1 //===-- SelectionDAGISel.cpp - Implement the SelectionDAGISel class -------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This implements the SelectionDAGISel class.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/CodeGen/SelectionDAGISel.h"
15 #include "ScheduleDAGSDNodes.h"
16 #include "SelectionDAGBuilder.h"
17 #include "llvm/ADT/PostOrderIterator.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/BranchProbabilityInfo.h"
21 #include "llvm/Analysis/CFG.h"
22 #include "llvm/CodeGen/FastISel.h"
23 #include "llvm/CodeGen/FunctionLoweringInfo.h"
24 #include "llvm/CodeGen/GCMetadata.h"
25 #include "llvm/CodeGen/GCStrategy.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineModuleInfo.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
32 #include "llvm/CodeGen/SchedulerRegistry.h"
33 #include "llvm/CodeGen/SelectionDAG.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DebugInfo.h"
36 #include "llvm/IR/Function.h"
37 #include "llvm/IR/InlineAsm.h"
38 #include "llvm/IR/Instructions.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/Intrinsics.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/IR/Module.h"
43 #include "llvm/Support/Compiler.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/ErrorHandling.h"
46 #include "llvm/Support/Timer.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Target/TargetInstrInfo.h"
49 #include "llvm/Target/TargetIntrinsicInfo.h"
50 #include "llvm/Target/TargetLibraryInfo.h"
51 #include "llvm/Target/TargetLowering.h"
52 #include "llvm/Target/TargetMachine.h"
53 #include "llvm/Target/TargetOptions.h"
54 #include "llvm/Target/TargetRegisterInfo.h"
55 #include "llvm/Target/TargetSubtargetInfo.h"
56 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
57 #include <algorithm>
58 using namespace llvm;
59
60 #define DEBUG_TYPE "isel"
61
62 STATISTIC(NumFastIselFailures, "Number of instructions fast isel failed on");
63 STATISTIC(NumFastIselSuccess, "Number of instructions fast isel selected");
64 STATISTIC(NumFastIselBlocks, "Number of blocks selected entirely by fast isel");
65 STATISTIC(NumDAGBlocks, "Number of blocks selected using DAG");
66 STATISTIC(NumDAGIselRetries,"Number of times dag isel has to try another path");
67 STATISTIC(NumEntryBlocks, "Number of entry blocks encountered");
68 STATISTIC(NumFastIselFailLowerArguments,
69 "Number of entry blocks where fast isel failed to lower arguments");
70
71 #ifndef NDEBUG
72 static cl::opt<bool>
73 EnableFastISelVerbose2("fast-isel-verbose2", cl::Hidden,
74 cl::desc("Enable extra verbose messages in the \"fast\" "
75 "instruction selector"));
76
77 // Terminators
78 STATISTIC(NumFastIselFailRet,"Fast isel fails on Ret");
79 STATISTIC(NumFastIselFailBr,"Fast isel fails on Br");
80 STATISTIC(NumFastIselFailSwitch,"Fast isel fails on Switch");
81 STATISTIC(NumFastIselFailIndirectBr,"Fast isel fails on IndirectBr");
82 STATISTIC(NumFastIselFailInvoke,"Fast isel fails on Invoke");
83 STATISTIC(NumFastIselFailResume,"Fast isel fails on Resume");
84 STATISTIC(NumFastIselFailUnreachable,"Fast isel fails on Unreachable");
85
86 // Standard binary operators...
87 STATISTIC(NumFastIselFailAdd,"Fast isel fails on Add");
88 STATISTIC(NumFastIselFailFAdd,"Fast isel fails on FAdd");
89 STATISTIC(NumFastIselFailSub,"Fast isel fails on Sub");
90 STATISTIC(NumFastIselFailFSub,"Fast isel fails on FSub");
91 STATISTIC(NumFastIselFailMul,"Fast isel fails on Mul");
92 STATISTIC(NumFastIselFailFMul,"Fast isel fails on FMul");
93 STATISTIC(NumFastIselFailUDiv,"Fast isel fails on UDiv");
94 STATISTIC(NumFastIselFailSDiv,"Fast isel fails on SDiv");
95 STATISTIC(NumFastIselFailFDiv,"Fast isel fails on FDiv");
96 STATISTIC(NumFastIselFailURem,"Fast isel fails on URem");
97 STATISTIC(NumFastIselFailSRem,"Fast isel fails on SRem");
98 STATISTIC(NumFastIselFailFRem,"Fast isel fails on FRem");
99
100 // Logical operators...
101 STATISTIC(NumFastIselFailAnd,"Fast isel fails on And");
102 STATISTIC(NumFastIselFailOr,"Fast isel fails on Or");
103 STATISTIC(NumFastIselFailXor,"Fast isel fails on Xor");
104
105 // Memory instructions...
106 STATISTIC(NumFastIselFailAlloca,"Fast isel fails on Alloca");
107 STATISTIC(NumFastIselFailLoad,"Fast isel fails on Load");
108 STATISTIC(NumFastIselFailStore,"Fast isel fails on Store");
109 STATISTIC(NumFastIselFailAtomicCmpXchg,"Fast isel fails on AtomicCmpXchg");
110 STATISTIC(NumFastIselFailAtomicRMW,"Fast isel fails on AtomicRWM");
111 STATISTIC(NumFastIselFailFence,"Fast isel fails on Frence");
112 STATISTIC(NumFastIselFailGetElementPtr,"Fast isel fails on GetElementPtr");
113
114 // Convert instructions...
115 STATISTIC(NumFastIselFailTrunc,"Fast isel fails on Trunc");
116 STATISTIC(NumFastIselFailZExt,"Fast isel fails on ZExt");
117 STATISTIC(NumFastIselFailSExt,"Fast isel fails on SExt");
118 STATISTIC(NumFastIselFailFPTrunc,"Fast isel fails on FPTrunc");
119 STATISTIC(NumFastIselFailFPExt,"Fast isel fails on FPExt");
120 STATISTIC(NumFastIselFailFPToUI,"Fast isel fails on FPToUI");
121 STATISTIC(NumFastIselFailFPToSI,"Fast isel fails on FPToSI");
122 STATISTIC(NumFastIselFailUIToFP,"Fast isel fails on UIToFP");
123 STATISTIC(NumFastIselFailSIToFP,"Fast isel fails on SIToFP");
124 STATISTIC(NumFastIselFailIntToPtr,"Fast isel fails on IntToPtr");
125 STATISTIC(NumFastIselFailPtrToInt,"Fast isel fails on PtrToInt");
126 STATISTIC(NumFastIselFailBitCast,"Fast isel fails on BitCast");
127
128 // Other instructions...
129 STATISTIC(NumFastIselFailICmp,"Fast isel fails on ICmp");
130 STATISTIC(NumFastIselFailFCmp,"Fast isel fails on FCmp");
131 STATISTIC(NumFastIselFailPHI,"Fast isel fails on PHI");
132 STATISTIC(NumFastIselFailSelect,"Fast isel fails on Select");
133 STATISTIC(NumFastIselFailCall,"Fast isel fails on Call");
134 STATISTIC(NumFastIselFailShl,"Fast isel fails on Shl");
135 STATISTIC(NumFastIselFailLShr,"Fast isel fails on LShr");
136 STATISTIC(NumFastIselFailAShr,"Fast isel fails on AShr");
137 STATISTIC(NumFastIselFailVAArg,"Fast isel fails on VAArg");
138 STATISTIC(NumFastIselFailExtractElement,"Fast isel fails on ExtractElement");
139 STATISTIC(NumFastIselFailInsertElement,"Fast isel fails on InsertElement");
140 STATISTIC(NumFastIselFailShuffleVector,"Fast isel fails on ShuffleVector");
141 STATISTIC(NumFastIselFailExtractValue,"Fast isel fails on ExtractValue");
142 STATISTIC(NumFastIselFailInsertValue,"Fast isel fails on InsertValue");
143 STATISTIC(NumFastIselFailLandingPad,"Fast isel fails on LandingPad");
144
145 // Intrinsic instructions...
146 STATISTIC(NumFastIselFailIntrinsicCall, "Fast isel fails on Intrinsic call");
147 STATISTIC(NumFastIselFailSAddWithOverflow,
148 "Fast isel fails on sadd.with.overflow");
149 STATISTIC(NumFastIselFailUAddWithOverflow,
150 "Fast isel fails on uadd.with.overflow");
151 STATISTIC(NumFastIselFailSSubWithOverflow,
152 "Fast isel fails on ssub.with.overflow");
153 STATISTIC(NumFastIselFailUSubWithOverflow,
154 "Fast isel fails on usub.with.overflow");
155 STATISTIC(NumFastIselFailSMulWithOverflow,
156 "Fast isel fails on smul.with.overflow");
157 STATISTIC(NumFastIselFailUMulWithOverflow,
158 "Fast isel fails on umul.with.overflow");
159 STATISTIC(NumFastIselFailFrameaddress, "Fast isel fails on Frameaddress");
160 STATISTIC(NumFastIselFailSqrt, "Fast isel fails on sqrt call");
161 STATISTIC(NumFastIselFailStackMap, "Fast isel fails on StackMap call");
162 STATISTIC(NumFastIselFailPatchPoint, "Fast isel fails on PatchPoint call");
163 #endif
164
165 static cl::opt<bool>
166 EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
167 cl::desc("Enable verbose messages in the \"fast\" "
168 "instruction selector"));
169 static cl::opt<bool>
170 EnableFastISelAbort("fast-isel-abort", cl::Hidden,
171 cl::desc("Enable abort calls when \"fast\" instruction selection "
172 "fails to lower an instruction"));
173 static cl::opt<bool>
174 EnableFastISelAbortArgs("fast-isel-abort-args", cl::Hidden,
175 cl::desc("Enable abort calls when \"fast\" instruction selection "
176 "fails to lower a formal argument"));
177
178 static cl::opt<bool>
179 UseMBPI("use-mbpi",
180 cl::desc("use Machine Branch Probability Info"),
181 cl::init(true), cl::Hidden);
182
183 #ifndef NDEBUG
184 static cl::opt<bool>
185 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
186 cl::desc("Pop up a window to show dags before the first "
187 "dag combine pass"));
188 static cl::opt<bool>
189 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
190 cl::desc("Pop up a window to show dags before legalize types"));
191 static cl::opt<bool>
192 ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
193 cl::desc("Pop up a window to show dags before legalize"));
194 static cl::opt<bool>
195 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
196 cl::desc("Pop up a window to show dags before the second "
197 "dag combine pass"));
198 static cl::opt<bool>
199 ViewDAGCombineLT("view-dag-combine-lt-dags", cl::Hidden,
200 cl::desc("Pop up a window to show dags before the post legalize types"
201 " dag combine pass"));
202 static cl::opt<bool>
203 ViewISelDAGs("view-isel-dags", cl::Hidden,
204 cl::desc("Pop up a window to show isel dags as they are selected"));
205 static cl::opt<bool>
206 ViewSchedDAGs("view-sched-dags", cl::Hidden,
207 cl::desc("Pop up a window to show sched dags as they are processed"));
208 static cl::opt<bool>
209 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
210 cl::desc("Pop up a window to show SUnit dags after they are processed"));
211 #else
212 static const bool ViewDAGCombine1 = false,
213 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false,
214 ViewDAGCombine2 = false,
215 ViewDAGCombineLT = false,
216 ViewISelDAGs = false, ViewSchedDAGs = false,
217 ViewSUnitDAGs = false;
218 #endif
219
220 //===---------------------------------------------------------------------===//
221 ///
222 /// RegisterScheduler class - Track the registration of instruction schedulers.
223 ///
224 //===---------------------------------------------------------------------===//
225 MachinePassRegistry RegisterScheduler::Registry;
226
227 //===---------------------------------------------------------------------===//
228 ///
229 /// ISHeuristic command line option for instruction schedulers.
230 ///
231 //===---------------------------------------------------------------------===//
232 static cl::opt<RegisterScheduler::FunctionPassCtor, false,
233 RegisterPassParser<RegisterScheduler> >
234 ISHeuristic("pre-RA-sched",
235 cl::init(&createDefaultScheduler), cl::Hidden,
236 cl::desc("Instruction schedulers available (before register"
237 " allocation):"));
238
239 static RegisterScheduler
240 defaultListDAGScheduler("default", "Best scheduler for the target",
241 createDefaultScheduler);
242
243 namespace llvm {
244 //===--------------------------------------------------------------------===//
245 /// \brief This class is used by SelectionDAGISel to temporarily override
246 /// the optimization level on a per-function basis.
247 class OptLevelChanger {
248 SelectionDAGISel &IS;
249 CodeGenOpt::Level SavedOptLevel;
250 bool SavedFastISel;
251
252 public:
OptLevelChanger(SelectionDAGISel & ISel,CodeGenOpt::Level NewOptLevel)253 OptLevelChanger(SelectionDAGISel &ISel,
254 CodeGenOpt::Level NewOptLevel) : IS(ISel) {
255 SavedOptLevel = IS.OptLevel;
256 if (NewOptLevel == SavedOptLevel)
257 return;
258 IS.OptLevel = NewOptLevel;
259 IS.TM.setOptLevel(NewOptLevel);
260 SavedFastISel = IS.TM.Options.EnableFastISel;
261 if (NewOptLevel == CodeGenOpt::None)
262 IS.TM.setFastISel(true);
263 DEBUG(dbgs() << "\nChanging optimization level for Function "
264 << IS.MF->getFunction()->getName() << "\n");
265 DEBUG(dbgs() << "\tBefore: -O" << SavedOptLevel
266 << " ; After: -O" << NewOptLevel << "\n");
267 }
268
~OptLevelChanger()269 ~OptLevelChanger() {
270 if (IS.OptLevel == SavedOptLevel)
271 return;
272 DEBUG(dbgs() << "\nRestoring optimization level for Function "
273 << IS.MF->getFunction()->getName() << "\n");
274 DEBUG(dbgs() << "\tBefore: -O" << IS.OptLevel
275 << " ; After: -O" << SavedOptLevel << "\n");
276 IS.OptLevel = SavedOptLevel;
277 IS.TM.setOptLevel(SavedOptLevel);
278 IS.TM.setFastISel(SavedFastISel);
279 }
280 };
281
282 //===--------------------------------------------------------------------===//
283 /// createDefaultScheduler - This creates an instruction scheduler appropriate
284 /// for the target.
createDefaultScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)285 ScheduleDAGSDNodes* createDefaultScheduler(SelectionDAGISel *IS,
286 CodeGenOpt::Level OptLevel) {
287 const TargetLowering *TLI = IS->getTargetLowering();
288 const TargetSubtargetInfo &ST = IS->TM.getSubtarget<TargetSubtargetInfo>();
289
290 if (OptLevel == CodeGenOpt::None || ST.useMachineScheduler() ||
291 TLI->getSchedulingPreference() == Sched::Source)
292 return createSourceListDAGScheduler(IS, OptLevel);
293 if (TLI->getSchedulingPreference() == Sched::RegPressure)
294 return createBURRListDAGScheduler(IS, OptLevel);
295 if (TLI->getSchedulingPreference() == Sched::Hybrid)
296 return createHybridListDAGScheduler(IS, OptLevel);
297 if (TLI->getSchedulingPreference() == Sched::VLIW)
298 return createVLIWDAGScheduler(IS, OptLevel);
299 assert(TLI->getSchedulingPreference() == Sched::ILP &&
300 "Unknown sched type!");
301 return createILPListDAGScheduler(IS, OptLevel);
302 }
303 }
304
305 // EmitInstrWithCustomInserter - This method should be implemented by targets
306 // that mark instructions with the 'usesCustomInserter' flag. These
307 // instructions are special in various ways, which require special support to
308 // insert. The specified MachineInstr is created but not inserted into any
309 // basic blocks, and this method is called to expand it into a sequence of
310 // instructions, potentially also creating new basic blocks and control flow.
311 // When new basic blocks are inserted and the edges from MBB to its successors
312 // are modified, the method should insert pairs of <OldSucc, NewSucc> into the
313 // DenseMap.
314 MachineBasicBlock *
EmitInstrWithCustomInserter(MachineInstr * MI,MachineBasicBlock * MBB) const315 TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
316 MachineBasicBlock *MBB) const {
317 #ifndef NDEBUG
318 dbgs() << "If a target marks an instruction with "
319 "'usesCustomInserter', it must implement "
320 "TargetLowering::EmitInstrWithCustomInserter!";
321 #endif
322 llvm_unreachable(nullptr);
323 }
324
AdjustInstrPostInstrSelection(MachineInstr * MI,SDNode * Node) const325 void TargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
326 SDNode *Node) const {
327 assert(!MI->hasPostISelHook() &&
328 "If a target marks an instruction with 'hasPostISelHook', "
329 "it must implement TargetLowering::AdjustInstrPostInstrSelection!");
330 }
331
332 //===----------------------------------------------------------------------===//
333 // SelectionDAGISel code
334 //===----------------------------------------------------------------------===//
335
SelectionDAGISel(TargetMachine & tm,CodeGenOpt::Level OL)336 SelectionDAGISel::SelectionDAGISel(TargetMachine &tm,
337 CodeGenOpt::Level OL) :
338 MachineFunctionPass(ID), TM(tm),
339 FuncInfo(new FunctionLoweringInfo(TM)),
340 CurDAG(new SelectionDAG(tm, OL)),
341 SDB(new SelectionDAGBuilder(*CurDAG, *FuncInfo, OL)),
342 GFI(),
343 OptLevel(OL),
344 DAGSize(0) {
345 initializeGCModuleInfoPass(*PassRegistry::getPassRegistry());
346 initializeAliasAnalysisAnalysisGroup(*PassRegistry::getPassRegistry());
347 initializeBranchProbabilityInfoPass(*PassRegistry::getPassRegistry());
348 initializeTargetLibraryInfoPass(*PassRegistry::getPassRegistry());
349 }
350
~SelectionDAGISel()351 SelectionDAGISel::~SelectionDAGISel() {
352 delete SDB;
353 delete CurDAG;
354 delete FuncInfo;
355 }
356
getAnalysisUsage(AnalysisUsage & AU) const357 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
358 AU.addRequired<AliasAnalysis>();
359 AU.addPreserved<AliasAnalysis>();
360 AU.addRequired<GCModuleInfo>();
361 AU.addPreserved<GCModuleInfo>();
362 AU.addRequired<TargetLibraryInfo>();
363 if (UseMBPI && OptLevel != CodeGenOpt::None)
364 AU.addRequired<BranchProbabilityInfo>();
365 MachineFunctionPass::getAnalysisUsage(AU);
366 }
367
368 /// SplitCriticalSideEffectEdges - Look for critical edges with a PHI value that
369 /// may trap on it. In this case we have to split the edge so that the path
370 /// through the predecessor block that doesn't go to the phi block doesn't
371 /// execute the possibly trapping instruction.
372 ///
373 /// This is required for correctness, so it must be done at -O0.
374 ///
SplitCriticalSideEffectEdges(Function & Fn,Pass * SDISel)375 static void SplitCriticalSideEffectEdges(Function &Fn, Pass *SDISel) {
376 // Loop for blocks with phi nodes.
377 for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
378 PHINode *PN = dyn_cast<PHINode>(BB->begin());
379 if (!PN) continue;
380
381 ReprocessBlock:
382 // For each block with a PHI node, check to see if any of the input values
383 // are potentially trapping constant expressions. Constant expressions are
384 // the only potentially trapping value that can occur as the argument to a
385 // PHI.
386 for (BasicBlock::iterator I = BB->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
387 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
388 ConstantExpr *CE = dyn_cast<ConstantExpr>(PN->getIncomingValue(i));
389 if (!CE || !CE->canTrap()) continue;
390
391 // The only case we have to worry about is when the edge is critical.
392 // Since this block has a PHI Node, we assume it has multiple input
393 // edges: check to see if the pred has multiple successors.
394 BasicBlock *Pred = PN->getIncomingBlock(i);
395 if (Pred->getTerminator()->getNumSuccessors() == 1)
396 continue;
397
398 // Okay, we have to split this edge.
399 SplitCriticalEdge(Pred->getTerminator(),
400 GetSuccessorNumber(Pred, BB), SDISel, true);
401 goto ReprocessBlock;
402 }
403 }
404 }
405
runOnMachineFunction(MachineFunction & mf)406 bool SelectionDAGISel::runOnMachineFunction(MachineFunction &mf) {
407 // Do some sanity-checking on the command-line options.
408 assert((!EnableFastISelVerbose || TM.Options.EnableFastISel) &&
409 "-fast-isel-verbose requires -fast-isel");
410 assert((!EnableFastISelAbort || TM.Options.EnableFastISel) &&
411 "-fast-isel-abort requires -fast-isel");
412
413 const Function &Fn = *mf.getFunction();
414 const TargetInstrInfo &TII = *TM.getInstrInfo();
415 const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
416 const TargetLowering *TLI = TM.getTargetLowering();
417
418 MF = &mf;
419 RegInfo = &MF->getRegInfo();
420 AA = &getAnalysis<AliasAnalysis>();
421 LibInfo = &getAnalysis<TargetLibraryInfo>();
422 GFI = Fn.hasGC() ? &getAnalysis<GCModuleInfo>().getFunctionInfo(Fn) : nullptr;
423
424 TargetSubtargetInfo &ST =
425 const_cast<TargetSubtargetInfo&>(TM.getSubtarget<TargetSubtargetInfo>());
426 ST.resetSubtargetFeatures(MF);
427 TM.resetTargetOptions(MF);
428
429 // Reset OptLevel to None for optnone functions.
430 CodeGenOpt::Level NewOptLevel = OptLevel;
431 if (Fn.hasFnAttribute(Attribute::OptimizeNone))
432 NewOptLevel = CodeGenOpt::None;
433 OptLevelChanger OLC(*this, NewOptLevel);
434
435 DEBUG(dbgs() << "\n\n\n=== " << Fn.getName() << "\n");
436
437 SplitCriticalSideEffectEdges(const_cast<Function&>(Fn), this);
438
439 CurDAG->init(*MF, TLI);
440 FuncInfo->set(Fn, *MF, CurDAG);
441
442 if (UseMBPI && OptLevel != CodeGenOpt::None)
443 FuncInfo->BPI = &getAnalysis<BranchProbabilityInfo>();
444 else
445 FuncInfo->BPI = nullptr;
446
447 SDB->init(GFI, *AA, LibInfo);
448
449 MF->setHasInlineAsm(false);
450
451 SelectAllBasicBlocks(Fn);
452
453 // If the first basic block in the function has live ins that need to be
454 // copied into vregs, emit the copies into the top of the block before
455 // emitting the code for the block.
456 MachineBasicBlock *EntryMBB = MF->begin();
457 RegInfo->EmitLiveInCopies(EntryMBB, TRI, TII);
458
459 DenseMap<unsigned, unsigned> LiveInMap;
460 if (!FuncInfo->ArgDbgValues.empty())
461 for (MachineRegisterInfo::livein_iterator LI = RegInfo->livein_begin(),
462 E = RegInfo->livein_end(); LI != E; ++LI)
463 if (LI->second)
464 LiveInMap.insert(std::make_pair(LI->first, LI->second));
465
466 // Insert DBG_VALUE instructions for function arguments to the entry block.
467 for (unsigned i = 0, e = FuncInfo->ArgDbgValues.size(); i != e; ++i) {
468 MachineInstr *MI = FuncInfo->ArgDbgValues[e-i-1];
469 bool hasFI = MI->getOperand(0).isFI();
470 unsigned Reg =
471 hasFI ? TRI.getFrameRegister(*MF) : MI->getOperand(0).getReg();
472 if (TargetRegisterInfo::isPhysicalRegister(Reg))
473 EntryMBB->insert(EntryMBB->begin(), MI);
474 else {
475 MachineInstr *Def = RegInfo->getVRegDef(Reg);
476 if (Def) {
477 MachineBasicBlock::iterator InsertPos = Def;
478 // FIXME: VR def may not be in entry block.
479 Def->getParent()->insert(std::next(InsertPos), MI);
480 } else
481 DEBUG(dbgs() << "Dropping debug info for dead vreg"
482 << TargetRegisterInfo::virtReg2Index(Reg) << "\n");
483 }
484
485 // If Reg is live-in then update debug info to track its copy in a vreg.
486 DenseMap<unsigned, unsigned>::iterator LDI = LiveInMap.find(Reg);
487 if (LDI != LiveInMap.end()) {
488 assert(!hasFI && "There's no handling of frame pointer updating here yet "
489 "- add if needed");
490 MachineInstr *Def = RegInfo->getVRegDef(LDI->second);
491 MachineBasicBlock::iterator InsertPos = Def;
492 const MDNode *Variable =
493 MI->getOperand(MI->getNumOperands()-1).getMetadata();
494 bool IsIndirect = MI->isIndirectDebugValue();
495 unsigned Offset = IsIndirect ? MI->getOperand(1).getImm() : 0;
496 // Def is never a terminator here, so it is ok to increment InsertPos.
497 BuildMI(*EntryMBB, ++InsertPos, MI->getDebugLoc(),
498 TII.get(TargetOpcode::DBG_VALUE),
499 IsIndirect,
500 LDI->second, Offset, Variable);
501
502 // If this vreg is directly copied into an exported register then
503 // that COPY instructions also need DBG_VALUE, if it is the only
504 // user of LDI->second.
505 MachineInstr *CopyUseMI = nullptr;
506 for (MachineRegisterInfo::use_instr_iterator
507 UI = RegInfo->use_instr_begin(LDI->second),
508 E = RegInfo->use_instr_end(); UI != E; ) {
509 MachineInstr *UseMI = &*(UI++);
510 if (UseMI->isDebugValue()) continue;
511 if (UseMI->isCopy() && !CopyUseMI && UseMI->getParent() == EntryMBB) {
512 CopyUseMI = UseMI; continue;
513 }
514 // Otherwise this is another use or second copy use.
515 CopyUseMI = nullptr; break;
516 }
517 if (CopyUseMI) {
518 MachineInstr *NewMI =
519 BuildMI(*MF, CopyUseMI->getDebugLoc(),
520 TII.get(TargetOpcode::DBG_VALUE),
521 IsIndirect,
522 CopyUseMI->getOperand(0).getReg(),
523 Offset, Variable);
524 MachineBasicBlock::iterator Pos = CopyUseMI;
525 EntryMBB->insertAfter(Pos, NewMI);
526 }
527 }
528 }
529
530 // Determine if there are any calls in this machine function.
531 MachineFrameInfo *MFI = MF->getFrameInfo();
532 for (const auto &MBB : *MF) {
533 if (MFI->hasCalls() && MF->hasInlineAsm())
534 break;
535
536 for (const auto &MI : MBB) {
537 const MCInstrDesc &MCID = TM.getInstrInfo()->get(MI.getOpcode());
538 if ((MCID.isCall() && !MCID.isReturn()) ||
539 MI.isStackAligningInlineAsm()) {
540 MFI->setHasCalls(true);
541 }
542 if (MI.isInlineAsm()) {
543 MF->setHasInlineAsm(true);
544 }
545 }
546 }
547
548 // Determine if there is a call to setjmp in the machine function.
549 MF->setExposesReturnsTwice(Fn.callsFunctionThatReturnsTwice());
550
551 // Replace forward-declared registers with the registers containing
552 // the desired value.
553 MachineRegisterInfo &MRI = MF->getRegInfo();
554 for (DenseMap<unsigned, unsigned>::iterator
555 I = FuncInfo->RegFixups.begin(), E = FuncInfo->RegFixups.end();
556 I != E; ++I) {
557 unsigned From = I->first;
558 unsigned To = I->second;
559 // If To is also scheduled to be replaced, find what its ultimate
560 // replacement is.
561 for (;;) {
562 DenseMap<unsigned, unsigned>::iterator J = FuncInfo->RegFixups.find(To);
563 if (J == E) break;
564 To = J->second;
565 }
566 // Make sure the new register has a sufficiently constrained register class.
567 if (TargetRegisterInfo::isVirtualRegister(From) &&
568 TargetRegisterInfo::isVirtualRegister(To))
569 MRI.constrainRegClass(To, MRI.getRegClass(From));
570 // Replace it.
571 MRI.replaceRegWith(From, To);
572 }
573
574 // Freeze the set of reserved registers now that MachineFrameInfo has been
575 // set up. All the information required by getReservedRegs() should be
576 // available now.
577 MRI.freezeReservedRegs(*MF);
578
579 // Release function-specific state. SDB and CurDAG are already cleared
580 // at this point.
581 FuncInfo->clear();
582
583 DEBUG(dbgs() << "*** MachineFunction at end of ISel ***\n");
584 DEBUG(MF->print(dbgs()));
585
586 return true;
587 }
588
SelectBasicBlock(BasicBlock::const_iterator Begin,BasicBlock::const_iterator End,bool & HadTailCall)589 void SelectionDAGISel::SelectBasicBlock(BasicBlock::const_iterator Begin,
590 BasicBlock::const_iterator End,
591 bool &HadTailCall) {
592 // Lower all of the non-terminator instructions. If a call is emitted
593 // as a tail call, cease emitting nodes for this block. Terminators
594 // are handled below.
595 for (BasicBlock::const_iterator I = Begin; I != End && !SDB->HasTailCall; ++I)
596 SDB->visit(*I);
597
598 // Make sure the root of the DAG is up-to-date.
599 CurDAG->setRoot(SDB->getControlRoot());
600 HadTailCall = SDB->HasTailCall;
601 SDB->clear();
602
603 // Final step, emit the lowered DAG as machine code.
604 CodeGenAndEmitDAG();
605 }
606
ComputeLiveOutVRegInfo()607 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
608 SmallPtrSet<SDNode*, 128> VisitedNodes;
609 SmallVector<SDNode*, 128> Worklist;
610
611 Worklist.push_back(CurDAG->getRoot().getNode());
612
613 APInt KnownZero;
614 APInt KnownOne;
615
616 do {
617 SDNode *N = Worklist.pop_back_val();
618
619 // If we've already seen this node, ignore it.
620 if (!VisitedNodes.insert(N))
621 continue;
622
623 // Otherwise, add all chain operands to the worklist.
624 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
625 if (N->getOperand(i).getValueType() == MVT::Other)
626 Worklist.push_back(N->getOperand(i).getNode());
627
628 // If this is a CopyToReg with a vreg dest, process it.
629 if (N->getOpcode() != ISD::CopyToReg)
630 continue;
631
632 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
633 if (!TargetRegisterInfo::isVirtualRegister(DestReg))
634 continue;
635
636 // Ignore non-scalar or non-integer values.
637 SDValue Src = N->getOperand(2);
638 EVT SrcVT = Src.getValueType();
639 if (!SrcVT.isInteger() || SrcVT.isVector())
640 continue;
641
642 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
643 CurDAG->computeKnownBits(Src, KnownZero, KnownOne);
644 FuncInfo->AddLiveOutRegInfo(DestReg, NumSignBits, KnownZero, KnownOne);
645 } while (!Worklist.empty());
646 }
647
CodeGenAndEmitDAG()648 void SelectionDAGISel::CodeGenAndEmitDAG() {
649 std::string GroupName;
650 if (TimePassesIsEnabled)
651 GroupName = "Instruction Selection and Scheduling";
652 std::string BlockName;
653 int BlockNumber = -1;
654 (void)BlockNumber;
655 #ifdef NDEBUG
656 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs ||
657 ViewDAGCombine2 || ViewDAGCombineLT || ViewISelDAGs || ViewSchedDAGs ||
658 ViewSUnitDAGs)
659 #endif
660 {
661 BlockNumber = FuncInfo->MBB->getNumber();
662 BlockName = MF->getName().str() + ":" +
663 FuncInfo->MBB->getBasicBlock()->getName().str();
664 }
665 DEBUG(dbgs() << "Initial selection DAG: BB#" << BlockNumber
666 << " '" << BlockName << "'\n"; CurDAG->dump());
667
668 if (ViewDAGCombine1) CurDAG->viewGraph("dag-combine1 input for " + BlockName);
669
670 // Run the DAG combiner in pre-legalize mode.
671 {
672 NamedRegionTimer T("DAG Combining 1", GroupName, TimePassesIsEnabled);
673 CurDAG->Combine(BeforeLegalizeTypes, *AA, OptLevel);
674 }
675
676 DEBUG(dbgs() << "Optimized lowered selection DAG: BB#" << BlockNumber
677 << " '" << BlockName << "'\n"; CurDAG->dump());
678
679 // Second step, hack on the DAG until it only uses operations and types that
680 // the target supports.
681 if (ViewLegalizeTypesDAGs) CurDAG->viewGraph("legalize-types input for " +
682 BlockName);
683
684 bool Changed;
685 {
686 NamedRegionTimer T("Type Legalization", GroupName, TimePassesIsEnabled);
687 Changed = CurDAG->LegalizeTypes();
688 }
689
690 DEBUG(dbgs() << "Type-legalized selection DAG: BB#" << BlockNumber
691 << " '" << BlockName << "'\n"; CurDAG->dump());
692
693 CurDAG->NewNodesMustHaveLegalTypes = true;
694
695 if (Changed) {
696 if (ViewDAGCombineLT)
697 CurDAG->viewGraph("dag-combine-lt input for " + BlockName);
698
699 // Run the DAG combiner in post-type-legalize mode.
700 {
701 NamedRegionTimer T("DAG Combining after legalize types", GroupName,
702 TimePassesIsEnabled);
703 CurDAG->Combine(AfterLegalizeTypes, *AA, OptLevel);
704 }
705
706 DEBUG(dbgs() << "Optimized type-legalized selection DAG: BB#" << BlockNumber
707 << " '" << BlockName << "'\n"; CurDAG->dump());
708
709 }
710
711 {
712 NamedRegionTimer T("Vector Legalization", GroupName, TimePassesIsEnabled);
713 Changed = CurDAG->LegalizeVectors();
714 }
715
716 if (Changed) {
717 {
718 NamedRegionTimer T("Type Legalization 2", GroupName, TimePassesIsEnabled);
719 CurDAG->LegalizeTypes();
720 }
721
722 if (ViewDAGCombineLT)
723 CurDAG->viewGraph("dag-combine-lv input for " + BlockName);
724
725 // Run the DAG combiner in post-type-legalize mode.
726 {
727 NamedRegionTimer T("DAG Combining after legalize vectors", GroupName,
728 TimePassesIsEnabled);
729 CurDAG->Combine(AfterLegalizeVectorOps, *AA, OptLevel);
730 }
731
732 DEBUG(dbgs() << "Optimized vector-legalized selection DAG: BB#"
733 << BlockNumber << " '" << BlockName << "'\n"; CurDAG->dump());
734 }
735
736 if (ViewLegalizeDAGs) CurDAG->viewGraph("legalize input for " + BlockName);
737
738 {
739 NamedRegionTimer T("DAG Legalization", GroupName, TimePassesIsEnabled);
740 CurDAG->Legalize();
741 }
742
743 DEBUG(dbgs() << "Legalized selection DAG: BB#" << BlockNumber
744 << " '" << BlockName << "'\n"; CurDAG->dump());
745
746 if (ViewDAGCombine2) CurDAG->viewGraph("dag-combine2 input for " + BlockName);
747
748 // Run the DAG combiner in post-legalize mode.
749 {
750 NamedRegionTimer T("DAG Combining 2", GroupName, TimePassesIsEnabled);
751 CurDAG->Combine(AfterLegalizeDAG, *AA, OptLevel);
752 }
753
754 DEBUG(dbgs() << "Optimized legalized selection DAG: BB#" << BlockNumber
755 << " '" << BlockName << "'\n"; CurDAG->dump());
756
757 if (OptLevel != CodeGenOpt::None)
758 ComputeLiveOutVRegInfo();
759
760 if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName);
761
762 // Third, instruction select all of the operations to machine code, adding the
763 // code to the MachineBasicBlock.
764 {
765 NamedRegionTimer T("Instruction Selection", GroupName, TimePassesIsEnabled);
766 DoInstructionSelection();
767 }
768
769 DEBUG(dbgs() << "Selected selection DAG: BB#" << BlockNumber
770 << " '" << BlockName << "'\n"; CurDAG->dump());
771
772 if (ViewSchedDAGs) CurDAG->viewGraph("scheduler input for " + BlockName);
773
774 // Schedule machine code.
775 ScheduleDAGSDNodes *Scheduler = CreateScheduler();
776 {
777 NamedRegionTimer T("Instruction Scheduling", GroupName,
778 TimePassesIsEnabled);
779 Scheduler->Run(CurDAG, FuncInfo->MBB);
780 }
781
782 if (ViewSUnitDAGs) Scheduler->viewGraph();
783
784 // Emit machine code to BB. This can change 'BB' to the last block being
785 // inserted into.
786 MachineBasicBlock *FirstMBB = FuncInfo->MBB, *LastMBB;
787 {
788 NamedRegionTimer T("Instruction Creation", GroupName, TimePassesIsEnabled);
789
790 // FuncInfo->InsertPt is passed by reference and set to the end of the
791 // scheduled instructions.
792 LastMBB = FuncInfo->MBB = Scheduler->EmitSchedule(FuncInfo->InsertPt);
793 }
794
795 // If the block was split, make sure we update any references that are used to
796 // update PHI nodes later on.
797 if (FirstMBB != LastMBB)
798 SDB->UpdateSplitBlock(FirstMBB, LastMBB);
799
800 // Free the scheduler state.
801 {
802 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName,
803 TimePassesIsEnabled);
804 delete Scheduler;
805 }
806
807 // Free the SelectionDAG state, now that we're finished with it.
808 CurDAG->clear();
809 }
810
811 namespace {
812 /// ISelUpdater - helper class to handle updates of the instruction selection
813 /// graph.
814 class ISelUpdater : public SelectionDAG::DAGUpdateListener {
815 SelectionDAG::allnodes_iterator &ISelPosition;
816 public:
ISelUpdater(SelectionDAG & DAG,SelectionDAG::allnodes_iterator & isp)817 ISelUpdater(SelectionDAG &DAG, SelectionDAG::allnodes_iterator &isp)
818 : SelectionDAG::DAGUpdateListener(DAG), ISelPosition(isp) {}
819
820 /// NodeDeleted - Handle nodes deleted from the graph. If the node being
821 /// deleted is the current ISelPosition node, update ISelPosition.
822 ///
NodeDeleted(SDNode * N,SDNode * E)823 void NodeDeleted(SDNode *N, SDNode *E) override {
824 if (ISelPosition == SelectionDAG::allnodes_iterator(N))
825 ++ISelPosition;
826 }
827 };
828 } // end anonymous namespace
829
DoInstructionSelection()830 void SelectionDAGISel::DoInstructionSelection() {
831 DEBUG(dbgs() << "===== Instruction selection begins: BB#"
832 << FuncInfo->MBB->getNumber()
833 << " '" << FuncInfo->MBB->getName() << "'\n");
834
835 PreprocessISelDAG();
836
837 // Select target instructions for the DAG.
838 {
839 // Number all nodes with a topological order and set DAGSize.
840 DAGSize = CurDAG->AssignTopologicalOrder();
841
842 // Create a dummy node (which is not added to allnodes), that adds
843 // a reference to the root node, preventing it from being deleted,
844 // and tracking any changes of the root.
845 HandleSDNode Dummy(CurDAG->getRoot());
846 SelectionDAG::allnodes_iterator ISelPosition (CurDAG->getRoot().getNode());
847 ++ISelPosition;
848
849 // Make sure that ISelPosition gets properly updated when nodes are deleted
850 // in calls made from this function.
851 ISelUpdater ISU(*CurDAG, ISelPosition);
852
853 // The AllNodes list is now topological-sorted. Visit the
854 // nodes by starting at the end of the list (the root of the
855 // graph) and preceding back toward the beginning (the entry
856 // node).
857 while (ISelPosition != CurDAG->allnodes_begin()) {
858 SDNode *Node = --ISelPosition;
859 // Skip dead nodes. DAGCombiner is expected to eliminate all dead nodes,
860 // but there are currently some corner cases that it misses. Also, this
861 // makes it theoretically possible to disable the DAGCombiner.
862 if (Node->use_empty())
863 continue;
864
865 SDNode *ResNode = Select(Node);
866
867 // FIXME: This is pretty gross. 'Select' should be changed to not return
868 // anything at all and this code should be nuked with a tactical strike.
869
870 // If node should not be replaced, continue with the next one.
871 if (ResNode == Node || Node->getOpcode() == ISD::DELETED_NODE)
872 continue;
873 // Replace node.
874 if (ResNode) {
875 ReplaceUses(Node, ResNode);
876 }
877
878 // If after the replacement this node is not used any more,
879 // remove this dead node.
880 if (Node->use_empty()) // Don't delete EntryToken, etc.
881 CurDAG->RemoveDeadNode(Node);
882 }
883
884 CurDAG->setRoot(Dummy.getValue());
885 }
886
887 DEBUG(dbgs() << "===== Instruction selection ends:\n");
888
889 PostprocessISelDAG();
890 }
891
892 /// PrepareEHLandingPad - Emit an EH_LABEL, set up live-in registers, and
893 /// do other setup for EH landing-pad blocks.
PrepareEHLandingPad()894 void SelectionDAGISel::PrepareEHLandingPad() {
895 MachineBasicBlock *MBB = FuncInfo->MBB;
896
897 // Add a label to mark the beginning of the landing pad. Deletion of the
898 // landing pad can thus be detected via the MachineModuleInfo.
899 MCSymbol *Label = MF->getMMI().addLandingPad(MBB);
900
901 // Assign the call site to the landing pad's begin label.
902 MF->getMMI().setCallSiteLandingPad(Label, SDB->LPadToCallSiteMap[MBB]);
903
904 const MCInstrDesc &II = TM.getInstrInfo()->get(TargetOpcode::EH_LABEL);
905 BuildMI(*MBB, FuncInfo->InsertPt, SDB->getCurDebugLoc(), II)
906 .addSym(Label);
907
908 // Mark exception register as live in.
909 const TargetLowering *TLI = getTargetLowering();
910 const TargetRegisterClass *PtrRC = TLI->getRegClassFor(TLI->getPointerTy());
911 if (unsigned Reg = TLI->getExceptionPointerRegister())
912 FuncInfo->ExceptionPointerVirtReg = MBB->addLiveIn(Reg, PtrRC);
913
914 // Mark exception selector register as live in.
915 if (unsigned Reg = TLI->getExceptionSelectorRegister())
916 FuncInfo->ExceptionSelectorVirtReg = MBB->addLiveIn(Reg, PtrRC);
917 }
918
919 /// isFoldedOrDeadInstruction - Return true if the specified instruction is
920 /// side-effect free and is either dead or folded into a generated instruction.
921 /// Return false if it needs to be emitted.
isFoldedOrDeadInstruction(const Instruction * I,FunctionLoweringInfo * FuncInfo)922 static bool isFoldedOrDeadInstruction(const Instruction *I,
923 FunctionLoweringInfo *FuncInfo) {
924 return !I->mayWriteToMemory() && // Side-effecting instructions aren't folded.
925 !isa<TerminatorInst>(I) && // Terminators aren't folded.
926 !isa<DbgInfoIntrinsic>(I) && // Debug instructions aren't folded.
927 !isa<LandingPadInst>(I) && // Landingpad instructions aren't folded.
928 !FuncInfo->isExportedInst(I); // Exported instrs must be computed.
929 }
930
931 #ifndef NDEBUG
932 // Collect per Instruction statistics for fast-isel misses. Only those
933 // instructions that cause the bail are accounted for. It does not account for
934 // instructions higher in the block. Thus, summing the per instructions stats
935 // will not add up to what is reported by NumFastIselFailures.
collectFailStats(const Instruction * I)936 static void collectFailStats(const Instruction *I) {
937 switch (I->getOpcode()) {
938 default: assert (0 && "<Invalid operator> ");
939
940 // Terminators
941 case Instruction::Ret: NumFastIselFailRet++; return;
942 case Instruction::Br: NumFastIselFailBr++; return;
943 case Instruction::Switch: NumFastIselFailSwitch++; return;
944 case Instruction::IndirectBr: NumFastIselFailIndirectBr++; return;
945 case Instruction::Invoke: NumFastIselFailInvoke++; return;
946 case Instruction::Resume: NumFastIselFailResume++; return;
947 case Instruction::Unreachable: NumFastIselFailUnreachable++; return;
948
949 // Standard binary operators...
950 case Instruction::Add: NumFastIselFailAdd++; return;
951 case Instruction::FAdd: NumFastIselFailFAdd++; return;
952 case Instruction::Sub: NumFastIselFailSub++; return;
953 case Instruction::FSub: NumFastIselFailFSub++; return;
954 case Instruction::Mul: NumFastIselFailMul++; return;
955 case Instruction::FMul: NumFastIselFailFMul++; return;
956 case Instruction::UDiv: NumFastIselFailUDiv++; return;
957 case Instruction::SDiv: NumFastIselFailSDiv++; return;
958 case Instruction::FDiv: NumFastIselFailFDiv++; return;
959 case Instruction::URem: NumFastIselFailURem++; return;
960 case Instruction::SRem: NumFastIselFailSRem++; return;
961 case Instruction::FRem: NumFastIselFailFRem++; return;
962
963 // Logical operators...
964 case Instruction::And: NumFastIselFailAnd++; return;
965 case Instruction::Or: NumFastIselFailOr++; return;
966 case Instruction::Xor: NumFastIselFailXor++; return;
967
968 // Memory instructions...
969 case Instruction::Alloca: NumFastIselFailAlloca++; return;
970 case Instruction::Load: NumFastIselFailLoad++; return;
971 case Instruction::Store: NumFastIselFailStore++; return;
972 case Instruction::AtomicCmpXchg: NumFastIselFailAtomicCmpXchg++; return;
973 case Instruction::AtomicRMW: NumFastIselFailAtomicRMW++; return;
974 case Instruction::Fence: NumFastIselFailFence++; return;
975 case Instruction::GetElementPtr: NumFastIselFailGetElementPtr++; return;
976
977 // Convert instructions...
978 case Instruction::Trunc: NumFastIselFailTrunc++; return;
979 case Instruction::ZExt: NumFastIselFailZExt++; return;
980 case Instruction::SExt: NumFastIselFailSExt++; return;
981 case Instruction::FPTrunc: NumFastIselFailFPTrunc++; return;
982 case Instruction::FPExt: NumFastIselFailFPExt++; return;
983 case Instruction::FPToUI: NumFastIselFailFPToUI++; return;
984 case Instruction::FPToSI: NumFastIselFailFPToSI++; return;
985 case Instruction::UIToFP: NumFastIselFailUIToFP++; return;
986 case Instruction::SIToFP: NumFastIselFailSIToFP++; return;
987 case Instruction::IntToPtr: NumFastIselFailIntToPtr++; return;
988 case Instruction::PtrToInt: NumFastIselFailPtrToInt++; return;
989 case Instruction::BitCast: NumFastIselFailBitCast++; return;
990
991 // Other instructions...
992 case Instruction::ICmp: NumFastIselFailICmp++; return;
993 case Instruction::FCmp: NumFastIselFailFCmp++; return;
994 case Instruction::PHI: NumFastIselFailPHI++; return;
995 case Instruction::Select: NumFastIselFailSelect++; return;
996 case Instruction::Call: {
997 if (auto const *Intrinsic = dyn_cast<IntrinsicInst>(I)) {
998 switch (Intrinsic->getIntrinsicID()) {
999 default:
1000 NumFastIselFailIntrinsicCall++; return;
1001 case Intrinsic::sadd_with_overflow:
1002 NumFastIselFailSAddWithOverflow++; return;
1003 case Intrinsic::uadd_with_overflow:
1004 NumFastIselFailUAddWithOverflow++; return;
1005 case Intrinsic::ssub_with_overflow:
1006 NumFastIselFailSSubWithOverflow++; return;
1007 case Intrinsic::usub_with_overflow:
1008 NumFastIselFailUSubWithOverflow++; return;
1009 case Intrinsic::smul_with_overflow:
1010 NumFastIselFailSMulWithOverflow++; return;
1011 case Intrinsic::umul_with_overflow:
1012 NumFastIselFailUMulWithOverflow++; return;
1013 case Intrinsic::frameaddress:
1014 NumFastIselFailFrameaddress++; return;
1015 case Intrinsic::sqrt:
1016 NumFastIselFailSqrt++; return;
1017 case Intrinsic::experimental_stackmap:
1018 NumFastIselFailStackMap++; return;
1019 case Intrinsic::experimental_patchpoint_void: // fall-through
1020 case Intrinsic::experimental_patchpoint_i64:
1021 NumFastIselFailPatchPoint++; return;
1022 }
1023 }
1024 NumFastIselFailCall++;
1025 return;
1026 }
1027 case Instruction::Shl: NumFastIselFailShl++; return;
1028 case Instruction::LShr: NumFastIselFailLShr++; return;
1029 case Instruction::AShr: NumFastIselFailAShr++; return;
1030 case Instruction::VAArg: NumFastIselFailVAArg++; return;
1031 case Instruction::ExtractElement: NumFastIselFailExtractElement++; return;
1032 case Instruction::InsertElement: NumFastIselFailInsertElement++; return;
1033 case Instruction::ShuffleVector: NumFastIselFailShuffleVector++; return;
1034 case Instruction::ExtractValue: NumFastIselFailExtractValue++; return;
1035 case Instruction::InsertValue: NumFastIselFailInsertValue++; return;
1036 case Instruction::LandingPad: NumFastIselFailLandingPad++; return;
1037 }
1038 }
1039 #endif
1040
SelectAllBasicBlocks(const Function & Fn)1041 void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
1042 // Initialize the Fast-ISel state, if needed.
1043 FastISel *FastIS = nullptr;
1044 if (TM.Options.EnableFastISel)
1045 FastIS = getTargetLowering()->createFastISel(*FuncInfo, LibInfo);
1046
1047 // Iterate over all basic blocks in the function.
1048 ReversePostOrderTraversal<const Function*> RPOT(&Fn);
1049 for (ReversePostOrderTraversal<const Function*>::rpo_iterator
1050 I = RPOT.begin(), E = RPOT.end(); I != E; ++I) {
1051 const BasicBlock *LLVMBB = *I;
1052
1053 if (OptLevel != CodeGenOpt::None) {
1054 bool AllPredsVisited = true;
1055 for (const_pred_iterator PI = pred_begin(LLVMBB), PE = pred_end(LLVMBB);
1056 PI != PE; ++PI) {
1057 if (!FuncInfo->VisitedBBs.count(*PI)) {
1058 AllPredsVisited = false;
1059 break;
1060 }
1061 }
1062
1063 if (AllPredsVisited) {
1064 for (BasicBlock::const_iterator I = LLVMBB->begin();
1065 const PHINode *PN = dyn_cast<PHINode>(I); ++I)
1066 FuncInfo->ComputePHILiveOutRegInfo(PN);
1067 } else {
1068 for (BasicBlock::const_iterator I = LLVMBB->begin();
1069 const PHINode *PN = dyn_cast<PHINode>(I); ++I)
1070 FuncInfo->InvalidatePHILiveOutRegInfo(PN);
1071 }
1072
1073 FuncInfo->VisitedBBs.insert(LLVMBB);
1074 }
1075
1076 BasicBlock::const_iterator const Begin = LLVMBB->getFirstNonPHI();
1077 BasicBlock::const_iterator const End = LLVMBB->end();
1078 BasicBlock::const_iterator BI = End;
1079
1080 FuncInfo->MBB = FuncInfo->MBBMap[LLVMBB];
1081 FuncInfo->InsertPt = FuncInfo->MBB->getFirstNonPHI();
1082
1083 // Setup an EH landing-pad block.
1084 FuncInfo->ExceptionPointerVirtReg = 0;
1085 FuncInfo->ExceptionSelectorVirtReg = 0;
1086 if (FuncInfo->MBB->isLandingPad())
1087 PrepareEHLandingPad();
1088
1089 // Before doing SelectionDAG ISel, see if FastISel has been requested.
1090 if (FastIS) {
1091 FastIS->startNewBlock();
1092
1093 // Emit code for any incoming arguments. This must happen before
1094 // beginning FastISel on the entry block.
1095 if (LLVMBB == &Fn.getEntryBlock()) {
1096 ++NumEntryBlocks;
1097
1098 // Lower any arguments needed in this block if this is the entry block.
1099 if (!FastIS->LowerArguments()) {
1100 // Fast isel failed to lower these arguments
1101 ++NumFastIselFailLowerArguments;
1102 if (EnableFastISelAbortArgs)
1103 llvm_unreachable("FastISel didn't lower all arguments");
1104
1105 // Use SelectionDAG argument lowering
1106 LowerArguments(Fn);
1107 CurDAG->setRoot(SDB->getControlRoot());
1108 SDB->clear();
1109 CodeGenAndEmitDAG();
1110 }
1111
1112 // If we inserted any instructions at the beginning, make a note of
1113 // where they are, so we can be sure to emit subsequent instructions
1114 // after them.
1115 if (FuncInfo->InsertPt != FuncInfo->MBB->begin())
1116 FastIS->setLastLocalValue(std::prev(FuncInfo->InsertPt));
1117 else
1118 FastIS->setLastLocalValue(nullptr);
1119 }
1120
1121 unsigned NumFastIselRemaining = std::distance(Begin, End);
1122 // Do FastISel on as many instructions as possible.
1123 for (; BI != Begin; --BI) {
1124 const Instruction *Inst = std::prev(BI);
1125
1126 // If we no longer require this instruction, skip it.
1127 if (isFoldedOrDeadInstruction(Inst, FuncInfo)) {
1128 --NumFastIselRemaining;
1129 continue;
1130 }
1131
1132 // Bottom-up: reset the insert pos at the top, after any local-value
1133 // instructions.
1134 FastIS->recomputeInsertPt();
1135
1136 // Try to select the instruction with FastISel.
1137 if (FastIS->SelectInstruction(Inst)) {
1138 --NumFastIselRemaining;
1139 ++NumFastIselSuccess;
1140 // If fast isel succeeded, skip over all the folded instructions, and
1141 // then see if there is a load right before the selected instructions.
1142 // Try to fold the load if so.
1143 const Instruction *BeforeInst = Inst;
1144 while (BeforeInst != Begin) {
1145 BeforeInst = std::prev(BasicBlock::const_iterator(BeforeInst));
1146 if (!isFoldedOrDeadInstruction(BeforeInst, FuncInfo))
1147 break;
1148 }
1149 if (BeforeInst != Inst && isa<LoadInst>(BeforeInst) &&
1150 BeforeInst->hasOneUse() &&
1151 FastIS->tryToFoldLoad(cast<LoadInst>(BeforeInst), Inst)) {
1152 // If we succeeded, don't re-select the load.
1153 BI = std::next(BasicBlock::const_iterator(BeforeInst));
1154 --NumFastIselRemaining;
1155 ++NumFastIselSuccess;
1156 }
1157 continue;
1158 }
1159
1160 #ifndef NDEBUG
1161 if (EnableFastISelVerbose2)
1162 collectFailStats(Inst);
1163 #endif
1164
1165 // Then handle certain instructions as single-LLVM-Instruction blocks.
1166 if (isa<CallInst>(Inst)) {
1167
1168 if (EnableFastISelVerbose || EnableFastISelAbort) {
1169 dbgs() << "FastISel missed call: ";
1170 Inst->dump();
1171 }
1172
1173 if (!Inst->getType()->isVoidTy() && !Inst->use_empty()) {
1174 unsigned &R = FuncInfo->ValueMap[Inst];
1175 if (!R)
1176 R = FuncInfo->CreateRegs(Inst->getType());
1177 }
1178
1179 bool HadTailCall = false;
1180 MachineBasicBlock::iterator SavedInsertPt = FuncInfo->InsertPt;
1181 SelectBasicBlock(Inst, BI, HadTailCall);
1182
1183 // If the call was emitted as a tail call, we're done with the block.
1184 // We also need to delete any previously emitted instructions.
1185 if (HadTailCall) {
1186 FastIS->removeDeadCode(SavedInsertPt, FuncInfo->MBB->end());
1187 --BI;
1188 break;
1189 }
1190
1191 // Recompute NumFastIselRemaining as Selection DAG instruction
1192 // selection may have handled the call, input args, etc.
1193 unsigned RemainingNow = std::distance(Begin, BI);
1194 NumFastIselFailures += NumFastIselRemaining - RemainingNow;
1195 NumFastIselRemaining = RemainingNow;
1196 continue;
1197 }
1198
1199 if (isa<TerminatorInst>(Inst) && !isa<BranchInst>(Inst)) {
1200 // Don't abort, and use a different message for terminator misses.
1201 NumFastIselFailures += NumFastIselRemaining;
1202 if (EnableFastISelVerbose || EnableFastISelAbort) {
1203 dbgs() << "FastISel missed terminator: ";
1204 Inst->dump();
1205 }
1206 } else {
1207 NumFastIselFailures += NumFastIselRemaining;
1208 if (EnableFastISelVerbose || EnableFastISelAbort) {
1209 dbgs() << "FastISel miss: ";
1210 Inst->dump();
1211 }
1212 if (EnableFastISelAbort)
1213 // The "fast" selector couldn't handle something and bailed.
1214 // For the purpose of debugging, just abort.
1215 llvm_unreachable("FastISel didn't select the entire block");
1216 }
1217 break;
1218 }
1219
1220 FastIS->recomputeInsertPt();
1221 } else {
1222 // Lower any arguments needed in this block if this is the entry block.
1223 if (LLVMBB == &Fn.getEntryBlock()) {
1224 ++NumEntryBlocks;
1225 LowerArguments(Fn);
1226 }
1227 }
1228
1229 if (Begin != BI)
1230 ++NumDAGBlocks;
1231 else
1232 ++NumFastIselBlocks;
1233
1234 if (Begin != BI) {
1235 // Run SelectionDAG instruction selection on the remainder of the block
1236 // not handled by FastISel. If FastISel is not run, this is the entire
1237 // block.
1238 bool HadTailCall;
1239 SelectBasicBlock(Begin, BI, HadTailCall);
1240 }
1241
1242 FinishBasicBlock();
1243 FuncInfo->PHINodesToUpdate.clear();
1244 }
1245
1246 delete FastIS;
1247 SDB->clearDanglingDebugInfo();
1248 SDB->SPDescriptor.resetPerFunctionState();
1249 }
1250
1251 /// Given that the input MI is before a partial terminator sequence TSeq, return
1252 /// true if M + TSeq also a partial terminator sequence.
1253 ///
1254 /// A Terminator sequence is a sequence of MachineInstrs which at this point in
1255 /// lowering copy vregs into physical registers, which are then passed into
1256 /// terminator instructors so we can satisfy ABI constraints. A partial
1257 /// terminator sequence is an improper subset of a terminator sequence (i.e. it
1258 /// may be the whole terminator sequence).
MIIsInTerminatorSequence(const MachineInstr * MI)1259 static bool MIIsInTerminatorSequence(const MachineInstr *MI) {
1260 // If we do not have a copy or an implicit def, we return true if and only if
1261 // MI is a debug value.
1262 if (!MI->isCopy() && !MI->isImplicitDef())
1263 // Sometimes DBG_VALUE MI sneak in between the copies from the vregs to the
1264 // physical registers if there is debug info associated with the terminator
1265 // of our mbb. We want to include said debug info in our terminator
1266 // sequence, so we return true in that case.
1267 return MI->isDebugValue();
1268
1269 // We have left the terminator sequence if we are not doing one of the
1270 // following:
1271 //
1272 // 1. Copying a vreg into a physical register.
1273 // 2. Copying a vreg into a vreg.
1274 // 3. Defining a register via an implicit def.
1275
1276 // OPI should always be a register definition...
1277 MachineInstr::const_mop_iterator OPI = MI->operands_begin();
1278 if (!OPI->isReg() || !OPI->isDef())
1279 return false;
1280
1281 // Defining any register via an implicit def is always ok.
1282 if (MI->isImplicitDef())
1283 return true;
1284
1285 // Grab the copy source...
1286 MachineInstr::const_mop_iterator OPI2 = OPI;
1287 ++OPI2;
1288 assert(OPI2 != MI->operands_end()
1289 && "Should have a copy implying we should have 2 arguments.");
1290
1291 // Make sure that the copy dest is not a vreg when the copy source is a
1292 // physical register.
1293 if (!OPI2->isReg() ||
1294 (!TargetRegisterInfo::isPhysicalRegister(OPI->getReg()) &&
1295 TargetRegisterInfo::isPhysicalRegister(OPI2->getReg())))
1296 return false;
1297
1298 return true;
1299 }
1300
1301 /// Find the split point at which to splice the end of BB into its success stack
1302 /// protector check machine basic block.
1303 ///
1304 /// On many platforms, due to ABI constraints, terminators, even before register
1305 /// allocation, use physical registers. This creates an issue for us since
1306 /// physical registers at this point can not travel across basic
1307 /// blocks. Luckily, selectiondag always moves physical registers into vregs
1308 /// when they enter functions and moves them through a sequence of copies back
1309 /// into the physical registers right before the terminator creating a
1310 /// ``Terminator Sequence''. This function is searching for the beginning of the
1311 /// terminator sequence so that we can ensure that we splice off not just the
1312 /// terminator, but additionally the copies that move the vregs into the
1313 /// physical registers.
1314 static MachineBasicBlock::iterator
FindSplitPointForStackProtector(MachineBasicBlock * BB,DebugLoc DL)1315 FindSplitPointForStackProtector(MachineBasicBlock *BB, DebugLoc DL) {
1316 MachineBasicBlock::iterator SplitPoint = BB->getFirstTerminator();
1317 //
1318 if (SplitPoint == BB->begin())
1319 return SplitPoint;
1320
1321 MachineBasicBlock::iterator Start = BB->begin();
1322 MachineBasicBlock::iterator Previous = SplitPoint;
1323 --Previous;
1324
1325 while (MIIsInTerminatorSequence(Previous)) {
1326 SplitPoint = Previous;
1327 if (Previous == Start)
1328 break;
1329 --Previous;
1330 }
1331
1332 return SplitPoint;
1333 }
1334
1335 void
FinishBasicBlock()1336 SelectionDAGISel::FinishBasicBlock() {
1337
1338 DEBUG(dbgs() << "Total amount of phi nodes to update: "
1339 << FuncInfo->PHINodesToUpdate.size() << "\n";
1340 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i)
1341 dbgs() << "Node " << i << " : ("
1342 << FuncInfo->PHINodesToUpdate[i].first
1343 << ", " << FuncInfo->PHINodesToUpdate[i].second << ")\n");
1344
1345 const bool MustUpdatePHINodes = SDB->SwitchCases.empty() &&
1346 SDB->JTCases.empty() &&
1347 SDB->BitTestCases.empty();
1348
1349 // Next, now that we know what the last MBB the LLVM BB expanded is, update
1350 // PHI nodes in successors.
1351 if (MustUpdatePHINodes) {
1352 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i) {
1353 MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[i].first);
1354 assert(PHI->isPHI() &&
1355 "This is not a machine PHI node that we are updating!");
1356 if (!FuncInfo->MBB->isSuccessor(PHI->getParent()))
1357 continue;
1358 PHI.addReg(FuncInfo->PHINodesToUpdate[i].second).addMBB(FuncInfo->MBB);
1359 }
1360 }
1361
1362 // Handle stack protector.
1363 if (SDB->SPDescriptor.shouldEmitStackProtector()) {
1364 MachineBasicBlock *ParentMBB = SDB->SPDescriptor.getParentMBB();
1365 MachineBasicBlock *SuccessMBB = SDB->SPDescriptor.getSuccessMBB();
1366
1367 // Find the split point to split the parent mbb. At the same time copy all
1368 // physical registers used in the tail of parent mbb into virtual registers
1369 // before the split point and back into physical registers after the split
1370 // point. This prevents us needing to deal with Live-ins and many other
1371 // register allocation issues caused by us splitting the parent mbb. The
1372 // register allocator will clean up said virtual copies later on.
1373 MachineBasicBlock::iterator SplitPoint =
1374 FindSplitPointForStackProtector(ParentMBB, SDB->getCurDebugLoc());
1375
1376 // Splice the terminator of ParentMBB into SuccessMBB.
1377 SuccessMBB->splice(SuccessMBB->end(), ParentMBB,
1378 SplitPoint,
1379 ParentMBB->end());
1380
1381 // Add compare/jump on neq/jump to the parent BB.
1382 FuncInfo->MBB = ParentMBB;
1383 FuncInfo->InsertPt = ParentMBB->end();
1384 SDB->visitSPDescriptorParent(SDB->SPDescriptor, ParentMBB);
1385 CurDAG->setRoot(SDB->getRoot());
1386 SDB->clear();
1387 CodeGenAndEmitDAG();
1388
1389 // CodeGen Failure MBB if we have not codegened it yet.
1390 MachineBasicBlock *FailureMBB = SDB->SPDescriptor.getFailureMBB();
1391 if (!FailureMBB->size()) {
1392 FuncInfo->MBB = FailureMBB;
1393 FuncInfo->InsertPt = FailureMBB->end();
1394 SDB->visitSPDescriptorFailure(SDB->SPDescriptor);
1395 CurDAG->setRoot(SDB->getRoot());
1396 SDB->clear();
1397 CodeGenAndEmitDAG();
1398 }
1399
1400 // Clear the Per-BB State.
1401 SDB->SPDescriptor.resetPerBBState();
1402 }
1403
1404 // If we updated PHI Nodes, return early.
1405 if (MustUpdatePHINodes)
1406 return;
1407
1408 for (unsigned i = 0, e = SDB->BitTestCases.size(); i != e; ++i) {
1409 // Lower header first, if it wasn't already lowered
1410 if (!SDB->BitTestCases[i].Emitted) {
1411 // Set the current basic block to the mbb we wish to insert the code into
1412 FuncInfo->MBB = SDB->BitTestCases[i].Parent;
1413 FuncInfo->InsertPt = FuncInfo->MBB->end();
1414 // Emit the code
1415 SDB->visitBitTestHeader(SDB->BitTestCases[i], FuncInfo->MBB);
1416 CurDAG->setRoot(SDB->getRoot());
1417 SDB->clear();
1418 CodeGenAndEmitDAG();
1419 }
1420
1421 uint32_t UnhandledWeight = 0;
1422 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j)
1423 UnhandledWeight += SDB->BitTestCases[i].Cases[j].ExtraWeight;
1424
1425 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j) {
1426 UnhandledWeight -= SDB->BitTestCases[i].Cases[j].ExtraWeight;
1427 // Set the current basic block to the mbb we wish to insert the code into
1428 FuncInfo->MBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1429 FuncInfo->InsertPt = FuncInfo->MBB->end();
1430 // Emit the code
1431 if (j+1 != ej)
1432 SDB->visitBitTestCase(SDB->BitTestCases[i],
1433 SDB->BitTestCases[i].Cases[j+1].ThisBB,
1434 UnhandledWeight,
1435 SDB->BitTestCases[i].Reg,
1436 SDB->BitTestCases[i].Cases[j],
1437 FuncInfo->MBB);
1438 else
1439 SDB->visitBitTestCase(SDB->BitTestCases[i],
1440 SDB->BitTestCases[i].Default,
1441 UnhandledWeight,
1442 SDB->BitTestCases[i].Reg,
1443 SDB->BitTestCases[i].Cases[j],
1444 FuncInfo->MBB);
1445
1446
1447 CurDAG->setRoot(SDB->getRoot());
1448 SDB->clear();
1449 CodeGenAndEmitDAG();
1450 }
1451
1452 // Update PHI Nodes
1453 for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1454 pi != pe; ++pi) {
1455 MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[pi].first);
1456 MachineBasicBlock *PHIBB = PHI->getParent();
1457 assert(PHI->isPHI() &&
1458 "This is not a machine PHI node that we are updating!");
1459 // This is "default" BB. We have two jumps to it. From "header" BB and
1460 // from last "case" BB.
1461 if (PHIBB == SDB->BitTestCases[i].Default)
1462 PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second)
1463 .addMBB(SDB->BitTestCases[i].Parent)
1464 .addReg(FuncInfo->PHINodesToUpdate[pi].second)
1465 .addMBB(SDB->BitTestCases[i].Cases.back().ThisBB);
1466 // One of "cases" BB.
1467 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size();
1468 j != ej; ++j) {
1469 MachineBasicBlock* cBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1470 if (cBB->isSuccessor(PHIBB))
1471 PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second).addMBB(cBB);
1472 }
1473 }
1474 }
1475 SDB->BitTestCases.clear();
1476
1477 // If the JumpTable record is filled in, then we need to emit a jump table.
1478 // Updating the PHI nodes is tricky in this case, since we need to determine
1479 // whether the PHI is a successor of the range check MBB or the jump table MBB
1480 for (unsigned i = 0, e = SDB->JTCases.size(); i != e; ++i) {
1481 // Lower header first, if it wasn't already lowered
1482 if (!SDB->JTCases[i].first.Emitted) {
1483 // Set the current basic block to the mbb we wish to insert the code into
1484 FuncInfo->MBB = SDB->JTCases[i].first.HeaderBB;
1485 FuncInfo->InsertPt = FuncInfo->MBB->end();
1486 // Emit the code
1487 SDB->visitJumpTableHeader(SDB->JTCases[i].second, SDB->JTCases[i].first,
1488 FuncInfo->MBB);
1489 CurDAG->setRoot(SDB->getRoot());
1490 SDB->clear();
1491 CodeGenAndEmitDAG();
1492 }
1493
1494 // Set the current basic block to the mbb we wish to insert the code into
1495 FuncInfo->MBB = SDB->JTCases[i].second.MBB;
1496 FuncInfo->InsertPt = FuncInfo->MBB->end();
1497 // Emit the code
1498 SDB->visitJumpTable(SDB->JTCases[i].second);
1499 CurDAG->setRoot(SDB->getRoot());
1500 SDB->clear();
1501 CodeGenAndEmitDAG();
1502
1503 // Update PHI Nodes
1504 for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1505 pi != pe; ++pi) {
1506 MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[pi].first);
1507 MachineBasicBlock *PHIBB = PHI->getParent();
1508 assert(PHI->isPHI() &&
1509 "This is not a machine PHI node that we are updating!");
1510 // "default" BB. We can go there only from header BB.
1511 if (PHIBB == SDB->JTCases[i].second.Default)
1512 PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second)
1513 .addMBB(SDB->JTCases[i].first.HeaderBB);
1514 // JT BB. Just iterate over successors here
1515 if (FuncInfo->MBB->isSuccessor(PHIBB))
1516 PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second).addMBB(FuncInfo->MBB);
1517 }
1518 }
1519 SDB->JTCases.clear();
1520
1521 // If the switch block involved a branch to one of the actual successors, we
1522 // need to update PHI nodes in that block.
1523 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i) {
1524 MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[i].first);
1525 assert(PHI->isPHI() &&
1526 "This is not a machine PHI node that we are updating!");
1527 if (FuncInfo->MBB->isSuccessor(PHI->getParent()))
1528 PHI.addReg(FuncInfo->PHINodesToUpdate[i].second).addMBB(FuncInfo->MBB);
1529 }
1530
1531 // If we generated any switch lowering information, build and codegen any
1532 // additional DAGs necessary.
1533 for (unsigned i = 0, e = SDB->SwitchCases.size(); i != e; ++i) {
1534 // Set the current basic block to the mbb we wish to insert the code into
1535 FuncInfo->MBB = SDB->SwitchCases[i].ThisBB;
1536 FuncInfo->InsertPt = FuncInfo->MBB->end();
1537
1538 // Determine the unique successors.
1539 SmallVector<MachineBasicBlock *, 2> Succs;
1540 Succs.push_back(SDB->SwitchCases[i].TrueBB);
1541 if (SDB->SwitchCases[i].TrueBB != SDB->SwitchCases[i].FalseBB)
1542 Succs.push_back(SDB->SwitchCases[i].FalseBB);
1543
1544 // Emit the code. Note that this could result in FuncInfo->MBB being split.
1545 SDB->visitSwitchCase(SDB->SwitchCases[i], FuncInfo->MBB);
1546 CurDAG->setRoot(SDB->getRoot());
1547 SDB->clear();
1548 CodeGenAndEmitDAG();
1549
1550 // Remember the last block, now that any splitting is done, for use in
1551 // populating PHI nodes in successors.
1552 MachineBasicBlock *ThisBB = FuncInfo->MBB;
1553
1554 // Handle any PHI nodes in successors of this chunk, as if we were coming
1555 // from the original BB before switch expansion. Note that PHI nodes can
1556 // occur multiple times in PHINodesToUpdate. We have to be very careful to
1557 // handle them the right number of times.
1558 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1559 FuncInfo->MBB = Succs[i];
1560 FuncInfo->InsertPt = FuncInfo->MBB->end();
1561 // FuncInfo->MBB may have been removed from the CFG if a branch was
1562 // constant folded.
1563 if (ThisBB->isSuccessor(FuncInfo->MBB)) {
1564 for (MachineBasicBlock::iterator
1565 MBBI = FuncInfo->MBB->begin(), MBBE = FuncInfo->MBB->end();
1566 MBBI != MBBE && MBBI->isPHI(); ++MBBI) {
1567 MachineInstrBuilder PHI(*MF, MBBI);
1568 // This value for this PHI node is recorded in PHINodesToUpdate.
1569 for (unsigned pn = 0; ; ++pn) {
1570 assert(pn != FuncInfo->PHINodesToUpdate.size() &&
1571 "Didn't find PHI entry!");
1572 if (FuncInfo->PHINodesToUpdate[pn].first == PHI) {
1573 PHI.addReg(FuncInfo->PHINodesToUpdate[pn].second).addMBB(ThisBB);
1574 break;
1575 }
1576 }
1577 }
1578 }
1579 }
1580 }
1581 SDB->SwitchCases.clear();
1582 }
1583
1584
1585 /// Create the scheduler. If a specific scheduler was specified
1586 /// via the SchedulerRegistry, use it, otherwise select the
1587 /// one preferred by the target.
1588 ///
CreateScheduler()1589 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
1590 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1591
1592 if (!Ctor) {
1593 Ctor = ISHeuristic;
1594 RegisterScheduler::setDefault(Ctor);
1595 }
1596
1597 return Ctor(this, OptLevel);
1598 }
1599
1600 //===----------------------------------------------------------------------===//
1601 // Helper functions used by the generated instruction selector.
1602 //===----------------------------------------------------------------------===//
1603 // Calls to these methods are generated by tblgen.
1604
1605 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1606 /// the dag combiner simplified the 255, we still want to match. RHS is the
1607 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1608 /// specified in the .td file (e.g. 255).
CheckAndMask(SDValue LHS,ConstantSDNode * RHS,int64_t DesiredMaskS) const1609 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1610 int64_t DesiredMaskS) const {
1611 const APInt &ActualMask = RHS->getAPIntValue();
1612 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1613
1614 // If the actual mask exactly matches, success!
1615 if (ActualMask == DesiredMask)
1616 return true;
1617
1618 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1619 if (ActualMask.intersects(~DesiredMask))
1620 return false;
1621
1622 // Otherwise, the DAG Combiner may have proven that the value coming in is
1623 // either already zero or is not demanded. Check for known zero input bits.
1624 APInt NeededMask = DesiredMask & ~ActualMask;
1625 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1626 return true;
1627
1628 // TODO: check to see if missing bits are just not demanded.
1629
1630 // Otherwise, this pattern doesn't match.
1631 return false;
1632 }
1633
1634 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1635 /// the dag combiner simplified the 255, we still want to match. RHS is the
1636 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1637 /// specified in the .td file (e.g. 255).
CheckOrMask(SDValue LHS,ConstantSDNode * RHS,int64_t DesiredMaskS) const1638 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1639 int64_t DesiredMaskS) const {
1640 const APInt &ActualMask = RHS->getAPIntValue();
1641 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1642
1643 // If the actual mask exactly matches, success!
1644 if (ActualMask == DesiredMask)
1645 return true;
1646
1647 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1648 if (ActualMask.intersects(~DesiredMask))
1649 return false;
1650
1651 // Otherwise, the DAG Combiner may have proven that the value coming in is
1652 // either already zero or is not demanded. Check for known zero input bits.
1653 APInt NeededMask = DesiredMask & ~ActualMask;
1654
1655 APInt KnownZero, KnownOne;
1656 CurDAG->computeKnownBits(LHS, KnownZero, KnownOne);
1657
1658 // If all the missing bits in the or are already known to be set, match!
1659 if ((NeededMask & KnownOne) == NeededMask)
1660 return true;
1661
1662 // TODO: check to see if missing bits are just not demanded.
1663
1664 // Otherwise, this pattern doesn't match.
1665 return false;
1666 }
1667
1668
1669 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1670 /// by tblgen. Others should not call it.
1671 void SelectionDAGISel::
SelectInlineAsmMemoryOperands(std::vector<SDValue> & Ops)1672 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1673 std::vector<SDValue> InOps;
1674 std::swap(InOps, Ops);
1675
1676 Ops.push_back(InOps[InlineAsm::Op_InputChain]); // 0
1677 Ops.push_back(InOps[InlineAsm::Op_AsmString]); // 1
1678 Ops.push_back(InOps[InlineAsm::Op_MDNode]); // 2, !srcloc
1679 Ops.push_back(InOps[InlineAsm::Op_ExtraInfo]); // 3 (SideEffect, AlignStack)
1680
1681 unsigned i = InlineAsm::Op_FirstOperand, e = InOps.size();
1682 if (InOps[e-1].getValueType() == MVT::Glue)
1683 --e; // Don't process a glue operand if it is here.
1684
1685 while (i != e) {
1686 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1687 if (!InlineAsm::isMemKind(Flags)) {
1688 // Just skip over this operand, copying the operands verbatim.
1689 Ops.insert(Ops.end(), InOps.begin()+i,
1690 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
1691 i += InlineAsm::getNumOperandRegisters(Flags) + 1;
1692 } else {
1693 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 &&
1694 "Memory operand with multiple values?");
1695 // Otherwise, this is a memory operand. Ask the target to select it.
1696 std::vector<SDValue> SelOps;
1697 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps))
1698 report_fatal_error("Could not match memory address. Inline asm"
1699 " failure!");
1700
1701 // Add this to the output node.
1702 unsigned NewFlags =
1703 InlineAsm::getFlagWord(InlineAsm::Kind_Mem, SelOps.size());
1704 Ops.push_back(CurDAG->getTargetConstant(NewFlags, MVT::i32));
1705 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1706 i += 2;
1707 }
1708 }
1709
1710 // Add the glue input back if present.
1711 if (e != InOps.size())
1712 Ops.push_back(InOps.back());
1713 }
1714
1715 /// findGlueUse - Return use of MVT::Glue value produced by the specified
1716 /// SDNode.
1717 ///
findGlueUse(SDNode * N)1718 static SDNode *findGlueUse(SDNode *N) {
1719 unsigned FlagResNo = N->getNumValues()-1;
1720 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1721 SDUse &Use = I.getUse();
1722 if (Use.getResNo() == FlagResNo)
1723 return Use.getUser();
1724 }
1725 return nullptr;
1726 }
1727
1728 /// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
1729 /// This function recursively traverses up the operand chain, ignoring
1730 /// certain nodes.
findNonImmUse(SDNode * Use,SDNode * Def,SDNode * ImmedUse,SDNode * Root,SmallPtrSet<SDNode *,16> & Visited,bool IgnoreChains)1731 static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
1732 SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited,
1733 bool IgnoreChains) {
1734 // The NodeID's are given uniques ID's where a node ID is guaranteed to be
1735 // greater than all of its (recursive) operands. If we scan to a point where
1736 // 'use' is smaller than the node we're scanning for, then we know we will
1737 // never find it.
1738 //
1739 // The Use may be -1 (unassigned) if it is a newly allocated node. This can
1740 // happen because we scan down to newly selected nodes in the case of glue
1741 // uses.
1742 if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1))
1743 return false;
1744
1745 // Don't revisit nodes if we already scanned it and didn't fail, we know we
1746 // won't fail if we scan it again.
1747 if (!Visited.insert(Use))
1748 return false;
1749
1750 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
1751 // Ignore chain uses, they are validated by HandleMergeInputChains.
1752 if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains)
1753 continue;
1754
1755 SDNode *N = Use->getOperand(i).getNode();
1756 if (N == Def) {
1757 if (Use == ImmedUse || Use == Root)
1758 continue; // We are not looking for immediate use.
1759 assert(N != Root);
1760 return true;
1761 }
1762
1763 // Traverse up the operand chain.
1764 if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
1765 return true;
1766 }
1767 return false;
1768 }
1769
1770 /// IsProfitableToFold - Returns true if it's profitable to fold the specific
1771 /// operand node N of U during instruction selection that starts at Root.
IsProfitableToFold(SDValue N,SDNode * U,SDNode * Root) const1772 bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
1773 SDNode *Root) const {
1774 if (OptLevel == CodeGenOpt::None) return false;
1775 return N.hasOneUse();
1776 }
1777
1778 /// IsLegalToFold - Returns true if the specific operand node N of
1779 /// U can be folded during instruction selection that starts at Root.
IsLegalToFold(SDValue N,SDNode * U,SDNode * Root,CodeGenOpt::Level OptLevel,bool IgnoreChains)1780 bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
1781 CodeGenOpt::Level OptLevel,
1782 bool IgnoreChains) {
1783 if (OptLevel == CodeGenOpt::None) return false;
1784
1785 // If Root use can somehow reach N through a path that that doesn't contain
1786 // U then folding N would create a cycle. e.g. In the following
1787 // diagram, Root can reach N through X. If N is folded into into Root, then
1788 // X is both a predecessor and a successor of U.
1789 //
1790 // [N*] //
1791 // ^ ^ //
1792 // / \ //
1793 // [U*] [X]? //
1794 // ^ ^ //
1795 // \ / //
1796 // \ / //
1797 // [Root*] //
1798 //
1799 // * indicates nodes to be folded together.
1800 //
1801 // If Root produces glue, then it gets (even more) interesting. Since it
1802 // will be "glued" together with its glue use in the scheduler, we need to
1803 // check if it might reach N.
1804 //
1805 // [N*] //
1806 // ^ ^ //
1807 // / \ //
1808 // [U*] [X]? //
1809 // ^ ^ //
1810 // \ \ //
1811 // \ | //
1812 // [Root*] | //
1813 // ^ | //
1814 // f | //
1815 // | / //
1816 // [Y] / //
1817 // ^ / //
1818 // f / //
1819 // | / //
1820 // [GU] //
1821 //
1822 // If GU (glue use) indirectly reaches N (the load), and Root folds N
1823 // (call it Fold), then X is a predecessor of GU and a successor of
1824 // Fold. But since Fold and GU are glued together, this will create
1825 // a cycle in the scheduling graph.
1826
1827 // If the node has glue, walk down the graph to the "lowest" node in the
1828 // glueged set.
1829 EVT VT = Root->getValueType(Root->getNumValues()-1);
1830 while (VT == MVT::Glue) {
1831 SDNode *GU = findGlueUse(Root);
1832 if (!GU)
1833 break;
1834 Root = GU;
1835 VT = Root->getValueType(Root->getNumValues()-1);
1836
1837 // If our query node has a glue result with a use, we've walked up it. If
1838 // the user (which has already been selected) has a chain or indirectly uses
1839 // the chain, our WalkChainUsers predicate will not consider it. Because of
1840 // this, we cannot ignore chains in this predicate.
1841 IgnoreChains = false;
1842 }
1843
1844
1845 SmallPtrSet<SDNode*, 16> Visited;
1846 return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains);
1847 }
1848
Select_INLINEASM(SDNode * N)1849 SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
1850 std::vector<SDValue> Ops(N->op_begin(), N->op_end());
1851 SelectInlineAsmMemoryOperands(Ops);
1852
1853 EVT VTs[] = { MVT::Other, MVT::Glue };
1854 SDValue New = CurDAG->getNode(ISD::INLINEASM, SDLoc(N), VTs, Ops);
1855 New->setNodeId(-1);
1856 return New.getNode();
1857 }
1858
1859 SDNode
Select_READ_REGISTER(SDNode * Op)1860 *SelectionDAGISel::Select_READ_REGISTER(SDNode *Op) {
1861 SDLoc dl(Op);
1862 MDNodeSDNode *MD = dyn_cast<MDNodeSDNode>(Op->getOperand(0));
1863 const MDString *RegStr = dyn_cast<MDString>(MD->getMD()->getOperand(0));
1864 unsigned Reg = getTargetLowering()->getRegisterByName(
1865 RegStr->getString().data(), Op->getValueType(0));
1866 SDValue New = CurDAG->getCopyFromReg(
1867 CurDAG->getEntryNode(), dl, Reg, Op->getValueType(0));
1868 New->setNodeId(-1);
1869 return New.getNode();
1870 }
1871
1872 SDNode
Select_WRITE_REGISTER(SDNode * Op)1873 *SelectionDAGISel::Select_WRITE_REGISTER(SDNode *Op) {
1874 SDLoc dl(Op);
1875 MDNodeSDNode *MD = dyn_cast<MDNodeSDNode>(Op->getOperand(1));
1876 const MDString *RegStr = dyn_cast<MDString>(MD->getMD()->getOperand(0));
1877 unsigned Reg = getTargetLowering()->getRegisterByName(
1878 RegStr->getString().data(), Op->getOperand(2).getValueType());
1879 SDValue New = CurDAG->getCopyToReg(
1880 CurDAG->getEntryNode(), dl, Reg, Op->getOperand(2));
1881 New->setNodeId(-1);
1882 return New.getNode();
1883 }
1884
1885
1886
Select_UNDEF(SDNode * N)1887 SDNode *SelectionDAGISel::Select_UNDEF(SDNode *N) {
1888 return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0));
1889 }
1890
1891 /// GetVBR - decode a vbr encoding whose top bit is set.
1892 LLVM_ATTRIBUTE_ALWAYS_INLINE static uint64_t
GetVBR(uint64_t Val,const unsigned char * MatcherTable,unsigned & Idx)1893 GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
1894 assert(Val >= 128 && "Not a VBR");
1895 Val &= 127; // Remove first vbr bit.
1896
1897 unsigned Shift = 7;
1898 uint64_t NextBits;
1899 do {
1900 NextBits = MatcherTable[Idx++];
1901 Val |= (NextBits&127) << Shift;
1902 Shift += 7;
1903 } while (NextBits & 128);
1904
1905 return Val;
1906 }
1907
1908
1909 /// UpdateChainsAndGlue - When a match is complete, this method updates uses of
1910 /// interior glue and chain results to use the new glue and chain results.
1911 void SelectionDAGISel::
UpdateChainsAndGlue(SDNode * NodeToMatch,SDValue InputChain,const SmallVectorImpl<SDNode * > & ChainNodesMatched,SDValue InputGlue,const SmallVectorImpl<SDNode * > & GlueResultNodesMatched,bool isMorphNodeTo)1912 UpdateChainsAndGlue(SDNode *NodeToMatch, SDValue InputChain,
1913 const SmallVectorImpl<SDNode*> &ChainNodesMatched,
1914 SDValue InputGlue,
1915 const SmallVectorImpl<SDNode*> &GlueResultNodesMatched,
1916 bool isMorphNodeTo) {
1917 SmallVector<SDNode*, 4> NowDeadNodes;
1918
1919 // Now that all the normal results are replaced, we replace the chain and
1920 // glue results if present.
1921 if (!ChainNodesMatched.empty()) {
1922 assert(InputChain.getNode() &&
1923 "Matched input chains but didn't produce a chain");
1924 // Loop over all of the nodes we matched that produced a chain result.
1925 // Replace all the chain results with the final chain we ended up with.
1926 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1927 SDNode *ChainNode = ChainNodesMatched[i];
1928
1929 // If this node was already deleted, don't look at it.
1930 if (ChainNode->getOpcode() == ISD::DELETED_NODE)
1931 continue;
1932
1933 // Don't replace the results of the root node if we're doing a
1934 // MorphNodeTo.
1935 if (ChainNode == NodeToMatch && isMorphNodeTo)
1936 continue;
1937
1938 SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1);
1939 if (ChainVal.getValueType() == MVT::Glue)
1940 ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
1941 assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
1942 CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain);
1943
1944 // If the node became dead and we haven't already seen it, delete it.
1945 if (ChainNode->use_empty() &&
1946 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), ChainNode))
1947 NowDeadNodes.push_back(ChainNode);
1948 }
1949 }
1950
1951 // If the result produces glue, update any glue results in the matched
1952 // pattern with the glue result.
1953 if (InputGlue.getNode()) {
1954 // Handle any interior nodes explicitly marked.
1955 for (unsigned i = 0, e = GlueResultNodesMatched.size(); i != e; ++i) {
1956 SDNode *FRN = GlueResultNodesMatched[i];
1957
1958 // If this node was already deleted, don't look at it.
1959 if (FRN->getOpcode() == ISD::DELETED_NODE)
1960 continue;
1961
1962 assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Glue &&
1963 "Doesn't have a glue result");
1964 CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1),
1965 InputGlue);
1966
1967 // If the node became dead and we haven't already seen it, delete it.
1968 if (FRN->use_empty() &&
1969 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), FRN))
1970 NowDeadNodes.push_back(FRN);
1971 }
1972 }
1973
1974 if (!NowDeadNodes.empty())
1975 CurDAG->RemoveDeadNodes(NowDeadNodes);
1976
1977 DEBUG(dbgs() << "ISEL: Match complete!\n");
1978 }
1979
1980 enum ChainResult {
1981 CR_Simple,
1982 CR_InducesCycle,
1983 CR_LeadsToInteriorNode
1984 };
1985
1986 /// WalkChainUsers - Walk down the users of the specified chained node that is
1987 /// part of the pattern we're matching, looking at all of the users we find.
1988 /// This determines whether something is an interior node, whether we have a
1989 /// non-pattern node in between two pattern nodes (which prevent folding because
1990 /// it would induce a cycle) and whether we have a TokenFactor node sandwiched
1991 /// between pattern nodes (in which case the TF becomes part of the pattern).
1992 ///
1993 /// The walk we do here is guaranteed to be small because we quickly get down to
1994 /// already selected nodes "below" us.
1995 static ChainResult
WalkChainUsers(const SDNode * ChainedNode,SmallVectorImpl<SDNode * > & ChainedNodesInPattern,SmallVectorImpl<SDNode * > & InteriorChainedNodes)1996 WalkChainUsers(const SDNode *ChainedNode,
1997 SmallVectorImpl<SDNode*> &ChainedNodesInPattern,
1998 SmallVectorImpl<SDNode*> &InteriorChainedNodes) {
1999 ChainResult Result = CR_Simple;
2000
2001 for (SDNode::use_iterator UI = ChainedNode->use_begin(),
2002 E = ChainedNode->use_end(); UI != E; ++UI) {
2003 // Make sure the use is of the chain, not some other value we produce.
2004 if (UI.getUse().getValueType() != MVT::Other) continue;
2005
2006 SDNode *User = *UI;
2007
2008 if (User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
2009 continue;
2010
2011 // If we see an already-selected machine node, then we've gone beyond the
2012 // pattern that we're selecting down into the already selected chunk of the
2013 // DAG.
2014 unsigned UserOpcode = User->getOpcode();
2015 if (User->isMachineOpcode() ||
2016 UserOpcode == ISD::CopyToReg ||
2017 UserOpcode == ISD::CopyFromReg ||
2018 UserOpcode == ISD::INLINEASM ||
2019 UserOpcode == ISD::EH_LABEL ||
2020 UserOpcode == ISD::LIFETIME_START ||
2021 UserOpcode == ISD::LIFETIME_END) {
2022 // If their node ID got reset to -1 then they've already been selected.
2023 // Treat them like a MachineOpcode.
2024 if (User->getNodeId() == -1)
2025 continue;
2026 }
2027
2028 // If we have a TokenFactor, we handle it specially.
2029 if (User->getOpcode() != ISD::TokenFactor) {
2030 // If the node isn't a token factor and isn't part of our pattern, then it
2031 // must be a random chained node in between two nodes we're selecting.
2032 // This happens when we have something like:
2033 // x = load ptr
2034 // call
2035 // y = x+4
2036 // store y -> ptr
2037 // Because we structurally match the load/store as a read/modify/write,
2038 // but the call is chained between them. We cannot fold in this case
2039 // because it would induce a cycle in the graph.
2040 if (!std::count(ChainedNodesInPattern.begin(),
2041 ChainedNodesInPattern.end(), User))
2042 return CR_InducesCycle;
2043
2044 // Otherwise we found a node that is part of our pattern. For example in:
2045 // x = load ptr
2046 // y = x+4
2047 // store y -> ptr
2048 // This would happen when we're scanning down from the load and see the
2049 // store as a user. Record that there is a use of ChainedNode that is
2050 // part of the pattern and keep scanning uses.
2051 Result = CR_LeadsToInteriorNode;
2052 InteriorChainedNodes.push_back(User);
2053 continue;
2054 }
2055
2056 // If we found a TokenFactor, there are two cases to consider: first if the
2057 // TokenFactor is just hanging "below" the pattern we're matching (i.e. no
2058 // uses of the TF are in our pattern) we just want to ignore it. Second,
2059 // the TokenFactor can be sandwiched in between two chained nodes, like so:
2060 // [Load chain]
2061 // ^
2062 // |
2063 // [Load]
2064 // ^ ^
2065 // | \ DAG's like cheese
2066 // / \ do you?
2067 // / |
2068 // [TokenFactor] [Op]
2069 // ^ ^
2070 // | |
2071 // \ /
2072 // \ /
2073 // [Store]
2074 //
2075 // In this case, the TokenFactor becomes part of our match and we rewrite it
2076 // as a new TokenFactor.
2077 //
2078 // To distinguish these two cases, do a recursive walk down the uses.
2079 switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) {
2080 case CR_Simple:
2081 // If the uses of the TokenFactor are just already-selected nodes, ignore
2082 // it, it is "below" our pattern.
2083 continue;
2084 case CR_InducesCycle:
2085 // If the uses of the TokenFactor lead to nodes that are not part of our
2086 // pattern that are not selected, folding would turn this into a cycle,
2087 // bail out now.
2088 return CR_InducesCycle;
2089 case CR_LeadsToInteriorNode:
2090 break; // Otherwise, keep processing.
2091 }
2092
2093 // Okay, we know we're in the interesting interior case. The TokenFactor
2094 // is now going to be considered part of the pattern so that we rewrite its
2095 // uses (it may have uses that are not part of the pattern) with the
2096 // ultimate chain result of the generated code. We will also add its chain
2097 // inputs as inputs to the ultimate TokenFactor we create.
2098 Result = CR_LeadsToInteriorNode;
2099 ChainedNodesInPattern.push_back(User);
2100 InteriorChainedNodes.push_back(User);
2101 continue;
2102 }
2103
2104 return Result;
2105 }
2106
2107 /// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
2108 /// operation for when the pattern matched at least one node with a chains. The
2109 /// input vector contains a list of all of the chained nodes that we match. We
2110 /// must determine if this is a valid thing to cover (i.e. matching it won't
2111 /// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
2112 /// be used as the input node chain for the generated nodes.
2113 static SDValue
HandleMergeInputChains(SmallVectorImpl<SDNode * > & ChainNodesMatched,SelectionDAG * CurDAG)2114 HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
2115 SelectionDAG *CurDAG) {
2116 // Walk all of the chained nodes we've matched, recursively scanning down the
2117 // users of the chain result. This adds any TokenFactor nodes that are caught
2118 // in between chained nodes to the chained and interior nodes list.
2119 SmallVector<SDNode*, 3> InteriorChainedNodes;
2120 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
2121 if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched,
2122 InteriorChainedNodes) == CR_InducesCycle)
2123 return SDValue(); // Would induce a cycle.
2124 }
2125
2126 // Okay, we have walked all the matched nodes and collected TokenFactor nodes
2127 // that we are interested in. Form our input TokenFactor node.
2128 SmallVector<SDValue, 3> InputChains;
2129 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
2130 // Add the input chain of this node to the InputChains list (which will be
2131 // the operands of the generated TokenFactor) if it's not an interior node.
2132 SDNode *N = ChainNodesMatched[i];
2133 if (N->getOpcode() != ISD::TokenFactor) {
2134 if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N))
2135 continue;
2136
2137 // Otherwise, add the input chain.
2138 SDValue InChain = ChainNodesMatched[i]->getOperand(0);
2139 assert(InChain.getValueType() == MVT::Other && "Not a chain");
2140 InputChains.push_back(InChain);
2141 continue;
2142 }
2143
2144 // If we have a token factor, we want to add all inputs of the token factor
2145 // that are not part of the pattern we're matching.
2146 for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) {
2147 if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(),
2148 N->getOperand(op).getNode()))
2149 InputChains.push_back(N->getOperand(op));
2150 }
2151 }
2152
2153 if (InputChains.size() == 1)
2154 return InputChains[0];
2155 return CurDAG->getNode(ISD::TokenFactor, SDLoc(ChainNodesMatched[0]),
2156 MVT::Other, InputChains);
2157 }
2158
2159 /// MorphNode - Handle morphing a node in place for the selector.
2160 SDNode *SelectionDAGISel::
MorphNode(SDNode * Node,unsigned TargetOpc,SDVTList VTList,ArrayRef<SDValue> Ops,unsigned EmitNodeInfo)2161 MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList,
2162 ArrayRef<SDValue> Ops, unsigned EmitNodeInfo) {
2163 // It is possible we're using MorphNodeTo to replace a node with no
2164 // normal results with one that has a normal result (or we could be
2165 // adding a chain) and the input could have glue and chains as well.
2166 // In this case we need to shift the operands down.
2167 // FIXME: This is a horrible hack and broken in obscure cases, no worse
2168 // than the old isel though.
2169 int OldGlueResultNo = -1, OldChainResultNo = -1;
2170
2171 unsigned NTMNumResults = Node->getNumValues();
2172 if (Node->getValueType(NTMNumResults-1) == MVT::Glue) {
2173 OldGlueResultNo = NTMNumResults-1;
2174 if (NTMNumResults != 1 &&
2175 Node->getValueType(NTMNumResults-2) == MVT::Other)
2176 OldChainResultNo = NTMNumResults-2;
2177 } else if (Node->getValueType(NTMNumResults-1) == MVT::Other)
2178 OldChainResultNo = NTMNumResults-1;
2179
2180 // Call the underlying SelectionDAG routine to do the transmogrification. Note
2181 // that this deletes operands of the old node that become dead.
2182 SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops);
2183
2184 // MorphNodeTo can operate in two ways: if an existing node with the
2185 // specified operands exists, it can just return it. Otherwise, it
2186 // updates the node in place to have the requested operands.
2187 if (Res == Node) {
2188 // If we updated the node in place, reset the node ID. To the isel,
2189 // this should be just like a newly allocated machine node.
2190 Res->setNodeId(-1);
2191 }
2192
2193 unsigned ResNumResults = Res->getNumValues();
2194 // Move the glue if needed.
2195 if ((EmitNodeInfo & OPFL_GlueOutput) && OldGlueResultNo != -1 &&
2196 (unsigned)OldGlueResultNo != ResNumResults-1)
2197 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldGlueResultNo),
2198 SDValue(Res, ResNumResults-1));
2199
2200 if ((EmitNodeInfo & OPFL_GlueOutput) != 0)
2201 --ResNumResults;
2202
2203 // Move the chain reference if needed.
2204 if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 &&
2205 (unsigned)OldChainResultNo != ResNumResults-1)
2206 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo),
2207 SDValue(Res, ResNumResults-1));
2208
2209 // Otherwise, no replacement happened because the node already exists. Replace
2210 // Uses of the old node with the new one.
2211 if (Res != Node)
2212 CurDAG->ReplaceAllUsesWith(Node, Res);
2213
2214 return Res;
2215 }
2216
2217 /// CheckSame - Implements OP_CheckSame.
2218 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckSame(const unsigned char * MatcherTable,unsigned & MatcherIndex,SDValue N,const SmallVectorImpl<std::pair<SDValue,SDNode * >> & RecordedNodes)2219 CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2220 SDValue N,
2221 const SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes) {
2222 // Accept if it is exactly the same as a previously recorded node.
2223 unsigned RecNo = MatcherTable[MatcherIndex++];
2224 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2225 return N == RecordedNodes[RecNo].first;
2226 }
2227
2228 /// CheckChildSame - Implements OP_CheckChildXSame.
2229 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckChildSame(const unsigned char * MatcherTable,unsigned & MatcherIndex,SDValue N,const SmallVectorImpl<std::pair<SDValue,SDNode * >> & RecordedNodes,unsigned ChildNo)2230 CheckChildSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2231 SDValue N,
2232 const SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes,
2233 unsigned ChildNo) {
2234 if (ChildNo >= N.getNumOperands())
2235 return false; // Match fails if out of range child #.
2236 return ::CheckSame(MatcherTable, MatcherIndex, N.getOperand(ChildNo),
2237 RecordedNodes);
2238 }
2239
2240 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
2241 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckPatternPredicate(const unsigned char * MatcherTable,unsigned & MatcherIndex,const SelectionDAGISel & SDISel)2242 CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2243 const SelectionDAGISel &SDISel) {
2244 return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]);
2245 }
2246
2247 /// CheckNodePredicate - Implements OP_CheckNodePredicate.
2248 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckNodePredicate(const unsigned char * MatcherTable,unsigned & MatcherIndex,const SelectionDAGISel & SDISel,SDNode * N)2249 CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2250 const SelectionDAGISel &SDISel, SDNode *N) {
2251 return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]);
2252 }
2253
2254 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckOpcode(const unsigned char * MatcherTable,unsigned & MatcherIndex,SDNode * N)2255 CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2256 SDNode *N) {
2257 uint16_t Opc = MatcherTable[MatcherIndex++];
2258 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2259 return N->getOpcode() == Opc;
2260 }
2261
2262 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckType(const unsigned char * MatcherTable,unsigned & MatcherIndex,SDValue N,const TargetLowering * TLI)2263 CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2264 SDValue N, const TargetLowering *TLI) {
2265 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2266 if (N.getValueType() == VT) return true;
2267
2268 // Handle the case when VT is iPTR.
2269 return VT == MVT::iPTR && N.getValueType() == TLI->getPointerTy();
2270 }
2271
2272 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckChildType(const unsigned char * MatcherTable,unsigned & MatcherIndex,SDValue N,const TargetLowering * TLI,unsigned ChildNo)2273 CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2274 SDValue N, const TargetLowering *TLI, unsigned ChildNo) {
2275 if (ChildNo >= N.getNumOperands())
2276 return false; // Match fails if out of range child #.
2277 return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI);
2278 }
2279
2280 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckCondCode(const unsigned char * MatcherTable,unsigned & MatcherIndex,SDValue N)2281 CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2282 SDValue N) {
2283 return cast<CondCodeSDNode>(N)->get() ==
2284 (ISD::CondCode)MatcherTable[MatcherIndex++];
2285 }
2286
2287 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckValueType(const unsigned char * MatcherTable,unsigned & MatcherIndex,SDValue N,const TargetLowering * TLI)2288 CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2289 SDValue N, const TargetLowering *TLI) {
2290 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2291 if (cast<VTSDNode>(N)->getVT() == VT)
2292 return true;
2293
2294 // Handle the case when VT is iPTR.
2295 return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI->getPointerTy();
2296 }
2297
2298 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckInteger(const unsigned char * MatcherTable,unsigned & MatcherIndex,SDValue N)2299 CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2300 SDValue N) {
2301 int64_t Val = MatcherTable[MatcherIndex++];
2302 if (Val & 128)
2303 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2304
2305 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
2306 return C && C->getSExtValue() == Val;
2307 }
2308
2309 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckChildInteger(const unsigned char * MatcherTable,unsigned & MatcherIndex,SDValue N,unsigned ChildNo)2310 CheckChildInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2311 SDValue N, unsigned ChildNo) {
2312 if (ChildNo >= N.getNumOperands())
2313 return false; // Match fails if out of range child #.
2314 return ::CheckInteger(MatcherTable, MatcherIndex, N.getOperand(ChildNo));
2315 }
2316
2317 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckAndImm(const unsigned char * MatcherTable,unsigned & MatcherIndex,SDValue N,const SelectionDAGISel & SDISel)2318 CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2319 SDValue N, const SelectionDAGISel &SDISel) {
2320 int64_t Val = MatcherTable[MatcherIndex++];
2321 if (Val & 128)
2322 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2323
2324 if (N->getOpcode() != ISD::AND) return false;
2325
2326 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
2327 return C && SDISel.CheckAndMask(N.getOperand(0), C, Val);
2328 }
2329
2330 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
CheckOrImm(const unsigned char * MatcherTable,unsigned & MatcherIndex,SDValue N,const SelectionDAGISel & SDISel)2331 CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2332 SDValue N, const SelectionDAGISel &SDISel) {
2333 int64_t Val = MatcherTable[MatcherIndex++];
2334 if (Val & 128)
2335 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2336
2337 if (N->getOpcode() != ISD::OR) return false;
2338
2339 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
2340 return C && SDISel.CheckOrMask(N.getOperand(0), C, Val);
2341 }
2342
2343 /// IsPredicateKnownToFail - If we know how and can do so without pushing a
2344 /// scope, evaluate the current node. If the current predicate is known to
2345 /// fail, set Result=true and return anything. If the current predicate is
2346 /// known to pass, set Result=false and return the MatcherIndex to continue
2347 /// with. If the current predicate is unknown, set Result=false and return the
2348 /// MatcherIndex to continue with.
IsPredicateKnownToFail(const unsigned char * Table,unsigned Index,SDValue N,bool & Result,const SelectionDAGISel & SDISel,SmallVectorImpl<std::pair<SDValue,SDNode * >> & RecordedNodes)2349 static unsigned IsPredicateKnownToFail(const unsigned char *Table,
2350 unsigned Index, SDValue N,
2351 bool &Result,
2352 const SelectionDAGISel &SDISel,
2353 SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes) {
2354 switch (Table[Index++]) {
2355 default:
2356 Result = false;
2357 return Index-1; // Could not evaluate this predicate.
2358 case SelectionDAGISel::OPC_CheckSame:
2359 Result = !::CheckSame(Table, Index, N, RecordedNodes);
2360 return Index;
2361 case SelectionDAGISel::OPC_CheckChild0Same:
2362 case SelectionDAGISel::OPC_CheckChild1Same:
2363 case SelectionDAGISel::OPC_CheckChild2Same:
2364 case SelectionDAGISel::OPC_CheckChild3Same:
2365 Result = !::CheckChildSame(Table, Index, N, RecordedNodes,
2366 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Same);
2367 return Index;
2368 case SelectionDAGISel::OPC_CheckPatternPredicate:
2369 Result = !::CheckPatternPredicate(Table, Index, SDISel);
2370 return Index;
2371 case SelectionDAGISel::OPC_CheckPredicate:
2372 Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode());
2373 return Index;
2374 case SelectionDAGISel::OPC_CheckOpcode:
2375 Result = !::CheckOpcode(Table, Index, N.getNode());
2376 return Index;
2377 case SelectionDAGISel::OPC_CheckType:
2378 Result = !::CheckType(Table, Index, N, SDISel.getTargetLowering());
2379 return Index;
2380 case SelectionDAGISel::OPC_CheckChild0Type:
2381 case SelectionDAGISel::OPC_CheckChild1Type:
2382 case SelectionDAGISel::OPC_CheckChild2Type:
2383 case SelectionDAGISel::OPC_CheckChild3Type:
2384 case SelectionDAGISel::OPC_CheckChild4Type:
2385 case SelectionDAGISel::OPC_CheckChild5Type:
2386 case SelectionDAGISel::OPC_CheckChild6Type:
2387 case SelectionDAGISel::OPC_CheckChild7Type:
2388 Result = !::CheckChildType(Table, Index, N, SDISel.getTargetLowering(),
2389 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Type);
2390 return Index;
2391 case SelectionDAGISel::OPC_CheckCondCode:
2392 Result = !::CheckCondCode(Table, Index, N);
2393 return Index;
2394 case SelectionDAGISel::OPC_CheckValueType:
2395 Result = !::CheckValueType(Table, Index, N, SDISel.getTargetLowering());
2396 return Index;
2397 case SelectionDAGISel::OPC_CheckInteger:
2398 Result = !::CheckInteger(Table, Index, N);
2399 return Index;
2400 case SelectionDAGISel::OPC_CheckChild0Integer:
2401 case SelectionDAGISel::OPC_CheckChild1Integer:
2402 case SelectionDAGISel::OPC_CheckChild2Integer:
2403 case SelectionDAGISel::OPC_CheckChild3Integer:
2404 case SelectionDAGISel::OPC_CheckChild4Integer:
2405 Result = !::CheckChildInteger(Table, Index, N,
2406 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Integer);
2407 return Index;
2408 case SelectionDAGISel::OPC_CheckAndImm:
2409 Result = !::CheckAndImm(Table, Index, N, SDISel);
2410 return Index;
2411 case SelectionDAGISel::OPC_CheckOrImm:
2412 Result = !::CheckOrImm(Table, Index, N, SDISel);
2413 return Index;
2414 }
2415 }
2416
2417 namespace {
2418
2419 struct MatchScope {
2420 /// FailIndex - If this match fails, this is the index to continue with.
2421 unsigned FailIndex;
2422
2423 /// NodeStack - The node stack when the scope was formed.
2424 SmallVector<SDValue, 4> NodeStack;
2425
2426 /// NumRecordedNodes - The number of recorded nodes when the scope was formed.
2427 unsigned NumRecordedNodes;
2428
2429 /// NumMatchedMemRefs - The number of matched memref entries.
2430 unsigned NumMatchedMemRefs;
2431
2432 /// InputChain/InputGlue - The current chain/glue
2433 SDValue InputChain, InputGlue;
2434
2435 /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty.
2436 bool HasChainNodesMatched, HasGlueResultNodesMatched;
2437 };
2438
2439 }
2440
2441 SDNode *SelectionDAGISel::
SelectCodeCommon(SDNode * NodeToMatch,const unsigned char * MatcherTable,unsigned TableSize)2442 SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
2443 unsigned TableSize) {
2444 // FIXME: Should these even be selected? Handle these cases in the caller?
2445 switch (NodeToMatch->getOpcode()) {
2446 default:
2447 break;
2448 case ISD::EntryToken: // These nodes remain the same.
2449 case ISD::BasicBlock:
2450 case ISD::Register:
2451 case ISD::RegisterMask:
2452 //case ISD::VALUETYPE:
2453 //case ISD::CONDCODE:
2454 case ISD::HANDLENODE:
2455 case ISD::MDNODE_SDNODE:
2456 case ISD::TargetConstant:
2457 case ISD::TargetConstantFP:
2458 case ISD::TargetConstantPool:
2459 case ISD::TargetFrameIndex:
2460 case ISD::TargetExternalSymbol:
2461 case ISD::TargetBlockAddress:
2462 case ISD::TargetJumpTable:
2463 case ISD::TargetGlobalTLSAddress:
2464 case ISD::TargetGlobalAddress:
2465 case ISD::TokenFactor:
2466 case ISD::CopyFromReg:
2467 case ISD::CopyToReg:
2468 case ISD::EH_LABEL:
2469 case ISD::LIFETIME_START:
2470 case ISD::LIFETIME_END:
2471 NodeToMatch->setNodeId(-1); // Mark selected.
2472 return nullptr;
2473 case ISD::AssertSext:
2474 case ISD::AssertZext:
2475 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0),
2476 NodeToMatch->getOperand(0));
2477 return nullptr;
2478 case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch);
2479 case ISD::READ_REGISTER: return Select_READ_REGISTER(NodeToMatch);
2480 case ISD::WRITE_REGISTER: return Select_WRITE_REGISTER(NodeToMatch);
2481 case ISD::UNDEF: return Select_UNDEF(NodeToMatch);
2482 }
2483
2484 assert(!NodeToMatch->isMachineOpcode() && "Node already selected!");
2485
2486 // Set up the node stack with NodeToMatch as the only node on the stack.
2487 SmallVector<SDValue, 8> NodeStack;
2488 SDValue N = SDValue(NodeToMatch, 0);
2489 NodeStack.push_back(N);
2490
2491 // MatchScopes - Scopes used when matching, if a match failure happens, this
2492 // indicates where to continue checking.
2493 SmallVector<MatchScope, 8> MatchScopes;
2494
2495 // RecordedNodes - This is the set of nodes that have been recorded by the
2496 // state machine. The second value is the parent of the node, or null if the
2497 // root is recorded.
2498 SmallVector<std::pair<SDValue, SDNode*>, 8> RecordedNodes;
2499
2500 // MatchedMemRefs - This is the set of MemRef's we've seen in the input
2501 // pattern.
2502 SmallVector<MachineMemOperand*, 2> MatchedMemRefs;
2503
2504 // These are the current input chain and glue for use when generating nodes.
2505 // Various Emit operations change these. For example, emitting a copytoreg
2506 // uses and updates these.
2507 SDValue InputChain, InputGlue;
2508
2509 // ChainNodesMatched - If a pattern matches nodes that have input/output
2510 // chains, the OPC_EmitMergeInputChains operation is emitted which indicates
2511 // which ones they are. The result is captured into this list so that we can
2512 // update the chain results when the pattern is complete.
2513 SmallVector<SDNode*, 3> ChainNodesMatched;
2514 SmallVector<SDNode*, 3> GlueResultNodesMatched;
2515
2516 DEBUG(dbgs() << "ISEL: Starting pattern match on root node: ";
2517 NodeToMatch->dump(CurDAG);
2518 dbgs() << '\n');
2519
2520 // Determine where to start the interpreter. Normally we start at opcode #0,
2521 // but if the state machine starts with an OPC_SwitchOpcode, then we
2522 // accelerate the first lookup (which is guaranteed to be hot) with the
2523 // OpcodeOffset table.
2524 unsigned MatcherIndex = 0;
2525
2526 if (!OpcodeOffset.empty()) {
2527 // Already computed the OpcodeOffset table, just index into it.
2528 if (N.getOpcode() < OpcodeOffset.size())
2529 MatcherIndex = OpcodeOffset[N.getOpcode()];
2530 DEBUG(dbgs() << " Initial Opcode index to " << MatcherIndex << "\n");
2531
2532 } else if (MatcherTable[0] == OPC_SwitchOpcode) {
2533 // Otherwise, the table isn't computed, but the state machine does start
2534 // with an OPC_SwitchOpcode instruction. Populate the table now, since this
2535 // is the first time we're selecting an instruction.
2536 unsigned Idx = 1;
2537 while (1) {
2538 // Get the size of this case.
2539 unsigned CaseSize = MatcherTable[Idx++];
2540 if (CaseSize & 128)
2541 CaseSize = GetVBR(CaseSize, MatcherTable, Idx);
2542 if (CaseSize == 0) break;
2543
2544 // Get the opcode, add the index to the table.
2545 uint16_t Opc = MatcherTable[Idx++];
2546 Opc |= (unsigned short)MatcherTable[Idx++] << 8;
2547 if (Opc >= OpcodeOffset.size())
2548 OpcodeOffset.resize((Opc+1)*2);
2549 OpcodeOffset[Opc] = Idx;
2550 Idx += CaseSize;
2551 }
2552
2553 // Okay, do the lookup for the first opcode.
2554 if (N.getOpcode() < OpcodeOffset.size())
2555 MatcherIndex = OpcodeOffset[N.getOpcode()];
2556 }
2557
2558 while (1) {
2559 assert(MatcherIndex < TableSize && "Invalid index");
2560 #ifndef NDEBUG
2561 unsigned CurrentOpcodeIndex = MatcherIndex;
2562 #endif
2563 BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++];
2564 switch (Opcode) {
2565 case OPC_Scope: {
2566 // Okay, the semantics of this operation are that we should push a scope
2567 // then evaluate the first child. However, pushing a scope only to have
2568 // the first check fail (which then pops it) is inefficient. If we can
2569 // determine immediately that the first check (or first several) will
2570 // immediately fail, don't even bother pushing a scope for them.
2571 unsigned FailIndex;
2572
2573 while (1) {
2574 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2575 if (NumToSkip & 128)
2576 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2577 // Found the end of the scope with no match.
2578 if (NumToSkip == 0) {
2579 FailIndex = 0;
2580 break;
2581 }
2582
2583 FailIndex = MatcherIndex+NumToSkip;
2584
2585 unsigned MatcherIndexOfPredicate = MatcherIndex;
2586 (void)MatcherIndexOfPredicate; // silence warning.
2587
2588 // If we can't evaluate this predicate without pushing a scope (e.g. if
2589 // it is a 'MoveParent') or if the predicate succeeds on this node, we
2590 // push the scope and evaluate the full predicate chain.
2591 bool Result;
2592 MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
2593 Result, *this, RecordedNodes);
2594 if (!Result)
2595 break;
2596
2597 DEBUG(dbgs() << " Skipped scope entry (due to false predicate) at "
2598 << "index " << MatcherIndexOfPredicate
2599 << ", continuing at " << FailIndex << "\n");
2600 ++NumDAGIselRetries;
2601
2602 // Otherwise, we know that this case of the Scope is guaranteed to fail,
2603 // move to the next case.
2604 MatcherIndex = FailIndex;
2605 }
2606
2607 // If the whole scope failed to match, bail.
2608 if (FailIndex == 0) break;
2609
2610 // Push a MatchScope which indicates where to go if the first child fails
2611 // to match.
2612 MatchScope NewEntry;
2613 NewEntry.FailIndex = FailIndex;
2614 NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
2615 NewEntry.NumRecordedNodes = RecordedNodes.size();
2616 NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
2617 NewEntry.InputChain = InputChain;
2618 NewEntry.InputGlue = InputGlue;
2619 NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty();
2620 NewEntry.HasGlueResultNodesMatched = !GlueResultNodesMatched.empty();
2621 MatchScopes.push_back(NewEntry);
2622 continue;
2623 }
2624 case OPC_RecordNode: {
2625 // Remember this node, it may end up being an operand in the pattern.
2626 SDNode *Parent = nullptr;
2627 if (NodeStack.size() > 1)
2628 Parent = NodeStack[NodeStack.size()-2].getNode();
2629 RecordedNodes.push_back(std::make_pair(N, Parent));
2630 continue;
2631 }
2632
2633 case OPC_RecordChild0: case OPC_RecordChild1:
2634 case OPC_RecordChild2: case OPC_RecordChild3:
2635 case OPC_RecordChild4: case OPC_RecordChild5:
2636 case OPC_RecordChild6: case OPC_RecordChild7: {
2637 unsigned ChildNo = Opcode-OPC_RecordChild0;
2638 if (ChildNo >= N.getNumOperands())
2639 break; // Match fails if out of range child #.
2640
2641 RecordedNodes.push_back(std::make_pair(N->getOperand(ChildNo),
2642 N.getNode()));
2643 continue;
2644 }
2645 case OPC_RecordMemRef:
2646 MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand());
2647 continue;
2648
2649 case OPC_CaptureGlueInput:
2650 // If the current node has an input glue, capture it in InputGlue.
2651 if (N->getNumOperands() != 0 &&
2652 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue)
2653 InputGlue = N->getOperand(N->getNumOperands()-1);
2654 continue;
2655
2656 case OPC_MoveChild: {
2657 unsigned ChildNo = MatcherTable[MatcherIndex++];
2658 if (ChildNo >= N.getNumOperands())
2659 break; // Match fails if out of range child #.
2660 N = N.getOperand(ChildNo);
2661 NodeStack.push_back(N);
2662 continue;
2663 }
2664
2665 case OPC_MoveParent:
2666 // Pop the current node off the NodeStack.
2667 NodeStack.pop_back();
2668 assert(!NodeStack.empty() && "Node stack imbalance!");
2669 N = NodeStack.back();
2670 continue;
2671
2672 case OPC_CheckSame:
2673 if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
2674 continue;
2675
2676 case OPC_CheckChild0Same: case OPC_CheckChild1Same:
2677 case OPC_CheckChild2Same: case OPC_CheckChild3Same:
2678 if (!::CheckChildSame(MatcherTable, MatcherIndex, N, RecordedNodes,
2679 Opcode-OPC_CheckChild0Same))
2680 break;
2681 continue;
2682
2683 case OPC_CheckPatternPredicate:
2684 if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break;
2685 continue;
2686 case OPC_CheckPredicate:
2687 if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this,
2688 N.getNode()))
2689 break;
2690 continue;
2691 case OPC_CheckComplexPat: {
2692 unsigned CPNum = MatcherTable[MatcherIndex++];
2693 unsigned RecNo = MatcherTable[MatcherIndex++];
2694 assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
2695 if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo].second,
2696 RecordedNodes[RecNo].first, CPNum,
2697 RecordedNodes))
2698 break;
2699 continue;
2700 }
2701 case OPC_CheckOpcode:
2702 if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
2703 continue;
2704
2705 case OPC_CheckType:
2706 if (!::CheckType(MatcherTable, MatcherIndex, N, getTargetLowering()))
2707 break;
2708 continue;
2709
2710 case OPC_SwitchOpcode: {
2711 unsigned CurNodeOpcode = N.getOpcode();
2712 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2713 unsigned CaseSize;
2714 while (1) {
2715 // Get the size of this case.
2716 CaseSize = MatcherTable[MatcherIndex++];
2717 if (CaseSize & 128)
2718 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2719 if (CaseSize == 0) break;
2720
2721 uint16_t Opc = MatcherTable[MatcherIndex++];
2722 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2723
2724 // If the opcode matches, then we will execute this case.
2725 if (CurNodeOpcode == Opc)
2726 break;
2727
2728 // Otherwise, skip over this case.
2729 MatcherIndex += CaseSize;
2730 }
2731
2732 // If no cases matched, bail out.
2733 if (CaseSize == 0) break;
2734
2735 // Otherwise, execute the case we found.
2736 DEBUG(dbgs() << " OpcodeSwitch from " << SwitchStart
2737 << " to " << MatcherIndex << "\n");
2738 continue;
2739 }
2740
2741 case OPC_SwitchType: {
2742 MVT CurNodeVT = N.getSimpleValueType();
2743 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2744 unsigned CaseSize;
2745 while (1) {
2746 // Get the size of this case.
2747 CaseSize = MatcherTable[MatcherIndex++];
2748 if (CaseSize & 128)
2749 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2750 if (CaseSize == 0) break;
2751
2752 MVT CaseVT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2753 if (CaseVT == MVT::iPTR)
2754 CaseVT = getTargetLowering()->getPointerTy();
2755
2756 // If the VT matches, then we will execute this case.
2757 if (CurNodeVT == CaseVT)
2758 break;
2759
2760 // Otherwise, skip over this case.
2761 MatcherIndex += CaseSize;
2762 }
2763
2764 // If no cases matched, bail out.
2765 if (CaseSize == 0) break;
2766
2767 // Otherwise, execute the case we found.
2768 DEBUG(dbgs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
2769 << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
2770 continue;
2771 }
2772 case OPC_CheckChild0Type: case OPC_CheckChild1Type:
2773 case OPC_CheckChild2Type: case OPC_CheckChild3Type:
2774 case OPC_CheckChild4Type: case OPC_CheckChild5Type:
2775 case OPC_CheckChild6Type: case OPC_CheckChild7Type:
2776 if (!::CheckChildType(MatcherTable, MatcherIndex, N, getTargetLowering(),
2777 Opcode-OPC_CheckChild0Type))
2778 break;
2779 continue;
2780 case OPC_CheckCondCode:
2781 if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
2782 continue;
2783 case OPC_CheckValueType:
2784 if (!::CheckValueType(MatcherTable, MatcherIndex, N, getTargetLowering()))
2785 break;
2786 continue;
2787 case OPC_CheckInteger:
2788 if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
2789 continue;
2790 case OPC_CheckChild0Integer: case OPC_CheckChild1Integer:
2791 case OPC_CheckChild2Integer: case OPC_CheckChild3Integer:
2792 case OPC_CheckChild4Integer:
2793 if (!::CheckChildInteger(MatcherTable, MatcherIndex, N,
2794 Opcode-OPC_CheckChild0Integer)) break;
2795 continue;
2796 case OPC_CheckAndImm:
2797 if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
2798 continue;
2799 case OPC_CheckOrImm:
2800 if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
2801 continue;
2802
2803 case OPC_CheckFoldableChainNode: {
2804 assert(NodeStack.size() != 1 && "No parent node");
2805 // Verify that all intermediate nodes between the root and this one have
2806 // a single use.
2807 bool HasMultipleUses = false;
2808 for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i)
2809 if (!NodeStack[i].hasOneUse()) {
2810 HasMultipleUses = true;
2811 break;
2812 }
2813 if (HasMultipleUses) break;
2814
2815 // Check to see that the target thinks this is profitable to fold and that
2816 // we can fold it without inducing cycles in the graph.
2817 if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2818 NodeToMatch) ||
2819 !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2820 NodeToMatch, OptLevel,
2821 true/*We validate our own chains*/))
2822 break;
2823
2824 continue;
2825 }
2826 case OPC_EmitInteger: {
2827 MVT::SimpleValueType VT =
2828 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2829 int64_t Val = MatcherTable[MatcherIndex++];
2830 if (Val & 128)
2831 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2832 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2833 CurDAG->getTargetConstant(Val, VT), nullptr));
2834 continue;
2835 }
2836 case OPC_EmitRegister: {
2837 MVT::SimpleValueType VT =
2838 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2839 unsigned RegNo = MatcherTable[MatcherIndex++];
2840 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2841 CurDAG->getRegister(RegNo, VT), nullptr));
2842 continue;
2843 }
2844 case OPC_EmitRegister2: {
2845 // For targets w/ more than 256 register names, the register enum
2846 // values are stored in two bytes in the matcher table (just like
2847 // opcodes).
2848 MVT::SimpleValueType VT =
2849 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2850 unsigned RegNo = MatcherTable[MatcherIndex++];
2851 RegNo |= MatcherTable[MatcherIndex++] << 8;
2852 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2853 CurDAG->getRegister(RegNo, VT), nullptr));
2854 continue;
2855 }
2856
2857 case OPC_EmitConvertToTarget: {
2858 // Convert from IMM/FPIMM to target version.
2859 unsigned RecNo = MatcherTable[MatcherIndex++];
2860 assert(RecNo < RecordedNodes.size() && "Invalid EmitConvertToTarget");
2861 SDValue Imm = RecordedNodes[RecNo].first;
2862
2863 if (Imm->getOpcode() == ISD::Constant) {
2864 const ConstantInt *Val=cast<ConstantSDNode>(Imm)->getConstantIntValue();
2865 Imm = CurDAG->getConstant(*Val, Imm.getValueType(), true);
2866 } else if (Imm->getOpcode() == ISD::ConstantFP) {
2867 const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue();
2868 Imm = CurDAG->getConstantFP(*Val, Imm.getValueType(), true);
2869 }
2870
2871 RecordedNodes.push_back(std::make_pair(Imm, RecordedNodes[RecNo].second));
2872 continue;
2873 }
2874
2875 case OPC_EmitMergeInputChains1_0: // OPC_EmitMergeInputChains, 1, 0
2876 case OPC_EmitMergeInputChains1_1: { // OPC_EmitMergeInputChains, 1, 1
2877 // These are space-optimized forms of OPC_EmitMergeInputChains.
2878 assert(!InputChain.getNode() &&
2879 "EmitMergeInputChains should be the first chain producing node");
2880 assert(ChainNodesMatched.empty() &&
2881 "Should only have one EmitMergeInputChains per match");
2882
2883 // Read all of the chained nodes.
2884 unsigned RecNo = Opcode == OPC_EmitMergeInputChains1_1;
2885 assert(RecNo < RecordedNodes.size() && "Invalid EmitMergeInputChains");
2886 ChainNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
2887
2888 // FIXME: What if other value results of the node have uses not matched
2889 // by this pattern?
2890 if (ChainNodesMatched.back() != NodeToMatch &&
2891 !RecordedNodes[RecNo].first.hasOneUse()) {
2892 ChainNodesMatched.clear();
2893 break;
2894 }
2895
2896 // Merge the input chains if they are not intra-pattern references.
2897 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2898
2899 if (!InputChain.getNode())
2900 break; // Failed to merge.
2901 continue;
2902 }
2903
2904 case OPC_EmitMergeInputChains: {
2905 assert(!InputChain.getNode() &&
2906 "EmitMergeInputChains should be the first chain producing node");
2907 // This node gets a list of nodes we matched in the input that have
2908 // chains. We want to token factor all of the input chains to these nodes
2909 // together. However, if any of the input chains is actually one of the
2910 // nodes matched in this pattern, then we have an intra-match reference.
2911 // Ignore these because the newly token factored chain should not refer to
2912 // the old nodes.
2913 unsigned NumChains = MatcherTable[MatcherIndex++];
2914 assert(NumChains != 0 && "Can't TF zero chains");
2915
2916 assert(ChainNodesMatched.empty() &&
2917 "Should only have one EmitMergeInputChains per match");
2918
2919 // Read all of the chained nodes.
2920 for (unsigned i = 0; i != NumChains; ++i) {
2921 unsigned RecNo = MatcherTable[MatcherIndex++];
2922 assert(RecNo < RecordedNodes.size() && "Invalid EmitMergeInputChains");
2923 ChainNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
2924
2925 // FIXME: What if other value results of the node have uses not matched
2926 // by this pattern?
2927 if (ChainNodesMatched.back() != NodeToMatch &&
2928 !RecordedNodes[RecNo].first.hasOneUse()) {
2929 ChainNodesMatched.clear();
2930 break;
2931 }
2932 }
2933
2934 // If the inner loop broke out, the match fails.
2935 if (ChainNodesMatched.empty())
2936 break;
2937
2938 // Merge the input chains if they are not intra-pattern references.
2939 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2940
2941 if (!InputChain.getNode())
2942 break; // Failed to merge.
2943
2944 continue;
2945 }
2946
2947 case OPC_EmitCopyToReg: {
2948 unsigned RecNo = MatcherTable[MatcherIndex++];
2949 assert(RecNo < RecordedNodes.size() && "Invalid EmitCopyToReg");
2950 unsigned DestPhysReg = MatcherTable[MatcherIndex++];
2951
2952 if (!InputChain.getNode())
2953 InputChain = CurDAG->getEntryNode();
2954
2955 InputChain = CurDAG->getCopyToReg(InputChain, SDLoc(NodeToMatch),
2956 DestPhysReg, RecordedNodes[RecNo].first,
2957 InputGlue);
2958
2959 InputGlue = InputChain.getValue(1);
2960 continue;
2961 }
2962
2963 case OPC_EmitNodeXForm: {
2964 unsigned XFormNo = MatcherTable[MatcherIndex++];
2965 unsigned RecNo = MatcherTable[MatcherIndex++];
2966 assert(RecNo < RecordedNodes.size() && "Invalid EmitNodeXForm");
2967 SDValue Res = RunSDNodeXForm(RecordedNodes[RecNo].first, XFormNo);
2968 RecordedNodes.push_back(std::pair<SDValue,SDNode*>(Res, nullptr));
2969 continue;
2970 }
2971
2972 case OPC_EmitNode:
2973 case OPC_MorphNodeTo: {
2974 uint16_t TargetOpc = MatcherTable[MatcherIndex++];
2975 TargetOpc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2976 unsigned EmitNodeInfo = MatcherTable[MatcherIndex++];
2977 // Get the result VT list.
2978 unsigned NumVTs = MatcherTable[MatcherIndex++];
2979 SmallVector<EVT, 4> VTs;
2980 for (unsigned i = 0; i != NumVTs; ++i) {
2981 MVT::SimpleValueType VT =
2982 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2983 if (VT == MVT::iPTR) VT = getTargetLowering()->getPointerTy().SimpleTy;
2984 VTs.push_back(VT);
2985 }
2986
2987 if (EmitNodeInfo & OPFL_Chain)
2988 VTs.push_back(MVT::Other);
2989 if (EmitNodeInfo & OPFL_GlueOutput)
2990 VTs.push_back(MVT::Glue);
2991
2992 // This is hot code, so optimize the two most common cases of 1 and 2
2993 // results.
2994 SDVTList VTList;
2995 if (VTs.size() == 1)
2996 VTList = CurDAG->getVTList(VTs[0]);
2997 else if (VTs.size() == 2)
2998 VTList = CurDAG->getVTList(VTs[0], VTs[1]);
2999 else
3000 VTList = CurDAG->getVTList(VTs);
3001
3002 // Get the operand list.
3003 unsigned NumOps = MatcherTable[MatcherIndex++];
3004 SmallVector<SDValue, 8> Ops;
3005 for (unsigned i = 0; i != NumOps; ++i) {
3006 unsigned RecNo = MatcherTable[MatcherIndex++];
3007 if (RecNo & 128)
3008 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
3009
3010 assert(RecNo < RecordedNodes.size() && "Invalid EmitNode");
3011 Ops.push_back(RecordedNodes[RecNo].first);
3012 }
3013
3014 // If there are variadic operands to add, handle them now.
3015 if (EmitNodeInfo & OPFL_VariadicInfo) {
3016 // Determine the start index to copy from.
3017 unsigned FirstOpToCopy = getNumFixedFromVariadicInfo(EmitNodeInfo);
3018 FirstOpToCopy += (EmitNodeInfo & OPFL_Chain) ? 1 : 0;
3019 assert(NodeToMatch->getNumOperands() >= FirstOpToCopy &&
3020 "Invalid variadic node");
3021 // Copy all of the variadic operands, not including a potential glue
3022 // input.
3023 for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands();
3024 i != e; ++i) {
3025 SDValue V = NodeToMatch->getOperand(i);
3026 if (V.getValueType() == MVT::Glue) break;
3027 Ops.push_back(V);
3028 }
3029 }
3030
3031 // If this has chain/glue inputs, add them.
3032 if (EmitNodeInfo & OPFL_Chain)
3033 Ops.push_back(InputChain);
3034 if ((EmitNodeInfo & OPFL_GlueInput) && InputGlue.getNode() != nullptr)
3035 Ops.push_back(InputGlue);
3036
3037 // Create the node.
3038 SDNode *Res = nullptr;
3039 if (Opcode != OPC_MorphNodeTo) {
3040 // If this is a normal EmitNode command, just create the new node and
3041 // add the results to the RecordedNodes list.
3042 Res = CurDAG->getMachineNode(TargetOpc, SDLoc(NodeToMatch),
3043 VTList, Ops);
3044
3045 // Add all the non-glue/non-chain results to the RecordedNodes list.
3046 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
3047 if (VTs[i] == MVT::Other || VTs[i] == MVT::Glue) break;
3048 RecordedNodes.push_back(std::pair<SDValue,SDNode*>(SDValue(Res, i),
3049 nullptr));
3050 }
3051
3052 } else if (NodeToMatch->getOpcode() != ISD::DELETED_NODE) {
3053 Res = MorphNode(NodeToMatch, TargetOpc, VTList, Ops, EmitNodeInfo);
3054 } else {
3055 // NodeToMatch was eliminated by CSE when the target changed the DAG.
3056 // We will visit the equivalent node later.
3057 DEBUG(dbgs() << "Node was eliminated by CSE\n");
3058 return nullptr;
3059 }
3060
3061 // If the node had chain/glue results, update our notion of the current
3062 // chain and glue.
3063 if (EmitNodeInfo & OPFL_GlueOutput) {
3064 InputGlue = SDValue(Res, VTs.size()-1);
3065 if (EmitNodeInfo & OPFL_Chain)
3066 InputChain = SDValue(Res, VTs.size()-2);
3067 } else if (EmitNodeInfo & OPFL_Chain)
3068 InputChain = SDValue(Res, VTs.size()-1);
3069
3070 // If the OPFL_MemRefs glue is set on this node, slap all of the
3071 // accumulated memrefs onto it.
3072 //
3073 // FIXME: This is vastly incorrect for patterns with multiple outputs
3074 // instructions that access memory and for ComplexPatterns that match
3075 // loads.
3076 if (EmitNodeInfo & OPFL_MemRefs) {
3077 // Only attach load or store memory operands if the generated
3078 // instruction may load or store.
3079 const MCInstrDesc &MCID = TM.getInstrInfo()->get(TargetOpc);
3080 bool mayLoad = MCID.mayLoad();
3081 bool mayStore = MCID.mayStore();
3082
3083 unsigned NumMemRefs = 0;
3084 for (SmallVectorImpl<MachineMemOperand *>::const_iterator I =
3085 MatchedMemRefs.begin(), E = MatchedMemRefs.end(); I != E; ++I) {
3086 if ((*I)->isLoad()) {
3087 if (mayLoad)
3088 ++NumMemRefs;
3089 } else if ((*I)->isStore()) {
3090 if (mayStore)
3091 ++NumMemRefs;
3092 } else {
3093 ++NumMemRefs;
3094 }
3095 }
3096
3097 MachineSDNode::mmo_iterator MemRefs =
3098 MF->allocateMemRefsArray(NumMemRefs);
3099
3100 MachineSDNode::mmo_iterator MemRefsPos = MemRefs;
3101 for (SmallVectorImpl<MachineMemOperand *>::const_iterator I =
3102 MatchedMemRefs.begin(), E = MatchedMemRefs.end(); I != E; ++I) {
3103 if ((*I)->isLoad()) {
3104 if (mayLoad)
3105 *MemRefsPos++ = *I;
3106 } else if ((*I)->isStore()) {
3107 if (mayStore)
3108 *MemRefsPos++ = *I;
3109 } else {
3110 *MemRefsPos++ = *I;
3111 }
3112 }
3113
3114 cast<MachineSDNode>(Res)
3115 ->setMemRefs(MemRefs, MemRefs + NumMemRefs);
3116 }
3117
3118 DEBUG(dbgs() << " "
3119 << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created")
3120 << " node: "; Res->dump(CurDAG); dbgs() << "\n");
3121
3122 // If this was a MorphNodeTo then we're completely done!
3123 if (Opcode == OPC_MorphNodeTo) {
3124 // Update chain and glue uses.
3125 UpdateChainsAndGlue(NodeToMatch, InputChain, ChainNodesMatched,
3126 InputGlue, GlueResultNodesMatched, true);
3127 return Res;
3128 }
3129
3130 continue;
3131 }
3132
3133 case OPC_MarkGlueResults: {
3134 unsigned NumNodes = MatcherTable[MatcherIndex++];
3135
3136 // Read and remember all the glue-result nodes.
3137 for (unsigned i = 0; i != NumNodes; ++i) {
3138 unsigned RecNo = MatcherTable[MatcherIndex++];
3139 if (RecNo & 128)
3140 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
3141
3142 assert(RecNo < RecordedNodes.size() && "Invalid MarkGlueResults");
3143 GlueResultNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
3144 }
3145 continue;
3146 }
3147
3148 case OPC_CompleteMatch: {
3149 // The match has been completed, and any new nodes (if any) have been
3150 // created. Patch up references to the matched dag to use the newly
3151 // created nodes.
3152 unsigned NumResults = MatcherTable[MatcherIndex++];
3153
3154 for (unsigned i = 0; i != NumResults; ++i) {
3155 unsigned ResSlot = MatcherTable[MatcherIndex++];
3156 if (ResSlot & 128)
3157 ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex);
3158
3159 assert(ResSlot < RecordedNodes.size() && "Invalid CompleteMatch");
3160 SDValue Res = RecordedNodes[ResSlot].first;
3161
3162 assert(i < NodeToMatch->getNumValues() &&
3163 NodeToMatch->getValueType(i) != MVT::Other &&
3164 NodeToMatch->getValueType(i) != MVT::Glue &&
3165 "Invalid number of results to complete!");
3166 assert((NodeToMatch->getValueType(i) == Res.getValueType() ||
3167 NodeToMatch->getValueType(i) == MVT::iPTR ||
3168 Res.getValueType() == MVT::iPTR ||
3169 NodeToMatch->getValueType(i).getSizeInBits() ==
3170 Res.getValueType().getSizeInBits()) &&
3171 "invalid replacement");
3172 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, i), Res);
3173 }
3174
3175 // If the root node defines glue, add it to the glue nodes to update list.
3176 if (NodeToMatch->getValueType(NodeToMatch->getNumValues()-1) == MVT::Glue)
3177 GlueResultNodesMatched.push_back(NodeToMatch);
3178
3179 // Update chain and glue uses.
3180 UpdateChainsAndGlue(NodeToMatch, InputChain, ChainNodesMatched,
3181 InputGlue, GlueResultNodesMatched, false);
3182
3183 assert(NodeToMatch->use_empty() &&
3184 "Didn't replace all uses of the node?");
3185
3186 // FIXME: We just return here, which interacts correctly with SelectRoot
3187 // above. We should fix this to not return an SDNode* anymore.
3188 return nullptr;
3189 }
3190 }
3191
3192 // If the code reached this point, then the match failed. See if there is
3193 // another child to try in the current 'Scope', otherwise pop it until we
3194 // find a case to check.
3195 DEBUG(dbgs() << " Match failed at index " << CurrentOpcodeIndex << "\n");
3196 ++NumDAGIselRetries;
3197 while (1) {
3198 if (MatchScopes.empty()) {
3199 CannotYetSelect(NodeToMatch);
3200 return nullptr;
3201 }
3202
3203 // Restore the interpreter state back to the point where the scope was
3204 // formed.
3205 MatchScope &LastScope = MatchScopes.back();
3206 RecordedNodes.resize(LastScope.NumRecordedNodes);
3207 NodeStack.clear();
3208 NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end());
3209 N = NodeStack.back();
3210
3211 if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size())
3212 MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
3213 MatcherIndex = LastScope.FailIndex;
3214
3215 DEBUG(dbgs() << " Continuing at " << MatcherIndex << "\n");
3216
3217 InputChain = LastScope.InputChain;
3218 InputGlue = LastScope.InputGlue;
3219 if (!LastScope.HasChainNodesMatched)
3220 ChainNodesMatched.clear();
3221 if (!LastScope.HasGlueResultNodesMatched)
3222 GlueResultNodesMatched.clear();
3223
3224 // Check to see what the offset is at the new MatcherIndex. If it is zero
3225 // we have reached the end of this scope, otherwise we have another child
3226 // in the current scope to try.
3227 unsigned NumToSkip = MatcherTable[MatcherIndex++];
3228 if (NumToSkip & 128)
3229 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
3230
3231 // If we have another child in this scope to match, update FailIndex and
3232 // try it.
3233 if (NumToSkip != 0) {
3234 LastScope.FailIndex = MatcherIndex+NumToSkip;
3235 break;
3236 }
3237
3238 // End of this scope, pop it and try the next child in the containing
3239 // scope.
3240 MatchScopes.pop_back();
3241 }
3242 }
3243 }
3244
3245
3246
CannotYetSelect(SDNode * N)3247 void SelectionDAGISel::CannotYetSelect(SDNode *N) {
3248 std::string msg;
3249 raw_string_ostream Msg(msg);
3250 Msg << "Cannot select: ";
3251
3252 if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN &&
3253 N->getOpcode() != ISD::INTRINSIC_WO_CHAIN &&
3254 N->getOpcode() != ISD::INTRINSIC_VOID) {
3255 N->printrFull(Msg, CurDAG);
3256 Msg << "\nIn function: " << MF->getName();
3257 } else {
3258 bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other;
3259 unsigned iid =
3260 cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue();
3261 if (iid < Intrinsic::num_intrinsics)
3262 Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid);
3263 else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo())
3264 Msg << "target intrinsic %" << TII->getName(iid);
3265 else
3266 Msg << "unknown intrinsic #" << iid;
3267 }
3268 report_fatal_error(Msg.str());
3269 }
3270
3271 char SelectionDAGISel::ID = 0;
3272