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");
64 STATISTIC(NumDAGIselRetries,"Number of times dag isel has to try another path");
67 EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
68 cl::desc("Enable verbose messages in the \"fast\" "
69 "instruction selector"));
71 EnableFastISelAbort("fast-isel-abort", cl::Hidden,
72 cl::desc("Enable abort calls when \"fast\" instruction fails"));
74 SchedLiveInCopies("schedule-livein-copies", cl::Hidden,
75 cl::desc("Schedule copies of livein registers"),
80 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
81 cl::desc("Pop up a window to show dags before the first "
84 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
85 cl::desc("Pop up a window to show dags before legalize types"));
87 ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
88 cl::desc("Pop up a window to show dags before legalize"));
90 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
91 cl::desc("Pop up a window to show dags before the second "
94 ViewDAGCombineLT("view-dag-combine-lt-dags", cl::Hidden,
95 cl::desc("Pop up a window to show dags before the post legalize types"
96 " dag combine pass"));
98 ViewISelDAGs("view-isel-dags", cl::Hidden,
99 cl::desc("Pop up a window to show isel dags as they are selected"));
101 ViewSchedDAGs("view-sched-dags", cl::Hidden,
102 cl::desc("Pop up a window to show sched dags as they are processed"));
104 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
105 cl::desc("Pop up a window to show SUnit dags after they are processed"));
107 static const bool ViewDAGCombine1 = false,
108 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false,
109 ViewDAGCombine2 = false,
110 ViewDAGCombineLT = false,
111 ViewISelDAGs = false, ViewSchedDAGs = false,
112 ViewSUnitDAGs = false;
115 //===---------------------------------------------------------------------===//
117 /// RegisterScheduler class - Track the registration of instruction schedulers.
119 //===---------------------------------------------------------------------===//
120 MachinePassRegistry RegisterScheduler::Registry;
122 //===---------------------------------------------------------------------===//
124 /// ISHeuristic command line option for instruction schedulers.
126 //===---------------------------------------------------------------------===//
127 static cl::opt<RegisterScheduler::FunctionPassCtor, false,
128 RegisterPassParser<RegisterScheduler> >
129 ISHeuristic("pre-RA-sched",
130 cl::init(&createDefaultScheduler),
131 cl::desc("Instruction schedulers available (before register"
134 static RegisterScheduler
135 defaultListDAGScheduler("default", "Best scheduler for the target",
136 createDefaultScheduler);
139 //===--------------------------------------------------------------------===//
140 /// createDefaultScheduler - This creates an instruction scheduler appropriate
142 ScheduleDAGSDNodes* createDefaultScheduler(SelectionDAGISel *IS,
143 CodeGenOpt::Level OptLevel) {
144 const TargetLowering &TLI = IS->getTargetLowering();
146 if (OptLevel == CodeGenOpt::None)
147 return createFastDAGScheduler(IS, OptLevel);
148 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency)
149 return createTDListDAGScheduler(IS, OptLevel);
150 assert(TLI.getSchedulingPreference() ==
151 TargetLowering::SchedulingForRegPressure && "Unknown sched type!");
152 return createBURRListDAGScheduler(IS, OptLevel);
156 // EmitInstrWithCustomInserter - This method should be implemented by targets
157 // that mark instructions with the 'usesCustomInserter' flag. These
158 // instructions are special in various ways, which require special support to
159 // insert. The specified MachineInstr is created but not inserted into any
160 // basic blocks, and this method is called to expand it into a sequence of
161 // instructions, potentially also creating new basic blocks and control flow.
162 // When new basic blocks are inserted and the edges from MBB to its successors
163 // are modified, the method should insert pairs of <OldSucc, NewSucc> into the
165 MachineBasicBlock *TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
166 MachineBasicBlock *MBB,
167 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
169 dbgs() << "If a target marks an instruction with "
170 "'usesCustomInserter', it must implement "
171 "TargetLowering::EmitInstrWithCustomInserter!";
177 /// EmitLiveInCopy - Emit a copy for a live in physical register. If the
178 /// physical register has only a single copy use, then coalesced the copy
180 static void EmitLiveInCopy(MachineBasicBlock *MBB,
181 MachineBasicBlock::iterator &InsertPos,
182 unsigned VirtReg, unsigned PhysReg,
183 const TargetRegisterClass *RC,
184 DenseMap<MachineInstr*, unsigned> &CopyRegMap,
185 const MachineRegisterInfo &MRI,
186 const TargetRegisterInfo &TRI,
187 const TargetInstrInfo &TII) {
188 unsigned NumUses = 0;
189 MachineInstr *UseMI = NULL;
190 for (MachineRegisterInfo::use_iterator UI = MRI.use_begin(VirtReg),
191 UE = MRI.use_end(); UI != UE; ++UI) {
197 // If the number of uses is not one, or the use is not a move instruction,
198 // don't coalesce. Also, only coalesce away a virtual register to virtual
200 bool Coalesced = false;
201 unsigned SrcReg, DstReg, SrcSubReg, DstSubReg;
203 TII.isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubReg, DstSubReg) &&
204 TargetRegisterInfo::isVirtualRegister(DstReg)) {
209 // Now find an ideal location to insert the copy.
210 MachineBasicBlock::iterator Pos = InsertPos;
211 while (Pos != MBB->begin()) {
212 MachineInstr *PrevMI = prior(Pos);
213 DenseMap<MachineInstr*, unsigned>::iterator RI = CopyRegMap.find(PrevMI);
214 // copyRegToReg might emit multiple instructions to do a copy.
215 unsigned CopyDstReg = (RI == CopyRegMap.end()) ? 0 : RI->second;
216 if (CopyDstReg && !TRI.regsOverlap(CopyDstReg, PhysReg))
217 // This is what the BB looks like right now:
222 // We want to insert "r1025 = mov r1". Inserting this copy below the
223 // move to r1024 makes it impossible for that move to be coalesced.
230 break; // Woot! Found a good location.
234 bool Emitted = TII.copyRegToReg(*MBB, Pos, VirtReg, PhysReg, RC, RC);
235 assert(Emitted && "Unable to issue a live-in copy instruction!\n");
238 CopyRegMap.insert(std::make_pair(prior(Pos), VirtReg));
240 if (&*InsertPos == UseMI) ++InsertPos;
245 /// EmitLiveInCopies - If this is the first basic block in the function,
246 /// and if it has live ins that need to be copied into vregs, emit the
247 /// copies into the block.
248 static void EmitLiveInCopies(MachineBasicBlock *EntryMBB,
249 const MachineRegisterInfo &MRI,
250 const TargetRegisterInfo &TRI,
251 const TargetInstrInfo &TII) {
252 if (SchedLiveInCopies) {
253 // Emit the copies at a heuristically-determined location in the block.
254 DenseMap<MachineInstr*, unsigned> CopyRegMap;
255 MachineBasicBlock::iterator InsertPos = EntryMBB->begin();
256 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
257 E = MRI.livein_end(); LI != E; ++LI)
259 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
260 EmitLiveInCopy(EntryMBB, InsertPos, LI->second, LI->first,
261 RC, CopyRegMap, MRI, TRI, TII);
264 // Emit the copies into the top of the block.
265 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
266 E = MRI.livein_end(); LI != E; ++LI)
268 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
269 bool Emitted = TII.copyRegToReg(*EntryMBB, EntryMBB->begin(),
270 LI->second, LI->first, RC, RC);
271 assert(Emitted && "Unable to issue a live-in copy instruction!\n");
277 //===----------------------------------------------------------------------===//
278 // SelectionDAGISel code
279 //===----------------------------------------------------------------------===//
281 SelectionDAGISel::SelectionDAGISel(TargetMachine &tm, CodeGenOpt::Level OL) :
282 MachineFunctionPass(&ID), TM(tm), TLI(*tm.getTargetLowering()),
283 FuncInfo(new FunctionLoweringInfo(TLI)),
284 CurDAG(new SelectionDAG(TLI, *FuncInfo)),
285 SDB(new SelectionDAGBuilder(*CurDAG, TLI, *FuncInfo, OL)),
291 SelectionDAGISel::~SelectionDAGISel() {
297 unsigned SelectionDAGISel::MakeReg(EVT VT) {
298 return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT));
301 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
302 AU.addRequired<AliasAnalysis>();
303 AU.addPreserved<AliasAnalysis>();
304 AU.addRequired<GCModuleInfo>();
305 AU.addPreserved<GCModuleInfo>();
306 AU.addRequired<DwarfWriter>();
307 AU.addPreserved<DwarfWriter>();
308 MachineFunctionPass::getAnalysisUsage(AU);
311 bool SelectionDAGISel::runOnMachineFunction(MachineFunction &mf) {
312 Function &Fn = *mf.getFunction();
314 // Do some sanity-checking on the command-line options.
315 assert((!EnableFastISelVerbose || EnableFastISel) &&
316 "-fast-isel-verbose requires -fast-isel");
317 assert((!EnableFastISelAbort || EnableFastISel) &&
318 "-fast-isel-abort requires -fast-isel");
320 // Get alias analysis for load/store combining.
321 AA = &getAnalysis<AliasAnalysis>();
324 const TargetInstrInfo &TII = *TM.getInstrInfo();
325 const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
328 GFI = &getAnalysis<GCModuleInfo>().getFunctionInfo(Fn);
331 RegInfo = &MF->getRegInfo();
332 DEBUG(dbgs() << "\n\n\n=== " << Fn.getName() << "\n");
334 MachineModuleInfo *MMI = getAnalysisIfAvailable<MachineModuleInfo>();
335 DwarfWriter *DW = getAnalysisIfAvailable<DwarfWriter>();
336 CurDAG->init(*MF, MMI, DW);
337 FuncInfo->set(Fn, *MF, EnableFastISel);
340 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
341 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(I->getTerminator()))
343 FuncInfo->MBBMap[Invoke->getSuccessor(1)]->setIsLandingPad();
345 SelectAllBasicBlocks(Fn, *MF, MMI, DW, TII);
347 // If the first basic block in the function has live ins that need to be
348 // copied into vregs, emit the copies into the top of the block before
349 // emitting the code for the block.
350 EmitLiveInCopies(MF->begin(), *RegInfo, TRI, TII);
352 // Add function live-ins to entry block live-in set.
353 for (MachineRegisterInfo::livein_iterator I = RegInfo->livein_begin(),
354 E = RegInfo->livein_end(); I != E; ++I)
355 MF->begin()->addLiveIn(I->first);
358 assert(FuncInfo->CatchInfoFound.size() == FuncInfo->CatchInfoLost.size() &&
359 "Not all catch info was assigned to a landing pad!");
367 /// SetDebugLoc - Update MF's and SDB's DebugLocs if debug information is
368 /// attached with this instruction.
369 static void SetDebugLoc(Instruction *I, SelectionDAGBuilder *SDB,
370 FastISel *FastIS, MachineFunction *MF) {
371 MDNode *Dbg = I->getDbgMetadata();
372 if (Dbg == 0) return;
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);
388 /// ResetDebugLoc - Set MF's and SDB's DebugLocs to Unknown.
389 static void ResetDebugLoc(SelectionDAGBuilder *SDB, FastISel *FastIS) {
390 SDB->setCurDebugLoc(DebugLoc::getUnknownLoc());
392 FastIS->setCurDebugLoc(DebugLoc::getUnknownLoc());
395 void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB,
396 BasicBlock::iterator Begin,
397 BasicBlock::iterator End,
399 SDB->setCurrentBasicBlock(BB);
401 // Lower all of the non-terminator instructions. If a call is emitted
402 // as a tail call, cease emitting nodes for this block.
403 for (BasicBlock::iterator I = Begin; I != End && !SDB->HasTailCall; ++I) {
404 SetDebugLoc(I, SDB, 0, MF);
406 if (!isa<TerminatorInst>(I)) {
409 // Set the current debug location back to "unknown" so that it doesn't
410 // spuriously apply to subsequent instructions.
411 ResetDebugLoc(SDB, 0);
415 if (!SDB->HasTailCall) {
416 // Ensure that all instructions which are used outside of their defining
417 // blocks are available as virtual registers. Invoke is handled elsewhere.
418 for (BasicBlock::iterator I = Begin; I != End; ++I)
419 if (!isa<PHINode>(I) && !isa<InvokeInst>(I))
420 SDB->CopyToExportRegsIfNeeded(I);
422 // Handle PHI nodes in successor blocks.
423 if (End == LLVMBB->end()) {
424 HandlePHINodesInSuccessorBlocks(LLVMBB);
426 // Lower the terminator after the copies are emitted.
427 SetDebugLoc(LLVMBB->getTerminator(), SDB, 0, MF);
428 SDB->visit(*LLVMBB->getTerminator());
429 ResetDebugLoc(SDB, 0);
433 // Make sure the root of the DAG is up-to-date.
434 CurDAG->setRoot(SDB->getControlRoot());
436 // Final step, emit the lowered DAG as machine code.
438 HadTailCall = SDB->HasTailCall;
443 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
444 /// nodes from the worklist.
445 class SDOPsWorkListRemover : public SelectionDAG::DAGUpdateListener {
446 SmallVector<SDNode*, 128> &Worklist;
447 SmallPtrSet<SDNode*, 128> &InWorklist;
449 SDOPsWorkListRemover(SmallVector<SDNode*, 128> &wl,
450 SmallPtrSet<SDNode*, 128> &inwl)
451 : Worklist(wl), InWorklist(inwl) {}
453 void RemoveFromWorklist(SDNode *N) {
454 if (!InWorklist.erase(N)) return;
456 SmallVector<SDNode*, 128>::iterator I =
457 std::find(Worklist.begin(), Worklist.end(), N);
458 assert(I != Worklist.end() && "Not in worklist");
460 *I = Worklist.back();
464 virtual void NodeDeleted(SDNode *N, SDNode *E) {
465 RemoveFromWorklist(N);
468 virtual void NodeUpdated(SDNode *N) {
474 /// TrivialTruncElim - Eliminate some trivial nops that can result from
475 /// ShrinkDemandedOps: (trunc (ext n)) -> n.
476 static bool TrivialTruncElim(SDValue Op,
477 TargetLowering::TargetLoweringOpt &TLO) {
478 SDValue N0 = Op.getOperand(0);
479 EVT VT = Op.getValueType();
480 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
481 N0.getOpcode() == ISD::SIGN_EXTEND ||
482 N0.getOpcode() == ISD::ANY_EXTEND) &&
483 N0.getOperand(0).getValueType() == VT) {
484 return TLO.CombineTo(Op, N0.getOperand(0));
489 /// ShrinkDemandedOps - A late transformation pass that shrink expressions
490 /// using TargetLowering::TargetLoweringOpt::ShrinkDemandedOp. It converts
491 /// x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free.
492 void SelectionDAGISel::ShrinkDemandedOps() {
493 SmallVector<SDNode*, 128> Worklist;
494 SmallPtrSet<SDNode*, 128> InWorklist;
496 // Add all the dag nodes to the worklist.
497 Worklist.reserve(CurDAG->allnodes_size());
498 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
499 E = CurDAG->allnodes_end(); I != E; ++I) {
500 Worklist.push_back(I);
501 InWorklist.insert(I);
504 TargetLowering::TargetLoweringOpt TLO(*CurDAG, true);
505 while (!Worklist.empty()) {
506 SDNode *N = Worklist.pop_back_val();
509 if (N->use_empty() && N != CurDAG->getRoot().getNode()) {
510 // Deleting this node may make its operands dead, add them to the worklist
511 // if they aren't already there.
512 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
513 if (InWorklist.insert(N->getOperand(i).getNode()))
514 Worklist.push_back(N->getOperand(i).getNode());
516 CurDAG->DeleteNode(N);
520 // Run ShrinkDemandedOp on scalar binary operations.
521 if (N->getNumValues() != 1 ||
522 !N->getValueType(0).isSimple() || !N->getValueType(0).isInteger())
525 unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
526 APInt Demanded = APInt::getAllOnesValue(BitWidth);
527 APInt KnownZero, KnownOne;
528 if (!TLI.SimplifyDemandedBits(SDValue(N, 0), Demanded,
529 KnownZero, KnownOne, TLO) &&
530 (N->getOpcode() != ISD::TRUNCATE ||
531 !TrivialTruncElim(SDValue(N, 0), TLO)))
535 assert(!InWorklist.count(N) && "Already in worklist");
536 Worklist.push_back(N);
537 InWorklist.insert(N);
539 // Replace the old value with the new one.
540 DEBUG(errs() << "\nShrinkDemandedOps replacing ";
541 TLO.Old.getNode()->dump(CurDAG);
542 errs() << "\nWith: ";
543 TLO.New.getNode()->dump(CurDAG);
546 if (InWorklist.insert(TLO.New.getNode()))
547 Worklist.push_back(TLO.New.getNode());
549 SDOPsWorkListRemover DeadNodes(Worklist, InWorklist);
550 CurDAG->ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, &DeadNodes);
552 if (!TLO.Old.getNode()->use_empty()) continue;
554 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands();
556 SDNode *OpNode = TLO.Old.getNode()->getOperand(i).getNode();
557 if (OpNode->hasOneUse()) {
558 // Add OpNode to the end of the list to revisit.
559 DeadNodes.RemoveFromWorklist(OpNode);
560 Worklist.push_back(OpNode);
561 InWorklist.insert(OpNode);
565 DeadNodes.RemoveFromWorklist(TLO.Old.getNode());
566 CurDAG->DeleteNode(TLO.Old.getNode());
570 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
571 SmallPtrSet<SDNode*, 128> VisitedNodes;
572 SmallVector<SDNode*, 128> Worklist;
574 Worklist.push_back(CurDAG->getRoot().getNode());
581 SDNode *N = Worklist.pop_back_val();
583 // If we've already seen this node, ignore it.
584 if (!VisitedNodes.insert(N))
587 // Otherwise, add all chain operands to the worklist.
588 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
589 if (N->getOperand(i).getValueType() == MVT::Other)
590 Worklist.push_back(N->getOperand(i).getNode());
592 // If this is a CopyToReg with a vreg dest, process it.
593 if (N->getOpcode() != ISD::CopyToReg)
596 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
597 if (!TargetRegisterInfo::isVirtualRegister(DestReg))
600 // Ignore non-scalar or non-integer values.
601 SDValue Src = N->getOperand(2);
602 EVT SrcVT = Src.getValueType();
603 if (!SrcVT.isInteger() || SrcVT.isVector())
606 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
607 Mask = APInt::getAllOnesValue(SrcVT.getSizeInBits());
608 CurDAG->ComputeMaskedBits(Src, Mask, KnownZero, KnownOne);
610 // Only install this information if it tells us something.
611 if (NumSignBits != 1 || KnownZero != 0 || KnownOne != 0) {
612 DestReg -= TargetRegisterInfo::FirstVirtualRegister;
613 if (DestReg >= FuncInfo->LiveOutRegInfo.size())
614 FuncInfo->LiveOutRegInfo.resize(DestReg+1);
615 FunctionLoweringInfo::LiveOutInfo &LOI =
616 FuncInfo->LiveOutRegInfo[DestReg];
617 LOI.NumSignBits = NumSignBits;
618 LOI.KnownOne = KnownOne;
619 LOI.KnownZero = KnownZero;
621 } while (!Worklist.empty());
624 void SelectionDAGISel::CodeGenAndEmitDAG() {
625 std::string GroupName;
626 if (TimePassesIsEnabled)
627 GroupName = "Instruction Selection and Scheduling";
628 std::string BlockName;
629 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs ||
630 ViewDAGCombine2 || ViewDAGCombineLT || ViewISelDAGs || ViewSchedDAGs ||
632 BlockName = MF->getFunction()->getNameStr() + ":" +
633 BB->getBasicBlock()->getNameStr();
635 DEBUG(dbgs() << "Initial selection DAG:\n");
636 DEBUG(CurDAG->dump());
638 if (ViewDAGCombine1) CurDAG->viewGraph("dag-combine1 input for " + BlockName);
640 // Run the DAG combiner in pre-legalize mode.
641 if (TimePassesIsEnabled) {
642 NamedRegionTimer T("DAG Combining 1", GroupName);
643 CurDAG->Combine(Unrestricted, *AA, OptLevel);
645 CurDAG->Combine(Unrestricted, *AA, OptLevel);
648 DEBUG(dbgs() << "Optimized lowered selection DAG:\n");
649 DEBUG(CurDAG->dump());
651 // Second step, hack on the DAG until it only uses operations and types that
652 // the target supports.
653 if (ViewLegalizeTypesDAGs) CurDAG->viewGraph("legalize-types input for " +
657 if (TimePassesIsEnabled) {
658 NamedRegionTimer T("Type Legalization", GroupName);
659 Changed = CurDAG->LegalizeTypes();
661 Changed = CurDAG->LegalizeTypes();
664 DEBUG(dbgs() << "Type-legalized selection DAG:\n");
665 DEBUG(CurDAG->dump());
668 if (ViewDAGCombineLT)
669 CurDAG->viewGraph("dag-combine-lt input for " + BlockName);
671 // Run the DAG combiner in post-type-legalize mode.
672 if (TimePassesIsEnabled) {
673 NamedRegionTimer T("DAG Combining after legalize types", GroupName);
674 CurDAG->Combine(NoIllegalTypes, *AA, OptLevel);
676 CurDAG->Combine(NoIllegalTypes, *AA, OptLevel);
679 DEBUG(dbgs() << "Optimized type-legalized selection DAG:\n");
680 DEBUG(CurDAG->dump());
683 if (TimePassesIsEnabled) {
684 NamedRegionTimer T("Vector Legalization", GroupName);
685 Changed = CurDAG->LegalizeVectors();
687 Changed = CurDAG->LegalizeVectors();
691 if (TimePassesIsEnabled) {
692 NamedRegionTimer T("Type Legalization 2", GroupName);
693 CurDAG->LegalizeTypes();
695 CurDAG->LegalizeTypes();
698 if (ViewDAGCombineLT)
699 CurDAG->viewGraph("dag-combine-lv input for " + BlockName);
701 // Run the DAG combiner in post-type-legalize mode.
702 if (TimePassesIsEnabled) {
703 NamedRegionTimer T("DAG Combining after legalize vectors", GroupName);
704 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
706 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
709 DEBUG(dbgs() << "Optimized vector-legalized selection DAG:\n");
710 DEBUG(CurDAG->dump());
713 if (ViewLegalizeDAGs) CurDAG->viewGraph("legalize input for " + BlockName);
715 if (TimePassesIsEnabled) {
716 NamedRegionTimer T("DAG Legalization", GroupName);
717 CurDAG->Legalize(OptLevel);
719 CurDAG->Legalize(OptLevel);
722 DEBUG(dbgs() << "Legalized selection DAG:\n");
723 DEBUG(CurDAG->dump());
725 if (ViewDAGCombine2) CurDAG->viewGraph("dag-combine2 input for " + BlockName);
727 // Run the DAG combiner in post-legalize mode.
728 if (TimePassesIsEnabled) {
729 NamedRegionTimer T("DAG Combining 2", GroupName);
730 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
732 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
735 DEBUG(dbgs() << "Optimized legalized selection DAG:\n");
736 DEBUG(CurDAG->dump());
738 if (OptLevel != CodeGenOpt::None) {
740 ComputeLiveOutVRegInfo();
743 if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName);
745 // Third, instruction select all of the operations to machine code, adding the
746 // code to the MachineBasicBlock.
747 if (TimePassesIsEnabled) {
748 NamedRegionTimer T("Instruction Selection", GroupName);
749 DoInstructionSelection();
751 DoInstructionSelection();
754 DEBUG(dbgs() << "Selected selection DAG:\n");
755 DEBUG(CurDAG->dump());
757 if (ViewSchedDAGs) CurDAG->viewGraph("scheduler input for " + BlockName);
759 // Schedule machine code.
760 ScheduleDAGSDNodes *Scheduler = CreateScheduler();
761 if (TimePassesIsEnabled) {
762 NamedRegionTimer T("Instruction Scheduling", GroupName);
763 Scheduler->Run(CurDAG, BB, BB->end());
765 Scheduler->Run(CurDAG, BB, BB->end());
768 if (ViewSUnitDAGs) Scheduler->viewGraph();
770 // Emit machine code to BB. This can change 'BB' to the last block being
772 if (TimePassesIsEnabled) {
773 NamedRegionTimer T("Instruction Creation", GroupName);
774 BB = Scheduler->EmitSchedule(&SDB->EdgeMapping);
776 BB = Scheduler->EmitSchedule(&SDB->EdgeMapping);
779 // Free the scheduler state.
780 if (TimePassesIsEnabled) {
781 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName);
787 DEBUG(dbgs() << "Selected machine code:\n");
791 void SelectionDAGISel::DoInstructionSelection() {
792 DEBUG(errs() << "===== Instruction selection begins:\n");
796 // Select target instructions for the DAG.
798 // Number all nodes with a topological order and set DAGSize.
799 DAGSize = CurDAG->AssignTopologicalOrder();
801 // Create a dummy node (which is not added to allnodes), that adds
802 // a reference to the root node, preventing it from being deleted,
803 // and tracking any changes of the root.
804 HandleSDNode Dummy(CurDAG->getRoot());
805 ISelPosition = SelectionDAG::allnodes_iterator(CurDAG->getRoot().getNode());
808 // The AllNodes list is now topological-sorted. Visit the
809 // nodes by starting at the end of the list (the root of the
810 // graph) and preceding back toward the beginning (the entry
812 while (ISelPosition != CurDAG->allnodes_begin()) {
813 SDNode *Node = --ISelPosition;
814 // Skip dead nodes. DAGCombiner is expected to eliminate all dead nodes,
815 // but there are currently some corner cases that it misses. Also, this
816 // makes it theoretically possible to disable the DAGCombiner.
817 if (Node->use_empty())
820 SDNode *ResNode = Select(Node);
822 // FIXME: This is pretty gross. 'Select' should be changed to not return
823 // anything at all and this code should be nuked with a tactical strike.
825 // If node should not be replaced, continue with the next one.
826 if (ResNode == Node || Node->getOpcode() == ISD::DELETED_NODE)
830 ReplaceUses(Node, ResNode);
832 // If after the replacement this node is not used any more,
833 // remove this dead node.
834 if (Node->use_empty()) { // Don't delete EntryToken, etc.
835 ISelUpdater ISU(ISelPosition);
836 CurDAG->RemoveDeadNode(Node, &ISU);
840 CurDAG->setRoot(Dummy.getValue());
842 DEBUG(errs() << "===== Instruction selection ends:\n");
844 PostprocessISelDAG();
848 void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn,
850 MachineModuleInfo *MMI,
852 const TargetInstrInfo &TII) {
853 // Initialize the Fast-ISel state, if needed.
854 FastISel *FastIS = 0;
856 FastIS = TLI.createFastISel(MF, MMI, DW,
859 FuncInfo->StaticAllocaMap
861 , FuncInfo->CatchInfoLost
865 // Iterate over all basic blocks in the function.
866 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
867 BasicBlock *LLVMBB = &*I;
868 BB = FuncInfo->MBBMap[LLVMBB];
870 BasicBlock::iterator const Begin = LLVMBB->begin();
871 BasicBlock::iterator const End = LLVMBB->end();
872 BasicBlock::iterator BI = Begin;
874 // Lower any arguments needed in this block if this is the entry block.
875 bool SuppressFastISel = false;
876 if (LLVMBB == &Fn.getEntryBlock()) {
877 LowerArguments(LLVMBB);
879 // If any of the arguments has the byval attribute, forgo
880 // fast-isel in the entry block.
883 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end();
885 if (Fn.paramHasAttr(j, Attribute::ByVal)) {
886 if (EnableFastISelVerbose || EnableFastISelAbort)
887 dbgs() << "FastISel skips entry block due to byval argument\n";
888 SuppressFastISel = true;
894 if (MMI && BB->isLandingPad()) {
895 // Add a label to mark the beginning of the landing pad. Deletion of the
896 // landing pad can thus be detected via the MachineModuleInfo.
897 MCSymbol *Label = MMI->addLandingPad(BB);
899 const TargetInstrDesc &II = TII.get(TargetOpcode::EH_LABEL);
900 BuildMI(BB, SDB->getCurDebugLoc(), II).addSym(Label);
902 // Mark exception register as live in.
903 unsigned Reg = TLI.getExceptionAddressRegister();
904 if (Reg) BB->addLiveIn(Reg);
906 // Mark exception selector register as live in.
907 Reg = TLI.getExceptionSelectorRegister();
908 if (Reg) BB->addLiveIn(Reg);
910 // FIXME: Hack around an exception handling flaw (PR1508): the personality
911 // function and list of typeids logically belong to the invoke (or, if you
912 // like, the basic block containing the invoke), and need to be associated
913 // with it in the dwarf exception handling tables. Currently however the
914 // information is provided by an intrinsic (eh.selector) that can be moved
915 // to unexpected places by the optimizers: if the unwind edge is critical,
916 // then breaking it can result in the intrinsics being in the successor of
917 // the landing pad, not the landing pad itself. This results
918 // in exceptions not being caught because no typeids are associated with
919 // the invoke. This may not be the only way things can go wrong, but it
920 // is the only way we try to work around for the moment.
921 BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
923 if (Br && Br->isUnconditional()) { // Critical edge?
924 BasicBlock::iterator I, E;
925 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
926 if (isa<EHSelectorInst>(I))
930 // No catch info found - try to extract some from the successor.
931 CopyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, *FuncInfo);
935 // Before doing SelectionDAG ISel, see if FastISel has been requested.
936 if (FastIS && !SuppressFastISel) {
937 // Emit code for any incoming arguments. This must happen before
938 // beginning FastISel on the entry block.
939 if (LLVMBB == &Fn.getEntryBlock()) {
940 CurDAG->setRoot(SDB->getControlRoot());
944 FastIS->startNewBlock(BB);
945 // Do FastISel on as many instructions as possible.
946 for (; BI != End; ++BI) {
947 // Just before the terminator instruction, insert instructions to
948 // feed PHI nodes in successor blocks.
949 if (isa<TerminatorInst>(BI))
950 if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, FastIS)) {
951 ++NumFastIselFailures;
952 ResetDebugLoc(SDB, FastIS);
953 if (EnableFastISelVerbose || EnableFastISelAbort) {
954 dbgs() << "FastISel miss: ";
957 assert(!EnableFastISelAbort &&
958 "FastISel didn't handle a PHI in a successor");
962 SetDebugLoc(BI, SDB, FastIS, &MF);
964 // Try to select the instruction with FastISel.
965 if (FastIS->SelectInstruction(BI)) {
966 ResetDebugLoc(SDB, FastIS);
970 // Clear out the debug location so that it doesn't carry over to
971 // unrelated instructions.
972 ResetDebugLoc(SDB, FastIS);
974 // Then handle certain instructions as single-LLVM-Instruction blocks.
975 if (isa<CallInst>(BI)) {
976 ++NumFastIselFailures;
977 if (EnableFastISelVerbose || EnableFastISelAbort) {
978 dbgs() << "FastISel missed call: ";
982 if (!BI->getType()->isVoidTy()) {
983 unsigned &R = FuncInfo->ValueMap[BI];
985 R = FuncInfo->CreateRegForValue(BI);
988 bool HadTailCall = false;
989 SelectBasicBlock(LLVMBB, BI, llvm::next(BI), HadTailCall);
991 // If the call was emitted as a tail call, we're done with the block.
997 // If the instruction was codegen'd with multiple blocks,
998 // inform the FastISel object where to resume inserting.
999 FastIS->setCurrentBlock(BB);
1003 // Otherwise, give up on FastISel for the rest of the block.
1004 // For now, be a little lenient about non-branch terminators.
1005 if (!isa<TerminatorInst>(BI) || isa<BranchInst>(BI)) {
1006 ++NumFastIselFailures;
1007 if (EnableFastISelVerbose || EnableFastISelAbort) {
1008 dbgs() << "FastISel miss: ";
1011 if (EnableFastISelAbort)
1012 // The "fast" selector couldn't handle something and bailed.
1013 // For the purpose of debugging, just abort.
1014 llvm_unreachable("FastISel didn't select the entire block");
1020 // Run SelectionDAG instruction selection on the remainder of the block
1021 // not handled by FastISel. If FastISel is not run, this is the entire
1025 SelectBasicBlock(LLVMBB, BI, End, HadTailCall);
1035 SelectionDAGISel::FinishBasicBlock() {
1037 DEBUG(dbgs() << "Target-post-processed machine code:\n");
1040 DEBUG(dbgs() << "Total amount of phi nodes to update: "
1041 << SDB->PHINodesToUpdate.size() << "\n");
1042 DEBUG(for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i)
1043 dbgs() << "Node " << i << " : ("
1044 << SDB->PHINodesToUpdate[i].first
1045 << ", " << SDB->PHINodesToUpdate[i].second << ")\n");
1047 // Next, now that we know what the last MBB the LLVM BB expanded is, update
1048 // PHI nodes in successors.
1049 if (SDB->SwitchCases.empty() &&
1050 SDB->JTCases.empty() &&
1051 SDB->BitTestCases.empty()) {
1052 for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) {
1053 MachineInstr *PHI = SDB->PHINodesToUpdate[i].first;
1054 assert(PHI->isPHI() &&
1055 "This is not a machine PHI node that we are updating!");
1056 if (!BB->isSuccessor(PHI->getParent()))
1058 PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second,
1060 PHI->addOperand(MachineOperand::CreateMBB(BB));
1062 SDB->PHINodesToUpdate.clear();
1066 for (unsigned i = 0, e = SDB->BitTestCases.size(); i != e; ++i) {
1067 // Lower header first, if it wasn't already lowered
1068 if (!SDB->BitTestCases[i].Emitted) {
1069 // Set the current basic block to the mbb we wish to insert the code into
1070 BB = SDB->BitTestCases[i].Parent;
1071 SDB->setCurrentBasicBlock(BB);
1073 SDB->visitBitTestHeader(SDB->BitTestCases[i]);
1074 CurDAG->setRoot(SDB->getRoot());
1075 CodeGenAndEmitDAG();
1079 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j) {
1080 // Set the current basic block to the mbb we wish to insert the code into
1081 BB = SDB->BitTestCases[i].Cases[j].ThisBB;
1082 SDB->setCurrentBasicBlock(BB);
1085 SDB->visitBitTestCase(SDB->BitTestCases[i].Cases[j+1].ThisBB,
1086 SDB->BitTestCases[i].Reg,
1087 SDB->BitTestCases[i].Cases[j]);
1089 SDB->visitBitTestCase(SDB->BitTestCases[i].Default,
1090 SDB->BitTestCases[i].Reg,
1091 SDB->BitTestCases[i].Cases[j]);
1094 CurDAG->setRoot(SDB->getRoot());
1095 CodeGenAndEmitDAG();
1100 for (unsigned pi = 0, pe = SDB->PHINodesToUpdate.size(); pi != pe; ++pi) {
1101 MachineInstr *PHI = SDB->PHINodesToUpdate[pi].first;
1102 MachineBasicBlock *PHIBB = PHI->getParent();
1103 assert(PHI->isPHI() &&
1104 "This is not a machine PHI node that we are updating!");
1105 // This is "default" BB. We have two jumps to it. From "header" BB and
1106 // from last "case" BB.
1107 if (PHIBB == SDB->BitTestCases[i].Default) {
1108 PHI->addOperand(MachineOperand::
1109 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1110 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Parent));
1111 PHI->addOperand(MachineOperand::
1112 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1113 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Cases.
1116 // One of "cases" BB.
1117 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size();
1119 MachineBasicBlock* cBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1120 if (cBB->isSuccessor(PHIBB)) {
1121 PHI->addOperand(MachineOperand::
1122 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1123 PHI->addOperand(MachineOperand::CreateMBB(cBB));
1128 SDB->BitTestCases.clear();
1130 // If the JumpTable record is filled in, then we need to emit a jump table.
1131 // Updating the PHI nodes is tricky in this case, since we need to determine
1132 // whether the PHI is a successor of the range check MBB or the jump table MBB
1133 for (unsigned i = 0, e = SDB->JTCases.size(); i != e; ++i) {
1134 // Lower header first, if it wasn't already lowered
1135 if (!SDB->JTCases[i].first.Emitted) {
1136 // Set the current basic block to the mbb we wish to insert the code into
1137 BB = SDB->JTCases[i].first.HeaderBB;
1138 SDB->setCurrentBasicBlock(BB);
1140 SDB->visitJumpTableHeader(SDB->JTCases[i].second, SDB->JTCases[i].first);
1141 CurDAG->setRoot(SDB->getRoot());
1142 CodeGenAndEmitDAG();
1146 // Set the current basic block to the mbb we wish to insert the code into
1147 BB = SDB->JTCases[i].second.MBB;
1148 SDB->setCurrentBasicBlock(BB);
1150 SDB->visitJumpTable(SDB->JTCases[i].second);
1151 CurDAG->setRoot(SDB->getRoot());
1152 CodeGenAndEmitDAG();
1156 for (unsigned pi = 0, pe = SDB->PHINodesToUpdate.size(); pi != pe; ++pi) {
1157 MachineInstr *PHI = SDB->PHINodesToUpdate[pi].first;
1158 MachineBasicBlock *PHIBB = PHI->getParent();
1159 assert(PHI->isPHI() &&
1160 "This is not a machine PHI node that we are updating!");
1161 // "default" BB. We can go there only from header BB.
1162 if (PHIBB == SDB->JTCases[i].second.Default) {
1164 (MachineOperand::CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1166 (MachineOperand::CreateMBB(SDB->JTCases[i].first.HeaderBB));
1168 // JT BB. Just iterate over successors here
1169 if (BB->isSuccessor(PHIBB)) {
1171 (MachineOperand::CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1172 PHI->addOperand(MachineOperand::CreateMBB(BB));
1176 SDB->JTCases.clear();
1178 // If the switch block involved a branch to one of the actual successors, we
1179 // need to update PHI nodes in that block.
1180 for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) {
1181 MachineInstr *PHI = SDB->PHINodesToUpdate[i].first;
1182 assert(PHI->isPHI() &&
1183 "This is not a machine PHI node that we are updating!");
1184 if (BB->isSuccessor(PHI->getParent())) {
1185 PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second,
1187 PHI->addOperand(MachineOperand::CreateMBB(BB));
1191 // If we generated any switch lowering information, build and codegen any
1192 // additional DAGs necessary.
1193 for (unsigned i = 0, e = SDB->SwitchCases.size(); i != e; ++i) {
1194 // Set the current basic block to the mbb we wish to insert the code into
1195 MachineBasicBlock *ThisBB = BB = SDB->SwitchCases[i].ThisBB;
1196 SDB->setCurrentBasicBlock(BB);
1199 SDB->visitSwitchCase(SDB->SwitchCases[i]);
1200 CurDAG->setRoot(SDB->getRoot());
1201 CodeGenAndEmitDAG();
1203 // Handle any PHI nodes in successors of this chunk, as if we were coming
1204 // from the original BB before switch expansion. Note that PHI nodes can
1205 // occur multiple times in PHINodesToUpdate. We have to be very careful to
1206 // handle them the right number of times.
1207 while ((BB = SDB->SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
1208 // If new BB's are created during scheduling, the edges may have been
1209 // updated. That is, the edge from ThisBB to BB may have been split and
1210 // BB's predecessor is now another block.
1211 DenseMap<MachineBasicBlock*, MachineBasicBlock*>::iterator EI =
1212 SDB->EdgeMapping.find(BB);
1213 if (EI != SDB->EdgeMapping.end())
1214 ThisBB = EI->second;
1216 // BB may have been removed from the CFG if a branch was constant folded.
1217 if (ThisBB->isSuccessor(BB)) {
1218 for (MachineBasicBlock::iterator Phi = BB->begin();
1219 Phi != BB->end() && Phi->isPHI();
1221 // This value for this PHI node is recorded in PHINodesToUpdate.
1222 for (unsigned pn = 0; ; ++pn) {
1223 assert(pn != SDB->PHINodesToUpdate.size() &&
1224 "Didn't find PHI entry!");
1225 if (SDB->PHINodesToUpdate[pn].first == Phi) {
1226 Phi->addOperand(MachineOperand::
1227 CreateReg(SDB->PHINodesToUpdate[pn].second,
1229 Phi->addOperand(MachineOperand::CreateMBB(ThisBB));
1236 // Don't process RHS if same block as LHS.
1237 if (BB == SDB->SwitchCases[i].FalseBB)
1238 SDB->SwitchCases[i].FalseBB = 0;
1240 // If we haven't handled the RHS, do so now. Otherwise, we're done.
1241 SDB->SwitchCases[i].TrueBB = SDB->SwitchCases[i].FalseBB;
1242 SDB->SwitchCases[i].FalseBB = 0;
1244 assert(SDB->SwitchCases[i].TrueBB == 0 && SDB->SwitchCases[i].FalseBB == 0);
1247 SDB->SwitchCases.clear();
1249 SDB->PHINodesToUpdate.clear();
1253 /// Create the scheduler. If a specific scheduler was specified
1254 /// via the SchedulerRegistry, use it, otherwise select the
1255 /// one preferred by the target.
1257 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
1258 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1262 RegisterScheduler::setDefault(Ctor);
1265 return Ctor(this, OptLevel);
1268 ScheduleHazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
1269 return new ScheduleHazardRecognizer();
1272 //===----------------------------------------------------------------------===//
1273 // Helper functions used by the generated instruction selector.
1274 //===----------------------------------------------------------------------===//
1275 // Calls to these methods are generated by tblgen.
1277 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1278 /// the dag combiner simplified the 255, we still want to match. RHS is the
1279 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1280 /// specified in the .td file (e.g. 255).
1281 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1282 int64_t DesiredMaskS) const {
1283 const APInt &ActualMask = RHS->getAPIntValue();
1284 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1286 // If the actual mask exactly matches, success!
1287 if (ActualMask == DesiredMask)
1290 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1291 if (ActualMask.intersects(~DesiredMask))
1294 // Otherwise, the DAG Combiner may have proven that the value coming in is
1295 // either already zero or is not demanded. Check for known zero input bits.
1296 APInt NeededMask = DesiredMask & ~ActualMask;
1297 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1300 // TODO: check to see if missing bits are just not demanded.
1302 // Otherwise, this pattern doesn't match.
1306 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1307 /// the dag combiner simplified the 255, we still want to match. RHS is the
1308 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1309 /// specified in the .td file (e.g. 255).
1310 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1311 int64_t DesiredMaskS) const {
1312 const APInt &ActualMask = RHS->getAPIntValue();
1313 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1315 // If the actual mask exactly matches, success!
1316 if (ActualMask == DesiredMask)
1319 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1320 if (ActualMask.intersects(~DesiredMask))
1323 // Otherwise, the DAG Combiner may have proven that the value coming in is
1324 // either already zero or is not demanded. Check for known zero input bits.
1325 APInt NeededMask = DesiredMask & ~ActualMask;
1327 APInt KnownZero, KnownOne;
1328 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
1330 // If all the missing bits in the or are already known to be set, match!
1331 if ((NeededMask & KnownOne) == NeededMask)
1334 // TODO: check to see if missing bits are just not demanded.
1336 // Otherwise, this pattern doesn't match.
1341 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1342 /// by tblgen. Others should not call it.
1343 void SelectionDAGISel::
1344 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1345 std::vector<SDValue> InOps;
1346 std::swap(InOps, Ops);
1348 Ops.push_back(InOps[0]); // input chain.
1349 Ops.push_back(InOps[1]); // input asm string.
1351 unsigned i = 2, e = InOps.size();
1352 if (InOps[e-1].getValueType() == MVT::Flag)
1353 --e; // Don't process a flag operand if it is here.
1356 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1357 if ((Flags & 7) != 4 /*MEM*/) {
1358 // Just skip over this operand, copying the operands verbatim.
1359 Ops.insert(Ops.end(), InOps.begin()+i,
1360 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
1361 i += InlineAsm::getNumOperandRegisters(Flags) + 1;
1363 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 &&
1364 "Memory operand with multiple values?");
1365 // Otherwise, this is a memory operand. Ask the target to select it.
1366 std::vector<SDValue> SelOps;
1367 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps)) {
1368 llvm_report_error("Could not match memory address. Inline asm"
1372 // Add this to the output node.
1373 Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size()<< 3),
1375 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1380 // Add the flag input back if present.
1381 if (e != InOps.size())
1382 Ops.push_back(InOps.back());
1385 /// findFlagUse - Return use of EVT::Flag value produced by the specified
1388 static SDNode *findFlagUse(SDNode *N) {
1389 unsigned FlagResNo = N->getNumValues()-1;
1390 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1391 SDUse &Use = I.getUse();
1392 if (Use.getResNo() == FlagResNo)
1393 return Use.getUser();
1398 /// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
1399 /// This function recursively traverses up the operand chain, ignoring
1401 static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
1402 SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited,
1403 bool IgnoreChains) {
1404 // The NodeID's are given uniques ID's where a node ID is guaranteed to be
1405 // greater than all of its (recursive) operands. If we scan to a point where
1406 // 'use' is smaller than the node we're scanning for, then we know we will
1409 // The Use may be -1 (unassigned) if it is a newly allocated node. This can
1410 // happen because we scan down to newly selected nodes in the case of flag
1412 if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1))
1415 // Don't revisit nodes if we already scanned it and didn't fail, we know we
1416 // won't fail if we scan it again.
1417 if (!Visited.insert(Use))
1420 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
1421 // Ignore chain uses, they are validated by HandleMergeInputChains.
1422 if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains)
1425 SDNode *N = Use->getOperand(i).getNode();
1427 if (Use == ImmedUse || Use == Root)
1428 continue; // We are not looking for immediate use.
1433 // Traverse up the operand chain.
1434 if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
1440 /// IsProfitableToFold - Returns true if it's profitable to fold the specific
1441 /// operand node N of U during instruction selection that starts at Root.
1442 bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
1443 SDNode *Root) const {
1444 if (OptLevel == CodeGenOpt::None) return false;
1445 return N.hasOneUse();
1448 /// IsLegalToFold - Returns true if the specific operand node N of
1449 /// U can be folded during instruction selection that starts at Root.
1450 bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
1451 bool IgnoreChains) const {
1452 if (OptLevel == CodeGenOpt::None) return false;
1454 // If Root use can somehow reach N through a path that that doesn't contain
1455 // U then folding N would create a cycle. e.g. In the following
1456 // diagram, Root can reach N through X. If N is folded into into Root, then
1457 // X is both a predecessor and a successor of U.
1468 // * indicates nodes to be folded together.
1470 // If Root produces a flag, then it gets (even more) interesting. Since it
1471 // will be "glued" together with its flag use in the scheduler, we need to
1472 // check if it might reach N.
1491 // If FU (flag use) indirectly reaches N (the load), and Root folds N
1492 // (call it Fold), then X is a predecessor of FU and a successor of
1493 // Fold. But since Fold and FU are flagged together, this will create
1494 // a cycle in the scheduling graph.
1496 // If the node has flags, walk down the graph to the "lowest" node in the
1498 EVT VT = Root->getValueType(Root->getNumValues()-1);
1499 while (VT == MVT::Flag) {
1500 SDNode *FU = findFlagUse(Root);
1504 VT = Root->getValueType(Root->getNumValues()-1);
1506 // If our query node has a flag result with a use, we've walked up it. If
1507 // the user (which has already been selected) has a chain or indirectly uses
1508 // the chain, our WalkChainUsers predicate will not consider it. Because of
1509 // this, we cannot ignore chains in this predicate.
1510 IgnoreChains = false;
1514 SmallPtrSet<SDNode*, 16> Visited;
1515 return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains);
1518 SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
1519 std::vector<SDValue> Ops(N->op_begin(), N->op_end());
1520 SelectInlineAsmMemoryOperands(Ops);
1522 std::vector<EVT> VTs;
1523 VTs.push_back(MVT::Other);
1524 VTs.push_back(MVT::Flag);
1525 SDValue New = CurDAG->getNode(ISD::INLINEASM, N->getDebugLoc(),
1526 VTs, &Ops[0], Ops.size());
1528 return New.getNode();
1531 SDNode *SelectionDAGISel::Select_UNDEF(SDNode *N) {
1532 return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0));
1535 /// GetVBR - decode a vbr encoding whose top bit is set.
1536 ALWAYS_INLINE static uint64_t
1537 GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
1538 assert(Val >= 128 && "Not a VBR");
1539 Val &= 127; // Remove first vbr bit.
1544 NextBits = MatcherTable[Idx++];
1545 Val |= (NextBits&127) << Shift;
1547 } while (NextBits & 128);
1553 /// UpdateChainsAndFlags - When a match is complete, this method updates uses of
1554 /// interior flag and chain results to use the new flag and chain results.
1555 void SelectionDAGISel::
1556 UpdateChainsAndFlags(SDNode *NodeToMatch, SDValue InputChain,
1557 const SmallVectorImpl<SDNode*> &ChainNodesMatched,
1559 const SmallVectorImpl<SDNode*> &FlagResultNodesMatched,
1560 bool isMorphNodeTo) {
1561 SmallVector<SDNode*, 4> NowDeadNodes;
1563 ISelUpdater ISU(ISelPosition);
1565 // Now that all the normal results are replaced, we replace the chain and
1566 // flag results if present.
1567 if (!ChainNodesMatched.empty()) {
1568 assert(InputChain.getNode() != 0 &&
1569 "Matched input chains but didn't produce a chain");
1570 // Loop over all of the nodes we matched that produced a chain result.
1571 // Replace all the chain results with the final chain we ended up with.
1572 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1573 SDNode *ChainNode = ChainNodesMatched[i];
1575 // If this node was already deleted, don't look at it.
1576 if (ChainNode->getOpcode() == ISD::DELETED_NODE)
1579 // Don't replace the results of the root node if we're doing a
1581 if (ChainNode == NodeToMatch && isMorphNodeTo)
1584 SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1);
1585 if (ChainVal.getValueType() == MVT::Flag)
1586 ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
1587 assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
1588 CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain, &ISU);
1590 // If the node became dead and we haven't already seen it, delete it.
1591 if (ChainNode->use_empty() &&
1592 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), ChainNode))
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 and we haven't already seen it, delete it.
1614 if (FRN->use_empty() &&
1615 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), FRN))
1616 NowDeadNodes.push_back(FRN);
1620 if (!NowDeadNodes.empty())
1621 CurDAG->RemoveDeadNodes(NowDeadNodes, &ISU);
1623 DEBUG(errs() << "ISEL: Match complete!\n");
1629 CR_LeadsToInteriorNode
1632 /// WalkChainUsers - Walk down the users of the specified chained node that is
1633 /// part of the pattern we're matching, looking at all of the users we find.
1634 /// This determines whether something is an interior node, whether we have a
1635 /// non-pattern node in between two pattern nodes (which prevent folding because
1636 /// it would induce a cycle) and whether we have a TokenFactor node sandwiched
1637 /// between pattern nodes (in which case the TF becomes part of the pattern).
1639 /// The walk we do here is guaranteed to be small because we quickly get down to
1640 /// already selected nodes "below" us.
1642 WalkChainUsers(SDNode *ChainedNode,
1643 SmallVectorImpl<SDNode*> &ChainedNodesInPattern,
1644 SmallVectorImpl<SDNode*> &InteriorChainedNodes) {
1645 ChainResult Result = CR_Simple;
1647 for (SDNode::use_iterator UI = ChainedNode->use_begin(),
1648 E = ChainedNode->use_end(); UI != E; ++UI) {
1649 // Make sure the use is of the chain, not some other value we produce.
1650 if (UI.getUse().getValueType() != MVT::Other) continue;
1654 // If we see an already-selected machine node, then we've gone beyond the
1655 // pattern that we're selecting down into the already selected chunk of the
1657 if (User->isMachineOpcode() ||
1658 User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
1661 if (User->getOpcode() == ISD::CopyToReg ||
1662 User->getOpcode() == ISD::CopyFromReg ||
1663 User->getOpcode() == ISD::INLINEASM ||
1664 User->getOpcode() == ISD::EH_LABEL) {
1665 // If their node ID got reset to -1 then they've already been selected.
1666 // Treat them like a MachineOpcode.
1667 if (User->getNodeId() == -1)
1671 // If we have a TokenFactor, we handle it specially.
1672 if (User->getOpcode() != ISD::TokenFactor) {
1673 // If the node isn't a token factor and isn't part of our pattern, then it
1674 // must be a random chained node in between two nodes we're selecting.
1675 // This happens when we have something like:
1680 // Because we structurally match the load/store as a read/modify/write,
1681 // but the call is chained between them. We cannot fold in this case
1682 // because it would induce a cycle in the graph.
1683 if (!std::count(ChainedNodesInPattern.begin(),
1684 ChainedNodesInPattern.end(), User))
1685 return CR_InducesCycle;
1687 // Otherwise we found a node that is part of our pattern. For example in:
1691 // This would happen when we're scanning down from the load and see the
1692 // store as a user. Record that there is a use of ChainedNode that is
1693 // part of the pattern and keep scanning uses.
1694 Result = CR_LeadsToInteriorNode;
1695 InteriorChainedNodes.push_back(User);
1699 // If we found a TokenFactor, there are two cases to consider: first if the
1700 // TokenFactor is just hanging "below" the pattern we're matching (i.e. no
1701 // uses of the TF are in our pattern) we just want to ignore it. Second,
1702 // the TokenFactor can be sandwiched in between two chained nodes, like so:
1708 // | \ DAG's like cheese
1711 // [TokenFactor] [Op]
1718 // In this case, the TokenFactor becomes part of our match and we rewrite it
1719 // as a new TokenFactor.
1721 // To distinguish these two cases, do a recursive walk down the uses.
1722 switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) {
1724 // If the uses of the TokenFactor are just already-selected nodes, ignore
1725 // it, it is "below" our pattern.
1727 case CR_InducesCycle:
1728 // If the uses of the TokenFactor lead to nodes that are not part of our
1729 // pattern that are not selected, folding would turn this into a cycle,
1731 return CR_InducesCycle;
1732 case CR_LeadsToInteriorNode:
1733 break; // Otherwise, keep processing.
1736 // Okay, we know we're in the interesting interior case. The TokenFactor
1737 // is now going to be considered part of the pattern so that we rewrite its
1738 // uses (it may have uses that are not part of the pattern) with the
1739 // ultimate chain result of the generated code. We will also add its chain
1740 // inputs as inputs to the ultimate TokenFactor we create.
1741 Result = CR_LeadsToInteriorNode;
1742 ChainedNodesInPattern.push_back(User);
1743 InteriorChainedNodes.push_back(User);
1750 /// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
1751 /// operation for when the pattern matched at least one node with a chains. The
1752 /// input vector contains a list of all of the chained nodes that we match. We
1753 /// must determine if this is a valid thing to cover (i.e. matching it won't
1754 /// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
1755 /// be used as the input node chain for the generated nodes.
1757 HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
1758 SelectionDAG *CurDAG) {
1759 // Walk all of the chained nodes we've matched, recursively scanning down the
1760 // users of the chain result. This adds any TokenFactor nodes that are caught
1761 // in between chained nodes to the chained and interior nodes list.
1762 SmallVector<SDNode*, 3> InteriorChainedNodes;
1763 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1764 if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched,
1765 InteriorChainedNodes) == CR_InducesCycle)
1766 return SDValue(); // Would induce a cycle.
1769 // Okay, we have walked all the matched nodes and collected TokenFactor nodes
1770 // that we are interested in. Form our input TokenFactor node.
1771 SmallVector<SDValue, 3> InputChains;
1772 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1773 // Add the input chain of this node to the InputChains list (which will be
1774 // the operands of the generated TokenFactor) if it's not an interior node.
1775 SDNode *N = ChainNodesMatched[i];
1776 if (N->getOpcode() != ISD::TokenFactor) {
1777 if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N))
1780 // Otherwise, add the input chain.
1781 SDValue InChain = ChainNodesMatched[i]->getOperand(0);
1782 assert(InChain.getValueType() == MVT::Other && "Not a chain");
1783 InputChains.push_back(InChain);
1787 // If we have a token factor, we want to add all inputs of the token factor
1788 // that are not part of the pattern we're matching.
1789 for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) {
1790 if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(),
1791 N->getOperand(op).getNode()))
1792 InputChains.push_back(N->getOperand(op));
1797 if (InputChains.size() == 1)
1798 return InputChains[0];
1799 return CurDAG->getNode(ISD::TokenFactor, ChainNodesMatched[0]->getDebugLoc(),
1800 MVT::Other, &InputChains[0], InputChains.size());
1803 /// MorphNode - Handle morphing a node in place for the selector.
1804 SDNode *SelectionDAGISel::
1805 MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList,
1806 const SDValue *Ops, unsigned NumOps, unsigned EmitNodeInfo) {
1807 // It is possible we're using MorphNodeTo to replace a node with no
1808 // normal results with one that has a normal result (or we could be
1809 // adding a chain) and the input could have flags and chains as well.
1810 // In this case we need to shift the operands down.
1811 // FIXME: This is a horrible hack and broken in obscure cases, no worse
1812 // than the old isel though.
1813 int OldFlagResultNo = -1, OldChainResultNo = -1;
1815 unsigned NTMNumResults = Node->getNumValues();
1816 if (Node->getValueType(NTMNumResults-1) == MVT::Flag) {
1817 OldFlagResultNo = NTMNumResults-1;
1818 if (NTMNumResults != 1 &&
1819 Node->getValueType(NTMNumResults-2) == MVT::Other)
1820 OldChainResultNo = NTMNumResults-2;
1821 } else if (Node->getValueType(NTMNumResults-1) == MVT::Other)
1822 OldChainResultNo = NTMNumResults-1;
1824 // Call the underlying SelectionDAG routine to do the transmogrification. Note
1825 // that this deletes operands of the old node that become dead.
1826 SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops, NumOps);
1828 // MorphNodeTo can operate in two ways: if an existing node with the
1829 // specified operands exists, it can just return it. Otherwise, it
1830 // updates the node in place to have the requested operands.
1832 // If we updated the node in place, reset the node ID. To the isel,
1833 // this should be just like a newly allocated machine node.
1837 unsigned ResNumResults = Res->getNumValues();
1838 // Move the flag if needed.
1839 if ((EmitNodeInfo & OPFL_FlagOutput) && OldFlagResultNo != -1 &&
1840 (unsigned)OldFlagResultNo != ResNumResults-1)
1841 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldFlagResultNo),
1842 SDValue(Res, ResNumResults-1));
1844 if ((EmitNodeInfo & OPFL_FlagOutput) != 0)
1847 // Move the chain reference if needed.
1848 if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 &&
1849 (unsigned)OldChainResultNo != ResNumResults-1)
1850 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo),
1851 SDValue(Res, ResNumResults-1));
1853 // Otherwise, no replacement happened because the node already exists. Replace
1854 // Uses of the old node with the new one.
1856 CurDAG->ReplaceAllUsesWith(Node, Res);
1861 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1862 ALWAYS_INLINE static bool
1863 CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1864 SDValue N, const SmallVectorImpl<SDValue> &RecordedNodes) {
1865 // Accept if it is exactly the same as a previously recorded node.
1866 unsigned RecNo = MatcherTable[MatcherIndex++];
1867 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
1868 return N == RecordedNodes[RecNo];
1871 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1872 ALWAYS_INLINE static bool
1873 CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1874 SelectionDAGISel &SDISel) {
1875 return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]);
1878 /// CheckNodePredicate - Implements OP_CheckNodePredicate.
1879 ALWAYS_INLINE static bool
1880 CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1881 SelectionDAGISel &SDISel, SDNode *N) {
1882 return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]);
1885 ALWAYS_INLINE static bool
1886 CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1888 uint16_t Opc = MatcherTable[MatcherIndex++];
1889 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
1890 return N->getOpcode() == Opc;
1893 ALWAYS_INLINE static bool
1894 CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1895 SDValue N, const TargetLowering &TLI) {
1896 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1897 if (N.getValueType() == VT) return true;
1899 // Handle the case when VT is iPTR.
1900 return VT == MVT::iPTR && N.getValueType() == TLI.getPointerTy();
1903 ALWAYS_INLINE static bool
1904 CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1905 SDValue N, const TargetLowering &TLI,
1907 if (ChildNo >= N.getNumOperands())
1908 return false; // Match fails if out of range child #.
1909 return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI);
1913 ALWAYS_INLINE static bool
1914 CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1916 return cast<CondCodeSDNode>(N)->get() ==
1917 (ISD::CondCode)MatcherTable[MatcherIndex++];
1920 ALWAYS_INLINE static bool
1921 CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1922 SDValue N, const TargetLowering &TLI) {
1923 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1924 if (cast<VTSDNode>(N)->getVT() == VT)
1927 // Handle the case when VT is iPTR.
1928 return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI.getPointerTy();
1931 ALWAYS_INLINE static bool
1932 CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1934 int64_t Val = MatcherTable[MatcherIndex++];
1936 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1938 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
1939 return C != 0 && C->getSExtValue() == Val;
1942 ALWAYS_INLINE static bool
1943 CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1944 SDValue N, SelectionDAGISel &SDISel) {
1945 int64_t Val = MatcherTable[MatcherIndex++];
1947 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1949 if (N->getOpcode() != ISD::AND) return false;
1951 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1952 return C != 0 && SDISel.CheckAndMask(N.getOperand(0), C, Val);
1955 ALWAYS_INLINE static bool
1956 CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1957 SDValue N, SelectionDAGISel &SDISel) {
1958 int64_t Val = MatcherTable[MatcherIndex++];
1960 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1962 if (N->getOpcode() != ISD::OR) return false;
1964 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1965 return C != 0 && SDISel.CheckOrMask(N.getOperand(0), C, Val);
1968 /// IsPredicateKnownToFail - If we know how and can do so without pushing a
1969 /// scope, evaluate the current node. If the current predicate is known to
1970 /// fail, set Result=true and return anything. If the current predicate is
1971 /// known to pass, set Result=false and return the MatcherIndex to continue
1972 /// with. If the current predicate is unknown, set Result=false and return the
1973 /// MatcherIndex to continue with.
1974 static unsigned IsPredicateKnownToFail(const unsigned char *Table,
1975 unsigned Index, SDValue N,
1976 bool &Result, SelectionDAGISel &SDISel,
1977 SmallVectorImpl<SDValue> &RecordedNodes){
1978 switch (Table[Index++]) {
1981 return Index-1; // Could not evaluate this predicate.
1982 case SelectionDAGISel::OPC_CheckSame:
1983 Result = !::CheckSame(Table, Index, N, RecordedNodes);
1985 case SelectionDAGISel::OPC_CheckPatternPredicate:
1986 Result = !::CheckPatternPredicate(Table, Index, SDISel);
1988 case SelectionDAGISel::OPC_CheckPredicate:
1989 Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode());
1991 case SelectionDAGISel::OPC_CheckOpcode:
1992 Result = !::CheckOpcode(Table, Index, N.getNode());
1994 case SelectionDAGISel::OPC_CheckType:
1995 Result = !::CheckType(Table, Index, N, SDISel.TLI);
1997 case SelectionDAGISel::OPC_CheckChild0Type:
1998 case SelectionDAGISel::OPC_CheckChild1Type:
1999 case SelectionDAGISel::OPC_CheckChild2Type:
2000 case SelectionDAGISel::OPC_CheckChild3Type:
2001 case SelectionDAGISel::OPC_CheckChild4Type:
2002 case SelectionDAGISel::OPC_CheckChild5Type:
2003 case SelectionDAGISel::OPC_CheckChild6Type:
2004 case SelectionDAGISel::OPC_CheckChild7Type:
2005 Result = !::CheckChildType(Table, Index, N, SDISel.TLI,
2006 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Type);
2008 case SelectionDAGISel::OPC_CheckCondCode:
2009 Result = !::CheckCondCode(Table, Index, N);
2011 case SelectionDAGISel::OPC_CheckValueType:
2012 Result = !::CheckValueType(Table, Index, N, SDISel.TLI);
2014 case SelectionDAGISel::OPC_CheckInteger:
2015 Result = !::CheckInteger(Table, Index, N);
2017 case SelectionDAGISel::OPC_CheckAndImm:
2018 Result = !::CheckAndImm(Table, Index, N, SDISel);
2020 case SelectionDAGISel::OPC_CheckOrImm:
2021 Result = !::CheckOrImm(Table, Index, N, SDISel);
2028 /// FailIndex - If this match fails, this is the index to continue with.
2031 /// NodeStack - The node stack when the scope was formed.
2032 SmallVector<SDValue, 4> NodeStack;
2034 /// NumRecordedNodes - The number of recorded nodes when the scope was formed.
2035 unsigned NumRecordedNodes;
2037 /// NumMatchedMemRefs - The number of matched memref entries.
2038 unsigned NumMatchedMemRefs;
2040 /// InputChain/InputFlag - The current chain/flag
2041 SDValue InputChain, InputFlag;
2043 /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty.
2044 bool HasChainNodesMatched, HasFlagResultNodesMatched;
2047 SDNode *SelectionDAGISel::
2048 SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
2049 unsigned TableSize) {
2050 // FIXME: Should these even be selected? Handle these cases in the caller?
2051 switch (NodeToMatch->getOpcode()) {
2054 case ISD::EntryToken: // These nodes remain the same.
2055 case ISD::BasicBlock:
2057 //case ISD::VALUETYPE:
2058 //case ISD::CONDCODE:
2059 case ISD::HANDLENODE:
2060 case ISD::TargetConstant:
2061 case ISD::TargetConstantFP:
2062 case ISD::TargetConstantPool:
2063 case ISD::TargetFrameIndex:
2064 case ISD::TargetExternalSymbol:
2065 case ISD::TargetBlockAddress:
2066 case ISD::TargetJumpTable:
2067 case ISD::TargetGlobalTLSAddress:
2068 case ISD::TargetGlobalAddress:
2069 case ISD::TokenFactor:
2070 case ISD::CopyFromReg:
2071 case ISD::CopyToReg:
2073 NodeToMatch->setNodeId(-1); // Mark selected.
2075 case ISD::AssertSext:
2076 case ISD::AssertZext:
2077 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0),
2078 NodeToMatch->getOperand(0));
2080 case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch);
2081 case ISD::UNDEF: return Select_UNDEF(NodeToMatch);
2084 assert(!NodeToMatch->isMachineOpcode() && "Node already selected!");
2086 // Set up the node stack with NodeToMatch as the only node on the stack.
2087 SmallVector<SDValue, 8> NodeStack;
2088 SDValue N = SDValue(NodeToMatch, 0);
2089 NodeStack.push_back(N);
2091 // MatchScopes - Scopes used when matching, if a match failure happens, this
2092 // indicates where to continue checking.
2093 SmallVector<MatchScope, 8> MatchScopes;
2095 // RecordedNodes - This is the set of nodes that have been recorded by the
2097 SmallVector<SDValue, 8> RecordedNodes;
2099 // MatchedMemRefs - This is the set of MemRef's we've seen in the input
2101 SmallVector<MachineMemOperand*, 2> MatchedMemRefs;
2103 // These are the current input chain and flag for use when generating nodes.
2104 // Various Emit operations change these. For example, emitting a copytoreg
2105 // uses and updates these.
2106 SDValue InputChain, InputFlag;
2108 // ChainNodesMatched - If a pattern matches nodes that have input/output
2109 // chains, the OPC_EmitMergeInputChains operation is emitted which indicates
2110 // which ones they are. The result is captured into this list so that we can
2111 // update the chain results when the pattern is complete.
2112 SmallVector<SDNode*, 3> ChainNodesMatched;
2113 SmallVector<SDNode*, 3> FlagResultNodesMatched;
2115 DEBUG(errs() << "ISEL: Starting pattern match on root node: ";
2116 NodeToMatch->dump(CurDAG);
2119 // Determine where to start the interpreter. Normally we start at opcode #0,
2120 // but if the state machine starts with an OPC_SwitchOpcode, then we
2121 // accelerate the first lookup (which is guaranteed to be hot) with the
2122 // OpcodeOffset table.
2123 unsigned MatcherIndex = 0;
2125 if (!OpcodeOffset.empty()) {
2126 // Already computed the OpcodeOffset table, just index into it.
2127 if (N.getOpcode() < OpcodeOffset.size())
2128 MatcherIndex = OpcodeOffset[N.getOpcode()];
2129 DEBUG(errs() << " Initial Opcode index to " << MatcherIndex << "\n");
2131 } else if (MatcherTable[0] == OPC_SwitchOpcode) {
2132 // Otherwise, the table isn't computed, but the state machine does start
2133 // with an OPC_SwitchOpcode instruction. Populate the table now, since this
2134 // is the first time we're selecting an instruction.
2137 // Get the size of this case.
2138 unsigned CaseSize = MatcherTable[Idx++];
2140 CaseSize = GetVBR(CaseSize, MatcherTable, Idx);
2141 if (CaseSize == 0) break;
2143 // Get the opcode, add the index to the table.
2144 uint16_t Opc = MatcherTable[Idx++];
2145 Opc |= (unsigned short)MatcherTable[Idx++] << 8;
2146 if (Opc >= OpcodeOffset.size())
2147 OpcodeOffset.resize((Opc+1)*2);
2148 OpcodeOffset[Opc] = Idx;
2152 // Okay, do the lookup for the first opcode.
2153 if (N.getOpcode() < OpcodeOffset.size())
2154 MatcherIndex = OpcodeOffset[N.getOpcode()];
2158 assert(MatcherIndex < TableSize && "Invalid index");
2160 unsigned CurrentOpcodeIndex = MatcherIndex;
2162 BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++];
2165 // Okay, the semantics of this operation are that we should push a scope
2166 // then evaluate the first child. However, pushing a scope only to have
2167 // the first check fail (which then pops it) is inefficient. If we can
2168 // determine immediately that the first check (or first several) will
2169 // immediately fail, don't even bother pushing a scope for them.
2173 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2174 if (NumToSkip & 128)
2175 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2176 // Found the end of the scope with no match.
2177 if (NumToSkip == 0) {
2182 FailIndex = MatcherIndex+NumToSkip;
2184 unsigned MatcherIndexOfPredicate = MatcherIndex;
2185 (void)MatcherIndexOfPredicate; // silence warning.
2187 // If we can't evaluate this predicate without pushing a scope (e.g. if
2188 // it is a 'MoveParent') or if the predicate succeeds on this node, we
2189 // push the scope and evaluate the full predicate chain.
2191 MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
2192 Result, *this, RecordedNodes);
2196 DEBUG(errs() << " Skipped scope entry (due to false predicate) at "
2197 << "index " << MatcherIndexOfPredicate
2198 << ", continuing at " << FailIndex << "\n");
2199 ++NumDAGIselRetries;
2201 // Otherwise, we know that this case of the Scope is guaranteed to fail,
2202 // move to the next case.
2203 MatcherIndex = FailIndex;
2206 // If the whole scope failed to match, bail.
2207 if (FailIndex == 0) break;
2209 // Push a MatchScope which indicates where to go if the first child fails
2211 MatchScope NewEntry;
2212 NewEntry.FailIndex = FailIndex;
2213 NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
2214 NewEntry.NumRecordedNodes = RecordedNodes.size();
2215 NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
2216 NewEntry.InputChain = InputChain;
2217 NewEntry.InputFlag = InputFlag;
2218 NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty();
2219 NewEntry.HasFlagResultNodesMatched = !FlagResultNodesMatched.empty();
2220 MatchScopes.push_back(NewEntry);
2223 case OPC_RecordNode:
2224 // Remember this node, it may end up being an operand in the pattern.
2225 RecordedNodes.push_back(N);
2228 case OPC_RecordChild0: case OPC_RecordChild1:
2229 case OPC_RecordChild2: case OPC_RecordChild3:
2230 case OPC_RecordChild4: case OPC_RecordChild5:
2231 case OPC_RecordChild6: case OPC_RecordChild7: {
2232 unsigned ChildNo = Opcode-OPC_RecordChild0;
2233 if (ChildNo >= N.getNumOperands())
2234 break; // Match fails if out of range child #.
2236 RecordedNodes.push_back(N->getOperand(ChildNo));
2239 case OPC_RecordMemRef:
2240 MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand());
2243 case OPC_CaptureFlagInput:
2244 // If the current node has an input flag, capture it in InputFlag.
2245 if (N->getNumOperands() != 0 &&
2246 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag)
2247 InputFlag = N->getOperand(N->getNumOperands()-1);
2250 case OPC_MoveChild: {
2251 unsigned ChildNo = MatcherTable[MatcherIndex++];
2252 if (ChildNo >= N.getNumOperands())
2253 break; // Match fails if out of range child #.
2254 N = N.getOperand(ChildNo);
2255 NodeStack.push_back(N);
2259 case OPC_MoveParent:
2260 // Pop the current node off the NodeStack.
2261 NodeStack.pop_back();
2262 assert(!NodeStack.empty() && "Node stack imbalance!");
2263 N = NodeStack.back();
2267 if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
2269 case OPC_CheckPatternPredicate:
2270 if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break;
2272 case OPC_CheckPredicate:
2273 if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this,
2277 case OPC_CheckComplexPat: {
2278 unsigned CPNum = MatcherTable[MatcherIndex++];
2279 unsigned RecNo = MatcherTable[MatcherIndex++];
2280 assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
2281 if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo], CPNum,
2286 case OPC_CheckOpcode:
2287 if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
2291 if (!::CheckType(MatcherTable, MatcherIndex, N, TLI)) break;
2294 case OPC_SwitchOpcode: {
2295 unsigned CurNodeOpcode = N.getOpcode();
2296 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2299 // Get the size of this case.
2300 CaseSize = MatcherTable[MatcherIndex++];
2302 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2303 if (CaseSize == 0) break;
2305 uint16_t Opc = MatcherTable[MatcherIndex++];
2306 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2308 // If the opcode matches, then we will execute this case.
2309 if (CurNodeOpcode == Opc)
2312 // Otherwise, skip over this case.
2313 MatcherIndex += CaseSize;
2316 // If no cases matched, bail out.
2317 if (CaseSize == 0) break;
2319 // Otherwise, execute the case we found.
2320 DEBUG(errs() << " OpcodeSwitch from " << SwitchStart
2321 << " to " << MatcherIndex << "\n");
2325 case OPC_SwitchType: {
2326 MVT::SimpleValueType CurNodeVT = N.getValueType().getSimpleVT().SimpleTy;
2327 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2330 // Get the size of this case.
2331 CaseSize = MatcherTable[MatcherIndex++];
2333 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2334 if (CaseSize == 0) break;
2336 MVT::SimpleValueType CaseVT =
2337 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2338 if (CaseVT == MVT::iPTR)
2339 CaseVT = TLI.getPointerTy().SimpleTy;
2341 // If the VT matches, then we will execute this case.
2342 if (CurNodeVT == CaseVT)
2345 // Otherwise, skip over this case.
2346 MatcherIndex += CaseSize;
2349 // If no cases matched, bail out.
2350 if (CaseSize == 0) break;
2352 // Otherwise, execute the case we found.
2353 DEBUG(errs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
2354 << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
2357 case OPC_CheckChild0Type: case OPC_CheckChild1Type:
2358 case OPC_CheckChild2Type: case OPC_CheckChild3Type:
2359 case OPC_CheckChild4Type: case OPC_CheckChild5Type:
2360 case OPC_CheckChild6Type: case OPC_CheckChild7Type:
2361 if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI,
2362 Opcode-OPC_CheckChild0Type))
2365 case OPC_CheckCondCode:
2366 if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
2368 case OPC_CheckValueType:
2369 if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI)) break;
2371 case OPC_CheckInteger:
2372 if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
2374 case OPC_CheckAndImm:
2375 if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
2377 case OPC_CheckOrImm:
2378 if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
2381 case OPC_CheckFoldableChainNode: {
2382 assert(NodeStack.size() != 1 && "No parent node");
2383 // Verify that all intermediate nodes between the root and this one have
2385 bool HasMultipleUses = false;
2386 for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i)
2387 if (!NodeStack[i].hasOneUse()) {
2388 HasMultipleUses = true;
2391 if (HasMultipleUses) break;
2393 // Check to see that the target thinks this is profitable to fold and that
2394 // we can fold it without inducing cycles in the graph.
2395 if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2397 !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2398 NodeToMatch, true/*We validate our own chains*/))
2403 case OPC_EmitInteger: {
2404 MVT::SimpleValueType VT =
2405 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2406 int64_t Val = MatcherTable[MatcherIndex++];
2408 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2409 RecordedNodes.push_back(CurDAG->getTargetConstant(Val, VT));
2412 case OPC_EmitRegister: {
2413 MVT::SimpleValueType VT =
2414 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2415 unsigned RegNo = MatcherTable[MatcherIndex++];
2416 RecordedNodes.push_back(CurDAG->getRegister(RegNo, VT));
2420 case OPC_EmitConvertToTarget: {
2421 // Convert from IMM/FPIMM to target version.
2422 unsigned RecNo = MatcherTable[MatcherIndex++];
2423 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2424 SDValue Imm = RecordedNodes[RecNo];
2426 if (Imm->getOpcode() == ISD::Constant) {
2427 int64_t Val = cast<ConstantSDNode>(Imm)->getZExtValue();
2428 Imm = CurDAG->getTargetConstant(Val, Imm.getValueType());
2429 } else if (Imm->getOpcode() == ISD::ConstantFP) {
2430 const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue();
2431 Imm = CurDAG->getTargetConstantFP(*Val, Imm.getValueType());
2434 RecordedNodes.push_back(Imm);
2438 case OPC_EmitMergeInputChains1_0: // OPC_EmitMergeInputChains, 1, 0
2439 case OPC_EmitMergeInputChains1_1: { // OPC_EmitMergeInputChains, 1, 1
2440 // These are space-optimized forms of OPC_EmitMergeInputChains.
2441 assert(InputChain.getNode() == 0 &&
2442 "EmitMergeInputChains should be the first chain producing node");
2443 assert(ChainNodesMatched.empty() &&
2444 "Should only have one EmitMergeInputChains per match");
2446 // Read all of the chained nodes.
2447 unsigned RecNo = Opcode == OPC_EmitMergeInputChains1_1;
2448 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2449 ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode());
2451 // FIXME: What if other value results of the node have uses not matched
2453 if (ChainNodesMatched.back() != NodeToMatch &&
2454 !RecordedNodes[RecNo].hasOneUse()) {
2455 ChainNodesMatched.clear();
2459 // Merge the input chains if they are not intra-pattern references.
2460 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2462 if (InputChain.getNode() == 0)
2463 break; // Failed to merge.
2467 case OPC_EmitMergeInputChains: {
2468 assert(InputChain.getNode() == 0 &&
2469 "EmitMergeInputChains should be the first chain producing node");
2470 // This node gets a list of nodes we matched in the input that have
2471 // chains. We want to token factor all of the input chains to these nodes
2472 // together. However, if any of the input chains is actually one of the
2473 // nodes matched in this pattern, then we have an intra-match reference.
2474 // Ignore these because the newly token factored chain should not refer to
2476 unsigned NumChains = MatcherTable[MatcherIndex++];
2477 assert(NumChains != 0 && "Can't TF zero chains");
2479 assert(ChainNodesMatched.empty() &&
2480 "Should only have one EmitMergeInputChains per match");
2482 // Read all of the chained nodes.
2483 for (unsigned i = 0; i != NumChains; ++i) {
2484 unsigned RecNo = MatcherTable[MatcherIndex++];
2485 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2486 ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode());
2488 // FIXME: What if other value results of the node have uses not matched
2490 if (ChainNodesMatched.back() != NodeToMatch &&
2491 !RecordedNodes[RecNo].hasOneUse()) {
2492 ChainNodesMatched.clear();
2497 // If the inner loop broke out, the match fails.
2498 if (ChainNodesMatched.empty())
2501 // Merge the input chains if they are not intra-pattern references.
2502 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2504 if (InputChain.getNode() == 0)
2505 break; // Failed to merge.
2510 case OPC_EmitCopyToReg: {
2511 unsigned RecNo = MatcherTable[MatcherIndex++];
2512 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2513 unsigned DestPhysReg = MatcherTable[MatcherIndex++];
2515 if (InputChain.getNode() == 0)
2516 InputChain = CurDAG->getEntryNode();
2518 InputChain = CurDAG->getCopyToReg(InputChain, NodeToMatch->getDebugLoc(),
2519 DestPhysReg, RecordedNodes[RecNo],
2522 InputFlag = InputChain.getValue(1);
2526 case OPC_EmitNodeXForm: {
2527 unsigned XFormNo = MatcherTable[MatcherIndex++];
2528 unsigned RecNo = MatcherTable[MatcherIndex++];
2529 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2530 RecordedNodes.push_back(RunSDNodeXForm(RecordedNodes[RecNo], XFormNo));
2535 case OPC_MorphNodeTo: {
2536 uint16_t TargetOpc = MatcherTable[MatcherIndex++];
2537 TargetOpc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2538 unsigned EmitNodeInfo = MatcherTable[MatcherIndex++];
2539 // Get the result VT list.
2540 unsigned NumVTs = MatcherTable[MatcherIndex++];
2541 SmallVector<EVT, 4> VTs;
2542 for (unsigned i = 0; i != NumVTs; ++i) {
2543 MVT::SimpleValueType VT =
2544 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2545 if (VT == MVT::iPTR) VT = TLI.getPointerTy().SimpleTy;
2549 if (EmitNodeInfo & OPFL_Chain)
2550 VTs.push_back(MVT::Other);
2551 if (EmitNodeInfo & OPFL_FlagOutput)
2552 VTs.push_back(MVT::Flag);
2554 // This is hot code, so optimize the two most common cases of 1 and 2
2557 if (VTs.size() == 1)
2558 VTList = CurDAG->getVTList(VTs[0]);
2559 else if (VTs.size() == 2)
2560 VTList = CurDAG->getVTList(VTs[0], VTs[1]);
2562 VTList = CurDAG->getVTList(VTs.data(), VTs.size());
2564 // Get the operand list.
2565 unsigned NumOps = MatcherTable[MatcherIndex++];
2566 SmallVector<SDValue, 8> Ops;
2567 for (unsigned i = 0; i != NumOps; ++i) {
2568 unsigned RecNo = MatcherTable[MatcherIndex++];
2570 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2572 assert(RecNo < RecordedNodes.size() && "Invalid EmitNode");
2573 Ops.push_back(RecordedNodes[RecNo]);
2576 // If there are variadic operands to add, handle them now.
2577 if (EmitNodeInfo & OPFL_VariadicInfo) {
2578 // Determine the start index to copy from.
2579 unsigned FirstOpToCopy = getNumFixedFromVariadicInfo(EmitNodeInfo);
2580 FirstOpToCopy += (EmitNodeInfo & OPFL_Chain) ? 1 : 0;
2581 assert(NodeToMatch->getNumOperands() >= FirstOpToCopy &&
2582 "Invalid variadic node");
2583 // Copy all of the variadic operands, not including a potential flag
2585 for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands();
2587 SDValue V = NodeToMatch->getOperand(i);
2588 if (V.getValueType() == MVT::Flag) break;
2593 // If this has chain/flag inputs, add them.
2594 if (EmitNodeInfo & OPFL_Chain)
2595 Ops.push_back(InputChain);
2596 if ((EmitNodeInfo & OPFL_FlagInput) && InputFlag.getNode() != 0)
2597 Ops.push_back(InputFlag);
2601 if (Opcode != OPC_MorphNodeTo) {
2602 // If this is a normal EmitNode command, just create the new node and
2603 // add the results to the RecordedNodes list.
2604 Res = CurDAG->getMachineNode(TargetOpc, NodeToMatch->getDebugLoc(),
2605 VTList, Ops.data(), Ops.size());
2607 // Add all the non-flag/non-chain results to the RecordedNodes list.
2608 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
2609 if (VTs[i] == MVT::Other || VTs[i] == MVT::Flag) break;
2610 RecordedNodes.push_back(SDValue(Res, i));
2614 Res = MorphNode(NodeToMatch, TargetOpc, VTList, Ops.data(), Ops.size(),
2618 // If the node had chain/flag results, update our notion of the current
2620 if (EmitNodeInfo & OPFL_FlagOutput) {
2621 InputFlag = SDValue(Res, VTs.size()-1);
2622 if (EmitNodeInfo & OPFL_Chain)
2623 InputChain = SDValue(Res, VTs.size()-2);
2624 } else if (EmitNodeInfo & OPFL_Chain)
2625 InputChain = SDValue(Res, VTs.size()-1);
2627 // If the OPFL_MemRefs flag is set on this node, slap all of the
2628 // accumulated memrefs onto it.
2630 // FIXME: This is vastly incorrect for patterns with multiple outputs
2631 // instructions that access memory and for ComplexPatterns that match
2633 if (EmitNodeInfo & OPFL_MemRefs) {
2634 MachineSDNode::mmo_iterator MemRefs =
2635 MF->allocateMemRefsArray(MatchedMemRefs.size());
2636 std::copy(MatchedMemRefs.begin(), MatchedMemRefs.end(), MemRefs);
2637 cast<MachineSDNode>(Res)
2638 ->setMemRefs(MemRefs, MemRefs + MatchedMemRefs.size());
2642 << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created")
2643 << " node: "; Res->dump(CurDAG); errs() << "\n");
2645 // If this was a MorphNodeTo then we're completely done!
2646 if (Opcode == OPC_MorphNodeTo) {
2647 // Update chain and flag uses.
2648 UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched,
2649 InputFlag, FlagResultNodesMatched, true);
2656 case OPC_MarkFlagResults: {
2657 unsigned NumNodes = MatcherTable[MatcherIndex++];
2659 // Read and remember all the flag-result nodes.
2660 for (unsigned i = 0; i != NumNodes; ++i) {
2661 unsigned RecNo = MatcherTable[MatcherIndex++];
2663 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2665 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2666 FlagResultNodesMatched.push_back(RecordedNodes[RecNo].getNode());
2671 case OPC_CompleteMatch: {
2672 // The match has been completed, and any new nodes (if any) have been
2673 // created. Patch up references to the matched dag to use the newly
2675 unsigned NumResults = MatcherTable[MatcherIndex++];
2677 for (unsigned i = 0; i != NumResults; ++i) {
2678 unsigned ResSlot = MatcherTable[MatcherIndex++];
2680 ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex);
2682 assert(ResSlot < RecordedNodes.size() && "Invalid CheckSame");
2683 SDValue Res = RecordedNodes[ResSlot];
2685 assert(i < NodeToMatch->getNumValues() &&
2686 NodeToMatch->getValueType(i) != MVT::Other &&
2687 NodeToMatch->getValueType(i) != MVT::Flag &&
2688 "Invalid number of results to complete!");
2689 assert((NodeToMatch->getValueType(i) == Res.getValueType() ||
2690 NodeToMatch->getValueType(i) == MVT::iPTR ||
2691 Res.getValueType() == MVT::iPTR ||
2692 NodeToMatch->getValueType(i).getSizeInBits() ==
2693 Res.getValueType().getSizeInBits()) &&
2694 "invalid replacement");
2695 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, i), Res);
2698 // If the root node defines a flag, add it to the flag nodes to update
2700 if (NodeToMatch->getValueType(NodeToMatch->getNumValues()-1) == MVT::Flag)
2701 FlagResultNodesMatched.push_back(NodeToMatch);
2703 // Update chain and flag uses.
2704 UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched,
2705 InputFlag, FlagResultNodesMatched, false);
2707 assert(NodeToMatch->use_empty() &&
2708 "Didn't replace all uses of the node?");
2710 // FIXME: We just return here, which interacts correctly with SelectRoot
2711 // above. We should fix this to not return an SDNode* anymore.
2716 // If the code reached this point, then the match failed. See if there is
2717 // another child to try in the current 'Scope', otherwise pop it until we
2718 // find a case to check.
2719 DEBUG(errs() << " Match failed at index " << CurrentOpcodeIndex << "\n");
2720 ++NumDAGIselRetries;
2722 if (MatchScopes.empty()) {
2723 CannotYetSelect(NodeToMatch);
2727 // Restore the interpreter state back to the point where the scope was
2729 MatchScope &LastScope = MatchScopes.back();
2730 RecordedNodes.resize(LastScope.NumRecordedNodes);
2732 NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end());
2733 N = NodeStack.back();
2735 if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size())
2736 MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
2737 MatcherIndex = LastScope.FailIndex;
2739 DEBUG(errs() << " Continuing at " << MatcherIndex << "\n");
2741 InputChain = LastScope.InputChain;
2742 InputFlag = LastScope.InputFlag;
2743 if (!LastScope.HasChainNodesMatched)
2744 ChainNodesMatched.clear();
2745 if (!LastScope.HasFlagResultNodesMatched)
2746 FlagResultNodesMatched.clear();
2748 // Check to see what the offset is at the new MatcherIndex. If it is zero
2749 // we have reached the end of this scope, otherwise we have another child
2750 // in the current scope to try.
2751 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2752 if (NumToSkip & 128)
2753 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2755 // If we have another child in this scope to match, update FailIndex and
2757 if (NumToSkip != 0) {
2758 LastScope.FailIndex = MatcherIndex+NumToSkip;
2762 // End of this scope, pop it and try the next child in the containing
2764 MatchScopes.pop_back();
2771 void SelectionDAGISel::CannotYetSelect(SDNode *N) {
2773 raw_string_ostream Msg(msg);
2774 Msg << "Cannot yet select: ";
2776 if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN &&
2777 N->getOpcode() != ISD::INTRINSIC_WO_CHAIN &&
2778 N->getOpcode() != ISD::INTRINSIC_VOID) {
2779 N->printrFull(Msg, CurDAG);
2781 bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other;
2783 cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue();
2784 if (iid < Intrinsic::num_intrinsics)
2785 Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid);
2786 else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo())
2787 Msg << "target intrinsic %" << TII->getName(iid);
2789 Msg << "unknown intrinsic #" << iid;
2791 llvm_report_error(Msg.str());
2794 char SelectionDAGISel::ID = 0;