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::f32, Promote);
103 AddPromotedToType(ISD::SELECT, MVT::f32, MVT::i32);
104 setOperationAction(ISD::SELECT, MVT::i64, Custom);
105 setOperationAction(ISD::SELECT, MVT::f64, Promote);
106 AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
108 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
109 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
111 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
113 setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
114 setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
116 setOperationAction(ISD::ANY_EXTEND, MVT::i64, Custom);
117 setOperationAction(ISD::SIGN_EXTEND, MVT::i64, Custom);
118 setOperationAction(ISD::ZERO_EXTEND, MVT::i64, Custom);
120 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
121 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
122 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
124 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
125 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
128 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
129 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
130 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
132 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Custom);
134 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
136 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
137 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
138 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom);
139 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
141 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
143 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
144 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Custom);
145 setLoadExtAction(ISD::SEXTLOAD, MVT::i16, Custom);
146 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Expand);
147 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, Expand);
148 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, Expand);
150 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
151 setLoadExtAction(ISD::ZEXTLOAD, MVT::i8, Custom);
152 setLoadExtAction(ISD::ZEXTLOAD, MVT::i16, Custom);
153 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Expand);
155 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
156 setLoadExtAction(ISD::EXTLOAD, MVT::i8, Custom);
157 setLoadExtAction(ISD::EXTLOAD, MVT::i16, Custom);
158 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Expand);
159 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
161 setTruncStoreAction(MVT::i32, MVT::i8, Custom);
162 setTruncStoreAction(MVT::i32, MVT::i16, Custom);
163 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
164 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
165 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Expand);
166 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Expand);
168 setOperationAction(ISD::LOAD, MVT::i1, Custom);
170 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
171 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
172 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
174 // These should use UDIVREM, so set them to expand
175 setOperationAction(ISD::UDIV, MVT::i64, Expand);
176 setOperationAction(ISD::UREM, MVT::i64, Expand);
178 // We only support LOAD/STORE and vector manipulation ops for vectors
179 // with > 4 elements.
181 MVT::v8i32, MVT::v8f32, MVT::v16i32, MVT::v16f32
184 for (MVT VT : VecTypes) {
185 for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
189 case ISD::BUILD_VECTOR:
191 case ISD::EXTRACT_VECTOR_ELT:
192 case ISD::INSERT_VECTOR_ELT:
193 case ISD::CONCAT_VECTORS:
194 case ISD::INSERT_SUBVECTOR:
195 case ISD::EXTRACT_SUBVECTOR:
198 setOperationAction(Op, VT, Expand);
204 for (int I = MVT::v1f64; I <= MVT::v8f64; ++I) {
205 MVT::SimpleValueType VT = static_cast<MVT::SimpleValueType>(I);
206 setOperationAction(ISD::FTRUNC, VT, Expand);
207 setOperationAction(ISD::FCEIL, VT, Expand);
208 setOperationAction(ISD::FFLOOR, VT, Expand);
211 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
212 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
213 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
214 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
215 setOperationAction(ISD::FRINT, MVT::f64, Legal);
218 setTargetDAGCombine(ISD::SELECT_CC);
219 setTargetDAGCombine(ISD::SETCC);
221 setSchedulingPreference(Sched::RegPressure);
224 //===----------------------------------------------------------------------===//
225 // TargetLowering queries
226 //===----------------------------------------------------------------------===//
228 bool SITargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
230 bool *IsFast) const {
234 // XXX: This depends on the address space and also we may want to revist
235 // the alignment values we specify in the DataLayout.
237 // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
238 // which isn't a simple VT.
239 if (!VT.isSimple() || VT == MVT::Other)
242 // XXX - CI changes say "Support for unaligned memory accesses" but I don't
243 // see what for specifically. The wording everywhere else seems to be the
246 // 3.6.4 - Operations using pairs of VGPRs (for example: double-floats) have
247 // no alignment restrictions.
248 if (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS) {
249 // Using any pair of GPRs should be the same as any other pair.
252 return VT.bitsGE(MVT::i64);
255 // XXX - The only mention I see of this in the ISA manual is for LDS direct
256 // reads the "byte address and must be dword aligned". Is it also true for the
257 // normal loads and stores?
258 if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS)
261 // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
262 // byte-address are ignored, thus forcing Dword alignment.
265 return VT.bitsGT(MVT::i32);
268 bool SITargetLowering::shouldSplitVectorType(EVT VT) const {
269 return VT.getScalarType().bitsLE(MVT::i16);
272 bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
274 const SIInstrInfo *TII =
275 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
276 return TII->isInlineConstant(Imm);
279 SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
280 SDLoc DL, SDValue Chain,
281 unsigned Offset, bool Signed) const {
282 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
283 PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
284 AMDGPUAS::CONSTANT_ADDRESS);
285 SDValue BasePtr = DAG.getCopyFromReg(Chain, DL,
286 MRI.getLiveInVirtReg(AMDGPU::SGPR0_SGPR1), MVT::i64);
287 SDValue Ptr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
288 DAG.getConstant(Offset, MVT::i64));
289 return DAG.getExtLoad(Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD, DL, VT, Chain, Ptr,
290 MachinePointerInfo(UndefValue::get(PtrTy)), MemVT,
291 false, false, MemVT.getSizeInBits() >> 3);
295 SDValue SITargetLowering::LowerFormalArguments(
297 CallingConv::ID CallConv,
299 const SmallVectorImpl<ISD::InputArg> &Ins,
300 SDLoc DL, SelectionDAG &DAG,
301 SmallVectorImpl<SDValue> &InVals) const {
303 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
305 MachineFunction &MF = DAG.getMachineFunction();
306 FunctionType *FType = MF.getFunction()->getFunctionType();
307 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
309 assert(CallConv == CallingConv::C);
311 SmallVector<ISD::InputArg, 16> Splits;
312 uint32_t Skipped = 0;
314 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
315 const ISD::InputArg &Arg = Ins[i];
317 // First check if it's a PS input addr
318 if (Info->ShaderType == ShaderType::PIXEL && !Arg.Flags.isInReg() &&
319 !Arg.Flags.isByVal()) {
321 assert((PSInputNum <= 15) && "Too many PS inputs!");
324 // We can savely skip PS inputs
330 Info->PSInputAddr |= 1 << PSInputNum++;
333 // Second split vertices into their elements
334 if (Info->ShaderType != ShaderType::COMPUTE && Arg.VT.isVector()) {
335 ISD::InputArg NewArg = Arg;
336 NewArg.Flags.setSplit();
337 NewArg.VT = Arg.VT.getVectorElementType();
339 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
340 // three or five element vertex only needs three or five registers,
341 // NOT four or eigth.
342 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
343 unsigned NumElements = ParamType->getVectorNumElements();
345 for (unsigned j = 0; j != NumElements; ++j) {
346 Splits.push_back(NewArg);
347 NewArg.PartOffset += NewArg.VT.getStoreSize();
350 } else if (Info->ShaderType != ShaderType::COMPUTE) {
351 Splits.push_back(Arg);
355 SmallVector<CCValAssign, 16> ArgLocs;
356 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
357 getTargetMachine(), ArgLocs, *DAG.getContext());
359 // At least one interpolation mode must be enabled or else the GPU will hang.
360 if (Info->ShaderType == ShaderType::PIXEL && (Info->PSInputAddr & 0x7F) == 0) {
361 Info->PSInputAddr |= 1;
362 CCInfo.AllocateReg(AMDGPU::VGPR0);
363 CCInfo.AllocateReg(AMDGPU::VGPR1);
366 // The pointer to the list of arguments is stored in SGPR0, SGPR1
367 if (Info->ShaderType == ShaderType::COMPUTE) {
368 CCInfo.AllocateReg(AMDGPU::SGPR0);
369 CCInfo.AllocateReg(AMDGPU::SGPR1);
370 MF.addLiveIn(AMDGPU::SGPR0_SGPR1, &AMDGPU::SReg_64RegClass);
373 if (Info->ShaderType == ShaderType::COMPUTE) {
374 getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
378 AnalyzeFormalArguments(CCInfo, Splits);
380 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
382 const ISD::InputArg &Arg = Ins[i];
383 if (Skipped & (1 << i)) {
384 InVals.push_back(DAG.getUNDEF(Arg.VT));
388 CCValAssign &VA = ArgLocs[ArgIdx++];
389 EVT VT = VA.getLocVT();
393 EVT MemVT = Splits[i].VT;
394 // The first 36 bytes of the input buffer contains information about
395 // thread group and global sizes.
396 SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, DAG.getRoot(),
397 36 + VA.getLocMemOffset(),
398 Ins[i].Flags.isSExt());
399 InVals.push_back(Arg);
402 assert(VA.isRegLoc() && "Parameter must be in a register!");
404 unsigned Reg = VA.getLocReg();
406 if (VT == MVT::i64) {
407 // For now assume it is a pointer
408 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
409 &AMDGPU::SReg_64RegClass);
410 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
411 InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
415 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
417 Reg = MF.addLiveIn(Reg, RC);
418 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
420 if (Arg.VT.isVector()) {
422 // Build a vector from the registers
423 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
424 unsigned NumElements = ParamType->getVectorNumElements();
426 SmallVector<SDValue, 4> Regs;
428 for (unsigned j = 1; j != NumElements; ++j) {
429 Reg = ArgLocs[ArgIdx++].getLocReg();
430 Reg = MF.addLiveIn(Reg, RC);
431 Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
434 // Fill up the missing vector elements
435 NumElements = Arg.VT.getVectorNumElements() - NumElements;
436 for (unsigned j = 0; j != NumElements; ++j)
437 Regs.push_back(DAG.getUNDEF(VT));
439 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT, Regs));
443 InVals.push_back(Val);
448 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
449 MachineInstr * MI, MachineBasicBlock * BB) const {
451 MachineBasicBlock::iterator I = *MI;
452 const SIInstrInfo *TII =
453 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
454 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
456 switch (MI->getOpcode()) {
458 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
459 case AMDGPU::BRANCH: return BB;
460 case AMDGPU::SI_ADDR64_RSRC: {
461 unsigned SuperReg = MI->getOperand(0).getReg();
462 unsigned SubRegLo = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
463 unsigned SubRegHi = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
464 unsigned SubRegHiHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
465 unsigned SubRegHiLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
466 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B64), SubRegLo)
467 .addOperand(MI->getOperand(1));
468 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiLo)
470 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiHi)
471 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
472 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SubRegHi)
474 .addImm(AMDGPU::sub0)
476 .addImm(AMDGPU::sub1);
477 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SuperReg)
479 .addImm(AMDGPU::sub0_sub1)
481 .addImm(AMDGPU::sub2_sub3);
482 MI->eraseFromParent();
485 case AMDGPU::V_SUB_F64:
486 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F64),
487 MI->getOperand(0).getReg())
488 .addReg(MI->getOperand(1).getReg())
489 .addReg(MI->getOperand(2).getReg())
490 .addImm(0) /* src2 */
492 .addImm(0) /* CLAMP */
493 .addImm(0) /* OMOD */
494 .addImm(2); /* NEG */
495 MI->eraseFromParent();
498 case AMDGPU::SI_RegisterStorePseudo: {
499 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
500 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
501 MachineInstrBuilder MIB =
502 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::SI_RegisterStore),
504 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i)
505 MIB.addOperand(MI->getOperand(i));
507 MI->eraseFromParent();
510 case AMDGPU::FABS_SI: {
511 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
512 const SIInstrInfo *TII =
513 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
514 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
515 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
518 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_AND_B32_e32),
519 MI->getOperand(0).getReg())
520 .addReg(MI->getOperand(1).getReg())
522 MI->eraseFromParent();
525 case AMDGPU::FNEG_SI: {
526 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
527 const SIInstrInfo *TII =
528 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
529 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
530 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
533 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_XOR_B32_e32),
534 MI->getOperand(0).getReg())
535 .addReg(MI->getOperand(1).getReg())
537 MI->eraseFromParent();
540 case AMDGPU::FCLAMP_SI: {
541 const SIInstrInfo *TII =
542 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
543 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F32_e64),
544 MI->getOperand(0).getReg())
545 .addImm(0) // SRC0 modifiers
546 .addOperand(MI->getOperand(1))
547 .addImm(0) // SRC1 modifiers
551 MI->eraseFromParent();
557 EVT SITargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
558 if (!VT.isVector()) {
561 return MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
564 MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
568 bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
569 VT = VT.getScalarType();
574 switch (VT.getSimpleVT().SimpleTy) {
576 return false; /* There is V_MAD_F32 for f32 */
586 //===----------------------------------------------------------------------===//
587 // Custom DAG Lowering Operations
588 //===----------------------------------------------------------------------===//
590 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
591 MachineFunction &MF = DAG.getMachineFunction();
592 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
593 switch (Op.getOpcode()) {
594 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
595 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
597 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
598 if (Op.getValueType().isVector() &&
599 (Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
600 Load->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS ||
601 (Load->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
602 Op.getValueType().getVectorNumElements() > 4))) {
603 SDValue MergedValues[2] = {
604 SplitVectorLoad(Op, DAG),
607 return DAG.getMergeValues(MergedValues, SDLoc(Op));
609 return LowerLOAD(Op, DAG);
613 case ISD::SELECT: return LowerSELECT(Op, DAG);
614 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
615 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, DAG);
616 case ISD::STORE: return LowerSTORE(Op, DAG);
617 case ISD::ANY_EXTEND: // Fall-through
618 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, DAG);
619 case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
620 case ISD::INTRINSIC_WO_CHAIN: {
621 unsigned IntrinsicID =
622 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
623 EVT VT = Op.getValueType();
625 //XXX: Hardcoded we only use two to store the pointer to the parameters.
626 unsigned NumUserSGPRs = 2;
627 switch (IntrinsicID) {
628 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
629 case Intrinsic::r600_read_ngroups_x:
630 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 0, false);
631 case Intrinsic::r600_read_ngroups_y:
632 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 4, false);
633 case Intrinsic::r600_read_ngroups_z:
634 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 8, false);
635 case Intrinsic::r600_read_global_size_x:
636 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 12, false);
637 case Intrinsic::r600_read_global_size_y:
638 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 16, false);
639 case Intrinsic::r600_read_global_size_z:
640 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 20, false);
641 case Intrinsic::r600_read_local_size_x:
642 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 24, false);
643 case Intrinsic::r600_read_local_size_y:
644 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 28, false);
645 case Intrinsic::r600_read_local_size_z:
646 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 32, false);
647 case Intrinsic::r600_read_tgid_x:
648 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
649 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 0), VT);
650 case Intrinsic::r600_read_tgid_y:
651 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
652 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 1), VT);
653 case Intrinsic::r600_read_tgid_z:
654 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
655 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 2), VT);
656 case Intrinsic::r600_read_tidig_x:
657 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
659 case Intrinsic::r600_read_tidig_y:
660 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
662 case Intrinsic::r600_read_tidig_z:
663 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
665 case AMDGPUIntrinsic::SI_load_const: {
671 MachineMemOperand *MMO = MF.getMachineMemOperand(
672 MachinePointerInfo(),
673 MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
674 VT.getSizeInBits() / 8, 4);
675 return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
676 Op->getVTList(), Ops, VT, MMO);
678 case AMDGPUIntrinsic::SI_sample:
679 return LowerSampleIntrinsic(AMDGPUISD::SAMPLE, Op, DAG);
680 case AMDGPUIntrinsic::SI_sampleb:
681 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEB, Op, DAG);
682 case AMDGPUIntrinsic::SI_sampled:
683 return LowerSampleIntrinsic(AMDGPUISD::SAMPLED, Op, DAG);
684 case AMDGPUIntrinsic::SI_samplel:
685 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEL, Op, DAG);
686 case AMDGPUIntrinsic::SI_vs_load_input:
687 return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
694 case ISD::INTRINSIC_VOID:
695 SDValue Chain = Op.getOperand(0);
696 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
698 switch (IntrinsicID) {
699 case AMDGPUIntrinsic::SI_tbuffer_store: {
717 EVT VT = Op.getOperand(3).getValueType();
719 MachineMemOperand *MMO = MF.getMachineMemOperand(
720 MachinePointerInfo(),
721 MachineMemOperand::MOStore,
722 VT.getSizeInBits() / 8, 4);
723 return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL,
724 Op->getVTList(), Ops, VT, MMO);
733 /// \brief Helper function for LowerBRCOND
734 static SDNode *findUser(SDValue Value, unsigned Opcode) {
736 SDNode *Parent = Value.getNode();
737 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
740 if (I.getUse().get() != Value)
743 if (I->getOpcode() == Opcode)
749 /// This transforms the control flow intrinsics to get the branch destination as
750 /// last parameter, also switches branch target with BR if the need arise
751 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
752 SelectionDAG &DAG) const {
756 SDNode *Intr = BRCOND.getOperand(1).getNode();
757 SDValue Target = BRCOND.getOperand(2);
758 SDNode *BR = nullptr;
760 if (Intr->getOpcode() == ISD::SETCC) {
761 // As long as we negate the condition everything is fine
762 SDNode *SetCC = Intr;
763 assert(SetCC->getConstantOperandVal(1) == 1);
764 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
766 Intr = SetCC->getOperand(0).getNode();
769 // Get the target from BR if we don't negate the condition
770 BR = findUser(BRCOND, ISD::BR);
771 Target = BR->getOperand(1);
774 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
776 // Build the result and
777 SmallVector<EVT, 4> Res;
778 for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
779 Res.push_back(Intr->getValueType(i));
781 // operands of the new intrinsic call
782 SmallVector<SDValue, 4> Ops;
783 Ops.push_back(BRCOND.getOperand(0));
784 for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
785 Ops.push_back(Intr->getOperand(i));
786 Ops.push_back(Target);
788 // build the new intrinsic call
789 SDNode *Result = DAG.getNode(
790 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
791 DAG.getVTList(Res), Ops).getNode();
794 // Give the branch instruction our target
799 DAG.MorphNodeTo(BR, ISD::BR, BR->getVTList(), Ops);
802 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
804 // Copy the intrinsic results to registers
805 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
806 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
810 Chain = DAG.getCopyToReg(
812 CopyToReg->getOperand(1),
813 SDValue(Result, i - 1),
816 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
819 // Remove the old intrinsic from the chain
820 DAG.ReplaceAllUsesOfValueWith(
821 SDValue(Intr, Intr->getNumValues() - 1),
822 Intr->getOperand(0));
827 SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
829 LoadSDNode *Load = cast<LoadSDNode>(Op);
830 SDValue Ret = AMDGPUTargetLowering::LowerLOAD(Op, DAG);
831 SDValue MergedValues[2];
832 MergedValues[1] = Load->getChain();
834 MergedValues[0] = Ret;
835 return DAG.getMergeValues(MergedValues, DL);
838 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
842 EVT MemVT = Load->getMemoryVT();
844 assert(!MemVT.isVector() && "Private loads should be scalarized");
845 assert(!MemVT.isFloatingPoint() && "FP loads should be promoted to int");
847 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
848 DAG.getConstant(2, MVT::i32));
849 Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
850 Load->getChain(), Ptr,
851 DAG.getTargetConstant(0, MVT::i32),
853 if (MemVT.getSizeInBits() == 64) {
854 SDValue IncPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
855 DAG.getConstant(1, MVT::i32));
857 SDValue LoadUpper = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
858 Load->getChain(), IncPtr,
859 DAG.getTargetConstant(0, MVT::i32),
862 Ret = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ret, LoadUpper);
865 MergedValues[0] = Ret;
866 return DAG.getMergeValues(MergedValues, DL);
870 SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
872 SelectionDAG &DAG) const {
873 return DAG.getNode(Opcode, SDLoc(Op), Op.getValueType(), Op.getOperand(1),
879 SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
880 if (Op.getValueType() != MVT::i64)
884 SDValue Cond = Op.getOperand(0);
886 SDValue Zero = DAG.getConstant(0, MVT::i32);
887 SDValue One = DAG.getConstant(1, MVT::i32);
889 SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1));
890 SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2));
892 SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero);
893 SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero);
895 SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1);
897 SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One);
898 SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One);
900 SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1);
902 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i32, Lo, Hi);
903 return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
906 SDValue SITargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
907 SDValue LHS = Op.getOperand(0);
908 SDValue RHS = Op.getOperand(1);
909 SDValue True = Op.getOperand(2);
910 SDValue False = Op.getOperand(3);
911 SDValue CC = Op.getOperand(4);
912 EVT VT = Op.getValueType();
915 SDValue Cond = DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS, CC);
916 return DAG.getNode(ISD::SELECT, DL, VT, Cond, True, False);
919 SDValue SITargetLowering::LowerSIGN_EXTEND(SDValue Op,
920 SelectionDAG &DAG) const {
921 EVT VT = Op.getValueType();
924 if (VT != MVT::i64) {
928 SDValue Hi = DAG.getNode(ISD::SRA, DL, MVT::i32, Op.getOperand(0),
929 DAG.getConstant(31, MVT::i32));
931 return DAG.getNode(ISD::BUILD_PAIR, DL, VT, Op.getOperand(0), Hi);
934 SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
936 StoreSDNode *Store = cast<StoreSDNode>(Op);
937 EVT VT = Store->getMemoryVT();
939 SDValue Ret = AMDGPUTargetLowering::LowerSTORE(Op, DAG);
943 if (VT.isVector() && VT.getVectorNumElements() >= 8)
944 return SplitVectorStore(Op, DAG);
947 return DAG.getTruncStore(Store->getChain(), DL,
948 DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32),
949 Store->getBasePtr(), MVT::i1, Store->getMemOperand());
951 if (Store->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
954 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Store->getBasePtr(),
955 DAG.getConstant(2, MVT::i32));
956 SDValue Chain = Store->getChain();
957 SmallVector<SDValue, 8> Values;
959 if (Store->isTruncatingStore()) {
961 if (Store->getMemoryVT() == MVT::i8) {
963 } else if (Store->getMemoryVT() == MVT::i16) {
966 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
967 Chain, Store->getBasePtr(),
968 DAG.getConstant(0, MVT::i32));
969 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, Store->getBasePtr(),
970 DAG.getConstant(0x3, MVT::i32));
971 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
972 DAG.getConstant(3, MVT::i32));
973 SDValue MaskedValue = DAG.getNode(ISD::AND, DL, MVT::i32, Store->getValue(),
974 DAG.getConstant(Mask, MVT::i32));
975 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
976 MaskedValue, ShiftAmt);
977 SDValue RotrAmt = DAG.getNode(ISD::SUB, DL, MVT::i32,
978 DAG.getConstant(32, MVT::i32), ShiftAmt);
979 SDValue DstMask = DAG.getNode(ISD::ROTR, DL, MVT::i32,
980 DAG.getConstant(Mask, MVT::i32),
982 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
983 Dst = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
985 Values.push_back(Dst);
986 } else if (VT == MVT::i64) {
987 for (unsigned i = 0; i < 2; ++i) {
988 Values.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
989 Store->getValue(), DAG.getConstant(i, MVT::i32)));
991 } else if (VT == MVT::i128) {
992 for (unsigned i = 0; i < 2; ++i) {
993 for (unsigned j = 0; j < 2; ++j) {
994 Values.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
995 DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
996 Store->getValue(), DAG.getConstant(i, MVT::i32)),
997 DAG.getConstant(j, MVT::i32)));
1001 Values.push_back(Store->getValue());
1004 for (unsigned i = 0; i < Values.size(); ++i) {
1005 SDValue PartPtr = DAG.getNode(ISD::ADD, DL, MVT::i32,
1006 Ptr, DAG.getConstant(i, MVT::i32));
1007 Chain = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1008 Chain, Values[i], PartPtr,
1009 DAG.getTargetConstant(0, MVT::i32));
1015 SDValue SITargetLowering::LowerZERO_EXTEND(SDValue Op,
1016 SelectionDAG &DAG) const {
1017 EVT VT = Op.getValueType();
1020 if (VT != MVT::i64) {
1024 SDValue Src = Op.getOperand(0);
1025 if (Src.getValueType() != MVT::i32)
1026 Src = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Src);
1028 SDValue Zero = DAG.getConstant(0, MVT::i32);
1029 return DAG.getNode(ISD::BUILD_PAIR, DL, VT, Src, Zero);
1032 //===----------------------------------------------------------------------===//
1033 // Custom DAG optimizations
1034 //===----------------------------------------------------------------------===//
1036 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
1037 DAGCombinerInfo &DCI) const {
1038 SelectionDAG &DAG = DCI.DAG;
1040 EVT VT = N->getValueType(0);
1042 switch (N->getOpcode()) {
1043 default: return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1044 case ISD::SELECT_CC: {
1045 ConstantSDNode *True, *False;
1046 // i1 selectcc(l, r, -1, 0, cc) -> i1 setcc(l, r, cc)
1047 if ((True = dyn_cast<ConstantSDNode>(N->getOperand(2)))
1048 && (False = dyn_cast<ConstantSDNode>(N->getOperand(3)))
1049 && True->isAllOnesValue()
1050 && False->isNullValue()
1052 return DAG.getNode(ISD::SETCC, DL, VT, N->getOperand(0),
1053 N->getOperand(1), N->getOperand(4));
1059 SDValue Arg0 = N->getOperand(0);
1060 SDValue Arg1 = N->getOperand(1);
1061 SDValue CC = N->getOperand(2);
1062 ConstantSDNode * C = nullptr;
1063 ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
1065 // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
1067 && Arg0.getOpcode() == ISD::SIGN_EXTEND
1068 && Arg0.getOperand(0).getValueType() == MVT::i1
1069 && (C = dyn_cast<ConstantSDNode>(Arg1))
1071 && CCOp == ISD::SETNE) {
1072 return SimplifySetCC(VT, Arg0.getOperand(0),
1073 DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
1079 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1082 /// \brief Test if RegClass is one of the VSrc classes
1083 static bool isVSrc(unsigned RegClass) {
1084 return AMDGPU::VSrc_32RegClassID == RegClass ||
1085 AMDGPU::VSrc_64RegClassID == RegClass;
1088 /// \brief Test if RegClass is one of the SSrc classes
1089 static bool isSSrc(unsigned RegClass) {
1090 return AMDGPU::SSrc_32RegClassID == RegClass ||
1091 AMDGPU::SSrc_64RegClassID == RegClass;
1094 /// \brief Analyze the possible immediate value Op
1096 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
1097 /// and the immediate value if it's a literal immediate
1098 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
1105 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
1106 if (Node->getZExtValue() >> 32) {
1109 Imm.I = Node->getSExtValue();
1110 } else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
1111 if (N->getValueType(0) != MVT::f32)
1113 Imm.F = Node->getValueAPF().convertToFloat();
1115 return -1; // It isn't an immediate
1117 if ((Imm.I >= -16 && Imm.I <= 64) ||
1118 Imm.F == 0.5f || Imm.F == -0.5f ||
1119 Imm.F == 1.0f || Imm.F == -1.0f ||
1120 Imm.F == 2.0f || Imm.F == -2.0f ||
1121 Imm.F == 4.0f || Imm.F == -4.0f)
1122 return 0; // It's an inline immediate
1124 return Imm.I; // It's a literal immediate
1127 /// \brief Try to fold an immediate directly into an instruction
1128 bool SITargetLowering::foldImm(SDValue &Operand, int32_t &Immediate,
1129 bool &ScalarSlotUsed) const {
1131 MachineSDNode *Mov = dyn_cast<MachineSDNode>(Operand);
1132 const SIInstrInfo *TII =
1133 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1134 if (!Mov || !TII->isMov(Mov->getMachineOpcode()))
1137 const SDValue &Op = Mov->getOperand(0);
1138 int32_t Value = analyzeImmediate(Op.getNode());
1140 // Not an immediate at all
1143 } else if (Value == 0) {
1144 // Inline immediates can always be fold
1148 } else if (Value == Immediate) {
1149 // Already fold literal immediate
1153 } else if (!ScalarSlotUsed && !Immediate) {
1154 // Fold this literal immediate
1155 ScalarSlotUsed = true;
1165 const TargetRegisterClass *SITargetLowering::getRegClassForNode(
1166 SelectionDAG &DAG, const SDValue &Op) const {
1167 const SIInstrInfo *TII =
1168 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1169 const SIRegisterInfo &TRI = TII->getRegisterInfo();
1171 if (!Op->isMachineOpcode()) {
1172 switch(Op->getOpcode()) {
1173 case ISD::CopyFromReg: {
1174 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1175 unsigned Reg = cast<RegisterSDNode>(Op->getOperand(1))->getReg();
1176 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
1177 return MRI.getRegClass(Reg);
1179 return TRI.getPhysRegClass(Reg);
1181 default: return nullptr;
1184 const MCInstrDesc &Desc = TII->get(Op->getMachineOpcode());
1185 int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
1186 if (OpClassID != -1) {
1187 return TRI.getRegClass(OpClassID);
1189 switch(Op.getMachineOpcode()) {
1190 case AMDGPU::COPY_TO_REGCLASS:
1191 // Operand 1 is the register class id for COPY_TO_REGCLASS instructions.
1192 OpClassID = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1194 // If the COPY_TO_REGCLASS instruction is copying to a VSrc register
1195 // class, then the register class for the value could be either a
1196 // VReg or and SReg. In order to get a more accurate
1197 if (OpClassID == AMDGPU::VSrc_32RegClassID ||
1198 OpClassID == AMDGPU::VSrc_64RegClassID) {
1199 return getRegClassForNode(DAG, Op.getOperand(0));
1201 return TRI.getRegClass(OpClassID);
1202 case AMDGPU::EXTRACT_SUBREG: {
1203 int SubIdx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1204 const TargetRegisterClass *SuperClass =
1205 getRegClassForNode(DAG, Op.getOperand(0));
1206 return TRI.getSubClassWithSubReg(SuperClass, SubIdx);
1208 case AMDGPU::REG_SEQUENCE:
1209 // Operand 0 is the register class id for REG_SEQUENCE instructions.
1210 return TRI.getRegClass(
1211 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue());
1213 return getRegClassFor(Op.getSimpleValueType());
1217 /// \brief Does "Op" fit into register class "RegClass" ?
1218 bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, const SDValue &Op,
1219 unsigned RegClass) const {
1220 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1221 const TargetRegisterClass *RC = getRegClassForNode(DAG, Op);
1225 return TRI->getRegClass(RegClass)->hasSubClassEq(RC);
1228 /// \brief Make sure that we don't exeed the number of allowed scalars
1229 void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
1231 bool &ScalarSlotUsed) const {
1233 // First map the operands register class to a destination class
1234 if (RegClass == AMDGPU::VSrc_32RegClassID)
1235 RegClass = AMDGPU::VReg_32RegClassID;
1236 else if (RegClass == AMDGPU::VSrc_64RegClassID)
1237 RegClass = AMDGPU::VReg_64RegClassID;
1241 // Nothing to do if they fit naturally
1242 if (fitsRegClass(DAG, Operand, RegClass))
1245 // If the scalar slot isn't used yet use it now
1246 if (!ScalarSlotUsed) {
1247 ScalarSlotUsed = true;
1251 // This is a conservative aproach. It is possible that we can't determine the
1252 // correct register class and copy too often, but better safe than sorry.
1253 SDValue RC = DAG.getTargetConstant(RegClass, MVT::i32);
1254 SDNode *Node = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, SDLoc(),
1255 Operand.getValueType(), Operand, RC);
1256 Operand = SDValue(Node, 0);
1259 /// \returns true if \p Node's operands are different from the SDValue list
1261 static bool isNodeChanged(const SDNode *Node, const std::vector<SDValue> &Ops) {
1262 for (unsigned i = 0, e = Node->getNumOperands(); i < e; ++i) {
1263 if (Ops[i].getNode() != Node->getOperand(i).getNode()) {
1270 /// \brief Try to fold the Nodes operands into the Node
1271 SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
1272 SelectionDAG &DAG) const {
1274 // Original encoding (either e32 or e64)
1275 int Opcode = Node->getMachineOpcode();
1276 const SIInstrInfo *TII =
1277 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1278 const MCInstrDesc *Desc = &TII->get(Opcode);
1280 unsigned NumDefs = Desc->getNumDefs();
1281 unsigned NumOps = Desc->getNumOperands();
1283 // Commuted opcode if available
1284 int OpcodeRev = Desc->isCommutable() ? TII->commuteOpcode(Opcode) : -1;
1285 const MCInstrDesc *DescRev = OpcodeRev == -1 ? nullptr : &TII->get(OpcodeRev);
1287 assert(!DescRev || DescRev->getNumDefs() == NumDefs);
1288 assert(!DescRev || DescRev->getNumOperands() == NumOps);
1290 // e64 version if available, -1 otherwise
1291 int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
1292 const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? nullptr : &TII->get(OpcodeE64);
1293 int InputModifiers[3] = {0};
1295 assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
1297 int32_t Immediate = Desc->getSize() == 4 ? 0 : -1;
1298 bool HaveVSrc = false, HaveSSrc = false;
1300 // First figure out what we already have in this instruction.
1301 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1302 i != e && Op < NumOps; ++i, ++Op) {
1304 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1305 if (isVSrc(RegClass))
1307 else if (isSSrc(RegClass))
1312 int32_t Imm = analyzeImmediate(Node->getOperand(i).getNode());
1313 if (Imm != -1 && Imm != 0) {
1314 // Literal immediate
1319 // If we neither have VSrc nor SSrc, it makes no sense to continue.
1320 if (!HaveVSrc && !HaveSSrc)
1323 // No scalar allowed when we have both VSrc and SSrc
1324 bool ScalarSlotUsed = HaveVSrc && HaveSSrc;
1326 // Second go over the operands and try to fold them
1327 std::vector<SDValue> Ops;
1328 bool Promote2e64 = false;
1329 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1330 i != e && Op < NumOps; ++i, ++Op) {
1332 const SDValue &Operand = Node->getOperand(i);
1333 Ops.push_back(Operand);
1335 // Already folded immediate?
1336 if (isa<ConstantSDNode>(Operand.getNode()) ||
1337 isa<ConstantFPSDNode>(Operand.getNode()))
1340 // Is this a VSrc or SSrc operand?
1341 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1342 if (isVSrc(RegClass) || isSSrc(RegClass)) {
1343 // Try to fold the immediates
1344 if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
1345 // Folding didn't work, make sure we don't hit the SReg limit.
1346 ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
1351 if (i == 1 && DescRev && fitsRegClass(DAG, Ops[0], RegClass)) {
1353 unsigned OtherRegClass = Desc->OpInfo[NumDefs].RegClass;
1354 assert(isVSrc(OtherRegClass) || isSSrc(OtherRegClass));
1356 // Test if it makes sense to swap operands
1357 if (foldImm(Ops[1], Immediate, ScalarSlotUsed) ||
1358 (!fitsRegClass(DAG, Ops[1], RegClass) &&
1359 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1361 // Swap commutable operands
1362 std::swap(Ops[0], Ops[1]);
1374 // Test if it makes sense to switch to e64 encoding
1375 unsigned OtherRegClass = DescE64->OpInfo[Op].RegClass;
1376 if (!isVSrc(OtherRegClass) && !isSSrc(OtherRegClass))
1379 int32_t TmpImm = -1;
1380 if (foldImm(Ops[i], TmpImm, ScalarSlotUsed) ||
1381 (!fitsRegClass(DAG, Ops[i], RegClass) &&
1382 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1384 // Switch to e64 encoding
1392 if (!DescE64 && !Promote2e64)
1394 if (!Operand.isMachineOpcode())
1396 if (Operand.getMachineOpcode() == AMDGPU::FNEG_SI) {
1398 Ops.push_back(Operand.getOperand(0));
1399 InputModifiers[i] = 1;
1406 else if (Operand.getMachineOpcode() == AMDGPU::FABS_SI) {
1408 Ops.push_back(Operand.getOperand(0));
1409 InputModifiers[i] = 2;
1419 std::vector<SDValue> OldOps(Ops);
1421 for (unsigned i = 0; i < OldOps.size(); ++i) {
1423 Ops.push_back(DAG.getTargetConstant(InputModifiers[i], MVT::i32));
1424 Ops.push_back(OldOps[i]);
1426 // Add the modifier flags while promoting
1427 for (unsigned i = 0; i < 2; ++i)
1428 Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
1431 // Add optional chain and glue
1432 for (unsigned i = NumOps - NumDefs, e = Node->getNumOperands(); i < e; ++i)
1433 Ops.push_back(Node->getOperand(i));
1435 // Nodes that have a glue result are not CSE'd by getMachineNode(), so in
1436 // this case a brand new node is always be created, even if the operands
1437 // are the same as before. So, manually check if anything has been changed.
1438 if (Desc->Opcode == Opcode && !isNodeChanged(Node, Ops)) {
1442 // Create a complete new instruction
1443 return DAG.getMachineNode(Desc->Opcode, SDLoc(Node), Node->getVTList(), Ops);
1446 /// \brief Helper function for adjustWritemask
1447 static unsigned SubIdx2Lane(unsigned Idx) {
1450 case AMDGPU::sub0: return 0;
1451 case AMDGPU::sub1: return 1;
1452 case AMDGPU::sub2: return 2;
1453 case AMDGPU::sub3: return 3;
1457 /// \brief Adjust the writemask of MIMG instructions
1458 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
1459 SelectionDAG &DAG) const {
1460 SDNode *Users[4] = { };
1462 unsigned OldDmask = Node->getConstantOperandVal(0);
1463 unsigned NewDmask = 0;
1465 // Try to figure out the used register components
1466 for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
1469 // Abort if we can't understand the usage
1470 if (!I->isMachineOpcode() ||
1471 I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
1474 // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used.
1475 // Note that subregs are packed, i.e. Lane==0 is the first bit set
1476 // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit
1478 Lane = SubIdx2Lane(I->getConstantOperandVal(1));
1480 // Set which texture component corresponds to the lane.
1482 for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) {
1484 Comp = countTrailingZeros(Dmask);
1485 Dmask &= ~(1 << Comp);
1488 // Abort if we have more than one user per component
1493 NewDmask |= 1 << Comp;
1496 // Abort if there's no change
1497 if (NewDmask == OldDmask)
1500 // Adjust the writemask in the node
1501 std::vector<SDValue> Ops;
1502 Ops.push_back(DAG.getTargetConstant(NewDmask, MVT::i32));
1503 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
1504 Ops.push_back(Node->getOperand(i));
1505 Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops);
1507 // If we only got one lane, replace it with a copy
1508 // (if NewDmask has only one bit set...)
1509 if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
1510 SDValue RC = DAG.getTargetConstant(AMDGPU::VReg_32RegClassID, MVT::i32);
1511 SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1512 SDLoc(), Users[Lane]->getValueType(0),
1513 SDValue(Node, 0), RC);
1514 DAG.ReplaceAllUsesWith(Users[Lane], Copy);
1518 // Update the users of the node with the new indices
1519 for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
1521 SDNode *User = Users[i];
1525 SDValue Op = DAG.getTargetConstant(Idx, MVT::i32);
1526 DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
1530 case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
1531 case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
1532 case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
1537 /// \brief Fold the instructions after selecting them.
1538 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
1539 SelectionDAG &DAG) const {
1540 const SIInstrInfo *TII =
1541 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1542 Node = AdjustRegClass(Node, DAG);
1544 if (TII->isMIMG(Node->getMachineOpcode()))
1545 adjustWritemask(Node, DAG);
1547 return foldOperands(Node, DAG);
1550 /// \brief Assign the register class depending on the number of
1551 /// bits set in the writemask
1552 void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
1553 SDNode *Node) const {
1554 const SIInstrInfo *TII =
1555 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1556 if (!TII->isMIMG(MI->getOpcode()))
1559 unsigned VReg = MI->getOperand(0).getReg();
1560 unsigned Writemask = MI->getOperand(1).getImm();
1561 unsigned BitsSet = 0;
1562 for (unsigned i = 0; i < 4; ++i)
1563 BitsSet += Writemask & (1 << i) ? 1 : 0;
1565 const TargetRegisterClass *RC;
1568 case 1: RC = &AMDGPU::VReg_32RegClass; break;
1569 case 2: RC = &AMDGPU::VReg_64RegClass; break;
1570 case 3: RC = &AMDGPU::VReg_96RegClass; break;
1573 unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
1574 MI->setDesc(TII->get(NewOpcode));
1575 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1576 MRI.setRegClass(VReg, RC);
1579 MachineSDNode *SITargetLowering::AdjustRegClass(MachineSDNode *N,
1580 SelectionDAG &DAG) const {
1583 unsigned NewOpcode = N->getMachineOpcode();
1585 switch (N->getMachineOpcode()) {
1587 case AMDGPU::S_LOAD_DWORD_IMM:
1588 NewOpcode = AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1590 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1591 if (NewOpcode == N->getMachineOpcode()) {
1592 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1595 case AMDGPU::S_LOAD_DWORDX4_IMM:
1596 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1597 if (NewOpcode == N->getMachineOpcode()) {
1598 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1600 if (fitsRegClass(DAG, N->getOperand(0), AMDGPU::SReg_64RegClassID)) {
1603 ConstantSDNode *Offset = cast<ConstantSDNode>(N->getOperand(1));
1605 SDValue(DAG.getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::i128,
1606 DAG.getConstant(0, MVT::i64)), 0),
1608 DAG.getConstant(Offset->getSExtValue() << 2, MVT::i32)
1610 return DAG.getMachineNode(NewOpcode, DL, N->getVTList(), Ops);
1615 SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1616 const TargetRegisterClass *RC,
1617 unsigned Reg, EVT VT) const {
1618 SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT);
1620 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()),
1621 cast<RegisterSDNode>(VReg)->getReg(), VT);