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