1 //===-- SIISelLowering.cpp - SI DAG Lowering Implementation ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief Custom DAG lowering for SI
13 //===----------------------------------------------------------------------===//
15 #include "SIISelLowering.h"
17 #include "AMDGPUSubtarget.h"
18 #include "AMDILIntrinsicInfo.h"
19 #include "SIInstrInfo.h"
20 #include "SIMachineFunctionInfo.h"
21 #include "SIRegisterInfo.h"
22 #include "llvm/CodeGen/CallingConvLower.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/SelectionDAG.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/ADT/SmallString.h"
31 SITargetLowering::SITargetLowering(TargetMachine &TM) :
32 AMDGPUTargetLowering(TM) {
33 addRegisterClass(MVT::i1, &AMDGPU::VReg_1RegClass);
34 addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
36 addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
37 addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);
39 addRegisterClass(MVT::i32, &AMDGPU::SReg_32RegClass);
40 addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
42 addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
43 addRegisterClass(MVT::v2i32, &AMDGPU::SReg_64RegClass);
44 addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
46 addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
47 addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
49 addRegisterClass(MVT::v8i32, &AMDGPU::VReg_256RegClass);
50 addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
52 addRegisterClass(MVT::v16i32, &AMDGPU::VReg_512RegClass);
53 addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);
55 computeRegisterProperties();
58 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
59 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
60 setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
61 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
62 setCondCodeAction(ISD::SETULE, MVT::f32, Expand);
63 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
65 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
66 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
67 setCondCodeAction(ISD::SETUGE, MVT::f64, Expand);
68 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
69 setCondCodeAction(ISD::SETULE, MVT::f64, Expand);
70 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
72 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
73 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
74 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
75 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);
77 setOperationAction(ISD::ADD, MVT::i32, Legal);
78 setOperationAction(ISD::ADDC, MVT::i32, Legal);
79 setOperationAction(ISD::ADDE, MVT::i32, Legal);
81 // We need to custom lower vector stores from local memory
82 setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
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 // We need to custom lower loads/stores from private memory
91 setOperationAction(ISD::LOAD, MVT::i32, Custom);
92 setOperationAction(ISD::LOAD, MVT::i64, Custom);
93 setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
94 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
95 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
97 setOperationAction(ISD::STORE, MVT::i1, Custom);
98 setOperationAction(ISD::STORE, MVT::i32, Custom);
99 setOperationAction(ISD::STORE, MVT::i64, Custom);
100 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
101 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
103 setOperationAction(ISD::SELECT, MVT::f32, Promote);
104 AddPromotedToType(ISD::SELECT, MVT::f32, MVT::i32);
105 setOperationAction(ISD::SELECT, MVT::i64, Custom);
106 setOperationAction(ISD::SELECT, MVT::f64, Promote);
107 AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
109 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
110 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
111 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
112 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
114 setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
115 setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
117 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
118 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
119 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
121 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
122 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
123 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
125 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
127 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
129 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Custom);
131 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
133 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
134 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
135 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom);
136 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
138 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
140 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
141 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Custom);
142 setLoadExtAction(ISD::SEXTLOAD, MVT::i16, Custom);
143 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Expand);
144 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, Expand);
145 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, Expand);
147 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
148 setLoadExtAction(ISD::ZEXTLOAD, MVT::i8, Custom);
149 setLoadExtAction(ISD::ZEXTLOAD, MVT::i16, Custom);
150 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Expand);
152 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
153 setLoadExtAction(ISD::EXTLOAD, MVT::i8, Custom);
154 setLoadExtAction(ISD::EXTLOAD, MVT::i16, Custom);
155 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Expand);
156 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
158 setTruncStoreAction(MVT::i32, MVT::i8, Custom);
159 setTruncStoreAction(MVT::i32, MVT::i16, Custom);
160 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
161 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
162 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Expand);
163 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Expand);
165 setOperationAction(ISD::LOAD, MVT::i1, Custom);
167 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
168 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
169 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
171 // These should use UDIVREM, so set them to expand
172 setOperationAction(ISD::UDIV, MVT::i64, Expand);
173 setOperationAction(ISD::UREM, MVT::i64, Expand);
175 // We only support LOAD/STORE and vector manipulation ops for vectors
176 // with > 4 elements.
178 MVT::v8i32, MVT::v8f32, MVT::v16i32, MVT::v16f32
181 for (MVT VT : VecTypes) {
182 for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
186 case ISD::BUILD_VECTOR:
188 case ISD::EXTRACT_VECTOR_ELT:
189 case ISD::INSERT_VECTOR_ELT:
190 case ISD::CONCAT_VECTORS:
191 case ISD::INSERT_SUBVECTOR:
192 case ISD::EXTRACT_SUBVECTOR:
195 setOperationAction(Op, VT, Expand);
201 for (int I = MVT::v1f64; I <= MVT::v8f64; ++I) {
202 MVT::SimpleValueType VT = static_cast<MVT::SimpleValueType>(I);
203 setOperationAction(ISD::FTRUNC, VT, Expand);
204 setOperationAction(ISD::FCEIL, VT, Expand);
205 setOperationAction(ISD::FFLOOR, VT, Expand);
208 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
209 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
210 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
211 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
212 setOperationAction(ISD::FRINT, MVT::f64, Legal);
215 setTargetDAGCombine(ISD::SELECT_CC);
216 setTargetDAGCombine(ISD::SETCC);
218 setTargetDAGCombine(ISD::UINT_TO_FP);
220 setSchedulingPreference(Sched::RegPressure);
223 //===----------------------------------------------------------------------===//
224 // TargetLowering queries
225 //===----------------------------------------------------------------------===//
227 bool SITargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
229 bool *IsFast) const {
233 // XXX: This depends on the address space and also we may want to revist
234 // the alignment values we specify in the DataLayout.
236 // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
237 // which isn't a simple VT.
238 if (!VT.isSimple() || VT == MVT::Other)
241 // XXX - CI changes say "Support for unaligned memory accesses" but I don't
242 // see what for specifically. The wording everywhere else seems to be the
245 // 3.6.4 - Operations using pairs of VGPRs (for example: double-floats) have
246 // no alignment restrictions.
247 if (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS) {
248 // Using any pair of GPRs should be the same as any other pair.
251 return VT.bitsGE(MVT::i64);
254 // XXX - The only mention I see of this in the ISA manual is for LDS direct
255 // reads the "byte address and must be dword aligned". Is it also true for the
256 // normal loads and stores?
257 if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS)
260 // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
261 // byte-address are ignored, thus forcing Dword alignment.
264 return VT.bitsGT(MVT::i32);
267 bool SITargetLowering::shouldSplitVectorType(EVT VT) const {
268 return VT.getScalarType().bitsLE(MVT::i16);
271 bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
273 const SIInstrInfo *TII =
274 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
275 return TII->isInlineConstant(Imm);
278 SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
279 SDLoc DL, SDValue Chain,
280 unsigned Offset, bool Signed) const {
281 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
282 PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
283 AMDGPUAS::CONSTANT_ADDRESS);
284 SDValue BasePtr = DAG.getCopyFromReg(Chain, DL,
285 MRI.getLiveInVirtReg(AMDGPU::SGPR0_SGPR1), MVT::i64);
286 SDValue Ptr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
287 DAG.getConstant(Offset, MVT::i64));
288 return DAG.getExtLoad(Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD, DL, VT, Chain, Ptr,
289 MachinePointerInfo(UndefValue::get(PtrTy)), MemVT,
290 false, false, MemVT.getSizeInBits() >> 3);
294 SDValue SITargetLowering::LowerFormalArguments(
296 CallingConv::ID CallConv,
298 const SmallVectorImpl<ISD::InputArg> &Ins,
299 SDLoc DL, SelectionDAG &DAG,
300 SmallVectorImpl<SDValue> &InVals) const {
302 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
304 MachineFunction &MF = DAG.getMachineFunction();
305 FunctionType *FType = MF.getFunction()->getFunctionType();
306 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
308 assert(CallConv == CallingConv::C);
310 SmallVector<ISD::InputArg, 16> Splits;
311 uint32_t Skipped = 0;
313 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
314 const ISD::InputArg &Arg = Ins[i];
316 // First check if it's a PS input addr
317 if (Info->ShaderType == ShaderType::PIXEL && !Arg.Flags.isInReg() &&
318 !Arg.Flags.isByVal()) {
320 assert((PSInputNum <= 15) && "Too many PS inputs!");
323 // We can savely skip PS inputs
329 Info->PSInputAddr |= 1 << PSInputNum++;
332 // Second split vertices into their elements
333 if (Info->ShaderType != ShaderType::COMPUTE && Arg.VT.isVector()) {
334 ISD::InputArg NewArg = Arg;
335 NewArg.Flags.setSplit();
336 NewArg.VT = Arg.VT.getVectorElementType();
338 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
339 // three or five element vertex only needs three or five registers,
340 // NOT four or eigth.
341 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
342 unsigned NumElements = ParamType->getVectorNumElements();
344 for (unsigned j = 0; j != NumElements; ++j) {
345 Splits.push_back(NewArg);
346 NewArg.PartOffset += NewArg.VT.getStoreSize();
349 } else if (Info->ShaderType != ShaderType::COMPUTE) {
350 Splits.push_back(Arg);
354 SmallVector<CCValAssign, 16> ArgLocs;
355 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
356 getTargetMachine(), ArgLocs, *DAG.getContext());
358 // At least one interpolation mode must be enabled or else the GPU will hang.
359 if (Info->ShaderType == ShaderType::PIXEL && (Info->PSInputAddr & 0x7F) == 0) {
360 Info->PSInputAddr |= 1;
361 CCInfo.AllocateReg(AMDGPU::VGPR0);
362 CCInfo.AllocateReg(AMDGPU::VGPR1);
365 // The pointer to the list of arguments is stored in SGPR0, SGPR1
366 if (Info->ShaderType == ShaderType::COMPUTE) {
367 CCInfo.AllocateReg(AMDGPU::SGPR0);
368 CCInfo.AllocateReg(AMDGPU::SGPR1);
369 MF.addLiveIn(AMDGPU::SGPR0_SGPR1, &AMDGPU::SReg_64RegClass);
372 if (Info->ShaderType == ShaderType::COMPUTE) {
373 getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
377 AnalyzeFormalArguments(CCInfo, Splits);
379 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
381 const ISD::InputArg &Arg = Ins[i];
382 if (Skipped & (1 << i)) {
383 InVals.push_back(DAG.getUNDEF(Arg.VT));
387 CCValAssign &VA = ArgLocs[ArgIdx++];
388 EVT VT = VA.getLocVT();
392 EVT MemVT = Splits[i].VT;
393 // The first 36 bytes of the input buffer contains information about
394 // thread group and global sizes.
395 SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, DAG.getRoot(),
396 36 + VA.getLocMemOffset(),
397 Ins[i].Flags.isSExt());
398 InVals.push_back(Arg);
401 assert(VA.isRegLoc() && "Parameter must be in a register!");
403 unsigned Reg = VA.getLocReg();
405 if (VT == MVT::i64) {
406 // For now assume it is a pointer
407 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
408 &AMDGPU::SReg_64RegClass);
409 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
410 InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
414 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
416 Reg = MF.addLiveIn(Reg, RC);
417 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
419 if (Arg.VT.isVector()) {
421 // Build a vector from the registers
422 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
423 unsigned NumElements = ParamType->getVectorNumElements();
425 SmallVector<SDValue, 4> Regs;
427 for (unsigned j = 1; j != NumElements; ++j) {
428 Reg = ArgLocs[ArgIdx++].getLocReg();
429 Reg = MF.addLiveIn(Reg, RC);
430 Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
433 // Fill up the missing vector elements
434 NumElements = Arg.VT.getVectorNumElements() - NumElements;
435 for (unsigned j = 0; j != NumElements; ++j)
436 Regs.push_back(DAG.getUNDEF(VT));
438 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT, Regs));
442 InVals.push_back(Val);
447 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
448 MachineInstr * MI, MachineBasicBlock * BB) const {
450 MachineBasicBlock::iterator I = *MI;
451 const SIInstrInfo *TII =
452 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
453 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
455 switch (MI->getOpcode()) {
457 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
458 case AMDGPU::BRANCH: return BB;
459 case AMDGPU::SI_ADDR64_RSRC: {
460 unsigned SuperReg = MI->getOperand(0).getReg();
461 unsigned SubRegLo = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
462 unsigned SubRegHi = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
463 unsigned SubRegHiHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
464 unsigned SubRegHiLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
465 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B64), SubRegLo)
466 .addOperand(MI->getOperand(1));
467 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiLo)
469 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiHi)
470 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
471 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SubRegHi)
473 .addImm(AMDGPU::sub0)
475 .addImm(AMDGPU::sub1);
476 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SuperReg)
478 .addImm(AMDGPU::sub0_sub1)
480 .addImm(AMDGPU::sub2_sub3);
481 MI->eraseFromParent();
484 case AMDGPU::V_SUB_F64:
485 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F64),
486 MI->getOperand(0).getReg())
487 .addReg(MI->getOperand(1).getReg())
488 .addReg(MI->getOperand(2).getReg())
489 .addImm(0) /* src2 */
491 .addImm(0) /* CLAMP */
492 .addImm(0) /* OMOD */
493 .addImm(2); /* NEG */
494 MI->eraseFromParent();
497 case AMDGPU::SI_RegisterStorePseudo: {
498 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
499 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
500 MachineInstrBuilder MIB =
501 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::SI_RegisterStore),
503 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i)
504 MIB.addOperand(MI->getOperand(i));
506 MI->eraseFromParent();
509 case AMDGPU::FABS_SI: {
510 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
511 const SIInstrInfo *TII =
512 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
513 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
514 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
517 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_AND_B32_e32),
518 MI->getOperand(0).getReg())
519 .addReg(MI->getOperand(1).getReg())
521 MI->eraseFromParent();
524 case AMDGPU::FNEG_SI: {
525 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
526 const SIInstrInfo *TII =
527 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
528 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
529 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
532 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_XOR_B32_e32),
533 MI->getOperand(0).getReg())
534 .addReg(MI->getOperand(1).getReg())
536 MI->eraseFromParent();
539 case AMDGPU::FCLAMP_SI: {
540 const SIInstrInfo *TII =
541 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
542 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F32_e64),
543 MI->getOperand(0).getReg())
544 .addImm(0) // SRC0 modifiers
545 .addOperand(MI->getOperand(1))
546 .addImm(0) // SRC1 modifiers
550 MI->eraseFromParent();
556 EVT SITargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
557 if (!VT.isVector()) {
560 return MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
563 MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
567 bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
568 VT = VT.getScalarType();
573 switch (VT.getSimpleVT().SimpleTy) {
575 return false; /* There is V_MAD_F32 for f32 */
585 //===----------------------------------------------------------------------===//
586 // Custom DAG Lowering Operations
587 //===----------------------------------------------------------------------===//
589 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
590 MachineFunction &MF = DAG.getMachineFunction();
591 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
592 switch (Op.getOpcode()) {
593 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
594 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
596 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
597 if (Op.getValueType().isVector() &&
598 (Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
599 Load->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS ||
600 (Load->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
601 Op.getValueType().getVectorNumElements() > 4))) {
602 SDValue MergedValues[2] = {
603 SplitVectorLoad(Op, DAG),
606 return DAG.getMergeValues(MergedValues, SDLoc(Op));
608 return LowerLOAD(Op, DAG);
612 case ISD::SELECT: return LowerSELECT(Op, DAG);
613 case ISD::STORE: return LowerSTORE(Op, DAG);
614 case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
615 case ISD::INTRINSIC_WO_CHAIN: {
616 unsigned IntrinsicID =
617 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
618 EVT VT = Op.getValueType();
620 //XXX: Hardcoded we only use two to store the pointer to the parameters.
621 unsigned NumUserSGPRs = 2;
622 switch (IntrinsicID) {
623 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
624 case Intrinsic::r600_read_ngroups_x:
625 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 0, false);
626 case Intrinsic::r600_read_ngroups_y:
627 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 4, false);
628 case Intrinsic::r600_read_ngroups_z:
629 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 8, false);
630 case Intrinsic::r600_read_global_size_x:
631 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 12, false);
632 case Intrinsic::r600_read_global_size_y:
633 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 16, false);
634 case Intrinsic::r600_read_global_size_z:
635 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 20, false);
636 case Intrinsic::r600_read_local_size_x:
637 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 24, false);
638 case Intrinsic::r600_read_local_size_y:
639 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 28, false);
640 case Intrinsic::r600_read_local_size_z:
641 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 32, false);
642 case Intrinsic::r600_read_tgid_x:
643 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
644 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 0), VT);
645 case Intrinsic::r600_read_tgid_y:
646 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
647 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 1), VT);
648 case Intrinsic::r600_read_tgid_z:
649 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
650 AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 2), VT);
651 case Intrinsic::r600_read_tidig_x:
652 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
654 case Intrinsic::r600_read_tidig_y:
655 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
657 case Intrinsic::r600_read_tidig_z:
658 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
660 case AMDGPUIntrinsic::SI_load_const: {
666 MachineMemOperand *MMO = MF.getMachineMemOperand(
667 MachinePointerInfo(),
668 MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
669 VT.getSizeInBits() / 8, 4);
670 return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
671 Op->getVTList(), Ops, VT, MMO);
673 case AMDGPUIntrinsic::SI_sample:
674 return LowerSampleIntrinsic(AMDGPUISD::SAMPLE, Op, DAG);
675 case AMDGPUIntrinsic::SI_sampleb:
676 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEB, Op, DAG);
677 case AMDGPUIntrinsic::SI_sampled:
678 return LowerSampleIntrinsic(AMDGPUISD::SAMPLED, Op, DAG);
679 case AMDGPUIntrinsic::SI_samplel:
680 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEL, Op, DAG);
681 case AMDGPUIntrinsic::SI_vs_load_input:
682 return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
689 case ISD::INTRINSIC_VOID:
690 SDValue Chain = Op.getOperand(0);
691 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
693 switch (IntrinsicID) {
694 case AMDGPUIntrinsic::SI_tbuffer_store: {
712 EVT VT = Op.getOperand(3).getValueType();
714 MachineMemOperand *MMO = MF.getMachineMemOperand(
715 MachinePointerInfo(),
716 MachineMemOperand::MOStore,
717 VT.getSizeInBits() / 8, 4);
718 return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL,
719 Op->getVTList(), Ops, VT, MMO);
728 /// \brief Helper function for LowerBRCOND
729 static SDNode *findUser(SDValue Value, unsigned Opcode) {
731 SDNode *Parent = Value.getNode();
732 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
735 if (I.getUse().get() != Value)
738 if (I->getOpcode() == Opcode)
744 /// This transforms the control flow intrinsics to get the branch destination as
745 /// last parameter, also switches branch target with BR if the need arise
746 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
747 SelectionDAG &DAG) const {
751 SDNode *Intr = BRCOND.getOperand(1).getNode();
752 SDValue Target = BRCOND.getOperand(2);
753 SDNode *BR = nullptr;
755 if (Intr->getOpcode() == ISD::SETCC) {
756 // As long as we negate the condition everything is fine
757 SDNode *SetCC = Intr;
758 assert(SetCC->getConstantOperandVal(1) == 1);
759 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
761 Intr = SetCC->getOperand(0).getNode();
764 // Get the target from BR if we don't negate the condition
765 BR = findUser(BRCOND, ISD::BR);
766 Target = BR->getOperand(1);
769 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
771 // Build the result and
772 SmallVector<EVT, 4> Res;
773 for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
774 Res.push_back(Intr->getValueType(i));
776 // operands of the new intrinsic call
777 SmallVector<SDValue, 4> Ops;
778 Ops.push_back(BRCOND.getOperand(0));
779 for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
780 Ops.push_back(Intr->getOperand(i));
781 Ops.push_back(Target);
783 // build the new intrinsic call
784 SDNode *Result = DAG.getNode(
785 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
786 DAG.getVTList(Res), Ops).getNode();
789 // Give the branch instruction our target
794 DAG.MorphNodeTo(BR, ISD::BR, BR->getVTList(), Ops);
797 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
799 // Copy the intrinsic results to registers
800 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
801 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
805 Chain = DAG.getCopyToReg(
807 CopyToReg->getOperand(1),
808 SDValue(Result, i - 1),
811 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
814 // Remove the old intrinsic from the chain
815 DAG.ReplaceAllUsesOfValueWith(
816 SDValue(Intr, Intr->getNumValues() - 1),
817 Intr->getOperand(0));
822 SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
824 LoadSDNode *Load = cast<LoadSDNode>(Op);
825 SDValue Ret = AMDGPUTargetLowering::LowerLOAD(Op, DAG);
826 SDValue MergedValues[2];
827 MergedValues[1] = Load->getChain();
829 MergedValues[0] = Ret;
830 return DAG.getMergeValues(MergedValues, DL);
833 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
837 EVT MemVT = Load->getMemoryVT();
839 assert(!MemVT.isVector() && "Private loads should be scalarized");
840 assert(!MemVT.isFloatingPoint() && "FP loads should be promoted to int");
842 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
843 DAG.getConstant(2, MVT::i32));
844 Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
845 Load->getChain(), Ptr,
846 DAG.getTargetConstant(0, MVT::i32),
848 if (MemVT.getSizeInBits() == 64) {
849 SDValue IncPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
850 DAG.getConstant(1, MVT::i32));
852 SDValue LoadUpper = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
853 Load->getChain(), IncPtr,
854 DAG.getTargetConstant(0, MVT::i32),
857 Ret = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ret, LoadUpper);
860 MergedValues[0] = Ret;
861 return DAG.getMergeValues(MergedValues, DL);
865 SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
867 SelectionDAG &DAG) const {
868 return DAG.getNode(Opcode, SDLoc(Op), Op.getValueType(), Op.getOperand(1),
874 SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
875 if (Op.getValueType() != MVT::i64)
879 SDValue Cond = Op.getOperand(0);
881 SDValue Zero = DAG.getConstant(0, MVT::i32);
882 SDValue One = DAG.getConstant(1, MVT::i32);
884 SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1));
885 SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2));
887 SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero);
888 SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero);
890 SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1);
892 SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One);
893 SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One);
895 SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1);
897 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i32, Lo, Hi);
898 return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
901 SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
903 StoreSDNode *Store = cast<StoreSDNode>(Op);
904 EVT VT = Store->getMemoryVT();
906 SDValue Ret = AMDGPUTargetLowering::LowerSTORE(Op, DAG);
910 if (VT.isVector() && VT.getVectorNumElements() >= 8)
911 return SplitVectorStore(Op, DAG);
914 return DAG.getTruncStore(Store->getChain(), DL,
915 DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32),
916 Store->getBasePtr(), MVT::i1, Store->getMemOperand());
918 if (Store->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
921 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Store->getBasePtr(),
922 DAG.getConstant(2, MVT::i32));
923 SDValue Chain = Store->getChain();
924 SmallVector<SDValue, 8> Values;
926 if (Store->isTruncatingStore()) {
928 if (Store->getMemoryVT() == MVT::i8) {
930 } else if (Store->getMemoryVT() == MVT::i16) {
933 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
934 Chain, Store->getBasePtr(),
935 DAG.getConstant(0, MVT::i32));
936 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, Store->getBasePtr(),
937 DAG.getConstant(0x3, MVT::i32));
938 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
939 DAG.getConstant(3, MVT::i32));
940 SDValue MaskedValue = DAG.getNode(ISD::AND, DL, MVT::i32, Store->getValue(),
941 DAG.getConstant(Mask, MVT::i32));
942 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
943 MaskedValue, ShiftAmt);
944 SDValue RotrAmt = DAG.getNode(ISD::SUB, DL, MVT::i32,
945 DAG.getConstant(32, MVT::i32), ShiftAmt);
946 SDValue DstMask = DAG.getNode(ISD::ROTR, DL, MVT::i32,
947 DAG.getConstant(Mask, MVT::i32),
949 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
950 Dst = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
952 Values.push_back(Dst);
953 } else if (VT == MVT::i64) {
954 for (unsigned i = 0; i < 2; ++i) {
955 Values.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
956 Store->getValue(), DAG.getConstant(i, MVT::i32)));
958 } else if (VT == MVT::i128) {
959 for (unsigned i = 0; i < 2; ++i) {
960 for (unsigned j = 0; j < 2; ++j) {
961 Values.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
962 DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
963 Store->getValue(), DAG.getConstant(i, MVT::i32)),
964 DAG.getConstant(j, MVT::i32)));
968 Values.push_back(Store->getValue());
971 for (unsigned i = 0; i < Values.size(); ++i) {
972 SDValue PartPtr = DAG.getNode(ISD::ADD, DL, MVT::i32,
973 Ptr, DAG.getConstant(i, MVT::i32));
974 Chain = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
975 Chain, Values[i], PartPtr,
976 DAG.getTargetConstant(0, MVT::i32));
981 //===----------------------------------------------------------------------===//
982 // Custom DAG optimizations
983 //===----------------------------------------------------------------------===//
985 SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
986 DAGCombinerInfo &DCI) {
987 EVT VT = N->getValueType(0);
988 EVT ScalarVT = VT.getScalarType();
989 if (ScalarVT != MVT::f32)
992 SelectionDAG &DAG = DCI.DAG;
995 SDValue Src = N->getOperand(0);
996 EVT SrcVT = Src.getValueType();
998 // TODO: We could try to match extracting the higher bytes, which would be
999 // easier if i8 vectors weren't promoted to i32 vectors, particularly after
1000 // types are legalized. v4i8 -> v4f32 is probably the only case to worry
1001 // about in practice.
1002 if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) {
1003 if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) {
1004 SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src);
1005 DCI.AddToWorklist(Cvt.getNode());
1010 // We are primarily trying to catch operations on illegal vector types
1011 // before they are expanded.
1012 // For scalars, we can use the more flexible method of checking masked bits
1013 // after legalization.
1014 if (!DCI.isBeforeLegalize() ||
1015 !SrcVT.isVector() ||
1016 SrcVT.getVectorElementType() != MVT::i8) {
1020 assert(DCI.isBeforeLegalize() && "Unexpected legal type");
1022 // Weird sized vectors are a pain to handle, but we know 3 is really the same
1024 unsigned NElts = SrcVT.getVectorNumElements();
1025 if (!SrcVT.isSimple() && NElts != 3)
1028 // Handle v4i8 -> v4f32 extload. Replace the v4i8 with a legal i32 load to
1029 // prevent a mess from expanding to v4i32 and repacking.
1030 if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse()) {
1031 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), SrcVT);
1032 EVT RegVT = getEquivalentLoadRegType(*DAG.getContext(), SrcVT);
1033 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f32, NElts);
1035 LoadSDNode *Load = cast<LoadSDNode>(Src);
1036 SDValue NewLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegVT,
1040 Load->getMemOperand());
1042 // Make sure successors of the original load stay after it by updating
1043 // them to use the new Chain.
1044 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), NewLoad.getValue(1));
1046 SmallVector<SDValue, 4> Elts;
1047 if (RegVT.isVector())
1048 DAG.ExtractVectorElements(NewLoad, Elts);
1050 Elts.push_back(NewLoad);
1052 SmallVector<SDValue, 4> Ops;
1054 unsigned EltIdx = 0;
1055 for (SDValue Elt : Elts) {
1056 unsigned ComponentsInElt = std::min(4u, NElts - 4 * EltIdx);
1057 for (unsigned I = 0; I < ComponentsInElt; ++I) {
1058 unsigned Opc = AMDGPUISD::CVT_F32_UBYTE0 + I;
1059 SDValue Cvt = DAG.getNode(Opc, DL, MVT::f32, Elt);
1060 DCI.AddToWorklist(Cvt.getNode());
1067 assert(Ops.size() == NElts);
1069 return DAG.getNode(ISD::BUILD_VECTOR, DL, FloatVT, Ops);
1075 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
1076 DAGCombinerInfo &DCI) const {
1077 SelectionDAG &DAG = DCI.DAG;
1079 EVT VT = N->getValueType(0);
1081 switch (N->getOpcode()) {
1082 default: return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1083 case ISD::SELECT_CC: {
1084 ConstantSDNode *True, *False;
1085 // i1 selectcc(l, r, -1, 0, cc) -> i1 setcc(l, r, cc)
1086 if ((True = dyn_cast<ConstantSDNode>(N->getOperand(2)))
1087 && (False = dyn_cast<ConstantSDNode>(N->getOperand(3)))
1088 && True->isAllOnesValue()
1089 && False->isNullValue()
1091 return DAG.getNode(ISD::SETCC, DL, VT, N->getOperand(0),
1092 N->getOperand(1), N->getOperand(4));
1098 SDValue Arg0 = N->getOperand(0);
1099 SDValue Arg1 = N->getOperand(1);
1100 SDValue CC = N->getOperand(2);
1101 ConstantSDNode * C = nullptr;
1102 ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
1104 // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
1106 && Arg0.getOpcode() == ISD::SIGN_EXTEND
1107 && Arg0.getOperand(0).getValueType() == MVT::i1
1108 && (C = dyn_cast<ConstantSDNode>(Arg1))
1110 && CCOp == ISD::SETNE) {
1111 return SimplifySetCC(VT, Arg0.getOperand(0),
1112 DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
1117 case AMDGPUISD::CVT_F32_UBYTE0:
1118 case AMDGPUISD::CVT_F32_UBYTE1:
1119 case AMDGPUISD::CVT_F32_UBYTE2:
1120 case AMDGPUISD::CVT_F32_UBYTE3: {
1121 unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
1123 SDValue Src = N->getOperand(0);
1124 APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);
1126 APInt KnownZero, KnownOne;
1127 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1128 !DCI.isBeforeLegalizeOps());
1129 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1130 if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
1131 TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {
1132 DCI.CommitTargetLoweringOpt(TLO);
1138 case ISD::UINT_TO_FP: {
1139 return performUCharToFloatCombine(N, DCI);
1143 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1146 /// \brief Test if RegClass is one of the VSrc classes
1147 static bool isVSrc(unsigned RegClass) {
1148 return AMDGPU::VSrc_32RegClassID == RegClass ||
1149 AMDGPU::VSrc_64RegClassID == RegClass;
1152 /// \brief Test if RegClass is one of the SSrc classes
1153 static bool isSSrc(unsigned RegClass) {
1154 return AMDGPU::SSrc_32RegClassID == RegClass ||
1155 AMDGPU::SSrc_64RegClassID == RegClass;
1158 /// \brief Analyze the possible immediate value Op
1160 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
1161 /// and the immediate value if it's a literal immediate
1162 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
1169 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
1170 if (Node->getZExtValue() >> 32) {
1173 Imm.I = Node->getSExtValue();
1174 } else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
1175 if (N->getValueType(0) != MVT::f32)
1177 Imm.F = Node->getValueAPF().convertToFloat();
1179 return -1; // It isn't an immediate
1181 if ((Imm.I >= -16 && Imm.I <= 64) ||
1182 Imm.F == 0.5f || Imm.F == -0.5f ||
1183 Imm.F == 1.0f || Imm.F == -1.0f ||
1184 Imm.F == 2.0f || Imm.F == -2.0f ||
1185 Imm.F == 4.0f || Imm.F == -4.0f)
1186 return 0; // It's an inline immediate
1188 return Imm.I; // It's a literal immediate
1191 /// \brief Try to fold an immediate directly into an instruction
1192 bool SITargetLowering::foldImm(SDValue &Operand, int32_t &Immediate,
1193 bool &ScalarSlotUsed) const {
1195 MachineSDNode *Mov = dyn_cast<MachineSDNode>(Operand);
1196 const SIInstrInfo *TII =
1197 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1198 if (!Mov || !TII->isMov(Mov->getMachineOpcode()))
1201 const SDValue &Op = Mov->getOperand(0);
1202 int32_t Value = analyzeImmediate(Op.getNode());
1204 // Not an immediate at all
1207 } else if (Value == 0) {
1208 // Inline immediates can always be fold
1212 } else if (Value == Immediate) {
1213 // Already fold literal immediate
1217 } else if (!ScalarSlotUsed && !Immediate) {
1218 // Fold this literal immediate
1219 ScalarSlotUsed = true;
1229 const TargetRegisterClass *SITargetLowering::getRegClassForNode(
1230 SelectionDAG &DAG, const SDValue &Op) const {
1231 const SIInstrInfo *TII =
1232 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1233 const SIRegisterInfo &TRI = TII->getRegisterInfo();
1235 if (!Op->isMachineOpcode()) {
1236 switch(Op->getOpcode()) {
1237 case ISD::CopyFromReg: {
1238 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1239 unsigned Reg = cast<RegisterSDNode>(Op->getOperand(1))->getReg();
1240 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
1241 return MRI.getRegClass(Reg);
1243 return TRI.getPhysRegClass(Reg);
1245 default: return nullptr;
1248 const MCInstrDesc &Desc = TII->get(Op->getMachineOpcode());
1249 int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
1250 if (OpClassID != -1) {
1251 return TRI.getRegClass(OpClassID);
1253 switch(Op.getMachineOpcode()) {
1254 case AMDGPU::COPY_TO_REGCLASS:
1255 // Operand 1 is the register class id for COPY_TO_REGCLASS instructions.
1256 OpClassID = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
1258 // If the COPY_TO_REGCLASS instruction is copying to a VSrc register
1259 // class, then the register class for the value could be either a
1260 // VReg or and SReg. In order to get a more accurate
1261 if (OpClassID == AMDGPU::VSrc_32RegClassID ||
1262 OpClassID == AMDGPU::VSrc_64RegClassID) {
1263 return getRegClassForNode(DAG, Op.getOperand(0));
1265 return TRI.getRegClass(OpClassID);
1266 case AMDGPU::EXTRACT_SUBREG: {
1267 int SubIdx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1268 const TargetRegisterClass *SuperClass =
1269 getRegClassForNode(DAG, Op.getOperand(0));
1270 return TRI.getSubClassWithSubReg(SuperClass, SubIdx);
1272 case AMDGPU::REG_SEQUENCE:
1273 // Operand 0 is the register class id for REG_SEQUENCE instructions.
1274 return TRI.getRegClass(
1275 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue());
1277 return getRegClassFor(Op.getSimpleValueType());
1281 /// \brief Does "Op" fit into register class "RegClass" ?
1282 bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, const SDValue &Op,
1283 unsigned RegClass) const {
1284 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1285 const TargetRegisterClass *RC = getRegClassForNode(DAG, Op);
1289 return TRI->getRegClass(RegClass)->hasSubClassEq(RC);
1292 /// \brief Make sure that we don't exeed the number of allowed scalars
1293 void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
1295 bool &ScalarSlotUsed) const {
1297 // First map the operands register class to a destination class
1298 if (RegClass == AMDGPU::VSrc_32RegClassID)
1299 RegClass = AMDGPU::VReg_32RegClassID;
1300 else if (RegClass == AMDGPU::VSrc_64RegClassID)
1301 RegClass = AMDGPU::VReg_64RegClassID;
1305 // Nothing to do if they fit naturally
1306 if (fitsRegClass(DAG, Operand, RegClass))
1309 // If the scalar slot isn't used yet use it now
1310 if (!ScalarSlotUsed) {
1311 ScalarSlotUsed = true;
1315 // This is a conservative aproach. It is possible that we can't determine the
1316 // correct register class and copy too often, but better safe than sorry.
1317 SDValue RC = DAG.getTargetConstant(RegClass, MVT::i32);
1318 SDNode *Node = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, SDLoc(),
1319 Operand.getValueType(), Operand, RC);
1320 Operand = SDValue(Node, 0);
1323 /// \returns true if \p Node's operands are different from the SDValue list
1325 static bool isNodeChanged(const SDNode *Node, const std::vector<SDValue> &Ops) {
1326 for (unsigned i = 0, e = Node->getNumOperands(); i < e; ++i) {
1327 if (Ops[i].getNode() != Node->getOperand(i).getNode()) {
1334 /// \brief Try to fold the Nodes operands into the Node
1335 SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
1336 SelectionDAG &DAG) const {
1338 // Original encoding (either e32 or e64)
1339 int Opcode = Node->getMachineOpcode();
1340 const SIInstrInfo *TII =
1341 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1342 const MCInstrDesc *Desc = &TII->get(Opcode);
1344 unsigned NumDefs = Desc->getNumDefs();
1345 unsigned NumOps = Desc->getNumOperands();
1347 // Commuted opcode if available
1348 int OpcodeRev = Desc->isCommutable() ? TII->commuteOpcode(Opcode) : -1;
1349 const MCInstrDesc *DescRev = OpcodeRev == -1 ? nullptr : &TII->get(OpcodeRev);
1351 assert(!DescRev || DescRev->getNumDefs() == NumDefs);
1352 assert(!DescRev || DescRev->getNumOperands() == NumOps);
1354 // e64 version if available, -1 otherwise
1355 int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
1356 const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? nullptr : &TII->get(OpcodeE64);
1357 int InputModifiers[3] = {0};
1359 assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
1361 int32_t Immediate = Desc->getSize() == 4 ? 0 : -1;
1362 bool HaveVSrc = false, HaveSSrc = false;
1364 // First figure out what we already have in this instruction.
1365 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1366 i != e && Op < NumOps; ++i, ++Op) {
1368 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1369 if (isVSrc(RegClass))
1371 else if (isSSrc(RegClass))
1376 int32_t Imm = analyzeImmediate(Node->getOperand(i).getNode());
1377 if (Imm != -1 && Imm != 0) {
1378 // Literal immediate
1383 // If we neither have VSrc nor SSrc, it makes no sense to continue.
1384 if (!HaveVSrc && !HaveSSrc)
1387 // No scalar allowed when we have both VSrc and SSrc
1388 bool ScalarSlotUsed = HaveVSrc && HaveSSrc;
1390 // Second go over the operands and try to fold them
1391 std::vector<SDValue> Ops;
1392 bool Promote2e64 = false;
1393 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
1394 i != e && Op < NumOps; ++i, ++Op) {
1396 const SDValue &Operand = Node->getOperand(i);
1397 Ops.push_back(Operand);
1399 // Already folded immediate?
1400 if (isa<ConstantSDNode>(Operand.getNode()) ||
1401 isa<ConstantFPSDNode>(Operand.getNode()))
1404 // Is this a VSrc or SSrc operand?
1405 unsigned RegClass = Desc->OpInfo[Op].RegClass;
1406 if (isVSrc(RegClass) || isSSrc(RegClass)) {
1407 // Try to fold the immediates
1408 if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
1409 // Folding didn't work, make sure we don't hit the SReg limit.
1410 ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
1415 if (i == 1 && DescRev && fitsRegClass(DAG, Ops[0], RegClass)) {
1417 unsigned OtherRegClass = Desc->OpInfo[NumDefs].RegClass;
1418 assert(isVSrc(OtherRegClass) || isSSrc(OtherRegClass));
1420 // Test if it makes sense to swap operands
1421 if (foldImm(Ops[1], Immediate, ScalarSlotUsed) ||
1422 (!fitsRegClass(DAG, Ops[1], RegClass) &&
1423 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1425 // Swap commutable operands
1426 std::swap(Ops[0], Ops[1]);
1438 // Test if it makes sense to switch to e64 encoding
1439 unsigned OtherRegClass = DescE64->OpInfo[Op].RegClass;
1440 if (!isVSrc(OtherRegClass) && !isSSrc(OtherRegClass))
1443 int32_t TmpImm = -1;
1444 if (foldImm(Ops[i], TmpImm, ScalarSlotUsed) ||
1445 (!fitsRegClass(DAG, Ops[i], RegClass) &&
1446 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
1448 // Switch to e64 encoding
1456 if (!DescE64 && !Promote2e64)
1458 if (!Operand.isMachineOpcode())
1460 if (Operand.getMachineOpcode() == AMDGPU::FNEG_SI) {
1462 Ops.push_back(Operand.getOperand(0));
1463 InputModifiers[i] = 1;
1470 else if (Operand.getMachineOpcode() == AMDGPU::FABS_SI) {
1472 Ops.push_back(Operand.getOperand(0));
1473 InputModifiers[i] = 2;
1483 std::vector<SDValue> OldOps(Ops);
1485 for (unsigned i = 0; i < OldOps.size(); ++i) {
1487 Ops.push_back(DAG.getTargetConstant(InputModifiers[i], MVT::i32));
1488 Ops.push_back(OldOps[i]);
1490 // Add the modifier flags while promoting
1491 for (unsigned i = 0; i < 2; ++i)
1492 Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
1495 // Add optional chain and glue
1496 for (unsigned i = NumOps - NumDefs, e = Node->getNumOperands(); i < e; ++i)
1497 Ops.push_back(Node->getOperand(i));
1499 // Nodes that have a glue result are not CSE'd by getMachineNode(), so in
1500 // this case a brand new node is always be created, even if the operands
1501 // are the same as before. So, manually check if anything has been changed.
1502 if (Desc->Opcode == Opcode && !isNodeChanged(Node, Ops)) {
1506 // Create a complete new instruction
1507 return DAG.getMachineNode(Desc->Opcode, SDLoc(Node), Node->getVTList(), Ops);
1510 /// \brief Helper function for adjustWritemask
1511 static unsigned SubIdx2Lane(unsigned Idx) {
1514 case AMDGPU::sub0: return 0;
1515 case AMDGPU::sub1: return 1;
1516 case AMDGPU::sub2: return 2;
1517 case AMDGPU::sub3: return 3;
1521 /// \brief Adjust the writemask of MIMG instructions
1522 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
1523 SelectionDAG &DAG) const {
1524 SDNode *Users[4] = { };
1526 unsigned OldDmask = Node->getConstantOperandVal(0);
1527 unsigned NewDmask = 0;
1529 // Try to figure out the used register components
1530 for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
1533 // Abort if we can't understand the usage
1534 if (!I->isMachineOpcode() ||
1535 I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
1538 // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used.
1539 // Note that subregs are packed, i.e. Lane==0 is the first bit set
1540 // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit
1542 Lane = SubIdx2Lane(I->getConstantOperandVal(1));
1544 // Set which texture component corresponds to the lane.
1546 for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) {
1548 Comp = countTrailingZeros(Dmask);
1549 Dmask &= ~(1 << Comp);
1552 // Abort if we have more than one user per component
1557 NewDmask |= 1 << Comp;
1560 // Abort if there's no change
1561 if (NewDmask == OldDmask)
1564 // Adjust the writemask in the node
1565 std::vector<SDValue> Ops;
1566 Ops.push_back(DAG.getTargetConstant(NewDmask, MVT::i32));
1567 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
1568 Ops.push_back(Node->getOperand(i));
1569 Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops);
1571 // If we only got one lane, replace it with a copy
1572 // (if NewDmask has only one bit set...)
1573 if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
1574 SDValue RC = DAG.getTargetConstant(AMDGPU::VReg_32RegClassID, MVT::i32);
1575 SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1576 SDLoc(), Users[Lane]->getValueType(0),
1577 SDValue(Node, 0), RC);
1578 DAG.ReplaceAllUsesWith(Users[Lane], Copy);
1582 // Update the users of the node with the new indices
1583 for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
1585 SDNode *User = Users[i];
1589 SDValue Op = DAG.getTargetConstant(Idx, MVT::i32);
1590 DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
1594 case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
1595 case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
1596 case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
1601 /// \brief Fold the instructions after selecting them.
1602 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
1603 SelectionDAG &DAG) const {
1604 const SIInstrInfo *TII =
1605 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1606 Node = AdjustRegClass(Node, DAG);
1608 if (TII->isMIMG(Node->getMachineOpcode()))
1609 adjustWritemask(Node, DAG);
1611 return foldOperands(Node, DAG);
1614 /// \brief Assign the register class depending on the number of
1615 /// bits set in the writemask
1616 void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
1617 SDNode *Node) const {
1618 const SIInstrInfo *TII =
1619 static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
1620 if (!TII->isMIMG(MI->getOpcode()))
1623 unsigned VReg = MI->getOperand(0).getReg();
1624 unsigned Writemask = MI->getOperand(1).getImm();
1625 unsigned BitsSet = 0;
1626 for (unsigned i = 0; i < 4; ++i)
1627 BitsSet += Writemask & (1 << i) ? 1 : 0;
1629 const TargetRegisterClass *RC;
1632 case 1: RC = &AMDGPU::VReg_32RegClass; break;
1633 case 2: RC = &AMDGPU::VReg_64RegClass; break;
1634 case 3: RC = &AMDGPU::VReg_96RegClass; break;
1637 unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
1638 MI->setDesc(TII->get(NewOpcode));
1639 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1640 MRI.setRegClass(VReg, RC);
1643 MachineSDNode *SITargetLowering::AdjustRegClass(MachineSDNode *N,
1644 SelectionDAG &DAG) const {
1647 unsigned NewOpcode = N->getMachineOpcode();
1649 switch (N->getMachineOpcode()) {
1651 case AMDGPU::S_LOAD_DWORD_IMM:
1652 NewOpcode = AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1654 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1655 if (NewOpcode == N->getMachineOpcode()) {
1656 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1659 case AMDGPU::S_LOAD_DWORDX4_IMM:
1660 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1661 if (NewOpcode == N->getMachineOpcode()) {
1662 NewOpcode = AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1664 if (fitsRegClass(DAG, N->getOperand(0), AMDGPU::SReg_64RegClassID)) {
1667 ConstantSDNode *Offset = cast<ConstantSDNode>(N->getOperand(1));
1669 SDValue(DAG.getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::i128,
1670 DAG.getConstant(0, MVT::i64)), 0),
1672 DAG.getConstant(Offset->getSExtValue() << 2, MVT::i32)
1674 return DAG.getMachineNode(NewOpcode, DL, N->getVTList(), Ops);
1679 SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1680 const TargetRegisterClass *RC,
1681 unsigned Reg, EVT VT) const {
1682 SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT);
1684 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()),
1685 cast<RegisterSDNode>(VReg)->getReg(), VT);