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 EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
64 cl::desc("Enable verbose messages in the experimental \"fast\" "
65 "instruction selector"));
67 EnableFastISelAbort("fast-isel-abort", cl::Hidden,
68 cl::desc("Enable abort calls when \"fast\" instruction fails"));
70 SchedLiveInCopies("schedule-livein-copies",
71 cl::desc("Schedule copies of livein registers"),
76 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
77 cl::desc("Pop up a window to show dags before the first "
80 ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden,
81 cl::desc("Pop up a window to show dags before legalize types"));
83 ViewLegalizeDAGs("view-legalize-dags", cl::Hidden,
84 cl::desc("Pop up a window to show dags before legalize"));
86 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
87 cl::desc("Pop up a window to show dags before the second "
90 ViewISelDAGs("view-isel-dags", cl::Hidden,
91 cl::desc("Pop up a window to show isel dags as they are selected"));
93 ViewSchedDAGs("view-sched-dags", cl::Hidden,
94 cl::desc("Pop up a window to show sched dags as they are processed"));
96 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
97 cl::desc("Pop up a window to show SUnit dags after they are processed"));
99 static const bool ViewDAGCombine1 = false,
100 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false,
101 ViewDAGCombine2 = 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 ScheduleDAG* createDefaultScheduler(SelectionDAGISel *IS,
135 MachineBasicBlock *BB,
137 TargetLowering &TLI = IS->getTargetLowering();
139 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency) {
140 return createTDListDAGScheduler(IS, DAG, BB, Fast);
142 assert(TLI.getSchedulingPreference() ==
143 TargetLowering::SchedulingForRegPressure && "Unknown sched type!");
144 return createBURRListDAGScheduler(IS, DAG, BB, Fast);
149 // EmitInstrWithCustomInserter - This method should be implemented by targets
150 // that mark instructions with the 'usesCustomDAGSchedInserter' flag. These
151 // instructions are special in various ways, which require special support to
152 // insert. The specified MachineInstr is created but not inserted into any
153 // basic blocks, and the scheduler passes ownership of it to this method.
154 MachineBasicBlock *TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
155 MachineBasicBlock *MBB) {
156 cerr << "If a target marks an instruction with "
157 << "'usesCustomDAGSchedInserter', it must implement "
158 << "TargetLowering::EmitInstrWithCustomInserter!\n";
163 /// EmitLiveInCopy - Emit a copy for a live in physical register. If the
164 /// physical register has only a single copy use, then coalesced the copy
166 static void EmitLiveInCopy(MachineBasicBlock *MBB,
167 MachineBasicBlock::iterator &InsertPos,
168 unsigned VirtReg, unsigned PhysReg,
169 const TargetRegisterClass *RC,
170 DenseMap<MachineInstr*, unsigned> &CopyRegMap,
171 const MachineRegisterInfo &MRI,
172 const TargetRegisterInfo &TRI,
173 const TargetInstrInfo &TII) {
174 unsigned NumUses = 0;
175 MachineInstr *UseMI = NULL;
176 for (MachineRegisterInfo::use_iterator UI = MRI.use_begin(VirtReg),
177 UE = MRI.use_end(); UI != UE; ++UI) {
183 // If the number of uses is not one, or the use is not a move instruction,
184 // don't coalesce. Also, only coalesce away a virtual register to virtual
186 bool Coalesced = false;
187 unsigned SrcReg, DstReg;
189 TII.isMoveInstr(*UseMI, SrcReg, DstReg) &&
190 TargetRegisterInfo::isVirtualRegister(DstReg)) {
195 // Now find an ideal location to insert the copy.
196 MachineBasicBlock::iterator Pos = InsertPos;
197 while (Pos != MBB->begin()) {
198 MachineInstr *PrevMI = prior(Pos);
199 DenseMap<MachineInstr*, unsigned>::iterator RI = CopyRegMap.find(PrevMI);
200 // copyRegToReg might emit multiple instructions to do a copy.
201 unsigned CopyDstReg = (RI == CopyRegMap.end()) ? 0 : RI->second;
202 if (CopyDstReg && !TRI.regsOverlap(CopyDstReg, PhysReg))
203 // This is what the BB looks like right now:
208 // We want to insert "r1025 = mov r1". Inserting this copy below the
209 // move to r1024 makes it impossible for that move to be coalesced.
216 break; // Woot! Found a good location.
220 TII.copyRegToReg(*MBB, Pos, VirtReg, PhysReg, RC, RC);
221 CopyRegMap.insert(std::make_pair(prior(Pos), VirtReg));
223 if (&*InsertPos == UseMI) ++InsertPos;
228 /// EmitLiveInCopies - If this is the first basic block in the function,
229 /// and if it has live ins that need to be copied into vregs, emit the
230 /// copies into the block.
231 static void EmitLiveInCopies(MachineBasicBlock *EntryMBB,
232 const MachineRegisterInfo &MRI,
233 const TargetRegisterInfo &TRI,
234 const TargetInstrInfo &TII) {
235 if (SchedLiveInCopies) {
236 // Emit the copies at a heuristically-determined location in the block.
237 DenseMap<MachineInstr*, unsigned> CopyRegMap;
238 MachineBasicBlock::iterator InsertPos = EntryMBB->begin();
239 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
240 E = MRI.livein_end(); LI != E; ++LI)
242 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
243 EmitLiveInCopy(EntryMBB, InsertPos, LI->second, LI->first,
244 RC, CopyRegMap, MRI, TRI, TII);
247 // Emit the copies into the top of the block.
248 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
249 E = MRI.livein_end(); LI != E; ++LI)
251 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
252 TII.copyRegToReg(*EntryMBB, EntryMBB->begin(),
253 LI->second, LI->first, RC, RC);
258 //===----------------------------------------------------------------------===//
259 // SelectionDAGISel code
260 //===----------------------------------------------------------------------===//
262 SelectionDAGISel::SelectionDAGISel(TargetLowering &tli, bool fast) :
263 FunctionPass(&ID), TLI(tli),
264 FuncInfo(new FunctionLoweringInfo(TLI)),
265 CurDAG(new SelectionDAG(TLI, *FuncInfo)),
266 SDL(new SelectionDAGLowering(*CurDAG, TLI, *FuncInfo)),
272 SelectionDAGISel::~SelectionDAGISel() {
278 unsigned SelectionDAGISel::MakeReg(MVT VT) {
279 return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT));
282 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
283 AU.addRequired<AliasAnalysis>();
284 AU.addRequired<GCModuleInfo>();
285 AU.setPreservesAll();
288 bool SelectionDAGISel::runOnFunction(Function &Fn) {
289 // Do some sanity-checking on the command-line options.
290 assert((!EnableFastISelVerbose || EnableFastISel) &&
291 "-fast-isel-verbose requires -fast-isel");
292 assert((!EnableFastISelAbort || EnableFastISel) &&
293 "-fast-isel-abort requires -fast-isel");
295 // Get alias analysis for load/store combining.
296 AA = &getAnalysis<AliasAnalysis>();
298 TargetMachine &TM = TLI.getTargetMachine();
299 MachineFunction &MF = MachineFunction::construct(&Fn, TM);
300 const MachineRegisterInfo &MRI = MF.getRegInfo();
301 const TargetInstrInfo &TII = *TM.getInstrInfo();
302 const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
304 if (MF.getFunction()->hasGC())
305 GFI = &getAnalysis<GCModuleInfo>().getFunctionInfo(*MF.getFunction());
308 RegInfo = &MF.getRegInfo();
309 DOUT << "\n\n\n=== " << Fn.getName() << "\n";
311 FuncInfo->set(Fn, MF, EnableFastISel);
312 CurDAG->init(MF, getAnalysisToUpdate<MachineModuleInfo>());
315 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
316 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(I->getTerminator()))
318 FuncInfo->MBBMap[Invoke->getSuccessor(1)]->setIsLandingPad();
320 SelectAllBasicBlocks(Fn, MF);
322 // If the first basic block in the function has live ins that need to be
323 // copied into vregs, emit the copies into the top of the block before
324 // emitting the code for the block.
325 EmitLiveInCopies(MF.begin(), MRI, TRI, TII);
327 // Add function live-ins to entry block live-in set.
328 for (MachineRegisterInfo::livein_iterator I = RegInfo->livein_begin(),
329 E = RegInfo->livein_end(); I != E; ++I)
330 MF.begin()->addLiveIn(I->first);
333 assert(FuncInfo->CatchInfoFound.size() == FuncInfo->CatchInfoLost.size() &&
334 "Not all catch info was assigned to a landing pad!");
342 static void copyCatchInfo(BasicBlock *SrcBB, BasicBlock *DestBB,
343 MachineModuleInfo *MMI, FunctionLoweringInfo &FLI) {
344 for (BasicBlock::iterator I = SrcBB->begin(), E = --SrcBB->end(); I != E; ++I)
345 if (EHSelectorInst *EHSel = dyn_cast<EHSelectorInst>(I)) {
346 // Apply the catch info to DestBB.
347 AddCatchInfo(*EHSel, MMI, FLI.MBBMap[DestBB]);
349 if (!FLI.MBBMap[SrcBB]->isLandingPad())
350 FLI.CatchInfoFound.insert(EHSel);
355 /// IsFixedFrameObjectWithPosOffset - Check if object is a fixed frame object and
356 /// whether object offset >= 0.
358 IsFixedFrameObjectWithPosOffset(MachineFrameInfo * MFI, SDValue Op) {
359 if (!isa<FrameIndexSDNode>(Op)) return false;
361 FrameIndexSDNode * FrameIdxNode = dyn_cast<FrameIndexSDNode>(Op);
362 int FrameIdx = FrameIdxNode->getIndex();
363 return MFI->isFixedObjectIndex(FrameIdx) &&
364 MFI->getObjectOffset(FrameIdx) >= 0;
367 /// IsPossiblyOverwrittenArgumentOfTailCall - Check if the operand could
368 /// possibly be overwritten when lowering the outgoing arguments in a tail
369 /// call. Currently the implementation of this call is very conservative and
370 /// assumes all arguments sourcing from FORMAL_ARGUMENTS or a CopyFromReg with
371 /// virtual registers would be overwritten by direct lowering.
372 static bool IsPossiblyOverwrittenArgumentOfTailCall(SDValue Op,
373 MachineFrameInfo * MFI) {
374 RegisterSDNode * OpReg = NULL;
375 if (Op.getOpcode() == ISD::FORMAL_ARGUMENTS ||
376 (Op.getOpcode()== ISD::CopyFromReg &&
377 (OpReg = dyn_cast<RegisterSDNode>(Op.getOperand(1))) &&
378 (OpReg->getReg() >= TargetRegisterInfo::FirstVirtualRegister)) ||
379 (Op.getOpcode() == ISD::LOAD &&
380 IsFixedFrameObjectWithPosOffset(MFI, Op.getOperand(1))) ||
381 (Op.getOpcode() == ISD::MERGE_VALUES &&
382 Op.getOperand(Op.getResNo()).getOpcode() == ISD::LOAD &&
383 IsFixedFrameObjectWithPosOffset(MFI, Op.getOperand(Op.getResNo()).
389 /// CheckDAGForTailCallsAndFixThem - This Function looks for CALL nodes in the
390 /// DAG and fixes their tailcall attribute operand.
391 static void CheckDAGForTailCallsAndFixThem(SelectionDAG &DAG,
392 TargetLowering& TLI) {
394 SDValue Terminator = DAG.getRoot();
397 if (Terminator.getOpcode() == ISD::RET) {
398 Ret = Terminator.getNode();
401 // Fix tail call attribute of CALL nodes.
402 for (SelectionDAG::allnodes_iterator BE = DAG.allnodes_begin(),
403 BI = DAG.allnodes_end(); BI != BE; ) {
405 if (BI->getOpcode() == ISD::CALL) {
406 SDValue OpRet(Ret, 0);
407 SDValue OpCall(BI, 0);
408 bool isMarkedTailCall =
409 cast<ConstantSDNode>(OpCall.getOperand(3))->getZExtValue() != 0;
410 // If CALL node has tail call attribute set to true and the call is not
411 // eligible (no RET or the target rejects) the attribute is fixed to
412 // false. The TargetLowering::IsEligibleForTailCallOptimization function
413 // must correctly identify tail call optimizable calls.
414 if (!isMarkedTailCall) continue;
416 !TLI.IsEligibleForTailCallOptimization(OpCall, OpRet, DAG)) {
417 // Not eligible. Mark CALL node as non tail call.
418 SmallVector<SDValue, 32> Ops;
420 for(SDNode::op_iterator I =OpCall.getNode()->op_begin(),
421 E = OpCall.getNode()->op_end(); I != E; I++, idx++) {
425 Ops.push_back(DAG.getConstant(false, TLI.getPointerTy()));
427 DAG.UpdateNodeOperands(OpCall, Ops.begin(), Ops.size());
429 // Look for tail call clobbered arguments. Emit a series of
430 // copyto/copyfrom virtual register nodes to protect them.
431 SmallVector<SDValue, 32> Ops;
432 SDValue Chain = OpCall.getOperand(0), InFlag;
434 for(SDNode::op_iterator I = OpCall.getNode()->op_begin(),
435 E = OpCall.getNode()->op_end(); I != E; I++, idx++) {
437 if (idx > 4 && (idx % 2)) {
438 bool isByVal = cast<ARG_FLAGSSDNode>(OpCall.getOperand(idx+1))->
439 getArgFlags().isByVal();
440 MachineFunction &MF = DAG.getMachineFunction();
441 MachineFrameInfo *MFI = MF.getFrameInfo();
443 IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)) {
444 MVT VT = Arg.getValueType();
445 unsigned VReg = MF.getRegInfo().
446 createVirtualRegister(TLI.getRegClassFor(VT));
447 Chain = DAG.getCopyToReg(Chain, VReg, Arg, InFlag);
448 InFlag = Chain.getValue(1);
449 Arg = DAG.getCopyFromReg(Chain, VReg, VT, InFlag);
450 Chain = Arg.getValue(1);
451 InFlag = Arg.getValue(2);
456 // Link in chain of CopyTo/CopyFromReg.
458 DAG.UpdateNodeOperands(OpCall, Ops.begin(), Ops.size());
464 void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB,
465 BasicBlock::iterator Begin,
466 BasicBlock::iterator End) {
467 SDL->setCurrentBasicBlock(BB);
469 MachineModuleInfo *MMI = CurDAG->getMachineModuleInfo();
471 if (MMI && BB->isLandingPad()) {
472 // Add a label to mark the beginning of the landing pad. Deletion of the
473 // landing pad can thus be detected via the MachineModuleInfo.
474 unsigned LabelID = MMI->addLandingPad(BB);
475 CurDAG->setRoot(CurDAG->getLabel(ISD::EH_LABEL,
476 CurDAG->getEntryNode(), LabelID));
478 // Mark exception register as live in.
479 unsigned Reg = TLI.getExceptionAddressRegister();
480 if (Reg) BB->addLiveIn(Reg);
482 // Mark exception selector register as live in.
483 Reg = TLI.getExceptionSelectorRegister();
484 if (Reg) BB->addLiveIn(Reg);
486 // FIXME: Hack around an exception handling flaw (PR1508): the personality
487 // function and list of typeids logically belong to the invoke (or, if you
488 // like, the basic block containing the invoke), and need to be associated
489 // with it in the dwarf exception handling tables. Currently however the
490 // information is provided by an intrinsic (eh.selector) that can be moved
491 // to unexpected places by the optimizers: if the unwind edge is critical,
492 // then breaking it can result in the intrinsics being in the successor of
493 // the landing pad, not the landing pad itself. This results in exceptions
494 // not being caught because no typeids are associated with the invoke.
495 // This may not be the only way things can go wrong, but it is the only way
496 // we try to work around for the moment.
497 BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
499 if (Br && Br->isUnconditional()) { // Critical edge?
500 BasicBlock::iterator I, E;
501 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
502 if (isa<EHSelectorInst>(I))
506 // No catch info found - try to extract some from the successor.
507 copyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, *FuncInfo);
511 // Lower all of the non-terminator instructions.
512 for (BasicBlock::iterator I = Begin; I != End; ++I)
513 if (!isa<TerminatorInst>(I))
516 // Ensure that all instructions which are used outside of their defining
517 // blocks are available as virtual registers. Invoke is handled elsewhere.
518 for (BasicBlock::iterator I = Begin; I != End; ++I)
519 if (!I->use_empty() && !isa<PHINode>(I) && !isa<InvokeInst>(I)) {
520 DenseMap<const Value*,unsigned>::iterator VMI =FuncInfo->ValueMap.find(I);
521 if (VMI != FuncInfo->ValueMap.end())
522 SDL->CopyValueToVirtualRegister(I, VMI->second);
525 // Handle PHI nodes in successor blocks.
526 if (End == LLVMBB->end()) {
527 HandlePHINodesInSuccessorBlocks(LLVMBB);
529 // Lower the terminator after the copies are emitted.
530 SDL->visit(*LLVMBB->getTerminator());
533 // Make sure the root of the DAG is up-to-date.
534 CurDAG->setRoot(SDL->getControlRoot());
536 // Check whether calls in this block are real tail calls. Fix up CALL nodes
537 // with correct tailcall attribute so that the target can rely on the tailcall
538 // attribute indicating whether the call is really eligible for tail call
540 CheckDAGForTailCallsAndFixThem(*CurDAG, TLI);
542 // Final step, emit the lowered DAG as machine code.
547 void SelectionDAGISel::ComputeLiveOutVRegInfo() {
548 SmallPtrSet<SDNode*, 128> VisitedNodes;
549 SmallVector<SDNode*, 128> Worklist;
551 Worklist.push_back(CurDAG->getRoot().getNode());
557 while (!Worklist.empty()) {
558 SDNode *N = Worklist.back();
561 // If we've already seen this node, ignore it.
562 if (!VisitedNodes.insert(N))
565 // Otherwise, add all chain operands to the worklist.
566 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
567 if (N->getOperand(i).getValueType() == MVT::Other)
568 Worklist.push_back(N->getOperand(i).getNode());
570 // If this is a CopyToReg with a vreg dest, process it.
571 if (N->getOpcode() != ISD::CopyToReg)
574 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
575 if (!TargetRegisterInfo::isVirtualRegister(DestReg))
578 // Ignore non-scalar or non-integer values.
579 SDValue Src = N->getOperand(2);
580 MVT SrcVT = Src.getValueType();
581 if (!SrcVT.isInteger() || SrcVT.isVector())
584 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src);
585 Mask = APInt::getAllOnesValue(SrcVT.getSizeInBits());
586 CurDAG->ComputeMaskedBits(Src, Mask, KnownZero, KnownOne);
588 // Only install this information if it tells us something.
589 if (NumSignBits != 1 || KnownZero != 0 || KnownOne != 0) {
590 DestReg -= TargetRegisterInfo::FirstVirtualRegister;
591 FunctionLoweringInfo &FLI = CurDAG->getFunctionLoweringInfo();
592 if (DestReg >= FLI.LiveOutRegInfo.size())
593 FLI.LiveOutRegInfo.resize(DestReg+1);
594 FunctionLoweringInfo::LiveOutInfo &LOI = FLI.LiveOutRegInfo[DestReg];
595 LOI.NumSignBits = NumSignBits;
596 LOI.KnownOne = NumSignBits;
597 LOI.KnownZero = NumSignBits;
602 void SelectionDAGISel::CodeGenAndEmitDAG() {
603 std::string GroupName;
604 if (TimePassesIsEnabled)
605 GroupName = "Instruction Selection and Scheduling";
606 std::string BlockName;
607 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs ||
608 ViewDAGCombine2 || ViewISelDAGs || ViewSchedDAGs || ViewSUnitDAGs)
609 BlockName = CurDAG->getMachineFunction().getFunction()->getName() + ':' +
610 BB->getBasicBlock()->getName();
612 DOUT << "Initial selection DAG:\n";
613 DEBUG(CurDAG->dump());
615 if (ViewDAGCombine1) CurDAG->viewGraph("dag-combine1 input for " + BlockName);
617 // Run the DAG combiner in pre-legalize mode.
618 if (TimePassesIsEnabled) {
619 NamedRegionTimer T("DAG Combining 1", GroupName);
620 CurDAG->Combine(false, *AA, Fast);
622 CurDAG->Combine(false, *AA, Fast);
625 DOUT << "Optimized lowered selection DAG:\n";
626 DEBUG(CurDAG->dump());
628 // Second step, hack on the DAG until it only uses operations and types that
629 // the target supports.
630 if (EnableLegalizeTypes) {// Enable this some day.
631 if (ViewLegalizeTypesDAGs) CurDAG->viewGraph("legalize-types input for " +
634 if (TimePassesIsEnabled) {
635 NamedRegionTimer T("Type Legalization", GroupName);
636 CurDAG->LegalizeTypes();
638 CurDAG->LegalizeTypes();
641 DOUT << "Type-legalized selection DAG:\n";
642 DEBUG(CurDAG->dump());
644 // TODO: enable a dag combine pass here.
647 if (ViewLegalizeDAGs) CurDAG->viewGraph("legalize input for " + BlockName);
649 if (TimePassesIsEnabled) {
650 NamedRegionTimer T("DAG Legalization", GroupName);
656 DOUT << "Legalized selection DAG:\n";
657 DEBUG(CurDAG->dump());
659 if (ViewDAGCombine2) CurDAG->viewGraph("dag-combine2 input for " + BlockName);
661 // Run the DAG combiner in post-legalize mode.
662 if (TimePassesIsEnabled) {
663 NamedRegionTimer T("DAG Combining 2", GroupName);
664 CurDAG->Combine(true, *AA, Fast);
666 CurDAG->Combine(true, *AA, Fast);
669 DOUT << "Optimized legalized selection DAG:\n";
670 DEBUG(CurDAG->dump());
672 if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName);
674 if (!Fast && EnableValueProp)
675 ComputeLiveOutVRegInfo();
677 // Third, instruction select all of the operations to machine code, adding the
678 // code to the MachineBasicBlock.
679 if (TimePassesIsEnabled) {
680 NamedRegionTimer T("Instruction Selection", GroupName);
686 DOUT << "Selected selection DAG:\n";
687 DEBUG(CurDAG->dump());
689 if (ViewSchedDAGs) CurDAG->viewGraph("scheduler input for " + BlockName);
691 // Schedule machine code.
692 ScheduleDAG *Scheduler;
693 if (TimePassesIsEnabled) {
694 NamedRegionTimer T("Instruction Scheduling", GroupName);
695 Scheduler = Schedule();
697 Scheduler = Schedule();
700 if (ViewSUnitDAGs) Scheduler->viewGraph();
702 // Emit machine code to BB. This can change 'BB' to the last block being
704 if (TimePassesIsEnabled) {
705 NamedRegionTimer T("Instruction Creation", GroupName);
706 BB = Scheduler->EmitSchedule();
708 BB = Scheduler->EmitSchedule();
711 // Free the scheduler state.
712 if (TimePassesIsEnabled) {
713 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName);
719 DOUT << "Selected machine code:\n";
723 void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn, MachineFunction &MF) {
724 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
725 BasicBlock *LLVMBB = &*I;
726 BB = FuncInfo->MBBMap[LLVMBB];
728 BasicBlock::iterator const Begin = LLVMBB->begin();
729 BasicBlock::iterator const End = LLVMBB->end();
730 BasicBlock::iterator BI = Begin;
732 // Lower any arguments needed in this block if this is the entry block.
733 if (LLVMBB == &Fn.getEntryBlock())
734 LowerArguments(LLVMBB);
736 // Before doing SelectionDAG ISel, see if FastISel has been requested.
737 // FastISel doesn't support EH landing pads, which require special handling.
738 if (EnableFastISel && !BB->isLandingPad()) {
739 if (FastISel *F = TLI.createFastISel(*FuncInfo->MF, FuncInfo->ValueMap,
741 FuncInfo->StaticAllocaMap)) {
742 // Emit code for any incoming arguments. This must happen before
743 // beginning FastISel on the entry block.
744 if (LLVMBB == &Fn.getEntryBlock()) {
745 CurDAG->setRoot(SDL->getControlRoot());
749 F->setCurrentBlock(BB);
750 // Do FastISel on as many instructions as possible.
751 for (; BI != End; ++BI) {
752 // Just before the terminator instruction, insert instructions to
753 // feed PHI nodes in successor blocks.
754 if (isa<TerminatorInst>(BI))
755 if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, F)) {
756 if (EnableFastISelVerbose || EnableFastISelAbort) {
757 cerr << "FastISel miss: ";
760 if (EnableFastISelAbort)
761 assert(0 && "FastISel didn't handle a PHI in a successor");
765 // First try normal tablegen-generated "fast" selection.
766 if (F->SelectInstruction(BI))
769 // Next, try calling the target to attempt to handle the instruction.
770 if (F->TargetSelectInstruction(BI))
773 // Then handle certain instructions as single-LLVM-Instruction blocks.
774 if (isa<CallInst>(BI)) {
775 if (BI->getType() != Type::VoidTy) {
776 unsigned &R = FuncInfo->ValueMap[BI];
778 R = FuncInfo->CreateRegForValue(BI);
781 SelectBasicBlock(LLVMBB, BI, next(BI));
785 // Otherwise, give up on FastISel for the rest of the block.
786 // For now, be a little lenient about non-branch terminators.
787 if (!isa<TerminatorInst>(BI) || isa<BranchInst>(BI)) {
788 if (EnableFastISelVerbose || EnableFastISelAbort) {
789 cerr << "FastISel miss: ";
792 if (EnableFastISelAbort)
793 // The "fast" selector couldn't handle something and bailed.
794 // For the purpose of debugging, just abort.
795 assert(0 && "FastISel didn't select the entire block");
803 // Run SelectionDAG instruction selection on the remainder of the block
804 // not handled by FastISel. If FastISel is not run, this is the entire
807 SelectBasicBlock(LLVMBB, BI, End);
814 SelectionDAGISel::FinishBasicBlock() {
816 // Perform target specific isel post processing.
817 InstructionSelectPostProcessing();
819 DOUT << "Target-post-processed machine code:\n";
822 DOUT << "Total amount of phi nodes to update: "
823 << SDL->PHINodesToUpdate.size() << "\n";
824 DEBUG(for (unsigned i = 0, e = SDL->PHINodesToUpdate.size(); i != e; ++i)
825 DOUT << "Node " << i << " : (" << SDL->PHINodesToUpdate[i].first
826 << ", " << SDL->PHINodesToUpdate[i].second << ")\n";);
828 // Next, now that we know what the last MBB the LLVM BB expanded is, update
829 // PHI nodes in successors.
830 if (SDL->SwitchCases.empty() &&
831 SDL->JTCases.empty() &&
832 SDL->BitTestCases.empty()) {
833 for (unsigned i = 0, e = SDL->PHINodesToUpdate.size(); i != e; ++i) {
834 MachineInstr *PHI = SDL->PHINodesToUpdate[i].first;
835 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
836 "This is not a machine PHI node that we are updating!");
837 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[i].second,
839 PHI->addOperand(MachineOperand::CreateMBB(BB));
841 SDL->PHINodesToUpdate.clear();
845 for (unsigned i = 0, e = SDL->BitTestCases.size(); i != e; ++i) {
846 // Lower header first, if it wasn't already lowered
847 if (!SDL->BitTestCases[i].Emitted) {
848 // Set the current basic block to the mbb we wish to insert the code into
849 BB = SDL->BitTestCases[i].Parent;
850 SDL->setCurrentBasicBlock(BB);
852 SDL->visitBitTestHeader(SDL->BitTestCases[i]);
853 CurDAG->setRoot(SDL->getRoot());
858 for (unsigned j = 0, ej = SDL->BitTestCases[i].Cases.size(); j != ej; ++j) {
859 // Set the current basic block to the mbb we wish to insert the code into
860 BB = SDL->BitTestCases[i].Cases[j].ThisBB;
861 SDL->setCurrentBasicBlock(BB);
864 SDL->visitBitTestCase(SDL->BitTestCases[i].Cases[j+1].ThisBB,
865 SDL->BitTestCases[i].Reg,
866 SDL->BitTestCases[i].Cases[j]);
868 SDL->visitBitTestCase(SDL->BitTestCases[i].Default,
869 SDL->BitTestCases[i].Reg,
870 SDL->BitTestCases[i].Cases[j]);
873 CurDAG->setRoot(SDL->getRoot());
879 for (unsigned pi = 0, pe = SDL->PHINodesToUpdate.size(); pi != pe; ++pi) {
880 MachineInstr *PHI = SDL->PHINodesToUpdate[pi].first;
881 MachineBasicBlock *PHIBB = PHI->getParent();
882 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
883 "This is not a machine PHI node that we are updating!");
884 // This is "default" BB. We have two jumps to it. From "header" BB and
885 // from last "case" BB.
886 if (PHIBB == SDL->BitTestCases[i].Default) {
887 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
889 PHI->addOperand(MachineOperand::CreateMBB(SDL->BitTestCases[i].Parent));
890 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
892 PHI->addOperand(MachineOperand::CreateMBB(SDL->BitTestCases[i].Cases.
895 // One of "cases" BB.
896 for (unsigned j = 0, ej = SDL->BitTestCases[i].Cases.size();
898 MachineBasicBlock* cBB = SDL->BitTestCases[i].Cases[j].ThisBB;
899 if (cBB->succ_end() !=
900 std::find(cBB->succ_begin(),cBB->succ_end(), PHIBB)) {
901 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
903 PHI->addOperand(MachineOperand::CreateMBB(cBB));
908 SDL->BitTestCases.clear();
910 // If the JumpTable record is filled in, then we need to emit a jump table.
911 // Updating the PHI nodes is tricky in this case, since we need to determine
912 // whether the PHI is a successor of the range check MBB or the jump table MBB
913 for (unsigned i = 0, e = SDL->JTCases.size(); i != e; ++i) {
914 // Lower header first, if it wasn't already lowered
915 if (!SDL->JTCases[i].first.Emitted) {
916 // Set the current basic block to the mbb we wish to insert the code into
917 BB = SDL->JTCases[i].first.HeaderBB;
918 SDL->setCurrentBasicBlock(BB);
920 SDL->visitJumpTableHeader(SDL->JTCases[i].second, SDL->JTCases[i].first);
921 CurDAG->setRoot(SDL->getRoot());
926 // Set the current basic block to the mbb we wish to insert the code into
927 BB = SDL->JTCases[i].second.MBB;
928 SDL->setCurrentBasicBlock(BB);
930 SDL->visitJumpTable(SDL->JTCases[i].second);
931 CurDAG->setRoot(SDL->getRoot());
936 for (unsigned pi = 0, pe = SDL->PHINodesToUpdate.size(); pi != pe; ++pi) {
937 MachineInstr *PHI = SDL->PHINodesToUpdate[pi].first;
938 MachineBasicBlock *PHIBB = PHI->getParent();
939 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
940 "This is not a machine PHI node that we are updating!");
941 // "default" BB. We can go there only from header BB.
942 if (PHIBB == SDL->JTCases[i].second.Default) {
943 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
945 PHI->addOperand(MachineOperand::CreateMBB(SDL->JTCases[i].first.HeaderBB));
947 // JT BB. Just iterate over successors here
948 if (BB->succ_end() != std::find(BB->succ_begin(),BB->succ_end(), PHIBB)) {
949 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pi].second,
951 PHI->addOperand(MachineOperand::CreateMBB(BB));
955 SDL->JTCases.clear();
957 // If the switch block involved a branch to one of the actual successors, we
958 // need to update PHI nodes in that block.
959 for (unsigned i = 0, e = SDL->PHINodesToUpdate.size(); i != e; ++i) {
960 MachineInstr *PHI = SDL->PHINodesToUpdate[i].first;
961 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
962 "This is not a machine PHI node that we are updating!");
963 if (BB->isSuccessor(PHI->getParent())) {
964 PHI->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[i].second,
966 PHI->addOperand(MachineOperand::CreateMBB(BB));
970 // If we generated any switch lowering information, build and codegen any
971 // additional DAGs necessary.
972 for (unsigned i = 0, e = SDL->SwitchCases.size(); i != e; ++i) {
973 // Set the current basic block to the mbb we wish to insert the code into
974 BB = SDL->SwitchCases[i].ThisBB;
975 SDL->setCurrentBasicBlock(BB);
978 SDL->visitSwitchCase(SDL->SwitchCases[i]);
979 CurDAG->setRoot(SDL->getRoot());
983 // Handle any PHI nodes in successors of this chunk, as if we were coming
984 // from the original BB before switch expansion. Note that PHI nodes can
985 // occur multiple times in PHINodesToUpdate. We have to be very careful to
986 // handle them the right number of times.
987 while ((BB = SDL->SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
988 for (MachineBasicBlock::iterator Phi = BB->begin();
989 Phi != BB->end() && Phi->getOpcode() == TargetInstrInfo::PHI; ++Phi){
990 // This value for this PHI node is recorded in PHINodesToUpdate, get it.
991 for (unsigned pn = 0; ; ++pn) {
992 assert(pn != SDL->PHINodesToUpdate.size() &&
993 "Didn't find PHI entry!");
994 if (SDL->PHINodesToUpdate[pn].first == Phi) {
995 Phi->addOperand(MachineOperand::CreateReg(SDL->PHINodesToUpdate[pn].
997 Phi->addOperand(MachineOperand::CreateMBB(SDL->SwitchCases[i].ThisBB));
1003 // Don't process RHS if same block as LHS.
1004 if (BB == SDL->SwitchCases[i].FalseBB)
1005 SDL->SwitchCases[i].FalseBB = 0;
1007 // If we haven't handled the RHS, do so now. Otherwise, we're done.
1008 SDL->SwitchCases[i].TrueBB = SDL->SwitchCases[i].FalseBB;
1009 SDL->SwitchCases[i].FalseBB = 0;
1011 assert(SDL->SwitchCases[i].TrueBB == 0 && SDL->SwitchCases[i].FalseBB == 0);
1013 SDL->SwitchCases.clear();
1015 SDL->PHINodesToUpdate.clear();
1019 /// Schedule - Pick a safe ordering for instructions for each
1020 /// target node in the graph.
1022 ScheduleDAG *SelectionDAGISel::Schedule() {
1023 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
1027 RegisterScheduler::setDefault(Ctor);
1030 ScheduleDAG *Scheduler = Ctor(this, CurDAG, BB, Fast);
1037 HazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
1038 return new HazardRecognizer();
1041 //===----------------------------------------------------------------------===//
1042 // Helper functions used by the generated instruction selector.
1043 //===----------------------------------------------------------------------===//
1044 // Calls to these methods are generated by tblgen.
1046 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
1047 /// the dag combiner simplified the 255, we still want to match. RHS is the
1048 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
1049 /// specified in the .td file (e.g. 255).
1050 bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS,
1051 int64_t DesiredMaskS) const {
1052 const APInt &ActualMask = RHS->getAPIntValue();
1053 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1055 // If the actual mask exactly matches, success!
1056 if (ActualMask == DesiredMask)
1059 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1060 if (ActualMask.intersects(~DesiredMask))
1063 // Otherwise, the DAG Combiner may have proven that the value coming in is
1064 // either already zero or is not demanded. Check for known zero input bits.
1065 APInt NeededMask = DesiredMask & ~ActualMask;
1066 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
1069 // TODO: check to see if missing bits are just not demanded.
1071 // Otherwise, this pattern doesn't match.
1075 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
1076 /// the dag combiner simplified the 255, we still want to match. RHS is the
1077 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
1078 /// specified in the .td file (e.g. 255).
1079 bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS,
1080 int64_t DesiredMaskS) const {
1081 const APInt &ActualMask = RHS->getAPIntValue();
1082 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
1084 // If the actual mask exactly matches, success!
1085 if (ActualMask == DesiredMask)
1088 // If the actual AND mask is allowing unallowed bits, this doesn't match.
1089 if (ActualMask.intersects(~DesiredMask))
1092 // Otherwise, the DAG Combiner may have proven that the value coming in is
1093 // either already zero or is not demanded. Check for known zero input bits.
1094 APInt NeededMask = DesiredMask & ~ActualMask;
1096 APInt KnownZero, KnownOne;
1097 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
1099 // If all the missing bits in the or are already known to be set, match!
1100 if ((NeededMask & KnownOne) == NeededMask)
1103 // TODO: check to see if missing bits are just not demanded.
1105 // Otherwise, this pattern doesn't match.
1110 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
1111 /// by tblgen. Others should not call it.
1112 void SelectionDAGISel::
1113 SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
1114 std::vector<SDValue> InOps;
1115 std::swap(InOps, Ops);
1117 Ops.push_back(InOps[0]); // input chain.
1118 Ops.push_back(InOps[1]); // input asm string.
1120 unsigned i = 2, e = InOps.size();
1121 if (InOps[e-1].getValueType() == MVT::Flag)
1122 --e; // Don't process a flag operand if it is here.
1125 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue();
1126 if ((Flags & 7) != 4 /*MEM*/) {
1127 // Just skip over this operand, copying the operands verbatim.
1128 Ops.insert(Ops.end(), InOps.begin()+i, InOps.begin()+i+(Flags >> 3) + 1);
1129 i += (Flags >> 3) + 1;
1131 assert((Flags >> 3) == 1 && "Memory operand with multiple values?");
1132 // Otherwise, this is a memory operand. Ask the target to select it.
1133 std::vector<SDValue> SelOps;
1134 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps)) {
1135 cerr << "Could not match memory address. Inline asm failure!\n";
1139 // Add this to the output node.
1140 MVT IntPtrTy = CurDAG->getTargetLoweringInfo().getPointerTy();
1141 Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size() << 3),
1143 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
1148 // Add the flag input back if present.
1149 if (e != InOps.size())
1150 Ops.push_back(InOps.back());
1153 char SelectionDAGISel::ID = 0;