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 "llvm/CodeGen/SelectionDAGISel.h"
16 #include "SelectionDAGBuild.h"
17 #include "llvm/ADT/BitVector.h"
18 #include "llvm/Analysis/AliasAnalysis.h"
19 #include "llvm/Constants.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/InlineAsm.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/Intrinsics.h"
27 #include "llvm/IntrinsicInst.h"
28 #include "llvm/ParameterAttributes.h"
29 #include "llvm/CodeGen/FastISel.h"
30 #include "llvm/CodeGen/GCStrategy.h"
31 #include "llvm/CodeGen/GCMetadata.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineFrameInfo.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineJumpTableInfo.h"
36 #include "llvm/CodeGen/MachineModuleInfo.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/ScheduleDAG.h"
39 #include "llvm/CodeGen/SchedulerRegistry.h"
40 #include "llvm/CodeGen/SelectionDAG.h"
41 #include "llvm/Target/TargetRegisterInfo.h"
42 #include "llvm/Target/TargetData.h"
43 #include "llvm/Target/TargetFrameInfo.h"
44 #include "llvm/Target/TargetInstrInfo.h"
45 #include "llvm/Target/TargetLowering.h"
46 #include "llvm/Target/TargetMachine.h"
47 #include "llvm/Target/TargetOptions.h"
48 #include "llvm/Support/Compiler.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/Timer.h"
56 EnableValueProp("enable-value-prop", cl::Hidden);
58 EnableLegalizeTypes("enable-legalize-types", cl::Hidden);
60 EnableFastISel("fast-isel", cl::Hidden,
61 cl::desc("Enable the experimental \"fast\" instruction selector"));
63 DisableFastISelAbort("fast-isel-no-abort", cl::Hidden,
64 cl::desc("Use the SelectionDAGISel when \"fast\" instruction "
69 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
70 cl::desc("Pop up a window to show dags before the first "
73 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
74 cl::desc("Pop up a window to show dags before legalize types"));
76 ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
77 cl::desc("Pop up a window to show dags before legalize"));
79 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
80 cl::desc("Pop up a window to show dags before the second "
83 ViewISelDAGs("view-isel-dags", cl::Hidden,
84 cl::desc("Pop up a window to show isel dags as they are selected"));
86 ViewSchedDAGs("view-sched-dags", cl::Hidden,
87 cl::desc("Pop up a window to show sched dags as they are processed"));
89 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
90 cl::desc("Pop up a window to show SUnit dags after they are processed"));
92 static const bool ViewDAGCombine1 = false,
93 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false,
94 ViewDAGCombine2 = false,
95 ViewISelDAGs = false, ViewSchedDAGs = false,
96 ViewSUnitDAGs = false;
99 //===---------------------------------------------------------------------===//
101 /// RegisterScheduler class - Track the registration of instruction schedulers.
103 //===---------------------------------------------------------------------===//
104 MachinePassRegistry RegisterScheduler::Registry;
106 //===---------------------------------------------------------------------===//
108 /// ISHeuristic command line option for instruction schedulers.
110 //===---------------------------------------------------------------------===//
111 static cl::opt<RegisterScheduler::FunctionPassCtor, false,
112 RegisterPassParser<RegisterScheduler> >
113 ISHeuristic("pre-RA-sched",
114 cl::init(&createDefaultScheduler),
115 cl::desc("Instruction schedulers available (before register"
118 static RegisterScheduler
119 defaultListDAGScheduler("default", " Best scheduler for the target",
120 createDefaultScheduler);
123 //===--------------------------------------------------------------------===//
124 /// createDefaultScheduler - This creates an instruction scheduler appropriate
126 ScheduleDAG* createDefaultScheduler(SelectionDAGISel *IS,
128 MachineBasicBlock *BB,
130 TargetLowering &TLI = IS->getTargetLowering();
132 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency) {
133 return createTDListDAGScheduler(IS, DAG, BB, Fast);
135 assert(TLI.getSchedulingPreference() ==
136 TargetLowering::SchedulingForRegPressure && "Unknown sched type!");
137 return createBURRListDAGScheduler(IS, DAG, BB, Fast);
142 // EmitInstrWithCustomInserter - This method should be implemented by targets
143 // that mark instructions with the 'usesCustomDAGSchedInserter' flag. These
144 // instructions are special in various ways, which require special support to
145 // insert. The specified MachineInstr is created but not inserted into any
146 // basic blocks, and the scheduler passes ownership of it to this method.
147 MachineBasicBlock *TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
148 MachineBasicBlock *MBB) {
149 cerr << "If a target marks an instruction with "
150 << "'usesCustomDAGSchedInserter', it must implement "
151 << "TargetLowering::EmitInstrWithCustomInserter!\n";
156 //===----------------------------------------------------------------------===//
157 // SelectionDAGISel code
158 //===----------------------------------------------------------------------===//
160 SelectionDAGISel::SelectionDAGISel(TargetLowering &tli, bool fast) :
161 FunctionPass(&ID), TLI(tli),
162 FuncInfo(new FunctionLoweringInfo(TLI)),
163 CurDAG(new SelectionDAG(TLI, *FuncInfo)),
164 SDL(new SelectionDAGLowering(*CurDAG, TLI, *FuncInfo)),
170 SelectionDAGISel::~SelectionDAGISel() {
176 unsigned SelectionDAGISel::MakeReg(MVT VT) {
177 return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT));
180 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
181 AU.addRequired<AliasAnalysis>();
182 AU.addRequired<GCModuleInfo>();
183 AU.setPreservesAll();
186 bool SelectionDAGISel::runOnFunction(Function &Fn) {
187 // Get alias analysis for load/store combining.
188 AA = &getAnalysis<AliasAnalysis>();
190 MachineFunction &MF = MachineFunction::construct(&Fn, TLI.getTargetMachine());
191 if (MF.getFunction()->hasGC())
192 GFI = &getAnalysis<GCModuleInfo>().getFunctionInfo(*MF.getFunction());
195 RegInfo = &MF.getRegInfo();
196 DOUT << "\n\n\n=== " << Fn.getName() << "\n";
198 FuncInfo->set(Fn, MF, EnableFastISel);
199 CurDAG->init(MF, getAnalysisToUpdate<MachineModuleInfo>());
202 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
203 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(I->getTerminator()))
205 FuncInfo->MBBMap[Invoke->getSuccessor(1)]->setIsLandingPad();
207 SelectAllBasicBlocks(Fn, MF);
209 // Add function live-ins to entry block live-in set.
210 BasicBlock *EntryBB = &Fn.getEntryBlock();
211 BB = FuncInfo->MBBMap[EntryBB];
212 if (!RegInfo->livein_empty())
213 for (MachineRegisterInfo::livein_iterator I = RegInfo->livein_begin(),
214 E = RegInfo->livein_end(); I != E; ++I)
215 BB->addLiveIn(I->first);
218 assert(FuncInfo->CatchInfoFound.size() == FuncInfo->CatchInfoLost.size() &&
219 "Not all catch info was assigned to a landing pad!");
227 static void copyCatchInfo(BasicBlock *SrcBB, BasicBlock *DestBB,
228 MachineModuleInfo *MMI, FunctionLoweringInfo &FLI) {
229 for (BasicBlock::iterator I = SrcBB->begin(), E = --SrcBB->end(); I != E; ++I)
230 if (EHSelectorInst *EHSel = dyn_cast<EHSelectorInst>(I)) {
231 // Apply the catch info to DestBB.
232 AddCatchInfo(*EHSel, MMI, FLI.MBBMap[DestBB]);
234 if (!FLI.MBBMap[SrcBB]->isLandingPad())
235 FLI.CatchInfoFound.insert(EHSel);
240 /// IsFixedFrameObjectWithPosOffset - Check if object is a fixed frame object and
241 /// whether object offset >= 0.
243 IsFixedFrameObjectWithPosOffset(MachineFrameInfo * MFI, SDValue Op) {
244 if (!isa<FrameIndexSDNode>(Op)) return false;
246 FrameIndexSDNode * FrameIdxNode = dyn_cast<FrameIndexSDNode>(Op);
247 int FrameIdx = FrameIdxNode->getIndex();
248 return MFI->isFixedObjectIndex(FrameIdx) &&
249 MFI->getObjectOffset(FrameIdx) >= 0;
252 /// IsPossiblyOverwrittenArgumentOfTailCall - Check if the operand could
253 /// possibly be overwritten when lowering the outgoing arguments in a tail
254 /// call. Currently the implementation of this call is very conservative and
255 /// assumes all arguments sourcing from FORMAL_ARGUMENTS or a CopyFromReg with
256 /// virtual registers would be overwritten by direct lowering.
257 static bool IsPossiblyOverwrittenArgumentOfTailCall(SDValue Op,
258 MachineFrameInfo * MFI) {
259 RegisterSDNode * OpReg = NULL;
260 if (Op.getOpcode() == ISD::FORMAL_ARGUMENTS ||
261 (Op.getOpcode()== ISD::CopyFromReg &&
262 (OpReg = dyn_cast<RegisterSDNode>(Op.getOperand(1))) &&
263 (OpReg->getReg() >= TargetRegisterInfo::FirstVirtualRegister)) ||
264 (Op.getOpcode() == ISD::LOAD &&
265 IsFixedFrameObjectWithPosOffset(MFI, Op.getOperand(1))) ||
266 (Op.getOpcode() == ISD::MERGE_VALUES &&
267 Op.getOperand(Op.getResNo()).getOpcode() == ISD::LOAD &&
268 IsFixedFrameObjectWithPosOffset(MFI, Op.getOperand(Op.getResNo()).
274 /// CheckDAGForTailCallsAndFixThem - This Function looks for CALL nodes in the
275 /// DAG and fixes their tailcall attribute operand.
276 static void CheckDAGForTailCallsAndFixThem(SelectionDAG &DAG,
277 TargetLowering& TLI) {
279 SDValue Terminator = DAG.getRoot();
282 if (Terminator.getOpcode() == ISD::RET) {
283 Ret = Terminator.getNode();
286 // Fix tail call attribute of CALL nodes.
287 for (SelectionDAG::allnodes_iterator BE = DAG.allnodes_begin(),
288 BI = DAG.allnodes_end(); BI != BE; ) {
290 if (BI->getOpcode() == ISD::CALL) {
291 SDValue OpRet(Ret, 0);
292 SDValue OpCall(BI, 0);
293 bool isMarkedTailCall =
294 cast<ConstantSDNode>(OpCall.getOperand(3))->getValue() != 0;
295 // If CALL node has tail call attribute set to true and the call is not
296 // eligible (no RET or the target rejects) the attribute is fixed to
297 // false. The TargetLowering::IsEligibleForTailCallOptimization function
298 // must correctly identify tail call optimizable calls.
299 if (!isMarkedTailCall) continue;
301 !TLI.IsEligibleForTailCallOptimization(OpCall, OpRet, DAG)) {
302 // Not eligible. Mark CALL node as non tail call.
303 SmallVector<SDValue, 32> Ops;
305 for(SDNode::op_iterator I =OpCall.getNode()->op_begin(),
306 E = OpCall.getNode()->op_end(); I != E; I++, idx++) {
310 Ops.push_back(DAG.getConstant(false, TLI.getPointerTy()));
312 DAG.UpdateNodeOperands(OpCall, Ops.begin(), Ops.size());
314 // Look for tail call clobbered arguments. Emit a series of
315 // copyto/copyfrom virtual register nodes to protect them.
316 SmallVector<SDValue, 32> Ops;
317 SDValue Chain = OpCall.getOperand(0), InFlag;
319 for(SDNode::op_iterator I = OpCall.getNode()->op_begin(),
320 E = OpCall.getNode()->op_end(); I != E; I++, idx++) {
322 if (idx > 4 && (idx % 2)) {
323 bool isByVal = cast<ARG_FLAGSSDNode>(OpCall.getOperand(idx+1))->
324 getArgFlags().isByVal();
325 MachineFunction &MF = DAG.getMachineFunction();
326 MachineFrameInfo *MFI = MF.getFrameInfo();
328 IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)) {
329 MVT VT = Arg.getValueType();
330 unsigned VReg = MF.getRegInfo().
331 createVirtualRegister(TLI.getRegClassFor(VT));
332 Chain = DAG.getCopyToReg(Chain, VReg, Arg, InFlag);
333 InFlag = Chain.getValue(1);
334 Arg = DAG.getCopyFromReg(Chain, VReg, VT, InFlag);
335 Chain = Arg.getValue(1);
336 InFlag = Arg.getValue(2);
341 // Link in chain of CopyTo/CopyFromReg.
343 DAG.UpdateNodeOperands(OpCall, Ops.begin(), Ops.size());
349 void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB,
350 BasicBlock::iterator Begin,
351 BasicBlock::iterator End) {
352 SDL->setCurrentBasicBlock(BB);
354 MachineModuleInfo *MMI = CurDAG->getMachineModuleInfo();
356 if (MMI && BB->isLandingPad()) {
357 // Add a label to mark the beginning of the landing pad. Deletion of the
358 // landing pad can thus be detected via the MachineModuleInfo.
359 unsigned LabelID = MMI->addLandingPad(BB);
360 CurDAG->setRoot(CurDAG->getLabel(ISD::EH_LABEL,
361 CurDAG->getEntryNode(), LabelID));
363 // Mark exception register as live in.
364 unsigned Reg = TLI.getExceptionAddressRegister();
365 if (Reg) BB->addLiveIn(Reg);
367 // Mark exception selector register as live in.
368 Reg = TLI.getExceptionSelectorRegister();
369 if (Reg) BB->addLiveIn(Reg);
371 // FIXME: Hack around an exception handling flaw (PR1508): the personality
372 // function and list of typeids logically belong to the invoke (or, if you
373 // like, the basic block containing the invoke), and need to be associated
374 // with it in the dwarf exception handling tables. Currently however the
375 // information is provided by an intrinsic (eh.selector) that can be moved
376 // to unexpected places by the optimizers: if the unwind edge is critical,
377 // then breaking it can result in the intrinsics being in the successor of
378 // the landing pad, not the landing pad itself. This results in exceptions
379 // not being caught because no typeids are associated with the invoke.
380 // This may not be the only way things can go wrong, but it is the only way
381 // we try to work around for the moment.
382 BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
384 if (Br && Br->isUnconditional()) { // Critical edge?
385 BasicBlock::iterator I, E;
386 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
387 if (isa<EHSelectorInst>(I))
391 // No catch info found - try to extract some from the successor.
392 copyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, *FuncInfo);
396 // Lower all of the non-terminator instructions.
397 for (BasicBlock::iterator I = Begin; I != End; ++I)
398 if (!isa<TerminatorInst>(I))
401 // Ensure that all instructions which are used outside of their defining
402 // blocks are available as virtual registers. Invoke is handled elsewhere.
403 for (BasicBlock::iterator I = Begin; I != End; ++I)
404 if (!I->use_empty() && !isa<PHINode>(I) && !isa<InvokeInst>(I)) {
405 DenseMap<const Value*,unsigned>::iterator VMI =FuncInfo->ValueMap.find(I);
406 if (VMI != FuncInfo->ValueMap.end())
407 SDL->CopyValueToVirtualRegister(I, VMI->second);
410 // Handle PHI nodes in successor blocks.
411 if (End == LLVMBB->end()) {
412 HandlePHINodesInSuccessorBlocks(LLVMBB);
414 // Lower the terminator after the copies are emitted.
415 SDL->visit(*LLVMBB->getTerminator());
418 // Make sure the root of the DAG is up-to-date.
419 CurDAG->setRoot(SDL->getControlRoot());
421 // Check whether calls in this block are real tail calls. Fix up CALL nodes
422 // with correct tailcall attribute so that the target can rely on the tailcall
423 // attribute indicating whether the call is really eligible for tail call
425 CheckDAGForTailCallsAndFixThem(*CurDAG, TLI);
427 // Final step, emit the lowered DAG as machine code.
432 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
433 SmallPtrSet<SDNode*, 128> VisitedNodes;
434 SmallVector<SDNode*, 128> Worklist;
436 Worklist.push_back(CurDAG->getRoot().getNode());
442 while (!Worklist.empty()) {
443 SDNode *N = Worklist.back();
446 // If we've already seen this node, ignore it.
447 if (!VisitedNodes.insert(N))
450 // Otherwise, add all chain operands to the worklist.
451 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
452 if (N->getOperand(i).getValueType() == MVT::Other)
453 Worklist.push_back(N->getOperand(i).getNode());
455 // If this is a CopyToReg with a vreg dest, process it.
456 if (N->getOpcode() != ISD::CopyToReg)
459 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
460 if (!TargetRegisterInfo::isVirtualRegister(DestReg))
463 // Ignore non-scalar or non-integer values.
464 SDValue Src = N->getOperand(2);
465 MVT SrcVT = Src.getValueType();
466 if (!SrcVT.isInteger() || SrcVT.isVector())
469 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
470 Mask = APInt::getAllOnesValue(SrcVT.getSizeInBits());
471 CurDAG->ComputeMaskedBits(Src, Mask, KnownZero, KnownOne);
473 // Only install this information if it tells us something.
474 if (NumSignBits != 1 || KnownZero != 0 || KnownOne != 0) {
475 DestReg -= TargetRegisterInfo::FirstVirtualRegister;
476 FunctionLoweringInfo &FLI = CurDAG->getFunctionLoweringInfo();
477 if (DestReg >= FLI.LiveOutRegInfo.size())
478 FLI.LiveOutRegInfo.resize(DestReg+1);
479 FunctionLoweringInfo::LiveOutInfo &LOI = FLI.LiveOutRegInfo[DestReg];
480 LOI.NumSignBits = NumSignBits;
481 LOI.KnownOne = NumSignBits;
482 LOI.KnownZero = NumSignBits;
487 void SelectionDAGISel::CodeGenAndEmitDAG() {
488 std::string GroupName;
489 if (TimePassesIsEnabled)
490 GroupName = "Instruction Selection and Scheduling";
491 std::string BlockName;
492 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs ||
493 ViewDAGCombine2 || ViewISelDAGs || ViewSchedDAGs || ViewSUnitDAGs)
494 BlockName = CurDAG->getMachineFunction().getFunction()->getName() + ':' +
495 BB->getBasicBlock()->getName();
497 DOUT << "Initial selection DAG:\n";
498 DEBUG(CurDAG->dump());
500 if (ViewDAGCombine1) CurDAG->viewGraph("dag-combine1 input for " + BlockName);
502 // Run the DAG combiner in pre-legalize mode.
503 if (TimePassesIsEnabled) {
504 NamedRegionTimer T("DAG Combining 1", GroupName);
505 CurDAG->Combine(false, *AA, Fast);
507 CurDAG->Combine(false, *AA, Fast);
510 DOUT << "Optimized lowered selection DAG:\n";
511 DEBUG(CurDAG->dump());
513 // Second step, hack on the DAG until it only uses operations and types that
514 // the target supports.
515 if (EnableLegalizeTypes) {// Enable this some day.
516 if (ViewLegalizeTypesDAGs) CurDAG->viewGraph("legalize-types input for " +
519 if (TimePassesIsEnabled) {
520 NamedRegionTimer T("Type Legalization", GroupName);
521 CurDAG->LegalizeTypes();
523 CurDAG->LegalizeTypes();
526 DOUT << "Type-legalized selection DAG:\n";
527 DEBUG(CurDAG->dump());
529 // TODO: enable a dag combine pass here.
532 if (ViewLegalizeDAGs) CurDAG->viewGraph("legalize input for " + BlockName);
534 if (TimePassesIsEnabled) {
535 NamedRegionTimer T("DAG Legalization", GroupName);
541 DOUT << "Legalized selection DAG:\n";
542 DEBUG(CurDAG->dump());
544 if (ViewDAGCombine2) CurDAG->viewGraph("dag-combine2 input for " + BlockName);
546 // Run the DAG combiner in post-legalize mode.
547 if (TimePassesIsEnabled) {
548 NamedRegionTimer T("DAG Combining 2", GroupName);
549 CurDAG->Combine(true, *AA, Fast);
551 CurDAG->Combine(true, *AA, Fast);
554 DOUT << "Optimized legalized selection DAG:\n";
555 DEBUG(CurDAG->dump());
557 if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName);
559 if (!Fast && EnableValueProp)
560 ComputeLiveOutVRegInfo();
562 // Third, instruction select all of the operations to machine code, adding the
563 // code to the MachineBasicBlock.
564 if (TimePassesIsEnabled) {
565 NamedRegionTimer T("Instruction Selection", GroupName);
571 DOUT << "Selected selection DAG:\n";
572 DEBUG(CurDAG->dump());
574 if (ViewSchedDAGs) CurDAG->viewGraph("scheduler input for " + BlockName);
576 // Schedule machine code.
577 ScheduleDAG *Scheduler;
578 if (TimePassesIsEnabled) {
579 NamedRegionTimer T("Instruction Scheduling", GroupName);
580 Scheduler = Schedule();
582 Scheduler = Schedule();
585 if (ViewSUnitDAGs) Scheduler->viewGraph();
587 // Emit machine code to BB. This can change 'BB' to the last block being
589 if (TimePassesIsEnabled) {
590 NamedRegionTimer T("Instruction Creation", GroupName);
591 BB = Scheduler->EmitSchedule();
593 BB = Scheduler->EmitSchedule();
596 // Free the scheduler state.
597 if (TimePassesIsEnabled) {
598 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName);
604 DOUT << "Selected machine code:\n";
608 void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn, MachineFunction &MF) {
609 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
610 BasicBlock *LLVMBB = &*I;
611 BB = FuncInfo->MBBMap[LLVMBB];
613 BasicBlock::iterator const Begin = LLVMBB->begin();
614 BasicBlock::iterator const End = LLVMBB->end();
615 BasicBlock::iterator I = Begin;
617 // Lower any arguments needed in this block if this is the entry block.
618 if (LLVMBB == &Fn.getEntryBlock())
619 LowerArguments(LLVMBB);
621 // Before doing SelectionDAG ISel, see if FastISel has been requested.
622 // FastISel doesn't support EH landing pads, which require special handling.
623 if (EnableFastISel && !BB->isLandingPad()) {
624 if (FastISel *F = TLI.createFastISel(*FuncInfo->MF, FuncInfo->ValueMap,
626 // Emit code for any incoming arguments. This must happen before
627 // beginning FastISel on the entry block.
628 if (LLVMBB == &Fn.getEntryBlock()) {
629 CurDAG->setRoot(SDL->getControlRoot());
633 F->setCurrentBlock(BB);
634 // Do FastISel on as many instructions as possible.
635 for (; I != End; ++I) {
636 // Just before the terminator instruction, insert instructions to
637 // feed PHI nodes in successor blocks.
638 if (isa<TerminatorInst>(I))
639 if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, F)) {
640 if (DisableFastISelAbort)
645 assert(0 && "FastISel didn't handle a PHI in a successor");
648 // First try normal tablegen-generated "fast" selection.
649 if (F->SelectInstruction(I))
652 // Next, try calling the target to attempt to handle the instruction.
653 if (F->TargetSelectInstruction(I))
656 // Then handle certain instructions as single-LLVM-Instruction blocks.
657 if (isa<CallInst>(I) || isa<LoadInst>(I) ||
659 if (I->getType() != Type::VoidTy) {
660 unsigned &R = FuncInfo->ValueMap[I];
662 R = FuncInfo->CreateRegForValue(I);
665 SelectBasicBlock(LLVMBB, I, next(I));
669 if (!DisableFastISelAbort &&
670 // For now, don't abort on non-conditional-branch terminators.
671 (!isa<TerminatorInst>(I) ||
672 (isa<BranchInst>(I) &&
673 cast<BranchInst>(I)->isUnconditional()))) {
674 // The "fast" selector couldn't handle something and bailed.
675 // For the purpose of debugging, just abort.
679 assert(0 && "FastISel didn't select the entire block");
687 // Run SelectionDAG instruction selection on the remainder of the block
688 // not handled by FastISel. If FastISel is not run, this is the entire
691 SelectBasicBlock(LLVMBB, I, End);
698 SelectionDAGISel::FinishBasicBlock() {
700 // Perform target specific isel post processing.
701 InstructionSelectPostProcessing();
703 DOUT << "Target-post-processed machine code:\n";
706 DOUT << "Total amount of phi nodes to update: "
707 << SDL->PHINodesToUpdate.size() << "\n";
708 DEBUG(for (unsigned i = 0, e = SDL->PHINodesToUpdate.size(); i != e; ++i)
709 DOUT << "Node " << i << " : (" << SDL->PHINodesToUpdate[i].first
710 << ", " << SDL->PHINodesToUpdate[i].second << ")\n";);
712 // Next, now that we know what the last MBB the LLVM BB expanded is, update
713 // PHI nodes in successors.
714 if (SDL->SwitchCases.empty() &&
715 SDL->JTCases.empty() &&
716 SDL->BitTestCases.empty()) {
717 for (unsigned i = 0, e = SDL->PHINodesToUpdate.size(); i != e; ++i) {
718 MachineInstr *PHI = SDL->PHINodesToUpdate[i].first;
719 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
720 "This is not a machine PHI node that we are updating!");
721 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[i].second,
723 PHI->addOperand(MachineOperand::CreateMBB(BB));
725 SDL->PHINodesToUpdate.clear();
729 for (unsigned i = 0, e = SDL->BitTestCases.size(); i != e; ++i) {
730 // Lower header first, if it wasn't already lowered
731 if (!SDL->BitTestCases[i].Emitted) {
732 // Set the current basic block to the mbb we wish to insert the code into
733 BB = SDL->BitTestCases[i].Parent;
734 SDL->setCurrentBasicBlock(BB);
736 SDL->visitBitTestHeader(SDL->BitTestCases[i]);
737 CurDAG->setRoot(SDL->getRoot());
742 for (unsigned j = 0, ej = SDL->BitTestCases[i].Cases.size(); j != ej; ++j) {
743 // Set the current basic block to the mbb we wish to insert the code into
744 BB = SDL->BitTestCases[i].Cases[j].ThisBB;
745 SDL->setCurrentBasicBlock(BB);
748 SDL->visitBitTestCase(SDL->BitTestCases[i].Cases[j+1].ThisBB,
749 SDL->BitTestCases[i].Reg,
750 SDL->BitTestCases[i].Cases[j]);
752 SDL->visitBitTestCase(SDL->BitTestCases[i].Default,
753 SDL->BitTestCases[i].Reg,
754 SDL->BitTestCases[i].Cases[j]);
757 CurDAG->setRoot(SDL->getRoot());
763 for (unsigned pi = 0, pe = SDL->PHINodesToUpdate.size(); pi != pe; ++pi) {
764 MachineInstr *PHI = SDL->PHINodesToUpdate[pi].first;
765 MachineBasicBlock *PHIBB = PHI->getParent();
766 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
767 "This is not a machine PHI node that we are updating!");
768 // This is "default" BB. We have two jumps to it. From "header" BB and
769 // from last "case" BB.
770 if (PHIBB == SDL->BitTestCases[i].Default) {
771 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
773 PHI->addOperand(MachineOperand::CreateMBB(SDL->BitTestCases[i].Parent));
774 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
776 PHI->addOperand(MachineOperand::CreateMBB(SDL->BitTestCases[i].Cases.
779 // One of "cases" BB.
780 for (unsigned j = 0, ej = SDL->BitTestCases[i].Cases.size();
782 MachineBasicBlock* cBB = SDL->BitTestCases[i].Cases[j].ThisBB;
783 if (cBB->succ_end() !=
784 std::find(cBB->succ_begin(),cBB->succ_end(), PHIBB)) {
785 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
787 PHI->addOperand(MachineOperand::CreateMBB(cBB));
792 SDL->BitTestCases.clear();
794 // If the JumpTable record is filled in, then we need to emit a jump table.
795 // Updating the PHI nodes is tricky in this case, since we need to determine
796 // whether the PHI is a successor of the range check MBB or the jump table MBB
797 for (unsigned i = 0, e = SDL->JTCases.size(); i != e; ++i) {
798 // Lower header first, if it wasn't already lowered
799 if (!SDL->JTCases[i].first.Emitted) {
800 // Set the current basic block to the mbb we wish to insert the code into
801 BB = SDL->JTCases[i].first.HeaderBB;
802 SDL->setCurrentBasicBlock(BB);
804 SDL->visitJumpTableHeader(SDL->JTCases[i].second, SDL->JTCases[i].first);
805 CurDAG->setRoot(SDL->getRoot());
810 // Set the current basic block to the mbb we wish to insert the code into
811 BB = SDL->JTCases[i].second.MBB;
812 SDL->setCurrentBasicBlock(BB);
814 SDL->visitJumpTable(SDL->JTCases[i].second);
815 CurDAG->setRoot(SDL->getRoot());
820 for (unsigned pi = 0, pe = SDL->PHINodesToUpdate.size(); pi != pe; ++pi) {
821 MachineInstr *PHI = SDL->PHINodesToUpdate[pi].first;
822 MachineBasicBlock *PHIBB = PHI->getParent();
823 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
824 "This is not a machine PHI node that we are updating!");
825 // "default" BB. We can go there only from header BB.
826 if (PHIBB == SDL->JTCases[i].second.Default) {
827 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
829 PHI->addOperand(MachineOperand::CreateMBB(SDL->JTCases[i].first.HeaderBB));
831 // JT BB. Just iterate over successors here
832 if (BB->succ_end() != std::find(BB->succ_begin(),BB->succ_end(), PHIBB)) {
833 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
835 PHI->addOperand(MachineOperand::CreateMBB(BB));
839 SDL->JTCases.clear();
841 // If the switch block involved a branch to one of the actual successors, we
842 // need to update PHI nodes in that block.
843 for (unsigned i = 0, e = SDL->PHINodesToUpdate.size(); i != e; ++i) {
844 MachineInstr *PHI = SDL->PHINodesToUpdate[i].first;
845 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
846 "This is not a machine PHI node that we are updating!");
847 if (BB->isSuccessor(PHI->getParent())) {
848 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[i].second,
850 PHI->addOperand(MachineOperand::CreateMBB(BB));
854 // If we generated any switch lowering information, build and codegen any
855 // additional DAGs necessary.
856 for (unsigned i = 0, e = SDL->SwitchCases.size(); i != e; ++i) {
857 // Set the current basic block to the mbb we wish to insert the code into
858 BB = SDL->SwitchCases[i].ThisBB;
859 SDL->setCurrentBasicBlock(BB);
862 SDL->visitSwitchCase(SDL->SwitchCases[i]);
863 CurDAG->setRoot(SDL->getRoot());
867 // Handle any PHI nodes in successors of this chunk, as if we were coming
868 // from the original BB before switch expansion. Note that PHI nodes can
869 // occur multiple times in PHINodesToUpdate. We have to be very careful to
870 // handle them the right number of times.
871 while ((BB = SDL->SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
872 for (MachineBasicBlock::iterator Phi = BB->begin();
873 Phi != BB->end() && Phi->getOpcode() == TargetInstrInfo::PHI; ++Phi){
874 // This value for this PHI node is recorded in PHINodesToUpdate, get it.
875 for (unsigned pn = 0; ; ++pn) {
876 assert(pn != SDL->PHINodesToUpdate.size() &&
877 "Didn't find PHI entry!");
878 if (SDL->PHINodesToUpdate[pn].first == Phi) {
879 Phi->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pn].
881 Phi->addOperand(MachineOperand::CreateMBB(SDL->SwitchCases[i].ThisBB));
887 // Don't process RHS if same block as LHS.
888 if (BB == SDL->SwitchCases[i].FalseBB)
889 SDL->SwitchCases[i].FalseBB = 0;
891 // If we haven't handled the RHS, do so now. Otherwise, we're done.
892 SDL->SwitchCases[i].TrueBB = SDL->SwitchCases[i].FalseBB;
893 SDL->SwitchCases[i].FalseBB = 0;
895 assert(SDL->SwitchCases[i].TrueBB == 0 && SDL->SwitchCases[i].FalseBB == 0);
897 SDL->SwitchCases.clear();
899 SDL->PHINodesToUpdate.clear();
903 /// Schedule - Pick a safe ordering for instructions for each
904 /// target node in the graph.
906 ScheduleDAG *SelectionDAGISel::Schedule() {
907 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
911 RegisterScheduler::setDefault(Ctor);
914 ScheduleDAG *Scheduler = Ctor(this, CurDAG, BB, Fast);
921 HazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
922 return new HazardRecognizer();
925 //===----------------------------------------------------------------------===//
926 // Helper functions used by the generated instruction selector.
927 //===----------------------------------------------------------------------===//
928 // Calls to these methods are generated by tblgen.
930 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
931 /// the dag combiner simplified the 255, we still want to match. RHS is the
932 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
933 /// specified in the .td file (e.g. 255).
934 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
935 int64_t DesiredMaskS) const {
936 const APInt &ActualMask = RHS->getAPIntValue();
937 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
939 // If the actual mask exactly matches, success!
940 if (ActualMask == DesiredMask)
943 // If the actual AND mask is allowing unallowed bits, this doesn't match.
944 if (ActualMask.intersects(~DesiredMask))
947 // Otherwise, the DAG Combiner may have proven that the value coming in is
948 // either already zero or is not demanded. Check for known zero input bits.
949 APInt NeededMask = DesiredMask & ~ActualMask;
950 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
953 // TODO: check to see if missing bits are just not demanded.
955 // Otherwise, this pattern doesn't match.
959 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
960 /// the dag combiner simplified the 255, we still want to match. RHS is the
961 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
962 /// specified in the .td file (e.g. 255).
963 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
964 int64_t DesiredMaskS) const {
965 const APInt &ActualMask = RHS->getAPIntValue();
966 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
968 // If the actual mask exactly matches, success!
969 if (ActualMask == DesiredMask)
972 // If the actual AND mask is allowing unallowed bits, this doesn't match.
973 if (ActualMask.intersects(~DesiredMask))
976 // Otherwise, the DAG Combiner may have proven that the value coming in is
977 // either already zero or is not demanded. Check for known zero input bits.
978 APInt NeededMask = DesiredMask & ~ActualMask;
980 APInt KnownZero, KnownOne;
981 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
983 // If all the missing bits in the or are already known to be set, match!
984 if ((NeededMask & KnownOne) == NeededMask)
987 // TODO: check to see if missing bits are just not demanded.
989 // Otherwise, this pattern doesn't match.
994 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
995 /// by tblgen. Others should not call it.
996 void SelectionDAGISel::
997 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
998 std::vector<SDValue> InOps;
999 std::swap(InOps, Ops);
1001 Ops.push_back(InOps[0]); // input chain.
1002 Ops.push_back(InOps[1]); // input asm string.
1004 unsigned i = 2, e = InOps.size();
1005 if (InOps[e-1].getValueType() == MVT::Flag)
1006 --e; // Don't process a flag operand if it is here.
1009 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getValue();
1010 if ((Flags & 7) != 4 /*MEM*/) {
1011 // Just skip over this operand, copying the operands verbatim.
1012 Ops.insert(Ops.end(), InOps.begin()+i, InOps.begin()+i+(Flags >> 3) + 1);
1013 i += (Flags >> 3) + 1;
1015 assert((Flags >> 3) == 1 && "Memory operand with multiple values?");
1016 // Otherwise, this is a memory operand. Ask the target to select it.
1017 std::vector<SDValue> SelOps;
1018 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps)) {
1019 cerr << "Could not match memory address. Inline asm failure!\n";
1023 // Add this to the output node.
1024 MVT IntPtrTy = CurDAG->getTargetLoweringInfo().getPointerTy();
1025 Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size() << 3),
1027 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1032 // Add the flag input back if present.
1033 if (e != InOps.size())
1034 Ops.push_back(InOps.back());
1037 char SelectionDAGISel::ID = 0;