#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/DebugInfo.h"
#include "llvm/Constants.h"
-#include "llvm/CallingConv.h"
-#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
-#include "llvm/GlobalVariable.h"
#include "llvm/InlineAsm.h"
#include "llvm/Instructions.h"
#include "llvm/Intrinsics.h"
#include "llvm/CodeGen/GCStrategy.h"
#include "llvm/CodeGen/GCMetadata.h"
#include "llvm/CodeGen/MachineFunction.h"
-#include "llvm/CodeGen/MachineFunctionAnalysis.h"
-#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
-#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/ScheduleHazardRecognizer.h"
#include "llvm/CodeGen/SchedulerRegistry.h"
#include "llvm/CodeGen/SelectionDAG.h"
-#include "llvm/CodeGen/DwarfWriter.h"
#include "llvm/Target/TargetRegisterInfo.h"
-#include "llvm/Target/TargetData.h"
-#include "llvm/Target/TargetFrameInfo.h"
#include "llvm/Target/TargetIntrinsicInfo.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
-#include "llvm/Support/MathExtras.h"
#include "llvm/Support/Timer.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/ADT/Statistic.h"
#include <algorithm>
using namespace llvm;
+STATISTIC(NumFastIselFailures, "Number of instructions fast isel failed on");
+STATISTIC(NumDAGIselRetries,"Number of times dag isel has to try another path");
+
static cl::opt<bool>
EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
cl::desc("Enable verbose messages in the \"fast\" "
static cl::opt<bool>
EnableFastISelAbort("fast-isel-abort", cl::Hidden,
cl::desc("Enable abort calls when \"fast\" instruction fails"));
-static cl::opt<bool>
-SchedLiveInCopies("schedule-livein-copies", cl::Hidden,
- cl::desc("Schedule copies of livein registers"),
- cl::init(false));
#ifndef NDEBUG
static cl::opt<bool>
return 0;
}
-/// EmitLiveInCopy - Emit a copy for a live in physical register. If the
-/// physical register has only a single copy use, then coalesced the copy
-/// if possible.
-static void EmitLiveInCopy(MachineBasicBlock *MBB,
- MachineBasicBlock::iterator &InsertPos,
- unsigned VirtReg, unsigned PhysReg,
- const TargetRegisterClass *RC,
- DenseMap<MachineInstr*, unsigned> &CopyRegMap,
- const MachineRegisterInfo &MRI,
- const TargetRegisterInfo &TRI,
- const TargetInstrInfo &TII) {
- unsigned NumUses = 0;
- MachineInstr *UseMI = NULL;
- for (MachineRegisterInfo::use_iterator UI = MRI.use_begin(VirtReg),
- UE = MRI.use_end(); UI != UE; ++UI) {
- UseMI = &*UI;
- if (++NumUses > 1)
- break;
- }
-
- // If the number of uses is not one, or the use is not a move instruction,
- // don't coalesce. Also, only coalesce away a virtual register to virtual
- // register copy.
- bool Coalesced = false;
- unsigned SrcReg, DstReg, SrcSubReg, DstSubReg;
- if (NumUses == 1 &&
- TII.isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubReg, DstSubReg) &&
- TargetRegisterInfo::isVirtualRegister(DstReg)) {
- VirtReg = DstReg;
- Coalesced = true;
- }
-
- // Now find an ideal location to insert the copy.
- MachineBasicBlock::iterator Pos = InsertPos;
- while (Pos != MBB->begin()) {
- MachineInstr *PrevMI = prior(Pos);
- DenseMap<MachineInstr*, unsigned>::iterator RI = CopyRegMap.find(PrevMI);
- // copyRegToReg might emit multiple instructions to do a copy.
- unsigned CopyDstReg = (RI == CopyRegMap.end()) ? 0 : RI->second;
- if (CopyDstReg && !TRI.regsOverlap(CopyDstReg, PhysReg))
- // This is what the BB looks like right now:
- // r1024 = mov r0
- // ...
- // r1 = mov r1024
- //
- // We want to insert "r1025 = mov r1". Inserting this copy below the
- // move to r1024 makes it impossible for that move to be coalesced.
- //
- // r1025 = mov r1
- // r1024 = mov r0
- // ...
- // r1 = mov 1024
- // r2 = mov 1025
- break; // Woot! Found a good location.
- --Pos;
- }
-
- bool Emitted = TII.copyRegToReg(*MBB, Pos, VirtReg, PhysReg, RC, RC);
- assert(Emitted && "Unable to issue a live-in copy instruction!\n");
- (void) Emitted;
-
- CopyRegMap.insert(std::make_pair(prior(Pos), VirtReg));
- if (Coalesced) {
- if (&*InsertPos == UseMI) ++InsertPos;
- MBB->erase(UseMI);
- }
-}
-
-/// EmitLiveInCopies - If this is the first basic block in the function,
-/// and if it has live ins that need to be copied into vregs, emit the
-/// copies into the block.
-static void EmitLiveInCopies(MachineBasicBlock *EntryMBB,
- const MachineRegisterInfo &MRI,
- const TargetRegisterInfo &TRI,
- const TargetInstrInfo &TII) {
- if (SchedLiveInCopies) {
- // Emit the copies at a heuristically-determined location in the block.
- DenseMap<MachineInstr*, unsigned> CopyRegMap;
- MachineBasicBlock::iterator InsertPos = EntryMBB->begin();
- for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
- E = MRI.livein_end(); LI != E; ++LI)
- if (LI->second) {
- const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
- EmitLiveInCopy(EntryMBB, InsertPos, LI->second, LI->first,
- RC, CopyRegMap, MRI, TRI, TII);
- }
- } else {
- // Emit the copies into the top of the block.
- for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
- E = MRI.livein_end(); LI != E; ++LI)
- if (LI->second) {
- const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
- bool Emitted = TII.copyRegToReg(*EntryMBB, EntryMBB->begin(),
- LI->second, LI->first, RC, RC);
- assert(Emitted && "Unable to issue a live-in copy instruction!\n");
- (void) Emitted;
- }
- }
-}
-
//===----------------------------------------------------------------------===//
// SelectionDAGISel code
//===----------------------------------------------------------------------===//
delete FuncInfo;
}
-unsigned SelectionDAGISel::MakeReg(EVT VT) {
- return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT));
-}
-
void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<AliasAnalysis>();
AU.addPreserved<AliasAnalysis>();
AU.addRequired<GCModuleInfo>();
AU.addPreserved<GCModuleInfo>();
- AU.addRequired<DwarfWriter>();
- AU.addPreserved<DwarfWriter>();
MachineFunctionPass::getAnalysisUsage(AU);
}
bool SelectionDAGISel::runOnMachineFunction(MachineFunction &mf) {
- Function &Fn = *mf.getFunction();
-
// Do some sanity-checking on the command-line options.
assert((!EnableFastISelVerbose || EnableFastISel) &&
"-fast-isel-verbose requires -fast-isel");
assert((!EnableFastISelAbort || EnableFastISel) &&
"-fast-isel-abort requires -fast-isel");
- // Get alias analysis for load/store combining.
- AA = &getAnalysis<AliasAnalysis>();
-
- MF = &mf;
+ const Function &Fn = *mf.getFunction();
const TargetInstrInfo &TII = *TM.getInstrInfo();
const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
- if (Fn.hasGC())
- GFI = &getAnalysis<GCModuleInfo>().getFunctionInfo(Fn);
- else
- GFI = 0;
+ MF = &mf;
RegInfo = &MF->getRegInfo();
+ AA = &getAnalysis<AliasAnalysis>();
+ GFI = Fn.hasGC() ? &getAnalysis<GCModuleInfo>().getFunctionInfo(Fn) : 0;
+
DEBUG(dbgs() << "\n\n\n=== " << Fn.getName() << "\n");
- MachineModuleInfo *MMI = getAnalysisIfAvailable<MachineModuleInfo>();
- DwarfWriter *DW = getAnalysisIfAvailable<DwarfWriter>();
- CurDAG->init(*MF, MMI, DW);
+ CurDAG->init(*MF);
FuncInfo->set(Fn, *MF, EnableFastISel);
SDB->init(GFI, *AA);
- for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
- if (InvokeInst *Invoke = dyn_cast<InvokeInst>(I->getTerminator()))
- // Mark landing pad.
- FuncInfo->MBBMap[Invoke->getSuccessor(1)]->setIsLandingPad();
+ SelectAllBasicBlocks(Fn);
- SelectAllBasicBlocks(Fn, *MF, MMI, DW, TII);
+ // Release function-specific state. SDB and CurDAG are already cleared
+ // at this point.
+ FuncInfo->clear();
// If the first basic block in the function has live ins that need to be
// copied into vregs, emit the copies into the top of the block before
// emitting the code for the block.
- EmitLiveInCopies(MF->begin(), *RegInfo, TRI, TII);
-
- // Add function live-ins to entry block live-in set.
- for (MachineRegisterInfo::livein_iterator I = RegInfo->livein_begin(),
- E = RegInfo->livein_end(); I != E; ++I)
- MF->begin()->addLiveIn(I->first);
-
-#ifndef NDEBUG
- assert(FuncInfo->CatchInfoFound.size() == FuncInfo->CatchInfoLost.size() &&
- "Not all catch info was assigned to a landing pad!");
-#endif
-
- FuncInfo->clear();
+ RegInfo->EmitLiveInCopies(MF->begin(), TRI, TII);
return true;
}
/// SetDebugLoc - Update MF's and SDB's DebugLocs if debug information is
/// attached with this instruction.
-static void SetDebugLoc(unsigned MDDbgKind, Instruction *I,
- SelectionDAGBuilder *SDB,
+static void SetDebugLoc(const Instruction *I, SelectionDAGBuilder *SDB,
FastISel *FastIS, MachineFunction *MF) {
- if (isa<DbgInfoIntrinsic>(I)) return;
+ DebugLoc DL = I->getDebugLoc();
+ if (DL.isUnknown()) return;
- if (MDNode *Dbg = I->getMetadata(MDDbgKind)) {
- DILocation DILoc(Dbg);
- DebugLoc Loc = ExtractDebugLocation(DILoc, MF->getDebugLocInfo());
-
- SDB->setCurDebugLoc(Loc);
+ SDB->setCurDebugLoc(DL);
- if (FastIS)
- FastIS->setCurDebugLoc(Loc);
+ if (FastIS)
+ FastIS->setCurDebugLoc(DL);
- // If the function doesn't have a default debug location yet, set
- // it. This is kind of a hack.
- if (MF->getDefaultDebugLoc().isUnknown())
- MF->setDefaultDebugLoc(Loc);
- }
+ // If the function doesn't have a default debug location yet, set
+ // it. This is a total hack.
+ if (MF->getDefaultDebugLoc().isUnknown())
+ MF->setDefaultDebugLoc(DL);
}
/// ResetDebugLoc - Set MF's and SDB's DebugLocs to Unknown.
static void ResetDebugLoc(SelectionDAGBuilder *SDB, FastISel *FastIS) {
- SDB->setCurDebugLoc(DebugLoc::getUnknownLoc());
+ SDB->setCurDebugLoc(DebugLoc());
if (FastIS)
- FastIS->setCurDebugLoc(DebugLoc::getUnknownLoc());
+ FastIS->setCurDebugLoc(DebugLoc());
}
-void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB,
- BasicBlock::iterator Begin,
- BasicBlock::iterator End,
+void SelectionDAGISel::SelectBasicBlock(const BasicBlock *LLVMBB,
+ BasicBlock::const_iterator Begin,
+ BasicBlock::const_iterator End,
bool &HadTailCall) {
SDB->setCurrentBasicBlock(BB);
- unsigned MDDbgKind = LLVMBB->getContext().getMDKindID("dbg");
// Lower all of the non-terminator instructions. If a call is emitted
- // as a tail call, cease emitting nodes for this block.
- for (BasicBlock::iterator I = Begin; I != End && !SDB->HasTailCall; ++I) {
- SetDebugLoc(MDDbgKind, I, SDB, 0, MF);
-
- if (!isa<TerminatorInst>(I)) {
- SDB->visit(*I);
-
- // Set the current debug location back to "unknown" so that it doesn't
- // spuriously apply to subsequent instructions.
- ResetDebugLoc(SDB, 0);
- }
+ // as a tail call, cease emitting nodes for this block. Terminators
+ // are handled below.
+ for (BasicBlock::const_iterator I = Begin;
+ I != End && !SDB->HasTailCall && !isa<TerminatorInst>(I);
+ ++I) {
+ SetDebugLoc(I, SDB, 0, MF);
+ SDB->visit(*I);
+ ResetDebugLoc(SDB, 0);
}
if (!SDB->HasTailCall) {
// Ensure that all instructions which are used outside of their defining
// blocks are available as virtual registers. Invoke is handled elsewhere.
- for (BasicBlock::iterator I = Begin; I != End; ++I)
+ for (BasicBlock::const_iterator I = Begin; I != End; ++I)
if (!isa<PHINode>(I) && !isa<InvokeInst>(I))
SDB->CopyToExportRegsIfNeeded(I);
HandlePHINodesInSuccessorBlocks(LLVMBB);
// Lower the terminator after the copies are emitted.
- SetDebugLoc(MDDbgKind, LLVMBB->getTerminator(), SDB, 0, MF);
+ SetDebugLoc(LLVMBB->getTerminator(), SDB, 0, MF);
SDB->visit(*LLVMBB->getTerminator());
ResetDebugLoc(SDB, 0);
}
/// nodes from the worklist.
class SDOPsWorkListRemover : public SelectionDAG::DAGUpdateListener {
SmallVector<SDNode*, 128> &Worklist;
+ SmallPtrSet<SDNode*, 128> &InWorklist;
public:
- SDOPsWorkListRemover(SmallVector<SDNode*, 128> &wl) : Worklist(wl) {}
+ SDOPsWorkListRemover(SmallVector<SDNode*, 128> &wl,
+ SmallPtrSet<SDNode*, 128> &inwl)
+ : Worklist(wl), InWorklist(inwl) {}
+ void RemoveFromWorklist(SDNode *N) {
+ if (!InWorklist.erase(N)) return;
+
+ SmallVector<SDNode*, 128>::iterator I =
+ std::find(Worklist.begin(), Worklist.end(), N);
+ assert(I != Worklist.end() && "Not in worklist");
+
+ *I = Worklist.back();
+ Worklist.pop_back();
+ }
+
virtual void NodeDeleted(SDNode *N, SDNode *E) {
- Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), N),
- Worklist.end());
+ RemoveFromWorklist(N);
}
virtual void NodeUpdated(SDNode *N) {
/// x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free.
void SelectionDAGISel::ShrinkDemandedOps() {
SmallVector<SDNode*, 128> Worklist;
+ SmallPtrSet<SDNode*, 128> InWorklist;
// Add all the dag nodes to the worklist.
Worklist.reserve(CurDAG->allnodes_size());
for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
- E = CurDAG->allnodes_end(); I != E; ++I)
+ E = CurDAG->allnodes_end(); I != E; ++I) {
Worklist.push_back(I);
-
- APInt Mask;
- APInt KnownZero;
- APInt KnownOne;
+ InWorklist.insert(I);
+ }
TargetLowering::TargetLoweringOpt TLO(*CurDAG, true);
while (!Worklist.empty()) {
SDNode *N = Worklist.pop_back_val();
+ InWorklist.erase(N);
if (N->use_empty() && N != CurDAG->getRoot().getNode()) {
+ // Deleting this node may make its operands dead, add them to the worklist
+ // if they aren't already there.
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
+ if (InWorklist.insert(N->getOperand(i).getNode()))
+ Worklist.push_back(N->getOperand(i).getNode());
+
CurDAG->DeleteNode(N);
continue;
}
// Run ShrinkDemandedOp on scalar binary operations.
- if (N->getNumValues() == 1 &&
- N->getValueType(0).isSimple() && N->getValueType(0).isInteger()) {
- unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
- APInt Demanded = APInt::getAllOnesValue(BitWidth);
- APInt KnownZero, KnownOne;
- if (TLI.SimplifyDemandedBits(SDValue(N, 0), Demanded,
- KnownZero, KnownOne, TLO) ||
- (N->getOpcode() == ISD::TRUNCATE &&
- TrivialTruncElim(SDValue(N, 0), TLO))) {
- // Revisit the node.
- Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), N),
- Worklist.end());
- Worklist.push_back(N);
-
- // Replace the old value with the new one.
- DEBUG(errs() << "\nReplacing ";
- TLO.Old.getNode()->dump(CurDAG);
- errs() << "\nWith: ";
- TLO.New.getNode()->dump(CurDAG);
- errs() << '\n');
-
- Worklist.push_back(TLO.New.getNode());
-
- SDOPsWorkListRemover DeadNodes(Worklist);
- CurDAG->ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, &DeadNodes);
-
- if (TLO.Old.getNode()->use_empty()) {
- for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands();
- i != e; ++i) {
- SDNode *OpNode = TLO.Old.getNode()->getOperand(i).getNode();
- if (OpNode->hasOneUse()) {
- Worklist.erase(std::remove(Worklist.begin(), Worklist.end(),
- OpNode), Worklist.end());
- Worklist.push_back(OpNode);
- }
- }
+ if (N->getNumValues() != 1 ||
+ !N->getValueType(0).isSimple() || !N->getValueType(0).isInteger())
+ continue;
+
+ unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
+ APInt Demanded = APInt::getAllOnesValue(BitWidth);
+ APInt KnownZero, KnownOne;
+ if (!TLI.SimplifyDemandedBits(SDValue(N, 0), Demanded,
+ KnownZero, KnownOne, TLO) &&
+ (N->getOpcode() != ISD::TRUNCATE ||
+ !TrivialTruncElim(SDValue(N, 0), TLO)))
+ continue;
+
+ // Revisit the node.
+ assert(!InWorklist.count(N) && "Already in worklist");
+ Worklist.push_back(N);
+ InWorklist.insert(N);
- Worklist.erase(std::remove(Worklist.begin(), Worklist.end(),
- TLO.Old.getNode()), Worklist.end());
- CurDAG->DeleteNode(TLO.Old.getNode());
- }
+ // Replace the old value with the new one.
+ DEBUG(errs() << "\nShrinkDemandedOps replacing ";
+ TLO.Old.getNode()->dump(CurDAG);
+ errs() << "\nWith: ";
+ TLO.New.getNode()->dump(CurDAG);
+ errs() << '\n');
+
+ if (InWorklist.insert(TLO.New.getNode()))
+ Worklist.push_back(TLO.New.getNode());
+
+ SDOPsWorkListRemover DeadNodes(Worklist, InWorklist);
+ CurDAG->ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, &DeadNodes);
+
+ if (!TLO.Old.getNode()->use_empty()) continue;
+
+ for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands();
+ i != e; ++i) {
+ SDNode *OpNode = TLO.Old.getNode()->getOperand(i).getNode();
+ if (OpNode->hasOneUse()) {
+ // Add OpNode to the end of the list to revisit.
+ DeadNodes.RemoveFromWorklist(OpNode);
+ Worklist.push_back(OpNode);
+ InWorklist.insert(OpNode);
}
}
+
+ DeadNodes.RemoveFromWorklist(TLO.Old.getNode());
+ CurDAG->DeleteNode(TLO.Old.getNode());
}
}
DEBUG(dbgs() << "Optimized legalized selection DAG:\n");
DEBUG(CurDAG->dump());
- if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName);
-
if (OptLevel != CodeGenOpt::None) {
ShrinkDemandedOps();
ComputeLiveOutVRegInfo();
}
+ if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName);
+
// Third, instruction select all of the operations to machine code, adding the
// code to the MachineBasicBlock.
if (TimePassesIsEnabled) {
SDNode *ResNode = Select(Node);
+ // FIXME: This is pretty gross. 'Select' should be changed to not return
+ // anything at all and this code should be nuked with a tactical strike.
+
// If node should not be replaced, continue with the next one.
- if (ResNode == Node)
+ if (ResNode == Node || Node->getOpcode() == ISD::DELETED_NODE)
continue;
// Replace node.
if (ResNode)
DEBUG(errs() << "===== Instruction selection ends:\n");
PostprocessISelDAG();
-
- // FIXME: This shouldn't be needed, remove it.
- CurDAG->RemoveDeadNodes();
}
+/// PrepareEHLandingPad - Emit an EH_LABEL, set up live-in registers, and
+/// do other setup for EH landing-pad blocks.
+void SelectionDAGISel::PrepareEHLandingPad(MachineBasicBlock *BB) {
+ // Add a label to mark the beginning of the landing pad. Deletion of the
+ // landing pad can thus be detected via the MachineModuleInfo.
+ MCSymbol *Label = MF->getMMI().addLandingPad(BB);
+
+ const TargetInstrDesc &II =
+ TLI.getTargetMachine().getInstrInfo()->get(TargetOpcode::EH_LABEL);
+ BuildMI(BB, SDB->getCurDebugLoc(), II).addSym(Label);
+
+ // Mark exception register as live in.
+ unsigned Reg = TLI.getExceptionAddressRegister();
+ if (Reg) BB->addLiveIn(Reg);
+
+ // Mark exception selector register as live in.
+ Reg = TLI.getExceptionSelectorRegister();
+ if (Reg) BB->addLiveIn(Reg);
+
+ // FIXME: Hack around an exception handling flaw (PR1508): the personality
+ // function and list of typeids logically belong to the invoke (or, if you
+ // like, the basic block containing the invoke), and need to be associated
+ // with it in the dwarf exception handling tables. Currently however the
+ // information is provided by an intrinsic (eh.selector) that can be moved
+ // to unexpected places by the optimizers: if the unwind edge is critical,
+ // then breaking it can result in the intrinsics being in the successor of
+ // the landing pad, not the landing pad itself. This results
+ // in exceptions not being caught because no typeids are associated with
+ // the invoke. This may not be the only way things can go wrong, but it
+ // is the only way we try to work around for the moment.
+ const BasicBlock *LLVMBB = BB->getBasicBlock();
+ const BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
+
+ if (Br && Br->isUnconditional()) { // Critical edge?
+ BasicBlock::const_iterator I, E;
+ for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
+ if (isa<EHSelectorInst>(I))
+ break;
+
+ if (I == E)
+ // No catch info found - try to extract some from the successor.
+ CopyCatchInfo(Br->getSuccessor(0), LLVMBB, &MF->getMMI(), *FuncInfo);
+ }
+}
-void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn,
- MachineFunction &MF,
- MachineModuleInfo *MMI,
- DwarfWriter *DW,
- const TargetInstrInfo &TII) {
+void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
// Initialize the Fast-ISel state, if needed.
FastISel *FastIS = 0;
if (EnableFastISel)
- FastIS = TLI.createFastISel(MF, MMI, DW,
- FuncInfo->ValueMap,
- FuncInfo->MBBMap,
+ FastIS = TLI.createFastISel(*MF, FuncInfo->ValueMap, FuncInfo->MBBMap,
FuncInfo->StaticAllocaMap
#ifndef NDEBUG
, FuncInfo->CatchInfoLost
#endif
);
- unsigned MDDbgKind = Fn.getContext().getMDKindID("dbg");
-
// Iterate over all basic blocks in the function.
- for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
- BasicBlock *LLVMBB = &*I;
+ for (Function::const_iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
+ const BasicBlock *LLVMBB = &*I;
BB = FuncInfo->MBBMap[LLVMBB];
- BasicBlock::iterator const Begin = LLVMBB->begin();
- BasicBlock::iterator const End = LLVMBB->end();
- BasicBlock::iterator BI = Begin;
+ BasicBlock::const_iterator const Begin = LLVMBB->begin();
+ BasicBlock::const_iterator const End = LLVMBB->end();
+ BasicBlock::const_iterator BI = Begin;
// Lower any arguments needed in this block if this is the entry block.
bool SuppressFastISel = false;
// fast-isel in the entry block.
if (FastIS) {
unsigned j = 1;
- for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end();
+ for (Function::const_arg_iterator I = Fn.arg_begin(), E = Fn.arg_end();
I != E; ++I, ++j)
if (Fn.paramHasAttr(j, Attribute::ByVal)) {
if (EnableFastISelVerbose || EnableFastISelAbort)
}
}
- if (MMI && BB->isLandingPad()) {
- // Add a label to mark the beginning of the landing pad. Deletion of the
- // landing pad can thus be detected via the MachineModuleInfo.
- unsigned LabelID = MMI->addLandingPad(BB);
-
- const TargetInstrDesc &II = TII.get(TargetOpcode::EH_LABEL);
- BuildMI(BB, SDB->getCurDebugLoc(), II).addImm(LabelID);
-
- // Mark exception register as live in.
- unsigned Reg = TLI.getExceptionAddressRegister();
- if (Reg) BB->addLiveIn(Reg);
-
- // Mark exception selector register as live in.
- Reg = TLI.getExceptionSelectorRegister();
- if (Reg) BB->addLiveIn(Reg);
-
- // FIXME: Hack around an exception handling flaw (PR1508): the personality
- // function and list of typeids logically belong to the invoke (or, if you
- // like, the basic block containing the invoke), and need to be associated
- // with it in the dwarf exception handling tables. Currently however the
- // information is provided by an intrinsic (eh.selector) that can be moved
- // to unexpected places by the optimizers: if the unwind edge is critical,
- // then breaking it can result in the intrinsics being in the successor of
- // the landing pad, not the landing pad itself. This results
- // in exceptions not being caught because no typeids are associated with
- // the invoke. This may not be the only way things can go wrong, but it
- // is the only way we try to work around for the moment.
- BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
-
- if (Br && Br->isUnconditional()) { // Critical edge?
- BasicBlock::iterator I, E;
- for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
- if (isa<EHSelectorInst>(I))
- break;
-
- if (I == E)
- // No catch info found - try to extract some from the successor.
- CopyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, *FuncInfo);
- }
- }
-
+ // Setup an EH landing-pad block.
+ if (BB->isLandingPad())
+ PrepareEHLandingPad(BB);
+
// Before doing SelectionDAG ISel, see if FastISel has been requested.
if (FastIS && !SuppressFastISel) {
// Emit code for any incoming arguments. This must happen before
// feed PHI nodes in successor blocks.
if (isa<TerminatorInst>(BI))
if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, FastIS)) {
+ ++NumFastIselFailures;
ResetDebugLoc(SDB, FastIS);
if (EnableFastISelVerbose || EnableFastISelAbort) {
dbgs() << "FastISel miss: ";
break;
}
- SetDebugLoc(MDDbgKind, BI, SDB, FastIS, &MF);
+ SetDebugLoc(BI, SDB, FastIS, MF);
// Try to select the instruction with FastISel.
if (FastIS->SelectInstruction(BI)) {
// Then handle certain instructions as single-LLVM-Instruction blocks.
if (isa<CallInst>(BI)) {
+ ++NumFastIselFailures;
if (EnableFastISelVerbose || EnableFastISelAbort) {
dbgs() << "FastISel missed call: ";
BI->dump();
}
- if (!BI->getType()->isVoidTy()) {
+ if (!BI->getType()->isVoidTy() && !BI->use_empty()) {
unsigned &R = FuncInfo->ValueMap[BI];
if (!R)
R = FuncInfo->CreateRegForValue(BI);
// Otherwise, give up on FastISel for the rest of the block.
// For now, be a little lenient about non-branch terminators.
if (!isa<TerminatorInst>(BI) || isa<BranchInst>(BI)) {
+ ++NumFastIselFailures;
if (EnableFastISelVerbose || EnableFastISelAbort) {
dbgs() << "FastISel miss: ";
BI->dump();
MachineInstr *PHI = SDB->PHINodesToUpdate[i].first;
assert(PHI->isPHI() &&
"This is not a machine PHI node that we are updating!");
+ if (!BB->isSuccessor(PHI->getParent()))
+ continue;
PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second,
false));
PHI->addOperand(MachineOperand::CreateMBB(BB));
std::vector<SDValue> InOps;
std::swap(InOps, Ops);
- Ops.push_back(InOps[0]); // input chain.
- Ops.push_back(InOps[1]); // input asm string.
+ Ops.push_back(InOps[InlineAsm::Op_InputChain]); // 0
+ Ops.push_back(InOps[InlineAsm::Op_AsmString]); // 1
+ Ops.push_back(InOps[InlineAsm::Op_MDNode]); // 2, !srcloc
- unsigned i = 2, e = InOps.size();
+ unsigned i = InlineAsm::Op_FirstOperand, e = InOps.size();
if (InOps[e-1].getValueType() == MVT::Flag)
--e; // Don't process a flag operand if it is here.
while (i != e) {
unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
- if ((Flags & 7) != 4 /*MEM*/) {
+ if (!InlineAsm::isMemKind(Flags)) {
// Just skip over this operand, copying the operands verbatim.
Ops.insert(Ops.end(), InOps.begin()+i,
InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
"Memory operand with multiple values?");
// Otherwise, this is a memory operand. Ask the target to select it.
std::vector<SDValue> SelOps;
- if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps)) {
- llvm_report_error("Could not match memory address. Inline asm"
- " failure!");
- }
+ if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps))
+ report_fatal_error("Could not match memory address. Inline asm"
+ " failure!");
// Add this to the output node.
- Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size()<< 3),
- MVT::i32));
+ unsigned NewFlags =
+ InlineAsm::getFlagWord(InlineAsm::Kind_Mem, SelOps.size());
+ Ops.push_back(CurDAG->getTargetConstant(NewFlags, MVT::i32));
Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
i += 2;
}
/// This function recursively traverses up the operand chain, ignoring
/// certain nodes.
static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
- SDNode *Root,
- SmallPtrSet<SDNode*, 16> &Visited) {
+ SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited,
+ bool IgnoreChains) {
// The NodeID's are given uniques ID's where a node ID is guaranteed to be
// greater than all of its (recursive) operands. If we scan to a point where
// 'use' is smaller than the node we're scanning for, then we know we will
return false;
for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
+ // Ignore chain uses, they are validated by HandleMergeInputChains.
+ if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains)
+ continue;
+
SDNode *N = Use->getOperand(i).getNode();
if (N == Def) {
if (Use == ImmedUse || Use == Root)
}
// Traverse up the operand chain.
- if (findNonImmUse(N, Def, ImmedUse, Root, Visited))
+ if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
return true;
}
return false;
}
-/// isNonImmUse - Start searching from Root up the DAG to check is Def can
-/// be reached. Return true if that's the case. However, ignore direct uses
-/// by ImmedUse (which would be U in the example illustrated in
-/// IsLegalToFold) and by Root (which can happen in the store case).
-/// FIXME: to be really generic, we should allow direct use by any node
-/// that is being folded. But realisticly since we only fold loads which
-/// have one non-chain use, we only need to watch out for load/op/store
-/// and load/op/cmp case where the root (store / cmp) may reach the load via
-/// its chain operand.
-static inline bool isNonImmUse(SDNode *Root, SDNode *Def, SDNode *ImmedUse) {
- SmallPtrSet<SDNode*, 16> Visited;
- return findNonImmUse(Root, Def, ImmedUse, Root, Visited);
-}
-
/// IsProfitableToFold - Returns true if it's profitable to fold the specific
/// operand node N of U during instruction selection that starts at Root.
bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
/// IsLegalToFold - Returns true if the specific operand node N of
/// U can be folded during instruction selection that starts at Root.
-bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root) const {
+bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
+ bool IgnoreChains) const {
if (OptLevel == CodeGenOpt::None) return false;
// If Root use can somehow reach N through a path that that doesn't contain
// Fold. But since Fold and FU are flagged together, this will create
// a cycle in the scheduling graph.
+ // If the node has flags, walk down the graph to the "lowest" node in the
+ // flagged set.
EVT VT = Root->getValueType(Root->getNumValues()-1);
while (VT == MVT::Flag) {
SDNode *FU = findFlagUse(Root);
break;
Root = FU;
VT = Root->getValueType(Root->getNumValues()-1);
+
+ // If our query node has a flag result with a use, we've walked up it. If
+ // the user (which has already been selected) has a chain or indirectly uses
+ // the chain, our WalkChainUsers predicate will not consider it. Because of
+ // this, we cannot ignore chains in this predicate.
+ IgnoreChains = false;
}
+
- return !isNonImmUse(Root, N.getNode(), U);
+ SmallPtrSet<SDNode*, 16> Visited;
+ return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains);
}
SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
VTs.push_back(MVT::Flag);
SDValue New = CurDAG->getNode(ISD::INLINEASM, N->getDebugLoc(),
VTs, &Ops[0], Ops.size());
+ New->setNodeId(-1);
return New.getNode();
}
return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0));
}
-SDNode *SelectionDAGISel::Select_EH_LABEL(SDNode *N) {
- SDValue Chain = N->getOperand(0);
- unsigned C = cast<LabelSDNode>(N)->getLabelID();
- SDValue Tmp = CurDAG->getTargetConstant(C, MVT::i32);
- return CurDAG->SelectNodeTo(N, TargetOpcode::EH_LABEL,
- MVT::Other, Tmp, Chain);
-}
-
/// GetVBR - decode a vbr encoding whose top bit is set.
ALWAYS_INLINE static uint64_t
GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) {
/// UpdateChainsAndFlags - When a match is complete, this method updates uses of
/// interior flag and chain results to use the new flag and chain results.
-static void UpdateChainsAndFlags(SDNode *NodeToMatch, SDValue InputChain,
- const SmallVectorImpl<SDNode*> &ChainNodesMatched,
- SDValue InputFlag,
- const SmallVectorImpl<SDNode*> &FlagResultNodesMatched,
- bool isMorphNodeTo, SelectionDAG *CurDAG) {
+void SelectionDAGISel::
+UpdateChainsAndFlags(SDNode *NodeToMatch, SDValue InputChain,
+ const SmallVectorImpl<SDNode*> &ChainNodesMatched,
+ SDValue InputFlag,
+ const SmallVectorImpl<SDNode*> &FlagResultNodesMatched,
+ bool isMorphNodeTo) {
+ SmallVector<SDNode*, 4> NowDeadNodes;
+
+ ISelUpdater ISU(ISelPosition);
+
// Now that all the normal results are replaced, we replace the chain and
// flag results if present.
if (!ChainNodesMatched.empty()) {
for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) {
SDNode *ChainNode = ChainNodesMatched[i];
+ // If this node was already deleted, don't look at it.
+ if (ChainNode->getOpcode() == ISD::DELETED_NODE)
+ continue;
+
// Don't replace the results of the root node if we're doing a
// MorphNodeTo.
if (ChainNode == NodeToMatch && isMorphNodeTo)
if (ChainVal.getValueType() == MVT::Flag)
ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2);
assert(ChainVal.getValueType() == MVT::Other && "Not a chain?");
- CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain);
+ CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain, &ISU);
+
+ // If the node became dead and we haven't already seen it, delete it.
+ if (ChainNode->use_empty() &&
+ !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), ChainNode))
+ NowDeadNodes.push_back(ChainNode);
}
}
// Handle any interior nodes explicitly marked.
for (unsigned i = 0, e = FlagResultNodesMatched.size(); i != e; ++i) {
SDNode *FRN = FlagResultNodesMatched[i];
+
+ // If this node was already deleted, don't look at it.
+ if (FRN->getOpcode() == ISD::DELETED_NODE)
+ continue;
+
assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Flag &&
"Doesn't have a flag result");
CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1),
- InputFlag);
+ InputFlag, &ISU);
+
+ // If the node became dead and we haven't already seen it, delete it.
+ if (FRN->use_empty() &&
+ !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), FRN))
+ NowDeadNodes.push_back(FRN);
}
}
+ if (!NowDeadNodes.empty())
+ CurDAG->RemoveDeadNodes(NowDeadNodes, &ISU);
+
DEBUG(errs() << "ISEL: Match complete!\n");
}
// pattern that we're selecting down into the already selected chunk of the
// DAG.
if (User->isMachineOpcode() ||
- User->getOpcode() == ISD::CopyToReg ||
- User->getOpcode() == ISD::CopyFromReg ||
- User->getOpcode() == ISD::INLINEASM ||
User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
continue;
+
+ if (User->getOpcode() == ISD::CopyToReg ||
+ User->getOpcode() == ISD::CopyFromReg ||
+ User->getOpcode() == ISD::INLINEASM ||
+ User->getOpcode() == ISD::EH_LABEL) {
+ // If their node ID got reset to -1 then they've already been selected.
+ // Treat them like a MachineOpcode.
+ if (User->getNodeId() == -1)
+ continue;
+ }
// If we have a TokenFactor, we handle it specially.
if (User->getOpcode() != ISD::TokenFactor) {
}
/// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains
-/// operation for when the pattern matched multiple nodes with chains. The
+/// operation for when the pattern matched at least one node with a chains. The
/// input vector contains a list of all of the chained nodes that we match. We
/// must determine if this is a valid thing to cover (i.e. matching it won't
/// induce cycles in the DAG) and if so, creating a TokenFactor node. that will
static SDValue
HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched,
SelectionDAG *CurDAG) {
- assert(ChainNodesMatched.size() > 1 &&
- "Should only happen for multi chain node case");
-
// Walk all of the chained nodes we've matched, recursively scanning down the
// users of the chain result. This adds any TokenFactor nodes that are caught
// in between chained nodes to the chained and interior nodes list.
// It is possible we're using MorphNodeTo to replace a node with no
// normal results with one that has a normal result (or we could be
// adding a chain) and the input could have flags and chains as well.
- // In this case we need to shifting the operands down.
+ // In this case we need to shift the operands down.
// FIXME: This is a horrible hack and broken in obscure cases, no worse
- // than the old isel though. We should sink this into MorphNodeTo.
+ // than the old isel though.
int OldFlagResultNo = -1, OldChainResultNo = -1;
unsigned NTMNumResults = Node->getNumValues();
return Res;
}
+/// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
+ALWAYS_INLINE static bool
+CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex,
+ SDValue N, const SmallVectorImpl<SDValue> &RecordedNodes) {
+ // Accept if it is exactly the same as a previously recorded node.
+ unsigned RecNo = MatcherTable[MatcherIndex++];
+ assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
+ return N == RecordedNodes[RecNo];
+}
+
+/// CheckPatternPredicate - Implements OP_CheckPatternPredicate.
+ALWAYS_INLINE static bool
+CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
+ SelectionDAGISel &SDISel) {
+ return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]);
+}
+
+/// CheckNodePredicate - Implements OP_CheckNodePredicate.
+ALWAYS_INLINE static bool
+CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex,
+ SelectionDAGISel &SDISel, SDNode *N) {
+ return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]);
+}
+
+ALWAYS_INLINE static bool
+CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
+ SDNode *N) {
+ uint16_t Opc = MatcherTable[MatcherIndex++];
+ Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
+ return N->getOpcode() == Opc;
+}
+
+ALWAYS_INLINE static bool
+CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
+ SDValue N, const TargetLowering &TLI) {
+ MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
+ if (N.getValueType() == VT) return true;
+
+ // Handle the case when VT is iPTR.
+ return VT == MVT::iPTR && N.getValueType() == TLI.getPointerTy();
+}
+
+ALWAYS_INLINE static bool
+CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
+ SDValue N, const TargetLowering &TLI,
+ unsigned ChildNo) {
+ if (ChildNo >= N.getNumOperands())
+ return false; // Match fails if out of range child #.
+ return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI);
+}
+
+
+ALWAYS_INLINE static bool
+CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex,
+ SDValue N) {
+ return cast<CondCodeSDNode>(N)->get() ==
+ (ISD::CondCode)MatcherTable[MatcherIndex++];
+}
+
+ALWAYS_INLINE static bool
+CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex,
+ SDValue N, const TargetLowering &TLI) {
+ MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
+ if (cast<VTSDNode>(N)->getVT() == VT)
+ return true;
+
+ // Handle the case when VT is iPTR.
+ return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI.getPointerTy();
+}
+
+ALWAYS_INLINE static bool
+CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex,
+ SDValue N) {
+ int64_t Val = MatcherTable[MatcherIndex++];
+ if (Val & 128)
+ Val = GetVBR(Val, MatcherTable, MatcherIndex);
+
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
+ return C != 0 && C->getSExtValue() == Val;
+}
+
+ALWAYS_INLINE static bool
+CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
+ SDValue N, SelectionDAGISel &SDISel) {
+ int64_t Val = MatcherTable[MatcherIndex++];
+ if (Val & 128)
+ Val = GetVBR(Val, MatcherTable, MatcherIndex);
+
+ if (N->getOpcode() != ISD::AND) return false;
+
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
+ return C != 0 && SDISel.CheckAndMask(N.getOperand(0), C, Val);
+}
+
+ALWAYS_INLINE static bool
+CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex,
+ SDValue N, SelectionDAGISel &SDISel) {
+ int64_t Val = MatcherTable[MatcherIndex++];
+ if (Val & 128)
+ Val = GetVBR(Val, MatcherTable, MatcherIndex);
+
+ if (N->getOpcode() != ISD::OR) return false;
+
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
+ return C != 0 && SDISel.CheckOrMask(N.getOperand(0), C, Val);
+}
+
+/// IsPredicateKnownToFail - If we know how and can do so without pushing a
+/// scope, evaluate the current node. If the current predicate is known to
+/// fail, set Result=true and return anything. If the current predicate is
+/// known to pass, set Result=false and return the MatcherIndex to continue
+/// with. If the current predicate is unknown, set Result=false and return the
+/// MatcherIndex to continue with.
+static unsigned IsPredicateKnownToFail(const unsigned char *Table,
+ unsigned Index, SDValue N,
+ bool &Result, SelectionDAGISel &SDISel,
+ SmallVectorImpl<SDValue> &RecordedNodes){
+ switch (Table[Index++]) {
+ default:
+ Result = false;
+ return Index-1; // Could not evaluate this predicate.
+ case SelectionDAGISel::OPC_CheckSame:
+ Result = !::CheckSame(Table, Index, N, RecordedNodes);
+ return Index;
+ case SelectionDAGISel::OPC_CheckPatternPredicate:
+ Result = !::CheckPatternPredicate(Table, Index, SDISel);
+ return Index;
+ case SelectionDAGISel::OPC_CheckPredicate:
+ Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode());
+ return Index;
+ case SelectionDAGISel::OPC_CheckOpcode:
+ Result = !::CheckOpcode(Table, Index, N.getNode());
+ return Index;
+ case SelectionDAGISel::OPC_CheckType:
+ Result = !::CheckType(Table, Index, N, SDISel.TLI);
+ return Index;
+ case SelectionDAGISel::OPC_CheckChild0Type:
+ case SelectionDAGISel::OPC_CheckChild1Type:
+ case SelectionDAGISel::OPC_CheckChild2Type:
+ case SelectionDAGISel::OPC_CheckChild3Type:
+ case SelectionDAGISel::OPC_CheckChild4Type:
+ case SelectionDAGISel::OPC_CheckChild5Type:
+ case SelectionDAGISel::OPC_CheckChild6Type:
+ case SelectionDAGISel::OPC_CheckChild7Type:
+ Result = !::CheckChildType(Table, Index, N, SDISel.TLI,
+ Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Type);
+ return Index;
+ case SelectionDAGISel::OPC_CheckCondCode:
+ Result = !::CheckCondCode(Table, Index, N);
+ return Index;
+ case SelectionDAGISel::OPC_CheckValueType:
+ Result = !::CheckValueType(Table, Index, N, SDISel.TLI);
+ return Index;
+ case SelectionDAGISel::OPC_CheckInteger:
+ Result = !::CheckInteger(Table, Index, N);
+ return Index;
+ case SelectionDAGISel::OPC_CheckAndImm:
+ Result = !::CheckAndImm(Table, Index, N, SDISel);
+ return Index;
+ case SelectionDAGISel::OPC_CheckOrImm:
+ Result = !::CheckOrImm(Table, Index, N, SDISel);
+ return Index;
+ }
+}
+
+namespace {
struct MatchScope {
/// FailIndex - If this match fails, this is the index to continue with.
bool HasChainNodesMatched, HasFlagResultNodesMatched;
};
+}
+
SDNode *SelectionDAGISel::
SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
unsigned TableSize) {
case ISD::EntryToken: // These nodes remain the same.
case ISD::BasicBlock:
case ISD::Register:
+ //case ISD::VALUETYPE:
+ //case ISD::CONDCODE:
case ISD::HANDLENODE:
+ case ISD::MDNODE_SDNODE:
case ISD::TargetConstant:
case ISD::TargetConstantFP:
case ISD::TargetConstantPool:
case ISD::TokenFactor:
case ISD::CopyFromReg:
case ISD::CopyToReg:
+ case ISD::EH_LABEL:
+ NodeToMatch->setNodeId(-1); // Mark selected.
return 0;
case ISD::AssertSext:
case ISD::AssertZext:
NodeToMatch->getOperand(0));
return 0;
case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch);
- case ISD::EH_LABEL: return Select_EH_LABEL(NodeToMatch);
case ISD::UNDEF: return Select_UNDEF(NodeToMatch);
}
if (CaseSize == 0) break;
// Get the opcode, add the index to the table.
- unsigned Opc = MatcherTable[Idx++];
+ uint16_t Opc = MatcherTable[Idx++];
+ Opc |= (unsigned short)MatcherTable[Idx++] << 8;
if (Opc >= OpcodeOffset.size())
OpcodeOffset.resize((Opc+1)*2);
OpcodeOffset[Opc] = Idx;
while (1) {
assert(MatcherIndex < TableSize && "Invalid index");
+#ifndef NDEBUG
+ unsigned CurrentOpcodeIndex = MatcherIndex;
+#endif
BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++];
switch (Opcode) {
case OPC_Scope: {
- unsigned NumToSkip = MatcherTable[MatcherIndex++];
- if (NumToSkip & 128)
- NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
- assert(NumToSkip != 0 &&
- "First entry of OPC_Scope shouldn't be 0, scope has no children?");
-
+ // Okay, the semantics of this operation are that we should push a scope
+ // then evaluate the first child. However, pushing a scope only to have
+ // the first check fail (which then pops it) is inefficient. If we can
+ // determine immediately that the first check (or first several) will
+ // immediately fail, don't even bother pushing a scope for them.
+ unsigned FailIndex;
+
+ while (1) {
+ unsigned NumToSkip = MatcherTable[MatcherIndex++];
+ if (NumToSkip & 128)
+ NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex);
+ // Found the end of the scope with no match.
+ if (NumToSkip == 0) {
+ FailIndex = 0;
+ break;
+ }
+
+ FailIndex = MatcherIndex+NumToSkip;
+
+ unsigned MatcherIndexOfPredicate = MatcherIndex;
+ (void)MatcherIndexOfPredicate; // silence warning.
+
+ // If we can't evaluate this predicate without pushing a scope (e.g. if
+ // it is a 'MoveParent') or if the predicate succeeds on this node, we
+ // push the scope and evaluate the full predicate chain.
+ bool Result;
+ MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N,
+ Result, *this, RecordedNodes);
+ if (!Result)
+ break;
+
+ DEBUG(errs() << " Skipped scope entry (due to false predicate) at "
+ << "index " << MatcherIndexOfPredicate
+ << ", continuing at " << FailIndex << "\n");
+ ++NumDAGIselRetries;
+
+ // Otherwise, we know that this case of the Scope is guaranteed to fail,
+ // move to the next case.
+ MatcherIndex = FailIndex;
+ }
+
+ // If the whole scope failed to match, bail.
+ if (FailIndex == 0) break;
+
// Push a MatchScope which indicates where to go if the first child fails
// to match.
MatchScope NewEntry;
- NewEntry.FailIndex = MatcherIndex+NumToSkip;
+ NewEntry.FailIndex = FailIndex;
NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end());
NewEntry.NumRecordedNodes = RecordedNodes.size();
NewEntry.NumMatchedMemRefs = MatchedMemRefs.size();
N = NodeStack.back();
continue;
- case OPC_CheckSame: {
- // Accept if it is exactly the same as a previously recorded node.
- unsigned RecNo = MatcherTable[MatcherIndex++];
- assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
- if (N != RecordedNodes[RecNo]) break;
+ case OPC_CheckSame:
+ if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break;
continue;
- }
case OPC_CheckPatternPredicate:
- if (!CheckPatternPredicate(MatcherTable[MatcherIndex++])) break;
+ if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break;
continue;
case OPC_CheckPredicate:
- if (!CheckNodePredicate(N.getNode(), MatcherTable[MatcherIndex++])) break;
+ if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this,
+ N.getNode()))
+ break;
continue;
- case OPC_CheckComplexPat:
- if (!CheckComplexPattern(NodeToMatch, N,
- MatcherTable[MatcherIndex++], RecordedNodes))
+ case OPC_CheckComplexPat: {
+ unsigned CPNum = MatcherTable[MatcherIndex++];
+ unsigned RecNo = MatcherTable[MatcherIndex++];
+ assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
+ if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo], CPNum,
+ RecordedNodes))
break;
continue;
+ }
case OPC_CheckOpcode:
- if (N->getOpcode() != MatcherTable[MatcherIndex++]) break;
+ if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break;
+ continue;
+
+ case OPC_CheckType:
+ if (!::CheckType(MatcherTable, MatcherIndex, N, TLI)) break;
continue;
case OPC_SwitchOpcode: {
unsigned CurNodeOpcode = N.getOpcode();
-
unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
-
unsigned CaseSize;
while (1) {
// Get the size of this case.
CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
if (CaseSize == 0) break;
+ uint16_t Opc = MatcherTable[MatcherIndex++];
+ Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
+
// If the opcode matches, then we will execute this case.
- if (CurNodeOpcode == MatcherTable[MatcherIndex++])
+ if (CurNodeOpcode == Opc)
break;
// Otherwise, skip over this case.
MatcherIndex += CaseSize;
}
- // If we failed to match, bail out.
+ // If no cases matched, bail out.
if (CaseSize == 0) break;
// Otherwise, execute the case we found.
continue;
}
- case OPC_CheckType: {
- MVT::SimpleValueType VT =
- (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
- if (N.getValueType() != VT) {
- // Handle the case when VT is iPTR.
- if (VT != MVT::iPTR || N.getValueType() != TLI.getPointerTy())
+ case OPC_SwitchType: {
+ MVT::SimpleValueType CurNodeVT = N.getValueType().getSimpleVT().SimpleTy;
+ unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart;
+ unsigned CaseSize;
+ while (1) {
+ // Get the size of this case.
+ CaseSize = MatcherTable[MatcherIndex++];
+ if (CaseSize & 128)
+ CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex);
+ if (CaseSize == 0) break;
+
+ MVT::SimpleValueType CaseVT =
+ (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
+ if (CaseVT == MVT::iPTR)
+ CaseVT = TLI.getPointerTy().SimpleTy;
+
+ // If the VT matches, then we will execute this case.
+ if (CurNodeVT == CaseVT)
break;
+
+ // Otherwise, skip over this case.
+ MatcherIndex += CaseSize;
}
+
+ // If no cases matched, bail out.
+ if (CaseSize == 0) break;
+
+ // Otherwise, execute the case we found.
+ DEBUG(errs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
+ << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
continue;
}
case OPC_CheckChild0Type: case OPC_CheckChild1Type:
case OPC_CheckChild2Type: case OPC_CheckChild3Type:
case OPC_CheckChild4Type: case OPC_CheckChild5Type:
- case OPC_CheckChild6Type: case OPC_CheckChild7Type: {
- unsigned ChildNo = Opcode-OPC_CheckChild0Type;
- if (ChildNo >= N.getNumOperands())
- break; // Match fails if out of range child #.
-
- MVT::SimpleValueType VT =
- (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
- EVT ChildVT = N.getOperand(ChildNo).getValueType();
- if (ChildVT != VT) {
- // Handle the case when VT is iPTR.
- if (VT != MVT::iPTR || ChildVT != TLI.getPointerTy())
- break;
- }
+ case OPC_CheckChild6Type: case OPC_CheckChild7Type:
+ if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI,
+ Opcode-OPC_CheckChild0Type))
+ break;
continue;
- }
case OPC_CheckCondCode:
- if (cast<CondCodeSDNode>(N)->get() !=
- (ISD::CondCode)MatcherTable[MatcherIndex++]) break;
+ if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break;
continue;
- case OPC_CheckValueType: {
- MVT::SimpleValueType VT =
- (MVT::SimpleValueType)MatcherTable[MatcherIndex++];
- if (cast<VTSDNode>(N)->getVT() != VT) {
- // Handle the case when VT is iPTR.
- if (VT != MVT::iPTR || cast<VTSDNode>(N)->getVT() != TLI.getPointerTy())
- break;
- }
+ case OPC_CheckValueType:
+ if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI)) break;
continue;
- }
- case OPC_CheckInteger: {
- int64_t Val = MatcherTable[MatcherIndex++];
- if (Val & 128)
- Val = GetVBR(Val, MatcherTable, MatcherIndex);
-
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
- if (C == 0 || C->getSExtValue() != Val)
- break;
+ case OPC_CheckInteger:
+ if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break;
continue;
- }
- case OPC_CheckAndImm: {
- int64_t Val = MatcherTable[MatcherIndex++];
- if (Val & 128)
- Val = GetVBR(Val, MatcherTable, MatcherIndex);
-
- if (N->getOpcode() != ISD::AND) break;
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
- if (C == 0 || !CheckAndMask(N.getOperand(0), C, Val))
- break;
+ case OPC_CheckAndImm:
+ if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break;
continue;
- }
- case OPC_CheckOrImm: {
- int64_t Val = MatcherTable[MatcherIndex++];
- if (Val & 128)
- Val = GetVBR(Val, MatcherTable, MatcherIndex);
-
- if (N->getOpcode() != ISD::OR) break;
-
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
- if (C == 0 || !CheckOrMask(N.getOperand(0), C, Val))
- break;
+ case OPC_CheckOrImm:
+ if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break;
continue;
- }
case OPC_CheckFoldableChainNode: {
assert(NodeStack.size() != 1 && "No parent node");
if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
NodeToMatch) ||
!IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
- NodeToMatch))
+ NodeToMatch, true/*We validate our own chains*/))
break;
continue;
continue;
}
+ case OPC_EmitMergeInputChains1_0: // OPC_EmitMergeInputChains, 1, 0
+ case OPC_EmitMergeInputChains1_1: { // OPC_EmitMergeInputChains, 1, 1
+ // These are space-optimized forms of OPC_EmitMergeInputChains.
+ assert(InputChain.getNode() == 0 &&
+ "EmitMergeInputChains should be the first chain producing node");
+ assert(ChainNodesMatched.empty() &&
+ "Should only have one EmitMergeInputChains per match");
+
+ // Read all of the chained nodes.
+ unsigned RecNo = Opcode == OPC_EmitMergeInputChains1_1;
+ assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
+ ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode());
+
+ // FIXME: What if other value results of the node have uses not matched
+ // by this pattern?
+ if (ChainNodesMatched.back() != NodeToMatch &&
+ !RecordedNodes[RecNo].hasOneUse()) {
+ ChainNodesMatched.clear();
+ break;
+ }
+
+ // Merge the input chains if they are not intra-pattern references.
+ InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG);
+
+ if (InputChain.getNode() == 0)
+ break; // Failed to merge.
+ continue;
+ }
+
case OPC_EmitMergeInputChains: {
assert(InputChain.getNode() == 0 &&
"EmitMergeInputChains should be the first chain producing node");
assert(ChainNodesMatched.empty() &&
"Should only have one EmitMergeInputChains per match");
- // Handle the first chain.
- unsigned RecNo = MatcherTable[MatcherIndex++];
- assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
- ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode());
-
- // If the chained node is not the root, we can't fold it if it has
- // multiple uses.
- // FIXME: What if other value results of the node have uses not matched by
- // this pattern?
- if (ChainNodesMatched.back() != NodeToMatch &&
- !RecordedNodes[RecNo].hasOneUse()) {
- ChainNodesMatched.clear();
- break;
- }
-
- // The common case here is that we have exactly one chain, which is really
- // cheap to handle, just do it.
- if (NumChains == 1) {
- InputChain = RecordedNodes[RecNo].getOperand(0);
- assert(InputChain.getValueType() == MVT::Other && "Not a chain");
- continue;
- }
-
// Read all of the chained nodes.
- for (unsigned i = 1; i != NumChains; ++i) {
- RecNo = MatcherTable[MatcherIndex++];
+ for (unsigned i = 0; i != NumChains; ++i) {
+ unsigned RecNo = MatcherTable[MatcherIndex++];
assert(RecNo < RecordedNodes.size() && "Invalid CheckSame");
ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode());
// If the node had chain/flag results, update our notion of the current
// chain and flag.
- if (VTs.back() == MVT::Flag) {
+ if (EmitNodeInfo & OPFL_FlagOutput) {
InputFlag = SDValue(Res, VTs.size()-1);
if (EmitNodeInfo & OPFL_Chain)
InputChain = SDValue(Res, VTs.size()-2);
if (Opcode == OPC_MorphNodeTo) {
// Update chain and flag uses.
UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched,
- InputFlag, FlagResultNodesMatched, true, CurDAG);
- return 0;
+ InputFlag, FlagResultNodesMatched, true);
+ return Res;
}
continue;
assert(ResSlot < RecordedNodes.size() && "Invalid CheckSame");
SDValue Res = RecordedNodes[ResSlot];
- // FIXME2: Eliminate this horrible hack by fixing the 'Gen' program
- // after (parallel) on input patterns are removed. This would also
- // allow us to stop encoding #results in OPC_CompleteMatch's table
- // entry.
- if (NodeToMatch->getNumValues() <= i ||
- NodeToMatch->getValueType(i) == MVT::Other ||
- NodeToMatch->getValueType(i) == MVT::Flag)
- break;
+ assert(i < NodeToMatch->getNumValues() &&
+ NodeToMatch->getValueType(i) != MVT::Other &&
+ NodeToMatch->getValueType(i) != MVT::Flag &&
+ "Invalid number of results to complete!");
assert((NodeToMatch->getValueType(i) == Res.getValueType() ||
NodeToMatch->getValueType(i) == MVT::iPTR ||
Res.getValueType() == MVT::iPTR ||
// Update chain and flag uses.
UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched,
- InputFlag, FlagResultNodesMatched, false, CurDAG);
+ InputFlag, FlagResultNodesMatched, false);
assert(NodeToMatch->use_empty() &&
"Didn't replace all uses of the node?");
// If the code reached this point, then the match failed. See if there is
// another child to try in the current 'Scope', otherwise pop it until we
// find a case to check.
+ DEBUG(errs() << " Match failed at index " << CurrentOpcodeIndex << "\n");
+ ++NumDAGIselRetries;
while (1) {
if (MatchScopes.empty()) {
CannotYetSelect(NodeToMatch);
NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end());
N = NodeStack.back();
- DEBUG(errs() << " Match failed at index " << MatcherIndex
- << " continuing at " << LastScope.FailIndex << "\n");
-
if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size())
MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
MatcherIndex = LastScope.FailIndex;
+ DEBUG(errs() << " Continuing at " << MatcherIndex << "\n");
+
InputChain = LastScope.InputChain;
InputFlag = LastScope.InputFlag;
if (!LastScope.HasChainNodesMatched)
void SelectionDAGISel::CannotYetSelect(SDNode *N) {
- if (N->getOpcode() == ISD::INTRINSIC_W_CHAIN ||
- N->getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
- N->getOpcode() == ISD::INTRINSIC_VOID)
- return CannotYetSelectIntrinsic(N);
-
std::string msg;
raw_string_ostream Msg(msg);
Msg << "Cannot yet select: ";
- N->printrFull(Msg, CurDAG);
- llvm_report_error(Msg.str());
-}
-
-void SelectionDAGISel::CannotYetSelectIntrinsic(SDNode *N) {
- dbgs() << "Cannot yet select: ";
- unsigned iid =
- cast<ConstantSDNode>(N->getOperand(N->getOperand(0).getValueType() ==
- MVT::Other))->getZExtValue();
- if (iid < Intrinsic::num_intrinsics)
- llvm_report_error("Cannot yet select: intrinsic %" +
- Intrinsic::getName((Intrinsic::ID)iid));
- else if (const TargetIntrinsicInfo *tii = TM.getIntrinsicInfo())
- llvm_report_error(Twine("Cannot yet select: target intrinsic %") +
- tii->getName(iid));
+
+ if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN &&
+ N->getOpcode() != ISD::INTRINSIC_WO_CHAIN &&
+ N->getOpcode() != ISD::INTRINSIC_VOID) {
+ N->printrFull(Msg, CurDAG);
+ } else {
+ bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other;
+ unsigned iid =
+ cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue();
+ if (iid < Intrinsic::num_intrinsics)
+ Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid);
+ else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo())
+ Msg << "target intrinsic %" << TII->getName(iid);
+ else
+ Msg << "unknown intrinsic #" << iid;
+ }
+ report_fatal_error(Msg.str());
}
char SelectionDAGISel::ID = 0;