diff --git a/llvm/include/llvm/Analysis/Passes.h b/llvm/include/llvm/Analysis/Passes.h index e5b86e867364..e68765310fb8 100644 --- a/llvm/include/llvm/Analysis/Passes.h +++ b/llvm/include/llvm/Analysis/Passes.h @@ -72,6 +72,13 @@ namespace llvm { /// FunctionPass *createLibCallAliasAnalysisPass(LibCallInfo *LCI); + //===--------------------------------------------------------------------===// + // + // createScalarEvolutionAliasAnalysisPass - This pass implements a simple + // alias analysis using ScalarEvolution queries. + // + FunctionPass *createScalarEvolutionAliasAnalysisPass(); + //===--------------------------------------------------------------------===// // // createAndersensPass - This pass implements Andersen's interprocedural alias diff --git a/llvm/include/llvm/LinkAllPasses.h b/llvm/include/llvm/LinkAllPasses.h index 52bca9db6c33..347dba44477c 100644 --- a/llvm/include/llvm/LinkAllPasses.h +++ b/llvm/include/llvm/LinkAllPasses.h @@ -51,6 +51,7 @@ namespace { (void) llvm::createStructRetPromotionPass(); (void) llvm::createBasicAliasAnalysisPass(); (void) llvm::createLibCallAliasAnalysisPass(0); + (void) llvm::createScalarEvolutionAliasAnalysisPass(); (void) llvm::createBlockPlacementPass(); (void) llvm::createBlockProfilerPass(); (void) llvm::createBreakCriticalEdgesPass(); diff --git a/llvm/lib/Analysis/ScalarEvolutionAliasAnalysis.cpp b/llvm/lib/Analysis/ScalarEvolutionAliasAnalysis.cpp new file mode 100644 index 000000000000..381ea204864c --- /dev/null +++ b/llvm/lib/Analysis/ScalarEvolutionAliasAnalysis.cpp @@ -0,0 +1,128 @@ +//===- ScalarEvolutionAliasAnalysis.cpp - SCEV-based Alias Analysis -------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file defines the ScalarEvolutionAliasAnalysis pass, which implements a +// simple alias analysis implemented in terms of ScalarEvolution queries. +// +// ScalarEvolution has a more complete understanding of pointer arithmetic +// than BasicAliasAnalysis' collection of ad-hoc analyses. +// +//===----------------------------------------------------------------------===// + +#include "llvm/Analysis/AliasAnalysis.h" +#include "llvm/Analysis/ScalarEvolutionExpressions.h" +#include "llvm/Analysis/Passes.h" +#include "llvm/Pass.h" +#include "llvm/Support/Compiler.h" +using namespace llvm; + +namespace { + /// ScalarEvolutionAliasAnalysis - This is a simple alias analysis + /// implementation that uses ScalarEvolution to answer queries. + class VISIBILITY_HIDDEN ScalarEvolutionAliasAnalysis : public FunctionPass, + public AliasAnalysis { + ScalarEvolution *SE; + + public: + static char ID; // Class identification, replacement for typeinfo + ScalarEvolutionAliasAnalysis() : FunctionPass(&ID), SE(0) {} + + private: + virtual void getAnalysisUsage(AnalysisUsage &AU) const; + virtual bool runOnFunction(Function &F); + virtual AliasResult alias(const Value *V1, unsigned V1Size, + const Value *V2, unsigned V2Size); + + Value *GetUnderlyingIdentifiedObject(const SCEV *S); + }; +} // End of anonymous namespace + +// Register this pass... +char ScalarEvolutionAliasAnalysis::ID = 0; +static RegisterPass +X("scev-aa", "ScalarEvolution-based Alias Analysis", false, true); + +// Declare that we implement the AliasAnalysis interface +static RegisterAnalysisGroup Y(X); + +FunctionPass *llvm::createScalarEvolutionAliasAnalysisPass() { + return new ScalarEvolutionAliasAnalysis(); +} + +void +ScalarEvolutionAliasAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { + AU.addRequiredTransitive(); + AU.setPreservesAll(); + AliasAnalysis::getAnalysisUsage(AU); +} + +bool +ScalarEvolutionAliasAnalysis::runOnFunction(Function &F) { + InitializeAliasAnalysis(this); + SE = &getAnalysis(); + return false; +} + +Value * +ScalarEvolutionAliasAnalysis::GetUnderlyingIdentifiedObject(const SCEV *S) { + if (const SCEVAddRecExpr *AR = dyn_cast(S)) { + return GetUnderlyingIdentifiedObject(AR->getStart()); + } else if (const SCEVAddExpr *A = dyn_cast(S)) { + // If there's a pointer operand, it'll be sorted at the end of the list. + const SCEV *Last = A->getOperand(A->getNumOperands()-1); + if (isa(Last->getType())) + return GetUnderlyingIdentifiedObject(Last); + } else if (const SCEVUnknown *U = dyn_cast(S)) { + // Determine if we've found an Identified object. + Value *V = U->getValue(); + if (isIdentifiedObject(V)) + return V; + } + // No Identified object found. + return 0; +} + +AliasAnalysis::AliasResult +ScalarEvolutionAliasAnalysis::alias(const Value *A, unsigned ASize, + const Value *B, unsigned BSize) { + // This is ScalarEvolutionAliasAnalysis. Get the SCEVs! + const SCEV *AS = SE->getSCEV(const_cast(A)); + const SCEV *BS = SE->getSCEV(const_cast(B)); + + // If they evaluate to the same expression, it's a MustAlias. + if (AS == BS) return MustAlias; + + // If something is known about the difference between the two addresses, + // see if it's enough to prove a NoAlias. + if (SE->getEffectiveSCEVType(AS->getType()) == + SE->getEffectiveSCEVType(BS->getType())) { + unsigned BitWidth = SE->getTypeSizeInBits(AS->getType()); + APInt AI(BitWidth, ASize); + const SCEV *BA = SE->getMinusSCEV(BS, AS); + if (AI.ule(SE->getUnsignedRange(BA).getUnsignedMin())) { + APInt BI(BitWidth, BSize); + const SCEV *AB = SE->getMinusSCEV(AS, BS); + if (BI.ule(SE->getUnsignedRange(AB).getUnsignedMin())) + return NoAlias; + } + } + + // If ScalarEvolution can find an underlying object, form a new query. + // The correctness of this depends on ScalarEvolution not recognizing + // inttoptr and ptrtoint operators. + Value *AO = GetUnderlyingIdentifiedObject(AS); + Value *BO = GetUnderlyingIdentifiedObject(BS); + if ((AO && AO != A) || (BO && BO != B)) + if (alias(AO ? AO : A, AO ? ~0u : ASize, + BO ? BO : B, BO ? ~0u : BSize) == NoAlias) + return NoAlias; + + // Forward the query to the next analysis. + return AliasAnalysis::alias(A, ASize, B, BSize); +} diff --git a/llvm/test/Analysis/ScalarEvolution/scev-aa.ll b/llvm/test/Analysis/ScalarEvolution/scev-aa.ll new file mode 100644 index 000000000000..2208b3db4d75 --- /dev/null +++ b/llvm/test/Analysis/ScalarEvolution/scev-aa.ll @@ -0,0 +1,194 @@ +; RUN: llvm-as < %s | opt -scev-aa -aa-eval -print-all-alias-modref-info \ +; RUN: |& FileCheck %s + +; At the time of this writing, all of these CHECK lines are cases that +; plain -basicaa misses. + +target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64" + +; p[i] and p[i+1] don't alias. + +; CHECK: Function: loop: 3 pointers, 0 call sites +; CHECK: NoAlias: double* %pi, double* %pi.next + +define void @loop(double* nocapture %p, i64 %n) nounwind { +entry: + %j = icmp sgt i64 %n, 0 + br i1 %j, label %bb, label %return + +bb: + %i = phi i64 [ 0, %entry ], [ %i.next, %bb ] + %pi = getelementptr double* %p, i64 %i + %i.next = add i64 %i, 1 + %pi.next = getelementptr double* %p, i64 %i.next + %x = load double* %pi + %y = load double* %pi.next + %z = fmul double %x, %y + store double %z, double* %pi + %exitcond = icmp eq i64 %i.next, %n + br i1 %exitcond, label %return, label %bb + +return: + ret void +} + +; Slightly more involved: p[j][i], p[j][i+1], and p[j+1][i] don't alias. + +; CHECK: Function: nestedloop: 4 pointers, 0 call sites +; CHECK: NoAlias: double* %pi.j, double* %pi.next.j +; CHECK: NoAlias: double* %pi.j, double* %pi.j.next +; CHECK: NoAlias: double* %pi.j.next, double* %pi.next.j + +define void @nestedloop(double* nocapture %p, i64 %m) nounwind { +entry: + %k = icmp sgt i64 %m, 0 + br i1 %k, label %guard, label %return + +guard: + %l = icmp sgt i64 91, 0 + br i1 %l, label %outer.loop, label %return + +outer.loop: + %j = phi i64 [ 0, %guard ], [ %j.next, %outer.latch ] + br label %bb + +bb: + %i = phi i64 [ 0, %outer.loop ], [ %i.next, %bb ] + %i.next = add i64 %i, 1 + + %e = add i64 %i, %j + %pi.j = getelementptr double* %p, i64 %e + %f = add i64 %i.next, %j + %pi.next.j = getelementptr double* %p, i64 %f + %x = load double* %pi.j + %y = load double* %pi.next.j + %z = fmul double %x, %y + store double %z, double* %pi.j + + %o = add i64 %j, 91 + %g = add i64 %i, %o + %pi.j.next = getelementptr double* %p, i64 %g + %a = load double* %pi.j.next + %b = fmul double %x, %a + store double %b, double* %pi.j.next + + %exitcond = icmp eq i64 %i.next, 91 + br i1 %exitcond, label %outer.latch, label %bb + +outer.latch: + %j.next = add i64 %j, 91 + %h = icmp eq i64 %j.next, %m + br i1 %h, label %return, label %outer.loop + +return: + ret void +} + +; Even more involved: same as nestedloop, but with a variable extent. +; When n is 1, p[j+1][i] does alias p[j][i+1], and there's no way to +; prove whether n will be greater than 1, so that relation will always +; by MayAlias. The loop is guarded by a n > 0 test though, so +; p[j+1][i] and p[j][i] can theoretically be determined to be NoAlias, +; however the analysis currently doesn't do that. +; TODO: Make the analysis smarter and turn that MayAlias into a NoAlias. + +; CHECK: Function: nestedloop_more: 4 pointers, 0 call sites +; CHECK: NoAlias: double* %pi.j, double* %pi.next.j +; CHECK: MayAlias: double* %pi.j, double* %pi.j.next + +define void @nestedloop_more(double* nocapture %p, i64 %n, i64 %m) nounwind { +entry: + %k = icmp sgt i64 %m, 0 + br i1 %k, label %guard, label %return + +guard: + %l = icmp sgt i64 %n, 0 + br i1 %l, label %outer.loop, label %return + +outer.loop: + %j = phi i64 [ 0, %guard ], [ %j.next, %outer.latch ] + br label %bb + +bb: + %i = phi i64 [ 0, %outer.loop ], [ %i.next, %bb ] + %i.next = add i64 %i, 1 + + %e = add i64 %i, %j + %pi.j = getelementptr double* %p, i64 %e + %f = add i64 %i.next, %j + %pi.next.j = getelementptr double* %p, i64 %f + %x = load double* %pi.j + %y = load double* %pi.next.j + %z = fmul double %x, %y + store double %z, double* %pi.j + + %o = add i64 %j, %n + %g = add i64 %i, %o + %pi.j.next = getelementptr double* %p, i64 %g + %a = load double* %pi.j.next + %b = fmul double %x, %a + store double %b, double* %pi.j.next + + %exitcond = icmp eq i64 %i.next, %n + br i1 %exitcond, label %outer.latch, label %bb + +outer.latch: + %j.next = add i64 %j, %n + %h = icmp eq i64 %j.next, %m + br i1 %h, label %return, label %outer.loop + +return: + ret void +} + +; ScalarEvolution expands field offsets into constants, which allows it to +; do aggressive analysis. Contrast this with BasicAA, which works by +; recognizing GEP idioms. + +%struct.A = type { %struct.B, i32, i32 } +%struct.B = type { double } + +; CHECK: Function: foo: 7 pointers, 0 call sites +; CHECK: NoAlias: %struct.B* %B, i32* %Z +; CHECK: NoAlias: %struct.B* %B, %struct.B* %C +; CHECK: MustAlias: %struct.B* %C, i32* %Z +; CHECK: NoAlias: %struct.B* %B, i32* %X +; CHECK: MustAlias: i32* %X, i32* %Z +; CHECK: MustAlias: %struct.B* %C, i32* %Y +; CHECK: MustAlias: i32* %X, i32* %Y + +define void @foo() { +entry: + %A = alloca %struct.A + %B = getelementptr %struct.A* %A, i32 0, i32 0 + %Q = bitcast %struct.B* %B to %struct.A* + %Z = getelementptr %struct.A* %Q, i32 0, i32 1 + %C = getelementptr %struct.B* %B, i32 1 + %X = bitcast %struct.B* %C to i32* + %Y = getelementptr %struct.A* %A, i32 0, i32 1 + ret void +} + +; CHECK: Function: bar: 7 pointers, 0 call sites +; CHECK: NoAlias: %struct.B* %N, i32* %P +; CHECK: NoAlias: %struct.B* %N, %struct.B* %R +; CHECK: MustAlias: %struct.B* %R, i32* %P +; CHECK: NoAlias: %struct.B* %N, i32* %W +; CHECK: MustAlias: i32* %P, i32* %W +; CHECK: MustAlias: %struct.B* %R, i32* %V +; CHECK: MustAlias: i32* %V, i32* %W + +define void @bar() { + %M = alloca %struct.A + %N = getelementptr %struct.A* %M, i32 0, i32 0 + %O = bitcast %struct.B* %N to %struct.A* + %P = getelementptr %struct.A* %O, i32 0, i32 1 + %R = getelementptr %struct.B* %N, i32 1 + %W = bitcast %struct.B* %R to i32* + %V = getelementptr %struct.A* %M, i32 0, i32 1 + ret void +} + +; CHECK: 13 no alias responses +; CHECK: 26 may alias responses +; CHECK: 18 must alias responses