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 "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/CallingConv.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Function.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/InlineAsm.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/FastISel.h"
32 #include "llvm/CodeGen/GCStrategy.h"
33 #include "llvm/CodeGen/GCMetadata.h"
34 #include "llvm/CodeGen/MachineFunction.h"
35 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
36 #include "llvm/CodeGen/MachineFrameInfo.h"
37 #include "llvm/CodeGen/MachineInstrBuilder.h"
38 #include "llvm/CodeGen/MachineJumpTableInfo.h"
39 #include "llvm/CodeGen/MachineModuleInfo.h"
40 #include "llvm/CodeGen/MachineRegisterInfo.h"
41 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
42 #include "llvm/CodeGen/SchedulerRegistry.h"
43 #include "llvm/CodeGen/SelectionDAG.h"
44 #include "llvm/CodeGen/DwarfWriter.h"
45 #include "llvm/Target/TargetRegisterInfo.h"
46 #include "llvm/Target/TargetData.h"
47 #include "llvm/Target/TargetFrameInfo.h"
48 #include "llvm/Target/TargetIntrinsicInfo.h"
49 #include "llvm/Target/TargetInstrInfo.h"
50 #include "llvm/Target/TargetLowering.h"
51 #include "llvm/Target/TargetMachine.h"
52 #include "llvm/Target/TargetOptions.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/ErrorHandling.h"
56 #include "llvm/Support/MathExtras.h"
57 #include "llvm/Support/Timer.h"
58 #include "llvm/Support/raw_ostream.h"
59 #include "llvm/ADT/Statistic.h"
63 STATISTIC(NumFastIselFailures, "Number of instructions fast isel failed on");
66 EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
67 cl::desc("Enable verbose messages in the \"fast\" "
68 "instruction selector"));
70 EnableFastISelAbort("fast-isel-abort", cl::Hidden,
71 cl::desc("Enable abort calls when \"fast\" instruction fails"));
73 SchedLiveInCopies("schedule-livein-copies", cl::Hidden,
74 cl::desc("Schedule copies of livein registers"),
79 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
80 cl::desc("Pop up a window to show dags before the first "
83 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
84 cl::desc("Pop up a window to show dags before legalize types"));
86 ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
87 cl::desc("Pop up a window to show dags before legalize"));
89 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
90 cl::desc("Pop up a window to show dags before the second "
93 ViewDAGCombineLT("view-dag-combine-lt-dags", cl::Hidden,
94 cl::desc("Pop up a window to show dags before the post legalize types"
95 " dag combine pass"));
97 ViewISelDAGs("view-isel-dags", cl::Hidden,
98 cl::desc("Pop up a window to show isel dags as they are selected"));
100 ViewSchedDAGs("view-sched-dags", cl::Hidden,
101 cl::desc("Pop up a window to show sched dags as they are processed"));
103 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
104 cl::desc("Pop up a window to show SUnit dags after they are processed"));
106 static const bool ViewDAGCombine1 = false,
107 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false,
108 ViewDAGCombine2 = false,
109 ViewDAGCombineLT = false,
110 ViewISelDAGs = false, ViewSchedDAGs = false,
111 ViewSUnitDAGs = false;
114 //===---------------------------------------------------------------------===//
116 /// RegisterScheduler class - Track the registration of instruction schedulers.
118 //===---------------------------------------------------------------------===//
119 MachinePassRegistry RegisterScheduler::Registry;
121 //===---------------------------------------------------------------------===//
123 /// ISHeuristic command line option for instruction schedulers.
125 //===---------------------------------------------------------------------===//
126 static cl::opt<RegisterScheduler::FunctionPassCtor, false,
127 RegisterPassParser<RegisterScheduler> >
128 ISHeuristic("pre-RA-sched",
129 cl::init(&createDefaultScheduler),
130 cl::desc("Instruction schedulers available (before register"
133 static RegisterScheduler
134 defaultListDAGScheduler("default", "Best scheduler for the target",
135 createDefaultScheduler);
138 //===--------------------------------------------------------------------===//
139 /// createDefaultScheduler - This creates an instruction scheduler appropriate
141 ScheduleDAGSDNodes* createDefaultScheduler(SelectionDAGISel *IS,
142 CodeGenOpt::Level OptLevel) {
143 const TargetLowering &TLI = IS->getTargetLowering();
145 if (OptLevel == CodeGenOpt::None)
146 return createFastDAGScheduler(IS, OptLevel);
147 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency)
148 return createTDListDAGScheduler(IS, OptLevel);
149 assert(TLI.getSchedulingPreference() ==
150 TargetLowering::SchedulingForRegPressure && "Unknown sched type!");
151 return createBURRListDAGScheduler(IS, OptLevel);
155 // EmitInstrWithCustomInserter - This method should be implemented by targets
156 // that mark instructions with the 'usesCustomInserter' flag. These
157 // instructions are special in various ways, which require special support to
158 // insert. The specified MachineInstr is created but not inserted into any
159 // basic blocks, and this method is called to expand it into a sequence of
160 // instructions, potentially also creating new basic blocks and control flow.
161 // When new basic blocks are inserted and the edges from MBB to its successors
162 // are modified, the method should insert pairs of <OldSucc, NewSucc> into the
164 MachineBasicBlock *TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
165 MachineBasicBlock *MBB,
166 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
168 dbgs() << "If a target marks an instruction with "
169 "'usesCustomInserter', it must implement "
170 "TargetLowering::EmitInstrWithCustomInserter!";
176 /// EmitLiveInCopy - Emit a copy for a live in physical register. If the
177 /// physical register has only a single copy use, then coalesced the copy
179 static void EmitLiveInCopy(MachineBasicBlock *MBB,
180 MachineBasicBlock::iterator &InsertPos,
181 unsigned VirtReg, unsigned PhysReg,
182 const TargetRegisterClass *RC,
183 DenseMap<MachineInstr*, unsigned> &CopyRegMap,
184 const MachineRegisterInfo &MRI,
185 const TargetRegisterInfo &TRI,
186 const TargetInstrInfo &TII) {
187 unsigned NumUses = 0;
188 MachineInstr *UseMI = NULL;
189 for (MachineRegisterInfo::use_iterator UI = MRI.use_begin(VirtReg),
190 UE = MRI.use_end(); UI != UE; ++UI) {
196 // If the number of uses is not one, or the use is not a move instruction,
197 // don't coalesce. Also, only coalesce away a virtual register to virtual
199 bool Coalesced = false;
200 unsigned SrcReg, DstReg, SrcSubReg, DstSubReg;
202 TII.isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubReg, DstSubReg) &&
203 TargetRegisterInfo::isVirtualRegister(DstReg)) {
208 // Now find an ideal location to insert the copy.
209 MachineBasicBlock::iterator Pos = InsertPos;
210 while (Pos != MBB->begin()) {
211 MachineInstr *PrevMI = prior(Pos);
212 DenseMap<MachineInstr*, unsigned>::iterator RI = CopyRegMap.find(PrevMI);
213 // copyRegToReg might emit multiple instructions to do a copy.
214 unsigned CopyDstReg = (RI == CopyRegMap.end()) ? 0 : RI->second;
215 if (CopyDstReg && !TRI.regsOverlap(CopyDstReg, PhysReg))
216 // This is what the BB looks like right now:
221 // We want to insert "r1025 = mov r1". Inserting this copy below the
222 // move to r1024 makes it impossible for that move to be coalesced.
229 break; // Woot! Found a good location.
233 bool Emitted = TII.copyRegToReg(*MBB, Pos, VirtReg, PhysReg, RC, RC);
234 assert(Emitted && "Unable to issue a live-in copy instruction!\n");
237 CopyRegMap.insert(std::make_pair(prior(Pos), VirtReg));
239 if (&*InsertPos == UseMI) ++InsertPos;
244 /// EmitLiveInCopies - If this is the first basic block in the function,
245 /// and if it has live ins that need to be copied into vregs, emit the
246 /// copies into the block.
247 static void EmitLiveInCopies(MachineBasicBlock *EntryMBB,
248 const MachineRegisterInfo &MRI,
249 const TargetRegisterInfo &TRI,
250 const TargetInstrInfo &TII) {
251 if (SchedLiveInCopies) {
252 // Emit the copies at a heuristically-determined location in the block.
253 DenseMap<MachineInstr*, unsigned> CopyRegMap;
254 MachineBasicBlock::iterator InsertPos = EntryMBB->begin();
255 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
256 E = MRI.livein_end(); LI != E; ++LI)
258 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
259 EmitLiveInCopy(EntryMBB, InsertPos, LI->second, LI->first,
260 RC, CopyRegMap, MRI, TRI, TII);
263 // Emit the copies into the top of the block.
264 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
265 E = MRI.livein_end(); LI != E; ++LI)
267 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
268 bool Emitted = TII.copyRegToReg(*EntryMBB, EntryMBB->begin(),
269 LI->second, LI->first, RC, RC);
270 assert(Emitted && "Unable to issue a live-in copy instruction!\n");
276 //===----------------------------------------------------------------------===//
277 // SelectionDAGISel code
278 //===----------------------------------------------------------------------===//
280 SelectionDAGISel::SelectionDAGISel(TargetMachine &tm, CodeGenOpt::Level OL) :
281 MachineFunctionPass(&ID), TM(tm), TLI(*tm.getTargetLowering()),
282 FuncInfo(new FunctionLoweringInfo(TLI)),
283 CurDAG(new SelectionDAG(TLI, *FuncInfo)),
284 SDB(new SelectionDAGBuilder(*CurDAG, TLI, *FuncInfo, OL)),
290 SelectionDAGISel::~SelectionDAGISel() {
296 unsigned SelectionDAGISel::MakeReg(EVT VT) {
297 return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT));
300 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
301 AU.addRequired<AliasAnalysis>();
302 AU.addPreserved<AliasAnalysis>();
303 AU.addRequired<GCModuleInfo>();
304 AU.addPreserved<GCModuleInfo>();
305 AU.addRequired<DwarfWriter>();
306 AU.addPreserved<DwarfWriter>();
307 MachineFunctionPass::getAnalysisUsage(AU);
310 bool SelectionDAGISel::runOnMachineFunction(MachineFunction &mf) {
311 Function &Fn = *mf.getFunction();
313 // Do some sanity-checking on the command-line options.
314 assert((!EnableFastISelVerbose || EnableFastISel) &&
315 "-fast-isel-verbose requires -fast-isel");
316 assert((!EnableFastISelAbort || EnableFastISel) &&
317 "-fast-isel-abort requires -fast-isel");
319 // Get alias analysis for load/store combining.
320 AA = &getAnalysis<AliasAnalysis>();
323 const TargetInstrInfo &TII = *TM.getInstrInfo();
324 const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
327 GFI = &getAnalysis<GCModuleInfo>().getFunctionInfo(Fn);
330 RegInfo = &MF->getRegInfo();
331 DEBUG(dbgs() << "\n\n\n=== " << Fn.getName() << "\n");
333 MachineModuleInfo *MMI = getAnalysisIfAvailable<MachineModuleInfo>();
334 DwarfWriter *DW = getAnalysisIfAvailable<DwarfWriter>();
335 CurDAG->init(*MF, MMI, DW);
336 FuncInfo->set(Fn, *MF, EnableFastISel);
339 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
340 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(I->getTerminator()))
342 FuncInfo->MBBMap[Invoke->getSuccessor(1)]->setIsLandingPad();
344 SelectAllBasicBlocks(Fn, *MF, MMI, DW, TII);
346 // If the first basic block in the function has live ins that need to be
347 // copied into vregs, emit the copies into the top of the block before
348 // emitting the code for the block.
349 EmitLiveInCopies(MF->begin(), *RegInfo, TRI, TII);
351 // Add function live-ins to entry block live-in set.
352 for (MachineRegisterInfo::livein_iterator I = RegInfo->livein_begin(),
353 E = RegInfo->livein_end(); I != E; ++I)
354 MF->begin()->addLiveIn(I->first);
357 assert(FuncInfo->CatchInfoFound.size() == FuncInfo->CatchInfoLost.size() &&
358 "Not all catch info was assigned to a landing pad!");
366 /// SetDebugLoc - Update MF's and SDB's DebugLocs if debug information is
367 /// attached with this instruction.
368 static void SetDebugLoc(unsigned MDDbgKind, Instruction *I,
369 SelectionDAGBuilder *SDB,
370 FastISel *FastIS, MachineFunction *MF) {
371 if (isa<DbgInfoIntrinsic>(I)) return;
373 if (MDNode *Dbg = I->getMetadata(MDDbgKind)) {
374 DILocation DILoc(Dbg);
375 DebugLoc Loc = ExtractDebugLocation(DILoc, MF->getDebugLocInfo());
377 SDB->setCurDebugLoc(Loc);
380 FastIS->setCurDebugLoc(Loc);
382 // If the function doesn't have a default debug location yet, set
383 // it. This is kind of a hack.
384 if (MF->getDefaultDebugLoc().isUnknown())
385 MF->setDefaultDebugLoc(Loc);
389 /// ResetDebugLoc - Set MF's and SDB's DebugLocs to Unknown.
390 static void ResetDebugLoc(SelectionDAGBuilder *SDB, FastISel *FastIS) {
391 SDB->setCurDebugLoc(DebugLoc::getUnknownLoc());
393 FastIS->setCurDebugLoc(DebugLoc::getUnknownLoc());
396 void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB,
397 BasicBlock::iterator Begin,
398 BasicBlock::iterator End,
400 SDB->setCurrentBasicBlock(BB);
401 unsigned MDDbgKind = LLVMBB->getContext().getMDKindID("dbg");
403 // Lower all of the non-terminator instructions. If a call is emitted
404 // as a tail call, cease emitting nodes for this block.
405 for (BasicBlock::iterator I = Begin; I != End && !SDB->HasTailCall; ++I) {
406 SetDebugLoc(MDDbgKind, I, SDB, 0, MF);
408 if (!isa<TerminatorInst>(I)) {
411 // Set the current debug location back to "unknown" so that it doesn't
412 // spuriously apply to subsequent instructions.
413 ResetDebugLoc(SDB, 0);
417 if (!SDB->HasTailCall) {
418 // Ensure that all instructions which are used outside of their defining
419 // blocks are available as virtual registers. Invoke is handled elsewhere.
420 for (BasicBlock::iterator I = Begin; I != End; ++I)
421 if (!isa<PHINode>(I) && !isa<InvokeInst>(I))
422 SDB->CopyToExportRegsIfNeeded(I);
424 // Handle PHI nodes in successor blocks.
425 if (End == LLVMBB->end()) {
426 HandlePHINodesInSuccessorBlocks(LLVMBB);
428 // Lower the terminator after the copies are emitted.
429 SetDebugLoc(MDDbgKind, LLVMBB->getTerminator(), SDB, 0, MF);
430 SDB->visit(*LLVMBB->getTerminator());
431 ResetDebugLoc(SDB, 0);
435 // Make sure the root of the DAG is up-to-date.
436 CurDAG->setRoot(SDB->getControlRoot());
438 // Final step, emit the lowered DAG as machine code.
440 HadTailCall = SDB->HasTailCall;
445 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
446 /// nodes from the worklist.
447 class SDOPsWorkListRemover : public SelectionDAG::DAGUpdateListener {
448 SmallVector<SDNode*, 128> &Worklist;
449 SmallPtrSet<SDNode*, 128> &InWorklist;
451 SDOPsWorkListRemover(SmallVector<SDNode*, 128> &wl,
452 SmallPtrSet<SDNode*, 128> &inwl)
453 : Worklist(wl), InWorklist(inwl) {}
455 void RemoveFromWorklist(SDNode *N) {
456 if (!InWorklist.erase(N)) return;
458 SmallVector<SDNode*, 128>::iterator I =
459 std::find(Worklist.begin(), Worklist.end(), N);
460 assert(I != Worklist.end() && "Not in worklist");
462 *I = Worklist.back();
466 virtual void NodeDeleted(SDNode *N, SDNode *E) {
467 RemoveFromWorklist(N);
470 virtual void NodeUpdated(SDNode *N) {
476 /// TrivialTruncElim - Eliminate some trivial nops that can result from
477 /// ShrinkDemandedOps: (trunc (ext n)) -> n.
478 static bool TrivialTruncElim(SDValue Op,
479 TargetLowering::TargetLoweringOpt &TLO) {
480 SDValue N0 = Op.getOperand(0);
481 EVT VT = Op.getValueType();
482 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
483 N0.getOpcode() == ISD::SIGN_EXTEND ||
484 N0.getOpcode() == ISD::ANY_EXTEND) &&
485 N0.getOperand(0).getValueType() == VT) {
486 return TLO.CombineTo(Op, N0.getOperand(0));
491 /// ShrinkDemandedOps - A late transformation pass that shrink expressions
492 /// using TargetLowering::TargetLoweringOpt::ShrinkDemandedOp. It converts
493 /// x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free.
494 void SelectionDAGISel::ShrinkDemandedOps() {
495 SmallVector<SDNode*, 128> Worklist;
496 SmallPtrSet<SDNode*, 128> InWorklist;
498 // Add all the dag nodes to the worklist.
499 Worklist.reserve(CurDAG->allnodes_size());
500 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
501 E = CurDAG->allnodes_end(); I != E; ++I) {
502 Worklist.push_back(I);
503 InWorklist.insert(I);
506 TargetLowering::TargetLoweringOpt TLO(*CurDAG, true);
507 while (!Worklist.empty()) {
508 SDNode *N = Worklist.pop_back_val();
511 if (N->use_empty() && N != CurDAG->getRoot().getNode()) {
512 // Deleting this node may make its operands dead, add them to the worklist
513 // if they aren't already there.
514 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
515 if (InWorklist.insert(N->getOperand(i).getNode()))
516 Worklist.push_back(N->getOperand(i).getNode());
518 CurDAG->DeleteNode(N);
522 // Run ShrinkDemandedOp on scalar binary operations.
523 if (N->getNumValues() != 1 ||
524 !N->getValueType(0).isSimple() || !N->getValueType(0).isInteger())
527 unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
528 APInt Demanded = APInt::getAllOnesValue(BitWidth);
529 APInt KnownZero, KnownOne;
530 if (!TLI.SimplifyDemandedBits(SDValue(N, 0), Demanded,
531 KnownZero, KnownOne, TLO) &&
532 (N->getOpcode() != ISD::TRUNCATE ||
533 !TrivialTruncElim(SDValue(N, 0), TLO)))
537 assert(!InWorklist.count(N) && "Already in worklist");
538 Worklist.push_back(N);
539 InWorklist.insert(N);
541 // Replace the old value with the new one.
542 DEBUG(errs() << "\nShrinkDemandedOps replacing ";
543 TLO.Old.getNode()->dump(CurDAG);
544 errs() << "\nWith: ";
545 TLO.New.getNode()->dump(CurDAG);
548 if (InWorklist.insert(TLO.New.getNode()))
549 Worklist.push_back(TLO.New.getNode());
551 SDOPsWorkListRemover DeadNodes(Worklist, InWorklist);
552 CurDAG->ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, &DeadNodes);
554 if (!TLO.Old.getNode()->use_empty()) continue;
556 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands();
558 SDNode *OpNode = TLO.Old.getNode()->getOperand(i).getNode();
559 if (OpNode->hasOneUse()) {
560 // Add OpNode to the end of the list to revisit.
561 DeadNodes.RemoveFromWorklist(OpNode);
562 Worklist.push_back(OpNode);
563 InWorklist.insert(OpNode);
567 DeadNodes.RemoveFromWorklist(TLO.Old.getNode());
568 CurDAG->DeleteNode(TLO.Old.getNode());
572 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
573 SmallPtrSet<SDNode*, 128> VisitedNodes;
574 SmallVector<SDNode*, 128> Worklist;
576 Worklist.push_back(CurDAG->getRoot().getNode());
583 SDNode *N = Worklist.pop_back_val();
585 // If we've already seen this node, ignore it.
586 if (!VisitedNodes.insert(N))
589 // Otherwise, add all chain operands to the worklist.
590 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
591 if (N->getOperand(i).getValueType() == MVT::Other)
592 Worklist.push_back(N->getOperand(i).getNode());
594 // If this is a CopyToReg with a vreg dest, process it.
595 if (N->getOpcode() != ISD::CopyToReg)
598 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
599 if (!TargetRegisterInfo::isVirtualRegister(DestReg))
602 // Ignore non-scalar or non-integer values.
603 SDValue Src = N->getOperand(2);
604 EVT SrcVT = Src.getValueType();
605 if (!SrcVT.isInteger() || SrcVT.isVector())
608 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
609 Mask = APInt::getAllOnesValue(SrcVT.getSizeInBits());
610 CurDAG->ComputeMaskedBits(Src, Mask, KnownZero, KnownOne);
612 // Only install this information if it tells us something.
613 if (NumSignBits != 1 || KnownZero != 0 || KnownOne != 0) {
614 DestReg -= TargetRegisterInfo::FirstVirtualRegister;
615 if (DestReg >= FuncInfo->LiveOutRegInfo.size())
616 FuncInfo->LiveOutRegInfo.resize(DestReg+1);
617 FunctionLoweringInfo::LiveOutInfo &LOI =
618 FuncInfo->LiveOutRegInfo[DestReg];
619 LOI.NumSignBits = NumSignBits;
620 LOI.KnownOne = KnownOne;
621 LOI.KnownZero = KnownZero;
623 } while (!Worklist.empty());
626 void SelectionDAGISel::CodeGenAndEmitDAG() {
627 std::string GroupName;
628 if (TimePassesIsEnabled)
629 GroupName = "Instruction Selection and Scheduling";
630 std::string BlockName;
631 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs ||
632 ViewDAGCombine2 || ViewDAGCombineLT || ViewISelDAGs || ViewSchedDAGs ||
634 BlockName = MF->getFunction()->getNameStr() + ":" +
635 BB->getBasicBlock()->getNameStr();
637 DEBUG(dbgs() << "Initial selection DAG:\n");
638 DEBUG(CurDAG->dump());
640 if (ViewDAGCombine1) CurDAG->viewGraph("dag-combine1 input for " + BlockName);
642 // Run the DAG combiner in pre-legalize mode.
643 if (TimePassesIsEnabled) {
644 NamedRegionTimer T("DAG Combining 1", GroupName);
645 CurDAG->Combine(Unrestricted, *AA, OptLevel);
647 CurDAG->Combine(Unrestricted, *AA, OptLevel);
650 DEBUG(dbgs() << "Optimized lowered selection DAG:\n");
651 DEBUG(CurDAG->dump());
653 // Second step, hack on the DAG until it only uses operations and types that
654 // the target supports.
655 if (ViewLegalizeTypesDAGs) CurDAG->viewGraph("legalize-types input for " +
659 if (TimePassesIsEnabled) {
660 NamedRegionTimer T("Type Legalization", GroupName);
661 Changed = CurDAG->LegalizeTypes();
663 Changed = CurDAG->LegalizeTypes();
666 DEBUG(dbgs() << "Type-legalized selection DAG:\n");
667 DEBUG(CurDAG->dump());
670 if (ViewDAGCombineLT)
671 CurDAG->viewGraph("dag-combine-lt input for " + BlockName);
673 // Run the DAG combiner in post-type-legalize mode.
674 if (TimePassesIsEnabled) {
675 NamedRegionTimer T("DAG Combining after legalize types", GroupName);
676 CurDAG->Combine(NoIllegalTypes, *AA, OptLevel);
678 CurDAG->Combine(NoIllegalTypes, *AA, OptLevel);
681 DEBUG(dbgs() << "Optimized type-legalized selection DAG:\n");
682 DEBUG(CurDAG->dump());
685 if (TimePassesIsEnabled) {
686 NamedRegionTimer T("Vector Legalization", GroupName);
687 Changed = CurDAG->LegalizeVectors();
689 Changed = CurDAG->LegalizeVectors();
693 if (TimePassesIsEnabled) {
694 NamedRegionTimer T("Type Legalization 2", GroupName);
695 CurDAG->LegalizeTypes();
697 CurDAG->LegalizeTypes();
700 if (ViewDAGCombineLT)
701 CurDAG->viewGraph("dag-combine-lv input for " + BlockName);
703 // Run the DAG combiner in post-type-legalize mode.
704 if (TimePassesIsEnabled) {
705 NamedRegionTimer T("DAG Combining after legalize vectors", GroupName);
706 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
708 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
711 DEBUG(dbgs() << "Optimized vector-legalized selection DAG:\n");
712 DEBUG(CurDAG->dump());
715 if (ViewLegalizeDAGs) CurDAG->viewGraph("legalize input for " + BlockName);
717 if (TimePassesIsEnabled) {
718 NamedRegionTimer T("DAG Legalization", GroupName);
719 CurDAG->Legalize(OptLevel);
721 CurDAG->Legalize(OptLevel);
724 DEBUG(dbgs() << "Legalized selection DAG:\n");
725 DEBUG(CurDAG->dump());
727 if (ViewDAGCombine2) CurDAG->viewGraph("dag-combine2 input for " + BlockName);
729 // Run the DAG combiner in post-legalize mode.
730 if (TimePassesIsEnabled) {
731 NamedRegionTimer T("DAG Combining 2", GroupName);
732 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
734 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
737 DEBUG(dbgs() << "Optimized legalized selection DAG:\n");
738 DEBUG(CurDAG->dump());
740 if (OptLevel != CodeGenOpt::None) {
742 ComputeLiveOutVRegInfo();
745 if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName);
747 // Third, instruction select all of the operations to machine code, adding the
748 // code to the MachineBasicBlock.
749 if (TimePassesIsEnabled) {
750 NamedRegionTimer T("Instruction Selection", GroupName);
751 DoInstructionSelection();
753 DoInstructionSelection();
756 DEBUG(dbgs() << "Selected selection DAG:\n");
757 DEBUG(CurDAG->dump());
759 if (ViewSchedDAGs) CurDAG->viewGraph("scheduler input for " + BlockName);
761 // Schedule machine code.
762 ScheduleDAGSDNodes *Scheduler = CreateScheduler();
763 if (TimePassesIsEnabled) {
764 NamedRegionTimer T("Instruction Scheduling", GroupName);
765 Scheduler->Run(CurDAG, BB, BB->end());
767 Scheduler->Run(CurDAG, BB, BB->end());
770 if (ViewSUnitDAGs) Scheduler->viewGraph();
772 // Emit machine code to BB. This can change 'BB' to the last block being
774 if (TimePassesIsEnabled) {
775 NamedRegionTimer T("Instruction Creation", GroupName);
776 BB = Scheduler->EmitSchedule(&SDB->EdgeMapping);
778 BB = Scheduler->EmitSchedule(&SDB->EdgeMapping);
781 // Free the scheduler state.
782 if (TimePassesIsEnabled) {
783 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName);
789 DEBUG(dbgs() << "Selected machine code:\n");
793 void SelectionDAGISel::DoInstructionSelection() {
794 DEBUG(errs() << "===== Instruction selection begins:\n");
798 // Select target instructions for the DAG.
800 // Number all nodes with a topological order and set DAGSize.
801 DAGSize = CurDAG->AssignTopologicalOrder();
803 // Create a dummy node (which is not added to allnodes), that adds
804 // a reference to the root node, preventing it from being deleted,
805 // and tracking any changes of the root.
806 HandleSDNode Dummy(CurDAG->getRoot());
807 ISelPosition = SelectionDAG::allnodes_iterator(CurDAG->getRoot().getNode());
810 // The AllNodes list is now topological-sorted. Visit the
811 // nodes by starting at the end of the list (the root of the
812 // graph) and preceding back toward the beginning (the entry
814 while (ISelPosition != CurDAG->allnodes_begin()) {
815 SDNode *Node = --ISelPosition;
816 // Skip dead nodes. DAGCombiner is expected to eliminate all dead nodes,
817 // but there are currently some corner cases that it misses. Also, this
818 // makes it theoretically possible to disable the DAGCombiner.
819 if (Node->use_empty())
822 SDNode *ResNode = Select(Node);
824 // FIXME: This is pretty gross. 'Select' should be changed to not return
825 // anything at all and this code should be nuked with a tactical strike.
827 // If node should not be replaced, continue with the next one.
828 if (ResNode == Node || Node->getOpcode() == ISD::DELETED_NODE)
832 ReplaceUses(Node, ResNode);
834 // If after the replacement this node is not used any more,
835 // remove this dead node.
836 if (Node->use_empty()) { // Don't delete EntryToken, etc.
837 ISelUpdater ISU(ISelPosition);
838 CurDAG->RemoveDeadNode(Node, &ISU);
842 CurDAG->setRoot(Dummy.getValue());
844 DEBUG(errs() << "===== Instruction selection ends:\n");
846 PostprocessISelDAG();
848 // FIXME: This shouldn't be needed, remove it.
849 CurDAG->RemoveDeadNodes();
853 void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn,
855 MachineModuleInfo *MMI,
857 const TargetInstrInfo &TII) {
858 // Initialize the Fast-ISel state, if needed.
859 FastISel *FastIS = 0;
861 FastIS = TLI.createFastISel(MF, MMI, DW,
864 FuncInfo->StaticAllocaMap
866 , FuncInfo->CatchInfoLost
870 unsigned MDDbgKind = Fn.getContext().getMDKindID("dbg");
872 // Iterate over all basic blocks in the function.
873 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
874 BasicBlock *LLVMBB = &*I;
875 BB = FuncInfo->MBBMap[LLVMBB];
877 BasicBlock::iterator const Begin = LLVMBB->begin();
878 BasicBlock::iterator const End = LLVMBB->end();
879 BasicBlock::iterator BI = Begin;
881 // Lower any arguments needed in this block if this is the entry block.
882 bool SuppressFastISel = false;
883 if (LLVMBB == &Fn.getEntryBlock()) {
884 LowerArguments(LLVMBB);
886 // If any of the arguments has the byval attribute, forgo
887 // fast-isel in the entry block.
890 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end();
892 if (Fn.paramHasAttr(j, Attribute::ByVal)) {
893 if (EnableFastISelVerbose || EnableFastISelAbort)
894 dbgs() << "FastISel skips entry block due to byval argument\n";
895 SuppressFastISel = true;
901 if (MMI && BB->isLandingPad()) {
902 // Add a label to mark the beginning of the landing pad. Deletion of the
903 // landing pad can thus be detected via the MachineModuleInfo.
904 MCSymbol *Label = MMI->addLandingPad(BB);
906 const TargetInstrDesc &II = TII.get(TargetOpcode::EH_LABEL);
907 BuildMI(BB, SDB->getCurDebugLoc(), II).addSym(Label);
909 // Mark exception register as live in.
910 unsigned Reg = TLI.getExceptionAddressRegister();
911 if (Reg) BB->addLiveIn(Reg);
913 // Mark exception selector register as live in.
914 Reg = TLI.getExceptionSelectorRegister();
915 if (Reg) BB->addLiveIn(Reg);
917 // FIXME: Hack around an exception handling flaw (PR1508): the personality
918 // function and list of typeids logically belong to the invoke (or, if you
919 // like, the basic block containing the invoke), and need to be associated
920 // with it in the dwarf exception handling tables. Currently however the
921 // information is provided by an intrinsic (eh.selector) that can be moved
922 // to unexpected places by the optimizers: if the unwind edge is critical,
923 // then breaking it can result in the intrinsics being in the successor of
924 // the landing pad, not the landing pad itself. This results
925 // in exceptions not being caught because no typeids are associated with
926 // the invoke. This may not be the only way things can go wrong, but it
927 // is the only way we try to work around for the moment.
928 BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
930 if (Br && Br->isUnconditional()) { // Critical edge?
931 BasicBlock::iterator I, E;
932 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
933 if (isa<EHSelectorInst>(I))
937 // No catch info found - try to extract some from the successor.
938 CopyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, *FuncInfo);
942 // Before doing SelectionDAG ISel, see if FastISel has been requested.
943 if (FastIS && !SuppressFastISel) {
944 // Emit code for any incoming arguments. This must happen before
945 // beginning FastISel on the entry block.
946 if (LLVMBB == &Fn.getEntryBlock()) {
947 CurDAG->setRoot(SDB->getControlRoot());
951 FastIS->startNewBlock(BB);
952 // Do FastISel on as many instructions as possible.
953 for (; BI != End; ++BI) {
954 // Just before the terminator instruction, insert instructions to
955 // feed PHI nodes in successor blocks.
956 if (isa<TerminatorInst>(BI))
957 if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, FastIS)) {
958 ++NumFastIselFailures;
959 ResetDebugLoc(SDB, FastIS);
960 if (EnableFastISelVerbose || EnableFastISelAbort) {
961 dbgs() << "FastISel miss: ";
964 assert(!EnableFastISelAbort &&
965 "FastISel didn't handle a PHI in a successor");
969 SetDebugLoc(MDDbgKind, BI, SDB, FastIS, &MF);
971 // Try to select the instruction with FastISel.
972 if (FastIS->SelectInstruction(BI)) {
973 ResetDebugLoc(SDB, FastIS);
977 // Clear out the debug location so that it doesn't carry over to
978 // unrelated instructions.
979 ResetDebugLoc(SDB, FastIS);
981 // Then handle certain instructions as single-LLVM-Instruction blocks.
982 if (isa<CallInst>(BI)) {
983 ++NumFastIselFailures;
984 if (EnableFastISelVerbose || EnableFastISelAbort) {
985 dbgs() << "FastISel missed call: ";
989 if (!BI->getType()->isVoidTy()) {
990 unsigned &R = FuncInfo->ValueMap[BI];
992 R = FuncInfo->CreateRegForValue(BI);
995 bool HadTailCall = false;
996 SelectBasicBlock(LLVMBB, BI, llvm::next(BI), HadTailCall);
998 // If the call was emitted as a tail call, we're done with the block.
1004 // If the instruction was codegen'd with multiple blocks,
1005 // inform the FastISel object where to resume inserting.
1006 FastIS->setCurrentBlock(BB);
1010 // Otherwise, give up on FastISel for the rest of the block.
1011 // For now, be a little lenient about non-branch terminators.
1012 if (!isa<TerminatorInst>(BI) || isa<BranchInst>(BI)) {
1013 ++NumFastIselFailures;
1014 if (EnableFastISelVerbose || EnableFastISelAbort) {
1015 dbgs() << "FastISel miss: ";
1018 if (EnableFastISelAbort)
1019 // The "fast" selector couldn't handle something and bailed.
1020 // For the purpose of debugging, just abort.
1021 llvm_unreachable("FastISel didn't select the entire block");
1027 // Run SelectionDAG instruction selection on the remainder of the block
1028 // not handled by FastISel. If FastISel is not run, this is the entire
1032 SelectBasicBlock(LLVMBB, BI, End, HadTailCall);
1042 SelectionDAGISel::FinishBasicBlock() {
1044 DEBUG(dbgs() << "Target-post-processed machine code:\n");
1047 DEBUG(dbgs() << "Total amount of phi nodes to update: "
1048 << SDB->PHINodesToUpdate.size() << "\n");
1049 DEBUG(for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i)
1050 dbgs() << "Node " << i << " : ("
1051 << SDB->PHINodesToUpdate[i].first
1052 << ", " << SDB->PHINodesToUpdate[i].second << ")\n");
1054 // Next, now that we know what the last MBB the LLVM BB expanded is, update
1055 // PHI nodes in successors.
1056 if (SDB->SwitchCases.empty() &&
1057 SDB->JTCases.empty() &&
1058 SDB->BitTestCases.empty()) {
1059 for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) {
1060 MachineInstr *PHI = SDB->PHINodesToUpdate[i].first;
1061 assert(PHI->isPHI() &&
1062 "This is not a machine PHI node that we are updating!");
1063 if (!BB->isSuccessor(PHI->getParent()))
1065 PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second,
1067 PHI->addOperand(MachineOperand::CreateMBB(BB));
1069 SDB->PHINodesToUpdate.clear();
1073 for (unsigned i = 0, e = SDB->BitTestCases.size(); i != e; ++i) {
1074 // Lower header first, if it wasn't already lowered
1075 if (!SDB->BitTestCases[i].Emitted) {
1076 // Set the current basic block to the mbb we wish to insert the code into
1077 BB = SDB->BitTestCases[i].Parent;
1078 SDB->setCurrentBasicBlock(BB);
1080 SDB->visitBitTestHeader(SDB->BitTestCases[i]);
1081 CurDAG->setRoot(SDB->getRoot());
1082 CodeGenAndEmitDAG();
1086 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j) {
1087 // Set the current basic block to the mbb we wish to insert the code into
1088 BB = SDB->BitTestCases[i].Cases[j].ThisBB;
1089 SDB->setCurrentBasicBlock(BB);
1092 SDB->visitBitTestCase(SDB->BitTestCases[i].Cases[j+1].ThisBB,
1093 SDB->BitTestCases[i].Reg,
1094 SDB->BitTestCases[i].Cases[j]);
1096 SDB->visitBitTestCase(SDB->BitTestCases[i].Default,
1097 SDB->BitTestCases[i].Reg,
1098 SDB->BitTestCases[i].Cases[j]);
1101 CurDAG->setRoot(SDB->getRoot());
1102 CodeGenAndEmitDAG();
1107 for (unsigned pi = 0, pe = SDB->PHINodesToUpdate.size(); pi != pe; ++pi) {
1108 MachineInstr *PHI = SDB->PHINodesToUpdate[pi].first;
1109 MachineBasicBlock *PHIBB = PHI->getParent();
1110 assert(PHI->isPHI() &&
1111 "This is not a machine PHI node that we are updating!");
1112 // This is "default" BB. We have two jumps to it. From "header" BB and
1113 // from last "case" BB.
1114 if (PHIBB == SDB->BitTestCases[i].Default) {
1115 PHI->addOperand(MachineOperand::
1116 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1117 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Parent));
1118 PHI->addOperand(MachineOperand::
1119 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1120 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Cases.
1123 // One of "cases" BB.
1124 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size();
1126 MachineBasicBlock* cBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1127 if (cBB->isSuccessor(PHIBB)) {
1128 PHI->addOperand(MachineOperand::
1129 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1130 PHI->addOperand(MachineOperand::CreateMBB(cBB));
1135 SDB->BitTestCases.clear();
1137 // If the JumpTable record is filled in, then we need to emit a jump table.
1138 // Updating the PHI nodes is tricky in this case, since we need to determine
1139 // whether the PHI is a successor of the range check MBB or the jump table MBB
1140 for (unsigned i = 0, e = SDB->JTCases.size(); i != e; ++i) {
1141 // Lower header first, if it wasn't already lowered
1142 if (!SDB->JTCases[i].first.Emitted) {
1143 // Set the current basic block to the mbb we wish to insert the code into
1144 BB = SDB->JTCases[i].first.HeaderBB;
1145 SDB->setCurrentBasicBlock(BB);
1147 SDB->visitJumpTableHeader(SDB->JTCases[i].second, SDB->JTCases[i].first);
1148 CurDAG->setRoot(SDB->getRoot());
1149 CodeGenAndEmitDAG();
1153 // Set the current basic block to the mbb we wish to insert the code into
1154 BB = SDB->JTCases[i].second.MBB;
1155 SDB->setCurrentBasicBlock(BB);
1157 SDB->visitJumpTable(SDB->JTCases[i].second);
1158 CurDAG->setRoot(SDB->getRoot());
1159 CodeGenAndEmitDAG();
1163 for (unsigned pi = 0, pe = SDB->PHINodesToUpdate.size(); pi != pe; ++pi) {
1164 MachineInstr *PHI = SDB->PHINodesToUpdate[pi].first;
1165 MachineBasicBlock *PHIBB = PHI->getParent();
1166 assert(PHI->isPHI() &&
1167 "This is not a machine PHI node that we are updating!");
1168 // "default" BB. We can go there only from header BB.
1169 if (PHIBB == SDB->JTCases[i].second.Default) {
1171 (MachineOperand::CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1173 (MachineOperand::CreateMBB(SDB->JTCases[i].first.HeaderBB));
1175 // JT BB. Just iterate over successors here
1176 if (BB->isSuccessor(PHIBB)) {
1178 (MachineOperand::CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1179 PHI->addOperand(MachineOperand::CreateMBB(BB));
1183 SDB->JTCases.clear();
1185 // If the switch block involved a branch to one of the actual successors, we
1186 // need to update PHI nodes in that block.
1187 for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) {
1188 MachineInstr *PHI = SDB->PHINodesToUpdate[i].first;
1189 assert(PHI->isPHI() &&
1190 "This is not a machine PHI node that we are updating!");
1191 if (BB->isSuccessor(PHI->getParent())) {
1192 PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second,
1194 PHI->addOperand(MachineOperand::CreateMBB(BB));
1198 // If we generated any switch lowering information, build and codegen any
1199 // additional DAGs necessary.
1200 for (unsigned i = 0, e = SDB->SwitchCases.size(); i != e; ++i) {
1201 // Set the current basic block to the mbb we wish to insert the code into
1202 MachineBasicBlock *ThisBB = BB = SDB->SwitchCases[i].ThisBB;
1203 SDB->setCurrentBasicBlock(BB);
1206 SDB->visitSwitchCase(SDB->SwitchCases[i]);
1207 CurDAG->setRoot(SDB->getRoot());
1208 CodeGenAndEmitDAG();
1210 // Handle any PHI nodes in successors of this chunk, as if we were coming
1211 // from the original BB before switch expansion. Note that PHI nodes can
1212 // occur multiple times in PHINodesToUpdate. We have to be very careful to
1213 // handle them the right number of times.
1214 while ((BB = SDB->SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
1215 // If new BB's are created during scheduling, the edges may have been
1216 // updated. That is, the edge from ThisBB to BB may have been split and
1217 // BB's predecessor is now another block.
1218 DenseMap<MachineBasicBlock*, MachineBasicBlock*>::iterator EI =
1219 SDB->EdgeMapping.find(BB);
1220 if (EI != SDB->EdgeMapping.end())
1221 ThisBB = EI->second;
1223 // BB may have been removed from the CFG if a branch was constant folded.
1224 if (ThisBB->isSuccessor(BB)) {
1225 for (MachineBasicBlock::iterator Phi = BB->begin();
1226 Phi != BB->end() && Phi->isPHI();
1228 // This value for this PHI node is recorded in PHINodesToUpdate.
1229 for (unsigned pn = 0; ; ++pn) {
1230 assert(pn != SDB->PHINodesToUpdate.size() &&
1231 "Didn't find PHI entry!");
1232 if (SDB->PHINodesToUpdate[pn].first == Phi) {
1233 Phi->addOperand(MachineOperand::
1234 CreateReg(SDB->PHINodesToUpdate[pn].second,
1236 Phi->addOperand(MachineOperand::CreateMBB(ThisBB));
1243 // Don't process RHS if same block as LHS.
1244 if (BB == SDB->SwitchCases[i].FalseBB)
1245 SDB->SwitchCases[i].FalseBB = 0;
1247 // If we haven't handled the RHS, do so now. Otherwise, we're done.
1248 SDB->SwitchCases[i].TrueBB = SDB->SwitchCases[i].FalseBB;
1249 SDB->SwitchCases[i].FalseBB = 0;
1251 assert(SDB->SwitchCases[i].TrueBB == 0 && SDB->SwitchCases[i].FalseBB == 0);
1254 SDB->SwitchCases.clear();
1256 SDB->PHINodesToUpdate.clear();
1260 /// Create the scheduler. If a specific scheduler was specified
1261 /// via the SchedulerRegistry, use it, otherwise select the
1262 /// one preferred by the target.
1264 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
1265 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1269 RegisterScheduler::setDefault(Ctor);
1272 return Ctor(this, OptLevel);
1275 ScheduleHazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
1276 return new ScheduleHazardRecognizer();
1279 //===----------------------------------------------------------------------===//
1280 // Helper functions used by the generated instruction selector.
1281 //===----------------------------------------------------------------------===//
1282 // Calls to these methods are generated by tblgen.
1284 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1285 /// the dag combiner simplified the 255, we still want to match. RHS is the
1286 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1287 /// specified in the .td file (e.g. 255).
1288 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1289 int64_t DesiredMaskS) const {
1290 const APInt &ActualMask = RHS->getAPIntValue();
1291 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1293 // If the actual mask exactly matches, success!
1294 if (ActualMask == DesiredMask)
1297 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1298 if (ActualMask.intersects(~DesiredMask))
1301 // Otherwise, the DAG Combiner may have proven that the value coming in is
1302 // either already zero or is not demanded. Check for known zero input bits.
1303 APInt NeededMask = DesiredMask & ~ActualMask;
1304 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1307 // TODO: check to see if missing bits are just not demanded.
1309 // Otherwise, this pattern doesn't match.
1313 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1314 /// the dag combiner simplified the 255, we still want to match. RHS is the
1315 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1316 /// specified in the .td file (e.g. 255).
1317 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1318 int64_t DesiredMaskS) const {
1319 const APInt &ActualMask = RHS->getAPIntValue();
1320 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1322 // If the actual mask exactly matches, success!
1323 if (ActualMask == DesiredMask)
1326 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1327 if (ActualMask.intersects(~DesiredMask))
1330 // Otherwise, the DAG Combiner may have proven that the value coming in is
1331 // either already zero or is not demanded. Check for known zero input bits.
1332 APInt NeededMask = DesiredMask & ~ActualMask;
1334 APInt KnownZero, KnownOne;
1335 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
1337 // If all the missing bits in the or are already known to be set, match!
1338 if ((NeededMask & KnownOne) == NeededMask)
1341 // TODO: check to see if missing bits are just not demanded.
1343 // Otherwise, this pattern doesn't match.
1348 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1349 /// by tblgen. Others should not call it.
1350 void SelectionDAGISel::
1351 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1352 std::vector<SDValue> InOps;
1353 std::swap(InOps, Ops);
1355 Ops.push_back(InOps[0]); // input chain.
1356 Ops.push_back(InOps[1]); // input asm string.
1358 unsigned i = 2, e = InOps.size();
1359 if (InOps[e-1].getValueType() == MVT::Flag)
1360 --e; // Don't process a flag operand if it is here.
1363 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1364 if ((Flags & 7) != 4 /*MEM*/) {
1365 // Just skip over this operand, copying the operands verbatim.
1366 Ops.insert(Ops.end(), InOps.begin()+i,
1367 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
1368 i += InlineAsm::getNumOperandRegisters(Flags) + 1;
1370 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 &&
1371 "Memory operand with multiple values?");
1372 // Otherwise, this is a memory operand. Ask the target to select it.
1373 std::vector<SDValue> SelOps;
1374 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps)) {
1375 llvm_report_error("Could not match memory address. Inline asm"
1379 // Add this to the output node.
1380 Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size()<< 3),
1382 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1387 // Add the flag input back if present.
1388 if (e != InOps.size())
1389 Ops.push_back(InOps.back());
1392 /// findFlagUse - Return use of EVT::Flag value produced by the specified
1395 static SDNode *findFlagUse(SDNode *N) {
1396 unsigned FlagResNo = N->getNumValues()-1;
1397 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1398 SDUse &Use = I.getUse();
1399 if (Use.getResNo() == FlagResNo)
1400 return Use.getUser();
1405 /// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
1406 /// This function recursively traverses up the operand chain, ignoring
1408 static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
1409 SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited,
1410 bool IgnoreChains) {
1411 // The NodeID's are given uniques ID's where a node ID is guaranteed to be
1412 // greater than all of its (recursive) operands. If we scan to a point where
1413 // 'use' is smaller than the node we're scanning for, then we know we will
1416 // The Use may be -1 (unassigned) if it is a newly allocated node. This can
1417 // happen because we scan down to newly selected nodes in the case of flag
1419 if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1))
1422 // Don't revisit nodes if we already scanned it and didn't fail, we know we
1423 // won't fail if we scan it again.
1424 if (!Visited.insert(Use))
1427 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
1428 // Ignore chain uses, they are validated by HandleMergeInputChains.
1429 if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains)
1432 SDNode *N = Use->getOperand(i).getNode();
1434 if (Use == ImmedUse || Use == Root)
1435 continue; // We are not looking for immediate use.
1440 // Traverse up the operand chain.
1441 if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
1447 /// IsProfitableToFold - Returns true if it's profitable to fold the specific
1448 /// operand node N of U during instruction selection that starts at Root.
1449 bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
1450 SDNode *Root) const {
1451 if (OptLevel == CodeGenOpt::None) return false;
1452 return N.hasOneUse();
1455 /// IsLegalToFold - Returns true if the specific operand node N of
1456 /// U can be folded during instruction selection that starts at Root.
1457 bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
1458 bool IgnoreChains) const {
1459 if (OptLevel == CodeGenOpt::None) return false;
1461 // If Root use can somehow reach N through a path that that doesn't contain
1462 // U then folding N would create a cycle. e.g. In the following
1463 // diagram, Root can reach N through X. If N is folded into into Root, then
1464 // X is both a predecessor and a successor of U.
1475 // * indicates nodes to be folded together.
1477 // If Root produces a flag, then it gets (even more) interesting. Since it
1478 // will be "glued" together with its flag use in the scheduler, we need to
1479 // check if it might reach N.
1498 // If FU (flag use) indirectly reaches N (the load), and Root folds N
1499 // (call it Fold), then X is a predecessor of FU and a successor of
1500 // Fold. But since Fold and FU are flagged together, this will create
1501 // a cycle in the scheduling graph.
1503 // If the node has flags, walk down the graph to the "lowest" node in the
1505 EVT VT = Root->getValueType(Root->getNumValues()-1);
1506 while (VT == MVT::Flag) {
1507 SDNode *FU = findFlagUse(Root);
1511 VT = Root->getValueType(Root->getNumValues()-1);
1513 // If our query node has a flag result with a use, we've walked up it. If
1514 // the user (which has already been selected) has a chain or indirectly uses
1515 // the chain, our WalkChainUsers predicate will not consider it. Because of
1516 // this, we cannot ignore chains in this predicate.
1517 IgnoreChains = false;
1521 SmallPtrSet<SDNode*, 16> Visited;
1522 return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains);
1525 SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
1526 std::vector<SDValue> Ops(N->op_begin(), N->op_end());
1527 SelectInlineAsmMemoryOperands(Ops);
1529 std::vector<EVT> VTs;
1530 VTs.push_back(MVT::Other);
1531 VTs.push_back(MVT::Flag);
1532 SDValue New = CurDAG->getNode(ISD::INLINEASM, N->getDebugLoc(),
1533 VTs, &Ops[0], Ops.size());
1535 return New.getNode();
1538 SDNode *SelectionDAGISel::Select_UNDEF(SDNode *N) {
1539 return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0));
1542 /// GetVBR - decode a vbr encoding whose top bit is set.
1543 ALWAYS_INLINE static uint64_t
1544 GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
1545 assert(Val >= 128 && "Not a VBR");
1546 Val &= 127; // Remove first vbr bit.
1551 NextBits = MatcherTable[Idx++];
1552 Val |= (NextBits&127) << Shift;
1554 } while (NextBits & 128);
1560 /// UpdateChainsAndFlags - When a match is complete, this method updates uses of
1561 /// interior flag and chain results to use the new flag and chain results.
1562 void SelectionDAGISel::
1563 UpdateChainsAndFlags(SDNode *NodeToMatch, SDValue InputChain,
1564 const SmallVectorImpl<SDNode*> &ChainNodesMatched,
1566 const SmallVectorImpl<SDNode*> &FlagResultNodesMatched,
1567 bool isMorphNodeTo) {
1568 SmallVector<SDNode*, 4> NowDeadNodes;
1570 ISelUpdater ISU(ISelPosition);
1572 // Now that all the normal results are replaced, we replace the chain and
1573 // flag results if present.
1574 if (!ChainNodesMatched.empty()) {
1575 assert(InputChain.getNode() != 0 &&
1576 "Matched input chains but didn't produce a chain");
1577 // Loop over all of the nodes we matched that produced a chain result.
1578 // Replace all the chain results with the final chain we ended up with.
1579 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1580 SDNode *ChainNode = ChainNodesMatched[i];
1582 // If this node was already deleted, don't look at it.
1583 if (ChainNode->getOpcode() == ISD::DELETED_NODE)
1586 // Don't replace the results of the root node if we're doing a
1588 if (ChainNode == NodeToMatch && isMorphNodeTo)
1591 SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1);
1592 if (ChainVal.getValueType() == MVT::Flag)
1593 ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
1594 assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
1595 CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain, &ISU);
1597 // If the node became dead, delete it.
1598 if (ChainNode->use_empty())
1599 NowDeadNodes.push_back(ChainNode);
1603 // If the result produces a flag, update any flag results in the matched
1604 // pattern with the flag result.
1605 if (InputFlag.getNode() != 0) {
1606 // Handle any interior nodes explicitly marked.
1607 for (unsigned i = 0, e = FlagResultNodesMatched.size(); i != e; ++i) {
1608 SDNode *FRN = FlagResultNodesMatched[i];
1610 // If this node was already deleted, don't look at it.
1611 if (FRN->getOpcode() == ISD::DELETED_NODE)
1614 assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Flag &&
1615 "Doesn't have a flag result");
1616 CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1),
1619 // If the node became dead, delete it.
1620 if (FRN->use_empty())
1621 NowDeadNodes.push_back(FRN);
1625 if (!NowDeadNodes.empty())
1626 CurDAG->RemoveDeadNodes(NowDeadNodes, &ISU);
1628 DEBUG(errs() << "ISEL: Match complete!\n");
1634 CR_LeadsToInteriorNode
1637 /// WalkChainUsers - Walk down the users of the specified chained node that is
1638 /// part of the pattern we're matching, looking at all of the users we find.
1639 /// This determines whether something is an interior node, whether we have a
1640 /// non-pattern node in between two pattern nodes (which prevent folding because
1641 /// it would induce a cycle) and whether we have a TokenFactor node sandwiched
1642 /// between pattern nodes (in which case the TF becomes part of the pattern).
1644 /// The walk we do here is guaranteed to be small because we quickly get down to
1645 /// already selected nodes "below" us.
1647 WalkChainUsers(SDNode *ChainedNode,
1648 SmallVectorImpl<SDNode*> &ChainedNodesInPattern,
1649 SmallVectorImpl<SDNode*> &InteriorChainedNodes) {
1650 ChainResult Result = CR_Simple;
1652 for (SDNode::use_iterator UI = ChainedNode->use_begin(),
1653 E = ChainedNode->use_end(); UI != E; ++UI) {
1654 // Make sure the use is of the chain, not some other value we produce.
1655 if (UI.getUse().getValueType() != MVT::Other) continue;
1659 // If we see an already-selected machine node, then we've gone beyond the
1660 // pattern that we're selecting down into the already selected chunk of the
1662 if (User->isMachineOpcode() ||
1663 User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
1666 if (User->getOpcode() == ISD::CopyToReg ||
1667 User->getOpcode() == ISD::CopyFromReg ||
1668 User->getOpcode() == ISD::INLINEASM ||
1669 User->getOpcode() == ISD::EH_LABEL) {
1670 // If their node ID got reset to -1 then they've already been selected.
1671 // Treat them like a MachineOpcode.
1672 if (User->getNodeId() == -1)
1676 // If we have a TokenFactor, we handle it specially.
1677 if (User->getOpcode() != ISD::TokenFactor) {
1678 // If the node isn't a token factor and isn't part of our pattern, then it
1679 // must be a random chained node in between two nodes we're selecting.
1680 // This happens when we have something like:
1685 // Because we structurally match the load/store as a read/modify/write,
1686 // but the call is chained between them. We cannot fold in this case
1687 // because it would induce a cycle in the graph.
1688 if (!std::count(ChainedNodesInPattern.begin(),
1689 ChainedNodesInPattern.end(), User))
1690 return CR_InducesCycle;
1692 // Otherwise we found a node that is part of our pattern. For example in:
1696 // This would happen when we're scanning down from the load and see the
1697 // store as a user. Record that there is a use of ChainedNode that is
1698 // part of the pattern and keep scanning uses.
1699 Result = CR_LeadsToInteriorNode;
1700 InteriorChainedNodes.push_back(User);
1704 // If we found a TokenFactor, there are two cases to consider: first if the
1705 // TokenFactor is just hanging "below" the pattern we're matching (i.e. no
1706 // uses of the TF are in our pattern) we just want to ignore it. Second,
1707 // the TokenFactor can be sandwiched in between two chained nodes, like so:
1713 // | \ DAG's like cheese
1716 // [TokenFactor] [Op]
1723 // In this case, the TokenFactor becomes part of our match and we rewrite it
1724 // as a new TokenFactor.
1726 // To distinguish these two cases, do a recursive walk down the uses.
1727 switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) {
1729 // If the uses of the TokenFactor are just already-selected nodes, ignore
1730 // it, it is "below" our pattern.
1732 case CR_InducesCycle:
1733 // If the uses of the TokenFactor lead to nodes that are not part of our
1734 // pattern that are not selected, folding would turn this into a cycle,
1736 return CR_InducesCycle;
1737 case CR_LeadsToInteriorNode:
1738 break; // Otherwise, keep processing.
1741 // Okay, we know we're in the interesting interior case. The TokenFactor
1742 // is now going to be considered part of the pattern so that we rewrite its
1743 // uses (it may have uses that are not part of the pattern) with the
1744 // ultimate chain result of the generated code. We will also add its chain
1745 // inputs as inputs to the ultimate TokenFactor we create.
1746 Result = CR_LeadsToInteriorNode;
1747 ChainedNodesInPattern.push_back(User);
1748 InteriorChainedNodes.push_back(User);
1755 /// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
1756 /// operation for when the pattern matched at least one node with a chains. The
1757 /// input vector contains a list of all of the chained nodes that we match. We
1758 /// must determine if this is a valid thing to cover (i.e. matching it won't
1759 /// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
1760 /// be used as the input node chain for the generated nodes.
1762 HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
1763 SelectionDAG *CurDAG) {
1764 // Walk all of the chained nodes we've matched, recursively scanning down the
1765 // users of the chain result. This adds any TokenFactor nodes that are caught
1766 // in between chained nodes to the chained and interior nodes list.
1767 SmallVector<SDNode*, 3> InteriorChainedNodes;
1768 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1769 if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched,
1770 InteriorChainedNodes) == CR_InducesCycle)
1771 return SDValue(); // Would induce a cycle.
1774 // Okay, we have walked all the matched nodes and collected TokenFactor nodes
1775 // that we are interested in. Form our input TokenFactor node.
1776 SmallVector<SDValue, 3> InputChains;
1777 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1778 // Add the input chain of this node to the InputChains list (which will be
1779 // the operands of the generated TokenFactor) if it's not an interior node.
1780 SDNode *N = ChainNodesMatched[i];
1781 if (N->getOpcode() != ISD::TokenFactor) {
1782 if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N))
1785 // Otherwise, add the input chain.
1786 SDValue InChain = ChainNodesMatched[i]->getOperand(0);
1787 assert(InChain.getValueType() == MVT::Other && "Not a chain");
1788 InputChains.push_back(InChain);
1792 // If we have a token factor, we want to add all inputs of the token factor
1793 // that are not part of the pattern we're matching.
1794 for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) {
1795 if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(),
1796 N->getOperand(op).getNode()))
1797 InputChains.push_back(N->getOperand(op));
1802 if (InputChains.size() == 1)
1803 return InputChains[0];
1804 return CurDAG->getNode(ISD::TokenFactor, ChainNodesMatched[0]->getDebugLoc(),
1805 MVT::Other, &InputChains[0], InputChains.size());
1808 /// MorphNode - Handle morphing a node in place for the selector.
1809 SDNode *SelectionDAGISel::
1810 MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList,
1811 const SDValue *Ops, unsigned NumOps, unsigned EmitNodeInfo) {
1812 // It is possible we're using MorphNodeTo to replace a node with no
1813 // normal results with one that has a normal result (or we could be
1814 // adding a chain) and the input could have flags and chains as well.
1815 // In this case we need to shifting the operands down.
1816 // FIXME: This is a horrible hack and broken in obscure cases, no worse
1817 // than the old isel though. We should sink this into MorphNodeTo.
1818 int OldFlagResultNo = -1, OldChainResultNo = -1;
1820 unsigned NTMNumResults = Node->getNumValues();
1821 if (Node->getValueType(NTMNumResults-1) == MVT::Flag) {
1822 OldFlagResultNo = NTMNumResults-1;
1823 if (NTMNumResults != 1 &&
1824 Node->getValueType(NTMNumResults-2) == MVT::Other)
1825 OldChainResultNo = NTMNumResults-2;
1826 } else if (Node->getValueType(NTMNumResults-1) == MVT::Other)
1827 OldChainResultNo = NTMNumResults-1;
1829 // Call the underlying SelectionDAG routine to do the transmogrification. Note
1830 // that this deletes operands of the old node that become dead.
1831 SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops, NumOps);
1833 // MorphNodeTo can operate in two ways: if an existing node with the
1834 // specified operands exists, it can just return it. Otherwise, it
1835 // updates the node in place to have the requested operands.
1837 // If we updated the node in place, reset the node ID. To the isel,
1838 // this should be just like a newly allocated machine node.
1842 unsigned ResNumResults = Res->getNumValues();
1843 // Move the flag if needed.
1844 if ((EmitNodeInfo & OPFL_FlagOutput) && OldFlagResultNo != -1 &&
1845 (unsigned)OldFlagResultNo != ResNumResults-1)
1846 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldFlagResultNo),
1847 SDValue(Res, ResNumResults-1));
1849 if ((EmitNodeInfo & OPFL_FlagOutput) != 0)
1852 // Move the chain reference if needed.
1853 if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 &&
1854 (unsigned)OldChainResultNo != ResNumResults-1)
1855 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo),
1856 SDValue(Res, ResNumResults-1));
1858 // Otherwise, no replacement happened because the node already exists. Replace
1859 // Uses of the old node with the new one.
1861 CurDAG->ReplaceAllUsesWith(Node, Res);
1866 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1867 ALWAYS_INLINE static bool
1868 CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1869 SDValue N, const SmallVectorImpl<SDValue> &RecordedNodes) {
1870 // Accept if it is exactly the same as a previously recorded node.
1871 unsigned RecNo = MatcherTable[MatcherIndex++];
1872 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
1873 return N == RecordedNodes[RecNo];
1876 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1877 ALWAYS_INLINE static bool
1878 CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1879 SelectionDAGISel &SDISel) {
1880 return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]);
1883 /// CheckNodePredicate - Implements OP_CheckNodePredicate.
1884 ALWAYS_INLINE static bool
1885 CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1886 SelectionDAGISel &SDISel, SDNode *N) {
1887 return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]);
1890 ALWAYS_INLINE static bool
1891 CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1893 return N->getOpcode() == MatcherTable[MatcherIndex++];
1896 ALWAYS_INLINE static bool
1897 CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1898 SDValue N, const TargetLowering &TLI) {
1899 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1900 if (N.getValueType() == VT) return true;
1902 // Handle the case when VT is iPTR.
1903 return VT == MVT::iPTR && N.getValueType() == TLI.getPointerTy();
1906 ALWAYS_INLINE static bool
1907 CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1908 SDValue N, const TargetLowering &TLI,
1910 if (ChildNo >= N.getNumOperands())
1911 return false; // Match fails if out of range child #.
1912 return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI);
1916 ALWAYS_INLINE static bool
1917 CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1919 return cast<CondCodeSDNode>(N)->get() ==
1920 (ISD::CondCode)MatcherTable[MatcherIndex++];
1923 ALWAYS_INLINE static bool
1924 CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1925 SDValue N, const TargetLowering &TLI) {
1926 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1927 if (cast<VTSDNode>(N)->getVT() == VT)
1930 // Handle the case when VT is iPTR.
1931 return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI.getPointerTy();
1934 ALWAYS_INLINE static bool
1935 CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1937 int64_t Val = MatcherTable[MatcherIndex++];
1939 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1941 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
1942 return C != 0 && C->getSExtValue() == Val;
1945 ALWAYS_INLINE static bool
1946 CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1947 SDValue N, SelectionDAGISel &SDISel) {
1948 int64_t Val = MatcherTable[MatcherIndex++];
1950 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1952 if (N->getOpcode() != ISD::AND) return false;
1954 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1955 return C != 0 && SDISel.CheckAndMask(N.getOperand(0), C, Val);
1958 ALWAYS_INLINE static bool
1959 CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1960 SDValue N, SelectionDAGISel &SDISel) {
1961 int64_t Val = MatcherTable[MatcherIndex++];
1963 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1965 if (N->getOpcode() != ISD::OR) return false;
1967 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1968 return C != 0 && SDISel.CheckOrMask(N.getOperand(0), C, Val);
1971 /// IsPredicateKnownToFail - If we know how and can do so without pushing a
1972 /// scope, evaluate the current node. If the current predicate is known to
1973 /// fail, set Result=true and return anything. If the current predicate is
1974 /// known to pass, set Result=false and return the MatcherIndex to continue
1975 /// with. If the current predicate is unknown, set Result=false and return the
1976 /// MatcherIndex to continue with.
1977 static unsigned IsPredicateKnownToFail(const unsigned char *Table,
1978 unsigned Index, SDValue N,
1979 bool &Result, SelectionDAGISel &SDISel,
1980 SmallVectorImpl<SDValue> &RecordedNodes){
1981 switch (Table[Index++]) {
1984 return Index-1; // Could not evaluate this predicate.
1985 case SelectionDAGISel::OPC_CheckSame:
1986 Result = !::CheckSame(Table, Index, N, RecordedNodes);
1988 case SelectionDAGISel::OPC_CheckPatternPredicate:
1989 Result = !::CheckPatternPredicate(Table, Index, SDISel);
1991 case SelectionDAGISel::OPC_CheckPredicate:
1992 Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode());
1994 case SelectionDAGISel::OPC_CheckOpcode:
1995 Result = !::CheckOpcode(Table, Index, N.getNode());
1997 case SelectionDAGISel::OPC_CheckType:
1998 Result = !::CheckType(Table, Index, N, SDISel.TLI);
2000 case SelectionDAGISel::OPC_CheckChild0Type:
2001 case SelectionDAGISel::OPC_CheckChild1Type:
2002 case SelectionDAGISel::OPC_CheckChild2Type:
2003 case SelectionDAGISel::OPC_CheckChild3Type:
2004 case SelectionDAGISel::OPC_CheckChild4Type:
2005 case SelectionDAGISel::OPC_CheckChild5Type:
2006 case SelectionDAGISel::OPC_CheckChild6Type:
2007 case SelectionDAGISel::OPC_CheckChild7Type:
2008 Result = !::CheckChildType(Table, Index, N, SDISel.TLI,
2009 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Type);
2011 case SelectionDAGISel::OPC_CheckCondCode:
2012 Result = !::CheckCondCode(Table, Index, N);
2014 case SelectionDAGISel::OPC_CheckValueType:
2015 Result = !::CheckValueType(Table, Index, N, SDISel.TLI);
2017 case SelectionDAGISel::OPC_CheckInteger:
2018 Result = !::CheckInteger(Table, Index, N);
2020 case SelectionDAGISel::OPC_CheckAndImm:
2021 Result = !::CheckAndImm(Table, Index, N, SDISel);
2023 case SelectionDAGISel::OPC_CheckOrImm:
2024 Result = !::CheckOrImm(Table, Index, N, SDISel);
2031 /// FailIndex - If this match fails, this is the index to continue with.
2034 /// NodeStack - The node stack when the scope was formed.
2035 SmallVector<SDValue, 4> NodeStack;
2037 /// NumRecordedNodes - The number of recorded nodes when the scope was formed.
2038 unsigned NumRecordedNodes;
2040 /// NumMatchedMemRefs - The number of matched memref entries.
2041 unsigned NumMatchedMemRefs;
2043 /// InputChain/InputFlag - The current chain/flag
2044 SDValue InputChain, InputFlag;
2046 /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty.
2047 bool HasChainNodesMatched, HasFlagResultNodesMatched;
2050 SDNode *SelectionDAGISel::
2051 SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
2052 unsigned TableSize) {
2053 // FIXME: Should these even be selected? Handle these cases in the caller?
2054 switch (NodeToMatch->getOpcode()) {
2057 case ISD::EntryToken: // These nodes remain the same.
2058 case ISD::BasicBlock:
2060 //case ISD::VALUETYPE:
2061 //case ISD::CONDCODE:
2062 case ISD::HANDLENODE:
2063 case ISD::TargetConstant:
2064 case ISD::TargetConstantFP:
2065 case ISD::TargetConstantPool:
2066 case ISD::TargetFrameIndex:
2067 case ISD::TargetExternalSymbol:
2068 case ISD::TargetBlockAddress:
2069 case ISD::TargetJumpTable:
2070 case ISD::TargetGlobalTLSAddress:
2071 case ISD::TargetGlobalAddress:
2072 case ISD::TokenFactor:
2073 case ISD::CopyFromReg:
2074 case ISD::CopyToReg:
2076 NodeToMatch->setNodeId(-1); // Mark selected.
2078 case ISD::AssertSext:
2079 case ISD::AssertZext:
2080 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0),
2081 NodeToMatch->getOperand(0));
2083 case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch);
2084 case ISD::UNDEF: return Select_UNDEF(NodeToMatch);
2087 assert(!NodeToMatch->isMachineOpcode() && "Node already selected!");
2089 // Set up the node stack with NodeToMatch as the only node on the stack.
2090 SmallVector<SDValue, 8> NodeStack;
2091 SDValue N = SDValue(NodeToMatch, 0);
2092 NodeStack.push_back(N);
2094 // MatchScopes - Scopes used when matching, if a match failure happens, this
2095 // indicates where to continue checking.
2096 SmallVector<MatchScope, 8> MatchScopes;
2098 // RecordedNodes - This is the set of nodes that have been recorded by the
2100 SmallVector<SDValue, 8> RecordedNodes;
2102 // MatchedMemRefs - This is the set of MemRef's we've seen in the input
2104 SmallVector<MachineMemOperand*, 2> MatchedMemRefs;
2106 // These are the current input chain and flag for use when generating nodes.
2107 // Various Emit operations change these. For example, emitting a copytoreg
2108 // uses and updates these.
2109 SDValue InputChain, InputFlag;
2111 // ChainNodesMatched - If a pattern matches nodes that have input/output
2112 // chains, the OPC_EmitMergeInputChains operation is emitted which indicates
2113 // which ones they are. The result is captured into this list so that we can
2114 // update the chain results when the pattern is complete.
2115 SmallVector<SDNode*, 3> ChainNodesMatched;
2116 SmallVector<SDNode*, 3> FlagResultNodesMatched;
2118 DEBUG(errs() << "ISEL: Starting pattern match on root node: ";
2119 NodeToMatch->dump(CurDAG);
2122 // Determine where to start the interpreter. Normally we start at opcode #0,
2123 // but if the state machine starts with an OPC_SwitchOpcode, then we
2124 // accelerate the first lookup (which is guaranteed to be hot) with the
2125 // OpcodeOffset table.
2126 unsigned MatcherIndex = 0;
2128 if (!OpcodeOffset.empty()) {
2129 // Already computed the OpcodeOffset table, just index into it.
2130 if (N.getOpcode() < OpcodeOffset.size())
2131 MatcherIndex = OpcodeOffset[N.getOpcode()];
2132 DEBUG(errs() << " Initial Opcode index to " << MatcherIndex << "\n");
2134 } else if (MatcherTable[0] == OPC_SwitchOpcode) {
2135 // Otherwise, the table isn't computed, but the state machine does start
2136 // with an OPC_SwitchOpcode instruction. Populate the table now, since this
2137 // is the first time we're selecting an instruction.
2140 // Get the size of this case.
2141 unsigned CaseSize = MatcherTable[Idx++];
2143 CaseSize = GetVBR(CaseSize, MatcherTable, Idx);
2144 if (CaseSize == 0) break;
2146 // Get the opcode, add the index to the table.
2147 unsigned Opc = MatcherTable[Idx++];
2148 if (Opc >= OpcodeOffset.size())
2149 OpcodeOffset.resize((Opc+1)*2);
2150 OpcodeOffset[Opc] = Idx;
2154 // Okay, do the lookup for the first opcode.
2155 if (N.getOpcode() < OpcodeOffset.size())
2156 MatcherIndex = OpcodeOffset[N.getOpcode()];
2160 assert(MatcherIndex < TableSize && "Invalid index");
2162 unsigned CurrentOpcodeIndex = MatcherIndex;
2164 BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++];
2167 // Okay, the semantics of this operation are that we should push a scope
2168 // then evaluate the first child. However, pushing a scope only to have
2169 // the first check fail (which then pops it) is inefficient. If we can
2170 // determine immediately that the first check (or first several) will
2171 // immediately fail, don't even bother pushing a scope for them.
2175 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2176 if (NumToSkip & 128)
2177 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2178 // Found the end of the scope with no match.
2179 if (NumToSkip == 0) {
2184 FailIndex = MatcherIndex+NumToSkip;
2186 // If we can't evaluate this predicate without pushing a scope (e.g. if
2187 // it is a 'MoveParent') or if the predicate succeeds on this node, we
2188 // push the scope and evaluate the full predicate chain.
2190 MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
2191 Result, *this, RecordedNodes);
2195 DEBUG(errs() << " Skipped scope entry at index " << MatcherIndex
2196 << " continuing at " << FailIndex << "\n");
2199 // Otherwise, we know that this case of the Scope is guaranteed to fail,
2200 // move to the next case.
2201 MatcherIndex = FailIndex;
2204 // If the whole scope failed to match, bail.
2205 if (FailIndex == 0) break;
2207 // Push a MatchScope which indicates where to go if the first child fails
2209 MatchScope NewEntry;
2210 NewEntry.FailIndex = FailIndex;
2211 NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
2212 NewEntry.NumRecordedNodes = RecordedNodes.size();
2213 NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
2214 NewEntry.InputChain = InputChain;
2215 NewEntry.InputFlag = InputFlag;
2216 NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty();
2217 NewEntry.HasFlagResultNodesMatched = !FlagResultNodesMatched.empty();
2218 MatchScopes.push_back(NewEntry);
2221 case OPC_RecordNode:
2222 // Remember this node, it may end up being an operand in the pattern.
2223 RecordedNodes.push_back(N);
2226 case OPC_RecordChild0: case OPC_RecordChild1:
2227 case OPC_RecordChild2: case OPC_RecordChild3:
2228 case OPC_RecordChild4: case OPC_RecordChild5:
2229 case OPC_RecordChild6: case OPC_RecordChild7: {
2230 unsigned ChildNo = Opcode-OPC_RecordChild0;
2231 if (ChildNo >= N.getNumOperands())
2232 break; // Match fails if out of range child #.
2234 RecordedNodes.push_back(N->getOperand(ChildNo));
2237 case OPC_RecordMemRef:
2238 MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand());
2241 case OPC_CaptureFlagInput:
2242 // If the current node has an input flag, capture it in InputFlag.
2243 if (N->getNumOperands() != 0 &&
2244 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag)
2245 InputFlag = N->getOperand(N->getNumOperands()-1);
2248 case OPC_MoveChild: {
2249 unsigned ChildNo = MatcherTable[MatcherIndex++];
2250 if (ChildNo >= N.getNumOperands())
2251 break; // Match fails if out of range child #.
2252 N = N.getOperand(ChildNo);
2253 NodeStack.push_back(N);
2257 case OPC_MoveParent:
2258 // Pop the current node off the NodeStack.
2259 NodeStack.pop_back();
2260 assert(!NodeStack.empty() && "Node stack imbalance!");
2261 N = NodeStack.back();
2265 if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
2267 case OPC_CheckPatternPredicate:
2268 if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break;
2270 case OPC_CheckPredicate:
2271 if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this,
2275 case OPC_CheckComplexPat: {
2276 unsigned CPNum = MatcherTable[MatcherIndex++];
2277 unsigned RecNo = MatcherTable[MatcherIndex++];
2278 assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
2279 if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo], CPNum,
2284 case OPC_CheckOpcode:
2285 if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
2289 if (!::CheckType(MatcherTable, MatcherIndex, N, TLI)) break;
2292 case OPC_SwitchOpcode: {
2293 unsigned CurNodeOpcode = N.getOpcode();
2294 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2297 // Get the size of this case.
2298 CaseSize = MatcherTable[MatcherIndex++];
2300 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2301 if (CaseSize == 0) break;
2303 // If the opcode matches, then we will execute this case.
2304 if (CurNodeOpcode == MatcherTable[MatcherIndex++])
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() << " OpcodeSwitch from " << SwitchStart
2316 << " to " << MatcherIndex << "\n");
2320 case OPC_SwitchType: {
2321 MVT::SimpleValueType CurNodeVT = N.getValueType().getSimpleVT().SimpleTy;
2322 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2325 // Get the size of this case.
2326 CaseSize = MatcherTable[MatcherIndex++];
2328 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2329 if (CaseSize == 0) break;
2331 MVT::SimpleValueType CaseVT =
2332 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2333 if (CaseVT == MVT::iPTR)
2334 CaseVT = TLI.getPointerTy().SimpleTy;
2336 // If the VT matches, then we will execute this case.
2337 if (CurNodeVT == CaseVT)
2340 // Otherwise, skip over this case.
2341 MatcherIndex += CaseSize;
2344 // If no cases matched, bail out.
2345 if (CaseSize == 0) break;
2347 // Otherwise, execute the case we found.
2348 DEBUG(errs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
2349 << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
2352 case OPC_CheckChild0Type: case OPC_CheckChild1Type:
2353 case OPC_CheckChild2Type: case OPC_CheckChild3Type:
2354 case OPC_CheckChild4Type: case OPC_CheckChild5Type:
2355 case OPC_CheckChild6Type: case OPC_CheckChild7Type:
2356 if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI,
2357 Opcode-OPC_CheckChild0Type))
2360 case OPC_CheckCondCode:
2361 if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
2363 case OPC_CheckValueType:
2364 if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI)) break;
2366 case OPC_CheckInteger:
2367 if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
2369 case OPC_CheckAndImm:
2370 if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
2372 case OPC_CheckOrImm:
2373 if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
2376 case OPC_CheckFoldableChainNode: {
2377 assert(NodeStack.size() != 1 && "No parent node");
2378 // Verify that all intermediate nodes between the root and this one have
2380 bool HasMultipleUses = false;
2381 for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i)
2382 if (!NodeStack[i].hasOneUse()) {
2383 HasMultipleUses = true;
2386 if (HasMultipleUses) break;
2388 // Check to see that the target thinks this is profitable to fold and that
2389 // we can fold it without inducing cycles in the graph.
2390 if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2392 !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2393 NodeToMatch, true/*We validate our own chains*/))
2398 case OPC_EmitInteger: {
2399 MVT::SimpleValueType VT =
2400 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2401 int64_t Val = MatcherTable[MatcherIndex++];
2403 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2404 RecordedNodes.push_back(CurDAG->getTargetConstant(Val, VT));
2407 case OPC_EmitRegister: {
2408 MVT::SimpleValueType VT =
2409 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2410 unsigned RegNo = MatcherTable[MatcherIndex++];
2411 RecordedNodes.push_back(CurDAG->getRegister(RegNo, VT));
2415 case OPC_EmitConvertToTarget: {
2416 // Convert from IMM/FPIMM to target version.
2417 unsigned RecNo = MatcherTable[MatcherIndex++];
2418 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2419 SDValue Imm = RecordedNodes[RecNo];
2421 if (Imm->getOpcode() == ISD::Constant) {
2422 int64_t Val = cast<ConstantSDNode>(Imm)->getZExtValue();
2423 Imm = CurDAG->getTargetConstant(Val, Imm.getValueType());
2424 } else if (Imm->getOpcode() == ISD::ConstantFP) {
2425 const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue();
2426 Imm = CurDAG->getTargetConstantFP(*Val, Imm.getValueType());
2429 RecordedNodes.push_back(Imm);
2433 case OPC_EmitMergeInputChains: {
2434 assert(InputChain.getNode() == 0 &&
2435 "EmitMergeInputChains should be the first chain producing node");
2436 // This node gets a list of nodes we matched in the input that have
2437 // chains. We want to token factor all of the input chains to these nodes
2438 // together. However, if any of the input chains is actually one of the
2439 // nodes matched in this pattern, then we have an intra-match reference.
2440 // Ignore these because the newly token factored chain should not refer to
2442 unsigned NumChains = MatcherTable[MatcherIndex++];
2443 assert(NumChains != 0 && "Can't TF zero chains");
2445 assert(ChainNodesMatched.empty() &&
2446 "Should only have one EmitMergeInputChains per match");
2448 // Read all of the chained nodes.
2449 for (unsigned i = 0; i != NumChains; ++i) {
2450 unsigned RecNo = MatcherTable[MatcherIndex++];
2451 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2452 ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode());
2454 // FIXME: What if other value results of the node have uses not matched
2456 if (ChainNodesMatched.back() != NodeToMatch &&
2457 !RecordedNodes[RecNo].hasOneUse()) {
2458 ChainNodesMatched.clear();
2463 // If the inner loop broke out, the match fails.
2464 if (ChainNodesMatched.empty())
2467 // Merge the input chains if they are not intra-pattern references.
2468 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2470 if (InputChain.getNode() == 0)
2471 break; // Failed to merge.
2476 case OPC_EmitCopyToReg: {
2477 unsigned RecNo = MatcherTable[MatcherIndex++];
2478 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2479 unsigned DestPhysReg = MatcherTable[MatcherIndex++];
2481 if (InputChain.getNode() == 0)
2482 InputChain = CurDAG->getEntryNode();
2484 InputChain = CurDAG->getCopyToReg(InputChain, NodeToMatch->getDebugLoc(),
2485 DestPhysReg, RecordedNodes[RecNo],
2488 InputFlag = InputChain.getValue(1);
2492 case OPC_EmitNodeXForm: {
2493 unsigned XFormNo = MatcherTable[MatcherIndex++];
2494 unsigned RecNo = MatcherTable[MatcherIndex++];
2495 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2496 RecordedNodes.push_back(RunSDNodeXForm(RecordedNodes[RecNo], XFormNo));
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_FlagOutput)
2518 VTs.push_back(MVT::Flag);
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]);
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 flag
2551 for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands();
2553 SDValue V = NodeToMatch->getOperand(i);
2554 if (V.getValueType() == MVT::Flag) break;
2559 // If this has chain/flag inputs, add them.
2560 if (EmitNodeInfo & OPFL_Chain)
2561 Ops.push_back(InputChain);
2562 if ((EmitNodeInfo & OPFL_FlagInput) && InputFlag.getNode() != 0)
2563 Ops.push_back(InputFlag);
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-flag/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::Flag) break;
2576 RecordedNodes.push_back(SDValue(Res, i));
2580 Res = MorphNode(NodeToMatch, TargetOpc, VTList, Ops.data(), Ops.size(),
2584 // If the node had chain/flag results, update our notion of the current
2586 if (EmitNodeInfo & OPFL_FlagOutput) {
2587 InputFlag = SDValue(Res, VTs.size()-1);
2588 if (EmitNodeInfo & OPFL_Chain)
2589 InputChain = SDValue(Res, VTs.size()-2);
2590 } else if (EmitNodeInfo & OPFL_Chain)
2591 InputChain = SDValue(Res, VTs.size()-1);
2593 // If the OPFL_MemRefs flag is set on this node, slap all of the
2594 // accumulated memrefs onto it.
2596 // FIXME: This is vastly incorrect for patterns with multiple outputs
2597 // instructions that access memory and for ComplexPatterns that match
2599 if (EmitNodeInfo & OPFL_MemRefs) {
2600 MachineSDNode::mmo_iterator MemRefs =
2601 MF->allocateMemRefsArray(MatchedMemRefs.size());
2602 std::copy(MatchedMemRefs.begin(), MatchedMemRefs.end(), MemRefs);
2603 cast<MachineSDNode>(Res)
2604 ->setMemRefs(MemRefs, MemRefs + MatchedMemRefs.size());
2608 << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created")
2609 << " node: "; Res->dump(CurDAG); errs() << "\n");
2611 // If this was a MorphNodeTo then we're completely done!
2612 if (Opcode == OPC_MorphNodeTo) {
2613 // Update chain and flag uses.
2614 UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched,
2615 InputFlag, FlagResultNodesMatched, true);
2622 case OPC_MarkFlagResults: {
2623 unsigned NumNodes = MatcherTable[MatcherIndex++];
2625 // Read and remember all the flag-result nodes.
2626 for (unsigned i = 0; i != NumNodes; ++i) {
2627 unsigned RecNo = MatcherTable[MatcherIndex++];
2629 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2631 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2632 FlagResultNodesMatched.push_back(RecordedNodes[RecNo].getNode());
2637 case OPC_CompleteMatch: {
2638 // The match has been completed, and any new nodes (if any) have been
2639 // created. Patch up references to the matched dag to use the newly
2641 unsigned NumResults = MatcherTable[MatcherIndex++];
2643 for (unsigned i = 0; i != NumResults; ++i) {
2644 unsigned ResSlot = MatcherTable[MatcherIndex++];
2646 ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex);
2648 assert(ResSlot < RecordedNodes.size() && "Invalid CheckSame");
2649 SDValue Res = RecordedNodes[ResSlot];
2651 // FIXME2: Eliminate this horrible hack by fixing the 'Gen' program
2652 // after (parallel) on input patterns are removed. This would also
2653 // allow us to stop encoding #results in OPC_CompleteMatch's table
2655 if (NodeToMatch->getNumValues() <= i ||
2656 NodeToMatch->getValueType(i) == MVT::Other ||
2657 NodeToMatch->getValueType(i) == MVT::Flag)
2659 assert((NodeToMatch->getValueType(i) == Res.getValueType() ||
2660 NodeToMatch->getValueType(i) == MVT::iPTR ||
2661 Res.getValueType() == MVT::iPTR ||
2662 NodeToMatch->getValueType(i).getSizeInBits() ==
2663 Res.getValueType().getSizeInBits()) &&
2664 "invalid replacement");
2665 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, i), Res);
2668 // If the root node defines a flag, add it to the flag nodes to update
2670 if (NodeToMatch->getValueType(NodeToMatch->getNumValues()-1) == MVT::Flag)
2671 FlagResultNodesMatched.push_back(NodeToMatch);
2673 // Update chain and flag uses.
2674 UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched,
2675 InputFlag, FlagResultNodesMatched, false);
2677 assert(NodeToMatch->use_empty() &&
2678 "Didn't replace all uses of the node?");
2680 // FIXME: We just return here, which interacts correctly with SelectRoot
2681 // above. We should fix this to not return an SDNode* anymore.
2686 // If the code reached this point, then the match failed. See if there is
2687 // another child to try in the current 'Scope', otherwise pop it until we
2688 // find a case to check.
2689 DEBUG(errs() << " Match failed at index " << CurrentOpcodeIndex << "\n");
2691 if (MatchScopes.empty()) {
2692 CannotYetSelect(NodeToMatch);
2696 // Restore the interpreter state back to the point where the scope was
2698 MatchScope &LastScope = MatchScopes.back();
2699 RecordedNodes.resize(LastScope.NumRecordedNodes);
2701 NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end());
2702 N = NodeStack.back();
2704 if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size())
2705 MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
2706 MatcherIndex = LastScope.FailIndex;
2708 DEBUG(errs() << " Continuing at " << MatcherIndex << "\n");
2710 InputChain = LastScope.InputChain;
2711 InputFlag = LastScope.InputFlag;
2712 if (!LastScope.HasChainNodesMatched)
2713 ChainNodesMatched.clear();
2714 if (!LastScope.HasFlagResultNodesMatched)
2715 FlagResultNodesMatched.clear();
2717 // Check to see what the offset is at the new MatcherIndex. If it is zero
2718 // we have reached the end of this scope, otherwise we have another child
2719 // in the current scope to try.
2720 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2721 if (NumToSkip & 128)
2722 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2724 // If we have another child in this scope to match, update FailIndex and
2726 if (NumToSkip != 0) {
2727 LastScope.FailIndex = MatcherIndex+NumToSkip;
2731 // End of this scope, pop it and try the next child in the containing
2733 MatchScopes.pop_back();
2740 void SelectionDAGISel::CannotYetSelect(SDNode *N) {
2742 raw_string_ostream Msg(msg);
2743 Msg << "Cannot yet select: ";
2745 if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN &&
2746 N->getOpcode() != ISD::INTRINSIC_WO_CHAIN &&
2747 N->getOpcode() != ISD::INTRINSIC_VOID) {
2748 N->printrFull(Msg, CurDAG);
2750 bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other;
2752 cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue();
2753 if (iid < Intrinsic::num_intrinsics)
2754 Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid);
2755 else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo())
2756 Msg << "target intrinsic %" << TII->getName(iid);
2758 Msg << "unknown intrinsic #" << iid;
2760 llvm_report_error(Msg.str());
2763 char SelectionDAGISel::ID = 0;