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 #define DEBUG_TYPE "isel"
15 #include "ScheduleDAGSDNodes.h"
16 #include "SelectionDAGBuilder.h"
17 #include "llvm/CodeGen/FunctionLoweringInfo.h"
18 #include "llvm/CodeGen/SelectionDAGISel.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/DebugInfo.h"
21 #include "llvm/Constants.h"
22 #include "llvm/Function.h"
23 #include "llvm/InlineAsm.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/LLVMContext.h"
28 #include "llvm/Module.h"
29 #include "llvm/CodeGen/FastISel.h"
30 #include "llvm/CodeGen/GCStrategy.h"
31 #include "llvm/CodeGen/GCMetadata.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/CodeGen/MachineFunction.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
38 #include "llvm/CodeGen/SchedulerRegistry.h"
39 #include "llvm/CodeGen/SelectionDAG.h"
40 #include "llvm/Target/TargetRegisterInfo.h"
41 #include "llvm/Target/TargetIntrinsicInfo.h"
42 #include "llvm/Target/TargetInstrInfo.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Target/TargetMachine.h"
45 #include "llvm/Target/TargetOptions.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
47 #include "llvm/Support/Compiler.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/Timer.h"
51 #include "llvm/Support/raw_ostream.h"
52 #include "llvm/ADT/PostOrderIterator.h"
53 #include "llvm/ADT/Statistic.h"
57 STATISTIC(NumFastIselFailures, "Number of instructions fast isel failed on");
58 STATISTIC(NumFastIselBlocks, "Number of blocks selected entirely by fast isel");
59 STATISTIC(NumDAGBlocks, "Number of blocks selected using DAG");
60 STATISTIC(NumDAGIselRetries,"Number of times dag isel has to try another path");
63 STATISTIC(NumBBWithOutOfOrderLineInfo,
64 "Number of blocks with out of order line number info");
65 STATISTIC(NumMBBWithOutOfOrderLineInfo,
66 "Number of machine blocks with out of order line number info");
70 EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
71 cl::desc("Enable verbose messages in the \"fast\" "
72 "instruction selector"));
74 EnableFastISelAbort("fast-isel-abort", cl::Hidden,
75 cl::desc("Enable abort calls when \"fast\" instruction fails"));
79 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
80 cl::desc("Pop up a window to show dags before the first "
83 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
84 cl::desc("Pop up a window to show dags before legalize types"));
86 ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
87 cl::desc("Pop up a window to show dags before legalize"));
89 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
90 cl::desc("Pop up a window to show dags before the second "
93 ViewDAGCombineLT("view-dag-combine-lt-dags", cl::Hidden,
94 cl::desc("Pop up a window to show dags before the post legalize types"
95 " dag combine pass"));
97 ViewISelDAGs("view-isel-dags", cl::Hidden,
98 cl::desc("Pop up a window to show isel dags as they are selected"));
100 ViewSchedDAGs("view-sched-dags", cl::Hidden,
101 cl::desc("Pop up a window to show sched dags as they are processed"));
103 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
104 cl::desc("Pop up a window to show SUnit dags after they are processed"));
106 static const bool ViewDAGCombine1 = false,
107 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false,
108 ViewDAGCombine2 = false,
109 ViewDAGCombineLT = false,
110 ViewISelDAGs = false, ViewSchedDAGs = false,
111 ViewSUnitDAGs = false;
114 //===---------------------------------------------------------------------===//
116 /// RegisterScheduler class - Track the registration of instruction schedulers.
118 //===---------------------------------------------------------------------===//
119 MachinePassRegistry RegisterScheduler::Registry;
121 //===---------------------------------------------------------------------===//
123 /// ISHeuristic command line option for instruction schedulers.
125 //===---------------------------------------------------------------------===//
126 static cl::opt<RegisterScheduler::FunctionPassCtor, false,
127 RegisterPassParser<RegisterScheduler> >
128 ISHeuristic("pre-RA-sched",
129 cl::init(&createDefaultScheduler),
130 cl::desc("Instruction schedulers available (before register"
133 static RegisterScheduler
134 defaultListDAGScheduler("default", "Best scheduler for the target",
135 createDefaultScheduler);
138 //===--------------------------------------------------------------------===//
139 /// createDefaultScheduler - This creates an instruction scheduler appropriate
141 ScheduleDAGSDNodes* createDefaultScheduler(SelectionDAGISel *IS,
142 CodeGenOpt::Level OptLevel) {
143 const TargetLowering &TLI = IS->getTargetLowering();
145 if (OptLevel == CodeGenOpt::None)
146 return createSourceListDAGScheduler(IS, OptLevel);
147 if (TLI.getSchedulingPreference() == Sched::Latency)
148 return createTDListDAGScheduler(IS, OptLevel);
149 if (TLI.getSchedulingPreference() == Sched::RegPressure)
150 return createBURRListDAGScheduler(IS, OptLevel);
151 if (TLI.getSchedulingPreference() == Sched::Hybrid)
152 return createHybridListDAGScheduler(IS, OptLevel);
153 assert(TLI.getSchedulingPreference() == Sched::ILP &&
154 "Unknown sched type!");
155 return createILPListDAGScheduler(IS, OptLevel);
159 // EmitInstrWithCustomInserter - This method should be implemented by targets
160 // that mark instructions with the 'usesCustomInserter' flag. These
161 // instructions are special in various ways, which require special support to
162 // insert. The specified MachineInstr is created but not inserted into any
163 // basic blocks, and this method is called to expand it into a sequence of
164 // instructions, potentially also creating new basic blocks and control flow.
165 // When new basic blocks are inserted and the edges from MBB to its successors
166 // are modified, the method should insert pairs of <OldSucc, NewSucc> into the
169 TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
170 MachineBasicBlock *MBB) const {
172 dbgs() << "If a target marks an instruction with "
173 "'usesCustomInserter', it must implement "
174 "TargetLowering::EmitInstrWithCustomInserter!";
180 //===----------------------------------------------------------------------===//
181 // SelectionDAGISel code
182 //===----------------------------------------------------------------------===//
184 SelectionDAGISel::SelectionDAGISel(const TargetMachine &tm,
185 CodeGenOpt::Level OL) :
186 MachineFunctionPass(ID), TM(tm), TLI(*tm.getTargetLowering()),
187 FuncInfo(new FunctionLoweringInfo(TLI)),
188 CurDAG(new SelectionDAG(tm)),
189 SDB(new SelectionDAGBuilder(*CurDAG, *FuncInfo, OL)),
193 initializeGCModuleInfoPass(*PassRegistry::getPassRegistry());
194 initializeAliasAnalysisAnalysisGroup(*PassRegistry::getPassRegistry());
197 SelectionDAGISel::~SelectionDAGISel() {
203 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
204 AU.addRequired<AliasAnalysis>();
205 AU.addPreserved<AliasAnalysis>();
206 AU.addRequired<GCModuleInfo>();
207 AU.addPreserved<GCModuleInfo>();
208 MachineFunctionPass::getAnalysisUsage(AU);
211 /// FunctionCallsSetJmp - Return true if the function has a call to setjmp or
212 /// other function that gcc recognizes as "returning twice". This is used to
213 /// limit code-gen optimizations on the machine function.
215 /// FIXME: Remove after <rdar://problem/8031714> is fixed.
216 static bool FunctionCallsSetJmp(const Function *F) {
217 const Module *M = F->getParent();
218 static const char *ReturnsTwiceFns[] = {
228 #define NUM_RETURNS_TWICE_FNS sizeof(ReturnsTwiceFns) / sizeof(const char *)
230 for (unsigned I = 0; I < NUM_RETURNS_TWICE_FNS; ++I)
231 if (const Function *Callee = M->getFunction(ReturnsTwiceFns[I])) {
232 if (!Callee->use_empty())
233 for (Value::const_use_iterator
234 I = Callee->use_begin(), E = Callee->use_end();
236 if (const CallInst *CI = dyn_cast<CallInst>(*I))
237 if (CI->getParent()->getParent() == F)
242 #undef NUM_RETURNS_TWICE_FNS
245 /// SplitCriticalSideEffectEdges - Look for critical edges with a PHI value that
246 /// may trap on it. In this case we have to split the edge so that the path
247 /// through the predecessor block that doesn't go to the phi block doesn't
248 /// execute the possibly trapping instruction.
250 /// This is required for correctness, so it must be done at -O0.
252 static void SplitCriticalSideEffectEdges(Function &Fn, Pass *SDISel) {
253 // Loop for blocks with phi nodes.
254 for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
255 PHINode *PN = dyn_cast<PHINode>(BB->begin());
256 if (PN == 0) continue;
259 // For each block with a PHI node, check to see if any of the input values
260 // are potentially trapping constant expressions. Constant expressions are
261 // the only potentially trapping value that can occur as the argument to a
263 for (BasicBlock::iterator I = BB->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
264 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
265 ConstantExpr *CE = dyn_cast<ConstantExpr>(PN->getIncomingValue(i));
266 if (CE == 0 || !CE->canTrap()) continue;
268 // The only case we have to worry about is when the edge is critical.
269 // Since this block has a PHI Node, we assume it has multiple input
270 // edges: check to see if the pred has multiple successors.
271 BasicBlock *Pred = PN->getIncomingBlock(i);
272 if (Pred->getTerminator()->getNumSuccessors() == 1)
275 // Okay, we have to split this edge.
276 SplitCriticalEdge(Pred->getTerminator(),
277 GetSuccessorNumber(Pred, BB), SDISel, true);
283 bool SelectionDAGISel::runOnMachineFunction(MachineFunction &mf) {
284 // Do some sanity-checking on the command-line options.
285 assert((!EnableFastISelVerbose || EnableFastISel) &&
286 "-fast-isel-verbose requires -fast-isel");
287 assert((!EnableFastISelAbort || EnableFastISel) &&
288 "-fast-isel-abort requires -fast-isel");
290 const Function &Fn = *mf.getFunction();
291 const TargetInstrInfo &TII = *TM.getInstrInfo();
292 const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
295 RegInfo = &MF->getRegInfo();
296 AA = &getAnalysis<AliasAnalysis>();
297 GFI = Fn.hasGC() ? &getAnalysis<GCModuleInfo>().getFunctionInfo(Fn) : 0;
299 DEBUG(dbgs() << "\n\n\n=== " << Fn.getName() << "\n");
301 SplitCriticalSideEffectEdges(const_cast<Function&>(Fn), this);
304 FuncInfo->set(Fn, *MF);
307 SelectAllBasicBlocks(Fn);
309 // If the first basic block in the function has live ins that need to be
310 // copied into vregs, emit the copies into the top of the block before
311 // emitting the code for the block.
312 MachineBasicBlock *EntryMBB = MF->begin();
313 RegInfo->EmitLiveInCopies(EntryMBB, TRI, TII);
315 DenseMap<unsigned, unsigned> LiveInMap;
316 if (!FuncInfo->ArgDbgValues.empty())
317 for (MachineRegisterInfo::livein_iterator LI = RegInfo->livein_begin(),
318 E = RegInfo->livein_end(); LI != E; ++LI)
320 LiveInMap.insert(std::make_pair(LI->first, LI->second));
322 // Insert DBG_VALUE instructions for function arguments to the entry block.
323 for (unsigned i = 0, e = FuncInfo->ArgDbgValues.size(); i != e; ++i) {
324 MachineInstr *MI = FuncInfo->ArgDbgValues[e-i-1];
325 unsigned Reg = MI->getOperand(0).getReg();
326 if (TargetRegisterInfo::isPhysicalRegister(Reg))
327 EntryMBB->insert(EntryMBB->begin(), MI);
329 MachineInstr *Def = RegInfo->getVRegDef(Reg);
330 MachineBasicBlock::iterator InsertPos = Def;
331 // FIXME: VR def may not be in entry block.
332 Def->getParent()->insert(llvm::next(InsertPos), MI);
335 // If Reg is live-in then update debug info to track its copy in a vreg.
336 DenseMap<unsigned, unsigned>::iterator LDI = LiveInMap.find(Reg);
337 if (LDI != LiveInMap.end()) {
338 MachineInstr *Def = RegInfo->getVRegDef(LDI->second);
339 MachineBasicBlock::iterator InsertPos = Def;
340 const MDNode *Variable =
341 MI->getOperand(MI->getNumOperands()-1).getMetadata();
342 unsigned Offset = MI->getOperand(1).getImm();
343 // Def is never a terminator here, so it is ok to increment InsertPos.
344 BuildMI(*EntryMBB, ++InsertPos, MI->getDebugLoc(),
345 TII.get(TargetOpcode::DBG_VALUE))
346 .addReg(LDI->second, RegState::Debug)
347 .addImm(Offset).addMetadata(Variable);
349 // If this vreg is directly copied into an exported register then
350 // that COPY instructions also need DBG_VALUE, if it is the only
351 // user of LDI->second.
352 MachineInstr *CopyUseMI = NULL;
353 for (MachineRegisterInfo::use_iterator
354 UI = RegInfo->use_begin(LDI->second);
355 MachineInstr *UseMI = UI.skipInstruction();) {
356 if (UseMI->isDebugValue()) continue;
357 if (UseMI->isCopy() && !CopyUseMI && UseMI->getParent() == EntryMBB) {
358 CopyUseMI = UseMI; continue;
360 // Otherwise this is another use or second copy use.
361 CopyUseMI = NULL; break;
364 MachineInstr *NewMI =
365 BuildMI(*MF, CopyUseMI->getDebugLoc(),
366 TII.get(TargetOpcode::DBG_VALUE))
367 .addReg(CopyUseMI->getOperand(0).getReg(), RegState::Debug)
368 .addImm(Offset).addMetadata(Variable);
369 EntryMBB->insertAfter(CopyUseMI, NewMI);
374 // Determine if there are any calls in this machine function.
375 MachineFrameInfo *MFI = MF->getFrameInfo();
376 if (!MFI->hasCalls()) {
377 for (MachineFunction::const_iterator
378 I = MF->begin(), E = MF->end(); I != E; ++I) {
379 const MachineBasicBlock *MBB = I;
380 for (MachineBasicBlock::const_iterator
381 II = MBB->begin(), IE = MBB->end(); II != IE; ++II) {
382 const TargetInstrDesc &TID = TM.getInstrInfo()->get(II->getOpcode());
384 if ((TID.isCall() && !TID.isReturn()) ||
385 II->isStackAligningInlineAsm()) {
386 MFI->setHasCalls(true);
394 // Determine if there is a call to setjmp in the machine function.
395 MF->setCallsSetJmp(FunctionCallsSetJmp(&Fn));
397 // Replace forward-declared registers with the registers containing
398 // the desired value.
399 MachineRegisterInfo &MRI = MF->getRegInfo();
400 for (DenseMap<unsigned, unsigned>::iterator
401 I = FuncInfo->RegFixups.begin(), E = FuncInfo->RegFixups.end();
403 unsigned From = I->first;
404 unsigned To = I->second;
405 // If To is also scheduled to be replaced, find what its ultimate
408 DenseMap<unsigned, unsigned>::iterator J =
409 FuncInfo->RegFixups.find(To);
414 MRI.replaceRegWith(From, To);
417 // Release function-specific state. SDB and CurDAG are already cleared
425 SelectionDAGISel::SelectBasicBlock(BasicBlock::const_iterator Begin,
426 BasicBlock::const_iterator End,
428 // Lower all of the non-terminator instructions. If a call is emitted
429 // as a tail call, cease emitting nodes for this block. Terminators
430 // are handled below.
431 for (BasicBlock::const_iterator I = Begin; I != End && !SDB->HasTailCall; ++I)
434 // Make sure the root of the DAG is up-to-date.
435 CurDAG->setRoot(SDB->getControlRoot());
436 HadTailCall = SDB->HasTailCall;
439 // Final step, emit the lowered DAG as machine code.
444 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
445 SmallPtrSet<SDNode*, 128> VisitedNodes;
446 SmallVector<SDNode*, 128> Worklist;
448 Worklist.push_back(CurDAG->getRoot().getNode());
455 SDNode *N = Worklist.pop_back_val();
457 // If we've already seen this node, ignore it.
458 if (!VisitedNodes.insert(N))
461 // Otherwise, add all chain operands to the worklist.
462 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
463 if (N->getOperand(i).getValueType() == MVT::Other)
464 Worklist.push_back(N->getOperand(i).getNode());
466 // If this is a CopyToReg with a vreg dest, process it.
467 if (N->getOpcode() != ISD::CopyToReg)
470 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
471 if (!TargetRegisterInfo::isVirtualRegister(DestReg))
474 // Ignore non-scalar or non-integer values.
475 SDValue Src = N->getOperand(2);
476 EVT SrcVT = Src.getValueType();
477 if (!SrcVT.isInteger() || SrcVT.isVector())
480 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
481 Mask = APInt::getAllOnesValue(SrcVT.getSizeInBits());
482 CurDAG->ComputeMaskedBits(Src, Mask, KnownZero, KnownOne);
483 FuncInfo->AddLiveOutRegInfo(DestReg, NumSignBits, KnownZero, KnownOne);
484 } while (!Worklist.empty());
487 void SelectionDAGISel::CodeGenAndEmitDAG() {
488 std::string GroupName;
489 if (TimePassesIsEnabled)
490 GroupName = "Instruction Selection and Scheduling";
491 std::string BlockName;
492 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs ||
493 ViewDAGCombine2 || ViewDAGCombineLT || ViewISelDAGs || ViewSchedDAGs ||
495 BlockName = MF->getFunction()->getNameStr() + ":" +
496 FuncInfo->MBB->getBasicBlock()->getNameStr();
498 DEBUG(dbgs() << "Initial selection DAG:\n"; CurDAG->dump());
500 if (ViewDAGCombine1) CurDAG->viewGraph("dag-combine1 input for " + BlockName);
502 // Run the DAG combiner in pre-legalize mode.
504 NamedRegionTimer T("DAG Combining 1", GroupName, TimePassesIsEnabled);
505 CurDAG->Combine(Unrestricted, *AA, OptLevel);
508 DEBUG(dbgs() << "Optimized lowered selection DAG:\n"; CurDAG->dump());
510 // Second step, hack on the DAG until it only uses operations and types that
511 // the target supports.
512 if (ViewLegalizeTypesDAGs) CurDAG->viewGraph("legalize-types input for " +
517 NamedRegionTimer T("Type Legalization", GroupName, TimePassesIsEnabled);
518 Changed = CurDAG->LegalizeTypes();
521 DEBUG(dbgs() << "Type-legalized selection DAG:\n"; CurDAG->dump());
524 if (ViewDAGCombineLT)
525 CurDAG->viewGraph("dag-combine-lt input for " + BlockName);
527 // Run the DAG combiner in post-type-legalize mode.
529 NamedRegionTimer T("DAG Combining after legalize types", GroupName,
530 TimePassesIsEnabled);
531 CurDAG->Combine(NoIllegalTypes, *AA, OptLevel);
534 DEBUG(dbgs() << "Optimized type-legalized selection DAG:\n";
539 NamedRegionTimer T("Vector Legalization", GroupName, TimePassesIsEnabled);
540 Changed = CurDAG->LegalizeVectors();
545 NamedRegionTimer T("Type Legalization 2", GroupName, TimePassesIsEnabled);
546 CurDAG->LegalizeTypes();
549 if (ViewDAGCombineLT)
550 CurDAG->viewGraph("dag-combine-lv input for " + BlockName);
552 // Run the DAG combiner in post-type-legalize mode.
554 NamedRegionTimer T("DAG Combining after legalize vectors", GroupName,
555 TimePassesIsEnabled);
556 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
559 DEBUG(dbgs() << "Optimized vector-legalized selection DAG:\n";
563 if (ViewLegalizeDAGs) CurDAG->viewGraph("legalize input for " + BlockName);
566 NamedRegionTimer T("DAG Legalization", GroupName, TimePassesIsEnabled);
567 CurDAG->Legalize(OptLevel);
570 DEBUG(dbgs() << "Legalized selection DAG:\n"; CurDAG->dump());
572 if (ViewDAGCombine2) CurDAG->viewGraph("dag-combine2 input for " + BlockName);
574 // Run the DAG combiner in post-legalize mode.
576 NamedRegionTimer T("DAG Combining 2", GroupName, TimePassesIsEnabled);
577 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
580 DEBUG(dbgs() << "Optimized legalized selection DAG:\n"; CurDAG->dump());
582 if (OptLevel != CodeGenOpt::None)
583 ComputeLiveOutVRegInfo();
585 if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName);
587 // Third, instruction select all of the operations to machine code, adding the
588 // code to the MachineBasicBlock.
590 NamedRegionTimer T("Instruction Selection", GroupName, TimePassesIsEnabled);
591 DoInstructionSelection();
594 DEBUG(dbgs() << "Selected selection DAG:\n"; CurDAG->dump());
596 if (ViewSchedDAGs) CurDAG->viewGraph("scheduler input for " + BlockName);
598 // Schedule machine code.
599 ScheduleDAGSDNodes *Scheduler = CreateScheduler();
601 NamedRegionTimer T("Instruction Scheduling", GroupName,
602 TimePassesIsEnabled);
603 Scheduler->Run(CurDAG, FuncInfo->MBB, FuncInfo->InsertPt);
606 if (ViewSUnitDAGs) Scheduler->viewGraph();
608 // Emit machine code to BB. This can change 'BB' to the last block being
610 MachineBasicBlock *FirstMBB = FuncInfo->MBB, *LastMBB;
612 NamedRegionTimer T("Instruction Creation", GroupName, TimePassesIsEnabled);
614 LastMBB = FuncInfo->MBB = Scheduler->EmitSchedule();
615 FuncInfo->InsertPt = Scheduler->InsertPos;
618 // If the block was split, make sure we update any references that are used to
619 // update PHI nodes later on.
620 if (FirstMBB != LastMBB)
621 SDB->UpdateSplitBlock(FirstMBB, LastMBB);
623 // Free the scheduler state.
625 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName,
626 TimePassesIsEnabled);
630 // Free the SelectionDAG state, now that we're finished with it.
634 void SelectionDAGISel::DoInstructionSelection() {
635 DEBUG(errs() << "===== Instruction selection begins:\n");
639 // Select target instructions for the DAG.
641 // Number all nodes with a topological order and set DAGSize.
642 DAGSize = CurDAG->AssignTopologicalOrder();
644 // Create a dummy node (which is not added to allnodes), that adds
645 // a reference to the root node, preventing it from being deleted,
646 // and tracking any changes of the root.
647 HandleSDNode Dummy(CurDAG->getRoot());
648 ISelPosition = SelectionDAG::allnodes_iterator(CurDAG->getRoot().getNode());
651 // The AllNodes list is now topological-sorted. Visit the
652 // nodes by starting at the end of the list (the root of the
653 // graph) and preceding back toward the beginning (the entry
655 while (ISelPosition != CurDAG->allnodes_begin()) {
656 SDNode *Node = --ISelPosition;
657 // Skip dead nodes. DAGCombiner is expected to eliminate all dead nodes,
658 // but there are currently some corner cases that it misses. Also, this
659 // makes it theoretically possible to disable the DAGCombiner.
660 if (Node->use_empty())
663 SDNode *ResNode = Select(Node);
665 // FIXME: This is pretty gross. 'Select' should be changed to not return
666 // anything at all and this code should be nuked with a tactical strike.
668 // If node should not be replaced, continue with the next one.
669 if (ResNode == Node || Node->getOpcode() == ISD::DELETED_NODE)
673 ReplaceUses(Node, ResNode);
675 // If after the replacement this node is not used any more,
676 // remove this dead node.
677 if (Node->use_empty()) { // Don't delete EntryToken, etc.
678 ISelUpdater ISU(ISelPosition);
679 CurDAG->RemoveDeadNode(Node, &ISU);
683 CurDAG->setRoot(Dummy.getValue());
686 DEBUG(errs() << "===== Instruction selection ends:\n");
688 PostprocessISelDAG();
691 /// PrepareEHLandingPad - Emit an EH_LABEL, set up live-in registers, and
692 /// do other setup for EH landing-pad blocks.
693 void SelectionDAGISel::PrepareEHLandingPad() {
694 // Add a label to mark the beginning of the landing pad. Deletion of the
695 // landing pad can thus be detected via the MachineModuleInfo.
696 MCSymbol *Label = MF->getMMI().addLandingPad(FuncInfo->MBB);
698 const TargetInstrDesc &II = TM.getInstrInfo()->get(TargetOpcode::EH_LABEL);
699 BuildMI(*FuncInfo->MBB, FuncInfo->InsertPt, SDB->getCurDebugLoc(), II)
702 // Mark exception register as live in.
703 unsigned Reg = TLI.getExceptionAddressRegister();
704 if (Reg) FuncInfo->MBB->addLiveIn(Reg);
706 // Mark exception selector register as live in.
707 Reg = TLI.getExceptionSelectorRegister();
708 if (Reg) FuncInfo->MBB->addLiveIn(Reg);
710 // FIXME: Hack around an exception handling flaw (PR1508): the personality
711 // function and list of typeids logically belong to the invoke (or, if you
712 // like, the basic block containing the invoke), and need to be associated
713 // with it in the dwarf exception handling tables. Currently however the
714 // information is provided by an intrinsic (eh.selector) that can be moved
715 // to unexpected places by the optimizers: if the unwind edge is critical,
716 // then breaking it can result in the intrinsics being in the successor of
717 // the landing pad, not the landing pad itself. This results
718 // in exceptions not being caught because no typeids are associated with
719 // the invoke. This may not be the only way things can go wrong, but it
720 // is the only way we try to work around for the moment.
721 const BasicBlock *LLVMBB = FuncInfo->MBB->getBasicBlock();
722 const BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
724 if (Br && Br->isUnconditional()) { // Critical edge?
725 BasicBlock::const_iterator I, E;
726 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
727 if (isa<EHSelectorInst>(I))
731 // No catch info found - try to extract some from the successor.
732 CopyCatchInfo(Br->getSuccessor(0), LLVMBB, &MF->getMMI(), *FuncInfo);
739 bool SelectionDAGISel::TryToFoldFastISelLoad(const LoadInst *LI,
741 // Don't try to fold volatile loads. Target has to deal with alignment
743 if (LI->isVolatile()) return false;
745 // Figure out which vreg this is going into.
746 unsigned LoadReg = FastIS->getRegForValue(LI);
747 assert(LoadReg && "Load isn't already assigned a vreg? ");
749 // Check to see what the uses of this vreg are. If it has no uses, or more
750 // than one use (at the machine instr level) then we can't fold it.
751 MachineRegisterInfo::reg_iterator RI = RegInfo->reg_begin(LoadReg);
752 if (RI == RegInfo->reg_end())
755 // See if there is exactly one use of the vreg. If there are multiple uses,
756 // then the instruction got lowered to multiple machine instructions or the
757 // use of the loaded value ended up being multiple operands of the result, in
758 // either case, we can't fold this.
759 MachineRegisterInfo::reg_iterator PostRI = RI; ++PostRI;
760 if (PostRI != RegInfo->reg_end())
763 assert(RI.getOperand().isUse() &&
764 "The only use of the vreg must be a use, we haven't emitted the def!");
766 MachineInstr *User = &*RI;
768 // Set the insertion point properly. Folding the load can cause generation of
769 // other random instructions (like sign extends) for addressing modes, make
770 // sure they get inserted in a logical place before the new instruction.
771 FuncInfo->InsertPt = User;
772 FuncInfo->MBB = User->getParent();
774 // Ask the target to try folding the load.
775 return FastIS->TryToFoldLoad(User, RI.getOperandNo(), LI);
779 /// CheckLineNumbers - Check if basic block instructions follow source order
781 static void CheckLineNumbers(const BasicBlock *BB) {
784 for (BasicBlock::const_iterator BI = BB->begin(),
785 BE = BB->end(); BI != BE; ++BI) {
786 const DebugLoc DL = BI->getDebugLoc();
787 if (DL.isUnknown()) continue;
788 unsigned L = DL.getLine();
789 unsigned C = DL.getCol();
790 if (L < Line || (L == Line && C < Col)) {
791 ++NumBBWithOutOfOrderLineInfo;
799 /// CheckLineNumbers - Check if machine basic block instructions follow source
801 static void CheckLineNumbers(const MachineBasicBlock *MBB) {
804 for (MachineBasicBlock::const_iterator MBI = MBB->begin(),
805 MBE = MBB->end(); MBI != MBE; ++MBI) {
806 const DebugLoc DL = MBI->getDebugLoc();
807 if (DL.isUnknown()) continue;
808 unsigned L = DL.getLine();
809 unsigned C = DL.getCol();
810 if (L < Line || (L == Line && C < Col)) {
811 ++NumMBBWithOutOfOrderLineInfo;
820 void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
821 // Initialize the Fast-ISel state, if needed.
822 FastISel *FastIS = 0;
824 FastIS = TLI.createFastISel(*FuncInfo);
826 // Iterate over all basic blocks in the function.
827 ReversePostOrderTraversal<const Function*> RPOT(&Fn);
828 for (ReversePostOrderTraversal<const Function*>::rpo_iterator
829 I = RPOT.begin(), E = RPOT.end(); I != E; ++I) {
830 const BasicBlock *LLVMBB = *I;
832 CheckLineNumbers(LLVMBB);
835 if (OptLevel != CodeGenOpt::None) {
836 bool AllPredsVisited = true;
837 for (const_pred_iterator PI = pred_begin(LLVMBB), PE = pred_end(LLVMBB);
839 if (!FuncInfo->VisitedBBs.count(*PI)) {
840 AllPredsVisited = false;
845 if (!AllPredsVisited) {
846 for (BasicBlock::const_iterator I = LLVMBB->begin(), E = LLVMBB->end();
847 I != E && isa<PHINode>(I); ++I) {
848 FuncInfo->InvalidatePHILiveOutRegInfo(cast<PHINode>(I));
852 FuncInfo->VisitedBBs.insert(LLVMBB);
855 FuncInfo->MBB = FuncInfo->MBBMap[LLVMBB];
856 FuncInfo->InsertPt = FuncInfo->MBB->getFirstNonPHI();
858 BasicBlock::const_iterator const Begin = LLVMBB->getFirstNonPHI();
859 BasicBlock::const_iterator const End = LLVMBB->end();
860 BasicBlock::const_iterator BI = End;
862 FuncInfo->InsertPt = FuncInfo->MBB->getFirstNonPHI();
864 // Setup an EH landing-pad block.
865 if (FuncInfo->MBB->isLandingPad())
866 PrepareEHLandingPad();
868 // Lower any arguments needed in this block if this is the entry block.
869 if (LLVMBB == &Fn.getEntryBlock())
870 LowerArguments(LLVMBB);
872 // Before doing SelectionDAG ISel, see if FastISel has been requested.
874 FastIS->startNewBlock();
876 // Emit code for any incoming arguments. This must happen before
877 // beginning FastISel on the entry block.
878 if (LLVMBB == &Fn.getEntryBlock()) {
879 CurDAG->setRoot(SDB->getControlRoot());
883 // If we inserted any instructions at the beginning, make a note of
884 // where they are, so we can be sure to emit subsequent instructions
886 if (FuncInfo->InsertPt != FuncInfo->MBB->begin())
887 FastIS->setLastLocalValue(llvm::prior(FuncInfo->InsertPt));
889 FastIS->setLastLocalValue(0);
892 // Do FastISel on as many instructions as possible.
893 for (; BI != Begin; --BI) {
894 const Instruction *Inst = llvm::prior(BI);
896 // If we no longer require this instruction, skip it.
897 if (!Inst->mayWriteToMemory() &&
898 !isa<TerminatorInst>(Inst) &&
899 !isa<DbgInfoIntrinsic>(Inst) &&
900 !FuncInfo->isExportedInst(Inst))
903 // Bottom-up: reset the insert pos at the top, after any local-value
905 FastIS->recomputeInsertPt();
907 // Try to select the instruction with FastISel.
908 if (FastIS->SelectInstruction(Inst)) {
909 // If fast isel succeeded, check to see if there is a single-use
910 // non-volatile load right before the selected instruction, and see if
911 // the load is used by the instruction. If so, try to fold it.
912 const Instruction *BeforeInst = 0;
914 BeforeInst = llvm::prior(llvm::prior(BI));
915 if (BeforeInst && isa<LoadInst>(BeforeInst) &&
916 BeforeInst->hasOneUse() && *BeforeInst->use_begin() == Inst &&
917 TryToFoldFastISelLoad(cast<LoadInst>(BeforeInst), FastIS))
918 --BI; // If we succeeded, don't re-select the load.
922 // Then handle certain instructions as single-LLVM-Instruction blocks.
923 if (isa<CallInst>(Inst)) {
924 ++NumFastIselFailures;
925 if (EnableFastISelVerbose || EnableFastISelAbort) {
926 dbgs() << "FastISel missed call: ";
930 if (!Inst->getType()->isVoidTy() && !Inst->use_empty()) {
931 unsigned &R = FuncInfo->ValueMap[Inst];
933 R = FuncInfo->CreateRegs(Inst->getType());
936 bool HadTailCall = false;
937 SelectBasicBlock(Inst, BI, HadTailCall);
939 // If the call was emitted as a tail call, we're done with the block.
948 // Otherwise, give up on FastISel for the rest of the block.
949 // For now, be a little lenient about non-branch terminators.
950 if (!isa<TerminatorInst>(Inst) || isa<BranchInst>(Inst)) {
951 ++NumFastIselFailures;
952 if (EnableFastISelVerbose || EnableFastISelAbort) {
953 dbgs() << "FastISel miss: ";
956 if (EnableFastISelAbort)
957 // The "fast" selector couldn't handle something and bailed.
958 // For the purpose of debugging, just abort.
959 llvm_unreachable("FastISel didn't select the entire block");
964 FastIS->recomputeInsertPt();
972 // Run SelectionDAG instruction selection on the remainder of the block
973 // not handled by FastISel. If FastISel is not run, this is the entire
976 SelectBasicBlock(Begin, BI, HadTailCall);
979 FuncInfo->PHINodesToUpdate.clear();
984 for (MachineFunction::const_iterator MBI = MF->begin(), MBE = MF->end();
986 CheckLineNumbers(MBI);
991 SelectionDAGISel::FinishBasicBlock() {
993 DEBUG(dbgs() << "Total amount of phi nodes to update: "
994 << FuncInfo->PHINodesToUpdate.size() << "\n";
995 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i)
996 dbgs() << "Node " << i << " : ("
997 << FuncInfo->PHINodesToUpdate[i].first
998 << ", " << FuncInfo->PHINodesToUpdate[i].second << ")\n");
1000 // Next, now that we know what the last MBB the LLVM BB expanded is, update
1001 // PHI nodes in successors.
1002 if (SDB->SwitchCases.empty() &&
1003 SDB->JTCases.empty() &&
1004 SDB->BitTestCases.empty()) {
1005 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i) {
1006 MachineInstr *PHI = FuncInfo->PHINodesToUpdate[i].first;
1007 assert(PHI->isPHI() &&
1008 "This is not a machine PHI node that we are updating!");
1009 if (!FuncInfo->MBB->isSuccessor(PHI->getParent()))
1012 MachineOperand::CreateReg(FuncInfo->PHINodesToUpdate[i].second, false));
1013 PHI->addOperand(MachineOperand::CreateMBB(FuncInfo->MBB));
1018 for (unsigned i = 0, e = SDB->BitTestCases.size(); i != e; ++i) {
1019 // Lower header first, if it wasn't already lowered
1020 if (!SDB->BitTestCases[i].Emitted) {
1021 // Set the current basic block to the mbb we wish to insert the code into
1022 FuncInfo->MBB = SDB->BitTestCases[i].Parent;
1023 FuncInfo->InsertPt = FuncInfo->MBB->end();
1025 SDB->visitBitTestHeader(SDB->BitTestCases[i], FuncInfo->MBB);
1026 CurDAG->setRoot(SDB->getRoot());
1028 CodeGenAndEmitDAG();
1031 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j) {
1032 // Set the current basic block to the mbb we wish to insert the code into
1033 FuncInfo->MBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1034 FuncInfo->InsertPt = FuncInfo->MBB->end();
1037 SDB->visitBitTestCase(SDB->BitTestCases[i],
1038 SDB->BitTestCases[i].Cases[j+1].ThisBB,
1039 SDB->BitTestCases[i].Reg,
1040 SDB->BitTestCases[i].Cases[j],
1043 SDB->visitBitTestCase(SDB->BitTestCases[i],
1044 SDB->BitTestCases[i].Default,
1045 SDB->BitTestCases[i].Reg,
1046 SDB->BitTestCases[i].Cases[j],
1050 CurDAG->setRoot(SDB->getRoot());
1052 CodeGenAndEmitDAG();
1056 for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1058 MachineInstr *PHI = FuncInfo->PHINodesToUpdate[pi].first;
1059 MachineBasicBlock *PHIBB = PHI->getParent();
1060 assert(PHI->isPHI() &&
1061 "This is not a machine PHI node that we are updating!");
1062 // This is "default" BB. We have two jumps to it. From "header" BB and
1063 // from last "case" BB.
1064 if (PHIBB == SDB->BitTestCases[i].Default) {
1065 PHI->addOperand(MachineOperand::
1066 CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1068 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Parent));
1069 PHI->addOperand(MachineOperand::
1070 CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1072 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Cases.
1075 // One of "cases" BB.
1076 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size();
1078 MachineBasicBlock* cBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1079 if (cBB->isSuccessor(PHIBB)) {
1080 PHI->addOperand(MachineOperand::
1081 CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1083 PHI->addOperand(MachineOperand::CreateMBB(cBB));
1088 SDB->BitTestCases.clear();
1090 // If the JumpTable record is filled in, then we need to emit a jump table.
1091 // Updating the PHI nodes is tricky in this case, since we need to determine
1092 // whether the PHI is a successor of the range check MBB or the jump table MBB
1093 for (unsigned i = 0, e = SDB->JTCases.size(); i != e; ++i) {
1094 // Lower header first, if it wasn't already lowered
1095 if (!SDB->JTCases[i].first.Emitted) {
1096 // Set the current basic block to the mbb we wish to insert the code into
1097 FuncInfo->MBB = SDB->JTCases[i].first.HeaderBB;
1098 FuncInfo->InsertPt = FuncInfo->MBB->end();
1100 SDB->visitJumpTableHeader(SDB->JTCases[i].second, SDB->JTCases[i].first,
1102 CurDAG->setRoot(SDB->getRoot());
1104 CodeGenAndEmitDAG();
1107 // Set the current basic block to the mbb we wish to insert the code into
1108 FuncInfo->MBB = SDB->JTCases[i].second.MBB;
1109 FuncInfo->InsertPt = FuncInfo->MBB->end();
1111 SDB->visitJumpTable(SDB->JTCases[i].second);
1112 CurDAG->setRoot(SDB->getRoot());
1114 CodeGenAndEmitDAG();
1117 for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1119 MachineInstr *PHI = FuncInfo->PHINodesToUpdate[pi].first;
1120 MachineBasicBlock *PHIBB = PHI->getParent();
1121 assert(PHI->isPHI() &&
1122 "This is not a machine PHI node that we are updating!");
1123 // "default" BB. We can go there only from header BB.
1124 if (PHIBB == SDB->JTCases[i].second.Default) {
1126 (MachineOperand::CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1129 (MachineOperand::CreateMBB(SDB->JTCases[i].first.HeaderBB));
1131 // JT BB. Just iterate over successors here
1132 if (FuncInfo->MBB->isSuccessor(PHIBB)) {
1134 (MachineOperand::CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1136 PHI->addOperand(MachineOperand::CreateMBB(FuncInfo->MBB));
1140 SDB->JTCases.clear();
1142 // If the switch block involved a branch to one of the actual successors, we
1143 // need to update PHI nodes in that block.
1144 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i) {
1145 MachineInstr *PHI = FuncInfo->PHINodesToUpdate[i].first;
1146 assert(PHI->isPHI() &&
1147 "This is not a machine PHI node that we are updating!");
1148 if (FuncInfo->MBB->isSuccessor(PHI->getParent())) {
1150 MachineOperand::CreateReg(FuncInfo->PHINodesToUpdate[i].second, false));
1151 PHI->addOperand(MachineOperand::CreateMBB(FuncInfo->MBB));
1155 // If we generated any switch lowering information, build and codegen any
1156 // additional DAGs necessary.
1157 for (unsigned i = 0, e = SDB->SwitchCases.size(); i != e; ++i) {
1158 // Set the current basic block to the mbb we wish to insert the code into
1159 FuncInfo->MBB = SDB->SwitchCases[i].ThisBB;
1160 FuncInfo->InsertPt = FuncInfo->MBB->end();
1162 // Determine the unique successors.
1163 SmallVector<MachineBasicBlock *, 2> Succs;
1164 Succs.push_back(SDB->SwitchCases[i].TrueBB);
1165 if (SDB->SwitchCases[i].TrueBB != SDB->SwitchCases[i].FalseBB)
1166 Succs.push_back(SDB->SwitchCases[i].FalseBB);
1168 // Emit the code. Note that this could result in FuncInfo->MBB being split.
1169 SDB->visitSwitchCase(SDB->SwitchCases[i], FuncInfo->MBB);
1170 CurDAG->setRoot(SDB->getRoot());
1172 CodeGenAndEmitDAG();
1174 // Remember the last block, now that any splitting is done, for use in
1175 // populating PHI nodes in successors.
1176 MachineBasicBlock *ThisBB = FuncInfo->MBB;
1178 // Handle any PHI nodes in successors of this chunk, as if we were coming
1179 // from the original BB before switch expansion. Note that PHI nodes can
1180 // occur multiple times in PHINodesToUpdate. We have to be very careful to
1181 // handle them the right number of times.
1182 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1183 FuncInfo->MBB = Succs[i];
1184 FuncInfo->InsertPt = FuncInfo->MBB->end();
1185 // FuncInfo->MBB may have been removed from the CFG if a branch was
1187 if (ThisBB->isSuccessor(FuncInfo->MBB)) {
1188 for (MachineBasicBlock::iterator Phi = FuncInfo->MBB->begin();
1189 Phi != FuncInfo->MBB->end() && Phi->isPHI();
1191 // This value for this PHI node is recorded in PHINodesToUpdate.
1192 for (unsigned pn = 0; ; ++pn) {
1193 assert(pn != FuncInfo->PHINodesToUpdate.size() &&
1194 "Didn't find PHI entry!");
1195 if (FuncInfo->PHINodesToUpdate[pn].first == Phi) {
1196 Phi->addOperand(MachineOperand::
1197 CreateReg(FuncInfo->PHINodesToUpdate[pn].second,
1199 Phi->addOperand(MachineOperand::CreateMBB(ThisBB));
1207 SDB->SwitchCases.clear();
1211 /// Create the scheduler. If a specific scheduler was specified
1212 /// via the SchedulerRegistry, use it, otherwise select the
1213 /// one preferred by the target.
1215 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
1216 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1220 RegisterScheduler::setDefault(Ctor);
1223 return Ctor(this, OptLevel);
1226 //===----------------------------------------------------------------------===//
1227 // Helper functions used by the generated instruction selector.
1228 //===----------------------------------------------------------------------===//
1229 // Calls to these methods are generated by tblgen.
1231 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1232 /// the dag combiner simplified the 255, we still want to match. RHS is the
1233 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1234 /// specified in the .td file (e.g. 255).
1235 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1236 int64_t DesiredMaskS) const {
1237 const APInt &ActualMask = RHS->getAPIntValue();
1238 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1240 // If the actual mask exactly matches, success!
1241 if (ActualMask == DesiredMask)
1244 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1245 if (ActualMask.intersects(~DesiredMask))
1248 // Otherwise, the DAG Combiner may have proven that the value coming in is
1249 // either already zero or is not demanded. Check for known zero input bits.
1250 APInt NeededMask = DesiredMask & ~ActualMask;
1251 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1254 // TODO: check to see if missing bits are just not demanded.
1256 // Otherwise, this pattern doesn't match.
1260 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1261 /// the dag combiner simplified the 255, we still want to match. RHS is the
1262 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1263 /// specified in the .td file (e.g. 255).
1264 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1265 int64_t DesiredMaskS) const {
1266 const APInt &ActualMask = RHS->getAPIntValue();
1267 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1269 // If the actual mask exactly matches, success!
1270 if (ActualMask == DesiredMask)
1273 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1274 if (ActualMask.intersects(~DesiredMask))
1277 // Otherwise, the DAG Combiner may have proven that the value coming in is
1278 // either already zero or is not demanded. Check for known zero input bits.
1279 APInt NeededMask = DesiredMask & ~ActualMask;
1281 APInt KnownZero, KnownOne;
1282 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
1284 // If all the missing bits in the or are already known to be set, match!
1285 if ((NeededMask & KnownOne) == NeededMask)
1288 // TODO: check to see if missing bits are just not demanded.
1290 // Otherwise, this pattern doesn't match.
1295 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1296 /// by tblgen. Others should not call it.
1297 void SelectionDAGISel::
1298 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1299 std::vector<SDValue> InOps;
1300 std::swap(InOps, Ops);
1302 Ops.push_back(InOps[InlineAsm::Op_InputChain]); // 0
1303 Ops.push_back(InOps[InlineAsm::Op_AsmString]); // 1
1304 Ops.push_back(InOps[InlineAsm::Op_MDNode]); // 2, !srcloc
1305 Ops.push_back(InOps[InlineAsm::Op_ExtraInfo]); // 3 (SideEffect, AlignStack)
1307 unsigned i = InlineAsm::Op_FirstOperand, e = InOps.size();
1308 if (InOps[e-1].getValueType() == MVT::Glue)
1309 --e; // Don't process a glue operand if it is here.
1312 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1313 if (!InlineAsm::isMemKind(Flags)) {
1314 // Just skip over this operand, copying the operands verbatim.
1315 Ops.insert(Ops.end(), InOps.begin()+i,
1316 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
1317 i += InlineAsm::getNumOperandRegisters(Flags) + 1;
1319 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 &&
1320 "Memory operand with multiple values?");
1321 // Otherwise, this is a memory operand. Ask the target to select it.
1322 std::vector<SDValue> SelOps;
1323 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps))
1324 report_fatal_error("Could not match memory address. Inline asm"
1327 // Add this to the output node.
1329 InlineAsm::getFlagWord(InlineAsm::Kind_Mem, SelOps.size());
1330 Ops.push_back(CurDAG->getTargetConstant(NewFlags, MVT::i32));
1331 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1336 // Add the glue input back if present.
1337 if (e != InOps.size())
1338 Ops.push_back(InOps.back());
1341 /// findGlueUse - Return use of MVT::Glue value produced by the specified
1344 static SDNode *findGlueUse(SDNode *N) {
1345 unsigned FlagResNo = N->getNumValues()-1;
1346 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1347 SDUse &Use = I.getUse();
1348 if (Use.getResNo() == FlagResNo)
1349 return Use.getUser();
1354 /// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
1355 /// This function recursively traverses up the operand chain, ignoring
1357 static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
1358 SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited,
1359 bool IgnoreChains) {
1360 // The NodeID's are given uniques ID's where a node ID is guaranteed to be
1361 // greater than all of its (recursive) operands. If we scan to a point where
1362 // 'use' is smaller than the node we're scanning for, then we know we will
1365 // The Use may be -1 (unassigned) if it is a newly allocated node. This can
1366 // happen because we scan down to newly selected nodes in the case of glue
1368 if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1))
1371 // Don't revisit nodes if we already scanned it and didn't fail, we know we
1372 // won't fail if we scan it again.
1373 if (!Visited.insert(Use))
1376 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
1377 // Ignore chain uses, they are validated by HandleMergeInputChains.
1378 if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains)
1381 SDNode *N = Use->getOperand(i).getNode();
1383 if (Use == ImmedUse || Use == Root)
1384 continue; // We are not looking for immediate use.
1389 // Traverse up the operand chain.
1390 if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
1396 /// IsProfitableToFold - Returns true if it's profitable to fold the specific
1397 /// operand node N of U during instruction selection that starts at Root.
1398 bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
1399 SDNode *Root) const {
1400 if (OptLevel == CodeGenOpt::None) return false;
1401 return N.hasOneUse();
1404 /// IsLegalToFold - Returns true if the specific operand node N of
1405 /// U can be folded during instruction selection that starts at Root.
1406 bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
1407 CodeGenOpt::Level OptLevel,
1408 bool IgnoreChains) {
1409 if (OptLevel == CodeGenOpt::None) return false;
1411 // If Root use can somehow reach N through a path that that doesn't contain
1412 // U then folding N would create a cycle. e.g. In the following
1413 // diagram, Root can reach N through X. If N is folded into into Root, then
1414 // X is both a predecessor and a successor of U.
1425 // * indicates nodes to be folded together.
1427 // If Root produces glue, then it gets (even more) interesting. Since it
1428 // will be "glued" together with its glue use in the scheduler, we need to
1429 // check if it might reach N.
1448 // If GU (glue use) indirectly reaches N (the load), and Root folds N
1449 // (call it Fold), then X is a predecessor of GU and a successor of
1450 // Fold. But since Fold and GU are glued together, this will create
1451 // a cycle in the scheduling graph.
1453 // If the node has glue, walk down the graph to the "lowest" node in the
1455 EVT VT = Root->getValueType(Root->getNumValues()-1);
1456 while (VT == MVT::Glue) {
1457 SDNode *GU = findGlueUse(Root);
1461 VT = Root->getValueType(Root->getNumValues()-1);
1463 // If our query node has a glue result with a use, we've walked up it. If
1464 // the user (which has already been selected) has a chain or indirectly uses
1465 // the chain, our WalkChainUsers predicate will not consider it. Because of
1466 // this, we cannot ignore chains in this predicate.
1467 IgnoreChains = false;
1471 SmallPtrSet<SDNode*, 16> Visited;
1472 return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains);
1475 SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
1476 std::vector<SDValue> Ops(N->op_begin(), N->op_end());
1477 SelectInlineAsmMemoryOperands(Ops);
1479 std::vector<EVT> VTs;
1480 VTs.push_back(MVT::Other);
1481 VTs.push_back(MVT::Glue);
1482 SDValue New = CurDAG->getNode(ISD::INLINEASM, N->getDebugLoc(),
1483 VTs, &Ops[0], Ops.size());
1485 return New.getNode();
1488 SDNode *SelectionDAGISel::Select_UNDEF(SDNode *N) {
1489 return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0));
1492 /// GetVBR - decode a vbr encoding whose top bit is set.
1493 LLVM_ATTRIBUTE_ALWAYS_INLINE static uint64_t
1494 GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
1495 assert(Val >= 128 && "Not a VBR");
1496 Val &= 127; // Remove first vbr bit.
1501 NextBits = MatcherTable[Idx++];
1502 Val |= (NextBits&127) << Shift;
1504 } while (NextBits & 128);
1510 /// UpdateChainsAndGlue - When a match is complete, this method updates uses of
1511 /// interior glue and chain results to use the new glue and chain results.
1512 void SelectionDAGISel::
1513 UpdateChainsAndGlue(SDNode *NodeToMatch, SDValue InputChain,
1514 const SmallVectorImpl<SDNode*> &ChainNodesMatched,
1516 const SmallVectorImpl<SDNode*> &GlueResultNodesMatched,
1517 bool isMorphNodeTo) {
1518 SmallVector<SDNode*, 4> NowDeadNodes;
1520 ISelUpdater ISU(ISelPosition);
1522 // Now that all the normal results are replaced, we replace the chain and
1523 // glue results if present.
1524 if (!ChainNodesMatched.empty()) {
1525 assert(InputChain.getNode() != 0 &&
1526 "Matched input chains but didn't produce a chain");
1527 // Loop over all of the nodes we matched that produced a chain result.
1528 // Replace all the chain results with the final chain we ended up with.
1529 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1530 SDNode *ChainNode = ChainNodesMatched[i];
1532 // If this node was already deleted, don't look at it.
1533 if (ChainNode->getOpcode() == ISD::DELETED_NODE)
1536 // Don't replace the results of the root node if we're doing a
1538 if (ChainNode == NodeToMatch && isMorphNodeTo)
1541 SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1);
1542 if (ChainVal.getValueType() == MVT::Glue)
1543 ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
1544 assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
1545 CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain, &ISU);
1547 // If the node became dead and we haven't already seen it, delete it.
1548 if (ChainNode->use_empty() &&
1549 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), ChainNode))
1550 NowDeadNodes.push_back(ChainNode);
1554 // If the result produces glue, update any glue results in the matched
1555 // pattern with the glue result.
1556 if (InputGlue.getNode() != 0) {
1557 // Handle any interior nodes explicitly marked.
1558 for (unsigned i = 0, e = GlueResultNodesMatched.size(); i != e; ++i) {
1559 SDNode *FRN = GlueResultNodesMatched[i];
1561 // If this node was already deleted, don't look at it.
1562 if (FRN->getOpcode() == ISD::DELETED_NODE)
1565 assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Glue &&
1566 "Doesn't have a glue result");
1567 CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1),
1570 // If the node became dead and we haven't already seen it, delete it.
1571 if (FRN->use_empty() &&
1572 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), FRN))
1573 NowDeadNodes.push_back(FRN);
1577 if (!NowDeadNodes.empty())
1578 CurDAG->RemoveDeadNodes(NowDeadNodes, &ISU);
1580 DEBUG(errs() << "ISEL: Match complete!\n");
1586 CR_LeadsToInteriorNode
1589 /// WalkChainUsers - Walk down the users of the specified chained node that is
1590 /// part of the pattern we're matching, looking at all of the users we find.
1591 /// This determines whether something is an interior node, whether we have a
1592 /// non-pattern node in between two pattern nodes (which prevent folding because
1593 /// it would induce a cycle) and whether we have a TokenFactor node sandwiched
1594 /// between pattern nodes (in which case the TF becomes part of the pattern).
1596 /// The walk we do here is guaranteed to be small because we quickly get down to
1597 /// already selected nodes "below" us.
1599 WalkChainUsers(SDNode *ChainedNode,
1600 SmallVectorImpl<SDNode*> &ChainedNodesInPattern,
1601 SmallVectorImpl<SDNode*> &InteriorChainedNodes) {
1602 ChainResult Result = CR_Simple;
1604 for (SDNode::use_iterator UI = ChainedNode->use_begin(),
1605 E = ChainedNode->use_end(); UI != E; ++UI) {
1606 // Make sure the use is of the chain, not some other value we produce.
1607 if (UI.getUse().getValueType() != MVT::Other) continue;
1611 // If we see an already-selected machine node, then we've gone beyond the
1612 // pattern that we're selecting down into the already selected chunk of the
1614 if (User->isMachineOpcode() ||
1615 User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
1618 if (User->getOpcode() == ISD::CopyToReg ||
1619 User->getOpcode() == ISD::CopyFromReg ||
1620 User->getOpcode() == ISD::INLINEASM ||
1621 User->getOpcode() == ISD::EH_LABEL) {
1622 // If their node ID got reset to -1 then they've already been selected.
1623 // Treat them like a MachineOpcode.
1624 if (User->getNodeId() == -1)
1628 // If we have a TokenFactor, we handle it specially.
1629 if (User->getOpcode() != ISD::TokenFactor) {
1630 // If the node isn't a token factor and isn't part of our pattern, then it
1631 // must be a random chained node in between two nodes we're selecting.
1632 // This happens when we have something like:
1637 // Because we structurally match the load/store as a read/modify/write,
1638 // but the call is chained between them. We cannot fold in this case
1639 // because it would induce a cycle in the graph.
1640 if (!std::count(ChainedNodesInPattern.begin(),
1641 ChainedNodesInPattern.end(), User))
1642 return CR_InducesCycle;
1644 // Otherwise we found a node that is part of our pattern. For example in:
1648 // This would happen when we're scanning down from the load and see the
1649 // store as a user. Record that there is a use of ChainedNode that is
1650 // part of the pattern and keep scanning uses.
1651 Result = CR_LeadsToInteriorNode;
1652 InteriorChainedNodes.push_back(User);
1656 // If we found a TokenFactor, there are two cases to consider: first if the
1657 // TokenFactor is just hanging "below" the pattern we're matching (i.e. no
1658 // uses of the TF are in our pattern) we just want to ignore it. Second,
1659 // the TokenFactor can be sandwiched in between two chained nodes, like so:
1665 // | \ DAG's like cheese
1668 // [TokenFactor] [Op]
1675 // In this case, the TokenFactor becomes part of our match and we rewrite it
1676 // as a new TokenFactor.
1678 // To distinguish these two cases, do a recursive walk down the uses.
1679 switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) {
1681 // If the uses of the TokenFactor are just already-selected nodes, ignore
1682 // it, it is "below" our pattern.
1684 case CR_InducesCycle:
1685 // If the uses of the TokenFactor lead to nodes that are not part of our
1686 // pattern that are not selected, folding would turn this into a cycle,
1688 return CR_InducesCycle;
1689 case CR_LeadsToInteriorNode:
1690 break; // Otherwise, keep processing.
1693 // Okay, we know we're in the interesting interior case. The TokenFactor
1694 // is now going to be considered part of the pattern so that we rewrite its
1695 // uses (it may have uses that are not part of the pattern) with the
1696 // ultimate chain result of the generated code. We will also add its chain
1697 // inputs as inputs to the ultimate TokenFactor we create.
1698 Result = CR_LeadsToInteriorNode;
1699 ChainedNodesInPattern.push_back(User);
1700 InteriorChainedNodes.push_back(User);
1707 /// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
1708 /// operation for when the pattern matched at least one node with a chains. The
1709 /// input vector contains a list of all of the chained nodes that we match. We
1710 /// must determine if this is a valid thing to cover (i.e. matching it won't
1711 /// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
1712 /// be used as the input node chain for the generated nodes.
1714 HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
1715 SelectionDAG *CurDAG) {
1716 // Walk all of the chained nodes we've matched, recursively scanning down the
1717 // users of the chain result. This adds any TokenFactor nodes that are caught
1718 // in between chained nodes to the chained and interior nodes list.
1719 SmallVector<SDNode*, 3> InteriorChainedNodes;
1720 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1721 if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched,
1722 InteriorChainedNodes) == CR_InducesCycle)
1723 return SDValue(); // Would induce a cycle.
1726 // Okay, we have walked all the matched nodes and collected TokenFactor nodes
1727 // that we are interested in. Form our input TokenFactor node.
1728 SmallVector<SDValue, 3> InputChains;
1729 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1730 // Add the input chain of this node to the InputChains list (which will be
1731 // the operands of the generated TokenFactor) if it's not an interior node.
1732 SDNode *N = ChainNodesMatched[i];
1733 if (N->getOpcode() != ISD::TokenFactor) {
1734 if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N))
1737 // Otherwise, add the input chain.
1738 SDValue InChain = ChainNodesMatched[i]->getOperand(0);
1739 assert(InChain.getValueType() == MVT::Other && "Not a chain");
1740 InputChains.push_back(InChain);
1744 // If we have a token factor, we want to add all inputs of the token factor
1745 // that are not part of the pattern we're matching.
1746 for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) {
1747 if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(),
1748 N->getOperand(op).getNode()))
1749 InputChains.push_back(N->getOperand(op));
1754 if (InputChains.size() == 1)
1755 return InputChains[0];
1756 return CurDAG->getNode(ISD::TokenFactor, ChainNodesMatched[0]->getDebugLoc(),
1757 MVT::Other, &InputChains[0], InputChains.size());
1760 /// MorphNode - Handle morphing a node in place for the selector.
1761 SDNode *SelectionDAGISel::
1762 MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList,
1763 const SDValue *Ops, unsigned NumOps, unsigned EmitNodeInfo) {
1764 // It is possible we're using MorphNodeTo to replace a node with no
1765 // normal results with one that has a normal result (or we could be
1766 // adding a chain) and the input could have glue and chains as well.
1767 // In this case we need to shift the operands down.
1768 // FIXME: This is a horrible hack and broken in obscure cases, no worse
1769 // than the old isel though.
1770 int OldGlueResultNo = -1, OldChainResultNo = -1;
1772 unsigned NTMNumResults = Node->getNumValues();
1773 if (Node->getValueType(NTMNumResults-1) == MVT::Glue) {
1774 OldGlueResultNo = NTMNumResults-1;
1775 if (NTMNumResults != 1 &&
1776 Node->getValueType(NTMNumResults-2) == MVT::Other)
1777 OldChainResultNo = NTMNumResults-2;
1778 } else if (Node->getValueType(NTMNumResults-1) == MVT::Other)
1779 OldChainResultNo = NTMNumResults-1;
1781 // Call the underlying SelectionDAG routine to do the transmogrification. Note
1782 // that this deletes operands of the old node that become dead.
1783 SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops, NumOps);
1785 // MorphNodeTo can operate in two ways: if an existing node with the
1786 // specified operands exists, it can just return it. Otherwise, it
1787 // updates the node in place to have the requested operands.
1789 // If we updated the node in place, reset the node ID. To the isel,
1790 // this should be just like a newly allocated machine node.
1794 unsigned ResNumResults = Res->getNumValues();
1795 // Move the glue if needed.
1796 if ((EmitNodeInfo & OPFL_GlueOutput) && OldGlueResultNo != -1 &&
1797 (unsigned)OldGlueResultNo != ResNumResults-1)
1798 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldGlueResultNo),
1799 SDValue(Res, ResNumResults-1));
1801 if ((EmitNodeInfo & OPFL_GlueOutput) != 0)
1804 // Move the chain reference if needed.
1805 if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 &&
1806 (unsigned)OldChainResultNo != ResNumResults-1)
1807 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo),
1808 SDValue(Res, ResNumResults-1));
1810 // Otherwise, no replacement happened because the node already exists. Replace
1811 // Uses of the old node with the new one.
1813 CurDAG->ReplaceAllUsesWith(Node, Res);
1818 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1819 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1820 CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1822 const SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes) {
1823 // Accept if it is exactly the same as a previously recorded node.
1824 unsigned RecNo = MatcherTable[MatcherIndex++];
1825 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
1826 return N == RecordedNodes[RecNo].first;
1829 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1830 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1831 CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1832 SelectionDAGISel &SDISel) {
1833 return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]);
1836 /// CheckNodePredicate - Implements OP_CheckNodePredicate.
1837 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1838 CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1839 SelectionDAGISel &SDISel, SDNode *N) {
1840 return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]);
1843 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1844 CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1846 uint16_t Opc = MatcherTable[MatcherIndex++];
1847 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
1848 return N->getOpcode() == Opc;
1851 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1852 CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1853 SDValue N, const TargetLowering &TLI) {
1854 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1855 if (N.getValueType() == VT) return true;
1857 // Handle the case when VT is iPTR.
1858 return VT == MVT::iPTR && N.getValueType() == TLI.getPointerTy();
1861 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1862 CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1863 SDValue N, const TargetLowering &TLI,
1865 if (ChildNo >= N.getNumOperands())
1866 return false; // Match fails if out of range child #.
1867 return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI);
1871 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1872 CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1874 return cast<CondCodeSDNode>(N)->get() ==
1875 (ISD::CondCode)MatcherTable[MatcherIndex++];
1878 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1879 CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1880 SDValue N, const TargetLowering &TLI) {
1881 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1882 if (cast<VTSDNode>(N)->getVT() == VT)
1885 // Handle the case when VT is iPTR.
1886 return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI.getPointerTy();
1889 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1890 CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1892 int64_t Val = MatcherTable[MatcherIndex++];
1894 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1896 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
1897 return C != 0 && C->getSExtValue() == Val;
1900 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1901 CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1902 SDValue N, SelectionDAGISel &SDISel) {
1903 int64_t Val = MatcherTable[MatcherIndex++];
1905 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1907 if (N->getOpcode() != ISD::AND) return false;
1909 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1910 return C != 0 && SDISel.CheckAndMask(N.getOperand(0), C, Val);
1913 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1914 CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1915 SDValue N, SelectionDAGISel &SDISel) {
1916 int64_t Val = MatcherTable[MatcherIndex++];
1918 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1920 if (N->getOpcode() != ISD::OR) return false;
1922 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1923 return C != 0 && SDISel.CheckOrMask(N.getOperand(0), C, Val);
1926 /// IsPredicateKnownToFail - If we know how and can do so without pushing a
1927 /// scope, evaluate the current node. If the current predicate is known to
1928 /// fail, set Result=true and return anything. If the current predicate is
1929 /// known to pass, set Result=false and return the MatcherIndex to continue
1930 /// with. If the current predicate is unknown, set Result=false and return the
1931 /// MatcherIndex to continue with.
1932 static unsigned IsPredicateKnownToFail(const unsigned char *Table,
1933 unsigned Index, SDValue N,
1934 bool &Result, SelectionDAGISel &SDISel,
1935 SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes) {
1936 switch (Table[Index++]) {
1939 return Index-1; // Could not evaluate this predicate.
1940 case SelectionDAGISel::OPC_CheckSame:
1941 Result = !::CheckSame(Table, Index, N, RecordedNodes);
1943 case SelectionDAGISel::OPC_CheckPatternPredicate:
1944 Result = !::CheckPatternPredicate(Table, Index, SDISel);
1946 case SelectionDAGISel::OPC_CheckPredicate:
1947 Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode());
1949 case SelectionDAGISel::OPC_CheckOpcode:
1950 Result = !::CheckOpcode(Table, Index, N.getNode());
1952 case SelectionDAGISel::OPC_CheckType:
1953 Result = !::CheckType(Table, Index, N, SDISel.TLI);
1955 case SelectionDAGISel::OPC_CheckChild0Type:
1956 case SelectionDAGISel::OPC_CheckChild1Type:
1957 case SelectionDAGISel::OPC_CheckChild2Type:
1958 case SelectionDAGISel::OPC_CheckChild3Type:
1959 case SelectionDAGISel::OPC_CheckChild4Type:
1960 case SelectionDAGISel::OPC_CheckChild5Type:
1961 case SelectionDAGISel::OPC_CheckChild6Type:
1962 case SelectionDAGISel::OPC_CheckChild7Type:
1963 Result = !::CheckChildType(Table, Index, N, SDISel.TLI,
1964 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Type);
1966 case SelectionDAGISel::OPC_CheckCondCode:
1967 Result = !::CheckCondCode(Table, Index, N);
1969 case SelectionDAGISel::OPC_CheckValueType:
1970 Result = !::CheckValueType(Table, Index, N, SDISel.TLI);
1972 case SelectionDAGISel::OPC_CheckInteger:
1973 Result = !::CheckInteger(Table, Index, N);
1975 case SelectionDAGISel::OPC_CheckAndImm:
1976 Result = !::CheckAndImm(Table, Index, N, SDISel);
1978 case SelectionDAGISel::OPC_CheckOrImm:
1979 Result = !::CheckOrImm(Table, Index, N, SDISel);
1987 /// FailIndex - If this match fails, this is the index to continue with.
1990 /// NodeStack - The node stack when the scope was formed.
1991 SmallVector<SDValue, 4> NodeStack;
1993 /// NumRecordedNodes - The number of recorded nodes when the scope was formed.
1994 unsigned NumRecordedNodes;
1996 /// NumMatchedMemRefs - The number of matched memref entries.
1997 unsigned NumMatchedMemRefs;
1999 /// InputChain/InputGlue - The current chain/glue
2000 SDValue InputChain, InputGlue;
2002 /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty.
2003 bool HasChainNodesMatched, HasGlueResultNodesMatched;
2008 SDNode *SelectionDAGISel::
2009 SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
2010 unsigned TableSize) {
2011 // FIXME: Should these even be selected? Handle these cases in the caller?
2012 switch (NodeToMatch->getOpcode()) {
2015 case ISD::EntryToken: // These nodes remain the same.
2016 case ISD::BasicBlock:
2018 //case ISD::VALUETYPE:
2019 //case ISD::CONDCODE:
2020 case ISD::HANDLENODE:
2021 case ISD::MDNODE_SDNODE:
2022 case ISD::TargetConstant:
2023 case ISD::TargetConstantFP:
2024 case ISD::TargetConstantPool:
2025 case ISD::TargetFrameIndex:
2026 case ISD::TargetExternalSymbol:
2027 case ISD::TargetBlockAddress:
2028 case ISD::TargetJumpTable:
2029 case ISD::TargetGlobalTLSAddress:
2030 case ISD::TargetGlobalAddress:
2031 case ISD::TokenFactor:
2032 case ISD::CopyFromReg:
2033 case ISD::CopyToReg:
2035 NodeToMatch->setNodeId(-1); // Mark selected.
2037 case ISD::AssertSext:
2038 case ISD::AssertZext:
2039 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0),
2040 NodeToMatch->getOperand(0));
2042 case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch);
2043 case ISD::UNDEF: return Select_UNDEF(NodeToMatch);
2046 assert(!NodeToMatch->isMachineOpcode() && "Node already selected!");
2048 // Set up the node stack with NodeToMatch as the only node on the stack.
2049 SmallVector<SDValue, 8> NodeStack;
2050 SDValue N = SDValue(NodeToMatch, 0);
2051 NodeStack.push_back(N);
2053 // MatchScopes - Scopes used when matching, if a match failure happens, this
2054 // indicates where to continue checking.
2055 SmallVector<MatchScope, 8> MatchScopes;
2057 // RecordedNodes - This is the set of nodes that have been recorded by the
2058 // state machine. The second value is the parent of the node, or null if the
2059 // root is recorded.
2060 SmallVector<std::pair<SDValue, SDNode*>, 8> RecordedNodes;
2062 // MatchedMemRefs - This is the set of MemRef's we've seen in the input
2064 SmallVector<MachineMemOperand*, 2> MatchedMemRefs;
2066 // These are the current input chain and glue for use when generating nodes.
2067 // Various Emit operations change these. For example, emitting a copytoreg
2068 // uses and updates these.
2069 SDValue InputChain, InputGlue;
2071 // ChainNodesMatched - If a pattern matches nodes that have input/output
2072 // chains, the OPC_EmitMergeInputChains operation is emitted which indicates
2073 // which ones they are. The result is captured into this list so that we can
2074 // update the chain results when the pattern is complete.
2075 SmallVector<SDNode*, 3> ChainNodesMatched;
2076 SmallVector<SDNode*, 3> GlueResultNodesMatched;
2078 DEBUG(errs() << "ISEL: Starting pattern match on root node: ";
2079 NodeToMatch->dump(CurDAG);
2082 // Determine where to start the interpreter. Normally we start at opcode #0,
2083 // but if the state machine starts with an OPC_SwitchOpcode, then we
2084 // accelerate the first lookup (which is guaranteed to be hot) with the
2085 // OpcodeOffset table.
2086 unsigned MatcherIndex = 0;
2088 if (!OpcodeOffset.empty()) {
2089 // Already computed the OpcodeOffset table, just index into it.
2090 if (N.getOpcode() < OpcodeOffset.size())
2091 MatcherIndex = OpcodeOffset[N.getOpcode()];
2092 DEBUG(errs() << " Initial Opcode index to " << MatcherIndex << "\n");
2094 } else if (MatcherTable[0] == OPC_SwitchOpcode) {
2095 // Otherwise, the table isn't computed, but the state machine does start
2096 // with an OPC_SwitchOpcode instruction. Populate the table now, since this
2097 // is the first time we're selecting an instruction.
2100 // Get the size of this case.
2101 unsigned CaseSize = MatcherTable[Idx++];
2103 CaseSize = GetVBR(CaseSize, MatcherTable, Idx);
2104 if (CaseSize == 0) break;
2106 // Get the opcode, add the index to the table.
2107 uint16_t Opc = MatcherTable[Idx++];
2108 Opc |= (unsigned short)MatcherTable[Idx++] << 8;
2109 if (Opc >= OpcodeOffset.size())
2110 OpcodeOffset.resize((Opc+1)*2);
2111 OpcodeOffset[Opc] = Idx;
2115 // Okay, do the lookup for the first opcode.
2116 if (N.getOpcode() < OpcodeOffset.size())
2117 MatcherIndex = OpcodeOffset[N.getOpcode()];
2121 assert(MatcherIndex < TableSize && "Invalid index");
2123 unsigned CurrentOpcodeIndex = MatcherIndex;
2125 BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++];
2128 // Okay, the semantics of this operation are that we should push a scope
2129 // then evaluate the first child. However, pushing a scope only to have
2130 // the first check fail (which then pops it) is inefficient. If we can
2131 // determine immediately that the first check (or first several) will
2132 // immediately fail, don't even bother pushing a scope for them.
2136 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2137 if (NumToSkip & 128)
2138 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2139 // Found the end of the scope with no match.
2140 if (NumToSkip == 0) {
2145 FailIndex = MatcherIndex+NumToSkip;
2147 unsigned MatcherIndexOfPredicate = MatcherIndex;
2148 (void)MatcherIndexOfPredicate; // silence warning.
2150 // If we can't evaluate this predicate without pushing a scope (e.g. if
2151 // it is a 'MoveParent') or if the predicate succeeds on this node, we
2152 // push the scope and evaluate the full predicate chain.
2154 MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
2155 Result, *this, RecordedNodes);
2159 DEBUG(errs() << " Skipped scope entry (due to false predicate) at "
2160 << "index " << MatcherIndexOfPredicate
2161 << ", continuing at " << FailIndex << "\n");
2162 ++NumDAGIselRetries;
2164 // Otherwise, we know that this case of the Scope is guaranteed to fail,
2165 // move to the next case.
2166 MatcherIndex = FailIndex;
2169 // If the whole scope failed to match, bail.
2170 if (FailIndex == 0) break;
2172 // Push a MatchScope which indicates where to go if the first child fails
2174 MatchScope NewEntry;
2175 NewEntry.FailIndex = FailIndex;
2176 NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
2177 NewEntry.NumRecordedNodes = RecordedNodes.size();
2178 NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
2179 NewEntry.InputChain = InputChain;
2180 NewEntry.InputGlue = InputGlue;
2181 NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty();
2182 NewEntry.HasGlueResultNodesMatched = !GlueResultNodesMatched.empty();
2183 MatchScopes.push_back(NewEntry);
2186 case OPC_RecordNode: {
2187 // Remember this node, it may end up being an operand in the pattern.
2189 if (NodeStack.size() > 1)
2190 Parent = NodeStack[NodeStack.size()-2].getNode();
2191 RecordedNodes.push_back(std::make_pair(N, Parent));
2195 case OPC_RecordChild0: case OPC_RecordChild1:
2196 case OPC_RecordChild2: case OPC_RecordChild3:
2197 case OPC_RecordChild4: case OPC_RecordChild5:
2198 case OPC_RecordChild6: case OPC_RecordChild7: {
2199 unsigned ChildNo = Opcode-OPC_RecordChild0;
2200 if (ChildNo >= N.getNumOperands())
2201 break; // Match fails if out of range child #.
2203 RecordedNodes.push_back(std::make_pair(N->getOperand(ChildNo),
2207 case OPC_RecordMemRef:
2208 MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand());
2211 case OPC_CaptureGlueInput:
2212 // If the current node has an input glue, capture it in InputGlue.
2213 if (N->getNumOperands() != 0 &&
2214 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue)
2215 InputGlue = N->getOperand(N->getNumOperands()-1);
2218 case OPC_MoveChild: {
2219 unsigned ChildNo = MatcherTable[MatcherIndex++];
2220 if (ChildNo >= N.getNumOperands())
2221 break; // Match fails if out of range child #.
2222 N = N.getOperand(ChildNo);
2223 NodeStack.push_back(N);
2227 case OPC_MoveParent:
2228 // Pop the current node off the NodeStack.
2229 NodeStack.pop_back();
2230 assert(!NodeStack.empty() && "Node stack imbalance!");
2231 N = NodeStack.back();
2235 if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
2237 case OPC_CheckPatternPredicate:
2238 if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break;
2240 case OPC_CheckPredicate:
2241 if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this,
2245 case OPC_CheckComplexPat: {
2246 unsigned CPNum = MatcherTable[MatcherIndex++];
2247 unsigned RecNo = MatcherTable[MatcherIndex++];
2248 assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
2249 if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo].second,
2250 RecordedNodes[RecNo].first, CPNum,
2255 case OPC_CheckOpcode:
2256 if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
2260 if (!::CheckType(MatcherTable, MatcherIndex, N, TLI)) break;
2263 case OPC_SwitchOpcode: {
2264 unsigned CurNodeOpcode = N.getOpcode();
2265 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2268 // Get the size of this case.
2269 CaseSize = MatcherTable[MatcherIndex++];
2271 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2272 if (CaseSize == 0) break;
2274 uint16_t Opc = MatcherTable[MatcherIndex++];
2275 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2277 // If the opcode matches, then we will execute this case.
2278 if (CurNodeOpcode == Opc)
2281 // Otherwise, skip over this case.
2282 MatcherIndex += CaseSize;
2285 // If no cases matched, bail out.
2286 if (CaseSize == 0) break;
2288 // Otherwise, execute the case we found.
2289 DEBUG(errs() << " OpcodeSwitch from " << SwitchStart
2290 << " to " << MatcherIndex << "\n");
2294 case OPC_SwitchType: {
2295 MVT CurNodeVT = N.getValueType().getSimpleVT();
2296 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2299 // Get the size of this case.
2300 CaseSize = MatcherTable[MatcherIndex++];
2302 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2303 if (CaseSize == 0) break;
2305 MVT CaseVT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2306 if (CaseVT == MVT::iPTR)
2307 CaseVT = TLI.getPointerTy();
2309 // If the VT matches, then we will execute this case.
2310 if (CurNodeVT == CaseVT)
2313 // Otherwise, skip over this case.
2314 MatcherIndex += CaseSize;
2317 // If no cases matched, bail out.
2318 if (CaseSize == 0) break;
2320 // Otherwise, execute the case we found.
2321 DEBUG(errs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
2322 << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
2325 case OPC_CheckChild0Type: case OPC_CheckChild1Type:
2326 case OPC_CheckChild2Type: case OPC_CheckChild3Type:
2327 case OPC_CheckChild4Type: case OPC_CheckChild5Type:
2328 case OPC_CheckChild6Type: case OPC_CheckChild7Type:
2329 if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI,
2330 Opcode-OPC_CheckChild0Type))
2333 case OPC_CheckCondCode:
2334 if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
2336 case OPC_CheckValueType:
2337 if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI)) break;
2339 case OPC_CheckInteger:
2340 if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
2342 case OPC_CheckAndImm:
2343 if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
2345 case OPC_CheckOrImm:
2346 if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
2349 case OPC_CheckFoldableChainNode: {
2350 assert(NodeStack.size() != 1 && "No parent node");
2351 // Verify that all intermediate nodes between the root and this one have
2353 bool HasMultipleUses = false;
2354 for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i)
2355 if (!NodeStack[i].hasOneUse()) {
2356 HasMultipleUses = true;
2359 if (HasMultipleUses) break;
2361 // Check to see that the target thinks this is profitable to fold and that
2362 // we can fold it without inducing cycles in the graph.
2363 if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2365 !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2366 NodeToMatch, OptLevel,
2367 true/*We validate our own chains*/))
2372 case OPC_EmitInteger: {
2373 MVT::SimpleValueType VT =
2374 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2375 int64_t Val = MatcherTable[MatcherIndex++];
2377 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2378 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2379 CurDAG->getTargetConstant(Val, VT), (SDNode*)0));
2382 case OPC_EmitRegister: {
2383 MVT::SimpleValueType VT =
2384 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2385 unsigned RegNo = MatcherTable[MatcherIndex++];
2386 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2387 CurDAG->getRegister(RegNo, VT), (SDNode*)0));
2391 case OPC_EmitConvertToTarget: {
2392 // Convert from IMM/FPIMM to target version.
2393 unsigned RecNo = MatcherTable[MatcherIndex++];
2394 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2395 SDValue Imm = RecordedNodes[RecNo].first;
2397 if (Imm->getOpcode() == ISD::Constant) {
2398 int64_t Val = cast<ConstantSDNode>(Imm)->getZExtValue();
2399 Imm = CurDAG->getTargetConstant(Val, Imm.getValueType());
2400 } else if (Imm->getOpcode() == ISD::ConstantFP) {
2401 const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue();
2402 Imm = CurDAG->getTargetConstantFP(*Val, Imm.getValueType());
2405 RecordedNodes.push_back(std::make_pair(Imm, RecordedNodes[RecNo].second));
2409 case OPC_EmitMergeInputChains1_0: // OPC_EmitMergeInputChains, 1, 0
2410 case OPC_EmitMergeInputChains1_1: { // OPC_EmitMergeInputChains, 1, 1
2411 // These are space-optimized forms of OPC_EmitMergeInputChains.
2412 assert(InputChain.getNode() == 0 &&
2413 "EmitMergeInputChains should be the first chain producing node");
2414 assert(ChainNodesMatched.empty() &&
2415 "Should only have one EmitMergeInputChains per match");
2417 // Read all of the chained nodes.
2418 unsigned RecNo = Opcode == OPC_EmitMergeInputChains1_1;
2419 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2420 ChainNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
2422 // FIXME: What if other value results of the node have uses not matched
2424 if (ChainNodesMatched.back() != NodeToMatch &&
2425 !RecordedNodes[RecNo].first.hasOneUse()) {
2426 ChainNodesMatched.clear();
2430 // Merge the input chains if they are not intra-pattern references.
2431 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2433 if (InputChain.getNode() == 0)
2434 break; // Failed to merge.
2438 case OPC_EmitMergeInputChains: {
2439 assert(InputChain.getNode() == 0 &&
2440 "EmitMergeInputChains should be the first chain producing node");
2441 // This node gets a list of nodes we matched in the input that have
2442 // chains. We want to token factor all of the input chains to these nodes
2443 // together. However, if any of the input chains is actually one of the
2444 // nodes matched in this pattern, then we have an intra-match reference.
2445 // Ignore these because the newly token factored chain should not refer to
2447 unsigned NumChains = MatcherTable[MatcherIndex++];
2448 assert(NumChains != 0 && "Can't TF zero chains");
2450 assert(ChainNodesMatched.empty() &&
2451 "Should only have one EmitMergeInputChains per match");
2453 // Read all of the chained nodes.
2454 for (unsigned i = 0; i != NumChains; ++i) {
2455 unsigned RecNo = MatcherTable[MatcherIndex++];
2456 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2457 ChainNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
2459 // FIXME: What if other value results of the node have uses not matched
2461 if (ChainNodesMatched.back() != NodeToMatch &&
2462 !RecordedNodes[RecNo].first.hasOneUse()) {
2463 ChainNodesMatched.clear();
2468 // If the inner loop broke out, the match fails.
2469 if (ChainNodesMatched.empty())
2472 // Merge the input chains if they are not intra-pattern references.
2473 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2475 if (InputChain.getNode() == 0)
2476 break; // Failed to merge.
2481 case OPC_EmitCopyToReg: {
2482 unsigned RecNo = MatcherTable[MatcherIndex++];
2483 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2484 unsigned DestPhysReg = MatcherTable[MatcherIndex++];
2486 if (InputChain.getNode() == 0)
2487 InputChain = CurDAG->getEntryNode();
2489 InputChain = CurDAG->getCopyToReg(InputChain, NodeToMatch->getDebugLoc(),
2490 DestPhysReg, RecordedNodes[RecNo].first,
2493 InputGlue = InputChain.getValue(1);
2497 case OPC_EmitNodeXForm: {
2498 unsigned XFormNo = MatcherTable[MatcherIndex++];
2499 unsigned RecNo = MatcherTable[MatcherIndex++];
2500 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2501 SDValue Res = RunSDNodeXForm(RecordedNodes[RecNo].first, XFormNo);
2502 RecordedNodes.push_back(std::pair<SDValue,SDNode*>(Res, (SDNode*) 0));
2507 case OPC_MorphNodeTo: {
2508 uint16_t TargetOpc = MatcherTable[MatcherIndex++];
2509 TargetOpc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2510 unsigned EmitNodeInfo = MatcherTable[MatcherIndex++];
2511 // Get the result VT list.
2512 unsigned NumVTs = MatcherTable[MatcherIndex++];
2513 SmallVector<EVT, 4> VTs;
2514 for (unsigned i = 0; i != NumVTs; ++i) {
2515 MVT::SimpleValueType VT =
2516 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2517 if (VT == MVT::iPTR) VT = TLI.getPointerTy().SimpleTy;
2521 if (EmitNodeInfo & OPFL_Chain)
2522 VTs.push_back(MVT::Other);
2523 if (EmitNodeInfo & OPFL_GlueOutput)
2524 VTs.push_back(MVT::Glue);
2526 // This is hot code, so optimize the two most common cases of 1 and 2
2529 if (VTs.size() == 1)
2530 VTList = CurDAG->getVTList(VTs[0]);
2531 else if (VTs.size() == 2)
2532 VTList = CurDAG->getVTList(VTs[0], VTs[1]);
2534 VTList = CurDAG->getVTList(VTs.data(), VTs.size());
2536 // Get the operand list.
2537 unsigned NumOps = MatcherTable[MatcherIndex++];
2538 SmallVector<SDValue, 8> Ops;
2539 for (unsigned i = 0; i != NumOps; ++i) {
2540 unsigned RecNo = MatcherTable[MatcherIndex++];
2542 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2544 assert(RecNo < RecordedNodes.size() && "Invalid EmitNode");
2545 Ops.push_back(RecordedNodes[RecNo].first);
2548 // If there are variadic operands to add, handle them now.
2549 if (EmitNodeInfo & OPFL_VariadicInfo) {
2550 // Determine the start index to copy from.
2551 unsigned FirstOpToCopy = getNumFixedFromVariadicInfo(EmitNodeInfo);
2552 FirstOpToCopy += (EmitNodeInfo & OPFL_Chain) ? 1 : 0;
2553 assert(NodeToMatch->getNumOperands() >= FirstOpToCopy &&
2554 "Invalid variadic node");
2555 // Copy all of the variadic operands, not including a potential glue
2557 for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands();
2559 SDValue V = NodeToMatch->getOperand(i);
2560 if (V.getValueType() == MVT::Glue) break;
2565 // If this has chain/glue inputs, add them.
2566 if (EmitNodeInfo & OPFL_Chain)
2567 Ops.push_back(InputChain);
2568 if ((EmitNodeInfo & OPFL_GlueInput) && InputGlue.getNode() != 0)
2569 Ops.push_back(InputGlue);
2573 if (Opcode != OPC_MorphNodeTo) {
2574 // If this is a normal EmitNode command, just create the new node and
2575 // add the results to the RecordedNodes list.
2576 Res = CurDAG->getMachineNode(TargetOpc, NodeToMatch->getDebugLoc(),
2577 VTList, Ops.data(), Ops.size());
2579 // Add all the non-glue/non-chain results to the RecordedNodes list.
2580 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
2581 if (VTs[i] == MVT::Other || VTs[i] == MVT::Glue) break;
2582 RecordedNodes.push_back(std::pair<SDValue,SDNode*>(SDValue(Res, i),
2587 Res = MorphNode(NodeToMatch, TargetOpc, VTList, Ops.data(), Ops.size(),
2591 // If the node had chain/glue results, update our notion of the current
2593 if (EmitNodeInfo & OPFL_GlueOutput) {
2594 InputGlue = SDValue(Res, VTs.size()-1);
2595 if (EmitNodeInfo & OPFL_Chain)
2596 InputChain = SDValue(Res, VTs.size()-2);
2597 } else if (EmitNodeInfo & OPFL_Chain)
2598 InputChain = SDValue(Res, VTs.size()-1);
2600 // If the OPFL_MemRefs glue is set on this node, slap all of the
2601 // accumulated memrefs onto it.
2603 // FIXME: This is vastly incorrect for patterns with multiple outputs
2604 // instructions that access memory and for ComplexPatterns that match
2606 if (EmitNodeInfo & OPFL_MemRefs) {
2607 MachineSDNode::mmo_iterator MemRefs =
2608 MF->allocateMemRefsArray(MatchedMemRefs.size());
2609 std::copy(MatchedMemRefs.begin(), MatchedMemRefs.end(), MemRefs);
2610 cast<MachineSDNode>(Res)
2611 ->setMemRefs(MemRefs, MemRefs + MatchedMemRefs.size());
2615 << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created")
2616 << " node: "; Res->dump(CurDAG); errs() << "\n");
2618 // If this was a MorphNodeTo then we're completely done!
2619 if (Opcode == OPC_MorphNodeTo) {
2620 // Update chain and glue uses.
2621 UpdateChainsAndGlue(NodeToMatch, InputChain, ChainNodesMatched,
2622 InputGlue, GlueResultNodesMatched, true);
2629 case OPC_MarkGlueResults: {
2630 unsigned NumNodes = MatcherTable[MatcherIndex++];
2632 // Read and remember all the glue-result nodes.
2633 for (unsigned i = 0; i != NumNodes; ++i) {
2634 unsigned RecNo = MatcherTable[MatcherIndex++];
2636 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2638 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2639 GlueResultNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
2644 case OPC_CompleteMatch: {
2645 // The match has been completed, and any new nodes (if any) have been
2646 // created. Patch up references to the matched dag to use the newly
2648 unsigned NumResults = MatcherTable[MatcherIndex++];
2650 for (unsigned i = 0; i != NumResults; ++i) {
2651 unsigned ResSlot = MatcherTable[MatcherIndex++];
2653 ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex);
2655 assert(ResSlot < RecordedNodes.size() && "Invalid CheckSame");
2656 SDValue Res = RecordedNodes[ResSlot].first;
2658 assert(i < NodeToMatch->getNumValues() &&
2659 NodeToMatch->getValueType(i) != MVT::Other &&
2660 NodeToMatch->getValueType(i) != MVT::Glue &&
2661 "Invalid number of results to complete!");
2662 assert((NodeToMatch->getValueType(i) == Res.getValueType() ||
2663 NodeToMatch->getValueType(i) == MVT::iPTR ||
2664 Res.getValueType() == MVT::iPTR ||
2665 NodeToMatch->getValueType(i).getSizeInBits() ==
2666 Res.getValueType().getSizeInBits()) &&
2667 "invalid replacement");
2668 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, i), Res);
2671 // If the root node defines glue, add it to the glue nodes to update list.
2672 if (NodeToMatch->getValueType(NodeToMatch->getNumValues()-1) == MVT::Glue)
2673 GlueResultNodesMatched.push_back(NodeToMatch);
2675 // Update chain and glue uses.
2676 UpdateChainsAndGlue(NodeToMatch, InputChain, ChainNodesMatched,
2677 InputGlue, GlueResultNodesMatched, false);
2679 assert(NodeToMatch->use_empty() &&
2680 "Didn't replace all uses of the node?");
2682 // FIXME: We just return here, which interacts correctly with SelectRoot
2683 // above. We should fix this to not return an SDNode* anymore.
2688 // If the code reached this point, then the match failed. See if there is
2689 // another child to try in the current 'Scope', otherwise pop it until we
2690 // find a case to check.
2691 DEBUG(errs() << " Match failed at index " << CurrentOpcodeIndex << "\n");
2692 ++NumDAGIselRetries;
2694 if (MatchScopes.empty()) {
2695 CannotYetSelect(NodeToMatch);
2699 // Restore the interpreter state back to the point where the scope was
2701 MatchScope &LastScope = MatchScopes.back();
2702 RecordedNodes.resize(LastScope.NumRecordedNodes);
2704 NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end());
2705 N = NodeStack.back();
2707 if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size())
2708 MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
2709 MatcherIndex = LastScope.FailIndex;
2711 DEBUG(errs() << " Continuing at " << MatcherIndex << "\n");
2713 InputChain = LastScope.InputChain;
2714 InputGlue = LastScope.InputGlue;
2715 if (!LastScope.HasChainNodesMatched)
2716 ChainNodesMatched.clear();
2717 if (!LastScope.HasGlueResultNodesMatched)
2718 GlueResultNodesMatched.clear();
2720 // Check to see what the offset is at the new MatcherIndex. If it is zero
2721 // we have reached the end of this scope, otherwise we have another child
2722 // in the current scope to try.
2723 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2724 if (NumToSkip & 128)
2725 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2727 // If we have another child in this scope to match, update FailIndex and
2729 if (NumToSkip != 0) {
2730 LastScope.FailIndex = MatcherIndex+NumToSkip;
2734 // End of this scope, pop it and try the next child in the containing
2736 MatchScopes.pop_back();
2743 void SelectionDAGISel::CannotYetSelect(SDNode *N) {
2745 raw_string_ostream Msg(msg);
2746 Msg << "Cannot select: ";
2748 if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN &&
2749 N->getOpcode() != ISD::INTRINSIC_WO_CHAIN &&
2750 N->getOpcode() != ISD::INTRINSIC_VOID) {
2751 N->printrFull(Msg, CurDAG);
2753 bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other;
2755 cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue();
2756 if (iid < Intrinsic::num_intrinsics)
2757 Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid);
2758 else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo())
2759 Msg << "target intrinsic %" << TII->getName(iid);
2761 Msg << "unknown intrinsic #" << iid;
2763 report_fatal_error(Msg.str());
2766 char SelectionDAGISel::ID = 0;