1//===-- Analysis.cpp - CodeGen LLVM IR Analysis Utilities -----------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines several CodeGen-specific LLVM IR analysis utilties.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/CodeGen/Analysis.h"
15#include "llvm/Analysis/ValueTracking.h"
16#include "llvm/DerivedTypes.h"
17#include "llvm/Function.h"
18#include "llvm/Instructions.h"
19#include "llvm/IntrinsicInst.h"
20#include "llvm/LLVMContext.h"
21#include "llvm/Module.h"
22#include "llvm/CodeGen/MachineFunction.h"
23#include "llvm/CodeGen/SelectionDAG.h"
24#include "llvm/Target/TargetData.h"
25#include "llvm/Target/TargetLowering.h"
26#include "llvm/Target/TargetOptions.h"
27#include "llvm/Support/ErrorHandling.h"
28#include "llvm/Support/MathExtras.h"
29using namespace llvm;
30
31/// ComputeLinearIndex - Given an LLVM IR aggregate type and a sequence
32/// of insertvalue or extractvalue indices that identify a member, return
33/// the linearized index of the start of the member.
34///
35unsigned llvm::ComputeLinearIndex(Type *Ty,
36                                  const unsigned *Indices,
37                                  const unsigned *IndicesEnd,
38                                  unsigned CurIndex) {
39  // Base case: We're done.
40  if (Indices && Indices == IndicesEnd)
41    return CurIndex;
42
43  // Given a struct type, recursively traverse the elements.
44  if (StructType *STy = dyn_cast<StructType>(Ty)) {
45    for (StructType::element_iterator EB = STy->element_begin(),
46                                      EI = EB,
47                                      EE = STy->element_end();
48        EI != EE; ++EI) {
49      if (Indices && *Indices == unsigned(EI - EB))
50        return ComputeLinearIndex(*EI, Indices+1, IndicesEnd, CurIndex);
51      CurIndex = ComputeLinearIndex(*EI, 0, 0, CurIndex);
52    }
53    return CurIndex;
54  }
55  // Given an array type, recursively traverse the elements.
56  else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
57    Type *EltTy = ATy->getElementType();
58    for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i) {
59      if (Indices && *Indices == i)
60        return ComputeLinearIndex(EltTy, Indices+1, IndicesEnd, CurIndex);
61      CurIndex = ComputeLinearIndex(EltTy, 0, 0, CurIndex);
62    }
63    return CurIndex;
64  }
65  // We haven't found the type we're looking for, so keep searching.
66  return CurIndex + 1;
67}
68
69/// ComputeValueVTs - Given an LLVM IR type, compute a sequence of
70/// EVTs that represent all the individual underlying
71/// non-aggregate types that comprise it.
72///
73/// If Offsets is non-null, it points to a vector to be filled in
74/// with the in-memory offsets of each of the individual values.
75///
76void llvm::ComputeValueVTs(const TargetLowering &TLI, Type *Ty,
77                           SmallVectorImpl<EVT> &ValueVTs,
78                           SmallVectorImpl<uint64_t> *Offsets,
79                           uint64_t StartingOffset) {
80  // Given a struct type, recursively traverse the elements.
81  if (StructType *STy = dyn_cast<StructType>(Ty)) {
82    const StructLayout *SL = TLI.getTargetData()->getStructLayout(STy);
83    for (StructType::element_iterator EB = STy->element_begin(),
84                                      EI = EB,
85                                      EE = STy->element_end();
86         EI != EE; ++EI)
87      ComputeValueVTs(TLI, *EI, ValueVTs, Offsets,
88                      StartingOffset + SL->getElementOffset(EI - EB));
89    return;
90  }
91  // Given an array type, recursively traverse the elements.
92  if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
93    Type *EltTy = ATy->getElementType();
94    uint64_t EltSize = TLI.getTargetData()->getTypeAllocSize(EltTy);
95    for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
96      ComputeValueVTs(TLI, EltTy, ValueVTs, Offsets,
97                      StartingOffset + i * EltSize);
98    return;
99  }
100  // Interpret void as zero return values.
101  if (Ty->isVoidTy())
102    return;
103  // Base case: we can get an EVT for this LLVM IR type.
104  ValueVTs.push_back(TLI.getValueType(Ty));
105  if (Offsets)
106    Offsets->push_back(StartingOffset);
107}
108
109/// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
110GlobalVariable *llvm::ExtractTypeInfo(Value *V) {
111  V = V->stripPointerCasts();
112  GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
113
114  if (GV && GV->getName() == "llvm.eh.catch.all.value") {
115    assert(GV->hasInitializer() &&
116           "The EH catch-all value must have an initializer");
117    Value *Init = GV->getInitializer();
118    GV = dyn_cast<GlobalVariable>(Init);
119    if (!GV) V = cast<ConstantPointerNull>(Init);
120  }
121
122  assert((GV || isa<ConstantPointerNull>(V)) &&
123         "TypeInfo must be a global variable or NULL");
124  return GV;
125}
126
127/// hasInlineAsmMemConstraint - Return true if the inline asm instruction being
128/// processed uses a memory 'm' constraint.
129bool
130llvm::hasInlineAsmMemConstraint(InlineAsm::ConstraintInfoVector &CInfos,
131                                const TargetLowering &TLI) {
132  for (unsigned i = 0, e = CInfos.size(); i != e; ++i) {
133    InlineAsm::ConstraintInfo &CI = CInfos[i];
134    for (unsigned j = 0, ee = CI.Codes.size(); j != ee; ++j) {
135      TargetLowering::ConstraintType CType = TLI.getConstraintType(CI.Codes[j]);
136      if (CType == TargetLowering::C_Memory)
137        return true;
138    }
139
140    // Indirect operand accesses access memory.
141    if (CI.isIndirect)
142      return true;
143  }
144
145  return false;
146}
147
148/// getFCmpCondCode - Return the ISD condition code corresponding to
149/// the given LLVM IR floating-point condition code.  This includes
150/// consideration of global floating-point math flags.
151///
152ISD::CondCode llvm::getFCmpCondCode(FCmpInst::Predicate Pred) {
153  switch (Pred) {
154  case FCmpInst::FCMP_FALSE: return ISD::SETFALSE;
155  case FCmpInst::FCMP_OEQ:   return ISD::SETOEQ;
156  case FCmpInst::FCMP_OGT:   return ISD::SETOGT;
157  case FCmpInst::FCMP_OGE:   return ISD::SETOGE;
158  case FCmpInst::FCMP_OLT:   return ISD::SETOLT;
159  case FCmpInst::FCMP_OLE:   return ISD::SETOLE;
160  case FCmpInst::FCMP_ONE:   return ISD::SETONE;
161  case FCmpInst::FCMP_ORD:   return ISD::SETO;
162  case FCmpInst::FCMP_UNO:   return ISD::SETUO;
163  case FCmpInst::FCMP_UEQ:   return ISD::SETUEQ;
164  case FCmpInst::FCMP_UGT:   return ISD::SETUGT;
165  case FCmpInst::FCMP_UGE:   return ISD::SETUGE;
166  case FCmpInst::FCMP_ULT:   return ISD::SETULT;
167  case FCmpInst::FCMP_ULE:   return ISD::SETULE;
168  case FCmpInst::FCMP_UNE:   return ISD::SETUNE;
169  case FCmpInst::FCMP_TRUE:  return ISD::SETTRUE;
170  default: llvm_unreachable("Invalid FCmp predicate opcode!");
171  }
172}
173
174ISD::CondCode llvm::getFCmpCodeWithoutNaN(ISD::CondCode CC) {
175  switch (CC) {
176    case ISD::SETOEQ: case ISD::SETUEQ: return ISD::SETEQ;
177    case ISD::SETONE: case ISD::SETUNE: return ISD::SETNE;
178    case ISD::SETOLT: case ISD::SETULT: return ISD::SETLT;
179    case ISD::SETOLE: case ISD::SETULE: return ISD::SETLE;
180    case ISD::SETOGT: case ISD::SETUGT: return ISD::SETGT;
181    case ISD::SETOGE: case ISD::SETUGE: return ISD::SETGE;
182    default: return CC;
183  }
184}
185
186/// getICmpCondCode - Return the ISD condition code corresponding to
187/// the given LLVM IR integer condition code.
188///
189ISD::CondCode llvm::getICmpCondCode(ICmpInst::Predicate Pred) {
190  switch (Pred) {
191  case ICmpInst::ICMP_EQ:  return ISD::SETEQ;
192  case ICmpInst::ICMP_NE:  return ISD::SETNE;
193  case ICmpInst::ICMP_SLE: return ISD::SETLE;
194  case ICmpInst::ICMP_ULE: return ISD::SETULE;
195  case ICmpInst::ICMP_SGE: return ISD::SETGE;
196  case ICmpInst::ICMP_UGE: return ISD::SETUGE;
197  case ICmpInst::ICMP_SLT: return ISD::SETLT;
198  case ICmpInst::ICMP_ULT: return ISD::SETULT;
199  case ICmpInst::ICMP_SGT: return ISD::SETGT;
200  case ICmpInst::ICMP_UGT: return ISD::SETUGT;
201  default:
202    llvm_unreachable("Invalid ICmp predicate opcode!");
203  }
204}
205
206
207/// getNoopInput - If V is a noop (i.e., lowers to no machine code), look
208/// through it (and any transitive noop operands to it) and return its input
209/// value.  This is used to determine if a tail call can be formed.
210///
211static const Value *getNoopInput(const Value *V, const TargetLowering &TLI) {
212  // If V is not an instruction, it can't be looked through.
213  const Instruction *I = dyn_cast<Instruction>(V);
214  if (I == 0 || !I->hasOneUse() || I->getNumOperands() == 0) return V;
215
216  Value *Op = I->getOperand(0);
217
218  // Look through truly no-op truncates.
219  if (isa<TruncInst>(I) &&
220      TLI.isTruncateFree(I->getOperand(0)->getType(), I->getType()))
221    return getNoopInput(I->getOperand(0), TLI);
222
223  // Look through truly no-op bitcasts.
224  if (isa<BitCastInst>(I)) {
225    // No type change at all.
226    if (Op->getType() == I->getType())
227      return getNoopInput(Op, TLI);
228
229    // Pointer to pointer cast.
230    if (Op->getType()->isPointerTy() && I->getType()->isPointerTy())
231      return getNoopInput(Op, TLI);
232
233    if (isa<VectorType>(Op->getType()) && isa<VectorType>(I->getType()) &&
234        TLI.isTypeLegal(EVT::getEVT(Op->getType())) &&
235        TLI.isTypeLegal(EVT::getEVT(I->getType())))
236      return getNoopInput(Op, TLI);
237  }
238
239  // Look through inttoptr.
240  if (isa<IntToPtrInst>(I) && !isa<VectorType>(I->getType())) {
241    // Make sure this isn't a truncating or extending cast.  We could support
242    // this eventually, but don't bother for now.
243    if (TLI.getPointerTy().getSizeInBits() ==
244          cast<IntegerType>(Op->getType())->getBitWidth())
245      return getNoopInput(Op, TLI);
246  }
247
248  // Look through ptrtoint.
249  if (isa<PtrToIntInst>(I) && !isa<VectorType>(I->getType())) {
250    // Make sure this isn't a truncating or extending cast.  We could support
251    // this eventually, but don't bother for now.
252    if (TLI.getPointerTy().getSizeInBits() ==
253        cast<IntegerType>(I->getType())->getBitWidth())
254      return getNoopInput(Op, TLI);
255  }
256
257
258  // Otherwise it's not something we can look through.
259  return V;
260}
261
262
263/// Test if the given instruction is in a position to be optimized
264/// with a tail-call. This roughly means that it's in a block with
265/// a return and there's nothing that needs to be scheduled
266/// between it and the return.
267///
268/// This function only tests target-independent requirements.
269bool llvm::isInTailCallPosition(ImmutableCallSite CS, Attributes CalleeRetAttr,
270                                const TargetLowering &TLI) {
271  const Instruction *I = CS.getInstruction();
272  const BasicBlock *ExitBB = I->getParent();
273  const TerminatorInst *Term = ExitBB->getTerminator();
274  const ReturnInst *Ret = dyn_cast<ReturnInst>(Term);
275
276  // The block must end in a return statement or unreachable.
277  //
278  // FIXME: Decline tailcall if it's not guaranteed and if the block ends in
279  // an unreachable, for now. The way tailcall optimization is currently
280  // implemented means it will add an epilogue followed by a jump. That is
281  // not profitable. Also, if the callee is a special function (e.g.
282  // longjmp on x86), it can end up causing miscompilation that has not
283  // been fully understood.
284  if (!Ret &&
285      (!TLI.getTargetMachine().Options.GuaranteedTailCallOpt ||
286       !isa<UnreachableInst>(Term)))
287    return false;
288
289  // If I will have a chain, make sure no other instruction that will have a
290  // chain interposes between I and the return.
291  if (I->mayHaveSideEffects() || I->mayReadFromMemory() ||
292      !isSafeToSpeculativelyExecute(I))
293    for (BasicBlock::const_iterator BBI = prior(prior(ExitBB->end())); ;
294         --BBI) {
295      if (&*BBI == I)
296        break;
297      // Debug info intrinsics do not get in the way of tail call optimization.
298      if (isa<DbgInfoIntrinsic>(BBI))
299        continue;
300      if (BBI->mayHaveSideEffects() || BBI->mayReadFromMemory() ||
301          !isSafeToSpeculativelyExecute(BBI))
302        return false;
303    }
304
305  // If the block ends with a void return or unreachable, it doesn't matter
306  // what the call's return type is.
307  if (!Ret || Ret->getNumOperands() == 0) return true;
308
309  // If the return value is undef, it doesn't matter what the call's
310  // return type is.
311  if (isa<UndefValue>(Ret->getOperand(0))) return true;
312
313  // Conservatively require the attributes of the call to match those of
314  // the return. Ignore noalias because it doesn't affect the call sequence.
315  const Function *F = ExitBB->getParent();
316  Attributes CallerRetAttr = F->getAttributes().getRetAttributes();
317  if ((CalleeRetAttr ^ CallerRetAttr) & ~Attribute::NoAlias)
318    return false;
319
320  // It's not safe to eliminate the sign / zero extension of the return value.
321  if (CallerRetAttr.hasZExtAttr() || CallerRetAttr.hasSExtAttr())
322    return false;
323
324  // Otherwise, make sure the unmodified return value of I is the return value.
325  // We handle two cases: multiple return values + scalars.
326  Value *RetVal = Ret->getOperand(0);
327  if (!isa<InsertValueInst>(RetVal) || !isa<StructType>(RetVal->getType()))
328    // Handle scalars first.
329    return getNoopInput(Ret->getOperand(0), TLI) == I;
330
331  // If this is an aggregate return, look through the insert/extract values and
332  // see if each is transparent.
333  for (unsigned i = 0, e =cast<StructType>(RetVal->getType())->getNumElements();
334       i != e; ++i) {
335    const Value *InScalar = FindInsertedValue(RetVal, i);
336    if (InScalar == 0) return false;
337    InScalar = getNoopInput(InScalar, TLI);
338
339    // If the scalar value being inserted is an extractvalue of the right index
340    // from the call, then everything is good.
341    const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(InScalar);
342    if (EVI == 0 || EVI->getOperand(0) != I || EVI->getNumIndices() != 1 ||
343        EVI->getIndices()[0] != i)
344      return false;
345  }
346
347  return true;
348}
349
350bool llvm::isInTailCallPosition(SelectionDAG &DAG, SDNode *Node,
351                                SDValue &Chain, const TargetLowering &TLI) {
352  const Function *F = DAG.getMachineFunction().getFunction();
353
354  // Conservatively require the attributes of the call to match those of
355  // the return. Ignore noalias because it doesn't affect the call sequence.
356  Attributes CallerRetAttr = F->getAttributes().getRetAttributes();
357  if (CallerRetAttr & ~Attribute::NoAlias)
358    return false;
359
360  // It's not safe to eliminate the sign / zero extension of the return value.
361  if (CallerRetAttr.hasZExtAttr() || CallerRetAttr.hasSExtAttr())
362    return false;
363
364  // Check if the only use is a function return node.
365  return TLI.isUsedByReturnOnly(Node, Chain);
366}
367