1 //===-- SelectionDAGISel.cpp - Implement the SelectionDAGISel class -------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the SelectionDAGISel class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "isel"
15 #include "ScheduleDAGSDNodes.h"
16 #include "SelectionDAGBuilder.h"
17 #include "FunctionLoweringInfo.h"
18 #include "llvm/CodeGen/SelectionDAGISel.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/DebugInfo.h"
21 #include "llvm/Constants.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Function.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/InlineAsm.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/FastISel.h"
32 #include "llvm/CodeGen/GCStrategy.h"
33 #include "llvm/CodeGen/GCMetadata.h"
34 #include "llvm/CodeGen/MachineFunction.h"
35 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
36 #include "llvm/CodeGen/MachineFrameInfo.h"
37 #include "llvm/CodeGen/MachineInstrBuilder.h"
38 #include "llvm/CodeGen/MachineJumpTableInfo.h"
39 #include "llvm/CodeGen/MachineModuleInfo.h"
40 #include "llvm/CodeGen/MachineRegisterInfo.h"
41 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
42 #include "llvm/CodeGen/SchedulerRegistry.h"
43 #include "llvm/CodeGen/SelectionDAG.h"
44 #include "llvm/CodeGen/DwarfWriter.h"
45 #include "llvm/Target/TargetRegisterInfo.h"
46 #include "llvm/Target/TargetData.h"
47 #include "llvm/Target/TargetFrameInfo.h"
48 #include "llvm/Target/TargetIntrinsicInfo.h"
49 #include "llvm/Target/TargetInstrInfo.h"
50 #include "llvm/Target/TargetLowering.h"
51 #include "llvm/Target/TargetMachine.h"
52 #include "llvm/Target/TargetOptions.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/ErrorHandling.h"
56 #include "llvm/Support/MathExtras.h"
57 #include "llvm/Support/Timer.h"
58 #include "llvm/Support/raw_ostream.h"
59 #include "llvm/ADT/Statistic.h"
63 STATISTIC(NumFastIselFailures, "Number of instructions fast isel failed on");
66 EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
67 cl::desc("Enable verbose messages in the \"fast\" "
68 "instruction selector"));
70 EnableFastISelAbort("fast-isel-abort", cl::Hidden,
71 cl::desc("Enable abort calls when \"fast\" instruction fails"));
73 SchedLiveInCopies("schedule-livein-copies", cl::Hidden,
74 cl::desc("Schedule copies of livein registers"),
79 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
80 cl::desc("Pop up a window to show dags before the first "
83 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
84 cl::desc("Pop up a window to show dags before legalize types"));
86 ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
87 cl::desc("Pop up a window to show dags before legalize"));
89 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
90 cl::desc("Pop up a window to show dags before the second "
93 ViewDAGCombineLT("view-dag-combine-lt-dags", cl::Hidden,
94 cl::desc("Pop up a window to show dags before the post legalize types"
95 " dag combine pass"));
97 ViewISelDAGs("view-isel-dags", cl::Hidden,
98 cl::desc("Pop up a window to show isel dags as they are selected"));
100 ViewSchedDAGs("view-sched-dags", cl::Hidden,
101 cl::desc("Pop up a window to show sched dags as they are processed"));
103 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
104 cl::desc("Pop up a window to show SUnit dags after they are processed"));
106 static const bool ViewDAGCombine1 = false,
107 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false,
108 ViewDAGCombine2 = false,
109 ViewDAGCombineLT = false,
110 ViewISelDAGs = false, ViewSchedDAGs = false,
111 ViewSUnitDAGs = false;
114 //===---------------------------------------------------------------------===//
116 /// RegisterScheduler class - Track the registration of instruction schedulers.
118 //===---------------------------------------------------------------------===//
119 MachinePassRegistry RegisterScheduler::Registry;
121 //===---------------------------------------------------------------------===//
123 /// ISHeuristic command line option for instruction schedulers.
125 //===---------------------------------------------------------------------===//
126 static cl::opt<RegisterScheduler::FunctionPassCtor, false,
127 RegisterPassParser<RegisterScheduler> >
128 ISHeuristic("pre-RA-sched",
129 cl::init(&createDefaultScheduler),
130 cl::desc("Instruction schedulers available (before register"
133 static RegisterScheduler
134 defaultListDAGScheduler("default", "Best scheduler for the target",
135 createDefaultScheduler);
138 //===--------------------------------------------------------------------===//
139 /// createDefaultScheduler - This creates an instruction scheduler appropriate
141 ScheduleDAGSDNodes* createDefaultScheduler(SelectionDAGISel *IS,
142 CodeGenOpt::Level OptLevel) {
143 const TargetLowering &TLI = IS->getTargetLowering();
145 if (OptLevel == CodeGenOpt::None)
146 return createFastDAGScheduler(IS, OptLevel);
147 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency)
148 return createTDListDAGScheduler(IS, OptLevel);
149 assert(TLI.getSchedulingPreference() ==
150 TargetLowering::SchedulingForRegPressure && "Unknown sched type!");
151 return createBURRListDAGScheduler(IS, OptLevel);
155 // EmitInstrWithCustomInserter - This method should be implemented by targets
156 // that mark instructions with the 'usesCustomInserter' flag. These
157 // instructions are special in various ways, which require special support to
158 // insert. The specified MachineInstr is created but not inserted into any
159 // basic blocks, and this method is called to expand it into a sequence of
160 // instructions, potentially also creating new basic blocks and control flow.
161 // When new basic blocks are inserted and the edges from MBB to its successors
162 // are modified, the method should insert pairs of <OldSucc, NewSucc> into the
164 MachineBasicBlock *TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
165 MachineBasicBlock *MBB,
166 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
168 dbgs() << "If a target marks an instruction with "
169 "'usesCustomInserter', it must implement "
170 "TargetLowering::EmitInstrWithCustomInserter!";
176 /// EmitLiveInCopy - Emit a copy for a live in physical register. If the
177 /// physical register has only a single copy use, then coalesced the copy
179 static void EmitLiveInCopy(MachineBasicBlock *MBB,
180 MachineBasicBlock::iterator &InsertPos,
181 unsigned VirtReg, unsigned PhysReg,
182 const TargetRegisterClass *RC,
183 DenseMap<MachineInstr*, unsigned> &CopyRegMap,
184 const MachineRegisterInfo &MRI,
185 const TargetRegisterInfo &TRI,
186 const TargetInstrInfo &TII) {
187 unsigned NumUses = 0;
188 MachineInstr *UseMI = NULL;
189 for (MachineRegisterInfo::use_iterator UI = MRI.use_begin(VirtReg),
190 UE = MRI.use_end(); UI != UE; ++UI) {
196 // If the number of uses is not one, or the use is not a move instruction,
197 // don't coalesce. Also, only coalesce away a virtual register to virtual
199 bool Coalesced = false;
200 unsigned SrcReg, DstReg, SrcSubReg, DstSubReg;
202 TII.isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubReg, DstSubReg) &&
203 TargetRegisterInfo::isVirtualRegister(DstReg)) {
208 // Now find an ideal location to insert the copy.
209 MachineBasicBlock::iterator Pos = InsertPos;
210 while (Pos != MBB->begin()) {
211 MachineInstr *PrevMI = prior(Pos);
212 DenseMap<MachineInstr*, unsigned>::iterator RI = CopyRegMap.find(PrevMI);
213 // copyRegToReg might emit multiple instructions to do a copy.
214 unsigned CopyDstReg = (RI == CopyRegMap.end()) ? 0 : RI->second;
215 if (CopyDstReg && !TRI.regsOverlap(CopyDstReg, PhysReg))
216 // This is what the BB looks like right now:
221 // We want to insert "r1025 = mov r1". Inserting this copy below the
222 // move to r1024 makes it impossible for that move to be coalesced.
229 break; // Woot! Found a good location.
233 bool Emitted = TII.copyRegToReg(*MBB, Pos, VirtReg, PhysReg, RC, RC);
234 assert(Emitted && "Unable to issue a live-in copy instruction!\n");
237 CopyRegMap.insert(std::make_pair(prior(Pos), VirtReg));
239 if (&*InsertPos == UseMI) ++InsertPos;
244 /// EmitLiveInCopies - If this is the first basic block in the function,
245 /// and if it has live ins that need to be copied into vregs, emit the
246 /// copies into the block.
247 static void EmitLiveInCopies(MachineBasicBlock *EntryMBB,
248 const MachineRegisterInfo &MRI,
249 const TargetRegisterInfo &TRI,
250 const TargetInstrInfo &TII) {
251 if (SchedLiveInCopies) {
252 // Emit the copies at a heuristically-determined location in the block.
253 DenseMap<MachineInstr*, unsigned> CopyRegMap;
254 MachineBasicBlock::iterator InsertPos = EntryMBB->begin();
255 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
256 E = MRI.livein_end(); LI != E; ++LI)
258 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
259 EmitLiveInCopy(EntryMBB, InsertPos, LI->second, LI->first,
260 RC, CopyRegMap, MRI, TRI, TII);
263 // Emit the copies into the top of the block.
264 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
265 E = MRI.livein_end(); LI != E; ++LI)
267 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
268 bool Emitted = TII.copyRegToReg(*EntryMBB, EntryMBB->begin(),
269 LI->second, LI->first, RC, RC);
270 assert(Emitted && "Unable to issue a live-in copy instruction!\n");
276 //===----------------------------------------------------------------------===//
277 // SelectionDAGISel code
278 //===----------------------------------------------------------------------===//
280 SelectionDAGISel::SelectionDAGISel(TargetMachine &tm, CodeGenOpt::Level OL) :
281 MachineFunctionPass(&ID), TM(tm), TLI(*tm.getTargetLowering()),
282 FuncInfo(new FunctionLoweringInfo(TLI)),
283 CurDAG(new SelectionDAG(TLI, *FuncInfo)),
284 SDB(new SelectionDAGBuilder(*CurDAG, TLI, *FuncInfo, OL)),
290 SelectionDAGISel::~SelectionDAGISel() {
296 unsigned SelectionDAGISel::MakeReg(EVT VT) {
297 return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT));
300 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
301 AU.addRequired<AliasAnalysis>();
302 AU.addPreserved<AliasAnalysis>();
303 AU.addRequired<GCModuleInfo>();
304 AU.addPreserved<GCModuleInfo>();
305 AU.addRequired<DwarfWriter>();
306 AU.addPreserved<DwarfWriter>();
307 MachineFunctionPass::getAnalysisUsage(AU);
310 bool SelectionDAGISel::runOnMachineFunction(MachineFunction &mf) {
311 Function &Fn = *mf.getFunction();
313 // Do some sanity-checking on the command-line options.
314 assert((!EnableFastISelVerbose || EnableFastISel) &&
315 "-fast-isel-verbose requires -fast-isel");
316 assert((!EnableFastISelAbort || EnableFastISel) &&
317 "-fast-isel-abort requires -fast-isel");
319 // Get alias analysis for load/store combining.
320 AA = &getAnalysis<AliasAnalysis>();
323 const TargetInstrInfo &TII = *TM.getInstrInfo();
324 const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
327 GFI = &getAnalysis<GCModuleInfo>().getFunctionInfo(Fn);
330 RegInfo = &MF->getRegInfo();
331 DEBUG(dbgs() << "\n\n\n=== " << Fn.getName() << "\n");
333 MachineModuleInfo *MMI = getAnalysisIfAvailable<MachineModuleInfo>();
334 DwarfWriter *DW = getAnalysisIfAvailable<DwarfWriter>();
335 CurDAG->init(*MF, MMI, DW);
336 FuncInfo->set(Fn, *MF, EnableFastISel);
339 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
340 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(I->getTerminator()))
342 FuncInfo->MBBMap[Invoke->getSuccessor(1)]->setIsLandingPad();
344 SelectAllBasicBlocks(Fn, *MF, MMI, DW, TII);
346 // If the first basic block in the function has live ins that need to be
347 // copied into vregs, emit the copies into the top of the block before
348 // emitting the code for the block.
349 EmitLiveInCopies(MF->begin(), *RegInfo, TRI, TII);
351 // Add function live-ins to entry block live-in set.
352 for (MachineRegisterInfo::livein_iterator I = RegInfo->livein_begin(),
353 E = RegInfo->livein_end(); I != E; ++I)
354 MF->begin()->addLiveIn(I->first);
357 assert(FuncInfo->CatchInfoFound.size() == FuncInfo->CatchInfoLost.size() &&
358 "Not all catch info was assigned to a landing pad!");
366 /// SetDebugLoc - Update MF's and SDB's DebugLocs if debug information is
367 /// attached with this instruction.
368 static void SetDebugLoc(unsigned MDDbgKind, Instruction *I,
369 SelectionDAGBuilder *SDB,
370 FastISel *FastIS, MachineFunction *MF) {
371 if (MDNode *Dbg = I->getMetadata(MDDbgKind)) {
372 DILocation DILoc(Dbg);
373 DebugLoc Loc = ExtractDebugLocation(DILoc, MF->getDebugLocInfo());
375 SDB->setCurDebugLoc(Loc);
378 FastIS->setCurDebugLoc(Loc);
380 // If the function doesn't have a default debug location yet, set
381 // it. This is kind of a hack.
382 if (MF->getDefaultDebugLoc().isUnknown())
383 MF->setDefaultDebugLoc(Loc);
387 /// ResetDebugLoc - Set MF's and SDB's DebugLocs to Unknown.
388 static void ResetDebugLoc(SelectionDAGBuilder *SDB, FastISel *FastIS) {
389 SDB->setCurDebugLoc(DebugLoc::getUnknownLoc());
391 FastIS->setCurDebugLoc(DebugLoc::getUnknownLoc());
394 void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB,
395 BasicBlock::iterator Begin,
396 BasicBlock::iterator End,
398 SDB->setCurrentBasicBlock(BB);
399 unsigned MDDbgKind = LLVMBB->getContext().getMDKindID("dbg");
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(MDDbgKind, 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(MDDbgKind, 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 unsigned MDDbgKind = Fn.getContext().getMDKindID("dbg");
867 // Iterate over all basic blocks in the function.
868 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
869 BasicBlock *LLVMBB = &*I;
870 BB = FuncInfo->MBBMap[LLVMBB];
872 BasicBlock::iterator const Begin = LLVMBB->begin();
873 BasicBlock::iterator const End = LLVMBB->end();
874 BasicBlock::iterator BI = Begin;
876 // Lower any arguments needed in this block if this is the entry block.
877 bool SuppressFastISel = false;
878 if (LLVMBB == &Fn.getEntryBlock()) {
879 LowerArguments(LLVMBB);
881 // If any of the arguments has the byval attribute, forgo
882 // fast-isel in the entry block.
885 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end();
887 if (Fn.paramHasAttr(j, Attribute::ByVal)) {
888 if (EnableFastISelVerbose || EnableFastISelAbort)
889 dbgs() << "FastISel skips entry block due to byval argument\n";
890 SuppressFastISel = true;
896 if (MMI && BB->isLandingPad()) {
897 // Add a label to mark the beginning of the landing pad. Deletion of the
898 // landing pad can thus be detected via the MachineModuleInfo.
899 MCSymbol *Label = MMI->addLandingPad(BB);
901 const TargetInstrDesc &II = TII.get(TargetOpcode::EH_LABEL);
902 BuildMI(BB, SDB->getCurDebugLoc(), II).addSym(Label);
904 // Mark exception register as live in.
905 unsigned Reg = TLI.getExceptionAddressRegister();
906 if (Reg) BB->addLiveIn(Reg);
908 // Mark exception selector register as live in.
909 Reg = TLI.getExceptionSelectorRegister();
910 if (Reg) BB->addLiveIn(Reg);
912 // FIXME: Hack around an exception handling flaw (PR1508): the personality
913 // function and list of typeids logically belong to the invoke (or, if you
914 // like, the basic block containing the invoke), and need to be associated
915 // with it in the dwarf exception handling tables. Currently however the
916 // information is provided by an intrinsic (eh.selector) that can be moved
917 // to unexpected places by the optimizers: if the unwind edge is critical,
918 // then breaking it can result in the intrinsics being in the successor of
919 // the landing pad, not the landing pad itself. This results
920 // in exceptions not being caught because no typeids are associated with
921 // the invoke. This may not be the only way things can go wrong, but it
922 // is the only way we try to work around for the moment.
923 BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
925 if (Br && Br->isUnconditional()) { // Critical edge?
926 BasicBlock::iterator I, E;
927 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
928 if (isa<EHSelectorInst>(I))
932 // No catch info found - try to extract some from the successor.
933 CopyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, *FuncInfo);
937 // Before doing SelectionDAG ISel, see if FastISel has been requested.
938 if (FastIS && !SuppressFastISel) {
939 // Emit code for any incoming arguments. This must happen before
940 // beginning FastISel on the entry block.
941 if (LLVMBB == &Fn.getEntryBlock()) {
942 CurDAG->setRoot(SDB->getControlRoot());
946 FastIS->startNewBlock(BB);
947 // Do FastISel on as many instructions as possible.
948 for (; BI != End; ++BI) {
949 // Just before the terminator instruction, insert instructions to
950 // feed PHI nodes in successor blocks.
951 if (isa<TerminatorInst>(BI))
952 if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, FastIS)) {
953 ++NumFastIselFailures;
954 ResetDebugLoc(SDB, FastIS);
955 if (EnableFastISelVerbose || EnableFastISelAbort) {
956 dbgs() << "FastISel miss: ";
959 assert(!EnableFastISelAbort &&
960 "FastISel didn't handle a PHI in a successor");
964 SetDebugLoc(MDDbgKind, BI, SDB, FastIS, &MF);
966 // Try to select the instruction with FastISel.
967 if (FastIS->SelectInstruction(BI)) {
968 ResetDebugLoc(SDB, FastIS);
972 // Clear out the debug location so that it doesn't carry over to
973 // unrelated instructions.
974 ResetDebugLoc(SDB, FastIS);
976 // Then handle certain instructions as single-LLVM-Instruction blocks.
977 if (isa<CallInst>(BI)) {
978 ++NumFastIselFailures;
979 if (EnableFastISelVerbose || EnableFastISelAbort) {
980 dbgs() << "FastISel missed call: ";
984 if (!BI->getType()->isVoidTy()) {
985 unsigned &R = FuncInfo->ValueMap[BI];
987 R = FuncInfo->CreateRegForValue(BI);
990 bool HadTailCall = false;
991 SelectBasicBlock(LLVMBB, BI, llvm::next(BI), HadTailCall);
993 // If the call was emitted as a tail call, we're done with the block.
999 // If the instruction was codegen'd with multiple blocks,
1000 // inform the FastISel object where to resume inserting.
1001 FastIS->setCurrentBlock(BB);
1005 // Otherwise, give up on FastISel for the rest of the block.
1006 // For now, be a little lenient about non-branch terminators.
1007 if (!isa<TerminatorInst>(BI) || isa<BranchInst>(BI)) {
1008 ++NumFastIselFailures;
1009 if (EnableFastISelVerbose || EnableFastISelAbort) {
1010 dbgs() << "FastISel miss: ";
1013 if (EnableFastISelAbort)
1014 // The "fast" selector couldn't handle something and bailed.
1015 // For the purpose of debugging, just abort.
1016 llvm_unreachable("FastISel didn't select the entire block");
1022 // Run SelectionDAG instruction selection on the remainder of the block
1023 // not handled by FastISel. If FastISel is not run, this is the entire
1027 SelectBasicBlock(LLVMBB, BI, End, HadTailCall);
1037 SelectionDAGISel::FinishBasicBlock() {
1039 DEBUG(dbgs() << "Target-post-processed machine code:\n");
1042 DEBUG(dbgs() << "Total amount of phi nodes to update: "
1043 << SDB->PHINodesToUpdate.size() << "\n");
1044 DEBUG(for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i)
1045 dbgs() << "Node " << i << " : ("
1046 << SDB->PHINodesToUpdate[i].first
1047 << ", " << SDB->PHINodesToUpdate[i].second << ")\n");
1049 // Next, now that we know what the last MBB the LLVM BB expanded is, update
1050 // PHI nodes in successors.
1051 if (SDB->SwitchCases.empty() &&
1052 SDB->JTCases.empty() &&
1053 SDB->BitTestCases.empty()) {
1054 for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) {
1055 MachineInstr *PHI = SDB->PHINodesToUpdate[i].first;
1056 assert(PHI->isPHI() &&
1057 "This is not a machine PHI node that we are updating!");
1058 if (!BB->isSuccessor(PHI->getParent()))
1060 PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second,
1062 PHI->addOperand(MachineOperand::CreateMBB(BB));
1064 SDB->PHINodesToUpdate.clear();
1068 for (unsigned i = 0, e = SDB->BitTestCases.size(); i != e; ++i) {
1069 // Lower header first, if it wasn't already lowered
1070 if (!SDB->BitTestCases[i].Emitted) {
1071 // Set the current basic block to the mbb we wish to insert the code into
1072 BB = SDB->BitTestCases[i].Parent;
1073 SDB->setCurrentBasicBlock(BB);
1075 SDB->visitBitTestHeader(SDB->BitTestCases[i]);
1076 CurDAG->setRoot(SDB->getRoot());
1077 CodeGenAndEmitDAG();
1081 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j) {
1082 // Set the current basic block to the mbb we wish to insert the code into
1083 BB = SDB->BitTestCases[i].Cases[j].ThisBB;
1084 SDB->setCurrentBasicBlock(BB);
1087 SDB->visitBitTestCase(SDB->BitTestCases[i].Cases[j+1].ThisBB,
1088 SDB->BitTestCases[i].Reg,
1089 SDB->BitTestCases[i].Cases[j]);
1091 SDB->visitBitTestCase(SDB->BitTestCases[i].Default,
1092 SDB->BitTestCases[i].Reg,
1093 SDB->BitTestCases[i].Cases[j]);
1096 CurDAG->setRoot(SDB->getRoot());
1097 CodeGenAndEmitDAG();
1102 for (unsigned pi = 0, pe = SDB->PHINodesToUpdate.size(); pi != pe; ++pi) {
1103 MachineInstr *PHI = SDB->PHINodesToUpdate[pi].first;
1104 MachineBasicBlock *PHIBB = PHI->getParent();
1105 assert(PHI->isPHI() &&
1106 "This is not a machine PHI node that we are updating!");
1107 // This is "default" BB. We have two jumps to it. From "header" BB and
1108 // from last "case" BB.
1109 if (PHIBB == SDB->BitTestCases[i].Default) {
1110 PHI->addOperand(MachineOperand::
1111 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1112 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Parent));
1113 PHI->addOperand(MachineOperand::
1114 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1115 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Cases.
1118 // One of "cases" BB.
1119 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size();
1121 MachineBasicBlock* cBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1122 if (cBB->isSuccessor(PHIBB)) {
1123 PHI->addOperand(MachineOperand::
1124 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1125 PHI->addOperand(MachineOperand::CreateMBB(cBB));
1130 SDB->BitTestCases.clear();
1132 // If the JumpTable record is filled in, then we need to emit a jump table.
1133 // Updating the PHI nodes is tricky in this case, since we need to determine
1134 // whether the PHI is a successor of the range check MBB or the jump table MBB
1135 for (unsigned i = 0, e = SDB->JTCases.size(); i != e; ++i) {
1136 // Lower header first, if it wasn't already lowered
1137 if (!SDB->JTCases[i].first.Emitted) {
1138 // Set the current basic block to the mbb we wish to insert the code into
1139 BB = SDB->JTCases[i].first.HeaderBB;
1140 SDB->setCurrentBasicBlock(BB);
1142 SDB->visitJumpTableHeader(SDB->JTCases[i].second, SDB->JTCases[i].first);
1143 CurDAG->setRoot(SDB->getRoot());
1144 CodeGenAndEmitDAG();
1148 // Set the current basic block to the mbb we wish to insert the code into
1149 BB = SDB->JTCases[i].second.MBB;
1150 SDB->setCurrentBasicBlock(BB);
1152 SDB->visitJumpTable(SDB->JTCases[i].second);
1153 CurDAG->setRoot(SDB->getRoot());
1154 CodeGenAndEmitDAG();
1158 for (unsigned pi = 0, pe = SDB->PHINodesToUpdate.size(); pi != pe; ++pi) {
1159 MachineInstr *PHI = SDB->PHINodesToUpdate[pi].first;
1160 MachineBasicBlock *PHIBB = PHI->getParent();
1161 assert(PHI->isPHI() &&
1162 "This is not a machine PHI node that we are updating!");
1163 // "default" BB. We can go there only from header BB.
1164 if (PHIBB == SDB->JTCases[i].second.Default) {
1166 (MachineOperand::CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1168 (MachineOperand::CreateMBB(SDB->JTCases[i].first.HeaderBB));
1170 // JT BB. Just iterate over successors here
1171 if (BB->isSuccessor(PHIBB)) {
1173 (MachineOperand::CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1174 PHI->addOperand(MachineOperand::CreateMBB(BB));
1178 SDB->JTCases.clear();
1180 // If the switch block involved a branch to one of the actual successors, we
1181 // need to update PHI nodes in that block.
1182 for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) {
1183 MachineInstr *PHI = SDB->PHINodesToUpdate[i].first;
1184 assert(PHI->isPHI() &&
1185 "This is not a machine PHI node that we are updating!");
1186 if (BB->isSuccessor(PHI->getParent())) {
1187 PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second,
1189 PHI->addOperand(MachineOperand::CreateMBB(BB));
1193 // If we generated any switch lowering information, build and codegen any
1194 // additional DAGs necessary.
1195 for (unsigned i = 0, e = SDB->SwitchCases.size(); i != e; ++i) {
1196 // Set the current basic block to the mbb we wish to insert the code into
1197 MachineBasicBlock *ThisBB = BB = SDB->SwitchCases[i].ThisBB;
1198 SDB->setCurrentBasicBlock(BB);
1201 SDB->visitSwitchCase(SDB->SwitchCases[i]);
1202 CurDAG->setRoot(SDB->getRoot());
1203 CodeGenAndEmitDAG();
1205 // Handle any PHI nodes in successors of this chunk, as if we were coming
1206 // from the original BB before switch expansion. Note that PHI nodes can
1207 // occur multiple times in PHINodesToUpdate. We have to be very careful to
1208 // handle them the right number of times.
1209 while ((BB = SDB->SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
1210 // If new BB's are created during scheduling, the edges may have been
1211 // updated. That is, the edge from ThisBB to BB may have been split and
1212 // BB's predecessor is now another block.
1213 DenseMap<MachineBasicBlock*, MachineBasicBlock*>::iterator EI =
1214 SDB->EdgeMapping.find(BB);
1215 if (EI != SDB->EdgeMapping.end())
1216 ThisBB = EI->second;
1218 // BB may have been removed from the CFG if a branch was constant folded.
1219 if (ThisBB->isSuccessor(BB)) {
1220 for (MachineBasicBlock::iterator Phi = BB->begin();
1221 Phi != BB->end() && Phi->isPHI();
1223 // This value for this PHI node is recorded in PHINodesToUpdate.
1224 for (unsigned pn = 0; ; ++pn) {
1225 assert(pn != SDB->PHINodesToUpdate.size() &&
1226 "Didn't find PHI entry!");
1227 if (SDB->PHINodesToUpdate[pn].first == Phi) {
1228 Phi->addOperand(MachineOperand::
1229 CreateReg(SDB->PHINodesToUpdate[pn].second,
1231 Phi->addOperand(MachineOperand::CreateMBB(ThisBB));
1238 // Don't process RHS if same block as LHS.
1239 if (BB == SDB->SwitchCases[i].FalseBB)
1240 SDB->SwitchCases[i].FalseBB = 0;
1242 // If we haven't handled the RHS, do so now. Otherwise, we're done.
1243 SDB->SwitchCases[i].TrueBB = SDB->SwitchCases[i].FalseBB;
1244 SDB->SwitchCases[i].FalseBB = 0;
1246 assert(SDB->SwitchCases[i].TrueBB == 0 && SDB->SwitchCases[i].FalseBB == 0);
1249 SDB->SwitchCases.clear();
1251 SDB->PHINodesToUpdate.clear();
1255 /// Create the scheduler. If a specific scheduler was specified
1256 /// via the SchedulerRegistry, use it, otherwise select the
1257 /// one preferred by the target.
1259 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
1260 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1264 RegisterScheduler::setDefault(Ctor);
1267 return Ctor(this, OptLevel);
1270 ScheduleHazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
1271 return new ScheduleHazardRecognizer();
1274 //===----------------------------------------------------------------------===//
1275 // Helper functions used by the generated instruction selector.
1276 //===----------------------------------------------------------------------===//
1277 // Calls to these methods are generated by tblgen.
1279 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1280 /// the dag combiner simplified the 255, we still want to match. RHS is the
1281 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1282 /// specified in the .td file (e.g. 255).
1283 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1284 int64_t DesiredMaskS) const {
1285 const APInt &ActualMask = RHS->getAPIntValue();
1286 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1288 // If the actual mask exactly matches, success!
1289 if (ActualMask == DesiredMask)
1292 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1293 if (ActualMask.intersects(~DesiredMask))
1296 // Otherwise, the DAG Combiner may have proven that the value coming in is
1297 // either already zero or is not demanded. Check for known zero input bits.
1298 APInt NeededMask = DesiredMask & ~ActualMask;
1299 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1302 // TODO: check to see if missing bits are just not demanded.
1304 // Otherwise, this pattern doesn't match.
1308 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1309 /// the dag combiner simplified the 255, we still want to match. RHS is the
1310 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1311 /// specified in the .td file (e.g. 255).
1312 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1313 int64_t DesiredMaskS) const {
1314 const APInt &ActualMask = RHS->getAPIntValue();
1315 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1317 // If the actual mask exactly matches, success!
1318 if (ActualMask == DesiredMask)
1321 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1322 if (ActualMask.intersects(~DesiredMask))
1325 // Otherwise, the DAG Combiner may have proven that the value coming in is
1326 // either already zero or is not demanded. Check for known zero input bits.
1327 APInt NeededMask = DesiredMask & ~ActualMask;
1329 APInt KnownZero, KnownOne;
1330 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
1332 // If all the missing bits in the or are already known to be set, match!
1333 if ((NeededMask & KnownOne) == NeededMask)
1336 // TODO: check to see if missing bits are just not demanded.
1338 // Otherwise, this pattern doesn't match.
1343 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1344 /// by tblgen. Others should not call it.
1345 void SelectionDAGISel::
1346 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1347 std::vector<SDValue> InOps;
1348 std::swap(InOps, Ops);
1350 Ops.push_back(InOps[0]); // input chain.
1351 Ops.push_back(InOps[1]); // input asm string.
1353 unsigned i = 2, e = InOps.size();
1354 if (InOps[e-1].getValueType() == MVT::Flag)
1355 --e; // Don't process a flag operand if it is here.
1358 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1359 if ((Flags & 7) != 4 /*MEM*/) {
1360 // Just skip over this operand, copying the operands verbatim.
1361 Ops.insert(Ops.end(), InOps.begin()+i,
1362 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
1363 i += InlineAsm::getNumOperandRegisters(Flags) + 1;
1365 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 &&
1366 "Memory operand with multiple values?");
1367 // Otherwise, this is a memory operand. Ask the target to select it.
1368 std::vector<SDValue> SelOps;
1369 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps)) {
1370 llvm_report_error("Could not match memory address. Inline asm"
1374 // Add this to the output node.
1375 Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size()<< 3),
1377 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1382 // Add the flag input back if present.
1383 if (e != InOps.size())
1384 Ops.push_back(InOps.back());
1387 /// findFlagUse - Return use of EVT::Flag value produced by the specified
1390 static SDNode *findFlagUse(SDNode *N) {
1391 unsigned FlagResNo = N->getNumValues()-1;
1392 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1393 SDUse &Use = I.getUse();
1394 if (Use.getResNo() == FlagResNo)
1395 return Use.getUser();
1400 /// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
1401 /// This function recursively traverses up the operand chain, ignoring
1403 static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
1404 SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited,
1405 bool IgnoreChains) {
1406 // The NodeID's are given uniques ID's where a node ID is guaranteed to be
1407 // greater than all of its (recursive) operands. If we scan to a point where
1408 // 'use' is smaller than the node we're scanning for, then we know we will
1411 // The Use may be -1 (unassigned) if it is a newly allocated node. This can
1412 // happen because we scan down to newly selected nodes in the case of flag
1414 if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1))
1417 // Don't revisit nodes if we already scanned it and didn't fail, we know we
1418 // won't fail if we scan it again.
1419 if (!Visited.insert(Use))
1422 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
1423 // Ignore chain uses, they are validated by HandleMergeInputChains.
1424 if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains)
1427 SDNode *N = Use->getOperand(i).getNode();
1429 if (Use == ImmedUse || Use == Root)
1430 continue; // We are not looking for immediate use.
1435 // Traverse up the operand chain.
1436 if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
1442 /// IsProfitableToFold - Returns true if it's profitable to fold the specific
1443 /// operand node N of U during instruction selection that starts at Root.
1444 bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
1445 SDNode *Root) const {
1446 if (OptLevel == CodeGenOpt::None) return false;
1447 return N.hasOneUse();
1450 /// IsLegalToFold - Returns true if the specific operand node N of
1451 /// U can be folded during instruction selection that starts at Root.
1452 bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
1453 bool IgnoreChains) const {
1454 if (OptLevel == CodeGenOpt::None) return false;
1456 // If Root use can somehow reach N through a path that that doesn't contain
1457 // U then folding N would create a cycle. e.g. In the following
1458 // diagram, Root can reach N through X. If N is folded into into Root, then
1459 // X is both a predecessor and a successor of U.
1470 // * indicates nodes to be folded together.
1472 // If Root produces a flag, then it gets (even more) interesting. Since it
1473 // will be "glued" together with its flag use in the scheduler, we need to
1474 // check if it might reach N.
1493 // If FU (flag use) indirectly reaches N (the load), and Root folds N
1494 // (call it Fold), then X is a predecessor of FU and a successor of
1495 // Fold. But since Fold and FU are flagged together, this will create
1496 // a cycle in the scheduling graph.
1498 // If the node has flags, walk down the graph to the "lowest" node in the
1500 EVT VT = Root->getValueType(Root->getNumValues()-1);
1501 while (VT == MVT::Flag) {
1502 SDNode *FU = findFlagUse(Root);
1506 VT = Root->getValueType(Root->getNumValues()-1);
1508 // If our query node has a flag result with a use, we've walked up it. If
1509 // the user (which has already been selected) has a chain or indirectly uses
1510 // the chain, our WalkChainUsers predicate will not consider it. Because of
1511 // this, we cannot ignore chains in this predicate.
1512 IgnoreChains = false;
1516 SmallPtrSet<SDNode*, 16> Visited;
1517 return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains);
1520 SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
1521 std::vector<SDValue> Ops(N->op_begin(), N->op_end());
1522 SelectInlineAsmMemoryOperands(Ops);
1524 std::vector<EVT> VTs;
1525 VTs.push_back(MVT::Other);
1526 VTs.push_back(MVT::Flag);
1527 SDValue New = CurDAG->getNode(ISD::INLINEASM, N->getDebugLoc(),
1528 VTs, &Ops[0], Ops.size());
1530 return New.getNode();
1533 SDNode *SelectionDAGISel::Select_UNDEF(SDNode *N) {
1534 return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0));
1537 /// GetVBR - decode a vbr encoding whose top bit is set.
1538 ALWAYS_INLINE static uint64_t
1539 GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
1540 assert(Val >= 128 && "Not a VBR");
1541 Val &= 127; // Remove first vbr bit.
1546 NextBits = MatcherTable[Idx++];
1547 Val |= (NextBits&127) << Shift;
1549 } while (NextBits & 128);
1555 /// UpdateChainsAndFlags - When a match is complete, this method updates uses of
1556 /// interior flag and chain results to use the new flag and chain results.
1557 void SelectionDAGISel::
1558 UpdateChainsAndFlags(SDNode *NodeToMatch, SDValue InputChain,
1559 const SmallVectorImpl<SDNode*> &ChainNodesMatched,
1561 const SmallVectorImpl<SDNode*> &FlagResultNodesMatched,
1562 bool isMorphNodeTo) {
1563 SmallVector<SDNode*, 4> NowDeadNodes;
1565 ISelUpdater ISU(ISelPosition);
1567 // Now that all the normal results are replaced, we replace the chain and
1568 // flag results if present.
1569 if (!ChainNodesMatched.empty()) {
1570 assert(InputChain.getNode() != 0 &&
1571 "Matched input chains but didn't produce a chain");
1572 // Loop over all of the nodes we matched that produced a chain result.
1573 // Replace all the chain results with the final chain we ended up with.
1574 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1575 SDNode *ChainNode = ChainNodesMatched[i];
1577 // If this node was already deleted, don't look at it.
1578 if (ChainNode->getOpcode() == ISD::DELETED_NODE)
1581 // Don't replace the results of the root node if we're doing a
1583 if (ChainNode == NodeToMatch && isMorphNodeTo)
1586 SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1);
1587 if (ChainVal.getValueType() == MVT::Flag)
1588 ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
1589 assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
1590 CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain, &ISU);
1592 // If the node became dead and we haven't already seen it, delete it.
1593 if (ChainNode->use_empty() &&
1594 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), ChainNode))
1595 NowDeadNodes.push_back(ChainNode);
1599 // If the result produces a flag, update any flag results in the matched
1600 // pattern with the flag result.
1601 if (InputFlag.getNode() != 0) {
1602 // Handle any interior nodes explicitly marked.
1603 for (unsigned i = 0, e = FlagResultNodesMatched.size(); i != e; ++i) {
1604 SDNode *FRN = FlagResultNodesMatched[i];
1606 // If this node was already deleted, don't look at it.
1607 if (FRN->getOpcode() == ISD::DELETED_NODE)
1610 assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Flag &&
1611 "Doesn't have a flag result");
1612 CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1),
1615 // If the node became dead and we haven't already seen it, delete it.
1616 if (FRN->use_empty() &&
1617 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), FRN))
1618 NowDeadNodes.push_back(FRN);
1622 if (!NowDeadNodes.empty())
1623 CurDAG->RemoveDeadNodes(NowDeadNodes, &ISU);
1625 DEBUG(errs() << "ISEL: Match complete!\n");
1631 CR_LeadsToInteriorNode
1634 /// WalkChainUsers - Walk down the users of the specified chained node that is
1635 /// part of the pattern we're matching, looking at all of the users we find.
1636 /// This determines whether something is an interior node, whether we have a
1637 /// non-pattern node in between two pattern nodes (which prevent folding because
1638 /// it would induce a cycle) and whether we have a TokenFactor node sandwiched
1639 /// between pattern nodes (in which case the TF becomes part of the pattern).
1641 /// The walk we do here is guaranteed to be small because we quickly get down to
1642 /// already selected nodes "below" us.
1644 WalkChainUsers(SDNode *ChainedNode,
1645 SmallVectorImpl<SDNode*> &ChainedNodesInPattern,
1646 SmallVectorImpl<SDNode*> &InteriorChainedNodes) {
1647 ChainResult Result = CR_Simple;
1649 for (SDNode::use_iterator UI = ChainedNode->use_begin(),
1650 E = ChainedNode->use_end(); UI != E; ++UI) {
1651 // Make sure the use is of the chain, not some other value we produce.
1652 if (UI.getUse().getValueType() != MVT::Other) continue;
1656 // If we see an already-selected machine node, then we've gone beyond the
1657 // pattern that we're selecting down into the already selected chunk of the
1659 if (User->isMachineOpcode() ||
1660 User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
1663 if (User->getOpcode() == ISD::CopyToReg ||
1664 User->getOpcode() == ISD::CopyFromReg ||
1665 User->getOpcode() == ISD::INLINEASM ||
1666 User->getOpcode() == ISD::EH_LABEL) {
1667 // If their node ID got reset to -1 then they've already been selected.
1668 // Treat them like a MachineOpcode.
1669 if (User->getNodeId() == -1)
1673 // If we have a TokenFactor, we handle it specially.
1674 if (User->getOpcode() != ISD::TokenFactor) {
1675 // If the node isn't a token factor and isn't part of our pattern, then it
1676 // must be a random chained node in between two nodes we're selecting.
1677 // This happens when we have something like:
1682 // Because we structurally match the load/store as a read/modify/write,
1683 // but the call is chained between them. We cannot fold in this case
1684 // because it would induce a cycle in the graph.
1685 if (!std::count(ChainedNodesInPattern.begin(),
1686 ChainedNodesInPattern.end(), User))
1687 return CR_InducesCycle;
1689 // Otherwise we found a node that is part of our pattern. For example in:
1693 // This would happen when we're scanning down from the load and see the
1694 // store as a user. Record that there is a use of ChainedNode that is
1695 // part of the pattern and keep scanning uses.
1696 Result = CR_LeadsToInteriorNode;
1697 InteriorChainedNodes.push_back(User);
1701 // If we found a TokenFactor, there are two cases to consider: first if the
1702 // TokenFactor is just hanging "below" the pattern we're matching (i.e. no
1703 // uses of the TF are in our pattern) we just want to ignore it. Second,
1704 // the TokenFactor can be sandwiched in between two chained nodes, like so:
1710 // | \ DAG's like cheese
1713 // [TokenFactor] [Op]
1720 // In this case, the TokenFactor becomes part of our match and we rewrite it
1721 // as a new TokenFactor.
1723 // To distinguish these two cases, do a recursive walk down the uses.
1724 switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) {
1726 // If the uses of the TokenFactor are just already-selected nodes, ignore
1727 // it, it is "below" our pattern.
1729 case CR_InducesCycle:
1730 // If the uses of the TokenFactor lead to nodes that are not part of our
1731 // pattern that are not selected, folding would turn this into a cycle,
1733 return CR_InducesCycle;
1734 case CR_LeadsToInteriorNode:
1735 break; // Otherwise, keep processing.
1738 // Okay, we know we're in the interesting interior case. The TokenFactor
1739 // is now going to be considered part of the pattern so that we rewrite its
1740 // uses (it may have uses that are not part of the pattern) with the
1741 // ultimate chain result of the generated code. We will also add its chain
1742 // inputs as inputs to the ultimate TokenFactor we create.
1743 Result = CR_LeadsToInteriorNode;
1744 ChainedNodesInPattern.push_back(User);
1745 InteriorChainedNodes.push_back(User);
1752 /// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
1753 /// operation for when the pattern matched at least one node with a chains. The
1754 /// input vector contains a list of all of the chained nodes that we match. We
1755 /// must determine if this is a valid thing to cover (i.e. matching it won't
1756 /// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
1757 /// be used as the input node chain for the generated nodes.
1759 HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
1760 SelectionDAG *CurDAG) {
1761 // Walk all of the chained nodes we've matched, recursively scanning down the
1762 // users of the chain result. This adds any TokenFactor nodes that are caught
1763 // in between chained nodes to the chained and interior nodes list.
1764 SmallVector<SDNode*, 3> InteriorChainedNodes;
1765 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1766 if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched,
1767 InteriorChainedNodes) == CR_InducesCycle)
1768 return SDValue(); // Would induce a cycle.
1771 // Okay, we have walked all the matched nodes and collected TokenFactor nodes
1772 // that we are interested in. Form our input TokenFactor node.
1773 SmallVector<SDValue, 3> InputChains;
1774 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1775 // Add the input chain of this node to the InputChains list (which will be
1776 // the operands of the generated TokenFactor) if it's not an interior node.
1777 SDNode *N = ChainNodesMatched[i];
1778 if (N->getOpcode() != ISD::TokenFactor) {
1779 if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N))
1782 // Otherwise, add the input chain.
1783 SDValue InChain = ChainNodesMatched[i]->getOperand(0);
1784 assert(InChain.getValueType() == MVT::Other && "Not a chain");
1785 InputChains.push_back(InChain);
1789 // If we have a token factor, we want to add all inputs of the token factor
1790 // that are not part of the pattern we're matching.
1791 for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) {
1792 if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(),
1793 N->getOperand(op).getNode()))
1794 InputChains.push_back(N->getOperand(op));
1799 if (InputChains.size() == 1)
1800 return InputChains[0];
1801 return CurDAG->getNode(ISD::TokenFactor, ChainNodesMatched[0]->getDebugLoc(),
1802 MVT::Other, &InputChains[0], InputChains.size());
1805 /// MorphNode - Handle morphing a node in place for the selector.
1806 SDNode *SelectionDAGISel::
1807 MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList,
1808 const SDValue *Ops, unsigned NumOps, unsigned EmitNodeInfo) {
1809 // It is possible we're using MorphNodeTo to replace a node with no
1810 // normal results with one that has a normal result (or we could be
1811 // adding a chain) and the input could have flags and chains as well.
1812 // In this case we need to shift the operands down.
1813 // FIXME: This is a horrible hack and broken in obscure cases, no worse
1814 // than the old isel though.
1815 int OldFlagResultNo = -1, OldChainResultNo = -1;
1817 unsigned NTMNumResults = Node->getNumValues();
1818 if (Node->getValueType(NTMNumResults-1) == MVT::Flag) {
1819 OldFlagResultNo = NTMNumResults-1;
1820 if (NTMNumResults != 1 &&
1821 Node->getValueType(NTMNumResults-2) == MVT::Other)
1822 OldChainResultNo = NTMNumResults-2;
1823 } else if (Node->getValueType(NTMNumResults-1) == MVT::Other)
1824 OldChainResultNo = NTMNumResults-1;
1826 // Call the underlying SelectionDAG routine to do the transmogrification. Note
1827 // that this deletes operands of the old node that become dead.
1828 SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops, NumOps);
1830 // MorphNodeTo can operate in two ways: if an existing node with the
1831 // specified operands exists, it can just return it. Otherwise, it
1832 // updates the node in place to have the requested operands.
1834 // If we updated the node in place, reset the node ID. To the isel,
1835 // this should be just like a newly allocated machine node.
1839 unsigned ResNumResults = Res->getNumValues();
1840 // Move the flag if needed.
1841 if ((EmitNodeInfo & OPFL_FlagOutput) && OldFlagResultNo != -1 &&
1842 (unsigned)OldFlagResultNo != ResNumResults-1)
1843 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldFlagResultNo),
1844 SDValue(Res, ResNumResults-1));
1846 if ((EmitNodeInfo & OPFL_FlagOutput) != 0)
1849 // Move the chain reference if needed.
1850 if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 &&
1851 (unsigned)OldChainResultNo != ResNumResults-1)
1852 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo),
1853 SDValue(Res, ResNumResults-1));
1855 // Otherwise, no replacement happened because the node already exists. Replace
1856 // Uses of the old node with the new one.
1858 CurDAG->ReplaceAllUsesWith(Node, Res);
1863 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1864 ALWAYS_INLINE static bool
1865 CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1866 SDValue N, const SmallVectorImpl<SDValue> &RecordedNodes) {
1867 // Accept if it is exactly the same as a previously recorded node.
1868 unsigned RecNo = MatcherTable[MatcherIndex++];
1869 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
1870 return N == RecordedNodes[RecNo];
1873 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1874 ALWAYS_INLINE static bool
1875 CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1876 SelectionDAGISel &SDISel) {
1877 return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]);
1880 /// CheckNodePredicate - Implements OP_CheckNodePredicate.
1881 ALWAYS_INLINE static bool
1882 CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1883 SelectionDAGISel &SDISel, SDNode *N) {
1884 return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]);
1887 ALWAYS_INLINE static bool
1888 CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1890 uint16_t Opc = MatcherTable[MatcherIndex++];
1891 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
1892 return N->getOpcode() == Opc;
1895 ALWAYS_INLINE static bool
1896 CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1897 SDValue N, const TargetLowering &TLI) {
1898 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1899 if (N.getValueType() == VT) return true;
1901 // Handle the case when VT is iPTR.
1902 return VT == MVT::iPTR && N.getValueType() == TLI.getPointerTy();
1905 ALWAYS_INLINE static bool
1906 CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1907 SDValue N, const TargetLowering &TLI,
1909 if (ChildNo >= N.getNumOperands())
1910 return false; // Match fails if out of range child #.
1911 return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI);
1915 ALWAYS_INLINE static bool
1916 CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1918 return cast<CondCodeSDNode>(N)->get() ==
1919 (ISD::CondCode)MatcherTable[MatcherIndex++];
1922 ALWAYS_INLINE static bool
1923 CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1924 SDValue N, const TargetLowering &TLI) {
1925 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1926 if (cast<VTSDNode>(N)->getVT() == VT)
1929 // Handle the case when VT is iPTR.
1930 return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI.getPointerTy();
1933 ALWAYS_INLINE static bool
1934 CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1936 int64_t Val = MatcherTable[MatcherIndex++];
1938 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1940 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
1941 return C != 0 && C->getSExtValue() == Val;
1944 ALWAYS_INLINE static bool
1945 CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1946 SDValue N, SelectionDAGISel &SDISel) {
1947 int64_t Val = MatcherTable[MatcherIndex++];
1949 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1951 if (N->getOpcode() != ISD::AND) return false;
1953 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1954 return C != 0 && SDISel.CheckAndMask(N.getOperand(0), C, Val);
1957 ALWAYS_INLINE static bool
1958 CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1959 SDValue N, SelectionDAGISel &SDISel) {
1960 int64_t Val = MatcherTable[MatcherIndex++];
1962 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1964 if (N->getOpcode() != ISD::OR) return false;
1966 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1967 return C != 0 && SDISel.CheckOrMask(N.getOperand(0), C, Val);
1970 /// IsPredicateKnownToFail - If we know how and can do so without pushing a
1971 /// scope, evaluate the current node. If the current predicate is known to
1972 /// fail, set Result=true and return anything. If the current predicate is
1973 /// known to pass, set Result=false and return the MatcherIndex to continue
1974 /// with. If the current predicate is unknown, set Result=false and return the
1975 /// MatcherIndex to continue with.
1976 static unsigned IsPredicateKnownToFail(const unsigned char *Table,
1977 unsigned Index, SDValue N,
1978 bool &Result, SelectionDAGISel &SDISel,
1979 SmallVectorImpl<SDValue> &RecordedNodes){
1980 switch (Table[Index++]) {
1983 return Index-1; // Could not evaluate this predicate.
1984 case SelectionDAGISel::OPC_CheckSame:
1985 Result = !::CheckSame(Table, Index, N, RecordedNodes);
1987 case SelectionDAGISel::OPC_CheckPatternPredicate:
1988 Result = !::CheckPatternPredicate(Table, Index, SDISel);
1990 case SelectionDAGISel::OPC_CheckPredicate:
1991 Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode());
1993 case SelectionDAGISel::OPC_CheckOpcode:
1994 Result = !::CheckOpcode(Table, Index, N.getNode());
1996 case SelectionDAGISel::OPC_CheckType:
1997 Result = !::CheckType(Table, Index, N, SDISel.TLI);
1999 case SelectionDAGISel::OPC_CheckChild0Type:
2000 case SelectionDAGISel::OPC_CheckChild1Type:
2001 case SelectionDAGISel::OPC_CheckChild2Type:
2002 case SelectionDAGISel::OPC_CheckChild3Type:
2003 case SelectionDAGISel::OPC_CheckChild4Type:
2004 case SelectionDAGISel::OPC_CheckChild5Type:
2005 case SelectionDAGISel::OPC_CheckChild6Type:
2006 case SelectionDAGISel::OPC_CheckChild7Type:
2007 Result = !::CheckChildType(Table, Index, N, SDISel.TLI,
2008 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Type);
2010 case SelectionDAGISel::OPC_CheckCondCode:
2011 Result = !::CheckCondCode(Table, Index, N);
2013 case SelectionDAGISel::OPC_CheckValueType:
2014 Result = !::CheckValueType(Table, Index, N, SDISel.TLI);
2016 case SelectionDAGISel::OPC_CheckInteger:
2017 Result = !::CheckInteger(Table, Index, N);
2019 case SelectionDAGISel::OPC_CheckAndImm:
2020 Result = !::CheckAndImm(Table, Index, N, SDISel);
2022 case SelectionDAGISel::OPC_CheckOrImm:
2023 Result = !::CheckOrImm(Table, Index, N, SDISel);
2030 /// FailIndex - If this match fails, this is the index to continue with.
2033 /// NodeStack - The node stack when the scope was formed.
2034 SmallVector<SDValue, 4> NodeStack;
2036 /// NumRecordedNodes - The number of recorded nodes when the scope was formed.
2037 unsigned NumRecordedNodes;
2039 /// NumMatchedMemRefs - The number of matched memref entries.
2040 unsigned NumMatchedMemRefs;
2042 /// InputChain/InputFlag - The current chain/flag
2043 SDValue InputChain, InputFlag;
2045 /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty.
2046 bool HasChainNodesMatched, HasFlagResultNodesMatched;
2049 SDNode *SelectionDAGISel::
2050 SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
2051 unsigned TableSize) {
2052 // FIXME: Should these even be selected? Handle these cases in the caller?
2053 switch (NodeToMatch->getOpcode()) {
2056 case ISD::EntryToken: // These nodes remain the same.
2057 case ISD::BasicBlock:
2059 //case ISD::VALUETYPE:
2060 //case ISD::CONDCODE:
2061 case ISD::HANDLENODE:
2062 case ISD::TargetConstant:
2063 case ISD::TargetConstantFP:
2064 case ISD::TargetConstantPool:
2065 case ISD::TargetFrameIndex:
2066 case ISD::TargetExternalSymbol:
2067 case ISD::TargetBlockAddress:
2068 case ISD::TargetJumpTable:
2069 case ISD::TargetGlobalTLSAddress:
2070 case ISD::TargetGlobalAddress:
2071 case ISD::TokenFactor:
2072 case ISD::CopyFromReg:
2073 case ISD::CopyToReg:
2075 NodeToMatch->setNodeId(-1); // Mark selected.
2077 case ISD::AssertSext:
2078 case ISD::AssertZext:
2079 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0),
2080 NodeToMatch->getOperand(0));
2082 case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch);
2083 case ISD::UNDEF: return Select_UNDEF(NodeToMatch);
2086 assert(!NodeToMatch->isMachineOpcode() && "Node already selected!");
2088 // Set up the node stack with NodeToMatch as the only node on the stack.
2089 SmallVector<SDValue, 8> NodeStack;
2090 SDValue N = SDValue(NodeToMatch, 0);
2091 NodeStack.push_back(N);
2093 // MatchScopes - Scopes used when matching, if a match failure happens, this
2094 // indicates where to continue checking.
2095 SmallVector<MatchScope, 8> MatchScopes;
2097 // RecordedNodes - This is the set of nodes that have been recorded by the
2099 SmallVector<SDValue, 8> RecordedNodes;
2101 // MatchedMemRefs - This is the set of MemRef's we've seen in the input
2103 SmallVector<MachineMemOperand*, 2> MatchedMemRefs;
2105 // These are the current input chain and flag for use when generating nodes.
2106 // Various Emit operations change these. For example, emitting a copytoreg
2107 // uses and updates these.
2108 SDValue InputChain, InputFlag;
2110 // ChainNodesMatched - If a pattern matches nodes that have input/output
2111 // chains, the OPC_EmitMergeInputChains operation is emitted which indicates
2112 // which ones they are. The result is captured into this list so that we can
2113 // update the chain results when the pattern is complete.
2114 SmallVector<SDNode*, 3> ChainNodesMatched;
2115 SmallVector<SDNode*, 3> FlagResultNodesMatched;
2117 DEBUG(errs() << "ISEL: Starting pattern match on root node: ";
2118 NodeToMatch->dump(CurDAG);
2121 // Determine where to start the interpreter. Normally we start at opcode #0,
2122 // but if the state machine starts with an OPC_SwitchOpcode, then we
2123 // accelerate the first lookup (which is guaranteed to be hot) with the
2124 // OpcodeOffset table.
2125 unsigned MatcherIndex = 0;
2127 if (!OpcodeOffset.empty()) {
2128 // Already computed the OpcodeOffset table, just index into it.
2129 if (N.getOpcode() < OpcodeOffset.size())
2130 MatcherIndex = OpcodeOffset[N.getOpcode()];
2131 DEBUG(errs() << " Initial Opcode index to " << MatcherIndex << "\n");
2133 } else if (MatcherTable[0] == OPC_SwitchOpcode) {
2134 // Otherwise, the table isn't computed, but the state machine does start
2135 // with an OPC_SwitchOpcode instruction. Populate the table now, since this
2136 // is the first time we're selecting an instruction.
2139 // Get the size of this case.
2140 unsigned CaseSize = MatcherTable[Idx++];
2142 CaseSize = GetVBR(CaseSize, MatcherTable, Idx);
2143 if (CaseSize == 0) break;
2145 // Get the opcode, add the index to the table.
2146 uint16_t Opc = MatcherTable[Idx++];
2147 Opc |= (unsigned short)MatcherTable[Idx++] << 8;
2148 if (Opc >= OpcodeOffset.size())
2149 OpcodeOffset.resize((Opc+1)*2);
2150 OpcodeOffset[Opc] = Idx;
2154 // Okay, do the lookup for the first opcode.
2155 if (N.getOpcode() < OpcodeOffset.size())
2156 MatcherIndex = OpcodeOffset[N.getOpcode()];
2160 assert(MatcherIndex < TableSize && "Invalid index");
2162 unsigned CurrentOpcodeIndex = MatcherIndex;
2164 BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++];
2167 // Okay, the semantics of this operation are that we should push a scope
2168 // then evaluate the first child. However, pushing a scope only to have
2169 // the first check fail (which then pops it) is inefficient. If we can
2170 // determine immediately that the first check (or first several) will
2171 // immediately fail, don't even bother pushing a scope for them.
2175 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2176 if (NumToSkip & 128)
2177 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2178 // Found the end of the scope with no match.
2179 if (NumToSkip == 0) {
2184 FailIndex = MatcherIndex+NumToSkip;
2186 unsigned MatcherIndexOfPredicate = MatcherIndex;
2187 (void)MatcherIndexOfPredicate; // silence warning.
2189 // If we can't evaluate this predicate without pushing a scope (e.g. if
2190 // it is a 'MoveParent') or if the predicate succeeds on this node, we
2191 // push the scope and evaluate the full predicate chain.
2193 MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
2194 Result, *this, RecordedNodes);
2198 DEBUG(errs() << " Skipped scope entry (due to false predicate) at "
2199 << "index " << MatcherIndexOfPredicate
2200 << ", continuing at " << FailIndex << "\n");
2203 // Otherwise, we know that this case of the Scope is guaranteed to fail,
2204 // move to the next case.
2205 MatcherIndex = FailIndex;
2208 // If the whole scope failed to match, bail.
2209 if (FailIndex == 0) break;
2211 // Push a MatchScope which indicates where to go if the first child fails
2213 MatchScope NewEntry;
2214 NewEntry.FailIndex = FailIndex;
2215 NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
2216 NewEntry.NumRecordedNodes = RecordedNodes.size();
2217 NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
2218 NewEntry.InputChain = InputChain;
2219 NewEntry.InputFlag = InputFlag;
2220 NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty();
2221 NewEntry.HasFlagResultNodesMatched = !FlagResultNodesMatched.empty();
2222 MatchScopes.push_back(NewEntry);
2225 case OPC_RecordNode:
2226 // Remember this node, it may end up being an operand in the pattern.
2227 RecordedNodes.push_back(N);
2230 case OPC_RecordChild0: case OPC_RecordChild1:
2231 case OPC_RecordChild2: case OPC_RecordChild3:
2232 case OPC_RecordChild4: case OPC_RecordChild5:
2233 case OPC_RecordChild6: case OPC_RecordChild7: {
2234 unsigned ChildNo = Opcode-OPC_RecordChild0;
2235 if (ChildNo >= N.getNumOperands())
2236 break; // Match fails if out of range child #.
2238 RecordedNodes.push_back(N->getOperand(ChildNo));
2241 case OPC_RecordMemRef:
2242 MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand());
2245 case OPC_CaptureFlagInput:
2246 // If the current node has an input flag, capture it in InputFlag.
2247 if (N->getNumOperands() != 0 &&
2248 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag)
2249 InputFlag = N->getOperand(N->getNumOperands()-1);
2252 case OPC_MoveChild: {
2253 unsigned ChildNo = MatcherTable[MatcherIndex++];
2254 if (ChildNo >= N.getNumOperands())
2255 break; // Match fails if out of range child #.
2256 N = N.getOperand(ChildNo);
2257 NodeStack.push_back(N);
2261 case OPC_MoveParent:
2262 // Pop the current node off the NodeStack.
2263 NodeStack.pop_back();
2264 assert(!NodeStack.empty() && "Node stack imbalance!");
2265 N = NodeStack.back();
2269 if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
2271 case OPC_CheckPatternPredicate:
2272 if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break;
2274 case OPC_CheckPredicate:
2275 if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this,
2279 case OPC_CheckComplexPat: {
2280 unsigned CPNum = MatcherTable[MatcherIndex++];
2281 unsigned RecNo = MatcherTable[MatcherIndex++];
2282 assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
2283 if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo], CPNum,
2288 case OPC_CheckOpcode:
2289 if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
2293 if (!::CheckType(MatcherTable, MatcherIndex, N, TLI)) break;
2296 case OPC_SwitchOpcode: {
2297 unsigned CurNodeOpcode = N.getOpcode();
2298 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2301 // Get the size of this case.
2302 CaseSize = MatcherTable[MatcherIndex++];
2304 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2305 if (CaseSize == 0) break;
2307 uint16_t Opc = MatcherTable[MatcherIndex++];
2308 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2310 // If the opcode matches, then we will execute this case.
2311 if (CurNodeOpcode == Opc)
2314 // Otherwise, skip over this case.
2315 MatcherIndex += CaseSize;
2318 // If no cases matched, bail out.
2319 if (CaseSize == 0) break;
2321 // Otherwise, execute the case we found.
2322 DEBUG(errs() << " OpcodeSwitch from " << SwitchStart
2323 << " to " << MatcherIndex << "\n");
2327 case OPC_SwitchType: {
2328 MVT::SimpleValueType CurNodeVT = N.getValueType().getSimpleVT().SimpleTy;
2329 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2332 // Get the size of this case.
2333 CaseSize = MatcherTable[MatcherIndex++];
2335 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2336 if (CaseSize == 0) break;
2338 MVT::SimpleValueType CaseVT =
2339 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2340 if (CaseVT == MVT::iPTR)
2341 CaseVT = TLI.getPointerTy().SimpleTy;
2343 // If the VT matches, then we will execute this case.
2344 if (CurNodeVT == CaseVT)
2347 // Otherwise, skip over this case.
2348 MatcherIndex += CaseSize;
2351 // If no cases matched, bail out.
2352 if (CaseSize == 0) break;
2354 // Otherwise, execute the case we found.
2355 DEBUG(errs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
2356 << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
2359 case OPC_CheckChild0Type: case OPC_CheckChild1Type:
2360 case OPC_CheckChild2Type: case OPC_CheckChild3Type:
2361 case OPC_CheckChild4Type: case OPC_CheckChild5Type:
2362 case OPC_CheckChild6Type: case OPC_CheckChild7Type:
2363 if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI,
2364 Opcode-OPC_CheckChild0Type))
2367 case OPC_CheckCondCode:
2368 if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
2370 case OPC_CheckValueType:
2371 if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI)) break;
2373 case OPC_CheckInteger:
2374 if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
2376 case OPC_CheckAndImm:
2377 if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
2379 case OPC_CheckOrImm:
2380 if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
2383 case OPC_CheckFoldableChainNode: {
2384 assert(NodeStack.size() != 1 && "No parent node");
2385 // Verify that all intermediate nodes between the root and this one have
2387 bool HasMultipleUses = false;
2388 for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i)
2389 if (!NodeStack[i].hasOneUse()) {
2390 HasMultipleUses = true;
2393 if (HasMultipleUses) break;
2395 // Check to see that the target thinks this is profitable to fold and that
2396 // we can fold it without inducing cycles in the graph.
2397 if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2399 !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2400 NodeToMatch, true/*We validate our own chains*/))
2405 case OPC_EmitInteger: {
2406 MVT::SimpleValueType VT =
2407 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2408 int64_t Val = MatcherTable[MatcherIndex++];
2410 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2411 RecordedNodes.push_back(CurDAG->getTargetConstant(Val, VT));
2414 case OPC_EmitRegister: {
2415 MVT::SimpleValueType VT =
2416 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2417 unsigned RegNo = MatcherTable[MatcherIndex++];
2418 RecordedNodes.push_back(CurDAG->getRegister(RegNo, VT));
2422 case OPC_EmitConvertToTarget: {
2423 // Convert from IMM/FPIMM to target version.
2424 unsigned RecNo = MatcherTable[MatcherIndex++];
2425 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2426 SDValue Imm = RecordedNodes[RecNo];
2428 if (Imm->getOpcode() == ISD::Constant) {
2429 int64_t Val = cast<ConstantSDNode>(Imm)->getZExtValue();
2430 Imm = CurDAG->getTargetConstant(Val, Imm.getValueType());
2431 } else if (Imm->getOpcode() == ISD::ConstantFP) {
2432 const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue();
2433 Imm = CurDAG->getTargetConstantFP(*Val, Imm.getValueType());
2436 RecordedNodes.push_back(Imm);
2440 case OPC_EmitMergeInputChains1_0: // OPC_EmitMergeInputChains, 1, 0
2441 case OPC_EmitMergeInputChains1_1: { // OPC_EmitMergeInputChains, 1, 1
2442 // These are space-optimized forms of OPC_EmitMergeInputChains.
2443 assert(InputChain.getNode() == 0 &&
2444 "EmitMergeInputChains should be the first chain producing node");
2445 assert(ChainNodesMatched.empty() &&
2446 "Should only have one EmitMergeInputChains per match");
2448 // Read all of the chained nodes.
2449 unsigned RecNo = Opcode == OPC_EmitMergeInputChains1_1;
2450 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2451 ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode());
2453 // FIXME: What if other value results of the node have uses not matched
2455 if (ChainNodesMatched.back() != NodeToMatch &&
2456 !RecordedNodes[RecNo].hasOneUse()) {
2457 ChainNodesMatched.clear();
2461 // Merge the input chains if they are not intra-pattern references.
2462 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2464 if (InputChain.getNode() == 0)
2465 break; // Failed to merge.
2469 case OPC_EmitMergeInputChains: {
2470 assert(InputChain.getNode() == 0 &&
2471 "EmitMergeInputChains should be the first chain producing node");
2472 // This node gets a list of nodes we matched in the input that have
2473 // chains. We want to token factor all of the input chains to these nodes
2474 // together. However, if any of the input chains is actually one of the
2475 // nodes matched in this pattern, then we have an intra-match reference.
2476 // Ignore these because the newly token factored chain should not refer to
2478 unsigned NumChains = MatcherTable[MatcherIndex++];
2479 assert(NumChains != 0 && "Can't TF zero chains");
2481 assert(ChainNodesMatched.empty() &&
2482 "Should only have one EmitMergeInputChains per match");
2484 // Read all of the chained nodes.
2485 for (unsigned i = 0; i != NumChains; ++i) {
2486 unsigned RecNo = MatcherTable[MatcherIndex++];
2487 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2488 ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode());
2490 // FIXME: What if other value results of the node have uses not matched
2492 if (ChainNodesMatched.back() != NodeToMatch &&
2493 !RecordedNodes[RecNo].hasOneUse()) {
2494 ChainNodesMatched.clear();
2499 // If the inner loop broke out, the match fails.
2500 if (ChainNodesMatched.empty())
2503 // Merge the input chains if they are not intra-pattern references.
2504 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2506 if (InputChain.getNode() == 0)
2507 break; // Failed to merge.
2512 case OPC_EmitCopyToReg: {
2513 unsigned RecNo = MatcherTable[MatcherIndex++];
2514 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2515 unsigned DestPhysReg = MatcherTable[MatcherIndex++];
2517 if (InputChain.getNode() == 0)
2518 InputChain = CurDAG->getEntryNode();
2520 InputChain = CurDAG->getCopyToReg(InputChain, NodeToMatch->getDebugLoc(),
2521 DestPhysReg, RecordedNodes[RecNo],
2524 InputFlag = InputChain.getValue(1);
2528 case OPC_EmitNodeXForm: {
2529 unsigned XFormNo = MatcherTable[MatcherIndex++];
2530 unsigned RecNo = MatcherTable[MatcherIndex++];
2531 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2532 RecordedNodes.push_back(RunSDNodeXForm(RecordedNodes[RecNo], XFormNo));
2537 case OPC_MorphNodeTo: {
2538 uint16_t TargetOpc = MatcherTable[MatcherIndex++];
2539 TargetOpc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2540 unsigned EmitNodeInfo = MatcherTable[MatcherIndex++];
2541 // Get the result VT list.
2542 unsigned NumVTs = MatcherTable[MatcherIndex++];
2543 SmallVector<EVT, 4> VTs;
2544 for (unsigned i = 0; i != NumVTs; ++i) {
2545 MVT::SimpleValueType VT =
2546 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2547 if (VT == MVT::iPTR) VT = TLI.getPointerTy().SimpleTy;
2551 if (EmitNodeInfo & OPFL_Chain)
2552 VTs.push_back(MVT::Other);
2553 if (EmitNodeInfo & OPFL_FlagOutput)
2554 VTs.push_back(MVT::Flag);
2556 // This is hot code, so optimize the two most common cases of 1 and 2
2559 if (VTs.size() == 1)
2560 VTList = CurDAG->getVTList(VTs[0]);
2561 else if (VTs.size() == 2)
2562 VTList = CurDAG->getVTList(VTs[0], VTs[1]);
2564 VTList = CurDAG->getVTList(VTs.data(), VTs.size());
2566 // Get the operand list.
2567 unsigned NumOps = MatcherTable[MatcherIndex++];
2568 SmallVector<SDValue, 8> Ops;
2569 for (unsigned i = 0; i != NumOps; ++i) {
2570 unsigned RecNo = MatcherTable[MatcherIndex++];
2572 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2574 assert(RecNo < RecordedNodes.size() && "Invalid EmitNode");
2575 Ops.push_back(RecordedNodes[RecNo]);
2578 // If there are variadic operands to add, handle them now.
2579 if (EmitNodeInfo & OPFL_VariadicInfo) {
2580 // Determine the start index to copy from.
2581 unsigned FirstOpToCopy = getNumFixedFromVariadicInfo(EmitNodeInfo);
2582 FirstOpToCopy += (EmitNodeInfo & OPFL_Chain) ? 1 : 0;
2583 assert(NodeToMatch->getNumOperands() >= FirstOpToCopy &&
2584 "Invalid variadic node");
2585 // Copy all of the variadic operands, not including a potential flag
2587 for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands();
2589 SDValue V = NodeToMatch->getOperand(i);
2590 if (V.getValueType() == MVT::Flag) break;
2595 // If this has chain/flag inputs, add them.
2596 if (EmitNodeInfo & OPFL_Chain)
2597 Ops.push_back(InputChain);
2598 if ((EmitNodeInfo & OPFL_FlagInput) && InputFlag.getNode() != 0)
2599 Ops.push_back(InputFlag);
2603 if (Opcode != OPC_MorphNodeTo) {
2604 // If this is a normal EmitNode command, just create the new node and
2605 // add the results to the RecordedNodes list.
2606 Res = CurDAG->getMachineNode(TargetOpc, NodeToMatch->getDebugLoc(),
2607 VTList, Ops.data(), Ops.size());
2609 // Add all the non-flag/non-chain results to the RecordedNodes list.
2610 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
2611 if (VTs[i] == MVT::Other || VTs[i] == MVT::Flag) break;
2612 RecordedNodes.push_back(SDValue(Res, i));
2616 Res = MorphNode(NodeToMatch, TargetOpc, VTList, Ops.data(), Ops.size(),
2620 // If the node had chain/flag results, update our notion of the current
2622 if (EmitNodeInfo & OPFL_FlagOutput) {
2623 InputFlag = SDValue(Res, VTs.size()-1);
2624 if (EmitNodeInfo & OPFL_Chain)
2625 InputChain = SDValue(Res, VTs.size()-2);
2626 } else if (EmitNodeInfo & OPFL_Chain)
2627 InputChain = SDValue(Res, VTs.size()-1);
2629 // If the OPFL_MemRefs flag is set on this node, slap all of the
2630 // accumulated memrefs onto it.
2632 // FIXME: This is vastly incorrect for patterns with multiple outputs
2633 // instructions that access memory and for ComplexPatterns that match
2635 if (EmitNodeInfo & OPFL_MemRefs) {
2636 MachineSDNode::mmo_iterator MemRefs =
2637 MF->allocateMemRefsArray(MatchedMemRefs.size());
2638 std::copy(MatchedMemRefs.begin(), MatchedMemRefs.end(), MemRefs);
2639 cast<MachineSDNode>(Res)
2640 ->setMemRefs(MemRefs, MemRefs + MatchedMemRefs.size());
2644 << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created")
2645 << " node: "; Res->dump(CurDAG); errs() << "\n");
2647 // If this was a MorphNodeTo then we're completely done!
2648 if (Opcode == OPC_MorphNodeTo) {
2649 // Update chain and flag uses.
2650 UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched,
2651 InputFlag, FlagResultNodesMatched, true);
2658 case OPC_MarkFlagResults: {
2659 unsigned NumNodes = MatcherTable[MatcherIndex++];
2661 // Read and remember all the flag-result nodes.
2662 for (unsigned i = 0; i != NumNodes; ++i) {
2663 unsigned RecNo = MatcherTable[MatcherIndex++];
2665 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2667 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2668 FlagResultNodesMatched.push_back(RecordedNodes[RecNo].getNode());
2673 case OPC_CompleteMatch: {
2674 // The match has been completed, and any new nodes (if any) have been
2675 // created. Patch up references to the matched dag to use the newly
2677 unsigned NumResults = MatcherTable[MatcherIndex++];
2679 for (unsigned i = 0; i != NumResults; ++i) {
2680 unsigned ResSlot = MatcherTable[MatcherIndex++];
2682 ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex);
2684 assert(ResSlot < RecordedNodes.size() && "Invalid CheckSame");
2685 SDValue Res = RecordedNodes[ResSlot];
2687 assert(i < NodeToMatch->getNumValues() &&
2688 NodeToMatch->getValueType(i) != MVT::Other &&
2689 NodeToMatch->getValueType(i) != MVT::Flag &&
2690 "Invalid number of results to complete!");
2691 assert((NodeToMatch->getValueType(i) == Res.getValueType() ||
2692 NodeToMatch->getValueType(i) == MVT::iPTR ||
2693 Res.getValueType() == MVT::iPTR ||
2694 NodeToMatch->getValueType(i).getSizeInBits() ==
2695 Res.getValueType().getSizeInBits()) &&
2696 "invalid replacement");
2697 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, i), Res);
2700 // If the root node defines a flag, add it to the flag nodes to update
2702 if (NodeToMatch->getValueType(NodeToMatch->getNumValues()-1) == MVT::Flag)
2703 FlagResultNodesMatched.push_back(NodeToMatch);
2705 // Update chain and flag uses.
2706 UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched,
2707 InputFlag, FlagResultNodesMatched, false);
2709 assert(NodeToMatch->use_empty() &&
2710 "Didn't replace all uses of the node?");
2712 // FIXME: We just return here, which interacts correctly with SelectRoot
2713 // above. We should fix this to not return an SDNode* anymore.
2718 // If the code reached this point, then the match failed. See if there is
2719 // another child to try in the current 'Scope', otherwise pop it until we
2720 // find a case to check.
2721 DEBUG(errs() << " Match failed at index " << CurrentOpcodeIndex << "\n");
2723 if (MatchScopes.empty()) {
2724 CannotYetSelect(NodeToMatch);
2728 // Restore the interpreter state back to the point where the scope was
2730 MatchScope &LastScope = MatchScopes.back();
2731 RecordedNodes.resize(LastScope.NumRecordedNodes);
2733 NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end());
2734 N = NodeStack.back();
2736 if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size())
2737 MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
2738 MatcherIndex = LastScope.FailIndex;
2740 DEBUG(errs() << " Continuing at " << MatcherIndex << "\n");
2742 InputChain = LastScope.InputChain;
2743 InputFlag = LastScope.InputFlag;
2744 if (!LastScope.HasChainNodesMatched)
2745 ChainNodesMatched.clear();
2746 if (!LastScope.HasFlagResultNodesMatched)
2747 FlagResultNodesMatched.clear();
2749 // Check to see what the offset is at the new MatcherIndex. If it is zero
2750 // we have reached the end of this scope, otherwise we have another child
2751 // in the current scope to try.
2752 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2753 if (NumToSkip & 128)
2754 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2756 // If we have another child in this scope to match, update FailIndex and
2758 if (NumToSkip != 0) {
2759 LastScope.FailIndex = MatcherIndex+NumToSkip;
2763 // End of this scope, pop it and try the next child in the containing
2765 MatchScopes.pop_back();
2772 void SelectionDAGISel::CannotYetSelect(SDNode *N) {
2774 raw_string_ostream Msg(msg);
2775 Msg << "Cannot yet select: ";
2777 if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN &&
2778 N->getOpcode() != ISD::INTRINSIC_WO_CHAIN &&
2779 N->getOpcode() != ISD::INTRINSIC_VOID) {
2780 N->printrFull(Msg, CurDAG);
2782 bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other;
2784 cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue();
2785 if (iid < Intrinsic::num_intrinsics)
2786 Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid);
2787 else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo())
2788 Msg << "target intrinsic %" << TII->getName(iid);
2790 Msg << "unknown intrinsic #" << iid;
2792 llvm_report_error(Msg.str());
2795 char SelectionDAGISel::ID = 0;