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::i64, Custom);
95 setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
96 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
97 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
99 setOperationAction(ISD::STORE, MVT::i1, Custom);
100 setOperationAction(ISD::STORE, MVT::i32, Custom);
101 setOperationAction(ISD::STORE, MVT::i64, Custom);
102 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
103 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
105 setOperationAction(ISD::SELECT, MVT::f32, Promote);
106 AddPromotedToType(ISD::SELECT, MVT::f32, MVT::i32);
107 setOperationAction(ISD::SELECT, MVT::i64, Custom);
108 setOperationAction(ISD::SELECT, MVT::f64, Promote);
109 AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
111 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
112 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
113 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
114 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
116 setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
117 setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
119 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
120 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
121 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
123 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
124 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
125 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
127 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
128 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
129 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
131 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Custom);
133 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
135 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
136 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
137 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom);
138 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
140 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
142 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
143 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Custom);
144 setLoadExtAction(ISD::SEXTLOAD, MVT::i16, Custom);
145 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Expand);
146 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, Expand);
147 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, Expand);
149 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
150 setLoadExtAction(ISD::ZEXTLOAD, MVT::i8, Custom);
151 setLoadExtAction(ISD::ZEXTLOAD, MVT::i16, Custom);
152 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Expand);
154 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
155 setLoadExtAction(ISD::EXTLOAD, MVT::i8, Custom);
156 setLoadExtAction(ISD::EXTLOAD, MVT::i16, Custom);
157 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Expand);
158 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
160 setTruncStoreAction(MVT::i32, MVT::i8, Custom);
161 setTruncStoreAction(MVT::i32, MVT::i16, Custom);
162 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
163 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
164 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Expand);
165 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Expand);
167 setOperationAction(ISD::LOAD, MVT::i1, Custom);
169 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
170 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
171 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
173 // These should use UDIVREM, so set them to expand
174 setOperationAction(ISD::UDIV, MVT::i64, Expand);
175 setOperationAction(ISD::UREM, MVT::i64, Expand);
177 // We only support LOAD/STORE and vector manipulation ops for vectors
178 // with > 4 elements.
180 MVT::v8i32, MVT::v8f32, MVT::v16i32, MVT::v16f32
183 for (MVT VT : VecTypes) {
184 for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
188 case ISD::BUILD_VECTOR:
190 case ISD::EXTRACT_VECTOR_ELT:
191 case ISD::INSERT_VECTOR_ELT:
192 case ISD::CONCAT_VECTORS:
193 case ISD::INSERT_SUBVECTOR:
194 case ISD::EXTRACT_SUBVECTOR:
197 setOperationAction(Op, VT, Expand);
203 for (int I = MVT::v1f64; I <= MVT::v8f64; ++I) {
204 MVT::SimpleValueType VT = static_cast<MVT::SimpleValueType>(I);
205 setOperationAction(ISD::FTRUNC, VT, Expand);
206 setOperationAction(ISD::FCEIL, VT, Expand);
207 setOperationAction(ISD::FFLOOR, VT, Expand);
210 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
211 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
212 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
213 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
214 setOperationAction(ISD::FRINT, MVT::f64, Legal);
217 // FIXME: This should be removed and handled the same was as f32 fneg. Source
218 // modifiers also work for the double instructions.
219 setOperationAction(ISD::FNEG, 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 bool SITargetLowering::shouldSplitVectorType(EVT VT) const {
274 return VT.getScalarType().bitsLE(MVT::i16);
277 bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
279 const SIInstrInfo *TII =
280 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
281 return TII->isInlineConstant(Imm);
284 SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
285 SDLoc DL, SDValue Chain,
286 unsigned Offset, bool Signed) const {
287 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
288 PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
289 AMDGPUAS::CONSTANT_ADDRESS);
290 SDValue BasePtr = DAG.getCopyFromReg(Chain, DL,
291 MRI.getLiveInVirtReg(AMDGPU::SGPR0_SGPR1), MVT::i64);
292 SDValue Ptr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
293 DAG.getConstant(Offset, MVT::i64));
294 return DAG.getExtLoad(Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD, DL, VT, Chain, Ptr,
295 MachinePointerInfo(UndefValue::get(PtrTy)), MemVT,
296 false, false, MemVT.getSizeInBits() >> 3);
300 SDValue SITargetLowering::LowerFormalArguments(
302 CallingConv::ID CallConv,
304 const SmallVectorImpl<ISD::InputArg> &Ins,
305 SDLoc DL, SelectionDAG &DAG,
306 SmallVectorImpl<SDValue> &InVals) const {
308 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
310 MachineFunction &MF = DAG.getMachineFunction();
311 FunctionType *FType = MF.getFunction()->getFunctionType();
312 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
314 assert(CallConv == CallingConv::C);
316 SmallVector<ISD::InputArg, 16> Splits;
317 uint32_t Skipped = 0;
319 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
320 const ISD::InputArg &Arg = Ins[i];
322 // First check if it's a PS input addr
323 if (Info->ShaderType == ShaderType::PIXEL && !Arg.Flags.isInReg() &&
324 !Arg.Flags.isByVal()) {
326 assert((PSInputNum <= 15) && "Too many PS inputs!");
329 // We can savely skip PS inputs
335 Info->PSInputAddr |= 1 << PSInputNum++;
338 // Second split vertices into their elements
339 if (Info->ShaderType != ShaderType::COMPUTE && Arg.VT.isVector()) {
340 ISD::InputArg NewArg = Arg;
341 NewArg.Flags.setSplit();
342 NewArg.VT = Arg.VT.getVectorElementType();
344 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
345 // three or five element vertex only needs three or five registers,
346 // NOT four or eigth.
347 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
348 unsigned NumElements = ParamType->getVectorNumElements();
350 for (unsigned j = 0; j != NumElements; ++j) {
351 Splits.push_back(NewArg);
352 NewArg.PartOffset += NewArg.VT.getStoreSize();
355 } else if (Info->ShaderType != ShaderType::COMPUTE) {
356 Splits.push_back(Arg);
360 SmallVector<CCValAssign, 16> ArgLocs;
361 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
362 getTargetMachine(), ArgLocs, *DAG.getContext());
364 // At least one interpolation mode must be enabled or else the GPU will hang.
365 if (Info->ShaderType == ShaderType::PIXEL && (Info->PSInputAddr & 0x7F) == 0) {
366 Info->PSInputAddr |= 1;
367 CCInfo.AllocateReg(AMDGPU::VGPR0);
368 CCInfo.AllocateReg(AMDGPU::VGPR1);
371 // The pointer to the list of arguments is stored in SGPR0, SGPR1
372 if (Info->ShaderType == ShaderType::COMPUTE) {
373 CCInfo.AllocateReg(AMDGPU::SGPR0);
374 CCInfo.AllocateReg(AMDGPU::SGPR1);
375 MF.addLiveIn(AMDGPU::SGPR0_SGPR1, &AMDGPU::SReg_64RegClass);
378 if (Info->ShaderType == ShaderType::COMPUTE) {
379 getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
383 AnalyzeFormalArguments(CCInfo, Splits);
385 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
387 const ISD::InputArg &Arg = Ins[i];
388 if (Skipped & (1 << i)) {
389 InVals.push_back(DAG.getUNDEF(Arg.VT));
393 CCValAssign &VA = ArgLocs[ArgIdx++];
394 EVT VT = VA.getLocVT();
398 EVT MemVT = Splits[i].VT;
399 // The first 36 bytes of the input buffer contains information about
400 // thread group and global sizes.
401 SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, DAG.getRoot(),
402 36 + VA.getLocMemOffset(),
403 Ins[i].Flags.isSExt());
404 InVals.push_back(Arg);
407 assert(VA.isRegLoc() && "Parameter must be in a register!");
409 unsigned Reg = VA.getLocReg();
411 if (VT == MVT::i64) {
412 // For now assume it is a pointer
413 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
414 &AMDGPU::SReg_64RegClass);
415 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
416 InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
420 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
422 Reg = MF.addLiveIn(Reg, RC);
423 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
425 if (Arg.VT.isVector()) {
427 // Build a vector from the registers
428 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
429 unsigned NumElements = ParamType->getVectorNumElements();
431 SmallVector<SDValue, 4> Regs;
433 for (unsigned j = 1; j != NumElements; ++j) {
434 Reg = ArgLocs[ArgIdx++].getLocReg();
435 Reg = MF.addLiveIn(Reg, RC);
436 Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
439 // Fill up the missing vector elements
440 NumElements = Arg.VT.getVectorNumElements() - NumElements;
441 for (unsigned j = 0; j != NumElements; ++j)
442 Regs.push_back(DAG.getUNDEF(VT));
444 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT, Regs));
448 InVals.push_back(Val);
453 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
454 MachineInstr * MI, MachineBasicBlock * BB) const {
456 MachineBasicBlock::iterator I = *MI;
457 const SIInstrInfo *TII =
458 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
459 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
461 switch (MI->getOpcode()) {
463 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
464 case AMDGPU::BRANCH: return BB;
465 case AMDGPU::SI_ADDR64_RSRC: {
466 unsigned SuperReg = MI->getOperand(0).getReg();
467 unsigned SubRegLo = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
468 unsigned SubRegHi = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
469 unsigned SubRegHiHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
470 unsigned SubRegHiLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
471 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B64), SubRegLo)
472 .addOperand(MI->getOperand(1));
473 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiLo)
475 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiHi)
476 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
477 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SubRegHi)
479 .addImm(AMDGPU::sub0)
481 .addImm(AMDGPU::sub1);
482 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SuperReg)
484 .addImm(AMDGPU::sub0_sub1)
486 .addImm(AMDGPU::sub2_sub3);
487 MI->eraseFromParent();
490 case AMDGPU::V_SUB_F64: {
491 unsigned DestReg = MI->getOperand(0).getReg();
492 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F64), DestReg)
493 .addImm(0) // SRC0 modifiers
494 .addReg(MI->getOperand(1).getReg())
495 .addImm(1) // SRC1 modifiers
496 .addReg(MI->getOperand(2).getReg())
497 .addImm(0) // SRC2 modifiers
501 MI->eraseFromParent();
504 case AMDGPU::SI_RegisterStorePseudo: {
505 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
506 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
507 MachineInstrBuilder MIB =
508 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::SI_RegisterStore),
510 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i)
511 MIB.addOperand(MI->getOperand(i));
513 MI->eraseFromParent();
516 case AMDGPU::FABS_SI: {
517 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
518 const SIInstrInfo *TII =
519 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
520 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
521 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
524 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_AND_B32_e32),
525 MI->getOperand(0).getReg())
526 .addReg(MI->getOperand(1).getReg())
528 MI->eraseFromParent();
531 case AMDGPU::FNEG_SI: {
532 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
533 const SIInstrInfo *TII =
534 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
535 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
536 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
539 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_XOR_B32_e32),
540 MI->getOperand(0).getReg())
541 .addReg(MI->getOperand(1).getReg())
543 MI->eraseFromParent();
546 case AMDGPU::FCLAMP_SI: {
547 const SIInstrInfo *TII =
548 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
549 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F32_e64),
550 MI->getOperand(0).getReg())
551 .addImm(0) // SRC0 modifiers
552 .addOperand(MI->getOperand(1))
553 .addImm(0) // SRC1 modifiers
557 MI->eraseFromParent();
563 EVT SITargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
564 if (!VT.isVector()) {
567 return MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
570 MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
574 bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
575 VT = VT.getScalarType();
580 switch (VT.getSimpleVT().SimpleTy) {
582 return false; /* There is V_MAD_F32 for f32 */
592 //===----------------------------------------------------------------------===//
593 // Custom DAG Lowering Operations
594 //===----------------------------------------------------------------------===//
596 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
597 MachineFunction &MF = DAG.getMachineFunction();
598 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
599 switch (Op.getOpcode()) {
600 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
601 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
603 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
604 if (Op.getValueType().isVector() &&
605 (Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
606 Load->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS ||
607 (Load->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
608 Op.getValueType().getVectorNumElements() > 4))) {
609 SDValue MergedValues[2] = {
610 SplitVectorLoad(Op, DAG),
613 return DAG.getMergeValues(MergedValues, SDLoc(Op));
615 return LowerLOAD(Op, DAG);
619 case ISD::SELECT: return LowerSELECT(Op, DAG);
620 case ISD::STORE: return LowerSTORE(Op, DAG);
621 case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
622 case ISD::INTRINSIC_WO_CHAIN: {
623 unsigned IntrinsicID =
624 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
625 EVT VT = Op.getValueType();
627 //XXX: Hardcoded we only use two to store the pointer to the parameters.
628 unsigned NumUserSGPRs = 2;
629 switch (IntrinsicID) {
630 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
631 case Intrinsic::r600_read_ngroups_x:
632 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 0, false);
633 case Intrinsic::r600_read_ngroups_y:
634 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 4, false);
635 case Intrinsic::r600_read_ngroups_z:
636 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 8, false);
637 case Intrinsic::r600_read_global_size_x:
638 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 12, false);
639 case Intrinsic::r600_read_global_size_y:
640 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 16, false);
641 case Intrinsic::r600_read_global_size_z:
642 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 20, false);
643 case Intrinsic::r600_read_local_size_x:
644 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 24, false);
645 case Intrinsic::r600_read_local_size_y:
646 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 28, false);
647 case Intrinsic::r600_read_local_size_z:
648 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 32, false);
649 case Intrinsic::r600_read_tgid_x:
650 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
651 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 0), VT);
652 case Intrinsic::r600_read_tgid_y:
653 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
654 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 1), VT);
655 case Intrinsic::r600_read_tgid_z:
656 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
657 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 2), VT);
658 case Intrinsic::r600_read_tidig_x:
659 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
661 case Intrinsic::r600_read_tidig_y:
662 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
664 case Intrinsic::r600_read_tidig_z:
665 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
667 case AMDGPUIntrinsic::SI_load_const: {
673 MachineMemOperand *MMO = MF.getMachineMemOperand(
674 MachinePointerInfo(),
675 MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
676 VT.getSizeInBits() / 8, 4);
677 return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
678 Op->getVTList(), Ops, VT, MMO);
680 case AMDGPUIntrinsic::SI_sample:
681 return LowerSampleIntrinsic(AMDGPUISD::SAMPLE, Op, DAG);
682 case AMDGPUIntrinsic::SI_sampleb:
683 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEB, Op, DAG);
684 case AMDGPUIntrinsic::SI_sampled:
685 return LowerSampleIntrinsic(AMDGPUISD::SAMPLED, Op, DAG);
686 case AMDGPUIntrinsic::SI_samplel:
687 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEL, Op, DAG);
688 case AMDGPUIntrinsic::SI_vs_load_input:
689 return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
696 case ISD::INTRINSIC_VOID:
697 SDValue Chain = Op.getOperand(0);
698 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
700 switch (IntrinsicID) {
701 case AMDGPUIntrinsic::SI_tbuffer_store: {
719 EVT VT = Op.getOperand(3).getValueType();
721 MachineMemOperand *MMO = MF.getMachineMemOperand(
722 MachinePointerInfo(),
723 MachineMemOperand::MOStore,
724 VT.getSizeInBits() / 8, 4);
725 return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL,
726 Op->getVTList(), Ops, VT, MMO);
735 /// \brief Helper function for LowerBRCOND
736 static SDNode *findUser(SDValue Value, unsigned Opcode) {
738 SDNode *Parent = Value.getNode();
739 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
742 if (I.getUse().get() != Value)
745 if (I->getOpcode() == Opcode)
751 /// This transforms the control flow intrinsics to get the branch destination as
752 /// last parameter, also switches branch target with BR if the need arise
753 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
754 SelectionDAG &DAG) const {
758 SDNode *Intr = BRCOND.getOperand(1).getNode();
759 SDValue Target = BRCOND.getOperand(2);
760 SDNode *BR = nullptr;
762 if (Intr->getOpcode() == ISD::SETCC) {
763 // As long as we negate the condition everything is fine
764 SDNode *SetCC = Intr;
765 assert(SetCC->getConstantOperandVal(1) == 1);
766 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
768 Intr = SetCC->getOperand(0).getNode();
771 // Get the target from BR if we don't negate the condition
772 BR = findUser(BRCOND, ISD::BR);
773 Target = BR->getOperand(1);
776 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
778 // Build the result and
779 SmallVector<EVT, 4> Res;
780 for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
781 Res.push_back(Intr->getValueType(i));
783 // operands of the new intrinsic call
784 SmallVector<SDValue, 4> Ops;
785 Ops.push_back(BRCOND.getOperand(0));
786 for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
787 Ops.push_back(Intr->getOperand(i));
788 Ops.push_back(Target);
790 // build the new intrinsic call
791 SDNode *Result = DAG.getNode(
792 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
793 DAG.getVTList(Res), Ops).getNode();
796 // Give the branch instruction our target
801 DAG.MorphNodeTo(BR, ISD::BR, BR->getVTList(), Ops);
804 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
806 // Copy the intrinsic results to registers
807 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
808 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
812 Chain = DAG.getCopyToReg(
814 CopyToReg->getOperand(1),
815 SDValue(Result, i - 1),
818 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
821 // Remove the old intrinsic from the chain
822 DAG.ReplaceAllUsesOfValueWith(
823 SDValue(Intr, Intr->getNumValues() - 1),
824 Intr->getOperand(0));
829 SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
831 LoadSDNode *Load = cast<LoadSDNode>(Op);
832 SDValue Ret = AMDGPUTargetLowering::LowerLOAD(Op, DAG);
833 SDValue MergedValues[2];
834 MergedValues[1] = Load->getChain();
836 MergedValues[0] = Ret;
837 return DAG.getMergeValues(MergedValues, DL);
840 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
844 EVT MemVT = Load->getMemoryVT();
846 assert(!MemVT.isVector() && "Private loads should be scalarized");
847 assert(!MemVT.isFloatingPoint() && "FP loads should be promoted to int");
849 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
850 DAG.getConstant(2, MVT::i32));
851 Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
852 Load->getChain(), Ptr,
853 DAG.getTargetConstant(0, MVT::i32),
855 if (MemVT.getSizeInBits() == 64) {
856 SDValue IncPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
857 DAG.getConstant(1, MVT::i32));
859 SDValue LoadUpper = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
860 Load->getChain(), IncPtr,
861 DAG.getTargetConstant(0, MVT::i32),
864 Ret = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ret, LoadUpper);
867 MergedValues[0] = Ret;
868 return DAG.getMergeValues(MergedValues, DL);
872 SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
874 SelectionDAG &DAG) const {
875 return DAG.getNode(Opcode, SDLoc(Op), Op.getValueType(), Op.getOperand(1),
881 SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
882 if (Op.getValueType() != MVT::i64)
886 SDValue Cond = Op.getOperand(0);
888 SDValue Zero = DAG.getConstant(0, MVT::i32);
889 SDValue One = DAG.getConstant(1, MVT::i32);
891 SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1));
892 SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2));
894 SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero);
895 SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero);
897 SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1);
899 SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One);
900 SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One);
902 SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1);
904 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i32, Lo, Hi);
905 return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
908 SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
910 StoreSDNode *Store = cast<StoreSDNode>(Op);
911 EVT VT = Store->getMemoryVT();
913 SDValue Ret = AMDGPUTargetLowering::LowerSTORE(Op, DAG);
917 if (VT.isVector() && VT.getVectorNumElements() >= 8)
918 return SplitVectorStore(Op, DAG);
921 return DAG.getTruncStore(Store->getChain(), DL,
922 DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32),
923 Store->getBasePtr(), MVT::i1, Store->getMemOperand());
925 if (Store->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
928 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Store->getBasePtr(),
929 DAG.getConstant(2, MVT::i32));
930 SDValue Chain = Store->getChain();
931 SmallVector<SDValue, 8> Values;
933 if (Store->isTruncatingStore()) {
935 if (Store->getMemoryVT() == MVT::i8) {
937 } else if (Store->getMemoryVT() == MVT::i16) {
940 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
941 Chain, Store->getBasePtr(),
942 DAG.getConstant(0, MVT::i32));
943 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, Store->getBasePtr(),
944 DAG.getConstant(0x3, MVT::i32));
945 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
946 DAG.getConstant(3, MVT::i32));
947 SDValue MaskedValue = DAG.getNode(ISD::AND, DL, MVT::i32, Store->getValue(),
948 DAG.getConstant(Mask, MVT::i32));
949 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
950 MaskedValue, ShiftAmt);
951 SDValue RotrAmt = DAG.getNode(ISD::SUB, DL, MVT::i32,
952 DAG.getConstant(32, MVT::i32), ShiftAmt);
953 SDValue DstMask = DAG.getNode(ISD::ROTR, DL, MVT::i32,
954 DAG.getConstant(Mask, MVT::i32),
956 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
957 Dst = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
959 Values.push_back(Dst);
960 } else if (VT == MVT::i64) {
961 for (unsigned i = 0; i < 2; ++i) {
962 Values.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
963 Store->getValue(), DAG.getConstant(i, MVT::i32)));
965 } else if (VT == MVT::i128) {
966 for (unsigned i = 0; i < 2; ++i) {
967 for (unsigned j = 0; j < 2; ++j) {
968 Values.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
969 DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
970 Store->getValue(), DAG.getConstant(i, MVT::i32)),
971 DAG.getConstant(j, MVT::i32)));
975 Values.push_back(Store->getValue());
978 for (unsigned i = 0; i < Values.size(); ++i) {
979 SDValue PartPtr = DAG.getNode(ISD::ADD, DL, MVT::i32,
980 Ptr, DAG.getConstant(i, MVT::i32));
981 Chain = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
982 Chain, Values[i], PartPtr,
983 DAG.getTargetConstant(0, MVT::i32));
988 //===----------------------------------------------------------------------===//
989 // Custom DAG optimizations
990 //===----------------------------------------------------------------------===//
992 SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
993 DAGCombinerInfo &DCI) {
994 EVT VT = N->getValueType(0);
995 EVT ScalarVT = VT.getScalarType();
996 if (ScalarVT != MVT::f32)
999 SelectionDAG &DAG = DCI.DAG;
1002 SDValue Src = N->getOperand(0);
1003 EVT SrcVT = Src.getValueType();
1005 // TODO: We could try to match extracting the higher bytes, which would be
1006 // easier if i8 vectors weren't promoted to i32 vectors, particularly after
1007 // types are legalized. v4i8 -> v4f32 is probably the only case to worry
1008 // about in practice.
1009 if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) {
1010 if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) {
1011 SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src);
1012 DCI.AddToWorklist(Cvt.getNode());
1017 // We are primarily trying to catch operations on illegal vector types
1018 // before they are expanded.
1019 // For scalars, we can use the more flexible method of checking masked bits
1020 // after legalization.
1021 if (!DCI.isBeforeLegalize() ||
1022 !SrcVT.isVector() ||
1023 SrcVT.getVectorElementType() != MVT::i8) {
1027 assert(DCI.isBeforeLegalize() && "Unexpected legal type");
1029 // Weird sized vectors are a pain to handle, but we know 3 is really the same
1031 unsigned NElts = SrcVT.getVectorNumElements();
1032 if (!SrcVT.isSimple() && NElts != 3)
1035 // Handle v4i8 -> v4f32 extload. Replace the v4i8 with a legal i32 load to
1036 // prevent a mess from expanding to v4i32 and repacking.
1037 if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse()) {
1038 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), SrcVT);
1039 EVT RegVT = getEquivalentLoadRegType(*DAG.getContext(), SrcVT);
1040 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f32, NElts);
1042 LoadSDNode *Load = cast<LoadSDNode>(Src);
1043 SDValue NewLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegVT,
1047 Load->getMemOperand());
1049 // Make sure successors of the original load stay after it by updating
1050 // them to use the new Chain.
1051 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), NewLoad.getValue(1));
1053 SmallVector<SDValue, 4> Elts;
1054 if (RegVT.isVector())
1055 DAG.ExtractVectorElements(NewLoad, Elts);
1057 Elts.push_back(NewLoad);
1059 SmallVector<SDValue, 4> Ops;
1061 unsigned EltIdx = 0;
1062 for (SDValue Elt : Elts) {
1063 unsigned ComponentsInElt = std::min(4u, NElts - 4 * EltIdx);
1064 for (unsigned I = 0; I < ComponentsInElt; ++I) {
1065 unsigned Opc = AMDGPUISD::CVT_F32_UBYTE0 + I;
1066 SDValue Cvt = DAG.getNode(Opc, DL, MVT::f32, Elt);
1067 DCI.AddToWorklist(Cvt.getNode());
1074 assert(Ops.size() == NElts);
1076 return DAG.getNode(ISD::BUILD_VECTOR, DL, FloatVT, Ops);
1082 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
1083 DAGCombinerInfo &DCI) const {
1084 SelectionDAG &DAG = DCI.DAG;
1086 EVT VT = N->getValueType(0);
1088 switch (N->getOpcode()) {
1089 default: return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1090 case ISD::SELECT_CC: {
1091 ConstantSDNode *True, *False;
1092 // i1 selectcc(l, r, -1, 0, cc) -> i1 setcc(l, r, cc)
1093 if ((True = dyn_cast<ConstantSDNode>(N->getOperand(2)))
1094 && (False = dyn_cast<ConstantSDNode>(N->getOperand(3)))
1095 && True->isAllOnesValue()
1096 && False->isNullValue()
1098 return DAG.getNode(ISD::SETCC, DL, VT, N->getOperand(0),
1099 N->getOperand(1), N->getOperand(4));
1105 SDValue Arg0 = N->getOperand(0);
1106 SDValue Arg1 = N->getOperand(1);
1107 SDValue CC = N->getOperand(2);
1108 ConstantSDNode * C = nullptr;
1109 ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
1111 // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
1113 && Arg0.getOpcode() == ISD::SIGN_EXTEND
1114 && Arg0.getOperand(0).getValueType() == MVT::i1
1115 && (C = dyn_cast<ConstantSDNode>(Arg1))
1117 && CCOp == ISD::SETNE) {
1118 return SimplifySetCC(VT, Arg0.getOperand(0),
1119 DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
1124 case AMDGPUISD::CVT_F32_UBYTE0:
1125 case AMDGPUISD::CVT_F32_UBYTE1:
1126 case AMDGPUISD::CVT_F32_UBYTE2:
1127 case AMDGPUISD::CVT_F32_UBYTE3: {
1128 unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
1130 SDValue Src = N->getOperand(0);
1131 APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);
1133 APInt KnownZero, KnownOne;
1134 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1135 !DCI.isBeforeLegalizeOps());
1136 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1137 if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
1138 TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {
1139 DCI.CommitTargetLoweringOpt(TLO);
1145 case ISD::UINT_TO_FP: {
1146 return performUCharToFloatCombine(N, DCI);
1150 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1153 /// \brief Test if RegClass is one of the VSrc classes
1154 static bool isVSrc(unsigned RegClass) {
1155 return AMDGPU::VSrc_32RegClassID == RegClass ||
1156 AMDGPU::VSrc_64RegClassID == RegClass;
1159 /// \brief Test if RegClass is one of the SSrc classes
1160 static bool isSSrc(unsigned RegClass) {
1161 return AMDGPU::SSrc_32RegClassID == RegClass ||
1162 AMDGPU::SSrc_64RegClassID == RegClass;
1165 /// \brief Analyze the possible immediate value Op
1167 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
1168 /// and the immediate value if it's a literal immediate
1169 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
1176 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
1177 if (Node->getZExtValue() >> 32) {
1180 Imm.I = Node->getSExtValue();
1181 } else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
1182 if (N->getValueType(0) != MVT::f32)
1184 Imm.F = Node->getValueAPF().convertToFloat();
1186 return -1; // It isn't an immediate
1188 if ((Imm.I >= -16 && Imm.I <= 64) ||
1189 Imm.F == 0.5f || Imm.F == -0.5f ||
1190 Imm.F == 1.0f || Imm.F == -1.0f ||
1191 Imm.F == 2.0f || Imm.F == -2.0f ||
1192 Imm.F == 4.0f || Imm.F == -4.0f)
1193 return 0; // It's an inline immediate
1195 return Imm.I; // It's a literal immediate
1198 /// \brief Try to fold an immediate directly into an instruction
1199 bool SITargetLowering::foldImm(SDValue &Operand, int32_t &Immediate,
1200 bool &ScalarSlotUsed) const {
1202 MachineSDNode *Mov = dyn_cast<MachineSDNode>(Operand);
1203 const SIInstrInfo *TII =
1204 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1205 if (!Mov || !TII->isMov(Mov->getMachineOpcode()))
1208 const SDValue &Op = Mov->getOperand(0);
1209 int32_t Value = analyzeImmediate(Op.getNode());
1211 // Not an immediate at all
1214 } else if (Value == 0) {
1215 // Inline immediates can always be fold
1219 } else if (Value == Immediate) {
1220 // Already fold literal immediate
1224 } else if (!ScalarSlotUsed && !Immediate) {
1225 // Fold this literal immediate
1226 ScalarSlotUsed = true;
1236 const TargetRegisterClass *SITargetLowering::getRegClassForNode(
1237 SelectionDAG &DAG, const SDValue &Op) const {
1238 const SIInstrInfo *TII =
1239 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1240 const SIRegisterInfo &TRI = TII->getRegisterInfo();
1242 if (!Op->isMachineOpcode()) {
1243 switch(Op->getOpcode()) {
1244 case ISD::CopyFromReg: {
1245 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1246 unsigned Reg = cast<RegisterSDNode>(Op->getOperand(1))->getReg();
1247 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
1248 return MRI.getRegClass(Reg);
1250 return TRI.getPhysRegClass(Reg);
1252 default: return nullptr;
1255 const MCInstrDesc &Desc = TII->get(Op->getMachineOpcode());
1256 int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
1257 if (OpClassID != -1) {
1258 return TRI.getRegClass(OpClassID);
1260 switch(Op.getMachineOpcode()) {
1261 case AMDGPU::COPY_TO_REGCLASS:
1262 // Operand 1 is the register class id for COPY_TO_REGCLASS instructions.
1263 OpClassID = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1265 // If the COPY_TO_REGCLASS instruction is copying to a VSrc register
1266 // class, then the register class for the value could be either a
1267 // VReg or and SReg. In order to get a more accurate
1268 if (OpClassID == AMDGPU::VSrc_32RegClassID ||
1269 OpClassID == AMDGPU::VSrc_64RegClassID) {
1270 return getRegClassForNode(DAG, Op.getOperand(0));
1272 return TRI.getRegClass(OpClassID);
1273 case AMDGPU::EXTRACT_SUBREG: {
1274 int SubIdx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1275 const TargetRegisterClass *SuperClass =
1276 getRegClassForNode(DAG, Op.getOperand(0));
1277 return TRI.getSubClassWithSubReg(SuperClass, SubIdx);
1279 case AMDGPU::REG_SEQUENCE:
1280 // Operand 0 is the register class id for REG_SEQUENCE instructions.
1281 return TRI.getRegClass(
1282 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue());
1284 return getRegClassFor(Op.getSimpleValueType());
1288 /// \brief Does "Op" fit into register class "RegClass" ?
1289 bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, const SDValue &Op,
1290 unsigned RegClass) const {
1291 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1292 const TargetRegisterClass *RC = getRegClassForNode(DAG, Op);
1296 return TRI->getRegClass(RegClass)->hasSubClassEq(RC);
1299 /// \brief Make sure that we don't exeed the number of allowed scalars
1300 void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
1302 bool &ScalarSlotUsed) const {
1304 // First map the operands register class to a destination class
1305 if (RegClass == AMDGPU::VSrc_32RegClassID)
1306 RegClass = AMDGPU::VReg_32RegClassID;
1307 else if (RegClass == AMDGPU::VSrc_64RegClassID)
1308 RegClass = AMDGPU::VReg_64RegClassID;
1312 // Nothing to do if they fit naturally
1313 if (fitsRegClass(DAG, Operand, RegClass))
1316 // If the scalar slot isn't used yet use it now
1317 if (!ScalarSlotUsed) {
1318 ScalarSlotUsed = true;
1322 // This is a conservative aproach. It is possible that we can't determine the
1323 // correct register class and copy too often, but better safe than sorry.
1324 SDValue RC = DAG.getTargetConstant(RegClass, MVT::i32);
1325 SDNode *Node = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, SDLoc(),
1326 Operand.getValueType(), Operand, RC);
1327 Operand = SDValue(Node, 0);
1330 /// \returns true if \p Node's operands are different from the SDValue list
1332 static bool isNodeChanged(const SDNode *Node, const std::vector<SDValue> &Ops) {
1333 for (unsigned i = 0, e = Node->getNumOperands(); i < e; ++i) {
1334 if (Ops[i].getNode() != Node->getOperand(i).getNode()) {
1341 /// \brief Try to fold the Nodes operands into the Node
1342 SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
1343 SelectionDAG &DAG) const {
1345 // Original encoding (either e32 or e64)
1346 int Opcode = Node->getMachineOpcode();
1347 const SIInstrInfo *TII =
1348 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1349 const MCInstrDesc *Desc = &TII->get(Opcode);
1351 unsigned NumDefs = Desc->getNumDefs();
1352 unsigned NumOps = Desc->getNumOperands();
1354 // Commuted opcode if available
1355 int OpcodeRev = Desc->isCommutable() ? TII->commuteOpcode(Opcode) : -1;
1356 const MCInstrDesc *DescRev = OpcodeRev == -1 ? nullptr : &TII->get(OpcodeRev);
1358 assert(!DescRev || DescRev->getNumDefs() == NumDefs);
1359 assert(!DescRev || DescRev->getNumOperands() == NumOps);
1361 // e64 version if available, -1 otherwise
1362 int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
1363 const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? nullptr : &TII->get(OpcodeE64);
1364 int InputModifiers[3] = {0};
1366 assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
1368 int32_t Immediate = Desc->getSize() == 4 ? 0 : -1;
1369 bool HaveVSrc = false, HaveSSrc = false;
1371 // First figure out what we already have in this instruction.
1372 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1373 i != e && Op < NumOps; ++i, ++Op) {
1375 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1376 if (isVSrc(RegClass))
1378 else if (isSSrc(RegClass))
1383 int32_t Imm = analyzeImmediate(Node->getOperand(i).getNode());
1384 if (Imm != -1 && Imm != 0) {
1385 // Literal immediate
1390 // If we neither have VSrc nor SSrc, it makes no sense to continue.
1391 if (!HaveVSrc && !HaveSSrc)
1394 // No scalar allowed when we have both VSrc and SSrc
1395 bool ScalarSlotUsed = HaveVSrc && HaveSSrc;
1397 // Second go over the operands and try to fold them
1398 std::vector<SDValue> Ops;
1399 bool Promote2e64 = false;
1400 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1401 i != e && Op < NumOps; ++i, ++Op) {
1403 const SDValue &Operand = Node->getOperand(i);
1404 Ops.push_back(Operand);
1406 // Already folded immediate?
1407 if (isa<ConstantSDNode>(Operand.getNode()) ||
1408 isa<ConstantFPSDNode>(Operand.getNode()))
1411 // Is this a VSrc or SSrc operand?
1412 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1413 if (isVSrc(RegClass) || isSSrc(RegClass)) {
1414 // Try to fold the immediates
1415 if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
1416 // Folding didn't work, make sure we don't hit the SReg limit.
1417 ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
1422 if (i == 1 && DescRev && fitsRegClass(DAG, Ops[0], RegClass)) {
1424 unsigned OtherRegClass = Desc->OpInfo[NumDefs].RegClass;
1425 assert(isVSrc(OtherRegClass) || isSSrc(OtherRegClass));
1427 // Test if it makes sense to swap operands
1428 if (foldImm(Ops[1], Immediate, ScalarSlotUsed) ||
1429 (!fitsRegClass(DAG, Ops[1], RegClass) &&
1430 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1432 // Swap commutable operands
1433 std::swap(Ops[0], Ops[1]);
1445 // Test if it makes sense to switch to e64 encoding
1446 unsigned OtherRegClass = DescE64->OpInfo[Op].RegClass;
1447 if (!isVSrc(OtherRegClass) && !isSSrc(OtherRegClass))
1450 int32_t TmpImm = -1;
1451 if (foldImm(Ops[i], TmpImm, ScalarSlotUsed) ||
1452 (!fitsRegClass(DAG, Ops[i], RegClass) &&
1453 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1455 // Switch to e64 encoding
1463 if (!DescE64 && !Promote2e64)
1465 if (!Operand.isMachineOpcode())
1467 if (Operand.getMachineOpcode() == AMDGPU::FNEG_SI) {
1469 Ops.push_back(Operand.getOperand(0));
1470 InputModifiers[i] = 1;
1477 else if (Operand.getMachineOpcode() == AMDGPU::FABS_SI) {
1479 Ops.push_back(Operand.getOperand(0));
1480 InputModifiers[i] = 2;
1490 std::vector<SDValue> OldOps(Ops);
1492 for (unsigned i = 0; i < OldOps.size(); ++i) {
1494 Ops.push_back(DAG.getTargetConstant(InputModifiers[i], MVT::i32));
1495 Ops.push_back(OldOps[i]);
1497 // Add the modifier flags while promoting
1498 for (unsigned i = 0; i < 2; ++i)
1499 Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
1502 // Add optional chain and glue
1503 for (unsigned i = NumOps - NumDefs, e = Node->getNumOperands(); i < e; ++i)
1504 Ops.push_back(Node->getOperand(i));
1506 // Nodes that have a glue result are not CSE'd by getMachineNode(), so in
1507 // this case a brand new node is always be created, even if the operands
1508 // are the same as before. So, manually check if anything has been changed.
1509 if (Desc->Opcode == Opcode && !isNodeChanged(Node, Ops)) {
1513 // Create a complete new instruction
1514 return DAG.getMachineNode(Desc->Opcode, SDLoc(Node), Node->getVTList(), Ops);
1517 /// \brief Helper function for adjustWritemask
1518 static unsigned SubIdx2Lane(unsigned Idx) {
1521 case AMDGPU::sub0: return 0;
1522 case AMDGPU::sub1: return 1;
1523 case AMDGPU::sub2: return 2;
1524 case AMDGPU::sub3: return 3;
1528 /// \brief Adjust the writemask of MIMG instructions
1529 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
1530 SelectionDAG &DAG) const {
1531 SDNode *Users[4] = { };
1533 unsigned OldDmask = Node->getConstantOperandVal(0);
1534 unsigned NewDmask = 0;
1536 // Try to figure out the used register components
1537 for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
1540 // Abort if we can't understand the usage
1541 if (!I->isMachineOpcode() ||
1542 I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
1545 // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used.
1546 // Note that subregs are packed, i.e. Lane==0 is the first bit set
1547 // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit
1549 Lane = SubIdx2Lane(I->getConstantOperandVal(1));
1551 // Set which texture component corresponds to the lane.
1553 for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) {
1555 Comp = countTrailingZeros(Dmask);
1556 Dmask &= ~(1 << Comp);
1559 // Abort if we have more than one user per component
1564 NewDmask |= 1 << Comp;
1567 // Abort if there's no change
1568 if (NewDmask == OldDmask)
1571 // Adjust the writemask in the node
1572 std::vector<SDValue> Ops;
1573 Ops.push_back(DAG.getTargetConstant(NewDmask, MVT::i32));
1574 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
1575 Ops.push_back(Node->getOperand(i));
1576 Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops);
1578 // If we only got one lane, replace it with a copy
1579 // (if NewDmask has only one bit set...)
1580 if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
1581 SDValue RC = DAG.getTargetConstant(AMDGPU::VReg_32RegClassID, MVT::i32);
1582 SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1583 SDLoc(), Users[Lane]->getValueType(0),
1584 SDValue(Node, 0), RC);
1585 DAG.ReplaceAllUsesWith(Users[Lane], Copy);
1589 // Update the users of the node with the new indices
1590 for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
1592 SDNode *User = Users[i];
1596 SDValue Op = DAG.getTargetConstant(Idx, MVT::i32);
1597 DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
1601 case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
1602 case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
1603 case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
1608 /// \brief Fold the instructions after selecting them.
1609 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
1610 SelectionDAG &DAG) const {
1611 const SIInstrInfo *TII =
1612 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1613 Node = AdjustRegClass(Node, DAG);
1615 if (TII->isMIMG(Node->getMachineOpcode()))
1616 adjustWritemask(Node, DAG);
1618 return foldOperands(Node, DAG);
1621 /// \brief Assign the register class depending on the number of
1622 /// bits set in the writemask
1623 void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
1624 SDNode *Node) const {
1625 const SIInstrInfo *TII =
1626 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1627 if (!TII->isMIMG(MI->getOpcode()))
1630 unsigned VReg = MI->getOperand(0).getReg();
1631 unsigned Writemask = MI->getOperand(1).getImm();
1632 unsigned BitsSet = 0;
1633 for (unsigned i = 0; i < 4; ++i)
1634 BitsSet += Writemask & (1 << i) ? 1 : 0;
1636 const TargetRegisterClass *RC;
1639 case 1: RC = &AMDGPU::VReg_32RegClass; break;
1640 case 2: RC = &AMDGPU::VReg_64RegClass; break;
1641 case 3: RC = &AMDGPU::VReg_96RegClass; break;
1644 unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
1645 MI->setDesc(TII->get(NewOpcode));
1646 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1647 MRI.setRegClass(VReg, RC);
1650 MachineSDNode *SITargetLowering::AdjustRegClass(MachineSDNode *N,
1651 SelectionDAG &DAG) const {
1654 unsigned NewOpcode = N->getMachineOpcode();
1656 switch (N->getMachineOpcode()) {
1658 case AMDGPU::S_LOAD_DWORD_IMM:
1659 NewOpcode = AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1661 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1662 if (NewOpcode == N->getMachineOpcode()) {
1663 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1666 case AMDGPU::S_LOAD_DWORDX4_IMM:
1667 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1668 if (NewOpcode == N->getMachineOpcode()) {
1669 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1671 if (fitsRegClass(DAG, N->getOperand(0), AMDGPU::SReg_64RegClassID)) {
1674 ConstantSDNode *Offset = cast<ConstantSDNode>(N->getOperand(1));
1676 SDValue(DAG.getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::i128,
1677 DAG.getConstant(0, MVT::i64)), 0),
1679 DAG.getConstant(Offset->getSExtValue() << 2, MVT::i32)
1681 return DAG.getMachineNode(NewOpcode, DL, N->getVTList(), Ops);
1686 SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1687 const TargetRegisterClass *RC,
1688 unsigned Reg, EVT VT) const {
1689 SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT);
1691 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()),
1692 cast<RegisterSDNode>(VReg)->getReg(), VT);