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/Statistic.h"
56 STATISTIC(NumFastIselFailures, "Number of instructions fast isel failed on");
57 STATISTIC(NumFastIselBlocks, "Number of blocks selected entirely by fast isel");
58 STATISTIC(NumDAGBlocks, "Number of blocks selected using DAG");
59 STATISTIC(NumDAGIselRetries,"Number of times dag isel has to try another path");
62 STATISTIC(NumBBWithOutOfOrderLineInfo,
63 "Number of blocks with out of order line number info");
64 STATISTIC(NumMBBWithOutOfOrderLineInfo,
65 "Number of machine blocks with out of order line number info");
69 EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
70 cl::desc("Enable verbose messages in the \"fast\" "
71 "instruction selector"));
73 EnableFastISelAbort("fast-isel-abort", cl::Hidden,
74 cl::desc("Enable abort calls when \"fast\" instruction fails"));
78 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
79 cl::desc("Pop up a window to show dags before the first "
82 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
83 cl::desc("Pop up a window to show dags before legalize types"));
85 ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
86 cl::desc("Pop up a window to show dags before legalize"));
88 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
89 cl::desc("Pop up a window to show dags before the second "
92 ViewDAGCombineLT("view-dag-combine-lt-dags", cl::Hidden,
93 cl::desc("Pop up a window to show dags before the post legalize types"
94 " dag combine pass"));
96 ViewISelDAGs("view-isel-dags", cl::Hidden,
97 cl::desc("Pop up a window to show isel dags as they are selected"));
99 ViewSchedDAGs("view-sched-dags", cl::Hidden,
100 cl::desc("Pop up a window to show sched dags as they are processed"));
102 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
103 cl::desc("Pop up a window to show SUnit dags after they are processed"));
105 static const bool ViewDAGCombine1 = false,
106 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false,
107 ViewDAGCombine2 = false,
108 ViewDAGCombineLT = false,
109 ViewISelDAGs = false, ViewSchedDAGs = false,
110 ViewSUnitDAGs = false;
113 //===---------------------------------------------------------------------===//
115 /// RegisterScheduler class - Track the registration of instruction schedulers.
117 //===---------------------------------------------------------------------===//
118 MachinePassRegistry RegisterScheduler::Registry;
120 //===---------------------------------------------------------------------===//
122 /// ISHeuristic command line option for instruction schedulers.
124 //===---------------------------------------------------------------------===//
125 static cl::opt<RegisterScheduler::FunctionPassCtor, false,
126 RegisterPassParser<RegisterScheduler> >
127 ISHeuristic("pre-RA-sched",
128 cl::init(&createDefaultScheduler),
129 cl::desc("Instruction schedulers available (before register"
132 static RegisterScheduler
133 defaultListDAGScheduler("default", "Best scheduler for the target",
134 createDefaultScheduler);
137 //===--------------------------------------------------------------------===//
138 /// createDefaultScheduler - This creates an instruction scheduler appropriate
140 ScheduleDAGSDNodes* createDefaultScheduler(SelectionDAGISel *IS,
141 CodeGenOpt::Level OptLevel) {
142 const TargetLowering &TLI = IS->getTargetLowering();
144 if (OptLevel == CodeGenOpt::None)
145 return createSourceListDAGScheduler(IS, OptLevel);
146 if (TLI.getSchedulingPreference() == Sched::Latency)
147 return createTDListDAGScheduler(IS, OptLevel);
148 if (TLI.getSchedulingPreference() == Sched::RegPressure)
149 return createBURRListDAGScheduler(IS, OptLevel);
150 if (TLI.getSchedulingPreference() == Sched::Hybrid)
151 return createHybridListDAGScheduler(IS, OptLevel);
152 assert(TLI.getSchedulingPreference() == Sched::ILP &&
153 "Unknown sched type!");
154 return createILPListDAGScheduler(IS, OptLevel);
158 // EmitInstrWithCustomInserter - This method should be implemented by targets
159 // that mark instructions with the 'usesCustomInserter' flag. These
160 // instructions are special in various ways, which require special support to
161 // insert. The specified MachineInstr is created but not inserted into any
162 // basic blocks, and this method is called to expand it into a sequence of
163 // instructions, potentially also creating new basic blocks and control flow.
164 // When new basic blocks are inserted and the edges from MBB to its successors
165 // are modified, the method should insert pairs of <OldSucc, NewSucc> into the
168 TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
169 MachineBasicBlock *MBB) const {
171 dbgs() << "If a target marks an instruction with "
172 "'usesCustomInserter', it must implement "
173 "TargetLowering::EmitInstrWithCustomInserter!";
179 //===----------------------------------------------------------------------===//
180 // SelectionDAGISel code
181 //===----------------------------------------------------------------------===//
183 SelectionDAGISel::SelectionDAGISel(const TargetMachine &tm,
184 CodeGenOpt::Level OL) :
185 MachineFunctionPass(ID), TM(tm), TLI(*tm.getTargetLowering()),
186 FuncInfo(new FunctionLoweringInfo(TLI)),
187 CurDAG(new SelectionDAG(tm)),
188 SDB(new SelectionDAGBuilder(*CurDAG, *FuncInfo, OL)),
192 initializeGCModuleInfoPass(*PassRegistry::getPassRegistry());
193 initializeAliasAnalysisAnalysisGroup(*PassRegistry::getPassRegistry());
196 SelectionDAGISel::~SelectionDAGISel() {
202 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
203 AU.addRequired<AliasAnalysis>();
204 AU.addPreserved<AliasAnalysis>();
205 AU.addRequired<GCModuleInfo>();
206 AU.addPreserved<GCModuleInfo>();
207 MachineFunctionPass::getAnalysisUsage(AU);
210 /// FunctionCallsSetJmp - Return true if the function has a call to setjmp or
211 /// other function that gcc recognizes as "returning twice". This is used to
212 /// limit code-gen optimizations on the machine function.
214 /// FIXME: Remove after <rdar://problem/8031714> is fixed.
215 static bool FunctionCallsSetJmp(const Function *F) {
216 const Module *M = F->getParent();
217 static const char *ReturnsTwiceFns[] = {
227 #define NUM_RETURNS_TWICE_FNS sizeof(ReturnsTwiceFns) / sizeof(const char *)
229 for (unsigned I = 0; I < NUM_RETURNS_TWICE_FNS; ++I)
230 if (const Function *Callee = M->getFunction(ReturnsTwiceFns[I])) {
231 if (!Callee->use_empty())
232 for (Value::const_use_iterator
233 I = Callee->use_begin(), E = Callee->use_end();
235 if (const CallInst *CI = dyn_cast<CallInst>(*I))
236 if (CI->getParent()->getParent() == F)
241 #undef NUM_RETURNS_TWICE_FNS
244 /// SplitCriticalSideEffectEdges - Look for critical edges with a PHI value that
245 /// may trap on it. In this case we have to split the edge so that the path
246 /// through the predecessor block that doesn't go to the phi block doesn't
247 /// execute the possibly trapping instruction.
249 /// This is required for correctness, so it must be done at -O0.
251 static void SplitCriticalSideEffectEdges(Function &Fn, Pass *SDISel) {
252 // Loop for blocks with phi nodes.
253 for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
254 PHINode *PN = dyn_cast<PHINode>(BB->begin());
255 if (PN == 0) continue;
258 // For each block with a PHI node, check to see if any of the input values
259 // are potentially trapping constant expressions. Constant expressions are
260 // the only potentially trapping value that can occur as the argument to a
262 for (BasicBlock::iterator I = BB->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
263 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
264 ConstantExpr *CE = dyn_cast<ConstantExpr>(PN->getIncomingValue(i));
265 if (CE == 0 || !CE->canTrap()) continue;
267 // The only case we have to worry about is when the edge is critical.
268 // Since this block has a PHI Node, we assume it has multiple input
269 // edges: check to see if the pred has multiple successors.
270 BasicBlock *Pred = PN->getIncomingBlock(i);
271 if (Pred->getTerminator()->getNumSuccessors() == 1)
274 // Okay, we have to split this edge.
275 SplitCriticalEdge(Pred->getTerminator(),
276 GetSuccessorNumber(Pred, BB), SDISel, true);
282 bool SelectionDAGISel::runOnMachineFunction(MachineFunction &mf) {
283 // Do some sanity-checking on the command-line options.
284 assert((!EnableFastISelVerbose || EnableFastISel) &&
285 "-fast-isel-verbose requires -fast-isel");
286 assert((!EnableFastISelAbort || EnableFastISel) &&
287 "-fast-isel-abort requires -fast-isel");
289 const Function &Fn = *mf.getFunction();
290 const TargetInstrInfo &TII = *TM.getInstrInfo();
291 const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
294 RegInfo = &MF->getRegInfo();
295 AA = &getAnalysis<AliasAnalysis>();
296 GFI = Fn.hasGC() ? &getAnalysis<GCModuleInfo>().getFunctionInfo(Fn) : 0;
298 DEBUG(dbgs() << "\n\n\n=== " << Fn.getName() << "\n");
300 SplitCriticalSideEffectEdges(const_cast<Function&>(Fn), this);
303 FuncInfo->set(Fn, *MF);
306 SelectAllBasicBlocks(Fn);
308 // If the first basic block in the function has live ins that need to be
309 // copied into vregs, emit the copies into the top of the block before
310 // emitting the code for the block.
311 MachineBasicBlock *EntryMBB = MF->begin();
312 RegInfo->EmitLiveInCopies(EntryMBB, TRI, TII);
314 DenseMap<unsigned, unsigned> LiveInMap;
315 if (!FuncInfo->ArgDbgValues.empty())
316 for (MachineRegisterInfo::livein_iterator LI = RegInfo->livein_begin(),
317 E = RegInfo->livein_end(); LI != E; ++LI)
319 LiveInMap.insert(std::make_pair(LI->first, LI->second));
321 // Insert DBG_VALUE instructions for function arguments to the entry block.
322 for (unsigned i = 0, e = FuncInfo->ArgDbgValues.size(); i != e; ++i) {
323 MachineInstr *MI = FuncInfo->ArgDbgValues[e-i-1];
324 unsigned Reg = MI->getOperand(0).getReg();
325 if (TargetRegisterInfo::isPhysicalRegister(Reg))
326 EntryMBB->insert(EntryMBB->begin(), MI);
328 MachineInstr *Def = RegInfo->getVRegDef(Reg);
329 MachineBasicBlock::iterator InsertPos = Def;
330 // FIXME: VR def may not be in entry block.
331 Def->getParent()->insert(llvm::next(InsertPos), MI);
334 // If Reg is live-in then update debug info to track its copy in a vreg.
335 DenseMap<unsigned, unsigned>::iterator LDI = LiveInMap.find(Reg);
336 if (LDI != LiveInMap.end()) {
337 MachineInstr *Def = RegInfo->getVRegDef(LDI->second);
338 MachineBasicBlock::iterator InsertPos = Def;
339 const MDNode *Variable =
340 MI->getOperand(MI->getNumOperands()-1).getMetadata();
341 unsigned Offset = MI->getOperand(1).getImm();
342 // Def is never a terminator here, so it is ok to increment InsertPos.
343 BuildMI(*EntryMBB, ++InsertPos, MI->getDebugLoc(),
344 TII.get(TargetOpcode::DBG_VALUE))
345 .addReg(LDI->second, RegState::Debug)
346 .addImm(Offset).addMetadata(Variable);
348 // If this vreg is directly copied into an exported register then
349 // that COPY instructions also need DBG_VALUE, if it is the only
350 // user of LDI->second.
351 MachineInstr *CopyUseMI = NULL;
352 for (MachineRegisterInfo::use_iterator
353 UI = RegInfo->use_begin(LDI->second);
354 MachineInstr *UseMI = UI.skipInstruction();) {
355 if (UseMI->isDebugValue()) continue;
356 if (UseMI->isCopy() && !CopyUseMI && UseMI->getParent() == EntryMBB) {
357 CopyUseMI = UseMI; continue;
359 // Otherwise this is another use or second copy use.
360 CopyUseMI = NULL; break;
363 MachineInstr *NewMI =
364 BuildMI(*MF, CopyUseMI->getDebugLoc(),
365 TII.get(TargetOpcode::DBG_VALUE))
366 .addReg(CopyUseMI->getOperand(0).getReg(), RegState::Debug)
367 .addImm(Offset).addMetadata(Variable);
368 EntryMBB->insertAfter(CopyUseMI, NewMI);
373 // Determine if there are any calls in this machine function.
374 MachineFrameInfo *MFI = MF->getFrameInfo();
375 if (!MFI->hasCalls()) {
376 for (MachineFunction::const_iterator
377 I = MF->begin(), E = MF->end(); I != E; ++I) {
378 const MachineBasicBlock *MBB = I;
379 for (MachineBasicBlock::const_iterator
380 II = MBB->begin(), IE = MBB->end(); II != IE; ++II) {
381 const TargetInstrDesc &TID = TM.getInstrInfo()->get(II->getOpcode());
383 if ((TID.isCall() && !TID.isReturn()) ||
384 II->isStackAligningInlineAsm()) {
385 MFI->setHasCalls(true);
393 // Determine if there is a call to setjmp in the machine function.
394 MF->setCallsSetJmp(FunctionCallsSetJmp(&Fn));
396 // Replace forward-declared registers with the registers containing
397 // the desired value.
398 MachineRegisterInfo &MRI = MF->getRegInfo();
399 for (DenseMap<unsigned, unsigned>::iterator
400 I = FuncInfo->RegFixups.begin(), E = FuncInfo->RegFixups.end();
402 unsigned From = I->first;
403 unsigned To = I->second;
404 // If To is also scheduled to be replaced, find what its ultimate
407 DenseMap<unsigned, unsigned>::iterator J =
408 FuncInfo->RegFixups.find(To);
413 MRI.replaceRegWith(From, To);
416 // Release function-specific state. SDB and CurDAG are already cleared
424 SelectionDAGISel::SelectBasicBlock(BasicBlock::const_iterator Begin,
425 BasicBlock::const_iterator End,
427 // Lower all of the non-terminator instructions. If a call is emitted
428 // as a tail call, cease emitting nodes for this block. Terminators
429 // are handled below.
430 for (BasicBlock::const_iterator I = Begin; I != End && !SDB->HasTailCall; ++I)
433 // Make sure the root of the DAG is up-to-date.
434 CurDAG->setRoot(SDB->getControlRoot());
435 HadTailCall = SDB->HasTailCall;
438 // Final step, emit the lowered DAG as machine code.
443 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
444 SmallPtrSet<SDNode*, 128> VisitedNodes;
445 SmallVector<SDNode*, 128> Worklist;
447 Worklist.push_back(CurDAG->getRoot().getNode());
454 SDNode *N = Worklist.pop_back_val();
456 // If we've already seen this node, ignore it.
457 if (!VisitedNodes.insert(N))
460 // Otherwise, add all chain operands to the worklist.
461 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
462 if (N->getOperand(i).getValueType() == MVT::Other)
463 Worklist.push_back(N->getOperand(i).getNode());
465 // If this is a CopyToReg with a vreg dest, process it.
466 if (N->getOpcode() != ISD::CopyToReg)
469 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
470 if (!TargetRegisterInfo::isVirtualRegister(DestReg))
473 // Ignore non-scalar or non-integer values.
474 SDValue Src = N->getOperand(2);
475 EVT SrcVT = Src.getValueType();
476 if (!SrcVT.isInteger() || SrcVT.isVector())
479 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
480 Mask = APInt::getAllOnesValue(SrcVT.getSizeInBits());
481 CurDAG->ComputeMaskedBits(Src, Mask, KnownZero, KnownOne);
483 // Only install this information if it tells us something.
484 if (NumSignBits != 1 || KnownZero != 0 || KnownOne != 0) {
485 FuncInfo->LiveOutRegInfo.grow(DestReg);
486 FunctionLoweringInfo::LiveOutInfo &LOI =
487 FuncInfo->LiveOutRegInfo[DestReg];
488 LOI.NumSignBits = NumSignBits;
489 LOI.KnownOne = KnownOne;
490 LOI.KnownZero = KnownZero;
492 } while (!Worklist.empty());
495 void SelectionDAGISel::CodeGenAndEmitDAG() {
496 std::string GroupName;
497 if (TimePassesIsEnabled)
498 GroupName = "Instruction Selection and Scheduling";
499 std::string BlockName;
500 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs ||
501 ViewDAGCombine2 || ViewDAGCombineLT || ViewISelDAGs || ViewSchedDAGs ||
503 BlockName = MF->getFunction()->getNameStr() + ":" +
504 FuncInfo->MBB->getBasicBlock()->getNameStr();
506 DEBUG(dbgs() << "Initial selection DAG:\n"; CurDAG->dump());
508 if (ViewDAGCombine1) CurDAG->viewGraph("dag-combine1 input for " + BlockName);
510 // Run the DAG combiner in pre-legalize mode.
512 NamedRegionTimer T("DAG Combining 1", GroupName, TimePassesIsEnabled);
513 CurDAG->Combine(Unrestricted, *AA, OptLevel);
516 DEBUG(dbgs() << "Optimized lowered selection DAG:\n"; CurDAG->dump());
518 // Second step, hack on the DAG until it only uses operations and types that
519 // the target supports.
520 if (ViewLegalizeTypesDAGs) CurDAG->viewGraph("legalize-types input for " +
525 NamedRegionTimer T("Type Legalization", GroupName, TimePassesIsEnabled);
526 Changed = CurDAG->LegalizeTypes();
529 DEBUG(dbgs() << "Type-legalized selection DAG:\n"; CurDAG->dump());
532 if (ViewDAGCombineLT)
533 CurDAG->viewGraph("dag-combine-lt input for " + BlockName);
535 // Run the DAG combiner in post-type-legalize mode.
537 NamedRegionTimer T("DAG Combining after legalize types", GroupName,
538 TimePassesIsEnabled);
539 CurDAG->Combine(NoIllegalTypes, *AA, OptLevel);
542 DEBUG(dbgs() << "Optimized type-legalized selection DAG:\n";
547 NamedRegionTimer T("Vector Legalization", GroupName, TimePassesIsEnabled);
548 Changed = CurDAG->LegalizeVectors();
553 NamedRegionTimer T("Type Legalization 2", GroupName, TimePassesIsEnabled);
554 CurDAG->LegalizeTypes();
557 if (ViewDAGCombineLT)
558 CurDAG->viewGraph("dag-combine-lv input for " + BlockName);
560 // Run the DAG combiner in post-type-legalize mode.
562 NamedRegionTimer T("DAG Combining after legalize vectors", GroupName,
563 TimePassesIsEnabled);
564 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
567 DEBUG(dbgs() << "Optimized vector-legalized selection DAG:\n";
571 if (ViewLegalizeDAGs) CurDAG->viewGraph("legalize input for " + BlockName);
574 NamedRegionTimer T("DAG Legalization", GroupName, TimePassesIsEnabled);
575 CurDAG->Legalize(OptLevel);
578 DEBUG(dbgs() << "Legalized selection DAG:\n"; CurDAG->dump());
580 if (ViewDAGCombine2) CurDAG->viewGraph("dag-combine2 input for " + BlockName);
582 // Run the DAG combiner in post-legalize mode.
584 NamedRegionTimer T("DAG Combining 2", GroupName, TimePassesIsEnabled);
585 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
588 DEBUG(dbgs() << "Optimized legalized selection DAG:\n"; CurDAG->dump());
590 if (OptLevel != CodeGenOpt::None)
591 ComputeLiveOutVRegInfo();
593 if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName);
595 // Third, instruction select all of the operations to machine code, adding the
596 // code to the MachineBasicBlock.
598 NamedRegionTimer T("Instruction Selection", GroupName, TimePassesIsEnabled);
599 DoInstructionSelection();
602 DEBUG(dbgs() << "Selected selection DAG:\n"; CurDAG->dump());
604 if (ViewSchedDAGs) CurDAG->viewGraph("scheduler input for " + BlockName);
606 // Schedule machine code.
607 ScheduleDAGSDNodes *Scheduler = CreateScheduler();
609 NamedRegionTimer T("Instruction Scheduling", GroupName,
610 TimePassesIsEnabled);
611 Scheduler->Run(CurDAG, FuncInfo->MBB, FuncInfo->InsertPt);
614 if (ViewSUnitDAGs) Scheduler->viewGraph();
616 // Emit machine code to BB. This can change 'BB' to the last block being
618 MachineBasicBlock *FirstMBB = FuncInfo->MBB, *LastMBB;
620 NamedRegionTimer T("Instruction Creation", GroupName, TimePassesIsEnabled);
622 LastMBB = FuncInfo->MBB = Scheduler->EmitSchedule();
623 FuncInfo->InsertPt = Scheduler->InsertPos;
626 // If the block was split, make sure we update any references that are used to
627 // update PHI nodes later on.
628 if (FirstMBB != LastMBB)
629 SDB->UpdateSplitBlock(FirstMBB, LastMBB);
631 // Free the scheduler state.
633 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName,
634 TimePassesIsEnabled);
638 // Free the SelectionDAG state, now that we're finished with it.
642 void SelectionDAGISel::DoInstructionSelection() {
643 DEBUG(errs() << "===== Instruction selection begins:\n");
647 // Select target instructions for the DAG.
649 // Number all nodes with a topological order and set DAGSize.
650 DAGSize = CurDAG->AssignTopologicalOrder();
652 // Create a dummy node (which is not added to allnodes), that adds
653 // a reference to the root node, preventing it from being deleted,
654 // and tracking any changes of the root.
655 HandleSDNode Dummy(CurDAG->getRoot());
656 ISelPosition = SelectionDAG::allnodes_iterator(CurDAG->getRoot().getNode());
659 // The AllNodes list is now topological-sorted. Visit the
660 // nodes by starting at the end of the list (the root of the
661 // graph) and preceding back toward the beginning (the entry
663 while (ISelPosition != CurDAG->allnodes_begin()) {
664 SDNode *Node = --ISelPosition;
665 // Skip dead nodes. DAGCombiner is expected to eliminate all dead nodes,
666 // but there are currently some corner cases that it misses. Also, this
667 // makes it theoretically possible to disable the DAGCombiner.
668 if (Node->use_empty())
671 SDNode *ResNode = Select(Node);
673 // FIXME: This is pretty gross. 'Select' should be changed to not return
674 // anything at all and this code should be nuked with a tactical strike.
676 // If node should not be replaced, continue with the next one.
677 if (ResNode == Node || Node->getOpcode() == ISD::DELETED_NODE)
681 ReplaceUses(Node, ResNode);
683 // If after the replacement this node is not used any more,
684 // remove this dead node.
685 if (Node->use_empty()) { // Don't delete EntryToken, etc.
686 ISelUpdater ISU(ISelPosition);
687 CurDAG->RemoveDeadNode(Node, &ISU);
691 CurDAG->setRoot(Dummy.getValue());
694 DEBUG(errs() << "===== Instruction selection ends:\n");
696 PostprocessISelDAG();
699 /// PrepareEHLandingPad - Emit an EH_LABEL, set up live-in registers, and
700 /// do other setup for EH landing-pad blocks.
701 void SelectionDAGISel::PrepareEHLandingPad() {
702 // Add a label to mark the beginning of the landing pad. Deletion of the
703 // landing pad can thus be detected via the MachineModuleInfo.
704 MCSymbol *Label = MF->getMMI().addLandingPad(FuncInfo->MBB);
706 const TargetInstrDesc &II = TM.getInstrInfo()->get(TargetOpcode::EH_LABEL);
707 BuildMI(*FuncInfo->MBB, FuncInfo->InsertPt, SDB->getCurDebugLoc(), II)
710 // Mark exception register as live in.
711 unsigned Reg = TLI.getExceptionAddressRegister();
712 if (Reg) FuncInfo->MBB->addLiveIn(Reg);
714 // Mark exception selector register as live in.
715 Reg = TLI.getExceptionSelectorRegister();
716 if (Reg) FuncInfo->MBB->addLiveIn(Reg);
718 // FIXME: Hack around an exception handling flaw (PR1508): the personality
719 // function and list of typeids logically belong to the invoke (or, if you
720 // like, the basic block containing the invoke), and need to be associated
721 // with it in the dwarf exception handling tables. Currently however the
722 // information is provided by an intrinsic (eh.selector) that can be moved
723 // to unexpected places by the optimizers: if the unwind edge is critical,
724 // then breaking it can result in the intrinsics being in the successor of
725 // the landing pad, not the landing pad itself. This results
726 // in exceptions not being caught because no typeids are associated with
727 // the invoke. This may not be the only way things can go wrong, but it
728 // is the only way we try to work around for the moment.
729 const BasicBlock *LLVMBB = FuncInfo->MBB->getBasicBlock();
730 const BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
732 if (Br && Br->isUnconditional()) { // Critical edge?
733 BasicBlock::const_iterator I, E;
734 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
735 if (isa<EHSelectorInst>(I))
739 // No catch info found - try to extract some from the successor.
740 CopyCatchInfo(Br->getSuccessor(0), LLVMBB, &MF->getMMI(), *FuncInfo);
747 bool SelectionDAGISel::TryToFoldFastISelLoad(const LoadInst *LI,
749 // Don't try to fold volatile loads. Target has to deal with alignment
751 if (LI->isVolatile()) return false;
753 // Figure out which vreg this is going into.
754 unsigned LoadReg = FastIS->getRegForValue(LI);
755 assert(LoadReg && "Load isn't already assigned a vreg? ");
757 // Check to see what the uses of this vreg are. If it has no uses, or more
758 // than one use (at the machine instr level) then we can't fold it.
759 MachineRegisterInfo::reg_iterator RI = RegInfo->reg_begin(LoadReg);
760 if (RI == RegInfo->reg_end())
763 // See if there is exactly one use of the vreg. If there are multiple uses,
764 // then the instruction got lowered to multiple machine instructions or the
765 // use of the loaded value ended up being multiple operands of the result, in
766 // either case, we can't fold this.
767 MachineRegisterInfo::reg_iterator PostRI = RI; ++PostRI;
768 if (PostRI != RegInfo->reg_end())
771 assert(RI.getOperand().isUse() &&
772 "The only use of the vreg must be a use, we haven't emitted the def!");
774 MachineInstr *User = &*RI;
776 // Set the insertion point properly. Folding the load can cause generation of
777 // other random instructions (like sign extends) for addressing modes, make
778 // sure they get inserted in a logical place before the new instruction.
779 FuncInfo->InsertPt = User;
780 FuncInfo->MBB = User->getParent();
782 // Ask the target to try folding the load.
783 return FastIS->TryToFoldLoad(User, RI.getOperandNo(), LI);
787 /// CheckLineNumbers - Check if basic block instructions follow source order
789 static void CheckLineNumbers(const BasicBlock *BB) {
792 for (BasicBlock::const_iterator BI = BB->begin(),
793 BE = BB->end(); BI != BE; ++BI) {
794 const DebugLoc DL = BI->getDebugLoc();
795 if (DL.isUnknown()) continue;
796 unsigned L = DL.getLine();
797 unsigned C = DL.getCol();
798 if (L < Line || (L == Line && C < Col)) {
799 ++NumBBWithOutOfOrderLineInfo;
807 /// CheckLineNumbers - Check if machine basic block instructions follow source
809 static void CheckLineNumbers(const MachineBasicBlock *MBB) {
812 for (MachineBasicBlock::const_iterator MBI = MBB->begin(),
813 MBE = MBB->end(); MBI != MBE; ++MBI) {
814 const DebugLoc DL = MBI->getDebugLoc();
815 if (DL.isUnknown()) continue;
816 unsigned L = DL.getLine();
817 unsigned C = DL.getCol();
818 if (L < Line || (L == Line && C < Col)) {
819 ++NumMBBWithOutOfOrderLineInfo;
828 void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
829 // Initialize the Fast-ISel state, if needed.
830 FastISel *FastIS = 0;
832 FastIS = TLI.createFastISel(*FuncInfo);
834 // Iterate over all basic blocks in the function.
835 for (Function::const_iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
836 const BasicBlock *LLVMBB = &*I;
838 CheckLineNumbers(LLVMBB);
840 FuncInfo->MBB = FuncInfo->MBBMap[LLVMBB];
841 FuncInfo->InsertPt = FuncInfo->MBB->getFirstNonPHI();
843 BasicBlock::const_iterator const Begin = LLVMBB->getFirstNonPHI();
844 BasicBlock::const_iterator const End = LLVMBB->end();
845 BasicBlock::const_iterator BI = End;
847 FuncInfo->InsertPt = FuncInfo->MBB->getFirstNonPHI();
849 // Setup an EH landing-pad block.
850 if (FuncInfo->MBB->isLandingPad())
851 PrepareEHLandingPad();
853 // Lower any arguments needed in this block if this is the entry block.
854 if (LLVMBB == &Fn.getEntryBlock()) {
855 for (BasicBlock::const_iterator DBI = LLVMBB->begin(), DBE = LLVMBB->end();
857 if (const DbgInfoIntrinsic *DI = dyn_cast<DbgInfoIntrinsic>(DBI)) {
858 const DebugLoc DL = DI->getDebugLoc();
859 SDB->setCurDebugLoc(DL);
863 LowerArguments(LLVMBB);
866 // Before doing SelectionDAG ISel, see if FastISel has been requested.
868 FastIS->startNewBlock();
870 // Emit code for any incoming arguments. This must happen before
871 // beginning FastISel on the entry block.
872 if (LLVMBB == &Fn.getEntryBlock()) {
873 CurDAG->setRoot(SDB->getControlRoot());
877 // If we inserted any instructions at the beginning, make a note of
878 // where they are, so we can be sure to emit subsequent instructions
880 if (FuncInfo->InsertPt != FuncInfo->MBB->begin())
881 FastIS->setLastLocalValue(llvm::prior(FuncInfo->InsertPt));
883 FastIS->setLastLocalValue(0);
886 // Do FastISel on as many instructions as possible.
887 for (; BI != Begin; --BI) {
888 const Instruction *Inst = llvm::prior(BI);
890 // If we no longer require this instruction, skip it.
891 if (!Inst->mayWriteToMemory() &&
892 !isa<TerminatorInst>(Inst) &&
893 !isa<DbgInfoIntrinsic>(Inst) &&
894 !FuncInfo->isExportedInst(Inst))
897 // Bottom-up: reset the insert pos at the top, after any local-value
899 FastIS->recomputeInsertPt();
901 // Try to select the instruction with FastISel.
902 if (FastIS->SelectInstruction(Inst)) {
903 // If fast isel succeeded, check to see if there is a single-use
904 // non-volatile load right before the selected instruction, and see if
905 // the load is used by the instruction. If so, try to fold it.
906 const Instruction *BeforeInst = 0;
908 BeforeInst = llvm::prior(llvm::prior(BI));
909 if (BeforeInst && isa<LoadInst>(BeforeInst) &&
910 BeforeInst->hasOneUse() && *BeforeInst->use_begin() == Inst &&
911 TryToFoldFastISelLoad(cast<LoadInst>(BeforeInst), FastIS))
912 --BI; // If we succeeded, don't re-select the load.
916 // Then handle certain instructions as single-LLVM-Instruction blocks.
917 if (isa<CallInst>(Inst)) {
918 ++NumFastIselFailures;
919 if (EnableFastISelVerbose || EnableFastISelAbort) {
920 dbgs() << "FastISel missed call: ";
924 if (!Inst->getType()->isVoidTy() && !Inst->use_empty()) {
925 unsigned &R = FuncInfo->ValueMap[Inst];
927 R = FuncInfo->CreateRegs(Inst->getType());
930 bool HadTailCall = false;
931 SelectBasicBlock(Inst, BI, HadTailCall);
933 // If the call was emitted as a tail call, we're done with the block.
942 // Otherwise, give up on FastISel for the rest of the block.
943 // For now, be a little lenient about non-branch terminators.
944 if (!isa<TerminatorInst>(Inst) || isa<BranchInst>(Inst)) {
945 ++NumFastIselFailures;
946 if (EnableFastISelVerbose || EnableFastISelAbort) {
947 dbgs() << "FastISel miss: ";
950 if (EnableFastISelAbort)
951 // The "fast" selector couldn't handle something and bailed.
952 // For the purpose of debugging, just abort.
953 llvm_unreachable("FastISel didn't select the entire block");
958 FastIS->recomputeInsertPt();
966 // Run SelectionDAG instruction selection on the remainder of the block
967 // not handled by FastISel. If FastISel is not run, this is the entire
970 SelectBasicBlock(Begin, BI, HadTailCall);
973 FuncInfo->PHINodesToUpdate.clear();
978 for (MachineFunction::const_iterator MBI = MF->begin(), MBE = MF->end();
980 CheckLineNumbers(MBI);
985 SelectionDAGISel::FinishBasicBlock() {
987 DEBUG(dbgs() << "Total amount of phi nodes to update: "
988 << FuncInfo->PHINodesToUpdate.size() << "\n";
989 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i)
990 dbgs() << "Node " << i << " : ("
991 << FuncInfo->PHINodesToUpdate[i].first
992 << ", " << FuncInfo->PHINodesToUpdate[i].second << ")\n");
994 // Next, now that we know what the last MBB the LLVM BB expanded is, update
995 // PHI nodes in successors.
996 if (SDB->SwitchCases.empty() &&
997 SDB->JTCases.empty() &&
998 SDB->BitTestCases.empty()) {
999 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i) {
1000 MachineInstr *PHI = FuncInfo->PHINodesToUpdate[i].first;
1001 assert(PHI->isPHI() &&
1002 "This is not a machine PHI node that we are updating!");
1003 if (!FuncInfo->MBB->isSuccessor(PHI->getParent()))
1006 MachineOperand::CreateReg(FuncInfo->PHINodesToUpdate[i].second, false));
1007 PHI->addOperand(MachineOperand::CreateMBB(FuncInfo->MBB));
1012 for (unsigned i = 0, e = SDB->BitTestCases.size(); i != e; ++i) {
1013 // Lower header first, if it wasn't already lowered
1014 if (!SDB->BitTestCases[i].Emitted) {
1015 // Set the current basic block to the mbb we wish to insert the code into
1016 FuncInfo->MBB = SDB->BitTestCases[i].Parent;
1017 FuncInfo->InsertPt = FuncInfo->MBB->end();
1019 SDB->visitBitTestHeader(SDB->BitTestCases[i], FuncInfo->MBB);
1020 CurDAG->setRoot(SDB->getRoot());
1022 CodeGenAndEmitDAG();
1025 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j) {
1026 // Set the current basic block to the mbb we wish to insert the code into
1027 FuncInfo->MBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1028 FuncInfo->InsertPt = FuncInfo->MBB->end();
1031 SDB->visitBitTestCase(SDB->BitTestCases[i],
1032 SDB->BitTestCases[i].Cases[j+1].ThisBB,
1033 SDB->BitTestCases[i].Reg,
1034 SDB->BitTestCases[i].Cases[j],
1037 SDB->visitBitTestCase(SDB->BitTestCases[i],
1038 SDB->BitTestCases[i].Default,
1039 SDB->BitTestCases[i].Reg,
1040 SDB->BitTestCases[i].Cases[j],
1044 CurDAG->setRoot(SDB->getRoot());
1046 CodeGenAndEmitDAG();
1050 for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1052 MachineInstr *PHI = FuncInfo->PHINodesToUpdate[pi].first;
1053 MachineBasicBlock *PHIBB = PHI->getParent();
1054 assert(PHI->isPHI() &&
1055 "This is not a machine PHI node that we are updating!");
1056 // This is "default" BB. We have two jumps to it. From "header" BB and
1057 // from last "case" BB.
1058 if (PHIBB == SDB->BitTestCases[i].Default) {
1059 PHI->addOperand(MachineOperand::
1060 CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1062 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Parent));
1063 PHI->addOperand(MachineOperand::
1064 CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1066 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Cases.
1069 // One of "cases" BB.
1070 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size();
1072 MachineBasicBlock* cBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1073 if (cBB->isSuccessor(PHIBB)) {
1074 PHI->addOperand(MachineOperand::
1075 CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1077 PHI->addOperand(MachineOperand::CreateMBB(cBB));
1082 SDB->BitTestCases.clear();
1084 // If the JumpTable record is filled in, then we need to emit a jump table.
1085 // Updating the PHI nodes is tricky in this case, since we need to determine
1086 // whether the PHI is a successor of the range check MBB or the jump table MBB
1087 for (unsigned i = 0, e = SDB->JTCases.size(); i != e; ++i) {
1088 // Lower header first, if it wasn't already lowered
1089 if (!SDB->JTCases[i].first.Emitted) {
1090 // Set the current basic block to the mbb we wish to insert the code into
1091 FuncInfo->MBB = SDB->JTCases[i].first.HeaderBB;
1092 FuncInfo->InsertPt = FuncInfo->MBB->end();
1094 SDB->visitJumpTableHeader(SDB->JTCases[i].second, SDB->JTCases[i].first,
1096 CurDAG->setRoot(SDB->getRoot());
1098 CodeGenAndEmitDAG();
1101 // Set the current basic block to the mbb we wish to insert the code into
1102 FuncInfo->MBB = SDB->JTCases[i].second.MBB;
1103 FuncInfo->InsertPt = FuncInfo->MBB->end();
1105 SDB->visitJumpTable(SDB->JTCases[i].second);
1106 CurDAG->setRoot(SDB->getRoot());
1108 CodeGenAndEmitDAG();
1111 for (unsigned pi = 0, pe = FuncInfo->PHINodesToUpdate.size();
1113 MachineInstr *PHI = FuncInfo->PHINodesToUpdate[pi].first;
1114 MachineBasicBlock *PHIBB = PHI->getParent();
1115 assert(PHI->isPHI() &&
1116 "This is not a machine PHI node that we are updating!");
1117 // "default" BB. We can go there only from header BB.
1118 if (PHIBB == SDB->JTCases[i].second.Default) {
1120 (MachineOperand::CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1123 (MachineOperand::CreateMBB(SDB->JTCases[i].first.HeaderBB));
1125 // JT BB. Just iterate over successors here
1126 if (FuncInfo->MBB->isSuccessor(PHIBB)) {
1128 (MachineOperand::CreateReg(FuncInfo->PHINodesToUpdate[pi].second,
1130 PHI->addOperand(MachineOperand::CreateMBB(FuncInfo->MBB));
1134 SDB->JTCases.clear();
1136 // If the switch block involved a branch to one of the actual successors, we
1137 // need to update PHI nodes in that block.
1138 for (unsigned i = 0, e = FuncInfo->PHINodesToUpdate.size(); i != e; ++i) {
1139 MachineInstr *PHI = FuncInfo->PHINodesToUpdate[i].first;
1140 assert(PHI->isPHI() &&
1141 "This is not a machine PHI node that we are updating!");
1142 if (FuncInfo->MBB->isSuccessor(PHI->getParent())) {
1144 MachineOperand::CreateReg(FuncInfo->PHINodesToUpdate[i].second, false));
1145 PHI->addOperand(MachineOperand::CreateMBB(FuncInfo->MBB));
1149 // If we generated any switch lowering information, build and codegen any
1150 // additional DAGs necessary.
1151 for (unsigned i = 0, e = SDB->SwitchCases.size(); i != e; ++i) {
1152 // Set the current basic block to the mbb we wish to insert the code into
1153 FuncInfo->MBB = SDB->SwitchCases[i].ThisBB;
1154 FuncInfo->InsertPt = FuncInfo->MBB->end();
1156 // Determine the unique successors.
1157 SmallVector<MachineBasicBlock *, 2> Succs;
1158 Succs.push_back(SDB->SwitchCases[i].TrueBB);
1159 if (SDB->SwitchCases[i].TrueBB != SDB->SwitchCases[i].FalseBB)
1160 Succs.push_back(SDB->SwitchCases[i].FalseBB);
1162 // Emit the code. Note that this could result in FuncInfo->MBB being split.
1163 SDB->visitSwitchCase(SDB->SwitchCases[i], FuncInfo->MBB);
1164 CurDAG->setRoot(SDB->getRoot());
1166 CodeGenAndEmitDAG();
1168 // Remember the last block, now that any splitting is done, for use in
1169 // populating PHI nodes in successors.
1170 MachineBasicBlock *ThisBB = FuncInfo->MBB;
1172 // Handle any PHI nodes in successors of this chunk, as if we were coming
1173 // from the original BB before switch expansion. Note that PHI nodes can
1174 // occur multiple times in PHINodesToUpdate. We have to be very careful to
1175 // handle them the right number of times.
1176 for (unsigned i = 0, e = Succs.size(); i != e; ++i) {
1177 FuncInfo->MBB = Succs[i];
1178 FuncInfo->InsertPt = FuncInfo->MBB->end();
1179 // FuncInfo->MBB may have been removed from the CFG if a branch was
1181 if (ThisBB->isSuccessor(FuncInfo->MBB)) {
1182 for (MachineBasicBlock::iterator Phi = FuncInfo->MBB->begin();
1183 Phi != FuncInfo->MBB->end() && Phi->isPHI();
1185 // This value for this PHI node is recorded in PHINodesToUpdate.
1186 for (unsigned pn = 0; ; ++pn) {
1187 assert(pn != FuncInfo->PHINodesToUpdate.size() &&
1188 "Didn't find PHI entry!");
1189 if (FuncInfo->PHINodesToUpdate[pn].first == Phi) {
1190 Phi->addOperand(MachineOperand::
1191 CreateReg(FuncInfo->PHINodesToUpdate[pn].second,
1193 Phi->addOperand(MachineOperand::CreateMBB(ThisBB));
1201 SDB->SwitchCases.clear();
1205 /// Create the scheduler. If a specific scheduler was specified
1206 /// via the SchedulerRegistry, use it, otherwise select the
1207 /// one preferred by the target.
1209 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
1210 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1214 RegisterScheduler::setDefault(Ctor);
1217 return Ctor(this, OptLevel);
1220 //===----------------------------------------------------------------------===//
1221 // Helper functions used by the generated instruction selector.
1222 //===----------------------------------------------------------------------===//
1223 // Calls to these methods are generated by tblgen.
1225 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1226 /// the dag combiner simplified the 255, we still want to match. RHS is the
1227 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1228 /// specified in the .td file (e.g. 255).
1229 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1230 int64_t DesiredMaskS) const {
1231 const APInt &ActualMask = RHS->getAPIntValue();
1232 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1234 // If the actual mask exactly matches, success!
1235 if (ActualMask == DesiredMask)
1238 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1239 if (ActualMask.intersects(~DesiredMask))
1242 // Otherwise, the DAG Combiner may have proven that the value coming in is
1243 // either already zero or is not demanded. Check for known zero input bits.
1244 APInt NeededMask = DesiredMask & ~ActualMask;
1245 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1248 // TODO: check to see if missing bits are just not demanded.
1250 // Otherwise, this pattern doesn't match.
1254 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1255 /// the dag combiner simplified the 255, we still want to match. RHS is the
1256 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1257 /// specified in the .td file (e.g. 255).
1258 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1259 int64_t DesiredMaskS) const {
1260 const APInt &ActualMask = RHS->getAPIntValue();
1261 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1263 // If the actual mask exactly matches, success!
1264 if (ActualMask == DesiredMask)
1267 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1268 if (ActualMask.intersects(~DesiredMask))
1271 // Otherwise, the DAG Combiner may have proven that the value coming in is
1272 // either already zero or is not demanded. Check for known zero input bits.
1273 APInt NeededMask = DesiredMask & ~ActualMask;
1275 APInt KnownZero, KnownOne;
1276 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
1278 // If all the missing bits in the or are already known to be set, match!
1279 if ((NeededMask & KnownOne) == NeededMask)
1282 // TODO: check to see if missing bits are just not demanded.
1284 // Otherwise, this pattern doesn't match.
1289 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1290 /// by tblgen. Others should not call it.
1291 void SelectionDAGISel::
1292 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1293 std::vector<SDValue> InOps;
1294 std::swap(InOps, Ops);
1296 Ops.push_back(InOps[InlineAsm::Op_InputChain]); // 0
1297 Ops.push_back(InOps[InlineAsm::Op_AsmString]); // 1
1298 Ops.push_back(InOps[InlineAsm::Op_MDNode]); // 2, !srcloc
1299 Ops.push_back(InOps[InlineAsm::Op_ExtraInfo]); // 3 (SideEffect, AlignStack)
1301 unsigned i = InlineAsm::Op_FirstOperand, e = InOps.size();
1302 if (InOps[e-1].getValueType() == MVT::Glue)
1303 --e; // Don't process a glue operand if it is here.
1306 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1307 if (!InlineAsm::isMemKind(Flags)) {
1308 // Just skip over this operand, copying the operands verbatim.
1309 Ops.insert(Ops.end(), InOps.begin()+i,
1310 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
1311 i += InlineAsm::getNumOperandRegisters(Flags) + 1;
1313 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 &&
1314 "Memory operand with multiple values?");
1315 // Otherwise, this is a memory operand. Ask the target to select it.
1316 std::vector<SDValue> SelOps;
1317 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps))
1318 report_fatal_error("Could not match memory address. Inline asm"
1321 // Add this to the output node.
1323 InlineAsm::getFlagWord(InlineAsm::Kind_Mem, SelOps.size());
1324 Ops.push_back(CurDAG->getTargetConstant(NewFlags, MVT::i32));
1325 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1330 // Add the glue input back if present.
1331 if (e != InOps.size())
1332 Ops.push_back(InOps.back());
1335 /// findGlueUse - Return use of MVT::Glue value produced by the specified
1338 static SDNode *findGlueUse(SDNode *N) {
1339 unsigned FlagResNo = N->getNumValues()-1;
1340 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1341 SDUse &Use = I.getUse();
1342 if (Use.getResNo() == FlagResNo)
1343 return Use.getUser();
1348 /// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
1349 /// This function recursively traverses up the operand chain, ignoring
1351 static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
1352 SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited,
1353 bool IgnoreChains) {
1354 // The NodeID's are given uniques ID's where a node ID is guaranteed to be
1355 // greater than all of its (recursive) operands. If we scan to a point where
1356 // 'use' is smaller than the node we're scanning for, then we know we will
1359 // The Use may be -1 (unassigned) if it is a newly allocated node. This can
1360 // happen because we scan down to newly selected nodes in the case of glue
1362 if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1))
1365 // Don't revisit nodes if we already scanned it and didn't fail, we know we
1366 // won't fail if we scan it again.
1367 if (!Visited.insert(Use))
1370 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
1371 // Ignore chain uses, they are validated by HandleMergeInputChains.
1372 if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains)
1375 SDNode *N = Use->getOperand(i).getNode();
1377 if (Use == ImmedUse || Use == Root)
1378 continue; // We are not looking for immediate use.
1383 // Traverse up the operand chain.
1384 if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
1390 /// IsProfitableToFold - Returns true if it's profitable to fold the specific
1391 /// operand node N of U during instruction selection that starts at Root.
1392 bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
1393 SDNode *Root) const {
1394 if (OptLevel == CodeGenOpt::None) return false;
1395 return N.hasOneUse();
1398 /// IsLegalToFold - Returns true if the specific operand node N of
1399 /// U can be folded during instruction selection that starts at Root.
1400 bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
1401 CodeGenOpt::Level OptLevel,
1402 bool IgnoreChains) {
1403 if (OptLevel == CodeGenOpt::None) return false;
1405 // If Root use can somehow reach N through a path that that doesn't contain
1406 // U then folding N would create a cycle. e.g. In the following
1407 // diagram, Root can reach N through X. If N is folded into into Root, then
1408 // X is both a predecessor and a successor of U.
1419 // * indicates nodes to be folded together.
1421 // If Root produces glue, then it gets (even more) interesting. Since it
1422 // will be "glued" together with its glue use in the scheduler, we need to
1423 // check if it might reach N.
1442 // If GU (glue use) indirectly reaches N (the load), and Root folds N
1443 // (call it Fold), then X is a predecessor of GU and a successor of
1444 // Fold. But since Fold and GU are glued together, this will create
1445 // a cycle in the scheduling graph.
1447 // If the node has glue, walk down the graph to the "lowest" node in the
1449 EVT VT = Root->getValueType(Root->getNumValues()-1);
1450 while (VT == MVT::Glue) {
1451 SDNode *GU = findGlueUse(Root);
1455 VT = Root->getValueType(Root->getNumValues()-1);
1457 // If our query node has a glue result with a use, we've walked up it. If
1458 // the user (which has already been selected) has a chain or indirectly uses
1459 // the chain, our WalkChainUsers predicate will not consider it. Because of
1460 // this, we cannot ignore chains in this predicate.
1461 IgnoreChains = false;
1465 SmallPtrSet<SDNode*, 16> Visited;
1466 return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains);
1469 SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
1470 std::vector<SDValue> Ops(N->op_begin(), N->op_end());
1471 SelectInlineAsmMemoryOperands(Ops);
1473 std::vector<EVT> VTs;
1474 VTs.push_back(MVT::Other);
1475 VTs.push_back(MVT::Glue);
1476 SDValue New = CurDAG->getNode(ISD::INLINEASM, N->getDebugLoc(),
1477 VTs, &Ops[0], Ops.size());
1479 return New.getNode();
1482 SDNode *SelectionDAGISel::Select_UNDEF(SDNode *N) {
1483 return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0));
1486 /// GetVBR - decode a vbr encoding whose top bit is set.
1487 LLVM_ATTRIBUTE_ALWAYS_INLINE static uint64_t
1488 GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
1489 assert(Val >= 128 && "Not a VBR");
1490 Val &= 127; // Remove first vbr bit.
1495 NextBits = MatcherTable[Idx++];
1496 Val |= (NextBits&127) << Shift;
1498 } while (NextBits & 128);
1504 /// UpdateChainsAndGlue - When a match is complete, this method updates uses of
1505 /// interior glue and chain results to use the new glue and chain results.
1506 void SelectionDAGISel::
1507 UpdateChainsAndGlue(SDNode *NodeToMatch, SDValue InputChain,
1508 const SmallVectorImpl<SDNode*> &ChainNodesMatched,
1510 const SmallVectorImpl<SDNode*> &GlueResultNodesMatched,
1511 bool isMorphNodeTo) {
1512 SmallVector<SDNode*, 4> NowDeadNodes;
1514 ISelUpdater ISU(ISelPosition);
1516 // Now that all the normal results are replaced, we replace the chain and
1517 // glue results if present.
1518 if (!ChainNodesMatched.empty()) {
1519 assert(InputChain.getNode() != 0 &&
1520 "Matched input chains but didn't produce a chain");
1521 // Loop over all of the nodes we matched that produced a chain result.
1522 // Replace all the chain results with the final chain we ended up with.
1523 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1524 SDNode *ChainNode = ChainNodesMatched[i];
1526 // If this node was already deleted, don't look at it.
1527 if (ChainNode->getOpcode() == ISD::DELETED_NODE)
1530 // Don't replace the results of the root node if we're doing a
1532 if (ChainNode == NodeToMatch && isMorphNodeTo)
1535 SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1);
1536 if (ChainVal.getValueType() == MVT::Glue)
1537 ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
1538 assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
1539 CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain, &ISU);
1541 // If the node became dead and we haven't already seen it, delete it.
1542 if (ChainNode->use_empty() &&
1543 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), ChainNode))
1544 NowDeadNodes.push_back(ChainNode);
1548 // If the result produces glue, update any glue results in the matched
1549 // pattern with the glue result.
1550 if (InputGlue.getNode() != 0) {
1551 // Handle any interior nodes explicitly marked.
1552 for (unsigned i = 0, e = GlueResultNodesMatched.size(); i != e; ++i) {
1553 SDNode *FRN = GlueResultNodesMatched[i];
1555 // If this node was already deleted, don't look at it.
1556 if (FRN->getOpcode() == ISD::DELETED_NODE)
1559 assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Glue &&
1560 "Doesn't have a glue result");
1561 CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1),
1564 // If the node became dead and we haven't already seen it, delete it.
1565 if (FRN->use_empty() &&
1566 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), FRN))
1567 NowDeadNodes.push_back(FRN);
1571 if (!NowDeadNodes.empty())
1572 CurDAG->RemoveDeadNodes(NowDeadNodes, &ISU);
1574 DEBUG(errs() << "ISEL: Match complete!\n");
1580 CR_LeadsToInteriorNode
1583 /// WalkChainUsers - Walk down the users of the specified chained node that is
1584 /// part of the pattern we're matching, looking at all of the users we find.
1585 /// This determines whether something is an interior node, whether we have a
1586 /// non-pattern node in between two pattern nodes (which prevent folding because
1587 /// it would induce a cycle) and whether we have a TokenFactor node sandwiched
1588 /// between pattern nodes (in which case the TF becomes part of the pattern).
1590 /// The walk we do here is guaranteed to be small because we quickly get down to
1591 /// already selected nodes "below" us.
1593 WalkChainUsers(SDNode *ChainedNode,
1594 SmallVectorImpl<SDNode*> &ChainedNodesInPattern,
1595 SmallVectorImpl<SDNode*> &InteriorChainedNodes) {
1596 ChainResult Result = CR_Simple;
1598 for (SDNode::use_iterator UI = ChainedNode->use_begin(),
1599 E = ChainedNode->use_end(); UI != E; ++UI) {
1600 // Make sure the use is of the chain, not some other value we produce.
1601 if (UI.getUse().getValueType() != MVT::Other) continue;
1605 // If we see an already-selected machine node, then we've gone beyond the
1606 // pattern that we're selecting down into the already selected chunk of the
1608 if (User->isMachineOpcode() ||
1609 User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
1612 if (User->getOpcode() == ISD::CopyToReg ||
1613 User->getOpcode() == ISD::CopyFromReg ||
1614 User->getOpcode() == ISD::INLINEASM ||
1615 User->getOpcode() == ISD::EH_LABEL) {
1616 // If their node ID got reset to -1 then they've already been selected.
1617 // Treat them like a MachineOpcode.
1618 if (User->getNodeId() == -1)
1622 // If we have a TokenFactor, we handle it specially.
1623 if (User->getOpcode() != ISD::TokenFactor) {
1624 // If the node isn't a token factor and isn't part of our pattern, then it
1625 // must be a random chained node in between two nodes we're selecting.
1626 // This happens when we have something like:
1631 // Because we structurally match the load/store as a read/modify/write,
1632 // but the call is chained between them. We cannot fold in this case
1633 // because it would induce a cycle in the graph.
1634 if (!std::count(ChainedNodesInPattern.begin(),
1635 ChainedNodesInPattern.end(), User))
1636 return CR_InducesCycle;
1638 // Otherwise we found a node that is part of our pattern. For example in:
1642 // This would happen when we're scanning down from the load and see the
1643 // store as a user. Record that there is a use of ChainedNode that is
1644 // part of the pattern and keep scanning uses.
1645 Result = CR_LeadsToInteriorNode;
1646 InteriorChainedNodes.push_back(User);
1650 // If we found a TokenFactor, there are two cases to consider: first if the
1651 // TokenFactor is just hanging "below" the pattern we're matching (i.e. no
1652 // uses of the TF are in our pattern) we just want to ignore it. Second,
1653 // the TokenFactor can be sandwiched in between two chained nodes, like so:
1659 // | \ DAG's like cheese
1662 // [TokenFactor] [Op]
1669 // In this case, the TokenFactor becomes part of our match and we rewrite it
1670 // as a new TokenFactor.
1672 // To distinguish these two cases, do a recursive walk down the uses.
1673 switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) {
1675 // If the uses of the TokenFactor are just already-selected nodes, ignore
1676 // it, it is "below" our pattern.
1678 case CR_InducesCycle:
1679 // If the uses of the TokenFactor lead to nodes that are not part of our
1680 // pattern that are not selected, folding would turn this into a cycle,
1682 return CR_InducesCycle;
1683 case CR_LeadsToInteriorNode:
1684 break; // Otherwise, keep processing.
1687 // Okay, we know we're in the interesting interior case. The TokenFactor
1688 // is now going to be considered part of the pattern so that we rewrite its
1689 // uses (it may have uses that are not part of the pattern) with the
1690 // ultimate chain result of the generated code. We will also add its chain
1691 // inputs as inputs to the ultimate TokenFactor we create.
1692 Result = CR_LeadsToInteriorNode;
1693 ChainedNodesInPattern.push_back(User);
1694 InteriorChainedNodes.push_back(User);
1701 /// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
1702 /// operation for when the pattern matched at least one node with a chains. The
1703 /// input vector contains a list of all of the chained nodes that we match. We
1704 /// must determine if this is a valid thing to cover (i.e. matching it won't
1705 /// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
1706 /// be used as the input node chain for the generated nodes.
1708 HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
1709 SelectionDAG *CurDAG) {
1710 // Walk all of the chained nodes we've matched, recursively scanning down the
1711 // users of the chain result. This adds any TokenFactor nodes that are caught
1712 // in between chained nodes to the chained and interior nodes list.
1713 SmallVector<SDNode*, 3> InteriorChainedNodes;
1714 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1715 if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched,
1716 InteriorChainedNodes) == CR_InducesCycle)
1717 return SDValue(); // Would induce a cycle.
1720 // Okay, we have walked all the matched nodes and collected TokenFactor nodes
1721 // that we are interested in. Form our input TokenFactor node.
1722 SmallVector<SDValue, 3> InputChains;
1723 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1724 // Add the input chain of this node to the InputChains list (which will be
1725 // the operands of the generated TokenFactor) if it's not an interior node.
1726 SDNode *N = ChainNodesMatched[i];
1727 if (N->getOpcode() != ISD::TokenFactor) {
1728 if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N))
1731 // Otherwise, add the input chain.
1732 SDValue InChain = ChainNodesMatched[i]->getOperand(0);
1733 assert(InChain.getValueType() == MVT::Other && "Not a chain");
1734 InputChains.push_back(InChain);
1738 // If we have a token factor, we want to add all inputs of the token factor
1739 // that are not part of the pattern we're matching.
1740 for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) {
1741 if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(),
1742 N->getOperand(op).getNode()))
1743 InputChains.push_back(N->getOperand(op));
1748 if (InputChains.size() == 1)
1749 return InputChains[0];
1750 return CurDAG->getNode(ISD::TokenFactor, ChainNodesMatched[0]->getDebugLoc(),
1751 MVT::Other, &InputChains[0], InputChains.size());
1754 /// MorphNode - Handle morphing a node in place for the selector.
1755 SDNode *SelectionDAGISel::
1756 MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList,
1757 const SDValue *Ops, unsigned NumOps, unsigned EmitNodeInfo) {
1758 // It is possible we're using MorphNodeTo to replace a node with no
1759 // normal results with one that has a normal result (or we could be
1760 // adding a chain) and the input could have glue and chains as well.
1761 // In this case we need to shift the operands down.
1762 // FIXME: This is a horrible hack and broken in obscure cases, no worse
1763 // than the old isel though.
1764 int OldGlueResultNo = -1, OldChainResultNo = -1;
1766 unsigned NTMNumResults = Node->getNumValues();
1767 if (Node->getValueType(NTMNumResults-1) == MVT::Glue) {
1768 OldGlueResultNo = NTMNumResults-1;
1769 if (NTMNumResults != 1 &&
1770 Node->getValueType(NTMNumResults-2) == MVT::Other)
1771 OldChainResultNo = NTMNumResults-2;
1772 } else if (Node->getValueType(NTMNumResults-1) == MVT::Other)
1773 OldChainResultNo = NTMNumResults-1;
1775 // Call the underlying SelectionDAG routine to do the transmogrification. Note
1776 // that this deletes operands of the old node that become dead.
1777 SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops, NumOps);
1779 // MorphNodeTo can operate in two ways: if an existing node with the
1780 // specified operands exists, it can just return it. Otherwise, it
1781 // updates the node in place to have the requested operands.
1783 // If we updated the node in place, reset the node ID. To the isel,
1784 // this should be just like a newly allocated machine node.
1788 unsigned ResNumResults = Res->getNumValues();
1789 // Move the glue if needed.
1790 if ((EmitNodeInfo & OPFL_GlueOutput) && OldGlueResultNo != -1 &&
1791 (unsigned)OldGlueResultNo != ResNumResults-1)
1792 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldGlueResultNo),
1793 SDValue(Res, ResNumResults-1));
1795 if ((EmitNodeInfo & OPFL_GlueOutput) != 0)
1798 // Move the chain reference if needed.
1799 if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 &&
1800 (unsigned)OldChainResultNo != ResNumResults-1)
1801 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo),
1802 SDValue(Res, ResNumResults-1));
1804 // Otherwise, no replacement happened because the node already exists. Replace
1805 // Uses of the old node with the new one.
1807 CurDAG->ReplaceAllUsesWith(Node, Res);
1812 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1813 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1814 CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1816 const SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes) {
1817 // Accept if it is exactly the same as a previously recorded node.
1818 unsigned RecNo = MatcherTable[MatcherIndex++];
1819 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
1820 return N == RecordedNodes[RecNo].first;
1823 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1824 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1825 CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1826 SelectionDAGISel &SDISel) {
1827 return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]);
1830 /// CheckNodePredicate - Implements OP_CheckNodePredicate.
1831 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1832 CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1833 SelectionDAGISel &SDISel, SDNode *N) {
1834 return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]);
1837 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1838 CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1840 uint16_t Opc = MatcherTable[MatcherIndex++];
1841 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
1842 return N->getOpcode() == Opc;
1845 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1846 CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1847 SDValue N, const TargetLowering &TLI) {
1848 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1849 if (N.getValueType() == VT) return true;
1851 // Handle the case when VT is iPTR.
1852 return VT == MVT::iPTR && N.getValueType() == TLI.getPointerTy();
1855 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1856 CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1857 SDValue N, const TargetLowering &TLI,
1859 if (ChildNo >= N.getNumOperands())
1860 return false; // Match fails if out of range child #.
1861 return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI);
1865 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1866 CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1868 return cast<CondCodeSDNode>(N)->get() ==
1869 (ISD::CondCode)MatcherTable[MatcherIndex++];
1872 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1873 CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1874 SDValue N, const TargetLowering &TLI) {
1875 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1876 if (cast<VTSDNode>(N)->getVT() == VT)
1879 // Handle the case when VT is iPTR.
1880 return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI.getPointerTy();
1883 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1884 CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1886 int64_t Val = MatcherTable[MatcherIndex++];
1888 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1890 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
1891 return C != 0 && C->getSExtValue() == Val;
1894 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1895 CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1896 SDValue N, SelectionDAGISel &SDISel) {
1897 int64_t Val = MatcherTable[MatcherIndex++];
1899 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1901 if (N->getOpcode() != ISD::AND) return false;
1903 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1904 return C != 0 && SDISel.CheckAndMask(N.getOperand(0), C, Val);
1907 LLVM_ATTRIBUTE_ALWAYS_INLINE static bool
1908 CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1909 SDValue N, SelectionDAGISel &SDISel) {
1910 int64_t Val = MatcherTable[MatcherIndex++];
1912 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1914 if (N->getOpcode() != ISD::OR) return false;
1916 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1917 return C != 0 && SDISel.CheckOrMask(N.getOperand(0), C, Val);
1920 /// IsPredicateKnownToFail - If we know how and can do so without pushing a
1921 /// scope, evaluate the current node. If the current predicate is known to
1922 /// fail, set Result=true and return anything. If the current predicate is
1923 /// known to pass, set Result=false and return the MatcherIndex to continue
1924 /// with. If the current predicate is unknown, set Result=false and return the
1925 /// MatcherIndex to continue with.
1926 static unsigned IsPredicateKnownToFail(const unsigned char *Table,
1927 unsigned Index, SDValue N,
1928 bool &Result, SelectionDAGISel &SDISel,
1929 SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes) {
1930 switch (Table[Index++]) {
1933 return Index-1; // Could not evaluate this predicate.
1934 case SelectionDAGISel::OPC_CheckSame:
1935 Result = !::CheckSame(Table, Index, N, RecordedNodes);
1937 case SelectionDAGISel::OPC_CheckPatternPredicate:
1938 Result = !::CheckPatternPredicate(Table, Index, SDISel);
1940 case SelectionDAGISel::OPC_CheckPredicate:
1941 Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode());
1943 case SelectionDAGISel::OPC_CheckOpcode:
1944 Result = !::CheckOpcode(Table, Index, N.getNode());
1946 case SelectionDAGISel::OPC_CheckType:
1947 Result = !::CheckType(Table, Index, N, SDISel.TLI);
1949 case SelectionDAGISel::OPC_CheckChild0Type:
1950 case SelectionDAGISel::OPC_CheckChild1Type:
1951 case SelectionDAGISel::OPC_CheckChild2Type:
1952 case SelectionDAGISel::OPC_CheckChild3Type:
1953 case SelectionDAGISel::OPC_CheckChild4Type:
1954 case SelectionDAGISel::OPC_CheckChild5Type:
1955 case SelectionDAGISel::OPC_CheckChild6Type:
1956 case SelectionDAGISel::OPC_CheckChild7Type:
1957 Result = !::CheckChildType(Table, Index, N, SDISel.TLI,
1958 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Type);
1960 case SelectionDAGISel::OPC_CheckCondCode:
1961 Result = !::CheckCondCode(Table, Index, N);
1963 case SelectionDAGISel::OPC_CheckValueType:
1964 Result = !::CheckValueType(Table, Index, N, SDISel.TLI);
1966 case SelectionDAGISel::OPC_CheckInteger:
1967 Result = !::CheckInteger(Table, Index, N);
1969 case SelectionDAGISel::OPC_CheckAndImm:
1970 Result = !::CheckAndImm(Table, Index, N, SDISel);
1972 case SelectionDAGISel::OPC_CheckOrImm:
1973 Result = !::CheckOrImm(Table, Index, N, SDISel);
1981 /// FailIndex - If this match fails, this is the index to continue with.
1984 /// NodeStack - The node stack when the scope was formed.
1985 SmallVector<SDValue, 4> NodeStack;
1987 /// NumRecordedNodes - The number of recorded nodes when the scope was formed.
1988 unsigned NumRecordedNodes;
1990 /// NumMatchedMemRefs - The number of matched memref entries.
1991 unsigned NumMatchedMemRefs;
1993 /// InputChain/InputGlue - The current chain/glue
1994 SDValue InputChain, InputGlue;
1996 /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty.
1997 bool HasChainNodesMatched, HasGlueResultNodesMatched;
2002 SDNode *SelectionDAGISel::
2003 SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
2004 unsigned TableSize) {
2005 // FIXME: Should these even be selected? Handle these cases in the caller?
2006 switch (NodeToMatch->getOpcode()) {
2009 case ISD::EntryToken: // These nodes remain the same.
2010 case ISD::BasicBlock:
2012 //case ISD::VALUETYPE:
2013 //case ISD::CONDCODE:
2014 case ISD::HANDLENODE:
2015 case ISD::MDNODE_SDNODE:
2016 case ISD::TargetConstant:
2017 case ISD::TargetConstantFP:
2018 case ISD::TargetConstantPool:
2019 case ISD::TargetFrameIndex:
2020 case ISD::TargetExternalSymbol:
2021 case ISD::TargetBlockAddress:
2022 case ISD::TargetJumpTable:
2023 case ISD::TargetGlobalTLSAddress:
2024 case ISD::TargetGlobalAddress:
2025 case ISD::TokenFactor:
2026 case ISD::CopyFromReg:
2027 case ISD::CopyToReg:
2029 NodeToMatch->setNodeId(-1); // Mark selected.
2031 case ISD::AssertSext:
2032 case ISD::AssertZext:
2033 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0),
2034 NodeToMatch->getOperand(0));
2036 case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch);
2037 case ISD::UNDEF: return Select_UNDEF(NodeToMatch);
2040 assert(!NodeToMatch->isMachineOpcode() && "Node already selected!");
2042 // Set up the node stack with NodeToMatch as the only node on the stack.
2043 SmallVector<SDValue, 8> NodeStack;
2044 SDValue N = SDValue(NodeToMatch, 0);
2045 NodeStack.push_back(N);
2047 // MatchScopes - Scopes used when matching, if a match failure happens, this
2048 // indicates where to continue checking.
2049 SmallVector<MatchScope, 8> MatchScopes;
2051 // RecordedNodes - This is the set of nodes that have been recorded by the
2052 // state machine. The second value is the parent of the node, or null if the
2053 // root is recorded.
2054 SmallVector<std::pair<SDValue, SDNode*>, 8> RecordedNodes;
2056 // MatchedMemRefs - This is the set of MemRef's we've seen in the input
2058 SmallVector<MachineMemOperand*, 2> MatchedMemRefs;
2060 // These are the current input chain and glue for use when generating nodes.
2061 // Various Emit operations change these. For example, emitting a copytoreg
2062 // uses and updates these.
2063 SDValue InputChain, InputGlue;
2065 // ChainNodesMatched - If a pattern matches nodes that have input/output
2066 // chains, the OPC_EmitMergeInputChains operation is emitted which indicates
2067 // which ones they are. The result is captured into this list so that we can
2068 // update the chain results when the pattern is complete.
2069 SmallVector<SDNode*, 3> ChainNodesMatched;
2070 SmallVector<SDNode*, 3> GlueResultNodesMatched;
2072 DEBUG(errs() << "ISEL: Starting pattern match on root node: ";
2073 NodeToMatch->dump(CurDAG);
2076 // Determine where to start the interpreter. Normally we start at opcode #0,
2077 // but if the state machine starts with an OPC_SwitchOpcode, then we
2078 // accelerate the first lookup (which is guaranteed to be hot) with the
2079 // OpcodeOffset table.
2080 unsigned MatcherIndex = 0;
2082 if (!OpcodeOffset.empty()) {
2083 // Already computed the OpcodeOffset table, just index into it.
2084 if (N.getOpcode() < OpcodeOffset.size())
2085 MatcherIndex = OpcodeOffset[N.getOpcode()];
2086 DEBUG(errs() << " Initial Opcode index to " << MatcherIndex << "\n");
2088 } else if (MatcherTable[0] == OPC_SwitchOpcode) {
2089 // Otherwise, the table isn't computed, but the state machine does start
2090 // with an OPC_SwitchOpcode instruction. Populate the table now, since this
2091 // is the first time we're selecting an instruction.
2094 // Get the size of this case.
2095 unsigned CaseSize = MatcherTable[Idx++];
2097 CaseSize = GetVBR(CaseSize, MatcherTable, Idx);
2098 if (CaseSize == 0) break;
2100 // Get the opcode, add the index to the table.
2101 uint16_t Opc = MatcherTable[Idx++];
2102 Opc |= (unsigned short)MatcherTable[Idx++] << 8;
2103 if (Opc >= OpcodeOffset.size())
2104 OpcodeOffset.resize((Opc+1)*2);
2105 OpcodeOffset[Opc] = Idx;
2109 // Okay, do the lookup for the first opcode.
2110 if (N.getOpcode() < OpcodeOffset.size())
2111 MatcherIndex = OpcodeOffset[N.getOpcode()];
2115 assert(MatcherIndex < TableSize && "Invalid index");
2117 unsigned CurrentOpcodeIndex = MatcherIndex;
2119 BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++];
2122 // Okay, the semantics of this operation are that we should push a scope
2123 // then evaluate the first child. However, pushing a scope only to have
2124 // the first check fail (which then pops it) is inefficient. If we can
2125 // determine immediately that the first check (or first several) will
2126 // immediately fail, don't even bother pushing a scope for them.
2130 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2131 if (NumToSkip & 128)
2132 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2133 // Found the end of the scope with no match.
2134 if (NumToSkip == 0) {
2139 FailIndex = MatcherIndex+NumToSkip;
2141 unsigned MatcherIndexOfPredicate = MatcherIndex;
2142 (void)MatcherIndexOfPredicate; // silence warning.
2144 // If we can't evaluate this predicate without pushing a scope (e.g. if
2145 // it is a 'MoveParent') or if the predicate succeeds on this node, we
2146 // push the scope and evaluate the full predicate chain.
2148 MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
2149 Result, *this, RecordedNodes);
2153 DEBUG(errs() << " Skipped scope entry (due to false predicate) at "
2154 << "index " << MatcherIndexOfPredicate
2155 << ", continuing at " << FailIndex << "\n");
2156 ++NumDAGIselRetries;
2158 // Otherwise, we know that this case of the Scope is guaranteed to fail,
2159 // move to the next case.
2160 MatcherIndex = FailIndex;
2163 // If the whole scope failed to match, bail.
2164 if (FailIndex == 0) break;
2166 // Push a MatchScope which indicates where to go if the first child fails
2168 MatchScope NewEntry;
2169 NewEntry.FailIndex = FailIndex;
2170 NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
2171 NewEntry.NumRecordedNodes = RecordedNodes.size();
2172 NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
2173 NewEntry.InputChain = InputChain;
2174 NewEntry.InputGlue = InputGlue;
2175 NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty();
2176 NewEntry.HasGlueResultNodesMatched = !GlueResultNodesMatched.empty();
2177 MatchScopes.push_back(NewEntry);
2180 case OPC_RecordNode: {
2181 // Remember this node, it may end up being an operand in the pattern.
2183 if (NodeStack.size() > 1)
2184 Parent = NodeStack[NodeStack.size()-2].getNode();
2185 RecordedNodes.push_back(std::make_pair(N, Parent));
2189 case OPC_RecordChild0: case OPC_RecordChild1:
2190 case OPC_RecordChild2: case OPC_RecordChild3:
2191 case OPC_RecordChild4: case OPC_RecordChild5:
2192 case OPC_RecordChild6: case OPC_RecordChild7: {
2193 unsigned ChildNo = Opcode-OPC_RecordChild0;
2194 if (ChildNo >= N.getNumOperands())
2195 break; // Match fails if out of range child #.
2197 RecordedNodes.push_back(std::make_pair(N->getOperand(ChildNo),
2201 case OPC_RecordMemRef:
2202 MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand());
2205 case OPC_CaptureGlueInput:
2206 // If the current node has an input glue, capture it in InputGlue.
2207 if (N->getNumOperands() != 0 &&
2208 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue)
2209 InputGlue = N->getOperand(N->getNumOperands()-1);
2212 case OPC_MoveChild: {
2213 unsigned ChildNo = MatcherTable[MatcherIndex++];
2214 if (ChildNo >= N.getNumOperands())
2215 break; // Match fails if out of range child #.
2216 N = N.getOperand(ChildNo);
2217 NodeStack.push_back(N);
2221 case OPC_MoveParent:
2222 // Pop the current node off the NodeStack.
2223 NodeStack.pop_back();
2224 assert(!NodeStack.empty() && "Node stack imbalance!");
2225 N = NodeStack.back();
2229 if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
2231 case OPC_CheckPatternPredicate:
2232 if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break;
2234 case OPC_CheckPredicate:
2235 if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this,
2239 case OPC_CheckComplexPat: {
2240 unsigned CPNum = MatcherTable[MatcherIndex++];
2241 unsigned RecNo = MatcherTable[MatcherIndex++];
2242 assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
2243 if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo].second,
2244 RecordedNodes[RecNo].first, CPNum,
2249 case OPC_CheckOpcode:
2250 if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
2254 if (!::CheckType(MatcherTable, MatcherIndex, N, TLI)) break;
2257 case OPC_SwitchOpcode: {
2258 unsigned CurNodeOpcode = N.getOpcode();
2259 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2262 // Get the size of this case.
2263 CaseSize = MatcherTable[MatcherIndex++];
2265 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2266 if (CaseSize == 0) break;
2268 uint16_t Opc = MatcherTable[MatcherIndex++];
2269 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2271 // If the opcode matches, then we will execute this case.
2272 if (CurNodeOpcode == Opc)
2275 // Otherwise, skip over this case.
2276 MatcherIndex += CaseSize;
2279 // If no cases matched, bail out.
2280 if (CaseSize == 0) break;
2282 // Otherwise, execute the case we found.
2283 DEBUG(errs() << " OpcodeSwitch from " << SwitchStart
2284 << " to " << MatcherIndex << "\n");
2288 case OPC_SwitchType: {
2289 MVT CurNodeVT = N.getValueType().getSimpleVT();
2290 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2293 // Get the size of this case.
2294 CaseSize = MatcherTable[MatcherIndex++];
2296 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2297 if (CaseSize == 0) break;
2299 MVT CaseVT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2300 if (CaseVT == MVT::iPTR)
2301 CaseVT = TLI.getPointerTy();
2303 // If the VT matches, then we will execute this case.
2304 if (CurNodeVT == CaseVT)
2307 // Otherwise, skip over this case.
2308 MatcherIndex += CaseSize;
2311 // If no cases matched, bail out.
2312 if (CaseSize == 0) break;
2314 // Otherwise, execute the case we found.
2315 DEBUG(errs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
2316 << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
2319 case OPC_CheckChild0Type: case OPC_CheckChild1Type:
2320 case OPC_CheckChild2Type: case OPC_CheckChild3Type:
2321 case OPC_CheckChild4Type: case OPC_CheckChild5Type:
2322 case OPC_CheckChild6Type: case OPC_CheckChild7Type:
2323 if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI,
2324 Opcode-OPC_CheckChild0Type))
2327 case OPC_CheckCondCode:
2328 if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
2330 case OPC_CheckValueType:
2331 if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI)) break;
2333 case OPC_CheckInteger:
2334 if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
2336 case OPC_CheckAndImm:
2337 if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
2339 case OPC_CheckOrImm:
2340 if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
2343 case OPC_CheckFoldableChainNode: {
2344 assert(NodeStack.size() != 1 && "No parent node");
2345 // Verify that all intermediate nodes between the root and this one have
2347 bool HasMultipleUses = false;
2348 for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i)
2349 if (!NodeStack[i].hasOneUse()) {
2350 HasMultipleUses = true;
2353 if (HasMultipleUses) break;
2355 // Check to see that the target thinks this is profitable to fold and that
2356 // we can fold it without inducing cycles in the graph.
2357 if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2359 !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2360 NodeToMatch, OptLevel,
2361 true/*We validate our own chains*/))
2366 case OPC_EmitInteger: {
2367 MVT::SimpleValueType VT =
2368 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2369 int64_t Val = MatcherTable[MatcherIndex++];
2371 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2372 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2373 CurDAG->getTargetConstant(Val, VT), (SDNode*)0));
2376 case OPC_EmitRegister: {
2377 MVT::SimpleValueType VT =
2378 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2379 unsigned RegNo = MatcherTable[MatcherIndex++];
2380 RecordedNodes.push_back(std::pair<SDValue, SDNode*>(
2381 CurDAG->getRegister(RegNo, VT), (SDNode*)0));
2385 case OPC_EmitConvertToTarget: {
2386 // Convert from IMM/FPIMM to target version.
2387 unsigned RecNo = MatcherTable[MatcherIndex++];
2388 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2389 SDValue Imm = RecordedNodes[RecNo].first;
2391 if (Imm->getOpcode() == ISD::Constant) {
2392 int64_t Val = cast<ConstantSDNode>(Imm)->getZExtValue();
2393 Imm = CurDAG->getTargetConstant(Val, Imm.getValueType());
2394 } else if (Imm->getOpcode() == ISD::ConstantFP) {
2395 const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue();
2396 Imm = CurDAG->getTargetConstantFP(*Val, Imm.getValueType());
2399 RecordedNodes.push_back(std::make_pair(Imm, RecordedNodes[RecNo].second));
2403 case OPC_EmitMergeInputChains1_0: // OPC_EmitMergeInputChains, 1, 0
2404 case OPC_EmitMergeInputChains1_1: { // OPC_EmitMergeInputChains, 1, 1
2405 // These are space-optimized forms of OPC_EmitMergeInputChains.
2406 assert(InputChain.getNode() == 0 &&
2407 "EmitMergeInputChains should be the first chain producing node");
2408 assert(ChainNodesMatched.empty() &&
2409 "Should only have one EmitMergeInputChains per match");
2411 // Read all of the chained nodes.
2412 unsigned RecNo = Opcode == OPC_EmitMergeInputChains1_1;
2413 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2414 ChainNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
2416 // FIXME: What if other value results of the node have uses not matched
2418 if (ChainNodesMatched.back() != NodeToMatch &&
2419 !RecordedNodes[RecNo].first.hasOneUse()) {
2420 ChainNodesMatched.clear();
2424 // Merge the input chains if they are not intra-pattern references.
2425 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2427 if (InputChain.getNode() == 0)
2428 break; // Failed to merge.
2432 case OPC_EmitMergeInputChains: {
2433 assert(InputChain.getNode() == 0 &&
2434 "EmitMergeInputChains should be the first chain producing node");
2435 // This node gets a list of nodes we matched in the input that have
2436 // chains. We want to token factor all of the input chains to these nodes
2437 // together. However, if any of the input chains is actually one of the
2438 // nodes matched in this pattern, then we have an intra-match reference.
2439 // Ignore these because the newly token factored chain should not refer to
2441 unsigned NumChains = MatcherTable[MatcherIndex++];
2442 assert(NumChains != 0 && "Can't TF zero chains");
2444 assert(ChainNodesMatched.empty() &&
2445 "Should only have one EmitMergeInputChains per match");
2447 // Read all of the chained nodes.
2448 for (unsigned i = 0; i != NumChains; ++i) {
2449 unsigned RecNo = MatcherTable[MatcherIndex++];
2450 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2451 ChainNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
2453 // FIXME: What if other value results of the node have uses not matched
2455 if (ChainNodesMatched.back() != NodeToMatch &&
2456 !RecordedNodes[RecNo].first.hasOneUse()) {
2457 ChainNodesMatched.clear();
2462 // If the inner loop broke out, the match fails.
2463 if (ChainNodesMatched.empty())
2466 // Merge the input chains if they are not intra-pattern references.
2467 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2469 if (InputChain.getNode() == 0)
2470 break; // Failed to merge.
2475 case OPC_EmitCopyToReg: {
2476 unsigned RecNo = MatcherTable[MatcherIndex++];
2477 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2478 unsigned DestPhysReg = MatcherTable[MatcherIndex++];
2480 if (InputChain.getNode() == 0)
2481 InputChain = CurDAG->getEntryNode();
2483 InputChain = CurDAG->getCopyToReg(InputChain, NodeToMatch->getDebugLoc(),
2484 DestPhysReg, RecordedNodes[RecNo].first,
2487 InputGlue = InputChain.getValue(1);
2491 case OPC_EmitNodeXForm: {
2492 unsigned XFormNo = MatcherTable[MatcherIndex++];
2493 unsigned RecNo = MatcherTable[MatcherIndex++];
2494 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2495 SDValue Res = RunSDNodeXForm(RecordedNodes[RecNo].first, XFormNo);
2496 RecordedNodes.push_back(std::pair<SDValue,SDNode*>(Res, (SDNode*) 0));
2501 case OPC_MorphNodeTo: {
2502 uint16_t TargetOpc = MatcherTable[MatcherIndex++];
2503 TargetOpc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2504 unsigned EmitNodeInfo = MatcherTable[MatcherIndex++];
2505 // Get the result VT list.
2506 unsigned NumVTs = MatcherTable[MatcherIndex++];
2507 SmallVector<EVT, 4> VTs;
2508 for (unsigned i = 0; i != NumVTs; ++i) {
2509 MVT::SimpleValueType VT =
2510 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2511 if (VT == MVT::iPTR) VT = TLI.getPointerTy().SimpleTy;
2515 if (EmitNodeInfo & OPFL_Chain)
2516 VTs.push_back(MVT::Other);
2517 if (EmitNodeInfo & OPFL_GlueOutput)
2518 VTs.push_back(MVT::Glue);
2520 // This is hot code, so optimize the two most common cases of 1 and 2
2523 if (VTs.size() == 1)
2524 VTList = CurDAG->getVTList(VTs[0]);
2525 else if (VTs.size() == 2)
2526 VTList = CurDAG->getVTList(VTs[0], VTs[1]);
2528 VTList = CurDAG->getVTList(VTs.data(), VTs.size());
2530 // Get the operand list.
2531 unsigned NumOps = MatcherTable[MatcherIndex++];
2532 SmallVector<SDValue, 8> Ops;
2533 for (unsigned i = 0; i != NumOps; ++i) {
2534 unsigned RecNo = MatcherTable[MatcherIndex++];
2536 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2538 assert(RecNo < RecordedNodes.size() && "Invalid EmitNode");
2539 Ops.push_back(RecordedNodes[RecNo].first);
2542 // If there are variadic operands to add, handle them now.
2543 if (EmitNodeInfo & OPFL_VariadicInfo) {
2544 // Determine the start index to copy from.
2545 unsigned FirstOpToCopy = getNumFixedFromVariadicInfo(EmitNodeInfo);
2546 FirstOpToCopy += (EmitNodeInfo & OPFL_Chain) ? 1 : 0;
2547 assert(NodeToMatch->getNumOperands() >= FirstOpToCopy &&
2548 "Invalid variadic node");
2549 // Copy all of the variadic operands, not including a potential glue
2551 for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands();
2553 SDValue V = NodeToMatch->getOperand(i);
2554 if (V.getValueType() == MVT::Glue) break;
2559 // If this has chain/glue inputs, add them.
2560 if (EmitNodeInfo & OPFL_Chain)
2561 Ops.push_back(InputChain);
2562 if ((EmitNodeInfo & OPFL_GlueInput) && InputGlue.getNode() != 0)
2563 Ops.push_back(InputGlue);
2567 if (Opcode != OPC_MorphNodeTo) {
2568 // If this is a normal EmitNode command, just create the new node and
2569 // add the results to the RecordedNodes list.
2570 Res = CurDAG->getMachineNode(TargetOpc, NodeToMatch->getDebugLoc(),
2571 VTList, Ops.data(), Ops.size());
2573 // Add all the non-glue/non-chain results to the RecordedNodes list.
2574 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
2575 if (VTs[i] == MVT::Other || VTs[i] == MVT::Glue) break;
2576 RecordedNodes.push_back(std::pair<SDValue,SDNode*>(SDValue(Res, i),
2581 Res = MorphNode(NodeToMatch, TargetOpc, VTList, Ops.data(), Ops.size(),
2585 // If the node had chain/glue results, update our notion of the current
2587 if (EmitNodeInfo & OPFL_GlueOutput) {
2588 InputGlue = SDValue(Res, VTs.size()-1);
2589 if (EmitNodeInfo & OPFL_Chain)
2590 InputChain = SDValue(Res, VTs.size()-2);
2591 } else if (EmitNodeInfo & OPFL_Chain)
2592 InputChain = SDValue(Res, VTs.size()-1);
2594 // If the OPFL_MemRefs glue is set on this node, slap all of the
2595 // accumulated memrefs onto it.
2597 // FIXME: This is vastly incorrect for patterns with multiple outputs
2598 // instructions that access memory and for ComplexPatterns that match
2600 if (EmitNodeInfo & OPFL_MemRefs) {
2601 MachineSDNode::mmo_iterator MemRefs =
2602 MF->allocateMemRefsArray(MatchedMemRefs.size());
2603 std::copy(MatchedMemRefs.begin(), MatchedMemRefs.end(), MemRefs);
2604 cast<MachineSDNode>(Res)
2605 ->setMemRefs(MemRefs, MemRefs + MatchedMemRefs.size());
2609 << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created")
2610 << " node: "; Res->dump(CurDAG); errs() << "\n");
2612 // If this was a MorphNodeTo then we're completely done!
2613 if (Opcode == OPC_MorphNodeTo) {
2614 // Update chain and glue uses.
2615 UpdateChainsAndGlue(NodeToMatch, InputChain, ChainNodesMatched,
2616 InputGlue, GlueResultNodesMatched, true);
2623 case OPC_MarkGlueResults: {
2624 unsigned NumNodes = MatcherTable[MatcherIndex++];
2626 // Read and remember all the glue-result nodes.
2627 for (unsigned i = 0; i != NumNodes; ++i) {
2628 unsigned RecNo = MatcherTable[MatcherIndex++];
2630 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2632 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2633 GlueResultNodesMatched.push_back(RecordedNodes[RecNo].first.getNode());
2638 case OPC_CompleteMatch: {
2639 // The match has been completed, and any new nodes (if any) have been
2640 // created. Patch up references to the matched dag to use the newly
2642 unsigned NumResults = MatcherTable[MatcherIndex++];
2644 for (unsigned i = 0; i != NumResults; ++i) {
2645 unsigned ResSlot = MatcherTable[MatcherIndex++];
2647 ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex);
2649 assert(ResSlot < RecordedNodes.size() && "Invalid CheckSame");
2650 SDValue Res = RecordedNodes[ResSlot].first;
2652 assert(i < NodeToMatch->getNumValues() &&
2653 NodeToMatch->getValueType(i) != MVT::Other &&
2654 NodeToMatch->getValueType(i) != MVT::Glue &&
2655 "Invalid number of results to complete!");
2656 assert((NodeToMatch->getValueType(i) == Res.getValueType() ||
2657 NodeToMatch->getValueType(i) == MVT::iPTR ||
2658 Res.getValueType() == MVT::iPTR ||
2659 NodeToMatch->getValueType(i).getSizeInBits() ==
2660 Res.getValueType().getSizeInBits()) &&
2661 "invalid replacement");
2662 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, i), Res);
2665 // If the root node defines glue, add it to the glue nodes to update list.
2666 if (NodeToMatch->getValueType(NodeToMatch->getNumValues()-1) == MVT::Glue)
2667 GlueResultNodesMatched.push_back(NodeToMatch);
2669 // Update chain and glue uses.
2670 UpdateChainsAndGlue(NodeToMatch, InputChain, ChainNodesMatched,
2671 InputGlue, GlueResultNodesMatched, false);
2673 assert(NodeToMatch->use_empty() &&
2674 "Didn't replace all uses of the node?");
2676 // FIXME: We just return here, which interacts correctly with SelectRoot
2677 // above. We should fix this to not return an SDNode* anymore.
2682 // If the code reached this point, then the match failed. See if there is
2683 // another child to try in the current 'Scope', otherwise pop it until we
2684 // find a case to check.
2685 DEBUG(errs() << " Match failed at index " << CurrentOpcodeIndex << "\n");
2686 ++NumDAGIselRetries;
2688 if (MatchScopes.empty()) {
2689 CannotYetSelect(NodeToMatch);
2693 // Restore the interpreter state back to the point where the scope was
2695 MatchScope &LastScope = MatchScopes.back();
2696 RecordedNodes.resize(LastScope.NumRecordedNodes);
2698 NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end());
2699 N = NodeStack.back();
2701 if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size())
2702 MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
2703 MatcherIndex = LastScope.FailIndex;
2705 DEBUG(errs() << " Continuing at " << MatcherIndex << "\n");
2707 InputChain = LastScope.InputChain;
2708 InputGlue = LastScope.InputGlue;
2709 if (!LastScope.HasChainNodesMatched)
2710 ChainNodesMatched.clear();
2711 if (!LastScope.HasGlueResultNodesMatched)
2712 GlueResultNodesMatched.clear();
2714 // Check to see what the offset is at the new MatcherIndex. If it is zero
2715 // we have reached the end of this scope, otherwise we have another child
2716 // in the current scope to try.
2717 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2718 if (NumToSkip & 128)
2719 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2721 // If we have another child in this scope to match, update FailIndex and
2723 if (NumToSkip != 0) {
2724 LastScope.FailIndex = MatcherIndex+NumToSkip;
2728 // End of this scope, pop it and try the next child in the containing
2730 MatchScopes.pop_back();
2737 void SelectionDAGISel::CannotYetSelect(SDNode *N) {
2739 raw_string_ostream Msg(msg);
2740 Msg << "Cannot select: ";
2742 if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN &&
2743 N->getOpcode() != ISD::INTRINSIC_WO_CHAIN &&
2744 N->getOpcode() != ISD::INTRINSIC_VOID) {
2745 N->printrFull(Msg, CurDAG);
2747 bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other;
2749 cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue();
2750 if (iid < Intrinsic::num_intrinsics)
2751 Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid);
2752 else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo())
2753 Msg << "target intrinsic %" << TII->getName(iid);
2755 Msg << "unknown intrinsic #" << iid;
2757 report_fatal_error(Msg.str());
2760 char SelectionDAGISel::ID = 0;