,
350 // NOT four or eigth.
351 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
352 unsigned NumElements = ParamType->getVectorNumElements();
354 for (unsigned j = 0; j != NumElements; ++j) {
355 Splits.push_back(NewArg);
356 NewArg.PartOffset += NewArg.VT.getStoreSize();
359 } else if (Info->ShaderType != ShaderType::COMPUTE) {
360 Splits.push_back(Arg);
364 SmallVector<CCValAssign, 16> ArgLocs;
365 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
366 getTargetMachine(), ArgLocs, *DAG.getContext());
368 // At least one interpolation mode must be enabled or else the GPU will hang.
369 if (Info->ShaderType == ShaderType::PIXEL && (Info->PSInputAddr & 0x7F) == 0) {
370 Info->PSInputAddr |= 1;
371 CCInfo.AllocateReg(AMDGPU::VGPR0);
372 CCInfo.AllocateReg(AMDGPU::VGPR1);
375 // The pointer to the list of arguments is stored in SGPR0, SGPR1
376 if (Info->ShaderType == ShaderType::COMPUTE) {
377 CCInfo.AllocateReg(AMDGPU::SGPR0);
378 CCInfo.AllocateReg(AMDGPU::SGPR1);
379 MF.addLiveIn(AMDGPU::SGPR0_SGPR1, &AMDGPU::SReg_64RegClass);
382 if (Info->ShaderType == ShaderType::COMPUTE) {
383 getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
387 AnalyzeFormalArguments(CCInfo, Splits);
389 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
391 const ISD::InputArg &Arg = Ins[i];
392 if (Skipped & (1 << i)) {
393 InVals.push_back(DAG.getUNDEF(Arg.VT));
397 CCValAssign &VA = ArgLocs[ArgIdx++];
398 EVT VT = VA.getLocVT();
402 EVT MemVT = Splits[i].VT;
403 // The first 36 bytes of the input buffer contains information about
404 // thread group and global sizes.
405 SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, DAG.getRoot(),
406 36 + VA.getLocMemOffset(),
407 Ins[i].Flags.isSExt());
408 InVals.push_back(Arg);
411 assert(VA.isRegLoc() && "Parameter must be in a register!");
413 unsigned Reg = VA.getLocReg();
415 if (VT == MVT::i64) {
416 // For now assume it is a pointer
417 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
418 &AMDGPU::SReg_64RegClass);
419 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
420 InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
424 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
426 Reg = MF.addLiveIn(Reg, RC);
427 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
429 if (Arg.VT.isVector()) {
431 // Build a vector from the registers
432 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
433 unsigned NumElements = ParamType->getVectorNumElements();
435 SmallVector<SDValue, 4> Regs;
437 for (unsigned j = 1; j != NumElements; ++j) {
438 Reg = ArgLocs[ArgIdx++].getLocReg();
439 Reg = MF.addLiveIn(Reg, RC);
440 Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
443 // Fill up the missing vector elements
444 NumElements = Arg.VT.getVectorNumElements() - NumElements;
445 for (unsigned j = 0; j != NumElements; ++j)
446 Regs.push_back(DAG.getUNDEF(VT));
448 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT, Regs));
452 InVals.push_back(Val);
457 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
458 MachineInstr * MI, MachineBasicBlock * BB) const {
460 MachineBasicBlock::iterator I = *MI;
461 const SIInstrInfo *TII =
462 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
463 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
465 switch (MI->getOpcode()) {
467 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
468 case AMDGPU::BRANCH: return BB;
469 case AMDGPU::SI_ADDR64_RSRC: {
470 unsigned SuperReg = MI->getOperand(0).getReg();
471 unsigned SubRegLo = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
472 unsigned SubRegHi = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
473 unsigned SubRegHiHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
474 unsigned SubRegHiLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
475 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B64), SubRegLo)
476 .addOperand(MI->getOperand(1));
477 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiLo)
479 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiHi)
480 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
481 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SubRegHi)
483 .addImm(AMDGPU::sub0)
485 .addImm(AMDGPU::sub1);
486 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SuperReg)
488 .addImm(AMDGPU::sub0_sub1)
490 .addImm(AMDGPU::sub2_sub3);
491 MI->eraseFromParent();
494 case AMDGPU::V_SUB_F64: {
495 unsigned DestReg = MI->getOperand(0).getReg();
496 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F64), DestReg)
497 .addImm(0) // SRC0 modifiers
498 .addReg(MI->getOperand(1).getReg())
499 .addImm(1) // SRC1 modifiers
500 .addReg(MI->getOperand(2).getReg())
501 .addImm(0) // SRC2 modifiers
505 MI->eraseFromParent();
508 case AMDGPU::SI_RegisterStorePseudo: {
509 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
510 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
511 MachineInstrBuilder MIB =
512 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::SI_RegisterStore),
514 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i)
515 MIB.addOperand(MI->getOperand(i));
517 MI->eraseFromParent();
520 case AMDGPU::FABS_SI: {
521 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
522 const SIInstrInfo *TII =
523 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
524 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
525 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
528 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_AND_B32_e32),
529 MI->getOperand(0).getReg())
530 .addReg(MI->getOperand(1).getReg())
532 MI->eraseFromParent();
535 case AMDGPU::FNEG_SI: {
536 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
537 const SIInstrInfo *TII =
538 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
539 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
540 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
543 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_XOR_B32_e32),
544 MI->getOperand(0).getReg())
545 .addReg(MI->getOperand(1).getReg())
547 MI->eraseFromParent();
550 case AMDGPU::FCLAMP_SI: {
551 const SIInstrInfo *TII =
552 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
553 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F32_e64),
554 MI->getOperand(0).getReg())
555 .addImm(0) // SRC0 modifiers
556 .addOperand(MI->getOperand(1))
557 .addImm(0) // SRC1 modifiers
561 MI->eraseFromParent();
567 EVT SITargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
568 if (!VT.isVector()) {
571 return MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
574 MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
578 bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
579 VT = VT.getScalarType();
584 switch (VT.getSimpleVT().SimpleTy) {
586 return false; /* There is V_MAD_F32 for f32 */
596 //===----------------------------------------------------------------------===//
597 // Custom DAG Lowering Operations
598 //===----------------------------------------------------------------------===//
600 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
601 MachineFunction &MF = DAG.getMachineFunction();
602 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
603 switch (Op.getOpcode()) {
604 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
605 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
607 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
608 EVT VT = Op.getValueType();
610 // These loads are legal.
611 if (Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS &&
612 VT.isVector() && VT.getVectorNumElements() == 2 &&
613 VT.getVectorElementType() == MVT::i32)
616 if (Op.getValueType().isVector() &&
617 (Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
618 Load->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS ||
619 (Load->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
620 Op.getValueType().getVectorNumElements() > 4))) {
621 SDValue MergedValues[2] = {
622 SplitVectorLoad(Op, DAG),
625 return DAG.getMergeValues(MergedValues, SDLoc(Op));
627 return LowerLOAD(Op, DAG);
631 case ISD::SELECT: return LowerSELECT(Op, DAG);
632 case ISD::STORE: return LowerSTORE(Op, DAG);
633 case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
634 case ISD::INTRINSIC_WO_CHAIN: {
635 unsigned IntrinsicID =
636 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
637 EVT VT = Op.getValueType();
639 //XXX: Hardcoded we only use two to store the pointer to the parameters.
640 unsigned NumUserSGPRs = 2;
641 switch (IntrinsicID) {
642 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
643 case Intrinsic::r600_read_ngroups_x:
644 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 0, false);
645 case Intrinsic::r600_read_ngroups_y:
646 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 4, false);
647 case Intrinsic::r600_read_ngroups_z:
648 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 8, false);
649 case Intrinsic::r600_read_global_size_x:
650 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 12, false);
651 case Intrinsic::r600_read_global_size_y:
652 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 16, false);
653 case Intrinsic::r600_read_global_size_z:
654 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 20, false);
655 case Intrinsic::r600_read_local_size_x:
656 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 24, false);
657 case Intrinsic::r600_read_local_size_y:
658 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 28, false);
659 case Intrinsic::r600_read_local_size_z:
660 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 32, false);
661 case Intrinsic::r600_read_tgid_x:
662 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
663 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 0), VT);
664 case Intrinsic::r600_read_tgid_y:
665 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
666 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 1), VT);
667 case Intrinsic::r600_read_tgid_z:
668 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
669 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 2), VT);
670 case Intrinsic::r600_read_tidig_x:
671 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
673 case Intrinsic::r600_read_tidig_y:
674 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
676 case Intrinsic::r600_read_tidig_z:
677 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
679 case AMDGPUIntrinsic::SI_load_const: {
685 MachineMemOperand *MMO = MF.getMachineMemOperand(
686 MachinePointerInfo(),
687 MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
688 VT.getSizeInBits() / 8, 4);
689 return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
690 Op->getVTList(), Ops, VT, MMO);
692 case AMDGPUIntrinsic::SI_sample:
693 return LowerSampleIntrinsic(AMDGPUISD::SAMPLE, Op, DAG);
694 case AMDGPUIntrinsic::SI_sampleb:
695 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEB, Op, DAG);
696 case AMDGPUIntrinsic::SI_sampled:
697 return LowerSampleIntrinsic(AMDGPUISD::SAMPLED, Op, DAG);
698 case AMDGPUIntrinsic::SI_samplel:
699 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEL, Op, DAG);
700 case AMDGPUIntrinsic::SI_vs_load_input:
701 return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
708 case ISD::INTRINSIC_VOID:
709 SDValue Chain = Op.getOperand(0);
710 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
712 switch (IntrinsicID) {
713 case AMDGPUIntrinsic::SI_tbuffer_store: {
731 EVT VT = Op.getOperand(3).getValueType();
733 MachineMemOperand *MMO = MF.getMachineMemOperand(
734 MachinePointerInfo(),
735 MachineMemOperand::MOStore,
736 VT.getSizeInBits() / 8, 4);
737 return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL,
738 Op->getVTList(), Ops, VT, MMO);
747 /// \brief Helper function for LowerBRCOND
748 static SDNode *findUser(SDValue Value, unsigned Opcode) {
750 SDNode *Parent = Value.getNode();
751 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
754 if (I.getUse().get() != Value)
757 if (I->getOpcode() == Opcode)
763 /// This transforms the control flow intrinsics to get the branch destination as
764 /// last parameter, also switches branch target with BR if the need arise
765 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
766 SelectionDAG &DAG) const {
770 SDNode *Intr = BRCOND.getOperand(1).getNode();
771 SDValue Target = BRCOND.getOperand(2);
772 SDNode *BR = nullptr;
774 if (Intr->getOpcode() == ISD::SETCC) {
775 // As long as we negate the condition everything is fine
776 SDNode *SetCC = Intr;
777 assert(SetCC->getConstantOperandVal(1) == 1);
778 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
780 Intr = SetCC->getOperand(0).getNode();
783 // Get the target from BR if we don't negate the condition
784 BR = findUser(BRCOND, ISD::BR);
785 Target = BR->getOperand(1);
788 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
790 // Build the result and
791 SmallVector<EVT, 4> Res;
792 for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
793 Res.push_back(Intr->getValueType(i));
795 // operands of the new intrinsic call
796 SmallVector<SDValue, 4> Ops;
797 Ops.push_back(BRCOND.getOperand(0));
798 for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
799 Ops.push_back(Intr->getOperand(i));
800 Ops.push_back(Target);
802 // build the new intrinsic call
803 SDNode *Result = DAG.getNode(
804 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
805 DAG.getVTList(Res), Ops).getNode();
808 // Give the branch instruction our target
813 DAG.MorphNodeTo(BR, ISD::BR, BR->getVTList(), Ops);
816 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
818 // Copy the intrinsic results to registers
819 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
820 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
824 Chain = DAG.getCopyToReg(
826 CopyToReg->getOperand(1),
827 SDValue(Result, i - 1),
830 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
833 // Remove the old intrinsic from the chain
834 DAG.ReplaceAllUsesOfValueWith(
835 SDValue(Intr, Intr->getNumValues() - 1),
836 Intr->getOperand(0));
841 SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
843 LoadSDNode *Load = cast<LoadSDNode>(Op);
844 SDValue Ret = AMDGPUTargetLowering::LowerLOAD(Op, DAG);
845 SDValue MergedValues[2];
846 MergedValues[1] = Load->getChain();
848 MergedValues[0] = Ret;
849 return DAG.getMergeValues(MergedValues, DL);
852 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
856 EVT MemVT = Load->getMemoryVT();
858 assert(!MemVT.isVector() && "Private loads should be scalarized");
859 assert(!MemVT.isFloatingPoint() && "FP loads should be promoted to int");
861 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
862 DAG.getConstant(2, MVT::i32));
863 Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
864 Load->getChain(), Ptr,
865 DAG.getTargetConstant(0, MVT::i32),
867 if (MemVT.getSizeInBits() == 64) {
868 SDValue IncPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
869 DAG.getConstant(1, MVT::i32));
871 SDValue LoadUpper = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
872 Load->getChain(), IncPtr,
873 DAG.getTargetConstant(0, MVT::i32),
876 Ret = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ret, LoadUpper);
879 MergedValues[0] = Ret;
880 return DAG.getMergeValues(MergedValues, DL);
884 SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
886 SelectionDAG &DAG) const {
887 return DAG.getNode(Opcode, SDLoc(Op), Op.getValueType(), Op.getOperand(1),
893 SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
894 if (Op.getValueType() != MVT::i64)
898 SDValue Cond = Op.getOperand(0);
900 SDValue Zero = DAG.getConstant(0, MVT::i32);
901 SDValue One = DAG.getConstant(1, MVT::i32);
903 SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1));
904 SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2));
906 SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero);
907 SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero);
909 SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1);
911 SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One);
912 SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One);
914 SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1);
916 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i32, Lo, Hi);
917 return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
920 SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
922 StoreSDNode *Store = cast<StoreSDNode>(Op);
923 EVT VT = Store->getMemoryVT();
925 // These stores are legal.
926 if (Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS &&
927 VT.isVector() && VT.getVectorNumElements() == 2 &&
928 VT.getVectorElementType() == MVT::i32)
931 SDValue Ret = AMDGPUTargetLowering::LowerSTORE(Op, DAG);
935 if (VT.isVector() && VT.getVectorNumElements() >= 8)
936 return SplitVectorStore(Op, DAG);
939 return DAG.getTruncStore(Store->getChain(), DL,
940 DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32),
941 Store->getBasePtr(), MVT::i1, Store->getMemOperand());
943 if (Store->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
946 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Store->getBasePtr(),
947 DAG.getConstant(2, MVT::i32));
948 SDValue Chain = Store->getChain();
949 SmallVector<SDValue, 8> Values;
951 if (Store->isTruncatingStore()) {
953 if (Store->getMemoryVT() == MVT::i8) {
955 } else if (Store->getMemoryVT() == MVT::i16) {
958 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
959 Chain, Store->getBasePtr(),
960 DAG.getConstant(0, MVT::i32));
961 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, Store->getBasePtr(),
962 DAG.getConstant(0x3, MVT::i32));
963 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
964 DAG.getConstant(3, MVT::i32));
965 SDValue MaskedValue = DAG.getNode(ISD::AND, DL, MVT::i32, Store->getValue(),
966 DAG.getConstant(Mask, MVT::i32));
967 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
968 MaskedValue, ShiftAmt);
969 SDValue RotrAmt = DAG.getNode(ISD::SUB, DL, MVT::i32,
970 DAG.getConstant(32, MVT::i32), ShiftAmt);
971 SDValue DstMask = DAG.getNode(ISD::ROTR, DL, MVT::i32,
972 DAG.getConstant(Mask, MVT::i32),
974 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
975 Dst = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
977 Values.push_back(Dst);
978 } else if (VT == MVT::i64) {
979 for (unsigned i = 0; i < 2; ++i) {
980 Values.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
981 Store->getValue(), DAG.getConstant(i, MVT::i32)));
983 } else if (VT == MVT::i128) {
984 for (unsigned i = 0; i < 2; ++i) {
985 for (unsigned j = 0; j < 2; ++j) {
986 Values.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
987 DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
988 Store->getValue(), DAG.getConstant(i, MVT::i32)),
989 DAG.getConstant(j, MVT::i32)));
993 Values.push_back(Store->getValue());
996 for (unsigned i = 0; i < Values.size(); ++i) {
997 SDValue PartPtr = DAG.getNode(ISD::ADD, DL, MVT::i32,
998 Ptr, DAG.getConstant(i, MVT::i32));
999 Chain = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1000 Chain, Values[i], PartPtr,
1001 DAG.getTargetConstant(0, MVT::i32));
1006 //===----------------------------------------------------------------------===//
1007 // Custom DAG optimizations
1008 //===----------------------------------------------------------------------===//
1010 SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
1011 DAGCombinerInfo &DCI) {
1012 EVT VT = N->getValueType(0);
1013 EVT ScalarVT = VT.getScalarType();
1014 if (ScalarVT != MVT::f32)
1017 SelectionDAG &DAG = DCI.DAG;
1020 SDValue Src = N->getOperand(0);
1021 EVT SrcVT = Src.getValueType();
1023 // TODO: We could try to match extracting the higher bytes, which would be
1024 // easier if i8 vectors weren't promoted to i32 vectors, particularly after
1025 // types are legalized. v4i8 -> v4f32 is probably the only case to worry
1026 // about in practice.
1027 if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) {
1028 if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) {
1029 SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src);
1030 DCI.AddToWorklist(Cvt.getNode());
1035 // We are primarily trying to catch operations on illegal vector types
1036 // before they are expanded.
1037 // For scalars, we can use the more flexible method of checking masked bits
1038 // after legalization.
1039 if (!DCI.isBeforeLegalize() ||
1040 !SrcVT.isVector() ||
1041 SrcVT.getVectorElementType() != MVT::i8) {
1045 assert(DCI.isBeforeLegalize() && "Unexpected legal type");
1047 // Weird sized vectors are a pain to handle, but we know 3 is really the same
1049 unsigned NElts = SrcVT.getVectorNumElements();
1050 if (!SrcVT.isSimple() && NElts != 3)
1053 // Handle v4i8 -> v4f32 extload. Replace the v4i8 with a legal i32 load to
1054 // prevent a mess from expanding to v4i32 and repacking.
1055 if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse()) {
1056 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), SrcVT);
1057 EVT RegVT = getEquivalentLoadRegType(*DAG.getContext(), SrcVT);
1058 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f32, NElts);
1060 LoadSDNode *Load = cast<LoadSDNode>(Src);
1061 SDValue NewLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegVT,
1065 Load->getMemOperand());
1067 // Make sure successors of the original load stay after it by updating
1068 // them to use the new Chain.
1069 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), NewLoad.getValue(1));
1071 SmallVector<SDValue, 4> Elts;
1072 if (RegVT.isVector())
1073 DAG.ExtractVectorElements(NewLoad, Elts);
1075 Elts.push_back(NewLoad);
1077 SmallVector<SDValue, 4> Ops;
1079 unsigned EltIdx = 0;
1080 for (SDValue Elt : Elts) {
1081 unsigned ComponentsInElt = std::min(4u, NElts - 4 * EltIdx);
1082 for (unsigned I = 0; I < ComponentsInElt; ++I) {
1083 unsigned Opc = AMDGPUISD::CVT_F32_UBYTE0 + I;
1084 SDValue Cvt = DAG.getNode(Opc, DL, MVT::f32, Elt);
1085 DCI.AddToWorklist(Cvt.getNode());
1092 assert(Ops.size() == NElts);
1094 return DAG.getNode(ISD::BUILD_VECTOR, DL, FloatVT, Ops);
1100 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
1101 DAGCombinerInfo &DCI) const {
1102 SelectionDAG &DAG = DCI.DAG;
1104 EVT VT = N->getValueType(0);
1106 switch (N->getOpcode()) {
1107 default: return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1108 case ISD::SELECT_CC: {
1109 ConstantSDNode *True, *False;
1110 // i1 selectcc(l, r, -1, 0, cc) -> i1 setcc(l, r, cc)
1111 if ((True = dyn_cast<ConstantSDNode>(N->getOperand(2)))
1112 && (False = dyn_cast<ConstantSDNode>(N->getOperand(3)))
1113 && True->isAllOnesValue()
1114 && False->isNullValue()
1116 return DAG.getNode(ISD::SETCC, DL, VT, N->getOperand(0),
1117 N->getOperand(1), N->getOperand(4));
1123 SDValue Arg0 = N->getOperand(0);
1124 SDValue Arg1 = N->getOperand(1);
1125 SDValue CC = N->getOperand(2);
1126 ConstantSDNode * C = nullptr;
1127 ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
1129 // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
1131 && Arg0.getOpcode() == ISD::SIGN_EXTEND
1132 && Arg0.getOperand(0).getValueType() == MVT::i1
1133 && (C = dyn_cast<ConstantSDNode>(Arg1))
1135 && CCOp == ISD::SETNE) {
1136 return SimplifySetCC(VT, Arg0.getOperand(0),
1137 DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
1142 case AMDGPUISD::CVT_F32_UBYTE0:
1143 case AMDGPUISD::CVT_F32_UBYTE1:
1144 case AMDGPUISD::CVT_F32_UBYTE2:
1145 case AMDGPUISD::CVT_F32_UBYTE3: {
1146 unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
1148 SDValue Src = N->getOperand(0);
1149 APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);
1151 APInt KnownZero, KnownOne;
1152 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1153 !DCI.isBeforeLegalizeOps());
1154 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1155 if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
1156 TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {
1157 DCI.CommitTargetLoweringOpt(TLO);
1163 case ISD::UINT_TO_FP: {
1164 return performUCharToFloatCombine(N, DCI);
1168 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1171 /// \brief Test if RegClass is one of the VSrc classes
1172 static bool isVSrc(unsigned RegClass) {
1173 return AMDGPU::VSrc_32RegClassID == RegClass ||
1174 AMDGPU::VSrc_64RegClassID == RegClass;
1177 /// \brief Test if RegClass is one of the SSrc classes
1178 static bool isSSrc(unsigned RegClass) {
1179 return AMDGPU::SSrc_32RegClassID == RegClass ||
1180 AMDGPU::SSrc_64RegClassID == RegClass;
1183 /// \brief Analyze the possible immediate value Op
1185 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
1186 /// and the immediate value if it's a literal immediate
1187 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
1194 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
1195 if (Node->getZExtValue() >> 32) {
1198 Imm.I = Node->getSExtValue();
1199 } else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
1200 if (N->getValueType(0) != MVT::f32)
1202 Imm.F = Node->getValueAPF().convertToFloat();
1204 return -1; // It isn't an immediate
1206 if ((Imm.I >= -16 && Imm.I <= 64) ||
1207 Imm.F == 0.5f || Imm.F == -0.5f ||
1208 Imm.F == 1.0f || Imm.F == -1.0f ||
1209 Imm.F == 2.0f || Imm.F == -2.0f ||
1210 Imm.F == 4.0f || Imm.F == -4.0f)
1211 return 0; // It's an inline immediate
1213 return Imm.I; // It's a literal immediate
1216 /// \brief Try to fold an immediate directly into an instruction
1217 bool SITargetLowering::foldImm(SDValue &Operand, int32_t &Immediate,
1218 bool &ScalarSlotUsed) const {
1220 MachineSDNode *Mov = dyn_cast<MachineSDNode>(Operand);
1221 const SIInstrInfo *TII =
1222 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1223 if (!Mov || !TII->isMov(Mov->getMachineOpcode()))
1226 const SDValue &Op = Mov->getOperand(0);
1227 int32_t Value = analyzeImmediate(Op.getNode());
1229 // Not an immediate at all
1232 } else if (Value == 0) {
1233 // Inline immediates can always be fold
1237 } else if (Value == Immediate) {
1238 // Already fold literal immediate
1242 } else if (!ScalarSlotUsed && !Immediate) {
1243 // Fold this literal immediate
1244 ScalarSlotUsed = true;
1254 const TargetRegisterClass *SITargetLowering::getRegClassForNode(
1255 SelectionDAG &DAG, const SDValue &Op) const {
1256 const SIInstrInfo *TII =
1257 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1258 const SIRegisterInfo &TRI = TII->getRegisterInfo();
1260 if (!Op->isMachineOpcode()) {
1261 switch(Op->getOpcode()) {
1262 case ISD::CopyFromReg: {
1263 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1264 unsigned Reg = cast<RegisterSDNode>(Op->getOperand(1))->getReg();
1265 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
1266 return MRI.getRegClass(Reg);
1268 return TRI.getPhysRegClass(Reg);
1270 default: return nullptr;
1273 const MCInstrDesc &Desc = TII->get(Op->getMachineOpcode());
1274 int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
1275 if (OpClassID != -1) {
1276 return TRI.getRegClass(OpClassID);
1278 switch(Op.getMachineOpcode()) {
1279 case AMDGPU::COPY_TO_REGCLASS:
1280 // Operand 1 is the register class id for COPY_TO_REGCLASS instructions.
1281 OpClassID = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1283 // If the COPY_TO_REGCLASS instruction is copying to a VSrc register
1284 // class, then the register class for the value could be either a
1285 // VReg or and SReg. In order to get a more accurate
1286 if (OpClassID == AMDGPU::VSrc_32RegClassID ||
1287 OpClassID == AMDGPU::VSrc_64RegClassID) {
1288 return getRegClassForNode(DAG, Op.getOperand(0));
1290 return TRI.getRegClass(OpClassID);
1291 case AMDGPU::EXTRACT_SUBREG: {
1292 int SubIdx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1293 const TargetRegisterClass *SuperClass =
1294 getRegClassForNode(DAG, Op.getOperand(0));
1295 return TRI.getSubClassWithSubReg(SuperClass, SubIdx);
1297 case AMDGPU::REG_SEQUENCE:
1298 // Operand 0 is the register class id for REG_SEQUENCE instructions.
1299 return TRI.getRegClass(
1300 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue());
1302 return getRegClassFor(Op.getSimpleValueType());
1306 /// \brief Does "Op" fit into register class "RegClass" ?
1307 bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, const SDValue &Op,
1308 unsigned RegClass) const {
1309 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1310 const TargetRegisterClass *RC = getRegClassForNode(DAG, Op);
1314 return TRI->getRegClass(RegClass)->hasSubClassEq(RC);
1317 /// \brief Make sure that we don't exeed the number of allowed scalars
1318 void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
1320 bool &ScalarSlotUsed) const {
1322 // First map the operands register class to a destination class
1323 if (RegClass == AMDGPU::VSrc_32RegClassID)
1324 RegClass = AMDGPU::VReg_32RegClassID;
1325 else if (RegClass == AMDGPU::VSrc_64RegClassID)
1326 RegClass = AMDGPU::VReg_64RegClassID;
1330 // Nothing to do if they fit naturally
1331 if (fitsRegClass(DAG, Operand, RegClass))
1334 // If the scalar slot isn't used yet use it now
1335 if (!ScalarSlotUsed) {
1336 ScalarSlotUsed = true;
1340 // This is a conservative aproach. It is possible that we can't determine the
1341 // correct register class and copy too often, but better safe than sorry.
1342 SDValue RC = DAG.getTargetConstant(RegClass, MVT::i32);
1343 SDNode *Node = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, SDLoc(),
1344 Operand.getValueType(), Operand, RC);
1345 Operand = SDValue(Node, 0);
1348 /// \returns true if \p Node's operands are different from the SDValue list
1350 static bool isNodeChanged(const SDNode *Node, const std::vector<SDValue> &Ops) {
1351 for (unsigned i = 0, e = Node->getNumOperands(); i < e; ++i) {
1352 if (Ops[i].getNode() != Node->getOperand(i).getNode()) {
1359 /// \brief Try to fold the Nodes operands into the Node
1360 SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
1361 SelectionDAG &DAG) const {
1363 // Original encoding (either e32 or e64)
1364 int Opcode = Node->getMachineOpcode();
1365 const SIInstrInfo *TII =
1366 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1367 const MCInstrDesc *Desc = &TII->get(Opcode);
1369 unsigned NumDefs = Desc->getNumDefs();
1370 unsigned NumOps = Desc->getNumOperands();
1372 // Commuted opcode if available
1373 int OpcodeRev = Desc->isCommutable() ? TII->commuteOpcode(Opcode) : -1;
1374 const MCInstrDesc *DescRev = OpcodeRev == -1 ? nullptr : &TII->get(OpcodeRev);
1376 assert(!DescRev || DescRev->getNumDefs() == NumDefs);
1377 assert(!DescRev || DescRev->getNumOperands() == NumOps);
1379 // e64 version if available, -1 otherwise
1380 int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
1381 const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? nullptr : &TII->get(OpcodeE64);
1382 int InputModifiers[3] = {0};
1384 assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
1386 int32_t Immediate = Desc->getSize() == 4 ? 0 : -1;
1387 bool HaveVSrc = false, HaveSSrc = false;
1389 // First figure out what we already have in this instruction.
1390 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1391 i != e && Op < NumOps; ++i, ++Op) {
1393 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1394 if (isVSrc(RegClass))
1396 else if (isSSrc(RegClass))
1401 int32_t Imm = analyzeImmediate(Node->getOperand(i).getNode());
1402 if (Imm != -1 && Imm != 0) {
1403 // Literal immediate
1408 // If we neither have VSrc nor SSrc, it makes no sense to continue.
1409 if (!HaveVSrc && !HaveSSrc)
1412 // No scalar allowed when we have both VSrc and SSrc
1413 bool ScalarSlotUsed = HaveVSrc && HaveSSrc;
1415 // Second go over the operands and try to fold them
1416 std::vector<SDValue> Ops;
1417 bool Promote2e64 = false;
1418 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1419 i != e && Op < NumOps; ++i, ++Op) {
1421 const SDValue &Operand = Node->getOperand(i);
1422 Ops.push_back(Operand);
1424 // Already folded immediate?
1425 if (isa<ConstantSDNode>(Operand.getNode()) ||
1426 isa<ConstantFPSDNode>(Operand.getNode()))
1429 // Is this a VSrc or SSrc operand?
1430 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1431 if (isVSrc(RegClass) || isSSrc(RegClass)) {
1432 // Try to fold the immediates
1433 if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
1434 // Folding didn't work, make sure we don't hit the SReg limit.
1435 ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
1440 if (i == 1 && DescRev && fitsRegClass(DAG, Ops[0], RegClass)) {
1442 unsigned OtherRegClass = Desc->OpInfo[NumDefs].RegClass;
1443 assert(isVSrc(OtherRegClass) || isSSrc(OtherRegClass));
1445 // Test if it makes sense to swap operands
1446 if (foldImm(Ops[1], Immediate, ScalarSlotUsed) ||
1447 (!fitsRegClass(DAG, Ops[1], RegClass) &&
1448 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1450 // Swap commutable operands
1451 std::swap(Ops[0], Ops[1]);
1463 // Test if it makes sense to switch to e64 encoding
1464 unsigned OtherRegClass = DescE64->OpInfo[Op].RegClass;
1465 if (!isVSrc(OtherRegClass) && !isSSrc(OtherRegClass))
1468 int32_t TmpImm = -1;
1469 if (foldImm(Ops[i], TmpImm, ScalarSlotUsed) ||
1470 (!fitsRegClass(DAG, Ops[i], RegClass) &&
1471 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1473 // Switch to e64 encoding
1481 if (!DescE64 && !Promote2e64)
1483 if (!Operand.isMachineOpcode())
1485 if (Operand.getMachineOpcode() == AMDGPU::FNEG_SI) {
1487 Ops.push_back(Operand.getOperand(0));
1488 InputModifiers[i] = 1;
1495 else if (Operand.getMachineOpcode() == AMDGPU::FABS_SI) {
1497 Ops.push_back(Operand.getOperand(0));
1498 InputModifiers[i] = 2;
1508 std::vector<SDValue> OldOps(Ops);
1510 for (unsigned i = 0; i < OldOps.size(); ++i) {
1512 Ops.push_back(DAG.getTargetConstant(InputModifiers[i], MVT::i32));
1513 Ops.push_back(OldOps[i]);
1515 // Add the modifier flags while promoting
1516 for (unsigned i = 0; i < 2; ++i)
1517 Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
1520 // Add optional chain and glue
1521 for (unsigned i = NumOps - NumDefs, e = Node->getNumOperands(); i < e; ++i)
1522 Ops.push_back(Node->getOperand(i));
1524 // Nodes that have a glue result are not CSE'd by getMachineNode(), so in
1525 // this case a brand new node is always be created, even if the operands
1526 // are the same as before. So, manually check if anything has been changed.
1527 if (Desc->Opcode == Opcode && !isNodeChanged(Node, Ops)) {
1531 // Create a complete new instruction
1532 return DAG.getMachineNode(Desc->Opcode, SDLoc(Node), Node->getVTList(), Ops);
1535 /// \brief Helper function for adjustWritemask
1536 static unsigned SubIdx2Lane(unsigned Idx) {
1539 case AMDGPU::sub0: return 0;
1540 case AMDGPU::sub1: return 1;
1541 case AMDGPU::sub2: return 2;
1542 case AMDGPU::sub3: return 3;
1546 /// \brief Adjust the writemask of MIMG instructions
1547 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
1548 SelectionDAG &DAG) const {
1549 SDNode *Users[4] = { };
1551 unsigned OldDmask = Node->getConstantOperandVal(0);
1552 unsigned NewDmask = 0;
1554 // Try to figure out the used register components
1555 for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
1558 // Abort if we can't understand the usage
1559 if (!I->isMachineOpcode() ||
1560 I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
1563 // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used.
1564 // Note that subregs are packed, i.e. Lane==0 is the first bit set
1565 // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit
1567 Lane = SubIdx2Lane(I->getConstantOperandVal(1));
1569 // Set which texture component corresponds to the lane.
1571 for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) {
1573 Comp = countTrailingZeros(Dmask);
1574 Dmask &= ~(1 << Comp);
1577 // Abort if we have more than one user per component
1582 NewDmask |= 1 << Comp;
1585 // Abort if there's no change
1586 if (NewDmask == OldDmask)
1589 // Adjust the writemask in the node
1590 std::vector<SDValue> Ops;
1591 Ops.push_back(DAG.getTargetConstant(NewDmask, MVT::i32));
1592 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
1593 Ops.push_back(Node->getOperand(i));
1594 Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops);
1596 // If we only got one lane, replace it with a copy
1597 // (if NewDmask has only one bit set...)
1598 if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
1599 SDValue RC = DAG.getTargetConstant(AMDGPU::VReg_32RegClassID, MVT::i32);
1600 SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1601 SDLoc(), Users[Lane]->getValueType(0),
1602 SDValue(Node, 0), RC);
1603 DAG.ReplaceAllUsesWith(Users[Lane], Copy);
1607 // Update the users of the node with the new indices
1608 for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
1610 SDNode *User = Users[i];
1614 SDValue Op = DAG.getTargetConstant(Idx, MVT::i32);
1615 DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
1619 case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
1620 case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
1621 case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
1626 /// \brief Fold the instructions after selecting them.
1627 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
1628 SelectionDAG &DAG) const {
1629 const SIInstrInfo *TII =
1630 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1631 Node = AdjustRegClass(Node, DAG);
1633 if (TII->isMIMG(Node->getMachineOpcode()))
1634 adjustWritemask(Node, DAG);
1636 return foldOperands(Node, DAG);
1639 /// \brief Assign the register class depending on the number of
1640 /// bits set in the writemask
1641 void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
1642 SDNode *Node) const {
1643 const SIInstrInfo *TII =
1644 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1645 if (!TII->isMIMG(MI->getOpcode()))
1648 unsigned VReg = MI->getOperand(0).getReg();
1649 unsigned Writemask = MI->getOperand(1).getImm();
1650 unsigned BitsSet = 0;
1651 for (unsigned i = 0; i < 4; ++i)
1652 BitsSet += Writemask & (1 << i) ? 1 : 0;
1654 const TargetRegisterClass *RC;
1657 case 1: RC = &AMDGPU::VReg_32RegClass; break;
1658 case 2: RC = &AMDGPU::VReg_64RegClass; break;
1659 case 3: RC = &AMDGPU::VReg_96RegClass; break;
1662 unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
1663 MI->setDesc(TII->get(NewOpcode));
1664 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1665 MRI.setRegClass(VReg, RC);
1668 MachineSDNode *SITargetLowering::AdjustRegClass(MachineSDNode *N,
1669 SelectionDAG &DAG) const {
1672 unsigned NewOpcode = N->getMachineOpcode();
1674 switch (N->getMachineOpcode()) {
1676 case AMDGPU::S_LOAD_DWORD_IMM:
1677 NewOpcode = AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1679 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1680 if (NewOpcode == N->getMachineOpcode()) {
1681 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1684 case AMDGPU::S_LOAD_DWORDX4_IMM:
1685 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1686 if (NewOpcode == N->getMachineOpcode()) {
1687 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1689 if (fitsRegClass(DAG, N->getOperand(0), AMDGPU::SReg_64RegClassID)) {
1692 ConstantSDNode *Offset = cast<ConstantSDNode>(N->getOperand(1));
1694 SDValue(DAG.getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::i128,
1695 DAG.getConstant(0, MVT::i64)), 0),
1697 DAG.getConstant(Offset->getSExtValue() << 2, MVT::i32)
1699 return DAG.getMachineNode(NewOpcode, DL, N->getVTList(), Ops);
1704 SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1705 const TargetRegisterClass *RC,
1706 unsigned Reg, EVT VT) const {
1707 SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT);
1709 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()),
1710 cast<RegisterSDNode>(VReg)->getReg(), VT);