1 //===- LoopVR.cpp - Value Range analysis driven by loop information -------===//
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 // FIXME: What does this do?
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "loopvr"
15 #include "llvm/Analysis/LoopVR.h"
16 #include "llvm/Constants.h"
17 #include "llvm/Instructions.h"
18 #include "llvm/LLVMContext.h"
19 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
20 #include "llvm/Assembly/Writer.h"
21 #include "llvm/Support/CFG.h"
22 #include "llvm/Support/Debug.h"
23 #include "llvm/Support/raw_ostream.h"
27 static RegisterPass<LoopVR> X("loopvr", "Loop Value Ranges", false, true);
29 /// getRange - determine the range for a particular SCEV within a given Loop
30 ConstantRange LoopVR::getRange(const SCEV *S, Loop *L, ScalarEvolution &SE) {
31 const SCEV *T = SE.getBackedgeTakenCount(L);
32 if (isa<SCEVCouldNotCompute>(T))
33 return ConstantRange(cast<IntegerType>(S->getType())->getBitWidth(), true);
35 T = SE.getTruncateOrZeroExtend(T, S->getType());
36 return getRange(S, T, SE);
39 /// getRange - determine the range for a particular SCEV with a given trip count
40 ConstantRange LoopVR::getRange(const SCEV *S, const SCEV *T, ScalarEvolution &SE){
42 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
43 return ConstantRange(C->getValue()->getValue());
45 ConstantRange FullSet(cast<IntegerType>(S->getType())->getBitWidth(), true);
47 // {x,+,y,+,...z}. We detect overflow by checking the size of the set after
48 // summing the upper and lower.
49 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
50 ConstantRange X = getRange(Add->getOperand(0), T, SE);
51 if (X.isFullSet()) return FullSet;
52 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i) {
53 ConstantRange Y = getRange(Add->getOperand(i), T, SE);
54 if (Y.isFullSet()) return FullSet;
56 APInt Spread_X = X.getSetSize(), Spread_Y = Y.getSetSize();
57 APInt NewLower = X.getLower() + Y.getLower();
58 APInt NewUpper = X.getUpper() + Y.getUpper() - 1;
59 if (NewLower == NewUpper)
62 X = ConstantRange(NewLower, NewUpper);
63 if (X.getSetSize().ult(Spread_X) || X.getSetSize().ult(Spread_Y))
64 return FullSet; // we've wrapped, therefore, full set.
69 // {x,*,y,*,...,z}. In order to detect overflow, we use k*bitwidth where
70 // k is the number of terms being multiplied.
71 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
72 ConstantRange X = getRange(Mul->getOperand(0), T, SE);
73 if (X.isFullSet()) return FullSet;
75 const IntegerType *Ty = Context->getIntegerType(X.getBitWidth());
76 const IntegerType *ExTy = Context->getIntegerType(X.getBitWidth() *
77 Mul->getNumOperands());
78 ConstantRange XExt = X.zeroExtend(ExTy->getBitWidth());
80 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i) {
81 ConstantRange Y = getRange(Mul->getOperand(i), T, SE);
82 if (Y.isFullSet()) return FullSet;
84 ConstantRange YExt = Y.zeroExtend(ExTy->getBitWidth());
85 XExt = ConstantRange(XExt.getLower() * YExt.getLower(),
86 ((XExt.getUpper()-1) * (YExt.getUpper()-1)) + 1);
88 return XExt.truncate(Ty->getBitWidth());
91 // X smax Y smax ... Z is: range(smax(X_smin, Y_smin, ..., Z_smin),
92 // smax(X_smax, Y_smax, ..., Z_smax))
93 // It doesn't matter if one of the SCEVs has FullSet because we're taking
94 // a maximum of the minimums across all of them.
95 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
96 ConstantRange X = getRange(SMax->getOperand(0), T, SE);
97 if (X.isFullSet()) return FullSet;
99 APInt smin = X.getSignedMin(), smax = X.getSignedMax();
100 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i) {
101 ConstantRange Y = getRange(SMax->getOperand(i), T, SE);
102 smin = APIntOps::smax(smin, Y.getSignedMin());
103 smax = APIntOps::smax(smax, Y.getSignedMax());
105 if (smax + 1 == smin) return FullSet;
106 return ConstantRange(smin, smax + 1);
109 // X umax Y umax ... Z is: range(umax(X_umin, Y_umin, ..., Z_umin),
110 // umax(X_umax, Y_umax, ..., Z_umax))
111 // It doesn't matter if one of the SCEVs has FullSet because we're taking
112 // a maximum of the minimums across all of them.
113 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
114 ConstantRange X = getRange(UMax->getOperand(0), T, SE);
115 if (X.isFullSet()) return FullSet;
117 APInt umin = X.getUnsignedMin(), umax = X.getUnsignedMax();
118 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i) {
119 ConstantRange Y = getRange(UMax->getOperand(i), T, SE);
120 umin = APIntOps::umax(umin, Y.getUnsignedMin());
121 umax = APIntOps::umax(umax, Y.getUnsignedMax());
123 if (umax + 1 == umin) return FullSet;
124 return ConstantRange(umin, umax + 1);
127 // L udiv R. Luckily, there's only ever 2 sides to a udiv.
128 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
129 ConstantRange L = getRange(UDiv->getLHS(), T, SE);
130 ConstantRange R = getRange(UDiv->getRHS(), T, SE);
131 if (L.isFullSet() && R.isFullSet()) return FullSet;
133 if (R.getUnsignedMax() == 0) {
134 // RHS must be single-element zero. Return an empty set.
135 return ConstantRange(R.getBitWidth(), false);
138 APInt Lower = L.getUnsignedMin().udiv(R.getUnsignedMax());
142 if (R.getUnsignedMin() == 0) {
143 // Just because it contains zero, doesn't mean it will also contain one.
144 // Use maximalIntersectWith to get the right behaviour.
145 ConstantRange NotZero(APInt(L.getBitWidth(), 1),
146 APInt::getNullValue(L.getBitWidth()));
147 R = R.maximalIntersectWith(NotZero);
150 // But, the maximal intersection might still include zero. If it does, then
151 // we know it also included one.
152 if (R.contains(APInt::getNullValue(L.getBitWidth())))
153 Upper = L.getUnsignedMax();
155 Upper = L.getUnsignedMax().udiv(R.getUnsignedMin());
157 return ConstantRange(Lower, Upper);
160 // ConstantRange already implements the cast operators.
162 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
163 T = SE.getTruncateOrZeroExtend(T, ZExt->getOperand()->getType());
164 ConstantRange X = getRange(ZExt->getOperand(), T, SE);
165 return X.zeroExtend(cast<IntegerType>(ZExt->getType())->getBitWidth());
168 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
169 T = SE.getTruncateOrZeroExtend(T, SExt->getOperand()->getType());
170 ConstantRange X = getRange(SExt->getOperand(), T, SE);
171 return X.signExtend(cast<IntegerType>(SExt->getType())->getBitWidth());
174 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
175 T = SE.getTruncateOrZeroExtend(T, Trunc->getOperand()->getType());
176 ConstantRange X = getRange(Trunc->getOperand(), T, SE);
177 if (X.isFullSet()) return FullSet;
178 return X.truncate(cast<IntegerType>(Trunc->getType())->getBitWidth());
181 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
182 const SCEVConstant *Trip = dyn_cast<SCEVConstant>(T);
183 if (!Trip) return FullSet;
185 if (AddRec->isAffine()) {
186 const SCEV *StartHandle = AddRec->getStart();
187 const SCEV *StepHandle = AddRec->getOperand(1);
189 const SCEVConstant *Step = dyn_cast<SCEVConstant>(StepHandle);
190 if (!Step) return FullSet;
192 uint32_t ExWidth = 2 * Trip->getValue()->getBitWidth();
193 APInt TripExt = Trip->getValue()->getValue(); TripExt.zext(ExWidth);
194 APInt StepExt = Step->getValue()->getValue(); StepExt.zext(ExWidth);
195 if ((TripExt * StepExt).ugt(APInt::getLowBitsSet(ExWidth, ExWidth >> 1)))
198 const SCEV *EndHandle = SE.getAddExpr(StartHandle,
199 SE.getMulExpr(T, StepHandle));
200 const SCEVConstant *Start = dyn_cast<SCEVConstant>(StartHandle);
201 const SCEVConstant *End = dyn_cast<SCEVConstant>(EndHandle);
202 if (!Start || !End) return FullSet;
204 const APInt &StartInt = Start->getValue()->getValue();
205 const APInt &EndInt = End->getValue()->getValue();
206 const APInt &StepInt = Step->getValue()->getValue();
208 if (StepInt.isNegative()) {
209 if (EndInt == StartInt + 1) return FullSet;
210 return ConstantRange(EndInt, StartInt + 1);
212 if (StartInt == EndInt + 1) return FullSet;
213 return ConstantRange(StartInt, EndInt + 1);
218 // TODO: non-affine addrec, udiv, SCEVUnknown (narrowed from elsewhere)?
223 bool LoopVR::runOnFunction(Function &F) { Map.clear(); return false; }
225 void LoopVR::print(std::ostream &os, const Module *) const {
226 raw_os_ostream OS(os);
227 for (std::map<Value *, ConstantRange *>::const_iterator I = Map.begin(),
228 E = Map.end(); I != E; ++I) {
229 OS << *I->first << ": " << *I->second << '\n';
233 void LoopVR::releaseMemory() {
234 for (std::map<Value *, ConstantRange *>::iterator I = Map.begin(),
235 E = Map.end(); I != E; ++I) {
242 ConstantRange LoopVR::compute(Value *V) {
243 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
244 return ConstantRange(CI->getValue());
246 Instruction *I = dyn_cast<Instruction>(V);
248 return ConstantRange(cast<IntegerType>(V->getType())->getBitWidth(), false);
250 LoopInfo &LI = getAnalysis<LoopInfo>();
252 Loop *L = LI.getLoopFor(I->getParent());
253 if (!L || L->isLoopInvariant(I))
254 return ConstantRange(cast<IntegerType>(V->getType())->getBitWidth(), false);
256 ScalarEvolution &SE = getAnalysis<ScalarEvolution>();
258 const SCEV *S = SE.getSCEV(I);
259 if (isa<SCEVUnknown>(S) || isa<SCEVCouldNotCompute>(S))
260 return ConstantRange(cast<IntegerType>(V->getType())->getBitWidth(), false);
262 return ConstantRange(getRange(S, L, SE));
265 ConstantRange LoopVR::get(Value *V) {
266 std::map<Value *, ConstantRange *>::iterator I = Map.find(V);
267 if (I == Map.end()) {
268 ConstantRange *CR = new ConstantRange(compute(V));
276 void LoopVR::remove(Value *V) {
277 std::map<Value *, ConstantRange *>::iterator I = Map.find(V);
278 if (I != Map.end()) {
284 void LoopVR::narrow(Value *V, const ConstantRange &CR) {
285 if (CR.isFullSet()) return;
287 std::map<Value *, ConstantRange *>::iterator I = Map.find(V);
289 Map[V] = new ConstantRange(CR);
291 Map[V] = new ConstantRange(Map[V]->maximalIntersectWith(CR));