Create a ScalarEvolution-based AliasAnalysis implementation.

This is a simple AliasAnalysis implementation which works by making
ScalarEvolution queries. ScalarEvolution has a more complete understanding
of arithmetic than BasicAA's collection of ad-hoc checks, so it handles
some cases that BasicAA misses, for example p[i] and p[i+1] within the
same iteration of a loop.

This is currently experimental. It may be that the main use for this pass
will be to help find cases where BasicAA can be profitably extended, or
to help in the development of the overall AliasAnalysis infrastructure,
however it's also possible that it could grow up to become a directly
useful pass.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80098 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Dan Gohman 2009-08-26 14:53:06 +00:00
parent e2cb912039
commit 2385e0e22c
4 changed files with 330 additions and 0 deletions

View File

@ -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

View File

@ -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();

View File

@ -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<ScalarEvolutionAliasAnalysis>
X("scev-aa", "ScalarEvolution-based Alias Analysis", false, true);
// Declare that we implement the AliasAnalysis interface
static RegisterAnalysisGroup<AliasAnalysis> Y(X);
FunctionPass *llvm::createScalarEvolutionAliasAnalysisPass() {
return new ScalarEvolutionAliasAnalysis();
}
void
ScalarEvolutionAliasAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequiredTransitive<ScalarEvolution>();
AU.setPreservesAll();
AliasAnalysis::getAnalysisUsage(AU);
}
bool
ScalarEvolutionAliasAnalysis::runOnFunction(Function &F) {
InitializeAliasAnalysis(this);
SE = &getAnalysis<ScalarEvolution>();
return false;
}
Value *
ScalarEvolutionAliasAnalysis::GetUnderlyingIdentifiedObject(const SCEV *S) {
if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
return GetUnderlyingIdentifiedObject(AR->getStart());
} else if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(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<PointerType>(Last->getType()))
return GetUnderlyingIdentifiedObject(Last);
} else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(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<Value *>(A));
const SCEV *BS = SE->getSCEV(const_cast<Value *>(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);
}

View File

@ -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