1 //===-- SIISelLowering.cpp - SI DAG Lowering Implementation ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief Custom DAG lowering for SI
13 //===----------------------------------------------------------------------===//
17 #define _USE_MATH_DEFINES
21 #include "SIISelLowering.h"
23 #include "AMDGPUIntrinsicInfo.h"
24 #include "AMDGPUSubtarget.h"
25 #include "SIInstrInfo.h"
26 #include "SIMachineFunctionInfo.h"
27 #include "SIRegisterInfo.h"
28 #include "llvm/ADT/BitVector.h"
29 #include "llvm/CodeGen/CallingConvLower.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/CodeGen/SelectionDAG.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/ADT/SmallString.h"
38 SITargetLowering::SITargetLowering(TargetMachine &TM) :
39 AMDGPUTargetLowering(TM) {
40 addRegisterClass(MVT::i1, &AMDGPU::VReg_1RegClass);
41 addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
43 addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
44 addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);
46 addRegisterClass(MVT::i32, &AMDGPU::SReg_32RegClass);
47 addRegisterClass(MVT::f32, &AMDGPU::VGPR_32RegClass);
49 addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
50 addRegisterClass(MVT::v2i32, &AMDGPU::SReg_64RegClass);
51 addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
53 addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
54 addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
56 addRegisterClass(MVT::v8i32, &AMDGPU::SReg_256RegClass);
57 addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
59 addRegisterClass(MVT::v16i32, &AMDGPU::SReg_512RegClass);
60 addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);
62 computeRegisterProperties();
64 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
65 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
66 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
67 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);
69 setOperationAction(ISD::ADD, MVT::i32, Legal);
70 setOperationAction(ISD::ADDC, MVT::i32, Legal);
71 setOperationAction(ISD::ADDE, MVT::i32, Legal);
72 setOperationAction(ISD::SUBC, MVT::i32, Legal);
73 setOperationAction(ISD::SUBE, MVT::i32, Legal);
75 setOperationAction(ISD::FSIN, MVT::f32, Custom);
76 setOperationAction(ISD::FCOS, MVT::f32, Custom);
78 setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
79 setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
80 setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
81 setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
83 // We need to custom lower vector stores from local memory
84 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
85 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
86 setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
88 setOperationAction(ISD::STORE, MVT::v8i32, Custom);
89 setOperationAction(ISD::STORE, MVT::v16i32, Custom);
91 setOperationAction(ISD::STORE, MVT::i1, Custom);
92 setOperationAction(ISD::STORE, MVT::i32, Custom);
93 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
94 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
96 setOperationAction(ISD::SELECT, MVT::i64, Custom);
97 setOperationAction(ISD::SELECT, MVT::f64, Promote);
98 AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
100 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
101 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
102 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
103 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
105 setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
106 setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
108 setOperationAction(ISD::BSWAP, MVT::i32, Legal);
110 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
111 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
112 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
114 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
115 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
116 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
118 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
119 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
120 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
122 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
123 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
125 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
126 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
127 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom);
128 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
130 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
131 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
133 for (MVT VT : MVT::integer_valuetypes()) {
137 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
138 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Legal);
139 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Legal);
140 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand);
142 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
143 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i8, Legal);
144 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i16, Legal);
145 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand);
147 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
148 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i8, Legal);
149 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i16, Legal);
150 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand);
153 for (MVT VT : MVT::integer_vector_valuetypes()) {
154 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i16, Expand);
155 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v16i16, Expand);
158 for (MVT VT : MVT::fp_valuetypes())
159 setLoadExtAction(ISD::EXTLOAD, VT, 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 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
185 setOperationAction(ISD::SELECT, MVT::i1, Promote);
187 for (MVT VT : VecTypes) {
188 for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
192 case ISD::BUILD_VECTOR:
194 case ISD::EXTRACT_VECTOR_ELT:
195 case ISD::INSERT_VECTOR_ELT:
196 case ISD::INSERT_SUBVECTOR:
197 case ISD::EXTRACT_SUBVECTOR:
199 case ISD::CONCAT_VECTORS:
200 setOperationAction(Op, VT, Custom);
203 setOperationAction(Op, VT, Expand);
209 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
210 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
211 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
212 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
213 setOperationAction(ISD::FRINT, MVT::f64, Legal);
216 setOperationAction(ISD::FDIV, MVT::f32, Custom);
218 setTargetDAGCombine(ISD::FADD);
219 setTargetDAGCombine(ISD::FSUB);
220 setTargetDAGCombine(ISD::FMINNUM);
221 setTargetDAGCombine(ISD::FMAXNUM);
222 setTargetDAGCombine(ISD::SELECT_CC);
223 setTargetDAGCombine(ISD::SETCC);
224 setTargetDAGCombine(ISD::AND);
225 setTargetDAGCombine(ISD::OR);
226 setTargetDAGCombine(ISD::UINT_TO_FP);
228 // All memory operations. Some folding on the pointer operand is done to help
229 // matching the constant offsets in the addressing modes.
230 setTargetDAGCombine(ISD::LOAD);
231 setTargetDAGCombine(ISD::STORE);
232 setTargetDAGCombine(ISD::ATOMIC_LOAD);
233 setTargetDAGCombine(ISD::ATOMIC_STORE);
234 setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP);
235 setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS);
236 setTargetDAGCombine(ISD::ATOMIC_SWAP);
237 setTargetDAGCombine(ISD::ATOMIC_LOAD_ADD);
238 setTargetDAGCombine(ISD::ATOMIC_LOAD_SUB);
239 setTargetDAGCombine(ISD::ATOMIC_LOAD_AND);
240 setTargetDAGCombine(ISD::ATOMIC_LOAD_OR);
241 setTargetDAGCombine(ISD::ATOMIC_LOAD_XOR);
242 setTargetDAGCombine(ISD::ATOMIC_LOAD_NAND);
243 setTargetDAGCombine(ISD::ATOMIC_LOAD_MIN);
244 setTargetDAGCombine(ISD::ATOMIC_LOAD_MAX);
245 setTargetDAGCombine(ISD::ATOMIC_LOAD_UMIN);
246 setTargetDAGCombine(ISD::ATOMIC_LOAD_UMAX);
248 setSchedulingPreference(Sched::RegPressure);
251 //===----------------------------------------------------------------------===//
252 // TargetLowering queries
253 //===----------------------------------------------------------------------===//
255 bool SITargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &,
257 // SI has some legal vector types, but no legal vector operations. Say no
258 // shuffles are legal in order to prefer scalarizing some vector operations.
262 // FIXME: This really needs an address space argument. The immediate offset
263 // size is different for different sets of memory instruction sets.
265 // The single offset DS instructions have a 16-bit unsigned byte offset.
267 // MUBUF / MTBUF have a 12-bit unsigned byte offset, and additionally can do r +
268 // r + i with addr64. 32-bit has more addressing mode options. Depending on the
269 // resource constant, it can also do (i64 r0) + (i32 r1) * (i14 i).
271 // SMRD instructions have an 8-bit, dword offset.
273 bool SITargetLowering::isLegalAddressingMode(const AddrMode &AM,
275 // No global is ever allowed as a base.
279 // Allow a 16-bit unsigned immediate field, since this is what DS instructions
281 if (!isUInt<16>(AM.BaseOffs))
286 case 0: // "r+i" or just "i", depending on HasBaseReg.
289 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
291 // Otherwise we have r+r or r+i.
294 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
298 default: // Don't allow n * r
305 bool SITargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
308 bool *IsFast) const {
312 // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
313 // which isn't a simple VT.
314 if (!VT.isSimple() || VT == MVT::Other)
317 // XXX - CI changes say "Support for unaligned memory accesses" but I don't
318 // see what for specifically. The wording everywhere else seems to be the
321 // XXX - The only mention I see of this in the ISA manual is for LDS direct
322 // reads the "byte address and must be dword aligned". Is it also true for the
323 // normal loads and stores?
324 if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS) {
325 // ds_read/write_b64 require 8-byte alignment, but we can do a 4 byte
326 // aligned, 8 byte access in a single operation using ds_read2/write2_b32
327 // with adjacent offsets.
328 return Align % 4 == 0;
331 // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
332 // byte-address are ignored, thus forcing Dword alignment.
333 // This applies to private, global, and constant memory.
336 return VT.bitsGT(MVT::i32);
339 EVT SITargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
340 unsigned SrcAlign, bool IsMemset,
343 MachineFunction &MF) const {
344 // FIXME: Should account for address space here.
346 // The default fallback uses the private pointer size as a guess for a type to
347 // use. Make sure we switch these to 64-bit accesses.
349 if (Size >= 16 && DstAlign >= 4) // XXX: Should only do for global
352 if (Size >= 8 && DstAlign >= 4)
359 TargetLoweringBase::LegalizeTypeAction
360 SITargetLowering::getPreferredVectorAction(EVT VT) const {
361 if (VT.getVectorNumElements() != 1 && VT.getScalarType().bitsLE(MVT::i16))
362 return TypeSplitVector;
364 return TargetLoweringBase::getPreferredVectorAction(VT);
367 bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
369 const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
370 getTargetMachine().getSubtargetImpl()->getInstrInfo());
371 return TII->isInlineConstant(Imm);
374 SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
375 SDLoc SL, SDValue Chain,
376 unsigned Offset, bool Signed) const {
377 const DataLayout *DL = getDataLayout();
378 MachineFunction &MF = DAG.getMachineFunction();
379 const SIRegisterInfo *TRI =
380 static_cast<const SIRegisterInfo*>(Subtarget->getRegisterInfo());
381 unsigned InputPtrReg = TRI->getPreloadedValue(MF, SIRegisterInfo::INPUT_PTR);
383 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
385 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
386 PointerType *PtrTy = PointerType::get(Ty, AMDGPUAS::CONSTANT_ADDRESS);
387 SDValue BasePtr = DAG.getCopyFromReg(Chain, SL,
388 MRI.getLiveInVirtReg(InputPtrReg), MVT::i64);
389 SDValue Ptr = DAG.getNode(ISD::ADD, SL, MVT::i64, BasePtr,
390 DAG.getConstant(Offset, MVT::i64));
391 SDValue PtrOffset = DAG.getUNDEF(getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
392 MachinePointerInfo PtrInfo(UndefValue::get(PtrTy));
394 return DAG.getLoad(ISD::UNINDEXED, Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD,
395 VT, SL, Chain, Ptr, PtrOffset, PtrInfo, MemVT,
397 true, // isNonTemporal
399 DL->getABITypeAlignment(Ty)); // Alignment
402 SDValue SITargetLowering::LowerFormalArguments(
404 CallingConv::ID CallConv,
406 const SmallVectorImpl<ISD::InputArg> &Ins,
407 SDLoc DL, SelectionDAG &DAG,
408 SmallVectorImpl<SDValue> &InVals) const {
410 const TargetMachine &TM = getTargetMachine();
411 const SIRegisterInfo *TRI =
412 static_cast<const SIRegisterInfo*>(TM.getSubtargetImpl()->getRegisterInfo());
414 MachineFunction &MF = DAG.getMachineFunction();
415 FunctionType *FType = MF.getFunction()->getFunctionType();
416 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
418 assert(CallConv == CallingConv::C);
420 SmallVector<ISD::InputArg, 16> Splits;
421 BitVector Skipped(Ins.size());
423 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
424 const ISD::InputArg &Arg = Ins[i];
426 // First check if it's a PS input addr
427 if (Info->getShaderType() == ShaderType::PIXEL && !Arg.Flags.isInReg() &&
428 !Arg.Flags.isByVal()) {
430 assert((PSInputNum <= 15) && "Too many PS inputs!");
433 // We can savely skip PS inputs
439 Info->PSInputAddr |= 1 << PSInputNum++;
442 // Second split vertices into their elements
443 if (Info->getShaderType() != ShaderType::COMPUTE && Arg.VT.isVector()) {
444 ISD::InputArg NewArg = Arg;
445 NewArg.Flags.setSplit();
446 NewArg.VT = Arg.VT.getVectorElementType();
448 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
449 // three or five element vertex only needs three or five registers,
450 // NOT four or eigth.
451 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
452 unsigned NumElements = ParamType->getVectorNumElements();
454 for (unsigned j = 0; j != NumElements; ++j) {
455 Splits.push_back(NewArg);
456 NewArg.PartOffset += NewArg.VT.getStoreSize();
459 } else if (Info->getShaderType() != ShaderType::COMPUTE) {
460 Splits.push_back(Arg);
464 SmallVector<CCValAssign, 16> ArgLocs;
465 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
468 // At least one interpolation mode must be enabled or else the GPU will hang.
469 if (Info->getShaderType() == ShaderType::PIXEL &&
470 (Info->PSInputAddr & 0x7F) == 0) {
471 Info->PSInputAddr |= 1;
472 CCInfo.AllocateReg(AMDGPU::VGPR0);
473 CCInfo.AllocateReg(AMDGPU::VGPR1);
476 // The pointer to the list of arguments is stored in SGPR0, SGPR1
477 // The pointer to the scratch buffer is stored in SGPR2, SGPR3
478 if (Info->getShaderType() == ShaderType::COMPUTE) {
479 if (Subtarget->isAmdHsaOS())
480 Info->NumUserSGPRs = 2; // FIXME: Need to support scratch buffers.
482 Info->NumUserSGPRs = 4;
484 unsigned InputPtrReg =
485 TRI->getPreloadedValue(MF, SIRegisterInfo::INPUT_PTR);
486 unsigned InputPtrRegLo =
487 TRI->getPhysRegSubReg(InputPtrReg, &AMDGPU::SReg_32RegClass, 0);
488 unsigned InputPtrRegHi =
489 TRI->getPhysRegSubReg(InputPtrReg, &AMDGPU::SReg_32RegClass, 1);
491 unsigned ScratchPtrReg =
492 TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_PTR);
493 unsigned ScratchPtrRegLo =
494 TRI->getPhysRegSubReg(ScratchPtrReg, &AMDGPU::SReg_32RegClass, 0);
495 unsigned ScratchPtrRegHi =
496 TRI->getPhysRegSubReg(ScratchPtrReg, &AMDGPU::SReg_32RegClass, 1);
498 CCInfo.AllocateReg(InputPtrRegLo);
499 CCInfo.AllocateReg(InputPtrRegHi);
500 CCInfo.AllocateReg(ScratchPtrRegLo);
501 CCInfo.AllocateReg(ScratchPtrRegHi);
502 MF.addLiveIn(InputPtrReg, &AMDGPU::SReg_64RegClass);
503 MF.addLiveIn(ScratchPtrReg, &AMDGPU::SReg_64RegClass);
506 if (Info->getShaderType() == ShaderType::COMPUTE) {
507 getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
511 AnalyzeFormalArguments(CCInfo, Splits);
513 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
515 const ISD::InputArg &Arg = Ins[i];
517 InVals.push_back(DAG.getUNDEF(Arg.VT));
521 CCValAssign &VA = ArgLocs[ArgIdx++];
522 MVT VT = VA.getLocVT();
526 EVT MemVT = Splits[i].VT;
527 const unsigned Offset = 36 + VA.getLocMemOffset();
528 // The first 36 bytes of the input buffer contains information about
529 // thread group and global sizes.
530 SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, DAG.getRoot(),
531 Offset, Ins[i].Flags.isSExt());
533 const PointerType *ParamTy =
534 dyn_cast<PointerType>(FType->getParamType(Ins[i].OrigArgIndex));
535 if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS &&
536 ParamTy && ParamTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
537 // On SI local pointers are just offsets into LDS, so they are always
538 // less than 16-bits. On CI and newer they could potentially be
539 // real pointers, so we can't guarantee their size.
540 Arg = DAG.getNode(ISD::AssertZext, DL, Arg.getValueType(), Arg,
541 DAG.getValueType(MVT::i16));
544 InVals.push_back(Arg);
545 Info->ABIArgOffset = Offset + MemVT.getStoreSize();
548 assert(VA.isRegLoc() && "Parameter must be in a register!");
550 unsigned Reg = VA.getLocReg();
552 if (VT == MVT::i64) {
553 // For now assume it is a pointer
554 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
555 &AMDGPU::SReg_64RegClass);
556 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
557 InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
561 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
563 Reg = MF.addLiveIn(Reg, RC);
564 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
566 if (Arg.VT.isVector()) {
568 // Build a vector from the registers
569 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
570 unsigned NumElements = ParamType->getVectorNumElements();
572 SmallVector<SDValue, 4> Regs;
574 for (unsigned j = 1; j != NumElements; ++j) {
575 Reg = ArgLocs[ArgIdx++].getLocReg();
576 Reg = MF.addLiveIn(Reg, RC);
577 Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
580 // Fill up the missing vector elements
581 NumElements = Arg.VT.getVectorNumElements() - NumElements;
582 for (unsigned j = 0; j != NumElements; ++j)
583 Regs.push_back(DAG.getUNDEF(VT));
585 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT, Regs));
589 InVals.push_back(Val);
592 if (Info->getShaderType() != ShaderType::COMPUTE) {
593 unsigned ScratchIdx = CCInfo.getFirstUnallocated(
594 AMDGPU::SGPR_32RegClass.begin(), AMDGPU::SGPR_32RegClass.getNumRegs());
595 Info->ScratchOffsetReg = AMDGPU::SGPR_32RegClass.getRegister(ScratchIdx);
600 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
601 MachineInstr * MI, MachineBasicBlock * BB) const {
603 MachineBasicBlock::iterator I = *MI;
604 const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
605 getTargetMachine().getSubtargetImpl()->getInstrInfo());
607 switch (MI->getOpcode()) {
609 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
610 case AMDGPU::BRANCH: return BB;
611 case AMDGPU::V_SUB_F64: {
612 unsigned DestReg = MI->getOperand(0).getReg();
613 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F64), DestReg)
614 .addImm(0) // SRC0 modifiers
615 .addReg(MI->getOperand(1).getReg())
616 .addImm(1) // SRC1 modifiers
617 .addReg(MI->getOperand(2).getReg())
620 MI->eraseFromParent();
623 case AMDGPU::SI_RegisterStorePseudo: {
624 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
625 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
626 MachineInstrBuilder MIB =
627 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::SI_RegisterStore),
629 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i)
630 MIB.addOperand(MI->getOperand(i));
632 MI->eraseFromParent();
639 bool SITargetLowering::enableAggressiveFMAFusion(EVT VT) const {
640 // This currently forces unfolding various combinations of fsub into fma with
641 // free fneg'd operands. As long as we have fast FMA (controlled by
642 // isFMAFasterThanFMulAndFAdd), we should perform these.
644 // When fma is quarter rate, for f64 where add / sub are at best half rate,
645 // most of these combines appear to be cycle neutral but save on instruction
646 // count / code size.
650 EVT SITargetLowering::getSetCCResultType(LLVMContext &Ctx, EVT VT) const {
651 if (!VT.isVector()) {
654 return EVT::getVectorVT(Ctx, MVT::i1, VT.getVectorNumElements());
657 MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
661 // Answering this is somewhat tricky and depends on the specific device which
662 // have different rates for fma or all f64 operations.
664 // v_fma_f64 and v_mul_f64 always take the same number of cycles as each other
665 // regardless of which device (although the number of cycles differs between
666 // devices), so it is always profitable for f64.
668 // v_fma_f32 takes 4 or 16 cycles depending on the device, so it is profitable
669 // only on full rate devices. Normally, we should prefer selecting v_mad_f32
670 // which we can always do even without fused FP ops since it returns the same
671 // result as the separate operations and since it is always full
672 // rate. Therefore, we lie and report that it is not faster for f32. v_mad_f32
673 // however does not support denormals, so we do report fma as faster if we have
674 // a fast fma device and require denormals.
676 bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
677 VT = VT.getScalarType();
682 switch (VT.getSimpleVT().SimpleTy) {
684 // This is as fast on some subtargets. However, we always have full rate f32
685 // mad available which returns the same result as the separate operations
686 // which we should prefer over fma.
697 //===----------------------------------------------------------------------===//
698 // Custom DAG Lowering Operations
699 //===----------------------------------------------------------------------===//
701 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
702 switch (Op.getOpcode()) {
703 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
704 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
705 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
707 SDValue Result = LowerLOAD(Op, DAG);
708 assert((!Result.getNode() ||
709 Result.getNode()->getNumValues() == 2) &&
710 "Load should return a value and a chain");
716 return LowerTrig(Op, DAG);
717 case ISD::SELECT: return LowerSELECT(Op, DAG);
718 case ISD::FDIV: return LowerFDIV(Op, DAG);
719 case ISD::STORE: return LowerSTORE(Op, DAG);
720 case ISD::GlobalAddress: {
721 MachineFunction &MF = DAG.getMachineFunction();
722 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
723 return LowerGlobalAddress(MFI, Op, DAG);
725 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
726 case ISD::INTRINSIC_VOID: return LowerINTRINSIC_VOID(Op, DAG);
731 /// \brief Helper function for LowerBRCOND
732 static SDNode *findUser(SDValue Value, unsigned Opcode) {
734 SDNode *Parent = Value.getNode();
735 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
738 if (I.getUse().get() != Value)
741 if (I->getOpcode() == Opcode)
747 SDValue SITargetLowering::LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const {
749 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Op);
750 unsigned FrameIndex = FINode->getIndex();
752 return DAG.getTargetFrameIndex(FrameIndex, MVT::i32);
755 /// This transforms the control flow intrinsics to get the branch destination as
756 /// last parameter, also switches branch target with BR if the need arise
757 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
758 SelectionDAG &DAG) const {
762 SDNode *Intr = BRCOND.getOperand(1).getNode();
763 SDValue Target = BRCOND.getOperand(2);
764 SDNode *BR = nullptr;
766 if (Intr->getOpcode() == ISD::SETCC) {
767 // As long as we negate the condition everything is fine
768 SDNode *SetCC = Intr;
769 assert(SetCC->getConstantOperandVal(1) == 1);
770 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
772 Intr = SetCC->getOperand(0).getNode();
775 // Get the target from BR if we don't negate the condition
776 BR = findUser(BRCOND, ISD::BR);
777 Target = BR->getOperand(1);
780 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
782 // Build the result and
783 SmallVector<EVT, 4> Res;
784 for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
785 Res.push_back(Intr->getValueType(i));
787 // operands of the new intrinsic call
788 SmallVector<SDValue, 4> Ops;
789 Ops.push_back(BRCOND.getOperand(0));
790 for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
791 Ops.push_back(Intr->getOperand(i));
792 Ops.push_back(Target);
794 // build the new intrinsic call
795 SDNode *Result = DAG.getNode(
796 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
797 DAG.getVTList(Res), Ops).getNode();
800 // Give the branch instruction our target
805 SDValue NewBR = DAG.getNode(ISD::BR, DL, BR->getVTList(), Ops);
806 DAG.ReplaceAllUsesWith(BR, NewBR.getNode());
807 BR = NewBR.getNode();
810 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
812 // Copy the intrinsic results to registers
813 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
814 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
818 Chain = DAG.getCopyToReg(
820 CopyToReg->getOperand(1),
821 SDValue(Result, i - 1),
824 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
827 // Remove the old intrinsic from the chain
828 DAG.ReplaceAllUsesOfValueWith(
829 SDValue(Intr, Intr->getNumValues() - 1),
830 Intr->getOperand(0));
835 SDValue SITargetLowering::LowerGlobalAddress(AMDGPUMachineFunction *MFI,
837 SelectionDAG &DAG) const {
838 GlobalAddressSDNode *GSD = cast<GlobalAddressSDNode>(Op);
840 if (GSD->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS)
841 return AMDGPUTargetLowering::LowerGlobalAddress(MFI, Op, DAG);
844 const GlobalValue *GV = GSD->getGlobal();
845 MVT PtrVT = getPointerTy(GSD->getAddressSpace());
847 SDValue Ptr = DAG.getNode(AMDGPUISD::CONST_DATA_PTR, DL, PtrVT);
848 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32);
850 SDValue PtrLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Ptr,
851 DAG.getConstant(0, MVT::i32));
852 SDValue PtrHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Ptr,
853 DAG.getConstant(1, MVT::i32));
855 SDValue Lo = DAG.getNode(ISD::ADDC, DL, DAG.getVTList(MVT::i32, MVT::Glue),
857 SDValue Hi = DAG.getNode(ISD::ADDE, DL, DAG.getVTList(MVT::i32, MVT::Glue),
858 PtrHi, DAG.getConstant(0, MVT::i32),
859 SDValue(Lo.getNode(), 1));
860 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi);
863 SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
864 SelectionDAG &DAG) const {
865 MachineFunction &MF = DAG.getMachineFunction();
866 const SIRegisterInfo *TRI =
867 static_cast<const SIRegisterInfo*>(MF.getSubtarget().getRegisterInfo());
869 EVT VT = Op.getValueType();
871 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
873 switch (IntrinsicID) {
874 case Intrinsic::r600_read_ngroups_x:
875 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
876 SI::KernelInputOffsets::NGROUPS_X, false);
877 case Intrinsic::r600_read_ngroups_y:
878 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
879 SI::KernelInputOffsets::NGROUPS_Y, false);
880 case Intrinsic::r600_read_ngroups_z:
881 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
882 SI::KernelInputOffsets::NGROUPS_Z, false);
883 case Intrinsic::r600_read_global_size_x:
884 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
885 SI::KernelInputOffsets::GLOBAL_SIZE_X, false);
886 case Intrinsic::r600_read_global_size_y:
887 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
888 SI::KernelInputOffsets::GLOBAL_SIZE_Y, false);
889 case Intrinsic::r600_read_global_size_z:
890 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
891 SI::KernelInputOffsets::GLOBAL_SIZE_Z, false);
892 case Intrinsic::r600_read_local_size_x:
893 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
894 SI::KernelInputOffsets::LOCAL_SIZE_X, false);
895 case Intrinsic::r600_read_local_size_y:
896 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
897 SI::KernelInputOffsets::LOCAL_SIZE_Y, false);
898 case Intrinsic::r600_read_local_size_z:
899 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
900 SI::KernelInputOffsets::LOCAL_SIZE_Z, false);
902 case Intrinsic::AMDGPU_read_workdim:
903 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
904 MF.getInfo<SIMachineFunctionInfo>()->ABIArgOffset,
907 case Intrinsic::r600_read_tgid_x:
908 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
909 TRI->getPreloadedValue(MF, SIRegisterInfo::TGID_X), VT);
910 case Intrinsic::r600_read_tgid_y:
911 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
912 TRI->getPreloadedValue(MF, SIRegisterInfo::TGID_Y), VT);
913 case Intrinsic::r600_read_tgid_z:
914 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
915 TRI->getPreloadedValue(MF, SIRegisterInfo::TGID_Z), VT);
916 case Intrinsic::r600_read_tidig_x:
917 return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
918 TRI->getPreloadedValue(MF, SIRegisterInfo::TIDIG_X), VT);
919 case Intrinsic::r600_read_tidig_y:
920 return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
921 TRI->getPreloadedValue(MF, SIRegisterInfo::TIDIG_Y), VT);
922 case Intrinsic::r600_read_tidig_z:
923 return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
924 TRI->getPreloadedValue(MF, SIRegisterInfo::TIDIG_Z), VT);
925 case AMDGPUIntrinsic::SI_load_const: {
931 MachineMemOperand *MMO = MF.getMachineMemOperand(
932 MachinePointerInfo(),
933 MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
934 VT.getStoreSize(), 4);
935 return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
936 Op->getVTList(), Ops, VT, MMO);
938 case AMDGPUIntrinsic::SI_sample:
939 return LowerSampleIntrinsic(AMDGPUISD::SAMPLE, Op, DAG);
940 case AMDGPUIntrinsic::SI_sampleb:
941 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEB, Op, DAG);
942 case AMDGPUIntrinsic::SI_sampled:
943 return LowerSampleIntrinsic(AMDGPUISD::SAMPLED, Op, DAG);
944 case AMDGPUIntrinsic::SI_samplel:
945 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEL, Op, DAG);
946 case AMDGPUIntrinsic::SI_vs_load_input:
947 return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
952 return AMDGPUTargetLowering::LowerOperation(Op, DAG);
956 SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
957 SelectionDAG &DAG) const {
958 MachineFunction &MF = DAG.getMachineFunction();
959 SDValue Chain = Op.getOperand(0);
960 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
962 switch (IntrinsicID) {
963 case AMDGPUIntrinsic::SI_tbuffer_store: {
982 EVT VT = Op.getOperand(3).getValueType();
984 MachineMemOperand *MMO = MF.getMachineMemOperand(
985 MachinePointerInfo(),
986 MachineMemOperand::MOStore,
987 VT.getStoreSize(), 4);
988 return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL,
989 Op->getVTList(), Ops, VT, MMO);
996 SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
998 LoadSDNode *Load = cast<LoadSDNode>(Op);
1000 if (Op.getValueType().isVector()) {
1001 assert(Op.getValueType().getVectorElementType() == MVT::i32 &&
1002 "Custom lowering for non-i32 vectors hasn't been implemented.");
1003 unsigned NumElements = Op.getValueType().getVectorNumElements();
1004 assert(NumElements != 2 && "v2 loads are supported for all address spaces.");
1005 switch (Load->getAddressSpace()) {
1007 case AMDGPUAS::GLOBAL_ADDRESS:
1008 case AMDGPUAS::PRIVATE_ADDRESS:
1009 // v4 loads are supported for private and global memory.
1010 if (NumElements <= 4)
1013 case AMDGPUAS::LOCAL_ADDRESS:
1014 return ScalarizeVectorLoad(Op, DAG);
1018 return AMDGPUTargetLowering::LowerLOAD(Op, DAG);
1021 SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
1023 SelectionDAG &DAG) const {
1024 return DAG.getNode(Opcode, SDLoc(Op), Op.getValueType(), Op.getOperand(1),
1030 SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
1031 if (Op.getValueType() != MVT::i64)
1035 SDValue Cond = Op.getOperand(0);
1037 SDValue Zero = DAG.getConstant(0, MVT::i32);
1038 SDValue One = DAG.getConstant(1, MVT::i32);
1040 SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1));
1041 SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2));
1043 SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero);
1044 SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero);
1046 SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1);
1048 SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One);
1049 SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One);
1051 SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1);
1053 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i32, Lo, Hi);
1054 return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
1057 // Catch division cases where we can use shortcuts with rcp and rsq
1059 SDValue SITargetLowering::LowerFastFDIV(SDValue Op, SelectionDAG &DAG) const {
1061 SDValue LHS = Op.getOperand(0);
1062 SDValue RHS = Op.getOperand(1);
1063 EVT VT = Op.getValueType();
1064 bool Unsafe = DAG.getTarget().Options.UnsafeFPMath;
1066 if (const ConstantFPSDNode *CLHS = dyn_cast<ConstantFPSDNode>(LHS)) {
1067 if ((Unsafe || (VT == MVT::f32 && !Subtarget->hasFP32Denormals())) &&
1068 CLHS->isExactlyValue(1.0)) {
1069 // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to
1070 // the CI documentation has a worst case error of 1 ulp.
1071 // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to
1072 // use it as long as we aren't trying to use denormals.
1074 // 1.0 / sqrt(x) -> rsq(x)
1076 // XXX - Is UnsafeFPMath sufficient to do this for f64? The maximum ULP
1077 // error seems really high at 2^29 ULP.
1078 if (RHS.getOpcode() == ISD::FSQRT)
1079 return DAG.getNode(AMDGPUISD::RSQ, SL, VT, RHS.getOperand(0));
1081 // 1.0 / x -> rcp(x)
1082 return DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
1087 // Turn into multiply by the reciprocal.
1088 // x / y -> x * (1.0 / y)
1089 SDValue Recip = DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
1090 return DAG.getNode(ISD::FMUL, SL, VT, LHS, Recip);
1096 SDValue SITargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const {
1097 SDValue FastLowered = LowerFastFDIV(Op, DAG);
1098 if (FastLowered.getNode())
1101 // This uses v_rcp_f32 which does not handle denormals. Let this hit a
1102 // selection error for now rather than do something incorrect.
1103 if (Subtarget->hasFP32Denormals())
1107 SDValue LHS = Op.getOperand(0);
1108 SDValue RHS = Op.getOperand(1);
1110 SDValue r1 = DAG.getNode(ISD::FABS, SL, MVT::f32, RHS);
1112 const APFloat K0Val(BitsToFloat(0x6f800000));
1113 const SDValue K0 = DAG.getConstantFP(K0Val, MVT::f32);
1115 const APFloat K1Val(BitsToFloat(0x2f800000));
1116 const SDValue K1 = DAG.getConstantFP(K1Val, MVT::f32);
1118 const SDValue One = DAG.getConstantFP(1.0, MVT::f32);
1120 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f32);
1122 SDValue r2 = DAG.getSetCC(SL, SetCCVT, r1, K0, ISD::SETOGT);
1124 SDValue r3 = DAG.getNode(ISD::SELECT, SL, MVT::f32, r2, K1, One);
1126 r1 = DAG.getNode(ISD::FMUL, SL, MVT::f32, RHS, r3);
1128 SDValue r0 = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f32, r1);
1130 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f32, LHS, r0);
1132 return DAG.getNode(ISD::FMUL, SL, MVT::f32, r3, Mul);
1135 SDValue SITargetLowering::LowerFDIV64(SDValue Op, SelectionDAG &DAG) const {
1139 SDValue SITargetLowering::LowerFDIV(SDValue Op, SelectionDAG &DAG) const {
1140 EVT VT = Op.getValueType();
1143 return LowerFDIV32(Op, DAG);
1146 return LowerFDIV64(Op, DAG);
1148 llvm_unreachable("Unexpected type for fdiv");
1151 SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1153 StoreSDNode *Store = cast<StoreSDNode>(Op);
1154 EVT VT = Store->getMemoryVT();
1156 // These stores are legal.
1157 if (Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS &&
1158 VT.isVector() && VT.getVectorNumElements() == 2 &&
1159 VT.getVectorElementType() == MVT::i32)
1162 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) {
1163 if (VT.isVector() && VT.getVectorNumElements() > 4)
1164 return ScalarizeVectorStore(Op, DAG);
1168 SDValue Ret = AMDGPUTargetLowering::LowerSTORE(Op, DAG);
1172 if (VT.isVector() && VT.getVectorNumElements() >= 8)
1173 return ScalarizeVectorStore(Op, DAG);
1176 return DAG.getTruncStore(Store->getChain(), DL,
1177 DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32),
1178 Store->getBasePtr(), MVT::i1, Store->getMemOperand());
1183 SDValue SITargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const {
1184 EVT VT = Op.getValueType();
1185 SDValue Arg = Op.getOperand(0);
1186 SDValue FractPart = DAG.getNode(AMDGPUISD::FRACT, SDLoc(Op), VT,
1187 DAG.getNode(ISD::FMUL, SDLoc(Op), VT, Arg,
1188 DAG.getConstantFP(0.5 / M_PI, VT)));
1190 switch (Op.getOpcode()) {
1192 return DAG.getNode(AMDGPUISD::COS_HW, SDLoc(Op), VT, FractPart);
1194 return DAG.getNode(AMDGPUISD::SIN_HW, SDLoc(Op), VT, FractPart);
1196 llvm_unreachable("Wrong trig opcode");
1200 //===----------------------------------------------------------------------===//
1201 // Custom DAG optimizations
1202 //===----------------------------------------------------------------------===//
1204 SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
1205 DAGCombinerInfo &DCI) const {
1206 EVT VT = N->getValueType(0);
1207 EVT ScalarVT = VT.getScalarType();
1208 if (ScalarVT != MVT::f32)
1211 SelectionDAG &DAG = DCI.DAG;
1214 SDValue Src = N->getOperand(0);
1215 EVT SrcVT = Src.getValueType();
1217 // TODO: We could try to match extracting the higher bytes, which would be
1218 // easier if i8 vectors weren't promoted to i32 vectors, particularly after
1219 // types are legalized. v4i8 -> v4f32 is probably the only case to worry
1220 // about in practice.
1221 if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) {
1222 if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) {
1223 SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src);
1224 DCI.AddToWorklist(Cvt.getNode());
1229 // We are primarily trying to catch operations on illegal vector types
1230 // before they are expanded.
1231 // For scalars, we can use the more flexible method of checking masked bits
1232 // after legalization.
1233 if (!DCI.isBeforeLegalize() ||
1234 !SrcVT.isVector() ||
1235 SrcVT.getVectorElementType() != MVT::i8) {
1239 assert(DCI.isBeforeLegalize() && "Unexpected legal type");
1241 // Weird sized vectors are a pain to handle, but we know 3 is really the same
1243 unsigned NElts = SrcVT.getVectorNumElements();
1244 if (!SrcVT.isSimple() && NElts != 3)
1247 // Handle v4i8 -> v4f32 extload. Replace the v4i8 with a legal i32 load to
1248 // prevent a mess from expanding to v4i32 and repacking.
1249 if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse()) {
1250 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), SrcVT);
1251 EVT RegVT = getEquivalentLoadRegType(*DAG.getContext(), SrcVT);
1252 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f32, NElts);
1253 LoadSDNode *Load = cast<LoadSDNode>(Src);
1255 unsigned AS = Load->getAddressSpace();
1256 unsigned Align = Load->getAlignment();
1257 Type *Ty = LoadVT.getTypeForEVT(*DAG.getContext());
1258 unsigned ABIAlignment = getDataLayout()->getABITypeAlignment(Ty);
1260 // Don't try to replace the load if we have to expand it due to alignment
1261 // problems. Otherwise we will end up scalarizing the load, and trying to
1262 // repack into the vector for no real reason.
1263 if (Align < ABIAlignment &&
1264 !allowsMisalignedMemoryAccesses(LoadVT, AS, Align, nullptr)) {
1268 SDValue NewLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegVT,
1272 Load->getMemOperand());
1274 // Make sure successors of the original load stay after it by updating
1275 // them to use the new Chain.
1276 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), NewLoad.getValue(1));
1278 SmallVector<SDValue, 4> Elts;
1279 if (RegVT.isVector())
1280 DAG.ExtractVectorElements(NewLoad, Elts);
1282 Elts.push_back(NewLoad);
1284 SmallVector<SDValue, 4> Ops;
1286 unsigned EltIdx = 0;
1287 for (SDValue Elt : Elts) {
1288 unsigned ComponentsInElt = std::min(4u, NElts - 4 * EltIdx);
1289 for (unsigned I = 0; I < ComponentsInElt; ++I) {
1290 unsigned Opc = AMDGPUISD::CVT_F32_UBYTE0 + I;
1291 SDValue Cvt = DAG.getNode(Opc, DL, MVT::f32, Elt);
1292 DCI.AddToWorklist(Cvt.getNode());
1299 assert(Ops.size() == NElts);
1301 return DAG.getNode(ISD::BUILD_VECTOR, DL, FloatVT, Ops);
1307 // (shl (add x, c1), c2) -> add (shl x, c2), (shl c1, c2)
1309 // This is a variant of
1310 // (mul (add x, c1), c2) -> add (mul x, c2), (mul c1, c2),
1312 // The normal DAG combiner will do this, but only if the add has one use since
1313 // that would increase the number of instructions.
1315 // This prevents us from seeing a constant offset that can be folded into a
1316 // memory instruction's addressing mode. If we know the resulting add offset of
1317 // a pointer can be folded into an addressing offset, we can replace the pointer
1318 // operand with the add of new constant offset. This eliminates one of the uses,
1319 // and may allow the remaining use to also be simplified.
1321 SDValue SITargetLowering::performSHLPtrCombine(SDNode *N,
1323 DAGCombinerInfo &DCI) const {
1324 SDValue N0 = N->getOperand(0);
1325 SDValue N1 = N->getOperand(1);
1327 if (N0.getOpcode() != ISD::ADD)
1330 const ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N1);
1334 const ConstantSDNode *CAdd = dyn_cast<ConstantSDNode>(N0.getOperand(1));
1338 const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
1339 getTargetMachine().getSubtargetImpl()->getInstrInfo());
1341 // If the resulting offset is too large, we can't fold it into the addressing
1343 APInt Offset = CAdd->getAPIntValue() << CN1->getAPIntValue();
1344 if (!TII->canFoldOffset(Offset.getZExtValue(), AddrSpace))
1347 SelectionDAG &DAG = DCI.DAG;
1349 EVT VT = N->getValueType(0);
1351 SDValue ShlX = DAG.getNode(ISD::SHL, SL, VT, N0.getOperand(0), N1);
1352 SDValue COffset = DAG.getConstant(Offset, MVT::i32);
1354 return DAG.getNode(ISD::ADD, SL, VT, ShlX, COffset);
1357 SDValue SITargetLowering::performAndCombine(SDNode *N,
1358 DAGCombinerInfo &DCI) const {
1359 if (DCI.isBeforeLegalize())
1362 SelectionDAG &DAG = DCI.DAG;
1364 // (and (fcmp ord x, x), (fcmp une (fabs x), inf)) ->
1365 // fp_class x, ~(s_nan | q_nan | n_infinity | p_infinity)
1366 SDValue LHS = N->getOperand(0);
1367 SDValue RHS = N->getOperand(1);
1369 if (LHS.getOpcode() == ISD::SETCC &&
1370 RHS.getOpcode() == ISD::SETCC) {
1371 ISD::CondCode LCC = cast<CondCodeSDNode>(LHS.getOperand(2))->get();
1372 ISD::CondCode RCC = cast<CondCodeSDNode>(RHS.getOperand(2))->get();
1374 SDValue X = LHS.getOperand(0);
1375 SDValue Y = RHS.getOperand(0);
1376 if (Y.getOpcode() != ISD::FABS || Y.getOperand(0) != X)
1379 if (LCC == ISD::SETO) {
1380 if (X != LHS.getOperand(1))
1383 if (RCC == ISD::SETUNE) {
1384 const ConstantFPSDNode *C1 = dyn_cast<ConstantFPSDNode>(RHS.getOperand(1));
1385 if (!C1 || !C1->isInfinity() || C1->isNegative())
1388 const uint32_t Mask = SIInstrFlags::N_NORMAL |
1389 SIInstrFlags::N_SUBNORMAL |
1390 SIInstrFlags::N_ZERO |
1391 SIInstrFlags::P_ZERO |
1392 SIInstrFlags::P_SUBNORMAL |
1393 SIInstrFlags::P_NORMAL;
1395 static_assert(((~(SIInstrFlags::S_NAN |
1396 SIInstrFlags::Q_NAN |
1397 SIInstrFlags::N_INFINITY |
1398 SIInstrFlags::P_INFINITY)) & 0x3ff) == Mask,
1401 return DAG.getNode(AMDGPUISD::FP_CLASS, SDLoc(N), MVT::i1,
1402 X, DAG.getConstant(Mask, MVT::i32));
1410 SDValue SITargetLowering::performOrCombine(SDNode *N,
1411 DAGCombinerInfo &DCI) const {
1412 SelectionDAG &DAG = DCI.DAG;
1413 SDValue LHS = N->getOperand(0);
1414 SDValue RHS = N->getOperand(1);
1416 // or (fp_class x, c1), (fp_class x, c2) -> fp_class x, (c1 | c2)
1417 if (LHS.getOpcode() == AMDGPUISD::FP_CLASS &&
1418 RHS.getOpcode() == AMDGPUISD::FP_CLASS) {
1419 SDValue Src = LHS.getOperand(0);
1420 if (Src != RHS.getOperand(0))
1423 const ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(LHS.getOperand(1));
1424 const ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(RHS.getOperand(1));
1428 // Only 10 bits are used.
1429 static const uint32_t MaxMask = 0x3ff;
1431 uint32_t NewMask = (CLHS->getZExtValue() | CRHS->getZExtValue()) & MaxMask;
1432 return DAG.getNode(AMDGPUISD::FP_CLASS, SDLoc(N), MVT::i1,
1433 Src, DAG.getConstant(NewMask, MVT::i32));
1439 SDValue SITargetLowering::performClassCombine(SDNode *N,
1440 DAGCombinerInfo &DCI) const {
1441 SelectionDAG &DAG = DCI.DAG;
1442 SDValue Mask = N->getOperand(1);
1444 // fp_class x, 0 -> false
1445 if (const ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(Mask)) {
1446 if (CMask->isNullValue())
1447 return DAG.getConstant(0, MVT::i1);
1453 static unsigned minMaxOpcToMin3Max3Opc(unsigned Opc) {
1456 return AMDGPUISD::FMAX3;
1457 case AMDGPUISD::SMAX:
1458 return AMDGPUISD::SMAX3;
1459 case AMDGPUISD::UMAX:
1460 return AMDGPUISD::UMAX3;
1462 return AMDGPUISD::FMIN3;
1463 case AMDGPUISD::SMIN:
1464 return AMDGPUISD::SMIN3;
1465 case AMDGPUISD::UMIN:
1466 return AMDGPUISD::UMIN3;
1468 llvm_unreachable("Not a min/max opcode");
1472 SDValue SITargetLowering::performMin3Max3Combine(SDNode *N,
1473 DAGCombinerInfo &DCI) const {
1474 SelectionDAG &DAG = DCI.DAG;
1476 unsigned Opc = N->getOpcode();
1477 SDValue Op0 = N->getOperand(0);
1478 SDValue Op1 = N->getOperand(1);
1480 // Only do this if the inner op has one use since this will just increases
1481 // register pressure for no benefit.
1483 // max(max(a, b), c)
1484 if (Op0.getOpcode() == Opc && Op0.hasOneUse()) {
1486 return DAG.getNode(minMaxOpcToMin3Max3Opc(Opc),
1494 // max(a, max(b, c))
1495 if (Op1.getOpcode() == Opc && Op1.hasOneUse()) {
1497 return DAG.getNode(minMaxOpcToMin3Max3Opc(Opc),
1508 SDValue SITargetLowering::performSetCCCombine(SDNode *N,
1509 DAGCombinerInfo &DCI) const {
1510 SelectionDAG &DAG = DCI.DAG;
1513 SDValue LHS = N->getOperand(0);
1514 SDValue RHS = N->getOperand(1);
1515 EVT VT = LHS.getValueType();
1517 if (VT != MVT::f32 && VT != MVT::f64)
1520 // Match isinf pattern
1521 // (fcmp oeq (fabs x), inf) -> (fp_class x, (p_infinity | n_infinity))
1522 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
1523 if (CC == ISD::SETOEQ && LHS.getOpcode() == ISD::FABS) {
1524 const ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS);
1528 const APFloat &APF = CRHS->getValueAPF();
1529 if (APF.isInfinity() && !APF.isNegative()) {
1530 unsigned Mask = SIInstrFlags::P_INFINITY | SIInstrFlags::N_INFINITY;
1531 return DAG.getNode(AMDGPUISD::FP_CLASS, SL, MVT::i1,
1532 LHS.getOperand(0), DAG.getConstant(Mask, MVT::i32));
1539 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
1540 DAGCombinerInfo &DCI) const {
1541 SelectionDAG &DAG = DCI.DAG;
1544 switch (N->getOpcode()) {
1546 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1548 return performSetCCCombine(N, DCI);
1549 case ISD::FMAXNUM: // TODO: What about fmax_legacy?
1551 case AMDGPUISD::SMAX:
1552 case AMDGPUISD::SMIN:
1553 case AMDGPUISD::UMAX:
1554 case AMDGPUISD::UMIN: {
1555 if (DCI.getDAGCombineLevel() >= AfterLegalizeDAG &&
1556 getTargetMachine().getOptLevel() > CodeGenOpt::None)
1557 return performMin3Max3Combine(N, DCI);
1561 case AMDGPUISD::CVT_F32_UBYTE0:
1562 case AMDGPUISD::CVT_F32_UBYTE1:
1563 case AMDGPUISD::CVT_F32_UBYTE2:
1564 case AMDGPUISD::CVT_F32_UBYTE3: {
1565 unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
1567 SDValue Src = N->getOperand(0);
1568 APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);
1570 APInt KnownZero, KnownOne;
1571 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1572 !DCI.isBeforeLegalizeOps());
1573 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1574 if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
1575 TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {
1576 DCI.CommitTargetLoweringOpt(TLO);
1582 case ISD::UINT_TO_FP: {
1583 return performUCharToFloatCombine(N, DCI);
1586 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
1589 EVT VT = N->getValueType(0);
1593 SDValue LHS = N->getOperand(0);
1594 SDValue RHS = N->getOperand(1);
1596 // These should really be instruction patterns, but writing patterns with
1597 // source modiifiers is a pain.
1599 // fadd (fadd (a, a), b) -> mad 2.0, a, b
1600 if (LHS.getOpcode() == ISD::FADD) {
1601 SDValue A = LHS.getOperand(0);
1602 if (A == LHS.getOperand(1)) {
1603 const SDValue Two = DAG.getConstantFP(2.0, MVT::f32);
1604 return DAG.getNode(AMDGPUISD::MAD, DL, VT, Two, A, RHS);
1608 // fadd (b, fadd (a, a)) -> mad 2.0, a, b
1609 if (RHS.getOpcode() == ISD::FADD) {
1610 SDValue A = RHS.getOperand(0);
1611 if (A == RHS.getOperand(1)) {
1612 const SDValue Two = DAG.getConstantFP(2.0, MVT::f32);
1613 return DAG.getNode(AMDGPUISD::MAD, DL, VT, Two, A, LHS);
1620 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
1623 EVT VT = N->getValueType(0);
1625 // Try to get the fneg to fold into the source modifier. This undoes generic
1626 // DAG combines and folds them into the mad.
1627 if (VT == MVT::f32) {
1628 SDValue LHS = N->getOperand(0);
1629 SDValue RHS = N->getOperand(1);
1631 if (LHS.getOpcode() == ISD::FMUL) {
1632 // (fsub (fmul a, b), c) -> mad a, b, (fneg c)
1634 SDValue A = LHS.getOperand(0);
1635 SDValue B = LHS.getOperand(1);
1636 SDValue C = DAG.getNode(ISD::FNEG, DL, VT, RHS);
1638 return DAG.getNode(AMDGPUISD::MAD, DL, VT, A, B, C);
1641 if (RHS.getOpcode() == ISD::FMUL) {
1642 // (fsub c, (fmul a, b)) -> mad (fneg a), b, c
1644 SDValue A = DAG.getNode(ISD::FNEG, DL, VT, RHS.getOperand(0));
1645 SDValue B = RHS.getOperand(1);
1648 return DAG.getNode(AMDGPUISD::MAD, DL, VT, A, B, C);
1651 if (LHS.getOpcode() == ISD::FADD) {
1652 // (fsub (fadd a, a), c) -> mad 2.0, a, (fneg c)
1654 SDValue A = LHS.getOperand(0);
1655 if (A == LHS.getOperand(1)) {
1656 const SDValue Two = DAG.getConstantFP(2.0, MVT::f32);
1657 SDValue NegRHS = DAG.getNode(ISD::FNEG, DL, VT, RHS);
1659 return DAG.getNode(AMDGPUISD::MAD, DL, VT, Two, A, NegRHS);
1663 if (RHS.getOpcode() == ISD::FADD) {
1664 // (fsub c, (fadd a, a)) -> mad -2.0, a, c
1666 SDValue A = RHS.getOperand(0);
1667 if (A == RHS.getOperand(1)) {
1668 const SDValue NegTwo = DAG.getConstantFP(-2.0, MVT::f32);
1669 return DAG.getNode(AMDGPUISD::MAD, DL, VT, NegTwo, A, LHS);
1679 case ISD::ATOMIC_LOAD:
1680 case ISD::ATOMIC_STORE:
1681 case ISD::ATOMIC_CMP_SWAP:
1682 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
1683 case ISD::ATOMIC_SWAP:
1684 case ISD::ATOMIC_LOAD_ADD:
1685 case ISD::ATOMIC_LOAD_SUB:
1686 case ISD::ATOMIC_LOAD_AND:
1687 case ISD::ATOMIC_LOAD_OR:
1688 case ISD::ATOMIC_LOAD_XOR:
1689 case ISD::ATOMIC_LOAD_NAND:
1690 case ISD::ATOMIC_LOAD_MIN:
1691 case ISD::ATOMIC_LOAD_MAX:
1692 case ISD::ATOMIC_LOAD_UMIN:
1693 case ISD::ATOMIC_LOAD_UMAX: { // TODO: Target mem intrinsics.
1694 if (DCI.isBeforeLegalize())
1697 MemSDNode *MemNode = cast<MemSDNode>(N);
1698 SDValue Ptr = MemNode->getBasePtr();
1700 // TODO: We could also do this for multiplies.
1701 unsigned AS = MemNode->getAddressSpace();
1702 if (Ptr.getOpcode() == ISD::SHL && AS != AMDGPUAS::PRIVATE_ADDRESS) {
1703 SDValue NewPtr = performSHLPtrCombine(Ptr.getNode(), AS, DCI);
1705 SmallVector<SDValue, 8> NewOps;
1706 for (unsigned I = 0, E = MemNode->getNumOperands(); I != E; ++I)
1707 NewOps.push_back(MemNode->getOperand(I));
1709 NewOps[N->getOpcode() == ISD::STORE ? 2 : 1] = NewPtr;
1710 return SDValue(DAG.UpdateNodeOperands(MemNode, NewOps), 0);
1716 return performAndCombine(N, DCI);
1718 return performOrCombine(N, DCI);
1719 case AMDGPUISD::FP_CLASS:
1720 return performClassCombine(N, DCI);
1722 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1725 /// \brief Test if RegClass is one of the VSrc classes
1726 static bool isVSrc(unsigned RegClass) {
1728 default: return false;
1729 case AMDGPU::VS_32RegClassID:
1730 case AMDGPU::VS_64RegClassID:
1735 /// \brief Analyze the possible immediate value Op
1737 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
1738 /// and the immediate value if it's a literal immediate
1739 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
1741 const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
1742 getTargetMachine().getSubtargetImpl()->getInstrInfo());
1744 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
1745 if (Node->getZExtValue() >> 32)
1748 if (TII->isInlineConstant(Node->getAPIntValue()))
1751 return Node->getZExtValue();
1754 if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
1755 if (TII->isInlineConstant(Node->getValueAPF().bitcastToAPInt()))
1758 if (Node->getValueType(0) == MVT::f32)
1759 return FloatToBits(Node->getValueAPF().convertToFloat());
1767 const TargetRegisterClass *SITargetLowering::getRegClassForNode(
1768 SelectionDAG &DAG, const SDValue &Op) const {
1769 const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
1770 getTargetMachine().getSubtargetImpl()->getInstrInfo());
1771 const SIRegisterInfo &TRI = TII->getRegisterInfo();
1773 if (!Op->isMachineOpcode()) {
1774 switch(Op->getOpcode()) {
1775 case ISD::CopyFromReg: {
1776 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1777 unsigned Reg = cast<RegisterSDNode>(Op->getOperand(1))->getReg();
1778 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
1779 return MRI.getRegClass(Reg);
1781 return TRI.getPhysRegClass(Reg);
1783 default: return nullptr;
1786 const MCInstrDesc &Desc = TII->get(Op->getMachineOpcode());
1787 int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
1788 if (OpClassID != -1) {
1789 return TRI.getRegClass(OpClassID);
1791 switch(Op.getMachineOpcode()) {
1792 case AMDGPU::COPY_TO_REGCLASS:
1793 // Operand 1 is the register class id for COPY_TO_REGCLASS instructions.
1794 OpClassID = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1796 // If the COPY_TO_REGCLASS instruction is copying to a VSrc register
1797 // class, then the register class for the value could be either a
1798 // VReg or and SReg. In order to get a more accurate
1799 if (isVSrc(OpClassID))
1800 return getRegClassForNode(DAG, Op.getOperand(0));
1802 return TRI.getRegClass(OpClassID);
1803 case AMDGPU::EXTRACT_SUBREG: {
1804 int SubIdx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1805 const TargetRegisterClass *SuperClass =
1806 getRegClassForNode(DAG, Op.getOperand(0));
1807 return TRI.getSubClassWithSubReg(SuperClass, SubIdx);
1809 case AMDGPU::REG_SEQUENCE:
1810 // Operand 0 is the register class id for REG_SEQUENCE instructions.
1811 return TRI.getRegClass(
1812 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue());
1814 return getRegClassFor(Op.getSimpleValueType());
1818 /// \brief Does "Op" fit into register class "RegClass" ?
1819 bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, const SDValue &Op,
1820 unsigned RegClass) const {
1821 const TargetRegisterInfo *TRI =
1822 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
1823 const TargetRegisterClass *RC = getRegClassForNode(DAG, Op);
1827 return TRI->getRegClass(RegClass)->hasSubClassEq(RC);
1830 /// \brief Helper function for adjustWritemask
1831 static unsigned SubIdx2Lane(unsigned Idx) {
1834 case AMDGPU::sub0: return 0;
1835 case AMDGPU::sub1: return 1;
1836 case AMDGPU::sub2: return 2;
1837 case AMDGPU::sub3: return 3;
1841 /// \brief Adjust the writemask of MIMG instructions
1842 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
1843 SelectionDAG &DAG) const {
1844 SDNode *Users[4] = { };
1846 unsigned OldDmask = Node->getConstantOperandVal(0);
1847 unsigned NewDmask = 0;
1849 // Try to figure out the used register components
1850 for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
1853 // Abort if we can't understand the usage
1854 if (!I->isMachineOpcode() ||
1855 I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
1858 // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used.
1859 // Note that subregs are packed, i.e. Lane==0 is the first bit set
1860 // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit
1862 Lane = SubIdx2Lane(I->getConstantOperandVal(1));
1864 // Set which texture component corresponds to the lane.
1866 for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) {
1868 Comp = countTrailingZeros(Dmask);
1869 Dmask &= ~(1 << Comp);
1872 // Abort if we have more than one user per component
1877 NewDmask |= 1 << Comp;
1880 // Abort if there's no change
1881 if (NewDmask == OldDmask)
1884 // Adjust the writemask in the node
1885 std::vector<SDValue> Ops;
1886 Ops.push_back(DAG.getTargetConstant(NewDmask, MVT::i32));
1887 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
1888 Ops.push_back(Node->getOperand(i));
1889 Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops);
1891 // If we only got one lane, replace it with a copy
1892 // (if NewDmask has only one bit set...)
1893 if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
1894 SDValue RC = DAG.getTargetConstant(AMDGPU::VGPR_32RegClassID, MVT::i32);
1895 SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1896 SDLoc(), Users[Lane]->getValueType(0),
1897 SDValue(Node, 0), RC);
1898 DAG.ReplaceAllUsesWith(Users[Lane], Copy);
1902 // Update the users of the node with the new indices
1903 for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
1905 SDNode *User = Users[i];
1909 SDValue Op = DAG.getTargetConstant(Idx, MVT::i32);
1910 DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
1914 case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
1915 case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
1916 case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
1921 /// \brief Legalize target independent instructions (e.g. INSERT_SUBREG)
1922 /// with frame index operands.
1923 /// LLVM assumes that inputs are to these instructions are registers.
1924 void SITargetLowering::legalizeTargetIndependentNode(SDNode *Node,
1925 SelectionDAG &DAG) const {
1927 SmallVector<SDValue, 8> Ops;
1928 for (unsigned i = 0; i < Node->getNumOperands(); ++i) {
1929 if (!isa<FrameIndexSDNode>(Node->getOperand(i))) {
1930 Ops.push_back(Node->getOperand(i));
1935 Ops.push_back(SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL,
1936 Node->getOperand(i).getValueType(),
1937 Node->getOperand(i)), 0));
1940 DAG.UpdateNodeOperands(Node, Ops);
1943 /// \brief Fold the instructions after selecting them.
1944 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
1945 SelectionDAG &DAG) const {
1946 const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
1947 getTargetMachine().getSubtargetImpl()->getInstrInfo());
1948 Node = AdjustRegClass(Node, DAG);
1950 if (TII->isMIMG(Node->getMachineOpcode()))
1951 adjustWritemask(Node, DAG);
1953 if (Node->getMachineOpcode() == AMDGPU::INSERT_SUBREG ||
1954 Node->getMachineOpcode() == AMDGPU::REG_SEQUENCE) {
1955 legalizeTargetIndependentNode(Node, DAG);
1961 /// \brief Assign the register class depending on the number of
1962 /// bits set in the writemask
1963 void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
1964 SDNode *Node) const {
1965 const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
1966 getTargetMachine().getSubtargetImpl()->getInstrInfo());
1968 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1969 TII->legalizeOperands(MI);
1971 if (TII->isMIMG(MI->getOpcode())) {
1972 unsigned VReg = MI->getOperand(0).getReg();
1973 unsigned Writemask = MI->getOperand(1).getImm();
1974 unsigned BitsSet = 0;
1975 for (unsigned i = 0; i < 4; ++i)
1976 BitsSet += Writemask & (1 << i) ? 1 : 0;
1978 const TargetRegisterClass *RC;
1981 case 1: RC = &AMDGPU::VGPR_32RegClass; break;
1982 case 2: RC = &AMDGPU::VReg_64RegClass; break;
1983 case 3: RC = &AMDGPU::VReg_96RegClass; break;
1986 unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
1987 MI->setDesc(TII->get(NewOpcode));
1988 MRI.setRegClass(VReg, RC);
1992 // Replace unused atomics with the no return version.
1993 int NoRetAtomicOp = AMDGPU::getAtomicNoRetOp(MI->getOpcode());
1994 if (NoRetAtomicOp != -1) {
1995 if (!Node->hasAnyUseOfValue(0)) {
1996 MI->setDesc(TII->get(NoRetAtomicOp));
1997 MI->RemoveOperand(0);
2004 static SDValue buildSMovImm32(SelectionDAG &DAG, SDLoc DL, uint64_t Val) {
2005 SDValue K = DAG.getTargetConstant(Val, MVT::i32);
2006 return SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, K), 0);
2009 MachineSDNode *SITargetLowering::wrapAddr64Rsrc(SelectionDAG &DAG,
2011 SDValue Ptr) const {
2012 const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
2013 getTargetMachine().getSubtargetImpl()->getInstrInfo());
2015 // XXX - Workaround for moveToVALU not handling different register class
2016 // inserts for REG_SEQUENCE.
2018 // Build the half of the subregister with the constants.
2019 const SDValue Ops0[] = {
2020 DAG.getTargetConstant(AMDGPU::SGPR_64RegClassID, MVT::i32),
2021 buildSMovImm32(DAG, DL, 0),
2022 DAG.getTargetConstant(AMDGPU::sub0, MVT::i32),
2023 buildSMovImm32(DAG, DL, TII->getDefaultRsrcDataFormat() >> 32),
2024 DAG.getTargetConstant(AMDGPU::sub1, MVT::i32)
2027 SDValue SubRegHi = SDValue(DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL,
2028 MVT::v2i32, Ops0), 0);
2030 // Combine the constants and the pointer.
2031 const SDValue Ops1[] = {
2032 DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, MVT::i32),
2034 DAG.getTargetConstant(AMDGPU::sub0_sub1, MVT::i32),
2036 DAG.getTargetConstant(AMDGPU::sub2_sub3, MVT::i32)
2039 return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops1);
2041 const SDValue Ops[] = {
2042 DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, MVT::i32),
2044 DAG.getTargetConstant(AMDGPU::sub0_sub1, MVT::i32),
2045 buildSMovImm32(DAG, DL, 0),
2046 DAG.getTargetConstant(AMDGPU::sub2, MVT::i32),
2047 buildSMovImm32(DAG, DL, TII->getDefaultRsrcFormat() >> 32),
2048 DAG.getTargetConstant(AMDGPU::sub3, MVT::i32)
2051 return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops);
2056 /// \brief Return a resource descriptor with the 'Add TID' bit enabled
2057 /// The TID (Thread ID) is multipled by the stride value (bits [61:48]
2058 /// of the resource descriptor) to create an offset, which is added to the
2059 /// resource ponter.
2060 MachineSDNode *SITargetLowering::buildRSRC(SelectionDAG &DAG,
2063 uint32_t RsrcDword1,
2064 uint64_t RsrcDword2And3) const {
2065 SDValue PtrLo = DAG.getTargetExtractSubreg(AMDGPU::sub0, DL, MVT::i32, Ptr);
2066 SDValue PtrHi = DAG.getTargetExtractSubreg(AMDGPU::sub1, DL, MVT::i32, Ptr);
2068 PtrHi = SDValue(DAG.getMachineNode(AMDGPU::S_OR_B32, DL, MVT::i32, PtrHi,
2069 DAG.getConstant(RsrcDword1, MVT::i32)), 0);
2072 SDValue DataLo = buildSMovImm32(DAG, DL,
2073 RsrcDword2And3 & UINT64_C(0xFFFFFFFF));
2074 SDValue DataHi = buildSMovImm32(DAG, DL, RsrcDword2And3 >> 32);
2076 const SDValue Ops[] = {
2077 DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, MVT::i32),
2079 DAG.getTargetConstant(AMDGPU::sub0, MVT::i32),
2081 DAG.getTargetConstant(AMDGPU::sub1, MVT::i32),
2083 DAG.getTargetConstant(AMDGPU::sub2, MVT::i32),
2085 DAG.getTargetConstant(AMDGPU::sub3, MVT::i32)
2088 return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops);
2091 MachineSDNode *SITargetLowering::buildScratchRSRC(SelectionDAG &DAG,
2093 SDValue Ptr) const {
2094 const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
2095 getTargetMachine().getSubtargetImpl()->getInstrInfo());
2096 uint64_t Rsrc = TII->getDefaultRsrcDataFormat() | AMDGPU::RSRC_TID_ENABLE |
2099 return buildRSRC(DAG, DL, Ptr, 0, Rsrc);
2102 MachineSDNode *SITargetLowering::AdjustRegClass(MachineSDNode *N,
2103 SelectionDAG &DAG) const {
2106 unsigned NewOpcode = N->getMachineOpcode();
2108 switch (N->getMachineOpcode()) {
2110 case AMDGPU::S_LOAD_DWORD_IMM:
2111 NewOpcode = AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
2113 case AMDGPU::S_LOAD_DWORDX2_SGPR:
2114 if (NewOpcode == N->getMachineOpcode()) {
2115 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
2118 case AMDGPU::S_LOAD_DWORDX4_IMM:
2119 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
2120 if (NewOpcode == N->getMachineOpcode()) {
2121 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
2123 if (fitsRegClass(DAG, N->getOperand(0), AMDGPU::SReg_64RegClassID)) {
2126 ConstantSDNode *Offset = cast<ConstantSDNode>(N->getOperand(1));
2128 const SDValue Zero64 = DAG.getTargetConstant(0, MVT::i64);
2129 SDValue Ptr(DAG.getMachineNode(AMDGPU::S_MOV_B64, DL, MVT::i64, Zero64), 0);
2130 MachineSDNode *RSrc = wrapAddr64Rsrc(DAG, DL, Ptr);
2132 SmallVector<SDValue, 8> Ops;
2133 Ops.push_back(SDValue(RSrc, 0));
2134 Ops.push_back(N->getOperand(0));
2136 // The immediate offset is in dwords on SI and in bytes on VI.
2137 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
2138 Ops.push_back(DAG.getTargetConstant(Offset->getSExtValue(), MVT::i32));
2140 Ops.push_back(DAG.getTargetConstant(Offset->getSExtValue() << 2, MVT::i32));
2142 // Copy remaining operands so we keep any chain and glue nodes that follow
2143 // the normal operands.
2144 for (unsigned I = 2, E = N->getNumOperands(); I != E; ++I)
2145 Ops.push_back(N->getOperand(I));
2147 return DAG.getMachineNode(NewOpcode, DL, N->getVTList(), Ops);
2152 SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2153 const TargetRegisterClass *RC,
2154 unsigned Reg, EVT VT) const {
2155 SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT);
2157 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()),
2158 cast<RegisterSDNode>(VReg)->getReg(), VT);