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 "
67 SchedLiveInCopies("schedule-livein-copies",
68 cl::desc("Schedule copies of livein registers"),
73 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
74 cl::desc("Pop up a window to show dags before the first "
77 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
78 cl::desc("Pop up a window to show dags before legalize types"));
80 ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
81 cl::desc("Pop up a window to show dags before legalize"));
83 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
84 cl::desc("Pop up a window to show dags before the second "
87 ViewISelDAGs("view-isel-dags", cl::Hidden,
88 cl::desc("Pop up a window to show isel dags as they are selected"));
90 ViewSchedDAGs("view-sched-dags", cl::Hidden,
91 cl::desc("Pop up a window to show sched dags as they are processed"));
93 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
94 cl::desc("Pop up a window to show SUnit dags after they are processed"));
96 static const bool ViewDAGCombine1 = false,
97 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false,
98 ViewDAGCombine2 = false,
99 ViewISelDAGs = false, ViewSchedDAGs = false,
100 ViewSUnitDAGs = false;
103 //===---------------------------------------------------------------------===//
105 /// RegisterScheduler class - Track the registration of instruction schedulers.
107 //===---------------------------------------------------------------------===//
108 MachinePassRegistry RegisterScheduler::Registry;
110 //===---------------------------------------------------------------------===//
112 /// ISHeuristic command line option for instruction schedulers.
114 //===---------------------------------------------------------------------===//
115 static cl::opt<RegisterScheduler::FunctionPassCtor, false,
116 RegisterPassParser<RegisterScheduler> >
117 ISHeuristic("pre-RA-sched",
118 cl::init(&createDefaultScheduler),
119 cl::desc("Instruction schedulers available (before register"
122 static RegisterScheduler
123 defaultListDAGScheduler("default", " Best scheduler for the target",
124 createDefaultScheduler);
127 //===--------------------------------------------------------------------===//
128 /// createDefaultScheduler - This creates an instruction scheduler appropriate
130 ScheduleDAG* createDefaultScheduler(SelectionDAGISel *IS,
132 MachineBasicBlock *BB,
134 TargetLowering &TLI = IS->getTargetLowering();
136 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency) {
137 return createTDListDAGScheduler(IS, DAG, BB, Fast);
139 assert(TLI.getSchedulingPreference() ==
140 TargetLowering::SchedulingForRegPressure && "Unknown sched type!");
141 return createBURRListDAGScheduler(IS, DAG, BB, Fast);
146 // EmitInstrWithCustomInserter - This method should be implemented by targets
147 // that mark instructions with the 'usesCustomDAGSchedInserter' flag. These
148 // instructions are special in various ways, which require special support to
149 // insert. The specified MachineInstr is created but not inserted into any
150 // basic blocks, and the scheduler passes ownership of it to this method.
151 MachineBasicBlock *TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
152 MachineBasicBlock *MBB) {
153 cerr << "If a target marks an instruction with "
154 << "'usesCustomDAGSchedInserter', it must implement "
155 << "TargetLowering::EmitInstrWithCustomInserter!\n";
160 /// EmitLiveInCopy - Emit a copy for a live in physical register. If the
161 /// physical register has only a single copy use, then coalesced the copy
163 static void EmitLiveInCopy(MachineBasicBlock *MBB,
164 MachineBasicBlock::iterator &InsertPos,
165 unsigned VirtReg, unsigned PhysReg,
166 const TargetRegisterClass *RC,
167 DenseMap<MachineInstr*, unsigned> &CopyRegMap,
168 const MachineRegisterInfo &MRI,
169 const TargetRegisterInfo &TRI,
170 const TargetInstrInfo &TII) {
171 unsigned NumUses = 0;
172 MachineInstr *UseMI = NULL;
173 for (MachineRegisterInfo::use_iterator UI = MRI.use_begin(VirtReg),
174 UE = MRI.use_end(); UI != UE; ++UI) {
180 // If the number of uses is not one, or the use is not a move instruction,
181 // don't coalesce. Also, only coalesce away a virtual register to virtual
183 bool Coalesced = false;
184 unsigned SrcReg, DstReg;
186 TII.isMoveInstr(*UseMI, SrcReg, DstReg) &&
187 TargetRegisterInfo::isVirtualRegister(DstReg)) {
192 // Now find an ideal location to insert the copy.
193 MachineBasicBlock::iterator Pos = InsertPos;
194 while (Pos != MBB->begin()) {
195 MachineInstr *PrevMI = prior(Pos);
196 DenseMap<MachineInstr*, unsigned>::iterator RI = CopyRegMap.find(PrevMI);
197 // copyRegToReg might emit multiple instructions to do a copy.
198 unsigned CopyDstReg = (RI == CopyRegMap.end()) ? 0 : RI->second;
199 if (CopyDstReg && !TRI.regsOverlap(CopyDstReg, PhysReg))
200 // This is what the BB looks like right now:
205 // We want to insert "r1025 = mov r1". Inserting this copy below the
206 // move to r1024 makes it impossible for that move to be coalesced.
213 break; // Woot! Found a good location.
217 TII.copyRegToReg(*MBB, Pos, VirtReg, PhysReg, RC, RC);
218 CopyRegMap.insert(std::make_pair(prior(Pos), VirtReg));
220 if (&*InsertPos == UseMI) ++InsertPos;
225 /// EmitLiveInCopies - If this is the first basic block in the function,
226 /// and if it has live ins that need to be copied into vregs, emit the
227 /// copies into the block.
228 static void EmitLiveInCopies(MachineBasicBlock *EntryMBB,
229 const MachineRegisterInfo &MRI,
230 const TargetRegisterInfo &TRI,
231 const TargetInstrInfo &TII) {
232 if (SchedLiveInCopies) {
233 // Emit the copies at a heuristically-determined location in the block.
234 DenseMap<MachineInstr*, unsigned> CopyRegMap;
235 MachineBasicBlock::iterator InsertPos = EntryMBB->begin();
236 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
237 E = MRI.livein_end(); LI != E; ++LI)
239 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
240 EmitLiveInCopy(EntryMBB, InsertPos, LI->second, LI->first,
241 RC, CopyRegMap, MRI, TRI, TII);
244 // Emit the copies into the top of the block.
245 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
246 E = MRI.livein_end(); LI != E; ++LI)
248 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
249 TII.copyRegToReg(*EntryMBB, EntryMBB->begin(),
250 LI->second, LI->first, RC, RC);
255 //===----------------------------------------------------------------------===//
256 // SelectionDAGISel code
257 //===----------------------------------------------------------------------===//
259 SelectionDAGISel::SelectionDAGISel(TargetLowering &tli, bool fast) :
260 FunctionPass(&ID), TLI(tli),
261 FuncInfo(new FunctionLoweringInfo(TLI)),
262 CurDAG(new SelectionDAG(TLI, *FuncInfo)),
263 SDL(new SelectionDAGLowering(*CurDAG, TLI, *FuncInfo)),
269 SelectionDAGISel::~SelectionDAGISel() {
275 unsigned SelectionDAGISel::MakeReg(MVT VT) {
276 return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT));
279 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
280 AU.addRequired<AliasAnalysis>();
281 AU.addRequired<GCModuleInfo>();
282 AU.setPreservesAll();
285 bool SelectionDAGISel::runOnFunction(Function &Fn) {
286 // Get alias analysis for load/store combining.
287 AA = &getAnalysis<AliasAnalysis>();
289 TargetMachine &TM = TLI.getTargetMachine();
290 MachineFunction &MF = MachineFunction::construct(&Fn, TM);
291 const MachineRegisterInfo &MRI = MF.getRegInfo();
292 const TargetInstrInfo &TII = *TM.getInstrInfo();
293 const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
295 if (MF.getFunction()->hasGC())
296 GFI = &getAnalysis<GCModuleInfo>().getFunctionInfo(*MF.getFunction());
299 RegInfo = &MF.getRegInfo();
300 DOUT << "\n\n\n=== " << Fn.getName() << "\n";
302 FuncInfo->set(Fn, MF, EnableFastISel);
303 CurDAG->init(MF, getAnalysisToUpdate<MachineModuleInfo>());
306 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
307 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(I->getTerminator()))
309 FuncInfo->MBBMap[Invoke->getSuccessor(1)]->setIsLandingPad();
311 SelectAllBasicBlocks(Fn, MF);
313 // If the first basic block in the function has live ins that need to be
314 // copied into vregs, emit the copies into the top of the block before
315 // emitting the code for the block.
316 EmitLiveInCopies(MF.begin(), MRI, TRI, TII);
318 // Add function live-ins to entry block live-in set.
319 for (MachineRegisterInfo::livein_iterator I = RegInfo->livein_begin(),
320 E = RegInfo->livein_end(); I != E; ++I)
321 MF.begin()->addLiveIn(I->first);
324 assert(FuncInfo->CatchInfoFound.size() == FuncInfo->CatchInfoLost.size() &&
325 "Not all catch info was assigned to a landing pad!");
333 static void copyCatchInfo(BasicBlock *SrcBB, BasicBlock *DestBB,
334 MachineModuleInfo *MMI, FunctionLoweringInfo &FLI) {
335 for (BasicBlock::iterator I = SrcBB->begin(), E = --SrcBB->end(); I != E; ++I)
336 if (EHSelectorInst *EHSel = dyn_cast<EHSelectorInst>(I)) {
337 // Apply the catch info to DestBB.
338 AddCatchInfo(*EHSel, MMI, FLI.MBBMap[DestBB]);
340 if (!FLI.MBBMap[SrcBB]->isLandingPad())
341 FLI.CatchInfoFound.insert(EHSel);
346 /// IsFixedFrameObjectWithPosOffset - Check if object is a fixed frame object and
347 /// whether object offset >= 0.
349 IsFixedFrameObjectWithPosOffset(MachineFrameInfo * MFI, SDValue Op) {
350 if (!isa<FrameIndexSDNode>(Op)) return false;
352 FrameIndexSDNode * FrameIdxNode = dyn_cast<FrameIndexSDNode>(Op);
353 int FrameIdx = FrameIdxNode->getIndex();
354 return MFI->isFixedObjectIndex(FrameIdx) &&
355 MFI->getObjectOffset(FrameIdx) >= 0;
358 /// IsPossiblyOverwrittenArgumentOfTailCall - Check if the operand could
359 /// possibly be overwritten when lowering the outgoing arguments in a tail
360 /// call. Currently the implementation of this call is very conservative and
361 /// assumes all arguments sourcing from FORMAL_ARGUMENTS or a CopyFromReg with
362 /// virtual registers would be overwritten by direct lowering.
363 static bool IsPossiblyOverwrittenArgumentOfTailCall(SDValue Op,
364 MachineFrameInfo * MFI) {
365 RegisterSDNode * OpReg = NULL;
366 if (Op.getOpcode() == ISD::FORMAL_ARGUMENTS ||
367 (Op.getOpcode()== ISD::CopyFromReg &&
368 (OpReg = dyn_cast<RegisterSDNode>(Op.getOperand(1))) &&
369 (OpReg->getReg() >= TargetRegisterInfo::FirstVirtualRegister)) ||
370 (Op.getOpcode() == ISD::LOAD &&
371 IsFixedFrameObjectWithPosOffset(MFI, Op.getOperand(1))) ||
372 (Op.getOpcode() == ISD::MERGE_VALUES &&
373 Op.getOperand(Op.getResNo()).getOpcode() == ISD::LOAD &&
374 IsFixedFrameObjectWithPosOffset(MFI, Op.getOperand(Op.getResNo()).
380 /// CheckDAGForTailCallsAndFixThem - This Function looks for CALL nodes in the
381 /// DAG and fixes their tailcall attribute operand.
382 static void CheckDAGForTailCallsAndFixThem(SelectionDAG &DAG,
383 TargetLowering& TLI) {
385 SDValue Terminator = DAG.getRoot();
388 if (Terminator.getOpcode() == ISD::RET) {
389 Ret = Terminator.getNode();
392 // Fix tail call attribute of CALL nodes.
393 for (SelectionDAG::allnodes_iterator BE = DAG.allnodes_begin(),
394 BI = DAG.allnodes_end(); BI != BE; ) {
396 if (BI->getOpcode() == ISD::CALL) {
397 SDValue OpRet(Ret, 0);
398 SDValue OpCall(BI, 0);
399 bool isMarkedTailCall =
400 cast<ConstantSDNode>(OpCall.getOperand(3))->getValue() != 0;
401 // If CALL node has tail call attribute set to true and the call is not
402 // eligible (no RET or the target rejects) the attribute is fixed to
403 // false. The TargetLowering::IsEligibleForTailCallOptimization function
404 // must correctly identify tail call optimizable calls.
405 if (!isMarkedTailCall) continue;
407 !TLI.IsEligibleForTailCallOptimization(OpCall, OpRet, DAG)) {
408 // Not eligible. Mark CALL node as non tail call.
409 SmallVector<SDValue, 32> Ops;
411 for(SDNode::op_iterator I =OpCall.getNode()->op_begin(),
412 E = OpCall.getNode()->op_end(); I != E; I++, idx++) {
416 Ops.push_back(DAG.getConstant(false, TLI.getPointerTy()));
418 DAG.UpdateNodeOperands(OpCall, Ops.begin(), Ops.size());
420 // Look for tail call clobbered arguments. Emit a series of
421 // copyto/copyfrom virtual register nodes to protect them.
422 SmallVector<SDValue, 32> Ops;
423 SDValue Chain = OpCall.getOperand(0), InFlag;
425 for(SDNode::op_iterator I = OpCall.getNode()->op_begin(),
426 E = OpCall.getNode()->op_end(); I != E; I++, idx++) {
428 if (idx > 4 && (idx % 2)) {
429 bool isByVal = cast<ARG_FLAGSSDNode>(OpCall.getOperand(idx+1))->
430 getArgFlags().isByVal();
431 MachineFunction &MF = DAG.getMachineFunction();
432 MachineFrameInfo *MFI = MF.getFrameInfo();
434 IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)) {
435 MVT VT = Arg.getValueType();
436 unsigned VReg = MF.getRegInfo().
437 createVirtualRegister(TLI.getRegClassFor(VT));
438 Chain = DAG.getCopyToReg(Chain, VReg, Arg, InFlag);
439 InFlag = Chain.getValue(1);
440 Arg = DAG.getCopyFromReg(Chain, VReg, VT, InFlag);
441 Chain = Arg.getValue(1);
442 InFlag = Arg.getValue(2);
447 // Link in chain of CopyTo/CopyFromReg.
449 DAG.UpdateNodeOperands(OpCall, Ops.begin(), Ops.size());
455 void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB,
456 BasicBlock::iterator Begin,
457 BasicBlock::iterator End) {
458 SDL->setCurrentBasicBlock(BB);
460 MachineModuleInfo *MMI = CurDAG->getMachineModuleInfo();
462 if (MMI && BB->isLandingPad()) {
463 // Add a label to mark the beginning of the landing pad. Deletion of the
464 // landing pad can thus be detected via the MachineModuleInfo.
465 unsigned LabelID = MMI->addLandingPad(BB);
466 CurDAG->setRoot(CurDAG->getLabel(ISD::EH_LABEL,
467 CurDAG->getEntryNode(), LabelID));
469 // Mark exception register as live in.
470 unsigned Reg = TLI.getExceptionAddressRegister();
471 if (Reg) BB->addLiveIn(Reg);
473 // Mark exception selector register as live in.
474 Reg = TLI.getExceptionSelectorRegister();
475 if (Reg) BB->addLiveIn(Reg);
477 // FIXME: Hack around an exception handling flaw (PR1508): the personality
478 // function and list of typeids logically belong to the invoke (or, if you
479 // like, the basic block containing the invoke), and need to be associated
480 // with it in the dwarf exception handling tables. Currently however the
481 // information is provided by an intrinsic (eh.selector) that can be moved
482 // to unexpected places by the optimizers: if the unwind edge is critical,
483 // then breaking it can result in the intrinsics being in the successor of
484 // the landing pad, not the landing pad itself. This results in exceptions
485 // not being caught because no typeids are associated with the invoke.
486 // This may not be the only way things can go wrong, but it is the only way
487 // we try to work around for the moment.
488 BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
490 if (Br && Br->isUnconditional()) { // Critical edge?
491 BasicBlock::iterator I, E;
492 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
493 if (isa<EHSelectorInst>(I))
497 // No catch info found - try to extract some from the successor.
498 copyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, *FuncInfo);
502 // Lower all of the non-terminator instructions.
503 for (BasicBlock::iterator I = Begin; I != End; ++I)
504 if (!isa<TerminatorInst>(I))
507 // Ensure that all instructions which are used outside of their defining
508 // blocks are available as virtual registers. Invoke is handled elsewhere.
509 for (BasicBlock::iterator I = Begin; I != End; ++I)
510 if (!I->use_empty() && !isa<PHINode>(I) && !isa<InvokeInst>(I)) {
511 DenseMap<const Value*,unsigned>::iterator VMI =FuncInfo->ValueMap.find(I);
512 if (VMI != FuncInfo->ValueMap.end())
513 SDL->CopyValueToVirtualRegister(I, VMI->second);
516 // Handle PHI nodes in successor blocks.
517 if (End == LLVMBB->end()) {
518 HandlePHINodesInSuccessorBlocks(LLVMBB);
520 // Lower the terminator after the copies are emitted.
521 SDL->visit(*LLVMBB->getTerminator());
524 // Make sure the root of the DAG is up-to-date.
525 CurDAG->setRoot(SDL->getControlRoot());
527 // Check whether calls in this block are real tail calls. Fix up CALL nodes
528 // with correct tailcall attribute so that the target can rely on the tailcall
529 // attribute indicating whether the call is really eligible for tail call
531 CheckDAGForTailCallsAndFixThem(*CurDAG, TLI);
533 // Final step, emit the lowered DAG as machine code.
538 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
539 SmallPtrSet<SDNode*, 128> VisitedNodes;
540 SmallVector<SDNode*, 128> Worklist;
542 Worklist.push_back(CurDAG->getRoot().getNode());
548 while (!Worklist.empty()) {
549 SDNode *N = Worklist.back();
552 // If we've already seen this node, ignore it.
553 if (!VisitedNodes.insert(N))
556 // Otherwise, add all chain operands to the worklist.
557 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
558 if (N->getOperand(i).getValueType() == MVT::Other)
559 Worklist.push_back(N->getOperand(i).getNode());
561 // If this is a CopyToReg with a vreg dest, process it.
562 if (N->getOpcode() != ISD::CopyToReg)
565 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
566 if (!TargetRegisterInfo::isVirtualRegister(DestReg))
569 // Ignore non-scalar or non-integer values.
570 SDValue Src = N->getOperand(2);
571 MVT SrcVT = Src.getValueType();
572 if (!SrcVT.isInteger() || SrcVT.isVector())
575 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
576 Mask = APInt::getAllOnesValue(SrcVT.getSizeInBits());
577 CurDAG->ComputeMaskedBits(Src, Mask, KnownZero, KnownOne);
579 // Only install this information if it tells us something.
580 if (NumSignBits != 1 || KnownZero != 0 || KnownOne != 0) {
581 DestReg -= TargetRegisterInfo::FirstVirtualRegister;
582 FunctionLoweringInfo &FLI = CurDAG->getFunctionLoweringInfo();
583 if (DestReg >= FLI.LiveOutRegInfo.size())
584 FLI.LiveOutRegInfo.resize(DestReg+1);
585 FunctionLoweringInfo::LiveOutInfo &LOI = FLI.LiveOutRegInfo[DestReg];
586 LOI.NumSignBits = NumSignBits;
587 LOI.KnownOne = NumSignBits;
588 LOI.KnownZero = NumSignBits;
593 void SelectionDAGISel::CodeGenAndEmitDAG() {
594 std::string GroupName;
595 if (TimePassesIsEnabled)
596 GroupName = "Instruction Selection and Scheduling";
597 std::string BlockName;
598 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs ||
599 ViewDAGCombine2 || ViewISelDAGs || ViewSchedDAGs || ViewSUnitDAGs)
600 BlockName = CurDAG->getMachineFunction().getFunction()->getName() + ':' +
601 BB->getBasicBlock()->getName();
603 DOUT << "Initial selection DAG:\n";
604 DEBUG(CurDAG->dump());
606 if (ViewDAGCombine1) CurDAG->viewGraph("dag-combine1 input for " + BlockName);
608 // Run the DAG combiner in pre-legalize mode.
609 if (TimePassesIsEnabled) {
610 NamedRegionTimer T("DAG Combining 1", GroupName);
611 CurDAG->Combine(false, *AA, Fast);
613 CurDAG->Combine(false, *AA, Fast);
616 DOUT << "Optimized lowered selection DAG:\n";
617 DEBUG(CurDAG->dump());
619 // Second step, hack on the DAG until it only uses operations and types that
620 // the target supports.
621 if (EnableLegalizeTypes) {// Enable this some day.
622 if (ViewLegalizeTypesDAGs) CurDAG->viewGraph("legalize-types input for " +
625 if (TimePassesIsEnabled) {
626 NamedRegionTimer T("Type Legalization", GroupName);
627 CurDAG->LegalizeTypes();
629 CurDAG->LegalizeTypes();
632 DOUT << "Type-legalized selection DAG:\n";
633 DEBUG(CurDAG->dump());
635 // TODO: enable a dag combine pass here.
638 if (ViewLegalizeDAGs) CurDAG->viewGraph("legalize input for " + BlockName);
640 if (TimePassesIsEnabled) {
641 NamedRegionTimer T("DAG Legalization", GroupName);
647 DOUT << "Legalized selection DAG:\n";
648 DEBUG(CurDAG->dump());
650 if (ViewDAGCombine2) CurDAG->viewGraph("dag-combine2 input for " + BlockName);
652 // Run the DAG combiner in post-legalize mode.
653 if (TimePassesIsEnabled) {
654 NamedRegionTimer T("DAG Combining 2", GroupName);
655 CurDAG->Combine(true, *AA, Fast);
657 CurDAG->Combine(true, *AA, Fast);
660 DOUT << "Optimized legalized selection DAG:\n";
661 DEBUG(CurDAG->dump());
663 if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName);
665 if (!Fast && EnableValueProp)
666 ComputeLiveOutVRegInfo();
668 // Third, instruction select all of the operations to machine code, adding the
669 // code to the MachineBasicBlock.
670 if (TimePassesIsEnabled) {
671 NamedRegionTimer T("Instruction Selection", GroupName);
677 DOUT << "Selected selection DAG:\n";
678 DEBUG(CurDAG->dump());
680 if (ViewSchedDAGs) CurDAG->viewGraph("scheduler input for " + BlockName);
682 // Schedule machine code.
683 ScheduleDAG *Scheduler;
684 if (TimePassesIsEnabled) {
685 NamedRegionTimer T("Instruction Scheduling", GroupName);
686 Scheduler = Schedule();
688 Scheduler = Schedule();
691 if (ViewSUnitDAGs) Scheduler->viewGraph();
693 // Emit machine code to BB. This can change 'BB' to the last block being
695 if (TimePassesIsEnabled) {
696 NamedRegionTimer T("Instruction Creation", GroupName);
697 BB = Scheduler->EmitSchedule();
699 BB = Scheduler->EmitSchedule();
702 // Free the scheduler state.
703 if (TimePassesIsEnabled) {
704 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName);
710 DOUT << "Selected machine code:\n";
714 void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn, MachineFunction &MF) {
715 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
716 BasicBlock *LLVMBB = &*I;
717 BB = FuncInfo->MBBMap[LLVMBB];
719 BasicBlock::iterator const Begin = LLVMBB->begin();
720 BasicBlock::iterator const End = LLVMBB->end();
721 BasicBlock::iterator BI = Begin;
723 // Lower any arguments needed in this block if this is the entry block.
724 if (LLVMBB == &Fn.getEntryBlock())
725 LowerArguments(LLVMBB);
727 // Before doing SelectionDAG ISel, see if FastISel has been requested.
728 // FastISel doesn't support EH landing pads, which require special handling.
729 if (EnableFastISel && !BB->isLandingPad()) {
730 if (FastISel *F = TLI.createFastISel(*FuncInfo->MF, FuncInfo->ValueMap,
732 // Emit code for any incoming arguments. This must happen before
733 // beginning FastISel on the entry block.
734 if (LLVMBB == &Fn.getEntryBlock()) {
735 CurDAG->setRoot(SDL->getControlRoot());
739 F->setCurrentBlock(BB);
740 // Do FastISel on as many instructions as possible.
741 for (; BI != End; ++BI) {
742 // Just before the terminator instruction, insert instructions to
743 // feed PHI nodes in successor blocks.
744 if (isa<TerminatorInst>(BI))
745 if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, F)) {
746 if (DisableFastISelAbort)
751 assert(0 && "FastISel didn't handle a PHI in a successor");
754 // First try normal tablegen-generated "fast" selection.
755 if (F->SelectInstruction(BI))
758 // Next, try calling the target to attempt to handle the instruction.
759 if (F->TargetSelectInstruction(BI))
762 // Then handle certain instructions as single-LLVM-Instruction blocks.
763 if (isa<CallInst>(BI) || isa<LoadInst>(BI) ||
764 isa<StoreInst>(BI)) {
765 if (BI->getType() != Type::VoidTy) {
766 unsigned &R = FuncInfo->ValueMap[BI];
768 R = FuncInfo->CreateRegForValue(BI);
771 SelectBasicBlock(LLVMBB, BI, next(BI));
775 if (!DisableFastISelAbort &&
776 // For now, don't abort on non-conditional-branch terminators.
777 (!isa<TerminatorInst>(BI) ||
778 (isa<BranchInst>(BI) &&
779 cast<BranchInst>(BI)->isUnconditional()))) {
780 // The "fast" selector couldn't handle something and bailed.
781 // For the purpose of debugging, just abort.
785 assert(0 && "FastISel didn't select the entire block");
793 // Run SelectionDAG instruction selection on the remainder of the block
794 // not handled by FastISel. If FastISel is not run, this is the entire
797 SelectBasicBlock(LLVMBB, BI, End);
804 SelectionDAGISel::FinishBasicBlock() {
806 // Perform target specific isel post processing.
807 InstructionSelectPostProcessing();
809 DOUT << "Target-post-processed machine code:\n";
812 DOUT << "Total amount of phi nodes to update: "
813 << SDL->PHINodesToUpdate.size() << "\n";
814 DEBUG(for (unsigned i = 0, e = SDL->PHINodesToUpdate.size(); i != e; ++i)
815 DOUT << "Node " << i << " : (" << SDL->PHINodesToUpdate[i].first
816 << ", " << SDL->PHINodesToUpdate[i].second << ")\n";);
818 // Next, now that we know what the last MBB the LLVM BB expanded is, update
819 // PHI nodes in successors.
820 if (SDL->SwitchCases.empty() &&
821 SDL->JTCases.empty() &&
822 SDL->BitTestCases.empty()) {
823 for (unsigned i = 0, e = SDL->PHINodesToUpdate.size(); i != e; ++i) {
824 MachineInstr *PHI = SDL->PHINodesToUpdate[i].first;
825 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
826 "This is not a machine PHI node that we are updating!");
827 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[i].second,
829 PHI->addOperand(MachineOperand::CreateMBB(BB));
831 SDL->PHINodesToUpdate.clear();
835 for (unsigned i = 0, e = SDL->BitTestCases.size(); i != e; ++i) {
836 // Lower header first, if it wasn't already lowered
837 if (!SDL->BitTestCases[i].Emitted) {
838 // Set the current basic block to the mbb we wish to insert the code into
839 BB = SDL->BitTestCases[i].Parent;
840 SDL->setCurrentBasicBlock(BB);
842 SDL->visitBitTestHeader(SDL->BitTestCases[i]);
843 CurDAG->setRoot(SDL->getRoot());
848 for (unsigned j = 0, ej = SDL->BitTestCases[i].Cases.size(); j != ej; ++j) {
849 // Set the current basic block to the mbb we wish to insert the code into
850 BB = SDL->BitTestCases[i].Cases[j].ThisBB;
851 SDL->setCurrentBasicBlock(BB);
854 SDL->visitBitTestCase(SDL->BitTestCases[i].Cases[j+1].ThisBB,
855 SDL->BitTestCases[i].Reg,
856 SDL->BitTestCases[i].Cases[j]);
858 SDL->visitBitTestCase(SDL->BitTestCases[i].Default,
859 SDL->BitTestCases[i].Reg,
860 SDL->BitTestCases[i].Cases[j]);
863 CurDAG->setRoot(SDL->getRoot());
869 for (unsigned pi = 0, pe = SDL->PHINodesToUpdate.size(); pi != pe; ++pi) {
870 MachineInstr *PHI = SDL->PHINodesToUpdate[pi].first;
871 MachineBasicBlock *PHIBB = PHI->getParent();
872 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
873 "This is not a machine PHI node that we are updating!");
874 // This is "default" BB. We have two jumps to it. From "header" BB and
875 // from last "case" BB.
876 if (PHIBB == SDL->BitTestCases[i].Default) {
877 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
879 PHI->addOperand(MachineOperand::CreateMBB(SDL->BitTestCases[i].Parent));
880 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
882 PHI->addOperand(MachineOperand::CreateMBB(SDL->BitTestCases[i].Cases.
885 // One of "cases" BB.
886 for (unsigned j = 0, ej = SDL->BitTestCases[i].Cases.size();
888 MachineBasicBlock* cBB = SDL->BitTestCases[i].Cases[j].ThisBB;
889 if (cBB->succ_end() !=
890 std::find(cBB->succ_begin(),cBB->succ_end(), PHIBB)) {
891 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
893 PHI->addOperand(MachineOperand::CreateMBB(cBB));
898 SDL->BitTestCases.clear();
900 // If the JumpTable record is filled in, then we need to emit a jump table.
901 // Updating the PHI nodes is tricky in this case, since we need to determine
902 // whether the PHI is a successor of the range check MBB or the jump table MBB
903 for (unsigned i = 0, e = SDL->JTCases.size(); i != e; ++i) {
904 // Lower header first, if it wasn't already lowered
905 if (!SDL->JTCases[i].first.Emitted) {
906 // Set the current basic block to the mbb we wish to insert the code into
907 BB = SDL->JTCases[i].first.HeaderBB;
908 SDL->setCurrentBasicBlock(BB);
910 SDL->visitJumpTableHeader(SDL->JTCases[i].second, SDL->JTCases[i].first);
911 CurDAG->setRoot(SDL->getRoot());
916 // Set the current basic block to the mbb we wish to insert the code into
917 BB = SDL->JTCases[i].second.MBB;
918 SDL->setCurrentBasicBlock(BB);
920 SDL->visitJumpTable(SDL->JTCases[i].second);
921 CurDAG->setRoot(SDL->getRoot());
926 for (unsigned pi = 0, pe = SDL->PHINodesToUpdate.size(); pi != pe; ++pi) {
927 MachineInstr *PHI = SDL->PHINodesToUpdate[pi].first;
928 MachineBasicBlock *PHIBB = PHI->getParent();
929 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
930 "This is not a machine PHI node that we are updating!");
931 // "default" BB. We can go there only from header BB.
932 if (PHIBB == SDL->JTCases[i].second.Default) {
933 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
935 PHI->addOperand(MachineOperand::CreateMBB(SDL->JTCases[i].first.HeaderBB));
937 // JT BB. Just iterate over successors here
938 if (BB->succ_end() != std::find(BB->succ_begin(),BB->succ_end(), PHIBB)) {
939 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
941 PHI->addOperand(MachineOperand::CreateMBB(BB));
945 SDL->JTCases.clear();
947 // If the switch block involved a branch to one of the actual successors, we
948 // need to update PHI nodes in that block.
949 for (unsigned i = 0, e = SDL->PHINodesToUpdate.size(); i != e; ++i) {
950 MachineInstr *PHI = SDL->PHINodesToUpdate[i].first;
951 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
952 "This is not a machine PHI node that we are updating!");
953 if (BB->isSuccessor(PHI->getParent())) {
954 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[i].second,
956 PHI->addOperand(MachineOperand::CreateMBB(BB));
960 // If we generated any switch lowering information, build and codegen any
961 // additional DAGs necessary.
962 for (unsigned i = 0, e = SDL->SwitchCases.size(); i != e; ++i) {
963 // Set the current basic block to the mbb we wish to insert the code into
964 BB = SDL->SwitchCases[i].ThisBB;
965 SDL->setCurrentBasicBlock(BB);
968 SDL->visitSwitchCase(SDL->SwitchCases[i]);
969 CurDAG->setRoot(SDL->getRoot());
973 // Handle any PHI nodes in successors of this chunk, as if we were coming
974 // from the original BB before switch expansion. Note that PHI nodes can
975 // occur multiple times in PHINodesToUpdate. We have to be very careful to
976 // handle them the right number of times.
977 while ((BB = SDL->SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
978 for (MachineBasicBlock::iterator Phi = BB->begin();
979 Phi != BB->end() && Phi->getOpcode() == TargetInstrInfo::PHI; ++Phi){
980 // This value for this PHI node is recorded in PHINodesToUpdate, get it.
981 for (unsigned pn = 0; ; ++pn) {
982 assert(pn != SDL->PHINodesToUpdate.size() &&
983 "Didn't find PHI entry!");
984 if (SDL->PHINodesToUpdate[pn].first == Phi) {
985 Phi->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pn].
987 Phi->addOperand(MachineOperand::CreateMBB(SDL->SwitchCases[i].ThisBB));
993 // Don't process RHS if same block as LHS.
994 if (BB == SDL->SwitchCases[i].FalseBB)
995 SDL->SwitchCases[i].FalseBB = 0;
997 // If we haven't handled the RHS, do so now. Otherwise, we're done.
998 SDL->SwitchCases[i].TrueBB = SDL->SwitchCases[i].FalseBB;
999 SDL->SwitchCases[i].FalseBB = 0;
1001 assert(SDL->SwitchCases[i].TrueBB == 0 && SDL->SwitchCases[i].FalseBB == 0);
1003 SDL->SwitchCases.clear();
1005 SDL->PHINodesToUpdate.clear();
1009 /// Schedule - Pick a safe ordering for instructions for each
1010 /// target node in the graph.
1012 ScheduleDAG *SelectionDAGISel::Schedule() {
1013 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1017 RegisterScheduler::setDefault(Ctor);
1020 ScheduleDAG *Scheduler = Ctor(this, CurDAG, BB, Fast);
1027 HazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
1028 return new HazardRecognizer();
1031 //===----------------------------------------------------------------------===//
1032 // Helper functions used by the generated instruction selector.
1033 //===----------------------------------------------------------------------===//
1034 // Calls to these methods are generated by tblgen.
1036 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1037 /// the dag combiner simplified the 255, we still want to match. RHS is the
1038 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1039 /// specified in the .td file (e.g. 255).
1040 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1041 int64_t DesiredMaskS) const {
1042 const APInt &ActualMask = RHS->getAPIntValue();
1043 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1045 // If the actual mask exactly matches, success!
1046 if (ActualMask == DesiredMask)
1049 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1050 if (ActualMask.intersects(~DesiredMask))
1053 // Otherwise, the DAG Combiner may have proven that the value coming in is
1054 // either already zero or is not demanded. Check for known zero input bits.
1055 APInt NeededMask = DesiredMask & ~ActualMask;
1056 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1059 // TODO: check to see if missing bits are just not demanded.
1061 // Otherwise, this pattern doesn't match.
1065 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1066 /// the dag combiner simplified the 255, we still want to match. RHS is the
1067 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1068 /// specified in the .td file (e.g. 255).
1069 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1070 int64_t DesiredMaskS) const {
1071 const APInt &ActualMask = RHS->getAPIntValue();
1072 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1074 // If the actual mask exactly matches, success!
1075 if (ActualMask == DesiredMask)
1078 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1079 if (ActualMask.intersects(~DesiredMask))
1082 // Otherwise, the DAG Combiner may have proven that the value coming in is
1083 // either already zero or is not demanded. Check for known zero input bits.
1084 APInt NeededMask = DesiredMask & ~ActualMask;
1086 APInt KnownZero, KnownOne;
1087 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
1089 // If all the missing bits in the or are already known to be set, match!
1090 if ((NeededMask & KnownOne) == NeededMask)
1093 // TODO: check to see if missing bits are just not demanded.
1095 // Otherwise, this pattern doesn't match.
1100 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1101 /// by tblgen. Others should not call it.
1102 void SelectionDAGISel::
1103 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1104 std::vector<SDValue> InOps;
1105 std::swap(InOps, Ops);
1107 Ops.push_back(InOps[0]); // input chain.
1108 Ops.push_back(InOps[1]); // input asm string.
1110 unsigned i = 2, e = InOps.size();
1111 if (InOps[e-1].getValueType() == MVT::Flag)
1112 --e; // Don't process a flag operand if it is here.
1115 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getValue();
1116 if ((Flags & 7) != 4 /*MEM*/) {
1117 // Just skip over this operand, copying the operands verbatim.
1118 Ops.insert(Ops.end(), InOps.begin()+i, InOps.begin()+i+(Flags >> 3) + 1);
1119 i += (Flags >> 3) + 1;
1121 assert((Flags >> 3) == 1 && "Memory operand with multiple values?");
1122 // Otherwise, this is a memory operand. Ask the target to select it.
1123 std::vector<SDValue> SelOps;
1124 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps)) {
1125 cerr << "Could not match memory address. Inline asm failure!\n";
1129 // Add this to the output node.
1130 MVT IntPtrTy = CurDAG->getTargetLoweringInfo().getPointerTy();
1131 Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size() << 3),
1133 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1138 // Add the flag input back if present.
1139 if (e != InOps.size())
1140 Ops.push_back(InOps.back());
1143 char SelectionDAGISel::ID = 0;