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 const AMDGPUSubtarget &STI)
40 : AMDGPUTargetLowering(TM, STI) {
41 addRegisterClass(MVT::i1, &AMDGPU::VReg_1RegClass);
42 addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
44 addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
45 addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);
47 addRegisterClass(MVT::i32, &AMDGPU::SReg_32RegClass);
48 addRegisterClass(MVT::f32, &AMDGPU::VGPR_32RegClass);
50 addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
51 addRegisterClass(MVT::v2i32, &AMDGPU::SReg_64RegClass);
52 addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
54 addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
55 addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
57 addRegisterClass(MVT::v8i32, &AMDGPU::SReg_256RegClass);
58 addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
60 addRegisterClass(MVT::v16i32, &AMDGPU::SReg_512RegClass);
61 addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);
63 computeRegisterProperties(STI.getRegisterInfo());
65 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
66 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
67 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
68 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);
70 setOperationAction(ISD::ADD, MVT::i32, Legal);
71 setOperationAction(ISD::ADDC, MVT::i32, Legal);
72 setOperationAction(ISD::ADDE, MVT::i32, Legal);
73 setOperationAction(ISD::SUBC, MVT::i32, Legal);
74 setOperationAction(ISD::SUBE, MVT::i32, Legal);
76 setOperationAction(ISD::FSIN, MVT::f32, Custom);
77 setOperationAction(ISD::FCOS, MVT::f32, Custom);
79 setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
80 setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
82 // We need to custom lower vector stores from local memory
83 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
84 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
85 setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
87 setOperationAction(ISD::STORE, MVT::v8i32, Custom);
88 setOperationAction(ISD::STORE, MVT::v16i32, Custom);
90 setOperationAction(ISD::STORE, MVT::i1, Custom);
91 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
93 setOperationAction(ISD::SELECT, MVT::i64, Custom);
94 setOperationAction(ISD::SELECT, MVT::f64, Promote);
95 AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
97 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
98 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
99 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
100 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
102 setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
103 setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
105 setOperationAction(ISD::BSWAP, MVT::i32, Legal);
107 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
108 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
109 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
111 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
112 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
113 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
115 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
116 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
117 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
119 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
120 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
122 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
123 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
124 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom);
125 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
127 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
128 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
130 for (MVT VT : MVT::integer_valuetypes()) {
134 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
135 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Legal);
136 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Legal);
137 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand);
139 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
140 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i8, Legal);
141 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i16, Legal);
142 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand);
144 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
145 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i8, Legal);
146 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i16, Legal);
147 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand);
150 for (MVT VT : MVT::integer_vector_valuetypes()) {
151 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i16, Expand);
152 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v16i16, Expand);
155 for (MVT VT : MVT::fp_valuetypes())
156 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
158 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
159 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
160 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Expand);
161 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Expand);
163 setOperationAction(ISD::LOAD, MVT::i1, Custom);
165 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
166 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
167 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
169 // These should use UDIVREM, so set them to expand
170 setOperationAction(ISD::UDIV, MVT::i64, Expand);
171 setOperationAction(ISD::UREM, MVT::i64, Expand);
173 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
174 setOperationAction(ISD::SELECT, MVT::i1, Promote);
176 // We only support LOAD/STORE and vector manipulation ops for vectors
177 // with > 4 elements.
178 for (MVT VT : {MVT::v8i32, MVT::v8f32, MVT::v16i32, MVT::v16f32}) {
179 for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
183 case ISD::BUILD_VECTOR:
185 case ISD::EXTRACT_VECTOR_ELT:
186 case ISD::INSERT_VECTOR_ELT:
187 case ISD::INSERT_SUBVECTOR:
188 case ISD::EXTRACT_SUBVECTOR:
190 case ISD::CONCAT_VECTORS:
191 setOperationAction(Op, VT, Custom);
194 setOperationAction(Op, VT, Expand);
200 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
201 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
202 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
203 setOperationAction(ISD::FRINT, MVT::f64, Legal);
206 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
207 setOperationAction(ISD::FDIV, MVT::f32, Custom);
208 setOperationAction(ISD::FDIV, MVT::f64, Custom);
210 setTargetDAGCombine(ISD::FADD);
211 setTargetDAGCombine(ISD::FSUB);
212 setTargetDAGCombine(ISD::FMINNUM);
213 setTargetDAGCombine(ISD::FMAXNUM);
214 setTargetDAGCombine(ISD::SELECT_CC);
215 setTargetDAGCombine(ISD::SETCC);
216 setTargetDAGCombine(ISD::AND);
217 setTargetDAGCombine(ISD::OR);
218 setTargetDAGCombine(ISD::UINT_TO_FP);
220 // All memory operations. Some folding on the pointer operand is done to help
221 // matching the constant offsets in the addressing modes.
222 setTargetDAGCombine(ISD::LOAD);
223 setTargetDAGCombine(ISD::STORE);
224 setTargetDAGCombine(ISD::ATOMIC_LOAD);
225 setTargetDAGCombine(ISD::ATOMIC_STORE);
226 setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP);
227 setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS);
228 setTargetDAGCombine(ISD::ATOMIC_SWAP);
229 setTargetDAGCombine(ISD::ATOMIC_LOAD_ADD);
230 setTargetDAGCombine(ISD::ATOMIC_LOAD_SUB);
231 setTargetDAGCombine(ISD::ATOMIC_LOAD_AND);
232 setTargetDAGCombine(ISD::ATOMIC_LOAD_OR);
233 setTargetDAGCombine(ISD::ATOMIC_LOAD_XOR);
234 setTargetDAGCombine(ISD::ATOMIC_LOAD_NAND);
235 setTargetDAGCombine(ISD::ATOMIC_LOAD_MIN);
236 setTargetDAGCombine(ISD::ATOMIC_LOAD_MAX);
237 setTargetDAGCombine(ISD::ATOMIC_LOAD_UMIN);
238 setTargetDAGCombine(ISD::ATOMIC_LOAD_UMAX);
240 setSchedulingPreference(Sched::RegPressure);
243 //===----------------------------------------------------------------------===//
244 // TargetLowering queries
245 //===----------------------------------------------------------------------===//
247 bool SITargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &,
249 // SI has some legal vector types, but no legal vector operations. Say no
250 // shuffles are legal in order to prefer scalarizing some vector operations.
254 // FIXME: This really needs an address space argument. The immediate offset
255 // size is different for different sets of memory instruction sets.
257 // The single offset DS instructions have a 16-bit unsigned byte offset.
259 // MUBUF / MTBUF have a 12-bit unsigned byte offset, and additionally can do r +
260 // r + i with addr64. 32-bit has more addressing mode options. Depending on the
261 // resource constant, it can also do (i64 r0) + (i32 r1) * (i14 i).
263 // SMRD instructions have an 8-bit, dword offset.
265 bool SITargetLowering::isLegalAddressingMode(const AddrMode &AM,
267 // No global is ever allowed as a base.
271 // Allow a 16-bit unsigned immediate field, since this is what DS instructions
273 if (!isUInt<16>(AM.BaseOffs))
278 case 0: // "r+i" or just "i", depending on HasBaseReg.
281 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
283 // Otherwise we have r+r or r+i.
286 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
290 default: // Don't allow n * r
297 bool SITargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
300 bool *IsFast) const {
304 // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
305 // which isn't a simple VT.
306 if (!VT.isSimple() || VT == MVT::Other)
309 // TODO - CI+ supports unaligned memory accesses, but this requires driver
312 // XXX - The only mention I see of this in the ISA manual is for LDS direct
313 // reads the "byte address and must be dword aligned". Is it also true for the
314 // normal loads and stores?
315 if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS) {
316 // ds_read/write_b64 require 8-byte alignment, but we can do a 4 byte
317 // aligned, 8 byte access in a single operation using ds_read2/write2_b32
318 // with adjacent offsets.
319 return Align % 4 == 0;
322 // Smaller than dword value must be aligned.
323 // FIXME: This should be allowed on CI+
324 if (VT.bitsLT(MVT::i32))
327 // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
328 // byte-address are ignored, thus forcing Dword alignment.
329 // This applies to private, global, and constant memory.
333 return VT.bitsGT(MVT::i32) && Align % 4 == 0;
336 EVT SITargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
337 unsigned SrcAlign, bool IsMemset,
340 MachineFunction &MF) const {
341 // FIXME: Should account for address space here.
343 // The default fallback uses the private pointer size as a guess for a type to
344 // use. Make sure we switch these to 64-bit accesses.
346 if (Size >= 16 && DstAlign >= 4) // XXX: Should only do for global
349 if (Size >= 8 && DstAlign >= 4)
356 TargetLoweringBase::LegalizeTypeAction
357 SITargetLowering::getPreferredVectorAction(EVT VT) const {
358 if (VT.getVectorNumElements() != 1 && VT.getScalarType().bitsLE(MVT::i16))
359 return TypeSplitVector;
361 return TargetLoweringBase::getPreferredVectorAction(VT);
364 bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
366 const SIInstrInfo *TII =
367 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
368 return TII->isInlineConstant(Imm);
371 SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
372 SDLoc SL, SDValue Chain,
373 unsigned Offset, bool Signed) const {
374 const DataLayout *DL = getDataLayout();
375 MachineFunction &MF = DAG.getMachineFunction();
376 const SIRegisterInfo *TRI =
377 static_cast<const SIRegisterInfo*>(Subtarget->getRegisterInfo());
378 unsigned InputPtrReg = TRI->getPreloadedValue(MF, SIRegisterInfo::INPUT_PTR);
380 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
382 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
383 PointerType *PtrTy = PointerType::get(Ty, AMDGPUAS::CONSTANT_ADDRESS);
384 SDValue BasePtr = DAG.getCopyFromReg(Chain, SL,
385 MRI.getLiveInVirtReg(InputPtrReg), MVT::i64);
386 SDValue Ptr = DAG.getNode(ISD::ADD, SL, MVT::i64, BasePtr,
387 DAG.getConstant(Offset, SL, MVT::i64));
388 SDValue PtrOffset = DAG.getUNDEF(getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
389 MachinePointerInfo PtrInfo(UndefValue::get(PtrTy));
391 return DAG.getLoad(ISD::UNINDEXED, Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD,
392 VT, SL, Chain, Ptr, PtrOffset, PtrInfo, MemVT,
394 true, // isNonTemporal
396 DL->getABITypeAlignment(Ty)); // Alignment
399 SDValue SITargetLowering::LowerFormalArguments(
400 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
401 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
402 SmallVectorImpl<SDValue> &InVals) const {
403 const SIRegisterInfo *TRI =
404 static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());
406 MachineFunction &MF = DAG.getMachineFunction();
407 FunctionType *FType = MF.getFunction()->getFunctionType();
408 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
410 assert(CallConv == CallingConv::C);
412 SmallVector<ISD::InputArg, 16> Splits;
413 BitVector Skipped(Ins.size());
415 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
416 const ISD::InputArg &Arg = Ins[i];
418 // First check if it's a PS input addr
419 if (Info->getShaderType() == ShaderType::PIXEL && !Arg.Flags.isInReg() &&
420 !Arg.Flags.isByVal()) {
422 assert((PSInputNum <= 15) && "Too many PS inputs!");
425 // We can savely skip PS inputs
431 Info->PSInputAddr |= 1 << PSInputNum++;
434 // Second split vertices into their elements
435 if (Info->getShaderType() != ShaderType::COMPUTE && Arg.VT.isVector()) {
436 ISD::InputArg NewArg = Arg;
437 NewArg.Flags.setSplit();
438 NewArg.VT = Arg.VT.getVectorElementType();
440 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
441 // three or five element vertex only needs three or five registers,
442 // NOT four or eigth.
443 Type *ParamType = FType->getParamType(Arg.getOrigArgIndex());
444 unsigned NumElements = ParamType->getVectorNumElements();
446 for (unsigned j = 0; j != NumElements; ++j) {
447 Splits.push_back(NewArg);
448 NewArg.PartOffset += NewArg.VT.getStoreSize();
451 } else if (Info->getShaderType() != ShaderType::COMPUTE) {
452 Splits.push_back(Arg);
456 SmallVector<CCValAssign, 16> ArgLocs;
457 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
460 // At least one interpolation mode must be enabled or else the GPU will hang.
461 if (Info->getShaderType() == ShaderType::PIXEL &&
462 (Info->PSInputAddr & 0x7F) == 0) {
463 Info->PSInputAddr |= 1;
464 CCInfo.AllocateReg(AMDGPU::VGPR0);
465 CCInfo.AllocateReg(AMDGPU::VGPR1);
468 // The pointer to the list of arguments is stored in SGPR0, SGPR1
469 // The pointer to the scratch buffer is stored in SGPR2, SGPR3
470 if (Info->getShaderType() == ShaderType::COMPUTE) {
471 if (Subtarget->isAmdHsaOS())
472 Info->NumUserSGPRs = 2; // FIXME: Need to support scratch buffers.
474 Info->NumUserSGPRs = 4;
476 unsigned InputPtrReg =
477 TRI->getPreloadedValue(MF, SIRegisterInfo::INPUT_PTR);
478 unsigned InputPtrRegLo =
479 TRI->getPhysRegSubReg(InputPtrReg, &AMDGPU::SReg_32RegClass, 0);
480 unsigned InputPtrRegHi =
481 TRI->getPhysRegSubReg(InputPtrReg, &AMDGPU::SReg_32RegClass, 1);
483 unsigned ScratchPtrReg =
484 TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_PTR);
485 unsigned ScratchPtrRegLo =
486 TRI->getPhysRegSubReg(ScratchPtrReg, &AMDGPU::SReg_32RegClass, 0);
487 unsigned ScratchPtrRegHi =
488 TRI->getPhysRegSubReg(ScratchPtrReg, &AMDGPU::SReg_32RegClass, 1);
490 CCInfo.AllocateReg(InputPtrRegLo);
491 CCInfo.AllocateReg(InputPtrRegHi);
492 CCInfo.AllocateReg(ScratchPtrRegLo);
493 CCInfo.AllocateReg(ScratchPtrRegHi);
494 MF.addLiveIn(InputPtrReg, &AMDGPU::SReg_64RegClass);
495 MF.addLiveIn(ScratchPtrReg, &AMDGPU::SReg_64RegClass);
498 if (Info->getShaderType() == ShaderType::COMPUTE) {
499 getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
503 AnalyzeFormalArguments(CCInfo, Splits);
505 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
507 const ISD::InputArg &Arg = Ins[i];
509 InVals.push_back(DAG.getUNDEF(Arg.VT));
513 CCValAssign &VA = ArgLocs[ArgIdx++];
514 MVT VT = VA.getLocVT();
518 EVT MemVT = Splits[i].VT;
519 const unsigned Offset = 36 + VA.getLocMemOffset();
520 // The first 36 bytes of the input buffer contains information about
521 // thread group and global sizes.
522 SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, DAG.getRoot(),
523 Offset, Ins[i].Flags.isSExt());
525 const PointerType *ParamTy =
526 dyn_cast<PointerType>(FType->getParamType(Ins[i].getOrigArgIndex()));
527 if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS &&
528 ParamTy && ParamTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
529 // On SI local pointers are just offsets into LDS, so they are always
530 // less than 16-bits. On CI and newer they could potentially be
531 // real pointers, so we can't guarantee their size.
532 Arg = DAG.getNode(ISD::AssertZext, DL, Arg.getValueType(), Arg,
533 DAG.getValueType(MVT::i16));
536 InVals.push_back(Arg);
537 Info->ABIArgOffset = Offset + MemVT.getStoreSize();
540 assert(VA.isRegLoc() && "Parameter must be in a register!");
542 unsigned Reg = VA.getLocReg();
544 if (VT == MVT::i64) {
545 // For now assume it is a pointer
546 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
547 &AMDGPU::SReg_64RegClass);
548 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
549 InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
553 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
555 Reg = MF.addLiveIn(Reg, RC);
556 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
558 if (Arg.VT.isVector()) {
560 // Build a vector from the registers
561 Type *ParamType = FType->getParamType(Arg.getOrigArgIndex());
562 unsigned NumElements = ParamType->getVectorNumElements();
564 SmallVector<SDValue, 4> Regs;
566 for (unsigned j = 1; j != NumElements; ++j) {
567 Reg = ArgLocs[ArgIdx++].getLocReg();
568 Reg = MF.addLiveIn(Reg, RC);
569 Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
572 // Fill up the missing vector elements
573 NumElements = Arg.VT.getVectorNumElements() - NumElements;
574 Regs.append(NumElements, DAG.getUNDEF(VT));
576 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT, Regs));
580 InVals.push_back(Val);
583 if (Info->getShaderType() != ShaderType::COMPUTE) {
584 unsigned ScratchIdx = CCInfo.getFirstUnallocated(ArrayRef<MCPhysReg>(
585 AMDGPU::SGPR_32RegClass.begin(), AMDGPU::SGPR_32RegClass.getNumRegs()));
586 Info->ScratchOffsetReg = AMDGPU::SGPR_32RegClass.getRegister(ScratchIdx);
591 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
592 MachineInstr * MI, MachineBasicBlock * BB) const {
594 MachineBasicBlock::iterator I = *MI;
595 const SIInstrInfo *TII =
596 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
598 switch (MI->getOpcode()) {
600 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
603 case AMDGPU::SI_RegisterStorePseudo: {
604 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
605 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
606 MachineInstrBuilder MIB =
607 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::SI_RegisterStore),
609 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i)
610 MIB.addOperand(MI->getOperand(i));
612 MI->eraseFromParent();
619 bool SITargetLowering::enableAggressiveFMAFusion(EVT VT) const {
620 // This currently forces unfolding various combinations of fsub into fma with
621 // free fneg'd operands. As long as we have fast FMA (controlled by
622 // isFMAFasterThanFMulAndFAdd), we should perform these.
624 // When fma is quarter rate, for f64 where add / sub are at best half rate,
625 // most of these combines appear to be cycle neutral but save on instruction
626 // count / code size.
630 EVT SITargetLowering::getSetCCResultType(LLVMContext &Ctx, EVT VT) const {
631 if (!VT.isVector()) {
634 return EVT::getVectorVT(Ctx, MVT::i1, VT.getVectorNumElements());
637 MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
641 // Answering this is somewhat tricky and depends on the specific device which
642 // have different rates for fma or all f64 operations.
644 // v_fma_f64 and v_mul_f64 always take the same number of cycles as each other
645 // regardless of which device (although the number of cycles differs between
646 // devices), so it is always profitable for f64.
648 // v_fma_f32 takes 4 or 16 cycles depending on the device, so it is profitable
649 // only on full rate devices. Normally, we should prefer selecting v_mad_f32
650 // which we can always do even without fused FP ops since it returns the same
651 // result as the separate operations and since it is always full
652 // rate. Therefore, we lie and report that it is not faster for f32. v_mad_f32
653 // however does not support denormals, so we do report fma as faster if we have
654 // a fast fma device and require denormals.
656 bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
657 VT = VT.getScalarType();
662 switch (VT.getSimpleVT().SimpleTy) {
664 // This is as fast on some subtargets. However, we always have full rate f32
665 // mad available which returns the same result as the separate operations
666 // which we should prefer over fma. We can't use this if we want to support
667 // denormals, so only report this in these cases.
668 return Subtarget->hasFP32Denormals() && Subtarget->hasFastFMAF32();
678 //===----------------------------------------------------------------------===//
679 // Custom DAG Lowering Operations
680 //===----------------------------------------------------------------------===//
682 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
683 switch (Op.getOpcode()) {
684 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
685 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
686 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
688 SDValue Result = LowerLOAD(Op, DAG);
689 assert((!Result.getNode() ||
690 Result.getNode()->getNumValues() == 2) &&
691 "Load should return a value and a chain");
697 return LowerTrig(Op, DAG);
698 case ISD::SELECT: return LowerSELECT(Op, DAG);
699 case ISD::FDIV: return LowerFDIV(Op, DAG);
700 case ISD::STORE: return LowerSTORE(Op, DAG);
701 case ISD::GlobalAddress: {
702 MachineFunction &MF = DAG.getMachineFunction();
703 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
704 return LowerGlobalAddress(MFI, Op, DAG);
706 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
707 case ISD::INTRINSIC_VOID: return LowerINTRINSIC_VOID(Op, DAG);
712 /// \brief Helper function for LowerBRCOND
713 static SDNode *findUser(SDValue Value, unsigned Opcode) {
715 SDNode *Parent = Value.getNode();
716 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
719 if (I.getUse().get() != Value)
722 if (I->getOpcode() == Opcode)
728 SDValue SITargetLowering::LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const {
730 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Op);
731 unsigned FrameIndex = FINode->getIndex();
733 return DAG.getTargetFrameIndex(FrameIndex, MVT::i32);
736 /// This transforms the control flow intrinsics to get the branch destination as
737 /// last parameter, also switches branch target with BR if the need arise
738 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
739 SelectionDAG &DAG) const {
743 SDNode *Intr = BRCOND.getOperand(1).getNode();
744 SDValue Target = BRCOND.getOperand(2);
745 SDNode *BR = nullptr;
747 if (Intr->getOpcode() == ISD::SETCC) {
748 // As long as we negate the condition everything is fine
749 SDNode *SetCC = Intr;
750 assert(SetCC->getConstantOperandVal(1) == 1);
751 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
753 Intr = SetCC->getOperand(0).getNode();
756 // Get the target from BR if we don't negate the condition
757 BR = findUser(BRCOND, ISD::BR);
758 Target = BR->getOperand(1);
761 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
763 // Build the result and
764 ArrayRef<EVT> Res(Intr->value_begin() + 1, Intr->value_end());
766 // operands of the new intrinsic call
767 SmallVector<SDValue, 4> Ops;
768 Ops.push_back(BRCOND.getOperand(0));
769 Ops.append(Intr->op_begin() + 1, Intr->op_end());
770 Ops.push_back(Target);
772 // build the new intrinsic call
773 SDNode *Result = DAG.getNode(
774 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
775 DAG.getVTList(Res), Ops).getNode();
778 // Give the branch instruction our target
783 SDValue NewBR = DAG.getNode(ISD::BR, DL, BR->getVTList(), Ops);
784 DAG.ReplaceAllUsesWith(BR, NewBR.getNode());
785 BR = NewBR.getNode();
788 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
790 // Copy the intrinsic results to registers
791 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
792 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
796 Chain = DAG.getCopyToReg(
798 CopyToReg->getOperand(1),
799 SDValue(Result, i - 1),
802 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
805 // Remove the old intrinsic from the chain
806 DAG.ReplaceAllUsesOfValueWith(
807 SDValue(Intr, Intr->getNumValues() - 1),
808 Intr->getOperand(0));
813 SDValue SITargetLowering::LowerGlobalAddress(AMDGPUMachineFunction *MFI,
815 SelectionDAG &DAG) const {
816 GlobalAddressSDNode *GSD = cast<GlobalAddressSDNode>(Op);
818 if (GSD->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS)
819 return AMDGPUTargetLowering::LowerGlobalAddress(MFI, Op, DAG);
822 const GlobalValue *GV = GSD->getGlobal();
823 MVT PtrVT = getPointerTy(GSD->getAddressSpace());
825 SDValue Ptr = DAG.getNode(AMDGPUISD::CONST_DATA_PTR, DL, PtrVT);
826 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32);
828 SDValue PtrLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Ptr,
829 DAG.getConstant(0, DL, MVT::i32));
830 SDValue PtrHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Ptr,
831 DAG.getConstant(1, DL, MVT::i32));
833 SDValue Lo = DAG.getNode(ISD::ADDC, DL, DAG.getVTList(MVT::i32, MVT::Glue),
835 SDValue Hi = DAG.getNode(ISD::ADDE, DL, DAG.getVTList(MVT::i32, MVT::Glue),
836 PtrHi, DAG.getConstant(0, DL, MVT::i32),
837 SDValue(Lo.getNode(), 1));
838 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi);
841 SDValue SITargetLowering::copyToM0(SelectionDAG &DAG, SDValue Chain, SDLoc DL,
843 // We can't use CopyToReg, because MachineCSE won't combine COPY instructions,
844 // so we will end up with redundant moves to m0.
846 // We can't use S_MOV_B32, because there is no way to specify m0 as the
847 // destination register.
849 // We have to use them both. Machine cse will combine all the S_MOV_B32
850 // instructions and the register coalescer eliminate the extra copies.
851 SDNode *M0 = DAG.getMachineNode(AMDGPU::S_MOV_B32, DL, V.getValueType(), V);
852 return DAG.getCopyToReg(Chain, DL, DAG.getRegister(AMDGPU::M0, MVT::i32),
853 SDValue(M0, 0), SDValue()); // Glue
854 // A Null SDValue creates
858 SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
859 SelectionDAG &DAG) const {
860 MachineFunction &MF = DAG.getMachineFunction();
861 const SIRegisterInfo *TRI =
862 static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());
864 EVT VT = Op.getValueType();
866 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
868 switch (IntrinsicID) {
869 case Intrinsic::r600_read_ngroups_x:
870 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
871 SI::KernelInputOffsets::NGROUPS_X, false);
872 case Intrinsic::r600_read_ngroups_y:
873 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
874 SI::KernelInputOffsets::NGROUPS_Y, false);
875 case Intrinsic::r600_read_ngroups_z:
876 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
877 SI::KernelInputOffsets::NGROUPS_Z, false);
878 case Intrinsic::r600_read_global_size_x:
879 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
880 SI::KernelInputOffsets::GLOBAL_SIZE_X, false);
881 case Intrinsic::r600_read_global_size_y:
882 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
883 SI::KernelInputOffsets::GLOBAL_SIZE_Y, false);
884 case Intrinsic::r600_read_global_size_z:
885 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
886 SI::KernelInputOffsets::GLOBAL_SIZE_Z, false);
887 case Intrinsic::r600_read_local_size_x:
888 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
889 SI::KernelInputOffsets::LOCAL_SIZE_X, false);
890 case Intrinsic::r600_read_local_size_y:
891 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
892 SI::KernelInputOffsets::LOCAL_SIZE_Y, false);
893 case Intrinsic::r600_read_local_size_z:
894 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
895 SI::KernelInputOffsets::LOCAL_SIZE_Z, false);
897 case Intrinsic::AMDGPU_read_workdim:
898 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
899 MF.getInfo<SIMachineFunctionInfo>()->ABIArgOffset,
902 case Intrinsic::r600_read_tgid_x:
903 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
904 TRI->getPreloadedValue(MF, SIRegisterInfo::TGID_X), VT);
905 case Intrinsic::r600_read_tgid_y:
906 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
907 TRI->getPreloadedValue(MF, SIRegisterInfo::TGID_Y), VT);
908 case Intrinsic::r600_read_tgid_z:
909 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
910 TRI->getPreloadedValue(MF, SIRegisterInfo::TGID_Z), VT);
911 case Intrinsic::r600_read_tidig_x:
912 return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
913 TRI->getPreloadedValue(MF, SIRegisterInfo::TIDIG_X), VT);
914 case Intrinsic::r600_read_tidig_y:
915 return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
916 TRI->getPreloadedValue(MF, SIRegisterInfo::TIDIG_Y), VT);
917 case Intrinsic::r600_read_tidig_z:
918 return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
919 TRI->getPreloadedValue(MF, SIRegisterInfo::TIDIG_Z), VT);
920 case AMDGPUIntrinsic::SI_load_const: {
926 MachineMemOperand *MMO = MF.getMachineMemOperand(
927 MachinePointerInfo(),
928 MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
929 VT.getStoreSize(), 4);
930 return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
931 Op->getVTList(), Ops, VT, MMO);
933 case AMDGPUIntrinsic::SI_sample:
934 return LowerSampleIntrinsic(AMDGPUISD::SAMPLE, Op, DAG);
935 case AMDGPUIntrinsic::SI_sampleb:
936 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEB, Op, DAG);
937 case AMDGPUIntrinsic::SI_sampled:
938 return LowerSampleIntrinsic(AMDGPUISD::SAMPLED, Op, DAG);
939 case AMDGPUIntrinsic::SI_samplel:
940 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEL, Op, DAG);
941 case AMDGPUIntrinsic::SI_vs_load_input:
942 return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
947 case AMDGPUIntrinsic::AMDGPU_fract:
948 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
949 return DAG.getNode(ISD::FSUB, DL, VT, Op.getOperand(1),
950 DAG.getNode(ISD::FFLOOR, DL, VT, Op.getOperand(1)));
951 case AMDGPUIntrinsic::SI_fs_constant: {
952 SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(3));
953 SDValue Glue = M0.getValue(1);
954 return DAG.getNode(AMDGPUISD::INTERP_MOV, DL, MVT::f32,
955 DAG.getConstant(2, DL, MVT::i32), // P0
956 Op.getOperand(1), Op.getOperand(2), Glue);
958 case AMDGPUIntrinsic::SI_fs_interp: {
959 SDValue IJ = Op.getOperand(4);
960 SDValue I = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, IJ,
961 DAG.getConstant(0, DL, MVT::i32));
962 SDValue J = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, IJ,
963 DAG.getConstant(1, DL, MVT::i32));
964 SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(3));
965 SDValue Glue = M0.getValue(1);
966 SDValue P1 = DAG.getNode(AMDGPUISD::INTERP_P1, DL,
967 DAG.getVTList(MVT::f32, MVT::Glue),
968 I, Op.getOperand(1), Op.getOperand(2), Glue);
969 Glue = SDValue(P1.getNode(), 1);
970 return DAG.getNode(AMDGPUISD::INTERP_P2, DL, MVT::f32, P1, J,
971 Op.getOperand(1), Op.getOperand(2), Glue);
974 return AMDGPUTargetLowering::LowerOperation(Op, DAG);
978 SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
979 SelectionDAG &DAG) const {
980 MachineFunction &MF = DAG.getMachineFunction();
982 SDValue Chain = Op.getOperand(0);
983 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
985 switch (IntrinsicID) {
986 case AMDGPUIntrinsic::SI_sendmsg: {
987 Chain = copyToM0(DAG, Chain, DL, Op.getOperand(3));
988 SDValue Glue = Chain.getValue(1);
989 return DAG.getNode(AMDGPUISD::SENDMSG, DL, MVT::Other, Chain,
990 Op.getOperand(2), Glue);
992 case AMDGPUIntrinsic::SI_tbuffer_store: {
1010 EVT VT = Op.getOperand(3).getValueType();
1012 MachineMemOperand *MMO = MF.getMachineMemOperand(
1013 MachinePointerInfo(),
1014 MachineMemOperand::MOStore,
1015 VT.getStoreSize(), 4);
1016 return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL,
1017 Op->getVTList(), Ops, VT, MMO);
1024 SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1026 LoadSDNode *Load = cast<LoadSDNode>(Op);
1028 if (Op.getValueType().isVector()) {
1029 assert(Op.getValueType().getVectorElementType() == MVT::i32 &&
1030 "Custom lowering for non-i32 vectors hasn't been implemented.");
1031 unsigned NumElements = Op.getValueType().getVectorNumElements();
1032 assert(NumElements != 2 && "v2 loads are supported for all address spaces.");
1033 switch (Load->getAddressSpace()) {
1035 case AMDGPUAS::GLOBAL_ADDRESS:
1036 case AMDGPUAS::PRIVATE_ADDRESS:
1037 // v4 loads are supported for private and global memory.
1038 if (NumElements <= 4)
1041 case AMDGPUAS::LOCAL_ADDRESS:
1042 return ScalarizeVectorLoad(Op, DAG);
1046 return AMDGPUTargetLowering::LowerLOAD(Op, DAG);
1049 SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
1051 SelectionDAG &DAG) const {
1052 return DAG.getNode(Opcode, SDLoc(Op), Op.getValueType(), Op.getOperand(1),
1058 SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
1059 if (Op.getValueType() != MVT::i64)
1063 SDValue Cond = Op.getOperand(0);
1065 SDValue Zero = DAG.getConstant(0, DL, MVT::i32);
1066 SDValue One = DAG.getConstant(1, DL, MVT::i32);
1068 SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1));
1069 SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2));
1071 SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero);
1072 SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero);
1074 SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1);
1076 SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One);
1077 SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One);
1079 SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1);
1081 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i32, Lo, Hi);
1082 return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
1085 // Catch division cases where we can use shortcuts with rcp and rsq
1087 SDValue SITargetLowering::LowerFastFDIV(SDValue Op, SelectionDAG &DAG) const {
1089 SDValue LHS = Op.getOperand(0);
1090 SDValue RHS = Op.getOperand(1);
1091 EVT VT = Op.getValueType();
1092 bool Unsafe = DAG.getTarget().Options.UnsafeFPMath;
1094 if (const ConstantFPSDNode *CLHS = dyn_cast<ConstantFPSDNode>(LHS)) {
1095 if ((Unsafe || (VT == MVT::f32 && !Subtarget->hasFP32Denormals())) &&
1096 CLHS->isExactlyValue(1.0)) {
1097 // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to
1098 // the CI documentation has a worst case error of 1 ulp.
1099 // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to
1100 // use it as long as we aren't trying to use denormals.
1102 // 1.0 / sqrt(x) -> rsq(x)
1104 // XXX - Is UnsafeFPMath sufficient to do this for f64? The maximum ULP
1105 // error seems really high at 2^29 ULP.
1106 if (RHS.getOpcode() == ISD::FSQRT)
1107 return DAG.getNode(AMDGPUISD::RSQ, SL, VT, RHS.getOperand(0));
1109 // 1.0 / x -> rcp(x)
1110 return DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
1115 // Turn into multiply by the reciprocal.
1116 // x / y -> x * (1.0 / y)
1117 SDValue Recip = DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
1118 return DAG.getNode(ISD::FMUL, SL, VT, LHS, Recip);
1124 SDValue SITargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const {
1125 SDValue FastLowered = LowerFastFDIV(Op, DAG);
1126 if (FastLowered.getNode())
1129 // This uses v_rcp_f32 which does not handle denormals. Let this hit a
1130 // selection error for now rather than do something incorrect.
1131 if (Subtarget->hasFP32Denormals())
1135 SDValue LHS = Op.getOperand(0);
1136 SDValue RHS = Op.getOperand(1);
1138 SDValue r1 = DAG.getNode(ISD::FABS, SL, MVT::f32, RHS);
1140 const APFloat K0Val(BitsToFloat(0x6f800000));
1141 const SDValue K0 = DAG.getConstantFP(K0Val, SL, MVT::f32);
1143 const APFloat K1Val(BitsToFloat(0x2f800000));
1144 const SDValue K1 = DAG.getConstantFP(K1Val, SL, MVT::f32);
1146 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f32);
1148 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f32);
1150 SDValue r2 = DAG.getSetCC(SL, SetCCVT, r1, K0, ISD::SETOGT);
1152 SDValue r3 = DAG.getNode(ISD::SELECT, SL, MVT::f32, r2, K1, One);
1154 r1 = DAG.getNode(ISD::FMUL, SL, MVT::f32, RHS, r3);
1156 SDValue r0 = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f32, r1);
1158 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f32, LHS, r0);
1160 return DAG.getNode(ISD::FMUL, SL, MVT::f32, r3, Mul);
1163 SDValue SITargetLowering::LowerFDIV64(SDValue Op, SelectionDAG &DAG) const {
1164 if (DAG.getTarget().Options.UnsafeFPMath)
1165 return LowerFastFDIV(Op, DAG);
1168 SDValue X = Op.getOperand(0);
1169 SDValue Y = Op.getOperand(1);
1171 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64);
1173 SDVTList ScaleVT = DAG.getVTList(MVT::f64, MVT::i1);
1175 SDValue DivScale0 = DAG.getNode(AMDGPUISD::DIV_SCALE, SL, ScaleVT, Y, Y, X);
1177 SDValue NegDivScale0 = DAG.getNode(ISD::FNEG, SL, MVT::f64, DivScale0);
1179 SDValue Rcp = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f64, DivScale0);
1181 SDValue Fma0 = DAG.getNode(ISD::FMA, SL, MVT::f64, NegDivScale0, Rcp, One);
1183 SDValue Fma1 = DAG.getNode(ISD::FMA, SL, MVT::f64, Rcp, Fma0, Rcp);
1185 SDValue Fma2 = DAG.getNode(ISD::FMA, SL, MVT::f64, NegDivScale0, Fma1, One);
1187 SDValue DivScale1 = DAG.getNode(AMDGPUISD::DIV_SCALE, SL, ScaleVT, X, Y, X);
1189 SDValue Fma3 = DAG.getNode(ISD::FMA, SL, MVT::f64, Fma1, Fma2, Fma1);
1190 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, DivScale1, Fma3);
1192 SDValue Fma4 = DAG.getNode(ISD::FMA, SL, MVT::f64,
1193 NegDivScale0, Mul, DivScale1);
1197 if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS) {
1198 // Workaround a hardware bug on SI where the condition output from div_scale
1201 const SDValue Hi = DAG.getConstant(1, SL, MVT::i32);
1203 // Figure out if the scale to use for div_fmas.
1204 SDValue NumBC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X);
1205 SDValue DenBC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Y);
1206 SDValue Scale0BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, DivScale0);
1207 SDValue Scale1BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, DivScale1);
1209 SDValue NumHi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, NumBC, Hi);
1210 SDValue DenHi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, DenBC, Hi);
1213 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Scale0BC, Hi);
1215 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Scale1BC, Hi);
1217 SDValue CmpDen = DAG.getSetCC(SL, MVT::i1, DenHi, Scale0Hi, ISD::SETEQ);
1218 SDValue CmpNum = DAG.getSetCC(SL, MVT::i1, NumHi, Scale1Hi, ISD::SETEQ);
1219 Scale = DAG.getNode(ISD::XOR, SL, MVT::i1, CmpNum, CmpDen);
1221 Scale = DivScale1.getValue(1);
1224 SDValue Fmas = DAG.getNode(AMDGPUISD::DIV_FMAS, SL, MVT::f64,
1225 Fma4, Fma3, Mul, Scale);
1227 return DAG.getNode(AMDGPUISD::DIV_FIXUP, SL, MVT::f64, Fmas, Y, X);
1230 SDValue SITargetLowering::LowerFDIV(SDValue Op, SelectionDAG &DAG) const {
1231 EVT VT = Op.getValueType();
1234 return LowerFDIV32(Op, DAG);
1237 return LowerFDIV64(Op, DAG);
1239 llvm_unreachable("Unexpected type for fdiv");
1242 SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1244 StoreSDNode *Store = cast<StoreSDNode>(Op);
1245 EVT VT = Store->getMemoryVT();
1247 // These stores are legal.
1248 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) {
1249 if (VT.isVector() && VT.getVectorNumElements() > 4)
1250 return ScalarizeVectorStore(Op, DAG);
1254 SDValue Ret = AMDGPUTargetLowering::LowerSTORE(Op, DAG);
1258 if (VT.isVector() && VT.getVectorNumElements() >= 8)
1259 return ScalarizeVectorStore(Op, DAG);
1262 return DAG.getTruncStore(Store->getChain(), DL,
1263 DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32),
1264 Store->getBasePtr(), MVT::i1, Store->getMemOperand());
1269 SDValue SITargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const {
1271 EVT VT = Op.getValueType();
1272 SDValue Arg = Op.getOperand(0);
1273 SDValue FractPart = DAG.getNode(AMDGPUISD::FRACT, DL, VT,
1274 DAG.getNode(ISD::FMUL, DL, VT, Arg,
1275 DAG.getConstantFP(0.5/M_PI, DL,
1278 switch (Op.getOpcode()) {
1280 return DAG.getNode(AMDGPUISD::COS_HW, SDLoc(Op), VT, FractPart);
1282 return DAG.getNode(AMDGPUISD::SIN_HW, SDLoc(Op), VT, FractPart);
1284 llvm_unreachable("Wrong trig opcode");
1288 //===----------------------------------------------------------------------===//
1289 // Custom DAG optimizations
1290 //===----------------------------------------------------------------------===//
1292 SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
1293 DAGCombinerInfo &DCI) const {
1294 EVT VT = N->getValueType(0);
1295 EVT ScalarVT = VT.getScalarType();
1296 if (ScalarVT != MVT::f32)
1299 SelectionDAG &DAG = DCI.DAG;
1302 SDValue Src = N->getOperand(0);
1303 EVT SrcVT = Src.getValueType();
1305 // TODO: We could try to match extracting the higher bytes, which would be
1306 // easier if i8 vectors weren't promoted to i32 vectors, particularly after
1307 // types are legalized. v4i8 -> v4f32 is probably the only case to worry
1308 // about in practice.
1309 if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) {
1310 if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) {
1311 SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src);
1312 DCI.AddToWorklist(Cvt.getNode());
1317 // We are primarily trying to catch operations on illegal vector types
1318 // before they are expanded.
1319 // For scalars, we can use the more flexible method of checking masked bits
1320 // after legalization.
1321 if (!DCI.isBeforeLegalize() ||
1322 !SrcVT.isVector() ||
1323 SrcVT.getVectorElementType() != MVT::i8) {
1327 assert(DCI.isBeforeLegalize() && "Unexpected legal type");
1329 // Weird sized vectors are a pain to handle, but we know 3 is really the same
1331 unsigned NElts = SrcVT.getVectorNumElements();
1332 if (!SrcVT.isSimple() && NElts != 3)
1335 // Handle v4i8 -> v4f32 extload. Replace the v4i8 with a legal i32 load to
1336 // prevent a mess from expanding to v4i32 and repacking.
1337 if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse()) {
1338 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), SrcVT);
1339 EVT RegVT = getEquivalentLoadRegType(*DAG.getContext(), SrcVT);
1340 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f32, NElts);
1341 LoadSDNode *Load = cast<LoadSDNode>(Src);
1343 unsigned AS = Load->getAddressSpace();
1344 unsigned Align = Load->getAlignment();
1345 Type *Ty = LoadVT.getTypeForEVT(*DAG.getContext());
1346 unsigned ABIAlignment = getDataLayout()->getABITypeAlignment(Ty);
1348 // Don't try to replace the load if we have to expand it due to alignment
1349 // problems. Otherwise we will end up scalarizing the load, and trying to
1350 // repack into the vector for no real reason.
1351 if (Align < ABIAlignment &&
1352 !allowsMisalignedMemoryAccesses(LoadVT, AS, Align, nullptr)) {
1356 SDValue NewLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegVT,
1360 Load->getMemOperand());
1362 // Make sure successors of the original load stay after it by updating
1363 // them to use the new Chain.
1364 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), NewLoad.getValue(1));
1366 SmallVector<SDValue, 4> Elts;
1367 if (RegVT.isVector())
1368 DAG.ExtractVectorElements(NewLoad, Elts);
1370 Elts.push_back(NewLoad);
1372 SmallVector<SDValue, 4> Ops;
1374 unsigned EltIdx = 0;
1375 for (SDValue Elt : Elts) {
1376 unsigned ComponentsInElt = std::min(4u, NElts - 4 * EltIdx);
1377 for (unsigned I = 0; I < ComponentsInElt; ++I) {
1378 unsigned Opc = AMDGPUISD::CVT_F32_UBYTE0 + I;
1379 SDValue Cvt = DAG.getNode(Opc, DL, MVT::f32, Elt);
1380 DCI.AddToWorklist(Cvt.getNode());
1387 assert(Ops.size() == NElts);
1389 return DAG.getNode(ISD::BUILD_VECTOR, DL, FloatVT, Ops);
1395 /// \brief Return true if the given offset Size in bytes can be folded into
1396 /// the immediate offsets of a memory instruction for the given address space.
1397 static bool canFoldOffset(unsigned OffsetSize, unsigned AS,
1398 const AMDGPUSubtarget &STI) {
1400 case AMDGPUAS::GLOBAL_ADDRESS: {
1401 // MUBUF instructions a 12-bit offset in bytes.
1402 return isUInt<12>(OffsetSize);
1404 case AMDGPUAS::CONSTANT_ADDRESS: {
1405 // SMRD instructions have an 8-bit offset in dwords on SI and
1406 // a 20-bit offset in bytes on VI.
1407 if (STI.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1408 return isUInt<20>(OffsetSize);
1410 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
1412 case AMDGPUAS::LOCAL_ADDRESS:
1413 case AMDGPUAS::REGION_ADDRESS: {
1414 // The single offset versions have a 16-bit offset in bytes.
1415 return isUInt<16>(OffsetSize);
1417 case AMDGPUAS::PRIVATE_ADDRESS:
1418 // Indirect register addressing does not use any offsets.
1424 // (shl (add x, c1), c2) -> add (shl x, c2), (shl c1, c2)
1426 // This is a variant of
1427 // (mul (add x, c1), c2) -> add (mul x, c2), (mul c1, c2),
1429 // The normal DAG combiner will do this, but only if the add has one use since
1430 // that would increase the number of instructions.
1432 // This prevents us from seeing a constant offset that can be folded into a
1433 // memory instruction's addressing mode. If we know the resulting add offset of
1434 // a pointer can be folded into an addressing offset, we can replace the pointer
1435 // operand with the add of new constant offset. This eliminates one of the uses,
1436 // and may allow the remaining use to also be simplified.
1438 SDValue SITargetLowering::performSHLPtrCombine(SDNode *N,
1440 DAGCombinerInfo &DCI) const {
1441 SDValue N0 = N->getOperand(0);
1442 SDValue N1 = N->getOperand(1);
1444 if (N0.getOpcode() != ISD::ADD)
1447 const ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N1);
1451 const ConstantSDNode *CAdd = dyn_cast<ConstantSDNode>(N0.getOperand(1));
1455 // If the resulting offset is too large, we can't fold it into the addressing
1457 APInt Offset = CAdd->getAPIntValue() << CN1->getAPIntValue();
1458 if (!canFoldOffset(Offset.getZExtValue(), AddrSpace, *Subtarget))
1461 SelectionDAG &DAG = DCI.DAG;
1463 EVT VT = N->getValueType(0);
1465 SDValue ShlX = DAG.getNode(ISD::SHL, SL, VT, N0.getOperand(0), N1);
1466 SDValue COffset = DAG.getConstant(Offset, SL, MVT::i32);
1468 return DAG.getNode(ISD::ADD, SL, VT, ShlX, COffset);
1471 SDValue SITargetLowering::performAndCombine(SDNode *N,
1472 DAGCombinerInfo &DCI) const {
1473 if (DCI.isBeforeLegalize())
1476 SelectionDAG &DAG = DCI.DAG;
1478 // (and (fcmp ord x, x), (fcmp une (fabs x), inf)) ->
1479 // fp_class x, ~(s_nan | q_nan | n_infinity | p_infinity)
1480 SDValue LHS = N->getOperand(0);
1481 SDValue RHS = N->getOperand(1);
1483 if (LHS.getOpcode() == ISD::SETCC &&
1484 RHS.getOpcode() == ISD::SETCC) {
1485 ISD::CondCode LCC = cast<CondCodeSDNode>(LHS.getOperand(2))->get();
1486 ISD::CondCode RCC = cast<CondCodeSDNode>(RHS.getOperand(2))->get();
1488 SDValue X = LHS.getOperand(0);
1489 SDValue Y = RHS.getOperand(0);
1490 if (Y.getOpcode() != ISD::FABS || Y.getOperand(0) != X)
1493 if (LCC == ISD::SETO) {
1494 if (X != LHS.getOperand(1))
1497 if (RCC == ISD::SETUNE) {
1498 const ConstantFPSDNode *C1 = dyn_cast<ConstantFPSDNode>(RHS.getOperand(1));
1499 if (!C1 || !C1->isInfinity() || C1->isNegative())
1502 const uint32_t Mask = SIInstrFlags::N_NORMAL |
1503 SIInstrFlags::N_SUBNORMAL |
1504 SIInstrFlags::N_ZERO |
1505 SIInstrFlags::P_ZERO |
1506 SIInstrFlags::P_SUBNORMAL |
1507 SIInstrFlags::P_NORMAL;
1509 static_assert(((~(SIInstrFlags::S_NAN |
1510 SIInstrFlags::Q_NAN |
1511 SIInstrFlags::N_INFINITY |
1512 SIInstrFlags::P_INFINITY)) & 0x3ff) == Mask,
1516 return DAG.getNode(AMDGPUISD::FP_CLASS, DL, MVT::i1,
1517 X, DAG.getConstant(Mask, DL, MVT::i32));
1525 SDValue SITargetLowering::performOrCombine(SDNode *N,
1526 DAGCombinerInfo &DCI) const {
1527 SelectionDAG &DAG = DCI.DAG;
1528 SDValue LHS = N->getOperand(0);
1529 SDValue RHS = N->getOperand(1);
1531 // or (fp_class x, c1), (fp_class x, c2) -> fp_class x, (c1 | c2)
1532 if (LHS.getOpcode() == AMDGPUISD::FP_CLASS &&
1533 RHS.getOpcode() == AMDGPUISD::FP_CLASS) {
1534 SDValue Src = LHS.getOperand(0);
1535 if (Src != RHS.getOperand(0))
1538 const ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(LHS.getOperand(1));
1539 const ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(RHS.getOperand(1));
1543 // Only 10 bits are used.
1544 static const uint32_t MaxMask = 0x3ff;
1546 uint32_t NewMask = (CLHS->getZExtValue() | CRHS->getZExtValue()) & MaxMask;
1548 return DAG.getNode(AMDGPUISD::FP_CLASS, DL, MVT::i1,
1549 Src, DAG.getConstant(NewMask, DL, MVT::i32));
1555 SDValue SITargetLowering::performClassCombine(SDNode *N,
1556 DAGCombinerInfo &DCI) const {
1557 SelectionDAG &DAG = DCI.DAG;
1558 SDValue Mask = N->getOperand(1);
1560 // fp_class x, 0 -> false
1561 if (const ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(Mask)) {
1562 if (CMask->isNullValue())
1563 return DAG.getConstant(0, SDLoc(N), MVT::i1);
1569 static unsigned minMaxOpcToMin3Max3Opc(unsigned Opc) {
1572 return AMDGPUISD::FMAX3;
1573 case AMDGPUISD::SMAX:
1574 return AMDGPUISD::SMAX3;
1575 case AMDGPUISD::UMAX:
1576 return AMDGPUISD::UMAX3;
1578 return AMDGPUISD::FMIN3;
1579 case AMDGPUISD::SMIN:
1580 return AMDGPUISD::SMIN3;
1581 case AMDGPUISD::UMIN:
1582 return AMDGPUISD::UMIN3;
1584 llvm_unreachable("Not a min/max opcode");
1588 SDValue SITargetLowering::performMin3Max3Combine(SDNode *N,
1589 DAGCombinerInfo &DCI) const {
1590 SelectionDAG &DAG = DCI.DAG;
1592 unsigned Opc = N->getOpcode();
1593 SDValue Op0 = N->getOperand(0);
1594 SDValue Op1 = N->getOperand(1);
1596 // Only do this if the inner op has one use since this will just increases
1597 // register pressure for no benefit.
1599 // max(max(a, b), c)
1600 if (Op0.getOpcode() == Opc && Op0.hasOneUse()) {
1602 return DAG.getNode(minMaxOpcToMin3Max3Opc(Opc),
1610 // max(a, max(b, c))
1611 if (Op1.getOpcode() == Opc && Op1.hasOneUse()) {
1613 return DAG.getNode(minMaxOpcToMin3Max3Opc(Opc),
1624 SDValue SITargetLowering::performSetCCCombine(SDNode *N,
1625 DAGCombinerInfo &DCI) const {
1626 SelectionDAG &DAG = DCI.DAG;
1629 SDValue LHS = N->getOperand(0);
1630 SDValue RHS = N->getOperand(1);
1631 EVT VT = LHS.getValueType();
1633 if (VT != MVT::f32 && VT != MVT::f64)
1636 // Match isinf pattern
1637 // (fcmp oeq (fabs x), inf) -> (fp_class x, (p_infinity | n_infinity))
1638 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
1639 if (CC == ISD::SETOEQ && LHS.getOpcode() == ISD::FABS) {
1640 const ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS);
1644 const APFloat &APF = CRHS->getValueAPF();
1645 if (APF.isInfinity() && !APF.isNegative()) {
1646 unsigned Mask = SIInstrFlags::P_INFINITY | SIInstrFlags::N_INFINITY;
1647 return DAG.getNode(AMDGPUISD::FP_CLASS, SL, MVT::i1, LHS.getOperand(0),
1648 DAG.getConstant(Mask, SL, MVT::i32));
1655 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
1656 DAGCombinerInfo &DCI) const {
1657 SelectionDAG &DAG = DCI.DAG;
1660 switch (N->getOpcode()) {
1662 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1664 return performSetCCCombine(N, DCI);
1665 case ISD::FMAXNUM: // TODO: What about fmax_legacy?
1667 case AMDGPUISD::SMAX:
1668 case AMDGPUISD::SMIN:
1669 case AMDGPUISD::UMAX:
1670 case AMDGPUISD::UMIN: {
1671 if (DCI.getDAGCombineLevel() >= AfterLegalizeDAG &&
1672 N->getValueType(0) != MVT::f64 &&
1673 getTargetMachine().getOptLevel() > CodeGenOpt::None)
1674 return performMin3Max3Combine(N, DCI);
1678 case AMDGPUISD::CVT_F32_UBYTE0:
1679 case AMDGPUISD::CVT_F32_UBYTE1:
1680 case AMDGPUISD::CVT_F32_UBYTE2:
1681 case AMDGPUISD::CVT_F32_UBYTE3: {
1682 unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
1684 SDValue Src = N->getOperand(0);
1685 APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);
1687 APInt KnownZero, KnownOne;
1688 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1689 !DCI.isBeforeLegalizeOps());
1690 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1691 if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
1692 TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {
1693 DCI.CommitTargetLoweringOpt(TLO);
1699 case ISD::UINT_TO_FP: {
1700 return performUCharToFloatCombine(N, DCI);
1703 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
1706 EVT VT = N->getValueType(0);
1710 // Only do this if we are not trying to support denormals. v_mad_f32 does
1711 // not support denormals ever.
1712 if (Subtarget->hasFP32Denormals())
1715 SDValue LHS = N->getOperand(0);
1716 SDValue RHS = N->getOperand(1);
1718 // These should really be instruction patterns, but writing patterns with
1719 // source modiifiers is a pain.
1721 // fadd (fadd (a, a), b) -> mad 2.0, a, b
1722 if (LHS.getOpcode() == ISD::FADD) {
1723 SDValue A = LHS.getOperand(0);
1724 if (A == LHS.getOperand(1)) {
1725 const SDValue Two = DAG.getConstantFP(2.0, DL, MVT::f32);
1726 return DAG.getNode(ISD::FMAD, DL, VT, Two, A, RHS);
1730 // fadd (b, fadd (a, a)) -> mad 2.0, a, b
1731 if (RHS.getOpcode() == ISD::FADD) {
1732 SDValue A = RHS.getOperand(0);
1733 if (A == RHS.getOperand(1)) {
1734 const SDValue Two = DAG.getConstantFP(2.0, DL, MVT::f32);
1735 return DAG.getNode(ISD::FMAD, DL, VT, Two, A, LHS);
1742 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
1745 EVT VT = N->getValueType(0);
1747 // Try to get the fneg to fold into the source modifier. This undoes generic
1748 // DAG combines and folds them into the mad.
1750 // Only do this if we are not trying to support denormals. v_mad_f32 does
1751 // not support denormals ever.
1752 if (VT == MVT::f32 &&
1753 !Subtarget->hasFP32Denormals()) {
1754 SDValue LHS = N->getOperand(0);
1755 SDValue RHS = N->getOperand(1);
1756 if (LHS.getOpcode() == ISD::FADD) {
1757 // (fsub (fadd a, a), c) -> mad 2.0, a, (fneg c)
1759 SDValue A = LHS.getOperand(0);
1760 if (A == LHS.getOperand(1)) {
1761 const SDValue Two = DAG.getConstantFP(2.0, DL, MVT::f32);
1762 SDValue NegRHS = DAG.getNode(ISD::FNEG, DL, VT, RHS);
1764 return DAG.getNode(ISD::FMAD, DL, VT, Two, A, NegRHS);
1768 if (RHS.getOpcode() == ISD::FADD) {
1769 // (fsub c, (fadd a, a)) -> mad -2.0, a, c
1771 SDValue A = RHS.getOperand(0);
1772 if (A == RHS.getOperand(1)) {
1773 const SDValue NegTwo = DAG.getConstantFP(-2.0, DL, MVT::f32);
1774 return DAG.getNode(ISD::FMAD, DL, VT, NegTwo, A, LHS);
1786 case ISD::ATOMIC_LOAD:
1787 case ISD::ATOMIC_STORE:
1788 case ISD::ATOMIC_CMP_SWAP:
1789 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
1790 case ISD::ATOMIC_SWAP:
1791 case ISD::ATOMIC_LOAD_ADD:
1792 case ISD::ATOMIC_LOAD_SUB:
1793 case ISD::ATOMIC_LOAD_AND:
1794 case ISD::ATOMIC_LOAD_OR:
1795 case ISD::ATOMIC_LOAD_XOR:
1796 case ISD::ATOMIC_LOAD_NAND:
1797 case ISD::ATOMIC_LOAD_MIN:
1798 case ISD::ATOMIC_LOAD_MAX:
1799 case ISD::ATOMIC_LOAD_UMIN:
1800 case ISD::ATOMIC_LOAD_UMAX: { // TODO: Target mem intrinsics.
1801 if (DCI.isBeforeLegalize())
1804 MemSDNode *MemNode = cast<MemSDNode>(N);
1805 SDValue Ptr = MemNode->getBasePtr();
1807 // TODO: We could also do this for multiplies.
1808 unsigned AS = MemNode->getAddressSpace();
1809 if (Ptr.getOpcode() == ISD::SHL && AS != AMDGPUAS::PRIVATE_ADDRESS) {
1810 SDValue NewPtr = performSHLPtrCombine(Ptr.getNode(), AS, DCI);
1812 SmallVector<SDValue, 8> NewOps(MemNode->op_begin(), MemNode->op_end());
1814 NewOps[N->getOpcode() == ISD::STORE ? 2 : 1] = NewPtr;
1815 return SDValue(DAG.UpdateNodeOperands(MemNode, NewOps), 0);
1821 return performAndCombine(N, DCI);
1823 return performOrCombine(N, DCI);
1824 case AMDGPUISD::FP_CLASS:
1825 return performClassCombine(N, DCI);
1827 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1830 /// \brief Analyze the possible immediate value Op
1832 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
1833 /// and the immediate value if it's a literal immediate
1834 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
1836 const SIInstrInfo *TII =
1837 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
1839 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
1840 if (TII->isInlineConstant(Node->getAPIntValue()))
1843 uint64_t Val = Node->getZExtValue();
1844 return isUInt<32>(Val) ? Val : -1;
1847 if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
1848 if (TII->isInlineConstant(Node->getValueAPF().bitcastToAPInt()))
1851 if (Node->getValueType(0) == MVT::f32)
1852 return FloatToBits(Node->getValueAPF().convertToFloat());
1860 /// \brief Helper function for adjustWritemask
1861 static unsigned SubIdx2Lane(unsigned Idx) {
1864 case AMDGPU::sub0: return 0;
1865 case AMDGPU::sub1: return 1;
1866 case AMDGPU::sub2: return 2;
1867 case AMDGPU::sub3: return 3;
1871 /// \brief Adjust the writemask of MIMG instructions
1872 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
1873 SelectionDAG &DAG) const {
1874 SDNode *Users[4] = { };
1876 unsigned OldDmask = Node->getConstantOperandVal(0);
1877 unsigned NewDmask = 0;
1879 // Try to figure out the used register components
1880 for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
1883 // Abort if we can't understand the usage
1884 if (!I->isMachineOpcode() ||
1885 I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
1888 // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used.
1889 // Note that subregs are packed, i.e. Lane==0 is the first bit set
1890 // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit
1892 Lane = SubIdx2Lane(I->getConstantOperandVal(1));
1894 // Set which texture component corresponds to the lane.
1896 for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) {
1898 Comp = countTrailingZeros(Dmask);
1899 Dmask &= ~(1 << Comp);
1902 // Abort if we have more than one user per component
1907 NewDmask |= 1 << Comp;
1910 // Abort if there's no change
1911 if (NewDmask == OldDmask)
1914 // Adjust the writemask in the node
1915 std::vector<SDValue> Ops;
1916 Ops.push_back(DAG.getTargetConstant(NewDmask, SDLoc(Node), MVT::i32));
1917 Ops.insert(Ops.end(), Node->op_begin() + 1, Node->op_end());
1918 Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops);
1920 // If we only got one lane, replace it with a copy
1921 // (if NewDmask has only one bit set...)
1922 if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
1923 SDValue RC = DAG.getTargetConstant(AMDGPU::VGPR_32RegClassID, SDLoc(),
1925 SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1926 SDLoc(), Users[Lane]->getValueType(0),
1927 SDValue(Node, 0), RC);
1928 DAG.ReplaceAllUsesWith(Users[Lane], Copy);
1932 // Update the users of the node with the new indices
1933 for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
1935 SDNode *User = Users[i];
1939 SDValue Op = DAG.getTargetConstant(Idx, SDLoc(User), MVT::i32);
1940 DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
1944 case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
1945 case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
1946 case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
1951 /// \brief Legalize target independent instructions (e.g. INSERT_SUBREG)
1952 /// with frame index operands.
1953 /// LLVM assumes that inputs are to these instructions are registers.
1954 void SITargetLowering::legalizeTargetIndependentNode(SDNode *Node,
1955 SelectionDAG &DAG) const {
1957 SmallVector<SDValue, 8> Ops;
1958 for (unsigned i = 0; i < Node->getNumOperands(); ++i) {
1959 if (!isa<FrameIndexSDNode>(Node->getOperand(i))) {
1960 Ops.push_back(Node->getOperand(i));
1965 Ops.push_back(SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL,
1966 Node->getOperand(i).getValueType(),
1967 Node->getOperand(i)), 0));
1970 DAG.UpdateNodeOperands(Node, Ops);
1973 /// \brief Fold the instructions after selecting them.
1974 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
1975 SelectionDAG &DAG) const {
1976 const SIInstrInfo *TII =
1977 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
1979 if (TII->isMIMG(Node->getMachineOpcode()))
1980 adjustWritemask(Node, DAG);
1982 if (Node->getMachineOpcode() == AMDGPU::INSERT_SUBREG ||
1983 Node->getMachineOpcode() == AMDGPU::REG_SEQUENCE) {
1984 legalizeTargetIndependentNode(Node, DAG);
1990 /// \brief Assign the register class depending on the number of
1991 /// bits set in the writemask
1992 void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
1993 SDNode *Node) const {
1994 const SIInstrInfo *TII =
1995 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
1997 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1998 TII->legalizeOperands(MI);
2000 if (TII->isMIMG(MI->getOpcode())) {
2001 unsigned VReg = MI->getOperand(0).getReg();
2002 unsigned Writemask = MI->getOperand(1).getImm();
2003 unsigned BitsSet = 0;
2004 for (unsigned i = 0; i < 4; ++i)
2005 BitsSet += Writemask & (1 << i) ? 1 : 0;
2007 const TargetRegisterClass *RC;
2010 case 1: RC = &AMDGPU::VGPR_32RegClass; break;
2011 case 2: RC = &AMDGPU::VReg_64RegClass; break;
2012 case 3: RC = &AMDGPU::VReg_96RegClass; break;
2015 unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
2016 MI->setDesc(TII->get(NewOpcode));
2017 MRI.setRegClass(VReg, RC);
2021 // Replace unused atomics with the no return version.
2022 int NoRetAtomicOp = AMDGPU::getAtomicNoRetOp(MI->getOpcode());
2023 if (NoRetAtomicOp != -1) {
2024 if (!Node->hasAnyUseOfValue(0)) {
2025 MI->setDesc(TII->get(NoRetAtomicOp));
2026 MI->RemoveOperand(0);
2033 static SDValue buildSMovImm32(SelectionDAG &DAG, SDLoc DL, uint64_t Val) {
2034 SDValue K = DAG.getTargetConstant(Val, DL, MVT::i32);
2035 return SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, K), 0);
2038 MachineSDNode *SITargetLowering::wrapAddr64Rsrc(SelectionDAG &DAG,
2040 SDValue Ptr) const {
2041 const SIInstrInfo *TII =
2042 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
2044 // XXX - Workaround for moveToVALU not handling different register class
2045 // inserts for REG_SEQUENCE.
2047 // Build the half of the subregister with the constants.
2048 const SDValue Ops0[] = {
2049 DAG.getTargetConstant(AMDGPU::SGPR_64RegClassID, DL, MVT::i32),
2050 buildSMovImm32(DAG, DL, 0),
2051 DAG.getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
2052 buildSMovImm32(DAG, DL, TII->getDefaultRsrcDataFormat() >> 32),
2053 DAG.getTargetConstant(AMDGPU::sub1, DL, MVT::i32)
2056 SDValue SubRegHi = SDValue(DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL,
2057 MVT::v2i32, Ops0), 0);
2059 // Combine the constants and the pointer.
2060 const SDValue Ops1[] = {
2061 DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32),
2063 DAG.getTargetConstant(AMDGPU::sub0_sub1, DL, MVT::i32),
2065 DAG.getTargetConstant(AMDGPU::sub2_sub3, DL, MVT::i32)
2068 return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops1);
2070 const SDValue Ops[] = {
2071 DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, MVT::i32),
2073 DAG.getTargetConstant(AMDGPU::sub0_sub1, MVT::i32),
2074 buildSMovImm32(DAG, DL, 0),
2075 DAG.getTargetConstant(AMDGPU::sub2, MVT::i32),
2076 buildSMovImm32(DAG, DL, TII->getDefaultRsrcFormat() >> 32),
2077 DAG.getTargetConstant(AMDGPU::sub3, MVT::i32)
2080 return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops);
2085 /// \brief Return a resource descriptor with the 'Add TID' bit enabled
2086 /// The TID (Thread ID) is multipled by the stride value (bits [61:48]
2087 /// of the resource descriptor) to create an offset, which is added to the
2088 /// resource ponter.
2089 MachineSDNode *SITargetLowering::buildRSRC(SelectionDAG &DAG,
2092 uint32_t RsrcDword1,
2093 uint64_t RsrcDword2And3) const {
2094 SDValue PtrLo = DAG.getTargetExtractSubreg(AMDGPU::sub0, DL, MVT::i32, Ptr);
2095 SDValue PtrHi = DAG.getTargetExtractSubreg(AMDGPU::sub1, DL, MVT::i32, Ptr);
2097 PtrHi = SDValue(DAG.getMachineNode(AMDGPU::S_OR_B32, DL, MVT::i32, PtrHi,
2098 DAG.getConstant(RsrcDword1, DL, MVT::i32)),
2102 SDValue DataLo = buildSMovImm32(DAG, DL,
2103 RsrcDword2And3 & UINT64_C(0xFFFFFFFF));
2104 SDValue DataHi = buildSMovImm32(DAG, DL, RsrcDword2And3 >> 32);
2106 const SDValue Ops[] = {
2107 DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32),
2109 DAG.getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
2111 DAG.getTargetConstant(AMDGPU::sub1, DL, MVT::i32),
2113 DAG.getTargetConstant(AMDGPU::sub2, DL, MVT::i32),
2115 DAG.getTargetConstant(AMDGPU::sub3, DL, MVT::i32)
2118 return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops);
2121 MachineSDNode *SITargetLowering::buildScratchRSRC(SelectionDAG &DAG,
2123 SDValue Ptr) const {
2124 const SIInstrInfo *TII =
2125 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
2126 uint64_t Rsrc = TII->getDefaultRsrcDataFormat() | AMDGPU::RSRC_TID_ENABLE |
2129 return buildRSRC(DAG, DL, Ptr, 0, Rsrc);
2132 SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2133 const TargetRegisterClass *RC,
2134 unsigned Reg, EVT VT) const {
2135 SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT);
2137 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()),
2138 cast<RegisterSDNode>(VReg)->getReg(), VT);
2141 //===----------------------------------------------------------------------===//
2142 // SI Inline Assembly Support
2143 //===----------------------------------------------------------------------===//
2145 std::pair<unsigned, const TargetRegisterClass *>
2146 SITargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
2147 const std::string &Constraint,
2149 if (Constraint == "r") {
2150 switch(VT.SimpleTy) {
2151 default: llvm_unreachable("Unhandled type for 'r' inline asm constraint");
2153 return std::make_pair(0U, &AMDGPU::SGPR_64RegClass);
2155 return std::make_pair(0U, &AMDGPU::SGPR_32RegClass);
2159 if (Constraint.size() > 1) {
2160 const TargetRegisterClass *RC = nullptr;
2161 if (Constraint[1] == 'v') {
2162 RC = &AMDGPU::VGPR_32RegClass;
2163 } else if (Constraint[1] == 's') {
2164 RC = &AMDGPU::SGPR_32RegClass;
2168 unsigned Idx = std::atoi(Constraint.substr(2).c_str());
2169 if (Idx < RC->getNumRegs())
2170 return std::make_pair(RC->getRegister(Idx), RC);
2173 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);