[llvm] r198290 - BasicAA: Fix value equality and phi cycles

[Nick Lewycky]

On 01/01/2014 08:24 PM, Nick Lewycky wrote:
> On 01/01/2014 07:31 PM, Arnold Schwaighofer wrote:
>> Author: arnolds
>> Date: Wed Jan  1 21:31:36 2014
>> New Revision: 198290
>>
>
> I don't entirely understand why the fix isn't to speculate less often in
> aliasPHI?
>
>>     // Are we checking for alias of the same value?
>> -  if (V1 == V2) return MustAlias;
>> +  if (isValueEqual(V1, V2)) return MustAlias;
>
> This demands a massive explanation of what on earth is going on. Of
> course %x mustalias %x! (Yes, I've read the PR and understand it what's
> going on, but I'd much rather see code that explicitly dealt with the
> "we're in a loop and speculating on PHIs case" broken out.)

Er, I mean as a comment in the code.

Nick

>
>> +bool BasicAliasAnalysis::isValueEqual(const Value *V, const Value *V2) {
>> +  if (V != V2)
>> +    return false;
>> +
>> +  const Instruction *Inst = dyn_cast<Instruction>(V);
>> +  if (!Inst)
>> +    return true;
>> +
>> +  // Use the dominance if available.
>> +  DT = getAnalysisIfAvailable<DominatorTree>();
>> +  if (DT) {
>> +    if (VisitedPhiBBs.size() > MaxNumPhiBBsValueDominanceCheck)
>> +      return false;
>> +
>> +    // Make sure that the visited phis are dominated by the Value.
>> +    for (SmallPtrSet<const BasicBlock *, 8>::iterator
>> +             PI = VisitedPhiBBs.begin(),
>> +             PE = VisitedPhiBBs.end();
>> +         PI != PE; ++PI)
>> +      if (!DT->dominates(Inst, *PI))
>> +        return false;
>
> I haven't thought really hard about this, but did you want to use
> llvm::isPotentiallyReachable?
>
>> Modified:
>> llvm/trunk/test/Analysis/BasicAA/2006-03-03-BadArraySubscript.ll
>> URL:
>> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Analysis/BasicAA/2006-03-03-BadArraySubscript.ll?rev=198290&r1=198289&r2=198290&view=diff
>>
>> ==============================================================================
>>
>> --- llvm/trunk/test/Analysis/BasicAA/2006-03-03-BadArraySubscript.ll
>> (original)
>> +++ llvm/trunk/test/Analysis/BasicAA/2006-03-03-BadArraySubscript.ll
>> Wed Jan  1 21:31:36 2014
>> @@ -1,4 +1,4 @@
>> -; RUN: opt < %s -basicaa -aa-eval -disable-output 2>&1 | FileCheck %s
>> +; RUN: opt < %s -domtree -basicaa -aa-eval -disable-output 2>&1 |
>> FileCheck %s
>
> Adding -domtree to all these tests makes me very sad.
>
> Nick
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>