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 "ScheduleDAGSDNodes.h"
16 #include "SelectionDAGBuild.h"
17 #include "llvm/CodeGen/SelectionDAGISel.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/CodeGen/FastISel.h"
29 #include "llvm/CodeGen/GCStrategy.h"
30 #include "llvm/CodeGen/GCMetadata.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineJumpTableInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
38 #include "llvm/CodeGen/SchedulerRegistry.h"
39 #include "llvm/CodeGen/SelectionDAG.h"
40 #include "llvm/CodeGen/DwarfWriter.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 DisableLegalizeTypes("disable-legalize-types", cl::Hidden);
58 EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
59 cl::desc("Enable verbose messages in the \"fast\" "
60 "instruction selector"));
62 EnableFastISelAbort("fast-isel-abort", cl::Hidden,
63 cl::desc("Enable abort calls when \"fast\" instruction fails"));
65 SchedLiveInCopies("schedule-livein-copies",
66 cl::desc("Schedule copies of livein registers"),
71 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
72 cl::desc("Pop up a window to show dags before the first "
75 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
76 cl::desc("Pop up a window to show dags before legalize types"));
78 ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
79 cl::desc("Pop up a window to show dags before legalize"));
81 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
82 cl::desc("Pop up a window to show dags before the second "
85 ViewDAGCombineLT("view-dag-combine-lt-dags", cl::Hidden,
86 cl::desc("Pop up a window to show dags before the post legalize types"
87 " dag combine pass"));
89 ViewISelDAGs("view-isel-dags", cl::Hidden,
90 cl::desc("Pop up a window to show isel dags as they are selected"));
92 ViewSchedDAGs("view-sched-dags", cl::Hidden,
93 cl::desc("Pop up a window to show sched dags as they are processed"));
95 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
96 cl::desc("Pop up a window to show SUnit dags after they are processed"));
98 static const bool ViewDAGCombine1 = false,
99 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false,
100 ViewDAGCombine2 = false,
101 ViewDAGCombineLT = false,
102 ViewISelDAGs = false, ViewSchedDAGs = false,
103 ViewSUnitDAGs = false;
106 //===---------------------------------------------------------------------===//
108 /// RegisterScheduler class - Track the registration of instruction schedulers.
110 //===---------------------------------------------------------------------===//
111 MachinePassRegistry RegisterScheduler::Registry;
113 //===---------------------------------------------------------------------===//
115 /// ISHeuristic command line option for instruction schedulers.
117 //===---------------------------------------------------------------------===//
118 static cl::opt<RegisterScheduler::FunctionPassCtor, false,
119 RegisterPassParser<RegisterScheduler> >
120 ISHeuristic("pre-RA-sched",
121 cl::init(&createDefaultScheduler),
122 cl::desc("Instruction schedulers available (before register"
125 static RegisterScheduler
126 defaultListDAGScheduler("default", "Best scheduler for the target",
127 createDefaultScheduler);
130 //===--------------------------------------------------------------------===//
131 /// createDefaultScheduler - This creates an instruction scheduler appropriate
133 ScheduleDAGSDNodes* createDefaultScheduler(SelectionDAGISel *IS,
134 CodeGenOpt::Level OptLevel) {
135 const TargetLowering &TLI = IS->getTargetLowering();
137 if (OptLevel == CodeGenOpt::None)
138 return createFastDAGScheduler(IS, OptLevel);
139 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency)
140 return createTDListDAGScheduler(IS, OptLevel);
141 assert(TLI.getSchedulingPreference() ==
142 TargetLowering::SchedulingForRegPressure && "Unknown sched type!");
143 return createBURRListDAGScheduler(IS, OptLevel);
147 // EmitInstrWithCustomInserter - This method should be implemented by targets
148 // that mark instructions with the 'usesCustomDAGSchedInserter' flag. These
149 // instructions are special in various ways, which require special support to
150 // insert. The specified MachineInstr is created but not inserted into any
151 // basic blocks, and the scheduler passes ownership of it to this method.
152 MachineBasicBlock *TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
153 MachineBasicBlock *MBB) const {
154 cerr << "If a target marks an instruction with "
155 << "'usesCustomDAGSchedInserter', it must implement "
156 << "TargetLowering::EmitInstrWithCustomInserter!\n";
161 /// EmitLiveInCopy - Emit a copy for a live in physical register. If the
162 /// physical register has only a single copy use, then coalesced the copy
164 static void EmitLiveInCopy(MachineBasicBlock *MBB,
165 MachineBasicBlock::iterator &InsertPos,
166 unsigned VirtReg, unsigned PhysReg,
167 const TargetRegisterClass *RC,
168 DenseMap<MachineInstr*, unsigned> &CopyRegMap,
169 const MachineRegisterInfo &MRI,
170 const TargetRegisterInfo &TRI,
171 const TargetInstrInfo &TII) {
172 unsigned NumUses = 0;
173 MachineInstr *UseMI = NULL;
174 for (MachineRegisterInfo::use_iterator UI = MRI.use_begin(VirtReg),
175 UE = MRI.use_end(); UI != UE; ++UI) {
181 // If the number of uses is not one, or the use is not a move instruction,
182 // don't coalesce. Also, only coalesce away a virtual register to virtual
184 bool Coalesced = false;
185 unsigned SrcReg, DstReg, SrcSubReg, DstSubReg;
187 TII.isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubReg, DstSubReg) &&
188 TargetRegisterInfo::isVirtualRegister(DstReg)) {
193 // Now find an ideal location to insert the copy.
194 MachineBasicBlock::iterator Pos = InsertPos;
195 while (Pos != MBB->begin()) {
196 MachineInstr *PrevMI = prior(Pos);
197 DenseMap<MachineInstr*, unsigned>::iterator RI = CopyRegMap.find(PrevMI);
198 // copyRegToReg might emit multiple instructions to do a copy.
199 unsigned CopyDstReg = (RI == CopyRegMap.end()) ? 0 : RI->second;
200 if (CopyDstReg && !TRI.regsOverlap(CopyDstReg, PhysReg))
201 // This is what the BB looks like right now:
206 // We want to insert "r1025 = mov r1". Inserting this copy below the
207 // move to r1024 makes it impossible for that move to be coalesced.
214 break; // Woot! Found a good location.
218 TII.copyRegToReg(*MBB, Pos, VirtReg, PhysReg, RC, RC);
219 CopyRegMap.insert(std::make_pair(prior(Pos), VirtReg));
221 if (&*InsertPos == UseMI) ++InsertPos;
226 /// EmitLiveInCopies - If this is the first basic block in the function,
227 /// and if it has live ins that need to be copied into vregs, emit the
228 /// copies into the block.
229 static void EmitLiveInCopies(MachineBasicBlock *EntryMBB,
230 const MachineRegisterInfo &MRI,
231 const TargetRegisterInfo &TRI,
232 const TargetInstrInfo &TII) {
233 if (SchedLiveInCopies) {
234 // Emit the copies at a heuristically-determined location in the block.
235 DenseMap<MachineInstr*, unsigned> CopyRegMap;
236 MachineBasicBlock::iterator InsertPos = EntryMBB->begin();
237 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
238 E = MRI.livein_end(); LI != E; ++LI)
240 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
241 EmitLiveInCopy(EntryMBB, InsertPos, LI->second, LI->first,
242 RC, CopyRegMap, MRI, TRI, TII);
245 // Emit the copies into the top of the block.
246 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
247 E = MRI.livein_end(); LI != E; ++LI)
249 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
250 TII.copyRegToReg(*EntryMBB, EntryMBB->begin(),
251 LI->second, LI->first, RC, RC);
256 //===----------------------------------------------------------------------===//
257 // SelectionDAGISel code
258 //===----------------------------------------------------------------------===//
260 SelectionDAGISel::SelectionDAGISel(TargetMachine &tm, CodeGenOpt::Level OL) :
261 FunctionPass(&ID), TM(tm), TLI(*tm.getTargetLowering()),
262 FuncInfo(new FunctionLoweringInfo(TLI)),
263 CurDAG(new SelectionDAG(TLI, *FuncInfo)),
264 SDL(new SelectionDAGLowering(*CurDAG, TLI, *FuncInfo, OL)),
270 SelectionDAGISel::~SelectionDAGISel() {
276 unsigned SelectionDAGISel::MakeReg(MVT VT) {
277 return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT));
280 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
281 AU.addRequired<AliasAnalysis>();
282 AU.addRequired<GCModuleInfo>();
283 AU.addRequired<DwarfWriter>();
284 AU.setPreservesAll();
287 bool SelectionDAGISel::runOnFunction(Function &Fn) {
288 // Do some sanity-checking on the command-line options.
289 assert((!EnableFastISelVerbose || EnableFastISel) &&
290 "-fast-isel-verbose requires -fast-isel");
291 assert((!EnableFastISelAbort || EnableFastISel) &&
292 "-fast-isel-abort requires -fast-isel");
294 // Do not codegen any 'available_externally' functions at all, they have
295 // definitions outside the translation unit.
296 if (Fn.hasAvailableExternallyLinkage())
300 // Get alias analysis for load/store combining.
301 AA = &getAnalysis<AliasAnalysis>();
303 TargetMachine &TM = TLI.getTargetMachine();
304 MF = &MachineFunction::construct(&Fn, TM);
305 const TargetInstrInfo &TII = *TM.getInstrInfo();
306 const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
308 if (MF->getFunction()->hasGC())
309 GFI = &getAnalysis<GCModuleInfo>().getFunctionInfo(*MF->getFunction());
312 RegInfo = &MF->getRegInfo();
313 DOUT << "\n\n\n=== " << Fn.getName() << "\n";
315 MachineModuleInfo *MMI = getAnalysisIfAvailable<MachineModuleInfo>();
316 DwarfWriter *DW = getAnalysisIfAvailable<DwarfWriter>();
317 CurDAG->init(*MF, MMI, DW);
318 FuncInfo->set(Fn, *MF, *CurDAG, EnableFastISel);
321 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
322 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(I->getTerminator()))
324 FuncInfo->MBBMap[Invoke->getSuccessor(1)]->setIsLandingPad();
326 SelectAllBasicBlocks(Fn, *MF, MMI, DW, TII);
328 // If the first basic block in the function has live ins that need to be
329 // copied into vregs, emit the copies into the top of the block before
330 // emitting the code for the block.
331 EmitLiveInCopies(MF->begin(), *RegInfo, TRI, TII);
333 // Add function live-ins to entry block live-in set.
334 for (MachineRegisterInfo::livein_iterator I = RegInfo->livein_begin(),
335 E = RegInfo->livein_end(); I != E; ++I)
336 MF->begin()->addLiveIn(I->first);
339 assert(FuncInfo->CatchInfoFound.size() == FuncInfo->CatchInfoLost.size() &&
340 "Not all catch info was assigned to a landing pad!");
348 static void copyCatchInfo(BasicBlock *SrcBB, BasicBlock *DestBB,
349 MachineModuleInfo *MMI, FunctionLoweringInfo &FLI) {
350 for (BasicBlock::iterator I = SrcBB->begin(), E = --SrcBB->end(); I != E; ++I)
351 if (EHSelectorInst *EHSel = dyn_cast<EHSelectorInst>(I)) {
352 // Apply the catch info to DestBB.
353 AddCatchInfo(*EHSel, MMI, FLI.MBBMap[DestBB]);
355 if (!FLI.MBBMap[SrcBB]->isLandingPad())
356 FLI.CatchInfoFound.insert(EHSel);
361 /// IsFixedFrameObjectWithPosOffset - Check if object is a fixed frame object and
362 /// whether object offset >= 0.
364 IsFixedFrameObjectWithPosOffset(MachineFrameInfo *MFI, SDValue Op) {
365 if (!isa<FrameIndexSDNode>(Op)) return false;
367 FrameIndexSDNode * FrameIdxNode = dyn_cast<FrameIndexSDNode>(Op);
368 int FrameIdx = FrameIdxNode->getIndex();
369 return MFI->isFixedObjectIndex(FrameIdx) &&
370 MFI->getObjectOffset(FrameIdx) >= 0;
373 /// IsPossiblyOverwrittenArgumentOfTailCall - Check if the operand could
374 /// possibly be overwritten when lowering the outgoing arguments in a tail
375 /// call. Currently the implementation of this call is very conservative and
376 /// assumes all arguments sourcing from FORMAL_ARGUMENTS or a CopyFromReg with
377 /// virtual registers would be overwritten by direct lowering.
378 static bool IsPossiblyOverwrittenArgumentOfTailCall(SDValue Op,
379 MachineFrameInfo *MFI) {
380 RegisterSDNode * OpReg = NULL;
381 if (Op.getOpcode() == ISD::FORMAL_ARGUMENTS ||
382 (Op.getOpcode()== ISD::CopyFromReg &&
383 (OpReg = dyn_cast<RegisterSDNode>(Op.getOperand(1))) &&
384 (OpReg->getReg() >= TargetRegisterInfo::FirstVirtualRegister)) ||
385 (Op.getOpcode() == ISD::LOAD &&
386 IsFixedFrameObjectWithPosOffset(MFI, Op.getOperand(1))) ||
387 (Op.getOpcode() == ISD::MERGE_VALUES &&
388 Op.getOperand(Op.getResNo()).getOpcode() == ISD::LOAD &&
389 IsFixedFrameObjectWithPosOffset(MFI, Op.getOperand(Op.getResNo()).
395 /// CheckDAGForTailCallsAndFixThem - This Function looks for CALL nodes in the
396 /// DAG and fixes their tailcall attribute operand.
397 static void CheckDAGForTailCallsAndFixThem(SelectionDAG &DAG,
398 const TargetLowering& TLI) {
400 SDValue Terminator = DAG.getRoot();
403 if (Terminator.getOpcode() == ISD::RET) {
404 Ret = Terminator.getNode();
407 // Fix tail call attribute of CALL nodes.
408 for (SelectionDAG::allnodes_iterator BE = DAG.allnodes_begin(),
409 BI = DAG.allnodes_end(); BI != BE; ) {
411 if (CallSDNode *TheCall = dyn_cast<CallSDNode>(BI)) {
412 SDValue OpRet(Ret, 0);
413 SDValue OpCall(BI, 0);
414 bool isMarkedTailCall = TheCall->isTailCall();
415 // If CALL node has tail call attribute set to true and the call is not
416 // eligible (no RET or the target rejects) the attribute is fixed to
417 // false. The TargetLowering::IsEligibleForTailCallOptimization function
418 // must correctly identify tail call optimizable calls.
419 if (!isMarkedTailCall) continue;
421 !TLI.IsEligibleForTailCallOptimization(TheCall, OpRet, DAG)) {
422 // Not eligible. Mark CALL node as non tail call. Note that we
423 // can modify the call node in place since calls are not CSE'd.
424 TheCall->setNotTailCall();
426 // Look for tail call clobbered arguments. Emit a series of
427 // copyto/copyfrom virtual register nodes to protect them.
428 SmallVector<SDValue, 32> Ops;
429 SDValue Chain = TheCall->getChain(), InFlag;
430 Ops.push_back(Chain);
431 Ops.push_back(TheCall->getCallee());
432 for (unsigned i = 0, e = TheCall->getNumArgs(); i != e; ++i) {
433 SDValue Arg = TheCall->getArg(i);
434 bool isByVal = TheCall->getArgFlags(i).isByVal();
435 MachineFunction &MF = DAG.getMachineFunction();
436 MachineFrameInfo *MFI = MF.getFrameInfo();
438 IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)) {
439 MVT VT = Arg.getValueType();
440 unsigned VReg = MF.getRegInfo().
441 createVirtualRegister(TLI.getRegClassFor(VT));
442 Chain = DAG.getCopyToReg(Chain, Arg.getDebugLoc(),
444 InFlag = Chain.getValue(1);
445 Arg = DAG.getCopyFromReg(Chain, Arg.getDebugLoc(),
447 Chain = Arg.getValue(1);
448 InFlag = Arg.getValue(2);
451 Ops.push_back(TheCall->getArgFlagsVal(i));
453 // Link in chain of CopyTo/CopyFromReg.
455 DAG.UpdateNodeOperands(OpCall, Ops.begin(), Ops.size());
461 void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB,
462 BasicBlock::iterator Begin,
463 BasicBlock::iterator End) {
464 SDL->setCurrentBasicBlock(BB);
466 // Lower all of the non-terminator instructions.
467 for (BasicBlock::iterator I = Begin; I != End; ++I)
468 if (!isa<TerminatorInst>(I))
471 // Ensure that all instructions which are used outside of their defining
472 // blocks are available as virtual registers. Invoke is handled elsewhere.
473 for (BasicBlock::iterator I = Begin; I != End; ++I)
474 if (!isa<PHINode>(I) && !isa<InvokeInst>(I))
475 SDL->CopyToExportRegsIfNeeded(I);
477 // Handle PHI nodes in successor blocks.
478 if (End == LLVMBB->end()) {
479 HandlePHINodesInSuccessorBlocks(LLVMBB);
481 // Lower the terminator after the copies are emitted.
482 SDL->visit(*LLVMBB->getTerminator());
485 // Make sure the root of the DAG is up-to-date.
486 CurDAG->setRoot(SDL->getControlRoot());
488 // Check whether calls in this block are real tail calls. Fix up CALL nodes
489 // with correct tailcall attribute so that the target can rely on the tailcall
490 // attribute indicating whether the call is really eligible for tail call
492 if (PerformTailCallOpt)
493 CheckDAGForTailCallsAndFixThem(*CurDAG, TLI);
495 // Final step, emit the lowered DAG as machine code.
500 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
501 SmallPtrSet<SDNode*, 128> VisitedNodes;
502 SmallVector<SDNode*, 128> Worklist;
504 Worklist.push_back(CurDAG->getRoot().getNode());
510 while (!Worklist.empty()) {
511 SDNode *N = Worklist.back();
514 // If we've already seen this node, ignore it.
515 if (!VisitedNodes.insert(N))
518 // Otherwise, add all chain operands to the worklist.
519 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
520 if (N->getOperand(i).getValueType() == MVT::Other)
521 Worklist.push_back(N->getOperand(i).getNode());
523 // If this is a CopyToReg with a vreg dest, process it.
524 if (N->getOpcode() != ISD::CopyToReg)
527 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
528 if (!TargetRegisterInfo::isVirtualRegister(DestReg))
531 // Ignore non-scalar or non-integer values.
532 SDValue Src = N->getOperand(2);
533 MVT SrcVT = Src.getValueType();
534 if (!SrcVT.isInteger() || SrcVT.isVector())
537 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
538 Mask = APInt::getAllOnesValue(SrcVT.getSizeInBits());
539 CurDAG->ComputeMaskedBits(Src, Mask, KnownZero, KnownOne);
541 // Only install this information if it tells us something.
542 if (NumSignBits != 1 || KnownZero != 0 || KnownOne != 0) {
543 DestReg -= TargetRegisterInfo::FirstVirtualRegister;
544 FunctionLoweringInfo &FLI = CurDAG->getFunctionLoweringInfo();
545 if (DestReg >= FLI.LiveOutRegInfo.size())
546 FLI.LiveOutRegInfo.resize(DestReg+1);
547 FunctionLoweringInfo::LiveOutInfo &LOI = FLI.LiveOutRegInfo[DestReg];
548 LOI.NumSignBits = NumSignBits;
549 LOI.KnownOne = KnownOne;
550 LOI.KnownZero = KnownZero;
555 void SelectionDAGISel::CodeGenAndEmitDAG() {
556 std::string GroupName;
557 if (TimePassesIsEnabled)
558 GroupName = "Instruction Selection and Scheduling";
559 std::string BlockName;
560 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs ||
561 ViewDAGCombine2 || ViewDAGCombineLT || ViewISelDAGs || ViewSchedDAGs ||
563 BlockName = CurDAG->getMachineFunction().getFunction()->getName() + ':' +
564 BB->getBasicBlock()->getName();
566 DOUT << "Initial selection DAG:\n";
567 DEBUG(CurDAG->dump());
569 if (ViewDAGCombine1) CurDAG->viewGraph("dag-combine1 input for " + BlockName);
571 // Run the DAG combiner in pre-legalize mode.
572 if (TimePassesIsEnabled) {
573 NamedRegionTimer T("DAG Combining 1", GroupName);
574 CurDAG->Combine(Unrestricted, *AA, OptLevel);
576 CurDAG->Combine(Unrestricted, *AA, OptLevel);
579 DOUT << "Optimized lowered selection DAG:\n";
580 DEBUG(CurDAG->dump());
582 // Second step, hack on the DAG until it only uses operations and types that
583 // the target supports.
584 if (!DisableLegalizeTypes) {
585 if (ViewLegalizeTypesDAGs) CurDAG->viewGraph("legalize-types input for " +
589 if (TimePassesIsEnabled) {
590 NamedRegionTimer T("Type Legalization", GroupName);
591 Changed = CurDAG->LegalizeTypes();
593 Changed = CurDAG->LegalizeTypes();
596 DOUT << "Type-legalized selection DAG:\n";
597 DEBUG(CurDAG->dump());
600 if (ViewDAGCombineLT)
601 CurDAG->viewGraph("dag-combine-lt input for " + BlockName);
603 // Run the DAG combiner in post-type-legalize mode.
604 if (TimePassesIsEnabled) {
605 NamedRegionTimer T("DAG Combining after legalize types", GroupName);
606 CurDAG->Combine(NoIllegalTypes, *AA, OptLevel);
608 CurDAG->Combine(NoIllegalTypes, *AA, OptLevel);
611 DOUT << "Optimized type-legalized selection DAG:\n";
612 DEBUG(CurDAG->dump());
616 if (ViewLegalizeDAGs) CurDAG->viewGraph("legalize input for " + BlockName);
618 if (TimePassesIsEnabled) {
619 NamedRegionTimer T("DAG Legalization", GroupName);
620 CurDAG->Legalize(DisableLegalizeTypes, OptLevel);
622 CurDAG->Legalize(DisableLegalizeTypes, OptLevel);
625 DOUT << "Legalized selection DAG:\n";
626 DEBUG(CurDAG->dump());
628 if (ViewDAGCombine2) CurDAG->viewGraph("dag-combine2 input for " + BlockName);
630 // Run the DAG combiner in post-legalize mode.
631 if (TimePassesIsEnabled) {
632 NamedRegionTimer T("DAG Combining 2", GroupName);
633 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
635 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel);
638 DOUT << "Optimized legalized selection DAG:\n";
639 DEBUG(CurDAG->dump());
641 if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName);
643 if (OptLevel != CodeGenOpt::None)
644 ComputeLiveOutVRegInfo();
646 // Third, instruction select all of the operations to machine code, adding the
647 // code to the MachineBasicBlock.
648 if (TimePassesIsEnabled) {
649 NamedRegionTimer T("Instruction Selection", GroupName);
655 DOUT << "Selected selection DAG:\n";
656 DEBUG(CurDAG->dump());
658 if (ViewSchedDAGs) CurDAG->viewGraph("scheduler input for " + BlockName);
660 // Schedule machine code.
661 ScheduleDAGSDNodes *Scheduler = CreateScheduler();
662 if (TimePassesIsEnabled) {
663 NamedRegionTimer T("Instruction Scheduling", GroupName);
664 Scheduler->Run(CurDAG, BB, BB->end());
666 Scheduler->Run(CurDAG, BB, BB->end());
669 if (ViewSUnitDAGs) Scheduler->viewGraph();
671 // Emit machine code to BB. This can change 'BB' to the last block being
673 if (TimePassesIsEnabled) {
674 NamedRegionTimer T("Instruction Creation", GroupName);
675 BB = Scheduler->EmitSchedule();
677 BB = Scheduler->EmitSchedule();
680 // Free the scheduler state.
681 if (TimePassesIsEnabled) {
682 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName);
688 DOUT << "Selected machine code:\n";
692 void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn,
694 MachineModuleInfo *MMI,
696 const TargetInstrInfo &TII) {
697 // Initialize the Fast-ISel state, if needed.
698 FastISel *FastIS = 0;
700 FastIS = TLI.createFastISel(MF, MMI, DW,
703 FuncInfo->StaticAllocaMap
705 , FuncInfo->CatchInfoLost
709 // Iterate over all basic blocks in the function.
710 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
711 BasicBlock *LLVMBB = &*I;
712 BB = FuncInfo->MBBMap[LLVMBB];
714 BasicBlock::iterator const Begin = LLVMBB->begin();
715 BasicBlock::iterator const End = LLVMBB->end();
716 BasicBlock::iterator BI = Begin;
718 // Lower any arguments needed in this block if this is the entry block.
719 bool SuppressFastISel = false;
720 if (LLVMBB == &Fn.getEntryBlock()) {
721 LowerArguments(LLVMBB);
723 // If any of the arguments has the byval attribute, forgo
724 // fast-isel in the entry block.
727 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end();
729 if (Fn.paramHasAttr(j, Attribute::ByVal)) {
730 if (EnableFastISelVerbose || EnableFastISelAbort)
731 cerr << "FastISel skips entry block due to byval argument\n";
732 SuppressFastISel = true;
738 if (MMI && BB->isLandingPad()) {
739 // Add a label to mark the beginning of the landing pad. Deletion of the
740 // landing pad can thus be detected via the MachineModuleInfo.
741 unsigned LabelID = MMI->addLandingPad(BB);
743 const TargetInstrDesc &II = TII.get(TargetInstrInfo::EH_LABEL);
744 BuildMI(BB, SDL->getCurDebugLoc(), II).addImm(LabelID);
746 // Mark exception register as live in.
747 unsigned Reg = TLI.getExceptionAddressRegister();
748 if (Reg) BB->addLiveIn(Reg);
750 // Mark exception selector register as live in.
751 Reg = TLI.getExceptionSelectorRegister();
752 if (Reg) BB->addLiveIn(Reg);
754 // FIXME: Hack around an exception handling flaw (PR1508): the personality
755 // function and list of typeids logically belong to the invoke (or, if you
756 // like, the basic block containing the invoke), and need to be associated
757 // with it in the dwarf exception handling tables. Currently however the
758 // information is provided by an intrinsic (eh.selector) that can be moved
759 // to unexpected places by the optimizers: if the unwind edge is critical,
760 // then breaking it can result in the intrinsics being in the successor of
761 // the landing pad, not the landing pad itself. This results in exceptions
762 // not being caught because no typeids are associated with the invoke.
763 // This may not be the only way things can go wrong, but it is the only way
764 // we try to work around for the moment.
765 BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
767 if (Br && Br->isUnconditional()) { // Critical edge?
768 BasicBlock::iterator I, E;
769 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
770 if (isa<EHSelectorInst>(I))
774 // No catch info found - try to extract some from the successor.
775 copyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, *FuncInfo);
779 // Before doing SelectionDAG ISel, see if FastISel has been requested.
780 if (FastIS && !SuppressFastISel) {
781 // Emit code for any incoming arguments. This must happen before
782 // beginning FastISel on the entry block.
783 if (LLVMBB == &Fn.getEntryBlock()) {
784 CurDAG->setRoot(SDL->getControlRoot());
788 FastIS->startNewBlock(BB);
789 // Do FastISel on as many instructions as possible.
790 for (; BI != End; ++BI) {
791 // Just before the terminator instruction, insert instructions to
792 // feed PHI nodes in successor blocks.
793 if (isa<TerminatorInst>(BI))
794 if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, FastIS)) {
795 if (EnableFastISelVerbose || EnableFastISelAbort) {
796 cerr << "FastISel miss: ";
799 if (EnableFastISelAbort)
800 assert(0 && "FastISel didn't handle a PHI in a successor");
804 // First try normal tablegen-generated "fast" selection.
805 if (FastIS->SelectInstruction(BI))
808 // Next, try calling the target to attempt to handle the instruction.
809 if (FastIS->TargetSelectInstruction(BI))
812 // Then handle certain instructions as single-LLVM-Instruction blocks.
813 if (isa<CallInst>(BI)) {
814 if (EnableFastISelVerbose || EnableFastISelAbort) {
815 cerr << "FastISel missed call: ";
819 if (BI->getType() != Type::VoidTy) {
820 unsigned &R = FuncInfo->ValueMap[BI];
822 R = FuncInfo->CreateRegForValue(BI);
825 SDL->setCurDebugLoc(FastIS->getCurDebugLoc());
826 SelectBasicBlock(LLVMBB, BI, next(BI));
827 // If the instruction was codegen'd with multiple blocks,
828 // inform the FastISel object where to resume inserting.
829 FastIS->setCurrentBlock(BB);
833 // Otherwise, give up on FastISel for the rest of the block.
834 // For now, be a little lenient about non-branch terminators.
835 if (!isa<TerminatorInst>(BI) || isa<BranchInst>(BI)) {
836 if (EnableFastISelVerbose || EnableFastISelAbort) {
837 cerr << "FastISel miss: ";
840 if (EnableFastISelAbort)
841 // The "fast" selector couldn't handle something and bailed.
842 // For the purpose of debugging, just abort.
843 assert(0 && "FastISel didn't select the entire block");
849 // Run SelectionDAG instruction selection on the remainder of the block
850 // not handled by FastISel. If FastISel is not run, this is the entire
853 // If FastISel is run and it has known DebugLoc then use it.
854 if (FastIS && !FastIS->getCurDebugLoc().isUnknown())
855 SDL->setCurDebugLoc(FastIS->getCurDebugLoc());
856 SelectBasicBlock(LLVMBB, BI, End);
866 SelectionDAGISel::FinishBasicBlock() {
868 DOUT << "Target-post-processed machine code:\n";
871 DOUT << "Total amount of phi nodes to update: "
872 << SDL->PHINodesToUpdate.size() << "\n";
873 DEBUG(for (unsigned i = 0, e = SDL->PHINodesToUpdate.size(); i != e; ++i)
874 DOUT << "Node " << i << " : (" << SDL->PHINodesToUpdate[i].first
875 << ", " << SDL->PHINodesToUpdate[i].second << ")\n";);
877 // Next, now that we know what the last MBB the LLVM BB expanded is, update
878 // PHI nodes in successors.
879 if (SDL->SwitchCases.empty() &&
880 SDL->JTCases.empty() &&
881 SDL->BitTestCases.empty()) {
882 for (unsigned i = 0, e = SDL->PHINodesToUpdate.size(); i != e; ++i) {
883 MachineInstr *PHI = SDL->PHINodesToUpdate[i].first;
884 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
885 "This is not a machine PHI node that we are updating!");
886 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[i].second,
888 PHI->addOperand(MachineOperand::CreateMBB(BB));
890 SDL->PHINodesToUpdate.clear();
894 for (unsigned i = 0, e = SDL->BitTestCases.size(); i != e; ++i) {
895 // Lower header first, if it wasn't already lowered
896 if (!SDL->BitTestCases[i].Emitted) {
897 // Set the current basic block to the mbb we wish to insert the code into
898 BB = SDL->BitTestCases[i].Parent;
899 SDL->setCurrentBasicBlock(BB);
901 SDL->visitBitTestHeader(SDL->BitTestCases[i]);
902 CurDAG->setRoot(SDL->getRoot());
907 for (unsigned j = 0, ej = SDL->BitTestCases[i].Cases.size(); j != ej; ++j) {
908 // Set the current basic block to the mbb we wish to insert the code into
909 BB = SDL->BitTestCases[i].Cases[j].ThisBB;
910 SDL->setCurrentBasicBlock(BB);
913 SDL->visitBitTestCase(SDL->BitTestCases[i].Cases[j+1].ThisBB,
914 SDL->BitTestCases[i].Reg,
915 SDL->BitTestCases[i].Cases[j]);
917 SDL->visitBitTestCase(SDL->BitTestCases[i].Default,
918 SDL->BitTestCases[i].Reg,
919 SDL->BitTestCases[i].Cases[j]);
922 CurDAG->setRoot(SDL->getRoot());
928 for (unsigned pi = 0, pe = SDL->PHINodesToUpdate.size(); pi != pe; ++pi) {
929 MachineInstr *PHI = SDL->PHINodesToUpdate[pi].first;
930 MachineBasicBlock *PHIBB = PHI->getParent();
931 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
932 "This is not a machine PHI node that we are updating!");
933 // This is "default" BB. We have two jumps to it. From "header" BB and
934 // from last "case" BB.
935 if (PHIBB == SDL->BitTestCases[i].Default) {
936 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
938 PHI->addOperand(MachineOperand::CreateMBB(SDL->BitTestCases[i].Parent));
939 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
941 PHI->addOperand(MachineOperand::CreateMBB(SDL->BitTestCases[i].Cases.
944 // One of "cases" BB.
945 for (unsigned j = 0, ej = SDL->BitTestCases[i].Cases.size();
947 MachineBasicBlock* cBB = SDL->BitTestCases[i].Cases[j].ThisBB;
948 if (cBB->succ_end() !=
949 std::find(cBB->succ_begin(),cBB->succ_end(), PHIBB)) {
950 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
952 PHI->addOperand(MachineOperand::CreateMBB(cBB));
957 SDL->BitTestCases.clear();
959 // If the JumpTable record is filled in, then we need to emit a jump table.
960 // Updating the PHI nodes is tricky in this case, since we need to determine
961 // whether the PHI is a successor of the range check MBB or the jump table MBB
962 for (unsigned i = 0, e = SDL->JTCases.size(); i != e; ++i) {
963 // Lower header first, if it wasn't already lowered
964 if (!SDL->JTCases[i].first.Emitted) {
965 // Set the current basic block to the mbb we wish to insert the code into
966 BB = SDL->JTCases[i].first.HeaderBB;
967 SDL->setCurrentBasicBlock(BB);
969 SDL->visitJumpTableHeader(SDL->JTCases[i].second, SDL->JTCases[i].first);
970 CurDAG->setRoot(SDL->getRoot());
975 // Set the current basic block to the mbb we wish to insert the code into
976 BB = SDL->JTCases[i].second.MBB;
977 SDL->setCurrentBasicBlock(BB);
979 SDL->visitJumpTable(SDL->JTCases[i].second);
980 CurDAG->setRoot(SDL->getRoot());
985 for (unsigned pi = 0, pe = SDL->PHINodesToUpdate.size(); pi != pe; ++pi) {
986 MachineInstr *PHI = SDL->PHINodesToUpdate[pi].first;
987 MachineBasicBlock *PHIBB = PHI->getParent();
988 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
989 "This is not a machine PHI node that we are updating!");
990 // "default" BB. We can go there only from header BB.
991 if (PHIBB == SDL->JTCases[i].second.Default) {
992 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
994 PHI->addOperand(MachineOperand::CreateMBB(SDL->JTCases[i].first.HeaderBB));
996 // JT BB. Just iterate over successors here
997 if (BB->succ_end() != std::find(BB->succ_begin(),BB->succ_end(), PHIBB)) {
998 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
1000 PHI->addOperand(MachineOperand::CreateMBB(BB));
1004 SDL->JTCases.clear();
1006 // If the switch block involved a branch to one of the actual successors, we
1007 // need to update PHI nodes in that block.
1008 for (unsigned i = 0, e = SDL->PHINodesToUpdate.size(); i != e; ++i) {
1009 MachineInstr *PHI = SDL->PHINodesToUpdate[i].first;
1010 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
1011 "This is not a machine PHI node that we are updating!");
1012 if (BB->isSuccessor(PHI->getParent())) {
1013 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[i].second,
1015 PHI->addOperand(MachineOperand::CreateMBB(BB));
1019 // If we generated any switch lowering information, build and codegen any
1020 // additional DAGs necessary.
1021 for (unsigned i = 0, e = SDL->SwitchCases.size(); i != e; ++i) {
1022 // Set the current basic block to the mbb we wish to insert the code into
1023 BB = SDL->SwitchCases[i].ThisBB;
1024 SDL->setCurrentBasicBlock(BB);
1027 SDL->visitSwitchCase(SDL->SwitchCases[i]);
1028 CurDAG->setRoot(SDL->getRoot());
1029 CodeGenAndEmitDAG();
1032 // Handle any PHI nodes in successors of this chunk, as if we were coming
1033 // from the original BB before switch expansion. Note that PHI nodes can
1034 // occur multiple times in PHINodesToUpdate. We have to be very careful to
1035 // handle them the right number of times.
1036 while ((BB = SDL->SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
1037 for (MachineBasicBlock::iterator Phi = BB->begin();
1038 Phi != BB->end() && Phi->getOpcode() == TargetInstrInfo::PHI; ++Phi){
1039 // This value for this PHI node is recorded in PHINodesToUpdate, get it.
1040 for (unsigned pn = 0; ; ++pn) {
1041 assert(pn != SDL->PHINodesToUpdate.size() &&
1042 "Didn't find PHI entry!");
1043 if (SDL->PHINodesToUpdate[pn].first == Phi) {
1044 Phi->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pn].
1046 Phi->addOperand(MachineOperand::CreateMBB(SDL->SwitchCases[i].ThisBB));
1052 // Don't process RHS if same block as LHS.
1053 if (BB == SDL->SwitchCases[i].FalseBB)
1054 SDL->SwitchCases[i].FalseBB = 0;
1056 // If we haven't handled the RHS, do so now. Otherwise, we're done.
1057 SDL->SwitchCases[i].TrueBB = SDL->SwitchCases[i].FalseBB;
1058 SDL->SwitchCases[i].FalseBB = 0;
1060 assert(SDL->SwitchCases[i].TrueBB == 0 && SDL->SwitchCases[i].FalseBB == 0);
1062 SDL->SwitchCases.clear();
1064 SDL->PHINodesToUpdate.clear();
1068 /// Create the scheduler. If a specific scheduler was specified
1069 /// via the SchedulerRegistry, use it, otherwise select the
1070 /// one preferred by the target.
1072 ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() {
1073 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1077 RegisterScheduler::setDefault(Ctor);
1080 return Ctor(this, OptLevel);
1083 ScheduleHazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
1084 return new ScheduleHazardRecognizer();
1087 //===----------------------------------------------------------------------===//
1088 // Helper functions used by the generated instruction selector.
1089 //===----------------------------------------------------------------------===//
1090 // Calls to these methods are generated by tblgen.
1092 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1093 /// the dag combiner simplified the 255, we still want to match. RHS is the
1094 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1095 /// specified in the .td file (e.g. 255).
1096 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1097 int64_t DesiredMaskS) const {
1098 const APInt &ActualMask = RHS->getAPIntValue();
1099 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1101 // If the actual mask exactly matches, success!
1102 if (ActualMask == DesiredMask)
1105 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1106 if (ActualMask.intersects(~DesiredMask))
1109 // Otherwise, the DAG Combiner may have proven that the value coming in is
1110 // either already zero or is not demanded. Check for known zero input bits.
1111 APInt NeededMask = DesiredMask & ~ActualMask;
1112 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1115 // TODO: check to see if missing bits are just not demanded.
1117 // Otherwise, this pattern doesn't match.
1121 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1122 /// the dag combiner simplified the 255, we still want to match. RHS is the
1123 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1124 /// specified in the .td file (e.g. 255).
1125 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1126 int64_t DesiredMaskS) const {
1127 const APInt &ActualMask = RHS->getAPIntValue();
1128 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1130 // If the actual mask exactly matches, success!
1131 if (ActualMask == DesiredMask)
1134 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1135 if (ActualMask.intersects(~DesiredMask))
1138 // Otherwise, the DAG Combiner may have proven that the value coming in is
1139 // either already zero or is not demanded. Check for known zero input bits.
1140 APInt NeededMask = DesiredMask & ~ActualMask;
1142 APInt KnownZero, KnownOne;
1143 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
1145 // If all the missing bits in the or are already known to be set, match!
1146 if ((NeededMask & KnownOne) == NeededMask)
1149 // TODO: check to see if missing bits are just not demanded.
1151 // Otherwise, this pattern doesn't match.
1156 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1157 /// by tblgen. Others should not call it.
1158 void SelectionDAGISel::
1159 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1160 std::vector<SDValue> InOps;
1161 std::swap(InOps, Ops);
1163 Ops.push_back(InOps[0]); // input chain.
1164 Ops.push_back(InOps[1]); // input asm string.
1166 unsigned i = 2, e = InOps.size();
1167 if (InOps[e-1].getValueType() == MVT::Flag)
1168 --e; // Don't process a flag operand if it is here.
1171 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1172 if ((Flags & 7) != 4 /*MEM*/) {
1173 // Just skip over this operand, copying the operands verbatim.
1174 Ops.insert(Ops.end(), InOps.begin()+i,
1175 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1);
1176 i += InlineAsm::getNumOperandRegisters(Flags) + 1;
1178 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 &&
1179 "Memory operand with multiple values?");
1180 // Otherwise, this is a memory operand. Ask the target to select it.
1181 std::vector<SDValue> SelOps;
1182 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps)) {
1183 cerr << "Could not match memory address. Inline asm failure!\n";
1187 // Add this to the output node.
1188 MVT IntPtrTy = CurDAG->getTargetLoweringInfo().getPointerTy();
1189 Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size()<< 3),
1191 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1196 // Add the flag input back if present.
1197 if (e != InOps.size())
1198 Ops.push_back(InOps.back());
1201 /// findFlagUse - Return use of MVT::Flag value produced by the specified
1204 static SDNode *findFlagUse(SDNode *N) {
1205 unsigned FlagResNo = N->getNumValues()-1;
1206 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1207 SDUse &Use = I.getUse();
1208 if (Use.getResNo() == FlagResNo)
1209 return Use.getUser();
1214 /// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
1215 /// This function recursively traverses up the operand chain, ignoring
1217 static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
1219 SmallPtrSet<SDNode*, 16> &Visited) {
1220 if (Use->getNodeId() < Def->getNodeId() ||
1221 !Visited.insert(Use))
1224 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
1225 SDNode *N = Use->getOperand(i).getNode();
1227 if (Use == ImmedUse || Use == Root)
1228 continue; // We are not looking for immediate use.
1233 // Traverse up the operand chain.
1234 if (findNonImmUse(N, Def, ImmedUse, Root, Visited))
1240 /// isNonImmUse - Start searching from Root up the DAG to check is Def can
1241 /// be reached. Return true if that's the case. However, ignore direct uses
1242 /// by ImmedUse (which would be U in the example illustrated in
1243 /// IsLegalAndProfitableToFold) and by Root (which can happen in the store
1245 /// FIXME: to be really generic, we should allow direct use by any node
1246 /// that is being folded. But realisticly since we only fold loads which
1247 /// have one non-chain use, we only need to watch out for load/op/store
1248 /// and load/op/cmp case where the root (store / cmp) may reach the load via
1249 /// its chain operand.
1250 static inline bool isNonImmUse(SDNode *Root, SDNode *Def, SDNode *ImmedUse) {
1251 SmallPtrSet<SDNode*, 16> Visited;
1252 return findNonImmUse(Root, Def, ImmedUse, Root, Visited);
1255 /// IsLegalAndProfitableToFold - Returns true if the specific operand node N of
1256 /// U can be folded during instruction selection that starts at Root and
1257 /// folding N is profitable.
1258 bool SelectionDAGISel::IsLegalAndProfitableToFold(SDNode *N, SDNode *U,
1259 SDNode *Root) const {
1260 if (OptLevel == CodeGenOpt::None) return false;
1262 // If Root use can somehow reach N through a path that that doesn't contain
1263 // U then folding N would create a cycle. e.g. In the following
1264 // diagram, Root can reach N through X. If N is folded into into Root, then
1265 // X is both a predecessor and a successor of U.
1276 // * indicates nodes to be folded together.
1278 // If Root produces a flag, then it gets (even more) interesting. Since it
1279 // will be "glued" together with its flag use in the scheduler, we need to
1280 // check if it might reach N.
1299 // If FU (flag use) indirectly reaches N (the load), and Root folds N
1300 // (call it Fold), then X is a predecessor of FU and a successor of
1301 // Fold. But since Fold and FU are flagged together, this will create
1302 // a cycle in the scheduling graph.
1304 MVT VT = Root->getValueType(Root->getNumValues()-1);
1305 while (VT == MVT::Flag) {
1306 SDNode *FU = findFlagUse(Root);
1310 VT = Root->getValueType(Root->getNumValues()-1);
1313 return !isNonImmUse(Root, N, U);
1317 char SelectionDAGISel::ID = 0;