1 //===-- AMDGPUISelDAGToDAG.cpp - A dag to dag inst selector for AMDGPU ----===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //==-----------------------------------------------------------------------===//
8 //
9 /// \file
10 /// Defines an instruction selector for the AMDGPU target.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "AMDGPU.h"
15 #include "AMDGPUArgumentUsageInfo.h"
16 #include "AMDGPUISelLowering.h" // For AMDGPUISD
17 #include "AMDGPUInstrInfo.h"
18 #include "AMDGPUPerfHintAnalysis.h"
19 #include "AMDGPUSubtarget.h"
20 #include "AMDGPUTargetMachine.h"
21 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
22 #include "SIDefines.h"
23 #include "SIISelLowering.h"
24 #include "SIInstrInfo.h"
25 #include "SIMachineFunctionInfo.h"
26 #include "SIRegisterInfo.h"
27 #include "llvm/ADT/APInt.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/StringRef.h"
30 #include "llvm/Analysis/LegacyDivergenceAnalysis.h"
31 #include "llvm/Analysis/ValueTracking.h"
32 #include "llvm/CodeGen/FunctionLoweringInfo.h"
33 #include "llvm/CodeGen/ISDOpcodes.h"
34 #include "llvm/CodeGen/MachineFunction.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/CodeGen/SelectionDAGISel.h"
38 #include "llvm/CodeGen/SelectionDAGNodes.h"
39 #include "llvm/CodeGen/ValueTypes.h"
40 #include "llvm/IR/BasicBlock.h"
41 #include "llvm/InitializePasses.h"
42 #ifdef EXPENSIVE_CHECKS
43 #include "llvm/IR/Dominators.h"
44 #endif
45 #include "llvm/IR/Instruction.h"
46 #include "llvm/MC/MCInstrDesc.h"
47 #include "llvm/Support/Casting.h"
48 #include "llvm/Support/CodeGen.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MachineValueType.h"
51 #include "llvm/Support/MathExtras.h"
52 #include <cassert>
53 #include <cstdint>
54 #include <new>
55 #include <vector>
56 
57 #define DEBUG_TYPE "isel"
58 
59 using namespace llvm;
60 
61 namespace llvm {
62 
63 class R600InstrInfo;
64 
65 } // end namespace llvm
66 
67 //===----------------------------------------------------------------------===//
68 // Instruction Selector Implementation
69 //===----------------------------------------------------------------------===//
70 
71 namespace {
72 
73 static bool isNullConstantOrUndef(SDValue V) {
74   if (V.isUndef())
75     return true;
76 
77   ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
78   return Const != nullptr && Const->isNullValue();
79 }
80 
81 static bool getConstantValue(SDValue N, uint32_t &Out) {
82   // This is only used for packed vectors, where ussing 0 for undef should
83   // always be good.
84   if (N.isUndef()) {
85     Out = 0;
86     return true;
87   }
88 
89   if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
90     Out = C->getAPIntValue().getSExtValue();
91     return true;
92   }
93 
94   if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N)) {
95     Out = C->getValueAPF().bitcastToAPInt().getSExtValue();
96     return true;
97   }
98 
99   return false;
100 }
101 
102 // TODO: Handle undef as zero
103 static SDNode *packConstantV2I16(const SDNode *N, SelectionDAG &DAG,
104                                  bool Negate = false) {
105   assert(N->getOpcode() == ISD::BUILD_VECTOR && N->getNumOperands() == 2);
106   uint32_t LHSVal, RHSVal;
107   if (getConstantValue(N->getOperand(0), LHSVal) &&
108       getConstantValue(N->getOperand(1), RHSVal)) {
109     SDLoc SL(N);
110     uint32_t K = Negate ?
111       (-LHSVal & 0xffff) | (-RHSVal << 16) :
112       (LHSVal & 0xffff) | (RHSVal << 16);
113     return DAG.getMachineNode(AMDGPU::S_MOV_B32, SL, N->getValueType(0),
114                               DAG.getTargetConstant(K, SL, MVT::i32));
115   }
116 
117   return nullptr;
118 }
119 
120 static SDNode *packNegConstantV2I16(const SDNode *N, SelectionDAG &DAG) {
121   return packConstantV2I16(N, DAG, true);
122 }
123 
124 /// AMDGPU specific code to select AMDGPU machine instructions for
125 /// SelectionDAG operations.
126 class AMDGPUDAGToDAGISel : public SelectionDAGISel {
127   // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can
128   // make the right decision when generating code for different targets.
129   const GCNSubtarget *Subtarget;
130 
131   // Default FP mode for the current function.
132   AMDGPU::SIModeRegisterDefaults Mode;
133 
134   bool EnableLateStructurizeCFG;
135 
136 public:
137   explicit AMDGPUDAGToDAGISel(TargetMachine *TM = nullptr,
138                               CodeGenOpt::Level OptLevel = CodeGenOpt::Default)
139     : SelectionDAGISel(*TM, OptLevel) {
140     EnableLateStructurizeCFG = AMDGPUTargetMachine::EnableLateStructurizeCFG;
141   }
142   ~AMDGPUDAGToDAGISel() override = default;
143 
144   void getAnalysisUsage(AnalysisUsage &AU) const override {
145     AU.addRequired<AMDGPUArgumentUsageInfo>();
146     AU.addRequired<LegacyDivergenceAnalysis>();
147 #ifdef EXPENSIVE_CHECKS
148     AU.addRequired<DominatorTreeWrapperPass>();
149     AU.addRequired<LoopInfoWrapperPass>();
150 #endif
151     SelectionDAGISel::getAnalysisUsage(AU);
152   }
153 
154   bool matchLoadD16FromBuildVector(SDNode *N) const;
155 
156   bool runOnMachineFunction(MachineFunction &MF) override;
157   void PreprocessISelDAG() override;
158   void Select(SDNode *N) override;
159   StringRef getPassName() const override;
160   void PostprocessISelDAG() override;
161 
162 protected:
163   void SelectBuildVector(SDNode *N, unsigned RegClassID);
164 
165 private:
166   std::pair<SDValue, SDValue> foldFrameIndex(SDValue N) const;
167   bool isNoNanSrc(SDValue N) const;
168   bool isInlineImmediate(const SDNode *N, bool Negated = false) const;
169   bool isNegInlineImmediate(const SDNode *N) const {
170     return isInlineImmediate(N, true);
171   }
172 
173   bool isInlineImmediate16(int64_t Imm) const {
174     return AMDGPU::isInlinableLiteral16(Imm, Subtarget->hasInv2PiInlineImm());
175   }
176 
177   bool isInlineImmediate32(int64_t Imm) const {
178     return AMDGPU::isInlinableLiteral32(Imm, Subtarget->hasInv2PiInlineImm());
179   }
180 
181   bool isInlineImmediate64(int64_t Imm) const {
182     return AMDGPU::isInlinableLiteral64(Imm, Subtarget->hasInv2PiInlineImm());
183   }
184 
185   bool isInlineImmediate(const APFloat &Imm) const {
186     return Subtarget->getInstrInfo()->isInlineConstant(Imm);
187   }
188 
189   bool isVGPRImm(const SDNode *N) const;
190   bool isUniformLoad(const SDNode *N) const;
191   bool isUniformBr(const SDNode *N) const;
192 
193   MachineSDNode *buildSMovImm64(SDLoc &DL, uint64_t Val, EVT VT) const;
194 
195   SDNode *glueCopyToOp(SDNode *N, SDValue NewChain, SDValue Glue) const;
196   SDNode *glueCopyToM0(SDNode *N, SDValue Val) const;
197   SDNode *glueCopyToM0LDSInit(SDNode *N) const;
198 
199   const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const;
200   virtual bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset);
201   virtual bool SelectADDRIndirect(SDValue Addr, SDValue &Base, SDValue &Offset);
202   bool isDSOffsetLegal(SDValue Base, unsigned Offset,
203                        unsigned OffsetBits) const;
204   bool SelectDS1Addr1Offset(SDValue Ptr, SDValue &Base, SDValue &Offset) const;
205   bool SelectDS64Bit4ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0,
206                                  SDValue &Offset1) const;
207   bool SelectMUBUF(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
208                    SDValue &SOffset, SDValue &Offset, SDValue &Offen,
209                    SDValue &Idxen, SDValue &Addr64, SDValue &GLC, SDValue &SLC,
210                    SDValue &TFE, SDValue &DLC, SDValue &SWZ) const;
211   bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
212                          SDValue &SOffset, SDValue &Offset, SDValue &GLC,
213                          SDValue &SLC, SDValue &TFE, SDValue &DLC,
214                          SDValue &SWZ) const;
215   bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
216                          SDValue &VAddr, SDValue &SOffset, SDValue &Offset,
217                          SDValue &SLC) const;
218   bool SelectMUBUFScratchOffen(SDNode *Parent,
219                                SDValue Addr, SDValue &RSrc, SDValue &VAddr,
220                                SDValue &SOffset, SDValue &ImmOffset) const;
221   bool SelectMUBUFScratchOffset(SDNode *Parent,
222                                 SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
223                                 SDValue &Offset) const;
224 
225   bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset,
226                          SDValue &Offset, SDValue &GLC, SDValue &SLC,
227                          SDValue &TFE, SDValue &DLC, SDValue &SWZ) const;
228   bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
229                          SDValue &Offset, SDValue &SLC) const;
230   bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
231                          SDValue &Offset) const;
232 
233   template <bool IsSigned>
234   bool SelectFlatOffset(SDNode *N, SDValue Addr, SDValue &VAddr,
235                         SDValue &Offset, SDValue &SLC) const;
236   bool SelectFlatAtomic(SDNode *N, SDValue Addr, SDValue &VAddr,
237                         SDValue &Offset, SDValue &SLC) const;
238   bool SelectFlatAtomicSigned(SDNode *N, SDValue Addr, SDValue &VAddr,
239                               SDValue &Offset, SDValue &SLC) const;
240 
241   bool SelectSMRDOffset(SDValue ByteOffsetNode, SDValue &Offset,
242                         bool &Imm) const;
243   SDValue Expand32BitAddress(SDValue Addr) const;
244   bool SelectSMRD(SDValue Addr, SDValue &SBase, SDValue &Offset,
245                   bool &Imm) const;
246   bool SelectSMRDImm(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
247   bool SelectSMRDImm32(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
248   bool SelectSMRDSgpr(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
249   bool SelectSMRDBufferImm(SDValue Addr, SDValue &Offset) const;
250   bool SelectSMRDBufferImm32(SDValue Addr, SDValue &Offset) const;
251   bool SelectMOVRELOffset(SDValue Index, SDValue &Base, SDValue &Offset) const;
252 
253   bool SelectVOP3Mods_NNaN(SDValue In, SDValue &Src, SDValue &SrcMods) const;
254   bool SelectVOP3ModsImpl(SDValue In, SDValue &Src, unsigned &SrcMods) const;
255   bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
256   bool SelectVOP3NoMods(SDValue In, SDValue &Src) const;
257   bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
258                        SDValue &Clamp, SDValue &Omod) const;
259   bool SelectVOP3NoMods0(SDValue In, SDValue &Src, SDValue &SrcMods,
260                          SDValue &Clamp, SDValue &Omod) const;
261 
262   bool SelectVOP3OMods(SDValue In, SDValue &Src,
263                        SDValue &Clamp, SDValue &Omod) const;
264 
265   bool SelectVOP3PMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
266 
267   bool SelectVOP3OpSel(SDValue In, SDValue &Src, SDValue &SrcMods) const;
268 
269   bool SelectVOP3OpSelMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
270   bool SelectVOP3PMadMixModsImpl(SDValue In, SDValue &Src, unsigned &Mods) const;
271   bool SelectVOP3PMadMixMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
272 
273   SDValue getHi16Elt(SDValue In) const;
274 
275   SDValue getMaterializedScalarImm32(int64_t Val, const SDLoc &DL) const;
276 
277   void SelectADD_SUB_I64(SDNode *N);
278   void SelectAddcSubb(SDNode *N);
279   void SelectUADDO_USUBO(SDNode *N);
280   void SelectDIV_SCALE(SDNode *N);
281   void SelectDIV_FMAS(SDNode *N);
282   void SelectMAD_64_32(SDNode *N);
283   void SelectFMA_W_CHAIN(SDNode *N);
284   void SelectFMUL_W_CHAIN(SDNode *N);
285 
286   SDNode *getS_BFE(unsigned Opcode, const SDLoc &DL, SDValue Val,
287                    uint32_t Offset, uint32_t Width);
288   void SelectS_BFEFromShifts(SDNode *N);
289   void SelectS_BFE(SDNode *N);
290   bool isCBranchSCC(const SDNode *N) const;
291   void SelectBRCOND(SDNode *N);
292   void SelectFMAD_FMA(SDNode *N);
293   void SelectATOMIC_CMP_SWAP(SDNode *N);
294   void SelectDSAppendConsume(SDNode *N, unsigned IntrID);
295   void SelectDS_GWS(SDNode *N, unsigned IntrID);
296   void SelectInterpP1F16(SDNode *N);
297   void SelectINTRINSIC_W_CHAIN(SDNode *N);
298   void SelectINTRINSIC_WO_CHAIN(SDNode *N);
299   void SelectINTRINSIC_VOID(SDNode *N);
300 
301 protected:
302   // Include the pieces autogenerated from the target description.
303 #include "AMDGPUGenDAGISel.inc"
304 };
305 
306 class R600DAGToDAGISel : public AMDGPUDAGToDAGISel {
307   const R600Subtarget *Subtarget;
308 
309   bool isConstantLoad(const MemSDNode *N, int cbID) const;
310   bool SelectGlobalValueConstantOffset(SDValue Addr, SDValue& IntPtr);
311   bool SelectGlobalValueVariableOffset(SDValue Addr, SDValue &BaseReg,
312                                        SDValue& Offset);
313 public:
314   explicit R600DAGToDAGISel(TargetMachine *TM, CodeGenOpt::Level OptLevel) :
315       AMDGPUDAGToDAGISel(TM, OptLevel) {}
316 
317   void Select(SDNode *N) override;
318 
319   bool SelectADDRIndirect(SDValue Addr, SDValue &Base,
320                           SDValue &Offset) override;
321   bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
322                           SDValue &Offset) override;
323 
324   bool runOnMachineFunction(MachineFunction &MF) override;
325 
326   void PreprocessISelDAG() override {}
327 
328 protected:
329   // Include the pieces autogenerated from the target description.
330 #include "R600GenDAGISel.inc"
331 };
332 
333 static SDValue stripBitcast(SDValue Val) {
334   return Val.getOpcode() == ISD::BITCAST ? Val.getOperand(0) : Val;
335 }
336 
337 // Figure out if this is really an extract of the high 16-bits of a dword.
338 static bool isExtractHiElt(SDValue In, SDValue &Out) {
339   In = stripBitcast(In);
340   if (In.getOpcode() != ISD::TRUNCATE)
341     return false;
342 
343   SDValue Srl = In.getOperand(0);
344   if (Srl.getOpcode() == ISD::SRL) {
345     if (ConstantSDNode *ShiftAmt = dyn_cast<ConstantSDNode>(Srl.getOperand(1))) {
346       if (ShiftAmt->getZExtValue() == 16) {
347         Out = stripBitcast(Srl.getOperand(0));
348         return true;
349       }
350     }
351   }
352 
353   return false;
354 }
355 
356 // Look through operations that obscure just looking at the low 16-bits of the
357 // same register.
358 static SDValue stripExtractLoElt(SDValue In) {
359   if (In.getOpcode() == ISD::TRUNCATE) {
360     SDValue Src = In.getOperand(0);
361     if (Src.getValueType().getSizeInBits() == 32)
362       return stripBitcast(Src);
363   }
364 
365   return In;
366 }
367 
368 }  // end anonymous namespace
369 
370 INITIALIZE_PASS_BEGIN(AMDGPUDAGToDAGISel, "amdgpu-isel",
371                       "AMDGPU DAG->DAG Pattern Instruction Selection", false, false)
372 INITIALIZE_PASS_DEPENDENCY(AMDGPUArgumentUsageInfo)
373 INITIALIZE_PASS_DEPENDENCY(AMDGPUPerfHintAnalysis)
374 INITIALIZE_PASS_DEPENDENCY(LegacyDivergenceAnalysis)
375 #ifdef EXPENSIVE_CHECKS
376 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
377 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
378 #endif
379 INITIALIZE_PASS_END(AMDGPUDAGToDAGISel, "amdgpu-isel",
380                     "AMDGPU DAG->DAG Pattern Instruction Selection", false, false)
381 
382 /// This pass converts a legalized DAG into a AMDGPU-specific
383 // DAG, ready for instruction scheduling.
384 FunctionPass *llvm::createAMDGPUISelDag(TargetMachine *TM,
385                                         CodeGenOpt::Level OptLevel) {
386   return new AMDGPUDAGToDAGISel(TM, OptLevel);
387 }
388 
389 /// This pass converts a legalized DAG into a R600-specific
390 // DAG, ready for instruction scheduling.
391 FunctionPass *llvm::createR600ISelDag(TargetMachine *TM,
392                                       CodeGenOpt::Level OptLevel) {
393   return new R600DAGToDAGISel(TM, OptLevel);
394 }
395 
396 bool AMDGPUDAGToDAGISel::runOnMachineFunction(MachineFunction &MF) {
397 #ifdef EXPENSIVE_CHECKS
398   DominatorTree & DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
399   LoopInfo * LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
400   for (auto &L : LI->getLoopsInPreorder()) {
401     assert(L->isLCSSAForm(DT));
402   }
403 #endif
404   Subtarget = &MF.getSubtarget<GCNSubtarget>();
405   Mode = AMDGPU::SIModeRegisterDefaults(MF.getFunction(), *Subtarget);
406   return SelectionDAGISel::runOnMachineFunction(MF);
407 }
408 
409 bool AMDGPUDAGToDAGISel::matchLoadD16FromBuildVector(SDNode *N) const {
410   assert(Subtarget->d16PreservesUnusedBits());
411   MVT VT = N->getValueType(0).getSimpleVT();
412   if (VT != MVT::v2i16 && VT != MVT::v2f16)
413     return false;
414 
415   SDValue Lo = N->getOperand(0);
416   SDValue Hi = N->getOperand(1);
417 
418   LoadSDNode *LdHi = dyn_cast<LoadSDNode>(stripBitcast(Hi));
419 
420   // build_vector lo, (load ptr) -> load_d16_hi ptr, lo
421   // build_vector lo, (zextload ptr from i8) -> load_d16_hi_u8 ptr, lo
422   // build_vector lo, (sextload ptr from i8) -> load_d16_hi_i8 ptr, lo
423 
424   // Need to check for possible indirect dependencies on the other half of the
425   // vector to avoid introducing a cycle.
426   if (LdHi && Hi.hasOneUse() && !LdHi->isPredecessorOf(Lo.getNode())) {
427     SDVTList VTList = CurDAG->getVTList(VT, MVT::Other);
428 
429     SDValue TiedIn = CurDAG->getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), VT, Lo);
430     SDValue Ops[] = {
431       LdHi->getChain(), LdHi->getBasePtr(), TiedIn
432     };
433 
434     unsigned LoadOp = AMDGPUISD::LOAD_D16_HI;
435     if (LdHi->getMemoryVT() == MVT::i8) {
436       LoadOp = LdHi->getExtensionType() == ISD::SEXTLOAD ?
437         AMDGPUISD::LOAD_D16_HI_I8 : AMDGPUISD::LOAD_D16_HI_U8;
438     } else {
439       assert(LdHi->getMemoryVT() == MVT::i16);
440     }
441 
442     SDValue NewLoadHi =
443       CurDAG->getMemIntrinsicNode(LoadOp, SDLoc(LdHi), VTList,
444                                   Ops, LdHi->getMemoryVT(),
445                                   LdHi->getMemOperand());
446 
447     CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), NewLoadHi);
448     CurDAG->ReplaceAllUsesOfValueWith(SDValue(LdHi, 1), NewLoadHi.getValue(1));
449     return true;
450   }
451 
452   // build_vector (load ptr), hi -> load_d16_lo ptr, hi
453   // build_vector (zextload ptr from i8), hi -> load_d16_lo_u8 ptr, hi
454   // build_vector (sextload ptr from i8), hi -> load_d16_lo_i8 ptr, hi
455   LoadSDNode *LdLo = dyn_cast<LoadSDNode>(stripBitcast(Lo));
456   if (LdLo && Lo.hasOneUse()) {
457     SDValue TiedIn = getHi16Elt(Hi);
458     if (!TiedIn || LdLo->isPredecessorOf(TiedIn.getNode()))
459       return false;
460 
461     SDVTList VTList = CurDAG->getVTList(VT, MVT::Other);
462     unsigned LoadOp = AMDGPUISD::LOAD_D16_LO;
463     if (LdLo->getMemoryVT() == MVT::i8) {
464       LoadOp = LdLo->getExtensionType() == ISD::SEXTLOAD ?
465         AMDGPUISD::LOAD_D16_LO_I8 : AMDGPUISD::LOAD_D16_LO_U8;
466     } else {
467       assert(LdLo->getMemoryVT() == MVT::i16);
468     }
469 
470     TiedIn = CurDAG->getNode(ISD::BITCAST, SDLoc(N), VT, TiedIn);
471 
472     SDValue Ops[] = {
473       LdLo->getChain(), LdLo->getBasePtr(), TiedIn
474     };
475 
476     SDValue NewLoadLo =
477       CurDAG->getMemIntrinsicNode(LoadOp, SDLoc(LdLo), VTList,
478                                   Ops, LdLo->getMemoryVT(),
479                                   LdLo->getMemOperand());
480 
481     CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), NewLoadLo);
482     CurDAG->ReplaceAllUsesOfValueWith(SDValue(LdLo, 1), NewLoadLo.getValue(1));
483     return true;
484   }
485 
486   return false;
487 }
488 
489 void AMDGPUDAGToDAGISel::PreprocessISelDAG() {
490   if (!Subtarget->d16PreservesUnusedBits())
491     return;
492 
493   SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end();
494 
495   bool MadeChange = false;
496   while (Position != CurDAG->allnodes_begin()) {
497     SDNode *N = &*--Position;
498     if (N->use_empty())
499       continue;
500 
501     switch (N->getOpcode()) {
502     case ISD::BUILD_VECTOR:
503       MadeChange |= matchLoadD16FromBuildVector(N);
504       break;
505     default:
506       break;
507     }
508   }
509 
510   if (MadeChange) {
511     CurDAG->RemoveDeadNodes();
512     LLVM_DEBUG(dbgs() << "After PreProcess:\n";
513                CurDAG->dump(););
514   }
515 }
516 
517 bool AMDGPUDAGToDAGISel::isNoNanSrc(SDValue N) const {
518   if (TM.Options.NoNaNsFPMath)
519     return true;
520 
521   // TODO: Move into isKnownNeverNaN
522   if (N->getFlags().isDefined())
523     return N->getFlags().hasNoNaNs();
524 
525   return CurDAG->isKnownNeverNaN(N);
526 }
527 
528 bool AMDGPUDAGToDAGISel::isInlineImmediate(const SDNode *N,
529                                            bool Negated) const {
530   if (N->isUndef())
531     return true;
532 
533   const SIInstrInfo *TII = Subtarget->getInstrInfo();
534   if (Negated) {
535     if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N))
536       return TII->isInlineConstant(-C->getAPIntValue());
537 
538     if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N))
539       return TII->isInlineConstant(-C->getValueAPF().bitcastToAPInt());
540 
541   } else {
542     if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N))
543       return TII->isInlineConstant(C->getAPIntValue());
544 
545     if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N))
546       return TII->isInlineConstant(C->getValueAPF().bitcastToAPInt());
547   }
548 
549   return false;
550 }
551 
552 /// Determine the register class for \p OpNo
553 /// \returns The register class of the virtual register that will be used for
554 /// the given operand number \OpNo or NULL if the register class cannot be
555 /// determined.
556 const TargetRegisterClass *AMDGPUDAGToDAGISel::getOperandRegClass(SDNode *N,
557                                                           unsigned OpNo) const {
558   if (!N->isMachineOpcode()) {
559     if (N->getOpcode() == ISD::CopyToReg) {
560       unsigned Reg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
561       if (Register::isVirtualRegister(Reg)) {
562         MachineRegisterInfo &MRI = CurDAG->getMachineFunction().getRegInfo();
563         return MRI.getRegClass(Reg);
564       }
565 
566       const SIRegisterInfo *TRI
567         = static_cast<const GCNSubtarget *>(Subtarget)->getRegisterInfo();
568       return TRI->getPhysRegClass(Reg);
569     }
570 
571     return nullptr;
572   }
573 
574   switch (N->getMachineOpcode()) {
575   default: {
576     const MCInstrDesc &Desc =
577         Subtarget->getInstrInfo()->get(N->getMachineOpcode());
578     unsigned OpIdx = Desc.getNumDefs() + OpNo;
579     if (OpIdx >= Desc.getNumOperands())
580       return nullptr;
581     int RegClass = Desc.OpInfo[OpIdx].RegClass;
582     if (RegClass == -1)
583       return nullptr;
584 
585     return Subtarget->getRegisterInfo()->getRegClass(RegClass);
586   }
587   case AMDGPU::REG_SEQUENCE: {
588     unsigned RCID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
589     const TargetRegisterClass *SuperRC =
590         Subtarget->getRegisterInfo()->getRegClass(RCID);
591 
592     SDValue SubRegOp = N->getOperand(OpNo + 1);
593     unsigned SubRegIdx = cast<ConstantSDNode>(SubRegOp)->getZExtValue();
594     return Subtarget->getRegisterInfo()->getSubClassWithSubReg(SuperRC,
595                                                               SubRegIdx);
596   }
597   }
598 }
599 
600 SDNode *AMDGPUDAGToDAGISel::glueCopyToOp(SDNode *N, SDValue NewChain,
601                                          SDValue Glue) const {
602   SmallVector <SDValue, 8> Ops;
603   Ops.push_back(NewChain); // Replace the chain.
604   for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
605     Ops.push_back(N->getOperand(i));
606 
607   Ops.push_back(Glue);
608   return CurDAG->MorphNodeTo(N, N->getOpcode(), N->getVTList(), Ops);
609 }
610 
611 SDNode *AMDGPUDAGToDAGISel::glueCopyToM0(SDNode *N, SDValue Val) const {
612   const SITargetLowering& Lowering =
613     *static_cast<const SITargetLowering*>(getTargetLowering());
614 
615   assert(N->getOperand(0).getValueType() == MVT::Other && "Expected chain");
616 
617   SDValue M0 = Lowering.copyToM0(*CurDAG, N->getOperand(0), SDLoc(N), Val);
618   return glueCopyToOp(N, M0, M0.getValue(1));
619 }
620 
621 SDNode *AMDGPUDAGToDAGISel::glueCopyToM0LDSInit(SDNode *N) const {
622   unsigned AS = cast<MemSDNode>(N)->getAddressSpace();
623   if (AS == AMDGPUAS::LOCAL_ADDRESS) {
624     if (Subtarget->ldsRequiresM0Init())
625       return glueCopyToM0(N, CurDAG->getTargetConstant(-1, SDLoc(N), MVT::i32));
626   } else if (AS == AMDGPUAS::REGION_ADDRESS) {
627     MachineFunction &MF = CurDAG->getMachineFunction();
628     unsigned Value = MF.getInfo<SIMachineFunctionInfo>()->getGDSSize();
629     return
630         glueCopyToM0(N, CurDAG->getTargetConstant(Value, SDLoc(N), MVT::i32));
631   }
632   return N;
633 }
634 
635 MachineSDNode *AMDGPUDAGToDAGISel::buildSMovImm64(SDLoc &DL, uint64_t Imm,
636                                                   EVT VT) const {
637   SDNode *Lo = CurDAG->getMachineNode(
638       AMDGPU::S_MOV_B32, DL, MVT::i32,
639       CurDAG->getTargetConstant(Imm & 0xFFFFFFFF, DL, MVT::i32));
640   SDNode *Hi =
641       CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32,
642                              CurDAG->getTargetConstant(Imm >> 32, DL, MVT::i32));
643   const SDValue Ops[] = {
644       CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32),
645       SDValue(Lo, 0), CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
646       SDValue(Hi, 0), CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32)};
647 
648   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL, VT, Ops);
649 }
650 
651 static unsigned selectSGPRVectorRegClassID(unsigned NumVectorElts) {
652   switch (NumVectorElts) {
653   case 1:
654     return AMDGPU::SReg_32RegClassID;
655   case 2:
656     return AMDGPU::SReg_64RegClassID;
657   case 3:
658     return AMDGPU::SGPR_96RegClassID;
659   case 4:
660     return AMDGPU::SGPR_128RegClassID;
661   case 5:
662     return AMDGPU::SGPR_160RegClassID;
663   case 8:
664     return AMDGPU::SReg_256RegClassID;
665   case 16:
666     return AMDGPU::SReg_512RegClassID;
667   case 32:
668     return AMDGPU::SReg_1024RegClassID;
669   }
670 
671   llvm_unreachable("invalid vector size");
672 }
673 
674 void AMDGPUDAGToDAGISel::SelectBuildVector(SDNode *N, unsigned RegClassID) {
675   EVT VT = N->getValueType(0);
676   unsigned NumVectorElts = VT.getVectorNumElements();
677   EVT EltVT = VT.getVectorElementType();
678   SDLoc DL(N);
679   SDValue RegClass = CurDAG->getTargetConstant(RegClassID, DL, MVT::i32);
680 
681   if (NumVectorElts == 1) {
682     CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS, EltVT, N->getOperand(0),
683                          RegClass);
684     return;
685   }
686 
687   assert(NumVectorElts <= 32 && "Vectors with more than 32 elements not "
688                                   "supported yet");
689   // 32 = Max Num Vector Elements
690   // 2 = 2 REG_SEQUENCE operands per element (value, subreg index)
691   // 1 = Vector Register Class
692   SmallVector<SDValue, 32 * 2 + 1> RegSeqArgs(NumVectorElts * 2 + 1);
693 
694   bool IsGCN = CurDAG->getSubtarget().getTargetTriple().getArch() ==
695                Triple::amdgcn;
696   RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, DL, MVT::i32);
697   bool IsRegSeq = true;
698   unsigned NOps = N->getNumOperands();
699   for (unsigned i = 0; i < NOps; i++) {
700     // XXX: Why is this here?
701     if (isa<RegisterSDNode>(N->getOperand(i))) {
702       IsRegSeq = false;
703       break;
704     }
705     unsigned Sub = IsGCN ? SIRegisterInfo::getSubRegFromChannel(i)
706                          : R600RegisterInfo::getSubRegFromChannel(i);
707     RegSeqArgs[1 + (2 * i)] = N->getOperand(i);
708     RegSeqArgs[1 + (2 * i) + 1] = CurDAG->getTargetConstant(Sub, DL, MVT::i32);
709   }
710   if (NOps != NumVectorElts) {
711     // Fill in the missing undef elements if this was a scalar_to_vector.
712     assert(N->getOpcode() == ISD::SCALAR_TO_VECTOR && NOps < NumVectorElts);
713     MachineSDNode *ImpDef = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
714                                                    DL, EltVT);
715     for (unsigned i = NOps; i < NumVectorElts; ++i) {
716       unsigned Sub = IsGCN ? SIRegisterInfo::getSubRegFromChannel(i)
717                            : R600RegisterInfo::getSubRegFromChannel(i);
718       RegSeqArgs[1 + (2 * i)] = SDValue(ImpDef, 0);
719       RegSeqArgs[1 + (2 * i) + 1] =
720           CurDAG->getTargetConstant(Sub, DL, MVT::i32);
721     }
722   }
723 
724   if (!IsRegSeq)
725     SelectCode(N);
726   CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(), RegSeqArgs);
727 }
728 
729 void AMDGPUDAGToDAGISel::Select(SDNode *N) {
730   unsigned int Opc = N->getOpcode();
731   if (N->isMachineOpcode()) {
732     N->setNodeId(-1);
733     return;   // Already selected.
734   }
735 
736   // isa<MemSDNode> almost works but is slightly too permissive for some DS
737   // intrinsics.
738   if (Opc == ISD::LOAD || Opc == ISD::STORE || isa<AtomicSDNode>(N) ||
739       (Opc == AMDGPUISD::ATOMIC_INC || Opc == AMDGPUISD::ATOMIC_DEC ||
740        Opc == ISD::ATOMIC_LOAD_FADD ||
741        Opc == AMDGPUISD::ATOMIC_LOAD_FMIN ||
742        Opc == AMDGPUISD::ATOMIC_LOAD_FMAX)) {
743     N = glueCopyToM0LDSInit(N);
744     SelectCode(N);
745     return;
746   }
747 
748   switch (Opc) {
749   default:
750     break;
751   // We are selecting i64 ADD here instead of custom lower it during
752   // DAG legalization, so we can fold some i64 ADDs used for address
753   // calculation into the LOAD and STORE instructions.
754   case ISD::ADDC:
755   case ISD::ADDE:
756   case ISD::SUBC:
757   case ISD::SUBE: {
758     if (N->getValueType(0) != MVT::i64)
759       break;
760 
761     SelectADD_SUB_I64(N);
762     return;
763   }
764   case ISD::ADDCARRY:
765   case ISD::SUBCARRY:
766     if (N->getValueType(0) != MVT::i32)
767       break;
768 
769     SelectAddcSubb(N);
770     return;
771   case ISD::UADDO:
772   case ISD::USUBO: {
773     SelectUADDO_USUBO(N);
774     return;
775   }
776   case AMDGPUISD::FMUL_W_CHAIN: {
777     SelectFMUL_W_CHAIN(N);
778     return;
779   }
780   case AMDGPUISD::FMA_W_CHAIN: {
781     SelectFMA_W_CHAIN(N);
782     return;
783   }
784 
785   case ISD::SCALAR_TO_VECTOR:
786   case ISD::BUILD_VECTOR: {
787     EVT VT = N->getValueType(0);
788     unsigned NumVectorElts = VT.getVectorNumElements();
789     if (VT.getScalarSizeInBits() == 16) {
790       if (Opc == ISD::BUILD_VECTOR && NumVectorElts == 2) {
791         if (SDNode *Packed = packConstantV2I16(N, *CurDAG)) {
792           ReplaceNode(N, Packed);
793           return;
794         }
795       }
796 
797       break;
798     }
799 
800     assert(VT.getVectorElementType().bitsEq(MVT::i32));
801     unsigned RegClassID = selectSGPRVectorRegClassID(NumVectorElts);
802     SelectBuildVector(N, RegClassID);
803     return;
804   }
805   case ISD::BUILD_PAIR: {
806     SDValue RC, SubReg0, SubReg1;
807     SDLoc DL(N);
808     if (N->getValueType(0) == MVT::i128) {
809       RC = CurDAG->getTargetConstant(AMDGPU::SGPR_128RegClassID, DL, MVT::i32);
810       SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0_sub1, DL, MVT::i32);
811       SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub2_sub3, DL, MVT::i32);
812     } else if (N->getValueType(0) == MVT::i64) {
813       RC = CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32);
814       SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
815       SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
816     } else {
817       llvm_unreachable("Unhandled value type for BUILD_PAIR");
818     }
819     const SDValue Ops[] = { RC, N->getOperand(0), SubReg0,
820                             N->getOperand(1), SubReg1 };
821     ReplaceNode(N, CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL,
822                                           N->getValueType(0), Ops));
823     return;
824   }
825 
826   case ISD::Constant:
827   case ISD::ConstantFP: {
828     if (N->getValueType(0).getSizeInBits() != 64 || isInlineImmediate(N))
829       break;
830 
831     uint64_t Imm;
832     if (ConstantFPSDNode *FP = dyn_cast<ConstantFPSDNode>(N))
833       Imm = FP->getValueAPF().bitcastToAPInt().getZExtValue();
834     else {
835       ConstantSDNode *C = cast<ConstantSDNode>(N);
836       Imm = C->getZExtValue();
837     }
838 
839     SDLoc DL(N);
840     ReplaceNode(N, buildSMovImm64(DL, Imm, N->getValueType(0)));
841     return;
842   }
843   case AMDGPUISD::BFE_I32:
844   case AMDGPUISD::BFE_U32: {
845     // There is a scalar version available, but unlike the vector version which
846     // has a separate operand for the offset and width, the scalar version packs
847     // the width and offset into a single operand. Try to move to the scalar
848     // version if the offsets are constant, so that we can try to keep extended
849     // loads of kernel arguments in SGPRs.
850 
851     // TODO: Technically we could try to pattern match scalar bitshifts of
852     // dynamic values, but it's probably not useful.
853     ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
854     if (!Offset)
855       break;
856 
857     ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
858     if (!Width)
859       break;
860 
861     bool Signed = Opc == AMDGPUISD::BFE_I32;
862 
863     uint32_t OffsetVal = Offset->getZExtValue();
864     uint32_t WidthVal = Width->getZExtValue();
865 
866     ReplaceNode(N, getS_BFE(Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32,
867                             SDLoc(N), N->getOperand(0), OffsetVal, WidthVal));
868     return;
869   }
870   case AMDGPUISD::DIV_SCALE: {
871     SelectDIV_SCALE(N);
872     return;
873   }
874   case AMDGPUISD::DIV_FMAS: {
875     SelectDIV_FMAS(N);
876     return;
877   }
878   case AMDGPUISD::MAD_I64_I32:
879   case AMDGPUISD::MAD_U64_U32: {
880     SelectMAD_64_32(N);
881     return;
882   }
883   case ISD::CopyToReg: {
884     const SITargetLowering& Lowering =
885       *static_cast<const SITargetLowering*>(getTargetLowering());
886     N = Lowering.legalizeTargetIndependentNode(N, *CurDAG);
887     break;
888   }
889   case ISD::AND:
890   case ISD::SRL:
891   case ISD::SRA:
892   case ISD::SIGN_EXTEND_INREG:
893     if (N->getValueType(0) != MVT::i32)
894       break;
895 
896     SelectS_BFE(N);
897     return;
898   case ISD::BRCOND:
899     SelectBRCOND(N);
900     return;
901   case ISD::FMAD:
902   case ISD::FMA:
903     SelectFMAD_FMA(N);
904     return;
905   case AMDGPUISD::ATOMIC_CMP_SWAP:
906     SelectATOMIC_CMP_SWAP(N);
907     return;
908   case AMDGPUISD::CVT_PKRTZ_F16_F32:
909   case AMDGPUISD::CVT_PKNORM_I16_F32:
910   case AMDGPUISD::CVT_PKNORM_U16_F32:
911   case AMDGPUISD::CVT_PK_U16_U32:
912   case AMDGPUISD::CVT_PK_I16_I32: {
913     // Hack around using a legal type if f16 is illegal.
914     if (N->getValueType(0) == MVT::i32) {
915       MVT NewVT = Opc == AMDGPUISD::CVT_PKRTZ_F16_F32 ? MVT::v2f16 : MVT::v2i16;
916       N = CurDAG->MorphNodeTo(N, N->getOpcode(), CurDAG->getVTList(NewVT),
917                               { N->getOperand(0), N->getOperand(1) });
918       SelectCode(N);
919       return;
920     }
921 
922     break;
923   }
924   case ISD::INTRINSIC_W_CHAIN: {
925     SelectINTRINSIC_W_CHAIN(N);
926     return;
927   }
928   case ISD::INTRINSIC_WO_CHAIN: {
929     SelectINTRINSIC_WO_CHAIN(N);
930     return;
931   }
932   case ISD::INTRINSIC_VOID: {
933     SelectINTRINSIC_VOID(N);
934     return;
935   }
936   }
937 
938   SelectCode(N);
939 }
940 
941 bool AMDGPUDAGToDAGISel::isUniformBr(const SDNode *N) const {
942   const BasicBlock *BB = FuncInfo->MBB->getBasicBlock();
943   const Instruction *Term = BB->getTerminator();
944   return Term->getMetadata("amdgpu.uniform") ||
945          Term->getMetadata("structurizecfg.uniform");
946 }
947 
948 StringRef AMDGPUDAGToDAGISel::getPassName() const {
949   return "AMDGPU DAG->DAG Pattern Instruction Selection";
950 }
951 
952 //===----------------------------------------------------------------------===//
953 // Complex Patterns
954 //===----------------------------------------------------------------------===//
955 
956 bool AMDGPUDAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
957                                             SDValue &Offset) {
958   return false;
959 }
960 
961 bool AMDGPUDAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base,
962                                             SDValue &Offset) {
963   ConstantSDNode *C;
964   SDLoc DL(Addr);
965 
966   if ((C = dyn_cast<ConstantSDNode>(Addr))) {
967     Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32);
968     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
969   } else if ((Addr.getOpcode() == AMDGPUISD::DWORDADDR) &&
970              (C = dyn_cast<ConstantSDNode>(Addr.getOperand(0)))) {
971     Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32);
972     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
973   } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
974             (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) {
975     Base = Addr.getOperand(0);
976     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
977   } else {
978     Base = Addr;
979     Offset = CurDAG->getTargetConstant(0, DL, MVT::i32);
980   }
981 
982   return true;
983 }
984 
985 SDValue AMDGPUDAGToDAGISel::getMaterializedScalarImm32(int64_t Val,
986                                                        const SDLoc &DL) const {
987   SDNode *Mov = CurDAG->getMachineNode(
988     AMDGPU::S_MOV_B32, DL, MVT::i32,
989     CurDAG->getTargetConstant(Val, DL, MVT::i32));
990   return SDValue(Mov, 0);
991 }
992 
993 // FIXME: Should only handle addcarry/subcarry
994 void AMDGPUDAGToDAGISel::SelectADD_SUB_I64(SDNode *N) {
995   SDLoc DL(N);
996   SDValue LHS = N->getOperand(0);
997   SDValue RHS = N->getOperand(1);
998 
999   unsigned Opcode = N->getOpcode();
1000   bool ConsumeCarry = (Opcode == ISD::ADDE || Opcode == ISD::SUBE);
1001   bool ProduceCarry =
1002       ConsumeCarry || Opcode == ISD::ADDC || Opcode == ISD::SUBC;
1003   bool IsAdd = Opcode == ISD::ADD || Opcode == ISD::ADDC || Opcode == ISD::ADDE;
1004 
1005   SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
1006   SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
1007 
1008   SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
1009                                        DL, MVT::i32, LHS, Sub0);
1010   SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
1011                                        DL, MVT::i32, LHS, Sub1);
1012 
1013   SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
1014                                        DL, MVT::i32, RHS, Sub0);
1015   SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
1016                                        DL, MVT::i32, RHS, Sub1);
1017 
1018   SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue);
1019 
1020   unsigned Opc = IsAdd ? AMDGPU::S_ADD_U32 : AMDGPU::S_SUB_U32;
1021   unsigned CarryOpc = IsAdd ? AMDGPU::S_ADDC_U32 : AMDGPU::S_SUBB_U32;
1022 
1023   SDNode *AddLo;
1024   if (!ConsumeCarry) {
1025     SDValue Args[] = { SDValue(Lo0, 0), SDValue(Lo1, 0) };
1026     AddLo = CurDAG->getMachineNode(Opc, DL, VTList, Args);
1027   } else {
1028     SDValue Args[] = { SDValue(Lo0, 0), SDValue(Lo1, 0), N->getOperand(2) };
1029     AddLo = CurDAG->getMachineNode(CarryOpc, DL, VTList, Args);
1030   }
1031   SDValue AddHiArgs[] = {
1032     SDValue(Hi0, 0),
1033     SDValue(Hi1, 0),
1034     SDValue(AddLo, 1)
1035   };
1036   SDNode *AddHi = CurDAG->getMachineNode(CarryOpc, DL, VTList, AddHiArgs);
1037 
1038   SDValue RegSequenceArgs[] = {
1039     CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32),
1040     SDValue(AddLo,0),
1041     Sub0,
1042     SDValue(AddHi,0),
1043     Sub1,
1044   };
1045   SDNode *RegSequence = CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL,
1046                                                MVT::i64, RegSequenceArgs);
1047 
1048   if (ProduceCarry) {
1049     // Replace the carry-use
1050     ReplaceUses(SDValue(N, 1), SDValue(AddHi, 1));
1051   }
1052 
1053   // Replace the remaining uses.
1054   ReplaceNode(N, RegSequence);
1055 }
1056 
1057 void AMDGPUDAGToDAGISel::SelectAddcSubb(SDNode *N) {
1058   SDLoc DL(N);
1059   SDValue LHS = N->getOperand(0);
1060   SDValue RHS = N->getOperand(1);
1061   SDValue CI = N->getOperand(2);
1062 
1063   unsigned Opc = N->getOpcode() == ISD::ADDCARRY ? AMDGPU::V_ADDC_U32_e64
1064                                                  : AMDGPU::V_SUBB_U32_e64;
1065   CurDAG->SelectNodeTo(
1066       N, Opc, N->getVTList(),
1067       {LHS, RHS, CI, CurDAG->getTargetConstant(0, {}, MVT::i1) /*clamp bit*/});
1068 }
1069 
1070 void AMDGPUDAGToDAGISel::SelectUADDO_USUBO(SDNode *N) {
1071   // The name of the opcodes are misleading. v_add_i32/v_sub_i32 have unsigned
1072   // carry out despite the _i32 name. These were renamed in VI to _U32.
1073   // FIXME: We should probably rename the opcodes here.
1074   unsigned Opc = N->getOpcode() == ISD::UADDO ?
1075     AMDGPU::V_ADD_I32_e64 : AMDGPU::V_SUB_I32_e64;
1076 
1077   CurDAG->SelectNodeTo(
1078       N, Opc, N->getVTList(),
1079       {N->getOperand(0), N->getOperand(1),
1080        CurDAG->getTargetConstant(0, {}, MVT::i1) /*clamp bit*/});
1081 }
1082 
1083 void AMDGPUDAGToDAGISel::SelectFMA_W_CHAIN(SDNode *N) {
1084   SDLoc SL(N);
1085   //  src0_modifiers, src0,  src1_modifiers, src1, src2_modifiers, src2, clamp, omod
1086   SDValue Ops[10];
1087 
1088   SelectVOP3Mods0(N->getOperand(1), Ops[1], Ops[0], Ops[6], Ops[7]);
1089   SelectVOP3Mods(N->getOperand(2), Ops[3], Ops[2]);
1090   SelectVOP3Mods(N->getOperand(3), Ops[5], Ops[4]);
1091   Ops[8] = N->getOperand(0);
1092   Ops[9] = N->getOperand(4);
1093 
1094   CurDAG->SelectNodeTo(N, AMDGPU::V_FMA_F32, N->getVTList(), Ops);
1095 }
1096 
1097 void AMDGPUDAGToDAGISel::SelectFMUL_W_CHAIN(SDNode *N) {
1098   SDLoc SL(N);
1099   //    src0_modifiers, src0,  src1_modifiers, src1, clamp, omod
1100   SDValue Ops[8];
1101 
1102   SelectVOP3Mods0(N->getOperand(1), Ops[1], Ops[0], Ops[4], Ops[5]);
1103   SelectVOP3Mods(N->getOperand(2), Ops[3], Ops[2]);
1104   Ops[6] = N->getOperand(0);
1105   Ops[7] = N->getOperand(3);
1106 
1107   CurDAG->SelectNodeTo(N, AMDGPU::V_MUL_F32_e64, N->getVTList(), Ops);
1108 }
1109 
1110 // We need to handle this here because tablegen doesn't support matching
1111 // instructions with multiple outputs.
1112 void AMDGPUDAGToDAGISel::SelectDIV_SCALE(SDNode *N) {
1113   SDLoc SL(N);
1114   EVT VT = N->getValueType(0);
1115 
1116   assert(VT == MVT::f32 || VT == MVT::f64);
1117 
1118   unsigned Opc
1119     = (VT == MVT::f64) ? AMDGPU::V_DIV_SCALE_F64 : AMDGPU::V_DIV_SCALE_F32;
1120 
1121   SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2) };
1122   CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
1123 }
1124 
1125 void AMDGPUDAGToDAGISel::SelectDIV_FMAS(SDNode *N) {
1126   const GCNSubtarget *ST = static_cast<const GCNSubtarget *>(Subtarget);
1127   const SIRegisterInfo *TRI = ST->getRegisterInfo();
1128 
1129   SDLoc SL(N);
1130   EVT VT = N->getValueType(0);
1131 
1132   assert(VT == MVT::f32 || VT == MVT::f64);
1133 
1134   unsigned Opc
1135     = (VT == MVT::f64) ? AMDGPU::V_DIV_FMAS_F64 : AMDGPU::V_DIV_FMAS_F32;
1136 
1137   SDValue CarryIn = N->getOperand(3);
1138   // V_DIV_FMAS implicitly reads VCC.
1139   SDValue VCC = CurDAG->getCopyToReg(CurDAG->getEntryNode(), SL,
1140                                      TRI->getVCC(), CarryIn, SDValue());
1141 
1142   SDValue Ops[10];
1143 
1144   SelectVOP3Mods0(N->getOperand(0), Ops[1], Ops[0], Ops[6], Ops[7]);
1145   SelectVOP3Mods(N->getOperand(1), Ops[3], Ops[2]);
1146   SelectVOP3Mods(N->getOperand(2), Ops[5], Ops[4]);
1147 
1148   Ops[8] = VCC;
1149   Ops[9] = VCC.getValue(1);
1150 
1151   CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
1152 }
1153 
1154 // We need to handle this here because tablegen doesn't support matching
1155 // instructions with multiple outputs.
1156 void AMDGPUDAGToDAGISel::SelectMAD_64_32(SDNode *N) {
1157   SDLoc SL(N);
1158   bool Signed = N->getOpcode() == AMDGPUISD::MAD_I64_I32;
1159   unsigned Opc = Signed ? AMDGPU::V_MAD_I64_I32 : AMDGPU::V_MAD_U64_U32;
1160 
1161   SDValue Clamp = CurDAG->getTargetConstant(0, SL, MVT::i1);
1162   SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
1163                     Clamp };
1164   CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
1165 }
1166 
1167 bool AMDGPUDAGToDAGISel::isDSOffsetLegal(SDValue Base, unsigned Offset,
1168                                          unsigned OffsetBits) const {
1169   if ((OffsetBits == 16 && !isUInt<16>(Offset)) ||
1170       (OffsetBits == 8 && !isUInt<8>(Offset)))
1171     return false;
1172 
1173   if (Subtarget->hasUsableDSOffset() ||
1174       Subtarget->unsafeDSOffsetFoldingEnabled())
1175     return true;
1176 
1177   // On Southern Islands instruction with a negative base value and an offset
1178   // don't seem to work.
1179   return CurDAG->SignBitIsZero(Base);
1180 }
1181 
1182 bool AMDGPUDAGToDAGISel::SelectDS1Addr1Offset(SDValue Addr, SDValue &Base,
1183                                               SDValue &Offset) const {
1184   SDLoc DL(Addr);
1185   if (CurDAG->isBaseWithConstantOffset(Addr)) {
1186     SDValue N0 = Addr.getOperand(0);
1187     SDValue N1 = Addr.getOperand(1);
1188     ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
1189     if (isDSOffsetLegal(N0, C1->getSExtValue(), 16)) {
1190       // (add n0, c0)
1191       Base = N0;
1192       Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16);
1193       return true;
1194     }
1195   } else if (Addr.getOpcode() == ISD::SUB) {
1196     // sub C, x -> add (sub 0, x), C
1197     if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr.getOperand(0))) {
1198       int64_t ByteOffset = C->getSExtValue();
1199       if (isUInt<16>(ByteOffset)) {
1200         SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
1201 
1202         // XXX - This is kind of hacky. Create a dummy sub node so we can check
1203         // the known bits in isDSOffsetLegal. We need to emit the selected node
1204         // here, so this is thrown away.
1205         SDValue Sub = CurDAG->getNode(ISD::SUB, DL, MVT::i32,
1206                                       Zero, Addr.getOperand(1));
1207 
1208         if (isDSOffsetLegal(Sub, ByteOffset, 16)) {
1209           SmallVector<SDValue, 3> Opnds;
1210           Opnds.push_back(Zero);
1211           Opnds.push_back(Addr.getOperand(1));
1212 
1213           // FIXME: Select to VOP3 version for with-carry.
1214           unsigned SubOp = AMDGPU::V_SUB_I32_e32;
1215           if (Subtarget->hasAddNoCarry()) {
1216             SubOp = AMDGPU::V_SUB_U32_e64;
1217             Opnds.push_back(
1218                 CurDAG->getTargetConstant(0, {}, MVT::i1)); // clamp bit
1219           }
1220 
1221           MachineSDNode *MachineSub =
1222               CurDAG->getMachineNode(SubOp, DL, MVT::i32, Opnds);
1223 
1224           Base = SDValue(MachineSub, 0);
1225           Offset = CurDAG->getTargetConstant(ByteOffset, DL, MVT::i16);
1226           return true;
1227         }
1228       }
1229     }
1230   } else if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
1231     // If we have a constant address, prefer to put the constant into the
1232     // offset. This can save moves to load the constant address since multiple
1233     // operations can share the zero base address register, and enables merging
1234     // into read2 / write2 instructions.
1235 
1236     SDLoc DL(Addr);
1237 
1238     if (isUInt<16>(CAddr->getZExtValue())) {
1239       SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
1240       MachineSDNode *MovZero = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
1241                                  DL, MVT::i32, Zero);
1242       Base = SDValue(MovZero, 0);
1243       Offset = CurDAG->getTargetConstant(CAddr->getZExtValue(), DL, MVT::i16);
1244       return true;
1245     }
1246   }
1247 
1248   // default case
1249   Base = Addr;
1250   Offset = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i16);
1251   return true;
1252 }
1253 
1254 // TODO: If offset is too big, put low 16-bit into offset.
1255 bool AMDGPUDAGToDAGISel::SelectDS64Bit4ByteAligned(SDValue Addr, SDValue &Base,
1256                                                    SDValue &Offset0,
1257                                                    SDValue &Offset1) const {
1258   SDLoc DL(Addr);
1259 
1260   if (CurDAG->isBaseWithConstantOffset(Addr)) {
1261     SDValue N0 = Addr.getOperand(0);
1262     SDValue N1 = Addr.getOperand(1);
1263     ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
1264     unsigned DWordOffset0 = C1->getZExtValue() / 4;
1265     unsigned DWordOffset1 = DWordOffset0 + 1;
1266     // (add n0, c0)
1267     if (isDSOffsetLegal(N0, DWordOffset1, 8)) {
1268       Base = N0;
1269       Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8);
1270       Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8);
1271       return true;
1272     }
1273   } else if (Addr.getOpcode() == ISD::SUB) {
1274     // sub C, x -> add (sub 0, x), C
1275     if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr.getOperand(0))) {
1276       unsigned DWordOffset0 = C->getZExtValue() / 4;
1277       unsigned DWordOffset1 = DWordOffset0 + 1;
1278 
1279       if (isUInt<8>(DWordOffset0)) {
1280         SDLoc DL(Addr);
1281         SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
1282 
1283         // XXX - This is kind of hacky. Create a dummy sub node so we can check
1284         // the known bits in isDSOffsetLegal. We need to emit the selected node
1285         // here, so this is thrown away.
1286         SDValue Sub = CurDAG->getNode(ISD::SUB, DL, MVT::i32,
1287                                       Zero, Addr.getOperand(1));
1288 
1289         if (isDSOffsetLegal(Sub, DWordOffset1, 8)) {
1290           SmallVector<SDValue, 3> Opnds;
1291           Opnds.push_back(Zero);
1292           Opnds.push_back(Addr.getOperand(1));
1293           unsigned SubOp = AMDGPU::V_SUB_I32_e32;
1294           if (Subtarget->hasAddNoCarry()) {
1295             SubOp = AMDGPU::V_SUB_U32_e64;
1296             Opnds.push_back(
1297                 CurDAG->getTargetConstant(0, {}, MVT::i1)); // clamp bit
1298           }
1299 
1300           MachineSDNode *MachineSub
1301             = CurDAG->getMachineNode(SubOp, DL, MVT::i32, Opnds);
1302 
1303           Base = SDValue(MachineSub, 0);
1304           Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8);
1305           Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8);
1306           return true;
1307         }
1308       }
1309     }
1310   } else if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
1311     unsigned DWordOffset0 = CAddr->getZExtValue() / 4;
1312     unsigned DWordOffset1 = DWordOffset0 + 1;
1313     assert(4 * DWordOffset0 == CAddr->getZExtValue());
1314 
1315     if (isUInt<8>(DWordOffset0) && isUInt<8>(DWordOffset1)) {
1316       SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
1317       MachineSDNode *MovZero
1318         = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
1319                                  DL, MVT::i32, Zero);
1320       Base = SDValue(MovZero, 0);
1321       Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8);
1322       Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8);
1323       return true;
1324     }
1325   }
1326 
1327   // default case
1328 
1329   Base = Addr;
1330   Offset0 = CurDAG->getTargetConstant(0, DL, MVT::i8);
1331   Offset1 = CurDAG->getTargetConstant(1, DL, MVT::i8);
1332   return true;
1333 }
1334 
1335 bool AMDGPUDAGToDAGISel::SelectMUBUF(SDValue Addr, SDValue &Ptr,
1336                                      SDValue &VAddr, SDValue &SOffset,
1337                                      SDValue &Offset, SDValue &Offen,
1338                                      SDValue &Idxen, SDValue &Addr64,
1339                                      SDValue &GLC, SDValue &SLC,
1340                                      SDValue &TFE, SDValue &DLC,
1341                                      SDValue &SWZ) const {
1342   // Subtarget prefers to use flat instruction
1343   // FIXME: This should be a pattern predicate and not reach here
1344   if (Subtarget->useFlatForGlobal())
1345     return false;
1346 
1347   SDLoc DL(Addr);
1348 
1349   if (!GLC.getNode())
1350     GLC = CurDAG->getTargetConstant(0, DL, MVT::i1);
1351   if (!SLC.getNode())
1352     SLC = CurDAG->getTargetConstant(0, DL, MVT::i1);
1353   TFE = CurDAG->getTargetConstant(0, DL, MVT::i1);
1354   DLC = CurDAG->getTargetConstant(0, DL, MVT::i1);
1355   SWZ = CurDAG->getTargetConstant(0, DL, MVT::i1);
1356 
1357   Idxen = CurDAG->getTargetConstant(0, DL, MVT::i1);
1358   Offen = CurDAG->getTargetConstant(0, DL, MVT::i1);
1359   Addr64 = CurDAG->getTargetConstant(0, DL, MVT::i1);
1360   SOffset = CurDAG->getTargetConstant(0, DL, MVT::i32);
1361 
1362   ConstantSDNode *C1 = nullptr;
1363   SDValue N0 = Addr;
1364   if (CurDAG->isBaseWithConstantOffset(Addr)) {
1365     C1 = cast<ConstantSDNode>(Addr.getOperand(1));
1366     if (isUInt<32>(C1->getZExtValue()))
1367       N0 = Addr.getOperand(0);
1368     else
1369       C1 = nullptr;
1370   }
1371 
1372   if (N0.getOpcode() == ISD::ADD) {
1373     // (add N2, N3) -> addr64, or
1374     // (add (add N2, N3), C1) -> addr64
1375     SDValue N2 = N0.getOperand(0);
1376     SDValue N3 = N0.getOperand(1);
1377     Addr64 = CurDAG->getTargetConstant(1, DL, MVT::i1);
1378 
1379     if (N2->isDivergent()) {
1380       if (N3->isDivergent()) {
1381         // Both N2 and N3 are divergent. Use N0 (the result of the add) as the
1382         // addr64, and construct the resource from a 0 address.
1383         Ptr = SDValue(buildSMovImm64(DL, 0, MVT::v2i32), 0);
1384         VAddr = N0;
1385       } else {
1386         // N2 is divergent, N3 is not.
1387         Ptr = N3;
1388         VAddr = N2;
1389       }
1390     } else {
1391       // N2 is not divergent.
1392       Ptr = N2;
1393       VAddr = N3;
1394     }
1395     Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1396   } else if (N0->isDivergent()) {
1397     // N0 is divergent. Use it as the addr64, and construct the resource from a
1398     // 0 address.
1399     Ptr = SDValue(buildSMovImm64(DL, 0, MVT::v2i32), 0);
1400     VAddr = N0;
1401     Addr64 = CurDAG->getTargetConstant(1, DL, MVT::i1);
1402   } else {
1403     // N0 -> offset, or
1404     // (N0 + C1) -> offset
1405     VAddr = CurDAG->getTargetConstant(0, DL, MVT::i32);
1406     Ptr = N0;
1407   }
1408 
1409   if (!C1) {
1410     // No offset.
1411     Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1412     return true;
1413   }
1414 
1415   if (SIInstrInfo::isLegalMUBUFImmOffset(C1->getZExtValue())) {
1416     // Legal offset for instruction.
1417     Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16);
1418     return true;
1419   }
1420 
1421   // Illegal offset, store it in soffset.
1422   Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1423   SOffset =
1424       SDValue(CurDAG->getMachineNode(
1425                   AMDGPU::S_MOV_B32, DL, MVT::i32,
1426                   CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i32)),
1427               0);
1428   return true;
1429 }
1430 
1431 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
1432                                            SDValue &VAddr, SDValue &SOffset,
1433                                            SDValue &Offset, SDValue &GLC,
1434                                            SDValue &SLC, SDValue &TFE,
1435                                            SDValue &DLC, SDValue &SWZ) const {
1436   SDValue Ptr, Offen, Idxen, Addr64;
1437 
1438   // addr64 bit was removed for volcanic islands.
1439   // FIXME: This should be a pattern predicate and not reach here
1440   if (!Subtarget->hasAddr64())
1441     return false;
1442 
1443   if (!SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
1444               GLC, SLC, TFE, DLC, SWZ))
1445     return false;
1446 
1447   ConstantSDNode *C = cast<ConstantSDNode>(Addr64);
1448   if (C->getSExtValue()) {
1449     SDLoc DL(Addr);
1450 
1451     const SITargetLowering& Lowering =
1452       *static_cast<const SITargetLowering*>(getTargetLowering());
1453 
1454     SRsrc = SDValue(Lowering.wrapAddr64Rsrc(*CurDAG, DL, Ptr), 0);
1455     return true;
1456   }
1457 
1458   return false;
1459 }
1460 
1461 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
1462                                            SDValue &VAddr, SDValue &SOffset,
1463                                            SDValue &Offset,
1464                                            SDValue &SLC) const {
1465   SLC = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i1);
1466   SDValue GLC, TFE, DLC, SWZ;
1467 
1468   return SelectMUBUFAddr64(Addr, SRsrc, VAddr, SOffset, Offset, GLC, SLC, TFE, DLC, SWZ);
1469 }
1470 
1471 static bool isStackPtrRelative(const MachinePointerInfo &PtrInfo) {
1472   auto PSV = PtrInfo.V.dyn_cast<const PseudoSourceValue *>();
1473   return PSV && PSV->isStack();
1474 }
1475 
1476 std::pair<SDValue, SDValue> AMDGPUDAGToDAGISel::foldFrameIndex(SDValue N) const {
1477   const MachineFunction &MF = CurDAG->getMachineFunction();
1478   const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1479 
1480   if (auto FI = dyn_cast<FrameIndexSDNode>(N)) {
1481     SDValue TFI = CurDAG->getTargetFrameIndex(FI->getIndex(),
1482                                               FI->getValueType(0));
1483 
1484     // If we can resolve this to a frame index access, this will be relative to
1485     // either the stack or frame pointer SGPR.
1486     return std::make_pair(
1487         TFI, CurDAG->getRegister(Info->getStackPtrOffsetReg(), MVT::i32));
1488   }
1489 
1490   // If we don't know this private access is a local stack object, it needs to
1491   // be relative to the entry point's scratch wave offset register.
1492   return std::make_pair(N, CurDAG->getRegister(Info->getScratchWaveOffsetReg(),
1493                                                MVT::i32));
1494 }
1495 
1496 bool AMDGPUDAGToDAGISel::SelectMUBUFScratchOffen(SDNode *Parent,
1497                                                  SDValue Addr, SDValue &Rsrc,
1498                                                  SDValue &VAddr, SDValue &SOffset,
1499                                                  SDValue &ImmOffset) const {
1500 
1501   SDLoc DL(Addr);
1502   MachineFunction &MF = CurDAG->getMachineFunction();
1503   const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1504 
1505   Rsrc = CurDAG->getRegister(Info->getScratchRSrcReg(), MVT::v4i32);
1506 
1507   if (ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
1508     unsigned Imm = CAddr->getZExtValue();
1509 
1510     SDValue HighBits = CurDAG->getTargetConstant(Imm & ~4095, DL, MVT::i32);
1511     MachineSDNode *MovHighBits = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
1512                                                         DL, MVT::i32, HighBits);
1513     VAddr = SDValue(MovHighBits, 0);
1514 
1515     // In a call sequence, stores to the argument stack area are relative to the
1516     // stack pointer.
1517     const MachinePointerInfo &PtrInfo = cast<MemSDNode>(Parent)->getPointerInfo();
1518     unsigned SOffsetReg = isStackPtrRelative(PtrInfo) ?
1519       Info->getStackPtrOffsetReg() : Info->getScratchWaveOffsetReg();
1520 
1521     SOffset = CurDAG->getRegister(SOffsetReg, MVT::i32);
1522     ImmOffset = CurDAG->getTargetConstant(Imm & 4095, DL, MVT::i16);
1523     return true;
1524   }
1525 
1526   if (CurDAG->isBaseWithConstantOffset(Addr)) {
1527     // (add n0, c1)
1528 
1529     SDValue N0 = Addr.getOperand(0);
1530     SDValue N1 = Addr.getOperand(1);
1531 
1532     // Offsets in vaddr must be positive if range checking is enabled.
1533     //
1534     // The total computation of vaddr + soffset + offset must not overflow.  If
1535     // vaddr is negative, even if offset is 0 the sgpr offset add will end up
1536     // overflowing.
1537     //
1538     // Prior to gfx9, MUBUF instructions with the vaddr offset enabled would
1539     // always perform a range check. If a negative vaddr base index was used,
1540     // this would fail the range check. The overall address computation would
1541     // compute a valid address, but this doesn't happen due to the range
1542     // check. For out-of-bounds MUBUF loads, a 0 is returned.
1543     //
1544     // Therefore it should be safe to fold any VGPR offset on gfx9 into the
1545     // MUBUF vaddr, but not on older subtargets which can only do this if the
1546     // sign bit is known 0.
1547     ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
1548     if (SIInstrInfo::isLegalMUBUFImmOffset(C1->getZExtValue()) &&
1549         (!Subtarget->privateMemoryResourceIsRangeChecked() ||
1550          CurDAG->SignBitIsZero(N0))) {
1551       std::tie(VAddr, SOffset) = foldFrameIndex(N0);
1552       ImmOffset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16);
1553       return true;
1554     }
1555   }
1556 
1557   // (node)
1558   std::tie(VAddr, SOffset) = foldFrameIndex(Addr);
1559   ImmOffset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1560   return true;
1561 }
1562 
1563 bool AMDGPUDAGToDAGISel::SelectMUBUFScratchOffset(SDNode *Parent,
1564                                                   SDValue Addr,
1565                                                   SDValue &SRsrc,
1566                                                   SDValue &SOffset,
1567                                                   SDValue &Offset) const {
1568   ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr);
1569   if (!CAddr || !SIInstrInfo::isLegalMUBUFImmOffset(CAddr->getZExtValue()))
1570     return false;
1571 
1572   SDLoc DL(Addr);
1573   MachineFunction &MF = CurDAG->getMachineFunction();
1574   const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1575 
1576   SRsrc = CurDAG->getRegister(Info->getScratchRSrcReg(), MVT::v4i32);
1577 
1578   const MachinePointerInfo &PtrInfo = cast<MemSDNode>(Parent)->getPointerInfo();
1579   unsigned SOffsetReg = isStackPtrRelative(PtrInfo) ?
1580     Info->getStackPtrOffsetReg() : Info->getScratchWaveOffsetReg();
1581 
1582   // FIXME: Get from MachinePointerInfo? We should only be using the frame
1583   // offset if we know this is in a call sequence.
1584   SOffset = CurDAG->getRegister(SOffsetReg, MVT::i32);
1585 
1586   Offset = CurDAG->getTargetConstant(CAddr->getZExtValue(), DL, MVT::i16);
1587   return true;
1588 }
1589 
1590 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1591                                            SDValue &SOffset, SDValue &Offset,
1592                                            SDValue &GLC, SDValue &SLC,
1593                                            SDValue &TFE, SDValue &DLC,
1594                                            SDValue &SWZ) const {
1595   SDValue Ptr, VAddr, Offen, Idxen, Addr64;
1596   const SIInstrInfo *TII =
1597     static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
1598 
1599   if (!SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
1600               GLC, SLC, TFE, DLC, SWZ))
1601     return false;
1602 
1603   if (!cast<ConstantSDNode>(Offen)->getSExtValue() &&
1604       !cast<ConstantSDNode>(Idxen)->getSExtValue() &&
1605       !cast<ConstantSDNode>(Addr64)->getSExtValue()) {
1606     uint64_t Rsrc = TII->getDefaultRsrcDataFormat() |
1607                     APInt::getAllOnesValue(32).getZExtValue(); // Size
1608     SDLoc DL(Addr);
1609 
1610     const SITargetLowering& Lowering =
1611       *static_cast<const SITargetLowering*>(getTargetLowering());
1612 
1613     SRsrc = SDValue(Lowering.buildRSRC(*CurDAG, DL, Ptr, 0, Rsrc), 0);
1614     return true;
1615   }
1616   return false;
1617 }
1618 
1619 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1620                                            SDValue &Soffset, SDValue &Offset
1621                                            ) const {
1622   SDValue GLC, SLC, TFE, DLC, SWZ;
1623 
1624   return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE, DLC, SWZ);
1625 }
1626 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1627                                            SDValue &Soffset, SDValue &Offset,
1628                                            SDValue &SLC) const {
1629   SDValue GLC, TFE, DLC, SWZ;
1630 
1631   return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE, DLC, SWZ);
1632 }
1633 
1634 // Find a load or store from corresponding pattern root.
1635 // Roots may be build_vector, bitconvert or their combinations.
1636 static MemSDNode* findMemSDNode(SDNode *N) {
1637   N = AMDGPUTargetLowering::stripBitcast(SDValue(N,0)).getNode();
1638   if (MemSDNode *MN = dyn_cast<MemSDNode>(N))
1639     return MN;
1640   assert(isa<BuildVectorSDNode>(N));
1641   for (SDValue V : N->op_values())
1642     if (MemSDNode *MN =
1643           dyn_cast<MemSDNode>(AMDGPUTargetLowering::stripBitcast(V)))
1644       return MN;
1645   llvm_unreachable("cannot find MemSDNode in the pattern!");
1646 }
1647 
1648 template <bool IsSigned>
1649 bool AMDGPUDAGToDAGISel::SelectFlatOffset(SDNode *N,
1650                                           SDValue Addr,
1651                                           SDValue &VAddr,
1652                                           SDValue &Offset,
1653                                           SDValue &SLC) const {
1654   int64_t OffsetVal = 0;
1655 
1656   if (Subtarget->hasFlatInstOffsets() &&
1657       (!Subtarget->hasFlatSegmentOffsetBug() ||
1658        findMemSDNode(N)->getAddressSpace() != AMDGPUAS::FLAT_ADDRESS) &&
1659       CurDAG->isBaseWithConstantOffset(Addr)) {
1660     SDValue N0 = Addr.getOperand(0);
1661     SDValue N1 = Addr.getOperand(1);
1662     uint64_t COffsetVal = cast<ConstantSDNode>(N1)->getSExtValue();
1663 
1664     const SIInstrInfo *TII = Subtarget->getInstrInfo();
1665     unsigned AS = findMemSDNode(N)->getAddressSpace();
1666     if (TII->isLegalFLATOffset(COffsetVal, AS, IsSigned)) {
1667       Addr = N0;
1668       OffsetVal = COffsetVal;
1669     } else {
1670       // If the offset doesn't fit, put the low bits into the offset field and
1671       // add the rest.
1672 
1673       SDLoc DL(N);
1674       uint64_t ImmField;
1675       const unsigned NumBits = TII->getNumFlatOffsetBits(AS, IsSigned);
1676       if (IsSigned) {
1677         ImmField = SignExtend64(COffsetVal, NumBits);
1678 
1679         // Don't use a negative offset field if the base offset is positive.
1680         // Since the scheduler currently relies on the offset field, doing so
1681         // could result in strange scheduling decisions.
1682 
1683         // TODO: Should we not do this in the opposite direction as well?
1684         if (static_cast<int64_t>(COffsetVal) > 0) {
1685           if (static_cast<int64_t>(ImmField) < 0) {
1686             const uint64_t OffsetMask = maskTrailingOnes<uint64_t>(NumBits - 1);
1687             ImmField = COffsetVal & OffsetMask;
1688           }
1689         }
1690       } else {
1691         // TODO: Should we do this for a negative offset?
1692         const uint64_t OffsetMask = maskTrailingOnes<uint64_t>(NumBits);
1693         ImmField = COffsetVal & OffsetMask;
1694       }
1695 
1696       uint64_t RemainderOffset = COffsetVal - ImmField;
1697 
1698       assert(TII->isLegalFLATOffset(ImmField, AS, IsSigned));
1699       assert(RemainderOffset + ImmField == COffsetVal);
1700 
1701       OffsetVal = ImmField;
1702 
1703       // TODO: Should this try to use a scalar add pseudo if the base address is
1704       // uniform and saddr is usable?
1705       SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
1706       SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
1707 
1708       SDNode *N0Lo = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
1709                                             DL, MVT::i32, N0, Sub0);
1710       SDNode *N0Hi = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
1711                                             DL, MVT::i32, N0, Sub1);
1712 
1713       SDValue AddOffsetLo
1714         = getMaterializedScalarImm32(Lo_32(RemainderOffset), DL);
1715       SDValue AddOffsetHi
1716         = getMaterializedScalarImm32(Hi_32(RemainderOffset), DL);
1717 
1718       SDVTList VTs = CurDAG->getVTList(MVT::i32, MVT::i1);
1719       SDValue Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1);
1720 
1721       SDNode *Add = CurDAG->getMachineNode(
1722         AMDGPU::V_ADD_I32_e64, DL, VTs,
1723         {AddOffsetLo, SDValue(N0Lo, 0), Clamp});
1724 
1725       SDNode *Addc = CurDAG->getMachineNode(
1726         AMDGPU::V_ADDC_U32_e64, DL, VTs,
1727         {AddOffsetHi, SDValue(N0Hi, 0), SDValue(Add, 1), Clamp});
1728 
1729       SDValue RegSequenceArgs[] = {
1730         CurDAG->getTargetConstant(AMDGPU::VReg_64RegClassID, DL, MVT::i32),
1731         SDValue(Add, 0), Sub0, SDValue(Addc, 0), Sub1
1732       };
1733 
1734       Addr = SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL,
1735                                             MVT::i64, RegSequenceArgs), 0);
1736     }
1737   }
1738 
1739   VAddr = Addr;
1740   Offset = CurDAG->getTargetConstant(OffsetVal, SDLoc(), MVT::i16);
1741   SLC = CurDAG->getTargetConstant(0, SDLoc(), MVT::i1);
1742   return true;
1743 }
1744 
1745 bool AMDGPUDAGToDAGISel::SelectFlatAtomic(SDNode *N,
1746                                           SDValue Addr,
1747                                           SDValue &VAddr,
1748                                           SDValue &Offset,
1749                                           SDValue &SLC) const {
1750   return SelectFlatOffset<false>(N, Addr, VAddr, Offset, SLC);
1751 }
1752 
1753 bool AMDGPUDAGToDAGISel::SelectFlatAtomicSigned(SDNode *N,
1754                                                 SDValue Addr,
1755                                                 SDValue &VAddr,
1756                                                 SDValue &Offset,
1757                                                 SDValue &SLC) const {
1758   return SelectFlatOffset<true>(N, Addr, VAddr, Offset, SLC);
1759 }
1760 
1761 bool AMDGPUDAGToDAGISel::SelectSMRDOffset(SDValue ByteOffsetNode,
1762                                           SDValue &Offset, bool &Imm) const {
1763 
1764   // FIXME: Handle non-constant offsets.
1765   ConstantSDNode *C = dyn_cast<ConstantSDNode>(ByteOffsetNode);
1766   if (!C)
1767     return false;
1768 
1769   SDLoc SL(ByteOffsetNode);
1770   GCNSubtarget::Generation Gen = Subtarget->getGeneration();
1771   uint64_t ByteOffset = C->getZExtValue();
1772   Optional<int64_t> EncodedOffset =
1773       AMDGPU::getSMRDEncodedOffset(*Subtarget, ByteOffset);
1774   if (EncodedOffset) {
1775     Offset = CurDAG->getTargetConstant(*EncodedOffset, SL, MVT::i32);
1776     Imm = true;
1777     return true;
1778   }
1779 
1780   if (Gen == AMDGPUSubtarget::SEA_ISLANDS) {
1781     EncodedOffset =
1782         AMDGPU::getSMRDEncodedLiteralOffset32(*Subtarget, ByteOffset);
1783     if (EncodedOffset) {
1784       Offset = CurDAG->getTargetConstant(*EncodedOffset, SL, MVT::i32);
1785       return true;
1786     }
1787   }
1788 
1789   if (!isUInt<32>(ByteOffset) && !isInt<32>(ByteOffset))
1790     return false;
1791 
1792   SDValue C32Bit = CurDAG->getTargetConstant(ByteOffset, SL, MVT::i32);
1793   Offset = SDValue(
1794       CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SL, MVT::i32, C32Bit), 0);
1795 
1796   return true;
1797 }
1798 
1799 SDValue AMDGPUDAGToDAGISel::Expand32BitAddress(SDValue Addr) const {
1800   if (Addr.getValueType() != MVT::i32)
1801     return Addr;
1802 
1803   // Zero-extend a 32-bit address.
1804   SDLoc SL(Addr);
1805 
1806   const MachineFunction &MF = CurDAG->getMachineFunction();
1807   const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1808   unsigned AddrHiVal = Info->get32BitAddressHighBits();
1809   SDValue AddrHi = CurDAG->getTargetConstant(AddrHiVal, SL, MVT::i32);
1810 
1811   const SDValue Ops[] = {
1812     CurDAG->getTargetConstant(AMDGPU::SReg_64_XEXECRegClassID, SL, MVT::i32),
1813     Addr,
1814     CurDAG->getTargetConstant(AMDGPU::sub0, SL, MVT::i32),
1815     SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SL, MVT::i32, AddrHi),
1816             0),
1817     CurDAG->getTargetConstant(AMDGPU::sub1, SL, MVT::i32),
1818   };
1819 
1820   return SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, SL, MVT::i64,
1821                                         Ops), 0);
1822 }
1823 
1824 bool AMDGPUDAGToDAGISel::SelectSMRD(SDValue Addr, SDValue &SBase,
1825                                      SDValue &Offset, bool &Imm) const {
1826   SDLoc SL(Addr);
1827 
1828   // A 32-bit (address + offset) should not cause unsigned 32-bit integer
1829   // wraparound, because s_load instructions perform the addition in 64 bits.
1830   if ((Addr.getValueType() != MVT::i32 ||
1831        Addr->getFlags().hasNoUnsignedWrap()) &&
1832       CurDAG->isBaseWithConstantOffset(Addr)) {
1833     SDValue N0 = Addr.getOperand(0);
1834     SDValue N1 = Addr.getOperand(1);
1835 
1836     if (SelectSMRDOffset(N1, Offset, Imm)) {
1837       SBase = Expand32BitAddress(N0);
1838       return true;
1839     }
1840   }
1841   SBase = Expand32BitAddress(Addr);
1842   Offset = CurDAG->getTargetConstant(0, SL, MVT::i32);
1843   Imm = true;
1844   return true;
1845 }
1846 
1847 bool AMDGPUDAGToDAGISel::SelectSMRDImm(SDValue Addr, SDValue &SBase,
1848                                        SDValue &Offset) const {
1849   bool Imm = false;
1850   return SelectSMRD(Addr, SBase, Offset, Imm) && Imm;
1851 }
1852 
1853 bool AMDGPUDAGToDAGISel::SelectSMRDImm32(SDValue Addr, SDValue &SBase,
1854                                          SDValue &Offset) const {
1855 
1856   assert(Subtarget->getGeneration() == AMDGPUSubtarget::SEA_ISLANDS);
1857 
1858   bool Imm = false;
1859   if (!SelectSMRD(Addr, SBase, Offset, Imm))
1860     return false;
1861 
1862   return !Imm && isa<ConstantSDNode>(Offset);
1863 }
1864 
1865 bool AMDGPUDAGToDAGISel::SelectSMRDSgpr(SDValue Addr, SDValue &SBase,
1866                                         SDValue &Offset) const {
1867   bool Imm = false;
1868   return SelectSMRD(Addr, SBase, Offset, Imm) && !Imm &&
1869          !isa<ConstantSDNode>(Offset);
1870 }
1871 
1872 bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm(SDValue Addr,
1873                                              SDValue &Offset) const {
1874   if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr)) {
1875     if (auto Imm = AMDGPU::getSMRDEncodedOffset(*Subtarget,
1876                                                 C->getZExtValue())) {
1877       Offset = CurDAG->getTargetConstant(*Imm, SDLoc(Addr), MVT::i32);
1878       return true;
1879     }
1880   }
1881 
1882   return false;
1883 }
1884 
1885 bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm32(SDValue Addr,
1886                                                SDValue &Offset) const {
1887   assert(Subtarget->getGeneration() == AMDGPUSubtarget::SEA_ISLANDS);
1888 
1889   if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr)) {
1890     if (auto Imm = AMDGPU::getSMRDEncodedLiteralOffset32(*Subtarget,
1891                                                          C->getZExtValue())) {
1892       Offset = CurDAG->getTargetConstant(*Imm, SDLoc(Addr), MVT::i32);
1893       return true;
1894     }
1895   }
1896 
1897   return false;
1898 }
1899 
1900 bool AMDGPUDAGToDAGISel::SelectMOVRELOffset(SDValue Index,
1901                                             SDValue &Base,
1902                                             SDValue &Offset) const {
1903   SDLoc DL(Index);
1904 
1905   if (CurDAG->isBaseWithConstantOffset(Index)) {
1906     SDValue N0 = Index.getOperand(0);
1907     SDValue N1 = Index.getOperand(1);
1908     ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
1909 
1910     // (add n0, c0)
1911     // Don't peel off the offset (c0) if doing so could possibly lead
1912     // the base (n0) to be negative.
1913     if (C1->getSExtValue() <= 0 || CurDAG->SignBitIsZero(N0)) {
1914       Base = N0;
1915       Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i32);
1916       return true;
1917     }
1918   }
1919 
1920   if (isa<ConstantSDNode>(Index))
1921     return false;
1922 
1923   Base = Index;
1924   Offset = CurDAG->getTargetConstant(0, DL, MVT::i32);
1925   return true;
1926 }
1927 
1928 SDNode *AMDGPUDAGToDAGISel::getS_BFE(unsigned Opcode, const SDLoc &DL,
1929                                      SDValue Val, uint32_t Offset,
1930                                      uint32_t Width) {
1931   // Transformation function, pack the offset and width of a BFE into
1932   // the format expected by the S_BFE_I32 / S_BFE_U32. In the second
1933   // source, bits [5:0] contain the offset and bits [22:16] the width.
1934   uint32_t PackedVal = Offset | (Width << 16);
1935   SDValue PackedConst = CurDAG->getTargetConstant(PackedVal, DL, MVT::i32);
1936 
1937   return CurDAG->getMachineNode(Opcode, DL, MVT::i32, Val, PackedConst);
1938 }
1939 
1940 void AMDGPUDAGToDAGISel::SelectS_BFEFromShifts(SDNode *N) {
1941   // "(a << b) srl c)" ---> "BFE_U32 a, (c-b), (32-c)
1942   // "(a << b) sra c)" ---> "BFE_I32 a, (c-b), (32-c)
1943   // Predicate: 0 < b <= c < 32
1944 
1945   const SDValue &Shl = N->getOperand(0);
1946   ConstantSDNode *B = dyn_cast<ConstantSDNode>(Shl->getOperand(1));
1947   ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1948 
1949   if (B && C) {
1950     uint32_t BVal = B->getZExtValue();
1951     uint32_t CVal = C->getZExtValue();
1952 
1953     if (0 < BVal && BVal <= CVal && CVal < 32) {
1954       bool Signed = N->getOpcode() == ISD::SRA;
1955       unsigned Opcode = Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32;
1956 
1957       ReplaceNode(N, getS_BFE(Opcode, SDLoc(N), Shl.getOperand(0), CVal - BVal,
1958                               32 - CVal));
1959       return;
1960     }
1961   }
1962   SelectCode(N);
1963 }
1964 
1965 void AMDGPUDAGToDAGISel::SelectS_BFE(SDNode *N) {
1966   switch (N->getOpcode()) {
1967   case ISD::AND:
1968     if (N->getOperand(0).getOpcode() == ISD::SRL) {
1969       // "(a srl b) & mask" ---> "BFE_U32 a, b, popcount(mask)"
1970       // Predicate: isMask(mask)
1971       const SDValue &Srl = N->getOperand(0);
1972       ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(Srl.getOperand(1));
1973       ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(N->getOperand(1));
1974 
1975       if (Shift && Mask) {
1976         uint32_t ShiftVal = Shift->getZExtValue();
1977         uint32_t MaskVal = Mask->getZExtValue();
1978 
1979         if (isMask_32(MaskVal)) {
1980           uint32_t WidthVal = countPopulation(MaskVal);
1981 
1982           ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_U32, SDLoc(N),
1983                                   Srl.getOperand(0), ShiftVal, WidthVal));
1984           return;
1985         }
1986       }
1987     }
1988     break;
1989   case ISD::SRL:
1990     if (N->getOperand(0).getOpcode() == ISD::AND) {
1991       // "(a & mask) srl b)" ---> "BFE_U32 a, b, popcount(mask >> b)"
1992       // Predicate: isMask(mask >> b)
1993       const SDValue &And = N->getOperand(0);
1994       ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(N->getOperand(1));
1995       ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(And->getOperand(1));
1996 
1997       if (Shift && Mask) {
1998         uint32_t ShiftVal = Shift->getZExtValue();
1999         uint32_t MaskVal = Mask->getZExtValue() >> ShiftVal;
2000 
2001         if (isMask_32(MaskVal)) {
2002           uint32_t WidthVal = countPopulation(MaskVal);
2003 
2004           ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_U32, SDLoc(N),
2005                                   And.getOperand(0), ShiftVal, WidthVal));
2006           return;
2007         }
2008       }
2009     } else if (N->getOperand(0).getOpcode() == ISD::SHL) {
2010       SelectS_BFEFromShifts(N);
2011       return;
2012     }
2013     break;
2014   case ISD::SRA:
2015     if (N->getOperand(0).getOpcode() == ISD::SHL) {
2016       SelectS_BFEFromShifts(N);
2017       return;
2018     }
2019     break;
2020 
2021   case ISD::SIGN_EXTEND_INREG: {
2022     // sext_inreg (srl x, 16), i8 -> bfe_i32 x, 16, 8
2023     SDValue Src = N->getOperand(0);
2024     if (Src.getOpcode() != ISD::SRL)
2025       break;
2026 
2027     const ConstantSDNode *Amt = dyn_cast<ConstantSDNode>(Src.getOperand(1));
2028     if (!Amt)
2029       break;
2030 
2031     unsigned Width = cast<VTSDNode>(N->getOperand(1))->getVT().getSizeInBits();
2032     ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_I32, SDLoc(N), Src.getOperand(0),
2033                             Amt->getZExtValue(), Width));
2034     return;
2035   }
2036   }
2037 
2038   SelectCode(N);
2039 }
2040 
2041 bool AMDGPUDAGToDAGISel::isCBranchSCC(const SDNode *N) const {
2042   assert(N->getOpcode() == ISD::BRCOND);
2043   if (!N->hasOneUse())
2044     return false;
2045 
2046   SDValue Cond = N->getOperand(1);
2047   if (Cond.getOpcode() == ISD::CopyToReg)
2048     Cond = Cond.getOperand(2);
2049 
2050   if (Cond.getOpcode() != ISD::SETCC || !Cond.hasOneUse())
2051     return false;
2052 
2053   MVT VT = Cond.getOperand(0).getSimpleValueType();
2054   if (VT == MVT::i32)
2055     return true;
2056 
2057   if (VT == MVT::i64) {
2058     auto ST = static_cast<const GCNSubtarget *>(Subtarget);
2059 
2060     ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
2061     return (CC == ISD::SETEQ || CC == ISD::SETNE) && ST->hasScalarCompareEq64();
2062   }
2063 
2064   return false;
2065 }
2066 
2067 void AMDGPUDAGToDAGISel::SelectBRCOND(SDNode *N) {
2068   SDValue Cond = N->getOperand(1);
2069 
2070   if (Cond.isUndef()) {
2071     CurDAG->SelectNodeTo(N, AMDGPU::SI_BR_UNDEF, MVT::Other,
2072                          N->getOperand(2), N->getOperand(0));
2073     return;
2074   }
2075 
2076   const GCNSubtarget *ST = static_cast<const GCNSubtarget *>(Subtarget);
2077   const SIRegisterInfo *TRI = ST->getRegisterInfo();
2078 
2079   bool UseSCCBr = isCBranchSCC(N) && isUniformBr(N);
2080   unsigned BrOp = UseSCCBr ? AMDGPU::S_CBRANCH_SCC1 : AMDGPU::S_CBRANCH_VCCNZ;
2081   unsigned CondReg = UseSCCBr ? (unsigned)AMDGPU::SCC : TRI->getVCC();
2082   SDLoc SL(N);
2083 
2084   if (!UseSCCBr) {
2085     // This is the case that we are selecting to S_CBRANCH_VCCNZ.  We have not
2086     // analyzed what generates the vcc value, so we do not know whether vcc
2087     // bits for disabled lanes are 0.  Thus we need to mask out bits for
2088     // disabled lanes.
2089     //
2090     // For the case that we select S_CBRANCH_SCC1 and it gets
2091     // changed to S_CBRANCH_VCCNZ in SIFixSGPRCopies, SIFixSGPRCopies calls
2092     // SIInstrInfo::moveToVALU which inserts the S_AND).
2093     //
2094     // We could add an analysis of what generates the vcc value here and omit
2095     // the S_AND when is unnecessary. But it would be better to add a separate
2096     // pass after SIFixSGPRCopies to do the unnecessary S_AND removal, so it
2097     // catches both cases.
2098     Cond = SDValue(CurDAG->getMachineNode(ST->isWave32() ? AMDGPU::S_AND_B32
2099                                                          : AMDGPU::S_AND_B64,
2100                      SL, MVT::i1,
2101                      CurDAG->getRegister(ST->isWave32() ? AMDGPU::EXEC_LO
2102                                                         : AMDGPU::EXEC,
2103                                          MVT::i1),
2104                     Cond),
2105                    0);
2106   }
2107 
2108   SDValue VCC = CurDAG->getCopyToReg(N->getOperand(0), SL, CondReg, Cond);
2109   CurDAG->SelectNodeTo(N, BrOp, MVT::Other,
2110                        N->getOperand(2), // Basic Block
2111                        VCC.getValue(0));
2112 }
2113 
2114 void AMDGPUDAGToDAGISel::SelectFMAD_FMA(SDNode *N) {
2115   MVT VT = N->getSimpleValueType(0);
2116   bool IsFMA = N->getOpcode() == ISD::FMA;
2117   if (VT != MVT::f32 || (!Subtarget->hasMadMixInsts() &&
2118                          !Subtarget->hasFmaMixInsts()) ||
2119       ((IsFMA && Subtarget->hasMadMixInsts()) ||
2120        (!IsFMA && Subtarget->hasFmaMixInsts()))) {
2121     SelectCode(N);
2122     return;
2123   }
2124 
2125   SDValue Src0 = N->getOperand(0);
2126   SDValue Src1 = N->getOperand(1);
2127   SDValue Src2 = N->getOperand(2);
2128   unsigned Src0Mods, Src1Mods, Src2Mods;
2129 
2130   // Avoid using v_mad_mix_f32/v_fma_mix_f32 unless there is actually an operand
2131   // using the conversion from f16.
2132   bool Sel0 = SelectVOP3PMadMixModsImpl(Src0, Src0, Src0Mods);
2133   bool Sel1 = SelectVOP3PMadMixModsImpl(Src1, Src1, Src1Mods);
2134   bool Sel2 = SelectVOP3PMadMixModsImpl(Src2, Src2, Src2Mods);
2135 
2136   assert((IsFMA || !Mode.allFP32Denormals()) &&
2137          "fmad selected with denormals enabled");
2138   // TODO: We can select this with f32 denormals enabled if all the sources are
2139   // converted from f16 (in which case fmad isn't legal).
2140 
2141   if (Sel0 || Sel1 || Sel2) {
2142     // For dummy operands.
2143     SDValue Zero = CurDAG->getTargetConstant(0, SDLoc(), MVT::i32);
2144     SDValue Ops[] = {
2145       CurDAG->getTargetConstant(Src0Mods, SDLoc(), MVT::i32), Src0,
2146       CurDAG->getTargetConstant(Src1Mods, SDLoc(), MVT::i32), Src1,
2147       CurDAG->getTargetConstant(Src2Mods, SDLoc(), MVT::i32), Src2,
2148       CurDAG->getTargetConstant(0, SDLoc(), MVT::i1),
2149       Zero, Zero
2150     };
2151 
2152     CurDAG->SelectNodeTo(N,
2153                          IsFMA ? AMDGPU::V_FMA_MIX_F32 : AMDGPU::V_MAD_MIX_F32,
2154                          MVT::f32, Ops);
2155   } else {
2156     SelectCode(N);
2157   }
2158 }
2159 
2160 // This is here because there isn't a way to use the generated sub0_sub1 as the
2161 // subreg index to EXTRACT_SUBREG in tablegen.
2162 void AMDGPUDAGToDAGISel::SelectATOMIC_CMP_SWAP(SDNode *N) {
2163   MemSDNode *Mem = cast<MemSDNode>(N);
2164   unsigned AS = Mem->getAddressSpace();
2165   if (AS == AMDGPUAS::FLAT_ADDRESS) {
2166     SelectCode(N);
2167     return;
2168   }
2169 
2170   MVT VT = N->getSimpleValueType(0);
2171   bool Is32 = (VT == MVT::i32);
2172   SDLoc SL(N);
2173 
2174   MachineSDNode *CmpSwap = nullptr;
2175   if (Subtarget->hasAddr64()) {
2176     SDValue SRsrc, VAddr, SOffset, Offset, SLC;
2177 
2178     if (SelectMUBUFAddr64(Mem->getBasePtr(), SRsrc, VAddr, SOffset, Offset, SLC)) {
2179       unsigned Opcode = Is32 ? AMDGPU::BUFFER_ATOMIC_CMPSWAP_ADDR64_RTN :
2180         AMDGPU::BUFFER_ATOMIC_CMPSWAP_X2_ADDR64_RTN;
2181       SDValue CmpVal = Mem->getOperand(2);
2182 
2183       // XXX - Do we care about glue operands?
2184 
2185       SDValue Ops[] = {
2186         CmpVal, VAddr, SRsrc, SOffset, Offset, SLC, Mem->getChain()
2187       };
2188 
2189       CmpSwap = CurDAG->getMachineNode(Opcode, SL, Mem->getVTList(), Ops);
2190     }
2191   }
2192 
2193   if (!CmpSwap) {
2194     SDValue SRsrc, SOffset, Offset, SLC;
2195     if (SelectMUBUFOffset(Mem->getBasePtr(), SRsrc, SOffset, Offset, SLC)) {
2196       unsigned Opcode = Is32 ? AMDGPU::BUFFER_ATOMIC_CMPSWAP_OFFSET_RTN :
2197         AMDGPU::BUFFER_ATOMIC_CMPSWAP_X2_OFFSET_RTN;
2198 
2199       SDValue CmpVal = Mem->getOperand(2);
2200       SDValue Ops[] = {
2201         CmpVal, SRsrc, SOffset, Offset, SLC, Mem->getChain()
2202       };
2203 
2204       CmpSwap = CurDAG->getMachineNode(Opcode, SL, Mem->getVTList(), Ops);
2205     }
2206   }
2207 
2208   if (!CmpSwap) {
2209     SelectCode(N);
2210     return;
2211   }
2212 
2213   MachineMemOperand *MMO = Mem->getMemOperand();
2214   CurDAG->setNodeMemRefs(CmpSwap, {MMO});
2215 
2216   unsigned SubReg = Is32 ? AMDGPU::sub0 : AMDGPU::sub0_sub1;
2217   SDValue Extract
2218     = CurDAG->getTargetExtractSubreg(SubReg, SL, VT, SDValue(CmpSwap, 0));
2219 
2220   ReplaceUses(SDValue(N, 0), Extract);
2221   ReplaceUses(SDValue(N, 1), SDValue(CmpSwap, 1));
2222   CurDAG->RemoveDeadNode(N);
2223 }
2224 
2225 void AMDGPUDAGToDAGISel::SelectDSAppendConsume(SDNode *N, unsigned IntrID) {
2226   // The address is assumed to be uniform, so if it ends up in a VGPR, it will
2227   // be copied to an SGPR with readfirstlane.
2228   unsigned Opc = IntrID == Intrinsic::amdgcn_ds_append ?
2229     AMDGPU::DS_APPEND : AMDGPU::DS_CONSUME;
2230 
2231   SDValue Chain = N->getOperand(0);
2232   SDValue Ptr = N->getOperand(2);
2233   MemIntrinsicSDNode *M = cast<MemIntrinsicSDNode>(N);
2234   MachineMemOperand *MMO = M->getMemOperand();
2235   bool IsGDS = M->getAddressSpace() == AMDGPUAS::REGION_ADDRESS;
2236 
2237   SDValue Offset;
2238   if (CurDAG->isBaseWithConstantOffset(Ptr)) {
2239     SDValue PtrBase = Ptr.getOperand(0);
2240     SDValue PtrOffset = Ptr.getOperand(1);
2241 
2242     const APInt &OffsetVal = cast<ConstantSDNode>(PtrOffset)->getAPIntValue();
2243     if (isDSOffsetLegal(PtrBase, OffsetVal.getZExtValue(), 16)) {
2244       N = glueCopyToM0(N, PtrBase);
2245       Offset = CurDAG->getTargetConstant(OffsetVal, SDLoc(), MVT::i32);
2246     }
2247   }
2248 
2249   if (!Offset) {
2250     N = glueCopyToM0(N, Ptr);
2251     Offset = CurDAG->getTargetConstant(0, SDLoc(), MVT::i32);
2252   }
2253 
2254   SDValue Ops[] = {
2255     Offset,
2256     CurDAG->getTargetConstant(IsGDS, SDLoc(), MVT::i32),
2257     Chain,
2258     N->getOperand(N->getNumOperands() - 1) // New glue
2259   };
2260 
2261   SDNode *Selected = CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
2262   CurDAG->setNodeMemRefs(cast<MachineSDNode>(Selected), {MMO});
2263 }
2264 
2265 static unsigned gwsIntrinToOpcode(unsigned IntrID) {
2266   switch (IntrID) {
2267   case Intrinsic::amdgcn_ds_gws_init:
2268     return AMDGPU::DS_GWS_INIT;
2269   case Intrinsic::amdgcn_ds_gws_barrier:
2270     return AMDGPU::DS_GWS_BARRIER;
2271   case Intrinsic::amdgcn_ds_gws_sema_v:
2272     return AMDGPU::DS_GWS_SEMA_V;
2273   case Intrinsic::amdgcn_ds_gws_sema_br:
2274     return AMDGPU::DS_GWS_SEMA_BR;
2275   case Intrinsic::amdgcn_ds_gws_sema_p:
2276     return AMDGPU::DS_GWS_SEMA_P;
2277   case Intrinsic::amdgcn_ds_gws_sema_release_all:
2278     return AMDGPU::DS_GWS_SEMA_RELEASE_ALL;
2279   default:
2280     llvm_unreachable("not a gws intrinsic");
2281   }
2282 }
2283 
2284 void AMDGPUDAGToDAGISel::SelectDS_GWS(SDNode *N, unsigned IntrID) {
2285   if (IntrID == Intrinsic::amdgcn_ds_gws_sema_release_all &&
2286       !Subtarget->hasGWSSemaReleaseAll()) {
2287     // Let this error.
2288     SelectCode(N);
2289     return;
2290   }
2291 
2292   // Chain, intrinsic ID, vsrc, offset
2293   const bool HasVSrc = N->getNumOperands() == 4;
2294   assert(HasVSrc || N->getNumOperands() == 3);
2295 
2296   SDLoc SL(N);
2297   SDValue BaseOffset = N->getOperand(HasVSrc ? 3 : 2);
2298   int ImmOffset = 0;
2299   MemIntrinsicSDNode *M = cast<MemIntrinsicSDNode>(N);
2300   MachineMemOperand *MMO = M->getMemOperand();
2301 
2302   // Don't worry if the offset ends up in a VGPR. Only one lane will have
2303   // effect, so SIFixSGPRCopies will validly insert readfirstlane.
2304 
2305   // The resource id offset is computed as (<isa opaque base> + M0[21:16] +
2306   // offset field) % 64. Some versions of the programming guide omit the m0
2307   // part, or claim it's from offset 0.
2308   if (ConstantSDNode *ConstOffset = dyn_cast<ConstantSDNode>(BaseOffset)) {
2309     // If we have a constant offset, try to use the 0 in m0 as the base.
2310     // TODO: Look into changing the default m0 initialization value. If the
2311     // default -1 only set the low 16-bits, we could leave it as-is and add 1 to
2312     // the immediate offset.
2313     glueCopyToM0(N, CurDAG->getTargetConstant(0, SL, MVT::i32));
2314     ImmOffset = ConstOffset->getZExtValue();
2315   } else {
2316     if (CurDAG->isBaseWithConstantOffset(BaseOffset)) {
2317       ImmOffset = BaseOffset.getConstantOperandVal(1);
2318       BaseOffset = BaseOffset.getOperand(0);
2319     }
2320 
2321     // Prefer to do the shift in an SGPR since it should be possible to use m0
2322     // as the result directly. If it's already an SGPR, it will be eliminated
2323     // later.
2324     SDNode *SGPROffset
2325       = CurDAG->getMachineNode(AMDGPU::V_READFIRSTLANE_B32, SL, MVT::i32,
2326                                BaseOffset);
2327     // Shift to offset in m0
2328     SDNode *M0Base
2329       = CurDAG->getMachineNode(AMDGPU::S_LSHL_B32, SL, MVT::i32,
2330                                SDValue(SGPROffset, 0),
2331                                CurDAG->getTargetConstant(16, SL, MVT::i32));
2332     glueCopyToM0(N, SDValue(M0Base, 0));
2333   }
2334 
2335   SDValue Chain = N->getOperand(0);
2336   SDValue OffsetField = CurDAG->getTargetConstant(ImmOffset, SL, MVT::i32);
2337 
2338   // TODO: Can this just be removed from the instruction?
2339   SDValue GDS = CurDAG->getTargetConstant(1, SL, MVT::i1);
2340 
2341   const unsigned Opc = gwsIntrinToOpcode(IntrID);
2342   SmallVector<SDValue, 5> Ops;
2343   if (HasVSrc)
2344     Ops.push_back(N->getOperand(2));
2345   Ops.push_back(OffsetField);
2346   Ops.push_back(GDS);
2347   Ops.push_back(Chain);
2348 
2349   SDNode *Selected = CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
2350   CurDAG->setNodeMemRefs(cast<MachineSDNode>(Selected), {MMO});
2351 }
2352 
2353 void AMDGPUDAGToDAGISel::SelectInterpP1F16(SDNode *N) {
2354   if (Subtarget->getLDSBankCount() != 16) {
2355     // This is a single instruction with a pattern.
2356     SelectCode(N);
2357     return;
2358   }
2359 
2360   SDLoc DL(N);
2361 
2362   // This requires 2 instructions. It is possible to write a pattern to support
2363   // this, but the generated isel emitter doesn't correctly deal with multiple
2364   // output instructions using the same physical register input. The copy to m0
2365   // is incorrectly placed before the second instruction.
2366   //
2367   // TODO: Match source modifiers.
2368   //
2369   // def : Pat <
2370   //   (int_amdgcn_interp_p1_f16
2371   //    (VOP3Mods f32:$src0, i32:$src0_modifiers),
2372   //                             (i32 timm:$attrchan), (i32 timm:$attr),
2373   //                             (i1 timm:$high), M0),
2374   //   (V_INTERP_P1LV_F16 $src0_modifiers, VGPR_32:$src0, timm:$attr,
2375   //       timm:$attrchan, 0,
2376   //       (V_INTERP_MOV_F32 2, timm:$attr, timm:$attrchan), timm:$high)> {
2377   //   let Predicates = [has16BankLDS];
2378   // }
2379 
2380   // 16 bank LDS
2381   SDValue ToM0 = CurDAG->getCopyToReg(CurDAG->getEntryNode(), DL, AMDGPU::M0,
2382                                       N->getOperand(5), SDValue());
2383 
2384   SDVTList VTs = CurDAG->getVTList(MVT::f32, MVT::Other);
2385 
2386   SDNode *InterpMov =
2387     CurDAG->getMachineNode(AMDGPU::V_INTERP_MOV_F32, DL, VTs, {
2388         CurDAG->getTargetConstant(2, DL, MVT::i32), // P0
2389         N->getOperand(3),  // Attr
2390         N->getOperand(2),  // Attrchan
2391         ToM0.getValue(1) // In glue
2392   });
2393 
2394   SDNode *InterpP1LV =
2395     CurDAG->getMachineNode(AMDGPU::V_INTERP_P1LV_F16, DL, MVT::f32, {
2396         CurDAG->getTargetConstant(0, DL, MVT::i32), // $src0_modifiers
2397         N->getOperand(1), // Src0
2398         N->getOperand(3), // Attr
2399         N->getOperand(2), // Attrchan
2400         CurDAG->getTargetConstant(0, DL, MVT::i32), // $src2_modifiers
2401         SDValue(InterpMov, 0), // Src2 - holds two f16 values selected by high
2402         N->getOperand(4), // high
2403         CurDAG->getTargetConstant(0, DL, MVT::i1), // $clamp
2404         CurDAG->getTargetConstant(0, DL, MVT::i32), // $omod
2405         SDValue(InterpMov, 1)
2406   });
2407 
2408   CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), SDValue(InterpP1LV, 0));
2409 }
2410 
2411 void AMDGPUDAGToDAGISel::SelectINTRINSIC_W_CHAIN(SDNode *N) {
2412   unsigned IntrID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2413   switch (IntrID) {
2414   case Intrinsic::amdgcn_ds_append:
2415   case Intrinsic::amdgcn_ds_consume: {
2416     if (N->getValueType(0) != MVT::i32)
2417       break;
2418     SelectDSAppendConsume(N, IntrID);
2419     return;
2420   }
2421   }
2422 
2423   SelectCode(N);
2424 }
2425 
2426 void AMDGPUDAGToDAGISel::SelectINTRINSIC_WO_CHAIN(SDNode *N) {
2427   unsigned IntrID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2428   unsigned Opcode;
2429   switch (IntrID) {
2430   case Intrinsic::amdgcn_wqm:
2431     Opcode = AMDGPU::WQM;
2432     break;
2433   case Intrinsic::amdgcn_softwqm:
2434     Opcode = AMDGPU::SOFT_WQM;
2435     break;
2436   case Intrinsic::amdgcn_wwm:
2437     Opcode = AMDGPU::WWM;
2438     break;
2439   case Intrinsic::amdgcn_interp_p1_f16:
2440     SelectInterpP1F16(N);
2441     return;
2442   default:
2443     SelectCode(N);
2444     return;
2445   }
2446 
2447   SDValue Src = N->getOperand(1);
2448   CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), {Src});
2449 }
2450 
2451 void AMDGPUDAGToDAGISel::SelectINTRINSIC_VOID(SDNode *N) {
2452   unsigned IntrID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2453   switch (IntrID) {
2454   case Intrinsic::amdgcn_ds_gws_init:
2455   case Intrinsic::amdgcn_ds_gws_barrier:
2456   case Intrinsic::amdgcn_ds_gws_sema_v:
2457   case Intrinsic::amdgcn_ds_gws_sema_br:
2458   case Intrinsic::amdgcn_ds_gws_sema_p:
2459   case Intrinsic::amdgcn_ds_gws_sema_release_all:
2460     SelectDS_GWS(N, IntrID);
2461     return;
2462   default:
2463     break;
2464   }
2465 
2466   SelectCode(N);
2467 }
2468 
2469 bool AMDGPUDAGToDAGISel::SelectVOP3ModsImpl(SDValue In, SDValue &Src,
2470                                             unsigned &Mods) const {
2471   Mods = 0;
2472   Src = In;
2473 
2474   if (Src.getOpcode() == ISD::FNEG) {
2475     Mods |= SISrcMods::NEG;
2476     Src = Src.getOperand(0);
2477   }
2478 
2479   if (Src.getOpcode() == ISD::FABS) {
2480     Mods |= SISrcMods::ABS;
2481     Src = Src.getOperand(0);
2482   }
2483 
2484   return true;
2485 }
2486 
2487 bool AMDGPUDAGToDAGISel::SelectVOP3Mods(SDValue In, SDValue &Src,
2488                                         SDValue &SrcMods) const {
2489   unsigned Mods;
2490   if (SelectVOP3ModsImpl(In, Src, Mods)) {
2491     SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
2492     return true;
2493   }
2494 
2495   return false;
2496 }
2497 
2498 bool AMDGPUDAGToDAGISel::SelectVOP3Mods_NNaN(SDValue In, SDValue &Src,
2499                                              SDValue &SrcMods) const {
2500   SelectVOP3Mods(In, Src, SrcMods);
2501   return isNoNanSrc(Src);
2502 }
2503 
2504 bool AMDGPUDAGToDAGISel::SelectVOP3NoMods(SDValue In, SDValue &Src) const {
2505   if (In.getOpcode() == ISD::FABS || In.getOpcode() == ISD::FNEG)
2506     return false;
2507 
2508   Src = In;
2509   return true;
2510 }
2511 
2512 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0(SDValue In, SDValue &Src,
2513                                          SDValue &SrcMods, SDValue &Clamp,
2514                                          SDValue &Omod) const {
2515   SDLoc DL(In);
2516   Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1);
2517   Omod = CurDAG->getTargetConstant(0, DL, MVT::i1);
2518 
2519   return SelectVOP3Mods(In, Src, SrcMods);
2520 }
2521 
2522 bool AMDGPUDAGToDAGISel::SelectVOP3OMods(SDValue In, SDValue &Src,
2523                                          SDValue &Clamp, SDValue &Omod) const {
2524   Src = In;
2525 
2526   SDLoc DL(In);
2527   Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1);
2528   Omod = CurDAG->getTargetConstant(0, DL, MVT::i1);
2529 
2530   return true;
2531 }
2532 
2533 bool AMDGPUDAGToDAGISel::SelectVOP3PMods(SDValue In, SDValue &Src,
2534                                          SDValue &SrcMods) const {
2535   unsigned Mods = 0;
2536   Src = In;
2537 
2538   if (Src.getOpcode() == ISD::FNEG) {
2539     Mods ^= (SISrcMods::NEG | SISrcMods::NEG_HI);
2540     Src = Src.getOperand(0);
2541   }
2542 
2543   if (Src.getOpcode() == ISD::BUILD_VECTOR) {
2544     unsigned VecMods = Mods;
2545 
2546     SDValue Lo = stripBitcast(Src.getOperand(0));
2547     SDValue Hi = stripBitcast(Src.getOperand(1));
2548 
2549     if (Lo.getOpcode() == ISD::FNEG) {
2550       Lo = stripBitcast(Lo.getOperand(0));
2551       Mods ^= SISrcMods::NEG;
2552     }
2553 
2554     if (Hi.getOpcode() == ISD::FNEG) {
2555       Hi = stripBitcast(Hi.getOperand(0));
2556       Mods ^= SISrcMods::NEG_HI;
2557     }
2558 
2559     if (isExtractHiElt(Lo, Lo))
2560       Mods |= SISrcMods::OP_SEL_0;
2561 
2562     if (isExtractHiElt(Hi, Hi))
2563       Mods |= SISrcMods::OP_SEL_1;
2564 
2565     Lo = stripExtractLoElt(Lo);
2566     Hi = stripExtractLoElt(Hi);
2567 
2568     if (Lo == Hi && !isInlineImmediate(Lo.getNode())) {
2569       // Really a scalar input. Just select from the low half of the register to
2570       // avoid packing.
2571 
2572       Src = Lo;
2573       SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
2574       return true;
2575     }
2576 
2577     Mods = VecMods;
2578   }
2579 
2580   // Packed instructions do not have abs modifiers.
2581   Mods |= SISrcMods::OP_SEL_1;
2582 
2583   SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
2584   return true;
2585 }
2586 
2587 bool AMDGPUDAGToDAGISel::SelectVOP3OpSel(SDValue In, SDValue &Src,
2588                                          SDValue &SrcMods) const {
2589   Src = In;
2590   // FIXME: Handle op_sel
2591   SrcMods = CurDAG->getTargetConstant(0, SDLoc(In), MVT::i32);
2592   return true;
2593 }
2594 
2595 bool AMDGPUDAGToDAGISel::SelectVOP3OpSelMods(SDValue In, SDValue &Src,
2596                                              SDValue &SrcMods) const {
2597   // FIXME: Handle op_sel
2598   return SelectVOP3Mods(In, Src, SrcMods);
2599 }
2600 
2601 // The return value is not whether the match is possible (which it always is),
2602 // but whether or not it a conversion is really used.
2603 bool AMDGPUDAGToDAGISel::SelectVOP3PMadMixModsImpl(SDValue In, SDValue &Src,
2604                                                    unsigned &Mods) const {
2605   Mods = 0;
2606   SelectVOP3ModsImpl(In, Src, Mods);
2607 
2608   if (Src.getOpcode() == ISD::FP_EXTEND) {
2609     Src = Src.getOperand(0);
2610     assert(Src.getValueType() == MVT::f16);
2611     Src = stripBitcast(Src);
2612 
2613     // Be careful about folding modifiers if we already have an abs. fneg is
2614     // applied last, so we don't want to apply an earlier fneg.
2615     if ((Mods & SISrcMods::ABS) == 0) {
2616       unsigned ModsTmp;
2617       SelectVOP3ModsImpl(Src, Src, ModsTmp);
2618 
2619       if ((ModsTmp & SISrcMods::NEG) != 0)
2620         Mods ^= SISrcMods::NEG;
2621 
2622       if ((ModsTmp & SISrcMods::ABS) != 0)
2623         Mods |= SISrcMods::ABS;
2624     }
2625 
2626     // op_sel/op_sel_hi decide the source type and source.
2627     // If the source's op_sel_hi is set, it indicates to do a conversion from fp16.
2628     // If the sources's op_sel is set, it picks the high half of the source
2629     // register.
2630 
2631     Mods |= SISrcMods::OP_SEL_1;
2632     if (isExtractHiElt(Src, Src)) {
2633       Mods |= SISrcMods::OP_SEL_0;
2634 
2635       // TODO: Should we try to look for neg/abs here?
2636     }
2637 
2638     return true;
2639   }
2640 
2641   return false;
2642 }
2643 
2644 bool AMDGPUDAGToDAGISel::SelectVOP3PMadMixMods(SDValue In, SDValue &Src,
2645                                                SDValue &SrcMods) const {
2646   unsigned Mods = 0;
2647   SelectVOP3PMadMixModsImpl(In, Src, Mods);
2648   SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
2649   return true;
2650 }
2651 
2652 SDValue AMDGPUDAGToDAGISel::getHi16Elt(SDValue In) const {
2653   if (In.isUndef())
2654     return CurDAG->getUNDEF(MVT::i32);
2655 
2656   if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(In)) {
2657     SDLoc SL(In);
2658     return CurDAG->getConstant(C->getZExtValue() << 16, SL, MVT::i32);
2659   }
2660 
2661   if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(In)) {
2662     SDLoc SL(In);
2663     return CurDAG->getConstant(
2664       C->getValueAPF().bitcastToAPInt().getZExtValue() << 16, SL, MVT::i32);
2665   }
2666 
2667   SDValue Src;
2668   if (isExtractHiElt(In, Src))
2669     return Src;
2670 
2671   return SDValue();
2672 }
2673 
2674 bool AMDGPUDAGToDAGISel::isVGPRImm(const SDNode * N) const {
2675   assert(CurDAG->getTarget().getTargetTriple().getArch() == Triple::amdgcn);
2676 
2677   const SIRegisterInfo *SIRI =
2678     static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());
2679   const SIInstrInfo * SII =
2680     static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
2681 
2682   unsigned Limit = 0;
2683   bool AllUsesAcceptSReg = true;
2684   for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
2685     Limit < 10 && U != E; ++U, ++Limit) {
2686     const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo());
2687 
2688     // If the register class is unknown, it could be an unknown
2689     // register class that needs to be an SGPR, e.g. an inline asm
2690     // constraint
2691     if (!RC || SIRI->isSGPRClass(RC))
2692       return false;
2693 
2694     if (RC != &AMDGPU::VS_32RegClass) {
2695       AllUsesAcceptSReg = false;
2696       SDNode * User = *U;
2697       if (User->isMachineOpcode()) {
2698         unsigned Opc = User->getMachineOpcode();
2699         MCInstrDesc Desc = SII->get(Opc);
2700         if (Desc.isCommutable()) {
2701           unsigned OpIdx = Desc.getNumDefs() + U.getOperandNo();
2702           unsigned CommuteIdx1 = TargetInstrInfo::CommuteAnyOperandIndex;
2703           if (SII->findCommutedOpIndices(Desc, OpIdx, CommuteIdx1)) {
2704             unsigned CommutedOpNo = CommuteIdx1 - Desc.getNumDefs();
2705             const TargetRegisterClass *CommutedRC = getOperandRegClass(*U, CommutedOpNo);
2706             if (CommutedRC == &AMDGPU::VS_32RegClass)
2707               AllUsesAcceptSReg = true;
2708           }
2709         }
2710       }
2711       // If "AllUsesAcceptSReg == false" so far we haven't suceeded
2712       // commuting current user. This means have at least one use
2713       // that strictly require VGPR. Thus, we will not attempt to commute
2714       // other user instructions.
2715       if (!AllUsesAcceptSReg)
2716         break;
2717     }
2718   }
2719   return !AllUsesAcceptSReg && (Limit < 10);
2720 }
2721 
2722 bool AMDGPUDAGToDAGISel::isUniformLoad(const SDNode * N) const {
2723   auto Ld = cast<LoadSDNode>(N);
2724 
2725   return Ld->getAlignment() >= 4 &&
2726         (
2727           (
2728             (
2729               Ld->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS       ||
2730               Ld->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT
2731             )
2732             &&
2733             !N->isDivergent()
2734           )
2735           ||
2736           (
2737             Subtarget->getScalarizeGlobalBehavior() &&
2738             Ld->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
2739             !Ld->isVolatile() &&
2740             !N->isDivergent() &&
2741             static_cast<const SITargetLowering *>(
2742               getTargetLowering())->isMemOpHasNoClobberedMemOperand(N)
2743           )
2744         );
2745 }
2746 
2747 void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
2748   const AMDGPUTargetLowering& Lowering =
2749     *static_cast<const AMDGPUTargetLowering*>(getTargetLowering());
2750   bool IsModified = false;
2751   do {
2752     IsModified = false;
2753 
2754     // Go over all selected nodes and try to fold them a bit more
2755     SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_begin();
2756     while (Position != CurDAG->allnodes_end()) {
2757       SDNode *Node = &*Position++;
2758       MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(Node);
2759       if (!MachineNode)
2760         continue;
2761 
2762       SDNode *ResNode = Lowering.PostISelFolding(MachineNode, *CurDAG);
2763       if (ResNode != Node) {
2764         if (ResNode)
2765           ReplaceUses(Node, ResNode);
2766         IsModified = true;
2767       }
2768     }
2769     CurDAG->RemoveDeadNodes();
2770   } while (IsModified);
2771 }
2772 
2773 bool R600DAGToDAGISel::runOnMachineFunction(MachineFunction &MF) {
2774   Subtarget = &MF.getSubtarget<R600Subtarget>();
2775   return SelectionDAGISel::runOnMachineFunction(MF);
2776 }
2777 
2778 bool R600DAGToDAGISel::isConstantLoad(const MemSDNode *N, int CbId) const {
2779   if (!N->readMem())
2780     return false;
2781   if (CbId == -1)
2782     return N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS ||
2783            N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT;
2784 
2785   return N->getAddressSpace() == AMDGPUAS::CONSTANT_BUFFER_0 + CbId;
2786 }
2787 
2788 bool R600DAGToDAGISel::SelectGlobalValueConstantOffset(SDValue Addr,
2789                                                          SDValue& IntPtr) {
2790   if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Addr)) {
2791     IntPtr = CurDAG->getIntPtrConstant(Cst->getZExtValue() / 4, SDLoc(Addr),
2792                                        true);
2793     return true;
2794   }
2795   return false;
2796 }
2797 
2798 bool R600DAGToDAGISel::SelectGlobalValueVariableOffset(SDValue Addr,
2799     SDValue& BaseReg, SDValue &Offset) {
2800   if (!isa<ConstantSDNode>(Addr)) {
2801     BaseReg = Addr;
2802     Offset = CurDAG->getIntPtrConstant(0, SDLoc(Addr), true);
2803     return true;
2804   }
2805   return false;
2806 }
2807 
2808 void R600DAGToDAGISel::Select(SDNode *N) {
2809   unsigned int Opc = N->getOpcode();
2810   if (N->isMachineOpcode()) {
2811     N->setNodeId(-1);
2812     return;   // Already selected.
2813   }
2814 
2815   switch (Opc) {
2816   default: break;
2817   case AMDGPUISD::BUILD_VERTICAL_VECTOR:
2818   case ISD::SCALAR_TO_VECTOR:
2819   case ISD::BUILD_VECTOR: {
2820     EVT VT = N->getValueType(0);
2821     unsigned NumVectorElts = VT.getVectorNumElements();
2822     unsigned RegClassID;
2823     // BUILD_VECTOR was lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
2824     // that adds a 128 bits reg copy when going through TwoAddressInstructions
2825     // pass. We want to avoid 128 bits copies as much as possible because they
2826     // can't be bundled by our scheduler.
2827     switch(NumVectorElts) {
2828     case 2: RegClassID = R600::R600_Reg64RegClassID; break;
2829     case 4:
2830       if (Opc == AMDGPUISD::BUILD_VERTICAL_VECTOR)
2831         RegClassID = R600::R600_Reg128VerticalRegClassID;
2832       else
2833         RegClassID = R600::R600_Reg128RegClassID;
2834       break;
2835     default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
2836     }
2837     SelectBuildVector(N, RegClassID);
2838     return;
2839   }
2840   }
2841 
2842   SelectCode(N);
2843 }
2844 
2845 bool R600DAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base,
2846                                           SDValue &Offset) {
2847   ConstantSDNode *C;
2848   SDLoc DL(Addr);
2849 
2850   if ((C = dyn_cast<ConstantSDNode>(Addr))) {
2851     Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32);
2852     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
2853   } else if ((Addr.getOpcode() == AMDGPUISD::DWORDADDR) &&
2854              (C = dyn_cast<ConstantSDNode>(Addr.getOperand(0)))) {
2855     Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32);
2856     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
2857   } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
2858             (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) {
2859     Base = Addr.getOperand(0);
2860     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
2861   } else {
2862     Base = Addr;
2863     Offset = CurDAG->getTargetConstant(0, DL, MVT::i32);
2864   }
2865 
2866   return true;
2867 }
2868 
2869 bool R600DAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
2870                                           SDValue &Offset) {
2871   ConstantSDNode *IMMOffset;
2872 
2873   if (Addr.getOpcode() == ISD::ADD
2874       && (IMMOffset = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))
2875       && isInt<16>(IMMOffset->getZExtValue())) {
2876 
2877       Base = Addr.getOperand(0);
2878       Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), SDLoc(Addr),
2879                                          MVT::i32);
2880       return true;
2881   // If the pointer address is constant, we can move it to the offset field.
2882   } else if ((IMMOffset = dyn_cast<ConstantSDNode>(Addr))
2883              && isInt<16>(IMMOffset->getZExtValue())) {
2884     Base = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
2885                                   SDLoc(CurDAG->getEntryNode()),
2886                                   R600::ZERO, MVT::i32);
2887     Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), SDLoc(Addr),
2888                                        MVT::i32);
2889     return true;
2890   }
2891 
2892   // Default case, no offset
2893   Base = Addr;
2894   Offset = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i32);
2895   return true;
2896 }
2897