xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp (revision a90b9d0159070121c221b966469c3e36d912bf82)
1 //===- SelectionDAGISel.cpp - Implement the SelectionDAGISel class --------===//
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 // This implements the SelectionDAGISel class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/CodeGen/SelectionDAGISel.h"
14 #include "ScheduleDAGSDNodes.h"
15 #include "SelectionDAGBuilder.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/PostOrderIterator.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/ADT/StringRef.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/Analysis/AssumptionCache.h"
26 #include "llvm/Analysis/BranchProbabilityInfo.h"
27 #include "llvm/Analysis/CFG.h"
28 #include "llvm/Analysis/LazyBlockFrequencyInfo.h"
29 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
30 #include "llvm/Analysis/ProfileSummaryInfo.h"
31 #include "llvm/Analysis/TargetLibraryInfo.h"
32 #include "llvm/Analysis/TargetTransformInfo.h"
33 #include "llvm/Analysis/UniformityAnalysis.h"
34 #include "llvm/CodeGen/AssignmentTrackingAnalysis.h"
35 #include "llvm/CodeGen/CodeGenCommonISel.h"
36 #include "llvm/CodeGen/FastISel.h"
37 #include "llvm/CodeGen/FunctionLoweringInfo.h"
38 #include "llvm/CodeGen/GCMetadata.h"
39 #include "llvm/CodeGen/ISDOpcodes.h"
40 #include "llvm/CodeGen/MachineBasicBlock.h"
41 #include "llvm/CodeGen/MachineFrameInfo.h"
42 #include "llvm/CodeGen/MachineFunction.h"
43 #include "llvm/CodeGen/MachineFunctionPass.h"
44 #include "llvm/CodeGen/MachineInstr.h"
45 #include "llvm/CodeGen/MachineInstrBuilder.h"
46 #include "llvm/CodeGen/MachineMemOperand.h"
47 #include "llvm/CodeGen/MachineModuleInfo.h"
48 #include "llvm/CodeGen/MachineOperand.h"
49 #include "llvm/CodeGen/MachinePassRegistry.h"
50 #include "llvm/CodeGen/MachineRegisterInfo.h"
51 #include "llvm/CodeGen/MachineValueType.h"
52 #include "llvm/CodeGen/SchedulerRegistry.h"
53 #include "llvm/CodeGen/SelectionDAG.h"
54 #include "llvm/CodeGen/SelectionDAGNodes.h"
55 #include "llvm/CodeGen/StackMaps.h"
56 #include "llvm/CodeGen/StackProtector.h"
57 #include "llvm/CodeGen/SwiftErrorValueTracking.h"
58 #include "llvm/CodeGen/TargetInstrInfo.h"
59 #include "llvm/CodeGen/TargetLowering.h"
60 #include "llvm/CodeGen/TargetRegisterInfo.h"
61 #include "llvm/CodeGen/TargetSubtargetInfo.h"
62 #include "llvm/CodeGen/ValueTypes.h"
63 #include "llvm/CodeGen/WinEHFuncInfo.h"
64 #include "llvm/IR/BasicBlock.h"
65 #include "llvm/IR/Constants.h"
66 #include "llvm/IR/DataLayout.h"
67 #include "llvm/IR/DebugInfo.h"
68 #include "llvm/IR/DebugInfoMetadata.h"
69 #include "llvm/IR/DebugLoc.h"
70 #include "llvm/IR/DiagnosticInfo.h"
71 #include "llvm/IR/EHPersonalities.h"
72 #include "llvm/IR/Function.h"
73 #include "llvm/IR/InlineAsm.h"
74 #include "llvm/IR/InstIterator.h"
75 #include "llvm/IR/Instruction.h"
76 #include "llvm/IR/Instructions.h"
77 #include "llvm/IR/IntrinsicInst.h"
78 #include "llvm/IR/Intrinsics.h"
79 #include "llvm/IR/IntrinsicsWebAssembly.h"
80 #include "llvm/IR/Metadata.h"
81 #include "llvm/IR/PrintPasses.h"
82 #include "llvm/IR/Statepoint.h"
83 #include "llvm/IR/Type.h"
84 #include "llvm/IR/User.h"
85 #include "llvm/IR/Value.h"
86 #include "llvm/InitializePasses.h"
87 #include "llvm/MC/MCInstrDesc.h"
88 #include "llvm/Pass.h"
89 #include "llvm/Support/BranchProbability.h"
90 #include "llvm/Support/Casting.h"
91 #include "llvm/Support/CodeGen.h"
92 #include "llvm/Support/CommandLine.h"
93 #include "llvm/Support/Compiler.h"
94 #include "llvm/Support/Debug.h"
95 #include "llvm/Support/ErrorHandling.h"
96 #include "llvm/Support/KnownBits.h"
97 #include "llvm/Support/Timer.h"
98 #include "llvm/Support/raw_ostream.h"
99 #include "llvm/Target/TargetIntrinsicInfo.h"
100 #include "llvm/Target/TargetMachine.h"
101 #include "llvm/Target/TargetOptions.h"
102 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
103 #include <algorithm>
104 #include <cassert>
105 #include <cstdint>
106 #include <iterator>
107 #include <limits>
108 #include <memory>
109 #include <optional>
110 #include <string>
111 #include <utility>
112 #include <vector>
113 
114 using namespace llvm;
115 
116 #define DEBUG_TYPE "isel"
117 #define ISEL_DUMP_DEBUG_TYPE DEBUG_TYPE "-dump"
118 
119 STATISTIC(NumFastIselFailures, "Number of instructions fast isel failed on");
120 STATISTIC(NumFastIselSuccess, "Number of instructions fast isel selected");
121 STATISTIC(NumFastIselBlocks, "Number of blocks selected entirely by fast isel");
122 STATISTIC(NumDAGBlocks, "Number of blocks selected using DAG");
123 STATISTIC(NumDAGIselRetries,"Number of times dag isel has to try another path");
124 STATISTIC(NumEntryBlocks, "Number of entry blocks encountered");
125 STATISTIC(NumFastIselFailLowerArguments,
126           "Number of entry blocks where fast isel failed to lower arguments");
127 
128 static cl::opt<int> EnableFastISelAbort(
129     "fast-isel-abort", cl::Hidden,
130     cl::desc("Enable abort calls when \"fast\" instruction selection "
131              "fails to lower an instruction: 0 disable the abort, 1 will "
132              "abort but for args, calls and terminators, 2 will also "
133              "abort for argument lowering, and 3 will never fallback "
134              "to SelectionDAG."));
135 
136 static cl::opt<bool> EnableFastISelFallbackReport(
137     "fast-isel-report-on-fallback", cl::Hidden,
138     cl::desc("Emit a diagnostic when \"fast\" instruction selection "
139              "falls back to SelectionDAG."));
140 
141 static cl::opt<bool>
142 UseMBPI("use-mbpi",
143         cl::desc("use Machine Branch Probability Info"),
144         cl::init(true), cl::Hidden);
145 
146 #ifndef NDEBUG
147 static cl::opt<std::string>
148 FilterDAGBasicBlockName("filter-view-dags", cl::Hidden,
149                         cl::desc("Only display the basic block whose name "
150                                  "matches this for all view-*-dags options"));
151 static cl::opt<bool>
152 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
153           cl::desc("Pop up a window to show dags before the first "
154                    "dag combine pass"));
155 static cl::opt<bool>
156 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
157           cl::desc("Pop up a window to show dags before legalize types"));
158 static cl::opt<bool>
159     ViewDAGCombineLT("view-dag-combine-lt-dags", cl::Hidden,
160                      cl::desc("Pop up a window to show dags before the post "
161                               "legalize types dag combine pass"));
162 static cl::opt<bool>
163     ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
164                      cl::desc("Pop up a window to show dags before legalize"));
165 static cl::opt<bool>
166 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
167           cl::desc("Pop up a window to show dags before the second "
168                    "dag combine pass"));
169 static cl::opt<bool>
170 ViewISelDAGs("view-isel-dags", cl::Hidden,
171           cl::desc("Pop up a window to show isel dags as they are selected"));
172 static cl::opt<bool>
173 ViewSchedDAGs("view-sched-dags", cl::Hidden,
174           cl::desc("Pop up a window to show sched dags as they are processed"));
175 static cl::opt<bool>
176 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
177       cl::desc("Pop up a window to show SUnit dags after they are processed"));
178 #else
179 static const bool ViewDAGCombine1 = false, ViewLegalizeTypesDAGs = false,
180                   ViewDAGCombineLT = false, ViewLegalizeDAGs = false,
181                   ViewDAGCombine2 = false, ViewISelDAGs = false,
182                   ViewSchedDAGs = false, ViewSUnitDAGs = false;
183 #endif
184 
185 #ifndef NDEBUG
186 #define ISEL_DUMP(X)                                                           \
187   do {                                                                         \
188     if (llvm::DebugFlag &&                                                     \
189         (isCurrentDebugType(DEBUG_TYPE) ||                                     \
190          (isCurrentDebugType(ISEL_DUMP_DEBUG_TYPE) && MatchFilterFuncName))) { \
191       X;                                                                       \
192     }                                                                          \
193   } while (false)
194 #else
195 #define ISEL_DUMP(X) do { } while (false)
196 #endif
197 
198 //===---------------------------------------------------------------------===//
199 ///
200 /// RegisterScheduler class - Track the registration of instruction schedulers.
201 ///
202 //===---------------------------------------------------------------------===//
203 MachinePassRegistry<RegisterScheduler::FunctionPassCtor>
204     RegisterScheduler::Registry;
205 
206 //===---------------------------------------------------------------------===//
207 ///
208 /// ISHeuristic command line option for instruction schedulers.
209 ///
210 //===---------------------------------------------------------------------===//
211 static cl::opt<RegisterScheduler::FunctionPassCtor, false,
212                RegisterPassParser<RegisterScheduler>>
213 ISHeuristic("pre-RA-sched",
214             cl::init(&createDefaultScheduler), cl::Hidden,
215             cl::desc("Instruction schedulers available (before register"
216                      " allocation):"));
217 
218 static RegisterScheduler
219 defaultListDAGScheduler("default", "Best scheduler for the target",
220                         createDefaultScheduler);
221 
222 static bool dontUseFastISelFor(const Function &Fn) {
223   // Don't enable FastISel for functions with swiftasync Arguments.
224   // Debug info on those is reliant on good Argument lowering, and FastISel is
225   // not capable of lowering the entire function. Mixing the two selectors tend
226   // to result in poor lowering of Arguments.
227   return any_of(Fn.args(), [](const Argument &Arg) {
228     return Arg.hasAttribute(Attribute::AttrKind::SwiftAsync);
229   });
230 }
231 
232 namespace llvm {
233 
234   //===--------------------------------------------------------------------===//
235   /// This class is used by SelectionDAGISel to temporarily override
236   /// the optimization level on a per-function basis.
237   class OptLevelChanger {
238     SelectionDAGISel &IS;
239     CodeGenOptLevel SavedOptLevel;
240     bool SavedFastISel;
241 
242   public:
243     OptLevelChanger(SelectionDAGISel &ISel, CodeGenOptLevel NewOptLevel)
244         : IS(ISel) {
245       SavedOptLevel = IS.OptLevel;
246       SavedFastISel = IS.TM.Options.EnableFastISel;
247       if (NewOptLevel != SavedOptLevel) {
248         IS.OptLevel = NewOptLevel;
249         IS.TM.setOptLevel(NewOptLevel);
250         LLVM_DEBUG(dbgs() << "\nChanging optimization level for Function "
251                           << IS.MF->getFunction().getName() << "\n");
252         LLVM_DEBUG(dbgs() << "\tBefore: -O" << static_cast<int>(SavedOptLevel)
253                           << " ; After: -O" << static_cast<int>(NewOptLevel)
254                           << "\n");
255         if (NewOptLevel == CodeGenOptLevel::None)
256           IS.TM.setFastISel(IS.TM.getO0WantsFastISel());
257       }
258       if (dontUseFastISelFor(IS.MF->getFunction()))
259         IS.TM.setFastISel(false);
260       LLVM_DEBUG(
261           dbgs() << "\tFastISel is "
262                  << (IS.TM.Options.EnableFastISel ? "enabled" : "disabled")
263                  << "\n");
264     }
265 
266     ~OptLevelChanger() {
267       if (IS.OptLevel == SavedOptLevel)
268         return;
269       LLVM_DEBUG(dbgs() << "\nRestoring optimization level for Function "
270                         << IS.MF->getFunction().getName() << "\n");
271       LLVM_DEBUG(dbgs() << "\tBefore: -O" << static_cast<int>(IS.OptLevel)
272                         << " ; After: -O" << static_cast<int>(SavedOptLevel) << "\n");
273       IS.OptLevel = SavedOptLevel;
274       IS.TM.setOptLevel(SavedOptLevel);
275       IS.TM.setFastISel(SavedFastISel);
276     }
277   };
278 
279   //===--------------------------------------------------------------------===//
280   /// createDefaultScheduler - This creates an instruction scheduler appropriate
281   /// for the target.
282   ScheduleDAGSDNodes *createDefaultScheduler(SelectionDAGISel *IS,
283                                              CodeGenOptLevel OptLevel) {
284     const TargetLowering *TLI = IS->TLI;
285     const TargetSubtargetInfo &ST = IS->MF->getSubtarget();
286 
287     // Try first to see if the Target has its own way of selecting a scheduler
288     if (auto *SchedulerCtor = ST.getDAGScheduler(OptLevel)) {
289       return SchedulerCtor(IS, OptLevel);
290     }
291 
292     if (OptLevel == CodeGenOptLevel::None ||
293         (ST.enableMachineScheduler() && ST.enableMachineSchedDefaultSched()) ||
294         TLI->getSchedulingPreference() == Sched::Source)
295       return createSourceListDAGScheduler(IS, OptLevel);
296     if (TLI->getSchedulingPreference() == Sched::RegPressure)
297       return createBURRListDAGScheduler(IS, OptLevel);
298     if (TLI->getSchedulingPreference() == Sched::Hybrid)
299       return createHybridListDAGScheduler(IS, OptLevel);
300     if (TLI->getSchedulingPreference() == Sched::VLIW)
301       return createVLIWDAGScheduler(IS, OptLevel);
302     if (TLI->getSchedulingPreference() == Sched::Fast)
303       return createFastDAGScheduler(IS, OptLevel);
304     if (TLI->getSchedulingPreference() == Sched::Linearize)
305       return createDAGLinearizer(IS, OptLevel);
306     assert(TLI->getSchedulingPreference() == Sched::ILP &&
307            "Unknown sched type!");
308     return createILPListDAGScheduler(IS, OptLevel);
309   }
310 
311 } // end namespace llvm
312 
313 // EmitInstrWithCustomInserter - This method should be implemented by targets
314 // that mark instructions with the 'usesCustomInserter' flag.  These
315 // instructions are special in various ways, which require special support to
316 // insert.  The specified MachineInstr is created but not inserted into any
317 // basic blocks, and this method is called to expand it into a sequence of
318 // instructions, potentially also creating new basic blocks and control flow.
319 // When new basic blocks are inserted and the edges from MBB to its successors
320 // are modified, the method should insert pairs of <OldSucc, NewSucc> into the
321 // DenseMap.
322 MachineBasicBlock *
323 TargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
324                                             MachineBasicBlock *MBB) const {
325 #ifndef NDEBUG
326   dbgs() << "If a target marks an instruction with "
327           "'usesCustomInserter', it must implement "
328           "TargetLowering::EmitInstrWithCustomInserter!\n";
329 #endif
330   llvm_unreachable(nullptr);
331 }
332 
333 void TargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI,
334                                                    SDNode *Node) const {
335   assert(!MI.hasPostISelHook() &&
336          "If a target marks an instruction with 'hasPostISelHook', "
337          "it must implement TargetLowering::AdjustInstrPostInstrSelection!");
338 }
339 
340 //===----------------------------------------------------------------------===//
341 // SelectionDAGISel code
342 //===----------------------------------------------------------------------===//
343 
344 SelectionDAGISel::SelectionDAGISel(char &ID, TargetMachine &tm,
345                                    CodeGenOptLevel OL)
346     : MachineFunctionPass(ID), TM(tm), FuncInfo(new FunctionLoweringInfo()),
347       SwiftError(new SwiftErrorValueTracking()),
348       CurDAG(new SelectionDAG(tm, OL)),
349       SDB(std::make_unique<SelectionDAGBuilder>(*CurDAG, *FuncInfo, *SwiftError,
350                                                 OL)),
351       OptLevel(OL) {
352   initializeGCModuleInfoPass(*PassRegistry::getPassRegistry());
353   initializeBranchProbabilityInfoWrapperPassPass(
354       *PassRegistry::getPassRegistry());
355   initializeAAResultsWrapperPassPass(*PassRegistry::getPassRegistry());
356   initializeTargetLibraryInfoWrapperPassPass(*PassRegistry::getPassRegistry());
357 }
358 
359 SelectionDAGISel::~SelectionDAGISel() {
360   delete CurDAG;
361   delete SwiftError;
362 }
363 
364 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
365   if (OptLevel != CodeGenOptLevel::None)
366       AU.addRequired<AAResultsWrapperPass>();
367   AU.addRequired<GCModuleInfo>();
368   AU.addRequired<StackProtector>();
369   AU.addPreserved<GCModuleInfo>();
370   AU.addRequired<TargetLibraryInfoWrapperPass>();
371   AU.addRequired<TargetTransformInfoWrapperPass>();
372   AU.addRequired<AssumptionCacheTracker>();
373   if (UseMBPI && OptLevel != CodeGenOptLevel::None)
374       AU.addRequired<BranchProbabilityInfoWrapperPass>();
375   AU.addRequired<ProfileSummaryInfoWrapperPass>();
376   // AssignmentTrackingAnalysis only runs if assignment tracking is enabled for
377   // the module.
378   AU.addRequired<AssignmentTrackingAnalysis>();
379   AU.addPreserved<AssignmentTrackingAnalysis>();
380   if (OptLevel != CodeGenOptLevel::None)
381       LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU);
382   MachineFunctionPass::getAnalysisUsage(AU);
383 }
384 
385 static void computeUsesMSVCFloatingPoint(const Triple &TT, const Function &F,
386                                          MachineModuleInfo &MMI) {
387   // Only needed for MSVC
388   if (!TT.isWindowsMSVCEnvironment())
389     return;
390 
391   // If it's already set, nothing to do.
392   if (MMI.usesMSVCFloatingPoint())
393     return;
394 
395   for (const Instruction &I : instructions(F)) {
396     if (I.getType()->isFPOrFPVectorTy()) {
397       MMI.setUsesMSVCFloatingPoint(true);
398       return;
399     }
400     for (const auto &Op : I.operands()) {
401       if (Op->getType()->isFPOrFPVectorTy()) {
402         MMI.setUsesMSVCFloatingPoint(true);
403         return;
404       }
405     }
406   }
407 }
408 
409 bool SelectionDAGISel::runOnMachineFunction(MachineFunction &mf) {
410   // If we already selected that function, we do not need to run SDISel.
411   if (mf.getProperties().hasProperty(
412           MachineFunctionProperties::Property::Selected))
413     return false;
414   // Do some sanity-checking on the command-line options.
415   assert((!EnableFastISelAbort || TM.Options.EnableFastISel) &&
416          "-fast-isel-abort > 0 requires -fast-isel");
417 
418   const Function &Fn = mf.getFunction();
419   MF = &mf;
420 
421 #ifndef NDEBUG
422   StringRef FuncName = Fn.getName();
423   MatchFilterFuncName = isFunctionInPrintList(FuncName);
424 #else
425   (void)MatchFilterFuncName;
426 #endif
427 
428   // Decide what flavour of variable location debug-info will be used, before
429   // we change the optimisation level.
430   bool InstrRef = mf.shouldUseDebugInstrRef();
431   mf.setUseDebugInstrRef(InstrRef);
432 
433   // Reset the target options before resetting the optimization
434   // level below.
435   // FIXME: This is a horrible hack and should be processed via
436   // codegen looking at the optimization level explicitly when
437   // it wants to look at it.
438   TM.resetTargetOptions(Fn);
439   // Reset OptLevel to None for optnone functions.
440   CodeGenOptLevel NewOptLevel = OptLevel;
441   if (OptLevel != CodeGenOptLevel::None && skipFunction(Fn))
442     NewOptLevel = CodeGenOptLevel::None;
443   OptLevelChanger OLC(*this, NewOptLevel);
444 
445   TII = MF->getSubtarget().getInstrInfo();
446   TLI = MF->getSubtarget().getTargetLowering();
447   RegInfo = &MF->getRegInfo();
448   LibInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(Fn);
449   GFI = Fn.hasGC() ? &getAnalysis<GCModuleInfo>().getFunctionInfo(Fn) : nullptr;
450   ORE = std::make_unique<OptimizationRemarkEmitter>(&Fn);
451   AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(mf.getFunction());
452   auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
453   BlockFrequencyInfo *BFI = nullptr;
454   if (PSI && PSI->hasProfileSummary() && OptLevel != CodeGenOptLevel::None)
455     BFI = &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI();
456 
457   FunctionVarLocs const *FnVarLocs = nullptr;
458   if (isAssignmentTrackingEnabled(*Fn.getParent()))
459     FnVarLocs = getAnalysis<AssignmentTrackingAnalysis>().getResults();
460 
461   ISEL_DUMP(dbgs() << "\n\n\n=== " << FuncName << "\n");
462 
463   UniformityInfo *UA = nullptr;
464   if (auto *UAPass = getAnalysisIfAvailable<UniformityInfoWrapperPass>())
465     UA = &UAPass->getUniformityInfo();
466   CurDAG->init(*MF, *ORE, this, LibInfo, UA, PSI, BFI, FnVarLocs);
467   FuncInfo->set(Fn, *MF, CurDAG);
468   SwiftError->setFunction(*MF);
469 
470   // Now get the optional analyzes if we want to.
471   // This is based on the possibly changed OptLevel (after optnone is taken
472   // into account).  That's unfortunate but OK because it just means we won't
473   // ask for passes that have been required anyway.
474 
475   if (UseMBPI && OptLevel != CodeGenOptLevel::None)
476     FuncInfo->BPI = &getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
477   else
478     FuncInfo->BPI = nullptr;
479 
480   if (OptLevel != CodeGenOptLevel::None)
481     AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
482   else
483     AA = nullptr;
484 
485   SDB->init(GFI, AA, AC, LibInfo);
486 
487   MF->setHasInlineAsm(false);
488 
489   FuncInfo->SplitCSR = false;
490 
491   // We split CSR if the target supports it for the given function
492   // and the function has only return exits.
493   if (OptLevel != CodeGenOptLevel::None && TLI->supportSplitCSR(MF)) {
494     FuncInfo->SplitCSR = true;
495 
496     // Collect all the return blocks.
497     for (const BasicBlock &BB : Fn) {
498       if (!succ_empty(&BB))
499         continue;
500 
501       const Instruction *Term = BB.getTerminator();
502       if (isa<UnreachableInst>(Term) || isa<ReturnInst>(Term))
503         continue;
504 
505       // Bail out if the exit block is not Return nor Unreachable.
506       FuncInfo->SplitCSR = false;
507       break;
508     }
509   }
510 
511   MachineBasicBlock *EntryMBB = &MF->front();
512   if (FuncInfo->SplitCSR)
513     // This performs initialization so lowering for SplitCSR will be correct.
514     TLI->initializeSplitCSR(EntryMBB);
515 
516   SelectAllBasicBlocks(Fn);
517   if (FastISelFailed && EnableFastISelFallbackReport) {
518     DiagnosticInfoISelFallback DiagFallback(Fn);
519     Fn.getContext().diagnose(DiagFallback);
520   }
521 
522   // Replace forward-declared registers with the registers containing
523   // the desired value.
524   // Note: it is important that this happens **before** the call to
525   // EmitLiveInCopies, since implementations can skip copies of unused
526   // registers. If we don't apply the reg fixups before, some registers may
527   // appear as unused and will be skipped, resulting in bad MI.
528   MachineRegisterInfo &MRI = MF->getRegInfo();
529   for (DenseMap<Register, Register>::iterator I = FuncInfo->RegFixups.begin(),
530                                               E = FuncInfo->RegFixups.end();
531        I != E; ++I) {
532     Register From = I->first;
533     Register To = I->second;
534     // If To is also scheduled to be replaced, find what its ultimate
535     // replacement is.
536     while (true) {
537       DenseMap<Register, Register>::iterator J = FuncInfo->RegFixups.find(To);
538       if (J == E)
539         break;
540       To = J->second;
541     }
542     // Make sure the new register has a sufficiently constrained register class.
543     if (From.isVirtual() && To.isVirtual())
544       MRI.constrainRegClass(To, MRI.getRegClass(From));
545     // Replace it.
546 
547     // Replacing one register with another won't touch the kill flags.
548     // We need to conservatively clear the kill flags as a kill on the old
549     // register might dominate existing uses of the new register.
550     if (!MRI.use_empty(To))
551       MRI.clearKillFlags(From);
552     MRI.replaceRegWith(From, To);
553   }
554 
555   // If the first basic block in the function has live ins that need to be
556   // copied into vregs, emit the copies into the top of the block before
557   // emitting the code for the block.
558   const TargetRegisterInfo &TRI = *MF->getSubtarget().getRegisterInfo();
559   RegInfo->EmitLiveInCopies(EntryMBB, TRI, *TII);
560 
561   // Insert copies in the entry block and the return blocks.
562   if (FuncInfo->SplitCSR) {
563     SmallVector<MachineBasicBlock*, 4> Returns;
564     // Collect all the return blocks.
565     for (MachineBasicBlock &MBB : mf) {
566       if (!MBB.succ_empty())
567         continue;
568 
569       MachineBasicBlock::iterator Term = MBB.getFirstTerminator();
570       if (Term != MBB.end() && Term->isReturn()) {
571         Returns.push_back(&MBB);
572         continue;
573       }
574     }
575     TLI->insertCopiesSplitCSR(EntryMBB, Returns);
576   }
577 
578   DenseMap<unsigned, unsigned> LiveInMap;
579   if (!FuncInfo->ArgDbgValues.empty())
580     for (std::pair<unsigned, unsigned> LI : RegInfo->liveins())
581       if (LI.second)
582         LiveInMap.insert(LI);
583 
584   // Insert DBG_VALUE instructions for function arguments to the entry block.
585   for (unsigned i = 0, e = FuncInfo->ArgDbgValues.size(); i != e; ++i) {
586     MachineInstr *MI = FuncInfo->ArgDbgValues[e - i - 1];
587     assert(MI->getOpcode() != TargetOpcode::DBG_VALUE_LIST &&
588            "Function parameters should not be described by DBG_VALUE_LIST.");
589     bool hasFI = MI->getDebugOperand(0).isFI();
590     Register Reg =
591         hasFI ? TRI.getFrameRegister(*MF) : MI->getDebugOperand(0).getReg();
592     if (Reg.isPhysical())
593       EntryMBB->insert(EntryMBB->begin(), MI);
594     else {
595       MachineInstr *Def = RegInfo->getVRegDef(Reg);
596       if (Def) {
597         MachineBasicBlock::iterator InsertPos = Def;
598         // FIXME: VR def may not be in entry block.
599         Def->getParent()->insert(std::next(InsertPos), MI);
600       } else
601         LLVM_DEBUG(dbgs() << "Dropping debug info for dead vreg"
602                           << Register::virtReg2Index(Reg) << "\n");
603     }
604 
605     // Don't try and extend through copies in instruction referencing mode.
606     if (InstrRef)
607       continue;
608 
609     // If Reg is live-in then update debug info to track its copy in a vreg.
610     DenseMap<unsigned, unsigned>::iterator LDI = LiveInMap.find(Reg);
611     if (LDI != LiveInMap.end()) {
612       assert(!hasFI && "There's no handling of frame pointer updating here yet "
613                        "- add if needed");
614       MachineInstr *Def = RegInfo->getVRegDef(LDI->second);
615       MachineBasicBlock::iterator InsertPos = Def;
616       const MDNode *Variable = MI->getDebugVariable();
617       const MDNode *Expr = MI->getDebugExpression();
618       DebugLoc DL = MI->getDebugLoc();
619       bool IsIndirect = MI->isIndirectDebugValue();
620       if (IsIndirect)
621         assert(MI->getDebugOffset().getImm() == 0 &&
622                "DBG_VALUE with nonzero offset");
623       assert(cast<DILocalVariable>(Variable)->isValidLocationForIntrinsic(DL) &&
624              "Expected inlined-at fields to agree");
625       assert(MI->getOpcode() != TargetOpcode::DBG_VALUE_LIST &&
626              "Didn't expect to see a DBG_VALUE_LIST here");
627       // Def is never a terminator here, so it is ok to increment InsertPos.
628       BuildMI(*EntryMBB, ++InsertPos, DL, TII->get(TargetOpcode::DBG_VALUE),
629               IsIndirect, LDI->second, Variable, Expr);
630 
631       // If this vreg is directly copied into an exported register then
632       // that COPY instructions also need DBG_VALUE, if it is the only
633       // user of LDI->second.
634       MachineInstr *CopyUseMI = nullptr;
635       for (MachineRegisterInfo::use_instr_iterator
636            UI = RegInfo->use_instr_begin(LDI->second),
637            E = RegInfo->use_instr_end(); UI != E; ) {
638         MachineInstr *UseMI = &*(UI++);
639         if (UseMI->isDebugValue()) continue;
640         if (UseMI->isCopy() && !CopyUseMI && UseMI->getParent() == EntryMBB) {
641           CopyUseMI = UseMI; continue;
642         }
643         // Otherwise this is another use or second copy use.
644         CopyUseMI = nullptr; break;
645       }
646       if (CopyUseMI &&
647           TRI.getRegSizeInBits(LDI->second, MRI) ==
648               TRI.getRegSizeInBits(CopyUseMI->getOperand(0).getReg(), MRI)) {
649         // Use MI's debug location, which describes where Variable was
650         // declared, rather than whatever is attached to CopyUseMI.
651         MachineInstr *NewMI =
652             BuildMI(*MF, DL, TII->get(TargetOpcode::DBG_VALUE), IsIndirect,
653                     CopyUseMI->getOperand(0).getReg(), Variable, Expr);
654         MachineBasicBlock::iterator Pos = CopyUseMI;
655         EntryMBB->insertAfter(Pos, NewMI);
656       }
657     }
658   }
659 
660   // For debug-info, in instruction referencing mode, we need to perform some
661   // post-isel maintenence.
662   if (MF->useDebugInstrRef())
663     MF->finalizeDebugInstrRefs();
664 
665   // Determine if there are any calls in this machine function.
666   MachineFrameInfo &MFI = MF->getFrameInfo();
667   for (const auto &MBB : *MF) {
668     if (MFI.hasCalls() && MF->hasInlineAsm())
669       break;
670 
671     for (const auto &MI : MBB) {
672       const MCInstrDesc &MCID = TII->get(MI.getOpcode());
673       if ((MCID.isCall() && !MCID.isReturn()) ||
674           MI.isStackAligningInlineAsm()) {
675         MFI.setHasCalls(true);
676       }
677       if (MI.isInlineAsm()) {
678         MF->setHasInlineAsm(true);
679       }
680     }
681   }
682 
683   // Determine if there is a call to setjmp in the machine function.
684   MF->setExposesReturnsTwice(Fn.callsFunctionThatReturnsTwice());
685 
686   // Determine if floating point is used for msvc
687   computeUsesMSVCFloatingPoint(TM.getTargetTriple(), Fn, MF->getMMI());
688 
689   // Release function-specific state. SDB and CurDAG are already cleared
690   // at this point.
691   FuncInfo->clear();
692 
693   ISEL_DUMP(dbgs() << "*** MachineFunction at end of ISel ***\n");
694   ISEL_DUMP(MF->print(dbgs()));
695 
696   return true;
697 }
698 
699 static void reportFastISelFailure(MachineFunction &MF,
700                                   OptimizationRemarkEmitter &ORE,
701                                   OptimizationRemarkMissed &R,
702                                   bool ShouldAbort) {
703   // Print the function name explicitly if we don't have a debug location (which
704   // makes the diagnostic less useful) or if we're going to emit a raw error.
705   if (!R.getLocation().isValid() || ShouldAbort)
706     R << (" (in function: " + MF.getName() + ")").str();
707 
708   if (ShouldAbort)
709     report_fatal_error(Twine(R.getMsg()));
710 
711   ORE.emit(R);
712   LLVM_DEBUG(dbgs() << R.getMsg() << "\n");
713 }
714 
715 void SelectionDAGISel::SelectBasicBlock(BasicBlock::const_iterator Begin,
716                                         BasicBlock::const_iterator End,
717                                         bool &HadTailCall) {
718   // Allow creating illegal types during DAG building for the basic block.
719   CurDAG->NewNodesMustHaveLegalTypes = false;
720 
721   // Lower the instructions. If a call is emitted as a tail call, cease emitting
722   // nodes for this block. If an instruction is elided, don't emit it, but do
723   // handle any debug-info attached to it.
724   for (BasicBlock::const_iterator I = Begin; I != End && !SDB->HasTailCall; ++I) {
725     if (!ElidedArgCopyInstrs.count(&*I))
726       SDB->visit(*I);
727     else
728       SDB->visitDbgInfo(*I);
729   }
730 
731   // Make sure the root of the DAG is up-to-date.
732   CurDAG->setRoot(SDB->getControlRoot());
733   HadTailCall = SDB->HasTailCall;
734   SDB->resolveOrClearDbgInfo();
735   SDB->clear();
736 
737   // Final step, emit the lowered DAG as machine code.
738   CodeGenAndEmitDAG();
739 }
740 
741 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
742   SmallPtrSet<SDNode *, 16> Added;
743   SmallVector<SDNode*, 128> Worklist;
744 
745   Worklist.push_back(CurDAG->getRoot().getNode());
746   Added.insert(CurDAG->getRoot().getNode());
747 
748   KnownBits Known;
749 
750   do {
751     SDNode *N = Worklist.pop_back_val();
752 
753     // Otherwise, add all chain operands to the worklist.
754     for (const SDValue &Op : N->op_values())
755       if (Op.getValueType() == MVT::Other && Added.insert(Op.getNode()).second)
756         Worklist.push_back(Op.getNode());
757 
758     // If this is a CopyToReg with a vreg dest, process it.
759     if (N->getOpcode() != ISD::CopyToReg)
760       continue;
761 
762     unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
763     if (!Register::isVirtualRegister(DestReg))
764       continue;
765 
766     // Ignore non-integer values.
767     SDValue Src = N->getOperand(2);
768     EVT SrcVT = Src.getValueType();
769     if (!SrcVT.isInteger())
770       continue;
771 
772     unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
773     Known = CurDAG->computeKnownBits(Src);
774     FuncInfo->AddLiveOutRegInfo(DestReg, NumSignBits, Known);
775   } while (!Worklist.empty());
776 }
777 
778 void SelectionDAGISel::CodeGenAndEmitDAG() {
779   StringRef GroupName = "sdag";
780   StringRef GroupDescription = "Instruction Selection and Scheduling";
781   std::string BlockName;
782   bool MatchFilterBB = false; (void)MatchFilterBB;
783 #ifndef NDEBUG
784   TargetTransformInfo &TTI =
785       getAnalysis<TargetTransformInfoWrapperPass>().getTTI(*FuncInfo->Fn);
786 #endif
787 
788   // Pre-type legalization allow creation of any node types.
789   CurDAG->NewNodesMustHaveLegalTypes = false;
790 
791 #ifndef NDEBUG
792   MatchFilterBB = (FilterDAGBasicBlockName.empty() ||
793                    FilterDAGBasicBlockName ==
794                        FuncInfo->MBB->getBasicBlock()->getName());
795 #endif
796 #ifdef NDEBUG
797   if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewDAGCombineLT ||
798       ViewLegalizeDAGs || ViewDAGCombine2 || ViewISelDAGs || ViewSchedDAGs ||
799       ViewSUnitDAGs)
800 #endif
801   {
802     BlockName =
803         (MF->getName() + ":" + FuncInfo->MBB->getBasicBlock()->getName()).str();
804   }
805   ISEL_DUMP(dbgs() << "\nInitial selection DAG: "
806                    << printMBBReference(*FuncInfo->MBB) << " '" << BlockName
807                    << "'\n";
808             CurDAG->dump());
809 
810 #ifndef NDEBUG
811   if (TTI.hasBranchDivergence())
812     CurDAG->VerifyDAGDivergence();
813 #endif
814 
815   if (ViewDAGCombine1 && MatchFilterBB)
816     CurDAG->viewGraph("dag-combine1 input for " + BlockName);
817 
818   // Run the DAG combiner in pre-legalize mode.
819   {
820     NamedRegionTimer T("combine1", "DAG Combining 1", GroupName,
821                        GroupDescription, TimePassesIsEnabled);
822     CurDAG->Combine(BeforeLegalizeTypes, AA, OptLevel);
823   }
824 
825   ISEL_DUMP(dbgs() << "\nOptimized lowered selection DAG: "
826                    << printMBBReference(*FuncInfo->MBB) << " '" << BlockName
827                    << "'\n";
828             CurDAG->dump());
829 
830 #ifndef NDEBUG
831   if (TTI.hasBranchDivergence())
832     CurDAG->VerifyDAGDivergence();
833 #endif
834 
835   // Second step, hack on the DAG until it only uses operations and types that
836   // the target supports.
837   if (ViewLegalizeTypesDAGs && MatchFilterBB)
838     CurDAG->viewGraph("legalize-types input for " + BlockName);
839 
840   bool Changed;
841   {
842     NamedRegionTimer T("legalize_types", "Type Legalization", GroupName,
843                        GroupDescription, TimePassesIsEnabled);
844     Changed = CurDAG->LegalizeTypes();
845   }
846 
847   ISEL_DUMP(dbgs() << "\nType-legalized selection DAG: "
848                    << printMBBReference(*FuncInfo->MBB) << " '" << BlockName
849                    << "'\n";
850             CurDAG->dump());
851 
852 #ifndef NDEBUG
853   if (TTI.hasBranchDivergence())
854     CurDAG->VerifyDAGDivergence();
855 #endif
856 
857   // Only allow creation of legal node types.
858   CurDAG->NewNodesMustHaveLegalTypes = true;
859 
860   if (Changed) {
861     if (ViewDAGCombineLT && MatchFilterBB)
862       CurDAG->viewGraph("dag-combine-lt input for " + BlockName);
863 
864     // Run the DAG combiner in post-type-legalize mode.
865     {
866       NamedRegionTimer T("combine_lt", "DAG Combining after legalize types",
867                          GroupName, GroupDescription, TimePassesIsEnabled);
868       CurDAG->Combine(AfterLegalizeTypes, AA, OptLevel);
869     }
870 
871     ISEL_DUMP(dbgs() << "\nOptimized type-legalized selection DAG: "
872                      << printMBBReference(*FuncInfo->MBB) << " '" << BlockName
873                      << "'\n";
874               CurDAG->dump());
875 
876 #ifndef NDEBUG
877     if (TTI.hasBranchDivergence())
878       CurDAG->VerifyDAGDivergence();
879 #endif
880   }
881 
882   {
883     NamedRegionTimer T("legalize_vec", "Vector Legalization", GroupName,
884                        GroupDescription, TimePassesIsEnabled);
885     Changed = CurDAG->LegalizeVectors();
886   }
887 
888   if (Changed) {
889     ISEL_DUMP(dbgs() << "\nVector-legalized selection DAG: "
890                      << printMBBReference(*FuncInfo->MBB) << " '" << BlockName
891                      << "'\n";
892               CurDAG->dump());
893 
894 #ifndef NDEBUG
895     if (TTI.hasBranchDivergence())
896       CurDAG->VerifyDAGDivergence();
897 #endif
898 
899     {
900       NamedRegionTimer T("legalize_types2", "Type Legalization 2", GroupName,
901                          GroupDescription, TimePassesIsEnabled);
902       CurDAG->LegalizeTypes();
903     }
904 
905     ISEL_DUMP(dbgs() << "\nVector/type-legalized selection DAG: "
906                      << printMBBReference(*FuncInfo->MBB) << " '" << BlockName
907                      << "'\n";
908               CurDAG->dump());
909 
910 #ifndef NDEBUG
911     if (TTI.hasBranchDivergence())
912       CurDAG->VerifyDAGDivergence();
913 #endif
914 
915     if (ViewDAGCombineLT && MatchFilterBB)
916       CurDAG->viewGraph("dag-combine-lv input for " + BlockName);
917 
918     // Run the DAG combiner in post-type-legalize mode.
919     {
920       NamedRegionTimer T("combine_lv", "DAG Combining after legalize vectors",
921                          GroupName, GroupDescription, TimePassesIsEnabled);
922       CurDAG->Combine(AfterLegalizeVectorOps, AA, OptLevel);
923     }
924 
925     ISEL_DUMP(dbgs() << "\nOptimized vector-legalized selection DAG: "
926                      << printMBBReference(*FuncInfo->MBB) << " '" << BlockName
927                      << "'\n";
928               CurDAG->dump());
929 
930 #ifndef NDEBUG
931     if (TTI.hasBranchDivergence())
932       CurDAG->VerifyDAGDivergence();
933 #endif
934   }
935 
936   if (ViewLegalizeDAGs && MatchFilterBB)
937     CurDAG->viewGraph("legalize input for " + BlockName);
938 
939   {
940     NamedRegionTimer T("legalize", "DAG Legalization", GroupName,
941                        GroupDescription, TimePassesIsEnabled);
942     CurDAG->Legalize();
943   }
944 
945   ISEL_DUMP(dbgs() << "\nLegalized selection DAG: "
946                    << printMBBReference(*FuncInfo->MBB) << " '" << BlockName
947                    << "'\n";
948             CurDAG->dump());
949 
950 #ifndef NDEBUG
951   if (TTI.hasBranchDivergence())
952     CurDAG->VerifyDAGDivergence();
953 #endif
954 
955   if (ViewDAGCombine2 && MatchFilterBB)
956     CurDAG->viewGraph("dag-combine2 input for " + BlockName);
957 
958   // Run the DAG combiner in post-legalize mode.
959   {
960     NamedRegionTimer T("combine2", "DAG Combining 2", GroupName,
961                        GroupDescription, TimePassesIsEnabled);
962     CurDAG->Combine(AfterLegalizeDAG, AA, OptLevel);
963   }
964 
965   ISEL_DUMP(dbgs() << "\nOptimized legalized selection DAG: "
966                    << printMBBReference(*FuncInfo->MBB) << " '" << BlockName
967                    << "'\n";
968             CurDAG->dump());
969 
970 #ifndef NDEBUG
971   if (TTI.hasBranchDivergence())
972     CurDAG->VerifyDAGDivergence();
973 #endif
974 
975   if (OptLevel != CodeGenOptLevel::None)
976     ComputeLiveOutVRegInfo();
977 
978   if (ViewISelDAGs && MatchFilterBB)
979     CurDAG->viewGraph("isel input for " + BlockName);
980 
981   // Third, instruction select all of the operations to machine code, adding the
982   // code to the MachineBasicBlock.
983   {
984     NamedRegionTimer T("isel", "Instruction Selection", GroupName,
985                        GroupDescription, TimePassesIsEnabled);
986     DoInstructionSelection();
987   }
988 
989   ISEL_DUMP(dbgs() << "\nSelected selection DAG: "
990                    << printMBBReference(*FuncInfo->MBB) << " '" << BlockName
991                    << "'\n";
992             CurDAG->dump());
993 
994   if (ViewSchedDAGs && MatchFilterBB)
995     CurDAG->viewGraph("scheduler input for " + BlockName);
996 
997   // Schedule machine code.
998   ScheduleDAGSDNodes *Scheduler = CreateScheduler();
999   {
1000     NamedRegionTimer T("sched", "Instruction Scheduling", GroupName,
1001                        GroupDescription, TimePassesIsEnabled);
1002     Scheduler->Run(CurDAG, FuncInfo->MBB);
1003   }
1004 
1005   if (ViewSUnitDAGs && MatchFilterBB)
1006     Scheduler->viewGraph();
1007 
1008   // Emit machine code to BB.  This can change 'BB' to the last block being
1009   // inserted into.
1010   MachineBasicBlock *FirstMBB = FuncInfo->MBB, *LastMBB;
1011   {
1012     NamedRegionTimer T("emit", "Instruction Creation", GroupName,
1013                        GroupDescription, TimePassesIsEnabled);
1014 
1015     // FuncInfo->InsertPt is passed by reference and set to the end of the
1016     // scheduled instructions.
1017     LastMBB = FuncInfo->MBB = Scheduler->EmitSchedule(FuncInfo->InsertPt);
1018   }
1019 
1020   // If the block was split, make sure we update any references that are used to
1021   // update PHI nodes later on.
1022   if (FirstMBB != LastMBB)
1023     SDB->UpdateSplitBlock(FirstMBB, LastMBB);
1024 
1025   // Free the scheduler state.
1026   {
1027     NamedRegionTimer T("cleanup", "Instruction Scheduling Cleanup", GroupName,
1028                        GroupDescription, TimePassesIsEnabled);
1029     delete Scheduler;
1030   }
1031 
1032   // Free the SelectionDAG state, now that we're finished with it.
1033   CurDAG->clear();
1034 }
1035 
1036 namespace {
1037 
1038 /// ISelUpdater - helper class to handle updates of the instruction selection
1039 /// graph.
1040 class ISelUpdater : public SelectionDAG::DAGUpdateListener {
1041   SelectionDAG::allnodes_iterator &ISelPosition;
1042 
1043 public:
1044   ISelUpdater(SelectionDAG &DAG, SelectionDAG::allnodes_iterator &isp)
1045     : SelectionDAG::DAGUpdateListener(DAG), ISelPosition(isp) {}
1046 
1047   /// NodeDeleted - Handle nodes deleted from the graph. If the node being
1048   /// deleted is the current ISelPosition node, update ISelPosition.
1049   ///
1050   void NodeDeleted(SDNode *N, SDNode *E) override {
1051     if (ISelPosition == SelectionDAG::allnodes_iterator(N))
1052       ++ISelPosition;
1053   }
1054 
1055   /// NodeInserted - Handle new nodes inserted into the graph: propagate
1056   /// metadata from root nodes that also applies to new nodes, in case the root
1057   /// is later deleted.
1058   void NodeInserted(SDNode *N) override {
1059     SDNode *CurNode = &*ISelPosition;
1060     if (MDNode *MD = DAG.getPCSections(CurNode))
1061       DAG.addPCSections(N, MD);
1062   }
1063 };
1064 
1065 } // end anonymous namespace
1066 
1067 // This function is used to enforce the topological node id property
1068 // leveraged during instruction selection. Before the selection process all
1069 // nodes are given a non-negative id such that all nodes have a greater id than
1070 // their operands. As this holds transitively we can prune checks that a node N
1071 // is a predecessor of M another by not recursively checking through M's
1072 // operands if N's ID is larger than M's ID. This significantly improves
1073 // performance of various legality checks (e.g. IsLegalToFold / UpdateChains).
1074 
1075 // However, when we fuse multiple nodes into a single node during the
1076 // selection we may induce a predecessor relationship between inputs and
1077 // outputs of distinct nodes being merged, violating the topological property.
1078 // Should a fused node have a successor which has yet to be selected,
1079 // our legality checks would be incorrect. To avoid this we mark all unselected
1080 // successor nodes, i.e. id != -1, as invalid for pruning by bit-negating (x =>
1081 // (-(x+1))) the ids and modify our pruning check to ignore negative Ids of M.
1082 // We use bit-negation to more clearly enforce that node id -1 can only be
1083 // achieved by selected nodes. As the conversion is reversable to the original
1084 // Id, topological pruning can still be leveraged when looking for unselected
1085 // nodes. This method is called internally in all ISel replacement related
1086 // functions.
1087 void SelectionDAGISel::EnforceNodeIdInvariant(SDNode *Node) {
1088   SmallVector<SDNode *, 4> Nodes;
1089   Nodes.push_back(Node);
1090 
1091   while (!Nodes.empty()) {
1092     SDNode *N = Nodes.pop_back_val();
1093     for (auto *U : N->uses()) {
1094       auto UId = U->getNodeId();
1095       if (UId > 0) {
1096         InvalidateNodeId(U);
1097         Nodes.push_back(U);
1098       }
1099     }
1100   }
1101 }
1102 
1103 // InvalidateNodeId - As explained in EnforceNodeIdInvariant, mark a
1104 // NodeId with the equivalent node id which is invalid for topological
1105 // pruning.
1106 void SelectionDAGISel::InvalidateNodeId(SDNode *N) {
1107   int InvalidId = -(N->getNodeId() + 1);
1108   N->setNodeId(InvalidId);
1109 }
1110 
1111 // getUninvalidatedNodeId - get original uninvalidated node id.
1112 int SelectionDAGISel::getUninvalidatedNodeId(SDNode *N) {
1113   int Id = N->getNodeId();
1114   if (Id < -1)
1115     return -(Id + 1);
1116   return Id;
1117 }
1118 
1119 void SelectionDAGISel::DoInstructionSelection() {
1120   LLVM_DEBUG(dbgs() << "===== Instruction selection begins: "
1121                     << printMBBReference(*FuncInfo->MBB) << " '"
1122                     << FuncInfo->MBB->getName() << "'\n");
1123 
1124   PreprocessISelDAG();
1125 
1126   // Select target instructions for the DAG.
1127   {
1128     // Number all nodes with a topological order and set DAGSize.
1129     DAGSize = CurDAG->AssignTopologicalOrder();
1130 
1131     // Create a dummy node (which is not added to allnodes), that adds
1132     // a reference to the root node, preventing it from being deleted,
1133     // and tracking any changes of the root.
1134     HandleSDNode Dummy(CurDAG->getRoot());
1135     SelectionDAG::allnodes_iterator ISelPosition (CurDAG->getRoot().getNode());
1136     ++ISelPosition;
1137 
1138     // Make sure that ISelPosition gets properly updated when nodes are deleted
1139     // in calls made from this function. New nodes inherit relevant metadata.
1140     ISelUpdater ISU(*CurDAG, ISelPosition);
1141 
1142     // The AllNodes list is now topological-sorted. Visit the
1143     // nodes by starting at the end of the list (the root of the
1144     // graph) and preceding back toward the beginning (the entry
1145     // node).
1146     while (ISelPosition != CurDAG->allnodes_begin()) {
1147       SDNode *Node = &*--ISelPosition;
1148       // Skip dead nodes. DAGCombiner is expected to eliminate all dead nodes,
1149       // but there are currently some corner cases that it misses. Also, this
1150       // makes it theoretically possible to disable the DAGCombiner.
1151       if (Node->use_empty())
1152         continue;
1153 
1154 #ifndef NDEBUG
1155       SmallVector<SDNode *, 4> Nodes;
1156       Nodes.push_back(Node);
1157 
1158       while (!Nodes.empty()) {
1159         auto N = Nodes.pop_back_val();
1160         if (N->getOpcode() == ISD::TokenFactor || N->getNodeId() < 0)
1161           continue;
1162         for (const SDValue &Op : N->op_values()) {
1163           if (Op->getOpcode() == ISD::TokenFactor)
1164             Nodes.push_back(Op.getNode());
1165           else {
1166             // We rely on topological ordering of node ids for checking for
1167             // cycles when fusing nodes during selection. All unselected nodes
1168             // successors of an already selected node should have a negative id.
1169             // This assertion will catch such cases. If this assertion triggers
1170             // it is likely you using DAG-level Value/Node replacement functions
1171             // (versus equivalent ISEL replacement) in backend-specific
1172             // selections. See comment in EnforceNodeIdInvariant for more
1173             // details.
1174             assert(Op->getNodeId() != -1 &&
1175                    "Node has already selected predecessor node");
1176           }
1177         }
1178       }
1179 #endif
1180 
1181       // When we are using non-default rounding modes or FP exception behavior
1182       // FP operations are represented by StrictFP pseudo-operations.  For
1183       // targets that do not (yet) understand strict FP operations directly,
1184       // we convert them to normal FP opcodes instead at this point.  This
1185       // will allow them to be handled by existing target-specific instruction
1186       // selectors.
1187       if (!TLI->isStrictFPEnabled() && Node->isStrictFPOpcode()) {
1188         // For some opcodes, we need to call TLI->getOperationAction using
1189         // the first operand type instead of the result type.  Note that this
1190         // must match what SelectionDAGLegalize::LegalizeOp is doing.
1191         EVT ActionVT;
1192         switch (Node->getOpcode()) {
1193         case ISD::STRICT_SINT_TO_FP:
1194         case ISD::STRICT_UINT_TO_FP:
1195         case ISD::STRICT_LRINT:
1196         case ISD::STRICT_LLRINT:
1197         case ISD::STRICT_LROUND:
1198         case ISD::STRICT_LLROUND:
1199         case ISD::STRICT_FSETCC:
1200         case ISD::STRICT_FSETCCS:
1201           ActionVT = Node->getOperand(1).getValueType();
1202           break;
1203         default:
1204           ActionVT = Node->getValueType(0);
1205           break;
1206         }
1207         if (TLI->getOperationAction(Node->getOpcode(), ActionVT)
1208             == TargetLowering::Expand)
1209           Node = CurDAG->mutateStrictFPToFP(Node);
1210       }
1211 
1212       LLVM_DEBUG(dbgs() << "\nISEL: Starting selection on root node: ";
1213                  Node->dump(CurDAG));
1214 
1215       Select(Node);
1216     }
1217 
1218     CurDAG->setRoot(Dummy.getValue());
1219   }
1220 
1221   LLVM_DEBUG(dbgs() << "\n===== Instruction selection ends:\n");
1222 
1223   PostprocessISelDAG();
1224 }
1225 
1226 static bool hasExceptionPointerOrCodeUser(const CatchPadInst *CPI) {
1227   for (const User *U : CPI->users()) {
1228     if (const IntrinsicInst *EHPtrCall = dyn_cast<IntrinsicInst>(U)) {
1229       Intrinsic::ID IID = EHPtrCall->getIntrinsicID();
1230       if (IID == Intrinsic::eh_exceptionpointer ||
1231           IID == Intrinsic::eh_exceptioncode)
1232         return true;
1233     }
1234   }
1235   return false;
1236 }
1237 
1238 // wasm.landingpad.index intrinsic is for associating a landing pad index number
1239 // with a catchpad instruction. Retrieve the landing pad index in the intrinsic
1240 // and store the mapping in the function.
1241 static void mapWasmLandingPadIndex(MachineBasicBlock *MBB,
1242                                    const CatchPadInst *CPI) {
1243   MachineFunction *MF = MBB->getParent();
1244   // In case of single catch (...), we don't emit LSDA, so we don't need
1245   // this information.
1246   bool IsSingleCatchAllClause =
1247       CPI->arg_size() == 1 &&
1248       cast<Constant>(CPI->getArgOperand(0))->isNullValue();
1249   // cathchpads for longjmp use an empty type list, e.g. catchpad within %0 []
1250   // and they don't need LSDA info
1251   bool IsCatchLongjmp = CPI->arg_size() == 0;
1252   if (!IsSingleCatchAllClause && !IsCatchLongjmp) {
1253     // Create a mapping from landing pad label to landing pad index.
1254     bool IntrFound = false;
1255     for (const User *U : CPI->users()) {
1256       if (const auto *Call = dyn_cast<IntrinsicInst>(U)) {
1257         Intrinsic::ID IID = Call->getIntrinsicID();
1258         if (IID == Intrinsic::wasm_landingpad_index) {
1259           Value *IndexArg = Call->getArgOperand(1);
1260           int Index = cast<ConstantInt>(IndexArg)->getZExtValue();
1261           MF->setWasmLandingPadIndex(MBB, Index);
1262           IntrFound = true;
1263           break;
1264         }
1265       }
1266     }
1267     assert(IntrFound && "wasm.landingpad.index intrinsic not found!");
1268     (void)IntrFound;
1269   }
1270 }
1271 
1272 /// PrepareEHLandingPad - Emit an EH_LABEL, set up live-in registers, and
1273 /// do other setup for EH landing-pad blocks.
1274 bool SelectionDAGISel::PrepareEHLandingPad() {
1275   MachineBasicBlock *MBB = FuncInfo->MBB;
1276   const Constant *PersonalityFn = FuncInfo->Fn->getPersonalityFn();
1277   const BasicBlock *LLVMBB = MBB->getBasicBlock();
1278   const TargetRegisterClass *PtrRC =
1279       TLI->getRegClassFor(TLI->getPointerTy(CurDAG->getDataLayout()));
1280 
1281   auto Pers = classifyEHPersonality(PersonalityFn);
1282 
1283   // Catchpads have one live-in register, which typically holds the exception
1284   // pointer or code.
1285   if (isFuncletEHPersonality(Pers)) {
1286     if (const auto *CPI = dyn_cast<CatchPadInst>(LLVMBB->getFirstNonPHI())) {
1287       if (hasExceptionPointerOrCodeUser(CPI)) {
1288         // Get or create the virtual register to hold the pointer or code.  Mark
1289         // the live in physreg and copy into the vreg.
1290         MCPhysReg EHPhysReg = TLI->getExceptionPointerRegister(PersonalityFn);
1291         assert(EHPhysReg && "target lacks exception pointer register");
1292         MBB->addLiveIn(EHPhysReg);
1293         unsigned VReg = FuncInfo->getCatchPadExceptionPointerVReg(CPI, PtrRC);
1294         BuildMI(*MBB, FuncInfo->InsertPt, SDB->getCurDebugLoc(),
1295                 TII->get(TargetOpcode::COPY), VReg)
1296             .addReg(EHPhysReg, RegState::Kill);
1297       }
1298     }
1299     return true;
1300   }
1301 
1302   // Add a label to mark the beginning of the landing pad.  Deletion of the
1303   // landing pad can thus be detected via the MachineModuleInfo.
1304   MCSymbol *Label = MF->addLandingPad(MBB);
1305 
1306   const MCInstrDesc &II = TII->get(TargetOpcode::EH_LABEL);
1307   BuildMI(*MBB, FuncInfo->InsertPt, SDB->getCurDebugLoc(), II)
1308     .addSym(Label);
1309 
1310   // If the unwinder does not preserve all registers, ensure that the
1311   // function marks the clobbered registers as used.
1312   const TargetRegisterInfo &TRI = *MF->getSubtarget().getRegisterInfo();
1313   if (auto *RegMask = TRI.getCustomEHPadPreservedMask(*MF))
1314     MF->getRegInfo().addPhysRegsUsedFromRegMask(RegMask);
1315 
1316   if (Pers == EHPersonality::Wasm_CXX) {
1317     if (const auto *CPI = dyn_cast<CatchPadInst>(LLVMBB->getFirstNonPHI()))
1318       mapWasmLandingPadIndex(MBB, CPI);
1319   } else {
1320     // Assign the call site to the landing pad's begin label.
1321     MF->setCallSiteLandingPad(Label, SDB->LPadToCallSiteMap[MBB]);
1322     // Mark exception register as live in.
1323     if (unsigned Reg = TLI->getExceptionPointerRegister(PersonalityFn))
1324       FuncInfo->ExceptionPointerVirtReg = MBB->addLiveIn(Reg, PtrRC);
1325     // Mark exception selector register as live in.
1326     if (unsigned Reg = TLI->getExceptionSelectorRegister(PersonalityFn))
1327       FuncInfo->ExceptionSelectorVirtReg = MBB->addLiveIn(Reg, PtrRC);
1328   }
1329 
1330   return true;
1331 }
1332 
1333 // Mark and Report IPToState for each Block under IsEHa
1334 void SelectionDAGISel::reportIPToStateForBlocks(MachineFunction *MF) {
1335   MachineModuleInfo &MMI = MF->getMMI();
1336   llvm::WinEHFuncInfo *EHInfo = MF->getWinEHFuncInfo();
1337   if (!EHInfo)
1338     return;
1339   for (auto MBBI = MF->begin(), E = MF->end(); MBBI != E; ++MBBI) {
1340     MachineBasicBlock *MBB = &*MBBI;
1341     const BasicBlock *BB = MBB->getBasicBlock();
1342     int State = EHInfo->BlockToStateMap[BB];
1343     if (BB->getFirstMayFaultInst()) {
1344       // Report IP range only for blocks with Faulty inst
1345       auto MBBb = MBB->getFirstNonPHI();
1346       MachineInstr *MIb = &*MBBb;
1347       if (MIb->isTerminator())
1348         continue;
1349 
1350       // Insert EH Labels
1351       MCSymbol *BeginLabel = MMI.getContext().createTempSymbol();
1352       MCSymbol *EndLabel = MMI.getContext().createTempSymbol();
1353       EHInfo->addIPToStateRange(State, BeginLabel, EndLabel);
1354       BuildMI(*MBB, MBBb, SDB->getCurDebugLoc(),
1355               TII->get(TargetOpcode::EH_LABEL))
1356           .addSym(BeginLabel);
1357       auto MBBe = MBB->instr_end();
1358       MachineInstr *MIe = &*(--MBBe);
1359       // insert before (possible multiple) terminators
1360       while (MIe->isTerminator())
1361         MIe = &*(--MBBe);
1362       ++MBBe;
1363       BuildMI(*MBB, MBBe, SDB->getCurDebugLoc(),
1364               TII->get(TargetOpcode::EH_LABEL))
1365           .addSym(EndLabel);
1366     }
1367   }
1368 }
1369 
1370 /// isFoldedOrDeadInstruction - Return true if the specified instruction is
1371 /// side-effect free and is either dead or folded into a generated instruction.
1372 /// Return false if it needs to be emitted.
1373 static bool isFoldedOrDeadInstruction(const Instruction *I,
1374                                       const FunctionLoweringInfo &FuncInfo) {
1375   return !I->mayWriteToMemory() && // Side-effecting instructions aren't folded.
1376          !I->isTerminator() &&     // Terminators aren't folded.
1377          !isa<DbgInfoIntrinsic>(I) && // Debug instructions aren't folded.
1378          !I->isEHPad() &&             // EH pad instructions aren't folded.
1379          !FuncInfo.isExportedInst(I); // Exported instrs must be computed.
1380 }
1381 
1382 static bool processIfEntryValueDbgDeclare(FunctionLoweringInfo &FuncInfo,
1383                                           const Value *Arg, DIExpression *Expr,
1384                                           DILocalVariable *Var,
1385                                           DebugLoc DbgLoc) {
1386   if (!Expr->isEntryValue() || !isa<Argument>(Arg))
1387     return false;
1388 
1389   auto ArgIt = FuncInfo.ValueMap.find(Arg);
1390   if (ArgIt == FuncInfo.ValueMap.end())
1391     return false;
1392   Register ArgVReg = ArgIt->getSecond();
1393 
1394   // Find the corresponding livein physical register to this argument.
1395   for (auto [PhysReg, VirtReg] : FuncInfo.RegInfo->liveins())
1396     if (VirtReg == ArgVReg) {
1397       // Append an op deref to account for the fact that this is a dbg_declare.
1398       Expr = DIExpression::append(Expr, dwarf::DW_OP_deref);
1399       FuncInfo.MF->setVariableDbgInfo(Var, Expr, PhysReg, DbgLoc);
1400       LLVM_DEBUG(dbgs() << "processDbgDeclare: setVariableDbgInfo Var=" << *Var
1401                         << ", Expr=" << *Expr << ",  MCRegister=" << PhysReg
1402                         << ", DbgLoc=" << DbgLoc << "\n");
1403       return true;
1404     }
1405   return false;
1406 }
1407 
1408 static bool processDbgDeclare(FunctionLoweringInfo &FuncInfo,
1409                               const Value *Address, DIExpression *Expr,
1410                               DILocalVariable *Var, DebugLoc DbgLoc) {
1411   if (!Address) {
1412     LLVM_DEBUG(dbgs() << "processDbgDeclares skipping " << *Var
1413                       << " (bad address)\n");
1414     return false;
1415   }
1416 
1417   if (processIfEntryValueDbgDeclare(FuncInfo, Address, Expr, Var, DbgLoc))
1418     return true;
1419 
1420   MachineFunction *MF = FuncInfo.MF;
1421   const DataLayout &DL = MF->getDataLayout();
1422 
1423   assert(Var && "Missing variable");
1424   assert(DbgLoc && "Missing location");
1425 
1426   // Look through casts and constant offset GEPs. These mostly come from
1427   // inalloca.
1428   APInt Offset(DL.getTypeSizeInBits(Address->getType()), 0);
1429   Address = Address->stripAndAccumulateInBoundsConstantOffsets(DL, Offset);
1430 
1431   // Check if the variable is a static alloca or a byval or inalloca
1432   // argument passed in memory. If it is not, then we will ignore this
1433   // intrinsic and handle this during isel like dbg.value.
1434   int FI = std::numeric_limits<int>::max();
1435   if (const auto *AI = dyn_cast<AllocaInst>(Address)) {
1436     auto SI = FuncInfo.StaticAllocaMap.find(AI);
1437     if (SI != FuncInfo.StaticAllocaMap.end())
1438       FI = SI->second;
1439   } else if (const auto *Arg = dyn_cast<Argument>(Address))
1440     FI = FuncInfo.getArgumentFrameIndex(Arg);
1441 
1442   if (FI == std::numeric_limits<int>::max())
1443     return false;
1444 
1445   if (Offset.getBoolValue())
1446     Expr = DIExpression::prepend(Expr, DIExpression::ApplyOffset,
1447                                  Offset.getZExtValue());
1448 
1449   LLVM_DEBUG(dbgs() << "processDbgDeclare: setVariableDbgInfo Var=" << *Var
1450                     << ", Expr=" << *Expr << ",  FI=" << FI
1451                     << ", DbgLoc=" << DbgLoc << "\n");
1452   MF->setVariableDbgInfo(Var, Expr, FI, DbgLoc);
1453   return true;
1454 }
1455 
1456 /// Collect llvm.dbg.declare information. This is done after argument lowering
1457 /// in case the declarations refer to arguments.
1458 static void processDbgDeclares(FunctionLoweringInfo &FuncInfo) {
1459   for (const auto &I : instructions(*FuncInfo.Fn)) {
1460     const auto *DI = dyn_cast<DbgDeclareInst>(&I);
1461     if (DI && processDbgDeclare(FuncInfo, DI->getAddress(), DI->getExpression(),
1462                                 DI->getVariable(), DI->getDebugLoc()))
1463       FuncInfo.PreprocessedDbgDeclares.insert(DI);
1464 
1465     for (const DPValue &DPV : I.getDbgValueRange()) {
1466       if (DPV.getType() == DPValue::LocationType::Declare &&
1467           processDbgDeclare(FuncInfo, DPV.getVariableLocationOp(0),
1468                             DPV.getExpression(), DPV.getVariable(),
1469                             DPV.getDebugLoc()))
1470         FuncInfo.PreprocessedDPVDeclares.insert(&DPV);
1471     }
1472   }
1473 }
1474 
1475 /// Collect single location variable information generated with assignment
1476 /// tracking. This is done after argument lowering in case the declarations
1477 /// refer to arguments.
1478 static void processSingleLocVars(FunctionLoweringInfo &FuncInfo,
1479                                  FunctionVarLocs const *FnVarLocs) {
1480   for (auto It = FnVarLocs->single_locs_begin(),
1481             End = FnVarLocs->single_locs_end();
1482        It != End; ++It) {
1483     assert(!It->Values.hasArgList() && "Single loc variadic ops not supported");
1484     processDbgDeclare(FuncInfo, It->Values.getVariableLocationOp(0), It->Expr,
1485                       FnVarLocs->getDILocalVariable(It->VariableID), It->DL);
1486   }
1487 }
1488 
1489 void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
1490   FastISelFailed = false;
1491   // Initialize the Fast-ISel state, if needed.
1492   FastISel *FastIS = nullptr;
1493   if (TM.Options.EnableFastISel) {
1494     LLVM_DEBUG(dbgs() << "Enabling fast-isel\n");
1495     FastIS = TLI->createFastISel(*FuncInfo, LibInfo);
1496   }
1497 
1498   ReversePostOrderTraversal<const Function*> RPOT(&Fn);
1499 
1500   // Lower arguments up front. An RPO iteration always visits the entry block
1501   // first.
1502   assert(*RPOT.begin() == &Fn.getEntryBlock());
1503   ++NumEntryBlocks;
1504 
1505   // Set up FuncInfo for ISel. Entry blocks never have PHIs.
1506   FuncInfo->MBB = FuncInfo->MBBMap[&Fn.getEntryBlock()];
1507   FuncInfo->InsertPt = FuncInfo->MBB->begin();
1508 
1509   CurDAG->setFunctionLoweringInfo(FuncInfo.get());
1510 
1511   if (!FastIS) {
1512     LowerArguments(Fn);
1513   } else {
1514     // See if fast isel can lower the arguments.
1515     FastIS->startNewBlock();
1516     if (!FastIS->lowerArguments()) {
1517       FastISelFailed = true;
1518       // Fast isel failed to lower these arguments
1519       ++NumFastIselFailLowerArguments;
1520 
1521       OptimizationRemarkMissed R("sdagisel", "FastISelFailure",
1522                                  Fn.getSubprogram(),
1523                                  &Fn.getEntryBlock());
1524       R << "FastISel didn't lower all arguments: "
1525         << ore::NV("Prototype", Fn.getFunctionType());
1526       reportFastISelFailure(*MF, *ORE, R, EnableFastISelAbort > 1);
1527 
1528       // Use SelectionDAG argument lowering
1529       LowerArguments(Fn);
1530       CurDAG->setRoot(SDB->getControlRoot());
1531       SDB->clear();
1532       CodeGenAndEmitDAG();
1533     }
1534 
1535     // If we inserted any instructions at the beginning, make a note of
1536     // where they are, so we can be sure to emit subsequent instructions
1537     // after them.
1538     if (FuncInfo->InsertPt != FuncInfo->MBB->begin())
1539       FastIS->setLastLocalValue(&*std::prev(FuncInfo->InsertPt));
1540     else
1541       FastIS->setLastLocalValue(nullptr);
1542   }
1543 
1544   bool Inserted = SwiftError->createEntriesInEntryBlock(SDB->getCurDebugLoc());
1545 
1546   if (FastIS && Inserted)
1547     FastIS->setLastLocalValue(&*std::prev(FuncInfo->InsertPt));
1548 
1549   if (isAssignmentTrackingEnabled(*Fn.getParent())) {
1550     assert(CurDAG->getFunctionVarLocs() &&
1551            "expected AssignmentTrackingAnalysis pass results");
1552     processSingleLocVars(*FuncInfo, CurDAG->getFunctionVarLocs());
1553   } else {
1554     processDbgDeclares(*FuncInfo);
1555   }
1556 
1557   // Iterate over all basic blocks in the function.
1558   StackProtector &SP = getAnalysis<StackProtector>();
1559   for (const BasicBlock *LLVMBB : RPOT) {
1560     if (OptLevel != CodeGenOptLevel::None) {
1561       bool AllPredsVisited = true;
1562       for (const BasicBlock *Pred : predecessors(LLVMBB)) {
1563         if (!FuncInfo->VisitedBBs.count(Pred)) {
1564           AllPredsVisited = false;
1565           break;
1566         }
1567       }
1568 
1569       if (AllPredsVisited) {
1570         for (const PHINode &PN : LLVMBB->phis())
1571           FuncInfo->ComputePHILiveOutRegInfo(&PN);
1572       } else {
1573         for (const PHINode &PN : LLVMBB->phis())
1574           FuncInfo->InvalidatePHILiveOutRegInfo(&PN);
1575       }
1576 
1577       FuncInfo->VisitedBBs.insert(LLVMBB);
1578     }
1579 
1580     BasicBlock::const_iterator const Begin =
1581         LLVMBB->getFirstNonPHI()->getIterator();
1582     BasicBlock::const_iterator const End = LLVMBB->end();
1583     BasicBlock::const_iterator BI = End;
1584 
1585     FuncInfo->MBB = FuncInfo->MBBMap[LLVMBB];
1586     if (!FuncInfo->MBB)
1587       continue; // Some blocks like catchpads have no code or MBB.
1588 
1589     // Insert new instructions after any phi or argument setup code.
1590     FuncInfo->InsertPt = FuncInfo->MBB->end();
1591 
1592     // Setup an EH landing-pad block.
1593     FuncInfo->ExceptionPointerVirtReg = 0;
1594     FuncInfo->ExceptionSelectorVirtReg = 0;
1595     if (LLVMBB->isEHPad())
1596       if (!PrepareEHLandingPad())
1597         continue;
1598 
1599     // Before doing SelectionDAG ISel, see if FastISel has been requested.
1600     if (FastIS) {
1601       if (LLVMBB != &Fn.getEntryBlock())
1602         FastIS->startNewBlock();
1603 
1604       unsigned NumFastIselRemaining = std::distance(Begin, End);
1605 
1606       // Pre-assign swifterror vregs.
1607       SwiftError->preassignVRegs(FuncInfo->MBB, Begin, End);
1608 
1609       // Do FastISel on as many instructions as possible.
1610       for (; BI != Begin; --BI) {
1611         const Instruction *Inst = &*std::prev(BI);
1612 
1613         // If we no longer require this instruction, skip it.
1614         if (isFoldedOrDeadInstruction(Inst, *FuncInfo) ||
1615             ElidedArgCopyInstrs.count(Inst)) {
1616           --NumFastIselRemaining;
1617           FastIS->handleDbgInfo(Inst);
1618           continue;
1619         }
1620 
1621         // Bottom-up: reset the insert pos at the top, after any local-value
1622         // instructions.
1623         FastIS->recomputeInsertPt();
1624 
1625         // Try to select the instruction with FastISel.
1626         if (FastIS->selectInstruction(Inst)) {
1627           --NumFastIselRemaining;
1628           ++NumFastIselSuccess;
1629 
1630           FastIS->handleDbgInfo(Inst);
1631           // If fast isel succeeded, skip over all the folded instructions, and
1632           // then see if there is a load right before the selected instructions.
1633           // Try to fold the load if so.
1634           const Instruction *BeforeInst = Inst;
1635           while (BeforeInst != &*Begin) {
1636             BeforeInst = &*std::prev(BasicBlock::const_iterator(BeforeInst));
1637             if (!isFoldedOrDeadInstruction(BeforeInst, *FuncInfo))
1638               break;
1639           }
1640           if (BeforeInst != Inst && isa<LoadInst>(BeforeInst) &&
1641               BeforeInst->hasOneUse() &&
1642               FastIS->tryToFoldLoad(cast<LoadInst>(BeforeInst), Inst)) {
1643             // If we succeeded, don't re-select the load.
1644             LLVM_DEBUG(dbgs()
1645                        << "FastISel folded load: " << *BeforeInst << "\n");
1646             FastIS->handleDbgInfo(BeforeInst);
1647             BI = std::next(BasicBlock::const_iterator(BeforeInst));
1648             --NumFastIselRemaining;
1649             ++NumFastIselSuccess;
1650           }
1651           continue;
1652         }
1653 
1654         FastISelFailed = true;
1655 
1656         // Then handle certain instructions as single-LLVM-Instruction blocks.
1657         // We cannot separate out GCrelocates to their own blocks since we need
1658         // to keep track of gc-relocates for a particular gc-statepoint. This is
1659         // done by SelectionDAGBuilder::LowerAsSTATEPOINT, called before
1660         // visitGCRelocate.
1661         if (isa<CallInst>(Inst) && !isa<GCStatepointInst>(Inst) &&
1662             !isa<GCRelocateInst>(Inst) && !isa<GCResultInst>(Inst)) {
1663           OptimizationRemarkMissed R("sdagisel", "FastISelFailure",
1664                                      Inst->getDebugLoc(), LLVMBB);
1665 
1666           R << "FastISel missed call";
1667 
1668           if (R.isEnabled() || EnableFastISelAbort) {
1669             std::string InstStrStorage;
1670             raw_string_ostream InstStr(InstStrStorage);
1671             InstStr << *Inst;
1672 
1673             R << ": " << InstStr.str();
1674           }
1675 
1676           reportFastISelFailure(*MF, *ORE, R, EnableFastISelAbort > 2);
1677 
1678           if (!Inst->getType()->isVoidTy() && !Inst->getType()->isTokenTy() &&
1679               !Inst->use_empty()) {
1680             Register &R = FuncInfo->ValueMap[Inst];
1681             if (!R)
1682               R = FuncInfo->CreateRegs(Inst);
1683           }
1684 
1685           bool HadTailCall = false;
1686           MachineBasicBlock::iterator SavedInsertPt = FuncInfo->InsertPt;
1687           SelectBasicBlock(Inst->getIterator(), BI, HadTailCall);
1688 
1689           // If the call was emitted as a tail call, we're done with the block.
1690           // We also need to delete any previously emitted instructions.
1691           if (HadTailCall) {
1692             FastIS->removeDeadCode(SavedInsertPt, FuncInfo->MBB->end());
1693             --BI;
1694             break;
1695           }
1696 
1697           // Recompute NumFastIselRemaining as Selection DAG instruction
1698           // selection may have handled the call, input args, etc.
1699           unsigned RemainingNow = std::distance(Begin, BI);
1700           NumFastIselFailures += NumFastIselRemaining - RemainingNow;
1701           NumFastIselRemaining = RemainingNow;
1702           continue;
1703         }
1704 
1705         OptimizationRemarkMissed R("sdagisel", "FastISelFailure",
1706                                    Inst->getDebugLoc(), LLVMBB);
1707 
1708         bool ShouldAbort = EnableFastISelAbort;
1709         if (Inst->isTerminator()) {
1710           // Use a different message for terminator misses.
1711           R << "FastISel missed terminator";
1712           // Don't abort for terminator unless the level is really high
1713           ShouldAbort = (EnableFastISelAbort > 2);
1714         } else {
1715           R << "FastISel missed";
1716         }
1717 
1718         if (R.isEnabled() || EnableFastISelAbort) {
1719           std::string InstStrStorage;
1720           raw_string_ostream InstStr(InstStrStorage);
1721           InstStr << *Inst;
1722           R << ": " << InstStr.str();
1723         }
1724 
1725         reportFastISelFailure(*MF, *ORE, R, ShouldAbort);
1726 
1727         NumFastIselFailures += NumFastIselRemaining;
1728         break;
1729       }
1730 
1731       FastIS->recomputeInsertPt();
1732     }
1733 
1734     if (SP.shouldEmitSDCheck(*LLVMBB)) {
1735       bool FunctionBasedInstrumentation =
1736           TLI->getSSPStackGuardCheck(*Fn.getParent());
1737       SDB->SPDescriptor.initialize(LLVMBB, FuncInfo->MBBMap[LLVMBB],
1738                                    FunctionBasedInstrumentation);
1739     }
1740 
1741     if (Begin != BI)
1742       ++NumDAGBlocks;
1743     else
1744       ++NumFastIselBlocks;
1745 
1746     if (Begin != BI) {
1747       // Run SelectionDAG instruction selection on the remainder of the block
1748       // not handled by FastISel. If FastISel is not run, this is the entire
1749       // block.
1750       bool HadTailCall;
1751       SelectBasicBlock(Begin, BI, HadTailCall);
1752 
1753       // But if FastISel was run, we already selected some of the block.
1754       // If we emitted a tail-call, we need to delete any previously emitted
1755       // instruction that follows it.
1756       if (FastIS && HadTailCall && FuncInfo->InsertPt != FuncInfo->MBB->end())
1757         FastIS->removeDeadCode(FuncInfo->InsertPt, FuncInfo->MBB->end());
1758     }
1759 
1760     if (FastIS)
1761       FastIS->finishBasicBlock();
1762     FinishBasicBlock();
1763     FuncInfo->PHINodesToUpdate.clear();
1764     ElidedArgCopyInstrs.clear();
1765   }
1766 
1767   // AsynchEH: Report Block State under -AsynchEH
1768   if (Fn.getParent()->getModuleFlag("eh-asynch"))
1769     reportIPToStateForBlocks(MF);
1770 
1771   SP.copyToMachineFrameInfo(MF->getFrameInfo());
1772 
1773   SwiftError->propagateVRegs();
1774 
1775   delete FastIS;
1776   SDB->clearDanglingDebugInfo();
1777   SDB->SPDescriptor.resetPerFunctionState();
1778 }
1779 
1780 void
1781 SelectionDAGISel::FinishBasicBlock() {
1782   LLVM_DEBUG(dbgs() << "Total amount of phi nodes to update: "
1783                     << FuncInfo->PHINodesToUpdate.size() << "\n";
1784              for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e;
1785                   ++i) dbgs()
1786              << "Node " << i << " : (" << FuncInfo->PHINodesToUpdate[i].first
1787              << ", " << FuncInfo->PHINodesToUpdate[i].second << ")\n");
1788 
1789   // Next, now that we know what the last MBB the LLVM BB expanded is, update
1790   // PHI nodes in successors.
1791   for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i) {
1792     MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[i].first);
1793     assert(PHI->isPHI() &&
1794            "This is not a machine PHI node that we are updating!");
1795     if (!FuncInfo->MBB->isSuccessor(PHI->getParent()))
1796       continue;
1797     PHI.addReg(FuncInfo->PHINodesToUpdate[i].second).addMBB(FuncInfo->MBB);
1798   }
1799 
1800   // Handle stack protector.
1801   if (SDB->SPDescriptor.shouldEmitFunctionBasedCheckStackProtector()) {
1802     // The target provides a guard check function. There is no need to
1803     // generate error handling code or to split current basic block.
1804     MachineBasicBlock *ParentMBB = SDB->SPDescriptor.getParentMBB();
1805 
1806     // Add load and check to the basicblock.
1807     FuncInfo->MBB = ParentMBB;
1808     FuncInfo->InsertPt =
1809         findSplitPointForStackProtector(ParentMBB, *TII);
1810     SDB->visitSPDescriptorParent(SDB->SPDescriptor, ParentMBB);
1811     CurDAG->setRoot(SDB->getRoot());
1812     SDB->clear();
1813     CodeGenAndEmitDAG();
1814 
1815     // Clear the Per-BB State.
1816     SDB->SPDescriptor.resetPerBBState();
1817   } else if (SDB->SPDescriptor.shouldEmitStackProtector()) {
1818     MachineBasicBlock *ParentMBB = SDB->SPDescriptor.getParentMBB();
1819     MachineBasicBlock *SuccessMBB = SDB->SPDescriptor.getSuccessMBB();
1820 
1821     // Find the split point to split the parent mbb. At the same time copy all
1822     // physical registers used in the tail of parent mbb into virtual registers
1823     // before the split point and back into physical registers after the split
1824     // point. This prevents us needing to deal with Live-ins and many other
1825     // register allocation issues caused by us splitting the parent mbb. The
1826     // register allocator will clean up said virtual copies later on.
1827     MachineBasicBlock::iterator SplitPoint =
1828         findSplitPointForStackProtector(ParentMBB, *TII);
1829 
1830     // Splice the terminator of ParentMBB into SuccessMBB.
1831     SuccessMBB->splice(SuccessMBB->end(), ParentMBB,
1832                        SplitPoint,
1833                        ParentMBB->end());
1834 
1835     // Add compare/jump on neq/jump to the parent BB.
1836     FuncInfo->MBB = ParentMBB;
1837     FuncInfo->InsertPt = ParentMBB->end();
1838     SDB->visitSPDescriptorParent(SDB->SPDescriptor, ParentMBB);
1839     CurDAG->setRoot(SDB->getRoot());
1840     SDB->clear();
1841     CodeGenAndEmitDAG();
1842 
1843     // CodeGen Failure MBB if we have not codegened it yet.
1844     MachineBasicBlock *FailureMBB = SDB->SPDescriptor.getFailureMBB();
1845     if (FailureMBB->empty()) {
1846       FuncInfo->MBB = FailureMBB;
1847       FuncInfo->InsertPt = FailureMBB->end();
1848       SDB->visitSPDescriptorFailure(SDB->SPDescriptor);
1849       CurDAG->setRoot(SDB->getRoot());
1850       SDB->clear();
1851       CodeGenAndEmitDAG();
1852     }
1853 
1854     // Clear the Per-BB State.
1855     SDB->SPDescriptor.resetPerBBState();
1856   }
1857 
1858   // Lower each BitTestBlock.
1859   for (auto &BTB : SDB->SL->BitTestCases) {
1860     // Lower header first, if it wasn't already lowered
1861     if (!BTB.Emitted) {
1862       // Set the current basic block to the mbb we wish to insert the code into
1863       FuncInfo->MBB = BTB.Parent;
1864       FuncInfo->InsertPt = FuncInfo->MBB->end();
1865       // Emit the code
1866       SDB->visitBitTestHeader(BTB, FuncInfo->MBB);
1867       CurDAG->setRoot(SDB->getRoot());
1868       SDB->clear();
1869       CodeGenAndEmitDAG();
1870     }
1871 
1872     BranchProbability UnhandledProb = BTB.Prob;
1873     for (unsigned j = 0, ej = BTB.Cases.size(); j != ej; ++j) {
1874       UnhandledProb -= BTB.Cases[j].ExtraProb;
1875       // Set the current basic block to the mbb we wish to insert the code into
1876       FuncInfo->MBB = BTB.Cases[j].ThisBB;
1877       FuncInfo->InsertPt = FuncInfo->MBB->end();
1878       // Emit the code
1879 
1880       // If all cases cover a contiguous range, it is not necessary to jump to
1881       // the default block after the last bit test fails. This is because the
1882       // range check during bit test header creation has guaranteed that every
1883       // case here doesn't go outside the range. In this case, there is no need
1884       // to perform the last bit test, as it will always be true. Instead, make
1885       // the second-to-last bit-test fall through to the target of the last bit
1886       // test, and delete the last bit test.
1887 
1888       MachineBasicBlock *NextMBB;
1889       if ((BTB.ContiguousRange || BTB.FallthroughUnreachable) && j + 2 == ej) {
1890         // Second-to-last bit-test with contiguous range or omitted range
1891         // check: fall through to the target of the final bit test.
1892         NextMBB = BTB.Cases[j + 1].TargetBB;
1893       } else if (j + 1 == ej) {
1894         // For the last bit test, fall through to Default.
1895         NextMBB = BTB.Default;
1896       } else {
1897         // Otherwise, fall through to the next bit test.
1898         NextMBB = BTB.Cases[j + 1].ThisBB;
1899       }
1900 
1901       SDB->visitBitTestCase(BTB, NextMBB, UnhandledProb, BTB.Reg, BTB.Cases[j],
1902                             FuncInfo->MBB);
1903 
1904       CurDAG->setRoot(SDB->getRoot());
1905       SDB->clear();
1906       CodeGenAndEmitDAG();
1907 
1908       if ((BTB.ContiguousRange || BTB.FallthroughUnreachable) && j + 2 == ej) {
1909         // Since we're not going to use the final bit test, remove it.
1910         BTB.Cases.pop_back();
1911         break;
1912       }
1913     }
1914 
1915     // Update PHI Nodes
1916     for (const std::pair<MachineInstr *, unsigned> &P :
1917          FuncInfo->PHINodesToUpdate) {
1918       MachineInstrBuilder PHI(*MF, P.first);
1919       MachineBasicBlock *PHIBB = PHI->getParent();
1920       assert(PHI->isPHI() &&
1921              "This is not a machine PHI node that we are updating!");
1922       // This is "default" BB. We have two jumps to it. From "header" BB and
1923       // from last "case" BB, unless the latter was skipped.
1924       if (PHIBB == BTB.Default) {
1925         PHI.addReg(P.second).addMBB(BTB.Parent);
1926         if (!BTB.ContiguousRange) {
1927           PHI.addReg(P.second).addMBB(BTB.Cases.back().ThisBB);
1928          }
1929       }
1930       // One of "cases" BB.
1931       for (const SwitchCG::BitTestCase &BT : BTB.Cases) {
1932         MachineBasicBlock* cBB = BT.ThisBB;
1933         if (cBB->isSuccessor(PHIBB))
1934           PHI.addReg(P.second).addMBB(cBB);
1935       }
1936     }
1937   }
1938   SDB->SL->BitTestCases.clear();
1939 
1940   // If the JumpTable record is filled in, then we need to emit a jump table.
1941   // Updating the PHI nodes is tricky in this case, since we need to determine
1942   // whether the PHI is a successor of the range check MBB or the jump table MBB
1943   for (unsigned i = 0, e = SDB->SL->JTCases.size(); i != e; ++i) {
1944     // Lower header first, if it wasn't already lowered
1945     if (!SDB->SL->JTCases[i].first.Emitted) {
1946       // Set the current basic block to the mbb we wish to insert the code into
1947       FuncInfo->MBB = SDB->SL->JTCases[i].first.HeaderBB;
1948       FuncInfo->InsertPt = FuncInfo->MBB->end();
1949       // Emit the code
1950       SDB->visitJumpTableHeader(SDB->SL->JTCases[i].second,
1951                                 SDB->SL->JTCases[i].first, FuncInfo->MBB);
1952       CurDAG->setRoot(SDB->getRoot());
1953       SDB->clear();
1954       CodeGenAndEmitDAG();
1955     }
1956 
1957     // Set the current basic block to the mbb we wish to insert the code into
1958     FuncInfo->MBB = SDB->SL->JTCases[i].second.MBB;
1959     FuncInfo->InsertPt = FuncInfo->MBB->end();
1960     // Emit the code
1961     SDB->visitJumpTable(SDB->SL->JTCases[i].second);
1962     CurDAG->setRoot(SDB->getRoot());
1963     SDB->clear();
1964     CodeGenAndEmitDAG();
1965 
1966     // Update PHI Nodes
1967     for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1968          pi != pe; ++pi) {
1969       MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[pi].first);
1970       MachineBasicBlock *PHIBB = PHI->getParent();
1971       assert(PHI->isPHI() &&
1972              "This is not a machine PHI node that we are updating!");
1973       // "default" BB. We can go there only from header BB.
1974       if (PHIBB == SDB->SL->JTCases[i].second.Default)
1975         PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second)
1976            .addMBB(SDB->SL->JTCases[i].first.HeaderBB);
1977       // JT BB. Just iterate over successors here
1978       if (FuncInfo->MBB->isSuccessor(PHIBB))
1979         PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second).addMBB(FuncInfo->MBB);
1980     }
1981   }
1982   SDB->SL->JTCases.clear();
1983 
1984   // If we generated any switch lowering information, build and codegen any
1985   // additional DAGs necessary.
1986   for (unsigned i = 0, e = SDB->SL->SwitchCases.size(); i != e; ++i) {
1987     // Set the current basic block to the mbb we wish to insert the code into
1988     FuncInfo->MBB = SDB->SL->SwitchCases[i].ThisBB;
1989     FuncInfo->InsertPt = FuncInfo->MBB->end();
1990 
1991     // Determine the unique successors.
1992     SmallVector<MachineBasicBlock *, 2> Succs;
1993     Succs.push_back(SDB->SL->SwitchCases[i].TrueBB);
1994     if (SDB->SL->SwitchCases[i].TrueBB != SDB->SL->SwitchCases[i].FalseBB)
1995       Succs.push_back(SDB->SL->SwitchCases[i].FalseBB);
1996 
1997     // Emit the code. Note that this could result in FuncInfo->MBB being split.
1998     SDB->visitSwitchCase(SDB->SL->SwitchCases[i], FuncInfo->MBB);
1999     CurDAG->setRoot(SDB->getRoot());
2000     SDB->clear();
2001     CodeGenAndEmitDAG();
2002 
2003     // Remember the last block, now that any splitting is done, for use in
2004     // populating PHI nodes in successors.
2005     MachineBasicBlock *ThisBB = FuncInfo->MBB;
2006 
2007     // Handle any PHI nodes in successors of this chunk, as if we were coming
2008     // from the original BB before switch expansion.  Note that PHI nodes can
2009     // occur multiple times in PHINodesToUpdate.  We have to be very careful to
2010     // handle them the right number of times.
2011     for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
2012       FuncInfo->MBB = Succs[i];
2013       FuncInfo->InsertPt = FuncInfo->MBB->end();
2014       // FuncInfo->MBB may have been removed from the CFG if a branch was
2015       // constant folded.
2016       if (ThisBB->isSuccessor(FuncInfo->MBB)) {
2017         for (MachineBasicBlock::iterator
2018              MBBI = FuncInfo->MBB->begin(), MBBE = FuncInfo->MBB->end();
2019              MBBI != MBBE && MBBI->isPHI(); ++MBBI) {
2020           MachineInstrBuilder PHI(*MF, MBBI);
2021           // This value for this PHI node is recorded in PHINodesToUpdate.
2022           for (unsigned pn = 0; ; ++pn) {
2023             assert(pn != FuncInfo->PHINodesToUpdate.size() &&
2024                    "Didn't find PHI entry!");
2025             if (FuncInfo->PHINodesToUpdate[pn].first == PHI) {
2026               PHI.addReg(FuncInfo->PHINodesToUpdate[pn].second).addMBB(ThisBB);
2027               break;
2028             }
2029           }
2030         }
2031       }
2032     }
2033   }
2034   SDB->SL->SwitchCases.clear();
2035 }
2036 
2037 /// Create the scheduler. If a specific scheduler was specified
2038 /// via the SchedulerRegistry, use it, otherwise select the
2039 /// one preferred by the target.
2040 ///
2041 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
2042   return ISHeuristic(this, OptLevel);
2043 }
2044 
2045 //===----------------------------------------------------------------------===//
2046 // Helper functions used by the generated instruction selector.
2047 //===----------------------------------------------------------------------===//
2048 // Calls to these methods are generated by tblgen.
2049 
2050 /// CheckAndMask - The isel is trying to match something like (and X, 255).  If
2051 /// the dag combiner simplified the 255, we still want to match.  RHS is the
2052 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
2053 /// specified in the .td file (e.g. 255).
2054 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
2055                                     int64_t DesiredMaskS) const {
2056   const APInt &ActualMask = RHS->getAPIntValue();
2057   const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
2058 
2059   // If the actual mask exactly matches, success!
2060   if (ActualMask == DesiredMask)
2061     return true;
2062 
2063   // If the actual AND mask is allowing unallowed bits, this doesn't match.
2064   if (!ActualMask.isSubsetOf(DesiredMask))
2065     return false;
2066 
2067   // Otherwise, the DAG Combiner may have proven that the value coming in is
2068   // either already zero or is not demanded.  Check for known zero input bits.
2069   APInt NeededMask = DesiredMask & ~ActualMask;
2070   if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
2071     return true;
2072 
2073   // TODO: check to see if missing bits are just not demanded.
2074 
2075   // Otherwise, this pattern doesn't match.
2076   return false;
2077 }
2078 
2079 /// CheckOrMask - The isel is trying to match something like (or X, 255).  If
2080 /// the dag combiner simplified the 255, we still want to match.  RHS is the
2081 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
2082 /// specified in the .td file (e.g. 255).
2083 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
2084                                    int64_t DesiredMaskS) const {
2085   const APInt &ActualMask = RHS->getAPIntValue();
2086   const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
2087 
2088   // If the actual mask exactly matches, success!
2089   if (ActualMask == DesiredMask)
2090     return true;
2091 
2092   // If the actual AND mask is allowing unallowed bits, this doesn't match.
2093   if (!ActualMask.isSubsetOf(DesiredMask))
2094     return false;
2095 
2096   // Otherwise, the DAG Combiner may have proven that the value coming in is
2097   // either already zero or is not demanded.  Check for known zero input bits.
2098   APInt NeededMask = DesiredMask & ~ActualMask;
2099   KnownBits Known = CurDAG->computeKnownBits(LHS);
2100 
2101   // If all the missing bits in the or are already known to be set, match!
2102   if (NeededMask.isSubsetOf(Known.One))
2103     return true;
2104 
2105   // TODO: check to see if missing bits are just not demanded.
2106 
2107   // Otherwise, this pattern doesn't match.
2108   return false;
2109 }
2110 
2111 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
2112 /// by tblgen.  Others should not call it.
2113 void SelectionDAGISel::SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops,
2114                                                      const SDLoc &DL) {
2115   std::vector<SDValue> InOps;
2116   std::swap(InOps, Ops);
2117 
2118   Ops.push_back(InOps[InlineAsm::Op_InputChain]); // 0
2119   Ops.push_back(InOps[InlineAsm::Op_AsmString]);  // 1
2120   Ops.push_back(InOps[InlineAsm::Op_MDNode]);     // 2, !srcloc
2121   Ops.push_back(InOps[InlineAsm::Op_ExtraInfo]);  // 3 (SideEffect, AlignStack)
2122 
2123   unsigned i = InlineAsm::Op_FirstOperand, e = InOps.size();
2124   if (InOps[e-1].getValueType() == MVT::Glue)
2125     --e;  // Don't process a glue operand if it is here.
2126 
2127   while (i != e) {
2128     InlineAsm::Flag Flags(InOps[i]->getAsZExtVal());
2129     if (!Flags.isMemKind() && !Flags.isFuncKind()) {
2130       // Just skip over this operand, copying the operands verbatim.
2131       Ops.insert(Ops.end(), InOps.begin() + i,
2132                  InOps.begin() + i + Flags.getNumOperandRegisters() + 1);
2133       i += Flags.getNumOperandRegisters() + 1;
2134     } else {
2135       assert(Flags.getNumOperandRegisters() == 1 &&
2136              "Memory operand with multiple values?");
2137 
2138       unsigned TiedToOperand;
2139       if (Flags.isUseOperandTiedToDef(TiedToOperand)) {
2140         // We need the constraint ID from the operand this is tied to.
2141         unsigned CurOp = InlineAsm::Op_FirstOperand;
2142         Flags = InlineAsm::Flag(InOps[CurOp]->getAsZExtVal());
2143         for (; TiedToOperand; --TiedToOperand) {
2144           CurOp += Flags.getNumOperandRegisters() + 1;
2145           Flags = InlineAsm::Flag(InOps[CurOp]->getAsZExtVal());
2146         }
2147       }
2148 
2149       // Otherwise, this is a memory operand.  Ask the target to select it.
2150       std::vector<SDValue> SelOps;
2151       const InlineAsm::ConstraintCode ConstraintID =
2152           Flags.getMemoryConstraintID();
2153       if (SelectInlineAsmMemoryOperand(InOps[i+1], ConstraintID, SelOps))
2154         report_fatal_error("Could not match memory address.  Inline asm"
2155                            " failure!");
2156 
2157       // Add this to the output node.
2158       Flags = InlineAsm::Flag(Flags.isMemKind() ? InlineAsm::Kind::Mem
2159                                                 : InlineAsm::Kind::Func,
2160                               SelOps.size());
2161       Flags.setMemConstraint(ConstraintID);
2162       Ops.push_back(CurDAG->getTargetConstant(Flags, DL, MVT::i32));
2163       llvm::append_range(Ops, SelOps);
2164       i += 2;
2165     }
2166   }
2167 
2168   // Add the glue input back if present.
2169   if (e != InOps.size())
2170     Ops.push_back(InOps.back());
2171 }
2172 
2173 /// findGlueUse - Return use of MVT::Glue value produced by the specified
2174 /// SDNode.
2175 ///
2176 static SDNode *findGlueUse(SDNode *N) {
2177   unsigned FlagResNo = N->getNumValues()-1;
2178   for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
2179     SDUse &Use = I.getUse();
2180     if (Use.getResNo() == FlagResNo)
2181       return Use.getUser();
2182   }
2183   return nullptr;
2184 }
2185 
2186 /// findNonImmUse - Return true if "Def" is a predecessor of "Root" via a path
2187 /// beyond "ImmedUse".  We may ignore chains as they are checked separately.
2188 static bool findNonImmUse(SDNode *Root, SDNode *Def, SDNode *ImmedUse,
2189                           bool IgnoreChains) {
2190   SmallPtrSet<const SDNode *, 16> Visited;
2191   SmallVector<const SDNode *, 16> WorkList;
2192   // Only check if we have non-immediate uses of Def.
2193   if (ImmedUse->isOnlyUserOf(Def))
2194     return false;
2195 
2196   // We don't care about paths to Def that go through ImmedUse so mark it
2197   // visited and mark non-def operands as used.
2198   Visited.insert(ImmedUse);
2199   for (const SDValue &Op : ImmedUse->op_values()) {
2200     SDNode *N = Op.getNode();
2201     // Ignore chain deps (they are validated by
2202     // HandleMergeInputChains) and immediate uses
2203     if ((Op.getValueType() == MVT::Other && IgnoreChains) || N == Def)
2204       continue;
2205     if (!Visited.insert(N).second)
2206       continue;
2207     WorkList.push_back(N);
2208   }
2209 
2210   // Initialize worklist to operands of Root.
2211   if (Root != ImmedUse) {
2212     for (const SDValue &Op : Root->op_values()) {
2213       SDNode *N = Op.getNode();
2214       // Ignore chains (they are validated by HandleMergeInputChains)
2215       if ((Op.getValueType() == MVT::Other && IgnoreChains) || N == Def)
2216         continue;
2217       if (!Visited.insert(N).second)
2218         continue;
2219       WorkList.push_back(N);
2220     }
2221   }
2222 
2223   return SDNode::hasPredecessorHelper(Def, Visited, WorkList, 0, true);
2224 }
2225 
2226 /// IsProfitableToFold - Returns true if it's profitable to fold the specific
2227 /// operand node N of U during instruction selection that starts at Root.
2228 bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
2229                                           SDNode *Root) const {
2230   if (OptLevel == CodeGenOptLevel::None)
2231     return false;
2232   return N.hasOneUse();
2233 }
2234 
2235 /// IsLegalToFold - Returns true if the specific operand node N of
2236 /// U can be folded during instruction selection that starts at Root.
2237 bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
2238                                      CodeGenOptLevel OptLevel,
2239                                      bool IgnoreChains) {
2240   if (OptLevel == CodeGenOptLevel::None)
2241     return false;
2242 
2243   // If Root use can somehow reach N through a path that doesn't contain
2244   // U then folding N would create a cycle. e.g. In the following
2245   // diagram, Root can reach N through X. If N is folded into Root, then
2246   // X is both a predecessor and a successor of U.
2247   //
2248   //          [N*]           //
2249   //         ^   ^           //
2250   //        /     \          //
2251   //      [U*]    [X]?       //
2252   //        ^     ^          //
2253   //         \   /           //
2254   //          \ /            //
2255   //         [Root*]         //
2256   //
2257   // * indicates nodes to be folded together.
2258   //
2259   // If Root produces glue, then it gets (even more) interesting. Since it
2260   // will be "glued" together with its glue use in the scheduler, we need to
2261   // check if it might reach N.
2262   //
2263   //          [N*]           //
2264   //         ^   ^           //
2265   //        /     \          //
2266   //      [U*]    [X]?       //
2267   //        ^       ^        //
2268   //         \       \       //
2269   //          \      |       //
2270   //         [Root*] |       //
2271   //          ^      |       //
2272   //          f      |       //
2273   //          |      /       //
2274   //         [Y]    /        //
2275   //           ^   /         //
2276   //           f  /          //
2277   //           | /           //
2278   //          [GU]           //
2279   //
2280   // If GU (glue use) indirectly reaches N (the load), and Root folds N
2281   // (call it Fold), then X is a predecessor of GU and a successor of
2282   // Fold. But since Fold and GU are glued together, this will create
2283   // a cycle in the scheduling graph.
2284 
2285   // If the node has glue, walk down the graph to the "lowest" node in the
2286   // glueged set.
2287   EVT VT = Root->getValueType(Root->getNumValues()-1);
2288   while (VT == MVT::Glue) {
2289     SDNode *GU = findGlueUse(Root);
2290     if (!GU)
2291       break;
2292     Root = GU;
2293     VT = Root->getValueType(Root->getNumValues()-1);
2294 
2295     // If our query node has a glue result with a use, we've walked up it.  If
2296     // the user (which has already been selected) has a chain or indirectly uses
2297     // the chain, HandleMergeInputChains will not consider it.  Because of
2298     // this, we cannot ignore chains in this predicate.
2299     IgnoreChains = false;
2300   }
2301 
2302   return !findNonImmUse(Root, N.getNode(), U, IgnoreChains);
2303 }
2304 
2305 void SelectionDAGISel::Select_INLINEASM(SDNode *N) {
2306   SDLoc DL(N);
2307 
2308   std::vector<SDValue> Ops(N->op_begin(), N->op_end());
2309   SelectInlineAsmMemoryOperands(Ops, DL);
2310 
2311   const EVT VTs[] = {MVT::Other, MVT::Glue};
2312   SDValue New = CurDAG->getNode(N->getOpcode(), DL, VTs, Ops);
2313   New->setNodeId(-1);
2314   ReplaceUses(N, New.getNode());
2315   CurDAG->RemoveDeadNode(N);
2316 }
2317 
2318 void SelectionDAGISel::Select_READ_REGISTER(SDNode *Op) {
2319   SDLoc dl(Op);
2320   MDNodeSDNode *MD = cast<MDNodeSDNode>(Op->getOperand(1));
2321   const MDString *RegStr = cast<MDString>(MD->getMD()->getOperand(0));
2322 
2323   EVT VT = Op->getValueType(0);
2324   LLT Ty = VT.isSimple() ? getLLTForMVT(VT.getSimpleVT()) : LLT();
2325   Register Reg =
2326       TLI->getRegisterByName(RegStr->getString().data(), Ty,
2327                              CurDAG->getMachineFunction());
2328   SDValue New = CurDAG->getCopyFromReg(
2329                         Op->getOperand(0), dl, Reg, Op->getValueType(0));
2330   New->setNodeId(-1);
2331   ReplaceUses(Op, New.getNode());
2332   CurDAG->RemoveDeadNode(Op);
2333 }
2334 
2335 void SelectionDAGISel::Select_WRITE_REGISTER(SDNode *Op) {
2336   SDLoc dl(Op);
2337   MDNodeSDNode *MD = cast<MDNodeSDNode>(Op->getOperand(1));
2338   const MDString *RegStr = cast<MDString>(MD->getMD()->getOperand(0));
2339 
2340   EVT VT = Op->getOperand(2).getValueType();
2341   LLT Ty = VT.isSimple() ? getLLTForMVT(VT.getSimpleVT()) : LLT();
2342 
2343   Register Reg = TLI->getRegisterByName(RegStr->getString().data(), Ty,
2344                                         CurDAG->getMachineFunction());
2345   SDValue New = CurDAG->getCopyToReg(
2346                         Op->getOperand(0), dl, Reg, Op->getOperand(2));
2347   New->setNodeId(-1);
2348   ReplaceUses(Op, New.getNode());
2349   CurDAG->RemoveDeadNode(Op);
2350 }
2351 
2352 void SelectionDAGISel::Select_UNDEF(SDNode *N) {
2353   CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF, N->getValueType(0));
2354 }
2355 
2356 void SelectionDAGISel::Select_FREEZE(SDNode *N) {
2357   // TODO: We don't have FREEZE pseudo-instruction in MachineInstr-level now.
2358   // If FREEZE instruction is added later, the code below must be changed as
2359   // well.
2360   CurDAG->SelectNodeTo(N, TargetOpcode::COPY, N->getValueType(0),
2361                        N->getOperand(0));
2362 }
2363 
2364 void SelectionDAGISel::Select_ARITH_FENCE(SDNode *N) {
2365   CurDAG->SelectNodeTo(N, TargetOpcode::ARITH_FENCE, N->getValueType(0),
2366                        N->getOperand(0));
2367 }
2368 
2369 void SelectionDAGISel::Select_MEMBARRIER(SDNode *N) {
2370   CurDAG->SelectNodeTo(N, TargetOpcode::MEMBARRIER, N->getValueType(0),
2371                        N->getOperand(0));
2372 }
2373 
2374 void SelectionDAGISel::pushStackMapLiveVariable(SmallVectorImpl<SDValue> &Ops,
2375                                                 SDValue OpVal, SDLoc DL) {
2376   SDNode *OpNode = OpVal.getNode();
2377 
2378   // FrameIndex nodes should have been directly emitted to TargetFrameIndex
2379   // nodes at DAG-construction time.
2380   assert(OpNode->getOpcode() != ISD::FrameIndex);
2381 
2382   if (OpNode->getOpcode() == ISD::Constant) {
2383     Ops.push_back(
2384         CurDAG->getTargetConstant(StackMaps::ConstantOp, DL, MVT::i64));
2385     Ops.push_back(CurDAG->getTargetConstant(OpNode->getAsZExtVal(), DL,
2386                                             OpVal.getValueType()));
2387   } else {
2388     Ops.push_back(OpVal);
2389   }
2390 }
2391 
2392 void SelectionDAGISel::Select_STACKMAP(SDNode *N) {
2393   SmallVector<SDValue, 32> Ops;
2394   auto *It = N->op_begin();
2395   SDLoc DL(N);
2396 
2397   // Stash the chain and glue operands so we can move them to the end.
2398   SDValue Chain = *It++;
2399   SDValue InGlue = *It++;
2400 
2401   // <id> operand.
2402   SDValue ID = *It++;
2403   assert(ID.getValueType() == MVT::i64);
2404   Ops.push_back(ID);
2405 
2406   // <numShadowBytes> operand.
2407   SDValue Shad = *It++;
2408   assert(Shad.getValueType() == MVT::i32);
2409   Ops.push_back(Shad);
2410 
2411   // Live variable operands.
2412   for (; It != N->op_end(); It++)
2413     pushStackMapLiveVariable(Ops, *It, DL);
2414 
2415   Ops.push_back(Chain);
2416   Ops.push_back(InGlue);
2417 
2418   SDVTList NodeTys = CurDAG->getVTList(MVT::Other, MVT::Glue);
2419   CurDAG->SelectNodeTo(N, TargetOpcode::STACKMAP, NodeTys, Ops);
2420 }
2421 
2422 void SelectionDAGISel::Select_PATCHPOINT(SDNode *N) {
2423   SmallVector<SDValue, 32> Ops;
2424   auto *It = N->op_begin();
2425   SDLoc DL(N);
2426 
2427   // Cache arguments that will be moved to the end in the target node.
2428   SDValue Chain = *It++;
2429   std::optional<SDValue> Glue;
2430   if (It->getValueType() == MVT::Glue)
2431     Glue = *It++;
2432   SDValue RegMask = *It++;
2433 
2434   // <id> operand.
2435   SDValue ID = *It++;
2436   assert(ID.getValueType() == MVT::i64);
2437   Ops.push_back(ID);
2438 
2439   // <numShadowBytes> operand.
2440   SDValue Shad = *It++;
2441   assert(Shad.getValueType() == MVT::i32);
2442   Ops.push_back(Shad);
2443 
2444   // Add the callee.
2445   Ops.push_back(*It++);
2446 
2447   // Add <numArgs>.
2448   SDValue NumArgs = *It++;
2449   assert(NumArgs.getValueType() == MVT::i32);
2450   Ops.push_back(NumArgs);
2451 
2452   // Calling convention.
2453   Ops.push_back(*It++);
2454 
2455   // Push the args for the call.
2456   for (uint64_t I = NumArgs->getAsZExtVal(); I != 0; I--)
2457     Ops.push_back(*It++);
2458 
2459   // Now push the live variables.
2460   for (; It != N->op_end(); It++)
2461     pushStackMapLiveVariable(Ops, *It, DL);
2462 
2463   // Finally, the regmask, chain and (if present) glue are moved to the end.
2464   Ops.push_back(RegMask);
2465   Ops.push_back(Chain);
2466   if (Glue.has_value())
2467     Ops.push_back(*Glue);
2468 
2469   SDVTList NodeTys = N->getVTList();
2470   CurDAG->SelectNodeTo(N, TargetOpcode::PATCHPOINT, NodeTys, Ops);
2471 }
2472 
2473 /// GetVBR - decode a vbr encoding whose top bit is set.
2474 LLVM_ATTRIBUTE_ALWAYS_INLINE static uint64_t
2475 GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
2476   assert(Val >= 128 && "Not a VBR");
2477   Val &= 127;  // Remove first vbr bit.
2478 
2479   unsigned Shift = 7;
2480   uint64_t NextBits;
2481   do {
2482     NextBits = MatcherTable[Idx++];
2483     Val |= (NextBits&127) << Shift;
2484     Shift += 7;
2485   } while (NextBits & 128);
2486 
2487   return Val;
2488 }
2489 
2490 void SelectionDAGISel::Select_JUMP_TABLE_DEBUG_INFO(SDNode *N) {
2491   SDLoc dl(N);
2492   CurDAG->SelectNodeTo(N, TargetOpcode::JUMP_TABLE_DEBUG_INFO, MVT::Glue,
2493                        CurDAG->getTargetConstant(N->getConstantOperandVal(1),
2494                                                  dl, MVT::i64, true));
2495 }
2496 
2497 /// When a match is complete, this method updates uses of interior chain results
2498 /// to use the new results.
2499 void SelectionDAGISel::UpdateChains(
2500     SDNode *NodeToMatch, SDValue InputChain,
2501     SmallVectorImpl<SDNode *> &ChainNodesMatched, bool isMorphNodeTo) {
2502   SmallVector<SDNode*, 4> NowDeadNodes;
2503 
2504   // Now that all the normal results are replaced, we replace the chain and
2505   // glue results if present.
2506   if (!ChainNodesMatched.empty()) {
2507     assert(InputChain.getNode() &&
2508            "Matched input chains but didn't produce a chain");
2509     // Loop over all of the nodes we matched that produced a chain result.
2510     // Replace all the chain results with the final chain we ended up with.
2511     for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
2512       SDNode *ChainNode = ChainNodesMatched[i];
2513       // If ChainNode is null, it's because we replaced it on a previous
2514       // iteration and we cleared it out of the map. Just skip it.
2515       if (!ChainNode)
2516         continue;
2517 
2518       assert(ChainNode->getOpcode() != ISD::DELETED_NODE &&
2519              "Deleted node left in chain");
2520 
2521       // Don't replace the results of the root node if we're doing a
2522       // MorphNodeTo.
2523       if (ChainNode == NodeToMatch && isMorphNodeTo)
2524         continue;
2525 
2526       SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1);
2527       if (ChainVal.getValueType() == MVT::Glue)
2528         ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
2529       assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
2530       SelectionDAG::DAGNodeDeletedListener NDL(
2531           *CurDAG, [&](SDNode *N, SDNode *E) {
2532             std::replace(ChainNodesMatched.begin(), ChainNodesMatched.end(), N,
2533                          static_cast<SDNode *>(nullptr));
2534           });
2535       if (ChainNode->getOpcode() != ISD::TokenFactor)
2536         ReplaceUses(ChainVal, InputChain);
2537 
2538       // If the node became dead and we haven't already seen it, delete it.
2539       if (ChainNode != NodeToMatch && ChainNode->use_empty() &&
2540           !llvm::is_contained(NowDeadNodes, ChainNode))
2541         NowDeadNodes.push_back(ChainNode);
2542     }
2543   }
2544 
2545   if (!NowDeadNodes.empty())
2546     CurDAG->RemoveDeadNodes(NowDeadNodes);
2547 
2548   LLVM_DEBUG(dbgs() << "ISEL: Match complete!\n");
2549 }
2550 
2551 /// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
2552 /// operation for when the pattern matched at least one node with a chains.  The
2553 /// input vector contains a list of all of the chained nodes that we match.  We
2554 /// must determine if this is a valid thing to cover (i.e. matching it won't
2555 /// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
2556 /// be used as the input node chain for the generated nodes.
2557 static SDValue
2558 HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
2559                        SelectionDAG *CurDAG) {
2560 
2561   SmallPtrSet<const SDNode *, 16> Visited;
2562   SmallVector<const SDNode *, 8> Worklist;
2563   SmallVector<SDValue, 3> InputChains;
2564   unsigned int Max = 8192;
2565 
2566   // Quick exit on trivial merge.
2567   if (ChainNodesMatched.size() == 1)
2568     return ChainNodesMatched[0]->getOperand(0);
2569 
2570   // Add chains that aren't already added (internal). Peek through
2571   // token factors.
2572   std::function<void(const SDValue)> AddChains = [&](const SDValue V) {
2573     if (V.getValueType() != MVT::Other)
2574       return;
2575     if (V->getOpcode() == ISD::EntryToken)
2576       return;
2577     if (!Visited.insert(V.getNode()).second)
2578       return;
2579     if (V->getOpcode() == ISD::TokenFactor) {
2580       for (const SDValue &Op : V->op_values())
2581         AddChains(Op);
2582     } else
2583       InputChains.push_back(V);
2584   };
2585 
2586   for (auto *N : ChainNodesMatched) {
2587     Worklist.push_back(N);
2588     Visited.insert(N);
2589   }
2590 
2591   while (!Worklist.empty())
2592     AddChains(Worklist.pop_back_val()->getOperand(0));
2593 
2594   // Skip the search if there are no chain dependencies.
2595   if (InputChains.size() == 0)
2596     return CurDAG->getEntryNode();
2597 
2598   // If one of these chains is a successor of input, we must have a
2599   // node that is both the predecessor and successor of the
2600   // to-be-merged nodes. Fail.
2601   Visited.clear();
2602   for (SDValue V : InputChains)
2603     Worklist.push_back(V.getNode());
2604 
2605   for (auto *N : ChainNodesMatched)
2606     if (SDNode::hasPredecessorHelper(N, Visited, Worklist, Max, true))
2607       return SDValue();
2608 
2609   // Return merged chain.
2610   if (InputChains.size() == 1)
2611     return InputChains[0];
2612   return CurDAG->getNode(ISD::TokenFactor, SDLoc(ChainNodesMatched[0]),
2613                          MVT::Other, InputChains);
2614 }
2615 
2616 /// MorphNode - Handle morphing a node in place for the selector.
2617 SDNode *SelectionDAGISel::
2618 MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList,
2619           ArrayRef<SDValue> Ops, unsigned EmitNodeInfo) {
2620   // It is possible we're using MorphNodeTo to replace a node with no
2621   // normal results with one that has a normal result (or we could be
2622   // adding a chain) and the input could have glue and chains as well.
2623   // In this case we need to shift the operands down.
2624   // FIXME: This is a horrible hack and broken in obscure cases, no worse
2625   // than the old isel though.
2626   int OldGlueResultNo = -1, OldChainResultNo = -1;
2627 
2628   unsigned NTMNumResults = Node->getNumValues();
2629   if (Node->getValueType(NTMNumResults-1) == MVT::Glue) {
2630     OldGlueResultNo = NTMNumResults-1;
2631     if (NTMNumResults != 1 &&
2632         Node->getValueType(NTMNumResults-2) == MVT::Other)
2633       OldChainResultNo = NTMNumResults-2;
2634   } else if (Node->getValueType(NTMNumResults-1) == MVT::Other)
2635     OldChainResultNo = NTMNumResults-1;
2636 
2637   // Call the underlying SelectionDAG routine to do the transmogrification. Note
2638   // that this deletes operands of the old node that become dead.
2639   SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops);
2640 
2641   // MorphNodeTo can operate in two ways: if an existing node with the
2642   // specified operands exists, it can just return it.  Otherwise, it
2643   // updates the node in place to have the requested operands.
2644   if (Res == Node) {
2645     // If we updated the node in place, reset the node ID.  To the isel,
2646     // this should be just like a newly allocated machine node.
2647     Res->setNodeId(-1);
2648   }
2649 
2650   unsigned ResNumResults = Res->getNumValues();
2651   // Move the glue if needed.
2652   if ((EmitNodeInfo & OPFL_GlueOutput) && OldGlueResultNo != -1 &&
2653       static_cast<unsigned>(OldGlueResultNo) != ResNumResults - 1)
2654     ReplaceUses(SDValue(Node, OldGlueResultNo),
2655                 SDValue(Res, ResNumResults - 1));
2656 
2657   if ((EmitNodeInfo & OPFL_GlueOutput) != 0)
2658     --ResNumResults;
2659 
2660   // Move the chain reference if needed.
2661   if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 &&
2662       static_cast<unsigned>(OldChainResultNo) != ResNumResults - 1)
2663     ReplaceUses(SDValue(Node, OldChainResultNo),
2664                 SDValue(Res, ResNumResults - 1));
2665 
2666   // Otherwise, no replacement happened because the node already exists. Replace
2667   // Uses of the old node with the new one.
2668   if (Res != Node) {
2669     ReplaceNode(Node, Res);
2670   } else {
2671     EnforceNodeIdInvariant(Res);
2672   }
2673 
2674   return Res;
2675 }
2676 
2677 /// CheckSame - Implements OP_CheckSame.
2678 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2679 CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex, SDValue N,
2680           const SmallVectorImpl<std::pair<SDValue, SDNode *>> &RecordedNodes) {
2681   // Accept if it is exactly the same as a previously recorded node.
2682   unsigned RecNo = MatcherTable[MatcherIndex++];
2683   assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2684   return N == RecordedNodes[RecNo].first;
2685 }
2686 
2687 /// CheckChildSame - Implements OP_CheckChildXSame.
2688 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool CheckChildSame(
2689     const unsigned char *MatcherTable, unsigned &MatcherIndex, SDValue N,
2690     const SmallVectorImpl<std::pair<SDValue, SDNode *>> &RecordedNodes,
2691     unsigned ChildNo) {
2692   if (ChildNo >= N.getNumOperands())
2693     return false;  // Match fails if out of range child #.
2694   return ::CheckSame(MatcherTable, MatcherIndex, N.getOperand(ChildNo),
2695                      RecordedNodes);
2696 }
2697 
2698 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
2699 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2700 CheckPatternPredicate(unsigned Opcode, const unsigned char *MatcherTable,
2701                       unsigned &MatcherIndex, const SelectionDAGISel &SDISel) {
2702   bool TwoBytePredNo =
2703       Opcode == SelectionDAGISel::OPC_CheckPatternPredicateTwoByte;
2704   unsigned PredNo =
2705       TwoBytePredNo || Opcode == SelectionDAGISel::OPC_CheckPatternPredicate
2706           ? MatcherTable[MatcherIndex++]
2707           : Opcode - SelectionDAGISel::OPC_CheckPatternPredicate0;
2708   if (TwoBytePredNo)
2709     PredNo |= MatcherTable[MatcherIndex++] << 8;
2710   return SDISel.CheckPatternPredicate(PredNo);
2711 }
2712 
2713 /// CheckNodePredicate - Implements OP_CheckNodePredicate.
2714 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2715 CheckNodePredicate(unsigned Opcode, const unsigned char *MatcherTable,
2716                    unsigned &MatcherIndex, const SelectionDAGISel &SDISel,
2717                    SDNode *N) {
2718   unsigned PredNo = Opcode == SelectionDAGISel::OPC_CheckPredicate
2719                         ? MatcherTable[MatcherIndex++]
2720                         : Opcode - SelectionDAGISel::OPC_CheckPredicate0;
2721   return SDISel.CheckNodePredicate(N, PredNo);
2722 }
2723 
2724 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2725 CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2726             SDNode *N) {
2727   uint16_t Opc = MatcherTable[MatcherIndex++];
2728   Opc |= static_cast<uint16_t>(MatcherTable[MatcherIndex++]) << 8;
2729   return N->getOpcode() == Opc;
2730 }
2731 
2732 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool CheckType(MVT::SimpleValueType VT,
2733                                                    SDValue N,
2734                                                    const TargetLowering *TLI,
2735                                                    const DataLayout &DL) {
2736   if (N.getValueType() == VT)
2737     return true;
2738 
2739   // Handle the case when VT is iPTR.
2740   return VT == MVT::iPTR && N.getValueType() == TLI->getPointerTy(DL);
2741 }
2742 
2743 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2744 CheckChildType(MVT::SimpleValueType VT, SDValue N, const TargetLowering *TLI,
2745                const DataLayout &DL, unsigned ChildNo) {
2746   if (ChildNo >= N.getNumOperands())
2747     return false; // Match fails if out of range child #.
2748   return ::CheckType(VT, N.getOperand(ChildNo), TLI, DL);
2749 }
2750 
2751 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2752 CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2753               SDValue N) {
2754   return cast<CondCodeSDNode>(N)->get() ==
2755          static_cast<ISD::CondCode>(MatcherTable[MatcherIndex++]);
2756 }
2757 
2758 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2759 CheckChild2CondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2760                     SDValue N) {
2761   if (2 >= N.getNumOperands())
2762     return false;
2763   return ::CheckCondCode(MatcherTable, MatcherIndex, N.getOperand(2));
2764 }
2765 
2766 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2767 CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2768                SDValue N, const TargetLowering *TLI, const DataLayout &DL) {
2769   MVT::SimpleValueType VT =
2770       static_cast<MVT::SimpleValueType>(MatcherTable[MatcherIndex++]);
2771   if (cast<VTSDNode>(N)->getVT() == VT)
2772     return true;
2773 
2774   // Handle the case when VT is iPTR.
2775   return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI->getPointerTy(DL);
2776 }
2777 
2778 // Bit 0 stores the sign of the immediate. The upper bits contain the magnitude
2779 // shifted left by 1.
2780 static uint64_t decodeSignRotatedValue(uint64_t V) {
2781   if ((V & 1) == 0)
2782     return V >> 1;
2783   if (V != 1)
2784     return -(V >> 1);
2785   // There is no such thing as -0 with integers.  "-0" really means MININT.
2786   return 1ULL << 63;
2787 }
2788 
2789 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2790 CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2791              SDValue N) {
2792   int64_t Val = MatcherTable[MatcherIndex++];
2793   if (Val & 128)
2794     Val = GetVBR(Val, MatcherTable, MatcherIndex);
2795 
2796   Val = decodeSignRotatedValue(Val);
2797 
2798   ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
2799   return C && C->getAPIntValue().trySExtValue() == Val;
2800 }
2801 
2802 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2803 CheckChildInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2804                   SDValue N, unsigned ChildNo) {
2805   if (ChildNo >= N.getNumOperands())
2806     return false;  // Match fails if out of range child #.
2807   return ::CheckInteger(MatcherTable, MatcherIndex, N.getOperand(ChildNo));
2808 }
2809 
2810 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2811 CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2812             SDValue N, const SelectionDAGISel &SDISel) {
2813   int64_t Val = MatcherTable[MatcherIndex++];
2814   if (Val & 128)
2815     Val = GetVBR(Val, MatcherTable, MatcherIndex);
2816 
2817   if (N->getOpcode() != ISD::AND) return false;
2818 
2819   ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
2820   return C && SDISel.CheckAndMask(N.getOperand(0), C, Val);
2821 }
2822 
2823 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2824 CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex, SDValue N,
2825            const SelectionDAGISel &SDISel) {
2826   int64_t Val = MatcherTable[MatcherIndex++];
2827   if (Val & 128)
2828     Val = GetVBR(Val, MatcherTable, MatcherIndex);
2829 
2830   if (N->getOpcode() != ISD::OR) return false;
2831 
2832   ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
2833   return C && SDISel.CheckOrMask(N.getOperand(0), C, Val);
2834 }
2835 
2836 /// IsPredicateKnownToFail - If we know how and can do so without pushing a
2837 /// scope, evaluate the current node.  If the current predicate is known to
2838 /// fail, set Result=true and return anything.  If the current predicate is
2839 /// known to pass, set Result=false and return the MatcherIndex to continue
2840 /// with.  If the current predicate is unknown, set Result=false and return the
2841 /// MatcherIndex to continue with.
2842 static unsigned IsPredicateKnownToFail(const unsigned char *Table,
2843                                        unsigned Index, SDValue N,
2844                                        bool &Result,
2845                                        const SelectionDAGISel &SDISel,
2846                   SmallVectorImpl<std::pair<SDValue, SDNode*>> &RecordedNodes) {
2847   unsigned Opcode = Table[Index++];
2848   switch (Opcode) {
2849   default:
2850     Result = false;
2851     return Index-1;  // Could not evaluate this predicate.
2852   case SelectionDAGISel::OPC_CheckSame:
2853     Result = !::CheckSame(Table, Index, N, RecordedNodes);
2854     return Index;
2855   case SelectionDAGISel::OPC_CheckChild0Same:
2856   case SelectionDAGISel::OPC_CheckChild1Same:
2857   case SelectionDAGISel::OPC_CheckChild2Same:
2858   case SelectionDAGISel::OPC_CheckChild3Same:
2859     Result = !::CheckChildSame(Table, Index, N, RecordedNodes,
2860                         Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Same);
2861     return Index;
2862   case SelectionDAGISel::OPC_CheckPatternPredicate:
2863   case SelectionDAGISel::OPC_CheckPatternPredicate0:
2864   case SelectionDAGISel::OPC_CheckPatternPredicate1:
2865   case SelectionDAGISel::OPC_CheckPatternPredicate2:
2866   case SelectionDAGISel::OPC_CheckPatternPredicate3:
2867   case SelectionDAGISel::OPC_CheckPatternPredicate4:
2868   case SelectionDAGISel::OPC_CheckPatternPredicate5:
2869   case SelectionDAGISel::OPC_CheckPatternPredicate6:
2870   case SelectionDAGISel::OPC_CheckPatternPredicate7:
2871   case SelectionDAGISel::OPC_CheckPatternPredicateTwoByte:
2872     Result = !::CheckPatternPredicate(Opcode, Table, Index, SDISel);
2873     return Index;
2874   case SelectionDAGISel::OPC_CheckPredicate:
2875   case SelectionDAGISel::OPC_CheckPredicate0:
2876   case SelectionDAGISel::OPC_CheckPredicate1:
2877   case SelectionDAGISel::OPC_CheckPredicate2:
2878   case SelectionDAGISel::OPC_CheckPredicate3:
2879   case SelectionDAGISel::OPC_CheckPredicate4:
2880   case SelectionDAGISel::OPC_CheckPredicate5:
2881   case SelectionDAGISel::OPC_CheckPredicate6:
2882   case SelectionDAGISel::OPC_CheckPredicate7:
2883     Result = !::CheckNodePredicate(Opcode, Table, Index, SDISel, N.getNode());
2884     return Index;
2885   case SelectionDAGISel::OPC_CheckOpcode:
2886     Result = !::CheckOpcode(Table, Index, N.getNode());
2887     return Index;
2888   case SelectionDAGISel::OPC_CheckType:
2889   case SelectionDAGISel::OPC_CheckTypeI32:
2890   case SelectionDAGISel::OPC_CheckTypeI64: {
2891     MVT::SimpleValueType VT;
2892     switch (Opcode) {
2893     case SelectionDAGISel::OPC_CheckTypeI32:
2894       VT = MVT::i32;
2895       break;
2896     case SelectionDAGISel::OPC_CheckTypeI64:
2897       VT = MVT::i64;
2898       break;
2899     default:
2900       VT = static_cast<MVT::SimpleValueType>(Table[Index++]);
2901       break;
2902     }
2903     Result = !::CheckType(VT, N, SDISel.TLI, SDISel.CurDAG->getDataLayout());
2904     return Index;
2905   }
2906   case SelectionDAGISel::OPC_CheckTypeRes: {
2907     unsigned Res = Table[Index++];
2908     Result = !::CheckType(static_cast<MVT::SimpleValueType>(Table[Index++]),
2909                           N.getValue(Res), SDISel.TLI,
2910                           SDISel.CurDAG->getDataLayout());
2911     return Index;
2912   }
2913   case SelectionDAGISel::OPC_CheckChild0Type:
2914   case SelectionDAGISel::OPC_CheckChild1Type:
2915   case SelectionDAGISel::OPC_CheckChild2Type:
2916   case SelectionDAGISel::OPC_CheckChild3Type:
2917   case SelectionDAGISel::OPC_CheckChild4Type:
2918   case SelectionDAGISel::OPC_CheckChild5Type:
2919   case SelectionDAGISel::OPC_CheckChild6Type:
2920   case SelectionDAGISel::OPC_CheckChild7Type:
2921   case SelectionDAGISel::OPC_CheckChild0TypeI32:
2922   case SelectionDAGISel::OPC_CheckChild1TypeI32:
2923   case SelectionDAGISel::OPC_CheckChild2TypeI32:
2924   case SelectionDAGISel::OPC_CheckChild3TypeI32:
2925   case SelectionDAGISel::OPC_CheckChild4TypeI32:
2926   case SelectionDAGISel::OPC_CheckChild5TypeI32:
2927   case SelectionDAGISel::OPC_CheckChild6TypeI32:
2928   case SelectionDAGISel::OPC_CheckChild7TypeI32:
2929   case SelectionDAGISel::OPC_CheckChild0TypeI64:
2930   case SelectionDAGISel::OPC_CheckChild1TypeI64:
2931   case SelectionDAGISel::OPC_CheckChild2TypeI64:
2932   case SelectionDAGISel::OPC_CheckChild3TypeI64:
2933   case SelectionDAGISel::OPC_CheckChild4TypeI64:
2934   case SelectionDAGISel::OPC_CheckChild5TypeI64:
2935   case SelectionDAGISel::OPC_CheckChild6TypeI64:
2936   case SelectionDAGISel::OPC_CheckChild7TypeI64: {
2937     MVT::SimpleValueType VT;
2938     unsigned ChildNo;
2939     if (Opcode >= SelectionDAGISel::OPC_CheckChild0TypeI32 &&
2940         Opcode <= SelectionDAGISel::OPC_CheckChild7TypeI32) {
2941       VT = MVT::i32;
2942       ChildNo = Opcode - SelectionDAGISel::OPC_CheckChild0TypeI32;
2943     } else if (Opcode >= SelectionDAGISel::OPC_CheckChild0TypeI64 &&
2944                Opcode <= SelectionDAGISel::OPC_CheckChild7TypeI64) {
2945       VT = MVT::i64;
2946       ChildNo = Opcode - SelectionDAGISel::OPC_CheckChild0TypeI64;
2947     } else {
2948       VT = static_cast<MVT::SimpleValueType>(Table[Index++]);
2949       ChildNo = Opcode - SelectionDAGISel::OPC_CheckChild0Type;
2950     }
2951     Result = !::CheckChildType(VT, N, SDISel.TLI,
2952                                SDISel.CurDAG->getDataLayout(), ChildNo);
2953     return Index;
2954   }
2955   case SelectionDAGISel::OPC_CheckCondCode:
2956     Result = !::CheckCondCode(Table, Index, N);
2957     return Index;
2958   case SelectionDAGISel::OPC_CheckChild2CondCode:
2959     Result = !::CheckChild2CondCode(Table, Index, N);
2960     return Index;
2961   case SelectionDAGISel::OPC_CheckValueType:
2962     Result = !::CheckValueType(Table, Index, N, SDISel.TLI,
2963                                SDISel.CurDAG->getDataLayout());
2964     return Index;
2965   case SelectionDAGISel::OPC_CheckInteger:
2966     Result = !::CheckInteger(Table, Index, N);
2967     return Index;
2968   case SelectionDAGISel::OPC_CheckChild0Integer:
2969   case SelectionDAGISel::OPC_CheckChild1Integer:
2970   case SelectionDAGISel::OPC_CheckChild2Integer:
2971   case SelectionDAGISel::OPC_CheckChild3Integer:
2972   case SelectionDAGISel::OPC_CheckChild4Integer:
2973     Result = !::CheckChildInteger(Table, Index, N,
2974                      Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Integer);
2975     return Index;
2976   case SelectionDAGISel::OPC_CheckAndImm:
2977     Result = !::CheckAndImm(Table, Index, N, SDISel);
2978     return Index;
2979   case SelectionDAGISel::OPC_CheckOrImm:
2980     Result = !::CheckOrImm(Table, Index, N, SDISel);
2981     return Index;
2982   }
2983 }
2984 
2985 namespace {
2986 
2987 struct MatchScope {
2988   /// FailIndex - If this match fails, this is the index to continue with.
2989   unsigned FailIndex;
2990 
2991   /// NodeStack - The node stack when the scope was formed.
2992   SmallVector<SDValue, 4> NodeStack;
2993 
2994   /// NumRecordedNodes - The number of recorded nodes when the scope was formed.
2995   unsigned NumRecordedNodes;
2996 
2997   /// NumMatchedMemRefs - The number of matched memref entries.
2998   unsigned NumMatchedMemRefs;
2999 
3000   /// InputChain/InputGlue - The current chain/glue
3001   SDValue InputChain, InputGlue;
3002 
3003   /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty.
3004   bool HasChainNodesMatched;
3005 };
3006 
3007 /// \A DAG update listener to keep the matching state
3008 /// (i.e. RecordedNodes and MatchScope) uptodate if the target is allowed to
3009 /// change the DAG while matching.  X86 addressing mode matcher is an example
3010 /// for this.
3011 class MatchStateUpdater : public SelectionDAG::DAGUpdateListener
3012 {
3013   SDNode **NodeToMatch;
3014   SmallVectorImpl<std::pair<SDValue, SDNode *>> &RecordedNodes;
3015   SmallVectorImpl<MatchScope> &MatchScopes;
3016 
3017 public:
3018   MatchStateUpdater(SelectionDAG &DAG, SDNode **NodeToMatch,
3019                     SmallVectorImpl<std::pair<SDValue, SDNode *>> &RN,
3020                     SmallVectorImpl<MatchScope> &MS)
3021       : SelectionDAG::DAGUpdateListener(DAG), NodeToMatch(NodeToMatch),
3022         RecordedNodes(RN), MatchScopes(MS) {}
3023 
3024   void NodeDeleted(SDNode *N, SDNode *E) override {
3025     // Some early-returns here to avoid the search if we deleted the node or
3026     // if the update comes from MorphNodeTo (MorphNodeTo is the last thing we
3027     // do, so it's unnecessary to update matching state at that point).
3028     // Neither of these can occur currently because we only install this
3029     // update listener during matching a complex patterns.
3030     if (!E || E->isMachineOpcode())
3031       return;
3032     // Check if NodeToMatch was updated.
3033     if (N == *NodeToMatch)
3034       *NodeToMatch = E;
3035     // Performing linear search here does not matter because we almost never
3036     // run this code.  You'd have to have a CSE during complex pattern
3037     // matching.
3038     for (auto &I : RecordedNodes)
3039       if (I.first.getNode() == N)
3040         I.first.setNode(E);
3041 
3042     for (auto &I : MatchScopes)
3043       for (auto &J : I.NodeStack)
3044         if (J.getNode() == N)
3045           J.setNode(E);
3046   }
3047 };
3048 
3049 } // end anonymous namespace
3050 
3051 void SelectionDAGISel::SelectCodeCommon(SDNode *NodeToMatch,
3052                                         const unsigned char *MatcherTable,
3053                                         unsigned TableSize) {
3054   // FIXME: Should these even be selected?  Handle these cases in the caller?
3055   switch (NodeToMatch->getOpcode()) {
3056   default:
3057     break;
3058   case ISD::EntryToken:       // These nodes remain the same.
3059   case ISD::BasicBlock:
3060   case ISD::Register:
3061   case ISD::RegisterMask:
3062   case ISD::HANDLENODE:
3063   case ISD::MDNODE_SDNODE:
3064   case ISD::TargetConstant:
3065   case ISD::TargetConstantFP:
3066   case ISD::TargetConstantPool:
3067   case ISD::TargetFrameIndex:
3068   case ISD::TargetExternalSymbol:
3069   case ISD::MCSymbol:
3070   case ISD::TargetBlockAddress:
3071   case ISD::TargetJumpTable:
3072   case ISD::TargetGlobalTLSAddress:
3073   case ISD::TargetGlobalAddress:
3074   case ISD::TokenFactor:
3075   case ISD::CopyFromReg:
3076   case ISD::CopyToReg:
3077   case ISD::EH_LABEL:
3078   case ISD::ANNOTATION_LABEL:
3079   case ISD::LIFETIME_START:
3080   case ISD::LIFETIME_END:
3081   case ISD::PSEUDO_PROBE:
3082     NodeToMatch->setNodeId(-1); // Mark selected.
3083     return;
3084   case ISD::AssertSext:
3085   case ISD::AssertZext:
3086   case ISD::AssertAlign:
3087     ReplaceUses(SDValue(NodeToMatch, 0), NodeToMatch->getOperand(0));
3088     CurDAG->RemoveDeadNode(NodeToMatch);
3089     return;
3090   case ISD::INLINEASM:
3091   case ISD::INLINEASM_BR:
3092     Select_INLINEASM(NodeToMatch);
3093     return;
3094   case ISD::READ_REGISTER:
3095     Select_READ_REGISTER(NodeToMatch);
3096     return;
3097   case ISD::WRITE_REGISTER:
3098     Select_WRITE_REGISTER(NodeToMatch);
3099     return;
3100   case ISD::UNDEF:
3101     Select_UNDEF(NodeToMatch);
3102     return;
3103   case ISD::FREEZE:
3104     Select_FREEZE(NodeToMatch);
3105     return;
3106   case ISD::ARITH_FENCE:
3107     Select_ARITH_FENCE(NodeToMatch);
3108     return;
3109   case ISD::MEMBARRIER:
3110     Select_MEMBARRIER(NodeToMatch);
3111     return;
3112   case ISD::STACKMAP:
3113     Select_STACKMAP(NodeToMatch);
3114     return;
3115   case ISD::PATCHPOINT:
3116     Select_PATCHPOINT(NodeToMatch);
3117     return;
3118   case ISD::JUMP_TABLE_DEBUG_INFO:
3119     Select_JUMP_TABLE_DEBUG_INFO(NodeToMatch);
3120     return;
3121   }
3122 
3123   assert(!NodeToMatch->isMachineOpcode() && "Node already selected!");
3124 
3125   // Set up the node stack with NodeToMatch as the only node on the stack.
3126   SmallVector<SDValue, 8> NodeStack;
3127   SDValue N = SDValue(NodeToMatch, 0);
3128   NodeStack.push_back(N);
3129 
3130   // MatchScopes - Scopes used when matching, if a match failure happens, this
3131   // indicates where to continue checking.
3132   SmallVector<MatchScope, 8> MatchScopes;
3133 
3134   // RecordedNodes - This is the set of nodes that have been recorded by the
3135   // state machine.  The second value is the parent of the node, or null if the
3136   // root is recorded.
3137   SmallVector<std::pair<SDValue, SDNode*>, 8> RecordedNodes;
3138 
3139   // MatchedMemRefs - This is the set of MemRef's we've seen in the input
3140   // pattern.
3141   SmallVector<MachineMemOperand*, 2> MatchedMemRefs;
3142 
3143   // These are the current input chain and glue for use when generating nodes.
3144   // Various Emit operations change these.  For example, emitting a copytoreg
3145   // uses and updates these.
3146   SDValue InputChain, InputGlue;
3147 
3148   // ChainNodesMatched - If a pattern matches nodes that have input/output
3149   // chains, the OPC_EmitMergeInputChains operation is emitted which indicates
3150   // which ones they are.  The result is captured into this list so that we can
3151   // update the chain results when the pattern is complete.
3152   SmallVector<SDNode*, 3> ChainNodesMatched;
3153 
3154   LLVM_DEBUG(dbgs() << "ISEL: Starting pattern match\n");
3155 
3156   // Determine where to start the interpreter.  Normally we start at opcode #0,
3157   // but if the state machine starts with an OPC_SwitchOpcode, then we
3158   // accelerate the first lookup (which is guaranteed to be hot) with the
3159   // OpcodeOffset table.
3160   unsigned MatcherIndex = 0;
3161 
3162   if (!OpcodeOffset.empty()) {
3163     // Already computed the OpcodeOffset table, just index into it.
3164     if (N.getOpcode() < OpcodeOffset.size())
3165       MatcherIndex = OpcodeOffset[N.getOpcode()];
3166     LLVM_DEBUG(dbgs() << "  Initial Opcode index to " << MatcherIndex << "\n");
3167 
3168   } else if (MatcherTable[0] == OPC_SwitchOpcode) {
3169     // Otherwise, the table isn't computed, but the state machine does start
3170     // with an OPC_SwitchOpcode instruction.  Populate the table now, since this
3171     // is the first time we're selecting an instruction.
3172     unsigned Idx = 1;
3173     while (true) {
3174       // Get the size of this case.
3175       unsigned CaseSize = MatcherTable[Idx++];
3176       if (CaseSize & 128)
3177         CaseSize = GetVBR(CaseSize, MatcherTable, Idx);
3178       if (CaseSize == 0) break;
3179 
3180       // Get the opcode, add the index to the table.
3181       uint16_t Opc = MatcherTable[Idx++];
3182       Opc |= static_cast<uint16_t>(MatcherTable[Idx++]) << 8;
3183       if (Opc >= OpcodeOffset.size())
3184         OpcodeOffset.resize((Opc+1)*2);
3185       OpcodeOffset[Opc] = Idx;
3186       Idx += CaseSize;
3187     }
3188 
3189     // Okay, do the lookup for the first opcode.
3190     if (N.getOpcode() < OpcodeOffset.size())
3191       MatcherIndex = OpcodeOffset[N.getOpcode()];
3192   }
3193 
3194   while (true) {
3195     assert(MatcherIndex < TableSize && "Invalid index");
3196 #ifndef NDEBUG
3197     unsigned CurrentOpcodeIndex = MatcherIndex;
3198 #endif
3199     BuiltinOpcodes Opcode =
3200         static_cast<BuiltinOpcodes>(MatcherTable[MatcherIndex++]);
3201     switch (Opcode) {
3202     case OPC_Scope: {
3203       // Okay, the semantics of this operation are that we should push a scope
3204       // then evaluate the first child.  However, pushing a scope only to have
3205       // the first check fail (which then pops it) is inefficient.  If we can
3206       // determine immediately that the first check (or first several) will
3207       // immediately fail, don't even bother pushing a scope for them.
3208       unsigned FailIndex;
3209 
3210       while (true) {
3211         unsigned NumToSkip = MatcherTable[MatcherIndex++];
3212         if (NumToSkip & 128)
3213           NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
3214         // Found the end of the scope with no match.
3215         if (NumToSkip == 0) {
3216           FailIndex = 0;
3217           break;
3218         }
3219 
3220         FailIndex = MatcherIndex+NumToSkip;
3221 
3222         unsigned MatcherIndexOfPredicate = MatcherIndex;
3223         (void)MatcherIndexOfPredicate; // silence warning.
3224 
3225         // If we can't evaluate this predicate without pushing a scope (e.g. if
3226         // it is a 'MoveParent') or if the predicate succeeds on this node, we
3227         // push the scope and evaluate the full predicate chain.
3228         bool Result;
3229         MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
3230                                               Result, *this, RecordedNodes);
3231         if (!Result)
3232           break;
3233 
3234         LLVM_DEBUG(
3235             dbgs() << "  Skipped scope entry (due to false predicate) at "
3236                    << "index " << MatcherIndexOfPredicate << ", continuing at "
3237                    << FailIndex << "\n");
3238         ++NumDAGIselRetries;
3239 
3240         // Otherwise, we know that this case of the Scope is guaranteed to fail,
3241         // move to the next case.
3242         MatcherIndex = FailIndex;
3243       }
3244 
3245       // If the whole scope failed to match, bail.
3246       if (FailIndex == 0) break;
3247 
3248       // Push a MatchScope which indicates where to go if the first child fails
3249       // to match.
3250       MatchScope NewEntry;
3251       NewEntry.FailIndex = FailIndex;
3252       NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
3253       NewEntry.NumRecordedNodes = RecordedNodes.size();
3254       NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
3255       NewEntry.InputChain = InputChain;
3256       NewEntry.InputGlue = InputGlue;
3257       NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty();
3258       MatchScopes.push_back(NewEntry);
3259       continue;
3260     }
3261     case OPC_RecordNode: {
3262       // Remember this node, it may end up being an operand in the pattern.
3263       SDNode *Parent = nullptr;
3264       if (NodeStack.size() > 1)
3265         Parent = NodeStack[NodeStack.size()-2].getNode();
3266       RecordedNodes.push_back(std::make_pair(N, Parent));
3267       continue;
3268     }
3269 
3270     case OPC_RecordChild0: case OPC_RecordChild1:
3271     case OPC_RecordChild2: case OPC_RecordChild3:
3272     case OPC_RecordChild4: case OPC_RecordChild5:
3273     case OPC_RecordChild6: case OPC_RecordChild7: {
3274       unsigned ChildNo = Opcode-OPC_RecordChild0;
3275       if (ChildNo >= N.getNumOperands())
3276         break;  // Match fails if out of range child #.
3277 
3278       RecordedNodes.push_back(std::make_pair(N->getOperand(ChildNo),
3279                                              N.getNode()));
3280       continue;
3281     }
3282     case OPC_RecordMemRef:
3283       if (auto *MN = dyn_cast<MemSDNode>(N))
3284         MatchedMemRefs.push_back(MN->getMemOperand());
3285       else {
3286         LLVM_DEBUG(dbgs() << "Expected MemSDNode "; N->dump(CurDAG);
3287                    dbgs() << '\n');
3288       }
3289 
3290       continue;
3291 
3292     case OPC_CaptureGlueInput:
3293       // If the current node has an input glue, capture it in InputGlue.
3294       if (N->getNumOperands() != 0 &&
3295           N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue)
3296         InputGlue = N->getOperand(N->getNumOperands()-1);
3297       continue;
3298 
3299     case OPC_MoveChild: {
3300       unsigned ChildNo = MatcherTable[MatcherIndex++];
3301       if (ChildNo >= N.getNumOperands())
3302         break;  // Match fails if out of range child #.
3303       N = N.getOperand(ChildNo);
3304       NodeStack.push_back(N);
3305       continue;
3306     }
3307 
3308     case OPC_MoveChild0: case OPC_MoveChild1:
3309     case OPC_MoveChild2: case OPC_MoveChild3:
3310     case OPC_MoveChild4: case OPC_MoveChild5:
3311     case OPC_MoveChild6: case OPC_MoveChild7: {
3312       unsigned ChildNo = Opcode-OPC_MoveChild0;
3313       if (ChildNo >= N.getNumOperands())
3314         break;  // Match fails if out of range child #.
3315       N = N.getOperand(ChildNo);
3316       NodeStack.push_back(N);
3317       continue;
3318     }
3319 
3320     case OPC_MoveSibling:
3321     case OPC_MoveSibling0:
3322     case OPC_MoveSibling1:
3323     case OPC_MoveSibling2:
3324     case OPC_MoveSibling3:
3325     case OPC_MoveSibling4:
3326     case OPC_MoveSibling5:
3327     case OPC_MoveSibling6:
3328     case OPC_MoveSibling7: {
3329       // Pop the current node off the NodeStack.
3330       NodeStack.pop_back();
3331       assert(!NodeStack.empty() && "Node stack imbalance!");
3332       N = NodeStack.back();
3333 
3334       unsigned SiblingNo = Opcode == OPC_MoveSibling
3335                                ? MatcherTable[MatcherIndex++]
3336                                : Opcode - OPC_MoveSibling0;
3337       if (SiblingNo >= N.getNumOperands())
3338         break; // Match fails if out of range sibling #.
3339       N = N.getOperand(SiblingNo);
3340       NodeStack.push_back(N);
3341       continue;
3342     }
3343     case OPC_MoveParent:
3344       // Pop the current node off the NodeStack.
3345       NodeStack.pop_back();
3346       assert(!NodeStack.empty() && "Node stack imbalance!");
3347       N = NodeStack.back();
3348       continue;
3349 
3350     case OPC_CheckSame:
3351       if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
3352       continue;
3353 
3354     case OPC_CheckChild0Same: case OPC_CheckChild1Same:
3355     case OPC_CheckChild2Same: case OPC_CheckChild3Same:
3356       if (!::CheckChildSame(MatcherTable, MatcherIndex, N, RecordedNodes,
3357                             Opcode-OPC_CheckChild0Same))
3358         break;
3359       continue;
3360 
3361     case OPC_CheckPatternPredicate:
3362     case OPC_CheckPatternPredicate0:
3363     case OPC_CheckPatternPredicate1:
3364     case OPC_CheckPatternPredicate2:
3365     case OPC_CheckPatternPredicate3:
3366     case OPC_CheckPatternPredicate4:
3367     case OPC_CheckPatternPredicate5:
3368     case OPC_CheckPatternPredicate6:
3369     case OPC_CheckPatternPredicate7:
3370     case OPC_CheckPatternPredicateTwoByte:
3371       if (!::CheckPatternPredicate(Opcode, MatcherTable, MatcherIndex, *this))
3372         break;
3373       continue;
3374     case SelectionDAGISel::OPC_CheckPredicate0:
3375     case SelectionDAGISel::OPC_CheckPredicate1:
3376     case SelectionDAGISel::OPC_CheckPredicate2:
3377     case SelectionDAGISel::OPC_CheckPredicate3:
3378     case SelectionDAGISel::OPC_CheckPredicate4:
3379     case SelectionDAGISel::OPC_CheckPredicate5:
3380     case SelectionDAGISel::OPC_CheckPredicate6:
3381     case SelectionDAGISel::OPC_CheckPredicate7:
3382     case OPC_CheckPredicate:
3383       if (!::CheckNodePredicate(Opcode, MatcherTable, MatcherIndex, *this,
3384                                 N.getNode()))
3385         break;
3386       continue;
3387     case OPC_CheckPredicateWithOperands: {
3388       unsigned OpNum = MatcherTable[MatcherIndex++];
3389       SmallVector<SDValue, 8> Operands;
3390 
3391       for (unsigned i = 0; i < OpNum; ++i)
3392         Operands.push_back(RecordedNodes[MatcherTable[MatcherIndex++]].first);
3393 
3394       unsigned PredNo = MatcherTable[MatcherIndex++];
3395       if (!CheckNodePredicateWithOperands(N.getNode(), PredNo, Operands))
3396         break;
3397       continue;
3398     }
3399     case OPC_CheckComplexPat:
3400     case OPC_CheckComplexPat0:
3401     case OPC_CheckComplexPat1:
3402     case OPC_CheckComplexPat2:
3403     case OPC_CheckComplexPat3:
3404     case OPC_CheckComplexPat4:
3405     case OPC_CheckComplexPat5:
3406     case OPC_CheckComplexPat6:
3407     case OPC_CheckComplexPat7: {
3408       unsigned CPNum = Opcode == OPC_CheckComplexPat
3409                            ? MatcherTable[MatcherIndex++]
3410                            : Opcode - OPC_CheckComplexPat0;
3411       unsigned RecNo = MatcherTable[MatcherIndex++];
3412       assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
3413 
3414       // If target can modify DAG during matching, keep the matching state
3415       // consistent.
3416       std::unique_ptr<MatchStateUpdater> MSU;
3417       if (ComplexPatternFuncMutatesDAG())
3418         MSU.reset(new MatchStateUpdater(*CurDAG, &NodeToMatch, RecordedNodes,
3419                                         MatchScopes));
3420 
3421       if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo].second,
3422                                RecordedNodes[RecNo].first, CPNum,
3423                                RecordedNodes))
3424         break;
3425       continue;
3426     }
3427     case OPC_CheckOpcode:
3428       if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
3429       continue;
3430 
3431     case OPC_CheckType:
3432     case OPC_CheckTypeI32:
3433     case OPC_CheckTypeI64:
3434       MVT::SimpleValueType VT;
3435       switch (Opcode) {
3436       case OPC_CheckTypeI32:
3437         VT = MVT::i32;
3438         break;
3439       case OPC_CheckTypeI64:
3440         VT = MVT::i64;
3441         break;
3442       default:
3443         VT = static_cast<MVT::SimpleValueType>(MatcherTable[MatcherIndex++]);
3444         break;
3445       }
3446       if (!::CheckType(VT, N, TLI, CurDAG->getDataLayout()))
3447         break;
3448       continue;
3449 
3450     case OPC_CheckTypeRes: {
3451       unsigned Res = MatcherTable[MatcherIndex++];
3452       if (!::CheckType(
3453               static_cast<MVT::SimpleValueType>(MatcherTable[MatcherIndex++]),
3454               N.getValue(Res), TLI, CurDAG->getDataLayout()))
3455         break;
3456       continue;
3457     }
3458 
3459     case OPC_SwitchOpcode: {
3460       unsigned CurNodeOpcode = N.getOpcode();
3461       unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
3462       unsigned CaseSize;
3463       while (true) {
3464         // Get the size of this case.
3465         CaseSize = MatcherTable[MatcherIndex++];
3466         if (CaseSize & 128)
3467           CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
3468         if (CaseSize == 0) break;
3469 
3470         uint16_t Opc = MatcherTable[MatcherIndex++];
3471         Opc |= static_cast<uint16_t>(MatcherTable[MatcherIndex++]) << 8;
3472 
3473         // If the opcode matches, then we will execute this case.
3474         if (CurNodeOpcode == Opc)
3475           break;
3476 
3477         // Otherwise, skip over this case.
3478         MatcherIndex += CaseSize;
3479       }
3480 
3481       // If no cases matched, bail out.
3482       if (CaseSize == 0) break;
3483 
3484       // Otherwise, execute the case we found.
3485       LLVM_DEBUG(dbgs() << "  OpcodeSwitch from " << SwitchStart << " to "
3486                         << MatcherIndex << "\n");
3487       continue;
3488     }
3489 
3490     case OPC_SwitchType: {
3491       MVT CurNodeVT = N.getSimpleValueType();
3492       unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
3493       unsigned CaseSize;
3494       while (true) {
3495         // Get the size of this case.
3496         CaseSize = MatcherTable[MatcherIndex++];
3497         if (CaseSize & 128)
3498           CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
3499         if (CaseSize == 0) break;
3500 
3501         MVT CaseVT =
3502             static_cast<MVT::SimpleValueType>(MatcherTable[MatcherIndex++]);
3503         if (CaseVT == MVT::iPTR)
3504           CaseVT = TLI->getPointerTy(CurDAG->getDataLayout());
3505 
3506         // If the VT matches, then we will execute this case.
3507         if (CurNodeVT == CaseVT)
3508           break;
3509 
3510         // Otherwise, skip over this case.
3511         MatcherIndex += CaseSize;
3512       }
3513 
3514       // If no cases matched, bail out.
3515       if (CaseSize == 0) break;
3516 
3517       // Otherwise, execute the case we found.
3518       LLVM_DEBUG(dbgs() << "  TypeSwitch[" << CurNodeVT
3519                         << "] from " << SwitchStart << " to " << MatcherIndex
3520                         << '\n');
3521       continue;
3522     }
3523     case OPC_CheckChild0Type:
3524     case OPC_CheckChild1Type:
3525     case OPC_CheckChild2Type:
3526     case OPC_CheckChild3Type:
3527     case OPC_CheckChild4Type:
3528     case OPC_CheckChild5Type:
3529     case OPC_CheckChild6Type:
3530     case OPC_CheckChild7Type:
3531     case OPC_CheckChild0TypeI32:
3532     case OPC_CheckChild1TypeI32:
3533     case OPC_CheckChild2TypeI32:
3534     case OPC_CheckChild3TypeI32:
3535     case OPC_CheckChild4TypeI32:
3536     case OPC_CheckChild5TypeI32:
3537     case OPC_CheckChild6TypeI32:
3538     case OPC_CheckChild7TypeI32:
3539     case OPC_CheckChild0TypeI64:
3540     case OPC_CheckChild1TypeI64:
3541     case OPC_CheckChild2TypeI64:
3542     case OPC_CheckChild3TypeI64:
3543     case OPC_CheckChild4TypeI64:
3544     case OPC_CheckChild5TypeI64:
3545     case OPC_CheckChild6TypeI64:
3546     case OPC_CheckChild7TypeI64: {
3547       MVT::SimpleValueType VT;
3548       unsigned ChildNo;
3549       if (Opcode >= SelectionDAGISel::OPC_CheckChild0TypeI32 &&
3550           Opcode <= SelectionDAGISel::OPC_CheckChild7TypeI32) {
3551         VT = MVT::i32;
3552         ChildNo = Opcode - SelectionDAGISel::OPC_CheckChild0TypeI32;
3553       } else if (Opcode >= SelectionDAGISel::OPC_CheckChild0TypeI64 &&
3554                  Opcode <= SelectionDAGISel::OPC_CheckChild7TypeI64) {
3555         VT = MVT::i64;
3556         ChildNo = Opcode - SelectionDAGISel::OPC_CheckChild0TypeI64;
3557       } else {
3558         VT = static_cast<MVT::SimpleValueType>(MatcherTable[MatcherIndex++]);
3559         ChildNo = Opcode - SelectionDAGISel::OPC_CheckChild0Type;
3560       }
3561       if (!::CheckChildType(VT, N, TLI, CurDAG->getDataLayout(), ChildNo))
3562         break;
3563       continue;
3564     }
3565     case OPC_CheckCondCode:
3566       if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
3567       continue;
3568     case OPC_CheckChild2CondCode:
3569       if (!::CheckChild2CondCode(MatcherTable, MatcherIndex, N)) break;
3570       continue;
3571     case OPC_CheckValueType:
3572       if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI,
3573                             CurDAG->getDataLayout()))
3574         break;
3575       continue;
3576     case OPC_CheckInteger:
3577       if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
3578       continue;
3579     case OPC_CheckChild0Integer: case OPC_CheckChild1Integer:
3580     case OPC_CheckChild2Integer: case OPC_CheckChild3Integer:
3581     case OPC_CheckChild4Integer:
3582       if (!::CheckChildInteger(MatcherTable, MatcherIndex, N,
3583                                Opcode-OPC_CheckChild0Integer)) break;
3584       continue;
3585     case OPC_CheckAndImm:
3586       if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
3587       continue;
3588     case OPC_CheckOrImm:
3589       if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
3590       continue;
3591     case OPC_CheckImmAllOnesV:
3592       if (!ISD::isConstantSplatVectorAllOnes(N.getNode()))
3593         break;
3594       continue;
3595     case OPC_CheckImmAllZerosV:
3596       if (!ISD::isConstantSplatVectorAllZeros(N.getNode()))
3597         break;
3598       continue;
3599 
3600     case OPC_CheckFoldableChainNode: {
3601       assert(NodeStack.size() != 1 && "No parent node");
3602       // Verify that all intermediate nodes between the root and this one have
3603       // a single use (ignoring chains, which are handled in UpdateChains).
3604       bool HasMultipleUses = false;
3605       for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i) {
3606         unsigned NNonChainUses = 0;
3607         SDNode *NS = NodeStack[i].getNode();
3608         for (auto UI = NS->use_begin(), UE = NS->use_end(); UI != UE; ++UI)
3609           if (UI.getUse().getValueType() != MVT::Other)
3610             if (++NNonChainUses > 1) {
3611               HasMultipleUses = true;
3612               break;
3613             }
3614         if (HasMultipleUses) break;
3615       }
3616       if (HasMultipleUses) break;
3617 
3618       // Check to see that the target thinks this is profitable to fold and that
3619       // we can fold it without inducing cycles in the graph.
3620       if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
3621                               NodeToMatch) ||
3622           !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
3623                          NodeToMatch, OptLevel,
3624                          true/*We validate our own chains*/))
3625         break;
3626 
3627       continue;
3628     }
3629     case OPC_EmitInteger:
3630     case OPC_EmitInteger8:
3631     case OPC_EmitInteger16:
3632     case OPC_EmitInteger32:
3633     case OPC_EmitInteger64:
3634     case OPC_EmitStringInteger:
3635     case OPC_EmitStringInteger32: {
3636       MVT::SimpleValueType VT;
3637       switch (Opcode) {
3638       case OPC_EmitInteger8:
3639         VT = MVT::i8;
3640         break;
3641       case OPC_EmitInteger16:
3642         VT = MVT::i16;
3643         break;
3644       case OPC_EmitInteger32:
3645       case OPC_EmitStringInteger32:
3646         VT = MVT::i32;
3647         break;
3648       case OPC_EmitInteger64:
3649         VT = MVT::i64;
3650         break;
3651       default:
3652         VT = static_cast<MVT::SimpleValueType>(MatcherTable[MatcherIndex++]);
3653         break;
3654       }
3655       int64_t Val = MatcherTable[MatcherIndex++];
3656       if (Val & 128)
3657         Val = GetVBR(Val, MatcherTable, MatcherIndex);
3658       if (Opcode >= OPC_EmitInteger && Opcode <= OPC_EmitInteger64)
3659         Val = decodeSignRotatedValue(Val);
3660       RecordedNodes.push_back(std::pair<SDValue, SDNode *>(
3661           CurDAG->getTargetConstant(Val, SDLoc(NodeToMatch), VT), nullptr));
3662       continue;
3663     }
3664     case OPC_EmitRegister:
3665     case OPC_EmitRegisterI32:
3666     case OPC_EmitRegisterI64: {
3667       MVT::SimpleValueType VT;
3668       switch (Opcode) {
3669       case OPC_EmitRegisterI32:
3670         VT = MVT::i32;
3671         break;
3672       case OPC_EmitRegisterI64:
3673         VT = MVT::i64;
3674         break;
3675       default:
3676         VT = static_cast<MVT::SimpleValueType>(MatcherTable[MatcherIndex++]);
3677         break;
3678       }
3679       unsigned RegNo = MatcherTable[MatcherIndex++];
3680       RecordedNodes.push_back(std::pair<SDValue, SDNode *>(
3681           CurDAG->getRegister(RegNo, VT), nullptr));
3682       continue;
3683     }
3684     case OPC_EmitRegister2: {
3685       // For targets w/ more than 256 register names, the register enum
3686       // values are stored in two bytes in the matcher table (just like
3687       // opcodes).
3688       MVT::SimpleValueType VT =
3689           static_cast<MVT::SimpleValueType>(MatcherTable[MatcherIndex++]);
3690       unsigned RegNo = MatcherTable[MatcherIndex++];
3691       RegNo |= MatcherTable[MatcherIndex++] << 8;
3692       RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
3693                               CurDAG->getRegister(RegNo, VT), nullptr));
3694       continue;
3695     }
3696 
3697     case OPC_EmitConvertToTarget:
3698     case OPC_EmitConvertToTarget0:
3699     case OPC_EmitConvertToTarget1:
3700     case OPC_EmitConvertToTarget2:
3701     case OPC_EmitConvertToTarget3:
3702     case OPC_EmitConvertToTarget4:
3703     case OPC_EmitConvertToTarget5:
3704     case OPC_EmitConvertToTarget6:
3705     case OPC_EmitConvertToTarget7: {
3706       // Convert from IMM/FPIMM to target version.
3707       unsigned RecNo = Opcode == OPC_EmitConvertToTarget
3708                            ? MatcherTable[MatcherIndex++]
3709                            : Opcode - OPC_EmitConvertToTarget0;
3710       assert(RecNo < RecordedNodes.size() && "Invalid EmitConvertToTarget");
3711       SDValue Imm = RecordedNodes[RecNo].first;
3712 
3713       if (Imm->getOpcode() == ISD::Constant) {
3714         const ConstantInt *Val=cast<ConstantSDNode>(Imm)->getConstantIntValue();
3715         Imm = CurDAG->getTargetConstant(*Val, SDLoc(NodeToMatch),
3716                                         Imm.getValueType());
3717       } else if (Imm->getOpcode() == ISD::ConstantFP) {
3718         const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue();
3719         Imm = CurDAG->getTargetConstantFP(*Val, SDLoc(NodeToMatch),
3720                                           Imm.getValueType());
3721       }
3722 
3723       RecordedNodes.push_back(std::make_pair(Imm, RecordedNodes[RecNo].second));
3724       continue;
3725     }
3726 
3727     case OPC_EmitMergeInputChains1_0:    // OPC_EmitMergeInputChains, 1, 0
3728     case OPC_EmitMergeInputChains1_1:    // OPC_EmitMergeInputChains, 1, 1
3729     case OPC_EmitMergeInputChains1_2: {  // OPC_EmitMergeInputChains, 1, 2
3730       // These are space-optimized forms of OPC_EmitMergeInputChains.
3731       assert(!InputChain.getNode() &&
3732              "EmitMergeInputChains should be the first chain producing node");
3733       assert(ChainNodesMatched.empty() &&
3734              "Should only have one EmitMergeInputChains per match");
3735 
3736       // Read all of the chained nodes.
3737       unsigned RecNo = Opcode - OPC_EmitMergeInputChains1_0;
3738       assert(RecNo < RecordedNodes.size() && "Invalid EmitMergeInputChains");
3739       ChainNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
3740 
3741       // If the chained node is not the root, we can't fold it if it has
3742       // multiple uses.
3743       // FIXME: What if other value results of the node have uses not matched
3744       // by this pattern?
3745       if (ChainNodesMatched.back() != NodeToMatch &&
3746           !RecordedNodes[RecNo].first.hasOneUse()) {
3747         ChainNodesMatched.clear();
3748         break;
3749       }
3750 
3751       // Merge the input chains if they are not intra-pattern references.
3752       InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
3753 
3754       if (!InputChain.getNode())
3755         break;  // Failed to merge.
3756       continue;
3757     }
3758 
3759     case OPC_EmitMergeInputChains: {
3760       assert(!InputChain.getNode() &&
3761              "EmitMergeInputChains should be the first chain producing node");
3762       // This node gets a list of nodes we matched in the input that have
3763       // chains.  We want to token factor all of the input chains to these nodes
3764       // together.  However, if any of the input chains is actually one of the
3765       // nodes matched in this pattern, then we have an intra-match reference.
3766       // Ignore these because the newly token factored chain should not refer to
3767       // the old nodes.
3768       unsigned NumChains = MatcherTable[MatcherIndex++];
3769       assert(NumChains != 0 && "Can't TF zero chains");
3770 
3771       assert(ChainNodesMatched.empty() &&
3772              "Should only have one EmitMergeInputChains per match");
3773 
3774       // Read all of the chained nodes.
3775       for (unsigned i = 0; i != NumChains; ++i) {
3776         unsigned RecNo = MatcherTable[MatcherIndex++];
3777         assert(RecNo < RecordedNodes.size() && "Invalid EmitMergeInputChains");
3778         ChainNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
3779 
3780         // If the chained node is not the root, we can't fold it if it has
3781         // multiple uses.
3782         // FIXME: What if other value results of the node have uses not matched
3783         // by this pattern?
3784         if (ChainNodesMatched.back() != NodeToMatch &&
3785             !RecordedNodes[RecNo].first.hasOneUse()) {
3786           ChainNodesMatched.clear();
3787           break;
3788         }
3789       }
3790 
3791       // If the inner loop broke out, the match fails.
3792       if (ChainNodesMatched.empty())
3793         break;
3794 
3795       // Merge the input chains if they are not intra-pattern references.
3796       InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
3797 
3798       if (!InputChain.getNode())
3799         break;  // Failed to merge.
3800 
3801       continue;
3802     }
3803 
3804     case OPC_EmitCopyToReg:
3805     case OPC_EmitCopyToReg0:
3806     case OPC_EmitCopyToReg1:
3807     case OPC_EmitCopyToReg2:
3808     case OPC_EmitCopyToReg3:
3809     case OPC_EmitCopyToReg4:
3810     case OPC_EmitCopyToReg5:
3811     case OPC_EmitCopyToReg6:
3812     case OPC_EmitCopyToReg7:
3813     case OPC_EmitCopyToRegTwoByte: {
3814       unsigned RecNo =
3815           Opcode >= OPC_EmitCopyToReg0 && Opcode <= OPC_EmitCopyToReg7
3816               ? Opcode - OPC_EmitCopyToReg0
3817               : MatcherTable[MatcherIndex++];
3818       assert(RecNo < RecordedNodes.size() && "Invalid EmitCopyToReg");
3819       unsigned DestPhysReg = MatcherTable[MatcherIndex++];
3820       if (Opcode == OPC_EmitCopyToRegTwoByte)
3821         DestPhysReg |= MatcherTable[MatcherIndex++] << 8;
3822 
3823       if (!InputChain.getNode())
3824         InputChain = CurDAG->getEntryNode();
3825 
3826       InputChain = CurDAG->getCopyToReg(InputChain, SDLoc(NodeToMatch),
3827                                         DestPhysReg, RecordedNodes[RecNo].first,
3828                                         InputGlue);
3829 
3830       InputGlue = InputChain.getValue(1);
3831       continue;
3832     }
3833 
3834     case OPC_EmitNodeXForm: {
3835       unsigned XFormNo = MatcherTable[MatcherIndex++];
3836       unsigned RecNo = MatcherTable[MatcherIndex++];
3837       assert(RecNo < RecordedNodes.size() && "Invalid EmitNodeXForm");
3838       SDValue Res = RunSDNodeXForm(RecordedNodes[RecNo].first, XFormNo);
3839       RecordedNodes.push_back(std::pair<SDValue,SDNode*>(Res, nullptr));
3840       continue;
3841     }
3842     case OPC_Coverage: {
3843       // This is emitted right before MorphNode/EmitNode.
3844       // So it should be safe to assume that this node has been selected
3845       unsigned index = MatcherTable[MatcherIndex++];
3846       index |= (MatcherTable[MatcherIndex++] << 8);
3847       dbgs() << "COVERED: " << getPatternForIndex(index) << "\n";
3848       dbgs() << "INCLUDED: " << getIncludePathForIndex(index) << "\n";
3849       continue;
3850     }
3851 
3852     case OPC_EmitNode:
3853     case OPC_EmitNode0:
3854     case OPC_EmitNode1:
3855     case OPC_EmitNode2:
3856     case OPC_EmitNode0None:
3857     case OPC_EmitNode1None:
3858     case OPC_EmitNode2None:
3859     case OPC_EmitNode0Chain:
3860     case OPC_EmitNode1Chain:
3861     case OPC_EmitNode2Chain:
3862     case OPC_MorphNodeTo:
3863     case OPC_MorphNodeTo0:
3864     case OPC_MorphNodeTo1:
3865     case OPC_MorphNodeTo2:
3866     case OPC_MorphNodeTo0None:
3867     case OPC_MorphNodeTo1None:
3868     case OPC_MorphNodeTo2None:
3869     case OPC_MorphNodeTo0Chain:
3870     case OPC_MorphNodeTo1Chain:
3871     case OPC_MorphNodeTo2Chain:
3872     case OPC_MorphNodeTo0GlueInput:
3873     case OPC_MorphNodeTo1GlueInput:
3874     case OPC_MorphNodeTo2GlueInput:
3875     case OPC_MorphNodeTo0GlueOutput:
3876     case OPC_MorphNodeTo1GlueOutput:
3877     case OPC_MorphNodeTo2GlueOutput: {
3878       uint16_t TargetOpc = MatcherTable[MatcherIndex++];
3879       TargetOpc |= static_cast<uint16_t>(MatcherTable[MatcherIndex++]) << 8;
3880       unsigned EmitNodeInfo;
3881       if (Opcode >= OPC_EmitNode0None && Opcode <= OPC_EmitNode2Chain) {
3882         if (Opcode >= OPC_EmitNode0Chain && Opcode <= OPC_EmitNode2Chain)
3883           EmitNodeInfo = OPFL_Chain;
3884         else
3885           EmitNodeInfo = OPFL_None;
3886       } else if (Opcode >= OPC_MorphNodeTo0None &&
3887                  Opcode <= OPC_MorphNodeTo2GlueOutput) {
3888         if (Opcode >= OPC_MorphNodeTo0Chain && Opcode <= OPC_MorphNodeTo2Chain)
3889           EmitNodeInfo = OPFL_Chain;
3890         else if (Opcode >= OPC_MorphNodeTo0GlueInput &&
3891                  Opcode <= OPC_MorphNodeTo2GlueInput)
3892           EmitNodeInfo = OPFL_GlueInput;
3893         else if (Opcode >= OPC_MorphNodeTo0GlueOutput &&
3894                  Opcode <= OPC_MorphNodeTo2GlueOutput)
3895           EmitNodeInfo = OPFL_GlueOutput;
3896         else
3897           EmitNodeInfo = OPFL_None;
3898       } else
3899         EmitNodeInfo = MatcherTable[MatcherIndex++];
3900       // Get the result VT list.
3901       unsigned NumVTs;
3902       // If this is one of the compressed forms, get the number of VTs based
3903       // on the Opcode. Otherwise read the next byte from the table.
3904       if (Opcode >= OPC_MorphNodeTo0 && Opcode <= OPC_MorphNodeTo2)
3905         NumVTs = Opcode - OPC_MorphNodeTo0;
3906       else if (Opcode >= OPC_MorphNodeTo0None && Opcode <= OPC_MorphNodeTo2None)
3907         NumVTs = Opcode - OPC_MorphNodeTo0None;
3908       else if (Opcode >= OPC_MorphNodeTo0Chain &&
3909                Opcode <= OPC_MorphNodeTo2Chain)
3910         NumVTs = Opcode - OPC_MorphNodeTo0Chain;
3911       else if (Opcode >= OPC_MorphNodeTo0GlueInput &&
3912                Opcode <= OPC_MorphNodeTo2GlueInput)
3913         NumVTs = Opcode - OPC_MorphNodeTo0GlueInput;
3914       else if (Opcode >= OPC_MorphNodeTo0GlueOutput &&
3915                Opcode <= OPC_MorphNodeTo2GlueOutput)
3916         NumVTs = Opcode - OPC_MorphNodeTo0GlueOutput;
3917       else if (Opcode >= OPC_EmitNode0 && Opcode <= OPC_EmitNode2)
3918         NumVTs = Opcode - OPC_EmitNode0;
3919       else if (Opcode >= OPC_EmitNode0None && Opcode <= OPC_EmitNode2None)
3920         NumVTs = Opcode - OPC_EmitNode0None;
3921       else if (Opcode >= OPC_EmitNode0Chain && Opcode <= OPC_EmitNode2Chain)
3922         NumVTs = Opcode - OPC_EmitNode0Chain;
3923       else
3924         NumVTs = MatcherTable[MatcherIndex++];
3925       SmallVector<EVT, 4> VTs;
3926       for (unsigned i = 0; i != NumVTs; ++i) {
3927         MVT::SimpleValueType VT =
3928             static_cast<MVT::SimpleValueType>(MatcherTable[MatcherIndex++]);
3929         if (VT == MVT::iPTR)
3930           VT = TLI->getPointerTy(CurDAG->getDataLayout()).SimpleTy;
3931         VTs.push_back(VT);
3932       }
3933 
3934       if (EmitNodeInfo & OPFL_Chain)
3935         VTs.push_back(MVT::Other);
3936       if (EmitNodeInfo & OPFL_GlueOutput)
3937         VTs.push_back(MVT::Glue);
3938 
3939       // This is hot code, so optimize the two most common cases of 1 and 2
3940       // results.
3941       SDVTList VTList;
3942       if (VTs.size() == 1)
3943         VTList = CurDAG->getVTList(VTs[0]);
3944       else if (VTs.size() == 2)
3945         VTList = CurDAG->getVTList(VTs[0], VTs[1]);
3946       else
3947         VTList = CurDAG->getVTList(VTs);
3948 
3949       // Get the operand list.
3950       unsigned NumOps = MatcherTable[MatcherIndex++];
3951       SmallVector<SDValue, 8> Ops;
3952       for (unsigned i = 0; i != NumOps; ++i) {
3953         unsigned RecNo = MatcherTable[MatcherIndex++];
3954         if (RecNo & 128)
3955           RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
3956 
3957         assert(RecNo < RecordedNodes.size() && "Invalid EmitNode");
3958         Ops.push_back(RecordedNodes[RecNo].first);
3959       }
3960 
3961       // If there are variadic operands to add, handle them now.
3962       if (EmitNodeInfo & OPFL_VariadicInfo) {
3963         // Determine the start index to copy from.
3964         unsigned FirstOpToCopy = getNumFixedFromVariadicInfo(EmitNodeInfo);
3965         FirstOpToCopy += (EmitNodeInfo & OPFL_Chain) ? 1 : 0;
3966         assert(NodeToMatch->getNumOperands() >= FirstOpToCopy &&
3967                "Invalid variadic node");
3968         // Copy all of the variadic operands, not including a potential glue
3969         // input.
3970         for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands();
3971              i != e; ++i) {
3972           SDValue V = NodeToMatch->getOperand(i);
3973           if (V.getValueType() == MVT::Glue) break;
3974           Ops.push_back(V);
3975         }
3976       }
3977 
3978       // If this has chain/glue inputs, add them.
3979       if (EmitNodeInfo & OPFL_Chain)
3980         Ops.push_back(InputChain);
3981       if ((EmitNodeInfo & OPFL_GlueInput) && InputGlue.getNode() != nullptr)
3982         Ops.push_back(InputGlue);
3983 
3984       // Check whether any matched node could raise an FP exception.  Since all
3985       // such nodes must have a chain, it suffices to check ChainNodesMatched.
3986       // We need to perform this check before potentially modifying one of the
3987       // nodes via MorphNode.
3988       bool MayRaiseFPException =
3989           llvm::any_of(ChainNodesMatched, [this](SDNode *N) {
3990             return mayRaiseFPException(N) && !N->getFlags().hasNoFPExcept();
3991           });
3992 
3993       // Create the node.
3994       MachineSDNode *Res = nullptr;
3995       bool IsMorphNodeTo =
3996           Opcode == OPC_MorphNodeTo ||
3997           (Opcode >= OPC_MorphNodeTo0 && Opcode <= OPC_MorphNodeTo2GlueOutput);
3998       if (!IsMorphNodeTo) {
3999         // If this is a normal EmitNode command, just create the new node and
4000         // add the results to the RecordedNodes list.
4001         Res = CurDAG->getMachineNode(TargetOpc, SDLoc(NodeToMatch),
4002                                      VTList, Ops);
4003 
4004         // Add all the non-glue/non-chain results to the RecordedNodes list.
4005         for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
4006           if (VTs[i] == MVT::Other || VTs[i] == MVT::Glue) break;
4007           RecordedNodes.push_back(std::pair<SDValue,SDNode*>(SDValue(Res, i),
4008                                                              nullptr));
4009         }
4010       } else {
4011         assert(NodeToMatch->getOpcode() != ISD::DELETED_NODE &&
4012                "NodeToMatch was removed partway through selection");
4013         SelectionDAG::DAGNodeDeletedListener NDL(*CurDAG, [&](SDNode *N,
4014                                                               SDNode *E) {
4015           CurDAG->salvageDebugInfo(*N);
4016           auto &Chain = ChainNodesMatched;
4017           assert((!E || !is_contained(Chain, N)) &&
4018                  "Chain node replaced during MorphNode");
4019           llvm::erase(Chain, N);
4020         });
4021         Res = cast<MachineSDNode>(MorphNode(NodeToMatch, TargetOpc, VTList,
4022                                             Ops, EmitNodeInfo));
4023       }
4024 
4025       // Set the NoFPExcept flag when no original matched node could
4026       // raise an FP exception, but the new node potentially might.
4027       if (!MayRaiseFPException && mayRaiseFPException(Res)) {
4028         SDNodeFlags Flags = Res->getFlags();
4029         Flags.setNoFPExcept(true);
4030         Res->setFlags(Flags);
4031       }
4032 
4033       // If the node had chain/glue results, update our notion of the current
4034       // chain and glue.
4035       if (EmitNodeInfo & OPFL_GlueOutput) {
4036         InputGlue = SDValue(Res, VTs.size()-1);
4037         if (EmitNodeInfo & OPFL_Chain)
4038           InputChain = SDValue(Res, VTs.size()-2);
4039       } else if (EmitNodeInfo & OPFL_Chain)
4040         InputChain = SDValue(Res, VTs.size()-1);
4041 
4042       // If the OPFL_MemRefs glue is set on this node, slap all of the
4043       // accumulated memrefs onto it.
4044       //
4045       // FIXME: This is vastly incorrect for patterns with multiple outputs
4046       // instructions that access memory and for ComplexPatterns that match
4047       // loads.
4048       if (EmitNodeInfo & OPFL_MemRefs) {
4049         // Only attach load or store memory operands if the generated
4050         // instruction may load or store.
4051         const MCInstrDesc &MCID = TII->get(TargetOpc);
4052         bool mayLoad = MCID.mayLoad();
4053         bool mayStore = MCID.mayStore();
4054 
4055         // We expect to have relatively few of these so just filter them into a
4056         // temporary buffer so that we can easily add them to the instruction.
4057         SmallVector<MachineMemOperand *, 4> FilteredMemRefs;
4058         for (MachineMemOperand *MMO : MatchedMemRefs) {
4059           if (MMO->isLoad()) {
4060             if (mayLoad)
4061               FilteredMemRefs.push_back(MMO);
4062           } else if (MMO->isStore()) {
4063             if (mayStore)
4064               FilteredMemRefs.push_back(MMO);
4065           } else {
4066             FilteredMemRefs.push_back(MMO);
4067           }
4068         }
4069 
4070         CurDAG->setNodeMemRefs(Res, FilteredMemRefs);
4071       }
4072 
4073       LLVM_DEBUG(if (!MatchedMemRefs.empty() && Res->memoperands_empty()) dbgs()
4074                      << "  Dropping mem operands\n";
4075                  dbgs() << "  " << (IsMorphNodeTo ? "Morphed" : "Created")
4076                         << " node: ";
4077                  Res->dump(CurDAG););
4078 
4079       // If this was a MorphNodeTo then we're completely done!
4080       if (IsMorphNodeTo) {
4081         // Update chain uses.
4082         UpdateChains(Res, InputChain, ChainNodesMatched, true);
4083         return;
4084       }
4085       continue;
4086     }
4087 
4088     case OPC_CompleteMatch: {
4089       // The match has been completed, and any new nodes (if any) have been
4090       // created.  Patch up references to the matched dag to use the newly
4091       // created nodes.
4092       unsigned NumResults = MatcherTable[MatcherIndex++];
4093 
4094       for (unsigned i = 0; i != NumResults; ++i) {
4095         unsigned ResSlot = MatcherTable[MatcherIndex++];
4096         if (ResSlot & 128)
4097           ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex);
4098 
4099         assert(ResSlot < RecordedNodes.size() && "Invalid CompleteMatch");
4100         SDValue Res = RecordedNodes[ResSlot].first;
4101 
4102         assert(i < NodeToMatch->getNumValues() &&
4103                NodeToMatch->getValueType(i) != MVT::Other &&
4104                NodeToMatch->getValueType(i) != MVT::Glue &&
4105                "Invalid number of results to complete!");
4106         assert((NodeToMatch->getValueType(i) == Res.getValueType() ||
4107                 NodeToMatch->getValueType(i) == MVT::iPTR ||
4108                 Res.getValueType() == MVT::iPTR ||
4109                 NodeToMatch->getValueType(i).getSizeInBits() ==
4110                     Res.getValueSizeInBits()) &&
4111                "invalid replacement");
4112         ReplaceUses(SDValue(NodeToMatch, i), Res);
4113       }
4114 
4115       // Update chain uses.
4116       UpdateChains(NodeToMatch, InputChain, ChainNodesMatched, false);
4117 
4118       // If the root node defines glue, we need to update it to the glue result.
4119       // TODO: This never happens in our tests and I think it can be removed /
4120       // replaced with an assert, but if we do it this the way the change is
4121       // NFC.
4122       if (NodeToMatch->getValueType(NodeToMatch->getNumValues() - 1) ==
4123               MVT::Glue &&
4124           InputGlue.getNode())
4125         ReplaceUses(SDValue(NodeToMatch, NodeToMatch->getNumValues() - 1),
4126                     InputGlue);
4127 
4128       assert(NodeToMatch->use_empty() &&
4129              "Didn't replace all uses of the node?");
4130       CurDAG->RemoveDeadNode(NodeToMatch);
4131 
4132       return;
4133     }
4134     }
4135 
4136     // If the code reached this point, then the match failed.  See if there is
4137     // another child to try in the current 'Scope', otherwise pop it until we
4138     // find a case to check.
4139     LLVM_DEBUG(dbgs() << "  Match failed at index " << CurrentOpcodeIndex
4140                       << "\n");
4141     ++NumDAGIselRetries;
4142     while (true) {
4143       if (MatchScopes.empty()) {
4144         CannotYetSelect(NodeToMatch);
4145         return;
4146       }
4147 
4148       // Restore the interpreter state back to the point where the scope was
4149       // formed.
4150       MatchScope &LastScope = MatchScopes.back();
4151       RecordedNodes.resize(LastScope.NumRecordedNodes);
4152       NodeStack.clear();
4153       NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end());
4154       N = NodeStack.back();
4155 
4156       if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size())
4157         MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
4158       MatcherIndex = LastScope.FailIndex;
4159 
4160       LLVM_DEBUG(dbgs() << "  Continuing at " << MatcherIndex << "\n");
4161 
4162       InputChain = LastScope.InputChain;
4163       InputGlue = LastScope.InputGlue;
4164       if (!LastScope.HasChainNodesMatched)
4165         ChainNodesMatched.clear();
4166 
4167       // Check to see what the offset is at the new MatcherIndex.  If it is zero
4168       // we have reached the end of this scope, otherwise we have another child
4169       // in the current scope to try.
4170       unsigned NumToSkip = MatcherTable[MatcherIndex++];
4171       if (NumToSkip & 128)
4172         NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
4173 
4174       // If we have another child in this scope to match, update FailIndex and
4175       // try it.
4176       if (NumToSkip != 0) {
4177         LastScope.FailIndex = MatcherIndex+NumToSkip;
4178         break;
4179       }
4180 
4181       // End of this scope, pop it and try the next child in the containing
4182       // scope.
4183       MatchScopes.pop_back();
4184     }
4185   }
4186 }
4187 
4188 /// Return whether the node may raise an FP exception.
4189 bool SelectionDAGISel::mayRaiseFPException(SDNode *N) const {
4190   // For machine opcodes, consult the MCID flag.
4191   if (N->isMachineOpcode()) {
4192     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
4193     return MCID.mayRaiseFPException();
4194   }
4195 
4196   // For ISD opcodes, only StrictFP opcodes may raise an FP
4197   // exception.
4198   if (N->isTargetOpcode())
4199     return N->isTargetStrictFPOpcode();
4200   return N->isStrictFPOpcode();
4201 }
4202 
4203 bool SelectionDAGISel::isOrEquivalentToAdd(const SDNode *N) const {
4204   assert(N->getOpcode() == ISD::OR && "Unexpected opcode");
4205   auto *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4206   if (!C)
4207     return false;
4208 
4209   // Detect when "or" is used to add an offset to a stack object.
4210   if (auto *FN = dyn_cast<FrameIndexSDNode>(N->getOperand(0))) {
4211     MachineFrameInfo &MFI = MF->getFrameInfo();
4212     Align A = MFI.getObjectAlign(FN->getIndex());
4213     int32_t Off = C->getSExtValue();
4214     // If the alleged offset fits in the zero bits guaranteed by
4215     // the alignment, then this or is really an add.
4216     return (Off >= 0) && (((A.value() - 1) & Off) == unsigned(Off));
4217   }
4218   return false;
4219 }
4220 
4221 void SelectionDAGISel::CannotYetSelect(SDNode *N) {
4222   std::string msg;
4223   raw_string_ostream Msg(msg);
4224   Msg << "Cannot select: ";
4225 
4226   if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN &&
4227       N->getOpcode() != ISD::INTRINSIC_WO_CHAIN &&
4228       N->getOpcode() != ISD::INTRINSIC_VOID) {
4229     N->printrFull(Msg, CurDAG);
4230     Msg << "\nIn function: " << MF->getName();
4231   } else {
4232     bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other;
4233     unsigned iid = N->getConstantOperandVal(HasInputChain);
4234     if (iid < Intrinsic::num_intrinsics)
4235       Msg << "intrinsic %" << Intrinsic::getBaseName((Intrinsic::ID)iid);
4236     else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo())
4237       Msg << "target intrinsic %" << TII->getName(iid);
4238     else
4239       Msg << "unknown intrinsic #" << iid;
4240   }
4241   report_fatal_error(Twine(Msg.str()));
4242 }
4243