1 //===-- SelectionDAGISel.cpp - Implement the SelectionDAGISel class -------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the SelectionDAGISel class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/CodeGen/GCStrategy.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/Analysis/TargetLibraryInfo.h"
23 #include "llvm/CodeGen/Analysis.h"
24 #include "llvm/CodeGen/FastISel.h"
25 #include "llvm/CodeGen/FunctionLoweringInfo.h"
26 #include "llvm/CodeGen/GCMetadata.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineModuleInfo.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
33 #include "llvm/CodeGen/SchedulerRegistry.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/CodeGen/SelectionDAGISel.h"
36 #include "llvm/CodeGen/WinEHFuncInfo.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DebugInfo.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/InlineAsm.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/IntrinsicInst.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/IR/LLVMContext.h"
45 #include "llvm/IR/Module.h"
46 #include "llvm/MC/MCAsmInfo.h"
47 #include "llvm/Support/Compiler.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/Timer.h"
51 #include "llvm/Support/raw_ostream.h"
52 #include "llvm/Target/TargetInstrInfo.h"
53 #include "llvm/Target/TargetIntrinsicInfo.h"
54 #include "llvm/Target/TargetLowering.h"
55 #include "llvm/Target/TargetMachine.h"
56 #include "llvm/Target/TargetOptions.h"
57 #include "llvm/Target/TargetRegisterInfo.h"
58 #include "llvm/Target/TargetSubtargetInfo.h"
59 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
63 #define DEBUG_TYPE "isel"
65 STATISTIC(NumFastIselFailures, "Number of instructions fast isel failed on");
66 STATISTIC(NumFastIselSuccess, "Number of instructions fast isel selected");
67 STATISTIC(NumFastIselBlocks, "Number of blocks selected entirely by fast isel");
68 STATISTIC(NumDAGBlocks, "Number of blocks selected using DAG");
69 STATISTIC(NumDAGIselRetries,"Number of times dag isel has to try another path");
70 STATISTIC(NumEntryBlocks, "Number of entry blocks encountered");
71 STATISTIC(NumFastIselFailLowerArguments,
72 "Number of entry blocks where fast isel failed to lower arguments");
76 EnableFastISelVerbose2("fast-isel-verbose2", cl::Hidden,
77 cl::desc("Enable extra verbose messages in the \"fast\" "
78 "instruction selector"));
81 STATISTIC(NumFastIselFailRet,"Fast isel fails on Ret");
82 STATISTIC(NumFastIselFailBr,"Fast isel fails on Br");
83 STATISTIC(NumFastIselFailSwitch,"Fast isel fails on Switch");
84 STATISTIC(NumFastIselFailIndirectBr,"Fast isel fails on IndirectBr");
85 STATISTIC(NumFastIselFailInvoke,"Fast isel fails on Invoke");
86 STATISTIC(NumFastIselFailResume,"Fast isel fails on Resume");
87 STATISTIC(NumFastIselFailUnreachable,"Fast isel fails on Unreachable");
89 // Standard binary operators...
90 STATISTIC(NumFastIselFailAdd,"Fast isel fails on Add");
91 STATISTIC(NumFastIselFailFAdd,"Fast isel fails on FAdd");
92 STATISTIC(NumFastIselFailSub,"Fast isel fails on Sub");
93 STATISTIC(NumFastIselFailFSub,"Fast isel fails on FSub");
94 STATISTIC(NumFastIselFailMul,"Fast isel fails on Mul");
95 STATISTIC(NumFastIselFailFMul,"Fast isel fails on FMul");
96 STATISTIC(NumFastIselFailUDiv,"Fast isel fails on UDiv");
97 STATISTIC(NumFastIselFailSDiv,"Fast isel fails on SDiv");
98 STATISTIC(NumFastIselFailFDiv,"Fast isel fails on FDiv");
99 STATISTIC(NumFastIselFailURem,"Fast isel fails on URem");
100 STATISTIC(NumFastIselFailSRem,"Fast isel fails on SRem");
101 STATISTIC(NumFastIselFailFRem,"Fast isel fails on FRem");
103 // Logical operators...
104 STATISTIC(NumFastIselFailAnd,"Fast isel fails on And");
105 STATISTIC(NumFastIselFailOr,"Fast isel fails on Or");
106 STATISTIC(NumFastIselFailXor,"Fast isel fails on Xor");
108 // Memory instructions...
109 STATISTIC(NumFastIselFailAlloca,"Fast isel fails on Alloca");
110 STATISTIC(NumFastIselFailLoad,"Fast isel fails on Load");
111 STATISTIC(NumFastIselFailStore,"Fast isel fails on Store");
112 STATISTIC(NumFastIselFailAtomicCmpXchg,"Fast isel fails on AtomicCmpXchg");
113 STATISTIC(NumFastIselFailAtomicRMW,"Fast isel fails on AtomicRWM");
114 STATISTIC(NumFastIselFailFence,"Fast isel fails on Frence");
115 STATISTIC(NumFastIselFailGetElementPtr,"Fast isel fails on GetElementPtr");
117 // Convert instructions...
118 STATISTIC(NumFastIselFailTrunc,"Fast isel fails on Trunc");
119 STATISTIC(NumFastIselFailZExt,"Fast isel fails on ZExt");
120 STATISTIC(NumFastIselFailSExt,"Fast isel fails on SExt");
121 STATISTIC(NumFastIselFailFPTrunc,"Fast isel fails on FPTrunc");
122 STATISTIC(NumFastIselFailFPExt,"Fast isel fails on FPExt");
123 STATISTIC(NumFastIselFailFPToUI,"Fast isel fails on FPToUI");
124 STATISTIC(NumFastIselFailFPToSI,"Fast isel fails on FPToSI");
125 STATISTIC(NumFastIselFailUIToFP,"Fast isel fails on UIToFP");
126 STATISTIC(NumFastIselFailSIToFP,"Fast isel fails on SIToFP");
127 STATISTIC(NumFastIselFailIntToPtr,"Fast isel fails on IntToPtr");
128 STATISTIC(NumFastIselFailPtrToInt,"Fast isel fails on PtrToInt");
129 STATISTIC(NumFastIselFailBitCast,"Fast isel fails on BitCast");
131 // Other instructions...
132 STATISTIC(NumFastIselFailICmp,"Fast isel fails on ICmp");
133 STATISTIC(NumFastIselFailFCmp,"Fast isel fails on FCmp");
134 STATISTIC(NumFastIselFailPHI,"Fast isel fails on PHI");
135 STATISTIC(NumFastIselFailSelect,"Fast isel fails on Select");
136 STATISTIC(NumFastIselFailCall,"Fast isel fails on Call");
137 STATISTIC(NumFastIselFailShl,"Fast isel fails on Shl");
138 STATISTIC(NumFastIselFailLShr,"Fast isel fails on LShr");
139 STATISTIC(NumFastIselFailAShr,"Fast isel fails on AShr");
140 STATISTIC(NumFastIselFailVAArg,"Fast isel fails on VAArg");
141 STATISTIC(NumFastIselFailExtractElement,"Fast isel fails on ExtractElement");
142 STATISTIC(NumFastIselFailInsertElement,"Fast isel fails on InsertElement");
143 STATISTIC(NumFastIselFailShuffleVector,"Fast isel fails on ShuffleVector");
144 STATISTIC(NumFastIselFailExtractValue,"Fast isel fails on ExtractValue");
145 STATISTIC(NumFastIselFailInsertValue,"Fast isel fails on InsertValue");
146 STATISTIC(NumFastIselFailLandingPad,"Fast isel fails on LandingPad");
148 // Intrinsic instructions...
149 STATISTIC(NumFastIselFailIntrinsicCall, "Fast isel fails on Intrinsic call");
150 STATISTIC(NumFastIselFailSAddWithOverflow,
151 "Fast isel fails on sadd.with.overflow");
152 STATISTIC(NumFastIselFailUAddWithOverflow,
153 "Fast isel fails on uadd.with.overflow");
154 STATISTIC(NumFastIselFailSSubWithOverflow,
155 "Fast isel fails on ssub.with.overflow");
156 STATISTIC(NumFastIselFailUSubWithOverflow,
157 "Fast isel fails on usub.with.overflow");
158 STATISTIC(NumFastIselFailSMulWithOverflow,
159 "Fast isel fails on smul.with.overflow");
160 STATISTIC(NumFastIselFailUMulWithOverflow,
161 "Fast isel fails on umul.with.overflow");
162 STATISTIC(NumFastIselFailFrameaddress, "Fast isel fails on Frameaddress");
163 STATISTIC(NumFastIselFailSqrt, "Fast isel fails on sqrt call");
164 STATISTIC(NumFastIselFailStackMap, "Fast isel fails on StackMap call");
165 STATISTIC(NumFastIselFailPatchPoint, "Fast isel fails on PatchPoint call");
169 EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
170 cl::desc("Enable verbose messages in the \"fast\" "
171 "instruction selector"));
172 static cl::opt<int> EnableFastISelAbort(
173 "fast-isel-abort", cl::Hidden,
174 cl::desc("Enable abort calls when \"fast\" instruction selection "
175 "fails to lower an instruction: 0 disable the abort, 1 will "
176 "abort but for args, calls and terminators, 2 will also "
177 "abort for argument lowering, and 3 will never fallback "
178 "to SelectionDAG."));
182 cl::desc("use Machine Branch Probability Info"),
183 cl::init(true), cl::Hidden);
186 static cl::opt<std::string>
187 FilterDAGBasicBlockName("filter-view-dags", cl::Hidden,
188 cl::desc("Only display the basic block whose name "
189 "matches this for all view-*-dags options"));
191 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
192 cl::desc("Pop up a window to show dags before the first "
193 "dag combine pass"));
195 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
196 cl::desc("Pop up a window to show dags before legalize types"));
198 ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
199 cl::desc("Pop up a window to show dags before legalize"));
201 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
202 cl::desc("Pop up a window to show dags before the second "
203 "dag combine pass"));
205 ViewDAGCombineLT("view-dag-combine-lt-dags", cl::Hidden,
206 cl::desc("Pop up a window to show dags before the post legalize types"
207 " dag combine pass"));
209 ViewISelDAGs("view-isel-dags", cl::Hidden,
210 cl::desc("Pop up a window to show isel dags as they are selected"));
212 ViewSchedDAGs("view-sched-dags", cl::Hidden,
213 cl::desc("Pop up a window to show sched dags as they are processed"));
215 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
216 cl::desc("Pop up a window to show SUnit dags after they are processed"));
218 static const bool ViewDAGCombine1 = false,
219 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false,
220 ViewDAGCombine2 = false,
221 ViewDAGCombineLT = false,
222 ViewISelDAGs = false, ViewSchedDAGs = false,
223 ViewSUnitDAGs = false;
226 //===---------------------------------------------------------------------===//
228 /// RegisterScheduler class - Track the registration of instruction schedulers.
230 //===---------------------------------------------------------------------===//
231 MachinePassRegistry RegisterScheduler::Registry;
233 //===---------------------------------------------------------------------===//
235 /// ISHeuristic command line option for instruction schedulers.
237 //===---------------------------------------------------------------------===//
238 static cl::opt<RegisterScheduler::FunctionPassCtor, false,
239 RegisterPassParser<RegisterScheduler> >
240 ISHeuristic("pre-RA-sched",
241 cl::init(&createDefaultScheduler), cl::Hidden,
242 cl::desc("Instruction schedulers available (before register"
245 static RegisterScheduler
246 defaultListDAGScheduler("default", "Best scheduler for the target",
247 createDefaultScheduler);
250 //===--------------------------------------------------------------------===//
251 /// \brief This class is used by SelectionDAGISel to temporarily override
252 /// the optimization level on a per-function basis.
253 class OptLevelChanger {
254 SelectionDAGISel &IS;
255 CodeGenOpt::Level SavedOptLevel;
259 OptLevelChanger(SelectionDAGISel &ISel,
260 CodeGenOpt::Level NewOptLevel) : IS(ISel) {
261 SavedOptLevel = IS.OptLevel;
262 if (NewOptLevel == SavedOptLevel)
264 IS.OptLevel = NewOptLevel;
265 IS.TM.setOptLevel(NewOptLevel);
266 SavedFastISel = IS.TM.Options.EnableFastISel;
267 if (NewOptLevel == CodeGenOpt::None)
268 IS.TM.setFastISel(true);
269 DEBUG(dbgs() << "\nChanging optimization level for Function "
270 << IS.MF->getFunction()->getName() << "\n");
271 DEBUG(dbgs() << "\tBefore: -O" << SavedOptLevel
272 << " ; After: -O" << NewOptLevel << "\n");
276 if (IS.OptLevel == SavedOptLevel)
278 DEBUG(dbgs() << "\nRestoring optimization level for Function "
279 << IS.MF->getFunction()->getName() << "\n");
280 DEBUG(dbgs() << "\tBefore: -O" << IS.OptLevel
281 << " ; After: -O" << SavedOptLevel << "\n");
282 IS.OptLevel = SavedOptLevel;
283 IS.TM.setOptLevel(SavedOptLevel);
284 IS.TM.setFastISel(SavedFastISel);
288 //===--------------------------------------------------------------------===//
289 /// createDefaultScheduler - This creates an instruction scheduler appropriate
291 ScheduleDAGSDNodes* createDefaultScheduler(SelectionDAGISel *IS,
292 CodeGenOpt::Level OptLevel) {
293 const TargetLowering *TLI = IS->TLI;
294 const TargetSubtargetInfo &ST = IS->MF->getSubtarget();
296 if (OptLevel == CodeGenOpt::None ||
297 (ST.enableMachineScheduler() && ST.enableMachineSchedDefaultSched()) ||
298 TLI->getSchedulingPreference() == Sched::Source)
299 return createSourceListDAGScheduler(IS, OptLevel);
300 if (TLI->getSchedulingPreference() == Sched::RegPressure)
301 return createBURRListDAGScheduler(IS, OptLevel);
302 if (TLI->getSchedulingPreference() == Sched::Hybrid)
303 return createHybridListDAGScheduler(IS, OptLevel);
304 if (TLI->getSchedulingPreference() == Sched::VLIW)
305 return createVLIWDAGScheduler(IS, OptLevel);
306 assert(TLI->getSchedulingPreference() == Sched::ILP &&
307 "Unknown sched type!");
308 return createILPListDAGScheduler(IS, OptLevel);
312 // EmitInstrWithCustomInserter - This method should be implemented by targets
313 // that mark instructions with the 'usesCustomInserter' flag. These
314 // instructions are special in various ways, which require special support to
315 // insert. The specified MachineInstr is created but not inserted into any
316 // basic blocks, and this method is called to expand it into a sequence of
317 // instructions, potentially also creating new basic blocks and control flow.
318 // When new basic blocks are inserted and the edges from MBB to its successors
319 // are modified, the method should insert pairs of <OldSucc, NewSucc> into the
322 TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
323 MachineBasicBlock *MBB) const {
325 dbgs() << "If a target marks an instruction with "
326 "'usesCustomInserter', it must implement "
327 "TargetLowering::EmitInstrWithCustomInserter!";
329 llvm_unreachable(nullptr);
332 void TargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
333 SDNode *Node) const {
334 assert(!MI->hasPostISelHook() &&
335 "If a target marks an instruction with 'hasPostISelHook', "
336 "it must implement TargetLowering::AdjustInstrPostInstrSelection!");
339 //===----------------------------------------------------------------------===//
340 // SelectionDAGISel code
341 //===----------------------------------------------------------------------===//
343 SelectionDAGISel::SelectionDAGISel(TargetMachine &tm,
344 CodeGenOpt::Level OL) :
345 MachineFunctionPass(ID), TM(tm),
346 FuncInfo(new FunctionLoweringInfo()),
347 CurDAG(new SelectionDAG(tm, OL)),
348 SDB(new SelectionDAGBuilder(*CurDAG, *FuncInfo, OL)),
352 initializeGCModuleInfoPass(*PassRegistry::getPassRegistry());
353 initializeAliasAnalysisAnalysisGroup(*PassRegistry::getPassRegistry());
354 initializeBranchProbabilityInfoPass(*PassRegistry::getPassRegistry());
355 initializeTargetLibraryInfoWrapperPassPass(
356 *PassRegistry::getPassRegistry());
359 SelectionDAGISel::~SelectionDAGISel() {
365 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
366 AU.addRequired<AliasAnalysis>();
367 AU.addPreserved<AliasAnalysis>();
368 AU.addRequired<GCModuleInfo>();
369 AU.addPreserved<GCModuleInfo>();
370 AU.addRequired<TargetLibraryInfoWrapperPass>();
371 if (UseMBPI && OptLevel != CodeGenOpt::None)
372 AU.addRequired<BranchProbabilityInfo>();
373 MachineFunctionPass::getAnalysisUsage(AU);
376 /// SplitCriticalSideEffectEdges - Look for critical edges with a PHI value that
377 /// may trap on it. In this case we have to split the edge so that the path
378 /// through the predecessor block that doesn't go to the phi block doesn't
379 /// execute the possibly trapping instruction.
381 /// This is required for correctness, so it must be done at -O0.
383 static void SplitCriticalSideEffectEdges(Function &Fn, AliasAnalysis *AA) {
384 // Loop for blocks with phi nodes.
385 for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
386 PHINode *PN = dyn_cast<PHINode>(BB->begin());
390 // For each block with a PHI node, check to see if any of the input values
391 // are potentially trapping constant expressions. Constant expressions are
392 // the only potentially trapping value that can occur as the argument to a
394 for (BasicBlock::iterator I = BB->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
395 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
396 ConstantExpr *CE = dyn_cast<ConstantExpr>(PN->getIncomingValue(i));
397 if (!CE || !CE->canTrap()) continue;
399 // The only case we have to worry about is when the edge is critical.
400 // Since this block has a PHI Node, we assume it has multiple input
401 // edges: check to see if the pred has multiple successors.
402 BasicBlock *Pred = PN->getIncomingBlock(i);
403 if (Pred->getTerminator()->getNumSuccessors() == 1)
406 // Okay, we have to split this edge.
408 Pred->getTerminator(), GetSuccessorNumber(Pred, BB),
409 CriticalEdgeSplittingOptions(AA).setMergeIdenticalEdges());
415 bool SelectionDAGISel::runOnMachineFunction(MachineFunction &mf) {
416 // Do some sanity-checking on the command-line options.
417 assert((!EnableFastISelVerbose || TM.Options.EnableFastISel) &&
418 "-fast-isel-verbose requires -fast-isel");
419 assert((!EnableFastISelAbort || TM.Options.EnableFastISel) &&
420 "-fast-isel-abort > 0 requires -fast-isel");
422 const Function &Fn = *mf.getFunction();
425 // Reset the target options before resetting the optimization
427 // FIXME: This is a horrible hack and should be processed via
428 // codegen looking at the optimization level explicitly when
429 // it wants to look at it.
430 TM.resetTargetOptions(Fn);
431 // Reset OptLevel to None for optnone functions.
432 CodeGenOpt::Level NewOptLevel = OptLevel;
433 if (Fn.hasFnAttribute(Attribute::OptimizeNone))
434 NewOptLevel = CodeGenOpt::None;
435 OptLevelChanger OLC(*this, NewOptLevel);
437 TII = MF->getSubtarget().getInstrInfo();
438 TLI = MF->getSubtarget().getTargetLowering();
439 RegInfo = &MF->getRegInfo();
440 AA = &getAnalysis<AliasAnalysis>();
441 LibInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
442 GFI = Fn.hasGC() ? &getAnalysis<GCModuleInfo>().getFunctionInfo(Fn) : nullptr;
444 DEBUG(dbgs() << "\n\n\n=== " << Fn.getName() << "\n");
446 SplitCriticalSideEffectEdges(const_cast<Function&>(Fn), AA);
449 FuncInfo->set(Fn, *MF, CurDAG);
451 if (UseMBPI && OptLevel != CodeGenOpt::None)
452 FuncInfo->BPI = &getAnalysis<BranchProbabilityInfo>();
454 FuncInfo->BPI = nullptr;
456 SDB->init(GFI, *AA, LibInfo);
458 MF->setHasInlineAsm(false);
460 SelectAllBasicBlocks(Fn);
462 // If the first basic block in the function has live ins that need to be
463 // copied into vregs, emit the copies into the top of the block before
464 // emitting the code for the block.
465 MachineBasicBlock *EntryMBB = MF->begin();
466 const TargetRegisterInfo &TRI = *MF->getSubtarget().getRegisterInfo();
467 RegInfo->EmitLiveInCopies(EntryMBB, TRI, *TII);
469 DenseMap<unsigned, unsigned> LiveInMap;
470 if (!FuncInfo->ArgDbgValues.empty())
471 for (MachineRegisterInfo::livein_iterator LI = RegInfo->livein_begin(),
472 E = RegInfo->livein_end(); LI != E; ++LI)
474 LiveInMap.insert(std::make_pair(LI->first, LI->second));
476 // Insert DBG_VALUE instructions for function arguments to the entry block.
477 for (unsigned i = 0, e = FuncInfo->ArgDbgValues.size(); i != e; ++i) {
478 MachineInstr *MI = FuncInfo->ArgDbgValues[e-i-1];
479 bool hasFI = MI->getOperand(0).isFI();
481 hasFI ? TRI.getFrameRegister(*MF) : MI->getOperand(0).getReg();
482 if (TargetRegisterInfo::isPhysicalRegister(Reg))
483 EntryMBB->insert(EntryMBB->begin(), MI);
485 MachineInstr *Def = RegInfo->getVRegDef(Reg);
487 MachineBasicBlock::iterator InsertPos = Def;
488 // FIXME: VR def may not be in entry block.
489 Def->getParent()->insert(std::next(InsertPos), MI);
491 DEBUG(dbgs() << "Dropping debug info for dead vreg"
492 << TargetRegisterInfo::virtReg2Index(Reg) << "\n");
495 // If Reg is live-in then update debug info to track its copy in a vreg.
496 DenseMap<unsigned, unsigned>::iterator LDI = LiveInMap.find(Reg);
497 if (LDI != LiveInMap.end()) {
498 assert(!hasFI && "There's no handling of frame pointer updating here yet "
500 MachineInstr *Def = RegInfo->getVRegDef(LDI->second);
501 MachineBasicBlock::iterator InsertPos = Def;
502 const MDNode *Variable = MI->getDebugVariable();
503 const MDNode *Expr = MI->getDebugExpression();
504 DebugLoc DL = MI->getDebugLoc();
505 bool IsIndirect = MI->isIndirectDebugValue();
506 unsigned Offset = IsIndirect ? MI->getOperand(1).getImm() : 0;
507 assert(cast<DILocalVariable>(Variable)->isValidLocationForIntrinsic(DL) &&
508 "Expected inlined-at fields to agree");
509 // Def is never a terminator here, so it is ok to increment InsertPos.
510 BuildMI(*EntryMBB, ++InsertPos, DL, TII->get(TargetOpcode::DBG_VALUE),
511 IsIndirect, LDI->second, Offset, Variable, Expr);
513 // If this vreg is directly copied into an exported register then
514 // that COPY instructions also need DBG_VALUE, if it is the only
515 // user of LDI->second.
516 MachineInstr *CopyUseMI = nullptr;
517 for (MachineRegisterInfo::use_instr_iterator
518 UI = RegInfo->use_instr_begin(LDI->second),
519 E = RegInfo->use_instr_end(); UI != E; ) {
520 MachineInstr *UseMI = &*(UI++);
521 if (UseMI->isDebugValue()) continue;
522 if (UseMI->isCopy() && !CopyUseMI && UseMI->getParent() == EntryMBB) {
523 CopyUseMI = UseMI; continue;
525 // Otherwise this is another use or second copy use.
526 CopyUseMI = nullptr; break;
529 // Use MI's debug location, which describes where Variable was
530 // declared, rather than whatever is attached to CopyUseMI.
531 MachineInstr *NewMI =
532 BuildMI(*MF, DL, TII->get(TargetOpcode::DBG_VALUE), IsIndirect,
533 CopyUseMI->getOperand(0).getReg(), Offset, Variable, Expr);
534 MachineBasicBlock::iterator Pos = CopyUseMI;
535 EntryMBB->insertAfter(Pos, NewMI);
540 // Determine if there are any calls in this machine function.
541 MachineFrameInfo *MFI = MF->getFrameInfo();
542 for (const auto &MBB : *MF) {
543 if (MFI->hasCalls() && MF->hasInlineAsm())
546 for (const auto &MI : MBB) {
547 const MCInstrDesc &MCID = TII->get(MI.getOpcode());
548 if ((MCID.isCall() && !MCID.isReturn()) ||
549 MI.isStackAligningInlineAsm()) {
550 MFI->setHasCalls(true);
552 if (MI.isInlineAsm()) {
553 MF->setHasInlineAsm(true);
558 // Determine if there is a call to setjmp in the machine function.
559 MF->setExposesReturnsTwice(Fn.callsFunctionThatReturnsTwice());
561 // Replace forward-declared registers with the registers containing
562 // the desired value.
563 MachineRegisterInfo &MRI = MF->getRegInfo();
564 for (DenseMap<unsigned, unsigned>::iterator
565 I = FuncInfo->RegFixups.begin(), E = FuncInfo->RegFixups.end();
567 unsigned From = I->first;
568 unsigned To = I->second;
569 // If To is also scheduled to be replaced, find what its ultimate
572 DenseMap<unsigned, unsigned>::iterator J = FuncInfo->RegFixups.find(To);
576 // Make sure the new register has a sufficiently constrained register class.
577 if (TargetRegisterInfo::isVirtualRegister(From) &&
578 TargetRegisterInfo::isVirtualRegister(To))
579 MRI.constrainRegClass(To, MRI.getRegClass(From));
583 // Replacing one register with another won't touch the kill flags.
584 // We need to conservatively clear the kill flags as a kill on the old
585 // register might dominate existing uses of the new register.
586 if (!MRI.use_empty(To))
587 MRI.clearKillFlags(From);
588 MRI.replaceRegWith(From, To);
591 // Freeze the set of reserved registers now that MachineFrameInfo has been
592 // set up. All the information required by getReservedRegs() should be
594 MRI.freezeReservedRegs(*MF);
596 // Release function-specific state. SDB and CurDAG are already cleared
600 DEBUG(dbgs() << "*** MachineFunction at end of ISel ***\n");
601 DEBUG(MF->print(dbgs()));
606 void SelectionDAGISel::SelectBasicBlock(BasicBlock::const_iterator Begin,
607 BasicBlock::const_iterator End,
609 // Lower the instructions. If a call is emitted as a tail call, cease emitting
610 // nodes for this block.
611 for (BasicBlock::const_iterator I = Begin; I != End && !SDB->HasTailCall; ++I)
614 // Make sure the root of the DAG is up-to-date.
615 CurDAG->setRoot(SDB->getControlRoot());
616 HadTailCall = SDB->HasTailCall;
619 // Final step, emit the lowered DAG as machine code.
623 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
624 SmallPtrSet<SDNode*, 128> VisitedNodes;
625 SmallVector<SDNode*, 128> Worklist;
627 Worklist.push_back(CurDAG->getRoot().getNode());
633 SDNode *N = Worklist.pop_back_val();
635 // If we've already seen this node, ignore it.
636 if (!VisitedNodes.insert(N).second)
639 // Otherwise, add all chain operands to the worklist.
640 for (const SDValue &Op : N->op_values())
641 if (Op.getValueType() == MVT::Other)
642 Worklist.push_back(Op.getNode());
644 // If this is a CopyToReg with a vreg dest, process it.
645 if (N->getOpcode() != ISD::CopyToReg)
648 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
649 if (!TargetRegisterInfo::isVirtualRegister(DestReg))
652 // Ignore non-scalar or non-integer values.
653 SDValue Src = N->getOperand(2);
654 EVT SrcVT = Src.getValueType();
655 if (!SrcVT.isInteger() || SrcVT.isVector())
658 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
659 CurDAG->computeKnownBits(Src, KnownZero, KnownOne);
660 FuncInfo->AddLiveOutRegInfo(DestReg, NumSignBits, KnownZero, KnownOne);
661 } while (!Worklist.empty());
664 void SelectionDAGISel::CodeGenAndEmitDAG() {
665 std::string GroupName;
666 if (TimePassesIsEnabled)
667 GroupName = "Instruction Selection and Scheduling";
668 std::string BlockName;
669 int BlockNumber = -1;
671 bool MatchFilterBB = false; (void)MatchFilterBB;
673 MatchFilterBB = (FilterDAGBasicBlockName.empty() ||
674 FilterDAGBasicBlockName ==
675 FuncInfo->MBB->getBasicBlock()->getName().str());
678 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs ||
679 ViewDAGCombine2 || ViewDAGCombineLT || ViewISelDAGs || ViewSchedDAGs ||
683 BlockNumber = FuncInfo->MBB->getNumber();
685 (MF->getName() + ":" + FuncInfo->MBB->getBasicBlock()->getName()).str();
687 DEBUG(dbgs() << "Initial selection DAG: BB#" << BlockNumber
688 << " '" << BlockName << "'\n"; CurDAG->dump());
690 if (ViewDAGCombine1 && MatchFilterBB)
691 CurDAG->viewGraph("dag-combine1 input for " + BlockName);
693 // Run the DAG combiner in pre-legalize mode.
695 NamedRegionTimer T("DAG Combining 1", GroupName, TimePassesIsEnabled);
696 CurDAG->Combine(BeforeLegalizeTypes, *AA, OptLevel);
699 DEBUG(dbgs() << "Optimized lowered selection DAG: BB#" << BlockNumber
700 << " '" << BlockName << "'\n"; CurDAG->dump());
702 // Second step, hack on the DAG until it only uses operations and types that
703 // the target supports.
704 if (ViewLegalizeTypesDAGs && MatchFilterBB)
705 CurDAG->viewGraph("legalize-types input for " + BlockName);
709 NamedRegionTimer T("Type Legalization", GroupName, TimePassesIsEnabled);
710 Changed = CurDAG->LegalizeTypes();
713 DEBUG(dbgs() << "Type-legalized selection DAG: BB#" << BlockNumber
714 << " '" << BlockName << "'\n"; CurDAG->dump());
716 CurDAG->NewNodesMustHaveLegalTypes = true;
719 if (ViewDAGCombineLT && MatchFilterBB)
720 CurDAG->viewGraph("dag-combine-lt input for " + BlockName);
722 // Run the DAG combiner in post-type-legalize mode.
724 NamedRegionTimer T("DAG Combining after legalize types", GroupName,
725 TimePassesIsEnabled);
726 CurDAG->Combine(AfterLegalizeTypes, *AA, OptLevel);
729 DEBUG(dbgs() << "Optimized type-legalized selection DAG: BB#" << BlockNumber
730 << " '" << BlockName << "'\n"; CurDAG->dump());
735 NamedRegionTimer T("Vector Legalization", GroupName, TimePassesIsEnabled);
736 Changed = CurDAG->LegalizeVectors();
741 NamedRegionTimer T("Type Legalization 2", GroupName, TimePassesIsEnabled);
742 CurDAG->LegalizeTypes();
745 if (ViewDAGCombineLT && MatchFilterBB)
746 CurDAG->viewGraph("dag-combine-lv input for " + BlockName);
748 // Run the DAG combiner in post-type-legalize mode.
750 NamedRegionTimer T("DAG Combining after legalize vectors", GroupName,
751 TimePassesIsEnabled);
752 CurDAG->Combine(AfterLegalizeVectorOps, *AA, OptLevel);
755 DEBUG(dbgs() << "Optimized vector-legalized selection DAG: BB#"
756 << BlockNumber << " '" << BlockName << "'\n"; CurDAG->dump());
759 if (ViewLegalizeDAGs && MatchFilterBB)
760 CurDAG->viewGraph("legalize input for " + BlockName);
763 NamedRegionTimer T("DAG Legalization", GroupName, TimePassesIsEnabled);
767 DEBUG(dbgs() << "Legalized selection DAG: BB#" << BlockNumber
768 << " '" << BlockName << "'\n"; CurDAG->dump());
770 if (ViewDAGCombine2 && MatchFilterBB)
771 CurDAG->viewGraph("dag-combine2 input for " + BlockName);
773 // Run the DAG combiner in post-legalize mode.
775 NamedRegionTimer T("DAG Combining 2", GroupName, TimePassesIsEnabled);
776 CurDAG->Combine(AfterLegalizeDAG, *AA, OptLevel);
779 DEBUG(dbgs() << "Optimized legalized selection DAG: BB#" << BlockNumber
780 << " '" << BlockName << "'\n"; CurDAG->dump());
782 if (OptLevel != CodeGenOpt::None)
783 ComputeLiveOutVRegInfo();
785 if (ViewISelDAGs && MatchFilterBB)
786 CurDAG->viewGraph("isel input for " + BlockName);
788 // Third, instruction select all of the operations to machine code, adding the
789 // code to the MachineBasicBlock.
791 NamedRegionTimer T("Instruction Selection", GroupName, TimePassesIsEnabled);
792 DoInstructionSelection();
795 DEBUG(dbgs() << "Selected selection DAG: BB#" << BlockNumber
796 << " '" << BlockName << "'\n"; CurDAG->dump());
798 if (ViewSchedDAGs && MatchFilterBB)
799 CurDAG->viewGraph("scheduler input for " + BlockName);
801 // Schedule machine code.
802 ScheduleDAGSDNodes *Scheduler = CreateScheduler();
804 NamedRegionTimer T("Instruction Scheduling", GroupName,
805 TimePassesIsEnabled);
806 Scheduler->Run(CurDAG, FuncInfo->MBB);
809 if (ViewSUnitDAGs && MatchFilterBB) Scheduler->viewGraph();
811 // Emit machine code to BB. This can change 'BB' to the last block being
813 MachineBasicBlock *FirstMBB = FuncInfo->MBB, *LastMBB;
815 NamedRegionTimer T("Instruction Creation", GroupName, TimePassesIsEnabled);
817 // FuncInfo->InsertPt is passed by reference and set to the end of the
818 // scheduled instructions.
819 LastMBB = FuncInfo->MBB = Scheduler->EmitSchedule(FuncInfo->InsertPt);
822 // If the block was split, make sure we update any references that are used to
823 // update PHI nodes later on.
824 if (FirstMBB != LastMBB)
825 SDB->UpdateSplitBlock(FirstMBB, LastMBB);
827 // Free the scheduler state.
829 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName,
830 TimePassesIsEnabled);
834 // Free the SelectionDAG state, now that we're finished with it.
839 /// ISelUpdater - helper class to handle updates of the instruction selection
841 class ISelUpdater : public SelectionDAG::DAGUpdateListener {
842 SelectionDAG::allnodes_iterator &ISelPosition;
844 ISelUpdater(SelectionDAG &DAG, SelectionDAG::allnodes_iterator &isp)
845 : SelectionDAG::DAGUpdateListener(DAG), ISelPosition(isp) {}
847 /// NodeDeleted - Handle nodes deleted from the graph. If the node being
848 /// deleted is the current ISelPosition node, update ISelPosition.
850 void NodeDeleted(SDNode *N, SDNode *E) override {
851 if (ISelPosition == SelectionDAG::allnodes_iterator(N))
855 } // end anonymous namespace
857 void SelectionDAGISel::DoInstructionSelection() {
858 DEBUG(dbgs() << "===== Instruction selection begins: BB#"
859 << FuncInfo->MBB->getNumber()
860 << " '" << FuncInfo->MBB->getName() << "'\n");
864 // Select target instructions for the DAG.
866 // Number all nodes with a topological order and set DAGSize.
867 DAGSize = CurDAG->AssignTopologicalOrder();
869 // Create a dummy node (which is not added to allnodes), that adds
870 // a reference to the root node, preventing it from being deleted,
871 // and tracking any changes of the root.
872 HandleSDNode Dummy(CurDAG->getRoot());
873 SelectionDAG::allnodes_iterator ISelPosition (CurDAG->getRoot().getNode());
876 // Make sure that ISelPosition gets properly updated when nodes are deleted
877 // in calls made from this function.
878 ISelUpdater ISU(*CurDAG, ISelPosition);
880 // The AllNodes list is now topological-sorted. Visit the
881 // nodes by starting at the end of the list (the root of the
882 // graph) and preceding back toward the beginning (the entry
884 while (ISelPosition != CurDAG->allnodes_begin()) {
885 SDNode *Node = --ISelPosition;
886 // Skip dead nodes. DAGCombiner is expected to eliminate all dead nodes,
887 // but there are currently some corner cases that it misses. Also, this
888 // makes it theoretically possible to disable the DAGCombiner.
889 if (Node->use_empty())
892 SDNode *ResNode = Select(Node);
894 // FIXME: This is pretty gross. 'Select' should be changed to not return
895 // anything at all and this code should be nuked with a tactical strike.
897 // If node should not be replaced, continue with the next one.
898 if (ResNode == Node || Node->getOpcode() == ISD::DELETED_NODE)
902 ReplaceUses(Node, ResNode);
905 // If after the replacement this node is not used any more,
906 // remove this dead node.
907 if (Node->use_empty()) // Don't delete EntryToken, etc.
908 CurDAG->RemoveDeadNode(Node);
911 CurDAG->setRoot(Dummy.getValue());
914 DEBUG(dbgs() << "===== Instruction selection ends:\n");
916 PostprocessISelDAG();
919 /// PrepareEHLandingPad - Emit an EH_LABEL, set up live-in registers, and
920 /// do other setup for EH landing-pad blocks.
921 bool SelectionDAGISel::PrepareEHLandingPad() {
922 MachineBasicBlock *MBB = FuncInfo->MBB;
924 const TargetRegisterClass *PtrRC =
925 TLI->getRegClassFor(TLI->getPointerTy(CurDAG->getDataLayout()));
927 // Add a label to mark the beginning of the landing pad. Deletion of the
928 // landing pad can thus be detected via the MachineModuleInfo.
929 MCSymbol *Label = MF->getMMI().addLandingPad(MBB);
931 // Assign the call site to the landing pad's begin label.
932 MF->getMMI().setCallSiteLandingPad(Label, SDB->LPadToCallSiteMap[MBB]);
934 const MCInstrDesc &II = TII->get(TargetOpcode::EH_LABEL);
935 BuildMI(*MBB, FuncInfo->InsertPt, SDB->getCurDebugLoc(), II)
938 // If this is an MSVC-style personality function, we need to split the landing
939 // pad into several BBs.
940 const BasicBlock *LLVMBB = MBB->getBasicBlock();
941 const LandingPadInst *LPadInst = LLVMBB->getLandingPadInst();
942 MF->getMMI().addPersonality(MBB, cast<Function>(LPadInst->getParent()
945 ->stripPointerCasts()));
946 EHPersonality Personality = MF->getMMI().getPersonalityType();
948 if (isMSVCEHPersonality(Personality)) {
949 SmallVector<MachineBasicBlock *, 4> ClauseBBs;
950 const IntrinsicInst *ActionsCall =
951 dyn_cast<IntrinsicInst>(LLVMBB->getFirstInsertionPt());
952 // Get all invoke BBs that unwind to this landingpad.
953 SmallVector<MachineBasicBlock *, 4> InvokeBBs(MBB->pred_begin(),
955 if (ActionsCall && ActionsCall->getIntrinsicID() == Intrinsic::eh_actions) {
956 // If this is a call to llvm.eh.actions followed by indirectbr, then we've
957 // run WinEHPrepare, and we should remove this block from the machine CFG.
958 // Mark the targets of the indirectbr as landingpads instead.
959 for (const BasicBlock *LLVMSucc : successors(LLVMBB)) {
960 MachineBasicBlock *ClauseBB = FuncInfo->MBBMap[LLVMSucc];
961 // Add the edge from the invoke to the clause.
962 for (MachineBasicBlock *InvokeBB : InvokeBBs)
963 InvokeBB->addSuccessor(ClauseBB);
965 // Mark the clause as a landing pad or MI passes will delete it.
966 ClauseBB->setIsLandingPad();
970 // Remove the edge from the invoke to the lpad.
971 for (MachineBasicBlock *InvokeBB : InvokeBBs)
972 InvokeBB->removeSuccessor(MBB);
974 // Don't select instructions for the landingpad.
978 // Mark exception register as live in.
979 if (unsigned Reg = TLI->getExceptionPointerRegister())
980 FuncInfo->ExceptionPointerVirtReg = MBB->addLiveIn(Reg, PtrRC);
982 // Mark exception selector register as live in.
983 if (unsigned Reg = TLI->getExceptionSelectorRegister())
984 FuncInfo->ExceptionSelectorVirtReg = MBB->addLiveIn(Reg, PtrRC);
989 /// isFoldedOrDeadInstruction - Return true if the specified instruction is
990 /// side-effect free and is either dead or folded into a generated instruction.
991 /// Return false if it needs to be emitted.
992 static bool isFoldedOrDeadInstruction(const Instruction *I,
993 FunctionLoweringInfo *FuncInfo) {
994 return !I->mayWriteToMemory() && // Side-effecting instructions aren't folded.
995 !isa<TerminatorInst>(I) && // Terminators aren't folded.
996 !isa<DbgInfoIntrinsic>(I) && // Debug instructions aren't folded.
997 !isa<LandingPadInst>(I) && // Landingpad instructions aren't folded.
998 !FuncInfo->isExportedInst(I); // Exported instrs must be computed.
1002 // Collect per Instruction statistics for fast-isel misses. Only those
1003 // instructions that cause the bail are accounted for. It does not account for
1004 // instructions higher in the block. Thus, summing the per instructions stats
1005 // will not add up to what is reported by NumFastIselFailures.
1006 static void collectFailStats(const Instruction *I) {
1007 switch (I->getOpcode()) {
1008 default: assert (0 && "<Invalid operator> ");
1011 case Instruction::Ret: NumFastIselFailRet++; return;
1012 case Instruction::Br: NumFastIselFailBr++; return;
1013 case Instruction::Switch: NumFastIselFailSwitch++; return;
1014 case Instruction::IndirectBr: NumFastIselFailIndirectBr++; return;
1015 case Instruction::Invoke: NumFastIselFailInvoke++; return;
1016 case Instruction::Resume: NumFastIselFailResume++; return;
1017 case Instruction::Unreachable: NumFastIselFailUnreachable++; return;
1019 // Standard binary operators...
1020 case Instruction::Add: NumFastIselFailAdd++; return;
1021 case Instruction::FAdd: NumFastIselFailFAdd++; return;
1022 case Instruction::Sub: NumFastIselFailSub++; return;
1023 case Instruction::FSub: NumFastIselFailFSub++; return;
1024 case Instruction::Mul: NumFastIselFailMul++; return;
1025 case Instruction::FMul: NumFastIselFailFMul++; return;
1026 case Instruction::UDiv: NumFastIselFailUDiv++; return;
1027 case Instruction::SDiv: NumFastIselFailSDiv++; return;
1028 case Instruction::FDiv: NumFastIselFailFDiv++; return;
1029 case Instruction::URem: NumFastIselFailURem++; return;
1030 case Instruction::SRem: NumFastIselFailSRem++; return;
1031 case Instruction::FRem: NumFastIselFailFRem++; return;
1033 // Logical operators...
1034 case Instruction::And: NumFastIselFailAnd++; return;
1035 case Instruction::Or: NumFastIselFailOr++; return;
1036 case Instruction::Xor: NumFastIselFailXor++; return;
1038 // Memory instructions...
1039 case Instruction::Alloca: NumFastIselFailAlloca++; return;
1040 case Instruction::Load: NumFastIselFailLoad++; return;
1041 case Instruction::Store: NumFastIselFailStore++; return;
1042 case Instruction::AtomicCmpXchg: NumFastIselFailAtomicCmpXchg++; return;
1043 case Instruction::AtomicRMW: NumFastIselFailAtomicRMW++; return;
1044 case Instruction::Fence: NumFastIselFailFence++; return;
1045 case Instruction::GetElementPtr: NumFastIselFailGetElementPtr++; return;
1047 // Convert instructions...
1048 case Instruction::Trunc: NumFastIselFailTrunc++; return;
1049 case Instruction::ZExt: NumFastIselFailZExt++; return;
1050 case Instruction::SExt: NumFastIselFailSExt++; return;
1051 case Instruction::FPTrunc: NumFastIselFailFPTrunc++; return;
1052 case Instruction::FPExt: NumFastIselFailFPExt++; return;
1053 case Instruction::FPToUI: NumFastIselFailFPToUI++; return;
1054 case Instruction::FPToSI: NumFastIselFailFPToSI++; return;
1055 case Instruction::UIToFP: NumFastIselFailUIToFP++; return;
1056 case Instruction::SIToFP: NumFastIselFailSIToFP++; return;
1057 case Instruction::IntToPtr: NumFastIselFailIntToPtr++; return;
1058 case Instruction::PtrToInt: NumFastIselFailPtrToInt++; return;
1059 case Instruction::BitCast: NumFastIselFailBitCast++; return;
1061 // Other instructions...
1062 case Instruction::ICmp: NumFastIselFailICmp++; return;
1063 case Instruction::FCmp: NumFastIselFailFCmp++; return;
1064 case Instruction::PHI: NumFastIselFailPHI++; return;
1065 case Instruction::Select: NumFastIselFailSelect++; return;
1066 case Instruction::Call: {
1067 if (auto const *Intrinsic = dyn_cast<IntrinsicInst>(I)) {
1068 switch (Intrinsic->getIntrinsicID()) {
1070 NumFastIselFailIntrinsicCall++; return;
1071 case Intrinsic::sadd_with_overflow:
1072 NumFastIselFailSAddWithOverflow++; return;
1073 case Intrinsic::uadd_with_overflow:
1074 NumFastIselFailUAddWithOverflow++; return;
1075 case Intrinsic::ssub_with_overflow:
1076 NumFastIselFailSSubWithOverflow++; return;
1077 case Intrinsic::usub_with_overflow:
1078 NumFastIselFailUSubWithOverflow++; return;
1079 case Intrinsic::smul_with_overflow:
1080 NumFastIselFailSMulWithOverflow++; return;
1081 case Intrinsic::umul_with_overflow:
1082 NumFastIselFailUMulWithOverflow++; return;
1083 case Intrinsic::frameaddress:
1084 NumFastIselFailFrameaddress++; return;
1085 case Intrinsic::sqrt:
1086 NumFastIselFailSqrt++; return;
1087 case Intrinsic::experimental_stackmap:
1088 NumFastIselFailStackMap++; return;
1089 case Intrinsic::experimental_patchpoint_void: // fall-through
1090 case Intrinsic::experimental_patchpoint_i64:
1091 NumFastIselFailPatchPoint++; return;
1094 NumFastIselFailCall++;
1097 case Instruction::Shl: NumFastIselFailShl++; return;
1098 case Instruction::LShr: NumFastIselFailLShr++; return;
1099 case Instruction::AShr: NumFastIselFailAShr++; return;
1100 case Instruction::VAArg: NumFastIselFailVAArg++; return;
1101 case Instruction::ExtractElement: NumFastIselFailExtractElement++; return;
1102 case Instruction::InsertElement: NumFastIselFailInsertElement++; return;
1103 case Instruction::ShuffleVector: NumFastIselFailShuffleVector++; return;
1104 case Instruction::ExtractValue: NumFastIselFailExtractValue++; return;
1105 case Instruction::InsertValue: NumFastIselFailInsertValue++; return;
1106 case Instruction::LandingPad: NumFastIselFailLandingPad++; return;
1111 void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
1112 // Initialize the Fast-ISel state, if needed.
1113 FastISel *FastIS = nullptr;
1114 if (TM.Options.EnableFastISel)
1115 FastIS = TLI->createFastISel(*FuncInfo, LibInfo);
1117 // Iterate over all basic blocks in the function.
1118 ReversePostOrderTraversal<const Function*> RPOT(&Fn);
1119 for (ReversePostOrderTraversal<const Function*>::rpo_iterator
1120 I = RPOT.begin(), E = RPOT.end(); I != E; ++I) {
1121 const BasicBlock *LLVMBB = *I;
1123 if (OptLevel != CodeGenOpt::None) {
1124 bool AllPredsVisited = true;
1125 for (const_pred_iterator PI = pred_begin(LLVMBB), PE = pred_end(LLVMBB);
1127 if (!FuncInfo->VisitedBBs.count(*PI)) {
1128 AllPredsVisited = false;
1133 if (AllPredsVisited) {
1134 for (BasicBlock::const_iterator I = LLVMBB->begin();
1135 const PHINode *PN = dyn_cast<PHINode>(I); ++I)
1136 FuncInfo->ComputePHILiveOutRegInfo(PN);
1138 for (BasicBlock::const_iterator I = LLVMBB->begin();
1139 const PHINode *PN = dyn_cast<PHINode>(I); ++I)
1140 FuncInfo->InvalidatePHILiveOutRegInfo(PN);
1143 FuncInfo->VisitedBBs.insert(LLVMBB);
1146 BasicBlock::const_iterator const Begin = LLVMBB->getFirstNonPHI();
1147 BasicBlock::const_iterator const End = LLVMBB->end();
1148 BasicBlock::const_iterator BI = End;
1150 FuncInfo->MBB = FuncInfo->MBBMap[LLVMBB];
1151 FuncInfo->InsertPt = FuncInfo->MBB->getFirstNonPHI();
1153 // Setup an EH landing-pad block.
1154 FuncInfo->ExceptionPointerVirtReg = 0;
1155 FuncInfo->ExceptionSelectorVirtReg = 0;
1156 if (LLVMBB->isLandingPad())
1157 if (!PrepareEHLandingPad())
1160 // Before doing SelectionDAG ISel, see if FastISel has been requested.
1162 FastIS->startNewBlock();
1164 // Emit code for any incoming arguments. This must happen before
1165 // beginning FastISel on the entry block.
1166 if (LLVMBB == &Fn.getEntryBlock()) {
1169 // Lower any arguments needed in this block if this is the entry block.
1170 if (!FastIS->lowerArguments()) {
1171 // Fast isel failed to lower these arguments
1172 ++NumFastIselFailLowerArguments;
1173 if (EnableFastISelAbort > 1)
1174 report_fatal_error("FastISel didn't lower all arguments");
1176 // Use SelectionDAG argument lowering
1178 CurDAG->setRoot(SDB->getControlRoot());
1180 CodeGenAndEmitDAG();
1183 // If we inserted any instructions at the beginning, make a note of
1184 // where they are, so we can be sure to emit subsequent instructions
1186 if (FuncInfo->InsertPt != FuncInfo->MBB->begin())
1187 FastIS->setLastLocalValue(std::prev(FuncInfo->InsertPt));
1189 FastIS->setLastLocalValue(nullptr);
1192 unsigned NumFastIselRemaining = std::distance(Begin, End);
1193 // Do FastISel on as many instructions as possible.
1194 for (; BI != Begin; --BI) {
1195 const Instruction *Inst = std::prev(BI);
1197 // If we no longer require this instruction, skip it.
1198 if (isFoldedOrDeadInstruction(Inst, FuncInfo)) {
1199 --NumFastIselRemaining;
1203 // Bottom-up: reset the insert pos at the top, after any local-value
1205 FastIS->recomputeInsertPt();
1207 // Try to select the instruction with FastISel.
1208 if (FastIS->selectInstruction(Inst)) {
1209 --NumFastIselRemaining;
1210 ++NumFastIselSuccess;
1211 // If fast isel succeeded, skip over all the folded instructions, and
1212 // then see if there is a load right before the selected instructions.
1213 // Try to fold the load if so.
1214 const Instruction *BeforeInst = Inst;
1215 while (BeforeInst != Begin) {
1216 BeforeInst = std::prev(BasicBlock::const_iterator(BeforeInst));
1217 if (!isFoldedOrDeadInstruction(BeforeInst, FuncInfo))
1220 if (BeforeInst != Inst && isa<LoadInst>(BeforeInst) &&
1221 BeforeInst->hasOneUse() &&
1222 FastIS->tryToFoldLoad(cast<LoadInst>(BeforeInst), Inst)) {
1223 // If we succeeded, don't re-select the load.
1224 BI = std::next(BasicBlock::const_iterator(BeforeInst));
1225 --NumFastIselRemaining;
1226 ++NumFastIselSuccess;
1232 if (EnableFastISelVerbose2)
1233 collectFailStats(Inst);
1236 // Then handle certain instructions as single-LLVM-Instruction blocks.
1237 if (isa<CallInst>(Inst)) {
1239 if (EnableFastISelVerbose || EnableFastISelAbort) {
1240 dbgs() << "FastISel missed call: ";
1243 if (EnableFastISelAbort > 2)
1244 // FastISel selector couldn't handle something and bailed.
1245 // For the purpose of debugging, just abort.
1246 report_fatal_error("FastISel didn't select the entire block");
1248 if (!Inst->getType()->isVoidTy() && !Inst->use_empty()) {
1249 unsigned &R = FuncInfo->ValueMap[Inst];
1251 R = FuncInfo->CreateRegs(Inst->getType());
1254 bool HadTailCall = false;
1255 MachineBasicBlock::iterator SavedInsertPt = FuncInfo->InsertPt;
1256 SelectBasicBlock(Inst, BI, HadTailCall);
1258 // If the call was emitted as a tail call, we're done with the block.
1259 // We also need to delete any previously emitted instructions.
1261 FastIS->removeDeadCode(SavedInsertPt, FuncInfo->MBB->end());
1266 // Recompute NumFastIselRemaining as Selection DAG instruction
1267 // selection may have handled the call, input args, etc.
1268 unsigned RemainingNow = std::distance(Begin, BI);
1269 NumFastIselFailures += NumFastIselRemaining - RemainingNow;
1270 NumFastIselRemaining = RemainingNow;
1274 bool ShouldAbort = EnableFastISelAbort;
1275 if (EnableFastISelVerbose || EnableFastISelAbort) {
1276 if (isa<TerminatorInst>(Inst)) {
1277 // Use a different message for terminator misses.
1278 dbgs() << "FastISel missed terminator: ";
1279 // Don't abort unless for terminator unless the level is really high
1280 ShouldAbort = (EnableFastISelAbort > 2);
1282 dbgs() << "FastISel miss: ";
1287 // FastISel selector couldn't handle something and bailed.
1288 // For the purpose of debugging, just abort.
1289 report_fatal_error("FastISel didn't select the entire block");
1291 NumFastIselFailures += NumFastIselRemaining;
1295 FastIS->recomputeInsertPt();
1297 // Lower any arguments needed in this block if this is the entry block.
1298 if (LLVMBB == &Fn.getEntryBlock()) {
1307 ++NumFastIselBlocks;
1310 // Run SelectionDAG instruction selection on the remainder of the block
1311 // not handled by FastISel. If FastISel is not run, this is the entire
1314 SelectBasicBlock(Begin, BI, HadTailCall);
1318 FuncInfo->PHINodesToUpdate.clear();
1322 SDB->clearDanglingDebugInfo();
1323 SDB->SPDescriptor.resetPerFunctionState();
1326 /// Given that the input MI is before a partial terminator sequence TSeq, return
1327 /// true if M + TSeq also a partial terminator sequence.
1329 /// A Terminator sequence is a sequence of MachineInstrs which at this point in
1330 /// lowering copy vregs into physical registers, which are then passed into
1331 /// terminator instructors so we can satisfy ABI constraints. A partial
1332 /// terminator sequence is an improper subset of a terminator sequence (i.e. it
1333 /// may be the whole terminator sequence).
1334 static bool MIIsInTerminatorSequence(const MachineInstr *MI) {
1335 // If we do not have a copy or an implicit def, we return true if and only if
1336 // MI is a debug value.
1337 if (!MI->isCopy() && !MI->isImplicitDef())
1338 // Sometimes DBG_VALUE MI sneak in between the copies from the vregs to the
1339 // physical registers if there is debug info associated with the terminator
1340 // of our mbb. We want to include said debug info in our terminator
1341 // sequence, so we return true in that case.
1342 return MI->isDebugValue();
1344 // We have left the terminator sequence if we are not doing one of the
1347 // 1. Copying a vreg into a physical register.
1348 // 2. Copying a vreg into a vreg.
1349 // 3. Defining a register via an implicit def.
1351 // OPI should always be a register definition...
1352 MachineInstr::const_mop_iterator OPI = MI->operands_begin();
1353 if (!OPI->isReg() || !OPI->isDef())
1356 // Defining any register via an implicit def is always ok.
1357 if (MI->isImplicitDef())
1360 // Grab the copy source...
1361 MachineInstr::const_mop_iterator OPI2 = OPI;
1363 assert(OPI2 != MI->operands_end()
1364 && "Should have a copy implying we should have 2 arguments.");
1366 // Make sure that the copy dest is not a vreg when the copy source is a
1367 // physical register.
1368 if (!OPI2->isReg() ||
1369 (!TargetRegisterInfo::isPhysicalRegister(OPI->getReg()) &&
1370 TargetRegisterInfo::isPhysicalRegister(OPI2->getReg())))
1376 /// Find the split point at which to splice the end of BB into its success stack
1377 /// protector check machine basic block.
1379 /// On many platforms, due to ABI constraints, terminators, even before register
1380 /// allocation, use physical registers. This creates an issue for us since
1381 /// physical registers at this point can not travel across basic
1382 /// blocks. Luckily, selectiondag always moves physical registers into vregs
1383 /// when they enter functions and moves them through a sequence of copies back
1384 /// into the physical registers right before the terminator creating a
1385 /// ``Terminator Sequence''. This function is searching for the beginning of the
1386 /// terminator sequence so that we can ensure that we splice off not just the
1387 /// terminator, but additionally the copies that move the vregs into the
1388 /// physical registers.
1389 static MachineBasicBlock::iterator
1390 FindSplitPointForStackProtector(MachineBasicBlock *BB, DebugLoc DL) {
1391 MachineBasicBlock::iterator SplitPoint = BB->getFirstTerminator();
1393 if (SplitPoint == BB->begin())
1396 MachineBasicBlock::iterator Start = BB->begin();
1397 MachineBasicBlock::iterator Previous = SplitPoint;
1400 while (MIIsInTerminatorSequence(Previous)) {
1401 SplitPoint = Previous;
1402 if (Previous == Start)
1411 SelectionDAGISel::FinishBasicBlock() {
1413 DEBUG(dbgs() << "Total amount of phi nodes to update: "
1414 << FuncInfo->PHINodesToUpdate.size() << "\n";
1415 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i)
1416 dbgs() << "Node " << i << " : ("
1417 << FuncInfo->PHINodesToUpdate[i].first
1418 << ", " << FuncInfo->PHINodesToUpdate[i].second << ")\n");
1420 // Next, now that we know what the last MBB the LLVM BB expanded is, update
1421 // PHI nodes in successors.
1422 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i) {
1423 MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[i].first);
1424 assert(PHI->isPHI() &&
1425 "This is not a machine PHI node that we are updating!");
1426 if (!FuncInfo->MBB->isSuccessor(PHI->getParent()))
1428 PHI.addReg(FuncInfo->PHINodesToUpdate[i].second).addMBB(FuncInfo->MBB);
1431 // Handle stack protector.
1432 if (SDB->SPDescriptor.shouldEmitStackProtector()) {
1433 MachineBasicBlock *ParentMBB = SDB->SPDescriptor.getParentMBB();
1434 MachineBasicBlock *SuccessMBB = SDB->SPDescriptor.getSuccessMBB();
1436 // Find the split point to split the parent mbb. At the same time copy all
1437 // physical registers used in the tail of parent mbb into virtual registers
1438 // before the split point and back into physical registers after the split
1439 // point. This prevents us needing to deal with Live-ins and many other
1440 // register allocation issues caused by us splitting the parent mbb. The
1441 // register allocator will clean up said virtual copies later on.
1442 MachineBasicBlock::iterator SplitPoint =
1443 FindSplitPointForStackProtector(ParentMBB, SDB->getCurDebugLoc());
1445 // Splice the terminator of ParentMBB into SuccessMBB.
1446 SuccessMBB->splice(SuccessMBB->end(), ParentMBB,
1450 // Add compare/jump on neq/jump to the parent BB.
1451 FuncInfo->MBB = ParentMBB;
1452 FuncInfo->InsertPt = ParentMBB->end();
1453 SDB->visitSPDescriptorParent(SDB->SPDescriptor, ParentMBB);
1454 CurDAG->setRoot(SDB->getRoot());
1456 CodeGenAndEmitDAG();
1458 // CodeGen Failure MBB if we have not codegened it yet.
1459 MachineBasicBlock *FailureMBB = SDB->SPDescriptor.getFailureMBB();
1460 if (!FailureMBB->size()) {
1461 FuncInfo->MBB = FailureMBB;
1462 FuncInfo->InsertPt = FailureMBB->end();
1463 SDB->visitSPDescriptorFailure(SDB->SPDescriptor);
1464 CurDAG->setRoot(SDB->getRoot());
1466 CodeGenAndEmitDAG();
1469 // Clear the Per-BB State.
1470 SDB->SPDescriptor.resetPerBBState();
1473 for (unsigned i = 0, e = SDB->BitTestCases.size(); i != e; ++i) {
1474 // Lower header first, if it wasn't already lowered
1475 if (!SDB->BitTestCases[i].Emitted) {
1476 // Set the current basic block to the mbb we wish to insert the code into
1477 FuncInfo->MBB = SDB->BitTestCases[i].Parent;
1478 FuncInfo->InsertPt = FuncInfo->MBB->end();
1480 SDB->visitBitTestHeader(SDB->BitTestCases[i], FuncInfo->MBB);
1481 CurDAG->setRoot(SDB->getRoot());
1483 CodeGenAndEmitDAG();
1486 uint32_t UnhandledWeight = 0;
1487 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j)
1488 UnhandledWeight += SDB->BitTestCases[i].Cases[j].ExtraWeight;
1490 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j) {
1491 UnhandledWeight -= SDB->BitTestCases[i].Cases[j].ExtraWeight;
1492 // Set the current basic block to the mbb we wish to insert the code into
1493 FuncInfo->MBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1494 FuncInfo->InsertPt = FuncInfo->MBB->end();
1497 SDB->visitBitTestCase(SDB->BitTestCases[i],
1498 SDB->BitTestCases[i].Cases[j+1].ThisBB,
1500 SDB->BitTestCases[i].Reg,
1501 SDB->BitTestCases[i].Cases[j],
1504 SDB->visitBitTestCase(SDB->BitTestCases[i],
1505 SDB->BitTestCases[i].Default,
1507 SDB->BitTestCases[i].Reg,
1508 SDB->BitTestCases[i].Cases[j],
1512 CurDAG->setRoot(SDB->getRoot());
1514 CodeGenAndEmitDAG();
1518 for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1520 MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[pi].first);
1521 MachineBasicBlock *PHIBB = PHI->getParent();
1522 assert(PHI->isPHI() &&
1523 "This is not a machine PHI node that we are updating!");
1524 // This is "default" BB. We have two jumps to it. From "header" BB and
1525 // from last "case" BB.
1526 if (PHIBB == SDB->BitTestCases[i].Default)
1527 PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second)
1528 .addMBB(SDB->BitTestCases[i].Parent)
1529 .addReg(FuncInfo->PHINodesToUpdate[pi].second)
1530 .addMBB(SDB->BitTestCases[i].Cases.back().ThisBB);
1531 // One of "cases" BB.
1532 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size();
1534 MachineBasicBlock* cBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1535 if (cBB->isSuccessor(PHIBB))
1536 PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second).addMBB(cBB);
1540 SDB->BitTestCases.clear();
1542 // If the JumpTable record is filled in, then we need to emit a jump table.
1543 // Updating the PHI nodes is tricky in this case, since we need to determine
1544 // whether the PHI is a successor of the range check MBB or the jump table MBB
1545 for (unsigned i = 0, e = SDB->JTCases.size(); i != e; ++i) {
1546 // Lower header first, if it wasn't already lowered
1547 if (!SDB->JTCases[i].first.Emitted) {
1548 // Set the current basic block to the mbb we wish to insert the code into
1549 FuncInfo->MBB = SDB->JTCases[i].first.HeaderBB;
1550 FuncInfo->InsertPt = FuncInfo->MBB->end();
1552 SDB->visitJumpTableHeader(SDB->JTCases[i].second, SDB->JTCases[i].first,
1554 CurDAG->setRoot(SDB->getRoot());
1556 CodeGenAndEmitDAG();
1559 // Set the current basic block to the mbb we wish to insert the code into
1560 FuncInfo->MBB = SDB->JTCases[i].second.MBB;
1561 FuncInfo->InsertPt = FuncInfo->MBB->end();
1563 SDB->visitJumpTable(SDB->JTCases[i].second);
1564 CurDAG->setRoot(SDB->getRoot());
1566 CodeGenAndEmitDAG();
1569 for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1571 MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[pi].first);
1572 MachineBasicBlock *PHIBB = PHI->getParent();
1573 assert(PHI->isPHI() &&
1574 "This is not a machine PHI node that we are updating!");
1575 // "default" BB. We can go there only from header BB.
1576 if (PHIBB == SDB->JTCases[i].second.Default)
1577 PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second)
1578 .addMBB(SDB->JTCases[i].first.HeaderBB);
1579 // JT BB. Just iterate over successors here
1580 if (FuncInfo->MBB->isSuccessor(PHIBB))
1581 PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second).addMBB(FuncInfo->MBB);
1584 SDB->JTCases.clear();
1586 // If we generated any switch lowering information, build and codegen any
1587 // additional DAGs necessary.
1588 for (unsigned i = 0, e = SDB->SwitchCases.size(); i != e; ++i) {
1589 // Set the current basic block to the mbb we wish to insert the code into
1590 FuncInfo->MBB = SDB->SwitchCases[i].ThisBB;
1591 FuncInfo->InsertPt = FuncInfo->MBB->end();
1593 // Determine the unique successors.
1594 SmallVector<MachineBasicBlock *, 2> Succs;
1595 Succs.push_back(SDB->SwitchCases[i].TrueBB);
1596 if (SDB->SwitchCases[i].TrueBB != SDB->SwitchCases[i].FalseBB)
1597 Succs.push_back(SDB->SwitchCases[i].FalseBB);
1599 // Emit the code. Note that this could result in FuncInfo->MBB being split.
1600 SDB->visitSwitchCase(SDB->SwitchCases[i], FuncInfo->MBB);
1601 CurDAG->setRoot(SDB->getRoot());
1603 CodeGenAndEmitDAG();
1605 // Remember the last block, now that any splitting is done, for use in
1606 // populating PHI nodes in successors.
1607 MachineBasicBlock *ThisBB = FuncInfo->MBB;
1609 // Handle any PHI nodes in successors of this chunk, as if we were coming
1610 // from the original BB before switch expansion. Note that PHI nodes can
1611 // occur multiple times in PHINodesToUpdate. We have to be very careful to
1612 // handle them the right number of times.
1613 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1614 FuncInfo->MBB = Succs[i];
1615 FuncInfo->InsertPt = FuncInfo->MBB->end();
1616 // FuncInfo->MBB may have been removed from the CFG if a branch was
1618 if (ThisBB->isSuccessor(FuncInfo->MBB)) {
1619 for (MachineBasicBlock::iterator
1620 MBBI = FuncInfo->MBB->begin(), MBBE = FuncInfo->MBB->end();
1621 MBBI != MBBE && MBBI->isPHI(); ++MBBI) {
1622 MachineInstrBuilder PHI(*MF, MBBI);
1623 // This value for this PHI node is recorded in PHINodesToUpdate.
1624 for (unsigned pn = 0; ; ++pn) {
1625 assert(pn != FuncInfo->PHINodesToUpdate.size() &&
1626 "Didn't find PHI entry!");
1627 if (FuncInfo->PHINodesToUpdate[pn].first == PHI) {
1628 PHI.addReg(FuncInfo->PHINodesToUpdate[pn].second).addMBB(ThisBB);
1636 SDB->SwitchCases.clear();
1640 /// Create the scheduler. If a specific scheduler was specified
1641 /// via the SchedulerRegistry, use it, otherwise select the
1642 /// one preferred by the target.
1644 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
1645 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1649 RegisterScheduler::setDefault(Ctor);
1652 return Ctor(this, OptLevel);
1655 //===----------------------------------------------------------------------===//
1656 // Helper functions used by the generated instruction selector.
1657 //===----------------------------------------------------------------------===//
1658 // Calls to these methods are generated by tblgen.
1660 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1661 /// the dag combiner simplified the 255, we still want to match. RHS is the
1662 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1663 /// specified in the .td file (e.g. 255).
1664 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1665 int64_t DesiredMaskS) const {
1666 const APInt &ActualMask = RHS->getAPIntValue();
1667 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1669 // If the actual mask exactly matches, success!
1670 if (ActualMask == DesiredMask)
1673 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1674 if (ActualMask.intersects(~DesiredMask))
1677 // Otherwise, the DAG Combiner may have proven that the value coming in is
1678 // either already zero or is not demanded. Check for known zero input bits.
1679 APInt NeededMask = DesiredMask & ~ActualMask;
1680 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1683 // TODO: check to see if missing bits are just not demanded.
1685 // Otherwise, this pattern doesn't match.
1689 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1690 /// the dag combiner simplified the 255, we still want to match. RHS is the
1691 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1692 /// specified in the .td file (e.g. 255).
1693 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1694 int64_t DesiredMaskS) const {
1695 const APInt &ActualMask = RHS->getAPIntValue();
1696 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1698 // If the actual mask exactly matches, success!
1699 if (ActualMask == DesiredMask)
1702 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1703 if (ActualMask.intersects(~DesiredMask))
1706 // Otherwise, the DAG Combiner may have proven that the value coming in is
1707 // either already zero or is not demanded. Check for known zero input bits.
1708 APInt NeededMask = DesiredMask & ~ActualMask;
1710 APInt KnownZero, KnownOne;
1711 CurDAG->computeKnownBits(LHS, KnownZero, KnownOne);
1713 // If all the missing bits in the or are already known to be set, match!
1714 if ((NeededMask & KnownOne) == NeededMask)
1717 // TODO: check to see if missing bits are just not demanded.
1719 // Otherwise, this pattern doesn't match.
1723 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1724 /// by tblgen. Others should not call it.
1725 void SelectionDAGISel::
1726 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops, SDLoc DL) {
1727 std::vector<SDValue> InOps;
1728 std::swap(InOps, Ops);
1730 Ops.push_back(InOps[InlineAsm::Op_InputChain]); // 0
1731 Ops.push_back(InOps[InlineAsm::Op_AsmString]); // 1
1732 Ops.push_back(InOps[InlineAsm::Op_MDNode]); // 2, !srcloc
1733 Ops.push_back(InOps[InlineAsm::Op_ExtraInfo]); // 3 (SideEffect, AlignStack)
1735 unsigned i = InlineAsm::Op_FirstOperand, e = InOps.size();
1736 if (InOps[e-1].getValueType() == MVT::Glue)
1737 --e; // Don't process a glue operand if it is here.
1740 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1741 if (!InlineAsm::isMemKind(Flags)) {
1742 // Just skip over this operand, copying the operands verbatim.
1743 Ops.insert(Ops.end(), InOps.begin()+i,
1744 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
1745 i += InlineAsm::getNumOperandRegisters(Flags) + 1;
1747 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 &&
1748 "Memory operand with multiple values?");
1750 unsigned TiedToOperand;
1751 if (InlineAsm::isUseOperandTiedToDef(Flags, TiedToOperand)) {
1752 // We need the constraint ID from the operand this is tied to.
1753 unsigned CurOp = InlineAsm::Op_FirstOperand;
1754 Flags = cast<ConstantSDNode>(InOps[CurOp])->getZExtValue();
1755 for (; TiedToOperand; --TiedToOperand) {
1756 CurOp += InlineAsm::getNumOperandRegisters(Flags)+1;
1757 Flags = cast<ConstantSDNode>(InOps[CurOp])->getZExtValue();
1761 // Otherwise, this is a memory operand. Ask the target to select it.
1762 std::vector<SDValue> SelOps;
1763 if (SelectInlineAsmMemoryOperand(InOps[i+1],
1764 InlineAsm::getMemoryConstraintID(Flags),
1766 report_fatal_error("Could not match memory address. Inline asm"
1769 // Add this to the output node.
1771 InlineAsm::getFlagWord(InlineAsm::Kind_Mem, SelOps.size());
1772 Ops.push_back(CurDAG->getTargetConstant(NewFlags, DL, MVT::i32));
1773 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1778 // Add the glue input back if present.
1779 if (e != InOps.size())
1780 Ops.push_back(InOps.back());
1783 /// findGlueUse - Return use of MVT::Glue value produced by the specified
1786 static SDNode *findGlueUse(SDNode *N) {
1787 unsigned FlagResNo = N->getNumValues()-1;
1788 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1789 SDUse &Use = I.getUse();
1790 if (Use.getResNo() == FlagResNo)
1791 return Use.getUser();
1796 /// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
1797 /// This function recursively traverses up the operand chain, ignoring
1799 static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
1800 SDNode *Root, SmallPtrSetImpl<SDNode*> &Visited,
1801 bool IgnoreChains) {
1802 // The NodeID's are given uniques ID's where a node ID is guaranteed to be
1803 // greater than all of its (recursive) operands. If we scan to a point where
1804 // 'use' is smaller than the node we're scanning for, then we know we will
1807 // The Use may be -1 (unassigned) if it is a newly allocated node. This can
1808 // happen because we scan down to newly selected nodes in the case of glue
1810 if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1))
1813 // Don't revisit nodes if we already scanned it and didn't fail, we know we
1814 // won't fail if we scan it again.
1815 if (!Visited.insert(Use).second)
1818 for (const SDValue &Op : Use->op_values()) {
1819 // Ignore chain uses, they are validated by HandleMergeInputChains.
1820 if (Op.getValueType() == MVT::Other && IgnoreChains)
1823 SDNode *N = Op.getNode();
1825 if (Use == ImmedUse || Use == Root)
1826 continue; // We are not looking for immediate use.
1831 // Traverse up the operand chain.
1832 if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
1838 /// IsProfitableToFold - Returns true if it's profitable to fold the specific
1839 /// operand node N of U during instruction selection that starts at Root.
1840 bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
1841 SDNode *Root) const {
1842 if (OptLevel == CodeGenOpt::None) return false;
1843 return N.hasOneUse();
1846 /// IsLegalToFold - Returns true if the specific operand node N of
1847 /// U can be folded during instruction selection that starts at Root.
1848 bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
1849 CodeGenOpt::Level OptLevel,
1850 bool IgnoreChains) {
1851 if (OptLevel == CodeGenOpt::None) return false;
1853 // If Root use can somehow reach N through a path that that doesn't contain
1854 // U then folding N would create a cycle. e.g. In the following
1855 // diagram, Root can reach N through X. If N is folded into into Root, then
1856 // X is both a predecessor and a successor of U.
1867 // * indicates nodes to be folded together.
1869 // If Root produces glue, then it gets (even more) interesting. Since it
1870 // will be "glued" together with its glue use in the scheduler, we need to
1871 // check if it might reach N.
1890 // If GU (glue use) indirectly reaches N (the load), and Root folds N
1891 // (call it Fold), then X is a predecessor of GU and a successor of
1892 // Fold. But since Fold and GU are glued together, this will create
1893 // a cycle in the scheduling graph.
1895 // If the node has glue, walk down the graph to the "lowest" node in the
1897 EVT VT = Root->getValueType(Root->getNumValues()-1);
1898 while (VT == MVT::Glue) {
1899 SDNode *GU = findGlueUse(Root);
1903 VT = Root->getValueType(Root->getNumValues()-1);
1905 // If our query node has a glue result with a use, we've walked up it. If
1906 // the user (which has already been selected) has a chain or indirectly uses
1907 // the chain, our WalkChainUsers predicate will not consider it. Because of
1908 // this, we cannot ignore chains in this predicate.
1909 IgnoreChains = false;
1913 SmallPtrSet<SDNode*, 16> Visited;
1914 return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains);
1917 SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
1920 std::vector<SDValue> Ops(N->op_begin(), N->op_end());
1921 SelectInlineAsmMemoryOperands(Ops, DL);
1923 const EVT VTs[] = {MVT::Other, MVT::Glue};
1924 SDValue New = CurDAG->getNode(ISD::INLINEASM, DL, VTs, Ops);
1926 return New.getNode();
1930 *SelectionDAGISel::Select_READ_REGISTER(SDNode *Op) {
1932 MDNodeSDNode *MD = dyn_cast<MDNodeSDNode>(Op->getOperand(1));
1933 const MDString *RegStr = dyn_cast<MDString>(MD->getMD()->getOperand(0));
1935 TLI->getRegisterByName(RegStr->getString().data(), Op->getValueType(0),
1937 SDValue New = CurDAG->getCopyFromReg(
1938 Op->getOperand(0), dl, Reg, Op->getValueType(0));
1940 return New.getNode();
1944 *SelectionDAGISel::Select_WRITE_REGISTER(SDNode *Op) {
1946 MDNodeSDNode *MD = dyn_cast<MDNodeSDNode>(Op->getOperand(1));
1947 const MDString *RegStr = dyn_cast<MDString>(MD->getMD()->getOperand(0));
1948 unsigned Reg = TLI->getRegisterByName(RegStr->getString().data(),
1949 Op->getOperand(2).getValueType(),
1951 SDValue New = CurDAG->getCopyToReg(
1952 Op->getOperand(0), dl, Reg, Op->getOperand(2));
1954 return New.getNode();
1959 SDNode *SelectionDAGISel::Select_UNDEF(SDNode *N) {
1960 return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0));
1963 /// GetVBR - decode a vbr encoding whose top bit is set.
1964 LLVM_ATTRIBUTE_ALWAYS_INLINE static uint64_t
1965 GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
1966 assert(Val >= 128 && "Not a VBR");
1967 Val &= 127; // Remove first vbr bit.
1972 NextBits = MatcherTable[Idx++];
1973 Val |= (NextBits&127) << Shift;
1975 } while (NextBits & 128);
1981 /// UpdateChainsAndGlue - When a match is complete, this method updates uses of
1982 /// interior glue and chain results to use the new glue and chain results.
1983 void SelectionDAGISel::
1984 UpdateChainsAndGlue(SDNode *NodeToMatch, SDValue InputChain,
1985 const SmallVectorImpl<SDNode*> &ChainNodesMatched,
1987 const SmallVectorImpl<SDNode*> &GlueResultNodesMatched,
1988 bool isMorphNodeTo) {
1989 SmallVector<SDNode*, 4> NowDeadNodes;
1991 // Now that all the normal results are replaced, we replace the chain and
1992 // glue results if present.
1993 if (!ChainNodesMatched.empty()) {
1994 assert(InputChain.getNode() &&
1995 "Matched input chains but didn't produce a chain");
1996 // Loop over all of the nodes we matched that produced a chain result.
1997 // Replace all the chain results with the final chain we ended up with.
1998 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1999 SDNode *ChainNode = ChainNodesMatched[i];
2001 // If this node was already deleted, don't look at it.
2002 if (ChainNode->getOpcode() == ISD::DELETED_NODE)
2005 // Don't replace the results of the root node if we're doing a
2007 if (ChainNode == NodeToMatch && isMorphNodeTo)
2010 SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1);
2011 if (ChainVal.getValueType() == MVT::Glue)
2012 ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
2013 assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
2014 CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain);
2016 // If the node became dead and we haven't already seen it, delete it.
2017 if (ChainNode->use_empty() &&
2018 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), ChainNode))
2019 NowDeadNodes.push_back(ChainNode);
2023 // If the result produces glue, update any glue results in the matched
2024 // pattern with the glue result.
2025 if (InputGlue.getNode()) {
2026 // Handle any interior nodes explicitly marked.
2027 for (unsigned i = 0, e = GlueResultNodesMatched.size(); i != e; ++i) {
2028 SDNode *FRN = GlueResultNodesMatched[i];
2030 // If this node was already deleted, don't look at it.
2031 if (FRN->getOpcode() == ISD::DELETED_NODE)
2034 assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Glue &&
2035 "Doesn't have a glue result");
2036 CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1),
2039 // If the node became dead and we haven't already seen it, delete it.
2040 if (FRN->use_empty() &&
2041 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), FRN))
2042 NowDeadNodes.push_back(FRN);
2046 if (!NowDeadNodes.empty())
2047 CurDAG->RemoveDeadNodes(NowDeadNodes);
2049 DEBUG(dbgs() << "ISEL: Match complete!\n");
2055 CR_LeadsToInteriorNode
2058 /// WalkChainUsers - Walk down the users of the specified chained node that is
2059 /// part of the pattern we're matching, looking at all of the users we find.
2060 /// This determines whether something is an interior node, whether we have a
2061 /// non-pattern node in between two pattern nodes (which prevent folding because
2062 /// it would induce a cycle) and whether we have a TokenFactor node sandwiched
2063 /// between pattern nodes (in which case the TF becomes part of the pattern).
2065 /// The walk we do here is guaranteed to be small because we quickly get down to
2066 /// already selected nodes "below" us.
2068 WalkChainUsers(const SDNode *ChainedNode,
2069 SmallVectorImpl<SDNode*> &ChainedNodesInPattern,
2070 SmallVectorImpl<SDNode*> &InteriorChainedNodes) {
2071 ChainResult Result = CR_Simple;
2073 for (SDNode::use_iterator UI = ChainedNode->use_begin(),
2074 E = ChainedNode->use_end(); UI != E; ++UI) {
2075 // Make sure the use is of the chain, not some other value we produce.
2076 if (UI.getUse().getValueType() != MVT::Other) continue;
2080 if (User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
2083 // If we see an already-selected machine node, then we've gone beyond the
2084 // pattern that we're selecting down into the already selected chunk of the
2086 unsigned UserOpcode = User->getOpcode();
2087 if (User->isMachineOpcode() ||
2088 UserOpcode == ISD::CopyToReg ||
2089 UserOpcode == ISD::CopyFromReg ||
2090 UserOpcode == ISD::INLINEASM ||
2091 UserOpcode == ISD::EH_LABEL ||
2092 UserOpcode == ISD::LIFETIME_START ||
2093 UserOpcode == ISD::LIFETIME_END) {
2094 // If their node ID got reset to -1 then they've already been selected.
2095 // Treat them like a MachineOpcode.
2096 if (User->getNodeId() == -1)
2100 // If we have a TokenFactor, we handle it specially.
2101 if (User->getOpcode() != ISD::TokenFactor) {
2102 // If the node isn't a token factor and isn't part of our pattern, then it
2103 // must be a random chained node in between two nodes we're selecting.
2104 // This happens when we have something like:
2109 // Because we structurally match the load/store as a read/modify/write,
2110 // but the call is chained between them. We cannot fold in this case
2111 // because it would induce a cycle in the graph.
2112 if (!std::count(ChainedNodesInPattern.begin(),
2113 ChainedNodesInPattern.end(), User))
2114 return CR_InducesCycle;
2116 // Otherwise we found a node that is part of our pattern. For example in:
2120 // This would happen when we're scanning down from the load and see the
2121 // store as a user. Record that there is a use of ChainedNode that is
2122 // part of the pattern and keep scanning uses.
2123 Result = CR_LeadsToInteriorNode;
2124 InteriorChainedNodes.push_back(User);
2128 // If we found a TokenFactor, there are two cases to consider: first if the
2129 // TokenFactor is just hanging "below" the pattern we're matching (i.e. no
2130 // uses of the TF are in our pattern) we just want to ignore it. Second,
2131 // the TokenFactor can be sandwiched in between two chained nodes, like so:
2137 // | \ DAG's like cheese
2140 // [TokenFactor] [Op]
2147 // In this case, the TokenFactor becomes part of our match and we rewrite it
2148 // as a new TokenFactor.
2150 // To distinguish these two cases, do a recursive walk down the uses.
2151 switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) {
2153 // If the uses of the TokenFactor are just already-selected nodes, ignore
2154 // it, it is "below" our pattern.
2156 case CR_InducesCycle:
2157 // If the uses of the TokenFactor lead to nodes that are not part of our
2158 // pattern that are not selected, folding would turn this into a cycle,
2160 return CR_InducesCycle;
2161 case CR_LeadsToInteriorNode:
2162 break; // Otherwise, keep processing.
2165 // Okay, we know we're in the interesting interior case. The TokenFactor
2166 // is now going to be considered part of the pattern so that we rewrite its
2167 // uses (it may have uses that are not part of the pattern) with the
2168 // ultimate chain result of the generated code. We will also add its chain
2169 // inputs as inputs to the ultimate TokenFactor we create.
2170 Result = CR_LeadsToInteriorNode;
2171 ChainedNodesInPattern.push_back(User);
2172 InteriorChainedNodes.push_back(User);
2179 /// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
2180 /// operation for when the pattern matched at least one node with a chains. The
2181 /// input vector contains a list of all of the chained nodes that we match. We
2182 /// must determine if this is a valid thing to cover (i.e. matching it won't
2183 /// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
2184 /// be used as the input node chain for the generated nodes.
2186 HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
2187 SelectionDAG *CurDAG) {
2188 // Walk all of the chained nodes we've matched, recursively scanning down the
2189 // users of the chain result. This adds any TokenFactor nodes that are caught
2190 // in between chained nodes to the chained and interior nodes list.
2191 SmallVector<SDNode*, 3> InteriorChainedNodes;
2192 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
2193 if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched,
2194 InteriorChainedNodes) == CR_InducesCycle)
2195 return SDValue(); // Would induce a cycle.
2198 // Okay, we have walked all the matched nodes and collected TokenFactor nodes
2199 // that we are interested in. Form our input TokenFactor node.
2200 SmallVector<SDValue, 3> InputChains;
2201 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
2202 // Add the input chain of this node to the InputChains list (which will be
2203 // the operands of the generated TokenFactor) if it's not an interior node.
2204 SDNode *N = ChainNodesMatched[i];
2205 if (N->getOpcode() != ISD::TokenFactor) {
2206 if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N))
2209 // Otherwise, add the input chain.
2210 SDValue InChain = ChainNodesMatched[i]->getOperand(0);
2211 assert(InChain.getValueType() == MVT::Other && "Not a chain");
2212 InputChains.push_back(InChain);
2216 // If we have a token factor, we want to add all inputs of the token factor
2217 // that are not part of the pattern we're matching.
2218 for (const SDValue &Op : N->op_values()) {
2219 if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(),
2221 InputChains.push_back(Op);
2225 if (InputChains.size() == 1)
2226 return InputChains[0];
2227 return CurDAG->getNode(ISD::TokenFactor, SDLoc(ChainNodesMatched[0]),
2228 MVT::Other, InputChains);
2231 /// MorphNode - Handle morphing a node in place for the selector.
2232 SDNode *SelectionDAGISel::
2233 MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList,
2234 ArrayRef<SDValue> Ops, unsigned EmitNodeInfo) {
2235 // It is possible we're using MorphNodeTo to replace a node with no
2236 // normal results with one that has a normal result (or we could be
2237 // adding a chain) and the input could have glue and chains as well.
2238 // In this case we need to shift the operands down.
2239 // FIXME: This is a horrible hack and broken in obscure cases, no worse
2240 // than the old isel though.
2241 int OldGlueResultNo = -1, OldChainResultNo = -1;
2243 unsigned NTMNumResults = Node->getNumValues();
2244 if (Node->getValueType(NTMNumResults-1) == MVT::Glue) {
2245 OldGlueResultNo = NTMNumResults-1;
2246 if (NTMNumResults != 1 &&
2247 Node->getValueType(NTMNumResults-2) == MVT::Other)
2248 OldChainResultNo = NTMNumResults-2;
2249 } else if (Node->getValueType(NTMNumResults-1) == MVT::Other)
2250 OldChainResultNo = NTMNumResults-1;
2252 // Call the underlying SelectionDAG routine to do the transmogrification. Note
2253 // that this deletes operands of the old node that become dead.
2254 SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops);
2256 // MorphNodeTo can operate in two ways: if an existing node with the
2257 // specified operands exists, it can just return it. Otherwise, it
2258 // updates the node in place to have the requested operands.
2260 // If we updated the node in place, reset the node ID. To the isel,
2261 // this should be just like a newly allocated machine node.
2265 unsigned ResNumResults = Res->getNumValues();
2266 // Move the glue if needed.
2267 if ((EmitNodeInfo & OPFL_GlueOutput) && OldGlueResultNo != -1 &&
2268 (unsigned)OldGlueResultNo != ResNumResults-1)
2269 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldGlueResultNo),
2270 SDValue(Res, ResNumResults-1));
2272 if ((EmitNodeInfo & OPFL_GlueOutput) != 0)
2275 // Move the chain reference if needed.
2276 if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 &&
2277 (unsigned)OldChainResultNo != ResNumResults-1)
2278 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo),
2279 SDValue(Res, ResNumResults-1));
2281 // Otherwise, no replacement happened because the node already exists. Replace
2282 // Uses of the old node with the new one.
2284 CurDAG->ReplaceAllUsesWith(Node, Res);
2289 /// CheckSame - Implements OP_CheckSame.
2290 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2291 CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2293 const SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes) {
2294 // Accept if it is exactly the same as a previously recorded node.
2295 unsigned RecNo = MatcherTable[MatcherIndex++];
2296 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2297 return N == RecordedNodes[RecNo].first;
2300 /// CheckChildSame - Implements OP_CheckChildXSame.
2301 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2302 CheckChildSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2304 const SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes,
2306 if (ChildNo >= N.getNumOperands())
2307 return false; // Match fails if out of range child #.
2308 return ::CheckSame(MatcherTable, MatcherIndex, N.getOperand(ChildNo),
2312 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
2313 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2314 CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2315 const SelectionDAGISel &SDISel) {
2316 return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]);
2319 /// CheckNodePredicate - Implements OP_CheckNodePredicate.
2320 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2321 CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2322 const SelectionDAGISel &SDISel, SDNode *N) {
2323 return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]);
2326 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2327 CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2329 uint16_t Opc = MatcherTable[MatcherIndex++];
2330 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2331 return N->getOpcode() == Opc;
2334 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2335 CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex, SDValue N,
2336 const TargetLowering *TLI, const DataLayout &DL) {
2337 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2338 if (N.getValueType() == VT) return true;
2340 // Handle the case when VT is iPTR.
2341 return VT == MVT::iPTR && N.getValueType() == TLI->getPointerTy(DL);
2344 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2345 CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2346 SDValue N, const TargetLowering *TLI, const DataLayout &DL,
2348 if (ChildNo >= N.getNumOperands())
2349 return false; // Match fails if out of range child #.
2350 return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI,
2354 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2355 CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2357 return cast<CondCodeSDNode>(N)->get() ==
2358 (ISD::CondCode)MatcherTable[MatcherIndex++];
2361 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2362 CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2363 SDValue N, const TargetLowering *TLI, const DataLayout &DL) {
2364 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2365 if (cast<VTSDNode>(N)->getVT() == VT)
2368 // Handle the case when VT is iPTR.
2369 return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI->getPointerTy(DL);
2372 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2373 CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2375 int64_t Val = MatcherTable[MatcherIndex++];
2377 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2379 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
2380 return C && C->getSExtValue() == Val;
2383 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2384 CheckChildInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2385 SDValue N, unsigned ChildNo) {
2386 if (ChildNo >= N.getNumOperands())
2387 return false; // Match fails if out of range child #.
2388 return ::CheckInteger(MatcherTable, MatcherIndex, N.getOperand(ChildNo));
2391 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2392 CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2393 SDValue N, const SelectionDAGISel &SDISel) {
2394 int64_t Val = MatcherTable[MatcherIndex++];
2396 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2398 if (N->getOpcode() != ISD::AND) return false;
2400 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
2401 return C && SDISel.CheckAndMask(N.getOperand(0), C, Val);
2404 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2405 CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2406 SDValue N, const SelectionDAGISel &SDISel) {
2407 int64_t Val = MatcherTable[MatcherIndex++];
2409 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2411 if (N->getOpcode() != ISD::OR) return false;
2413 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
2414 return C && SDISel.CheckOrMask(N.getOperand(0), C, Val);
2417 /// IsPredicateKnownToFail - If we know how and can do so without pushing a
2418 /// scope, evaluate the current node. If the current predicate is known to
2419 /// fail, set Result=true and return anything. If the current predicate is
2420 /// known to pass, set Result=false and return the MatcherIndex to continue
2421 /// with. If the current predicate is unknown, set Result=false and return the
2422 /// MatcherIndex to continue with.
2423 static unsigned IsPredicateKnownToFail(const unsigned char *Table,
2424 unsigned Index, SDValue N,
2426 const SelectionDAGISel &SDISel,
2427 SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes) {
2428 switch (Table[Index++]) {
2431 return Index-1; // Could not evaluate this predicate.
2432 case SelectionDAGISel::OPC_CheckSame:
2433 Result = !::CheckSame(Table, Index, N, RecordedNodes);
2435 case SelectionDAGISel::OPC_CheckChild0Same:
2436 case SelectionDAGISel::OPC_CheckChild1Same:
2437 case SelectionDAGISel::OPC_CheckChild2Same:
2438 case SelectionDAGISel::OPC_CheckChild3Same:
2439 Result = !::CheckChildSame(Table, Index, N, RecordedNodes,
2440 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Same);
2442 case SelectionDAGISel::OPC_CheckPatternPredicate:
2443 Result = !::CheckPatternPredicate(Table, Index, SDISel);
2445 case SelectionDAGISel::OPC_CheckPredicate:
2446 Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode());
2448 case SelectionDAGISel::OPC_CheckOpcode:
2449 Result = !::CheckOpcode(Table, Index, N.getNode());
2451 case SelectionDAGISel::OPC_CheckType:
2452 Result = !::CheckType(Table, Index, N, SDISel.TLI,
2453 SDISel.CurDAG->getDataLayout());
2455 case SelectionDAGISel::OPC_CheckChild0Type:
2456 case SelectionDAGISel::OPC_CheckChild1Type:
2457 case SelectionDAGISel::OPC_CheckChild2Type:
2458 case SelectionDAGISel::OPC_CheckChild3Type:
2459 case SelectionDAGISel::OPC_CheckChild4Type:
2460 case SelectionDAGISel::OPC_CheckChild5Type:
2461 case SelectionDAGISel::OPC_CheckChild6Type:
2462 case SelectionDAGISel::OPC_CheckChild7Type:
2463 Result = !::CheckChildType(
2464 Table, Index, N, SDISel.TLI, SDISel.CurDAG->getDataLayout(),
2465 Table[Index - 1] - SelectionDAGISel::OPC_CheckChild0Type);
2467 case SelectionDAGISel::OPC_CheckCondCode:
2468 Result = !::CheckCondCode(Table, Index, N);
2470 case SelectionDAGISel::OPC_CheckValueType:
2471 Result = !::CheckValueType(Table, Index, N, SDISel.TLI,
2472 SDISel.CurDAG->getDataLayout());
2474 case SelectionDAGISel::OPC_CheckInteger:
2475 Result = !::CheckInteger(Table, Index, N);
2477 case SelectionDAGISel::OPC_CheckChild0Integer:
2478 case SelectionDAGISel::OPC_CheckChild1Integer:
2479 case SelectionDAGISel::OPC_CheckChild2Integer:
2480 case SelectionDAGISel::OPC_CheckChild3Integer:
2481 case SelectionDAGISel::OPC_CheckChild4Integer:
2482 Result = !::CheckChildInteger(Table, Index, N,
2483 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Integer);
2485 case SelectionDAGISel::OPC_CheckAndImm:
2486 Result = !::CheckAndImm(Table, Index, N, SDISel);
2488 case SelectionDAGISel::OPC_CheckOrImm:
2489 Result = !::CheckOrImm(Table, Index, N, SDISel);
2497 /// FailIndex - If this match fails, this is the index to continue with.
2500 /// NodeStack - The node stack when the scope was formed.
2501 SmallVector<SDValue, 4> NodeStack;
2503 /// NumRecordedNodes - The number of recorded nodes when the scope was formed.
2504 unsigned NumRecordedNodes;
2506 /// NumMatchedMemRefs - The number of matched memref entries.
2507 unsigned NumMatchedMemRefs;
2509 /// InputChain/InputGlue - The current chain/glue
2510 SDValue InputChain, InputGlue;
2512 /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty.
2513 bool HasChainNodesMatched, HasGlueResultNodesMatched;
2516 /// \\brief A DAG update listener to keep the matching state
2517 /// (i.e. RecordedNodes and MatchScope) uptodate if the target is allowed to
2518 /// change the DAG while matching. X86 addressing mode matcher is an example
2520 class MatchStateUpdater : public SelectionDAG::DAGUpdateListener
2522 SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes;
2523 SmallVectorImpl<MatchScope> &MatchScopes;
2525 MatchStateUpdater(SelectionDAG &DAG,
2526 SmallVectorImpl<std::pair<SDValue, SDNode*> > &RN,
2527 SmallVectorImpl<MatchScope> &MS) :
2528 SelectionDAG::DAGUpdateListener(DAG),
2529 RecordedNodes(RN), MatchScopes(MS) { }
2531 void NodeDeleted(SDNode *N, SDNode *E) override {
2532 // Some early-returns here to avoid the search if we deleted the node or
2533 // if the update comes from MorphNodeTo (MorphNodeTo is the last thing we
2534 // do, so it's unnecessary to update matching state at that point).
2535 // Neither of these can occur currently because we only install this
2536 // update listener during matching a complex patterns.
2537 if (!E || E->isMachineOpcode())
2539 // Performing linear search here does not matter because we almost never
2540 // run this code. You'd have to have a CSE during complex pattern
2542 for (auto &I : RecordedNodes)
2543 if (I.first.getNode() == N)
2546 for (auto &I : MatchScopes)
2547 for (auto &J : I.NodeStack)
2548 if (J.getNode() == N)
2554 SDNode *SelectionDAGISel::
2555 SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
2556 unsigned TableSize) {
2557 // FIXME: Should these even be selected? Handle these cases in the caller?
2558 switch (NodeToMatch->getOpcode()) {
2561 case ISD::EntryToken: // These nodes remain the same.
2562 case ISD::BasicBlock:
2564 case ISD::RegisterMask:
2565 case ISD::HANDLENODE:
2566 case ISD::MDNODE_SDNODE:
2567 case ISD::TargetConstant:
2568 case ISD::TargetConstantFP:
2569 case ISD::TargetConstantPool:
2570 case ISD::TargetFrameIndex:
2571 case ISD::TargetExternalSymbol:
2573 case ISD::TargetBlockAddress:
2574 case ISD::TargetJumpTable:
2575 case ISD::TargetGlobalTLSAddress:
2576 case ISD::TargetGlobalAddress:
2577 case ISD::TokenFactor:
2578 case ISD::CopyFromReg:
2579 case ISD::CopyToReg:
2581 case ISD::LIFETIME_START:
2582 case ISD::LIFETIME_END:
2583 NodeToMatch->setNodeId(-1); // Mark selected.
2585 case ISD::AssertSext:
2586 case ISD::AssertZext:
2587 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0),
2588 NodeToMatch->getOperand(0));
2590 case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch);
2591 case ISD::READ_REGISTER: return Select_READ_REGISTER(NodeToMatch);
2592 case ISD::WRITE_REGISTER: return Select_WRITE_REGISTER(NodeToMatch);
2593 case ISD::UNDEF: return Select_UNDEF(NodeToMatch);
2596 assert(!NodeToMatch->isMachineOpcode() && "Node already selected!");
2598 // Set up the node stack with NodeToMatch as the only node on the stack.
2599 SmallVector<SDValue, 8> NodeStack;
2600 SDValue N = SDValue(NodeToMatch, 0);
2601 NodeStack.push_back(N);
2603 // MatchScopes - Scopes used when matching, if a match failure happens, this
2604 // indicates where to continue checking.
2605 SmallVector<MatchScope, 8> MatchScopes;
2607 // RecordedNodes - This is the set of nodes that have been recorded by the
2608 // state machine. The second value is the parent of the node, or null if the
2609 // root is recorded.
2610 SmallVector<std::pair<SDValue, SDNode*>, 8> RecordedNodes;
2612 // MatchedMemRefs - This is the set of MemRef's we've seen in the input
2614 SmallVector<MachineMemOperand*, 2> MatchedMemRefs;
2616 // These are the current input chain and glue for use when generating nodes.
2617 // Various Emit operations change these. For example, emitting a copytoreg
2618 // uses and updates these.
2619 SDValue InputChain, InputGlue;
2621 // ChainNodesMatched - If a pattern matches nodes that have input/output
2622 // chains, the OPC_EmitMergeInputChains operation is emitted which indicates
2623 // which ones they are. The result is captured into this list so that we can
2624 // update the chain results when the pattern is complete.
2625 SmallVector<SDNode*, 3> ChainNodesMatched;
2626 SmallVector<SDNode*, 3> GlueResultNodesMatched;
2628 DEBUG(dbgs() << "ISEL: Starting pattern match on root node: ";
2629 NodeToMatch->dump(CurDAG);
2632 // Determine where to start the interpreter. Normally we start at opcode #0,
2633 // but if the state machine starts with an OPC_SwitchOpcode, then we
2634 // accelerate the first lookup (which is guaranteed to be hot) with the
2635 // OpcodeOffset table.
2636 unsigned MatcherIndex = 0;
2638 if (!OpcodeOffset.empty()) {
2639 // Already computed the OpcodeOffset table, just index into it.
2640 if (N.getOpcode() < OpcodeOffset.size())
2641 MatcherIndex = OpcodeOffset[N.getOpcode()];
2642 DEBUG(dbgs() << " Initial Opcode index to " << MatcherIndex << "\n");
2644 } else if (MatcherTable[0] == OPC_SwitchOpcode) {
2645 // Otherwise, the table isn't computed, but the state machine does start
2646 // with an OPC_SwitchOpcode instruction. Populate the table now, since this
2647 // is the first time we're selecting an instruction.
2650 // Get the size of this case.
2651 unsigned CaseSize = MatcherTable[Idx++];
2653 CaseSize = GetVBR(CaseSize, MatcherTable, Idx);
2654 if (CaseSize == 0) break;
2656 // Get the opcode, add the index to the table.
2657 uint16_t Opc = MatcherTable[Idx++];
2658 Opc |= (unsigned short)MatcherTable[Idx++] << 8;
2659 if (Opc >= OpcodeOffset.size())
2660 OpcodeOffset.resize((Opc+1)*2);
2661 OpcodeOffset[Opc] = Idx;
2665 // Okay, do the lookup for the first opcode.
2666 if (N.getOpcode() < OpcodeOffset.size())
2667 MatcherIndex = OpcodeOffset[N.getOpcode()];
2671 assert(MatcherIndex < TableSize && "Invalid index");
2673 unsigned CurrentOpcodeIndex = MatcherIndex;
2675 BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++];
2678 // Okay, the semantics of this operation are that we should push a scope
2679 // then evaluate the first child. However, pushing a scope only to have
2680 // the first check fail (which then pops it) is inefficient. If we can
2681 // determine immediately that the first check (or first several) will
2682 // immediately fail, don't even bother pushing a scope for them.
2686 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2687 if (NumToSkip & 128)
2688 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2689 // Found the end of the scope with no match.
2690 if (NumToSkip == 0) {
2695 FailIndex = MatcherIndex+NumToSkip;
2697 unsigned MatcherIndexOfPredicate = MatcherIndex;
2698 (void)MatcherIndexOfPredicate; // silence warning.
2700 // If we can't evaluate this predicate without pushing a scope (e.g. if
2701 // it is a 'MoveParent') or if the predicate succeeds on this node, we
2702 // push the scope and evaluate the full predicate chain.
2704 MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
2705 Result, *this, RecordedNodes);
2709 DEBUG(dbgs() << " Skipped scope entry (due to false predicate) at "
2710 << "index " << MatcherIndexOfPredicate
2711 << ", continuing at " << FailIndex << "\n");
2712 ++NumDAGIselRetries;
2714 // Otherwise, we know that this case of the Scope is guaranteed to fail,
2715 // move to the next case.
2716 MatcherIndex = FailIndex;
2719 // If the whole scope failed to match, bail.
2720 if (FailIndex == 0) break;
2722 // Push a MatchScope which indicates where to go if the first child fails
2724 MatchScope NewEntry;
2725 NewEntry.FailIndex = FailIndex;
2726 NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
2727 NewEntry.NumRecordedNodes = RecordedNodes.size();
2728 NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
2729 NewEntry.InputChain = InputChain;
2730 NewEntry.InputGlue = InputGlue;
2731 NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty();
2732 NewEntry.HasGlueResultNodesMatched = !GlueResultNodesMatched.empty();
2733 MatchScopes.push_back(NewEntry);
2736 case OPC_RecordNode: {
2737 // Remember this node, it may end up being an operand in the pattern.
2738 SDNode *Parent = nullptr;
2739 if (NodeStack.size() > 1)
2740 Parent = NodeStack[NodeStack.size()-2].getNode();
2741 RecordedNodes.push_back(std::make_pair(N, Parent));
2745 case OPC_RecordChild0: case OPC_RecordChild1:
2746 case OPC_RecordChild2: case OPC_RecordChild3:
2747 case OPC_RecordChild4: case OPC_RecordChild5:
2748 case OPC_RecordChild6: case OPC_RecordChild7: {
2749 unsigned ChildNo = Opcode-OPC_RecordChild0;
2750 if (ChildNo >= N.getNumOperands())
2751 break; // Match fails if out of range child #.
2753 RecordedNodes.push_back(std::make_pair(N->getOperand(ChildNo),
2757 case OPC_RecordMemRef:
2758 MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand());
2761 case OPC_CaptureGlueInput:
2762 // If the current node has an input glue, capture it in InputGlue.
2763 if (N->getNumOperands() != 0 &&
2764 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue)
2765 InputGlue = N->getOperand(N->getNumOperands()-1);
2768 case OPC_MoveChild: {
2769 unsigned ChildNo = MatcherTable[MatcherIndex++];
2770 if (ChildNo >= N.getNumOperands())
2771 break; // Match fails if out of range child #.
2772 N = N.getOperand(ChildNo);
2773 NodeStack.push_back(N);
2777 case OPC_MoveParent:
2778 // Pop the current node off the NodeStack.
2779 NodeStack.pop_back();
2780 assert(!NodeStack.empty() && "Node stack imbalance!");
2781 N = NodeStack.back();
2785 if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
2788 case OPC_CheckChild0Same: case OPC_CheckChild1Same:
2789 case OPC_CheckChild2Same: case OPC_CheckChild3Same:
2790 if (!::CheckChildSame(MatcherTable, MatcherIndex, N, RecordedNodes,
2791 Opcode-OPC_CheckChild0Same))
2795 case OPC_CheckPatternPredicate:
2796 if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break;
2798 case OPC_CheckPredicate:
2799 if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this,
2803 case OPC_CheckComplexPat: {
2804 unsigned CPNum = MatcherTable[MatcherIndex++];
2805 unsigned RecNo = MatcherTable[MatcherIndex++];
2806 assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
2808 // If target can modify DAG during matching, keep the matching state
2810 std::unique_ptr<MatchStateUpdater> MSU;
2811 if (ComplexPatternFuncMutatesDAG())
2812 MSU.reset(new MatchStateUpdater(*CurDAG, RecordedNodes,
2815 if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo].second,
2816 RecordedNodes[RecNo].first, CPNum,
2821 case OPC_CheckOpcode:
2822 if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
2826 if (!::CheckType(MatcherTable, MatcherIndex, N, TLI,
2827 CurDAG->getDataLayout()))
2831 case OPC_SwitchOpcode: {
2832 unsigned CurNodeOpcode = N.getOpcode();
2833 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2836 // Get the size of this case.
2837 CaseSize = MatcherTable[MatcherIndex++];
2839 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2840 if (CaseSize == 0) break;
2842 uint16_t Opc = MatcherTable[MatcherIndex++];
2843 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2845 // If the opcode matches, then we will execute this case.
2846 if (CurNodeOpcode == Opc)
2849 // Otherwise, skip over this case.
2850 MatcherIndex += CaseSize;
2853 // If no cases matched, bail out.
2854 if (CaseSize == 0) break;
2856 // Otherwise, execute the case we found.
2857 DEBUG(dbgs() << " OpcodeSwitch from " << SwitchStart
2858 << " to " << MatcherIndex << "\n");
2862 case OPC_SwitchType: {
2863 MVT CurNodeVT = N.getSimpleValueType();
2864 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2867 // Get the size of this case.
2868 CaseSize = MatcherTable[MatcherIndex++];
2870 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2871 if (CaseSize == 0) break;
2873 MVT CaseVT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2874 if (CaseVT == MVT::iPTR)
2875 CaseVT = TLI->getPointerTy(CurDAG->getDataLayout());
2877 // If the VT matches, then we will execute this case.
2878 if (CurNodeVT == CaseVT)
2881 // Otherwise, skip over this case.
2882 MatcherIndex += CaseSize;
2885 // If no cases matched, bail out.
2886 if (CaseSize == 0) break;
2888 // Otherwise, execute the case we found.
2889 DEBUG(dbgs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
2890 << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
2893 case OPC_CheckChild0Type: case OPC_CheckChild1Type:
2894 case OPC_CheckChild2Type: case OPC_CheckChild3Type:
2895 case OPC_CheckChild4Type: case OPC_CheckChild5Type:
2896 case OPC_CheckChild6Type: case OPC_CheckChild7Type:
2897 if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI,
2898 CurDAG->getDataLayout(),
2899 Opcode - OPC_CheckChild0Type))
2902 case OPC_CheckCondCode:
2903 if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
2905 case OPC_CheckValueType:
2906 if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI,
2907 CurDAG->getDataLayout()))
2910 case OPC_CheckInteger:
2911 if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
2913 case OPC_CheckChild0Integer: case OPC_CheckChild1Integer:
2914 case OPC_CheckChild2Integer: case OPC_CheckChild3Integer:
2915 case OPC_CheckChild4Integer:
2916 if (!::CheckChildInteger(MatcherTable, MatcherIndex, N,
2917 Opcode-OPC_CheckChild0Integer)) break;
2919 case OPC_CheckAndImm:
2920 if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
2922 case OPC_CheckOrImm:
2923 if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
2926 case OPC_CheckFoldableChainNode: {
2927 assert(NodeStack.size() != 1 && "No parent node");
2928 // Verify that all intermediate nodes between the root and this one have
2930 bool HasMultipleUses = false;
2931 for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i)
2932 if (!NodeStack[i].hasOneUse()) {
2933 HasMultipleUses = true;
2936 if (HasMultipleUses) break;
2938 // Check to see that the target thinks this is profitable to fold and that
2939 // we can fold it without inducing cycles in the graph.
2940 if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2942 !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2943 NodeToMatch, OptLevel,
2944 true/*We validate our own chains*/))
2949 case OPC_EmitInteger: {
2950 MVT::SimpleValueType VT =
2951 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2952 int64_t Val = MatcherTable[MatcherIndex++];
2954 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2955 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2956 CurDAG->getTargetConstant(Val, SDLoc(NodeToMatch),
2960 case OPC_EmitRegister: {
2961 MVT::SimpleValueType VT =
2962 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2963 unsigned RegNo = MatcherTable[MatcherIndex++];
2964 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2965 CurDAG->getRegister(RegNo, VT), nullptr));
2968 case OPC_EmitRegister2: {
2969 // For targets w/ more than 256 register names, the register enum
2970 // values are stored in two bytes in the matcher table (just like
2972 MVT::SimpleValueType VT =
2973 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2974 unsigned RegNo = MatcherTable[MatcherIndex++];
2975 RegNo |= MatcherTable[MatcherIndex++] << 8;
2976 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2977 CurDAG->getRegister(RegNo, VT), nullptr));
2981 case OPC_EmitConvertToTarget: {
2982 // Convert from IMM/FPIMM to target version.
2983 unsigned RecNo = MatcherTable[MatcherIndex++];
2984 assert(RecNo < RecordedNodes.size() && "Invalid EmitConvertToTarget");
2985 SDValue Imm = RecordedNodes[RecNo].first;
2987 if (Imm->getOpcode() == ISD::Constant) {
2988 const ConstantInt *Val=cast<ConstantSDNode>(Imm)->getConstantIntValue();
2989 Imm = CurDAG->getConstant(*Val, SDLoc(NodeToMatch), Imm.getValueType(),
2991 } else if (Imm->getOpcode() == ISD::ConstantFP) {
2992 const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue();
2993 Imm = CurDAG->getConstantFP(*Val, SDLoc(NodeToMatch),
2994 Imm.getValueType(), true);
2997 RecordedNodes.push_back(std::make_pair(Imm, RecordedNodes[RecNo].second));
3001 case OPC_EmitMergeInputChains1_0: // OPC_EmitMergeInputChains, 1, 0
3002 case OPC_EmitMergeInputChains1_1: { // OPC_EmitMergeInputChains, 1, 1
3003 // These are space-optimized forms of OPC_EmitMergeInputChains.
3004 assert(!InputChain.getNode() &&
3005 "EmitMergeInputChains should be the first chain producing node");
3006 assert(ChainNodesMatched.empty() &&
3007 "Should only have one EmitMergeInputChains per match");
3009 // Read all of the chained nodes.
3010 unsigned RecNo = Opcode == OPC_EmitMergeInputChains1_1;
3011 assert(RecNo < RecordedNodes.size() && "Invalid EmitMergeInputChains");
3012 ChainNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
3014 // FIXME: What if other value results of the node have uses not matched
3016 if (ChainNodesMatched.back() != NodeToMatch &&
3017 !RecordedNodes[RecNo].first.hasOneUse()) {
3018 ChainNodesMatched.clear();
3022 // Merge the input chains if they are not intra-pattern references.
3023 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
3025 if (!InputChain.getNode())
3026 break; // Failed to merge.
3030 case OPC_EmitMergeInputChains: {
3031 assert(!InputChain.getNode() &&
3032 "EmitMergeInputChains should be the first chain producing node");
3033 // This node gets a list of nodes we matched in the input that have
3034 // chains. We want to token factor all of the input chains to these nodes
3035 // together. However, if any of the input chains is actually one of the
3036 // nodes matched in this pattern, then we have an intra-match reference.
3037 // Ignore these because the newly token factored chain should not refer to
3039 unsigned NumChains = MatcherTable[MatcherIndex++];
3040 assert(NumChains != 0 && "Can't TF zero chains");
3042 assert(ChainNodesMatched.empty() &&
3043 "Should only have one EmitMergeInputChains per match");
3045 // Read all of the chained nodes.
3046 for (unsigned i = 0; i != NumChains; ++i) {
3047 unsigned RecNo = MatcherTable[MatcherIndex++];
3048 assert(RecNo < RecordedNodes.size() && "Invalid EmitMergeInputChains");
3049 ChainNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
3051 // FIXME: What if other value results of the node have uses not matched
3053 if (ChainNodesMatched.back() != NodeToMatch &&
3054 !RecordedNodes[RecNo].first.hasOneUse()) {
3055 ChainNodesMatched.clear();
3060 // If the inner loop broke out, the match fails.
3061 if (ChainNodesMatched.empty())
3064 // Merge the input chains if they are not intra-pattern references.
3065 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
3067 if (!InputChain.getNode())
3068 break; // Failed to merge.
3073 case OPC_EmitCopyToReg: {
3074 unsigned RecNo = MatcherTable[MatcherIndex++];
3075 assert(RecNo < RecordedNodes.size() && "Invalid EmitCopyToReg");
3076 unsigned DestPhysReg = MatcherTable[MatcherIndex++];
3078 if (!InputChain.getNode())
3079 InputChain = CurDAG->getEntryNode();
3081 InputChain = CurDAG->getCopyToReg(InputChain, SDLoc(NodeToMatch),
3082 DestPhysReg, RecordedNodes[RecNo].first,
3085 InputGlue = InputChain.getValue(1);
3089 case OPC_EmitNodeXForm: {
3090 unsigned XFormNo = MatcherTable[MatcherIndex++];
3091 unsigned RecNo = MatcherTable[MatcherIndex++];
3092 assert(RecNo < RecordedNodes.size() && "Invalid EmitNodeXForm");
3093 SDValue Res = RunSDNodeXForm(RecordedNodes[RecNo].first, XFormNo);
3094 RecordedNodes.push_back(std::pair<SDValue,SDNode*>(Res, nullptr));
3099 case OPC_MorphNodeTo: {
3100 uint16_t TargetOpc = MatcherTable[MatcherIndex++];
3101 TargetOpc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
3102 unsigned EmitNodeInfo = MatcherTable[MatcherIndex++];
3103 // Get the result VT list.
3104 unsigned NumVTs = MatcherTable[MatcherIndex++];
3105 SmallVector<EVT, 4> VTs;
3106 for (unsigned i = 0; i != NumVTs; ++i) {
3107 MVT::SimpleValueType VT =
3108 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
3109 if (VT == MVT::iPTR)
3110 VT = TLI->getPointerTy(CurDAG->getDataLayout()).SimpleTy;
3114 if (EmitNodeInfo & OPFL_Chain)
3115 VTs.push_back(MVT::Other);
3116 if (EmitNodeInfo & OPFL_GlueOutput)
3117 VTs.push_back(MVT::Glue);
3119 // This is hot code, so optimize the two most common cases of 1 and 2
3122 if (VTs.size() == 1)
3123 VTList = CurDAG->getVTList(VTs[0]);
3124 else if (VTs.size() == 2)
3125 VTList = CurDAG->getVTList(VTs[0], VTs[1]);
3127 VTList = CurDAG->getVTList(VTs);
3129 // Get the operand list.
3130 unsigned NumOps = MatcherTable[MatcherIndex++];
3131 SmallVector<SDValue, 8> Ops;
3132 for (unsigned i = 0; i != NumOps; ++i) {
3133 unsigned RecNo = MatcherTable[MatcherIndex++];
3135 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
3137 assert(RecNo < RecordedNodes.size() && "Invalid EmitNode");
3138 Ops.push_back(RecordedNodes[RecNo].first);
3141 // If there are variadic operands to add, handle them now.
3142 if (EmitNodeInfo & OPFL_VariadicInfo) {
3143 // Determine the start index to copy from.
3144 unsigned FirstOpToCopy = getNumFixedFromVariadicInfo(EmitNodeInfo);
3145 FirstOpToCopy += (EmitNodeInfo & OPFL_Chain) ? 1 : 0;
3146 assert(NodeToMatch->getNumOperands() >= FirstOpToCopy &&
3147 "Invalid variadic node");
3148 // Copy all of the variadic operands, not including a potential glue
3150 for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands();
3152 SDValue V = NodeToMatch->getOperand(i);
3153 if (V.getValueType() == MVT::Glue) break;
3158 // If this has chain/glue inputs, add them.
3159 if (EmitNodeInfo & OPFL_Chain)
3160 Ops.push_back(InputChain);
3161 if ((EmitNodeInfo & OPFL_GlueInput) && InputGlue.getNode() != nullptr)
3162 Ops.push_back(InputGlue);
3165 SDNode *Res = nullptr;
3166 if (Opcode != OPC_MorphNodeTo) {
3167 // If this is a normal EmitNode command, just create the new node and
3168 // add the results to the RecordedNodes list.
3169 Res = CurDAG->getMachineNode(TargetOpc, SDLoc(NodeToMatch),
3172 // Add all the non-glue/non-chain results to the RecordedNodes list.
3173 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
3174 if (VTs[i] == MVT::Other || VTs[i] == MVT::Glue) break;
3175 RecordedNodes.push_back(std::pair<SDValue,SDNode*>(SDValue(Res, i),
3179 } else if (NodeToMatch->getOpcode() != ISD::DELETED_NODE) {
3180 Res = MorphNode(NodeToMatch, TargetOpc, VTList, Ops, EmitNodeInfo);
3182 // NodeToMatch was eliminated by CSE when the target changed the DAG.
3183 // We will visit the equivalent node later.
3184 DEBUG(dbgs() << "Node was eliminated by CSE\n");
3188 // If the node had chain/glue results, update our notion of the current
3190 if (EmitNodeInfo & OPFL_GlueOutput) {
3191 InputGlue = SDValue(Res, VTs.size()-1);
3192 if (EmitNodeInfo & OPFL_Chain)
3193 InputChain = SDValue(Res, VTs.size()-2);
3194 } else if (EmitNodeInfo & OPFL_Chain)
3195 InputChain = SDValue(Res, VTs.size()-1);
3197 // If the OPFL_MemRefs glue is set on this node, slap all of the
3198 // accumulated memrefs onto it.
3200 // FIXME: This is vastly incorrect for patterns with multiple outputs
3201 // instructions that access memory and for ComplexPatterns that match
3203 if (EmitNodeInfo & OPFL_MemRefs) {
3204 // Only attach load or store memory operands if the generated
3205 // instruction may load or store.
3206 const MCInstrDesc &MCID = TII->get(TargetOpc);
3207 bool mayLoad = MCID.mayLoad();
3208 bool mayStore = MCID.mayStore();
3210 unsigned NumMemRefs = 0;
3211 for (SmallVectorImpl<MachineMemOperand *>::const_iterator I =
3212 MatchedMemRefs.begin(), E = MatchedMemRefs.end(); I != E; ++I) {
3213 if ((*I)->isLoad()) {
3216 } else if ((*I)->isStore()) {
3224 MachineSDNode::mmo_iterator MemRefs =
3225 MF->allocateMemRefsArray(NumMemRefs);
3227 MachineSDNode::mmo_iterator MemRefsPos = MemRefs;
3228 for (SmallVectorImpl<MachineMemOperand *>::const_iterator I =
3229 MatchedMemRefs.begin(), E = MatchedMemRefs.end(); I != E; ++I) {
3230 if ((*I)->isLoad()) {
3233 } else if ((*I)->isStore()) {
3241 cast<MachineSDNode>(Res)
3242 ->setMemRefs(MemRefs, MemRefs + NumMemRefs);
3246 << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created")
3247 << " node: "; Res->dump(CurDAG); dbgs() << "\n");
3249 // If this was a MorphNodeTo then we're completely done!
3250 if (Opcode == OPC_MorphNodeTo) {
3251 // Update chain and glue uses.
3252 UpdateChainsAndGlue(NodeToMatch, InputChain, ChainNodesMatched,
3253 InputGlue, GlueResultNodesMatched, true);
3260 case OPC_MarkGlueResults: {
3261 unsigned NumNodes = MatcherTable[MatcherIndex++];
3263 // Read and remember all the glue-result nodes.
3264 for (unsigned i = 0; i != NumNodes; ++i) {
3265 unsigned RecNo = MatcherTable[MatcherIndex++];
3267 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
3269 assert(RecNo < RecordedNodes.size() && "Invalid MarkGlueResults");
3270 GlueResultNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
3275 case OPC_CompleteMatch: {
3276 // The match has been completed, and any new nodes (if any) have been
3277 // created. Patch up references to the matched dag to use the newly
3279 unsigned NumResults = MatcherTable[MatcherIndex++];
3281 for (unsigned i = 0; i != NumResults; ++i) {
3282 unsigned ResSlot = MatcherTable[MatcherIndex++];
3284 ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex);
3286 assert(ResSlot < RecordedNodes.size() && "Invalid CompleteMatch");
3287 SDValue Res = RecordedNodes[ResSlot].first;
3289 assert(i < NodeToMatch->getNumValues() &&
3290 NodeToMatch->getValueType(i) != MVT::Other &&
3291 NodeToMatch->getValueType(i) != MVT::Glue &&
3292 "Invalid number of results to complete!");
3293 assert((NodeToMatch->getValueType(i) == Res.getValueType() ||
3294 NodeToMatch->getValueType(i) == MVT::iPTR ||
3295 Res.getValueType() == MVT::iPTR ||
3296 NodeToMatch->getValueType(i).getSizeInBits() ==
3297 Res.getValueType().getSizeInBits()) &&
3298 "invalid replacement");
3299 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, i), Res);
3302 // If the root node defines glue, add it to the glue nodes to update list.
3303 if (NodeToMatch->getValueType(NodeToMatch->getNumValues()-1) == MVT::Glue)
3304 GlueResultNodesMatched.push_back(NodeToMatch);
3306 // Update chain and glue uses.
3307 UpdateChainsAndGlue(NodeToMatch, InputChain, ChainNodesMatched,
3308 InputGlue, GlueResultNodesMatched, false);
3310 assert(NodeToMatch->use_empty() &&
3311 "Didn't replace all uses of the node?");
3313 // FIXME: We just return here, which interacts correctly with SelectRoot
3314 // above. We should fix this to not return an SDNode* anymore.
3319 // If the code reached this point, then the match failed. See if there is
3320 // another child to try in the current 'Scope', otherwise pop it until we
3321 // find a case to check.
3322 DEBUG(dbgs() << " Match failed at index " << CurrentOpcodeIndex << "\n");
3323 ++NumDAGIselRetries;
3325 if (MatchScopes.empty()) {
3326 CannotYetSelect(NodeToMatch);
3330 // Restore the interpreter state back to the point where the scope was
3332 MatchScope &LastScope = MatchScopes.back();
3333 RecordedNodes.resize(LastScope.NumRecordedNodes);
3335 NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end());
3336 N = NodeStack.back();
3338 if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size())
3339 MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
3340 MatcherIndex = LastScope.FailIndex;
3342 DEBUG(dbgs() << " Continuing at " << MatcherIndex << "\n");
3344 InputChain = LastScope.InputChain;
3345 InputGlue = LastScope.InputGlue;
3346 if (!LastScope.HasChainNodesMatched)
3347 ChainNodesMatched.clear();
3348 if (!LastScope.HasGlueResultNodesMatched)
3349 GlueResultNodesMatched.clear();
3351 // Check to see what the offset is at the new MatcherIndex. If it is zero
3352 // we have reached the end of this scope, otherwise we have another child
3353 // in the current scope to try.
3354 unsigned NumToSkip = MatcherTable[MatcherIndex++];
3355 if (NumToSkip & 128)
3356 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
3358 // If we have another child in this scope to match, update FailIndex and
3360 if (NumToSkip != 0) {
3361 LastScope.FailIndex = MatcherIndex+NumToSkip;
3365 // End of this scope, pop it and try the next child in the containing
3367 MatchScopes.pop_back();
3374 void SelectionDAGISel::CannotYetSelect(SDNode *N) {
3376 raw_string_ostream Msg(msg);
3377 Msg << "Cannot select: ";
3379 if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN &&
3380 N->getOpcode() != ISD::INTRINSIC_WO_CHAIN &&
3381 N->getOpcode() != ISD::INTRINSIC_VOID) {
3382 N->printrFull(Msg, CurDAG);
3383 Msg << "\nIn function: " << MF->getName();
3385 bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other;
3387 cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue();
3388 if (iid < Intrinsic::num_intrinsics)
3389 Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid);
3390 else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo())
3391 Msg << "target intrinsic %" << TII->getName(iid);
3393 Msg << "unknown intrinsic #" << iid;
3395 report_fatal_error(Msg.str());
3398 char SelectionDAGISel::ID = 0;