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->ComputePHILiveOutRegInfo(cast<PHINode>(I));
851 for (BasicBlock::const_iterator I = LLVMBB->begin(), E = LLVMBB->end();
852 I != E && isa<PHINode>(I); ++I) {
853 FuncInfo->InvalidatePHILiveOutRegInfo(cast<PHINode>(I));
857 FuncInfo->VisitedBBs.insert(LLVMBB);
860 FuncInfo->MBB = FuncInfo->MBBMap[LLVMBB];
861 FuncInfo->InsertPt = FuncInfo->MBB->getFirstNonPHI();
863 BasicBlock::const_iterator const Begin = LLVMBB->getFirstNonPHI();
864 BasicBlock::const_iterator const End = LLVMBB->end();
865 BasicBlock::const_iterator BI = End;
867 FuncInfo->InsertPt = FuncInfo->MBB->getFirstNonPHI();
869 // Setup an EH landing-pad block.
870 if (FuncInfo->MBB->isLandingPad())
871 PrepareEHLandingPad();
873 // Lower any arguments needed in this block if this is the entry block.
874 if (LLVMBB == &Fn.getEntryBlock())
875 LowerArguments(LLVMBB);
877 // Before doing SelectionDAG ISel, see if FastISel has been requested.
879 FastIS->startNewBlock();
881 // Emit code for any incoming arguments. This must happen before
882 // beginning FastISel on the entry block.
883 if (LLVMBB == &Fn.getEntryBlock()) {
884 CurDAG->setRoot(SDB->getControlRoot());
888 // If we inserted any instructions at the beginning, make a note of
889 // where they are, so we can be sure to emit subsequent instructions
891 if (FuncInfo->InsertPt != FuncInfo->MBB->begin())
892 FastIS->setLastLocalValue(llvm::prior(FuncInfo->InsertPt));
894 FastIS->setLastLocalValue(0);
897 // Do FastISel on as many instructions as possible.
898 for (; BI != Begin; --BI) {
899 const Instruction *Inst = llvm::prior(BI);
901 // If we no longer require this instruction, skip it.
902 if (!Inst->mayWriteToMemory() &&
903 !isa<TerminatorInst>(Inst) &&
904 !isa<DbgInfoIntrinsic>(Inst) &&
905 !FuncInfo->isExportedInst(Inst))
908 // Bottom-up: reset the insert pos at the top, after any local-value
910 FastIS->recomputeInsertPt();
912 // Try to select the instruction with FastISel.
913 if (FastIS->SelectInstruction(Inst)) {
914 // If fast isel succeeded, check to see if there is a single-use
915 // non-volatile load right before the selected instruction, and see if
916 // the load is used by the instruction. If so, try to fold it.
917 const Instruction *BeforeInst = 0;
919 BeforeInst = llvm::prior(llvm::prior(BI));
920 if (BeforeInst && isa<LoadInst>(BeforeInst) &&
921 BeforeInst->hasOneUse() && *BeforeInst->use_begin() == Inst &&
922 TryToFoldFastISelLoad(cast<LoadInst>(BeforeInst), FastIS))
923 --BI; // If we succeeded, don't re-select the load.
927 // Then handle certain instructions as single-LLVM-Instruction blocks.
928 if (isa<CallInst>(Inst)) {
929 ++NumFastIselFailures;
930 if (EnableFastISelVerbose || EnableFastISelAbort) {
931 dbgs() << "FastISel missed call: ";
935 if (!Inst->getType()->isVoidTy() && !Inst->use_empty()) {
936 unsigned &R = FuncInfo->ValueMap[Inst];
938 R = FuncInfo->CreateRegs(Inst->getType());
941 bool HadTailCall = false;
942 SelectBasicBlock(Inst, BI, HadTailCall);
944 // If the call was emitted as a tail call, we're done with the block.
953 // Otherwise, give up on FastISel for the rest of the block.
954 // For now, be a little lenient about non-branch terminators.
955 if (!isa<TerminatorInst>(Inst) || isa<BranchInst>(Inst)) {
956 ++NumFastIselFailures;
957 if (EnableFastISelVerbose || EnableFastISelAbort) {
958 dbgs() << "FastISel miss: ";
961 if (EnableFastISelAbort)
962 // The "fast" selector couldn't handle something and bailed.
963 // For the purpose of debugging, just abort.
964 llvm_unreachable("FastISel didn't select the entire block");
969 FastIS->recomputeInsertPt();
977 // Run SelectionDAG instruction selection on the remainder of the block
978 // not handled by FastISel. If FastISel is not run, this is the entire
981 SelectBasicBlock(Begin, BI, HadTailCall);
984 FuncInfo->PHINodesToUpdate.clear();
989 for (MachineFunction::const_iterator MBI = MF->begin(), MBE = MF->end();
991 CheckLineNumbers(MBI);
996 SelectionDAGISel::FinishBasicBlock() {
998 DEBUG(dbgs() << "Total amount of phi nodes to update: "
999 << FuncInfo->PHINodesToUpdate.size() << "\n";
1000 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i)
1001 dbgs() << "Node " << i << " : ("
1002 << FuncInfo->PHINodesToUpdate[i].first
1003 << ", " << FuncInfo->PHINodesToUpdate[i].second << ")\n");
1005 // Next, now that we know what the last MBB the LLVM BB expanded is, update
1006 // PHI nodes in successors.
1007 if (SDB->SwitchCases.empty() &&
1008 SDB->JTCases.empty() &&
1009 SDB->BitTestCases.empty()) {
1010 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i) {
1011 MachineInstr *PHI = FuncInfo->PHINodesToUpdate[i].first;
1012 assert(PHI->isPHI() &&
1013 "This is not a machine PHI node that we are updating!");
1014 if (!FuncInfo->MBB->isSuccessor(PHI->getParent()))
1017 MachineOperand::CreateReg(FuncInfo->PHINodesToUpdate[i].second, false));
1018 PHI->addOperand(MachineOperand::CreateMBB(FuncInfo->MBB));
1023 for (unsigned i = 0, e = SDB->BitTestCases.size(); i != e; ++i) {
1024 // Lower header first, if it wasn't already lowered
1025 if (!SDB->BitTestCases[i].Emitted) {
1026 // Set the current basic block to the mbb we wish to insert the code into
1027 FuncInfo->MBB = SDB->BitTestCases[i].Parent;
1028 FuncInfo->InsertPt = FuncInfo->MBB->end();
1030 SDB->visitBitTestHeader(SDB->BitTestCases[i], FuncInfo->MBB);
1031 CurDAG->setRoot(SDB->getRoot());
1033 CodeGenAndEmitDAG();
1036 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j) {
1037 // Set the current basic block to the mbb we wish to insert the code into
1038 FuncInfo->MBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1039 FuncInfo->InsertPt = FuncInfo->MBB->end();
1042 SDB->visitBitTestCase(SDB->BitTestCases[i],
1043 SDB->BitTestCases[i].Cases[j+1].ThisBB,
1044 SDB->BitTestCases[i].Reg,
1045 SDB->BitTestCases[i].Cases[j],
1048 SDB->visitBitTestCase(SDB->BitTestCases[i],
1049 SDB->BitTestCases[i].Default,
1050 SDB->BitTestCases[i].Reg,
1051 SDB->BitTestCases[i].Cases[j],
1055 CurDAG->setRoot(SDB->getRoot());
1057 CodeGenAndEmitDAG();
1061 for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1063 MachineInstr *PHI = FuncInfo->PHINodesToUpdate[pi].first;
1064 MachineBasicBlock *PHIBB = PHI->getParent();
1065 assert(PHI->isPHI() &&
1066 "This is not a machine PHI node that we are updating!");
1067 // This is "default" BB. We have two jumps to it. From "header" BB and
1068 // from last "case" BB.
1069 if (PHIBB == SDB->BitTestCases[i].Default) {
1070 PHI->addOperand(MachineOperand::
1071 CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1073 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Parent));
1074 PHI->addOperand(MachineOperand::
1075 CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1077 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Cases.
1080 // One of "cases" BB.
1081 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size();
1083 MachineBasicBlock* cBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1084 if (cBB->isSuccessor(PHIBB)) {
1085 PHI->addOperand(MachineOperand::
1086 CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1088 PHI->addOperand(MachineOperand::CreateMBB(cBB));
1093 SDB->BitTestCases.clear();
1095 // If the JumpTable record is filled in, then we need to emit a jump table.
1096 // Updating the PHI nodes is tricky in this case, since we need to determine
1097 // whether the PHI is a successor of the range check MBB or the jump table MBB
1098 for (unsigned i = 0, e = SDB->JTCases.size(); i != e; ++i) {
1099 // Lower header first, if it wasn't already lowered
1100 if (!SDB->JTCases[i].first.Emitted) {
1101 // Set the current basic block to the mbb we wish to insert the code into
1102 FuncInfo->MBB = SDB->JTCases[i].first.HeaderBB;
1103 FuncInfo->InsertPt = FuncInfo->MBB->end();
1105 SDB->visitJumpTableHeader(SDB->JTCases[i].second, SDB->JTCases[i].first,
1107 CurDAG->setRoot(SDB->getRoot());
1109 CodeGenAndEmitDAG();
1112 // Set the current basic block to the mbb we wish to insert the code into
1113 FuncInfo->MBB = SDB->JTCases[i].second.MBB;
1114 FuncInfo->InsertPt = FuncInfo->MBB->end();
1116 SDB->visitJumpTable(SDB->JTCases[i].second);
1117 CurDAG->setRoot(SDB->getRoot());
1119 CodeGenAndEmitDAG();
1122 for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1124 MachineInstr *PHI = FuncInfo->PHINodesToUpdate[pi].first;
1125 MachineBasicBlock *PHIBB = PHI->getParent();
1126 assert(PHI->isPHI() &&
1127 "This is not a machine PHI node that we are updating!");
1128 // "default" BB. We can go there only from header BB.
1129 if (PHIBB == SDB->JTCases[i].second.Default) {
1131 (MachineOperand::CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1134 (MachineOperand::CreateMBB(SDB->JTCases[i].first.HeaderBB));
1136 // JT BB. Just iterate over successors here
1137 if (FuncInfo->MBB->isSuccessor(PHIBB)) {
1139 (MachineOperand::CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1141 PHI->addOperand(MachineOperand::CreateMBB(FuncInfo->MBB));
1145 SDB->JTCases.clear();
1147 // If the switch block involved a branch to one of the actual successors, we
1148 // need to update PHI nodes in that block.
1149 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i) {
1150 MachineInstr *PHI = FuncInfo->PHINodesToUpdate[i].first;
1151 assert(PHI->isPHI() &&
1152 "This is not a machine PHI node that we are updating!");
1153 if (FuncInfo->MBB->isSuccessor(PHI->getParent())) {
1155 MachineOperand::CreateReg(FuncInfo->PHINodesToUpdate[i].second, false));
1156 PHI->addOperand(MachineOperand::CreateMBB(FuncInfo->MBB));
1160 // If we generated any switch lowering information, build and codegen any
1161 // additional DAGs necessary.
1162 for (unsigned i = 0, e = SDB->SwitchCases.size(); i != e; ++i) {
1163 // Set the current basic block to the mbb we wish to insert the code into
1164 FuncInfo->MBB = SDB->SwitchCases[i].ThisBB;
1165 FuncInfo->InsertPt = FuncInfo->MBB->end();
1167 // Determine the unique successors.
1168 SmallVector<MachineBasicBlock *, 2> Succs;
1169 Succs.push_back(SDB->SwitchCases[i].TrueBB);
1170 if (SDB->SwitchCases[i].TrueBB != SDB->SwitchCases[i].FalseBB)
1171 Succs.push_back(SDB->SwitchCases[i].FalseBB);
1173 // Emit the code. Note that this could result in FuncInfo->MBB being split.
1174 SDB->visitSwitchCase(SDB->SwitchCases[i], FuncInfo->MBB);
1175 CurDAG->setRoot(SDB->getRoot());
1177 CodeGenAndEmitDAG();
1179 // Remember the last block, now that any splitting is done, for use in
1180 // populating PHI nodes in successors.
1181 MachineBasicBlock *ThisBB = FuncInfo->MBB;
1183 // Handle any PHI nodes in successors of this chunk, as if we were coming
1184 // from the original BB before switch expansion. Note that PHI nodes can
1185 // occur multiple times in PHINodesToUpdate. We have to be very careful to
1186 // handle them the right number of times.
1187 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1188 FuncInfo->MBB = Succs[i];
1189 FuncInfo->InsertPt = FuncInfo->MBB->end();
1190 // FuncInfo->MBB may have been removed from the CFG if a branch was
1192 if (ThisBB->isSuccessor(FuncInfo->MBB)) {
1193 for (MachineBasicBlock::iterator Phi = FuncInfo->MBB->begin();
1194 Phi != FuncInfo->MBB->end() && Phi->isPHI();
1196 // This value for this PHI node is recorded in PHINodesToUpdate.
1197 for (unsigned pn = 0; ; ++pn) {
1198 assert(pn != FuncInfo->PHINodesToUpdate.size() &&
1199 "Didn't find PHI entry!");
1200 if (FuncInfo->PHINodesToUpdate[pn].first == Phi) {
1201 Phi->addOperand(MachineOperand::
1202 CreateReg(FuncInfo->PHINodesToUpdate[pn].second,
1204 Phi->addOperand(MachineOperand::CreateMBB(ThisBB));
1212 SDB->SwitchCases.clear();
1216 /// Create the scheduler. If a specific scheduler was specified
1217 /// via the SchedulerRegistry, use it, otherwise select the
1218 /// one preferred by the target.
1220 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
1221 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1225 RegisterScheduler::setDefault(Ctor);
1228 return Ctor(this, OptLevel);
1231 //===----------------------------------------------------------------------===//
1232 // Helper functions used by the generated instruction selector.
1233 //===----------------------------------------------------------------------===//
1234 // Calls to these methods are generated by tblgen.
1236 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1237 /// the dag combiner simplified the 255, we still want to match. RHS is the
1238 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1239 /// specified in the .td file (e.g. 255).
1240 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1241 int64_t DesiredMaskS) const {
1242 const APInt &ActualMask = RHS->getAPIntValue();
1243 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1245 // If the actual mask exactly matches, success!
1246 if (ActualMask == DesiredMask)
1249 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1250 if (ActualMask.intersects(~DesiredMask))
1253 // Otherwise, the DAG Combiner may have proven that the value coming in is
1254 // either already zero or is not demanded. Check for known zero input bits.
1255 APInt NeededMask = DesiredMask & ~ActualMask;
1256 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1259 // TODO: check to see if missing bits are just not demanded.
1261 // Otherwise, this pattern doesn't match.
1265 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1266 /// the dag combiner simplified the 255, we still want to match. RHS is the
1267 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1268 /// specified in the .td file (e.g. 255).
1269 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1270 int64_t DesiredMaskS) const {
1271 const APInt &ActualMask = RHS->getAPIntValue();
1272 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1274 // If the actual mask exactly matches, success!
1275 if (ActualMask == DesiredMask)
1278 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1279 if (ActualMask.intersects(~DesiredMask))
1282 // Otherwise, the DAG Combiner may have proven that the value coming in is
1283 // either already zero or is not demanded. Check for known zero input bits.
1284 APInt NeededMask = DesiredMask & ~ActualMask;
1286 APInt KnownZero, KnownOne;
1287 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
1289 // If all the missing bits in the or are already known to be set, match!
1290 if ((NeededMask & KnownOne) == NeededMask)
1293 // TODO: check to see if missing bits are just not demanded.
1295 // Otherwise, this pattern doesn't match.
1300 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1301 /// by tblgen. Others should not call it.
1302 void SelectionDAGISel::
1303 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1304 std::vector<SDValue> InOps;
1305 std::swap(InOps, Ops);
1307 Ops.push_back(InOps[InlineAsm::Op_InputChain]); // 0
1308 Ops.push_back(InOps[InlineAsm::Op_AsmString]); // 1
1309 Ops.push_back(InOps[InlineAsm::Op_MDNode]); // 2, !srcloc
1310 Ops.push_back(InOps[InlineAsm::Op_ExtraInfo]); // 3 (SideEffect, AlignStack)
1312 unsigned i = InlineAsm::Op_FirstOperand, e = InOps.size();
1313 if (InOps[e-1].getValueType() == MVT::Glue)
1314 --e; // Don't process a glue operand if it is here.
1317 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1318 if (!InlineAsm::isMemKind(Flags)) {
1319 // Just skip over this operand, copying the operands verbatim.
1320 Ops.insert(Ops.end(), InOps.begin()+i,
1321 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
1322 i += InlineAsm::getNumOperandRegisters(Flags) + 1;
1324 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 &&
1325 "Memory operand with multiple values?");
1326 // Otherwise, this is a memory operand. Ask the target to select it.
1327 std::vector<SDValue> SelOps;
1328 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps))
1329 report_fatal_error("Could not match memory address. Inline asm"
1332 // Add this to the output node.
1334 InlineAsm::getFlagWord(InlineAsm::Kind_Mem, SelOps.size());
1335 Ops.push_back(CurDAG->getTargetConstant(NewFlags, MVT::i32));
1336 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1341 // Add the glue input back if present.
1342 if (e != InOps.size())
1343 Ops.push_back(InOps.back());
1346 /// findGlueUse - Return use of MVT::Glue value produced by the specified
1349 static SDNode *findGlueUse(SDNode *N) {
1350 unsigned FlagResNo = N->getNumValues()-1;
1351 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1352 SDUse &Use = I.getUse();
1353 if (Use.getResNo() == FlagResNo)
1354 return Use.getUser();
1359 /// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
1360 /// This function recursively traverses up the operand chain, ignoring
1362 static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
1363 SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited,
1364 bool IgnoreChains) {
1365 // The NodeID's are given uniques ID's where a node ID is guaranteed to be
1366 // greater than all of its (recursive) operands. If we scan to a point where
1367 // 'use' is smaller than the node we're scanning for, then we know we will
1370 // The Use may be -1 (unassigned) if it is a newly allocated node. This can
1371 // happen because we scan down to newly selected nodes in the case of glue
1373 if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1))
1376 // Don't revisit nodes if we already scanned it and didn't fail, we know we
1377 // won't fail if we scan it again.
1378 if (!Visited.insert(Use))
1381 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
1382 // Ignore chain uses, they are validated by HandleMergeInputChains.
1383 if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains)
1386 SDNode *N = Use->getOperand(i).getNode();
1388 if (Use == ImmedUse || Use == Root)
1389 continue; // We are not looking for immediate use.
1394 // Traverse up the operand chain.
1395 if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
1401 /// IsProfitableToFold - Returns true if it's profitable to fold the specific
1402 /// operand node N of U during instruction selection that starts at Root.
1403 bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
1404 SDNode *Root) const {
1405 if (OptLevel == CodeGenOpt::None) return false;
1406 return N.hasOneUse();
1409 /// IsLegalToFold - Returns true if the specific operand node N of
1410 /// U can be folded during instruction selection that starts at Root.
1411 bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
1412 CodeGenOpt::Level OptLevel,
1413 bool IgnoreChains) {
1414 if (OptLevel == CodeGenOpt::None) return false;
1416 // If Root use can somehow reach N through a path that that doesn't contain
1417 // U then folding N would create a cycle. e.g. In the following
1418 // diagram, Root can reach N through X. If N is folded into into Root, then
1419 // X is both a predecessor and a successor of U.
1430 // * indicates nodes to be folded together.
1432 // If Root produces glue, then it gets (even more) interesting. Since it
1433 // will be "glued" together with its glue use in the scheduler, we need to
1434 // check if it might reach N.
1453 // If GU (glue use) indirectly reaches N (the load), and Root folds N
1454 // (call it Fold), then X is a predecessor of GU and a successor of
1455 // Fold. But since Fold and GU are glued together, this will create
1456 // a cycle in the scheduling graph.
1458 // If the node has glue, walk down the graph to the "lowest" node in the
1460 EVT VT = Root->getValueType(Root->getNumValues()-1);
1461 while (VT == MVT::Glue) {
1462 SDNode *GU = findGlueUse(Root);
1466 VT = Root->getValueType(Root->getNumValues()-1);
1468 // If our query node has a glue result with a use, we've walked up it. If
1469 // the user (which has already been selected) has a chain or indirectly uses
1470 // the chain, our WalkChainUsers predicate will not consider it. Because of
1471 // this, we cannot ignore chains in this predicate.
1472 IgnoreChains = false;
1476 SmallPtrSet<SDNode*, 16> Visited;
1477 return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains);
1480 SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
1481 std::vector<SDValue> Ops(N->op_begin(), N->op_end());
1482 SelectInlineAsmMemoryOperands(Ops);
1484 std::vector<EVT> VTs;
1485 VTs.push_back(MVT::Other);
1486 VTs.push_back(MVT::Glue);
1487 SDValue New = CurDAG->getNode(ISD::INLINEASM, N->getDebugLoc(),
1488 VTs, &Ops[0], Ops.size());
1490 return New.getNode();
1493 SDNode *SelectionDAGISel::Select_UNDEF(SDNode *N) {
1494 return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0));
1497 /// GetVBR - decode a vbr encoding whose top bit is set.
1498 LLVM_ATTRIBUTE_ALWAYS_INLINE static uint64_t
1499 GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
1500 assert(Val >= 128 && "Not a VBR");
1501 Val &= 127; // Remove first vbr bit.
1506 NextBits = MatcherTable[Idx++];
1507 Val |= (NextBits&127) << Shift;
1509 } while (NextBits & 128);
1515 /// UpdateChainsAndGlue - When a match is complete, this method updates uses of
1516 /// interior glue and chain results to use the new glue and chain results.
1517 void SelectionDAGISel::
1518 UpdateChainsAndGlue(SDNode *NodeToMatch, SDValue InputChain,
1519 const SmallVectorImpl<SDNode*> &ChainNodesMatched,
1521 const SmallVectorImpl<SDNode*> &GlueResultNodesMatched,
1522 bool isMorphNodeTo) {
1523 SmallVector<SDNode*, 4> NowDeadNodes;
1525 ISelUpdater ISU(ISelPosition);
1527 // Now that all the normal results are replaced, we replace the chain and
1528 // glue results if present.
1529 if (!ChainNodesMatched.empty()) {
1530 assert(InputChain.getNode() != 0 &&
1531 "Matched input chains but didn't produce a chain");
1532 // Loop over all of the nodes we matched that produced a chain result.
1533 // Replace all the chain results with the final chain we ended up with.
1534 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1535 SDNode *ChainNode = ChainNodesMatched[i];
1537 // If this node was already deleted, don't look at it.
1538 if (ChainNode->getOpcode() == ISD::DELETED_NODE)
1541 // Don't replace the results of the root node if we're doing a
1543 if (ChainNode == NodeToMatch && isMorphNodeTo)
1546 SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1);
1547 if (ChainVal.getValueType() == MVT::Glue)
1548 ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
1549 assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
1550 CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain, &ISU);
1552 // If the node became dead and we haven't already seen it, delete it.
1553 if (ChainNode->use_empty() &&
1554 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), ChainNode))
1555 NowDeadNodes.push_back(ChainNode);
1559 // If the result produces glue, update any glue results in the matched
1560 // pattern with the glue result.
1561 if (InputGlue.getNode() != 0) {
1562 // Handle any interior nodes explicitly marked.
1563 for (unsigned i = 0, e = GlueResultNodesMatched.size(); i != e; ++i) {
1564 SDNode *FRN = GlueResultNodesMatched[i];
1566 // If this node was already deleted, don't look at it.
1567 if (FRN->getOpcode() == ISD::DELETED_NODE)
1570 assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Glue &&
1571 "Doesn't have a glue result");
1572 CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1),
1575 // If the node became dead and we haven't already seen it, delete it.
1576 if (FRN->use_empty() &&
1577 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), FRN))
1578 NowDeadNodes.push_back(FRN);
1582 if (!NowDeadNodes.empty())
1583 CurDAG->RemoveDeadNodes(NowDeadNodes, &ISU);
1585 DEBUG(errs() << "ISEL: Match complete!\n");
1591 CR_LeadsToInteriorNode
1594 /// WalkChainUsers - Walk down the users of the specified chained node that is
1595 /// part of the pattern we're matching, looking at all of the users we find.
1596 /// This determines whether something is an interior node, whether we have a
1597 /// non-pattern node in between two pattern nodes (which prevent folding because
1598 /// it would induce a cycle) and whether we have a TokenFactor node sandwiched
1599 /// between pattern nodes (in which case the TF becomes part of the pattern).
1601 /// The walk we do here is guaranteed to be small because we quickly get down to
1602 /// already selected nodes "below" us.
1604 WalkChainUsers(SDNode *ChainedNode,
1605 SmallVectorImpl<SDNode*> &ChainedNodesInPattern,
1606 SmallVectorImpl<SDNode*> &InteriorChainedNodes) {
1607 ChainResult Result = CR_Simple;
1609 for (SDNode::use_iterator UI = ChainedNode->use_begin(),
1610 E = ChainedNode->use_end(); UI != E; ++UI) {
1611 // Make sure the use is of the chain, not some other value we produce.
1612 if (UI.getUse().getValueType() != MVT::Other) continue;
1616 // If we see an already-selected machine node, then we've gone beyond the
1617 // pattern that we're selecting down into the already selected chunk of the
1619 if (User->isMachineOpcode() ||
1620 User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
1623 if (User->getOpcode() == ISD::CopyToReg ||
1624 User->getOpcode() == ISD::CopyFromReg ||
1625 User->getOpcode() == ISD::INLINEASM ||
1626 User->getOpcode() == ISD::EH_LABEL) {
1627 // If their node ID got reset to -1 then they've already been selected.
1628 // Treat them like a MachineOpcode.
1629 if (User->getNodeId() == -1)
1633 // If we have a TokenFactor, we handle it specially.
1634 if (User->getOpcode() != ISD::TokenFactor) {
1635 // If the node isn't a token factor and isn't part of our pattern, then it
1636 // must be a random chained node in between two nodes we're selecting.
1637 // This happens when we have something like:
1642 // Because we structurally match the load/store as a read/modify/write,
1643 // but the call is chained between them. We cannot fold in this case
1644 // because it would induce a cycle in the graph.
1645 if (!std::count(ChainedNodesInPattern.begin(),
1646 ChainedNodesInPattern.end(), User))
1647 return CR_InducesCycle;
1649 // Otherwise we found a node that is part of our pattern. For example in:
1653 // This would happen when we're scanning down from the load and see the
1654 // store as a user. Record that there is a use of ChainedNode that is
1655 // part of the pattern and keep scanning uses.
1656 Result = CR_LeadsToInteriorNode;
1657 InteriorChainedNodes.push_back(User);
1661 // If we found a TokenFactor, there are two cases to consider: first if the
1662 // TokenFactor is just hanging "below" the pattern we're matching (i.e. no
1663 // uses of the TF are in our pattern) we just want to ignore it. Second,
1664 // the TokenFactor can be sandwiched in between two chained nodes, like so:
1670 // | \ DAG's like cheese
1673 // [TokenFactor] [Op]
1680 // In this case, the TokenFactor becomes part of our match and we rewrite it
1681 // as a new TokenFactor.
1683 // To distinguish these two cases, do a recursive walk down the uses.
1684 switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) {
1686 // If the uses of the TokenFactor are just already-selected nodes, ignore
1687 // it, it is "below" our pattern.
1689 case CR_InducesCycle:
1690 // If the uses of the TokenFactor lead to nodes that are not part of our
1691 // pattern that are not selected, folding would turn this into a cycle,
1693 return CR_InducesCycle;
1694 case CR_LeadsToInteriorNode:
1695 break; // Otherwise, keep processing.
1698 // Okay, we know we're in the interesting interior case. The TokenFactor
1699 // is now going to be considered part of the pattern so that we rewrite its
1700 // uses (it may have uses that are not part of the pattern) with the
1701 // ultimate chain result of the generated code. We will also add its chain
1702 // inputs as inputs to the ultimate TokenFactor we create.
1703 Result = CR_LeadsToInteriorNode;
1704 ChainedNodesInPattern.push_back(User);
1705 InteriorChainedNodes.push_back(User);
1712 /// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
1713 /// operation for when the pattern matched at least one node with a chains. The
1714 /// input vector contains a list of all of the chained nodes that we match. We
1715 /// must determine if this is a valid thing to cover (i.e. matching it won't
1716 /// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
1717 /// be used as the input node chain for the generated nodes.
1719 HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
1720 SelectionDAG *CurDAG) {
1721 // Walk all of the chained nodes we've matched, recursively scanning down the
1722 // users of the chain result. This adds any TokenFactor nodes that are caught
1723 // in between chained nodes to the chained and interior nodes list.
1724 SmallVector<SDNode*, 3> InteriorChainedNodes;
1725 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1726 if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched,
1727 InteriorChainedNodes) == CR_InducesCycle)
1728 return SDValue(); // Would induce a cycle.
1731 // Okay, we have walked all the matched nodes and collected TokenFactor nodes
1732 // that we are interested in. Form our input TokenFactor node.
1733 SmallVector<SDValue, 3> InputChains;
1734 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1735 // Add the input chain of this node to the InputChains list (which will be
1736 // the operands of the generated TokenFactor) if it's not an interior node.
1737 SDNode *N = ChainNodesMatched[i];
1738 if (N->getOpcode() != ISD::TokenFactor) {
1739 if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N))
1742 // Otherwise, add the input chain.
1743 SDValue InChain = ChainNodesMatched[i]->getOperand(0);
1744 assert(InChain.getValueType() == MVT::Other && "Not a chain");
1745 InputChains.push_back(InChain);
1749 // If we have a token factor, we want to add all inputs of the token factor
1750 // that are not part of the pattern we're matching.
1751 for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) {
1752 if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(),
1753 N->getOperand(op).getNode()))
1754 InputChains.push_back(N->getOperand(op));
1759 if (InputChains.size() == 1)
1760 return InputChains[0];
1761 return CurDAG->getNode(ISD::TokenFactor, ChainNodesMatched[0]->getDebugLoc(),
1762 MVT::Other, &InputChains[0], InputChains.size());
1765 /// MorphNode - Handle morphing a node in place for the selector.
1766 SDNode *SelectionDAGISel::
1767 MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList,
1768 const SDValue *Ops, unsigned NumOps, unsigned EmitNodeInfo) {
1769 // It is possible we're using MorphNodeTo to replace a node with no
1770 // normal results with one that has a normal result (or we could be
1771 // adding a chain) and the input could have glue and chains as well.
1772 // In this case we need to shift the operands down.
1773 // FIXME: This is a horrible hack and broken in obscure cases, no worse
1774 // than the old isel though.
1775 int OldGlueResultNo = -1, OldChainResultNo = -1;
1777 unsigned NTMNumResults = Node->getNumValues();
1778 if (Node->getValueType(NTMNumResults-1) == MVT::Glue) {
1779 OldGlueResultNo = NTMNumResults-1;
1780 if (NTMNumResults != 1 &&
1781 Node->getValueType(NTMNumResults-2) == MVT::Other)
1782 OldChainResultNo = NTMNumResults-2;
1783 } else if (Node->getValueType(NTMNumResults-1) == MVT::Other)
1784 OldChainResultNo = NTMNumResults-1;
1786 // Call the underlying SelectionDAG routine to do the transmogrification. Note
1787 // that this deletes operands of the old node that become dead.
1788 SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops, NumOps);
1790 // MorphNodeTo can operate in two ways: if an existing node with the
1791 // specified operands exists, it can just return it. Otherwise, it
1792 // updates the node in place to have the requested operands.
1794 // If we updated the node in place, reset the node ID. To the isel,
1795 // this should be just like a newly allocated machine node.
1799 unsigned ResNumResults = Res->getNumValues();
1800 // Move the glue if needed.
1801 if ((EmitNodeInfo & OPFL_GlueOutput) && OldGlueResultNo != -1 &&
1802 (unsigned)OldGlueResultNo != ResNumResults-1)
1803 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldGlueResultNo),
1804 SDValue(Res, ResNumResults-1));
1806 if ((EmitNodeInfo & OPFL_GlueOutput) != 0)
1809 // Move the chain reference if needed.
1810 if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 &&
1811 (unsigned)OldChainResultNo != ResNumResults-1)
1812 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo),
1813 SDValue(Res, ResNumResults-1));
1815 // Otherwise, no replacement happened because the node already exists. Replace
1816 // Uses of the old node with the new one.
1818 CurDAG->ReplaceAllUsesWith(Node, Res);
1823 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1824 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1825 CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1827 const SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes) {
1828 // Accept if it is exactly the same as a previously recorded node.
1829 unsigned RecNo = MatcherTable[MatcherIndex++];
1830 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
1831 return N == RecordedNodes[RecNo].first;
1834 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1835 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1836 CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1837 SelectionDAGISel &SDISel) {
1838 return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]);
1841 /// CheckNodePredicate - Implements OP_CheckNodePredicate.
1842 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1843 CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1844 SelectionDAGISel &SDISel, SDNode *N) {
1845 return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]);
1848 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1849 CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1851 uint16_t Opc = MatcherTable[MatcherIndex++];
1852 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
1853 return N->getOpcode() == Opc;
1856 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1857 CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1858 SDValue N, const TargetLowering &TLI) {
1859 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1860 if (N.getValueType() == VT) return true;
1862 // Handle the case when VT is iPTR.
1863 return VT == MVT::iPTR && N.getValueType() == TLI.getPointerTy();
1866 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1867 CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1868 SDValue N, const TargetLowering &TLI,
1870 if (ChildNo >= N.getNumOperands())
1871 return false; // Match fails if out of range child #.
1872 return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI);
1876 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1877 CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1879 return cast<CondCodeSDNode>(N)->get() ==
1880 (ISD::CondCode)MatcherTable[MatcherIndex++];
1883 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1884 CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1885 SDValue N, const TargetLowering &TLI) {
1886 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1887 if (cast<VTSDNode>(N)->getVT() == VT)
1890 // Handle the case when VT is iPTR.
1891 return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI.getPointerTy();
1894 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1895 CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1897 int64_t Val = MatcherTable[MatcherIndex++];
1899 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1901 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
1902 return C != 0 && C->getSExtValue() == Val;
1905 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1906 CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1907 SDValue N, SelectionDAGISel &SDISel) {
1908 int64_t Val = MatcherTable[MatcherIndex++];
1910 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1912 if (N->getOpcode() != ISD::AND) return false;
1914 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1915 return C != 0 && SDISel.CheckAndMask(N.getOperand(0), C, Val);
1918 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1919 CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1920 SDValue N, SelectionDAGISel &SDISel) {
1921 int64_t Val = MatcherTable[MatcherIndex++];
1923 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1925 if (N->getOpcode() != ISD::OR) return false;
1927 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1928 return C != 0 && SDISel.CheckOrMask(N.getOperand(0), C, Val);
1931 /// IsPredicateKnownToFail - If we know how and can do so without pushing a
1932 /// scope, evaluate the current node. If the current predicate is known to
1933 /// fail, set Result=true and return anything. If the current predicate is
1934 /// known to pass, set Result=false and return the MatcherIndex to continue
1935 /// with. If the current predicate is unknown, set Result=false and return the
1936 /// MatcherIndex to continue with.
1937 static unsigned IsPredicateKnownToFail(const unsigned char *Table,
1938 unsigned Index, SDValue N,
1939 bool &Result, SelectionDAGISel &SDISel,
1940 SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes) {
1941 switch (Table[Index++]) {
1944 return Index-1; // Could not evaluate this predicate.
1945 case SelectionDAGISel::OPC_CheckSame:
1946 Result = !::CheckSame(Table, Index, N, RecordedNodes);
1948 case SelectionDAGISel::OPC_CheckPatternPredicate:
1949 Result = !::CheckPatternPredicate(Table, Index, SDISel);
1951 case SelectionDAGISel::OPC_CheckPredicate:
1952 Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode());
1954 case SelectionDAGISel::OPC_CheckOpcode:
1955 Result = !::CheckOpcode(Table, Index, N.getNode());
1957 case SelectionDAGISel::OPC_CheckType:
1958 Result = !::CheckType(Table, Index, N, SDISel.TLI);
1960 case SelectionDAGISel::OPC_CheckChild0Type:
1961 case SelectionDAGISel::OPC_CheckChild1Type:
1962 case SelectionDAGISel::OPC_CheckChild2Type:
1963 case SelectionDAGISel::OPC_CheckChild3Type:
1964 case SelectionDAGISel::OPC_CheckChild4Type:
1965 case SelectionDAGISel::OPC_CheckChild5Type:
1966 case SelectionDAGISel::OPC_CheckChild6Type:
1967 case SelectionDAGISel::OPC_CheckChild7Type:
1968 Result = !::CheckChildType(Table, Index, N, SDISel.TLI,
1969 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Type);
1971 case SelectionDAGISel::OPC_CheckCondCode:
1972 Result = !::CheckCondCode(Table, Index, N);
1974 case SelectionDAGISel::OPC_CheckValueType:
1975 Result = !::CheckValueType(Table, Index, N, SDISel.TLI);
1977 case SelectionDAGISel::OPC_CheckInteger:
1978 Result = !::CheckInteger(Table, Index, N);
1980 case SelectionDAGISel::OPC_CheckAndImm:
1981 Result = !::CheckAndImm(Table, Index, N, SDISel);
1983 case SelectionDAGISel::OPC_CheckOrImm:
1984 Result = !::CheckOrImm(Table, Index, N, SDISel);
1992 /// FailIndex - If this match fails, this is the index to continue with.
1995 /// NodeStack - The node stack when the scope was formed.
1996 SmallVector<SDValue, 4> NodeStack;
1998 /// NumRecordedNodes - The number of recorded nodes when the scope was formed.
1999 unsigned NumRecordedNodes;
2001 /// NumMatchedMemRefs - The number of matched memref entries.
2002 unsigned NumMatchedMemRefs;
2004 /// InputChain/InputGlue - The current chain/glue
2005 SDValue InputChain, InputGlue;
2007 /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty.
2008 bool HasChainNodesMatched, HasGlueResultNodesMatched;
2013 SDNode *SelectionDAGISel::
2014 SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
2015 unsigned TableSize) {
2016 // FIXME: Should these even be selected? Handle these cases in the caller?
2017 switch (NodeToMatch->getOpcode()) {
2020 case ISD::EntryToken: // These nodes remain the same.
2021 case ISD::BasicBlock:
2023 //case ISD::VALUETYPE:
2024 //case ISD::CONDCODE:
2025 case ISD::HANDLENODE:
2026 case ISD::MDNODE_SDNODE:
2027 case ISD::TargetConstant:
2028 case ISD::TargetConstantFP:
2029 case ISD::TargetConstantPool:
2030 case ISD::TargetFrameIndex:
2031 case ISD::TargetExternalSymbol:
2032 case ISD::TargetBlockAddress:
2033 case ISD::TargetJumpTable:
2034 case ISD::TargetGlobalTLSAddress:
2035 case ISD::TargetGlobalAddress:
2036 case ISD::TokenFactor:
2037 case ISD::CopyFromReg:
2038 case ISD::CopyToReg:
2040 NodeToMatch->setNodeId(-1); // Mark selected.
2042 case ISD::AssertSext:
2043 case ISD::AssertZext:
2044 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0),
2045 NodeToMatch->getOperand(0));
2047 case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch);
2048 case ISD::UNDEF: return Select_UNDEF(NodeToMatch);
2051 assert(!NodeToMatch->isMachineOpcode() && "Node already selected!");
2053 // Set up the node stack with NodeToMatch as the only node on the stack.
2054 SmallVector<SDValue, 8> NodeStack;
2055 SDValue N = SDValue(NodeToMatch, 0);
2056 NodeStack.push_back(N);
2058 // MatchScopes - Scopes used when matching, if a match failure happens, this
2059 // indicates where to continue checking.
2060 SmallVector<MatchScope, 8> MatchScopes;
2062 // RecordedNodes - This is the set of nodes that have been recorded by the
2063 // state machine. The second value is the parent of the node, or null if the
2064 // root is recorded.
2065 SmallVector<std::pair<SDValue, SDNode*>, 8> RecordedNodes;
2067 // MatchedMemRefs - This is the set of MemRef's we've seen in the input
2069 SmallVector<MachineMemOperand*, 2> MatchedMemRefs;
2071 // These are the current input chain and glue for use when generating nodes.
2072 // Various Emit operations change these. For example, emitting a copytoreg
2073 // uses and updates these.
2074 SDValue InputChain, InputGlue;
2076 // ChainNodesMatched - If a pattern matches nodes that have input/output
2077 // chains, the OPC_EmitMergeInputChains operation is emitted which indicates
2078 // which ones they are. The result is captured into this list so that we can
2079 // update the chain results when the pattern is complete.
2080 SmallVector<SDNode*, 3> ChainNodesMatched;
2081 SmallVector<SDNode*, 3> GlueResultNodesMatched;
2083 DEBUG(errs() << "ISEL: Starting pattern match on root node: ";
2084 NodeToMatch->dump(CurDAG);
2087 // Determine where to start the interpreter. Normally we start at opcode #0,
2088 // but if the state machine starts with an OPC_SwitchOpcode, then we
2089 // accelerate the first lookup (which is guaranteed to be hot) with the
2090 // OpcodeOffset table.
2091 unsigned MatcherIndex = 0;
2093 if (!OpcodeOffset.empty()) {
2094 // Already computed the OpcodeOffset table, just index into it.
2095 if (N.getOpcode() < OpcodeOffset.size())
2096 MatcherIndex = OpcodeOffset[N.getOpcode()];
2097 DEBUG(errs() << " Initial Opcode index to " << MatcherIndex << "\n");
2099 } else if (MatcherTable[0] == OPC_SwitchOpcode) {
2100 // Otherwise, the table isn't computed, but the state machine does start
2101 // with an OPC_SwitchOpcode instruction. Populate the table now, since this
2102 // is the first time we're selecting an instruction.
2105 // Get the size of this case.
2106 unsigned CaseSize = MatcherTable[Idx++];
2108 CaseSize = GetVBR(CaseSize, MatcherTable, Idx);
2109 if (CaseSize == 0) break;
2111 // Get the opcode, add the index to the table.
2112 uint16_t Opc = MatcherTable[Idx++];
2113 Opc |= (unsigned short)MatcherTable[Idx++] << 8;
2114 if (Opc >= OpcodeOffset.size())
2115 OpcodeOffset.resize((Opc+1)*2);
2116 OpcodeOffset[Opc] = Idx;
2120 // Okay, do the lookup for the first opcode.
2121 if (N.getOpcode() < OpcodeOffset.size())
2122 MatcherIndex = OpcodeOffset[N.getOpcode()];
2126 assert(MatcherIndex < TableSize && "Invalid index");
2128 unsigned CurrentOpcodeIndex = MatcherIndex;
2130 BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++];
2133 // Okay, the semantics of this operation are that we should push a scope
2134 // then evaluate the first child. However, pushing a scope only to have
2135 // the first check fail (which then pops it) is inefficient. If we can
2136 // determine immediately that the first check (or first several) will
2137 // immediately fail, don't even bother pushing a scope for them.
2141 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2142 if (NumToSkip & 128)
2143 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2144 // Found the end of the scope with no match.
2145 if (NumToSkip == 0) {
2150 FailIndex = MatcherIndex+NumToSkip;
2152 unsigned MatcherIndexOfPredicate = MatcherIndex;
2153 (void)MatcherIndexOfPredicate; // silence warning.
2155 // If we can't evaluate this predicate without pushing a scope (e.g. if
2156 // it is a 'MoveParent') or if the predicate succeeds on this node, we
2157 // push the scope and evaluate the full predicate chain.
2159 MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
2160 Result, *this, RecordedNodes);
2164 DEBUG(errs() << " Skipped scope entry (due to false predicate) at "
2165 << "index " << MatcherIndexOfPredicate
2166 << ", continuing at " << FailIndex << "\n");
2167 ++NumDAGIselRetries;
2169 // Otherwise, we know that this case of the Scope is guaranteed to fail,
2170 // move to the next case.
2171 MatcherIndex = FailIndex;
2174 // If the whole scope failed to match, bail.
2175 if (FailIndex == 0) break;
2177 // Push a MatchScope which indicates where to go if the first child fails
2179 MatchScope NewEntry;
2180 NewEntry.FailIndex = FailIndex;
2181 NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
2182 NewEntry.NumRecordedNodes = RecordedNodes.size();
2183 NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
2184 NewEntry.InputChain = InputChain;
2185 NewEntry.InputGlue = InputGlue;
2186 NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty();
2187 NewEntry.HasGlueResultNodesMatched = !GlueResultNodesMatched.empty();
2188 MatchScopes.push_back(NewEntry);
2191 case OPC_RecordNode: {
2192 // Remember this node, it may end up being an operand in the pattern.
2194 if (NodeStack.size() > 1)
2195 Parent = NodeStack[NodeStack.size()-2].getNode();
2196 RecordedNodes.push_back(std::make_pair(N, Parent));
2200 case OPC_RecordChild0: case OPC_RecordChild1:
2201 case OPC_RecordChild2: case OPC_RecordChild3:
2202 case OPC_RecordChild4: case OPC_RecordChild5:
2203 case OPC_RecordChild6: case OPC_RecordChild7: {
2204 unsigned ChildNo = Opcode-OPC_RecordChild0;
2205 if (ChildNo >= N.getNumOperands())
2206 break; // Match fails if out of range child #.
2208 RecordedNodes.push_back(std::make_pair(N->getOperand(ChildNo),
2212 case OPC_RecordMemRef:
2213 MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand());
2216 case OPC_CaptureGlueInput:
2217 // If the current node has an input glue, capture it in InputGlue.
2218 if (N->getNumOperands() != 0 &&
2219 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue)
2220 InputGlue = N->getOperand(N->getNumOperands()-1);
2223 case OPC_MoveChild: {
2224 unsigned ChildNo = MatcherTable[MatcherIndex++];
2225 if (ChildNo >= N.getNumOperands())
2226 break; // Match fails if out of range child #.
2227 N = N.getOperand(ChildNo);
2228 NodeStack.push_back(N);
2232 case OPC_MoveParent:
2233 // Pop the current node off the NodeStack.
2234 NodeStack.pop_back();
2235 assert(!NodeStack.empty() && "Node stack imbalance!");
2236 N = NodeStack.back();
2240 if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
2242 case OPC_CheckPatternPredicate:
2243 if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break;
2245 case OPC_CheckPredicate:
2246 if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this,
2250 case OPC_CheckComplexPat: {
2251 unsigned CPNum = MatcherTable[MatcherIndex++];
2252 unsigned RecNo = MatcherTable[MatcherIndex++];
2253 assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
2254 if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo].second,
2255 RecordedNodes[RecNo].first, CPNum,
2260 case OPC_CheckOpcode:
2261 if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
2265 if (!::CheckType(MatcherTable, MatcherIndex, N, TLI)) break;
2268 case OPC_SwitchOpcode: {
2269 unsigned CurNodeOpcode = N.getOpcode();
2270 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2273 // Get the size of this case.
2274 CaseSize = MatcherTable[MatcherIndex++];
2276 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2277 if (CaseSize == 0) break;
2279 uint16_t Opc = MatcherTable[MatcherIndex++];
2280 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2282 // If the opcode matches, then we will execute this case.
2283 if (CurNodeOpcode == Opc)
2286 // Otherwise, skip over this case.
2287 MatcherIndex += CaseSize;
2290 // If no cases matched, bail out.
2291 if (CaseSize == 0) break;
2293 // Otherwise, execute the case we found.
2294 DEBUG(errs() << " OpcodeSwitch from " << SwitchStart
2295 << " to " << MatcherIndex << "\n");
2299 case OPC_SwitchType: {
2300 MVT CurNodeVT = N.getValueType().getSimpleVT();
2301 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2304 // Get the size of this case.
2305 CaseSize = MatcherTable[MatcherIndex++];
2307 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2308 if (CaseSize == 0) break;
2310 MVT CaseVT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2311 if (CaseVT == MVT::iPTR)
2312 CaseVT = TLI.getPointerTy();
2314 // If the VT matches, then we will execute this case.
2315 if (CurNodeVT == CaseVT)
2318 // Otherwise, skip over this case.
2319 MatcherIndex += CaseSize;
2322 // If no cases matched, bail out.
2323 if (CaseSize == 0) break;
2325 // Otherwise, execute the case we found.
2326 DEBUG(errs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
2327 << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
2330 case OPC_CheckChild0Type: case OPC_CheckChild1Type:
2331 case OPC_CheckChild2Type: case OPC_CheckChild3Type:
2332 case OPC_CheckChild4Type: case OPC_CheckChild5Type:
2333 case OPC_CheckChild6Type: case OPC_CheckChild7Type:
2334 if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI,
2335 Opcode-OPC_CheckChild0Type))
2338 case OPC_CheckCondCode:
2339 if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
2341 case OPC_CheckValueType:
2342 if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI)) break;
2344 case OPC_CheckInteger:
2345 if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
2347 case OPC_CheckAndImm:
2348 if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
2350 case OPC_CheckOrImm:
2351 if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
2354 case OPC_CheckFoldableChainNode: {
2355 assert(NodeStack.size() != 1 && "No parent node");
2356 // Verify that all intermediate nodes between the root and this one have
2358 bool HasMultipleUses = false;
2359 for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i)
2360 if (!NodeStack[i].hasOneUse()) {
2361 HasMultipleUses = true;
2364 if (HasMultipleUses) break;
2366 // Check to see that the target thinks this is profitable to fold and that
2367 // we can fold it without inducing cycles in the graph.
2368 if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2370 !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2371 NodeToMatch, OptLevel,
2372 true/*We validate our own chains*/))
2377 case OPC_EmitInteger: {
2378 MVT::SimpleValueType VT =
2379 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2380 int64_t Val = MatcherTable[MatcherIndex++];
2382 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2383 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2384 CurDAG->getTargetConstant(Val, VT), (SDNode*)0));
2387 case OPC_EmitRegister: {
2388 MVT::SimpleValueType VT =
2389 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2390 unsigned RegNo = MatcherTable[MatcherIndex++];
2391 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2392 CurDAG->getRegister(RegNo, VT), (SDNode*)0));
2396 case OPC_EmitConvertToTarget: {
2397 // Convert from IMM/FPIMM to target version.
2398 unsigned RecNo = MatcherTable[MatcherIndex++];
2399 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2400 SDValue Imm = RecordedNodes[RecNo].first;
2402 if (Imm->getOpcode() == ISD::Constant) {
2403 int64_t Val = cast<ConstantSDNode>(Imm)->getZExtValue();
2404 Imm = CurDAG->getTargetConstant(Val, Imm.getValueType());
2405 } else if (Imm->getOpcode() == ISD::ConstantFP) {
2406 const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue();
2407 Imm = CurDAG->getTargetConstantFP(*Val, Imm.getValueType());
2410 RecordedNodes.push_back(std::make_pair(Imm, RecordedNodes[RecNo].second));
2414 case OPC_EmitMergeInputChains1_0: // OPC_EmitMergeInputChains, 1, 0
2415 case OPC_EmitMergeInputChains1_1: { // OPC_EmitMergeInputChains, 1, 1
2416 // These are space-optimized forms of OPC_EmitMergeInputChains.
2417 assert(InputChain.getNode() == 0 &&
2418 "EmitMergeInputChains should be the first chain producing node");
2419 assert(ChainNodesMatched.empty() &&
2420 "Should only have one EmitMergeInputChains per match");
2422 // Read all of the chained nodes.
2423 unsigned RecNo = Opcode == OPC_EmitMergeInputChains1_1;
2424 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2425 ChainNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
2427 // FIXME: What if other value results of the node have uses not matched
2429 if (ChainNodesMatched.back() != NodeToMatch &&
2430 !RecordedNodes[RecNo].first.hasOneUse()) {
2431 ChainNodesMatched.clear();
2435 // Merge the input chains if they are not intra-pattern references.
2436 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2438 if (InputChain.getNode() == 0)
2439 break; // Failed to merge.
2443 case OPC_EmitMergeInputChains: {
2444 assert(InputChain.getNode() == 0 &&
2445 "EmitMergeInputChains should be the first chain producing node");
2446 // This node gets a list of nodes we matched in the input that have
2447 // chains. We want to token factor all of the input chains to these nodes
2448 // together. However, if any of the input chains is actually one of the
2449 // nodes matched in this pattern, then we have an intra-match reference.
2450 // Ignore these because the newly token factored chain should not refer to
2452 unsigned NumChains = MatcherTable[MatcherIndex++];
2453 assert(NumChains != 0 && "Can't TF zero chains");
2455 assert(ChainNodesMatched.empty() &&
2456 "Should only have one EmitMergeInputChains per match");
2458 // Read all of the chained nodes.
2459 for (unsigned i = 0; i != NumChains; ++i) {
2460 unsigned RecNo = MatcherTable[MatcherIndex++];
2461 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2462 ChainNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
2464 // FIXME: What if other value results of the node have uses not matched
2466 if (ChainNodesMatched.back() != NodeToMatch &&
2467 !RecordedNodes[RecNo].first.hasOneUse()) {
2468 ChainNodesMatched.clear();
2473 // If the inner loop broke out, the match fails.
2474 if (ChainNodesMatched.empty())
2477 // Merge the input chains if they are not intra-pattern references.
2478 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2480 if (InputChain.getNode() == 0)
2481 break; // Failed to merge.
2486 case OPC_EmitCopyToReg: {
2487 unsigned RecNo = MatcherTable[MatcherIndex++];
2488 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2489 unsigned DestPhysReg = MatcherTable[MatcherIndex++];
2491 if (InputChain.getNode() == 0)
2492 InputChain = CurDAG->getEntryNode();
2494 InputChain = CurDAG->getCopyToReg(InputChain, NodeToMatch->getDebugLoc(),
2495 DestPhysReg, RecordedNodes[RecNo].first,
2498 InputGlue = InputChain.getValue(1);
2502 case OPC_EmitNodeXForm: {
2503 unsigned XFormNo = MatcherTable[MatcherIndex++];
2504 unsigned RecNo = MatcherTable[MatcherIndex++];
2505 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2506 SDValue Res = RunSDNodeXForm(RecordedNodes[RecNo].first, XFormNo);
2507 RecordedNodes.push_back(std::pair<SDValue,SDNode*>(Res, (SDNode*) 0));
2512 case OPC_MorphNodeTo: {
2513 uint16_t TargetOpc = MatcherTable[MatcherIndex++];
2514 TargetOpc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2515 unsigned EmitNodeInfo = MatcherTable[MatcherIndex++];
2516 // Get the result VT list.
2517 unsigned NumVTs = MatcherTable[MatcherIndex++];
2518 SmallVector<EVT, 4> VTs;
2519 for (unsigned i = 0; i != NumVTs; ++i) {
2520 MVT::SimpleValueType VT =
2521 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2522 if (VT == MVT::iPTR) VT = TLI.getPointerTy().SimpleTy;
2526 if (EmitNodeInfo & OPFL_Chain)
2527 VTs.push_back(MVT::Other);
2528 if (EmitNodeInfo & OPFL_GlueOutput)
2529 VTs.push_back(MVT::Glue);
2531 // This is hot code, so optimize the two most common cases of 1 and 2
2534 if (VTs.size() == 1)
2535 VTList = CurDAG->getVTList(VTs[0]);
2536 else if (VTs.size() == 2)
2537 VTList = CurDAG->getVTList(VTs[0], VTs[1]);
2539 VTList = CurDAG->getVTList(VTs.data(), VTs.size());
2541 // Get the operand list.
2542 unsigned NumOps = MatcherTable[MatcherIndex++];
2543 SmallVector<SDValue, 8> Ops;
2544 for (unsigned i = 0; i != NumOps; ++i) {
2545 unsigned RecNo = MatcherTable[MatcherIndex++];
2547 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2549 assert(RecNo < RecordedNodes.size() && "Invalid EmitNode");
2550 Ops.push_back(RecordedNodes[RecNo].first);
2553 // If there are variadic operands to add, handle them now.
2554 if (EmitNodeInfo & OPFL_VariadicInfo) {
2555 // Determine the start index to copy from.
2556 unsigned FirstOpToCopy = getNumFixedFromVariadicInfo(EmitNodeInfo);
2557 FirstOpToCopy += (EmitNodeInfo & OPFL_Chain) ? 1 : 0;
2558 assert(NodeToMatch->getNumOperands() >= FirstOpToCopy &&
2559 "Invalid variadic node");
2560 // Copy all of the variadic operands, not including a potential glue
2562 for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands();
2564 SDValue V = NodeToMatch->getOperand(i);
2565 if (V.getValueType() == MVT::Glue) break;
2570 // If this has chain/glue inputs, add them.
2571 if (EmitNodeInfo & OPFL_Chain)
2572 Ops.push_back(InputChain);
2573 if ((EmitNodeInfo & OPFL_GlueInput) && InputGlue.getNode() != 0)
2574 Ops.push_back(InputGlue);
2578 if (Opcode != OPC_MorphNodeTo) {
2579 // If this is a normal EmitNode command, just create the new node and
2580 // add the results to the RecordedNodes list.
2581 Res = CurDAG->getMachineNode(TargetOpc, NodeToMatch->getDebugLoc(),
2582 VTList, Ops.data(), Ops.size());
2584 // Add all the non-glue/non-chain results to the RecordedNodes list.
2585 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
2586 if (VTs[i] == MVT::Other || VTs[i] == MVT::Glue) break;
2587 RecordedNodes.push_back(std::pair<SDValue,SDNode*>(SDValue(Res, i),
2592 Res = MorphNode(NodeToMatch, TargetOpc, VTList, Ops.data(), Ops.size(),
2596 // If the node had chain/glue results, update our notion of the current
2598 if (EmitNodeInfo & OPFL_GlueOutput) {
2599 InputGlue = SDValue(Res, VTs.size()-1);
2600 if (EmitNodeInfo & OPFL_Chain)
2601 InputChain = SDValue(Res, VTs.size()-2);
2602 } else if (EmitNodeInfo & OPFL_Chain)
2603 InputChain = SDValue(Res, VTs.size()-1);
2605 // If the OPFL_MemRefs glue is set on this node, slap all of the
2606 // accumulated memrefs onto it.
2608 // FIXME: This is vastly incorrect for patterns with multiple outputs
2609 // instructions that access memory and for ComplexPatterns that match
2611 if (EmitNodeInfo & OPFL_MemRefs) {
2612 MachineSDNode::mmo_iterator MemRefs =
2613 MF->allocateMemRefsArray(MatchedMemRefs.size());
2614 std::copy(MatchedMemRefs.begin(), MatchedMemRefs.end(), MemRefs);
2615 cast<MachineSDNode>(Res)
2616 ->setMemRefs(MemRefs, MemRefs + MatchedMemRefs.size());
2620 << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created")
2621 << " node: "; Res->dump(CurDAG); errs() << "\n");
2623 // If this was a MorphNodeTo then we're completely done!
2624 if (Opcode == OPC_MorphNodeTo) {
2625 // Update chain and glue uses.
2626 UpdateChainsAndGlue(NodeToMatch, InputChain, ChainNodesMatched,
2627 InputGlue, GlueResultNodesMatched, true);
2634 case OPC_MarkGlueResults: {
2635 unsigned NumNodes = MatcherTable[MatcherIndex++];
2637 // Read and remember all the glue-result nodes.
2638 for (unsigned i = 0; i != NumNodes; ++i) {
2639 unsigned RecNo = MatcherTable[MatcherIndex++];
2641 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2643 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2644 GlueResultNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
2649 case OPC_CompleteMatch: {
2650 // The match has been completed, and any new nodes (if any) have been
2651 // created. Patch up references to the matched dag to use the newly
2653 unsigned NumResults = MatcherTable[MatcherIndex++];
2655 for (unsigned i = 0; i != NumResults; ++i) {
2656 unsigned ResSlot = MatcherTable[MatcherIndex++];
2658 ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex);
2660 assert(ResSlot < RecordedNodes.size() && "Invalid CheckSame");
2661 SDValue Res = RecordedNodes[ResSlot].first;
2663 assert(i < NodeToMatch->getNumValues() &&
2664 NodeToMatch->getValueType(i) != MVT::Other &&
2665 NodeToMatch->getValueType(i) != MVT::Glue &&
2666 "Invalid number of results to complete!");
2667 assert((NodeToMatch->getValueType(i) == Res.getValueType() ||
2668 NodeToMatch->getValueType(i) == MVT::iPTR ||
2669 Res.getValueType() == MVT::iPTR ||
2670 NodeToMatch->getValueType(i).getSizeInBits() ==
2671 Res.getValueType().getSizeInBits()) &&
2672 "invalid replacement");
2673 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, i), Res);
2676 // If the root node defines glue, add it to the glue nodes to update list.
2677 if (NodeToMatch->getValueType(NodeToMatch->getNumValues()-1) == MVT::Glue)
2678 GlueResultNodesMatched.push_back(NodeToMatch);
2680 // Update chain and glue uses.
2681 UpdateChainsAndGlue(NodeToMatch, InputChain, ChainNodesMatched,
2682 InputGlue, GlueResultNodesMatched, false);
2684 assert(NodeToMatch->use_empty() &&
2685 "Didn't replace all uses of the node?");
2687 // FIXME: We just return here, which interacts correctly with SelectRoot
2688 // above. We should fix this to not return an SDNode* anymore.
2693 // If the code reached this point, then the match failed. See if there is
2694 // another child to try in the current 'Scope', otherwise pop it until we
2695 // find a case to check.
2696 DEBUG(errs() << " Match failed at index " << CurrentOpcodeIndex << "\n");
2697 ++NumDAGIselRetries;
2699 if (MatchScopes.empty()) {
2700 CannotYetSelect(NodeToMatch);
2704 // Restore the interpreter state back to the point where the scope was
2706 MatchScope &LastScope = MatchScopes.back();
2707 RecordedNodes.resize(LastScope.NumRecordedNodes);
2709 NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end());
2710 N = NodeStack.back();
2712 if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size())
2713 MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
2714 MatcherIndex = LastScope.FailIndex;
2716 DEBUG(errs() << " Continuing at " << MatcherIndex << "\n");
2718 InputChain = LastScope.InputChain;
2719 InputGlue = LastScope.InputGlue;
2720 if (!LastScope.HasChainNodesMatched)
2721 ChainNodesMatched.clear();
2722 if (!LastScope.HasGlueResultNodesMatched)
2723 GlueResultNodesMatched.clear();
2725 // Check to see what the offset is at the new MatcherIndex. If it is zero
2726 // we have reached the end of this scope, otherwise we have another child
2727 // in the current scope to try.
2728 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2729 if (NumToSkip & 128)
2730 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2732 // If we have another child in this scope to match, update FailIndex and
2734 if (NumToSkip != 0) {
2735 LastScope.FailIndex = MatcherIndex+NumToSkip;
2739 // End of this scope, pop it and try the next child in the containing
2741 MatchScopes.pop_back();
2748 void SelectionDAGISel::CannotYetSelect(SDNode *N) {
2750 raw_string_ostream Msg(msg);
2751 Msg << "Cannot select: ";
2753 if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN &&
2754 N->getOpcode() != ISD::INTRINSIC_WO_CHAIN &&
2755 N->getOpcode() != ISD::INTRINSIC_VOID) {
2756 N->printrFull(Msg, CurDAG);
2758 bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other;
2760 cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue();
2761 if (iid < Intrinsic::num_intrinsics)
2762 Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid);
2763 else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo())
2764 Msg << "target intrinsic %" << TII->getName(iid);
2766 Msg << "unknown intrinsic #" << iid;
2768 report_fatal_error(Msg.str());
2771 char SelectionDAGISel::ID = 0;