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 CurDAG->DeleteNode(N);
516 // Run ShrinkDemandedOp on scalar binary operations.
517 if (N->getNumValues() != 1 ||
518 !N->getValueType(0).isSimple() || !N->getValueType(0).isInteger())
521 unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
522 APInt Demanded = APInt::getAllOnesValue(BitWidth);
523 APInt KnownZero, KnownOne;
524 if (!TLI.SimplifyDemandedBits(SDValue(N, 0), Demanded,
525 KnownZero, KnownOne, TLO) &&
526 (N->getOpcode() != ISD::TRUNCATE ||
527 !TrivialTruncElim(SDValue(N, 0), TLO)))
531 assert(!InWorklist.count(N) && "Already in worklist");
532 Worklist.push_back(N);
533 InWorklist.insert(N);
535 // Replace the old value with the new one.
536 DEBUG(errs() << "\nShrinkDemandedOps replacing ";
537 TLO.Old.getNode()->dump(CurDAG);
538 errs() << "\nWith: ";
539 TLO.New.getNode()->dump(CurDAG);
542 if (InWorklist.insert(TLO.New.getNode()))
543 Worklist.push_back(TLO.New.getNode());
545 SDOPsWorkListRemover DeadNodes(Worklist, InWorklist);
546 CurDAG->ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, &DeadNodes);
548 if (!TLO.Old.getNode()->use_empty()) continue;
550 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands();
552 SDNode *OpNode = TLO.Old.getNode()->getOperand(i).getNode();
553 if (OpNode->hasOneUse()) {
554 // Add OpNode to the end of the list to revisit.
555 DeadNodes.RemoveFromWorklist(OpNode);
556 Worklist.push_back(OpNode);
557 InWorklist.insert(OpNode);
561 DeadNodes.RemoveFromWorklist(TLO.Old.getNode());
562 CurDAG->DeleteNode(TLO.Old.getNode());
566 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
567 SmallPtrSet<SDNode*, 128> VisitedNodes;
568 SmallVector<SDNode*, 128> Worklist;
570 Worklist.push_back(CurDAG->getRoot().getNode());
577 SDNode *N = Worklist.pop_back_val();
579 // If we've already seen this node, ignore it.
580 if (!VisitedNodes.insert(N))
583 // Otherwise, add all chain operands to the worklist.
584 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
585 if (N->getOperand(i).getValueType() == MVT::Other)
586 Worklist.push_back(N->getOperand(i).getNode());
588 // If this is a CopyToReg with a vreg dest, process it.
589 if (N->getOpcode() != ISD::CopyToReg)
592 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
593 if (!TargetRegisterInfo::isVirtualRegister(DestReg))
596 // Ignore non-scalar or non-integer values.
597 SDValue Src = N->getOperand(2);
598 EVT SrcVT = Src.getValueType();
599 if (!SrcVT.isInteger() || SrcVT.isVector())
602 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
603 Mask = APInt::getAllOnesValue(SrcVT.getSizeInBits());
604 CurDAG->ComputeMaskedBits(Src, Mask, KnownZero, KnownOne);
606 // Only install this information if it tells us something.
607 if (NumSignBits != 1 || KnownZero != 0 || KnownOne != 0) {
608 DestReg -= TargetRegisterInfo::FirstVirtualRegister;
609 if (DestReg >= FuncInfo->LiveOutRegInfo.size())
610 FuncInfo->LiveOutRegInfo.resize(DestReg+1);
611 FunctionLoweringInfo::LiveOutInfo &LOI =
612 FuncInfo->LiveOutRegInfo[DestReg];
613 LOI.NumSignBits = NumSignBits;
614 LOI.KnownOne = KnownOne;
615 LOI.KnownZero = KnownZero;
617 } while (!Worklist.empty());
620 void SelectionDAGISel::CodeGenAndEmitDAG() {
621 std::string GroupName;
622 if (TimePassesIsEnabled)
623 GroupName = "Instruction Selection and Scheduling";
624 std::string BlockName;
625 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs ||
626 ViewDAGCombine2 || ViewDAGCombineLT || ViewISelDAGs || ViewSchedDAGs ||
628 BlockName = MF->getFunction()->getNameStr() + ":" +
629 BB->getBasicBlock()->getNameStr();
631 DEBUG(dbgs() << "Initial selection DAG:\n");
632 DEBUG(CurDAG->dump());
634 if (ViewDAGCombine1) CurDAG->viewGraph("dag-combine1 input for " + BlockName);
636 // Run the DAG combiner in pre-legalize mode.
637 if (TimePassesIsEnabled) {
638 NamedRegionTimer T("DAG Combining 1", GroupName);
639 CurDAG->Combine(Unrestricted, *AA, OptLevel);
641 CurDAG->Combine(Unrestricted, *AA, OptLevel);
644 DEBUG(dbgs() << "Optimized lowered selection DAG:\n");
645 DEBUG(CurDAG->dump());
647 // Second step, hack on the DAG until it only uses operations and types that
648 // the target supports.
649 if (ViewLegalizeTypesDAGs) CurDAG->viewGraph("legalize-types input for " +
653 if (TimePassesIsEnabled) {
654 NamedRegionTimer T("Type Legalization", GroupName);
655 Changed = CurDAG->LegalizeTypes();
657 Changed = CurDAG->LegalizeTypes();
660 DEBUG(dbgs() << "Type-legalized selection DAG:\n");
661 DEBUG(CurDAG->dump());
664 if (ViewDAGCombineLT)
665 CurDAG->viewGraph("dag-combine-lt input for " + BlockName);
667 // Run the DAG combiner in post-type-legalize mode.
668 if (TimePassesIsEnabled) {
669 NamedRegionTimer T("DAG Combining after legalize types", GroupName);
670 CurDAG->Combine(NoIllegalTypes, *AA, OptLevel);
672 CurDAG->Combine(NoIllegalTypes, *AA, OptLevel);
675 DEBUG(dbgs() << "Optimized type-legalized selection DAG:\n");
676 DEBUG(CurDAG->dump());
679 if (TimePassesIsEnabled) {
680 NamedRegionTimer T("Vector Legalization", GroupName);
681 Changed = CurDAG->LegalizeVectors();
683 Changed = CurDAG->LegalizeVectors();
687 if (TimePassesIsEnabled) {
688 NamedRegionTimer T("Type Legalization 2", GroupName);
689 CurDAG->LegalizeTypes();
691 CurDAG->LegalizeTypes();
694 if (ViewDAGCombineLT)
695 CurDAG->viewGraph("dag-combine-lv input for " + BlockName);
697 // Run the DAG combiner in post-type-legalize mode.
698 if (TimePassesIsEnabled) {
699 NamedRegionTimer T("DAG Combining after legalize vectors", GroupName);
700 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
702 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
705 DEBUG(dbgs() << "Optimized vector-legalized selection DAG:\n");
706 DEBUG(CurDAG->dump());
709 if (ViewLegalizeDAGs) CurDAG->viewGraph("legalize input for " + BlockName);
711 if (TimePassesIsEnabled) {
712 NamedRegionTimer T("DAG Legalization", GroupName);
713 CurDAG->Legalize(OptLevel);
715 CurDAG->Legalize(OptLevel);
718 DEBUG(dbgs() << "Legalized selection DAG:\n");
719 DEBUG(CurDAG->dump());
721 if (ViewDAGCombine2) CurDAG->viewGraph("dag-combine2 input for " + BlockName);
723 // Run the DAG combiner in post-legalize mode.
724 if (TimePassesIsEnabled) {
725 NamedRegionTimer T("DAG Combining 2", GroupName);
726 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
728 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
731 DEBUG(dbgs() << "Optimized legalized selection DAG:\n");
732 DEBUG(CurDAG->dump());
734 if (OptLevel != CodeGenOpt::None) {
736 ComputeLiveOutVRegInfo();
739 if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName);
741 // Third, instruction select all of the operations to machine code, adding the
742 // code to the MachineBasicBlock.
743 if (TimePassesIsEnabled) {
744 NamedRegionTimer T("Instruction Selection", GroupName);
745 DoInstructionSelection();
747 DoInstructionSelection();
750 DEBUG(dbgs() << "Selected selection DAG:\n");
751 DEBUG(CurDAG->dump());
753 if (ViewSchedDAGs) CurDAG->viewGraph("scheduler input for " + BlockName);
755 // Schedule machine code.
756 ScheduleDAGSDNodes *Scheduler = CreateScheduler();
757 if (TimePassesIsEnabled) {
758 NamedRegionTimer T("Instruction Scheduling", GroupName);
759 Scheduler->Run(CurDAG, BB, BB->end());
761 Scheduler->Run(CurDAG, BB, BB->end());
764 if (ViewSUnitDAGs) Scheduler->viewGraph();
766 // Emit machine code to BB. This can change 'BB' to the last block being
768 if (TimePassesIsEnabled) {
769 NamedRegionTimer T("Instruction Creation", GroupName);
770 BB = Scheduler->EmitSchedule(&SDB->EdgeMapping);
772 BB = Scheduler->EmitSchedule(&SDB->EdgeMapping);
775 // Free the scheduler state.
776 if (TimePassesIsEnabled) {
777 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName);
783 DEBUG(dbgs() << "Selected machine code:\n");
787 void SelectionDAGISel::DoInstructionSelection() {
788 DEBUG(errs() << "===== Instruction selection begins:\n");
792 // Select target instructions for the DAG.
794 // Number all nodes with a topological order and set DAGSize.
795 DAGSize = CurDAG->AssignTopologicalOrder();
797 // Create a dummy node (which is not added to allnodes), that adds
798 // a reference to the root node, preventing it from being deleted,
799 // and tracking any changes of the root.
800 HandleSDNode Dummy(CurDAG->getRoot());
801 ISelPosition = SelectionDAG::allnodes_iterator(CurDAG->getRoot().getNode());
804 // The AllNodes list is now topological-sorted. Visit the
805 // nodes by starting at the end of the list (the root of the
806 // graph) and preceding back toward the beginning (the entry
808 while (ISelPosition != CurDAG->allnodes_begin()) {
809 SDNode *Node = --ISelPosition;
810 // Skip dead nodes. DAGCombiner is expected to eliminate all dead nodes,
811 // but there are currently some corner cases that it misses. Also, this
812 // makes it theoretically possible to disable the DAGCombiner.
813 if (Node->use_empty())
816 SDNode *ResNode = Select(Node);
818 // FIXME: This is pretty gross. 'Select' should be changed to not return
819 // anything at all and this code should be nuked with a tactical strike.
821 // If node should not be replaced, continue with the next one.
822 if (ResNode == Node || Node->getOpcode() == ISD::DELETED_NODE)
826 ReplaceUses(Node, ResNode);
828 // If after the replacement this node is not used any more,
829 // remove this dead node.
830 if (Node->use_empty()) { // Don't delete EntryToken, etc.
831 ISelUpdater ISU(ISelPosition);
832 CurDAG->RemoveDeadNode(Node, &ISU);
836 CurDAG->setRoot(Dummy.getValue());
838 DEBUG(errs() << "===== Instruction selection ends:\n");
840 PostprocessISelDAG();
842 // FIXME: This shouldn't be needed, remove it.
843 CurDAG->RemoveDeadNodes();
847 void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn,
849 MachineModuleInfo *MMI,
851 const TargetInstrInfo &TII) {
852 // Initialize the Fast-ISel state, if needed.
853 FastISel *FastIS = 0;
855 FastIS = TLI.createFastISel(MF, MMI, DW,
858 FuncInfo->StaticAllocaMap
860 , FuncInfo->CatchInfoLost
864 unsigned MDDbgKind = Fn.getContext().getMDKindID("dbg");
866 // Iterate over all basic blocks in the function.
867 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
868 BasicBlock *LLVMBB = &*I;
869 BB = FuncInfo->MBBMap[LLVMBB];
871 BasicBlock::iterator const Begin = LLVMBB->begin();
872 BasicBlock::iterator const End = LLVMBB->end();
873 BasicBlock::iterator BI = Begin;
875 // Lower any arguments needed in this block if this is the entry block.
876 bool SuppressFastISel = false;
877 if (LLVMBB == &Fn.getEntryBlock()) {
878 LowerArguments(LLVMBB);
880 // If any of the arguments has the byval attribute, forgo
881 // fast-isel in the entry block.
884 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end();
886 if (Fn.paramHasAttr(j, Attribute::ByVal)) {
887 if (EnableFastISelVerbose || EnableFastISelAbort)
888 dbgs() << "FastISel skips entry block due to byval argument\n";
889 SuppressFastISel = true;
895 if (MMI && BB->isLandingPad()) {
896 // Add a label to mark the beginning of the landing pad. Deletion of the
897 // landing pad can thus be detected via the MachineModuleInfo.
898 MCSymbol *Label = MMI->addLandingPad(BB);
900 const TargetInstrDesc &II = TII.get(TargetOpcode::EH_LABEL);
901 BuildMI(BB, SDB->getCurDebugLoc(), II).addSym(Label);
903 // Mark exception register as live in.
904 unsigned Reg = TLI.getExceptionAddressRegister();
905 if (Reg) BB->addLiveIn(Reg);
907 // Mark exception selector register as live in.
908 Reg = TLI.getExceptionSelectorRegister();
909 if (Reg) BB->addLiveIn(Reg);
911 // FIXME: Hack around an exception handling flaw (PR1508): the personality
912 // function and list of typeids logically belong to the invoke (or, if you
913 // like, the basic block containing the invoke), and need to be associated
914 // with it in the dwarf exception handling tables. Currently however the
915 // information is provided by an intrinsic (eh.selector) that can be moved
916 // to unexpected places by the optimizers: if the unwind edge is critical,
917 // then breaking it can result in the intrinsics being in the successor of
918 // the landing pad, not the landing pad itself. This results
919 // in exceptions not being caught because no typeids are associated with
920 // the invoke. This may not be the only way things can go wrong, but it
921 // is the only way we try to work around for the moment.
922 BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
924 if (Br && Br->isUnconditional()) { // Critical edge?
925 BasicBlock::iterator I, E;
926 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
927 if (isa<EHSelectorInst>(I))
931 // No catch info found - try to extract some from the successor.
932 CopyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, *FuncInfo);
936 // Before doing SelectionDAG ISel, see if FastISel has been requested.
937 if (FastIS && !SuppressFastISel) {
938 // Emit code for any incoming arguments. This must happen before
939 // beginning FastISel on the entry block.
940 if (LLVMBB == &Fn.getEntryBlock()) {
941 CurDAG->setRoot(SDB->getControlRoot());
945 FastIS->startNewBlock(BB);
946 // Do FastISel on as many instructions as possible.
947 for (; BI != End; ++BI) {
948 // Just before the terminator instruction, insert instructions to
949 // feed PHI nodes in successor blocks.
950 if (isa<TerminatorInst>(BI))
951 if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, FastIS)) {
952 ++NumFastIselFailures;
953 ResetDebugLoc(SDB, FastIS);
954 if (EnableFastISelVerbose || EnableFastISelAbort) {
955 dbgs() << "FastISel miss: ";
958 assert(!EnableFastISelAbort &&
959 "FastISel didn't handle a PHI in a successor");
963 SetDebugLoc(MDDbgKind, BI, SDB, FastIS, &MF);
965 // Try to select the instruction with FastISel.
966 if (FastIS->SelectInstruction(BI)) {
967 ResetDebugLoc(SDB, FastIS);
971 // Clear out the debug location so that it doesn't carry over to
972 // unrelated instructions.
973 ResetDebugLoc(SDB, FastIS);
975 // Then handle certain instructions as single-LLVM-Instruction blocks.
976 if (isa<CallInst>(BI)) {
977 ++NumFastIselFailures;
978 if (EnableFastISelVerbose || EnableFastISelAbort) {
979 dbgs() << "FastISel missed call: ";
983 if (!BI->getType()->isVoidTy()) {
984 unsigned &R = FuncInfo->ValueMap[BI];
986 R = FuncInfo->CreateRegForValue(BI);
989 bool HadTailCall = false;
990 SelectBasicBlock(LLVMBB, BI, llvm::next(BI), HadTailCall);
992 // If the call was emitted as a tail call, we're done with the block.
998 // If the instruction was codegen'd with multiple blocks,
999 // inform the FastISel object where to resume inserting.
1000 FastIS->setCurrentBlock(BB);
1004 // Otherwise, give up on FastISel for the rest of the block.
1005 // For now, be a little lenient about non-branch terminators.
1006 if (!isa<TerminatorInst>(BI) || isa<BranchInst>(BI)) {
1007 ++NumFastIselFailures;
1008 if (EnableFastISelVerbose || EnableFastISelAbort) {
1009 dbgs() << "FastISel miss: ";
1012 if (EnableFastISelAbort)
1013 // The "fast" selector couldn't handle something and bailed.
1014 // For the purpose of debugging, just abort.
1015 llvm_unreachable("FastISel didn't select the entire block");
1021 // Run SelectionDAG instruction selection on the remainder of the block
1022 // not handled by FastISel. If FastISel is not run, this is the entire
1026 SelectBasicBlock(LLVMBB, BI, End, HadTailCall);
1036 SelectionDAGISel::FinishBasicBlock() {
1038 DEBUG(dbgs() << "Target-post-processed machine code:\n");
1041 DEBUG(dbgs() << "Total amount of phi nodes to update: "
1042 << SDB->PHINodesToUpdate.size() << "\n");
1043 DEBUG(for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i)
1044 dbgs() << "Node " << i << " : ("
1045 << SDB->PHINodesToUpdate[i].first
1046 << ", " << SDB->PHINodesToUpdate[i].second << ")\n");
1048 // Next, now that we know what the last MBB the LLVM BB expanded is, update
1049 // PHI nodes in successors.
1050 if (SDB->SwitchCases.empty() &&
1051 SDB->JTCases.empty() &&
1052 SDB->BitTestCases.empty()) {
1053 for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) {
1054 MachineInstr *PHI = SDB->PHINodesToUpdate[i].first;
1055 assert(PHI->isPHI() &&
1056 "This is not a machine PHI node that we are updating!");
1057 if (!BB->isSuccessor(PHI->getParent()))
1059 PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second,
1061 PHI->addOperand(MachineOperand::CreateMBB(BB));
1063 SDB->PHINodesToUpdate.clear();
1067 for (unsigned i = 0, e = SDB->BitTestCases.size(); i != e; ++i) {
1068 // Lower header first, if it wasn't already lowered
1069 if (!SDB->BitTestCases[i].Emitted) {
1070 // Set the current basic block to the mbb we wish to insert the code into
1071 BB = SDB->BitTestCases[i].Parent;
1072 SDB->setCurrentBasicBlock(BB);
1074 SDB->visitBitTestHeader(SDB->BitTestCases[i]);
1075 CurDAG->setRoot(SDB->getRoot());
1076 CodeGenAndEmitDAG();
1080 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j) {
1081 // Set the current basic block to the mbb we wish to insert the code into
1082 BB = SDB->BitTestCases[i].Cases[j].ThisBB;
1083 SDB->setCurrentBasicBlock(BB);
1086 SDB->visitBitTestCase(SDB->BitTestCases[i].Cases[j+1].ThisBB,
1087 SDB->BitTestCases[i].Reg,
1088 SDB->BitTestCases[i].Cases[j]);
1090 SDB->visitBitTestCase(SDB->BitTestCases[i].Default,
1091 SDB->BitTestCases[i].Reg,
1092 SDB->BitTestCases[i].Cases[j]);
1095 CurDAG->setRoot(SDB->getRoot());
1096 CodeGenAndEmitDAG();
1101 for (unsigned pi = 0, pe = SDB->PHINodesToUpdate.size(); pi != pe; ++pi) {
1102 MachineInstr *PHI = SDB->PHINodesToUpdate[pi].first;
1103 MachineBasicBlock *PHIBB = PHI->getParent();
1104 assert(PHI->isPHI() &&
1105 "This is not a machine PHI node that we are updating!");
1106 // This is "default" BB. We have two jumps to it. From "header" BB and
1107 // from last "case" BB.
1108 if (PHIBB == SDB->BitTestCases[i].Default) {
1109 PHI->addOperand(MachineOperand::
1110 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1111 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Parent));
1112 PHI->addOperand(MachineOperand::
1113 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1114 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Cases.
1117 // One of "cases" BB.
1118 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size();
1120 MachineBasicBlock* cBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1121 if (cBB->isSuccessor(PHIBB)) {
1122 PHI->addOperand(MachineOperand::
1123 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1124 PHI->addOperand(MachineOperand::CreateMBB(cBB));
1129 SDB->BitTestCases.clear();
1131 // If the JumpTable record is filled in, then we need to emit a jump table.
1132 // Updating the PHI nodes is tricky in this case, since we need to determine
1133 // whether the PHI is a successor of the range check MBB or the jump table MBB
1134 for (unsigned i = 0, e = SDB->JTCases.size(); i != e; ++i) {
1135 // Lower header first, if it wasn't already lowered
1136 if (!SDB->JTCases[i].first.Emitted) {
1137 // Set the current basic block to the mbb we wish to insert the code into
1138 BB = SDB->JTCases[i].first.HeaderBB;
1139 SDB->setCurrentBasicBlock(BB);
1141 SDB->visitJumpTableHeader(SDB->JTCases[i].second, SDB->JTCases[i].first);
1142 CurDAG->setRoot(SDB->getRoot());
1143 CodeGenAndEmitDAG();
1147 // Set the current basic block to the mbb we wish to insert the code into
1148 BB = SDB->JTCases[i].second.MBB;
1149 SDB->setCurrentBasicBlock(BB);
1151 SDB->visitJumpTable(SDB->JTCases[i].second);
1152 CurDAG->setRoot(SDB->getRoot());
1153 CodeGenAndEmitDAG();
1157 for (unsigned pi = 0, pe = SDB->PHINodesToUpdate.size(); pi != pe; ++pi) {
1158 MachineInstr *PHI = SDB->PHINodesToUpdate[pi].first;
1159 MachineBasicBlock *PHIBB = PHI->getParent();
1160 assert(PHI->isPHI() &&
1161 "This is not a machine PHI node that we are updating!");
1162 // "default" BB. We can go there only from header BB.
1163 if (PHIBB == SDB->JTCases[i].second.Default) {
1165 (MachineOperand::CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1167 (MachineOperand::CreateMBB(SDB->JTCases[i].first.HeaderBB));
1169 // JT BB. Just iterate over successors here
1170 if (BB->isSuccessor(PHIBB)) {
1172 (MachineOperand::CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1173 PHI->addOperand(MachineOperand::CreateMBB(BB));
1177 SDB->JTCases.clear();
1179 // If the switch block involved a branch to one of the actual successors, we
1180 // need to update PHI nodes in that block.
1181 for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) {
1182 MachineInstr *PHI = SDB->PHINodesToUpdate[i].first;
1183 assert(PHI->isPHI() &&
1184 "This is not a machine PHI node that we are updating!");
1185 if (BB->isSuccessor(PHI->getParent())) {
1186 PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second,
1188 PHI->addOperand(MachineOperand::CreateMBB(BB));
1192 // If we generated any switch lowering information, build and codegen any
1193 // additional DAGs necessary.
1194 for (unsigned i = 0, e = SDB->SwitchCases.size(); i != e; ++i) {
1195 // Set the current basic block to the mbb we wish to insert the code into
1196 MachineBasicBlock *ThisBB = BB = SDB->SwitchCases[i].ThisBB;
1197 SDB->setCurrentBasicBlock(BB);
1200 SDB->visitSwitchCase(SDB->SwitchCases[i]);
1201 CurDAG->setRoot(SDB->getRoot());
1202 CodeGenAndEmitDAG();
1204 // Handle any PHI nodes in successors of this chunk, as if we were coming
1205 // from the original BB before switch expansion. Note that PHI nodes can
1206 // occur multiple times in PHINodesToUpdate. We have to be very careful to
1207 // handle them the right number of times.
1208 while ((BB = SDB->SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
1209 // If new BB's are created during scheduling, the edges may have been
1210 // updated. That is, the edge from ThisBB to BB may have been split and
1211 // BB's predecessor is now another block.
1212 DenseMap<MachineBasicBlock*, MachineBasicBlock*>::iterator EI =
1213 SDB->EdgeMapping.find(BB);
1214 if (EI != SDB->EdgeMapping.end())
1215 ThisBB = EI->second;
1217 // BB may have been removed from the CFG if a branch was constant folded.
1218 if (ThisBB->isSuccessor(BB)) {
1219 for (MachineBasicBlock::iterator Phi = BB->begin();
1220 Phi != BB->end() && Phi->isPHI();
1222 // This value for this PHI node is recorded in PHINodesToUpdate.
1223 for (unsigned pn = 0; ; ++pn) {
1224 assert(pn != SDB->PHINodesToUpdate.size() &&
1225 "Didn't find PHI entry!");
1226 if (SDB->PHINodesToUpdate[pn].first == Phi) {
1227 Phi->addOperand(MachineOperand::
1228 CreateReg(SDB->PHINodesToUpdate[pn].second,
1230 Phi->addOperand(MachineOperand::CreateMBB(ThisBB));
1237 // Don't process RHS if same block as LHS.
1238 if (BB == SDB->SwitchCases[i].FalseBB)
1239 SDB->SwitchCases[i].FalseBB = 0;
1241 // If we haven't handled the RHS, do so now. Otherwise, we're done.
1242 SDB->SwitchCases[i].TrueBB = SDB->SwitchCases[i].FalseBB;
1243 SDB->SwitchCases[i].FalseBB = 0;
1245 assert(SDB->SwitchCases[i].TrueBB == 0 && SDB->SwitchCases[i].FalseBB == 0);
1248 SDB->SwitchCases.clear();
1250 SDB->PHINodesToUpdate.clear();
1254 /// Create the scheduler. If a specific scheduler was specified
1255 /// via the SchedulerRegistry, use it, otherwise select the
1256 /// one preferred by the target.
1258 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
1259 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1263 RegisterScheduler::setDefault(Ctor);
1266 return Ctor(this, OptLevel);
1269 ScheduleHazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
1270 return new ScheduleHazardRecognizer();
1273 //===----------------------------------------------------------------------===//
1274 // Helper functions used by the generated instruction selector.
1275 //===----------------------------------------------------------------------===//
1276 // Calls to these methods are generated by tblgen.
1278 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1279 /// the dag combiner simplified the 255, we still want to match. RHS is the
1280 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1281 /// specified in the .td file (e.g. 255).
1282 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1283 int64_t DesiredMaskS) const {
1284 const APInt &ActualMask = RHS->getAPIntValue();
1285 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1287 // If the actual mask exactly matches, success!
1288 if (ActualMask == DesiredMask)
1291 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1292 if (ActualMask.intersects(~DesiredMask))
1295 // Otherwise, the DAG Combiner may have proven that the value coming in is
1296 // either already zero or is not demanded. Check for known zero input bits.
1297 APInt NeededMask = DesiredMask & ~ActualMask;
1298 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1301 // TODO: check to see if missing bits are just not demanded.
1303 // Otherwise, this pattern doesn't match.
1307 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1308 /// the dag combiner simplified the 255, we still want to match. RHS is the
1309 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1310 /// specified in the .td file (e.g. 255).
1311 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1312 int64_t DesiredMaskS) const {
1313 const APInt &ActualMask = RHS->getAPIntValue();
1314 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1316 // If the actual mask exactly matches, success!
1317 if (ActualMask == DesiredMask)
1320 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1321 if (ActualMask.intersects(~DesiredMask))
1324 // Otherwise, the DAG Combiner may have proven that the value coming in is
1325 // either already zero or is not demanded. Check for known zero input bits.
1326 APInt NeededMask = DesiredMask & ~ActualMask;
1328 APInt KnownZero, KnownOne;
1329 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
1331 // If all the missing bits in the or are already known to be set, match!
1332 if ((NeededMask & KnownOne) == NeededMask)
1335 // TODO: check to see if missing bits are just not demanded.
1337 // Otherwise, this pattern doesn't match.
1342 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1343 /// by tblgen. Others should not call it.
1344 void SelectionDAGISel::
1345 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1346 std::vector<SDValue> InOps;
1347 std::swap(InOps, Ops);
1349 Ops.push_back(InOps[0]); // input chain.
1350 Ops.push_back(InOps[1]); // input asm string.
1352 unsigned i = 2, e = InOps.size();
1353 if (InOps[e-1].getValueType() == MVT::Flag)
1354 --e; // Don't process a flag operand if it is here.
1357 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1358 if ((Flags & 7) != 4 /*MEM*/) {
1359 // Just skip over this operand, copying the operands verbatim.
1360 Ops.insert(Ops.end(), InOps.begin()+i,
1361 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
1362 i += InlineAsm::getNumOperandRegisters(Flags) + 1;
1364 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 &&
1365 "Memory operand with multiple values?");
1366 // Otherwise, this is a memory operand. Ask the target to select it.
1367 std::vector<SDValue> SelOps;
1368 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps)) {
1369 llvm_report_error("Could not match memory address. Inline asm"
1373 // Add this to the output node.
1374 Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size()<< 3),
1376 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1381 // Add the flag input back if present.
1382 if (e != InOps.size())
1383 Ops.push_back(InOps.back());
1386 /// findFlagUse - Return use of EVT::Flag value produced by the specified
1389 static SDNode *findFlagUse(SDNode *N) {
1390 unsigned FlagResNo = N->getNumValues()-1;
1391 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1392 SDUse &Use = I.getUse();
1393 if (Use.getResNo() == FlagResNo)
1394 return Use.getUser();
1399 /// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
1400 /// This function recursively traverses up the operand chain, ignoring
1402 static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
1403 SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited,
1404 bool IgnoreChains) {
1405 // The NodeID's are given uniques ID's where a node ID is guaranteed to be
1406 // greater than all of its (recursive) operands. If we scan to a point where
1407 // 'use' is smaller than the node we're scanning for, then we know we will
1410 // The Use may be -1 (unassigned) if it is a newly allocated node. This can
1411 // happen because we scan down to newly selected nodes in the case of flag
1413 if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1))
1416 // Don't revisit nodes if we already scanned it and didn't fail, we know we
1417 // won't fail if we scan it again.
1418 if (!Visited.insert(Use))
1421 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
1422 // Ignore chain uses, they are validated by HandleMergeInputChains.
1423 if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains)
1426 SDNode *N = Use->getOperand(i).getNode();
1428 if (Use == ImmedUse || Use == Root)
1429 continue; // We are not looking for immediate use.
1434 // Traverse up the operand chain.
1435 if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
1441 /// IsProfitableToFold - Returns true if it's profitable to fold the specific
1442 /// operand node N of U during instruction selection that starts at Root.
1443 bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
1444 SDNode *Root) const {
1445 if (OptLevel == CodeGenOpt::None) return false;
1446 return N.hasOneUse();
1449 /// IsLegalToFold - Returns true if the specific operand node N of
1450 /// U can be folded during instruction selection that starts at Root.
1451 bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
1452 bool IgnoreChains) const {
1453 if (OptLevel == CodeGenOpt::None) return false;
1455 // If Root use can somehow reach N through a path that that doesn't contain
1456 // U then folding N would create a cycle. e.g. In the following
1457 // diagram, Root can reach N through X. If N is folded into into Root, then
1458 // X is both a predecessor and a successor of U.
1469 // * indicates nodes to be folded together.
1471 // If Root produces a flag, then it gets (even more) interesting. Since it
1472 // will be "glued" together with its flag use in the scheduler, we need to
1473 // check if it might reach N.
1492 // If FU (flag use) indirectly reaches N (the load), and Root folds N
1493 // (call it Fold), then X is a predecessor of FU and a successor of
1494 // Fold. But since Fold and FU are flagged together, this will create
1495 // a cycle in the scheduling graph.
1497 // If the node has flags, walk down the graph to the "lowest" node in the
1499 EVT VT = Root->getValueType(Root->getNumValues()-1);
1500 while (VT == MVT::Flag) {
1501 SDNode *FU = findFlagUse(Root);
1505 VT = Root->getValueType(Root->getNumValues()-1);
1507 // If our query node has a flag result with a use, we've walked up it. If
1508 // the user (which has already been selected) has a chain or indirectly uses
1509 // the chain, our WalkChainUsers predicate will not consider it. Because of
1510 // this, we cannot ignore chains in this predicate.
1511 IgnoreChains = false;
1515 SmallPtrSet<SDNode*, 16> Visited;
1516 return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains);
1519 SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
1520 std::vector<SDValue> Ops(N->op_begin(), N->op_end());
1521 SelectInlineAsmMemoryOperands(Ops);
1523 std::vector<EVT> VTs;
1524 VTs.push_back(MVT::Other);
1525 VTs.push_back(MVT::Flag);
1526 SDValue New = CurDAG->getNode(ISD::INLINEASM, N->getDebugLoc(),
1527 VTs, &Ops[0], Ops.size());
1529 return New.getNode();
1532 SDNode *SelectionDAGISel::Select_UNDEF(SDNode *N) {
1533 return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0));
1536 /// GetVBR - decode a vbr encoding whose top bit is set.
1537 ALWAYS_INLINE static uint64_t
1538 GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
1539 assert(Val >= 128 && "Not a VBR");
1540 Val &= 127; // Remove first vbr bit.
1545 NextBits = MatcherTable[Idx++];
1546 Val |= (NextBits&127) << Shift;
1548 } while (NextBits & 128);
1554 /// UpdateChainsAndFlags - When a match is complete, this method updates uses of
1555 /// interior flag and chain results to use the new flag and chain results.
1556 void SelectionDAGISel::
1557 UpdateChainsAndFlags(SDNode *NodeToMatch, SDValue InputChain,
1558 const SmallVectorImpl<SDNode*> &ChainNodesMatched,
1560 const SmallVectorImpl<SDNode*> &FlagResultNodesMatched,
1561 bool isMorphNodeTo) {
1562 SmallVector<SDNode*, 4> NowDeadNodes;
1564 ISelUpdater ISU(ISelPosition);
1566 // Now that all the normal results are replaced, we replace the chain and
1567 // flag results if present.
1568 if (!ChainNodesMatched.empty()) {
1569 assert(InputChain.getNode() != 0 &&
1570 "Matched input chains but didn't produce a chain");
1571 // Loop over all of the nodes we matched that produced a chain result.
1572 // Replace all the chain results with the final chain we ended up with.
1573 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1574 SDNode *ChainNode = ChainNodesMatched[i];
1576 // If this node was already deleted, don't look at it.
1577 if (ChainNode->getOpcode() == ISD::DELETED_NODE)
1580 // Don't replace the results of the root node if we're doing a
1582 if (ChainNode == NodeToMatch && isMorphNodeTo)
1585 SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1);
1586 if (ChainVal.getValueType() == MVT::Flag)
1587 ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
1588 assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
1589 CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain, &ISU);
1591 // If the node became dead, delete it.
1592 if (ChainNode->use_empty())
1593 NowDeadNodes.push_back(ChainNode);
1597 // If the result produces a flag, update any flag results in the matched
1598 // pattern with the flag result.
1599 if (InputFlag.getNode() != 0) {
1600 // Handle any interior nodes explicitly marked.
1601 for (unsigned i = 0, e = FlagResultNodesMatched.size(); i != e; ++i) {
1602 SDNode *FRN = FlagResultNodesMatched[i];
1604 // If this node was already deleted, don't look at it.
1605 if (FRN->getOpcode() == ISD::DELETED_NODE)
1608 assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Flag &&
1609 "Doesn't have a flag result");
1610 CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1),
1613 // If the node became dead, delete it.
1614 if (FRN->use_empty())
1615 NowDeadNodes.push_back(FRN);
1619 if (!NowDeadNodes.empty())
1620 CurDAG->RemoveDeadNodes(NowDeadNodes, &ISU);
1622 DEBUG(errs() << "ISEL: Match complete!\n");
1628 CR_LeadsToInteriorNode
1631 /// WalkChainUsers - Walk down the users of the specified chained node that is
1632 /// part of the pattern we're matching, looking at all of the users we find.
1633 /// This determines whether something is an interior node, whether we have a
1634 /// non-pattern node in between two pattern nodes (which prevent folding because
1635 /// it would induce a cycle) and whether we have a TokenFactor node sandwiched
1636 /// between pattern nodes (in which case the TF becomes part of the pattern).
1638 /// The walk we do here is guaranteed to be small because we quickly get down to
1639 /// already selected nodes "below" us.
1641 WalkChainUsers(SDNode *ChainedNode,
1642 SmallVectorImpl<SDNode*> &ChainedNodesInPattern,
1643 SmallVectorImpl<SDNode*> &InteriorChainedNodes) {
1644 ChainResult Result = CR_Simple;
1646 for (SDNode::use_iterator UI = ChainedNode->use_begin(),
1647 E = ChainedNode->use_end(); UI != E; ++UI) {
1648 // Make sure the use is of the chain, not some other value we produce.
1649 if (UI.getUse().getValueType() != MVT::Other) continue;
1653 // If we see an already-selected machine node, then we've gone beyond the
1654 // pattern that we're selecting down into the already selected chunk of the
1656 if (User->isMachineOpcode() ||
1657 User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
1660 if (User->getOpcode() == ISD::CopyToReg ||
1661 User->getOpcode() == ISD::CopyFromReg ||
1662 User->getOpcode() == ISD::INLINEASM ||
1663 User->getOpcode() == ISD::EH_LABEL) {
1664 // If their node ID got reset to -1 then they've already been selected.
1665 // Treat them like a MachineOpcode.
1666 if (User->getNodeId() == -1)
1670 // If we have a TokenFactor, we handle it specially.
1671 if (User->getOpcode() != ISD::TokenFactor) {
1672 // If the node isn't a token factor and isn't part of our pattern, then it
1673 // must be a random chained node in between two nodes we're selecting.
1674 // This happens when we have something like:
1679 // Because we structurally match the load/store as a read/modify/write,
1680 // but the call is chained between them. We cannot fold in this case
1681 // because it would induce a cycle in the graph.
1682 if (!std::count(ChainedNodesInPattern.begin(),
1683 ChainedNodesInPattern.end(), User))
1684 return CR_InducesCycle;
1686 // Otherwise we found a node that is part of our pattern. For example in:
1690 // This would happen when we're scanning down from the load and see the
1691 // store as a user. Record that there is a use of ChainedNode that is
1692 // part of the pattern and keep scanning uses.
1693 Result = CR_LeadsToInteriorNode;
1694 InteriorChainedNodes.push_back(User);
1698 // If we found a TokenFactor, there are two cases to consider: first if the
1699 // TokenFactor is just hanging "below" the pattern we're matching (i.e. no
1700 // uses of the TF are in our pattern) we just want to ignore it. Second,
1701 // the TokenFactor can be sandwiched in between two chained nodes, like so:
1707 // | \ DAG's like cheese
1710 // [TokenFactor] [Op]
1717 // In this case, the TokenFactor becomes part of our match and we rewrite it
1718 // as a new TokenFactor.
1720 // To distinguish these two cases, do a recursive walk down the uses.
1721 switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) {
1723 // If the uses of the TokenFactor are just already-selected nodes, ignore
1724 // it, it is "below" our pattern.
1726 case CR_InducesCycle:
1727 // If the uses of the TokenFactor lead to nodes that are not part of our
1728 // pattern that are not selected, folding would turn this into a cycle,
1730 return CR_InducesCycle;
1731 case CR_LeadsToInteriorNode:
1732 break; // Otherwise, keep processing.
1735 // Okay, we know we're in the interesting interior case. The TokenFactor
1736 // is now going to be considered part of the pattern so that we rewrite its
1737 // uses (it may have uses that are not part of the pattern) with the
1738 // ultimate chain result of the generated code. We will also add its chain
1739 // inputs as inputs to the ultimate TokenFactor we create.
1740 Result = CR_LeadsToInteriorNode;
1741 ChainedNodesInPattern.push_back(User);
1742 InteriorChainedNodes.push_back(User);
1749 /// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
1750 /// operation for when the pattern matched at least one node with a chains. The
1751 /// input vector contains a list of all of the chained nodes that we match. We
1752 /// must determine if this is a valid thing to cover (i.e. matching it won't
1753 /// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
1754 /// be used as the input node chain for the generated nodes.
1756 HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
1757 SelectionDAG *CurDAG) {
1758 // Walk all of the chained nodes we've matched, recursively scanning down the
1759 // users of the chain result. This adds any TokenFactor nodes that are caught
1760 // in between chained nodes to the chained and interior nodes list.
1761 SmallVector<SDNode*, 3> InteriorChainedNodes;
1762 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1763 if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched,
1764 InteriorChainedNodes) == CR_InducesCycle)
1765 return SDValue(); // Would induce a cycle.
1768 // Okay, we have walked all the matched nodes and collected TokenFactor nodes
1769 // that we are interested in. Form our input TokenFactor node.
1770 SmallVector<SDValue, 3> InputChains;
1771 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1772 // Add the input chain of this node to the InputChains list (which will be
1773 // the operands of the generated TokenFactor) if it's not an interior node.
1774 SDNode *N = ChainNodesMatched[i];
1775 if (N->getOpcode() != ISD::TokenFactor) {
1776 if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N))
1779 // Otherwise, add the input chain.
1780 SDValue InChain = ChainNodesMatched[i]->getOperand(0);
1781 assert(InChain.getValueType() == MVT::Other && "Not a chain");
1782 InputChains.push_back(InChain);
1786 // If we have a token factor, we want to add all inputs of the token factor
1787 // that are not part of the pattern we're matching.
1788 for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) {
1789 if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(),
1790 N->getOperand(op).getNode()))
1791 InputChains.push_back(N->getOperand(op));
1796 if (InputChains.size() == 1)
1797 return InputChains[0];
1798 return CurDAG->getNode(ISD::TokenFactor, ChainNodesMatched[0]->getDebugLoc(),
1799 MVT::Other, &InputChains[0], InputChains.size());
1802 /// MorphNode - Handle morphing a node in place for the selector.
1803 SDNode *SelectionDAGISel::
1804 MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList,
1805 const SDValue *Ops, unsigned NumOps, unsigned EmitNodeInfo) {
1806 // It is possible we're using MorphNodeTo to replace a node with no
1807 // normal results with one that has a normal result (or we could be
1808 // adding a chain) and the input could have flags and chains as well.
1809 // In this case we need to shifting the operands down.
1810 // FIXME: This is a horrible hack and broken in obscure cases, no worse
1811 // than the old isel though. We should sink this into MorphNodeTo.
1812 int OldFlagResultNo = -1, OldChainResultNo = -1;
1814 unsigned NTMNumResults = Node->getNumValues();
1815 if (Node->getValueType(NTMNumResults-1) == MVT::Flag) {
1816 OldFlagResultNo = NTMNumResults-1;
1817 if (NTMNumResults != 1 &&
1818 Node->getValueType(NTMNumResults-2) == MVT::Other)
1819 OldChainResultNo = NTMNumResults-2;
1820 } else if (Node->getValueType(NTMNumResults-1) == MVT::Other)
1821 OldChainResultNo = NTMNumResults-1;
1823 // Call the underlying SelectionDAG routine to do the transmogrification. Note
1824 // that this deletes operands of the old node that become dead.
1825 SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops, NumOps);
1827 // MorphNodeTo can operate in two ways: if an existing node with the
1828 // specified operands exists, it can just return it. Otherwise, it
1829 // updates the node in place to have the requested operands.
1831 // If we updated the node in place, reset the node ID. To the isel,
1832 // this should be just like a newly allocated machine node.
1836 unsigned ResNumResults = Res->getNumValues();
1837 // Move the flag if needed.
1838 if ((EmitNodeInfo & OPFL_FlagOutput) && OldFlagResultNo != -1 &&
1839 (unsigned)OldFlagResultNo != ResNumResults-1)
1840 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldFlagResultNo),
1841 SDValue(Res, ResNumResults-1));
1843 if ((EmitNodeInfo & OPFL_FlagOutput) != 0)
1846 // Move the chain reference if needed.
1847 if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 &&
1848 (unsigned)OldChainResultNo != ResNumResults-1)
1849 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo),
1850 SDValue(Res, ResNumResults-1));
1852 // Otherwise, no replacement happened because the node already exists. Replace
1853 // Uses of the old node with the new one.
1855 CurDAG->ReplaceAllUsesWith(Node, Res);
1860 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1861 ALWAYS_INLINE static bool
1862 CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1863 SDValue N, const SmallVectorImpl<SDValue> &RecordedNodes) {
1864 // Accept if it is exactly the same as a previously recorded node.
1865 unsigned RecNo = MatcherTable[MatcherIndex++];
1866 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
1867 return N == RecordedNodes[RecNo];
1870 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1871 ALWAYS_INLINE static bool
1872 CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1873 SelectionDAGISel &SDISel) {
1874 return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]);
1877 /// CheckNodePredicate - Implements OP_CheckNodePredicate.
1878 ALWAYS_INLINE static bool
1879 CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1880 SelectionDAGISel &SDISel, SDNode *N) {
1881 return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]);
1884 ALWAYS_INLINE static bool
1885 CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1887 return N->getOpcode() == MatcherTable[MatcherIndex++];
1890 ALWAYS_INLINE static bool
1891 CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1892 SDValue N, const TargetLowering &TLI) {
1893 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1894 if (N.getValueType() == VT) return true;
1896 // Handle the case when VT is iPTR.
1897 return VT == MVT::iPTR && N.getValueType() == TLI.getPointerTy();
1900 ALWAYS_INLINE static bool
1901 CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1902 SDValue N, const TargetLowering &TLI,
1904 if (ChildNo >= N.getNumOperands())
1905 return false; // Match fails if out of range child #.
1906 return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI);
1910 ALWAYS_INLINE static bool
1911 CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1913 return cast<CondCodeSDNode>(N)->get() ==
1914 (ISD::CondCode)MatcherTable[MatcherIndex++];
1917 ALWAYS_INLINE static bool
1918 CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1919 SDValue N, const TargetLowering &TLI) {
1920 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1921 if (cast<VTSDNode>(N)->getVT() == VT)
1924 // Handle the case when VT is iPTR.
1925 return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI.getPointerTy();
1928 ALWAYS_INLINE static bool
1929 CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1931 int64_t Val = MatcherTable[MatcherIndex++];
1933 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1935 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
1936 return C != 0 && C->getSExtValue() == Val;
1939 ALWAYS_INLINE static bool
1940 CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1941 SDValue N, SelectionDAGISel &SDISel) {
1942 int64_t Val = MatcherTable[MatcherIndex++];
1944 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1946 if (N->getOpcode() != ISD::AND) return false;
1948 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1949 return C != 0 && SDISel.CheckAndMask(N.getOperand(0), C, Val);
1952 ALWAYS_INLINE static bool
1953 CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1954 SDValue N, SelectionDAGISel &SDISel) {
1955 int64_t Val = MatcherTable[MatcherIndex++];
1957 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1959 if (N->getOpcode() != ISD::OR) return false;
1961 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1962 return C != 0 && SDISel.CheckOrMask(N.getOperand(0), C, Val);
1965 /// IsPredicateKnownToFail - If we know how and can do so without pushing a
1966 /// scope, evaluate the current node. If the current predicate is known to
1967 /// fail, set Result=true and return anything. If the current predicate is
1968 /// known to pass, set Result=false and return the MatcherIndex to continue
1969 /// with. If the current predicate is unknown, set Result=false and return the
1970 /// MatcherIndex to continue with.
1971 static unsigned IsPredicateKnownToFail(const unsigned char *Table,
1972 unsigned Index, SDValue N,
1973 bool &Result, SelectionDAGISel &SDISel,
1974 SmallVectorImpl<SDValue> &RecordedNodes){
1975 switch (Table[Index++]) {
1978 return Index-1; // Could not evaluate this predicate.
1979 case SelectionDAGISel::OPC_CheckSame:
1980 Result = !::CheckSame(Table, Index, N, RecordedNodes);
1982 case SelectionDAGISel::OPC_CheckPatternPredicate:
1983 Result = !::CheckPatternPredicate(Table, Index, SDISel);
1985 case SelectionDAGISel::OPC_CheckPredicate:
1986 Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode());
1988 case SelectionDAGISel::OPC_CheckOpcode:
1989 Result = !::CheckOpcode(Table, Index, N.getNode());
1991 case SelectionDAGISel::OPC_CheckType:
1992 Result = !::CheckType(Table, Index, N, SDISel.TLI);
1994 case SelectionDAGISel::OPC_CheckChild0Type:
1995 case SelectionDAGISel::OPC_CheckChild1Type:
1996 case SelectionDAGISel::OPC_CheckChild2Type:
1997 case SelectionDAGISel::OPC_CheckChild3Type:
1998 case SelectionDAGISel::OPC_CheckChild4Type:
1999 case SelectionDAGISel::OPC_CheckChild5Type:
2000 case SelectionDAGISel::OPC_CheckChild6Type:
2001 case SelectionDAGISel::OPC_CheckChild7Type:
2002 Result = !::CheckChildType(Table, Index, N, SDISel.TLI,
2003 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Type);
2005 case SelectionDAGISel::OPC_CheckCondCode:
2006 Result = !::CheckCondCode(Table, Index, N);
2008 case SelectionDAGISel::OPC_CheckValueType:
2009 Result = !::CheckValueType(Table, Index, N, SDISel.TLI);
2011 case SelectionDAGISel::OPC_CheckInteger:
2012 Result = !::CheckInteger(Table, Index, N);
2014 case SelectionDAGISel::OPC_CheckAndImm:
2015 Result = !::CheckAndImm(Table, Index, N, SDISel);
2017 case SelectionDAGISel::OPC_CheckOrImm:
2018 Result = !::CheckOrImm(Table, Index, N, SDISel);
2025 /// FailIndex - If this match fails, this is the index to continue with.
2028 /// NodeStack - The node stack when the scope was formed.
2029 SmallVector<SDValue, 4> NodeStack;
2031 /// NumRecordedNodes - The number of recorded nodes when the scope was formed.
2032 unsigned NumRecordedNodes;
2034 /// NumMatchedMemRefs - The number of matched memref entries.
2035 unsigned NumMatchedMemRefs;
2037 /// InputChain/InputFlag - The current chain/flag
2038 SDValue InputChain, InputFlag;
2040 /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty.
2041 bool HasChainNodesMatched, HasFlagResultNodesMatched;
2044 SDNode *SelectionDAGISel::
2045 SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
2046 unsigned TableSize) {
2047 // FIXME: Should these even be selected? Handle these cases in the caller?
2048 switch (NodeToMatch->getOpcode()) {
2051 case ISD::EntryToken: // These nodes remain the same.
2052 case ISD::BasicBlock:
2054 //case ISD::VALUETYPE:
2055 //case ISD::CONDCODE:
2056 case ISD::HANDLENODE:
2057 case ISD::TargetConstant:
2058 case ISD::TargetConstantFP:
2059 case ISD::TargetConstantPool:
2060 case ISD::TargetFrameIndex:
2061 case ISD::TargetExternalSymbol:
2062 case ISD::TargetBlockAddress:
2063 case ISD::TargetJumpTable:
2064 case ISD::TargetGlobalTLSAddress:
2065 case ISD::TargetGlobalAddress:
2066 case ISD::TokenFactor:
2067 case ISD::CopyFromReg:
2068 case ISD::CopyToReg:
2070 NodeToMatch->setNodeId(-1); // Mark selected.
2072 case ISD::AssertSext:
2073 case ISD::AssertZext:
2074 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0),
2075 NodeToMatch->getOperand(0));
2077 case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch);
2078 case ISD::UNDEF: return Select_UNDEF(NodeToMatch);
2081 assert(!NodeToMatch->isMachineOpcode() && "Node already selected!");
2083 // Set up the node stack with NodeToMatch as the only node on the stack.
2084 SmallVector<SDValue, 8> NodeStack;
2085 SDValue N = SDValue(NodeToMatch, 0);
2086 NodeStack.push_back(N);
2088 // MatchScopes - Scopes used when matching, if a match failure happens, this
2089 // indicates where to continue checking.
2090 SmallVector<MatchScope, 8> MatchScopes;
2092 // RecordedNodes - This is the set of nodes that have been recorded by the
2094 SmallVector<SDValue, 8> RecordedNodes;
2096 // MatchedMemRefs - This is the set of MemRef's we've seen in the input
2098 SmallVector<MachineMemOperand*, 2> MatchedMemRefs;
2100 // These are the current input chain and flag for use when generating nodes.
2101 // Various Emit operations change these. For example, emitting a copytoreg
2102 // uses and updates these.
2103 SDValue InputChain, InputFlag;
2105 // ChainNodesMatched - If a pattern matches nodes that have input/output
2106 // chains, the OPC_EmitMergeInputChains operation is emitted which indicates
2107 // which ones they are. The result is captured into this list so that we can
2108 // update the chain results when the pattern is complete.
2109 SmallVector<SDNode*, 3> ChainNodesMatched;
2110 SmallVector<SDNode*, 3> FlagResultNodesMatched;
2112 DEBUG(errs() << "ISEL: Starting pattern match on root node: ";
2113 NodeToMatch->dump(CurDAG);
2116 // Determine where to start the interpreter. Normally we start at opcode #0,
2117 // but if the state machine starts with an OPC_SwitchOpcode, then we
2118 // accelerate the first lookup (which is guaranteed to be hot) with the
2119 // OpcodeOffset table.
2120 unsigned MatcherIndex = 0;
2122 if (!OpcodeOffset.empty()) {
2123 // Already computed the OpcodeOffset table, just index into it.
2124 if (N.getOpcode() < OpcodeOffset.size())
2125 MatcherIndex = OpcodeOffset[N.getOpcode()];
2126 DEBUG(errs() << " Initial Opcode index to " << MatcherIndex << "\n");
2128 } else if (MatcherTable[0] == OPC_SwitchOpcode) {
2129 // Otherwise, the table isn't computed, but the state machine does start
2130 // with an OPC_SwitchOpcode instruction. Populate the table now, since this
2131 // is the first time we're selecting an instruction.
2134 // Get the size of this case.
2135 unsigned CaseSize = MatcherTable[Idx++];
2137 CaseSize = GetVBR(CaseSize, MatcherTable, Idx);
2138 if (CaseSize == 0) break;
2140 // Get the opcode, add the index to the table.
2141 unsigned Opc = MatcherTable[Idx++];
2142 if (Opc >= OpcodeOffset.size())
2143 OpcodeOffset.resize((Opc+1)*2);
2144 OpcodeOffset[Opc] = Idx;
2148 // Okay, do the lookup for the first opcode.
2149 if (N.getOpcode() < OpcodeOffset.size())
2150 MatcherIndex = OpcodeOffset[N.getOpcode()];
2154 assert(MatcherIndex < TableSize && "Invalid index");
2156 unsigned CurrentOpcodeIndex = MatcherIndex;
2158 BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++];
2161 // Okay, the semantics of this operation are that we should push a scope
2162 // then evaluate the first child. However, pushing a scope only to have
2163 // the first check fail (which then pops it) is inefficient. If we can
2164 // determine immediately that the first check (or first several) will
2165 // immediately fail, don't even bother pushing a scope for them.
2169 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2170 if (NumToSkip & 128)
2171 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2172 // Found the end of the scope with no match.
2173 if (NumToSkip == 0) {
2178 FailIndex = MatcherIndex+NumToSkip;
2180 // If we can't evaluate this predicate without pushing a scope (e.g. if
2181 // it is a 'MoveParent') or if the predicate succeeds on this node, we
2182 // push the scope and evaluate the full predicate chain.
2184 MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
2185 Result, *this, RecordedNodes);
2189 DEBUG(errs() << " Skipped scope entry at index " << MatcherIndex
2190 << " continuing at " << FailIndex << "\n");
2193 // Otherwise, we know that this case of the Scope is guaranteed to fail,
2194 // move to the next case.
2195 MatcherIndex = FailIndex;
2198 // If the whole scope failed to match, bail.
2199 if (FailIndex == 0) break;
2201 // Push a MatchScope which indicates where to go if the first child fails
2203 MatchScope NewEntry;
2204 NewEntry.FailIndex = FailIndex;
2205 NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
2206 NewEntry.NumRecordedNodes = RecordedNodes.size();
2207 NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
2208 NewEntry.InputChain = InputChain;
2209 NewEntry.InputFlag = InputFlag;
2210 NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty();
2211 NewEntry.HasFlagResultNodesMatched = !FlagResultNodesMatched.empty();
2212 MatchScopes.push_back(NewEntry);
2215 case OPC_RecordNode:
2216 // Remember this node, it may end up being an operand in the pattern.
2217 RecordedNodes.push_back(N);
2220 case OPC_RecordChild0: case OPC_RecordChild1:
2221 case OPC_RecordChild2: case OPC_RecordChild3:
2222 case OPC_RecordChild4: case OPC_RecordChild5:
2223 case OPC_RecordChild6: case OPC_RecordChild7: {
2224 unsigned ChildNo = Opcode-OPC_RecordChild0;
2225 if (ChildNo >= N.getNumOperands())
2226 break; // Match fails if out of range child #.
2228 RecordedNodes.push_back(N->getOperand(ChildNo));
2231 case OPC_RecordMemRef:
2232 MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand());
2235 case OPC_CaptureFlagInput:
2236 // If the current node has an input flag, capture it in InputFlag.
2237 if (N->getNumOperands() != 0 &&
2238 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag)
2239 InputFlag = N->getOperand(N->getNumOperands()-1);
2242 case OPC_MoveChild: {
2243 unsigned ChildNo = MatcherTable[MatcherIndex++];
2244 if (ChildNo >= N.getNumOperands())
2245 break; // Match fails if out of range child #.
2246 N = N.getOperand(ChildNo);
2247 NodeStack.push_back(N);
2251 case OPC_MoveParent:
2252 // Pop the current node off the NodeStack.
2253 NodeStack.pop_back();
2254 assert(!NodeStack.empty() && "Node stack imbalance!");
2255 N = NodeStack.back();
2259 if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
2261 case OPC_CheckPatternPredicate:
2262 if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break;
2264 case OPC_CheckPredicate:
2265 if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this,
2269 case OPC_CheckComplexPat: {
2270 unsigned CPNum = MatcherTable[MatcherIndex++];
2271 unsigned RecNo = MatcherTable[MatcherIndex++];
2272 assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
2273 if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo], CPNum,
2278 case OPC_CheckOpcode:
2279 if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
2283 if (!::CheckType(MatcherTable, MatcherIndex, N, TLI)) break;
2286 case OPC_SwitchOpcode: {
2287 unsigned CurNodeOpcode = N.getOpcode();
2288 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2291 // Get the size of this case.
2292 CaseSize = MatcherTable[MatcherIndex++];
2294 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2295 if (CaseSize == 0) break;
2297 // If the opcode matches, then we will execute this case.
2298 if (CurNodeOpcode == MatcherTable[MatcherIndex++])
2301 // Otherwise, skip over this case.
2302 MatcherIndex += CaseSize;
2305 // If no cases matched, bail out.
2306 if (CaseSize == 0) break;
2308 // Otherwise, execute the case we found.
2309 DEBUG(errs() << " OpcodeSwitch from " << SwitchStart
2310 << " to " << MatcherIndex << "\n");
2314 case OPC_SwitchType: {
2315 MVT::SimpleValueType CurNodeVT = N.getValueType().getSimpleVT().SimpleTy;
2316 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2319 // Get the size of this case.
2320 CaseSize = MatcherTable[MatcherIndex++];
2322 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2323 if (CaseSize == 0) break;
2325 MVT::SimpleValueType CaseVT =
2326 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2327 if (CaseVT == MVT::iPTR)
2328 CaseVT = TLI.getPointerTy().SimpleTy;
2330 // If the VT matches, then we will execute this case.
2331 if (CurNodeVT == CaseVT)
2334 // Otherwise, skip over this case.
2335 MatcherIndex += CaseSize;
2338 // If no cases matched, bail out.
2339 if (CaseSize == 0) break;
2341 // Otherwise, execute the case we found.
2342 DEBUG(errs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
2343 << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
2346 case OPC_CheckChild0Type: case OPC_CheckChild1Type:
2347 case OPC_CheckChild2Type: case OPC_CheckChild3Type:
2348 case OPC_CheckChild4Type: case OPC_CheckChild5Type:
2349 case OPC_CheckChild6Type: case OPC_CheckChild7Type:
2350 if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI,
2351 Opcode-OPC_CheckChild0Type))
2354 case OPC_CheckCondCode:
2355 if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
2357 case OPC_CheckValueType:
2358 if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI)) break;
2360 case OPC_CheckInteger:
2361 if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
2363 case OPC_CheckAndImm:
2364 if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
2366 case OPC_CheckOrImm:
2367 if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
2370 case OPC_CheckFoldableChainNode: {
2371 assert(NodeStack.size() != 1 && "No parent node");
2372 // Verify that all intermediate nodes between the root and this one have
2374 bool HasMultipleUses = false;
2375 for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i)
2376 if (!NodeStack[i].hasOneUse()) {
2377 HasMultipleUses = true;
2380 if (HasMultipleUses) break;
2382 // Check to see that the target thinks this is profitable to fold and that
2383 // we can fold it without inducing cycles in the graph.
2384 if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2386 !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2387 NodeToMatch, true/*We validate our own chains*/))
2392 case OPC_EmitInteger: {
2393 MVT::SimpleValueType VT =
2394 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2395 int64_t Val = MatcherTable[MatcherIndex++];
2397 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2398 RecordedNodes.push_back(CurDAG->getTargetConstant(Val, VT));
2401 case OPC_EmitRegister: {
2402 MVT::SimpleValueType VT =
2403 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2404 unsigned RegNo = MatcherTable[MatcherIndex++];
2405 RecordedNodes.push_back(CurDAG->getRegister(RegNo, VT));
2409 case OPC_EmitConvertToTarget: {
2410 // Convert from IMM/FPIMM to target version.
2411 unsigned RecNo = MatcherTable[MatcherIndex++];
2412 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2413 SDValue Imm = RecordedNodes[RecNo];
2415 if (Imm->getOpcode() == ISD::Constant) {
2416 int64_t Val = cast<ConstantSDNode>(Imm)->getZExtValue();
2417 Imm = CurDAG->getTargetConstant(Val, Imm.getValueType());
2418 } else if (Imm->getOpcode() == ISD::ConstantFP) {
2419 const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue();
2420 Imm = CurDAG->getTargetConstantFP(*Val, Imm.getValueType());
2423 RecordedNodes.push_back(Imm);
2427 case OPC_EmitMergeInputChains: {
2428 assert(InputChain.getNode() == 0 &&
2429 "EmitMergeInputChains should be the first chain producing node");
2430 // This node gets a list of nodes we matched in the input that have
2431 // chains. We want to token factor all of the input chains to these nodes
2432 // together. However, if any of the input chains is actually one of the
2433 // nodes matched in this pattern, then we have an intra-match reference.
2434 // Ignore these because the newly token factored chain should not refer to
2436 unsigned NumChains = MatcherTable[MatcherIndex++];
2437 assert(NumChains != 0 && "Can't TF zero chains");
2439 assert(ChainNodesMatched.empty() &&
2440 "Should only have one EmitMergeInputChains per match");
2442 // Read all of the chained nodes.
2443 for (unsigned i = 0; i != NumChains; ++i) {
2444 unsigned RecNo = MatcherTable[MatcherIndex++];
2445 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2446 ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode());
2448 // FIXME: What if other value results of the node have uses not matched
2450 if (ChainNodesMatched.back() != NodeToMatch &&
2451 !RecordedNodes[RecNo].hasOneUse()) {
2452 ChainNodesMatched.clear();
2457 // If the inner loop broke out, the match fails.
2458 if (ChainNodesMatched.empty())
2461 // Merge the input chains if they are not intra-pattern references.
2462 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2464 if (InputChain.getNode() == 0)
2465 break; // Failed to merge.
2470 case OPC_EmitCopyToReg: {
2471 unsigned RecNo = MatcherTable[MatcherIndex++];
2472 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2473 unsigned DestPhysReg = MatcherTable[MatcherIndex++];
2475 if (InputChain.getNode() == 0)
2476 InputChain = CurDAG->getEntryNode();
2478 InputChain = CurDAG->getCopyToReg(InputChain, NodeToMatch->getDebugLoc(),
2479 DestPhysReg, RecordedNodes[RecNo],
2482 InputFlag = InputChain.getValue(1);
2486 case OPC_EmitNodeXForm: {
2487 unsigned XFormNo = MatcherTable[MatcherIndex++];
2488 unsigned RecNo = MatcherTable[MatcherIndex++];
2489 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2490 RecordedNodes.push_back(RunSDNodeXForm(RecordedNodes[RecNo], XFormNo));
2495 case OPC_MorphNodeTo: {
2496 uint16_t TargetOpc = MatcherTable[MatcherIndex++];
2497 TargetOpc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2498 unsigned EmitNodeInfo = MatcherTable[MatcherIndex++];
2499 // Get the result VT list.
2500 unsigned NumVTs = MatcherTable[MatcherIndex++];
2501 SmallVector<EVT, 4> VTs;
2502 for (unsigned i = 0; i != NumVTs; ++i) {
2503 MVT::SimpleValueType VT =
2504 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2505 if (VT == MVT::iPTR) VT = TLI.getPointerTy().SimpleTy;
2509 if (EmitNodeInfo & OPFL_Chain)
2510 VTs.push_back(MVT::Other);
2511 if (EmitNodeInfo & OPFL_FlagOutput)
2512 VTs.push_back(MVT::Flag);
2514 // This is hot code, so optimize the two most common cases of 1 and 2
2517 if (VTs.size() == 1)
2518 VTList = CurDAG->getVTList(VTs[0]);
2519 else if (VTs.size() == 2)
2520 VTList = CurDAG->getVTList(VTs[0], VTs[1]);
2522 VTList = CurDAG->getVTList(VTs.data(), VTs.size());
2524 // Get the operand list.
2525 unsigned NumOps = MatcherTable[MatcherIndex++];
2526 SmallVector<SDValue, 8> Ops;
2527 for (unsigned i = 0; i != NumOps; ++i) {
2528 unsigned RecNo = MatcherTable[MatcherIndex++];
2530 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2532 assert(RecNo < RecordedNodes.size() && "Invalid EmitNode");
2533 Ops.push_back(RecordedNodes[RecNo]);
2536 // If there are variadic operands to add, handle them now.
2537 if (EmitNodeInfo & OPFL_VariadicInfo) {
2538 // Determine the start index to copy from.
2539 unsigned FirstOpToCopy = getNumFixedFromVariadicInfo(EmitNodeInfo);
2540 FirstOpToCopy += (EmitNodeInfo & OPFL_Chain) ? 1 : 0;
2541 assert(NodeToMatch->getNumOperands() >= FirstOpToCopy &&
2542 "Invalid variadic node");
2543 // Copy all of the variadic operands, not including a potential flag
2545 for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands();
2547 SDValue V = NodeToMatch->getOperand(i);
2548 if (V.getValueType() == MVT::Flag) break;
2553 // If this has chain/flag inputs, add them.
2554 if (EmitNodeInfo & OPFL_Chain)
2555 Ops.push_back(InputChain);
2556 if ((EmitNodeInfo & OPFL_FlagInput) && InputFlag.getNode() != 0)
2557 Ops.push_back(InputFlag);
2561 if (Opcode != OPC_MorphNodeTo) {
2562 // If this is a normal EmitNode command, just create the new node and
2563 // add the results to the RecordedNodes list.
2564 Res = CurDAG->getMachineNode(TargetOpc, NodeToMatch->getDebugLoc(),
2565 VTList, Ops.data(), Ops.size());
2567 // Add all the non-flag/non-chain results to the RecordedNodes list.
2568 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
2569 if (VTs[i] == MVT::Other || VTs[i] == MVT::Flag) break;
2570 RecordedNodes.push_back(SDValue(Res, i));
2574 Res = MorphNode(NodeToMatch, TargetOpc, VTList, Ops.data(), Ops.size(),
2578 // If the node had chain/flag results, update our notion of the current
2580 if (EmitNodeInfo & OPFL_FlagOutput) {
2581 InputFlag = SDValue(Res, VTs.size()-1);
2582 if (EmitNodeInfo & OPFL_Chain)
2583 InputChain = SDValue(Res, VTs.size()-2);
2584 } else if (EmitNodeInfo & OPFL_Chain)
2585 InputChain = SDValue(Res, VTs.size()-1);
2587 // If the OPFL_MemRefs flag is set on this node, slap all of the
2588 // accumulated memrefs onto it.
2590 // FIXME: This is vastly incorrect for patterns with multiple outputs
2591 // instructions that access memory and for ComplexPatterns that match
2593 if (EmitNodeInfo & OPFL_MemRefs) {
2594 MachineSDNode::mmo_iterator MemRefs =
2595 MF->allocateMemRefsArray(MatchedMemRefs.size());
2596 std::copy(MatchedMemRefs.begin(), MatchedMemRefs.end(), MemRefs);
2597 cast<MachineSDNode>(Res)
2598 ->setMemRefs(MemRefs, MemRefs + MatchedMemRefs.size());
2602 << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created")
2603 << " node: "; Res->dump(CurDAG); errs() << "\n");
2605 // If this was a MorphNodeTo then we're completely done!
2606 if (Opcode == OPC_MorphNodeTo) {
2607 // Update chain and flag uses.
2608 UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched,
2609 InputFlag, FlagResultNodesMatched, true);
2616 case OPC_MarkFlagResults: {
2617 unsigned NumNodes = MatcherTable[MatcherIndex++];
2619 // Read and remember all the flag-result nodes.
2620 for (unsigned i = 0; i != NumNodes; ++i) {
2621 unsigned RecNo = MatcherTable[MatcherIndex++];
2623 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2625 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2626 FlagResultNodesMatched.push_back(RecordedNodes[RecNo].getNode());
2631 case OPC_CompleteMatch: {
2632 // The match has been completed, and any new nodes (if any) have been
2633 // created. Patch up references to the matched dag to use the newly
2635 unsigned NumResults = MatcherTable[MatcherIndex++];
2637 for (unsigned i = 0; i != NumResults; ++i) {
2638 unsigned ResSlot = MatcherTable[MatcherIndex++];
2640 ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex);
2642 assert(ResSlot < RecordedNodes.size() && "Invalid CheckSame");
2643 SDValue Res = RecordedNodes[ResSlot];
2645 // FIXME2: Eliminate this horrible hack by fixing the 'Gen' program
2646 // after (parallel) on input patterns are removed. This would also
2647 // allow us to stop encoding #results in OPC_CompleteMatch's table
2649 if (NodeToMatch->getNumValues() <= i ||
2650 NodeToMatch->getValueType(i) == MVT::Other ||
2651 NodeToMatch->getValueType(i) == MVT::Flag)
2653 assert((NodeToMatch->getValueType(i) == Res.getValueType() ||
2654 NodeToMatch->getValueType(i) == MVT::iPTR ||
2655 Res.getValueType() == MVT::iPTR ||
2656 NodeToMatch->getValueType(i).getSizeInBits() ==
2657 Res.getValueType().getSizeInBits()) &&
2658 "invalid replacement");
2659 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, i), Res);
2662 // If the root node defines a flag, add it to the flag nodes to update
2664 if (NodeToMatch->getValueType(NodeToMatch->getNumValues()-1) == MVT::Flag)
2665 FlagResultNodesMatched.push_back(NodeToMatch);
2667 // Update chain and flag uses.
2668 UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched,
2669 InputFlag, FlagResultNodesMatched, false);
2671 assert(NodeToMatch->use_empty() &&
2672 "Didn't replace all uses of the node?");
2674 // FIXME: We just return here, which interacts correctly with SelectRoot
2675 // above. We should fix this to not return an SDNode* anymore.
2680 // If the code reached this point, then the match failed. See if there is
2681 // another child to try in the current 'Scope', otherwise pop it until we
2682 // find a case to check.
2683 DEBUG(errs() << " Match failed at index " << CurrentOpcodeIndex << "\n");
2685 if (MatchScopes.empty()) {
2686 CannotYetSelect(NodeToMatch);
2690 // Restore the interpreter state back to the point where the scope was
2692 MatchScope &LastScope = MatchScopes.back();
2693 RecordedNodes.resize(LastScope.NumRecordedNodes);
2695 NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end());
2696 N = NodeStack.back();
2698 if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size())
2699 MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
2700 MatcherIndex = LastScope.FailIndex;
2702 DEBUG(errs() << " Continuing at " << MatcherIndex << "\n");
2704 InputChain = LastScope.InputChain;
2705 InputFlag = LastScope.InputFlag;
2706 if (!LastScope.HasChainNodesMatched)
2707 ChainNodesMatched.clear();
2708 if (!LastScope.HasFlagResultNodesMatched)
2709 FlagResultNodesMatched.clear();
2711 // Check to see what the offset is at the new MatcherIndex. If it is zero
2712 // we have reached the end of this scope, otherwise we have another child
2713 // in the current scope to try.
2714 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2715 if (NumToSkip & 128)
2716 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2718 // If we have another child in this scope to match, update FailIndex and
2720 if (NumToSkip != 0) {
2721 LastScope.FailIndex = MatcherIndex+NumToSkip;
2725 // End of this scope, pop it and try the next child in the containing
2727 MatchScopes.pop_back();
2734 void SelectionDAGISel::CannotYetSelect(SDNode *N) {
2736 raw_string_ostream Msg(msg);
2737 Msg << "Cannot yet select: ";
2739 if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN &&
2740 N->getOpcode() != ISD::INTRINSIC_WO_CHAIN &&
2741 N->getOpcode() != ISD::INTRINSIC_VOID) {
2742 N->printrFull(Msg, CurDAG);
2744 bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other;
2746 cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue();
2747 if (iid < Intrinsic::num_intrinsics)
2748 Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid);
2749 else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo())
2750 Msg << "target intrinsic %" << TII->getName(iid);
2752 Msg << "unknown intrinsic #" << iid;
2754 llvm_report_error(Msg.str());
2757 char SelectionDAGISel::ID = 0;