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();
846 // FIXME: This shouldn't be needed, remove it.
847 CurDAG->RemoveDeadNodes();
851 void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn,
853 MachineModuleInfo *MMI,
855 const TargetInstrInfo &TII) {
856 // Initialize the Fast-ISel state, if needed.
857 FastISel *FastIS = 0;
859 FastIS = TLI.createFastISel(MF, MMI, DW,
862 FuncInfo->StaticAllocaMap
864 , FuncInfo->CatchInfoLost
868 unsigned MDDbgKind = Fn.getContext().getMDKindID("dbg");
870 // Iterate over all basic blocks in the function.
871 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
872 BasicBlock *LLVMBB = &*I;
873 BB = FuncInfo->MBBMap[LLVMBB];
875 BasicBlock::iterator const Begin = LLVMBB->begin();
876 BasicBlock::iterator const End = LLVMBB->end();
877 BasicBlock::iterator BI = Begin;
879 // Lower any arguments needed in this block if this is the entry block.
880 bool SuppressFastISel = false;
881 if (LLVMBB == &Fn.getEntryBlock()) {
882 LowerArguments(LLVMBB);
884 // If any of the arguments has the byval attribute, forgo
885 // fast-isel in the entry block.
888 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end();
890 if (Fn.paramHasAttr(j, Attribute::ByVal)) {
891 if (EnableFastISelVerbose || EnableFastISelAbort)
892 dbgs() << "FastISel skips entry block due to byval argument\n";
893 SuppressFastISel = true;
899 if (MMI && BB->isLandingPad()) {
900 // Add a label to mark the beginning of the landing pad. Deletion of the
901 // landing pad can thus be detected via the MachineModuleInfo.
902 MCSymbol *Label = MMI->addLandingPad(BB);
904 const TargetInstrDesc &II = TII.get(TargetOpcode::EH_LABEL);
905 BuildMI(BB, SDB->getCurDebugLoc(), II).addSym(Label);
907 // Mark exception register as live in.
908 unsigned Reg = TLI.getExceptionAddressRegister();
909 if (Reg) BB->addLiveIn(Reg);
911 // Mark exception selector register as live in.
912 Reg = TLI.getExceptionSelectorRegister();
913 if (Reg) BB->addLiveIn(Reg);
915 // FIXME: Hack around an exception handling flaw (PR1508): the personality
916 // function and list of typeids logically belong to the invoke (or, if you
917 // like, the basic block containing the invoke), and need to be associated
918 // with it in the dwarf exception handling tables. Currently however the
919 // information is provided by an intrinsic (eh.selector) that can be moved
920 // to unexpected places by the optimizers: if the unwind edge is critical,
921 // then breaking it can result in the intrinsics being in the successor of
922 // the landing pad, not the landing pad itself. This results
923 // in exceptions not being caught because no typeids are associated with
924 // the invoke. This may not be the only way things can go wrong, but it
925 // is the only way we try to work around for the moment.
926 BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
928 if (Br && Br->isUnconditional()) { // Critical edge?
929 BasicBlock::iterator I, E;
930 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
931 if (isa<EHSelectorInst>(I))
935 // No catch info found - try to extract some from the successor.
936 CopyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, *FuncInfo);
940 // Before doing SelectionDAG ISel, see if FastISel has been requested.
941 if (FastIS && !SuppressFastISel) {
942 // Emit code for any incoming arguments. This must happen before
943 // beginning FastISel on the entry block.
944 if (LLVMBB == &Fn.getEntryBlock()) {
945 CurDAG->setRoot(SDB->getControlRoot());
949 FastIS->startNewBlock(BB);
950 // Do FastISel on as many instructions as possible.
951 for (; BI != End; ++BI) {
952 // Just before the terminator instruction, insert instructions to
953 // feed PHI nodes in successor blocks.
954 if (isa<TerminatorInst>(BI))
955 if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, FastIS)) {
956 ++NumFastIselFailures;
957 ResetDebugLoc(SDB, FastIS);
958 if (EnableFastISelVerbose || EnableFastISelAbort) {
959 dbgs() << "FastISel miss: ";
962 assert(!EnableFastISelAbort &&
963 "FastISel didn't handle a PHI in a successor");
967 SetDebugLoc(MDDbgKind, BI, SDB, FastIS, &MF);
969 // Try to select the instruction with FastISel.
970 if (FastIS->SelectInstruction(BI)) {
971 ResetDebugLoc(SDB, FastIS);
975 // Clear out the debug location so that it doesn't carry over to
976 // unrelated instructions.
977 ResetDebugLoc(SDB, FastIS);
979 // Then handle certain instructions as single-LLVM-Instruction blocks.
980 if (isa<CallInst>(BI)) {
981 ++NumFastIselFailures;
982 if (EnableFastISelVerbose || EnableFastISelAbort) {
983 dbgs() << "FastISel missed call: ";
987 if (!BI->getType()->isVoidTy()) {
988 unsigned &R = FuncInfo->ValueMap[BI];
990 R = FuncInfo->CreateRegForValue(BI);
993 bool HadTailCall = false;
994 SelectBasicBlock(LLVMBB, BI, llvm::next(BI), HadTailCall);
996 // If the call was emitted as a tail call, we're done with the block.
1002 // If the instruction was codegen'd with multiple blocks,
1003 // inform the FastISel object where to resume inserting.
1004 FastIS->setCurrentBlock(BB);
1008 // Otherwise, give up on FastISel for the rest of the block.
1009 // For now, be a little lenient about non-branch terminators.
1010 if (!isa<TerminatorInst>(BI) || isa<BranchInst>(BI)) {
1011 ++NumFastIselFailures;
1012 if (EnableFastISelVerbose || EnableFastISelAbort) {
1013 dbgs() << "FastISel miss: ";
1016 if (EnableFastISelAbort)
1017 // The "fast" selector couldn't handle something and bailed.
1018 // For the purpose of debugging, just abort.
1019 llvm_unreachable("FastISel didn't select the entire block");
1025 // Run SelectionDAG instruction selection on the remainder of the block
1026 // not handled by FastISel. If FastISel is not run, this is the entire
1030 SelectBasicBlock(LLVMBB, BI, End, HadTailCall);
1040 SelectionDAGISel::FinishBasicBlock() {
1042 DEBUG(dbgs() << "Target-post-processed machine code:\n");
1045 DEBUG(dbgs() << "Total amount of phi nodes to update: "
1046 << SDB->PHINodesToUpdate.size() << "\n");
1047 DEBUG(for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i)
1048 dbgs() << "Node " << i << " : ("
1049 << SDB->PHINodesToUpdate[i].first
1050 << ", " << SDB->PHINodesToUpdate[i].second << ")\n");
1052 // Next, now that we know what the last MBB the LLVM BB expanded is, update
1053 // PHI nodes in successors.
1054 if (SDB->SwitchCases.empty() &&
1055 SDB->JTCases.empty() &&
1056 SDB->BitTestCases.empty()) {
1057 for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) {
1058 MachineInstr *PHI = SDB->PHINodesToUpdate[i].first;
1059 assert(PHI->isPHI() &&
1060 "This is not a machine PHI node that we are updating!");
1061 if (!BB->isSuccessor(PHI->getParent()))
1063 PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second,
1065 PHI->addOperand(MachineOperand::CreateMBB(BB));
1067 SDB->PHINodesToUpdate.clear();
1071 for (unsigned i = 0, e = SDB->BitTestCases.size(); i != e; ++i) {
1072 // Lower header first, if it wasn't already lowered
1073 if (!SDB->BitTestCases[i].Emitted) {
1074 // Set the current basic block to the mbb we wish to insert the code into
1075 BB = SDB->BitTestCases[i].Parent;
1076 SDB->setCurrentBasicBlock(BB);
1078 SDB->visitBitTestHeader(SDB->BitTestCases[i]);
1079 CurDAG->setRoot(SDB->getRoot());
1080 CodeGenAndEmitDAG();
1084 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j) {
1085 // Set the current basic block to the mbb we wish to insert the code into
1086 BB = SDB->BitTestCases[i].Cases[j].ThisBB;
1087 SDB->setCurrentBasicBlock(BB);
1090 SDB->visitBitTestCase(SDB->BitTestCases[i].Cases[j+1].ThisBB,
1091 SDB->BitTestCases[i].Reg,
1092 SDB->BitTestCases[i].Cases[j]);
1094 SDB->visitBitTestCase(SDB->BitTestCases[i].Default,
1095 SDB->BitTestCases[i].Reg,
1096 SDB->BitTestCases[i].Cases[j]);
1099 CurDAG->setRoot(SDB->getRoot());
1100 CodeGenAndEmitDAG();
1105 for (unsigned pi = 0, pe = SDB->PHINodesToUpdate.size(); pi != pe; ++pi) {
1106 MachineInstr *PHI = SDB->PHINodesToUpdate[pi].first;
1107 MachineBasicBlock *PHIBB = PHI->getParent();
1108 assert(PHI->isPHI() &&
1109 "This is not a machine PHI node that we are updating!");
1110 // This is "default" BB. We have two jumps to it. From "header" BB and
1111 // from last "case" BB.
1112 if (PHIBB == SDB->BitTestCases[i].Default) {
1113 PHI->addOperand(MachineOperand::
1114 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1115 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Parent));
1116 PHI->addOperand(MachineOperand::
1117 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1118 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Cases.
1121 // One of "cases" BB.
1122 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size();
1124 MachineBasicBlock* cBB = SDB->BitTestCases[i].Cases[j].ThisBB;
1125 if (cBB->isSuccessor(PHIBB)) {
1126 PHI->addOperand(MachineOperand::
1127 CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1128 PHI->addOperand(MachineOperand::CreateMBB(cBB));
1133 SDB->BitTestCases.clear();
1135 // If the JumpTable record is filled in, then we need to emit a jump table.
1136 // Updating the PHI nodes is tricky in this case, since we need to determine
1137 // whether the PHI is a successor of the range check MBB or the jump table MBB
1138 for (unsigned i = 0, e = SDB->JTCases.size(); i != e; ++i) {
1139 // Lower header first, if it wasn't already lowered
1140 if (!SDB->JTCases[i].first.Emitted) {
1141 // Set the current basic block to the mbb we wish to insert the code into
1142 BB = SDB->JTCases[i].first.HeaderBB;
1143 SDB->setCurrentBasicBlock(BB);
1145 SDB->visitJumpTableHeader(SDB->JTCases[i].second, SDB->JTCases[i].first);
1146 CurDAG->setRoot(SDB->getRoot());
1147 CodeGenAndEmitDAG();
1151 // Set the current basic block to the mbb we wish to insert the code into
1152 BB = SDB->JTCases[i].second.MBB;
1153 SDB->setCurrentBasicBlock(BB);
1155 SDB->visitJumpTable(SDB->JTCases[i].second);
1156 CurDAG->setRoot(SDB->getRoot());
1157 CodeGenAndEmitDAG();
1161 for (unsigned pi = 0, pe = SDB->PHINodesToUpdate.size(); pi != pe; ++pi) {
1162 MachineInstr *PHI = SDB->PHINodesToUpdate[pi].first;
1163 MachineBasicBlock *PHIBB = PHI->getParent();
1164 assert(PHI->isPHI() &&
1165 "This is not a machine PHI node that we are updating!");
1166 // "default" BB. We can go there only from header BB.
1167 if (PHIBB == SDB->JTCases[i].second.Default) {
1169 (MachineOperand::CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1171 (MachineOperand::CreateMBB(SDB->JTCases[i].first.HeaderBB));
1173 // JT BB. Just iterate over successors here
1174 if (BB->isSuccessor(PHIBB)) {
1176 (MachineOperand::CreateReg(SDB->PHINodesToUpdate[pi].second, false));
1177 PHI->addOperand(MachineOperand::CreateMBB(BB));
1181 SDB->JTCases.clear();
1183 // If the switch block involved a branch to one of the actual successors, we
1184 // need to update PHI nodes in that block.
1185 for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) {
1186 MachineInstr *PHI = SDB->PHINodesToUpdate[i].first;
1187 assert(PHI->isPHI() &&
1188 "This is not a machine PHI node that we are updating!");
1189 if (BB->isSuccessor(PHI->getParent())) {
1190 PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second,
1192 PHI->addOperand(MachineOperand::CreateMBB(BB));
1196 // If we generated any switch lowering information, build and codegen any
1197 // additional DAGs necessary.
1198 for (unsigned i = 0, e = SDB->SwitchCases.size(); i != e; ++i) {
1199 // Set the current basic block to the mbb we wish to insert the code into
1200 MachineBasicBlock *ThisBB = BB = SDB->SwitchCases[i].ThisBB;
1201 SDB->setCurrentBasicBlock(BB);
1204 SDB->visitSwitchCase(SDB->SwitchCases[i]);
1205 CurDAG->setRoot(SDB->getRoot());
1206 CodeGenAndEmitDAG();
1208 // Handle any PHI nodes in successors of this chunk, as if we were coming
1209 // from the original BB before switch expansion. Note that PHI nodes can
1210 // occur multiple times in PHINodesToUpdate. We have to be very careful to
1211 // handle them the right number of times.
1212 while ((BB = SDB->SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
1213 // If new BB's are created during scheduling, the edges may have been
1214 // updated. That is, the edge from ThisBB to BB may have been split and
1215 // BB's predecessor is now another block.
1216 DenseMap<MachineBasicBlock*, MachineBasicBlock*>::iterator EI =
1217 SDB->EdgeMapping.find(BB);
1218 if (EI != SDB->EdgeMapping.end())
1219 ThisBB = EI->second;
1221 // BB may have been removed from the CFG if a branch was constant folded.
1222 if (ThisBB->isSuccessor(BB)) {
1223 for (MachineBasicBlock::iterator Phi = BB->begin();
1224 Phi != BB->end() && Phi->isPHI();
1226 // This value for this PHI node is recorded in PHINodesToUpdate.
1227 for (unsigned pn = 0; ; ++pn) {
1228 assert(pn != SDB->PHINodesToUpdate.size() &&
1229 "Didn't find PHI entry!");
1230 if (SDB->PHINodesToUpdate[pn].first == Phi) {
1231 Phi->addOperand(MachineOperand::
1232 CreateReg(SDB->PHINodesToUpdate[pn].second,
1234 Phi->addOperand(MachineOperand::CreateMBB(ThisBB));
1241 // Don't process RHS if same block as LHS.
1242 if (BB == SDB->SwitchCases[i].FalseBB)
1243 SDB->SwitchCases[i].FalseBB = 0;
1245 // If we haven't handled the RHS, do so now. Otherwise, we're done.
1246 SDB->SwitchCases[i].TrueBB = SDB->SwitchCases[i].FalseBB;
1247 SDB->SwitchCases[i].FalseBB = 0;
1249 assert(SDB->SwitchCases[i].TrueBB == 0 && SDB->SwitchCases[i].FalseBB == 0);
1252 SDB->SwitchCases.clear();
1254 SDB->PHINodesToUpdate.clear();
1258 /// Create the scheduler. If a specific scheduler was specified
1259 /// via the SchedulerRegistry, use it, otherwise select the
1260 /// one preferred by the target.
1262 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
1263 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1267 RegisterScheduler::setDefault(Ctor);
1270 return Ctor(this, OptLevel);
1273 ScheduleHazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
1274 return new ScheduleHazardRecognizer();
1277 //===----------------------------------------------------------------------===//
1278 // Helper functions used by the generated instruction selector.
1279 //===----------------------------------------------------------------------===//
1280 // Calls to these methods are generated by tblgen.
1282 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1283 /// the dag combiner simplified the 255, we still want to match. RHS is the
1284 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1285 /// specified in the .td file (e.g. 255).
1286 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1287 int64_t DesiredMaskS) const {
1288 const APInt &ActualMask = RHS->getAPIntValue();
1289 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1291 // If the actual mask exactly matches, success!
1292 if (ActualMask == DesiredMask)
1295 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1296 if (ActualMask.intersects(~DesiredMask))
1299 // Otherwise, the DAG Combiner may have proven that the value coming in is
1300 // either already zero or is not demanded. Check for known zero input bits.
1301 APInt NeededMask = DesiredMask & ~ActualMask;
1302 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1305 // TODO: check to see if missing bits are just not demanded.
1307 // Otherwise, this pattern doesn't match.
1311 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1312 /// the dag combiner simplified the 255, we still want to match. RHS is the
1313 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1314 /// specified in the .td file (e.g. 255).
1315 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1316 int64_t DesiredMaskS) const {
1317 const APInt &ActualMask = RHS->getAPIntValue();
1318 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1320 // If the actual mask exactly matches, success!
1321 if (ActualMask == DesiredMask)
1324 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1325 if (ActualMask.intersects(~DesiredMask))
1328 // Otherwise, the DAG Combiner may have proven that the value coming in is
1329 // either already zero or is not demanded. Check for known zero input bits.
1330 APInt NeededMask = DesiredMask & ~ActualMask;
1332 APInt KnownZero, KnownOne;
1333 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
1335 // If all the missing bits in the or are already known to be set, match!
1336 if ((NeededMask & KnownOne) == NeededMask)
1339 // TODO: check to see if missing bits are just not demanded.
1341 // Otherwise, this pattern doesn't match.
1346 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1347 /// by tblgen. Others should not call it.
1348 void SelectionDAGISel::
1349 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1350 std::vector<SDValue> InOps;
1351 std::swap(InOps, Ops);
1353 Ops.push_back(InOps[0]); // input chain.
1354 Ops.push_back(InOps[1]); // input asm string.
1356 unsigned i = 2, e = InOps.size();
1357 if (InOps[e-1].getValueType() == MVT::Flag)
1358 --e; // Don't process a flag operand if it is here.
1361 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1362 if ((Flags & 7) != 4 /*MEM*/) {
1363 // Just skip over this operand, copying the operands verbatim.
1364 Ops.insert(Ops.end(), InOps.begin()+i,
1365 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
1366 i += InlineAsm::getNumOperandRegisters(Flags) + 1;
1368 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 &&
1369 "Memory operand with multiple values?");
1370 // Otherwise, this is a memory operand. Ask the target to select it.
1371 std::vector<SDValue> SelOps;
1372 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps)) {
1373 llvm_report_error("Could not match memory address. Inline asm"
1377 // Add this to the output node.
1378 Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size()<< 3),
1380 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1385 // Add the flag input back if present.
1386 if (e != InOps.size())
1387 Ops.push_back(InOps.back());
1390 /// findFlagUse - Return use of EVT::Flag value produced by the specified
1393 static SDNode *findFlagUse(SDNode *N) {
1394 unsigned FlagResNo = N->getNumValues()-1;
1395 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1396 SDUse &Use = I.getUse();
1397 if (Use.getResNo() == FlagResNo)
1398 return Use.getUser();
1403 /// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
1404 /// This function recursively traverses up the operand chain, ignoring
1406 static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
1407 SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited,
1408 bool IgnoreChains) {
1409 // The NodeID's are given uniques ID's where a node ID is guaranteed to be
1410 // greater than all of its (recursive) operands. If we scan to a point where
1411 // 'use' is smaller than the node we're scanning for, then we know we will
1414 // The Use may be -1 (unassigned) if it is a newly allocated node. This can
1415 // happen because we scan down to newly selected nodes in the case of flag
1417 if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1))
1420 // Don't revisit nodes if we already scanned it and didn't fail, we know we
1421 // won't fail if we scan it again.
1422 if (!Visited.insert(Use))
1425 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
1426 // Ignore chain uses, they are validated by HandleMergeInputChains.
1427 if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains)
1430 SDNode *N = Use->getOperand(i).getNode();
1432 if (Use == ImmedUse || Use == Root)
1433 continue; // We are not looking for immediate use.
1438 // Traverse up the operand chain.
1439 if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
1445 /// IsProfitableToFold - Returns true if it's profitable to fold the specific
1446 /// operand node N of U during instruction selection that starts at Root.
1447 bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
1448 SDNode *Root) const {
1449 if (OptLevel == CodeGenOpt::None) return false;
1450 return N.hasOneUse();
1453 /// IsLegalToFold - Returns true if the specific operand node N of
1454 /// U can be folded during instruction selection that starts at Root.
1455 bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
1456 bool IgnoreChains) const {
1457 if (OptLevel == CodeGenOpt::None) return false;
1459 // If Root use can somehow reach N through a path that that doesn't contain
1460 // U then folding N would create a cycle. e.g. In the following
1461 // diagram, Root can reach N through X. If N is folded into into Root, then
1462 // X is both a predecessor and a successor of U.
1473 // * indicates nodes to be folded together.
1475 // If Root produces a flag, then it gets (even more) interesting. Since it
1476 // will be "glued" together with its flag use in the scheduler, we need to
1477 // check if it might reach N.
1496 // If FU (flag use) indirectly reaches N (the load), and Root folds N
1497 // (call it Fold), then X is a predecessor of FU and a successor of
1498 // Fold. But since Fold and FU are flagged together, this will create
1499 // a cycle in the scheduling graph.
1501 // If the node has flags, walk down the graph to the "lowest" node in the
1503 EVT VT = Root->getValueType(Root->getNumValues()-1);
1504 while (VT == MVT::Flag) {
1505 SDNode *FU = findFlagUse(Root);
1509 VT = Root->getValueType(Root->getNumValues()-1);
1511 // If our query node has a flag result with a use, we've walked up it. If
1512 // the user (which has already been selected) has a chain or indirectly uses
1513 // the chain, our WalkChainUsers predicate will not consider it. Because of
1514 // this, we cannot ignore chains in this predicate.
1515 IgnoreChains = false;
1519 SmallPtrSet<SDNode*, 16> Visited;
1520 return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains);
1523 SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
1524 std::vector<SDValue> Ops(N->op_begin(), N->op_end());
1525 SelectInlineAsmMemoryOperands(Ops);
1527 std::vector<EVT> VTs;
1528 VTs.push_back(MVT::Other);
1529 VTs.push_back(MVT::Flag);
1530 SDValue New = CurDAG->getNode(ISD::INLINEASM, N->getDebugLoc(),
1531 VTs, &Ops[0], Ops.size());
1533 return New.getNode();
1536 SDNode *SelectionDAGISel::Select_UNDEF(SDNode *N) {
1537 return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0));
1540 /// GetVBR - decode a vbr encoding whose top bit is set.
1541 ALWAYS_INLINE static uint64_t
1542 GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
1543 assert(Val >= 128 && "Not a VBR");
1544 Val &= 127; // Remove first vbr bit.
1549 NextBits = MatcherTable[Idx++];
1550 Val |= (NextBits&127) << Shift;
1552 } while (NextBits & 128);
1558 /// UpdateChainsAndFlags - When a match is complete, this method updates uses of
1559 /// interior flag and chain results to use the new flag and chain results.
1560 void SelectionDAGISel::
1561 UpdateChainsAndFlags(SDNode *NodeToMatch, SDValue InputChain,
1562 const SmallVectorImpl<SDNode*> &ChainNodesMatched,
1564 const SmallVectorImpl<SDNode*> &FlagResultNodesMatched,
1565 bool isMorphNodeTo) {
1566 SmallVector<SDNode*, 4> NowDeadNodes;
1568 ISelUpdater ISU(ISelPosition);
1570 // Now that all the normal results are replaced, we replace the chain and
1571 // flag results if present.
1572 if (!ChainNodesMatched.empty()) {
1573 assert(InputChain.getNode() != 0 &&
1574 "Matched input chains but didn't produce a chain");
1575 // Loop over all of the nodes we matched that produced a chain result.
1576 // Replace all the chain results with the final chain we ended up with.
1577 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1578 SDNode *ChainNode = ChainNodesMatched[i];
1580 // If this node was already deleted, don't look at it.
1581 if (ChainNode->getOpcode() == ISD::DELETED_NODE)
1584 // Don't replace the results of the root node if we're doing a
1586 if (ChainNode == NodeToMatch && isMorphNodeTo)
1589 SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1);
1590 if (ChainVal.getValueType() == MVT::Flag)
1591 ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
1592 assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
1593 CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain, &ISU);
1595 // If the node became dead, delete it.
1596 if (ChainNode->use_empty())
1597 NowDeadNodes.push_back(ChainNode);
1601 // If the result produces a flag, update any flag results in the matched
1602 // pattern with the flag result.
1603 if (InputFlag.getNode() != 0) {
1604 // Handle any interior nodes explicitly marked.
1605 for (unsigned i = 0, e = FlagResultNodesMatched.size(); i != e; ++i) {
1606 SDNode *FRN = FlagResultNodesMatched[i];
1608 // If this node was already deleted, don't look at it.
1609 if (FRN->getOpcode() == ISD::DELETED_NODE)
1612 assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Flag &&
1613 "Doesn't have a flag result");
1614 CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1),
1617 // If the node became dead, delete it.
1618 if (FRN->use_empty())
1619 NowDeadNodes.push_back(FRN);
1623 if (!NowDeadNodes.empty())
1624 CurDAG->RemoveDeadNodes(NowDeadNodes, &ISU);
1626 DEBUG(errs() << "ISEL: Match complete!\n");
1632 CR_LeadsToInteriorNode
1635 /// WalkChainUsers - Walk down the users of the specified chained node that is
1636 /// part of the pattern we're matching, looking at all of the users we find.
1637 /// This determines whether something is an interior node, whether we have a
1638 /// non-pattern node in between two pattern nodes (which prevent folding because
1639 /// it would induce a cycle) and whether we have a TokenFactor node sandwiched
1640 /// between pattern nodes (in which case the TF becomes part of the pattern).
1642 /// The walk we do here is guaranteed to be small because we quickly get down to
1643 /// already selected nodes "below" us.
1645 WalkChainUsers(SDNode *ChainedNode,
1646 SmallVectorImpl<SDNode*> &ChainedNodesInPattern,
1647 SmallVectorImpl<SDNode*> &InteriorChainedNodes) {
1648 ChainResult Result = CR_Simple;
1650 for (SDNode::use_iterator UI = ChainedNode->use_begin(),
1651 E = ChainedNode->use_end(); UI != E; ++UI) {
1652 // Make sure the use is of the chain, not some other value we produce.
1653 if (UI.getUse().getValueType() != MVT::Other) continue;
1657 // If we see an already-selected machine node, then we've gone beyond the
1658 // pattern that we're selecting down into the already selected chunk of the
1660 if (User->isMachineOpcode() ||
1661 User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
1664 if (User->getOpcode() == ISD::CopyToReg ||
1665 User->getOpcode() == ISD::CopyFromReg ||
1666 User->getOpcode() == ISD::INLINEASM ||
1667 User->getOpcode() == ISD::EH_LABEL) {
1668 // If their node ID got reset to -1 then they've already been selected.
1669 // Treat them like a MachineOpcode.
1670 if (User->getNodeId() == -1)
1674 // If we have a TokenFactor, we handle it specially.
1675 if (User->getOpcode() != ISD::TokenFactor) {
1676 // If the node isn't a token factor and isn't part of our pattern, then it
1677 // must be a random chained node in between two nodes we're selecting.
1678 // This happens when we have something like:
1683 // Because we structurally match the load/store as a read/modify/write,
1684 // but the call is chained between them. We cannot fold in this case
1685 // because it would induce a cycle in the graph.
1686 if (!std::count(ChainedNodesInPattern.begin(),
1687 ChainedNodesInPattern.end(), User))
1688 return CR_InducesCycle;
1690 // Otherwise we found a node that is part of our pattern. For example in:
1694 // This would happen when we're scanning down from the load and see the
1695 // store as a user. Record that there is a use of ChainedNode that is
1696 // part of the pattern and keep scanning uses.
1697 Result = CR_LeadsToInteriorNode;
1698 InteriorChainedNodes.push_back(User);
1702 // If we found a TokenFactor, there are two cases to consider: first if the
1703 // TokenFactor is just hanging "below" the pattern we're matching (i.e. no
1704 // uses of the TF are in our pattern) we just want to ignore it. Second,
1705 // the TokenFactor can be sandwiched in between two chained nodes, like so:
1711 // | \ DAG's like cheese
1714 // [TokenFactor] [Op]
1721 // In this case, the TokenFactor becomes part of our match and we rewrite it
1722 // as a new TokenFactor.
1724 // To distinguish these two cases, do a recursive walk down the uses.
1725 switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) {
1727 // If the uses of the TokenFactor are just already-selected nodes, ignore
1728 // it, it is "below" our pattern.
1730 case CR_InducesCycle:
1731 // If the uses of the TokenFactor lead to nodes that are not part of our
1732 // pattern that are not selected, folding would turn this into a cycle,
1734 return CR_InducesCycle;
1735 case CR_LeadsToInteriorNode:
1736 break; // Otherwise, keep processing.
1739 // Okay, we know we're in the interesting interior case. The TokenFactor
1740 // is now going to be considered part of the pattern so that we rewrite its
1741 // uses (it may have uses that are not part of the pattern) with the
1742 // ultimate chain result of the generated code. We will also add its chain
1743 // inputs as inputs to the ultimate TokenFactor we create.
1744 Result = CR_LeadsToInteriorNode;
1745 ChainedNodesInPattern.push_back(User);
1746 InteriorChainedNodes.push_back(User);
1753 /// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
1754 /// operation for when the pattern matched at least one node with a chains. The
1755 /// input vector contains a list of all of the chained nodes that we match. We
1756 /// must determine if this is a valid thing to cover (i.e. matching it won't
1757 /// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
1758 /// be used as the input node chain for the generated nodes.
1760 HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
1761 SelectionDAG *CurDAG) {
1762 // Walk all of the chained nodes we've matched, recursively scanning down the
1763 // users of the chain result. This adds any TokenFactor nodes that are caught
1764 // in between chained nodes to the chained and interior nodes list.
1765 SmallVector<SDNode*, 3> InteriorChainedNodes;
1766 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1767 if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched,
1768 InteriorChainedNodes) == CR_InducesCycle)
1769 return SDValue(); // Would induce a cycle.
1772 // Okay, we have walked all the matched nodes and collected TokenFactor nodes
1773 // that we are interested in. Form our input TokenFactor node.
1774 SmallVector<SDValue, 3> InputChains;
1775 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
1776 // Add the input chain of this node to the InputChains list (which will be
1777 // the operands of the generated TokenFactor) if it's not an interior node.
1778 SDNode *N = ChainNodesMatched[i];
1779 if (N->getOpcode() != ISD::TokenFactor) {
1780 if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N))
1783 // Otherwise, add the input chain.
1784 SDValue InChain = ChainNodesMatched[i]->getOperand(0);
1785 assert(InChain.getValueType() == MVT::Other && "Not a chain");
1786 InputChains.push_back(InChain);
1790 // If we have a token factor, we want to add all inputs of the token factor
1791 // that are not part of the pattern we're matching.
1792 for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) {
1793 if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(),
1794 N->getOperand(op).getNode()))
1795 InputChains.push_back(N->getOperand(op));
1800 if (InputChains.size() == 1)
1801 return InputChains[0];
1802 return CurDAG->getNode(ISD::TokenFactor, ChainNodesMatched[0]->getDebugLoc(),
1803 MVT::Other, &InputChains[0], InputChains.size());
1806 /// MorphNode - Handle morphing a node in place for the selector.
1807 SDNode *SelectionDAGISel::
1808 MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList,
1809 const SDValue *Ops, unsigned NumOps, unsigned EmitNodeInfo) {
1810 // It is possible we're using MorphNodeTo to replace a node with no
1811 // normal results with one that has a normal result (or we could be
1812 // adding a chain) and the input could have flags and chains as well.
1813 // In this case we need to shifting the operands down.
1814 // FIXME: This is a horrible hack and broken in obscure cases, no worse
1815 // than the old isel though. We should sink this into MorphNodeTo.
1816 int OldFlagResultNo = -1, OldChainResultNo = -1;
1818 unsigned NTMNumResults = Node->getNumValues();
1819 if (Node->getValueType(NTMNumResults-1) == MVT::Flag) {
1820 OldFlagResultNo = NTMNumResults-1;
1821 if (NTMNumResults != 1 &&
1822 Node->getValueType(NTMNumResults-2) == MVT::Other)
1823 OldChainResultNo = NTMNumResults-2;
1824 } else if (Node->getValueType(NTMNumResults-1) == MVT::Other)
1825 OldChainResultNo = NTMNumResults-1;
1827 // Call the underlying SelectionDAG routine to do the transmogrification. Note
1828 // that this deletes operands of the old node that become dead.
1829 SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops, NumOps);
1831 // MorphNodeTo can operate in two ways: if an existing node with the
1832 // specified operands exists, it can just return it. Otherwise, it
1833 // updates the node in place to have the requested operands.
1835 // If we updated the node in place, reset the node ID. To the isel,
1836 // this should be just like a newly allocated machine node.
1840 unsigned ResNumResults = Res->getNumValues();
1841 // Move the flag if needed.
1842 if ((EmitNodeInfo & OPFL_FlagOutput) && OldFlagResultNo != -1 &&
1843 (unsigned)OldFlagResultNo != ResNumResults-1)
1844 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldFlagResultNo),
1845 SDValue(Res, ResNumResults-1));
1847 if ((EmitNodeInfo & OPFL_FlagOutput) != 0)
1850 // Move the chain reference if needed.
1851 if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 &&
1852 (unsigned)OldChainResultNo != ResNumResults-1)
1853 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo),
1854 SDValue(Res, ResNumResults-1));
1856 // Otherwise, no replacement happened because the node already exists. Replace
1857 // Uses of the old node with the new one.
1859 CurDAG->ReplaceAllUsesWith(Node, Res);
1864 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1865 ALWAYS_INLINE static bool
1866 CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1867 SDValue N, const SmallVectorImpl<SDValue> &RecordedNodes) {
1868 // Accept if it is exactly the same as a previously recorded node.
1869 unsigned RecNo = MatcherTable[MatcherIndex++];
1870 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
1871 return N == RecordedNodes[RecNo];
1874 /// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
1875 ALWAYS_INLINE static bool
1876 CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1877 SelectionDAGISel &SDISel) {
1878 return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]);
1881 /// CheckNodePredicate - Implements OP_CheckNodePredicate.
1882 ALWAYS_INLINE static bool
1883 CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1884 SelectionDAGISel &SDISel, SDNode *N) {
1885 return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]);
1888 ALWAYS_INLINE static bool
1889 CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1891 return N->getOpcode() == MatcherTable[MatcherIndex++];
1894 ALWAYS_INLINE static bool
1895 CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1896 SDValue N, const TargetLowering &TLI) {
1897 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1898 if (N.getValueType() == VT) return true;
1900 // Handle the case when VT is iPTR.
1901 return VT == MVT::iPTR && N.getValueType() == TLI.getPointerTy();
1904 ALWAYS_INLINE static bool
1905 CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1906 SDValue N, const TargetLowering &TLI,
1908 if (ChildNo >= N.getNumOperands())
1909 return false; // Match fails if out of range child #.
1910 return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI);
1914 ALWAYS_INLINE static bool
1915 CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1917 return cast<CondCodeSDNode>(N)->get() ==
1918 (ISD::CondCode)MatcherTable[MatcherIndex++];
1921 ALWAYS_INLINE static bool
1922 CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1923 SDValue N, const TargetLowering &TLI) {
1924 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
1925 if (cast<VTSDNode>(N)->getVT() == VT)
1928 // Handle the case when VT is iPTR.
1929 return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI.getPointerTy();
1932 ALWAYS_INLINE static bool
1933 CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1935 int64_t Val = MatcherTable[MatcherIndex++];
1937 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1939 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
1940 return C != 0 && C->getSExtValue() == Val;
1943 ALWAYS_INLINE static bool
1944 CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1945 SDValue N, SelectionDAGISel &SDISel) {
1946 int64_t Val = MatcherTable[MatcherIndex++];
1948 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1950 if (N->getOpcode() != ISD::AND) return false;
1952 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1953 return C != 0 && SDISel.CheckAndMask(N.getOperand(0), C, Val);
1956 ALWAYS_INLINE static bool
1957 CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
1958 SDValue N, SelectionDAGISel &SDISel) {
1959 int64_t Val = MatcherTable[MatcherIndex++];
1961 Val = GetVBR(Val, MatcherTable, MatcherIndex);
1963 if (N->getOpcode() != ISD::OR) return false;
1965 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1966 return C != 0 && SDISel.CheckOrMask(N.getOperand(0), C, Val);
1969 /// IsPredicateKnownToFail - If we know how and can do so without pushing a
1970 /// scope, evaluate the current node. If the current predicate is known to
1971 /// fail, set Result=true and return anything. If the current predicate is
1972 /// known to pass, set Result=false and return the MatcherIndex to continue
1973 /// with. If the current predicate is unknown, set Result=false and return the
1974 /// MatcherIndex to continue with.
1975 static unsigned IsPredicateKnownToFail(const unsigned char *Table,
1976 unsigned Index, SDValue N,
1977 bool &Result, SelectionDAGISel &SDISel,
1978 SmallVectorImpl<SDValue> &RecordedNodes){
1979 switch (Table[Index++]) {
1982 return Index-1; // Could not evaluate this predicate.
1983 case SelectionDAGISel::OPC_CheckSame:
1984 Result = !::CheckSame(Table, Index, N, RecordedNodes);
1986 case SelectionDAGISel::OPC_CheckPatternPredicate:
1987 Result = !::CheckPatternPredicate(Table, Index, SDISel);
1989 case SelectionDAGISel::OPC_CheckPredicate:
1990 Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode());
1992 case SelectionDAGISel::OPC_CheckOpcode:
1993 Result = !::CheckOpcode(Table, Index, N.getNode());
1995 case SelectionDAGISel::OPC_CheckType:
1996 Result = !::CheckType(Table, Index, N, SDISel.TLI);
1998 case SelectionDAGISel::OPC_CheckChild0Type:
1999 case SelectionDAGISel::OPC_CheckChild1Type:
2000 case SelectionDAGISel::OPC_CheckChild2Type:
2001 case SelectionDAGISel::OPC_CheckChild3Type:
2002 case SelectionDAGISel::OPC_CheckChild4Type:
2003 case SelectionDAGISel::OPC_CheckChild5Type:
2004 case SelectionDAGISel::OPC_CheckChild6Type:
2005 case SelectionDAGISel::OPC_CheckChild7Type:
2006 Result = !::CheckChildType(Table, Index, N, SDISel.TLI,
2007 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Type);
2009 case SelectionDAGISel::OPC_CheckCondCode:
2010 Result = !::CheckCondCode(Table, Index, N);
2012 case SelectionDAGISel::OPC_CheckValueType:
2013 Result = !::CheckValueType(Table, Index, N, SDISel.TLI);
2015 case SelectionDAGISel::OPC_CheckInteger:
2016 Result = !::CheckInteger(Table, Index, N);
2018 case SelectionDAGISel::OPC_CheckAndImm:
2019 Result = !::CheckAndImm(Table, Index, N, SDISel);
2021 case SelectionDAGISel::OPC_CheckOrImm:
2022 Result = !::CheckOrImm(Table, Index, N, SDISel);
2029 /// FailIndex - If this match fails, this is the index to continue with.
2032 /// NodeStack - The node stack when the scope was formed.
2033 SmallVector<SDValue, 4> NodeStack;
2035 /// NumRecordedNodes - The number of recorded nodes when the scope was formed.
2036 unsigned NumRecordedNodes;
2038 /// NumMatchedMemRefs - The number of matched memref entries.
2039 unsigned NumMatchedMemRefs;
2041 /// InputChain/InputFlag - The current chain/flag
2042 SDValue InputChain, InputFlag;
2044 /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty.
2045 bool HasChainNodesMatched, HasFlagResultNodesMatched;
2048 SDNode *SelectionDAGISel::
2049 SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
2050 unsigned TableSize) {
2051 // FIXME: Should these even be selected? Handle these cases in the caller?
2052 switch (NodeToMatch->getOpcode()) {
2055 case ISD::EntryToken: // These nodes remain the same.
2056 case ISD::BasicBlock:
2058 //case ISD::VALUETYPE:
2059 //case ISD::CONDCODE:
2060 case ISD::HANDLENODE:
2061 case ISD::TargetConstant:
2062 case ISD::TargetConstantFP:
2063 case ISD::TargetConstantPool:
2064 case ISD::TargetFrameIndex:
2065 case ISD::TargetExternalSymbol:
2066 case ISD::TargetBlockAddress:
2067 case ISD::TargetJumpTable:
2068 case ISD::TargetGlobalTLSAddress:
2069 case ISD::TargetGlobalAddress:
2070 case ISD::TokenFactor:
2071 case ISD::CopyFromReg:
2072 case ISD::CopyToReg:
2074 NodeToMatch->setNodeId(-1); // Mark selected.
2076 case ISD::AssertSext:
2077 case ISD::AssertZext:
2078 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0),
2079 NodeToMatch->getOperand(0));
2081 case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch);
2082 case ISD::UNDEF: return Select_UNDEF(NodeToMatch);
2085 assert(!NodeToMatch->isMachineOpcode() && "Node already selected!");
2087 // Set up the node stack with NodeToMatch as the only node on the stack.
2088 SmallVector<SDValue, 8> NodeStack;
2089 SDValue N = SDValue(NodeToMatch, 0);
2090 NodeStack.push_back(N);
2092 // MatchScopes - Scopes used when matching, if a match failure happens, this
2093 // indicates where to continue checking.
2094 SmallVector<MatchScope, 8> MatchScopes;
2096 // RecordedNodes - This is the set of nodes that have been recorded by the
2098 SmallVector<SDValue, 8> RecordedNodes;
2100 // MatchedMemRefs - This is the set of MemRef's we've seen in the input
2102 SmallVector<MachineMemOperand*, 2> MatchedMemRefs;
2104 // These are the current input chain and flag for use when generating nodes.
2105 // Various Emit operations change these. For example, emitting a copytoreg
2106 // uses and updates these.
2107 SDValue InputChain, InputFlag;
2109 // ChainNodesMatched - If a pattern matches nodes that have input/output
2110 // chains, the OPC_EmitMergeInputChains operation is emitted which indicates
2111 // which ones they are. The result is captured into this list so that we can
2112 // update the chain results when the pattern is complete.
2113 SmallVector<SDNode*, 3> ChainNodesMatched;
2114 SmallVector<SDNode*, 3> FlagResultNodesMatched;
2116 DEBUG(errs() << "ISEL: Starting pattern match on root node: ";
2117 NodeToMatch->dump(CurDAG);
2120 // Determine where to start the interpreter. Normally we start at opcode #0,
2121 // but if the state machine starts with an OPC_SwitchOpcode, then we
2122 // accelerate the first lookup (which is guaranteed to be hot) with the
2123 // OpcodeOffset table.
2124 unsigned MatcherIndex = 0;
2126 if (!OpcodeOffset.empty()) {
2127 // Already computed the OpcodeOffset table, just index into it.
2128 if (N.getOpcode() < OpcodeOffset.size())
2129 MatcherIndex = OpcodeOffset[N.getOpcode()];
2130 DEBUG(errs() << " Initial Opcode index to " << MatcherIndex << "\n");
2132 } else if (MatcherTable[0] == OPC_SwitchOpcode) {
2133 // Otherwise, the table isn't computed, but the state machine does start
2134 // with an OPC_SwitchOpcode instruction. Populate the table now, since this
2135 // is the first time we're selecting an instruction.
2138 // Get the size of this case.
2139 unsigned CaseSize = MatcherTable[Idx++];
2141 CaseSize = GetVBR(CaseSize, MatcherTable, Idx);
2142 if (CaseSize == 0) break;
2144 // Get the opcode, add the index to the table.
2145 unsigned Opc = MatcherTable[Idx++];
2146 if (Opc >= OpcodeOffset.size())
2147 OpcodeOffset.resize((Opc+1)*2);
2148 OpcodeOffset[Opc] = Idx;
2152 // Okay, do the lookup for the first opcode.
2153 if (N.getOpcode() < OpcodeOffset.size())
2154 MatcherIndex = OpcodeOffset[N.getOpcode()];
2158 assert(MatcherIndex < TableSize && "Invalid index");
2160 unsigned CurrentOpcodeIndex = MatcherIndex;
2162 BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++];
2165 // Okay, the semantics of this operation are that we should push a scope
2166 // then evaluate the first child. However, pushing a scope only to have
2167 // the first check fail (which then pops it) is inefficient. If we can
2168 // determine immediately that the first check (or first several) will
2169 // immediately fail, don't even bother pushing a scope for them.
2173 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2174 if (NumToSkip & 128)
2175 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2176 // Found the end of the scope with no match.
2177 if (NumToSkip == 0) {
2182 FailIndex = MatcherIndex+NumToSkip;
2184 // If we can't evaluate this predicate without pushing a scope (e.g. if
2185 // it is a 'MoveParent') or if the predicate succeeds on this node, we
2186 // push the scope and evaluate the full predicate chain.
2188 MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
2189 Result, *this, RecordedNodes);
2193 DEBUG(errs() << " Skipped scope entry at index " << MatcherIndex
2194 << " continuing at " << FailIndex << "\n");
2197 // Otherwise, we know that this case of the Scope is guaranteed to fail,
2198 // move to the next case.
2199 MatcherIndex = FailIndex;
2202 // If the whole scope failed to match, bail.
2203 if (FailIndex == 0) break;
2205 // Push a MatchScope which indicates where to go if the first child fails
2207 MatchScope NewEntry;
2208 NewEntry.FailIndex = FailIndex;
2209 NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
2210 NewEntry.NumRecordedNodes = RecordedNodes.size();
2211 NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
2212 NewEntry.InputChain = InputChain;
2213 NewEntry.InputFlag = InputFlag;
2214 NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty();
2215 NewEntry.HasFlagResultNodesMatched = !FlagResultNodesMatched.empty();
2216 MatchScopes.push_back(NewEntry);
2219 case OPC_RecordNode:
2220 // Remember this node, it may end up being an operand in the pattern.
2221 RecordedNodes.push_back(N);
2224 case OPC_RecordChild0: case OPC_RecordChild1:
2225 case OPC_RecordChild2: case OPC_RecordChild3:
2226 case OPC_RecordChild4: case OPC_RecordChild5:
2227 case OPC_RecordChild6: case OPC_RecordChild7: {
2228 unsigned ChildNo = Opcode-OPC_RecordChild0;
2229 if (ChildNo >= N.getNumOperands())
2230 break; // Match fails if out of range child #.
2232 RecordedNodes.push_back(N->getOperand(ChildNo));
2235 case OPC_RecordMemRef:
2236 MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand());
2239 case OPC_CaptureFlagInput:
2240 // If the current node has an input flag, capture it in InputFlag.
2241 if (N->getNumOperands() != 0 &&
2242 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag)
2243 InputFlag = N->getOperand(N->getNumOperands()-1);
2246 case OPC_MoveChild: {
2247 unsigned ChildNo = MatcherTable[MatcherIndex++];
2248 if (ChildNo >= N.getNumOperands())
2249 break; // Match fails if out of range child #.
2250 N = N.getOperand(ChildNo);
2251 NodeStack.push_back(N);
2255 case OPC_MoveParent:
2256 // Pop the current node off the NodeStack.
2257 NodeStack.pop_back();
2258 assert(!NodeStack.empty() && "Node stack imbalance!");
2259 N = NodeStack.back();
2263 if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
2265 case OPC_CheckPatternPredicate:
2266 if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break;
2268 case OPC_CheckPredicate:
2269 if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this,
2273 case OPC_CheckComplexPat: {
2274 unsigned CPNum = MatcherTable[MatcherIndex++];
2275 unsigned RecNo = MatcherTable[MatcherIndex++];
2276 assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
2277 if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo], CPNum,
2282 case OPC_CheckOpcode:
2283 if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
2287 if (!::CheckType(MatcherTable, MatcherIndex, N, TLI)) break;
2290 case OPC_SwitchOpcode: {
2291 unsigned CurNodeOpcode = N.getOpcode();
2292 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2295 // Get the size of this case.
2296 CaseSize = MatcherTable[MatcherIndex++];
2298 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2299 if (CaseSize == 0) break;
2301 // If the opcode matches, then we will execute this case.
2302 if (CurNodeOpcode == MatcherTable[MatcherIndex++])
2305 // Otherwise, skip over this case.
2306 MatcherIndex += CaseSize;
2309 // If no cases matched, bail out.
2310 if (CaseSize == 0) break;
2312 // Otherwise, execute the case we found.
2313 DEBUG(errs() << " OpcodeSwitch from " << SwitchStart
2314 << " to " << MatcherIndex << "\n");
2318 case OPC_SwitchType: {
2319 MVT::SimpleValueType CurNodeVT = N.getValueType().getSimpleVT().SimpleTy;
2320 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
2323 // Get the size of this case.
2324 CaseSize = MatcherTable[MatcherIndex++];
2326 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
2327 if (CaseSize == 0) break;
2329 MVT::SimpleValueType CaseVT =
2330 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2331 if (CaseVT == MVT::iPTR)
2332 CaseVT = TLI.getPointerTy().SimpleTy;
2334 // If the VT matches, then we will execute this case.
2335 if (CurNodeVT == CaseVT)
2338 // Otherwise, skip over this case.
2339 MatcherIndex += CaseSize;
2342 // If no cases matched, bail out.
2343 if (CaseSize == 0) break;
2345 // Otherwise, execute the case we found.
2346 DEBUG(errs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
2347 << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
2350 case OPC_CheckChild0Type: case OPC_CheckChild1Type:
2351 case OPC_CheckChild2Type: case OPC_CheckChild3Type:
2352 case OPC_CheckChild4Type: case OPC_CheckChild5Type:
2353 case OPC_CheckChild6Type: case OPC_CheckChild7Type:
2354 if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI,
2355 Opcode-OPC_CheckChild0Type))
2358 case OPC_CheckCondCode:
2359 if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
2361 case OPC_CheckValueType:
2362 if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI)) break;
2364 case OPC_CheckInteger:
2365 if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
2367 case OPC_CheckAndImm:
2368 if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
2370 case OPC_CheckOrImm:
2371 if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
2374 case OPC_CheckFoldableChainNode: {
2375 assert(NodeStack.size() != 1 && "No parent node");
2376 // Verify that all intermediate nodes between the root and this one have
2378 bool HasMultipleUses = false;
2379 for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i)
2380 if (!NodeStack[i].hasOneUse()) {
2381 HasMultipleUses = true;
2384 if (HasMultipleUses) break;
2386 // Check to see that the target thinks this is profitable to fold and that
2387 // we can fold it without inducing cycles in the graph.
2388 if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2390 !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
2391 NodeToMatch, true/*We validate our own chains*/))
2396 case OPC_EmitInteger: {
2397 MVT::SimpleValueType VT =
2398 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2399 int64_t Val = MatcherTable[MatcherIndex++];
2401 Val = GetVBR(Val, MatcherTable, MatcherIndex);
2402 RecordedNodes.push_back(CurDAG->getTargetConstant(Val, VT));
2405 case OPC_EmitRegister: {
2406 MVT::SimpleValueType VT =
2407 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2408 unsigned RegNo = MatcherTable[MatcherIndex++];
2409 RecordedNodes.push_back(CurDAG->getRegister(RegNo, VT));
2413 case OPC_EmitConvertToTarget: {
2414 // Convert from IMM/FPIMM to target version.
2415 unsigned RecNo = MatcherTable[MatcherIndex++];
2416 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2417 SDValue Imm = RecordedNodes[RecNo];
2419 if (Imm->getOpcode() == ISD::Constant) {
2420 int64_t Val = cast<ConstantSDNode>(Imm)->getZExtValue();
2421 Imm = CurDAG->getTargetConstant(Val, Imm.getValueType());
2422 } else if (Imm->getOpcode() == ISD::ConstantFP) {
2423 const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue();
2424 Imm = CurDAG->getTargetConstantFP(*Val, Imm.getValueType());
2427 RecordedNodes.push_back(Imm);
2431 case OPC_EmitMergeInputChains: {
2432 assert(InputChain.getNode() == 0 &&
2433 "EmitMergeInputChains should be the first chain producing node");
2434 // This node gets a list of nodes we matched in the input that have
2435 // chains. We want to token factor all of the input chains to these nodes
2436 // together. However, if any of the input chains is actually one of the
2437 // nodes matched in this pattern, then we have an intra-match reference.
2438 // Ignore these because the newly token factored chain should not refer to
2440 unsigned NumChains = MatcherTable[MatcherIndex++];
2441 assert(NumChains != 0 && "Can't TF zero chains");
2443 assert(ChainNodesMatched.empty() &&
2444 "Should only have one EmitMergeInputChains per match");
2446 // Read all of the chained nodes.
2447 for (unsigned i = 0; i != NumChains; ++i) {
2448 unsigned RecNo = MatcherTable[MatcherIndex++];
2449 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2450 ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode());
2452 // FIXME: What if other value results of the node have uses not matched
2454 if (ChainNodesMatched.back() != NodeToMatch &&
2455 !RecordedNodes[RecNo].hasOneUse()) {
2456 ChainNodesMatched.clear();
2461 // If the inner loop broke out, the match fails.
2462 if (ChainNodesMatched.empty())
2465 // Merge the input chains if they are not intra-pattern references.
2466 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
2468 if (InputChain.getNode() == 0)
2469 break; // Failed to merge.
2474 case OPC_EmitCopyToReg: {
2475 unsigned RecNo = MatcherTable[MatcherIndex++];
2476 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2477 unsigned DestPhysReg = MatcherTable[MatcherIndex++];
2479 if (InputChain.getNode() == 0)
2480 InputChain = CurDAG->getEntryNode();
2482 InputChain = CurDAG->getCopyToReg(InputChain, NodeToMatch->getDebugLoc(),
2483 DestPhysReg, RecordedNodes[RecNo],
2486 InputFlag = InputChain.getValue(1);
2490 case OPC_EmitNodeXForm: {
2491 unsigned XFormNo = MatcherTable[MatcherIndex++];
2492 unsigned RecNo = MatcherTable[MatcherIndex++];
2493 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2494 RecordedNodes.push_back(RunSDNodeXForm(RecordedNodes[RecNo], XFormNo));
2499 case OPC_MorphNodeTo: {
2500 uint16_t TargetOpc = MatcherTable[MatcherIndex++];
2501 TargetOpc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
2502 unsigned EmitNodeInfo = MatcherTable[MatcherIndex++];
2503 // Get the result VT list.
2504 unsigned NumVTs = MatcherTable[MatcherIndex++];
2505 SmallVector<EVT, 4> VTs;
2506 for (unsigned i = 0; i != NumVTs; ++i) {
2507 MVT::SimpleValueType VT =
2508 (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
2509 if (VT == MVT::iPTR) VT = TLI.getPointerTy().SimpleTy;
2513 if (EmitNodeInfo & OPFL_Chain)
2514 VTs.push_back(MVT::Other);
2515 if (EmitNodeInfo & OPFL_FlagOutput)
2516 VTs.push_back(MVT::Flag);
2518 // This is hot code, so optimize the two most common cases of 1 and 2
2521 if (VTs.size() == 1)
2522 VTList = CurDAG->getVTList(VTs[0]);
2523 else if (VTs.size() == 2)
2524 VTList = CurDAG->getVTList(VTs[0], VTs[1]);
2526 VTList = CurDAG->getVTList(VTs.data(), VTs.size());
2528 // Get the operand list.
2529 unsigned NumOps = MatcherTable[MatcherIndex++];
2530 SmallVector<SDValue, 8> Ops;
2531 for (unsigned i = 0; i != NumOps; ++i) {
2532 unsigned RecNo = MatcherTable[MatcherIndex++];
2534 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2536 assert(RecNo < RecordedNodes.size() && "Invalid EmitNode");
2537 Ops.push_back(RecordedNodes[RecNo]);
2540 // If there are variadic operands to add, handle them now.
2541 if (EmitNodeInfo & OPFL_VariadicInfo) {
2542 // Determine the start index to copy from.
2543 unsigned FirstOpToCopy = getNumFixedFromVariadicInfo(EmitNodeInfo);
2544 FirstOpToCopy += (EmitNodeInfo & OPFL_Chain) ? 1 : 0;
2545 assert(NodeToMatch->getNumOperands() >= FirstOpToCopy &&
2546 "Invalid variadic node");
2547 // Copy all of the variadic operands, not including a potential flag
2549 for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands();
2551 SDValue V = NodeToMatch->getOperand(i);
2552 if (V.getValueType() == MVT::Flag) break;
2557 // If this has chain/flag inputs, add them.
2558 if (EmitNodeInfo & OPFL_Chain)
2559 Ops.push_back(InputChain);
2560 if ((EmitNodeInfo & OPFL_FlagInput) && InputFlag.getNode() != 0)
2561 Ops.push_back(InputFlag);
2565 if (Opcode != OPC_MorphNodeTo) {
2566 // If this is a normal EmitNode command, just create the new node and
2567 // add the results to the RecordedNodes list.
2568 Res = CurDAG->getMachineNode(TargetOpc, NodeToMatch->getDebugLoc(),
2569 VTList, Ops.data(), Ops.size());
2571 // Add all the non-flag/non-chain results to the RecordedNodes list.
2572 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
2573 if (VTs[i] == MVT::Other || VTs[i] == MVT::Flag) break;
2574 RecordedNodes.push_back(SDValue(Res, i));
2578 Res = MorphNode(NodeToMatch, TargetOpc, VTList, Ops.data(), Ops.size(),
2582 // If the node had chain/flag results, update our notion of the current
2584 if (EmitNodeInfo & OPFL_FlagOutput) {
2585 InputFlag = SDValue(Res, VTs.size()-1);
2586 if (EmitNodeInfo & OPFL_Chain)
2587 InputChain = SDValue(Res, VTs.size()-2);
2588 } else if (EmitNodeInfo & OPFL_Chain)
2589 InputChain = SDValue(Res, VTs.size()-1);
2591 // If the OPFL_MemRefs flag is set on this node, slap all of the
2592 // accumulated memrefs onto it.
2594 // FIXME: This is vastly incorrect for patterns with multiple outputs
2595 // instructions that access memory and for ComplexPatterns that match
2597 if (EmitNodeInfo & OPFL_MemRefs) {
2598 MachineSDNode::mmo_iterator MemRefs =
2599 MF->allocateMemRefsArray(MatchedMemRefs.size());
2600 std::copy(MatchedMemRefs.begin(), MatchedMemRefs.end(), MemRefs);
2601 cast<MachineSDNode>(Res)
2602 ->setMemRefs(MemRefs, MemRefs + MatchedMemRefs.size());
2606 << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created")
2607 << " node: "; Res->dump(CurDAG); errs() << "\n");
2609 // If this was a MorphNodeTo then we're completely done!
2610 if (Opcode == OPC_MorphNodeTo) {
2611 // Update chain and flag uses.
2612 UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched,
2613 InputFlag, FlagResultNodesMatched, true);
2620 case OPC_MarkFlagResults: {
2621 unsigned NumNodes = MatcherTable[MatcherIndex++];
2623 // Read and remember all the flag-result nodes.
2624 for (unsigned i = 0; i != NumNodes; ++i) {
2625 unsigned RecNo = MatcherTable[MatcherIndex++];
2627 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex);
2629 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
2630 FlagResultNodesMatched.push_back(RecordedNodes[RecNo].getNode());
2635 case OPC_CompleteMatch: {
2636 // The match has been completed, and any new nodes (if any) have been
2637 // created. Patch up references to the matched dag to use the newly
2639 unsigned NumResults = MatcherTable[MatcherIndex++];
2641 for (unsigned i = 0; i != NumResults; ++i) {
2642 unsigned ResSlot = MatcherTable[MatcherIndex++];
2644 ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex);
2646 assert(ResSlot < RecordedNodes.size() && "Invalid CheckSame");
2647 SDValue Res = RecordedNodes[ResSlot];
2649 // FIXME2: Eliminate this horrible hack by fixing the 'Gen' program
2650 // after (parallel) on input patterns are removed. This would also
2651 // allow us to stop encoding #results in OPC_CompleteMatch's table
2653 if (NodeToMatch->getNumValues() <= i ||
2654 NodeToMatch->getValueType(i) == MVT::Other ||
2655 NodeToMatch->getValueType(i) == MVT::Flag)
2657 assert((NodeToMatch->getValueType(i) == Res.getValueType() ||
2658 NodeToMatch->getValueType(i) == MVT::iPTR ||
2659 Res.getValueType() == MVT::iPTR ||
2660 NodeToMatch->getValueType(i).getSizeInBits() ==
2661 Res.getValueType().getSizeInBits()) &&
2662 "invalid replacement");
2663 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, i), Res);
2666 // If the root node defines a flag, add it to the flag nodes to update
2668 if (NodeToMatch->getValueType(NodeToMatch->getNumValues()-1) == MVT::Flag)
2669 FlagResultNodesMatched.push_back(NodeToMatch);
2671 // Update chain and flag uses.
2672 UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched,
2673 InputFlag, FlagResultNodesMatched, false);
2675 assert(NodeToMatch->use_empty() &&
2676 "Didn't replace all uses of the node?");
2678 // FIXME: We just return here, which interacts correctly with SelectRoot
2679 // above. We should fix this to not return an SDNode* anymore.
2684 // If the code reached this point, then the match failed. See if there is
2685 // another child to try in the current 'Scope', otherwise pop it until we
2686 // find a case to check.
2687 DEBUG(errs() << " Match failed at index " << CurrentOpcodeIndex << "\n");
2689 if (MatchScopes.empty()) {
2690 CannotYetSelect(NodeToMatch);
2694 // Restore the interpreter state back to the point where the scope was
2696 MatchScope &LastScope = MatchScopes.back();
2697 RecordedNodes.resize(LastScope.NumRecordedNodes);
2699 NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end());
2700 N = NodeStack.back();
2702 if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size())
2703 MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
2704 MatcherIndex = LastScope.FailIndex;
2706 DEBUG(errs() << " Continuing at " << MatcherIndex << "\n");
2708 InputChain = LastScope.InputChain;
2709 InputFlag = LastScope.InputFlag;
2710 if (!LastScope.HasChainNodesMatched)
2711 ChainNodesMatched.clear();
2712 if (!LastScope.HasFlagResultNodesMatched)
2713 FlagResultNodesMatched.clear();
2715 // Check to see what the offset is at the new MatcherIndex. If it is zero
2716 // we have reached the end of this scope, otherwise we have another child
2717 // in the current scope to try.
2718 unsigned NumToSkip = MatcherTable[MatcherIndex++];
2719 if (NumToSkip & 128)
2720 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
2722 // If we have another child in this scope to match, update FailIndex and
2724 if (NumToSkip != 0) {
2725 LastScope.FailIndex = MatcherIndex+NumToSkip;
2729 // End of this scope, pop it and try the next child in the containing
2731 MatchScopes.pop_back();
2738 void SelectionDAGISel::CannotYetSelect(SDNode *N) {
2740 raw_string_ostream Msg(msg);
2741 Msg << "Cannot yet select: ";
2743 if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN &&
2744 N->getOpcode() != ISD::INTRINSIC_WO_CHAIN &&
2745 N->getOpcode() != ISD::INTRINSIC_VOID) {
2746 N->printrFull(Msg, CurDAG);
2748 bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other;
2750 cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue();
2751 if (iid < Intrinsic::num_intrinsics)
2752 Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid);
2753 else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo())
2754 Msg << "target intrinsic %" << TII->getName(iid);
2756 Msg << "unknown intrinsic #" << iid;
2758 llvm_report_error(Msg.str());
2761 char SelectionDAGISel::ID = 0;