1 //===-- AMDGPUISelDAGToDAG.cpp - A dag to dag inst selector for AMDGPU ----===//
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 /// Defines an instruction selector for the AMDGPU target.
13 //===----------------------------------------------------------------------===//
16 #include "AMDGPUArgumentUsageInfo.h"
17 #include "AMDGPUISelLowering.h" // For AMDGPUISD
18 #include "AMDGPUInstrInfo.h"
19 #include "AMDGPUPerfHintAnalysis.h"
20 #include "AMDGPURegisterInfo.h"
21 #include "AMDGPUSubtarget.h"
22 #include "AMDGPUTargetMachine.h"
23 #include "SIDefines.h"
24 #include "SIISelLowering.h"
25 #include "SIInstrInfo.h"
26 #include "SIMachineFunctionInfo.h"
27 #include "SIRegisterInfo.h"
28 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
29 #include "llvm/ADT/APInt.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/StringRef.h"
32 #include "llvm/Analysis/LegacyDivergenceAnalysis.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/CodeGen/FunctionLoweringInfo.h"
35 #include "llvm/CodeGen/ISDOpcodes.h"
36 #include "llvm/CodeGen/MachineFunction.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/SelectionDAG.h"
39 #include "llvm/CodeGen/SelectionDAGISel.h"
40 #include "llvm/CodeGen/SelectionDAGNodes.h"
41 #include "llvm/CodeGen/ValueTypes.h"
42 #include "llvm/IR/BasicBlock.h"
43 #include "llvm/IR/Instruction.h"
44 #include "llvm/MC/MCInstrDesc.h"
45 #include "llvm/Support/Casting.h"
46 #include "llvm/Support/CodeGen.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/MachineValueType.h"
49 #include "llvm/Support/MathExtras.h"
61 } // end namespace llvm
63 //===----------------------------------------------------------------------===//
64 // Instruction Selector Implementation
65 //===----------------------------------------------------------------------===//
69 /// AMDGPU specific code to select AMDGPU machine instructions for
70 /// SelectionDAG operations.
71 class AMDGPUDAGToDAGISel : public SelectionDAGISel {
72 // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can
73 // make the right decision when generating code for different targets.
74 const GCNSubtarget *Subtarget;
75 bool EnableLateStructurizeCFG;
78 explicit AMDGPUDAGToDAGISel(TargetMachine *TM = nullptr,
79 CodeGenOpt::Level OptLevel = CodeGenOpt::Default)
80 : SelectionDAGISel(*TM, OptLevel) {
81 EnableLateStructurizeCFG = AMDGPUTargetMachine::EnableLateStructurizeCFG;
83 ~AMDGPUDAGToDAGISel() override = default;
85 void getAnalysisUsage(AnalysisUsage &AU) const override {
86 AU.addRequired<AMDGPUArgumentUsageInfo>();
87 AU.addRequired<AMDGPUPerfHintAnalysis>();
88 AU.addRequired<LegacyDivergenceAnalysis>();
89 SelectionDAGISel::getAnalysisUsage(AU);
92 bool runOnMachineFunction(MachineFunction &MF) override;
93 void Select(SDNode *N) override;
94 StringRef getPassName() const override;
95 void PostprocessISelDAG() override;
98 void SelectBuildVector(SDNode *N, unsigned RegClassID);
101 std::pair<SDValue, SDValue> foldFrameIndex(SDValue N) const;
102 bool isNoNanSrc(SDValue N) const;
103 bool isInlineImmediate(const SDNode *N) const;
104 bool isVGPRImm(const SDNode *N) const;
105 bool isUniformLoad(const SDNode *N) const;
106 bool isUniformBr(const SDNode *N) const;
108 MachineSDNode *buildSMovImm64(SDLoc &DL, uint64_t Val, EVT VT) const;
110 SDNode *glueCopyToM0(SDNode *N) const;
112 const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const;
113 virtual bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset);
114 virtual bool SelectADDRIndirect(SDValue Addr, SDValue &Base, SDValue &Offset);
115 bool isDSOffsetLegal(const SDValue &Base, unsigned Offset,
116 unsigned OffsetBits) const;
117 bool SelectDS1Addr1Offset(SDValue Ptr, SDValue &Base, SDValue &Offset) const;
118 bool SelectDS64Bit4ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0,
119 SDValue &Offset1) const;
120 bool SelectMUBUF(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
121 SDValue &SOffset, SDValue &Offset, SDValue &Offen,
122 SDValue &Idxen, SDValue &Addr64, SDValue &GLC, SDValue &SLC,
124 bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
125 SDValue &SOffset, SDValue &Offset, SDValue &GLC,
126 SDValue &SLC, SDValue &TFE) const;
127 bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
128 SDValue &VAddr, SDValue &SOffset, SDValue &Offset,
130 bool SelectMUBUFScratchOffen(SDNode *Parent,
131 SDValue Addr, SDValue &RSrc, SDValue &VAddr,
132 SDValue &SOffset, SDValue &ImmOffset) const;
133 bool SelectMUBUFScratchOffset(SDNode *Parent,
134 SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
135 SDValue &Offset) const;
137 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset,
138 SDValue &Offset, SDValue &GLC, SDValue &SLC,
140 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
141 SDValue &Offset, SDValue &SLC) const;
142 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
143 SDValue &Offset) const;
145 bool SelectFlatAtomic(SDValue Addr, SDValue &VAddr,
146 SDValue &Offset, SDValue &SLC) const;
147 bool SelectFlatAtomicSigned(SDValue Addr, SDValue &VAddr,
148 SDValue &Offset, SDValue &SLC) const;
150 template <bool IsSigned>
151 bool SelectFlatOffset(SDValue Addr, SDValue &VAddr,
152 SDValue &Offset, SDValue &SLC) const;
154 bool SelectSMRDOffset(SDValue ByteOffsetNode, SDValue &Offset,
156 SDValue Expand32BitAddress(SDValue Addr) const;
157 bool SelectSMRD(SDValue Addr, SDValue &SBase, SDValue &Offset,
159 bool SelectSMRDImm(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
160 bool SelectSMRDImm32(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
161 bool SelectSMRDSgpr(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
162 bool SelectSMRDBufferImm(SDValue Addr, SDValue &Offset) const;
163 bool SelectSMRDBufferImm32(SDValue Addr, SDValue &Offset) const;
164 bool SelectMOVRELOffset(SDValue Index, SDValue &Base, SDValue &Offset) const;
166 bool SelectVOP3Mods_NNaN(SDValue In, SDValue &Src, SDValue &SrcMods) const;
167 bool SelectVOP3ModsImpl(SDValue In, SDValue &Src, unsigned &SrcMods) const;
168 bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
169 bool SelectVOP3NoMods(SDValue In, SDValue &Src) const;
170 bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
171 SDValue &Clamp, SDValue &Omod) const;
172 bool SelectVOP3NoMods0(SDValue In, SDValue &Src, SDValue &SrcMods,
173 SDValue &Clamp, SDValue &Omod) const;
175 bool SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src, SDValue &SrcMods,
177 SDValue &Omod) const;
179 bool SelectVOP3OMods(SDValue In, SDValue &Src,
180 SDValue &Clamp, SDValue &Omod) const;
182 bool SelectVOP3PMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
183 bool SelectVOP3PMods0(SDValue In, SDValue &Src, SDValue &SrcMods,
184 SDValue &Clamp) const;
186 bool SelectVOP3OpSel(SDValue In, SDValue &Src, SDValue &SrcMods) const;
187 bool SelectVOP3OpSel0(SDValue In, SDValue &Src, SDValue &SrcMods,
188 SDValue &Clamp) const;
190 bool SelectVOP3OpSelMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
191 bool SelectVOP3OpSelMods0(SDValue In, SDValue &Src, SDValue &SrcMods,
192 SDValue &Clamp) const;
193 bool SelectVOP3PMadMixModsImpl(SDValue In, SDValue &Src, unsigned &Mods) const;
194 bool SelectVOP3PMadMixMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
196 bool SelectHi16Elt(SDValue In, SDValue &Src) const;
198 void SelectADD_SUB_I64(SDNode *N);
199 void SelectUADDO_USUBO(SDNode *N);
200 void SelectDIV_SCALE(SDNode *N);
201 void SelectMAD_64_32(SDNode *N);
202 void SelectFMA_W_CHAIN(SDNode *N);
203 void SelectFMUL_W_CHAIN(SDNode *N);
205 SDNode *getS_BFE(unsigned Opcode, const SDLoc &DL, SDValue Val,
206 uint32_t Offset, uint32_t Width);
207 void SelectS_BFEFromShifts(SDNode *N);
208 void SelectS_BFE(SDNode *N);
209 bool isCBranchSCC(const SDNode *N) const;
210 void SelectBRCOND(SDNode *N);
211 void SelectFMAD_FMA(SDNode *N);
212 void SelectATOMIC_CMP_SWAP(SDNode *N);
215 // Include the pieces autogenerated from the target description.
216 #include "AMDGPUGenDAGISel.inc"
219 class R600DAGToDAGISel : public AMDGPUDAGToDAGISel {
220 const R600Subtarget *Subtarget;
222 bool isConstantLoad(const MemSDNode *N, int cbID) const;
223 bool SelectGlobalValueConstantOffset(SDValue Addr, SDValue& IntPtr);
224 bool SelectGlobalValueVariableOffset(SDValue Addr, SDValue &BaseReg,
227 explicit R600DAGToDAGISel(TargetMachine *TM, CodeGenOpt::Level OptLevel) :
228 AMDGPUDAGToDAGISel(TM, OptLevel) {}
230 void Select(SDNode *N) override;
232 bool SelectADDRIndirect(SDValue Addr, SDValue &Base,
233 SDValue &Offset) override;
234 bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
235 SDValue &Offset) override;
237 bool runOnMachineFunction(MachineFunction &MF) override;
239 // Include the pieces autogenerated from the target description.
240 #include "R600GenDAGISel.inc"
243 } // end anonymous namespace
245 INITIALIZE_PASS_BEGIN(AMDGPUDAGToDAGISel, "amdgpu-isel",
246 "AMDGPU DAG->DAG Pattern Instruction Selection", false, false)
247 INITIALIZE_PASS_DEPENDENCY(AMDGPUArgumentUsageInfo)
248 INITIALIZE_PASS_DEPENDENCY(AMDGPUPerfHintAnalysis)
249 INITIALIZE_PASS_DEPENDENCY(LegacyDivergenceAnalysis)
250 INITIALIZE_PASS_END(AMDGPUDAGToDAGISel, "amdgpu-isel",
251 "AMDGPU DAG->DAG Pattern Instruction Selection", false, false)
253 /// This pass converts a legalized DAG into a AMDGPU-specific
254 // DAG, ready for instruction scheduling.
255 FunctionPass *llvm::createAMDGPUISelDag(TargetMachine *TM,
256 CodeGenOpt::Level OptLevel) {
257 return new AMDGPUDAGToDAGISel(TM, OptLevel);
260 /// This pass converts a legalized DAG into a R600-specific
261 // DAG, ready for instruction scheduling.
262 FunctionPass *llvm::createR600ISelDag(TargetMachine *TM,
263 CodeGenOpt::Level OptLevel) {
264 return new R600DAGToDAGISel(TM, OptLevel);
267 bool AMDGPUDAGToDAGISel::runOnMachineFunction(MachineFunction &MF) {
268 Subtarget = &MF.getSubtarget<GCNSubtarget>();
269 return SelectionDAGISel::runOnMachineFunction(MF);
272 bool AMDGPUDAGToDAGISel::isNoNanSrc(SDValue N) const {
273 if (TM.Options.NoNaNsFPMath)
276 // TODO: Move into isKnownNeverNaN
277 if (N->getFlags().isDefined())
278 return N->getFlags().hasNoNaNs();
280 return CurDAG->isKnownNeverNaN(N);
283 bool AMDGPUDAGToDAGISel::isInlineImmediate(const SDNode *N) const {
284 const SIInstrInfo *TII = Subtarget->getInstrInfo();
286 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N))
287 return TII->isInlineConstant(C->getAPIntValue());
289 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N))
290 return TII->isInlineConstant(C->getValueAPF().bitcastToAPInt());
295 /// Determine the register class for \p OpNo
296 /// \returns The register class of the virtual register that will be used for
297 /// the given operand number \OpNo or NULL if the register class cannot be
299 const TargetRegisterClass *AMDGPUDAGToDAGISel::getOperandRegClass(SDNode *N,
300 unsigned OpNo) const {
301 if (!N->isMachineOpcode()) {
302 if (N->getOpcode() == ISD::CopyToReg) {
303 unsigned Reg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
304 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
305 MachineRegisterInfo &MRI = CurDAG->getMachineFunction().getRegInfo();
306 return MRI.getRegClass(Reg);
309 const SIRegisterInfo *TRI
310 = static_cast<const GCNSubtarget *>(Subtarget)->getRegisterInfo();
311 return TRI->getPhysRegClass(Reg);
317 switch (N->getMachineOpcode()) {
319 const MCInstrDesc &Desc =
320 Subtarget->getInstrInfo()->get(N->getMachineOpcode());
321 unsigned OpIdx = Desc.getNumDefs() + OpNo;
322 if (OpIdx >= Desc.getNumOperands())
324 int RegClass = Desc.OpInfo[OpIdx].RegClass;
328 return Subtarget->getRegisterInfo()->getRegClass(RegClass);
330 case AMDGPU::REG_SEQUENCE: {
331 unsigned RCID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
332 const TargetRegisterClass *SuperRC =
333 Subtarget->getRegisterInfo()->getRegClass(RCID);
335 SDValue SubRegOp = N->getOperand(OpNo + 1);
336 unsigned SubRegIdx = cast<ConstantSDNode>(SubRegOp)->getZExtValue();
337 return Subtarget->getRegisterInfo()->getSubClassWithSubReg(SuperRC,
343 SDNode *AMDGPUDAGToDAGISel::glueCopyToM0(SDNode *N) const {
344 if (cast<MemSDNode>(N)->getAddressSpace() != AMDGPUAS::LOCAL_ADDRESS ||
345 !Subtarget->ldsRequiresM0Init())
348 const SITargetLowering& Lowering =
349 *static_cast<const SITargetLowering*>(getTargetLowering());
351 // Write max value to m0 before each load operation
353 SDValue M0 = Lowering.copyToM0(*CurDAG, CurDAG->getEntryNode(), SDLoc(N),
354 CurDAG->getTargetConstant(-1, SDLoc(N), MVT::i32));
356 SDValue Glue = M0.getValue(1);
358 SmallVector <SDValue, 8> Ops;
359 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
360 Ops.push_back(N->getOperand(i));
363 return CurDAG->MorphNodeTo(N, N->getOpcode(), N->getVTList(), Ops);
366 MachineSDNode *AMDGPUDAGToDAGISel::buildSMovImm64(SDLoc &DL, uint64_t Imm,
368 SDNode *Lo = CurDAG->getMachineNode(
369 AMDGPU::S_MOV_B32, DL, MVT::i32,
370 CurDAG->getConstant(Imm & 0xFFFFFFFF, DL, MVT::i32));
372 CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32,
373 CurDAG->getConstant(Imm >> 32, DL, MVT::i32));
374 const SDValue Ops[] = {
375 CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32),
376 SDValue(Lo, 0), CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
377 SDValue(Hi, 0), CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32)};
379 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL, VT, Ops);
382 static unsigned selectSGPRVectorRegClassID(unsigned NumVectorElts) {
383 switch (NumVectorElts) {
385 return AMDGPU::SReg_32_XM0RegClassID;
387 return AMDGPU::SReg_64RegClassID;
389 return AMDGPU::SReg_128RegClassID;
391 return AMDGPU::SReg_256RegClassID;
393 return AMDGPU::SReg_512RegClassID;
396 llvm_unreachable("invalid vector size");
399 static bool getConstantValue(SDValue N, uint32_t &Out) {
400 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
401 Out = C->getAPIntValue().getZExtValue();
405 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N)) {
406 Out = C->getValueAPF().bitcastToAPInt().getZExtValue();
413 void AMDGPUDAGToDAGISel::SelectBuildVector(SDNode *N, unsigned RegClassID) {
414 EVT VT = N->getValueType(0);
415 unsigned NumVectorElts = VT.getVectorNumElements();
416 EVT EltVT = VT.getVectorElementType();
418 SDValue RegClass = CurDAG->getTargetConstant(RegClassID, DL, MVT::i32);
420 if (NumVectorElts == 1) {
421 CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS, EltVT, N->getOperand(0),
426 assert(NumVectorElts <= 16 && "Vectors with more than 16 elements not "
428 // 16 = Max Num Vector Elements
429 // 2 = 2 REG_SEQUENCE operands per element (value, subreg index)
430 // 1 = Vector Register Class
431 SmallVector<SDValue, 16 * 2 + 1> RegSeqArgs(NumVectorElts * 2 + 1);
433 RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, DL, MVT::i32);
434 bool IsRegSeq = true;
435 unsigned NOps = N->getNumOperands();
436 for (unsigned i = 0; i < NOps; i++) {
437 // XXX: Why is this here?
438 if (isa<RegisterSDNode>(N->getOperand(i))) {
442 unsigned Sub = AMDGPURegisterInfo::getSubRegFromChannel(i);
443 RegSeqArgs[1 + (2 * i)] = N->getOperand(i);
444 RegSeqArgs[1 + (2 * i) + 1] = CurDAG->getTargetConstant(Sub, DL, MVT::i32);
446 if (NOps != NumVectorElts) {
447 // Fill in the missing undef elements if this was a scalar_to_vector.
448 assert(N->getOpcode() == ISD::SCALAR_TO_VECTOR && NOps < NumVectorElts);
449 MachineSDNode *ImpDef = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
451 for (unsigned i = NOps; i < NumVectorElts; ++i) {
452 unsigned Sub = AMDGPURegisterInfo::getSubRegFromChannel(i);
453 RegSeqArgs[1 + (2 * i)] = SDValue(ImpDef, 0);
454 RegSeqArgs[1 + (2 * i) + 1] =
455 CurDAG->getTargetConstant(Sub, DL, MVT::i32);
461 CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(), RegSeqArgs);
464 void AMDGPUDAGToDAGISel::Select(SDNode *N) {
465 unsigned int Opc = N->getOpcode();
466 if (N->isMachineOpcode()) {
468 return; // Already selected.
471 if (isa<AtomicSDNode>(N) ||
472 (Opc == AMDGPUISD::ATOMIC_INC || Opc == AMDGPUISD::ATOMIC_DEC ||
473 Opc == AMDGPUISD::ATOMIC_LOAD_FADD ||
474 Opc == AMDGPUISD::ATOMIC_LOAD_FMIN ||
475 Opc == AMDGPUISD::ATOMIC_LOAD_FMAX))
481 // We are selecting i64 ADD here instead of custom lower it during
482 // DAG legalization, so we can fold some i64 ADDs used for address
483 // calculation into the LOAD and STORE instructions.
488 if (N->getValueType(0) != MVT::i64)
491 SelectADD_SUB_I64(N);
496 SelectUADDO_USUBO(N);
499 case AMDGPUISD::FMUL_W_CHAIN: {
500 SelectFMUL_W_CHAIN(N);
503 case AMDGPUISD::FMA_W_CHAIN: {
504 SelectFMA_W_CHAIN(N);
508 case ISD::SCALAR_TO_VECTOR:
509 case ISD::BUILD_VECTOR: {
510 EVT VT = N->getValueType(0);
511 unsigned NumVectorElts = VT.getVectorNumElements();
512 if (VT.getScalarSizeInBits() == 16) {
513 if (Opc == ISD::BUILD_VECTOR && NumVectorElts == 2) {
514 uint32_t LHSVal, RHSVal;
515 if (getConstantValue(N->getOperand(0), LHSVal) &&
516 getConstantValue(N->getOperand(1), RHSVal)) {
517 uint32_t K = LHSVal | (RHSVal << 16);
518 CurDAG->SelectNodeTo(N, AMDGPU::S_MOV_B32, VT,
519 CurDAG->getTargetConstant(K, SDLoc(N), MVT::i32));
527 assert(VT.getVectorElementType().bitsEq(MVT::i32));
528 unsigned RegClassID = selectSGPRVectorRegClassID(NumVectorElts);
529 SelectBuildVector(N, RegClassID);
532 case ISD::BUILD_PAIR: {
533 SDValue RC, SubReg0, SubReg1;
535 if (N->getValueType(0) == MVT::i128) {
536 RC = CurDAG->getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32);
537 SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0_sub1, DL, MVT::i32);
538 SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub2_sub3, DL, MVT::i32);
539 } else if (N->getValueType(0) == MVT::i64) {
540 RC = CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32);
541 SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
542 SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
544 llvm_unreachable("Unhandled value type for BUILD_PAIR");
546 const SDValue Ops[] = { RC, N->getOperand(0), SubReg0,
547 N->getOperand(1), SubReg1 };
548 ReplaceNode(N, CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL,
549 N->getValueType(0), Ops));
554 case ISD::ConstantFP: {
555 if (N->getValueType(0).getSizeInBits() != 64 || isInlineImmediate(N))
559 if (ConstantFPSDNode *FP = dyn_cast<ConstantFPSDNode>(N))
560 Imm = FP->getValueAPF().bitcastToAPInt().getZExtValue();
562 ConstantSDNode *C = cast<ConstantSDNode>(N);
563 Imm = C->getZExtValue();
567 ReplaceNode(N, buildSMovImm64(DL, Imm, N->getValueType(0)));
572 case ISD::ATOMIC_LOAD:
573 case ISD::ATOMIC_STORE: {
578 case AMDGPUISD::BFE_I32:
579 case AMDGPUISD::BFE_U32: {
580 // There is a scalar version available, but unlike the vector version which
581 // has a separate operand for the offset and width, the scalar version packs
582 // the width and offset into a single operand. Try to move to the scalar
583 // version if the offsets are constant, so that we can try to keep extended
584 // loads of kernel arguments in SGPRs.
586 // TODO: Technically we could try to pattern match scalar bitshifts of
587 // dynamic values, but it's probably not useful.
588 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
592 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
596 bool Signed = Opc == AMDGPUISD::BFE_I32;
598 uint32_t OffsetVal = Offset->getZExtValue();
599 uint32_t WidthVal = Width->getZExtValue();
601 ReplaceNode(N, getS_BFE(Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32,
602 SDLoc(N), N->getOperand(0), OffsetVal, WidthVal));
605 case AMDGPUISD::DIV_SCALE: {
609 case AMDGPUISD::MAD_I64_I32:
610 case AMDGPUISD::MAD_U64_U32: {
614 case ISD::CopyToReg: {
615 const SITargetLowering& Lowering =
616 *static_cast<const SITargetLowering*>(getTargetLowering());
617 N = Lowering.legalizeTargetIndependentNode(N, *CurDAG);
623 case ISD::SIGN_EXTEND_INREG:
624 if (N->getValueType(0) != MVT::i32)
636 case AMDGPUISD::ATOMIC_CMP_SWAP:
637 SelectATOMIC_CMP_SWAP(N);
639 case AMDGPUISD::CVT_PKRTZ_F16_F32:
640 case AMDGPUISD::CVT_PKNORM_I16_F32:
641 case AMDGPUISD::CVT_PKNORM_U16_F32:
642 case AMDGPUISD::CVT_PK_U16_U32:
643 case AMDGPUISD::CVT_PK_I16_I32: {
644 // Hack around using a legal type if f16 is illegal.
645 if (N->getValueType(0) == MVT::i32) {
646 MVT NewVT = Opc == AMDGPUISD::CVT_PKRTZ_F16_F32 ? MVT::v2f16 : MVT::v2i16;
647 N = CurDAG->MorphNodeTo(N, N->getOpcode(), CurDAG->getVTList(NewVT),
648 { N->getOperand(0), N->getOperand(1) });
658 bool AMDGPUDAGToDAGISel::isUniformBr(const SDNode *N) const {
659 const BasicBlock *BB = FuncInfo->MBB->getBasicBlock();
660 const Instruction *Term = BB->getTerminator();
661 return Term->getMetadata("amdgpu.uniform") ||
662 Term->getMetadata("structurizecfg.uniform");
665 StringRef AMDGPUDAGToDAGISel::getPassName() const {
666 return "AMDGPU DAG->DAG Pattern Instruction Selection";
669 //===----------------------------------------------------------------------===//
671 //===----------------------------------------------------------------------===//
673 bool AMDGPUDAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
678 bool AMDGPUDAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base,
683 if ((C = dyn_cast<ConstantSDNode>(Addr))) {
684 Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32);
685 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
686 } else if ((Addr.getOpcode() == AMDGPUISD::DWORDADDR) &&
687 (C = dyn_cast<ConstantSDNode>(Addr.getOperand(0)))) {
688 Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32);
689 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
690 } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
691 (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) {
692 Base = Addr.getOperand(0);
693 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
696 Offset = CurDAG->getTargetConstant(0, DL, MVT::i32);
702 // FIXME: Should only handle addcarry/subcarry
703 void AMDGPUDAGToDAGISel::SelectADD_SUB_I64(SDNode *N) {
705 SDValue LHS = N->getOperand(0);
706 SDValue RHS = N->getOperand(1);
708 unsigned Opcode = N->getOpcode();
709 bool ConsumeCarry = (Opcode == ISD::ADDE || Opcode == ISD::SUBE);
711 ConsumeCarry || Opcode == ISD::ADDC || Opcode == ISD::SUBC;
712 bool IsAdd = Opcode == ISD::ADD || Opcode == ISD::ADDC || Opcode == ISD::ADDE;
714 SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
715 SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
717 SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
718 DL, MVT::i32, LHS, Sub0);
719 SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
720 DL, MVT::i32, LHS, Sub1);
722 SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
723 DL, MVT::i32, RHS, Sub0);
724 SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
725 DL, MVT::i32, RHS, Sub1);
727 SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue);
729 unsigned Opc = IsAdd ? AMDGPU::S_ADD_U32 : AMDGPU::S_SUB_U32;
730 unsigned CarryOpc = IsAdd ? AMDGPU::S_ADDC_U32 : AMDGPU::S_SUBB_U32;
734 SDValue Args[] = { SDValue(Lo0, 0), SDValue(Lo1, 0) };
735 AddLo = CurDAG->getMachineNode(Opc, DL, VTList, Args);
737 SDValue Args[] = { SDValue(Lo0, 0), SDValue(Lo1, 0), N->getOperand(2) };
738 AddLo = CurDAG->getMachineNode(CarryOpc, DL, VTList, Args);
740 SDValue AddHiArgs[] = {
745 SDNode *AddHi = CurDAG->getMachineNode(CarryOpc, DL, VTList, AddHiArgs);
747 SDValue RegSequenceArgs[] = {
748 CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32),
754 SDNode *RegSequence = CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL,
755 MVT::i64, RegSequenceArgs);
758 // Replace the carry-use
759 ReplaceUses(SDValue(N, 1), SDValue(AddHi, 1));
762 // Replace the remaining uses.
763 ReplaceNode(N, RegSequence);
766 void AMDGPUDAGToDAGISel::SelectUADDO_USUBO(SDNode *N) {
767 // The name of the opcodes are misleading. v_add_i32/v_sub_i32 have unsigned
768 // carry out despite the _i32 name. These were renamed in VI to _U32.
769 // FIXME: We should probably rename the opcodes here.
770 unsigned Opc = N->getOpcode() == ISD::UADDO ?
771 AMDGPU::V_ADD_I32_e64 : AMDGPU::V_SUB_I32_e64;
773 CurDAG->SelectNodeTo(N, Opc, N->getVTList(),
774 { N->getOperand(0), N->getOperand(1) });
777 void AMDGPUDAGToDAGISel::SelectFMA_W_CHAIN(SDNode *N) {
779 // src0_modifiers, src0, src1_modifiers, src1, src2_modifiers, src2, clamp, omod
782 SelectVOP3Mods0(N->getOperand(1), Ops[1], Ops[0], Ops[6], Ops[7]);
783 SelectVOP3Mods(N->getOperand(2), Ops[3], Ops[2]);
784 SelectVOP3Mods(N->getOperand(3), Ops[5], Ops[4]);
785 Ops[8] = N->getOperand(0);
786 Ops[9] = N->getOperand(4);
788 CurDAG->SelectNodeTo(N, AMDGPU::V_FMA_F32, N->getVTList(), Ops);
791 void AMDGPUDAGToDAGISel::SelectFMUL_W_CHAIN(SDNode *N) {
793 // src0_modifiers, src0, src1_modifiers, src1, clamp, omod
796 SelectVOP3Mods0(N->getOperand(1), Ops[1], Ops[0], Ops[4], Ops[5]);
797 SelectVOP3Mods(N->getOperand(2), Ops[3], Ops[2]);
798 Ops[6] = N->getOperand(0);
799 Ops[7] = N->getOperand(3);
801 CurDAG->SelectNodeTo(N, AMDGPU::V_MUL_F32_e64, N->getVTList(), Ops);
804 // We need to handle this here because tablegen doesn't support matching
805 // instructions with multiple outputs.
806 void AMDGPUDAGToDAGISel::SelectDIV_SCALE(SDNode *N) {
808 EVT VT = N->getValueType(0);
810 assert(VT == MVT::f32 || VT == MVT::f64);
813 = (VT == MVT::f64) ? AMDGPU::V_DIV_SCALE_F64 : AMDGPU::V_DIV_SCALE_F32;
815 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2) };
816 CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
819 // We need to handle this here because tablegen doesn't support matching
820 // instructions with multiple outputs.
821 void AMDGPUDAGToDAGISel::SelectMAD_64_32(SDNode *N) {
823 bool Signed = N->getOpcode() == AMDGPUISD::MAD_I64_I32;
824 unsigned Opc = Signed ? AMDGPU::V_MAD_I64_I32 : AMDGPU::V_MAD_U64_U32;
826 SDValue Clamp = CurDAG->getTargetConstant(0, SL, MVT::i1);
827 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
829 CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
832 bool AMDGPUDAGToDAGISel::isDSOffsetLegal(const SDValue &Base, unsigned Offset,
833 unsigned OffsetBits) const {
834 if ((OffsetBits == 16 && !isUInt<16>(Offset)) ||
835 (OffsetBits == 8 && !isUInt<8>(Offset)))
838 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS ||
839 Subtarget->unsafeDSOffsetFoldingEnabled())
842 // On Southern Islands instruction with a negative base value and an offset
843 // don't seem to work.
844 return CurDAG->SignBitIsZero(Base);
847 bool AMDGPUDAGToDAGISel::SelectDS1Addr1Offset(SDValue Addr, SDValue &Base,
848 SDValue &Offset) const {
850 if (CurDAG->isBaseWithConstantOffset(Addr)) {
851 SDValue N0 = Addr.getOperand(0);
852 SDValue N1 = Addr.getOperand(1);
853 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
854 if (isDSOffsetLegal(N0, C1->getSExtValue(), 16)) {
857 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16);
860 } else if (Addr.getOpcode() == ISD::SUB) {
861 // sub C, x -> add (sub 0, x), C
862 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr.getOperand(0))) {
863 int64_t ByteOffset = C->getSExtValue();
864 if (isUInt<16>(ByteOffset)) {
865 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
867 // XXX - This is kind of hacky. Create a dummy sub node so we can check
868 // the known bits in isDSOffsetLegal. We need to emit the selected node
869 // here, so this is thrown away.
870 SDValue Sub = CurDAG->getNode(ISD::SUB, DL, MVT::i32,
871 Zero, Addr.getOperand(1));
873 if (isDSOffsetLegal(Sub, ByteOffset, 16)) {
874 // FIXME: Select to VOP3 version for with-carry.
875 unsigned SubOp = Subtarget->hasAddNoCarry() ?
876 AMDGPU::V_SUB_U32_e64 : AMDGPU::V_SUB_I32_e32;
878 MachineSDNode *MachineSub
879 = CurDAG->getMachineNode(SubOp, DL, MVT::i32,
880 Zero, Addr.getOperand(1));
882 Base = SDValue(MachineSub, 0);
883 Offset = CurDAG->getTargetConstant(ByteOffset, DL, MVT::i16);
888 } else if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
889 // If we have a constant address, prefer to put the constant into the
890 // offset. This can save moves to load the constant address since multiple
891 // operations can share the zero base address register, and enables merging
892 // into read2 / write2 instructions.
896 if (isUInt<16>(CAddr->getZExtValue())) {
897 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
898 MachineSDNode *MovZero = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
900 Base = SDValue(MovZero, 0);
901 Offset = CurDAG->getTargetConstant(CAddr->getZExtValue(), DL, MVT::i16);
908 Offset = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i16);
912 // TODO: If offset is too big, put low 16-bit into offset.
913 bool AMDGPUDAGToDAGISel::SelectDS64Bit4ByteAligned(SDValue Addr, SDValue &Base,
915 SDValue &Offset1) const {
918 if (CurDAG->isBaseWithConstantOffset(Addr)) {
919 SDValue N0 = Addr.getOperand(0);
920 SDValue N1 = Addr.getOperand(1);
921 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
922 unsigned DWordOffset0 = C1->getZExtValue() / 4;
923 unsigned DWordOffset1 = DWordOffset0 + 1;
925 if (isDSOffsetLegal(N0, DWordOffset1, 8)) {
927 Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8);
928 Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8);
931 } else if (Addr.getOpcode() == ISD::SUB) {
932 // sub C, x -> add (sub 0, x), C
933 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr.getOperand(0))) {
934 unsigned DWordOffset0 = C->getZExtValue() / 4;
935 unsigned DWordOffset1 = DWordOffset0 + 1;
937 if (isUInt<8>(DWordOffset0)) {
939 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
941 // XXX - This is kind of hacky. Create a dummy sub node so we can check
942 // the known bits in isDSOffsetLegal. We need to emit the selected node
943 // here, so this is thrown away.
944 SDValue Sub = CurDAG->getNode(ISD::SUB, DL, MVT::i32,
945 Zero, Addr.getOperand(1));
947 if (isDSOffsetLegal(Sub, DWordOffset1, 8)) {
948 unsigned SubOp = Subtarget->hasAddNoCarry() ?
949 AMDGPU::V_SUB_U32_e64 : AMDGPU::V_SUB_I32_e32;
951 MachineSDNode *MachineSub
952 = CurDAG->getMachineNode(SubOp, DL, MVT::i32,
953 Zero, Addr.getOperand(1));
955 Base = SDValue(MachineSub, 0);
956 Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8);
957 Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8);
962 } else if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
963 unsigned DWordOffset0 = CAddr->getZExtValue() / 4;
964 unsigned DWordOffset1 = DWordOffset0 + 1;
965 assert(4 * DWordOffset0 == CAddr->getZExtValue());
967 if (isUInt<8>(DWordOffset0) && isUInt<8>(DWordOffset1)) {
968 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
969 MachineSDNode *MovZero
970 = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
972 Base = SDValue(MovZero, 0);
973 Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8);
974 Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8);
982 Offset0 = CurDAG->getTargetConstant(0, DL, MVT::i8);
983 Offset1 = CurDAG->getTargetConstant(1, DL, MVT::i8);
987 bool AMDGPUDAGToDAGISel::SelectMUBUF(SDValue Addr, SDValue &Ptr,
988 SDValue &VAddr, SDValue &SOffset,
989 SDValue &Offset, SDValue &Offen,
990 SDValue &Idxen, SDValue &Addr64,
991 SDValue &GLC, SDValue &SLC,
992 SDValue &TFE) const {
993 // Subtarget prefers to use flat instruction
994 if (Subtarget->useFlatForGlobal())
1000 GLC = CurDAG->getTargetConstant(0, DL, MVT::i1);
1002 SLC = CurDAG->getTargetConstant(0, DL, MVT::i1);
1003 TFE = CurDAG->getTargetConstant(0, DL, MVT::i1);
1005 Idxen = CurDAG->getTargetConstant(0, DL, MVT::i1);
1006 Offen = CurDAG->getTargetConstant(0, DL, MVT::i1);
1007 Addr64 = CurDAG->getTargetConstant(0, DL, MVT::i1);
1008 SOffset = CurDAG->getTargetConstant(0, DL, MVT::i32);
1010 ConstantSDNode *C1 = nullptr;
1012 if (CurDAG->isBaseWithConstantOffset(Addr)) {
1013 C1 = cast<ConstantSDNode>(Addr.getOperand(1));
1014 if (isUInt<32>(C1->getZExtValue()))
1015 N0 = Addr.getOperand(0);
1020 if (N0.getOpcode() == ISD::ADD) {
1021 // (add N2, N3) -> addr64, or
1022 // (add (add N2, N3), C1) -> addr64
1023 SDValue N2 = N0.getOperand(0);
1024 SDValue N3 = N0.getOperand(1);
1025 Addr64 = CurDAG->getTargetConstant(1, DL, MVT::i1);
1027 if (N2->isDivergent()) {
1028 if (N3->isDivergent()) {
1029 // Both N2 and N3 are divergent. Use N0 (the result of the add) as the
1030 // addr64, and construct the resource from a 0 address.
1031 Ptr = SDValue(buildSMovImm64(DL, 0, MVT::v2i32), 0);
1034 // N2 is divergent, N3 is not.
1039 // N2 is not divergent.
1043 Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1044 } else if (N0->isDivergent()) {
1045 // N0 is divergent. Use it as the addr64, and construct the resource from a
1047 Ptr = SDValue(buildSMovImm64(DL, 0, MVT::v2i32), 0);
1049 Addr64 = CurDAG->getTargetConstant(1, DL, MVT::i1);
1052 // (N0 + C1) -> offset
1053 VAddr = CurDAG->getTargetConstant(0, DL, MVT::i32);
1059 Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1063 if (SIInstrInfo::isLegalMUBUFImmOffset(C1->getZExtValue())) {
1064 // Legal offset for instruction.
1065 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16);
1069 // Illegal offset, store it in soffset.
1070 Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1072 SDValue(CurDAG->getMachineNode(
1073 AMDGPU::S_MOV_B32, DL, MVT::i32,
1074 CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i32)),
1079 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
1080 SDValue &VAddr, SDValue &SOffset,
1081 SDValue &Offset, SDValue &GLC,
1082 SDValue &SLC, SDValue &TFE) const {
1083 SDValue Ptr, Offen, Idxen, Addr64;
1085 // addr64 bit was removed for volcanic islands.
1086 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1089 if (!SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
1093 ConstantSDNode *C = cast<ConstantSDNode>(Addr64);
1094 if (C->getSExtValue()) {
1097 const SITargetLowering& Lowering =
1098 *static_cast<const SITargetLowering*>(getTargetLowering());
1100 SRsrc = SDValue(Lowering.wrapAddr64Rsrc(*CurDAG, DL, Ptr), 0);
1107 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
1108 SDValue &VAddr, SDValue &SOffset,
1110 SDValue &SLC) const {
1111 SLC = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i1);
1114 return SelectMUBUFAddr64(Addr, SRsrc, VAddr, SOffset, Offset, GLC, SLC, TFE);
1117 static bool isStackPtrRelative(const MachinePointerInfo &PtrInfo) {
1118 auto PSV = PtrInfo.V.dyn_cast<const PseudoSourceValue *>();
1119 return PSV && PSV->isStack();
1122 std::pair<SDValue, SDValue> AMDGPUDAGToDAGISel::foldFrameIndex(SDValue N) const {
1123 const MachineFunction &MF = CurDAG->getMachineFunction();
1124 const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1126 if (auto FI = dyn_cast<FrameIndexSDNode>(N)) {
1127 SDValue TFI = CurDAG->getTargetFrameIndex(FI->getIndex(),
1128 FI->getValueType(0));
1130 // If we can resolve this to a frame index access, this is relative to the
1131 // frame pointer SGPR.
1132 return std::make_pair(TFI, CurDAG->getRegister(Info->getFrameOffsetReg(),
1136 // If we don't know this private access is a local stack object, it needs to
1137 // be relative to the entry point's scratch wave offset register.
1138 return std::make_pair(N, CurDAG->getRegister(Info->getScratchWaveOffsetReg(),
1142 bool AMDGPUDAGToDAGISel::SelectMUBUFScratchOffen(SDNode *Parent,
1143 SDValue Addr, SDValue &Rsrc,
1144 SDValue &VAddr, SDValue &SOffset,
1145 SDValue &ImmOffset) const {
1148 MachineFunction &MF = CurDAG->getMachineFunction();
1149 const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1151 Rsrc = CurDAG->getRegister(Info->getScratchRSrcReg(), MVT::v4i32);
1153 if (ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
1154 unsigned Imm = CAddr->getZExtValue();
1156 SDValue HighBits = CurDAG->getTargetConstant(Imm & ~4095, DL, MVT::i32);
1157 MachineSDNode *MovHighBits = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
1158 DL, MVT::i32, HighBits);
1159 VAddr = SDValue(MovHighBits, 0);
1161 // In a call sequence, stores to the argument stack area are relative to the
1163 const MachinePointerInfo &PtrInfo = cast<MemSDNode>(Parent)->getPointerInfo();
1164 unsigned SOffsetReg = isStackPtrRelative(PtrInfo) ?
1165 Info->getStackPtrOffsetReg() : Info->getScratchWaveOffsetReg();
1167 SOffset = CurDAG->getRegister(SOffsetReg, MVT::i32);
1168 ImmOffset = CurDAG->getTargetConstant(Imm & 4095, DL, MVT::i16);
1172 if (CurDAG->isBaseWithConstantOffset(Addr)) {
1175 SDValue N0 = Addr.getOperand(0);
1176 SDValue N1 = Addr.getOperand(1);
1178 // Offsets in vaddr must be positive if range checking is enabled.
1180 // The total computation of vaddr + soffset + offset must not overflow. If
1181 // vaddr is negative, even if offset is 0 the sgpr offset add will end up
1184 // Prior to gfx9, MUBUF instructions with the vaddr offset enabled would
1185 // always perform a range check. If a negative vaddr base index was used,
1186 // this would fail the range check. The overall address computation would
1187 // compute a valid address, but this doesn't happen due to the range
1188 // check. For out-of-bounds MUBUF loads, a 0 is returned.
1190 // Therefore it should be safe to fold any VGPR offset on gfx9 into the
1191 // MUBUF vaddr, but not on older subtargets which can only do this if the
1192 // sign bit is known 0.
1193 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
1194 if (SIInstrInfo::isLegalMUBUFImmOffset(C1->getZExtValue()) &&
1195 (!Subtarget->privateMemoryResourceIsRangeChecked() ||
1196 CurDAG->SignBitIsZero(N0))) {
1197 std::tie(VAddr, SOffset) = foldFrameIndex(N0);
1198 ImmOffset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16);
1204 std::tie(VAddr, SOffset) = foldFrameIndex(Addr);
1205 ImmOffset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1209 bool AMDGPUDAGToDAGISel::SelectMUBUFScratchOffset(SDNode *Parent,
1213 SDValue &Offset) const {
1214 ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr);
1215 if (!CAddr || !SIInstrInfo::isLegalMUBUFImmOffset(CAddr->getZExtValue()))
1219 MachineFunction &MF = CurDAG->getMachineFunction();
1220 const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1222 SRsrc = CurDAG->getRegister(Info->getScratchRSrcReg(), MVT::v4i32);
1224 const MachinePointerInfo &PtrInfo = cast<MemSDNode>(Parent)->getPointerInfo();
1225 unsigned SOffsetReg = isStackPtrRelative(PtrInfo) ?
1226 Info->getStackPtrOffsetReg() : Info->getScratchWaveOffsetReg();
1228 // FIXME: Get from MachinePointerInfo? We should only be using the frame
1229 // offset if we know this is in a call sequence.
1230 SOffset = CurDAG->getRegister(SOffsetReg, MVT::i32);
1232 Offset = CurDAG->getTargetConstant(CAddr->getZExtValue(), DL, MVT::i16);
1236 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1237 SDValue &SOffset, SDValue &Offset,
1238 SDValue &GLC, SDValue &SLC,
1239 SDValue &TFE) const {
1240 SDValue Ptr, VAddr, Offen, Idxen, Addr64;
1241 const SIInstrInfo *TII =
1242 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
1244 if (!SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
1248 if (!cast<ConstantSDNode>(Offen)->getSExtValue() &&
1249 !cast<ConstantSDNode>(Idxen)->getSExtValue() &&
1250 !cast<ConstantSDNode>(Addr64)->getSExtValue()) {
1251 uint64_t Rsrc = TII->getDefaultRsrcDataFormat() |
1252 APInt::getAllOnesValue(32).getZExtValue(); // Size
1255 const SITargetLowering& Lowering =
1256 *static_cast<const SITargetLowering*>(getTargetLowering());
1258 SRsrc = SDValue(Lowering.buildRSRC(*CurDAG, DL, Ptr, 0, Rsrc), 0);
1264 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1265 SDValue &Soffset, SDValue &Offset
1267 SDValue GLC, SLC, TFE;
1269 return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE);
1271 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1272 SDValue &Soffset, SDValue &Offset,
1273 SDValue &SLC) const {
1276 return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE);
1279 template <bool IsSigned>
1280 bool AMDGPUDAGToDAGISel::SelectFlatOffset(SDValue Addr,
1283 SDValue &SLC) const {
1284 int64_t OffsetVal = 0;
1286 if (Subtarget->hasFlatInstOffsets() &&
1287 CurDAG->isBaseWithConstantOffset(Addr)) {
1288 SDValue N0 = Addr.getOperand(0);
1289 SDValue N1 = Addr.getOperand(1);
1290 int64_t COffsetVal = cast<ConstantSDNode>(N1)->getSExtValue();
1292 if ((IsSigned && isInt<13>(COffsetVal)) ||
1293 (!IsSigned && isUInt<12>(COffsetVal))) {
1295 OffsetVal = COffsetVal;
1300 Offset = CurDAG->getTargetConstant(OffsetVal, SDLoc(), MVT::i16);
1301 SLC = CurDAG->getTargetConstant(0, SDLoc(), MVT::i1);
1306 bool AMDGPUDAGToDAGISel::SelectFlatAtomic(SDValue Addr,
1309 SDValue &SLC) const {
1310 return SelectFlatOffset<false>(Addr, VAddr, Offset, SLC);
1313 bool AMDGPUDAGToDAGISel::SelectFlatAtomicSigned(SDValue Addr,
1316 SDValue &SLC) const {
1317 return SelectFlatOffset<true>(Addr, VAddr, Offset, SLC);
1320 bool AMDGPUDAGToDAGISel::SelectSMRDOffset(SDValue ByteOffsetNode,
1321 SDValue &Offset, bool &Imm) const {
1323 // FIXME: Handle non-constant offsets.
1324 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ByteOffsetNode);
1328 SDLoc SL(ByteOffsetNode);
1329 GCNSubtarget::Generation Gen = Subtarget->getGeneration();
1330 int64_t ByteOffset = C->getSExtValue();
1331 int64_t EncodedOffset = AMDGPU::getSMRDEncodedOffset(*Subtarget, ByteOffset);
1333 if (AMDGPU::isLegalSMRDImmOffset(*Subtarget, ByteOffset)) {
1334 Offset = CurDAG->getTargetConstant(EncodedOffset, SL, MVT::i32);
1339 if (!isUInt<32>(EncodedOffset) || !isUInt<32>(ByteOffset))
1342 if (Gen == AMDGPUSubtarget::SEA_ISLANDS && isUInt<32>(EncodedOffset)) {
1343 // 32-bit Immediates are supported on Sea Islands.
1344 Offset = CurDAG->getTargetConstant(EncodedOffset, SL, MVT::i32);
1346 SDValue C32Bit = CurDAG->getTargetConstant(ByteOffset, SL, MVT::i32);
1347 Offset = SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SL, MVT::i32,
1354 SDValue AMDGPUDAGToDAGISel::Expand32BitAddress(SDValue Addr) const {
1355 if (Addr.getValueType() != MVT::i32)
1358 // Zero-extend a 32-bit address.
1361 const MachineFunction &MF = CurDAG->getMachineFunction();
1362 const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1363 unsigned AddrHiVal = Info->get32BitAddressHighBits();
1364 SDValue AddrHi = CurDAG->getTargetConstant(AddrHiVal, SL, MVT::i32);
1366 const SDValue Ops[] = {
1367 CurDAG->getTargetConstant(AMDGPU::SReg_64_XEXECRegClassID, SL, MVT::i32),
1369 CurDAG->getTargetConstant(AMDGPU::sub0, SL, MVT::i32),
1370 SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SL, MVT::i32, AddrHi),
1372 CurDAG->getTargetConstant(AMDGPU::sub1, SL, MVT::i32),
1375 return SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, SL, MVT::i64,
1379 bool AMDGPUDAGToDAGISel::SelectSMRD(SDValue Addr, SDValue &SBase,
1380 SDValue &Offset, bool &Imm) const {
1383 // A 32-bit (address + offset) should not cause unsigned 32-bit integer
1384 // wraparound, because s_load instructions perform the addition in 64 bits.
1385 if ((Addr.getValueType() != MVT::i32 ||
1386 Addr->getFlags().hasNoUnsignedWrap()) &&
1387 CurDAG->isBaseWithConstantOffset(Addr)) {
1388 SDValue N0 = Addr.getOperand(0);
1389 SDValue N1 = Addr.getOperand(1);
1391 if (SelectSMRDOffset(N1, Offset, Imm)) {
1392 SBase = Expand32BitAddress(N0);
1396 SBase = Expand32BitAddress(Addr);
1397 Offset = CurDAG->getTargetConstant(0, SL, MVT::i32);
1402 bool AMDGPUDAGToDAGISel::SelectSMRDImm(SDValue Addr, SDValue &SBase,
1403 SDValue &Offset) const {
1405 return SelectSMRD(Addr, SBase, Offset, Imm) && Imm;
1408 bool AMDGPUDAGToDAGISel::SelectSMRDImm32(SDValue Addr, SDValue &SBase,
1409 SDValue &Offset) const {
1411 if (Subtarget->getGeneration() != AMDGPUSubtarget::SEA_ISLANDS)
1415 if (!SelectSMRD(Addr, SBase, Offset, Imm))
1418 return !Imm && isa<ConstantSDNode>(Offset);
1421 bool AMDGPUDAGToDAGISel::SelectSMRDSgpr(SDValue Addr, SDValue &SBase,
1422 SDValue &Offset) const {
1424 return SelectSMRD(Addr, SBase, Offset, Imm) && !Imm &&
1425 !isa<ConstantSDNode>(Offset);
1428 bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm(SDValue Addr,
1429 SDValue &Offset) const {
1431 return SelectSMRDOffset(Addr, Offset, Imm) && Imm;
1434 bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm32(SDValue Addr,
1435 SDValue &Offset) const {
1436 if (Subtarget->getGeneration() != AMDGPUSubtarget::SEA_ISLANDS)
1440 if (!SelectSMRDOffset(Addr, Offset, Imm))
1443 return !Imm && isa<ConstantSDNode>(Offset);
1446 bool AMDGPUDAGToDAGISel::SelectMOVRELOffset(SDValue Index,
1448 SDValue &Offset) const {
1451 if (CurDAG->isBaseWithConstantOffset(Index)) {
1452 SDValue N0 = Index.getOperand(0);
1453 SDValue N1 = Index.getOperand(1);
1454 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
1457 // Don't peel off the offset (c0) if doing so could possibly lead
1458 // the base (n0) to be negative.
1459 if (C1->getSExtValue() <= 0 || CurDAG->SignBitIsZero(N0)) {
1461 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i32);
1466 if (isa<ConstantSDNode>(Index))
1470 Offset = CurDAG->getTargetConstant(0, DL, MVT::i32);
1474 SDNode *AMDGPUDAGToDAGISel::getS_BFE(unsigned Opcode, const SDLoc &DL,
1475 SDValue Val, uint32_t Offset,
1477 // Transformation function, pack the offset and width of a BFE into
1478 // the format expected by the S_BFE_I32 / S_BFE_U32. In the second
1479 // source, bits [5:0] contain the offset and bits [22:16] the width.
1480 uint32_t PackedVal = Offset | (Width << 16);
1481 SDValue PackedConst = CurDAG->getTargetConstant(PackedVal, DL, MVT::i32);
1483 return CurDAG->getMachineNode(Opcode, DL, MVT::i32, Val, PackedConst);
1486 void AMDGPUDAGToDAGISel::SelectS_BFEFromShifts(SDNode *N) {
1487 // "(a << b) srl c)" ---> "BFE_U32 a, (c-b), (32-c)
1488 // "(a << b) sra c)" ---> "BFE_I32 a, (c-b), (32-c)
1489 // Predicate: 0 < b <= c < 32
1491 const SDValue &Shl = N->getOperand(0);
1492 ConstantSDNode *B = dyn_cast<ConstantSDNode>(Shl->getOperand(1));
1493 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1496 uint32_t BVal = B->getZExtValue();
1497 uint32_t CVal = C->getZExtValue();
1499 if (0 < BVal && BVal <= CVal && CVal < 32) {
1500 bool Signed = N->getOpcode() == ISD::SRA;
1501 unsigned Opcode = Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32;
1503 ReplaceNode(N, getS_BFE(Opcode, SDLoc(N), Shl.getOperand(0), CVal - BVal,
1511 void AMDGPUDAGToDAGISel::SelectS_BFE(SDNode *N) {
1512 switch (N->getOpcode()) {
1514 if (N->getOperand(0).getOpcode() == ISD::SRL) {
1515 // "(a srl b) & mask" ---> "BFE_U32 a, b, popcount(mask)"
1516 // Predicate: isMask(mask)
1517 const SDValue &Srl = N->getOperand(0);
1518 ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(Srl.getOperand(1));
1519 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(N->getOperand(1));
1521 if (Shift && Mask) {
1522 uint32_t ShiftVal = Shift->getZExtValue();
1523 uint32_t MaskVal = Mask->getZExtValue();
1525 if (isMask_32(MaskVal)) {
1526 uint32_t WidthVal = countPopulation(MaskVal);
1528 ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_U32, SDLoc(N),
1529 Srl.getOperand(0), ShiftVal, WidthVal));
1536 if (N->getOperand(0).getOpcode() == ISD::AND) {
1537 // "(a & mask) srl b)" ---> "BFE_U32 a, b, popcount(mask >> b)"
1538 // Predicate: isMask(mask >> b)
1539 const SDValue &And = N->getOperand(0);
1540 ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(N->getOperand(1));
1541 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(And->getOperand(1));
1543 if (Shift && Mask) {
1544 uint32_t ShiftVal = Shift->getZExtValue();
1545 uint32_t MaskVal = Mask->getZExtValue() >> ShiftVal;
1547 if (isMask_32(MaskVal)) {
1548 uint32_t WidthVal = countPopulation(MaskVal);
1550 ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_U32, SDLoc(N),
1551 And.getOperand(0), ShiftVal, WidthVal));
1555 } else if (N->getOperand(0).getOpcode() == ISD::SHL) {
1556 SelectS_BFEFromShifts(N);
1561 if (N->getOperand(0).getOpcode() == ISD::SHL) {
1562 SelectS_BFEFromShifts(N);
1567 case ISD::SIGN_EXTEND_INREG: {
1568 // sext_inreg (srl x, 16), i8 -> bfe_i32 x, 16, 8
1569 SDValue Src = N->getOperand(0);
1570 if (Src.getOpcode() != ISD::SRL)
1573 const ConstantSDNode *Amt = dyn_cast<ConstantSDNode>(Src.getOperand(1));
1577 unsigned Width = cast<VTSDNode>(N->getOperand(1))->getVT().getSizeInBits();
1578 ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_I32, SDLoc(N), Src.getOperand(0),
1579 Amt->getZExtValue(), Width));
1587 bool AMDGPUDAGToDAGISel::isCBranchSCC(const SDNode *N) const {
1588 assert(N->getOpcode() == ISD::BRCOND);
1589 if (!N->hasOneUse())
1592 SDValue Cond = N->getOperand(1);
1593 if (Cond.getOpcode() == ISD::CopyToReg)
1594 Cond = Cond.getOperand(2);
1596 if (Cond.getOpcode() != ISD::SETCC || !Cond.hasOneUse())
1599 MVT VT = Cond.getOperand(0).getSimpleValueType();
1603 if (VT == MVT::i64) {
1604 auto ST = static_cast<const GCNSubtarget *>(Subtarget);
1606 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
1607 return (CC == ISD::SETEQ || CC == ISD::SETNE) && ST->hasScalarCompareEq64();
1613 void AMDGPUDAGToDAGISel::SelectBRCOND(SDNode *N) {
1614 SDValue Cond = N->getOperand(1);
1616 if (Cond.isUndef()) {
1617 CurDAG->SelectNodeTo(N, AMDGPU::SI_BR_UNDEF, MVT::Other,
1618 N->getOperand(2), N->getOperand(0));
1622 bool UseSCCBr = isCBranchSCC(N) && isUniformBr(N);
1623 unsigned BrOp = UseSCCBr ? AMDGPU::S_CBRANCH_SCC1 : AMDGPU::S_CBRANCH_VCCNZ;
1624 unsigned CondReg = UseSCCBr ? AMDGPU::SCC : AMDGPU::VCC;
1628 // This is the case that we are selecting to S_CBRANCH_VCCNZ. We have not
1629 // analyzed what generates the vcc value, so we do not know whether vcc
1630 // bits for disabled lanes are 0. Thus we need to mask out bits for
1633 // For the case that we select S_CBRANCH_SCC1 and it gets
1634 // changed to S_CBRANCH_VCCNZ in SIFixSGPRCopies, SIFixSGPRCopies calls
1635 // SIInstrInfo::moveToVALU which inserts the S_AND).
1637 // We could add an analysis of what generates the vcc value here and omit
1638 // the S_AND when is unnecessary. But it would be better to add a separate
1639 // pass after SIFixSGPRCopies to do the unnecessary S_AND removal, so it
1640 // catches both cases.
1641 Cond = SDValue(CurDAG->getMachineNode(AMDGPU::S_AND_B64, SL, MVT::i1,
1642 CurDAG->getRegister(AMDGPU::EXEC, MVT::i1),
1647 SDValue VCC = CurDAG->getCopyToReg(N->getOperand(0), SL, CondReg, Cond);
1648 CurDAG->SelectNodeTo(N, BrOp, MVT::Other,
1649 N->getOperand(2), // Basic Block
1653 void AMDGPUDAGToDAGISel::SelectFMAD_FMA(SDNode *N) {
1654 MVT VT = N->getSimpleValueType(0);
1655 bool IsFMA = N->getOpcode() == ISD::FMA;
1656 if (VT != MVT::f32 || (!Subtarget->hasMadMixInsts() &&
1657 !Subtarget->hasFmaMixInsts()) ||
1658 ((IsFMA && Subtarget->hasMadMixInsts()) ||
1659 (!IsFMA && Subtarget->hasFmaMixInsts()))) {
1664 SDValue Src0 = N->getOperand(0);
1665 SDValue Src1 = N->getOperand(1);
1666 SDValue Src2 = N->getOperand(2);
1667 unsigned Src0Mods, Src1Mods, Src2Mods;
1669 // Avoid using v_mad_mix_f32/v_fma_mix_f32 unless there is actually an operand
1670 // using the conversion from f16.
1671 bool Sel0 = SelectVOP3PMadMixModsImpl(Src0, Src0, Src0Mods);
1672 bool Sel1 = SelectVOP3PMadMixModsImpl(Src1, Src1, Src1Mods);
1673 bool Sel2 = SelectVOP3PMadMixModsImpl(Src2, Src2, Src2Mods);
1675 assert((IsFMA || !Subtarget->hasFP32Denormals()) &&
1676 "fmad selected with denormals enabled");
1677 // TODO: We can select this with f32 denormals enabled if all the sources are
1678 // converted from f16 (in which case fmad isn't legal).
1680 if (Sel0 || Sel1 || Sel2) {
1681 // For dummy operands.
1682 SDValue Zero = CurDAG->getTargetConstant(0, SDLoc(), MVT::i32);
1684 CurDAG->getTargetConstant(Src0Mods, SDLoc(), MVT::i32), Src0,
1685 CurDAG->getTargetConstant(Src1Mods, SDLoc(), MVT::i32), Src1,
1686 CurDAG->getTargetConstant(Src2Mods, SDLoc(), MVT::i32), Src2,
1687 CurDAG->getTargetConstant(0, SDLoc(), MVT::i1),
1691 CurDAG->SelectNodeTo(N,
1692 IsFMA ? AMDGPU::V_FMA_MIX_F32 : AMDGPU::V_MAD_MIX_F32,
1699 // This is here because there isn't a way to use the generated sub0_sub1 as the
1700 // subreg index to EXTRACT_SUBREG in tablegen.
1701 void AMDGPUDAGToDAGISel::SelectATOMIC_CMP_SWAP(SDNode *N) {
1702 MemSDNode *Mem = cast<MemSDNode>(N);
1703 unsigned AS = Mem->getAddressSpace();
1704 if (AS == AMDGPUAS::FLAT_ADDRESS) {
1709 MVT VT = N->getSimpleValueType(0);
1710 bool Is32 = (VT == MVT::i32);
1713 MachineSDNode *CmpSwap = nullptr;
1714 if (Subtarget->hasAddr64()) {
1715 SDValue SRsrc, VAddr, SOffset, Offset, SLC;
1717 if (SelectMUBUFAddr64(Mem->getBasePtr(), SRsrc, VAddr, SOffset, Offset, SLC)) {
1718 unsigned Opcode = Is32 ? AMDGPU::BUFFER_ATOMIC_CMPSWAP_ADDR64_RTN :
1719 AMDGPU::BUFFER_ATOMIC_CMPSWAP_X2_ADDR64_RTN;
1720 SDValue CmpVal = Mem->getOperand(2);
1722 // XXX - Do we care about glue operands?
1725 CmpVal, VAddr, SRsrc, SOffset, Offset, SLC, Mem->getChain()
1728 CmpSwap = CurDAG->getMachineNode(Opcode, SL, Mem->getVTList(), Ops);
1733 SDValue SRsrc, SOffset, Offset, SLC;
1734 if (SelectMUBUFOffset(Mem->getBasePtr(), SRsrc, SOffset, Offset, SLC)) {
1735 unsigned Opcode = Is32 ? AMDGPU::BUFFER_ATOMIC_CMPSWAP_OFFSET_RTN :
1736 AMDGPU::BUFFER_ATOMIC_CMPSWAP_X2_OFFSET_RTN;
1738 SDValue CmpVal = Mem->getOperand(2);
1740 CmpVal, SRsrc, SOffset, Offset, SLC, Mem->getChain()
1743 CmpSwap = CurDAG->getMachineNode(Opcode, SL, Mem->getVTList(), Ops);
1752 MachineMemOperand *MMO = Mem->getMemOperand();
1753 CurDAG->setNodeMemRefs(CmpSwap, {MMO});
1755 unsigned SubReg = Is32 ? AMDGPU::sub0 : AMDGPU::sub0_sub1;
1757 = CurDAG->getTargetExtractSubreg(SubReg, SL, VT, SDValue(CmpSwap, 0));
1759 ReplaceUses(SDValue(N, 0), Extract);
1760 ReplaceUses(SDValue(N, 1), SDValue(CmpSwap, 1));
1761 CurDAG->RemoveDeadNode(N);
1764 bool AMDGPUDAGToDAGISel::SelectVOP3ModsImpl(SDValue In, SDValue &Src,
1765 unsigned &Mods) const {
1769 if (Src.getOpcode() == ISD::FNEG) {
1770 Mods |= SISrcMods::NEG;
1771 Src = Src.getOperand(0);
1774 if (Src.getOpcode() == ISD::FABS) {
1775 Mods |= SISrcMods::ABS;
1776 Src = Src.getOperand(0);
1782 bool AMDGPUDAGToDAGISel::SelectVOP3Mods(SDValue In, SDValue &Src,
1783 SDValue &SrcMods) const {
1785 if (SelectVOP3ModsImpl(In, Src, Mods)) {
1786 SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
1793 bool AMDGPUDAGToDAGISel::SelectVOP3Mods_NNaN(SDValue In, SDValue &Src,
1794 SDValue &SrcMods) const {
1795 SelectVOP3Mods(In, Src, SrcMods);
1796 return isNoNanSrc(Src);
1799 bool AMDGPUDAGToDAGISel::SelectVOP3NoMods(SDValue In, SDValue &Src) const {
1800 if (In.getOpcode() == ISD::FABS || In.getOpcode() == ISD::FNEG)
1807 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0(SDValue In, SDValue &Src,
1808 SDValue &SrcMods, SDValue &Clamp,
1809 SDValue &Omod) const {
1811 Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1);
1812 Omod = CurDAG->getTargetConstant(0, DL, MVT::i1);
1814 return SelectVOP3Mods(In, Src, SrcMods);
1817 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src,
1820 SDValue &Omod) const {
1821 Clamp = Omod = CurDAG->getTargetConstant(0, SDLoc(In), MVT::i32);
1822 return SelectVOP3Mods(In, Src, SrcMods);
1825 bool AMDGPUDAGToDAGISel::SelectVOP3OMods(SDValue In, SDValue &Src,
1826 SDValue &Clamp, SDValue &Omod) const {
1830 Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1);
1831 Omod = CurDAG->getTargetConstant(0, DL, MVT::i1);
1836 static SDValue stripBitcast(SDValue Val) {
1837 return Val.getOpcode() == ISD::BITCAST ? Val.getOperand(0) : Val;
1840 // Figure out if this is really an extract of the high 16-bits of a dword.
1841 static bool isExtractHiElt(SDValue In, SDValue &Out) {
1842 In = stripBitcast(In);
1843 if (In.getOpcode() != ISD::TRUNCATE)
1846 SDValue Srl = In.getOperand(0);
1847 if (Srl.getOpcode() == ISD::SRL) {
1848 if (ConstantSDNode *ShiftAmt = dyn_cast<ConstantSDNode>(Srl.getOperand(1))) {
1849 if (ShiftAmt->getZExtValue() == 16) {
1850 Out = stripBitcast(Srl.getOperand(0));
1859 // Look through operations that obscure just looking at the low 16-bits of the
1861 static SDValue stripExtractLoElt(SDValue In) {
1862 if (In.getOpcode() == ISD::TRUNCATE) {
1863 SDValue Src = In.getOperand(0);
1864 if (Src.getValueType().getSizeInBits() == 32)
1865 return stripBitcast(Src);
1871 bool AMDGPUDAGToDAGISel::SelectVOP3PMods(SDValue In, SDValue &Src,
1872 SDValue &SrcMods) const {
1876 if (Src.getOpcode() == ISD::FNEG) {
1877 Mods ^= (SISrcMods::NEG | SISrcMods::NEG_HI);
1878 Src = Src.getOperand(0);
1881 if (Src.getOpcode() == ISD::BUILD_VECTOR) {
1882 unsigned VecMods = Mods;
1884 SDValue Lo = stripBitcast(Src.getOperand(0));
1885 SDValue Hi = stripBitcast(Src.getOperand(1));
1887 if (Lo.getOpcode() == ISD::FNEG) {
1888 Lo = stripBitcast(Lo.getOperand(0));
1889 Mods ^= SISrcMods::NEG;
1892 if (Hi.getOpcode() == ISD::FNEG) {
1893 Hi = stripBitcast(Hi.getOperand(0));
1894 Mods ^= SISrcMods::NEG_HI;
1897 if (isExtractHiElt(Lo, Lo))
1898 Mods |= SISrcMods::OP_SEL_0;
1900 if (isExtractHiElt(Hi, Hi))
1901 Mods |= SISrcMods::OP_SEL_1;
1903 Lo = stripExtractLoElt(Lo);
1904 Hi = stripExtractLoElt(Hi);
1906 if (Lo == Hi && !isInlineImmediate(Lo.getNode())) {
1907 // Really a scalar input. Just select from the low half of the register to
1911 SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
1918 // Packed instructions do not have abs modifiers.
1919 Mods |= SISrcMods::OP_SEL_1;
1921 SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
1925 bool AMDGPUDAGToDAGISel::SelectVOP3PMods0(SDValue In, SDValue &Src,
1927 SDValue &Clamp) const {
1930 // FIXME: Handle clamp and op_sel
1931 Clamp = CurDAG->getTargetConstant(0, SL, MVT::i32);
1933 return SelectVOP3PMods(In, Src, SrcMods);
1936 bool AMDGPUDAGToDAGISel::SelectVOP3OpSel(SDValue In, SDValue &Src,
1937 SDValue &SrcMods) const {
1939 // FIXME: Handle op_sel
1940 SrcMods = CurDAG->getTargetConstant(0, SDLoc(In), MVT::i32);
1944 bool AMDGPUDAGToDAGISel::SelectVOP3OpSel0(SDValue In, SDValue &Src,
1946 SDValue &Clamp) const {
1949 // FIXME: Handle clamp
1950 Clamp = CurDAG->getTargetConstant(0, SL, MVT::i32);
1952 return SelectVOP3OpSel(In, Src, SrcMods);
1955 bool AMDGPUDAGToDAGISel::SelectVOP3OpSelMods(SDValue In, SDValue &Src,
1956 SDValue &SrcMods) const {
1957 // FIXME: Handle op_sel
1958 return SelectVOP3Mods(In, Src, SrcMods);
1961 bool AMDGPUDAGToDAGISel::SelectVOP3OpSelMods0(SDValue In, SDValue &Src,
1963 SDValue &Clamp) const {
1966 // FIXME: Handle clamp
1967 Clamp = CurDAG->getTargetConstant(0, SL, MVT::i32);
1969 return SelectVOP3OpSelMods(In, Src, SrcMods);
1972 // The return value is not whether the match is possible (which it always is),
1973 // but whether or not it a conversion is really used.
1974 bool AMDGPUDAGToDAGISel::SelectVOP3PMadMixModsImpl(SDValue In, SDValue &Src,
1975 unsigned &Mods) const {
1977 SelectVOP3ModsImpl(In, Src, Mods);
1979 if (Src.getOpcode() == ISD::FP_EXTEND) {
1980 Src = Src.getOperand(0);
1981 assert(Src.getValueType() == MVT::f16);
1982 Src = stripBitcast(Src);
1984 // Be careful about folding modifiers if we already have an abs. fneg is
1985 // applied last, so we don't want to apply an earlier fneg.
1986 if ((Mods & SISrcMods::ABS) == 0) {
1988 SelectVOP3ModsImpl(Src, Src, ModsTmp);
1990 if ((ModsTmp & SISrcMods::NEG) != 0)
1991 Mods ^= SISrcMods::NEG;
1993 if ((ModsTmp & SISrcMods::ABS) != 0)
1994 Mods |= SISrcMods::ABS;
1997 // op_sel/op_sel_hi decide the source type and source.
1998 // If the source's op_sel_hi is set, it indicates to do a conversion from fp16.
1999 // If the sources's op_sel is set, it picks the high half of the source
2002 Mods |= SISrcMods::OP_SEL_1;
2003 if (isExtractHiElt(Src, Src)) {
2004 Mods |= SISrcMods::OP_SEL_0;
2006 // TODO: Should we try to look for neg/abs here?
2015 bool AMDGPUDAGToDAGISel::SelectVOP3PMadMixMods(SDValue In, SDValue &Src,
2016 SDValue &SrcMods) const {
2018 SelectVOP3PMadMixModsImpl(In, Src, Mods);
2019 SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
2023 // TODO: Can we identify things like v_mad_mixhi_f16?
2024 bool AMDGPUDAGToDAGISel::SelectHi16Elt(SDValue In, SDValue &Src) const {
2030 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(In)) {
2032 SDValue K = CurDAG->getTargetConstant(C->getZExtValue() << 16, SL, MVT::i32);
2033 MachineSDNode *MovK = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
2035 Src = SDValue(MovK, 0);
2039 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(In)) {
2041 SDValue K = CurDAG->getTargetConstant(
2042 C->getValueAPF().bitcastToAPInt().getZExtValue() << 16, SL, MVT::i32);
2043 MachineSDNode *MovK = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
2045 Src = SDValue(MovK, 0);
2049 return isExtractHiElt(In, Src);
2052 bool AMDGPUDAGToDAGISel::isVGPRImm(const SDNode * N) const {
2053 if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS) {
2056 const SIRegisterInfo *SIRI =
2057 static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());
2058 const SIInstrInfo * SII =
2059 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
2062 bool AllUsesAcceptSReg = true;
2063 for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
2064 Limit < 10 && U != E; ++U, ++Limit) {
2065 const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo());
2067 // If the register class is unknown, it could be an unknown
2068 // register class that needs to be an SGPR, e.g. an inline asm
2070 if (!RC || SIRI->isSGPRClass(RC))
2073 if (RC != &AMDGPU::VS_32RegClass) {
2074 AllUsesAcceptSReg = false;
2076 if (User->isMachineOpcode()) {
2077 unsigned Opc = User->getMachineOpcode();
2078 MCInstrDesc Desc = SII->get(Opc);
2079 if (Desc.isCommutable()) {
2080 unsigned OpIdx = Desc.getNumDefs() + U.getOperandNo();
2081 unsigned CommuteIdx1 = TargetInstrInfo::CommuteAnyOperandIndex;
2082 if (SII->findCommutedOpIndices(Desc, OpIdx, CommuteIdx1)) {
2083 unsigned CommutedOpNo = CommuteIdx1 - Desc.getNumDefs();
2084 const TargetRegisterClass *CommutedRC = getOperandRegClass(*U, CommutedOpNo);
2085 if (CommutedRC == &AMDGPU::VS_32RegClass)
2086 AllUsesAcceptSReg = true;
2090 // If "AllUsesAcceptSReg == false" so far we haven't suceeded
2091 // commuting current user. This means have at least one use
2092 // that strictly require VGPR. Thus, we will not attempt to commute
2093 // other user instructions.
2094 if (!AllUsesAcceptSReg)
2098 return !AllUsesAcceptSReg && (Limit < 10);
2101 bool AMDGPUDAGToDAGISel::isUniformLoad(const SDNode * N) const {
2102 auto Ld = cast<LoadSDNode>(N);
2104 return Ld->getAlignment() >= 4 &&
2108 Ld->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS ||
2109 Ld->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT
2116 Subtarget->getScalarizeGlobalBehavior() &&
2117 Ld->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
2118 !Ld->isVolatile() &&
2119 !N->isDivergent() &&
2120 static_cast<const SITargetLowering *>(
2121 getTargetLowering())->isMemOpHasNoClobberedMemOperand(N)
2126 void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
2127 const AMDGPUTargetLowering& Lowering =
2128 *static_cast<const AMDGPUTargetLowering*>(getTargetLowering());
2129 bool IsModified = false;
2133 // Go over all selected nodes and try to fold them a bit more
2134 SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_begin();
2135 while (Position != CurDAG->allnodes_end()) {
2136 SDNode *Node = &*Position++;
2137 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(Node);
2141 SDNode *ResNode = Lowering.PostISelFolding(MachineNode, *CurDAG);
2142 if (ResNode != Node) {
2144 ReplaceUses(Node, ResNode);
2148 CurDAG->RemoveDeadNodes();
2149 } while (IsModified);
2152 bool R600DAGToDAGISel::runOnMachineFunction(MachineFunction &MF) {
2153 Subtarget = &MF.getSubtarget<R600Subtarget>();
2154 return SelectionDAGISel::runOnMachineFunction(MF);
2157 bool R600DAGToDAGISel::isConstantLoad(const MemSDNode *N, int CbId) const {
2161 return N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS ||
2162 N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT;
2164 return N->getAddressSpace() == AMDGPUAS::CONSTANT_BUFFER_0 + CbId;
2167 bool R600DAGToDAGISel::SelectGlobalValueConstantOffset(SDValue Addr,
2169 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Addr)) {
2170 IntPtr = CurDAG->getIntPtrConstant(Cst->getZExtValue() / 4, SDLoc(Addr),
2177 bool R600DAGToDAGISel::SelectGlobalValueVariableOffset(SDValue Addr,
2178 SDValue& BaseReg, SDValue &Offset) {
2179 if (!isa<ConstantSDNode>(Addr)) {
2181 Offset = CurDAG->getIntPtrConstant(0, SDLoc(Addr), true);
2187 void R600DAGToDAGISel::Select(SDNode *N) {
2188 unsigned int Opc = N->getOpcode();
2189 if (N->isMachineOpcode()) {
2191 return; // Already selected.
2196 case AMDGPUISD::BUILD_VERTICAL_VECTOR:
2197 case ISD::SCALAR_TO_VECTOR:
2198 case ISD::BUILD_VECTOR: {
2199 EVT VT = N->getValueType(0);
2200 unsigned NumVectorElts = VT.getVectorNumElements();
2201 unsigned RegClassID;
2202 // BUILD_VECTOR was lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
2203 // that adds a 128 bits reg copy when going through TwoAddressInstructions
2204 // pass. We want to avoid 128 bits copies as much as possible because they
2205 // can't be bundled by our scheduler.
2206 switch(NumVectorElts) {
2207 case 2: RegClassID = R600::R600_Reg64RegClassID; break;
2209 if (Opc == AMDGPUISD::BUILD_VERTICAL_VECTOR)
2210 RegClassID = R600::R600_Reg128VerticalRegClassID;
2212 RegClassID = R600::R600_Reg128RegClassID;
2214 default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
2216 SelectBuildVector(N, RegClassID);
2224 bool R600DAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base,
2229 if ((C = dyn_cast<ConstantSDNode>(Addr))) {
2230 Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32);
2231 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
2232 } else if ((Addr.getOpcode() == AMDGPUISD::DWORDADDR) &&
2233 (C = dyn_cast<ConstantSDNode>(Addr.getOperand(0)))) {
2234 Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32);
2235 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
2236 } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
2237 (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) {
2238 Base = Addr.getOperand(0);
2239 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
2242 Offset = CurDAG->getTargetConstant(0, DL, MVT::i32);
2248 bool R600DAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
2250 ConstantSDNode *IMMOffset;
2252 if (Addr.getOpcode() == ISD::ADD
2253 && (IMMOffset = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))
2254 && isInt<16>(IMMOffset->getZExtValue())) {
2256 Base = Addr.getOperand(0);
2257 Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), SDLoc(Addr),
2260 // If the pointer address is constant, we can move it to the offset field.
2261 } else if ((IMMOffset = dyn_cast<ConstantSDNode>(Addr))
2262 && isInt<16>(IMMOffset->getZExtValue())) {
2263 Base = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
2264 SDLoc(CurDAG->getEntryNode()),
2265 R600::ZERO, MVT::i32);
2266 Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), SDLoc(Addr),
2271 // Default case, no offset
2273 Offset = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i32);