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/IR/Constants.h"
37 #include "llvm/IR/DebugInfo.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/InlineAsm.h"
40 #include "llvm/IR/Instructions.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/Intrinsics.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/Module.h"
45 #include "llvm/MC/MCAsmInfo.h"
46 #include "llvm/Support/Compiler.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/Timer.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Target/TargetInstrInfo.h"
52 #include "llvm/Target/TargetIntrinsicInfo.h"
53 #include "llvm/Target/TargetLowering.h"
54 #include "llvm/Target/TargetMachine.h"
55 #include "llvm/Target/TargetOptions.h"
56 #include "llvm/Target/TargetRegisterInfo.h"
57 #include "llvm/Target/TargetSubtargetInfo.h"
58 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
62 #define DEBUG_TYPE "isel"
64 STATISTIC(NumFastIselFailures, "Number of instructions fast isel failed on");
65 STATISTIC(NumFastIselSuccess, "Number of instructions fast isel selected");
66 STATISTIC(NumFastIselBlocks, "Number of blocks selected entirely by fast isel");
67 STATISTIC(NumDAGBlocks, "Number of blocks selected using DAG");
68 STATISTIC(NumDAGIselRetries,"Number of times dag isel has to try another path");
69 STATISTIC(NumEntryBlocks, "Number of entry blocks encountered");
70 STATISTIC(NumFastIselFailLowerArguments,
71 "Number of entry blocks where fast isel failed to lower arguments");
75 EnableFastISelVerbose2("fast-isel-verbose2", cl::Hidden,
76 cl::desc("Enable extra verbose messages in the \"fast\" "
77 "instruction selector"));
80 STATISTIC(NumFastIselFailRet,"Fast isel fails on Ret");
81 STATISTIC(NumFastIselFailBr,"Fast isel fails on Br");
82 STATISTIC(NumFastIselFailSwitch,"Fast isel fails on Switch");
83 STATISTIC(NumFastIselFailIndirectBr,"Fast isel fails on IndirectBr");
84 STATISTIC(NumFastIselFailInvoke,"Fast isel fails on Invoke");
85 STATISTIC(NumFastIselFailResume,"Fast isel fails on Resume");
86 STATISTIC(NumFastIselFailUnreachable,"Fast isel fails on Unreachable");
88 // Standard binary operators...
89 STATISTIC(NumFastIselFailAdd,"Fast isel fails on Add");
90 STATISTIC(NumFastIselFailFAdd,"Fast isel fails on FAdd");
91 STATISTIC(NumFastIselFailSub,"Fast isel fails on Sub");
92 STATISTIC(NumFastIselFailFSub,"Fast isel fails on FSub");
93 STATISTIC(NumFastIselFailMul,"Fast isel fails on Mul");
94 STATISTIC(NumFastIselFailFMul,"Fast isel fails on FMul");
95 STATISTIC(NumFastIselFailUDiv,"Fast isel fails on UDiv");
96 STATISTIC(NumFastIselFailSDiv,"Fast isel fails on SDiv");
97 STATISTIC(NumFastIselFailFDiv,"Fast isel fails on FDiv");
98 STATISTIC(NumFastIselFailURem,"Fast isel fails on URem");
99 STATISTIC(NumFastIselFailSRem,"Fast isel fails on SRem");
100 STATISTIC(NumFastIselFailFRem,"Fast isel fails on FRem");
102 // Logical operators...
103 STATISTIC(NumFastIselFailAnd,"Fast isel fails on And");
104 STATISTIC(NumFastIselFailOr,"Fast isel fails on Or");
105 STATISTIC(NumFastIselFailXor,"Fast isel fails on Xor");
107 // Memory instructions...
108 STATISTIC(NumFastIselFailAlloca,"Fast isel fails on Alloca");
109 STATISTIC(NumFastIselFailLoad,"Fast isel fails on Load");
110 STATISTIC(NumFastIselFailStore,"Fast isel fails on Store");
111 STATISTIC(NumFastIselFailAtomicCmpXchg,"Fast isel fails on AtomicCmpXchg");
112 STATISTIC(NumFastIselFailAtomicRMW,"Fast isel fails on AtomicRWM");
113 STATISTIC(NumFastIselFailFence,"Fast isel fails on Frence");
114 STATISTIC(NumFastIselFailGetElementPtr,"Fast isel fails on GetElementPtr");
116 // Convert instructions...
117 STATISTIC(NumFastIselFailTrunc,"Fast isel fails on Trunc");
118 STATISTIC(NumFastIselFailZExt,"Fast isel fails on ZExt");
119 STATISTIC(NumFastIselFailSExt,"Fast isel fails on SExt");
120 STATISTIC(NumFastIselFailFPTrunc,"Fast isel fails on FPTrunc");
121 STATISTIC(NumFastIselFailFPExt,"Fast isel fails on FPExt");
122 STATISTIC(NumFastIselFailFPToUI,"Fast isel fails on FPToUI");
123 STATISTIC(NumFastIselFailFPToSI,"Fast isel fails on FPToSI");
124 STATISTIC(NumFastIselFailUIToFP,"Fast isel fails on UIToFP");
125 STATISTIC(NumFastIselFailSIToFP,"Fast isel fails on SIToFP");
126 STATISTIC(NumFastIselFailIntToPtr,"Fast isel fails on IntToPtr");
127 STATISTIC(NumFastIselFailPtrToInt,"Fast isel fails on PtrToInt");
128 STATISTIC(NumFastIselFailBitCast,"Fast isel fails on BitCast");
130 // Other instructions...
131 STATISTIC(NumFastIselFailICmp,"Fast isel fails on ICmp");
132 STATISTIC(NumFastIselFailFCmp,"Fast isel fails on FCmp");
133 STATISTIC(NumFastIselFailPHI,"Fast isel fails on PHI");
134 STATISTIC(NumFastIselFailSelect,"Fast isel fails on Select");
135 STATISTIC(NumFastIselFailCall,"Fast isel fails on Call");
136 STATISTIC(NumFastIselFailShl,"Fast isel fails on Shl");
137 STATISTIC(NumFastIselFailLShr,"Fast isel fails on LShr");
138 STATISTIC(NumFastIselFailAShr,"Fast isel fails on AShr");
139 STATISTIC(NumFastIselFailVAArg,"Fast isel fails on VAArg");
140 STATISTIC(NumFastIselFailExtractElement,"Fast isel fails on ExtractElement");
141 STATISTIC(NumFastIselFailInsertElement,"Fast isel fails on InsertElement");
142 STATISTIC(NumFastIselFailShuffleVector,"Fast isel fails on ShuffleVector");
143 STATISTIC(NumFastIselFailExtractValue,"Fast isel fails on ExtractValue");
144 STATISTIC(NumFastIselFailInsertValue,"Fast isel fails on InsertValue");
145 STATISTIC(NumFastIselFailLandingPad,"Fast isel fails on LandingPad");
147 // Intrinsic instructions...
148 STATISTIC(NumFastIselFailIntrinsicCall, "Fast isel fails on Intrinsic call");
149 STATISTIC(NumFastIselFailSAddWithOverflow,
150 "Fast isel fails on sadd.with.overflow");
151 STATISTIC(NumFastIselFailUAddWithOverflow,
152 "Fast isel fails on uadd.with.overflow");
153 STATISTIC(NumFastIselFailSSubWithOverflow,
154 "Fast isel fails on ssub.with.overflow");
155 STATISTIC(NumFastIselFailUSubWithOverflow,
156 "Fast isel fails on usub.with.overflow");
157 STATISTIC(NumFastIselFailSMulWithOverflow,
158 "Fast isel fails on smul.with.overflow");
159 STATISTIC(NumFastIselFailUMulWithOverflow,
160 "Fast isel fails on umul.with.overflow");
161 STATISTIC(NumFastIselFailFrameaddress, "Fast isel fails on Frameaddress");
162 STATISTIC(NumFastIselFailSqrt, "Fast isel fails on sqrt call");
163 STATISTIC(NumFastIselFailStackMap, "Fast isel fails on StackMap call");
164 STATISTIC(NumFastIselFailPatchPoint, "Fast isel fails on PatchPoint call");
168 EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
169 cl::desc("Enable verbose messages in the \"fast\" "
170 "instruction selector"));
171 static cl::opt<int> EnableFastISelAbort(
172 "fast-isel-abort", cl::Hidden,
173 cl::desc("Enable abort calls when \"fast\" instruction selection "
174 "fails to lower an instruction: 0 disable the abort, 1 will "
175 "abort but for args, calls and terminators, 2 will also "
176 "abort for argument lowering, and 3 will never fallback "
177 "to SelectionDAG."));
181 cl::desc("use Machine Branch Probability Info"),
182 cl::init(true), cl::Hidden);
185 static cl::opt<std::string>
186 FilterDAGBasicBlockName("filter-view-dags", cl::Hidden,
187 cl::desc("Only display the basic block whose name "
188 "matches this for all view-*-dags options"));
190 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
191 cl::desc("Pop up a window to show dags before the first "
192 "dag combine pass"));
194 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
195 cl::desc("Pop up a window to show dags before legalize types"));
197 ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
198 cl::desc("Pop up a window to show dags before legalize"));
200 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
201 cl::desc("Pop up a window to show dags before the second "
202 "dag combine pass"));
204 ViewDAGCombineLT("view-dag-combine-lt-dags", cl::Hidden,
205 cl::desc("Pop up a window to show dags before the post legalize types"
206 " dag combine pass"));
208 ViewISelDAGs("view-isel-dags", cl::Hidden,
209 cl::desc("Pop up a window to show isel dags as they are selected"));
211 ViewSchedDAGs("view-sched-dags", cl::Hidden,
212 cl::desc("Pop up a window to show sched dags as they are processed"));
214 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
215 cl::desc("Pop up a window to show SUnit dags after they are processed"));
217 static const bool ViewDAGCombine1 = false,
218 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false,
219 ViewDAGCombine2 = false,
220 ViewDAGCombineLT = false,
221 ViewISelDAGs = false, ViewSchedDAGs = false,
222 ViewSUnitDAGs = false;
225 //===---------------------------------------------------------------------===//
227 /// RegisterScheduler class - Track the registration of instruction schedulers.
229 //===---------------------------------------------------------------------===//
230 MachinePassRegistry RegisterScheduler::Registry;
232 //===---------------------------------------------------------------------===//
234 /// ISHeuristic command line option for instruction schedulers.
236 //===---------------------------------------------------------------------===//
237 static cl::opt<RegisterScheduler::FunctionPassCtor, false,
238 RegisterPassParser<RegisterScheduler> >
239 ISHeuristic("pre-RA-sched",
240 cl::init(&createDefaultScheduler), cl::Hidden,
241 cl::desc("Instruction schedulers available (before register"
244 static RegisterScheduler
245 defaultListDAGScheduler("default", "Best scheduler for the target",
246 createDefaultScheduler);
249 //===--------------------------------------------------------------------===//
250 /// \brief This class is used by SelectionDAGISel to temporarily override
251 /// the optimization level on a per-function basis.
252 class OptLevelChanger {
253 SelectionDAGISel &IS;
254 CodeGenOpt::Level SavedOptLevel;
258 OptLevelChanger(SelectionDAGISel &ISel,
259 CodeGenOpt::Level NewOptLevel) : IS(ISel) {
260 SavedOptLevel = IS.OptLevel;
261 if (NewOptLevel == SavedOptLevel)
263 IS.OptLevel = NewOptLevel;
264 IS.TM.setOptLevel(NewOptLevel);
265 SavedFastISel = IS.TM.Options.EnableFastISel;
266 if (NewOptLevel == CodeGenOpt::None)
267 IS.TM.setFastISel(true);
268 DEBUG(dbgs() << "\nChanging optimization level for Function "
269 << IS.MF->getFunction()->getName() << "\n");
270 DEBUG(dbgs() << "\tBefore: -O" << SavedOptLevel
271 << " ; After: -O" << NewOptLevel << "\n");
275 if (IS.OptLevel == SavedOptLevel)
277 DEBUG(dbgs() << "\nRestoring optimization level for Function "
278 << IS.MF->getFunction()->getName() << "\n");
279 DEBUG(dbgs() << "\tBefore: -O" << IS.OptLevel
280 << " ; After: -O" << SavedOptLevel << "\n");
281 IS.OptLevel = SavedOptLevel;
282 IS.TM.setOptLevel(SavedOptLevel);
283 IS.TM.setFastISel(SavedFastISel);
287 //===--------------------------------------------------------------------===//
288 /// createDefaultScheduler - This creates an instruction scheduler appropriate
290 ScheduleDAGSDNodes* createDefaultScheduler(SelectionDAGISel *IS,
291 CodeGenOpt::Level OptLevel) {
292 const TargetLowering *TLI = IS->TLI;
293 const TargetSubtargetInfo &ST = IS->MF->getSubtarget();
295 if (OptLevel == CodeGenOpt::None || ST.useMachineScheduler() ||
296 TLI->getSchedulingPreference() == Sched::Source)
297 return createSourceListDAGScheduler(IS, OptLevel);
298 if (TLI->getSchedulingPreference() == Sched::RegPressure)
299 return createBURRListDAGScheduler(IS, OptLevel);
300 if (TLI->getSchedulingPreference() == Sched::Hybrid)
301 return createHybridListDAGScheduler(IS, OptLevel);
302 if (TLI->getSchedulingPreference() == Sched::VLIW)
303 return createVLIWDAGScheduler(IS, OptLevel);
304 assert(TLI->getSchedulingPreference() == Sched::ILP &&
305 "Unknown sched type!");
306 return createILPListDAGScheduler(IS, OptLevel);
310 // EmitInstrWithCustomInserter - This method should be implemented by targets
311 // that mark instructions with the 'usesCustomInserter' flag. These
312 // instructions are special in various ways, which require special support to
313 // insert. The specified MachineInstr is created but not inserted into any
314 // basic blocks, and this method is called to expand it into a sequence of
315 // instructions, potentially also creating new basic blocks and control flow.
316 // When new basic blocks are inserted and the edges from MBB to its successors
317 // are modified, the method should insert pairs of <OldSucc, NewSucc> into the
320 TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
321 MachineBasicBlock *MBB) const {
323 dbgs() << "If a target marks an instruction with "
324 "'usesCustomInserter', it must implement "
325 "TargetLowering::EmitInstrWithCustomInserter!";
327 llvm_unreachable(nullptr);
330 void TargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
331 SDNode *Node) const {
332 assert(!MI->hasPostISelHook() &&
333 "If a target marks an instruction with 'hasPostISelHook', "
334 "it must implement TargetLowering::AdjustInstrPostInstrSelection!");
337 //===----------------------------------------------------------------------===//
338 // SelectionDAGISel code
339 //===----------------------------------------------------------------------===//
341 SelectionDAGISel::SelectionDAGISel(TargetMachine &tm,
342 CodeGenOpt::Level OL) :
343 MachineFunctionPass(ID), TM(tm),
344 FuncInfo(new FunctionLoweringInfo()),
345 CurDAG(new SelectionDAG(tm, OL)),
346 SDB(new SelectionDAGBuilder(*CurDAG, *FuncInfo, OL)),
350 initializeGCModuleInfoPass(*PassRegistry::getPassRegistry());
351 initializeAliasAnalysisAnalysisGroup(*PassRegistry::getPassRegistry());
352 initializeBranchProbabilityInfoPass(*PassRegistry::getPassRegistry());
353 initializeTargetLibraryInfoWrapperPassPass(
354 *PassRegistry::getPassRegistry());
357 SelectionDAGISel::~SelectionDAGISel() {
363 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
364 AU.addRequired<AliasAnalysis>();
365 AU.addPreserved<AliasAnalysis>();
366 AU.addRequired<GCModuleInfo>();
367 AU.addPreserved<GCModuleInfo>();
368 AU.addRequired<TargetLibraryInfoWrapperPass>();
369 if (UseMBPI && OptLevel != CodeGenOpt::None)
370 AU.addRequired<BranchProbabilityInfo>();
371 MachineFunctionPass::getAnalysisUsage(AU);
374 /// SplitCriticalSideEffectEdges - Look for critical edges with a PHI value that
375 /// may trap on it. In this case we have to split the edge so that the path
376 /// through the predecessor block that doesn't go to the phi block doesn't
377 /// execute the possibly trapping instruction.
379 /// This is required for correctness, so it must be done at -O0.
381 static void SplitCriticalSideEffectEdges(Function &Fn, AliasAnalysis *AA) {
382 // Loop for blocks with phi nodes.
383 for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
384 PHINode *PN = dyn_cast<PHINode>(BB->begin());
388 // For each block with a PHI node, check to see if any of the input values
389 // are potentially trapping constant expressions. Constant expressions are
390 // the only potentially trapping value that can occur as the argument to a
392 for (BasicBlock::iterator I = BB->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
393 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
394 ConstantExpr *CE = dyn_cast<ConstantExpr>(PN->getIncomingValue(i));
395 if (!CE || !CE->canTrap()) continue;
397 // The only case we have to worry about is when the edge is critical.
398 // Since this block has a PHI Node, we assume it has multiple input
399 // edges: check to see if the pred has multiple successors.
400 BasicBlock *Pred = PN->getIncomingBlock(i);
401 if (Pred->getTerminator()->getNumSuccessors() == 1)
404 // Okay, we have to split this edge.
406 Pred->getTerminator(), GetSuccessorNumber(Pred, BB),
407 CriticalEdgeSplittingOptions(AA).setMergeIdenticalEdges());
413 bool SelectionDAGISel::runOnMachineFunction(MachineFunction &mf) {
414 // Do some sanity-checking on the command-line options.
415 assert((!EnableFastISelVerbose || TM.Options.EnableFastISel) &&
416 "-fast-isel-verbose requires -fast-isel");
417 assert((!EnableFastISelAbort || TM.Options.EnableFastISel) &&
418 "-fast-isel-abort > 0 requires -fast-isel");
420 const Function &Fn = *mf.getFunction();
423 // Reset the target options before resetting the optimization
425 // FIXME: This is a horrible hack and should be processed via
426 // codegen looking at the optimization level explicitly when
427 // it wants to look at it.
428 TM.resetTargetOptions(Fn);
429 // Reset OptLevel to None for optnone functions.
430 CodeGenOpt::Level NewOptLevel = OptLevel;
431 if (Fn.hasFnAttribute(Attribute::OptimizeNone))
432 NewOptLevel = CodeGenOpt::None;
433 OptLevelChanger OLC(*this, NewOptLevel);
435 TII = MF->getSubtarget().getInstrInfo();
436 TLI = MF->getSubtarget().getTargetLowering();
437 RegInfo = &MF->getRegInfo();
438 AA = &getAnalysis<AliasAnalysis>();
439 LibInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
440 GFI = Fn.hasGC() ? &getAnalysis<GCModuleInfo>().getFunctionInfo(Fn) : nullptr;
442 DEBUG(dbgs() << "\n\n\n=== " << Fn.getName() << "\n");
444 SplitCriticalSideEffectEdges(const_cast<Function&>(Fn), AA);
447 FuncInfo->set(Fn, *MF, CurDAG);
449 if (UseMBPI && OptLevel != CodeGenOpt::None)
450 FuncInfo->BPI = &getAnalysis<BranchProbabilityInfo>();
452 FuncInfo->BPI = nullptr;
454 SDB->init(GFI, *AA, LibInfo);
456 MF->setHasInlineAsm(false);
458 SelectAllBasicBlocks(Fn);
460 // If the first basic block in the function has live ins that need to be
461 // copied into vregs, emit the copies into the top of the block before
462 // emitting the code for the block.
463 MachineBasicBlock *EntryMBB = MF->begin();
464 const TargetRegisterInfo &TRI = *MF->getSubtarget().getRegisterInfo();
465 RegInfo->EmitLiveInCopies(EntryMBB, TRI, *TII);
467 DenseMap<unsigned, unsigned> LiveInMap;
468 if (!FuncInfo->ArgDbgValues.empty())
469 for (MachineRegisterInfo::livein_iterator LI = RegInfo->livein_begin(),
470 E = RegInfo->livein_end(); LI != E; ++LI)
472 LiveInMap.insert(std::make_pair(LI->first, LI->second));
474 // Insert DBG_VALUE instructions for function arguments to the entry block.
475 for (unsigned i = 0, e = FuncInfo->ArgDbgValues.size(); i != e; ++i) {
476 MachineInstr *MI = FuncInfo->ArgDbgValues[e-i-1];
477 bool hasFI = MI->getOperand(0).isFI();
479 hasFI ? TRI.getFrameRegister(*MF) : MI->getOperand(0).getReg();
480 if (TargetRegisterInfo::isPhysicalRegister(Reg))
481 EntryMBB->insert(EntryMBB->begin(), MI);
483 MachineInstr *Def = RegInfo->getVRegDef(Reg);
485 MachineBasicBlock::iterator InsertPos = Def;
486 // FIXME: VR def may not be in entry block.
487 Def->getParent()->insert(std::next(InsertPos), MI);
489 DEBUG(dbgs() << "Dropping debug info for dead vreg"
490 << TargetRegisterInfo::virtReg2Index(Reg) << "\n");
493 // If Reg is live-in then update debug info to track its copy in a vreg.
494 DenseMap<unsigned, unsigned>::iterator LDI = LiveInMap.find(Reg);
495 if (LDI != LiveInMap.end()) {
496 assert(!hasFI && "There's no handling of frame pointer updating here yet "
498 MachineInstr *Def = RegInfo->getVRegDef(LDI->second);
499 MachineBasicBlock::iterator InsertPos = Def;
500 const MDNode *Variable = MI->getDebugVariable();
501 const MDNode *Expr = MI->getDebugExpression();
502 bool IsIndirect = MI->isIndirectDebugValue();
503 unsigned Offset = IsIndirect ? MI->getOperand(1).getImm() : 0;
504 // Def is never a terminator here, so it is ok to increment InsertPos.
505 BuildMI(*EntryMBB, ++InsertPos, MI->getDebugLoc(),
506 TII->get(TargetOpcode::DBG_VALUE), IsIndirect, LDI->second, Offset,
509 // If this vreg is directly copied into an exported register then
510 // that COPY instructions also need DBG_VALUE, if it is the only
511 // user of LDI->second.
512 MachineInstr *CopyUseMI = nullptr;
513 for (MachineRegisterInfo::use_instr_iterator
514 UI = RegInfo->use_instr_begin(LDI->second),
515 E = RegInfo->use_instr_end(); UI != E; ) {
516 MachineInstr *UseMI = &*(UI++);
517 if (UseMI->isDebugValue()) continue;
518 if (UseMI->isCopy() && !CopyUseMI && UseMI->getParent() == EntryMBB) {
519 CopyUseMI = UseMI; continue;
521 // Otherwise this is another use or second copy use.
522 CopyUseMI = nullptr; break;
525 MachineInstr *NewMI =
526 BuildMI(*MF, CopyUseMI->getDebugLoc(),
527 TII->get(TargetOpcode::DBG_VALUE), IsIndirect,
528 CopyUseMI->getOperand(0).getReg(), Offset, Variable, Expr);
529 MachineBasicBlock::iterator Pos = CopyUseMI;
530 EntryMBB->insertAfter(Pos, NewMI);
535 // Determine if there are any calls in this machine function.
536 MachineFrameInfo *MFI = MF->getFrameInfo();
537 for (const auto &MBB : *MF) {
538 if (MFI->hasCalls() && MF->hasInlineAsm())
541 for (const auto &MI : MBB) {
542 const MCInstrDesc &MCID = TII->get(MI.getOpcode());
543 if ((MCID.isCall() && !MCID.isReturn()) ||
544 MI.isStackAligningInlineAsm()) {
545 MFI->setHasCalls(true);
547 if (MI.isInlineAsm()) {
548 MF->setHasInlineAsm(true);
553 // Determine if there is a call to setjmp in the machine function.
554 MF->setExposesReturnsTwice(Fn.callsFunctionThatReturnsTwice());
556 // Replace forward-declared registers with the registers containing
557 // the desired value.
558 MachineRegisterInfo &MRI = MF->getRegInfo();
559 for (DenseMap<unsigned, unsigned>::iterator
560 I = FuncInfo->RegFixups.begin(), E = FuncInfo->RegFixups.end();
562 unsigned From = I->first;
563 unsigned To = I->second;
564 // If To is also scheduled to be replaced, find what its ultimate
567 DenseMap<unsigned, unsigned>::iterator J = FuncInfo->RegFixups.find(To);
571 // Make sure the new register has a sufficiently constrained register class.
572 if (TargetRegisterInfo::isVirtualRegister(From) &&
573 TargetRegisterInfo::isVirtualRegister(To))
574 MRI.constrainRegClass(To, MRI.getRegClass(From));
576 MRI.replaceRegWith(From, To);
579 // Freeze the set of reserved registers now that MachineFrameInfo has been
580 // set up. All the information required by getReservedRegs() should be
582 MRI.freezeReservedRegs(*MF);
584 // Release function-specific state. SDB and CurDAG are already cleared
588 DEBUG(dbgs() << "*** MachineFunction at end of ISel ***\n");
589 DEBUG(MF->print(dbgs()));
594 void SelectionDAGISel::SelectBasicBlock(BasicBlock::const_iterator Begin,
595 BasicBlock::const_iterator End,
597 // Lower all of the non-terminator instructions. If a call is emitted
598 // as a tail call, cease emitting nodes for this block. Terminators
599 // are handled below.
600 for (BasicBlock::const_iterator I = Begin; I != End && !SDB->HasTailCall; ++I)
603 // Make sure the root of the DAG is up-to-date.
604 CurDAG->setRoot(SDB->getControlRoot());
605 HadTailCall = SDB->HasTailCall;
608 // Final step, emit the lowered DAG as machine code.
612 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
613 SmallPtrSet<SDNode*, 128> VisitedNodes;
614 SmallVector<SDNode*, 128> Worklist;
616 Worklist.push_back(CurDAG->getRoot().getNode());
622 SDNode *N = Worklist.pop_back_val();
624 // If we've already seen this node, ignore it.
625 if (!VisitedNodes.insert(N).second)
628 // Otherwise, add all chain operands to the worklist.
629 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
630 if (N->getOperand(i).getValueType() == MVT::Other)
631 Worklist.push_back(N->getOperand(i).getNode());
633 // If this is a CopyToReg with a vreg dest, process it.
634 if (N->getOpcode() != ISD::CopyToReg)
637 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
638 if (!TargetRegisterInfo::isVirtualRegister(DestReg))
641 // Ignore non-scalar or non-integer values.
642 SDValue Src = N->getOperand(2);
643 EVT SrcVT = Src.getValueType();
644 if (!SrcVT.isInteger() || SrcVT.isVector())
647 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
648 CurDAG->computeKnownBits(Src, KnownZero, KnownOne);
649 FuncInfo->AddLiveOutRegInfo(DestReg, NumSignBits, KnownZero, KnownOne);
650 } while (!Worklist.empty());
653 void SelectionDAGISel::CodeGenAndEmitDAG() {
654 std::string GroupName;
655 if (TimePassesIsEnabled)
656 GroupName = "Instruction Selection and Scheduling";
657 std::string BlockName;
658 int BlockNumber = -1;
660 bool MatchFilterBB = false; (void)MatchFilterBB;
662 MatchFilterBB = (FilterDAGBasicBlockName.empty() ||
663 FilterDAGBasicBlockName ==
664 FuncInfo->MBB->getBasicBlock()->getName().str());
667 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs ||
668 ViewDAGCombine2 || ViewDAGCombineLT || ViewISelDAGs || ViewSchedDAGs ||
672 BlockNumber = FuncInfo->MBB->getNumber();
673 BlockName = MF->getName().str() + ":" +
674 FuncInfo->MBB->getBasicBlock()->getName().str();
676 DEBUG(dbgs() << "Initial selection DAG: BB#" << BlockNumber
677 << " '" << BlockName << "'\n"; CurDAG->dump());
679 if (ViewDAGCombine1 && MatchFilterBB)
680 CurDAG->viewGraph("dag-combine1 input for " + BlockName);
682 // Run the DAG combiner in pre-legalize mode.
684 NamedRegionTimer T("DAG Combining 1", GroupName, TimePassesIsEnabled);
685 CurDAG->Combine(BeforeLegalizeTypes, *AA, OptLevel);
688 DEBUG(dbgs() << "Optimized lowered selection DAG: BB#" << BlockNumber
689 << " '" << BlockName << "'\n"; CurDAG->dump());
691 // Second step, hack on the DAG until it only uses operations and types that
692 // the target supports.
693 if (ViewLegalizeTypesDAGs && MatchFilterBB)
694 CurDAG->viewGraph("legalize-types input for " + BlockName);
698 NamedRegionTimer T("Type Legalization", GroupName, TimePassesIsEnabled);
699 Changed = CurDAG->LegalizeTypes();
702 DEBUG(dbgs() << "Type-legalized selection DAG: BB#" << BlockNumber
703 << " '" << BlockName << "'\n"; CurDAG->dump());
705 CurDAG->NewNodesMustHaveLegalTypes = true;
708 if (ViewDAGCombineLT && MatchFilterBB)
709 CurDAG->viewGraph("dag-combine-lt input for " + BlockName);
711 // Run the DAG combiner in post-type-legalize mode.
713 NamedRegionTimer T("DAG Combining after legalize types", GroupName,
714 TimePassesIsEnabled);
715 CurDAG->Combine(AfterLegalizeTypes, *AA, OptLevel);
718 DEBUG(dbgs() << "Optimized type-legalized selection DAG: BB#" << BlockNumber
719 << " '" << BlockName << "'\n"; CurDAG->dump());
724 NamedRegionTimer T("Vector Legalization", GroupName, TimePassesIsEnabled);
725 Changed = CurDAG->LegalizeVectors();
730 NamedRegionTimer T("Type Legalization 2", GroupName, TimePassesIsEnabled);
731 CurDAG->LegalizeTypes();
734 if (ViewDAGCombineLT && MatchFilterBB)
735 CurDAG->viewGraph("dag-combine-lv input for " + BlockName);
737 // Run the DAG combiner in post-type-legalize mode.
739 NamedRegionTimer T("DAG Combining after legalize vectors", GroupName,
740 TimePassesIsEnabled);
741 CurDAG->Combine(AfterLegalizeVectorOps, *AA, OptLevel);
744 DEBUG(dbgs() << "Optimized vector-legalized selection DAG: BB#"
745 << BlockNumber << " '" << BlockName << "'\n"; CurDAG->dump());
748 if (ViewLegalizeDAGs && MatchFilterBB)
749 CurDAG->viewGraph("legalize input for " + BlockName);
752 NamedRegionTimer T("DAG Legalization", GroupName, TimePassesIsEnabled);
756 DEBUG(dbgs() << "Legalized selection DAG: BB#" << BlockNumber
757 << " '" << BlockName << "'\n"; CurDAG->dump());
759 if (ViewDAGCombine2 && MatchFilterBB)
760 CurDAG->viewGraph("dag-combine2 input for " + BlockName);
762 // Run the DAG combiner in post-legalize mode.
764 NamedRegionTimer T("DAG Combining 2", GroupName, TimePassesIsEnabled);
765 CurDAG->Combine(AfterLegalizeDAG, *AA, OptLevel);
768 DEBUG(dbgs() << "Optimized legalized selection DAG: BB#" << BlockNumber
769 << " '" << BlockName << "'\n"; CurDAG->dump());
771 if (OptLevel != CodeGenOpt::None)
772 ComputeLiveOutVRegInfo();
774 if (ViewISelDAGs && MatchFilterBB)
775 CurDAG->viewGraph("isel input for " + BlockName);
777 // Third, instruction select all of the operations to machine code, adding the
778 // code to the MachineBasicBlock.
780 NamedRegionTimer T("Instruction Selection", GroupName, TimePassesIsEnabled);
781 DoInstructionSelection();
784 DEBUG(dbgs() << "Selected selection DAG: BB#" << BlockNumber
785 << " '" << BlockName << "'\n"; CurDAG->dump());
787 if (ViewSchedDAGs && MatchFilterBB)
788 CurDAG->viewGraph("scheduler input for " + BlockName);
790 // Schedule machine code.
791 ScheduleDAGSDNodes *Scheduler = CreateScheduler();
793 NamedRegionTimer T("Instruction Scheduling", GroupName,
794 TimePassesIsEnabled);
795 Scheduler->Run(CurDAG, FuncInfo->MBB);
798 if (ViewSUnitDAGs && MatchFilterBB) Scheduler->viewGraph();
800 // Emit machine code to BB. This can change 'BB' to the last block being
802 MachineBasicBlock *FirstMBB = FuncInfo->MBB, *LastMBB;
804 NamedRegionTimer T("Instruction Creation", GroupName, TimePassesIsEnabled);
806 // FuncInfo->InsertPt is passed by reference and set to the end of the
807 // scheduled instructions.
808 LastMBB = FuncInfo->MBB = Scheduler->EmitSchedule(FuncInfo->InsertPt);
811 // If the block was split, make sure we update any references that are used to
812 // update PHI nodes later on.
813 if (FirstMBB != LastMBB)
814 SDB->UpdateSplitBlock(FirstMBB, LastMBB);
816 // Free the scheduler state.
818 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName,
819 TimePassesIsEnabled);
823 // Free the SelectionDAG state, now that we're finished with it.
828 /// ISelUpdater - helper class to handle updates of the instruction selection
830 class ISelUpdater : public SelectionDAG::DAGUpdateListener {
831 SelectionDAG::allnodes_iterator &ISelPosition;
833 ISelUpdater(SelectionDAG &DAG, SelectionDAG::allnodes_iterator &isp)
834 : SelectionDAG::DAGUpdateListener(DAG), ISelPosition(isp) {}
836 /// NodeDeleted - Handle nodes deleted from the graph. If the node being
837 /// deleted is the current ISelPosition node, update ISelPosition.
839 void NodeDeleted(SDNode *N, SDNode *E) override {
840 if (ISelPosition == SelectionDAG::allnodes_iterator(N))
844 } // end anonymous namespace
846 void SelectionDAGISel::DoInstructionSelection() {
847 DEBUG(dbgs() << "===== Instruction selection begins: BB#"
848 << FuncInfo->MBB->getNumber()
849 << " '" << FuncInfo->MBB->getName() << "'\n");
853 // Select target instructions for the DAG.
855 // Number all nodes with a topological order and set DAGSize.
856 DAGSize = CurDAG->AssignTopologicalOrder();
858 // Create a dummy node (which is not added to allnodes), that adds
859 // a reference to the root node, preventing it from being deleted,
860 // and tracking any changes of the root.
861 HandleSDNode Dummy(CurDAG->getRoot());
862 SelectionDAG::allnodes_iterator ISelPosition (CurDAG->getRoot().getNode());
865 // Make sure that ISelPosition gets properly updated when nodes are deleted
866 // in calls made from this function.
867 ISelUpdater ISU(*CurDAG, ISelPosition);
869 // The AllNodes list is now topological-sorted. Visit the
870 // nodes by starting at the end of the list (the root of the
871 // graph) and preceding back toward the beginning (the entry
873 while (ISelPosition != CurDAG->allnodes_begin()) {
874 SDNode *Node = --ISelPosition;
875 // Skip dead nodes. DAGCombiner is expected to eliminate all dead nodes,
876 // but there are currently some corner cases that it misses. Also, this
877 // makes it theoretically possible to disable the DAGCombiner.
878 if (Node->use_empty())
881 SDNode *ResNode = Select(Node);
883 // FIXME: This is pretty gross. 'Select' should be changed to not return
884 // anything at all and this code should be nuked with a tactical strike.
886 // If node should not be replaced, continue with the next one.
887 if (ResNode == Node || Node->getOpcode() == ISD::DELETED_NODE)
891 ReplaceUses(Node, ResNode);
894 // If after the replacement this node is not used any more,
895 // remove this dead node.
896 if (Node->use_empty()) // Don't delete EntryToken, etc.
897 CurDAG->RemoveDeadNode(Node);
900 CurDAG->setRoot(Dummy.getValue());
903 DEBUG(dbgs() << "===== Instruction selection ends:\n");
905 PostprocessISelDAG();
908 /// PrepareEHLandingPad - Emit an EH_LABEL, set up live-in registers, and
909 /// do other setup for EH landing-pad blocks.
910 void SelectionDAGISel::PrepareEHLandingPad() {
911 MachineBasicBlock *MBB = FuncInfo->MBB;
913 const TargetRegisterClass *PtrRC = TLI->getRegClassFor(TLI->getPointerTy());
915 // Add a label to mark the beginning of the landing pad. Deletion of the
916 // landing pad can thus be detected via the MachineModuleInfo.
917 MCSymbol *Label = MF->getMMI().addLandingPad(MBB);
919 // Assign the call site to the landing pad's begin label.
920 MF->getMMI().setCallSiteLandingPad(Label, SDB->LPadToCallSiteMap[MBB]);
922 const MCInstrDesc &II = TII->get(TargetOpcode::EH_LABEL);
923 BuildMI(*MBB, FuncInfo->InsertPt, SDB->getCurDebugLoc(), II)
926 // If this is an MSVC-style personality function, we need to split the landing
927 // pad into several BBs.
928 const BasicBlock *LLVMBB = MBB->getBasicBlock();
929 const LandingPadInst *LPadInst = LLVMBB->getLandingPadInst();
930 MF->getMMI().addPersonality(
931 MBB, cast<Function>(LPadInst->getPersonalityFn()->stripPointerCasts()));
932 if (MF->getMMI().getPersonalityType() == EHPersonality::MSVC_Win64SEH) {
933 // Make virtual registers and a series of labels that fill in values for the
935 auto &RI = MF->getRegInfo();
936 FuncInfo->ExceptionSelectorVirtReg = RI.createVirtualRegister(PtrRC);
938 // Get all invoke BBs that will unwind into the clause BBs.
939 SmallVector<MachineBasicBlock *, 4> InvokeBBs(MBB->pred_begin(),
942 // Emit separate machine basic blocks with separate labels for each clause
943 // before the main landing pad block.
944 MachineInstrBuilder SelectorPHI = BuildMI(
945 *MBB, MBB->begin(), SDB->getCurDebugLoc(), TII->get(TargetOpcode::PHI),
946 FuncInfo->ExceptionSelectorVirtReg);
947 for (unsigned I = 0, E = LPadInst->getNumClauses(); I != E; ++I) {
948 // Skip filter clauses, we can't implement them yet.
949 if (LPadInst->isFilter(I))
952 MachineBasicBlock *ClauseBB = MF->CreateMachineBasicBlock(LLVMBB);
953 MF->insert(MBB, ClauseBB);
955 // Add the edge from the invoke to the clause.
956 for (MachineBasicBlock *InvokeBB : InvokeBBs)
957 InvokeBB->addSuccessor(ClauseBB);
959 // Mark the clause as a landing pad or MI passes will delete it.
960 ClauseBB->setIsLandingPad();
962 GlobalValue *ClauseGV = ExtractTypeInfo(LPadInst->getClause(I));
964 // Start the BB with a label.
965 MCSymbol *ClauseLabel = MF->getMMI().addClauseForLandingPad(MBB);
966 BuildMI(*ClauseBB, ClauseBB->begin(), SDB->getCurDebugLoc(), II)
967 .addSym(ClauseLabel);
969 // Construct a simple BB that defines a register with the typeid constant.
970 FuncInfo->MBB = ClauseBB;
971 FuncInfo->InsertPt = ClauseBB->end();
972 unsigned VReg = SDB->visitLandingPadClauseBB(ClauseGV, MBB);
973 CurDAG->setRoot(SDB->getRoot());
977 // Add the typeid virtual register to the phi in the main landing pad.
978 SelectorPHI.addReg(VReg).addMBB(ClauseBB);
981 // Remove the edge from the invoke to the lpad.
982 for (MachineBasicBlock *InvokeBB : InvokeBBs)
983 InvokeBB->removeSuccessor(MBB);
985 // Restore FuncInfo back to its previous state and select the main landing
988 FuncInfo->InsertPt = MBB->end();
992 // Mark exception register as live in.
993 if (unsigned Reg = TLI->getExceptionPointerRegister())
994 FuncInfo->ExceptionPointerVirtReg = MBB->addLiveIn(Reg, PtrRC);
996 // Mark exception selector register as live in.
997 if (unsigned Reg = TLI->getExceptionSelectorRegister())
998 FuncInfo->ExceptionSelectorVirtReg = MBB->addLiveIn(Reg, PtrRC);
1001 /// isFoldedOrDeadInstruction - Return true if the specified instruction is
1002 /// side-effect free and is either dead or folded into a generated instruction.
1003 /// Return false if it needs to be emitted.
1004 static bool isFoldedOrDeadInstruction(const Instruction *I,
1005 FunctionLoweringInfo *FuncInfo) {
1006 return !I->mayWriteToMemory() && // Side-effecting instructions aren't folded.
1007 !isa<TerminatorInst>(I) && // Terminators aren't folded.
1008 !isa<DbgInfoIntrinsic>(I) && // Debug instructions aren't folded.
1009 !isa<LandingPadInst>(I) && // Landingpad instructions aren't folded.
1010 !FuncInfo->isExportedInst(I); // Exported instrs must be computed.
1014 // Collect per Instruction statistics for fast-isel misses. Only those
1015 // instructions that cause the bail are accounted for. It does not account for
1016 // instructions higher in the block. Thus, summing the per instructions stats
1017 // will not add up to what is reported by NumFastIselFailures.
1018 static void collectFailStats(const Instruction *I) {
1019 switch (I->getOpcode()) {
1020 default: assert (0 && "<Invalid operator> ");
1023 case Instruction::Ret: NumFastIselFailRet++; return;
1024 case Instruction::Br: NumFastIselFailBr++; return;
1025 case Instruction::Switch: NumFastIselFailSwitch++; return;
1026 case Instruction::IndirectBr: NumFastIselFailIndirectBr++; return;
1027 case Instruction::Invoke: NumFastIselFailInvoke++; return;
1028 case Instruction::Resume: NumFastIselFailResume++; return;
1029 case Instruction::Unreachable: NumFastIselFailUnreachable++; return;
1031 // Standard binary operators...
1032 case Instruction::Add: NumFastIselFailAdd++; return;
1033 case Instruction::FAdd: NumFastIselFailFAdd++; return;
1034 case Instruction::Sub: NumFastIselFailSub++; return;
1035 case Instruction::FSub: NumFastIselFailFSub++; return;
1036 case Instruction::Mul: NumFastIselFailMul++; return;
1037 case Instruction::FMul: NumFastIselFailFMul++; return;
1038 case Instruction::UDiv: NumFastIselFailUDiv++; return;
1039 case Instruction::SDiv: NumFastIselFailSDiv++; return;
1040 case Instruction::FDiv: NumFastIselFailFDiv++; return;
1041 case Instruction::URem: NumFastIselFailURem++; return;
1042 case Instruction::SRem: NumFastIselFailSRem++; return;
1043 case Instruction::FRem: NumFastIselFailFRem++; return;
1045 // Logical operators...
1046 case Instruction::And: NumFastIselFailAnd++; return;
1047 case Instruction::Or: NumFastIselFailOr++; return;
1048 case Instruction::Xor: NumFastIselFailXor++; return;
1050 // Memory instructions...
1051 case Instruction::Alloca: NumFastIselFailAlloca++; return;
1052 case Instruction::Load: NumFastIselFailLoad++; return;
1053 case Instruction::Store: NumFastIselFailStore++; return;
1054 case Instruction::AtomicCmpXchg: NumFastIselFailAtomicCmpXchg++; return;
1055 case Instruction::AtomicRMW: NumFastIselFailAtomicRMW++; return;
1056 case Instruction::Fence: NumFastIselFailFence++; return;
1057 case Instruction::GetElementPtr: NumFastIselFailGetElementPtr++; return;
1059 // Convert instructions...
1060 case Instruction::Trunc: NumFastIselFailTrunc++; return;
1061 case Instruction::ZExt: NumFastIselFailZExt++; return;
1062 case Instruction::SExt: NumFastIselFailSExt++; return;
1063 case Instruction::FPTrunc: NumFastIselFailFPTrunc++; return;
1064 case Instruction::FPExt: NumFastIselFailFPExt++; return;
1065 case Instruction::FPToUI: NumFastIselFailFPToUI++; return;
1066 case Instruction::FPToSI: NumFastIselFailFPToSI++; return;
1067 case Instruction::UIToFP: NumFastIselFailUIToFP++; return;
1068 case Instruction::SIToFP: NumFastIselFailSIToFP++; return;
1069 case Instruction::IntToPtr: NumFastIselFailIntToPtr++; return;
1070 case Instruction::PtrToInt: NumFastIselFailPtrToInt++; return;
1071 case Instruction::BitCast: NumFastIselFailBitCast++; return;
1073 // Other instructions...
1074 case Instruction::ICmp: NumFastIselFailICmp++; return;
1075 case Instruction::FCmp: NumFastIselFailFCmp++; return;
1076 case Instruction::PHI: NumFastIselFailPHI++; return;
1077 case Instruction::Select: NumFastIselFailSelect++; return;
1078 case Instruction::Call: {
1079 if (auto const *Intrinsic = dyn_cast<IntrinsicInst>(I)) {
1080 switch (Intrinsic->getIntrinsicID()) {
1082 NumFastIselFailIntrinsicCall++; return;
1083 case Intrinsic::sadd_with_overflow:
1084 NumFastIselFailSAddWithOverflow++; return;
1085 case Intrinsic::uadd_with_overflow:
1086 NumFastIselFailUAddWithOverflow++; return;
1087 case Intrinsic::ssub_with_overflow:
1088 NumFastIselFailSSubWithOverflow++; return;
1089 case Intrinsic::usub_with_overflow:
1090 NumFastIselFailUSubWithOverflow++; return;
1091 case Intrinsic::smul_with_overflow:
1092 NumFastIselFailSMulWithOverflow++; return;
1093 case Intrinsic::umul_with_overflow:
1094 NumFastIselFailUMulWithOverflow++; return;
1095 case Intrinsic::frameaddress:
1096 NumFastIselFailFrameaddress++; return;
1097 case Intrinsic::sqrt:
1098 NumFastIselFailSqrt++; return;
1099 case Intrinsic::experimental_stackmap:
1100 NumFastIselFailStackMap++; return;
1101 case Intrinsic::experimental_patchpoint_void: // fall-through
1102 case Intrinsic::experimental_patchpoint_i64:
1103 NumFastIselFailPatchPoint++; return;
1106 NumFastIselFailCall++;
1109 case Instruction::Shl: NumFastIselFailShl++; return;
1110 case Instruction::LShr: NumFastIselFailLShr++; return;
1111 case Instruction::AShr: NumFastIselFailAShr++; return;
1112 case Instruction::VAArg: NumFastIselFailVAArg++; return;
1113 case Instruction::ExtractElement: NumFastIselFailExtractElement++; return;
1114 case Instruction::InsertElement: NumFastIselFailInsertElement++; return;
1115 case Instruction::ShuffleVector: NumFastIselFailShuffleVector++; return;
1116 case Instruction::ExtractValue: NumFastIselFailExtractValue++; return;
1117 case Instruction::InsertValue: NumFastIselFailInsertValue++; return;
1118 case Instruction::LandingPad: NumFastIselFailLandingPad++; return;
1123 void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
1124 // Initialize the Fast-ISel state, if needed.
1125 FastISel *FastIS = nullptr;
1126 if (TM.Options.EnableFastISel)
1127 FastIS = TLI->createFastISel(*FuncInfo, LibInfo);
1129 // Iterate over all basic blocks in the function.
1130 ReversePostOrderTraversal<const Function*> RPOT(&Fn);
1131 for (ReversePostOrderTraversal<const Function*>::rpo_iterator
1132 I = RPOT.begin(), E = RPOT.end(); I != E; ++I) {
1133 const BasicBlock *LLVMBB = *I;
1135 if (OptLevel != CodeGenOpt::None) {
1136 bool AllPredsVisited = true;
1137 for (const_pred_iterator PI = pred_begin(LLVMBB), PE = pred_end(LLVMBB);
1139 if (!FuncInfo->VisitedBBs.count(*PI)) {
1140 AllPredsVisited = false;
1145 if (AllPredsVisited) {
1146 for (BasicBlock::const_iterator I = LLVMBB->begin();
1147 const PHINode *PN = dyn_cast<PHINode>(I); ++I)
1148 FuncInfo->ComputePHILiveOutRegInfo(PN);
1150 for (BasicBlock::const_iterator I = LLVMBB->begin();
1151 const PHINode *PN = dyn_cast<PHINode>(I); ++I)
1152 FuncInfo->InvalidatePHILiveOutRegInfo(PN);
1155 FuncInfo->VisitedBBs.insert(LLVMBB);
1158 BasicBlock::const_iterator const Begin = LLVMBB->getFirstNonPHI();
1159 BasicBlock::const_iterator const End = LLVMBB->end();
1160 BasicBlock::const_iterator BI = End;
1162 FuncInfo->MBB = FuncInfo->MBBMap[LLVMBB];
1163 FuncInfo->InsertPt = FuncInfo->MBB->getFirstNonPHI();
1165 // Setup an EH landing-pad block.
1166 FuncInfo->ExceptionPointerVirtReg = 0;
1167 FuncInfo->ExceptionSelectorVirtReg = 0;
1168 if (FuncInfo->MBB->isLandingPad())
1169 PrepareEHLandingPad();
1171 // Before doing SelectionDAG ISel, see if FastISel has been requested.
1173 FastIS->startNewBlock();
1175 // Emit code for any incoming arguments. This must happen before
1176 // beginning FastISel on the entry block.
1177 if (LLVMBB == &Fn.getEntryBlock()) {
1180 // Lower any arguments needed in this block if this is the entry block.
1181 if (!FastIS->lowerArguments()) {
1182 // Fast isel failed to lower these arguments
1183 ++NumFastIselFailLowerArguments;
1184 if (EnableFastISelAbort > 1)
1185 llvm_unreachable("FastISel didn't lower all arguments");
1187 // Use SelectionDAG argument lowering
1189 CurDAG->setRoot(SDB->getControlRoot());
1191 CodeGenAndEmitDAG();
1194 // If we inserted any instructions at the beginning, make a note of
1195 // where they are, so we can be sure to emit subsequent instructions
1197 if (FuncInfo->InsertPt != FuncInfo->MBB->begin())
1198 FastIS->setLastLocalValue(std::prev(FuncInfo->InsertPt));
1200 FastIS->setLastLocalValue(nullptr);
1203 unsigned NumFastIselRemaining = std::distance(Begin, End);
1204 // Do FastISel on as many instructions as possible.
1205 for (; BI != Begin; --BI) {
1206 const Instruction *Inst = std::prev(BI);
1208 // If we no longer require this instruction, skip it.
1209 if (isFoldedOrDeadInstruction(Inst, FuncInfo)) {
1210 --NumFastIselRemaining;
1214 // Bottom-up: reset the insert pos at the top, after any local-value
1216 FastIS->recomputeInsertPt();
1218 // Try to select the instruction with FastISel.
1219 if (FastIS->selectInstruction(Inst)) {
1220 --NumFastIselRemaining;
1221 ++NumFastIselSuccess;
1222 // If fast isel succeeded, skip over all the folded instructions, and
1223 // then see if there is a load right before the selected instructions.
1224 // Try to fold the load if so.
1225 const Instruction *BeforeInst = Inst;
1226 while (BeforeInst != Begin) {
1227 BeforeInst = std::prev(BasicBlock::const_iterator(BeforeInst));
1228 if (!isFoldedOrDeadInstruction(BeforeInst, FuncInfo))
1231 if (BeforeInst != Inst && isa<LoadInst>(BeforeInst) &&
1232 BeforeInst->hasOneUse() &&
1233 FastIS->tryToFoldLoad(cast<LoadInst>(BeforeInst), Inst)) {
1234 // If we succeeded, don't re-select the load.
1235 BI = std::next(BasicBlock::const_iterator(BeforeInst));
1236 --NumFastIselRemaining;
1237 ++NumFastIselSuccess;
1243 if (EnableFastISelVerbose2)
1244 collectFailStats(Inst);
1247 // Then handle certain instructions as single-LLVM-Instruction blocks.
1248 if (isa<CallInst>(Inst)) {
1250 if (EnableFastISelVerbose || EnableFastISelAbort) {
1251 dbgs() << "FastISel missed call: ";
1254 if (EnableFastISelAbort > 2)
1255 // FastISel selector couldn't handle something and bailed.
1256 // For the purpose of debugging, just abort.
1257 llvm_unreachable("FastISel didn't select the entire block");
1259 if (!Inst->getType()->isVoidTy() && !Inst->use_empty()) {
1260 unsigned &R = FuncInfo->ValueMap[Inst];
1262 R = FuncInfo->CreateRegs(Inst->getType());
1265 bool HadTailCall = false;
1266 MachineBasicBlock::iterator SavedInsertPt = FuncInfo->InsertPt;
1267 SelectBasicBlock(Inst, BI, HadTailCall);
1269 // If the call was emitted as a tail call, we're done with the block.
1270 // We also need to delete any previously emitted instructions.
1272 FastIS->removeDeadCode(SavedInsertPt, FuncInfo->MBB->end());
1277 // Recompute NumFastIselRemaining as Selection DAG instruction
1278 // selection may have handled the call, input args, etc.
1279 unsigned RemainingNow = std::distance(Begin, BI);
1280 NumFastIselFailures += NumFastIselRemaining - RemainingNow;
1281 NumFastIselRemaining = RemainingNow;
1285 if (EnableFastISelVerbose || EnableFastISelAbort) {
1286 if (isa<TerminatorInst>(Inst)) {
1287 // Use a different message for terminator misses.
1288 dbgs() << "FastISel missed terminator: ";
1290 dbgs() << "FastISel miss: ";
1294 if (EnableFastISelAbort > 2)
1295 // FastISel selector couldn't handle something and bailed.
1296 // For the purpose of debugging, just abort.
1297 llvm_unreachable("FastISel didn't select the entire block");
1299 NumFastIselFailures += NumFastIselRemaining;
1303 FastIS->recomputeInsertPt();
1305 // Lower any arguments needed in this block if this is the entry block.
1306 if (LLVMBB == &Fn.getEntryBlock()) {
1315 ++NumFastIselBlocks;
1318 // Run SelectionDAG instruction selection on the remainder of the block
1319 // not handled by FastISel. If FastISel is not run, this is the entire
1322 SelectBasicBlock(Begin, BI, HadTailCall);
1326 FuncInfo->PHINodesToUpdate.clear();
1330 SDB->clearDanglingDebugInfo();
1331 SDB->SPDescriptor.resetPerFunctionState();
1334 /// Given that the input MI is before a partial terminator sequence TSeq, return
1335 /// true if M + TSeq also a partial terminator sequence.
1337 /// A Terminator sequence is a sequence of MachineInstrs which at this point in
1338 /// lowering copy vregs into physical registers, which are then passed into
1339 /// terminator instructors so we can satisfy ABI constraints. A partial
1340 /// terminator sequence is an improper subset of a terminator sequence (i.e. it
1341 /// may be the whole terminator sequence).
1342 static bool MIIsInTerminatorSequence(const MachineInstr *MI) {
1343 // If we do not have a copy or an implicit def, we return true if and only if
1344 // MI is a debug value.
1345 if (!MI->isCopy() && !MI->isImplicitDef())
1346 // Sometimes DBG_VALUE MI sneak in between the copies from the vregs to the
1347 // physical registers if there is debug info associated with the terminator
1348 // of our mbb. We want to include said debug info in our terminator
1349 // sequence, so we return true in that case.
1350 return MI->isDebugValue();
1352 // We have left the terminator sequence if we are not doing one of the
1355 // 1. Copying a vreg into a physical register.
1356 // 2. Copying a vreg into a vreg.
1357 // 3. Defining a register via an implicit def.
1359 // OPI should always be a register definition...
1360 MachineInstr::const_mop_iterator OPI = MI->operands_begin();
1361 if (!OPI->isReg() || !OPI->isDef())
1364 // Defining any register via an implicit def is always ok.
1365 if (MI->isImplicitDef())
1368 // Grab the copy source...
1369 MachineInstr::const_mop_iterator OPI2 = OPI;
1371 assert(OPI2 != MI->operands_end()
1372 && "Should have a copy implying we should have 2 arguments.");
1374 // Make sure that the copy dest is not a vreg when the copy source is a
1375 // physical register.
1376 if (!OPI2->isReg() ||
1377 (!TargetRegisterInfo::isPhysicalRegister(OPI->getReg()) &&
1378 TargetRegisterInfo::isPhysicalRegister(OPI2->getReg())))
1384 /// Find the split point at which to splice the end of BB into its success stack
1385 /// protector check machine basic block.
1387 /// On many platforms, due to ABI constraints, terminators, even before register
1388 /// allocation, use physical registers. This creates an issue for us since
1389 /// physical registers at this point can not travel across basic
1390 /// blocks. Luckily, selectiondag always moves physical registers into vregs
1391 /// when they enter functions and moves them through a sequence of copies back
1392 /// into the physical registers right before the terminator creating a
1393 /// ``Terminator Sequence''. This function is searching for the beginning of the
1394 /// terminator sequence so that we can ensure that we splice off not just the
1395 /// terminator, but additionally the copies that move the vregs into the
1396 /// physical registers.
1397 static MachineBasicBlock::iterator
1398 FindSplitPointForStackProtector(MachineBasicBlock *BB, DebugLoc DL) {
1399 MachineBasicBlock::iterator SplitPoint = BB->getFirstTerminator();
1401 if (SplitPoint == BB->begin())
1404 MachineBasicBlock::iterator Start = BB->begin();
1405 MachineBasicBlock::iterator Previous = SplitPoint;
1408 while (MIIsInTerminatorSequence(Previous)) {
1409 SplitPoint = Previous;
1410 if (Previous == Start)
1419 SelectionDAGISel::FinishBasicBlock() {
1421 DEBUG(dbgs() << "Total amount of phi nodes to update: "
1422 << FuncInfo->PHINodesToUpdate.size() << "\n";
1423 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i)
1424 dbgs() << "Node " << i << " : ("
1425 << FuncInfo->PHINodesToUpdate[i].first
1426 << ", " << FuncInfo->PHINodesToUpdate[i].second << ")\n");
1428 const bool MustUpdatePHINodes = SDB->SwitchCases.empty() &&
1429 SDB->JTCases.empty() &&
1430 SDB->BitTestCases.empty();
1432 // Next, now that we know what the last MBB the LLVM BB expanded is, update
1433 // PHI nodes in successors.
1434 if (MustUpdatePHINodes) {
1435 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i) {
1436 MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[i].first);
1437 assert(PHI->isPHI() &&
1438 "This is not a machine PHI node that we are updating!");
1439 if (!FuncInfo->MBB->isSuccessor(PHI->getParent()))
1441 PHI.addReg(FuncInfo->PHINodesToUpdate[i].second).addMBB(FuncInfo->MBB);
1445 // Handle stack protector.
1446 if (SDB->SPDescriptor.shouldEmitStackProtector()) {
1447 MachineBasicBlock *ParentMBB = SDB->SPDescriptor.getParentMBB();
1448 MachineBasicBlock *SuccessMBB = SDB->SPDescriptor.getSuccessMBB();
1450 // Find the split point to split the parent mbb. At the same time copy all
1451 // physical registers used in the tail of parent mbb into virtual registers
1452 // before the split point and back into physical registers after the split
1453 // point. This prevents us needing to deal with Live-ins and many other
1454 // register allocation issues caused by us splitting the parent mbb. The
1455 // register allocator will clean up said virtual copies later on.
1456 MachineBasicBlock::iterator SplitPoint =
1457 FindSplitPointForStackProtector(ParentMBB, SDB->getCurDebugLoc());
1459 // Splice the terminator of ParentMBB into SuccessMBB.
1460 SuccessMBB->splice(SuccessMBB->end(), ParentMBB,
1464 // Add compare/jump on neq/jump to the parent BB.
1465 FuncInfo->MBB = ParentMBB;
1466 FuncInfo->InsertPt = ParentMBB->end();
1467 SDB->visitSPDescriptorParent(SDB->SPDescriptor, ParentMBB);
1468 CurDAG->setRoot(SDB->getRoot());
1470 CodeGenAndEmitDAG();
1472 // CodeGen Failure MBB if we have not codegened it yet.
1473 MachineBasicBlock *FailureMBB = SDB->SPDescriptor.getFailureMBB();
1474 if (!FailureMBB->size()) {
1475 FuncInfo->MBB = FailureMBB;
1476 FuncInfo->InsertPt = FailureMBB->end();
1477 SDB->visitSPDescriptorFailure(SDB->SPDescriptor);
1478 CurDAG->setRoot(SDB->getRoot());
1480 CodeGenAndEmitDAG();
1483 // Clear the Per-BB State.
1484 SDB->SPDescriptor.resetPerBBState();
1487 // If we updated PHI Nodes, return early.
1488 if (MustUpdatePHINodes)
1491 for (unsigned i = 0, e = SDB->BitTestCases.size(); i != e; ++i) {
1492 // Lower header first, if it wasn't already lowered
1493 if (!SDB->BitTestCases[i].Emitted) {
1494 // Set the current basic block to the mbb we wish to insert the code into
1495 FuncInfo->MBB = SDB->BitTestCases[i].Parent;
1496 FuncInfo->InsertPt = FuncInfo->MBB->end();
1498 SDB->visitBitTestHeader(SDB->BitTestCases[i], FuncInfo->MBB);
1499 CurDAG->setRoot(SDB->getRoot());
1501 CodeGenAndEmitDAG();
1504 uint32_t UnhandledWeight = 0;
1505 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j)
1506 UnhandledWeight += SDB->BitTestCases[i].Cases[j].ExtraWeight;
1508 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j) {
1509 UnhandledWeight -= SDB->BitTestCases[i].Cases[j].ExtraWeight;
1510 // Set the current basic block to the mbb we wish to insert the code into
1511 FuncInfo->MBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1512 FuncInfo->InsertPt = FuncInfo->MBB->end();
1515 SDB->visitBitTestCase(SDB->BitTestCases[i],
1516 SDB->BitTestCases[i].Cases[j+1].ThisBB,
1518 SDB->BitTestCases[i].Reg,
1519 SDB->BitTestCases[i].Cases[j],
1522 SDB->visitBitTestCase(SDB->BitTestCases[i],
1523 SDB->BitTestCases[i].Default,
1525 SDB->BitTestCases[i].Reg,
1526 SDB->BitTestCases[i].Cases[j],
1530 CurDAG->setRoot(SDB->getRoot());
1532 CodeGenAndEmitDAG();
1536 for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1538 MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[pi].first);
1539 MachineBasicBlock *PHIBB = PHI->getParent();
1540 assert(PHI->isPHI() &&
1541 "This is not a machine PHI node that we are updating!");
1542 // This is "default" BB. We have two jumps to it. From "header" BB and
1543 // from last "case" BB.
1544 if (PHIBB == SDB->BitTestCases[i].Default)
1545 PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second)
1546 .addMBB(SDB->BitTestCases[i].Parent)
1547 .addReg(FuncInfo->PHINodesToUpdate[pi].second)
1548 .addMBB(SDB->BitTestCases[i].Cases.back().ThisBB);
1549 // One of "cases" BB.
1550 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size();
1552 MachineBasicBlock* cBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1553 if (cBB->isSuccessor(PHIBB))
1554 PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second).addMBB(cBB);
1558 SDB->BitTestCases.clear();
1560 // If the JumpTable record is filled in, then we need to emit a jump table.
1561 // Updating the PHI nodes is tricky in this case, since we need to determine
1562 // whether the PHI is a successor of the range check MBB or the jump table MBB
1563 for (unsigned i = 0, e = SDB->JTCases.size(); i != e; ++i) {
1564 // Lower header first, if it wasn't already lowered
1565 if (!SDB->JTCases[i].first.Emitted) {
1566 // Set the current basic block to the mbb we wish to insert the code into
1567 FuncInfo->MBB = SDB->JTCases[i].first.HeaderBB;
1568 FuncInfo->InsertPt = FuncInfo->MBB->end();
1570 SDB->visitJumpTableHeader(SDB->JTCases[i].second, SDB->JTCases[i].first,
1572 CurDAG->setRoot(SDB->getRoot());
1574 CodeGenAndEmitDAG();
1577 // Set the current basic block to the mbb we wish to insert the code into
1578 FuncInfo->MBB = SDB->JTCases[i].second.MBB;
1579 FuncInfo->InsertPt = FuncInfo->MBB->end();
1581 SDB->visitJumpTable(SDB->JTCases[i].second);
1582 CurDAG->setRoot(SDB->getRoot());
1584 CodeGenAndEmitDAG();
1587 for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1589 MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[pi].first);
1590 MachineBasicBlock *PHIBB = PHI->getParent();
1591 assert(PHI->isPHI() &&
1592 "This is not a machine PHI node that we are updating!");
1593 // "default" BB. We can go there only from header BB.
1594 if (PHIBB == SDB->JTCases[i].second.Default)
1595 PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second)
1596 .addMBB(SDB->JTCases[i].first.HeaderBB);
1597 // JT BB. Just iterate over successors here
1598 if (FuncInfo->MBB->isSuccessor(PHIBB))
1599 PHI.addReg(FuncInfo->PHINodesToUpdate[pi].second).addMBB(FuncInfo->MBB);
1602 SDB->JTCases.clear();
1604 // If the switch block involved a branch to one of the actual successors, we
1605 // need to update PHI nodes in that block.
1606 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i) {
1607 MachineInstrBuilder PHI(*MF, FuncInfo->PHINodesToUpdate[i].first);
1608 assert(PHI->isPHI() &&
1609 "This is not a machine PHI node that we are updating!");
1610 if (FuncInfo->MBB->isSuccessor(PHI->getParent()))
1611 PHI.addReg(FuncInfo->PHINodesToUpdate[i].second).addMBB(FuncInfo->MBB);
1614 // If we generated any switch lowering information, build and codegen any
1615 // additional DAGs necessary.
1616 for (unsigned i = 0, e = SDB->SwitchCases.size(); i != e; ++i) {
1617 // Set the current basic block to the mbb we wish to insert the code into
1618 FuncInfo->MBB = SDB->SwitchCases[i].ThisBB;
1619 FuncInfo->InsertPt = FuncInfo->MBB->end();
1621 // Determine the unique successors.
1622 SmallVector<MachineBasicBlock *, 2> Succs;
1623 Succs.push_back(SDB->SwitchCases[i].TrueBB);
1624 if (SDB->SwitchCases[i].TrueBB != SDB->SwitchCases[i].FalseBB)
1625 Succs.push_back(SDB->SwitchCases[i].FalseBB);
1627 // Emit the code. Note that this could result in FuncInfo->MBB being split.
1628 SDB->visitSwitchCase(SDB->SwitchCases[i], FuncInfo->MBB);
1629 CurDAG->setRoot(SDB->getRoot());
1631 CodeGenAndEmitDAG();
1633 // Remember the last block, now that any splitting is done, for use in
1634 // populating PHI nodes in successors.
1635 MachineBasicBlock *ThisBB = FuncInfo->MBB;
1637 // Handle any PHI nodes in successors of this chunk, as if we were coming
1638 // from the original BB before switch expansion. Note that PHI nodes can
1639 // occur multiple times in PHINodesToUpdate. We have to be very careful to
1640 // handle them the right number of times.
1641 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1642 FuncInfo->MBB = Succs[i];
1643 FuncInfo->InsertPt = FuncInfo->MBB->end();
1644 // FuncInfo->MBB may have been removed from the CFG if a branch was
1646 if (ThisBB->isSuccessor(FuncInfo->MBB)) {
1647 for (MachineBasicBlock::iterator
1648 MBBI = FuncInfo->MBB->begin(), MBBE = FuncInfo->MBB->end();
1649 MBBI != MBBE && MBBI->isPHI(); ++MBBI) {
1650 MachineInstrBuilder PHI(*MF, MBBI);
1651 // This value for this PHI node is recorded in PHINodesToUpdate.
1652 for (unsigned pn = 0; ; ++pn) {
1653 assert(pn != FuncInfo->PHINodesToUpdate.size() &&
1654 "Didn't find PHI entry!");
1655 if (FuncInfo->PHINodesToUpdate[pn].first == PHI) {
1656 PHI.addReg(FuncInfo->PHINodesToUpdate[pn].second).addMBB(ThisBB);
1664 SDB->SwitchCases.clear();
1668 /// Create the scheduler. If a specific scheduler was specified
1669 /// via the SchedulerRegistry, use it, otherwise select the
1670 /// one preferred by the target.
1672 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
1673 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1677 RegisterScheduler::setDefault(Ctor);
1680 return Ctor(this, OptLevel);
1683 //===----------------------------------------------------------------------===//
1684 // Helper functions used by the generated instruction selector.
1685 //===----------------------------------------------------------------------===//
1686 // Calls to these methods are generated by tblgen.
1688 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1689 /// the dag combiner simplified the 255, we still want to match. RHS is the
1690 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1691 /// specified in the .td file (e.g. 255).
1692 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1693 int64_t DesiredMaskS) const {
1694 const APInt &ActualMask = RHS->getAPIntValue();
1695 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1697 // If the actual mask exactly matches, success!
1698 if (ActualMask == DesiredMask)
1701 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1702 if (ActualMask.intersects(~DesiredMask))
1705 // Otherwise, the DAG Combiner may have proven that the value coming in is
1706 // either already zero or is not demanded. Check for known zero input bits.
1707 APInt NeededMask = DesiredMask & ~ActualMask;
1708 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1711 // TODO: check to see if missing bits are just not demanded.
1713 // Otherwise, this pattern doesn't match.
1717 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1718 /// the dag combiner simplified the 255, we still want to match. RHS is the
1719 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1720 /// specified in the .td file (e.g. 255).
1721 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1722 int64_t DesiredMaskS) const {
1723 const APInt &ActualMask = RHS->getAPIntValue();
1724 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1726 // If the actual mask exactly matches, success!
1727 if (ActualMask == DesiredMask)
1730 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1731 if (ActualMask.intersects(~DesiredMask))
1734 // Otherwise, the DAG Combiner may have proven that the value coming in is
1735 // either already zero or is not demanded. Check for known zero input bits.
1736 APInt NeededMask = DesiredMask & ~ActualMask;
1738 APInt KnownZero, KnownOne;
1739 CurDAG->computeKnownBits(LHS, KnownZero, KnownOne);
1741 // If all the missing bits in the or are already known to be set, match!
1742 if ((NeededMask & KnownOne) == NeededMask)
1745 // TODO: check to see if missing bits are just not demanded.
1747 // Otherwise, this pattern doesn't match.
1752 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1753 /// by tblgen. Others should not call it.
1754 void SelectionDAGISel::
1755 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1756 std::vector<SDValue> InOps;
1757 std::swap(InOps, Ops);
1759 Ops.push_back(InOps[InlineAsm::Op_InputChain]); // 0
1760 Ops.push_back(InOps[InlineAsm::Op_AsmString]); // 1
1761 Ops.push_back(InOps[InlineAsm::Op_MDNode]); // 2, !srcloc
1762 Ops.push_back(InOps[InlineAsm::Op_ExtraInfo]); // 3 (SideEffect, AlignStack)
1764 unsigned i = InlineAsm::Op_FirstOperand, e = InOps.size();
1765 if (InOps[e-1].getValueType() == MVT::Glue)
1766 --e; // Don't process a glue operand if it is here.
1769 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1770 if (!InlineAsm::isMemKind(Flags)) {
1771 // Just skip over this operand, copying the operands verbatim.
1772 Ops.insert(Ops.end(), InOps.begin()+i,
1773 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
1774 i += InlineAsm::getNumOperandRegisters(Flags) + 1;
1776 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 &&
1777 "Memory operand with multiple values?");
1778 // Otherwise, this is a memory operand. Ask the target to select it.
1779 std::vector<SDValue> SelOps;
1780 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps))
1781 report_fatal_error("Could not match memory address. Inline asm"
1784 // Add this to the output node.
1786 InlineAsm::getFlagWord(InlineAsm::Kind_Mem, SelOps.size());
1787 Ops.push_back(CurDAG->getTargetConstant(NewFlags, MVT::i32));
1788 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1793 // Add the glue input back if present.
1794 if (e != InOps.size())
1795 Ops.push_back(InOps.back());
1798 /// findGlueUse - Return use of MVT::Glue value produced by the specified
1801 static SDNode *findGlueUse(SDNode *N) {
1802 unsigned FlagResNo = N->getNumValues()-1;
1803 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1804 SDUse &Use = I.getUse();
1805 if (Use.getResNo() == FlagResNo)
1806 return Use.getUser();
1811 /// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
1812 /// This function recursively traverses up the operand chain, ignoring
1814 static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
1815 SDNode *Root, SmallPtrSetImpl<SDNode*> &Visited,
1816 bool IgnoreChains) {
1817 // The NodeID's are given uniques ID's where a node ID is guaranteed to be
1818 // greater than all of its (recursive) operands. If we scan to a point where
1819 // 'use' is smaller than the node we're scanning for, then we know we will
1822 // The Use may be -1 (unassigned) if it is a newly allocated node. This can
1823 // happen because we scan down to newly selected nodes in the case of glue
1825 if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1))
1828 // Don't revisit nodes if we already scanned it and didn't fail, we know we
1829 // won't fail if we scan it again.
1830 if (!Visited.insert(Use).second)
1833 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
1834 // Ignore chain uses, they are validated by HandleMergeInputChains.
1835 if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains)
1838 SDNode *N = Use->getOperand(i).getNode();
1840 if (Use == ImmedUse || Use == Root)
1841 continue; // We are not looking for immediate use.
1846 // Traverse up the operand chain.
1847 if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
1853 /// IsProfitableToFold - Returns true if it's profitable to fold the specific
1854 /// operand node N of U during instruction selection that starts at Root.
1855 bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
1856 SDNode *Root) const {
1857 if (OptLevel == CodeGenOpt::None) return false;
1858 return N.hasOneUse();
1861 /// IsLegalToFold - Returns true if the specific operand node N of
1862 /// U can be folded during instruction selection that starts at Root.
1863 bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
1864 CodeGenOpt::Level OptLevel,
1865 bool IgnoreChains) {
1866 if (OptLevel == CodeGenOpt::None) return false;
1868 // If Root use can somehow reach N through a path that that doesn't contain
1869 // U then folding N would create a cycle. e.g. In the following
1870 // diagram, Root can reach N through X. If N is folded into into Root, then
1871 // X is both a predecessor and a successor of U.
1882 // * indicates nodes to be folded together.
1884 // If Root produces glue, then it gets (even more) interesting. Since it
1885 // will be "glued" together with its glue use in the scheduler, we need to
1886 // check if it might reach N.
1905 // If GU (glue use) indirectly reaches N (the load), and Root folds N
1906 // (call it Fold), then X is a predecessor of GU and a successor of
1907 // Fold. But since Fold and GU are glued together, this will create
1908 // a cycle in the scheduling graph.
1910 // If the node has glue, walk down the graph to the "lowest" node in the
1912 EVT VT = Root->getValueType(Root->getNumValues()-1);
1913 while (VT == MVT::Glue) {
1914 SDNode *GU = findGlueUse(Root);
1918 VT = Root->getValueType(Root->getNumValues()-1);
1920 // If our query node has a glue result with a use, we've walked up it. If
1921 // the user (which has already been selected) has a chain or indirectly uses
1922 // the chain, our WalkChainUsers predicate will not consider it. Because of
1923 // this, we cannot ignore chains in this predicate.
1924 IgnoreChains = false;
1928 SmallPtrSet<SDNode*, 16> Visited;
1929 return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains);
1932 SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
1933 std::vector<SDValue> Ops(N->op_begin(), N->op_end());
1934 SelectInlineAsmMemoryOperands(Ops);
1936 EVT VTs[] = { MVT::Other, MVT::Glue };
1937 SDValue New = CurDAG->getNode(ISD::INLINEASM, SDLoc(N), VTs, Ops);
1939 return New.getNode();
1943 *SelectionDAGISel::Select_READ_REGISTER(SDNode *Op) {
1945 MDNodeSDNode *MD = dyn_cast<MDNodeSDNode>(Op->getOperand(0));
1946 const MDString *RegStr = dyn_cast<MDString>(MD->getMD()->getOperand(0));
1948 TLI->getRegisterByName(RegStr->getString().data(), Op->getValueType(0));
1949 SDValue New = CurDAG->getCopyFromReg(
1950 CurDAG->getEntryNode(), dl, Reg, Op->getValueType(0));
1952 return New.getNode();
1956 *SelectionDAGISel::Select_WRITE_REGISTER(SDNode *Op) {
1958 MDNodeSDNode *MD = dyn_cast<MDNodeSDNode>(Op->getOperand(1));
1959 const MDString *RegStr = dyn_cast<MDString>(MD->getMD()->getOperand(0));
1960 unsigned Reg = TLI->getRegisterByName(RegStr->getString().data(),
1961 Op->getOperand(2).getValueType());
1962 SDValue New = CurDAG->getCopyToReg(
1963 CurDAG->getEntryNode(), dl, Reg, Op->getOperand(2));
1965 return New.getNode();
1970 SDNode *SelectionDAGISel::Select_UNDEF(SDNode *N) {
1971 return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0));
1974 /// GetVBR - decode a vbr encoding whose top bit is set.
1975 LLVM_ATTRIBUTE_ALWAYS_INLINE static uint64_t
1976 GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
1977 assert(Val >= 128 && "Not a VBR");
1978 Val &= 127; // Remove first vbr bit.
1983 NextBits = MatcherTable[Idx++];
1984 Val |= (NextBits&127) << Shift;
1986 } while (NextBits & 128);
1992 /// UpdateChainsAndGlue - When a match is complete, this method updates uses of
1993 /// interior glue and chain results to use the new glue and chain results.
1994 void SelectionDAGISel::
1995 UpdateChainsAndGlue(SDNode *NodeToMatch, SDValue InputChain,
1996 const SmallVectorImpl<SDNode*> &ChainNodesMatched,
1998 const SmallVectorImpl<SDNode*> &GlueResultNodesMatched,
1999 bool isMorphNodeTo) {
2000 SmallVector<SDNode*, 4> NowDeadNodes;
2002 // Now that all the normal results are replaced, we replace the chain and
2003 // glue results if present.
2004 if (!ChainNodesMatched.empty()) {
2005 assert(InputChain.getNode() &&
2006 "Matched input chains but didn't produce a chain");
2007 // Loop over all of the nodes we matched that produced a chain result.
2008 // Replace all the chain results with the final chain we ended up with.
2009 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
2010 SDNode *ChainNode = ChainNodesMatched[i];
2012 // If this node was already deleted, don't look at it.
2013 if (ChainNode->getOpcode() == ISD::DELETED_NODE)
2016 // Don't replace the results of the root node if we're doing a
2018 if (ChainNode == NodeToMatch && isMorphNodeTo)
2021 SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1);
2022 if (ChainVal.getValueType() == MVT::Glue)
2023 ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
2024 assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
2025 CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain);
2027 // If the node became dead and we haven't already seen it, delete it.
2028 if (ChainNode->use_empty() &&
2029 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), ChainNode))
2030 NowDeadNodes.push_back(ChainNode);
2034 // If the result produces glue, update any glue results in the matched
2035 // pattern with the glue result.
2036 if (InputGlue.getNode()) {
2037 // Handle any interior nodes explicitly marked.
2038 for (unsigned i = 0, e = GlueResultNodesMatched.size(); i != e; ++i) {
2039 SDNode *FRN = GlueResultNodesMatched[i];
2041 // If this node was already deleted, don't look at it.
2042 if (FRN->getOpcode() == ISD::DELETED_NODE)
2045 assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Glue &&
2046 "Doesn't have a glue result");
2047 CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1),
2050 // If the node became dead and we haven't already seen it, delete it.
2051 if (FRN->use_empty() &&
2052 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), FRN))
2053 NowDeadNodes.push_back(FRN);
2057 if (!NowDeadNodes.empty())
2058 CurDAG->RemoveDeadNodes(NowDeadNodes);
2060 DEBUG(dbgs() << "ISEL: Match complete!\n");
2066 CR_LeadsToInteriorNode
2069 /// WalkChainUsers - Walk down the users of the specified chained node that is
2070 /// part of the pattern we're matching, looking at all of the users we find.
2071 /// This determines whether something is an interior node, whether we have a
2072 /// non-pattern node in between two pattern nodes (which prevent folding because
2073 /// it would induce a cycle) and whether we have a TokenFactor node sandwiched
2074 /// between pattern nodes (in which case the TF becomes part of the pattern).
2076 /// The walk we do here is guaranteed to be small because we quickly get down to
2077 /// already selected nodes "below" us.
2079 WalkChainUsers(const SDNode *ChainedNode,
2080 SmallVectorImpl<SDNode*> &ChainedNodesInPattern,
2081 SmallVectorImpl<SDNode*> &InteriorChainedNodes) {
2082 ChainResult Result = CR_Simple;
2084 for (SDNode::use_iterator UI = ChainedNode->use_begin(),
2085 E = ChainedNode->use_end(); UI != E; ++UI) {
2086 // Make sure the use is of the chain, not some other value we produce.
2087 if (UI.getUse().getValueType() != MVT::Other) continue;
2091 if (User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
2094 // If we see an already-selected machine node, then we've gone beyond the
2095 // pattern that we're selecting down into the already selected chunk of the
2097 unsigned UserOpcode = User->getOpcode();
2098 if (User->isMachineOpcode() ||
2099 UserOpcode == ISD::CopyToReg ||
2100 UserOpcode == ISD::CopyFromReg ||
2101 UserOpcode == ISD::INLINEASM ||
2102 UserOpcode == ISD::EH_LABEL ||
2103 UserOpcode == ISD::LIFETIME_START ||
2104 UserOpcode == ISD::LIFETIME_END) {
2105 // If their node ID got reset to -1 then they've already been selected.
2106 // Treat them like a MachineOpcode.
2107 if (User->getNodeId() == -1)
2111 // If we have a TokenFactor, we handle it specially.
2112 if (User->getOpcode() != ISD::TokenFactor) {
2113 // If the node isn't a token factor and isn't part of our pattern, then it
2114 // must be a random chained node in between two nodes we're selecting.
2115 // This happens when we have something like:
2120 // Because we structurally match the load/store as a read/modify/write,
2121 // but the call is chained between them. We cannot fold in this case
2122 // because it would induce a cycle in the graph.
2123 if (!std::count(ChainedNodesInPattern.begin(),
2124 ChainedNodesInPattern.end(), User))
2125 return CR_InducesCycle;
2127 // Otherwise we found a node that is part of our pattern. For example in:
2131 // This would happen when we're scanning down from the load and see the
2132 // store as a user. Record that there is a use of ChainedNode that is
2133 // part of the pattern and keep scanning uses.
2134 Result = CR_LeadsToInteriorNode;
2135 InteriorChainedNodes.push_back(User);
2139 // If we found a TokenFactor, there are two cases to consider: first if the
2140 // TokenFactor is just hanging "below" the pattern we're matching (i.e. no
2141 // uses of the TF are in our pattern) we just want to ignore it. Second,
2142 // the TokenFactor can be sandwiched in between two chained nodes, like so:
2148 // | \ DAG's like cheese
2151 // [TokenFactor] [Op]
2158 // In this case, the TokenFactor becomes part of our match and we rewrite it
2159 // as a new TokenFactor.
2161 // To distinguish these two cases, do a recursive walk down the uses.
2162 switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) {
2164 // If the uses of the TokenFactor are just already-selected nodes, ignore
2165 // it, it is "below" our pattern.
2167 case CR_InducesCycle:
2168 // If the uses of the TokenFactor lead to nodes that are not part of our
2169 // pattern that are not selected, folding would turn this into a cycle,
2171 return CR_InducesCycle;
2172 case CR_LeadsToInteriorNode:
2173 break; // Otherwise, keep processing.
2176 // Okay, we know we're in the interesting interior case. The TokenFactor
2177 // is now going to be considered part of the pattern so that we rewrite its
2178 // uses (it may have uses that are not part of the pattern) with the
2179 // ultimate chain result of the generated code. We will also add its chain
2180 // inputs as inputs to the ultimate TokenFactor we create.
2181 Result = CR_LeadsToInteriorNode;
2182 ChainedNodesInPattern.push_back(User);
2183 InteriorChainedNodes.push_back(User);
2190 /// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
2191 /// operation for when the pattern matched at least one node with a chains. The
2192 /// input vector contains a list of all of the chained nodes that we match. We
2193 /// must determine if this is a valid thing to cover (i.e. matching it won't
2194 /// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
2195 /// be used as the input node chain for the generated nodes.
2197 HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
2198 SelectionDAG *CurDAG) {
2199 // Walk all of the chained nodes we've matched, recursively scanning down the
2200 // users of the chain result. This adds any TokenFactor nodes that are caught
2201 // in between chained nodes to the chained and interior nodes list.
2202 SmallVector<SDNode*, 3> InteriorChainedNodes;
2203 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
2204 if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched,
2205 InteriorChainedNodes) == CR_InducesCycle)
2206 return SDValue(); // Would induce a cycle.
2209 // Okay, we have walked all the matched nodes and collected TokenFactor nodes
2210 // that we are interested in. Form our input TokenFactor node.
2211 SmallVector<SDValue, 3> InputChains;
2212 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
2213 // Add the input chain of this node to the InputChains list (which will be
2214 // the operands of the generated TokenFactor) if it's not an interior node.
2215 SDNode *N = ChainNodesMatched[i];
2216 if (N->getOpcode() != ISD::TokenFactor) {
2217 if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N))
2220 // Otherwise, add the input chain.
2221 SDValue InChain = ChainNodesMatched[i]->getOperand(0);
2222 assert(InChain.getValueType() == MVT::Other && "Not a chain");
2223 InputChains.push_back(InChain);
2227 // If we have a token factor, we want to add all inputs of the token factor
2228 // that are not part of the pattern we're matching.
2229 for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) {
2230 if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(),
2231 N->getOperand(op).getNode()))
2232 InputChains.push_back(N->getOperand(op));
2236 if (InputChains.size() == 1)
2237 return InputChains[0];
2238 return CurDAG->getNode(ISD::TokenFactor, SDLoc(ChainNodesMatched[0]),
2239 MVT::Other, InputChains);
2242 /// MorphNode - Handle morphing a node in place for the selector.
2243 SDNode *SelectionDAGISel::
2244 MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList,
2245 ArrayRef<SDValue> Ops, unsigned EmitNodeInfo) {
2246 // It is possible we're using MorphNodeTo to replace a node with no
2247 // normal results with one that has a normal result (or we could be
2248 // adding a chain) and the input could have glue and chains as well.
2249 // In this case we need to shift the operands down.
2250 // FIXME: This is a horrible hack and broken in obscure cases, no worse
2251 // than the old isel though.
2252 int OldGlueResultNo = -1, OldChainResultNo = -1;
2254 unsigned NTMNumResults = Node->getNumValues();
2255 if (Node->getValueType(NTMNumResults-1) == MVT::Glue) {
2256 OldGlueResultNo = NTMNumResults-1;
2257 if (NTMNumResults != 1 &&
2258 Node->getValueType(NTMNumResults-2) == MVT::Other)
2259 OldChainResultNo = NTMNumResults-2;
2260 } else if (Node->getValueType(NTMNumResults-1) == MVT::Other)
2261 OldChainResultNo = NTMNumResults-1;
2263 // Call the underlying SelectionDAG routine to do the transmogrification. Note
2264 // that this deletes operands of the old node that become dead.
2265 SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops);
2267 // MorphNodeTo can operate in two ways: if an existing node with the
2268 // specified operands exists, it can just return it. Otherwise, it
2269 // updates the node in place to have the requested operands.
2271 // If we updated the node in place, reset the node ID. To the isel,
2272 // this should be just like a newly allocated machine node.
2276 unsigned ResNumResults = Res->getNumValues();
2277 // Move the glue if needed.
2278 if ((EmitNodeInfo & OPFL_GlueOutput) && OldGlueResultNo != -1 &&
2279 (unsigned)OldGlueResultNo != ResNumResults-1)
2280 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldGlueResultNo),
2281 SDValue(Res, ResNumResults-1));
2283 if ((EmitNodeInfo & OPFL_GlueOutput) != 0)
2286 // Move the chain reference if needed.
2287 if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 &&
2288 (unsigned)OldChainResultNo != ResNumResults-1)
2289 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo),
2290 SDValue(Res, ResNumResults-1));
2292 // Otherwise, no replacement happened because the node already exists. Replace
2293 // Uses of the old node with the new one.
2295 CurDAG->ReplaceAllUsesWith(Node, Res);
2300 /// CheckSame - Implements OP_CheckSame.
2301 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2302 CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2304 const SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes) {
2305 // Accept if it is exactly the same as a previously recorded node.
2306 unsigned RecNo = MatcherTable[MatcherIndex++];
2307 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2308 return N == RecordedNodes[RecNo].first;
2311 /// CheckChildSame - Implements OP_CheckChildXSame.
2312 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2313 CheckChildSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2315 const SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes,
2317 if (ChildNo >= N.getNumOperands())
2318 return false; // Match fails if out of range child #.
2319 return ::CheckSame(MatcherTable, MatcherIndex, N.getOperand(ChildNo),
2323 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
2324 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2325 CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2326 const SelectionDAGISel &SDISel) {
2327 return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]);
2330 /// CheckNodePredicate - Implements OP_CheckNodePredicate.
2331 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2332 CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2333 const SelectionDAGISel &SDISel, SDNode *N) {
2334 return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]);
2337 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2338 CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2340 uint16_t Opc = MatcherTable[MatcherIndex++];
2341 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2342 return N->getOpcode() == Opc;
2345 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2346 CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2347 SDValue N, const TargetLowering *TLI) {
2348 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2349 if (N.getValueType() == VT) return true;
2351 // Handle the case when VT is iPTR.
2352 return VT == MVT::iPTR && N.getValueType() == TLI->getPointerTy();
2355 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2356 CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2357 SDValue N, const TargetLowering *TLI, unsigned ChildNo) {
2358 if (ChildNo >= N.getNumOperands())
2359 return false; // Match fails if out of range child #.
2360 return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI);
2363 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2364 CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2366 return cast<CondCodeSDNode>(N)->get() ==
2367 (ISD::CondCode)MatcherTable[MatcherIndex++];
2370 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2371 CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2372 SDValue N, const TargetLowering *TLI) {
2373 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2374 if (cast<VTSDNode>(N)->getVT() == VT)
2377 // Handle the case when VT is iPTR.
2378 return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI->getPointerTy();
2381 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2382 CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2384 int64_t Val = MatcherTable[MatcherIndex++];
2386 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2388 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
2389 return C && C->getSExtValue() == Val;
2392 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2393 CheckChildInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2394 SDValue N, unsigned ChildNo) {
2395 if (ChildNo >= N.getNumOperands())
2396 return false; // Match fails if out of range child #.
2397 return ::CheckInteger(MatcherTable, MatcherIndex, N.getOperand(ChildNo));
2400 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2401 CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2402 SDValue N, const SelectionDAGISel &SDISel) {
2403 int64_t Val = MatcherTable[MatcherIndex++];
2405 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2407 if (N->getOpcode() != ISD::AND) return false;
2409 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
2410 return C && SDISel.CheckAndMask(N.getOperand(0), C, Val);
2413 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
2414 CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
2415 SDValue N, const SelectionDAGISel &SDISel) {
2416 int64_t Val = MatcherTable[MatcherIndex++];
2418 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2420 if (N->getOpcode() != ISD::OR) return false;
2422 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
2423 return C && SDISel.CheckOrMask(N.getOperand(0), C, Val);
2426 /// IsPredicateKnownToFail - If we know how and can do so without pushing a
2427 /// scope, evaluate the current node. If the current predicate is known to
2428 /// fail, set Result=true and return anything. If the current predicate is
2429 /// known to pass, set Result=false and return the MatcherIndex to continue
2430 /// with. If the current predicate is unknown, set Result=false and return the
2431 /// MatcherIndex to continue with.
2432 static unsigned IsPredicateKnownToFail(const unsigned char *Table,
2433 unsigned Index, SDValue N,
2435 const SelectionDAGISel &SDISel,
2436 SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes) {
2437 switch (Table[Index++]) {
2440 return Index-1; // Could not evaluate this predicate.
2441 case SelectionDAGISel::OPC_CheckSame:
2442 Result = !::CheckSame(Table, Index, N, RecordedNodes);
2444 case SelectionDAGISel::OPC_CheckChild0Same:
2445 case SelectionDAGISel::OPC_CheckChild1Same:
2446 case SelectionDAGISel::OPC_CheckChild2Same:
2447 case SelectionDAGISel::OPC_CheckChild3Same:
2448 Result = !::CheckChildSame(Table, Index, N, RecordedNodes,
2449 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Same);
2451 case SelectionDAGISel::OPC_CheckPatternPredicate:
2452 Result = !::CheckPatternPredicate(Table, Index, SDISel);
2454 case SelectionDAGISel::OPC_CheckPredicate:
2455 Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode());
2457 case SelectionDAGISel::OPC_CheckOpcode:
2458 Result = !::CheckOpcode(Table, Index, N.getNode());
2460 case SelectionDAGISel::OPC_CheckType:
2461 Result = !::CheckType(Table, Index, N, SDISel.TLI);
2463 case SelectionDAGISel::OPC_CheckChild0Type:
2464 case SelectionDAGISel::OPC_CheckChild1Type:
2465 case SelectionDAGISel::OPC_CheckChild2Type:
2466 case SelectionDAGISel::OPC_CheckChild3Type:
2467 case SelectionDAGISel::OPC_CheckChild4Type:
2468 case SelectionDAGISel::OPC_CheckChild5Type:
2469 case SelectionDAGISel::OPC_CheckChild6Type:
2470 case SelectionDAGISel::OPC_CheckChild7Type:
2471 Result = !::CheckChildType(Table, Index, N, SDISel.TLI,
2473 SelectionDAGISel::OPC_CheckChild0Type);
2475 case SelectionDAGISel::OPC_CheckCondCode:
2476 Result = !::CheckCondCode(Table, Index, N);
2478 case SelectionDAGISel::OPC_CheckValueType:
2479 Result = !::CheckValueType(Table, Index, N, SDISel.TLI);
2481 case SelectionDAGISel::OPC_CheckInteger:
2482 Result = !::CheckInteger(Table, Index, N);
2484 case SelectionDAGISel::OPC_CheckChild0Integer:
2485 case SelectionDAGISel::OPC_CheckChild1Integer:
2486 case SelectionDAGISel::OPC_CheckChild2Integer:
2487 case SelectionDAGISel::OPC_CheckChild3Integer:
2488 case SelectionDAGISel::OPC_CheckChild4Integer:
2489 Result = !::CheckChildInteger(Table, Index, N,
2490 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Integer);
2492 case SelectionDAGISel::OPC_CheckAndImm:
2493 Result = !::CheckAndImm(Table, Index, N, SDISel);
2495 case SelectionDAGISel::OPC_CheckOrImm:
2496 Result = !::CheckOrImm(Table, Index, N, SDISel);
2504 /// FailIndex - If this match fails, this is the index to continue with.
2507 /// NodeStack - The node stack when the scope was formed.
2508 SmallVector<SDValue, 4> NodeStack;
2510 /// NumRecordedNodes - The number of recorded nodes when the scope was formed.
2511 unsigned NumRecordedNodes;
2513 /// NumMatchedMemRefs - The number of matched memref entries.
2514 unsigned NumMatchedMemRefs;
2516 /// InputChain/InputGlue - The current chain/glue
2517 SDValue InputChain, InputGlue;
2519 /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty.
2520 bool HasChainNodesMatched, HasGlueResultNodesMatched;
2523 /// \\brief A DAG update listener to keep the matching state
2524 /// (i.e. RecordedNodes and MatchScope) uptodate if the target is allowed to
2525 /// change the DAG while matching. X86 addressing mode matcher is an example
2527 class MatchStateUpdater : public SelectionDAG::DAGUpdateListener
2529 SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes;
2530 SmallVectorImpl<MatchScope> &MatchScopes;
2532 MatchStateUpdater(SelectionDAG &DAG,
2533 SmallVectorImpl<std::pair<SDValue, SDNode*> > &RN,
2534 SmallVectorImpl<MatchScope> &MS) :
2535 SelectionDAG::DAGUpdateListener(DAG),
2536 RecordedNodes(RN), MatchScopes(MS) { }
2538 void NodeDeleted(SDNode *N, SDNode *E) {
2539 // Some early-returns here to avoid the search if we deleted the node or
2540 // if the update comes from MorphNodeTo (MorphNodeTo is the last thing we
2541 // do, so it's unnecessary to update matching state at that point).
2542 // Neither of these can occur currently because we only install this
2543 // update listener during matching a complex patterns.
2544 if (!E || E->isMachineOpcode())
2546 // Performing linear search here does not matter because we almost never
2547 // run this code. You'd have to have a CSE during complex pattern
2549 for (auto &I : RecordedNodes)
2550 if (I.first.getNode() == N)
2553 for (auto &I : MatchScopes)
2554 for (auto &J : I.NodeStack)
2555 if (J.getNode() == N)
2561 SDNode *SelectionDAGISel::
2562 SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
2563 unsigned TableSize) {
2564 // FIXME: Should these even be selected? Handle these cases in the caller?
2565 switch (NodeToMatch->getOpcode()) {
2568 case ISD::EntryToken: // These nodes remain the same.
2569 case ISD::BasicBlock:
2571 case ISD::RegisterMask:
2572 case ISD::HANDLENODE:
2573 case ISD::MDNODE_SDNODE:
2574 case ISD::TargetConstant:
2575 case ISD::TargetConstantFP:
2576 case ISD::TargetConstantPool:
2577 case ISD::TargetFrameIndex:
2578 case ISD::TargetExternalSymbol:
2579 case ISD::TargetBlockAddress:
2580 case ISD::TargetJumpTable:
2581 case ISD::TargetGlobalTLSAddress:
2582 case ISD::TargetGlobalAddress:
2583 case ISD::TokenFactor:
2584 case ISD::CopyFromReg:
2585 case ISD::CopyToReg:
2587 case ISD::LIFETIME_START:
2588 case ISD::LIFETIME_END:
2589 NodeToMatch->setNodeId(-1); // Mark selected.
2591 case ISD::AssertSext:
2592 case ISD::AssertZext:
2593 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0),
2594 NodeToMatch->getOperand(0));
2596 case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch);
2597 case ISD::READ_REGISTER: return Select_READ_REGISTER(NodeToMatch);
2598 case ISD::WRITE_REGISTER: return Select_WRITE_REGISTER(NodeToMatch);
2599 case ISD::UNDEF: return Select_UNDEF(NodeToMatch);
2602 assert(!NodeToMatch->isMachineOpcode() && "Node already selected!");
2604 // Set up the node stack with NodeToMatch as the only node on the stack.
2605 SmallVector<SDValue, 8> NodeStack;
2606 SDValue N = SDValue(NodeToMatch, 0);
2607 NodeStack.push_back(N);
2609 // MatchScopes - Scopes used when matching, if a match failure happens, this
2610 // indicates where to continue checking.
2611 SmallVector<MatchScope, 8> MatchScopes;
2613 // RecordedNodes - This is the set of nodes that have been recorded by the
2614 // state machine. The second value is the parent of the node, or null if the
2615 // root is recorded.
2616 SmallVector<std::pair<SDValue, SDNode*>, 8> RecordedNodes;
2618 // MatchedMemRefs - This is the set of MemRef's we've seen in the input
2620 SmallVector<MachineMemOperand*, 2> MatchedMemRefs;
2622 // These are the current input chain and glue for use when generating nodes.
2623 // Various Emit operations change these. For example, emitting a copytoreg
2624 // uses and updates these.
2625 SDValue InputChain, InputGlue;
2627 // ChainNodesMatched - If a pattern matches nodes that have input/output
2628 // chains, the OPC_EmitMergeInputChains operation is emitted which indicates
2629 // which ones they are. The result is captured into this list so that we can
2630 // update the chain results when the pattern is complete.
2631 SmallVector<SDNode*, 3> ChainNodesMatched;
2632 SmallVector<SDNode*, 3> GlueResultNodesMatched;
2634 DEBUG(dbgs() << "ISEL: Starting pattern match on root node: ";
2635 NodeToMatch->dump(CurDAG);
2638 // Determine where to start the interpreter. Normally we start at opcode #0,
2639 // but if the state machine starts with an OPC_SwitchOpcode, then we
2640 // accelerate the first lookup (which is guaranteed to be hot) with the
2641 // OpcodeOffset table.
2642 unsigned MatcherIndex = 0;
2644 if (!OpcodeOffset.empty()) {
2645 // Already computed the OpcodeOffset table, just index into it.
2646 if (N.getOpcode() < OpcodeOffset.size())
2647 MatcherIndex = OpcodeOffset[N.getOpcode()];
2648 DEBUG(dbgs() << " Initial Opcode index to " << MatcherIndex << "\n");
2650 } else if (MatcherTable[0] == OPC_SwitchOpcode) {
2651 // Otherwise, the table isn't computed, but the state machine does start
2652 // with an OPC_SwitchOpcode instruction. Populate the table now, since this
2653 // is the first time we're selecting an instruction.
2656 // Get the size of this case.
2657 unsigned CaseSize = MatcherTable[Idx++];
2659 CaseSize = GetVBR(CaseSize, MatcherTable, Idx);
2660 if (CaseSize == 0) break;
2662 // Get the opcode, add the index to the table.
2663 uint16_t Opc = MatcherTable[Idx++];
2664 Opc |= (unsigned short)MatcherTable[Idx++] << 8;
2665 if (Opc >= OpcodeOffset.size())
2666 OpcodeOffset.resize((Opc+1)*2);
2667 OpcodeOffset[Opc] = Idx;
2671 // Okay, do the lookup for the first opcode.
2672 if (N.getOpcode() < OpcodeOffset.size())
2673 MatcherIndex = OpcodeOffset[N.getOpcode()];
2677 assert(MatcherIndex < TableSize && "Invalid index");
2679 unsigned CurrentOpcodeIndex = MatcherIndex;
2681 BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++];
2684 // Okay, the semantics of this operation are that we should push a scope
2685 // then evaluate the first child. However, pushing a scope only to have
2686 // the first check fail (which then pops it) is inefficient. If we can
2687 // determine immediately that the first check (or first several) will
2688 // immediately fail, don't even bother pushing a scope for them.
2692 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2693 if (NumToSkip & 128)
2694 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2695 // Found the end of the scope with no match.
2696 if (NumToSkip == 0) {
2701 FailIndex = MatcherIndex+NumToSkip;
2703 unsigned MatcherIndexOfPredicate = MatcherIndex;
2704 (void)MatcherIndexOfPredicate; // silence warning.
2706 // If we can't evaluate this predicate without pushing a scope (e.g. if
2707 // it is a 'MoveParent') or if the predicate succeeds on this node, we
2708 // push the scope and evaluate the full predicate chain.
2710 MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
2711 Result, *this, RecordedNodes);
2715 DEBUG(dbgs() << " Skipped scope entry (due to false predicate) at "
2716 << "index " << MatcherIndexOfPredicate
2717 << ", continuing at " << FailIndex << "\n");
2718 ++NumDAGIselRetries;
2720 // Otherwise, we know that this case of the Scope is guaranteed to fail,
2721 // move to the next case.
2722 MatcherIndex = FailIndex;
2725 // If the whole scope failed to match, bail.
2726 if (FailIndex == 0) break;
2728 // Push a MatchScope which indicates where to go if the first child fails
2730 MatchScope NewEntry;
2731 NewEntry.FailIndex = FailIndex;
2732 NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
2733 NewEntry.NumRecordedNodes = RecordedNodes.size();
2734 NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
2735 NewEntry.InputChain = InputChain;
2736 NewEntry.InputGlue = InputGlue;
2737 NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty();
2738 NewEntry.HasGlueResultNodesMatched = !GlueResultNodesMatched.empty();
2739 MatchScopes.push_back(NewEntry);
2742 case OPC_RecordNode: {
2743 // Remember this node, it may end up being an operand in the pattern.
2744 SDNode *Parent = nullptr;
2745 if (NodeStack.size() > 1)
2746 Parent = NodeStack[NodeStack.size()-2].getNode();
2747 RecordedNodes.push_back(std::make_pair(N, Parent));
2751 case OPC_RecordChild0: case OPC_RecordChild1:
2752 case OPC_RecordChild2: case OPC_RecordChild3:
2753 case OPC_RecordChild4: case OPC_RecordChild5:
2754 case OPC_RecordChild6: case OPC_RecordChild7: {
2755 unsigned ChildNo = Opcode-OPC_RecordChild0;
2756 if (ChildNo >= N.getNumOperands())
2757 break; // Match fails if out of range child #.
2759 RecordedNodes.push_back(std::make_pair(N->getOperand(ChildNo),
2763 case OPC_RecordMemRef:
2764 MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand());
2767 case OPC_CaptureGlueInput:
2768 // If the current node has an input glue, capture it in InputGlue.
2769 if (N->getNumOperands() != 0 &&
2770 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue)
2771 InputGlue = N->getOperand(N->getNumOperands()-1);
2774 case OPC_MoveChild: {
2775 unsigned ChildNo = MatcherTable[MatcherIndex++];
2776 if (ChildNo >= N.getNumOperands())
2777 break; // Match fails if out of range child #.
2778 N = N.getOperand(ChildNo);
2779 NodeStack.push_back(N);
2783 case OPC_MoveParent:
2784 // Pop the current node off the NodeStack.
2785 NodeStack.pop_back();
2786 assert(!NodeStack.empty() && "Node stack imbalance!");
2787 N = NodeStack.back();
2791 if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
2794 case OPC_CheckChild0Same: case OPC_CheckChild1Same:
2795 case OPC_CheckChild2Same: case OPC_CheckChild3Same:
2796 if (!::CheckChildSame(MatcherTable, MatcherIndex, N, RecordedNodes,
2797 Opcode-OPC_CheckChild0Same))
2801 case OPC_CheckPatternPredicate:
2802 if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break;
2804 case OPC_CheckPredicate:
2805 if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this,
2809 case OPC_CheckComplexPat: {
2810 unsigned CPNum = MatcherTable[MatcherIndex++];
2811 unsigned RecNo = MatcherTable[MatcherIndex++];
2812 assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
2814 // If target can modify DAG during matching, keep the matching state
2816 std::unique_ptr<MatchStateUpdater> MSU;
2817 if (ComplexPatternFuncMutatesDAG())
2818 MSU.reset(new MatchStateUpdater(*CurDAG, RecordedNodes,
2821 if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo].second,
2822 RecordedNodes[RecNo].first, CPNum,
2827 case OPC_CheckOpcode:
2828 if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
2832 if (!::CheckType(MatcherTable, MatcherIndex, N, TLI))
2836 case OPC_SwitchOpcode: {
2837 unsigned CurNodeOpcode = N.getOpcode();
2838 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2841 // Get the size of this case.
2842 CaseSize = MatcherTable[MatcherIndex++];
2844 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2845 if (CaseSize == 0) break;
2847 uint16_t Opc = MatcherTable[MatcherIndex++];
2848 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2850 // If the opcode matches, then we will execute this case.
2851 if (CurNodeOpcode == Opc)
2854 // Otherwise, skip over this case.
2855 MatcherIndex += CaseSize;
2858 // If no cases matched, bail out.
2859 if (CaseSize == 0) break;
2861 // Otherwise, execute the case we found.
2862 DEBUG(dbgs() << " OpcodeSwitch from " << SwitchStart
2863 << " to " << MatcherIndex << "\n");
2867 case OPC_SwitchType: {
2868 MVT CurNodeVT = N.getSimpleValueType();
2869 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2872 // Get the size of this case.
2873 CaseSize = MatcherTable[MatcherIndex++];
2875 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2876 if (CaseSize == 0) break;
2878 MVT CaseVT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2879 if (CaseVT == MVT::iPTR)
2880 CaseVT = TLI->getPointerTy();
2882 // If the VT matches, then we will execute this case.
2883 if (CurNodeVT == CaseVT)
2886 // Otherwise, skip over this case.
2887 MatcherIndex += CaseSize;
2890 // If no cases matched, bail out.
2891 if (CaseSize == 0) break;
2893 // Otherwise, execute the case we found.
2894 DEBUG(dbgs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
2895 << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
2898 case OPC_CheckChild0Type: case OPC_CheckChild1Type:
2899 case OPC_CheckChild2Type: case OPC_CheckChild3Type:
2900 case OPC_CheckChild4Type: case OPC_CheckChild5Type:
2901 case OPC_CheckChild6Type: case OPC_CheckChild7Type:
2902 if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI,
2903 Opcode-OPC_CheckChild0Type))
2906 case OPC_CheckCondCode:
2907 if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
2909 case OPC_CheckValueType:
2910 if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI))
2913 case OPC_CheckInteger:
2914 if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
2916 case OPC_CheckChild0Integer: case OPC_CheckChild1Integer:
2917 case OPC_CheckChild2Integer: case OPC_CheckChild3Integer:
2918 case OPC_CheckChild4Integer:
2919 if (!::CheckChildInteger(MatcherTable, MatcherIndex, N,
2920 Opcode-OPC_CheckChild0Integer)) break;
2922 case OPC_CheckAndImm:
2923 if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
2925 case OPC_CheckOrImm:
2926 if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
2929 case OPC_CheckFoldableChainNode: {
2930 assert(NodeStack.size() != 1 && "No parent node");
2931 // Verify that all intermediate nodes between the root and this one have
2933 bool HasMultipleUses = false;
2934 for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i)
2935 if (!NodeStack[i].hasOneUse()) {
2936 HasMultipleUses = true;
2939 if (HasMultipleUses) break;
2941 // Check to see that the target thinks this is profitable to fold and that
2942 // we can fold it without inducing cycles in the graph.
2943 if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2945 !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2946 NodeToMatch, OptLevel,
2947 true/*We validate our own chains*/))
2952 case OPC_EmitInteger: {
2953 MVT::SimpleValueType VT =
2954 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2955 int64_t Val = MatcherTable[MatcherIndex++];
2957 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2958 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2959 CurDAG->getTargetConstant(Val, VT), nullptr));
2962 case OPC_EmitRegister: {
2963 MVT::SimpleValueType VT =
2964 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2965 unsigned RegNo = MatcherTable[MatcherIndex++];
2966 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2967 CurDAG->getRegister(RegNo, VT), nullptr));
2970 case OPC_EmitRegister2: {
2971 // For targets w/ more than 256 register names, the register enum
2972 // values are stored in two bytes in the matcher table (just like
2974 MVT::SimpleValueType VT =
2975 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2976 unsigned RegNo = MatcherTable[MatcherIndex++];
2977 RegNo |= MatcherTable[MatcherIndex++] << 8;
2978 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2979 CurDAG->getRegister(RegNo, VT), nullptr));
2983 case OPC_EmitConvertToTarget: {
2984 // Convert from IMM/FPIMM to target version.
2985 unsigned RecNo = MatcherTable[MatcherIndex++];
2986 assert(RecNo < RecordedNodes.size() && "Invalid EmitConvertToTarget");
2987 SDValue Imm = RecordedNodes[RecNo].first;
2989 if (Imm->getOpcode() == ISD::Constant) {
2990 const ConstantInt *Val=cast<ConstantSDNode>(Imm)->getConstantIntValue();
2991 Imm = CurDAG->getConstant(*Val, Imm.getValueType(), true);
2992 } else if (Imm->getOpcode() == ISD::ConstantFP) {
2993 const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue();
2994 Imm = CurDAG->getConstantFP(*Val, 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().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;