#include "llvm/Instructions.h"
#include "llvm/Intrinsics.h"
#include "llvm/IntrinsicInst.h"
-#include "llvm/ParameterAttributes.h"
#include "llvm/CodeGen/FastISel.h"
#include "llvm/CodeGen/GCStrategy.h"
#include "llvm/CodeGen/GCMetadata.h"
static cl::opt<bool>
EnableValueProp("enable-value-prop", cl::Hidden);
static cl::opt<bool>
-EnableLegalizeTypes("enable-legalize-types", cl::Hidden);
+DisableLegalizeTypes("disable-legalize-types", cl::Hidden);
+#ifndef NDEBUG
static cl::opt<bool>
-EnableFastISel("fast-isel", cl::Hidden,
- cl::desc("Enable the experimental \"fast\" instruction selector"));
+EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
+ cl::desc("Enable verbose messages in the \"fast\" "
+ "instruction selector"));
static cl::opt<bool>
-DisableFastISelAbort("fast-isel-no-abort", cl::Hidden,
- cl::desc("Use the SelectionDAGISel when \"fast\" instruction "
- "selection fails"));
+EnableFastISelAbort("fast-isel-abort", cl::Hidden,
+ cl::desc("Enable abort calls when \"fast\" instruction fails"));
+#else
+static const bool EnableFastISelVerbose = false,
+ EnableFastISelAbort = false;
+#endif
static cl::opt<bool>
SchedLiveInCopies("schedule-livein-copies",
cl::desc("Schedule copies of livein registers"),
" allocation):"));
static RegisterScheduler
-defaultListDAGScheduler("default", " Best scheduler for the target",
+defaultListDAGScheduler("default", "Best scheduler for the target",
createDefaultScheduler);
namespace llvm {
}
bool SelectionDAGISel::runOnFunction(Function &Fn) {
+ // 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>();
DOUT << "\n\n\n=== " << Fn.getName() << "\n";
FuncInfo->set(Fn, MF, EnableFastISel);
- CurDAG->init(MF, getAnalysisToUpdate<MachineModuleInfo>());
+ MachineModuleInfo *MMI = getAnalysisToUpdate<MachineModuleInfo>();
+ CurDAG->init(MF, MMI);
SDL->init(GFI, *AA);
for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
// Mark landing pad.
FuncInfo->MBBMap[Invoke->getSuccessor(1)]->setIsLandingPad();
- SelectAllBasicBlocks(Fn, MF);
+ SelectAllBasicBlocks(Fn, MF, MMI, TII);
// 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
for (SelectionDAG::allnodes_iterator BE = DAG.allnodes_begin(),
BI = DAG.allnodes_end(); BI != BE; ) {
--BI;
- if (BI->getOpcode() == ISD::CALL) {
+ if (CallSDNode *TheCall = dyn_cast<CallSDNode>(BI)) {
SDValue OpRet(Ret, 0);
SDValue OpCall(BI, 0);
- bool isMarkedTailCall =
- cast<ConstantSDNode>(OpCall.getOperand(3))->getValue() != 0;
+ bool isMarkedTailCall = TheCall->isTailCall();
// If CALL node has tail call attribute set to true and the call is not
// eligible (no RET or the target rejects) the attribute is fixed to
// false. The TargetLowering::IsEligibleForTailCallOptimization function
// must correctly identify tail call optimizable calls.
if (!isMarkedTailCall) continue;
if (Ret==NULL ||
- !TLI.IsEligibleForTailCallOptimization(OpCall, OpRet, DAG)) {
- // Not eligible. Mark CALL node as non tail call.
- SmallVector<SDValue, 32> Ops;
- unsigned idx=0;
- for(SDNode::op_iterator I =OpCall.getNode()->op_begin(),
- E = OpCall.getNode()->op_end(); I != E; I++, idx++) {
- if (idx!=3)
- Ops.push_back(*I);
- else
- Ops.push_back(DAG.getConstant(false, TLI.getPointerTy()));
- }
- DAG.UpdateNodeOperands(OpCall, Ops.begin(), Ops.size());
+ !TLI.IsEligibleForTailCallOptimization(TheCall, OpRet, DAG)) {
+ // Not eligible. Mark CALL node as non tail call. Note that we
+ // can modify the call node in place since calls are not CSE'd.
+ TheCall->setNotTailCall();
} else {
// Look for tail call clobbered arguments. Emit a series of
// copyto/copyfrom virtual register nodes to protect them.
SmallVector<SDValue, 32> Ops;
- SDValue Chain = OpCall.getOperand(0), InFlag;
- unsigned idx=0;
- for(SDNode::op_iterator I = OpCall.getNode()->op_begin(),
- E = OpCall.getNode()->op_end(); I != E; I++, idx++) {
- SDValue Arg = *I;
- if (idx > 4 && (idx % 2)) {
- bool isByVal = cast<ARG_FLAGSSDNode>(OpCall.getOperand(idx+1))->
- getArgFlags().isByVal();
- MachineFunction &MF = DAG.getMachineFunction();
- MachineFrameInfo *MFI = MF.getFrameInfo();
- if (!isByVal &&
- IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)) {
- MVT VT = Arg.getValueType();
- unsigned VReg = MF.getRegInfo().
- createVirtualRegister(TLI.getRegClassFor(VT));
- Chain = DAG.getCopyToReg(Chain, VReg, Arg, InFlag);
- InFlag = Chain.getValue(1);
- Arg = DAG.getCopyFromReg(Chain, VReg, VT, InFlag);
- Chain = Arg.getValue(1);
- InFlag = Arg.getValue(2);
- }
+ SDValue Chain = TheCall->getChain(), InFlag;
+ Ops.push_back(Chain);
+ Ops.push_back(TheCall->getCallee());
+ for (unsigned i = 0, e = TheCall->getNumArgs(); i != e; ++i) {
+ SDValue Arg = TheCall->getArg(i);
+ bool isByVal = TheCall->getArgFlags(i).isByVal();
+ MachineFunction &MF = DAG.getMachineFunction();
+ MachineFrameInfo *MFI = MF.getFrameInfo();
+ if (!isByVal &&
+ IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)) {
+ MVT VT = Arg.getValueType();
+ unsigned VReg = MF.getRegInfo().
+ createVirtualRegister(TLI.getRegClassFor(VT));
+ Chain = DAG.getCopyToReg(Chain, VReg, Arg, InFlag);
+ InFlag = Chain.getValue(1);
+ Arg = DAG.getCopyFromReg(Chain, VReg, VT, InFlag);
+ Chain = Arg.getValue(1);
+ InFlag = Arg.getValue(2);
}
Ops.push_back(Arg);
+ Ops.push_back(TheCall->getArgFlagsVal(i));
}
// Link in chain of CopyTo/CopyFromReg.
Ops[0] = Chain;
BasicBlock::iterator End) {
SDL->setCurrentBasicBlock(BB);
- MachineModuleInfo *MMI = CurDAG->getMachineModuleInfo();
-
- 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);
- CurDAG->setRoot(CurDAG->getLabel(ISD::EH_LABEL,
- CurDAG->getEntryNode(), 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);
- }
- }
-
// Lower all of the non-terminator instructions.
for (BasicBlock::iterator I = Begin; I != End; ++I)
if (!isa<TerminatorInst>(I))
// with correct tailcall attribute so that the target can rely on the tailcall
// attribute indicating whether the call is really eligible for tail call
// optimization.
- CheckDAGForTailCallsAndFixThem(*CurDAG, TLI);
+ if (PerformTailCallOpt)
+ CheckDAGForTailCallsAndFixThem(*CurDAG, TLI);
// Final step, emit the lowered DAG as machine code.
CodeGenAndEmitDAG();
// Second step, hack on the DAG until it only uses operations and types that
// the target supports.
- if (EnableLegalizeTypes) {// Enable this some day.
+ if (!DisableLegalizeTypes) {
if (ViewLegalizeTypesDAGs) CurDAG->viewGraph("legalize-types input for " +
BlockName);
DEBUG(BB->dump());
}
-void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn, MachineFunction &MF) {
+void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn, MachineFunction &MF,
+ MachineModuleInfo *MMI,
+ const TargetInstrInfo &TII) {
+ // Initialize the Fast-ISel state, if needed.
+ FastISel *FastIS = 0;
+ if (EnableFastISel)
+ FastIS = TLI.createFastISel(*FuncInfo->MF, MMI,
+ FuncInfo->ValueMap,
+ FuncInfo->MBBMap,
+ FuncInfo->StaticAllocaMap
+#ifndef NDEBUG
+ , FuncInfo->CatchInfoLost
+#endif
+ );
+
+ // Iterate over all basic blocks in the function.
for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
BasicBlock *LLVMBB = &*I;
BB = FuncInfo->MBBMap[LLVMBB];
BasicBlock::iterator const Begin = LLVMBB->begin();
BasicBlock::iterator const End = LLVMBB->end();
- BasicBlock::iterator I = Begin;
+ BasicBlock::iterator BI = Begin;
// Lower any arguments needed in this block if this is the entry block.
- if (LLVMBB == &Fn.getEntryBlock())
+ bool SuppressFastISel = false;
+ if (LLVMBB == &Fn.getEntryBlock()) {
LowerArguments(LLVMBB);
+ // If any of the arguments has the byval attribute, forgo
+ // fast-isel in the entry block.
+ if (FastIS) {
+ unsigned j = 1;
+ for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end();
+ I != E; ++I, ++j)
+ if (Fn.paramHasAttr(j, Attribute::ByVal)) {
+ if (EnableFastISelVerbose || EnableFastISelAbort)
+ cerr << "FastISel skips entry block due to byval argument\n";
+ SuppressFastISel = true;
+ break;
+ }
+ }
+ }
+
+ 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(TargetInstrInfo::EH_LABEL);
+ BuildMI(BB, 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);
+ }
+ }
+
// Before doing SelectionDAG ISel, see if FastISel has been requested.
- // FastISel doesn't support EH landing pads, which require special handling.
- if (EnableFastISel && !BB->isLandingPad()) {
- if (FastISel *F = TLI.createFastISel(*FuncInfo->MF, FuncInfo->ValueMap,
- FuncInfo->MBBMap)) {
- // Emit code for any incoming arguments. This must happen before
- // beginning FastISel on the entry block.
- if (LLVMBB == &Fn.getEntryBlock()) {
- CurDAG->setRoot(SDL->getControlRoot());
- CodeGenAndEmitDAG();
- SDL->clear();
- }
- F->setCurrentBlock(BB);
- // Do FastISel on as many instructions as possible.
- for (; I != End; ++I) {
- // Just before the terminator instruction, insert instructions to
- // feed PHI nodes in successor blocks.
- if (isa<TerminatorInst>(I))
- if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, F)) {
- if (DisableFastISelAbort)
- break;
-#ifndef NDEBUG
- I->dump();
-#endif
- assert(0 && "FastISel didn't handle a PHI in a successor");
+ if (FastIS && !SuppressFastISel) {
+ // Emit code for any incoming arguments. This must happen before
+ // beginning FastISel on the entry block.
+ if (LLVMBB == &Fn.getEntryBlock()) {
+ CurDAG->setRoot(SDL->getControlRoot());
+ CodeGenAndEmitDAG();
+ SDL->clear();
+ }
+ FastIS->startNewBlock(BB);
+ // Do FastISel on as many instructions as possible.
+ for (; BI != End; ++BI) {
+ // Just before the terminator instruction, insert instructions to
+ // feed PHI nodes in successor blocks.
+ if (isa<TerminatorInst>(BI))
+ if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, FastIS)) {
+ if (EnableFastISelVerbose || EnableFastISelAbort) {
+ cerr << "FastISel miss: ";
+ BI->dump();
}
+ if (EnableFastISelAbort)
+ assert(0 && "FastISel didn't handle a PHI in a successor");
+ break;
+ }
- // First try normal tablegen-generated "fast" selection.
- if (F->SelectInstruction(I))
- continue;
-
- // Next, try calling the target to attempt to handle the instruction.
- if (F->TargetSelectInstruction(I))
- continue;
-
- // Then handle certain instructions as single-LLVM-Instruction blocks.
- if (isa<CallInst>(I) || isa<LoadInst>(I) ||
- isa<StoreInst>(I)) {
- if (I->getType() != Type::VoidTy) {
- unsigned &R = FuncInfo->ValueMap[I];
- if (!R)
- R = FuncInfo->CreateRegForValue(I);
- }
+ // First try normal tablegen-generated "fast" selection.
+ if (FastIS->SelectInstruction(BI))
+ continue;
- SelectBasicBlock(LLVMBB, I, next(I));
- continue;
+ // Next, try calling the target to attempt to handle the instruction.
+ if (FastIS->TargetSelectInstruction(BI))
+ continue;
+
+ // Then handle certain instructions as single-LLVM-Instruction blocks.
+ if (isa<CallInst>(BI)) {
+ if (EnableFastISelVerbose || EnableFastISelAbort) {
+ cerr << "FastISel missed call: ";
+ BI->dump();
+ }
+
+ if (BI->getType() != Type::VoidTy) {
+ unsigned &R = FuncInfo->ValueMap[BI];
+ if (!R)
+ R = FuncInfo->CreateRegForValue(BI);
}
- if (!DisableFastISelAbort &&
- // For now, don't abort on non-conditional-branch terminators.
- (!isa<TerminatorInst>(I) ||
- (isa<BranchInst>(I) &&
- cast<BranchInst>(I)->isUnconditional()))) {
+ SelectBasicBlock(LLVMBB, BI, next(BI));
+ // If the instruction was codegen'd with multiple blocks,
+ // inform the FastISel object where to resume inserting.
+ FastIS->setCurrentBlock(BB);
+ continue;
+ }
+
+ // 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)) {
+ if (EnableFastISelVerbose || EnableFastISelAbort) {
+ cerr << "FastISel miss: ";
+ BI->dump();
+ }
+ if (EnableFastISelAbort)
// The "fast" selector couldn't handle something and bailed.
// For the purpose of debugging, just abort.
-#ifndef NDEBUG
- I->dump();
-#endif
assert(0 && "FastISel didn't select the entire block");
- }
- break;
}
- delete F;
+ break;
}
}
// Run SelectionDAG instruction selection on the remainder of the block
// not handled by FastISel. If FastISel is not run, this is the entire
// block.
- if (I != End)
- SelectBasicBlock(LLVMBB, I, End);
+ if (BI != End)
+ SelectBasicBlock(LLVMBB, BI, End);
FinishBasicBlock();
}
+
+ delete FastIS;
}
void
--e; // Don't process a flag operand if it is here.
while (i != e) {
- unsigned Flags = cast<ConstantSDNode>(InOps[i])->getValue();
+ unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
if ((Flags & 7) != 4 /*MEM*/) {
// Just skip over this operand, copying the operands verbatim.
Ops.insert(Ops.end(), InOps.begin()+i, InOps.begin()+i+(Flags >> 3) + 1);
// Add this to the output node.
MVT IntPtrTy = CurDAG->getTargetLoweringInfo().getPointerTy();
- Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size() << 3),
+ Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size()<< 3),
IntPtrTy));
Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
i += 2;