STATISTIC(NumFastIselBlocks, "Number of blocks selected entirely by fast isel");
STATISTIC(NumDAGBlocks, "Number of blocks selected using DAG");
STATISTIC(NumDAGIselRetries,"Number of times dag isel has to try another path");
+STATISTIC(NumEntryBlocks, "Number of entry blocks encountered");
+STATISTIC(NumFastIselFailLowerArguments,
+ "Number of entry blocks where fast isel failed to lower arguments");
#ifndef NDEBUG
static cl::opt<bool>
EnableFastISelVerbose2("fast-isel-verbose2", cl::Hidden,
cl::desc("Enable extra verbose messages in the \"fast\" "
"instruction selector"));
+
// Terminators
STATISTIC(NumFastIselFailRet,"Fast isel fails on Ret");
STATISTIC(NumFastIselFailBr,"Fast isel fails on Br");
} // end anonymous namespace
void SelectionDAGISel::DoInstructionSelection() {
- DEBUG(errs() << "===== Instruction selection begins: BB#"
+ DEBUG(dbgs() << "===== Instruction selection begins: BB#"
<< FuncInfo->MBB->getNumber()
<< " '" << FuncInfo->MBB->getName() << "'\n");
CurDAG->setRoot(Dummy.getValue());
}
- DEBUG(errs() << "===== Instruction selection ends:\n");
+ DEBUG(dbgs() << "===== Instruction selection ends:\n");
PostprocessISelDAG();
}
if (Reg) MBB->addLiveIn(Reg);
}
-/// TryToFoldFastISelLoad - We're checking to see if we can fold the specified
-/// load into the specified FoldInst. Note that we could have a sequence where
-/// multiple LLVM IR instructions are folded into the same machineinstr. For
-/// example we could have:
-/// A: x = load i32 *P
-/// B: y = icmp A, 42
-/// C: br y, ...
-///
-/// In this scenario, LI is "A", and FoldInst is "C". We know about "B" (and
-/// any other folded instructions) because it is between A and C.
-///
-/// If we succeed in folding the load into the operation, return true.
-///
-bool SelectionDAGISel::TryToFoldFastISelLoad(const LoadInst *LI,
- const Instruction *FoldInst,
- FastISel *FastIS) {
- // We know that the load has a single use, but don't know what it is. If it
- // isn't one of the folded instructions, then we can't succeed here. Handle
- // this by scanning the single-use users of the load until we get to FoldInst.
- unsigned MaxUsers = 6; // Don't scan down huge single-use chains of instrs.
-
- const Instruction *TheUser = LI->use_back();
- while (TheUser != FoldInst && // Scan up until we find FoldInst.
- // Stay in the right block.
- TheUser->getParent() == FoldInst->getParent() &&
- --MaxUsers) { // Don't scan too far.
- // If there are multiple or no uses of this instruction, then bail out.
- if (!TheUser->hasOneUse())
- return false;
-
- TheUser = TheUser->use_back();
- }
-
- // If we didn't find the fold instruction, then we failed to collapse the
- // sequence.
- if (TheUser != FoldInst)
- return false;
-
- // Don't try to fold volatile loads. Target has to deal with alignment
- // constraints.
- if (LI->isVolatile()) return false;
-
- // Figure out which vreg this is going into. If there is no assigned vreg yet
- // then there actually was no reference to it. Perhaps the load is referenced
- // by a dead instruction.
- unsigned LoadReg = FastIS->getRegForValue(LI);
- if (LoadReg == 0)
- return false;
-
- // Check to see what the uses of this vreg are. If it has no uses, or more
- // than one use (at the machine instr level) then we can't fold it.
- MachineRegisterInfo::reg_iterator RI = RegInfo->reg_begin(LoadReg);
- if (RI == RegInfo->reg_end())
- return false;
-
- // See if there is exactly one use of the vreg. If there are multiple uses,
- // then the instruction got lowered to multiple machine instructions or the
- // use of the loaded value ended up being multiple operands of the result, in
- // either case, we can't fold this.
- MachineRegisterInfo::reg_iterator PostRI = RI; ++PostRI;
- if (PostRI != RegInfo->reg_end())
- return false;
-
- assert(RI.getOperand().isUse() &&
- "The only use of the vreg must be a use, we haven't emitted the def!");
-
- MachineInstr *User = &*RI;
-
- // Set the insertion point properly. Folding the load can cause generation of
- // other random instructions (like sign extends) for addressing modes, make
- // sure they get inserted in a logical place before the new instruction.
- FuncInfo->InsertPt = User;
- FuncInfo->MBB = User->getParent();
-
- // Ask the target to try folding the load.
- return FastIS->TryToFoldLoad(User, RI.getOperandNo(), LI);
-}
-
/// isFoldedOrDeadInstruction - Return true if the specified instruction is
/// side-effect free and is either dead or folded into a generated instruction.
/// Return false if it needs to be emitted.
// Emit code for any incoming arguments. This must happen before
// beginning FastISel on the entry block.
if (LLVMBB == &Fn.getEntryBlock()) {
+ ++NumEntryBlocks;
+
// Lower any arguments needed in this block if this is the entry block.
if (!FastIS->LowerArguments()) {
// Fast isel failed to lower these arguments
+ ++NumFastIselFailLowerArguments;
if (EnableFastISelAbortArgs)
llvm_unreachable("FastISel didn't lower all arguments");
}
if (BeforeInst != Inst && isa<LoadInst>(BeforeInst) &&
BeforeInst->hasOneUse() &&
- TryToFoldFastISelLoad(cast<LoadInst>(BeforeInst), Inst, FastIS)) {
+ FastIS->tryToFoldLoad(cast<LoadInst>(BeforeInst), Inst)) {
// If we succeeded, don't re-select the load.
BI = llvm::next(BasicBlock::const_iterator(BeforeInst));
--NumFastIselRemaining;
FastIS->recomputeInsertPt();
} else {
// Lower any arguments needed in this block if this is the entry block.
- if (LLVMBB == &Fn.getEntryBlock())
+ if (LLVMBB == &Fn.getEntryBlock()) {
+ ++NumEntryBlocks;
LowerArguments(Fn);
+ }
}
if (Begin != BI)
if (!NowDeadNodes.empty())
CurDAG->RemoveDeadNodes(NowDeadNodes);
- DEBUG(errs() << "ISEL: Match complete!\n");
+ DEBUG(dbgs() << "ISEL: Match complete!\n");
}
enum ChainResult {
SmallVector<SDNode*, 3> ChainNodesMatched;
SmallVector<SDNode*, 3> GlueResultNodesMatched;
- DEBUG(errs() << "ISEL: Starting pattern match on root node: ";
+ DEBUG(dbgs() << "ISEL: Starting pattern match on root node: ";
NodeToMatch->dump(CurDAG);
- errs() << '\n');
+ dbgs() << '\n');
// Determine where to start the interpreter. Normally we start at opcode #0,
// but if the state machine starts with an OPC_SwitchOpcode, then we
// Already computed the OpcodeOffset table, just index into it.
if (N.getOpcode() < OpcodeOffset.size())
MatcherIndex = OpcodeOffset[N.getOpcode()];
- DEBUG(errs() << " Initial Opcode index to " << MatcherIndex << "\n");
+ DEBUG(dbgs() << " Initial Opcode index to " << MatcherIndex << "\n");
} else if (MatcherTable[0] == OPC_SwitchOpcode) {
// Otherwise, the table isn't computed, but the state machine does start
if (!Result)
break;
- DEBUG(errs() << " Skipped scope entry (due to false predicate) at "
+ DEBUG(dbgs() << " Skipped scope entry (due to false predicate) at "
<< "index " << MatcherIndexOfPredicate
<< ", continuing at " << FailIndex << "\n");
++NumDAGIselRetries;
if (CaseSize == 0) break;
// Otherwise, execute the case we found.
- DEBUG(errs() << " OpcodeSwitch from " << SwitchStart
+ DEBUG(dbgs() << " OpcodeSwitch from " << SwitchStart
<< " to " << MatcherIndex << "\n");
continue;
}
if (CaseSize == 0) break;
// Otherwise, execute the case we found.
- DEBUG(errs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
+ DEBUG(dbgs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString()
<< "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
continue;
}
// If this is a normal EmitNode command, just create the new node and
// add the results to the RecordedNodes list.
Res = CurDAG->getMachineNode(TargetOpc, NodeToMatch->getDebugLoc(),
- VTList, Ops.data(), Ops.size());
+ VTList, Ops);
// Add all the non-glue/non-chain results to the RecordedNodes list.
for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
->setMemRefs(MemRefs, MemRefs + NumMemRefs);
}
- DEBUG(errs() << " "
+ DEBUG(dbgs() << " "
<< (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created")
- << " node: "; Res->dump(CurDAG); errs() << "\n");
+ << " node: "; Res->dump(CurDAG); dbgs() << "\n");
// If this was a MorphNodeTo then we're completely done!
if (Opcode == OPC_MorphNodeTo) {
// 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");
+ DEBUG(dbgs() << " Match failed at index " << CurrentOpcodeIndex << "\n");
++NumDAGIselRetries;
while (1) {
if (MatchScopes.empty()) {
MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
MatcherIndex = LastScope.FailIndex;
- DEBUG(errs() << " Continuing at " << MatcherIndex << "\n");
+ DEBUG(dbgs() << " Continuing at " << MatcherIndex << "\n");
InputChain = LastScope.InputChain;
InputGlue = LastScope.InputGlue;
projects / oota-llvm.git / blobdiff